Commit | Line | Data |
---|---|---|
39de8281 AG |
1 | #ifndef _RDS_RDS_H |
2 | #define _RDS_RDS_H | |
3 | ||
4 | #include <net/sock.h> | |
5 | #include <linux/scatterlist.h> | |
6 | #include <linux/highmem.h> | |
7 | #include <rdma/rdma_cm.h> | |
8 | #include <linux/mutex.h> | |
9 | #include <linux/rds.h> | |
10 | ||
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 | ||
8cbd9606 AG |
31 | #ifdef ATOMIC64_INIT |
32 | #define KERNEL_HAS_ATOMIC64 | |
33 | #endif | |
34 | ||
39de8281 AG |
35 | #ifdef DEBUG |
36 | #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args) | |
37 | #else | |
38 | /* sigh, pr_debug() causes unused variable warnings */ | |
39 | static inline void __attribute__ ((format (printf, 1, 2))) | |
40 | rdsdebug(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 | ||
57 | struct 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 | */ | |
72 | enum { | |
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 | ||
84 | struct 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; | |
15133f6e | 100 | unsigned int c_xmit_atomic_sent; |
39de8281 AG |
101 | |
102 | spinlock_t c_lock; /* protect msg queues */ | |
103 | u64 c_next_tx_seq; | |
104 | struct list_head c_send_queue; | |
105 | struct list_head c_retrans; | |
106 | ||
107 | u64 c_next_rx_seq; | |
108 | ||
109 | struct rds_transport *c_trans; | |
110 | void *c_transport_data; | |
111 | ||
112 | atomic_t c_state; | |
113 | unsigned long c_flags; | |
114 | unsigned long c_reconnect_jiffies; | |
115 | struct delayed_work c_send_w; | |
116 | struct delayed_work c_recv_w; | |
117 | struct delayed_work c_conn_w; | |
118 | struct work_struct c_down_w; | |
119 | struct mutex c_cm_lock; /* protect conn state & cm */ | |
120 | ||
121 | struct list_head c_map_item; | |
122 | unsigned long c_map_queued; | |
123 | unsigned long c_map_offset; | |
124 | unsigned long c_map_bytes; | |
125 | ||
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 | |
7b70d033 | 136 | #define RDS_MAX_ADV_CREDIT 255 |
39de8281 AG |
137 | |
138 | /* | |
139 | * Maximum space available for extension headers. | |
140 | */ | |
141 | #define RDS_HEADER_EXT_SPACE 16 | |
142 | ||
143 | struct 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 | |
173 | struct 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 | |
182 | struct 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 | |
192 | struct 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 | ||
199 | struct 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 | ||
21f79afa AG |
210 | struct 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 | ||
21f79afa AG |
233 | static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset) |
234 | { | |
235 | return r_key | (((u64) offset) << 32); | |
236 | } | |
237 | ||
238 | static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie) | |
239 | { | |
240 | return cookie; | |
241 | } | |
242 | ||
243 | static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie) | |
244 | { | |
245 | return cookie >> 32; | |
246 | } | |
247 | ||
15133f6e AG |
248 | /* atomic operation types */ |
249 | #define RDS_ATOMIC_TYPE_CSWP 0 | |
250 | #define RDS_ATOMIC_TYPE_FADD 1 | |
251 | ||
39de8281 AG |
252 | /* |
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 | ||
289 | struct 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; | |
304 | struct rds_sock *m_rs; | |
39de8281 | 305 | rds_rdma_cookie_t m_rdma_cookie; |
e779137a | 306 | struct { |
15133f6e AG |
307 | struct rm_atomic_op { |
308 | int op_type; | |
309 | uint64_t op_swap_add; | |
310 | uint64_t op_compare; | |
311 | ||
312 | u32 op_rkey; | |
313 | u64 op_remote_addr; | |
