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> | |
7b565434 | 10 | #include <linux/rhashtable.h> |
39de8281 AG |
11 | |
12 | #include "info.h" | |
13 | ||
14 | /* | |
15 | * RDS Network protocol version | |
16 | */ | |
17 | #define RDS_PROTOCOL_3_0 0x0300 | |
18 | #define RDS_PROTOCOL_3_1 0x0301 | |
19 | #define RDS_PROTOCOL_VERSION RDS_PROTOCOL_3_1 | |
20 | #define RDS_PROTOCOL_MAJOR(v) ((v) >> 8) | |
21 | #define RDS_PROTOCOL_MINOR(v) ((v) & 255) | |
22 | #define RDS_PROTOCOL(maj, min) (((maj) << 8) | min) | |
23 | ||
24 | /* | |
25 | * XXX randomly chosen, but at least seems to be unused: | |
26 | * # 18464-18768 Unassigned | |
27 | * We should do better. We want a reserved port to discourage unpriv'ed | |
28 | * userspace from listening. | |
29 | */ | |
30 | #define RDS_PORT 18634 | |
31 | ||
8cbd9606 AG |
32 | #ifdef ATOMIC64_INIT |
33 | #define KERNEL_HAS_ATOMIC64 | |
34 | #endif | |
35 | ||
39de8281 AG |
36 | #ifdef DEBUG |
37 | #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args) | |
38 | #else | |
39 | /* sigh, pr_debug() causes unused variable warnings */ | |
b9075fa9 JP |
40 | static inline __printf(1, 2) |
41 | void rdsdebug(char *fmt, ...) | |
39de8281 AG |
42 | { |
43 | } | |
44 | #endif | |
45 | ||
46 | /* XXX is there one of these somewhere? */ | |
47 | #define ceil(x, y) \ | |
48 | ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; }) | |
49 | ||
50 | #define RDS_FRAG_SHIFT 12 | |
51 | #define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT)) | |
52 | ||
53 | #define RDS_CONG_MAP_BYTES (65536 / 8) | |
39de8281 AG |
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, | |
9c79440e | 77 | RDS_CONN_RESETTING, |
39de8281 AG |
78 | RDS_CONN_ERROR, |
79 | }; | |
80 | ||
81 | /* Bits for c_flags */ | |
82 | #define RDS_LL_SEND_FULL 0 | |
83 | #define RDS_RECONNECT_PENDING 1 | |
0f4b1c7e | 84 | #define RDS_IN_XMIT 2 |
73ce4317 | 85 | #define RDS_RECV_REFILL 3 |
39de8281 AG |
86 | |
87 | struct rds_connection { | |
88 | struct hlist_node c_hash_node; | |
89 | __be32 c_laddr; | |
90 | __be32 c_faddr; | |
3b20fc38 SV |
91 | unsigned int c_loopback:1, |
92 | c_outgoing:1, | |
93 | c_pad_to_32:30; | |
39de8281 AG |
94 | struct rds_connection *c_passive; |
95 | ||
96 | struct rds_cong_map *c_lcong; | |
97 | struct rds_cong_map *c_fcong; | |
98 | ||
39de8281 AG |
99 | struct rds_message *c_xmit_rm; |
100 | unsigned long c_xmit_sg; | |
101 | unsigned int c_xmit_hdr_off; | |
102 | unsigned int c_xmit_data_off; | |
15133f6e | 103 | unsigned int c_xmit_atomic_sent; |
5b2366bd AG |
104 | unsigned int c_xmit_rdma_sent; |
105 | unsigned int c_xmit_data_sent; | |
39de8281 AG |
106 | |
107 | spinlock_t c_lock; /* protect msg queues */ | |
108 | u64 c_next_tx_seq; | |
109 | struct list_head c_send_queue; | |
110 | struct list_head c_retrans; | |
111 | ||
112 | u64 c_next_rx_seq; | |
113 | ||
114 | struct rds_transport *c_trans; | |
115 | void *c_transport_data; | |
116 | ||
117 | atomic_t c_state; | |
443be0e5 | 118 | unsigned long c_send_gen; |
39de8281 AG |
119 | unsigned long c_flags; |
120 | unsigned long c_reconnect_jiffies; | |
121 | struct delayed_work c_send_w; | |
122 | struct delayed_work c_recv_w; | |
123 | struct delayed_work c_conn_w; | |
124 | struct work_struct c_down_w; | |
125 | struct mutex c_cm_lock; /* protect conn state & cm */ | |
0f4b1c7e | 126 | wait_queue_head_t c_waitq; |
39de8281 AG |
127 | |
128 | struct list_head c_map_item; | |
129 | unsigned long c_map_queued; | |
39de8281 AG |
130 | |
131 | unsigned int c_unacked_packets; | |
132 | unsigned int c_unacked_bytes; | |
133 | ||
134 | /* Protocol version */ | |
135 | unsigned int c_version; | |
d5a8ac28 | 136 | possible_net_t c_net; |
39de8281 AG |
137 | }; |
138 | ||
d5a8ac28 SV |
139 | static inline |
140 | struct net *rds_conn_net(struct rds_connection *conn) | |
141 | { | |
142 | return read_pnet(&conn->c_net); | |
143 | } | |
144 | ||
145 | static inline | |
146 | void rds_conn_net_set(struct rds_connection *conn, struct net *net) | |
147 | { | |
148 | write_pnet(&conn->c_net, net); | |
149 | } | |
150 | ||
39de8281 AG |
151 | #define RDS_FLAG_CONG_BITMAP 0x01 |
152 | #define RDS_FLAG_ACK_REQUIRED 0x02 | |
153 | #define RDS_FLAG_RETRANSMITTED 0x04 | |
7b70d033 | 154 | #define RDS_MAX_ADV_CREDIT 255 |
39de8281 AG |
155 | |
156 | /* | |
157 | * Maximum space available for extension headers. | |
158 | */ | |
159 | #define RDS_HEADER_EXT_SPACE 16 | |
160 | ||
161 | struct rds_header { | |
162 | __be64 h_sequence; | |
163 | __be64 h_ack; | |
164 | __be32 h_len; | |
165 | __be16 h_sport; | |
166 | __be16 h_dport; | |
167 | u8 h_flags; | |
168 | u8 h_credit; | |
169 | u8 h_padding[4]; | |
170 | __sum16 h_csum; | |
171 | ||
172 | u8 h_exthdr[RDS_HEADER_EXT_SPACE]; | |
173 | }; | |
174 | ||
175 | /* | |
176 | * Reserved - indicates end of extensions | |
177 | */ | |
178 | #define RDS_EXTHDR_NONE 0 | |
179 | ||
180 | /* | |
181 | * This extension header is included in the very | |
182 | * first message that is sent on a new connection, | |
183 | * and identifies the protocol level. This will help | |
184 | * rolling updates if a future change requires breaking | |
185 | * the protocol. | |
186 | * NB: This is no longer true for IB, where we do a version | |
187 | * negotiation during the connection setup phase (protocol | |
188 | * version information is included in the RDMA CM private data). | |
189 | */ | |
190 | #define RDS_EXTHDR_VERSION 1 | |
191 | struct rds_ext_header_version { | |
192 | __be32 h_version; | |
193 | }; | |
194 | ||
195 | /* | |
196 | * This extension header is included in the RDS message | |
197 | * chasing an RDMA operation. | |
198 | */ | |
199 | #define RDS_EXTHDR_RDMA 2 | |
200 | struct rds_ext_header_rdma { | |
201 | __be32 h_rdma_rkey; | |
202 | }; | |
203 | ||
204 | /* | |
205 | * This extension header tells the peer about the | |
206 | * destination <R_Key,offset> of the requested RDMA | |
207 | * operation. | |
208 | */ | |
209 | #define RDS_EXTHDR_RDMA_DEST 3 | |
210 | struct rds_ext_header_rdma_dest { | |
211 | __be32 h_rdma_rkey; | |
212 | __be32 h_rdma_offset; | |
213 | }; | |
214 | ||
215 | #define __RDS_EXTHDR_MAX 16 /* for now */ | |
216 | ||
217 | struct rds_incoming { | |
218 | atomic_t i_refcount; | |
219 | struct list_head i_item; | |
220 | struct rds_connection *i_conn; | |
221 | struct rds_header i_hdr; | |
222 | unsigned long i_rx_jiffies; | |
223 | __be32 i_saddr; | |
224 | ||
225 | rds_rdma_cookie_t i_rdma_cookie; | |
5711f8b3 | 226 | struct timeval i_rx_tstamp; |
39de8281 AG |
227 | }; |
228 | ||
21f79afa AG |
229 | struct rds_mr { |
230 | struct rb_node r_rb_node; | |
231 | atomic_t r_refcount; | |
232 | u32 r_key; | |
233 | ||
234 | /* A copy of the creation flags */ | |
235 | unsigned int r_use_once:1; | |
236 | unsigned int r_invalidate:1; | |
237 | unsigned int r_write:1; | |
238 | ||
239 | /* This is for RDS_MR_DEAD. | |
240 | * It would be nice & consistent to make this part of the above | |
241 | * bit field here, but we need to use test_and_set_bit. | |
242 | */ | |
243 | unsigned long r_state; | |
244 | struct rds_sock *r_sock; /* back pointer to the socket that owns us */ | |
245 | struct rds_transport *r_trans; | |
246 | void *r_trans_private; | |