314 | unsigned int op_notify:1; | |
315 | unsigned int op_recverr:1; | |
316 | unsigned int op_mapped:1; | |
317 | unsigned int op_active:1; | |
15133f6e | 318 | struct scatterlist *op_sg; |
f8b3aaf2 | 319 | struct rds_notifier *op_notifier; |
15133f6e AG |
320 | |
321 | struct rds_mr *op_rdma_mr; | |
322 | } atomic; | |
323 | struct rm_rdma_op { | |
f8b3aaf2 AG |
324 | u32 op_rkey; |
325 | u64 op_remote_addr; | |
326 | unsigned int op_write:1; | |
327 | unsigned int op_fence:1; | |
328 | unsigned int op_notify:1; | |
329 | unsigned int op_recverr:1; | |
330 | unsigned int op_mapped:1; | |
331 | unsigned int op_active:1; | |
332 | unsigned int op_bytes; | |
333 | unsigned int op_nents; | |
334 | unsigned int op_count; | |
335 | struct scatterlist *op_sg; | |
336 | struct rds_notifier *op_notifier; | |
337 | ||
338 | struct rds_mr *op_rdma_mr; | |
e779137a | 339 | } rdma; |
15133f6e | 340 | struct rm_data_op { |
241eef3e | 341 | unsigned int op_active:1; |
e779137a AG |
342 | unsigned int m_nents; |
343 | unsigned int m_count; | |
fc445084 | 344 | struct scatterlist *m_sg; |
e779137a AG |
345 | } data; |
346 | }; | |
fc445084 AG |
347 | unsigned int m_used_sgs; |
348 | unsigned int m_total_sgs; | |
39de8281 AG |
349 | }; |
350 | ||
351 | /* | |
352 | * The RDS notifier is used (optionally) to tell the application about | |
353 | * completed RDMA operations. Rather than keeping the whole rds message | |
354 | * around on the queue, we allocate a small notifier that is put on the | |
355 | * socket's notifier_list. Notifications are delivered to the application | |
356 | * through control messages. | |
357 | */ | |
358 | struct rds_notifier { | |
359 | struct list_head n_list; | |
360 | uint64_t n_user_token; | |
361 | int n_status; | |
362 | }; | |
363 | ||
364 | /** | |
365 | * struct rds_transport - transport specific behavioural hooks | |
366 | * | |
367 | * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send | |
368 | * part of a message. The caller serializes on the send_sem so this | |
369 | * doesn't need to be reentrant for a given conn. The header must be | |
370 | * sent before the data payload. .xmit must be prepared to send a | |
371 | * message with no data payload. .xmit should return the number of | |
372 | * bytes that were sent down the connection, including header bytes. | |
373 | * Returning 0 tells the caller that it doesn't need to perform any | |
374 | * additional work now. This is usually the case when the transport has | |
375 | * filled the sending queue for its connection and will handle | |
376 | * triggering the rds thread to continue the send when space becomes | |
377 | * available. Returning -EAGAIN tells the caller to retry the send | |
378 | * immediately. Returning -ENOMEM tells the caller to retry the send at | |
379 | * some point in the future. | |
380 | * | |
381 | * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once | |
382 | * it returns the connection can not call rds_recv_incoming(). | |
383 | * This will only be called once after conn_connect returns | |
384 | * non-zero success and will The caller serializes this with | |
385 | * the send and connecting paths (xmit_* and conn_*). The | |
386 | * transport is responsible for other serialization, including | |
387 | * rds_recv_incoming(). This is called in process context but | |
388 | * should try hard not to block. | |
389 | * | |
390 | * @xmit_cong_map: This asks the transport to send the local bitmap down the | |
391 | * given connection. XXX get a better story about the bitmap | |
392 | * flag and header. | |
393 | */ | |
394 | ||
335776bd AG |
395 | #define RDS_TRANS_IB 0 |
396 | #define RDS_TRANS_IWARP 1 | |
397 | #define RDS_TRANS_TCP 2 | |
398 | #define RDS_TRANS_COUNT 3 | |
399 | ||
39de8281 AG |
400 | struct rds_transport { |
401 | char t_name[TRANSNAMSIZ]; | |
402 | struct list_head t_item; | |
403 | struct module *t_owner; | |
404 | unsigned int t_prefer_loopback:1; | |
335776bd | 405 | unsigned int t_type; |
39de8281 AG |
406 | |
407 | int (*laddr_check)(__be32 addr); | |
408 | int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp); | |
409 | void (*conn_free)(void *data); | |
410 | int (*conn_connect)(struct rds_connection *conn); | |
411 | void (*conn_shutdown)(struct rds_connection *conn); | |
412 | void (*xmit_prepare)(struct rds_connection *conn); | |
413 | void (*xmit_complete)(struct rds_connection *conn); | |
414 | int (*xmit)(struct rds_connection *conn, struct rds_message *rm, | |
415 | unsigned int hdr_off, unsigned int sg, unsigned int off); | |
416 | int (*xmit_cong_map)(struct rds_connection *conn, | |
417 | struct rds_cong_map *map, unsigned long offset); | |
f8b3aaf2 | 418 | int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op); |
241eef3e | 419 | int (*xmit_atomic)(struct rds_connection *conn, struct rds_message *rm); |
39de8281 AG |
420 | int (*recv)(struct rds_connection *conn); |
421 | int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov, | |
422 | size_t size); | |
39de8281 AG |
423 | void (*inc_free)(struct rds_incoming *inc); |
424 | ||
425 | int (*cm_handle_connect)(struct rdma_cm_id *cm_id, | |
426 | struct rdma_cm_event *event); | |
427 | int (*cm_initiate_connect)(struct rdma_cm_id *cm_id); | |
428 | void (*cm_connect_complete)(struct rds_connection *conn, | |
429 | struct rdma_cm_event *event); | |
430 | ||
431 | unsigned int (*stats_info_copy)(struct rds_info_iterator *iter, | |
432 | unsigned int avail); | |
433 | void (*exit)(void); | |
434 | void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg, | |
435 | struct rds_sock *rs, u32 *key_ret); | |
436 | void (*sync_mr)(void *trans_private, int direction); | |
437 | void (*free_mr)(void *trans_private, int invalidate); | |
438 | void (*flush_mrs)(void); | |
439 | }; | |
440 | ||
441 | struct rds_sock { | |
442 | struct sock rs_sk; | |
443 | ||
444 | u64 rs_user_addr; | |
445 | u64 rs_user_bytes; | |
446 | ||
447 | /* | |
448 | * bound_addr used for both incoming and outgoing, no INADDR_ANY | |
449 | * support. | |
450 | */ | |
451 | struct rb_node rs_bound_node; | |
452 | __be32 rs_bound_addr; | |
453 | __be32 rs_conn_addr; | |
454 | __be16 rs_bound_port; | |
455 | __be16 rs_conn_port; | |
456 | ||
457 | /* | |
458 | * This is only used to communicate the transport between bind and | |
459 | * initiating connections. All other trans use is referenced through | |
460 | * the connection. | |
461 | */ | |
462 | struct rds_transport *rs_transport; | |
463 | ||
464 | /* | |
465 | * rds_sendmsg caches the conn it used the last time around. | |
466 | * This helps avoid costly lookups. | |
467 | */ | |
468 | struct rds_connection *rs_conn; | |
469 | ||
470 | /* flag indicating we were congested or not */ | |
471 | int rs_congested; | |
b98ba52f AG |
472 | /* seen congestion (ENOBUFS) when sending? */ |
473 | int rs_seen_congestion; | |
39de8281 AG |
474 | |
475 | /* rs_lock protects all these adjacent members before the newline */ | |
476 | spinlock_t rs_lock; | |
477 | struct list_head rs_send_queue; | |
478 | u32 rs_snd_bytes; | |
479 | int rs_rcv_bytes; | |
480 | struct list_head rs_notify_queue; /* currently used for failed RDMAs */ | |
481 | ||
482 | /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask | |
483 | * to decide whether the application should be woken up. | |
484 | * If not set, we use rs_cong_track to find out whether a cong map | |
485 | * update arrived. | |
486 | */ | |
487 | uint64_t rs_cong_mask; | |
488 | uint64_t rs_cong_notify; | |
489 | struct list_head rs_cong_list; | |
490 | unsigned long rs_cong_track; | |
491 | ||
492 | /* | |
493 | * rs_recv_lock protects the receive queue, and is | |
494 | * used to serialize with rds_release. | |
495 | */ | |
496 | rwlock_t rs_recv_lock; | |
497 | struct list_head rs_recv_queue; | |
498 | ||
499 | /* just for stats reporting */ | |
500 | struct list_head rs_item; | |
501 | ||
502 | /* these have their own lock */ | |
503 | spinlock_t rs_rdma_lock; | |
504 | struct rb_root rs_rdma_keys; | |
505 | ||
506 | /* Socket options - in case there will be more */ | |
507 | unsigned char rs_recverr, | |
508 | rs_cong_monitor; | |
509 | }; | |
510 | ||
511 | static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk) | |
512 | { | |
513 | return container_of(sk, struct rds_sock, rs_sk); | |
514 | } | |
515 | static inline struct sock *rds_rs_to_sk(struct rds_sock *rs) | |
516 | { | |
517 | return &rs->rs_sk; | |
518 | } | |
519 | ||
520 | /* | |
521 | * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value | |
522 | * to account for overhead. We don't account for overhead, we just apply | |
523 | * the number of payload bytes to the specified value. | |
524 | */ | |
525 | static inline int rds_sk_sndbuf(struct rds_sock *rs) | |
526 | { | |
527 | return rds_rs_to_sk(rs)->sk_sndbuf / 2; | |
528 | } | |
529 | static inline int rds_sk_rcvbuf(struct rds_sock *rs) | |
530 | { | |
531 | return rds_rs_to_sk(rs)->sk_rcvbuf / 2; | |
532 | } | |
533 | ||
534 | struct rds_statistics { | |
535 | uint64_t s_conn_reset; | |
536 | uint64_t s_recv_drop_bad_checksum; | |
537 | uint64_t s_recv_drop_old_seq; | |
538 | uint64_t s_recv_drop_no_sock; | |
539 | uint64_t s_recv_drop_dead_sock; | |
540 | uint64_t s_recv_deliver_raced; | |
541 | uint64_t s_recv_delivered; | |
542 | uint64_t s_recv_queued; | |
543 | uint64_t s_recv_immediate_retry; | |
544 | uint64_t s_recv_delayed_retry; | |
545 | uint64_t s_recv_ack_required; | |
546 | uint64_t s_recv_rdma_bytes; | |
547 | uint64_t s_recv_ping; | |
548 | uint64_t s_send_queue_empty; | |
549 | uint64_t s_send_queue_full; | |
550 | uint64_t s_send_sem_contention; | |
551 | uint64_t s_send_sem_queue_raced; | |
552 | uint64_t s_send_immediate_retry; | |
553 | uint64_t s_send_delayed_retry; | |
554 | uint64_t s_send_drop_acked; | |
555 | uint64_t s_send_ack_required; | |
556 | uint64_t s_send_queued; | |
557 | uint64_t s_send_rdma; | |
558 | uint64_t s_send_rdma_bytes; | |
559 | uint64_t s_send_pong; | |
560 | uint64_t s_page_remainder_hit; | |
561 | uint64_t s_page_remainder_miss; | |
562 | uint64_t s_copy_to_user; | |
563 | uint64_t s_copy_from_user; | |
564 | uint64_t s_cong_update_queued; | |
565 | uint64_t s_cong_update_received; | |
566 | uint64_t s_cong_send_error; | |
567 | uint64_t s_cong_send_blocked; | |
15133f6e AG |
568 | uint64_t s_atomic_cswp; |
569 | uint64_t s_atomic_fadd; | |
39de8281 AG |
570 | }; |
571 | ||
572 | /* af_rds.c */ | |
573 | void rds_sock_addref(struct rds_sock *rs); | |
574 | void rds_sock_put(struct rds_sock *rs); | |
575 | void rds_wake_sk_sleep(struct rds_sock *rs); | |
576 | static inline void __rds_wake_sk_sleep(struct sock *sk) | |
577 | { | |
aa395145 | 578 | wait_queue_head_t *waitq = sk_sleep(sk); |
39de8281 AG |
579 | |
580 | if (!sock_flag(sk, SOCK_DEAD) && waitq) | |
581 | wake_up(waitq); | |
582 | } | |
583 | extern wait_queue_head_t rds_poll_waitq; | |
584 | ||
585 | ||
586 | /* bind.c */ | |
587 | int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len); | |
588 | void rds_remove_bound(struct rds_sock *rs); | |
589 | struct rds_sock *rds_find_bound(__be32 addr, __be16 port); | |
590 | ||
591 | /* cong.c */ | |
592 | int rds_cong_get_maps(struct rds_connection *conn); | |
593 | void rds_cong_add_conn(struct rds_connection *conn); | |
594 | void rds_cong_remove_conn(struct rds_connection *conn); | |
595 | void rds_cong_set_bit(struct rds_cong_map *map, __be16 port); | |
596 | void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port); | |
597 | int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs); | |
598 | void rds_cong_queue_updates(struct rds_cong_map *map); | |
599 | void rds_cong_map_updated(struct rds_cong_map *map, uint64_t); | |
600 | int rds_cong_updated_since(unsigned long *recent); | |
601 | void rds_cong_add_socket(struct rds_sock *); | |
602 | void rds_cong_remove_socket(struct rds_sock *); | |
603 | void rds_cong_exit(void); | |
604 | struct rds_message *rds_cong_update_alloc(struct rds_connection *conn); | |
605 | ||
606 | /* conn.c */ | |
607 | int __init rds_conn_init(void); | |
608 | void rds_conn_exit(void); | |
609 | struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr, | |
610 | struct rds_transport *trans, gfp_t gfp); | |
611 | struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr, | |
612 | struct rds_transport *trans, gfp_t gfp); | |
2dc39357 | 613 | void rds_conn_shutdown(struct rds_connection *conn); |
39de8281 AG |
614 | void rds_conn_destroy(struct rds_connection *conn); |
615 | void rds_conn_reset(struct rds_connection *conn); | |
616 | void rds_conn_drop(struct rds_connection *conn); | |
617 | void rds_for_each_conn_info(struct socket *sock, unsigned int len, | |
618 | struct rds_info_iterator *iter, | |
619 | struct rds_info_lengths *lens, | |
620 | int (*visitor)(struct rds_connection *, void *), | |
621 | size_t item_len); | |
622 | void __rds_conn_error(struct rds_connection *conn, const char *, ...) | |
623 | __attribute__ ((format (printf, 2, 3))); | |
624 | #define rds_conn_error(conn, fmt...) \ | |
625 | __rds_conn_error(conn, KERN_WARNING "RDS: " fmt) | |
626 | ||
627 | static inline int | |
628 | rds_conn_transition(struct rds_connection *conn, int old, int new) | |
629 | { | |
630 | return atomic_cmpxchg(&conn->c_state, old, new) == old; | |
631 | } | |
632 | ||
633 | static inline int | |
634 | rds_conn_state(struct rds_connection *conn) | |
635 | { | |
636 | return atomic_read(&conn->c_state); | |
637 | } | |
638 | ||
639 | static inline int | |
640 | rds_conn_up(struct rds_connection *conn) | |
641 | { | |
642 | return atomic_read(&conn->c_state) == RDS_CONN_UP; | |
643 | } | |
644 | ||
645 | static inline int | |
646 | rds_conn_connecting(struct rds_connection *conn) | |
647 | { | |
648 | return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING; | |
649 | } | |
650 | ||
651 | /* message.c */ | |
652 | struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp); | |
fc445084 AG |
653 | struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents); |
654 | int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov, | |
39de8281 AG |
655 | size_t total_len); |
656 | struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len); | |
657 | void rds_message_populate_header(struct rds_header *hdr, __be16 sport, | |
658 | __be16 dport, u64 seq); | |
659 | int rds_message_add_extension(struct rds_header *hdr, | |
660 | unsigned int type, const void *data, unsigned int len); | |
661 | int rds_message_next_extension(struct rds_header *hdr, | |
662 | unsigned int *pos, void *buf, unsigned int *buflen); | |
663 | int rds_message_add_version_extension(struct rds_header *hdr, unsigned int version); | |
664 | int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *version); | |
665 | int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset); | |
666 | int rds_message_inc_copy_to_user(struct rds_incoming *inc, | |
667 | struct iovec *first_iov, size_t size); | |
39de8281 AG |
668 | void rds_message_inc_free(struct rds_incoming *inc); |
669 | void rds_message_addref(struct rds_message *rm); | |
670 | void rds_message_put(struct rds_message *rm); | |
671 | void rds_message_wait(struct rds_message *rm); | |
672 | void rds_message_unmapped(struct rds_message *rm); | |
673 | ||
674 | static inline void rds_message_make_checksum(struct rds_header *hdr) | |
675 | { | |
676 | hdr->h_csum = 0; | |
677 | hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2); | |
678 | } | |
679 | ||
680 | static inline int rds_message_verify_checksum(const struct rds_header *hdr) | |
681 | { | |
682 | return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0; | |
683 | } | |
684 | ||
685 | ||
686 | /* page.c */ | |
687 | int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes, | |
688 | gfp_t gfp); | |
689 | int rds_page_copy_user(struct page *page, unsigned long offset, | |
690 | void __user *ptr, unsigned long bytes, | |
691 | int to_user); | |
692 | #define rds_page_copy_to_user(page, offset, ptr, bytes) \ | |
693 | rds_page_copy_user(page, offset, ptr, bytes, 1) | |
694 | #define rds_page_copy_from_user(page, offset, ptr, bytes) \ | |
695 | rds_page_copy_user(page, offset, ptr, bytes, 0) | |
696 | void rds_page_exit(void); | |
697 | ||
698 | /* recv.c */ | |
699 | void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn, | |
700 | __be32 saddr); | |
701 | void rds_inc_addref(struct rds_incoming *inc); | |
702 | void rds_inc_put(struct rds_incoming *inc); | |
703 | void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr, | |
704 | struct rds_incoming *inc, gfp_t gfp, enum km_type km); | |
705 | int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg, | |
706 | size_t size, int msg_flags); | |
707 | void rds_clear_recv_queue(struct rds_sock *rs); | |
708 | int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg); | |
709 | void rds_inc_info_copy(struct rds_incoming *inc, | |
710 | struct rds_info_iterator *iter, | |
711 | __be32 saddr, __be32 daddr, int flip); | |
712 | ||
713 | /* send.c */ | |
714 | int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg, | |
715 | size_t payload_len); | |
716 | void rds_send_reset(struct rds_connection *conn); | |
717 | int rds_send_xmit(struct rds_connection *conn); | |
718 | struct sockaddr_in; | |
719 | void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest); | |
720 | typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack); | |
721 | void rds_send_drop_acked(struct rds_connection *conn, u64 ack, | |
722 | is_acked_func is_acked); | |
723 | int rds_send_acked_before(struct rds_connection *conn, u64 seq); | |
724 | void rds_send_remove_from_sock(struct list_head *messages, int status); | |
725 | int rds_send_pong(struct rds_connection *conn, __be16 dport); | |
726 | struct rds_message *rds_send_get_message(struct rds_connection *, | |
f8b3aaf2 | 727 | struct rm_rdma_op *); |
39de8281 AG |
728 | |
729 | /* rdma.c */ | |
730 | void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force); | |
21f79afa AG |
731 | int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen); |
732 | int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen); | |
733 | int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen); | |
734 | void rds_rdma_drop_keys(struct rds_sock *rs); | |
735 | int rds_rdma_extra_size(struct rds_rdma_args *args); | |
736 | int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, | |
737 | struct cmsghdr *cmsg); | |
738 | int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm, | |
739 | struct cmsghdr *cmsg); | |
740 | int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, | |
741 | struct cmsghdr *cmsg); | |
742 | int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm, | |
743 | struct cmsghdr *cmsg); | |
f8b3aaf2 | 744 | void rds_rdma_free_op(struct rm_rdma_op *ro); |
d0ab25a8 | 745 | void rds_atomic_free_op(struct rm_atomic_op *ao); |
15133f6e AG |
746 | void rds_rdma_send_complete(struct rds_message *rm, int wc_status); |
747 | void rds_atomic_send_complete(struct rds_message *rm, int wc_status); | |
748 | int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm, | |
749 | struct cmsghdr *cmsg); | |
21f79afa AG |
750 | |
751 | extern void __rds_put_mr_final(struct rds_mr *mr); | |
752 | static inline void rds_mr_put(struct rds_mr *mr) | |
753 | { | |
754 | if (atomic_dec_and_test(&mr->r_refcount)) | |
755 | __rds_put_mr_final(mr); | |
756 | } | |
39de8281 AG |
757 | |
758 | /* stats.c */ | |
9b8de747 | 759 | DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats); |
39de8281 AG |
760 | #define rds_stats_inc_which(which, member) do { \ |
761 | per_cpu(which, get_cpu()).member++; \ | |
762 | put_cpu(); \ | |
763 | } while (0) | |
764 | #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member) | |
765 | #define rds_stats_add_which(which, member, count) do { \ | |
766 | per_cpu(which, get_cpu()).member += count; \ | |
767 | put_cpu(); \ | |
768 | } while (0) | |
769 | #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count) | |
770 | int __init rds_stats_init(void); | |
771 | void rds_stats_exit(void); | |
772 | void rds_stats_info_copy(struct rds_info_iterator *iter, | |
36cbd3dc JE |
773 | uint64_t *values, const char *const *names, |
774 | size_t nr); | |
39de8281 AG |
775 | |
776 | /* sysctl.c */ | |
777 | int __init rds_sysctl_init(void); | |
778 | void rds_sysctl_exit(void); | |
779 | extern unsigned long rds_sysctl_sndbuf_min; | |
780 | extern unsigned long rds_sysctl_sndbuf_default; | |
781 | extern unsigned long rds_sysctl_sndbuf_max; | |
782 | extern unsigned long rds_sysctl_reconnect_min_jiffies; | |
783 | extern unsigned long rds_sysctl_reconnect_max_jiffies; | |
784 | extern unsigned int rds_sysctl_max_unacked_packets; | |
785 | extern unsigned int rds_sysctl_max_unacked_bytes; | |
786 | extern unsigned int rds_sysctl_ping_enable; | |
787 | extern unsigned long rds_sysctl_trace_flags; | |
788 | extern unsigned int rds_sysctl_trace_level; | |
789 | ||
790 | /* threads.c */ | |
791 | int __init rds_threads_init(void); | |
792 | void rds_threads_exit(void); | |
793 | extern struct workqueue_struct *rds_wq; | |
2dc39357 | 794 | void rds_queue_reconnect(struct rds_connection *conn); |
39de8281 AG |
795 | void rds_connect_worker(struct work_struct *); |
796 | void rds_shutdown_worker(struct work_struct *); | |
797 | void rds_send_worker(struct work_struct *); | |
798 | void rds_recv_worker(struct work_struct *); | |
799 | void rds_connect_complete(struct rds_connection *conn); | |
800 | ||
801 | /* transport.c */ | |
802 | int rds_trans_register(struct rds_transport *trans); | |
803 | void rds_trans_unregister(struct rds_transport *trans); | |
804 | struct rds_transport *rds_trans_get_preferred(__be32 addr); | |
805 | unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter, | |
806 | unsigned int avail); | |
807 | int __init rds_trans_init(void); | |
808 | void rds_trans_exit(void); | |
809 | ||
810 | #endif |