247 | }; | |
248 | ||
249 | /* Flags for mr->r_state */ | |
250 | #define RDS_MR_DEAD 0 | |
251 | ||
21f79afa AG |
252 | static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset) |
253 | { | |
254 | return r_key | (((u64) offset) << 32); | |
255 | } | |
256 | ||
257 | static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie) | |
258 | { | |
259 | return cookie; | |
260 | } | |
261 | ||
262 | static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie) | |
263 | { | |
264 | return cookie >> 32; | |
265 | } | |
266 | ||
15133f6e AG |
267 | /* atomic operation types */ |
268 | #define RDS_ATOMIC_TYPE_CSWP 0 | |
269 | #define RDS_ATOMIC_TYPE_FADD 1 | |
270 | ||
39de8281 AG |
271 | /* |
272 | * m_sock_item and m_conn_item are on lists that are serialized under | |
273 | * conn->c_lock. m_sock_item has additional meaning in that once it is empty | |
274 | * the message will not be put back on the retransmit list after being sent. | |
275 | * messages that are canceled while being sent rely on this. | |
276 | * | |
277 | * m_inc is used by loopback so that it can pass an incoming message straight | |
278 | * back up into the rx path. It embeds a wire header which is also used by | |
279 | * the send path, which is kind of awkward. | |
280 | * | |
281 | * m_sock_item indicates the message's presence on a socket's send or receive | |
282 | * queue. m_rs will point to that socket. | |
283 | * | |
284 | * m_daddr is used by cancellation to prune messages to a given destination. | |
285 | * | |
286 | * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock | |
287 | * nesting. As paths iterate over messages on a sock, or conn, they must | |
288 | * also lock the conn, or sock, to remove the message from those lists too. | |
289 | * Testing the flag to determine if the message is still on the lists lets | |
290 | * us avoid testing the list_head directly. That means each path can use | |
291 | * the message's list_head to keep it on a local list while juggling locks | |
292 | * without confusing the other path. | |
293 | * | |
294 | * m_ack_seq is an optional field set by transports who need a different | |
295 | * sequence number range to invalidate. They can use this in a callback | |
296 | * that they pass to rds_send_drop_acked() to see if each message has been | |
297 | * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't | |
298 | * had ack_seq set yet. | |
299 | */ | |
300 | #define RDS_MSG_ON_SOCK 1 | |
301 | #define RDS_MSG_ON_CONN 2 | |
302 | #define RDS_MSG_HAS_ACK_SEQ 3 | |
303 | #define RDS_MSG_ACK_REQUIRED 4 | |
304 | #define RDS_MSG_RETRANSMITTED 5 | |
305 | #define RDS_MSG_MAPPED 6 | |
306 | #define RDS_MSG_PAGEVEC 7 | |
307 | ||
308 | struct rds_message { | |
309 | atomic_t m_refcount; | |
310 | struct list_head m_sock_item; | |
311 | struct list_head m_conn_item; | |
312 | struct rds_incoming m_inc; | |
313 | u64 m_ack_seq; | |
314 | __be32 m_daddr; | |
315 | unsigned long m_flags; | |
316 | ||
317 | /* Never access m_rs without holding m_rs_lock. | |
318 | * Lock nesting is | |
319 | * rm->m_rs_lock | |
320 | * -> rs->rs_lock | |
321 | */ | |
322 | spinlock_t m_rs_lock; | |
c83188dc CM |
323 | wait_queue_head_t m_flush_wait; |
324 | ||
39de8281 | 325 | struct rds_sock *m_rs; |
7e3bd65e AG |
326 | |
327 | /* cookie to send to remote, in rds header */ | |
39de8281 | 328 | rds_rdma_cookie_t m_rdma_cookie; |
7e3bd65e AG |
329 | |
330 | unsigned int m_used_sgs; | |
331 | unsigned int m_total_sgs; | |
332 | ||
ff3d7d36 AG |
333 | void *m_final_op; |
334 | ||
e779137a | 335 | struct { |
15133f6e AG |
336 | struct rm_atomic_op { |
337 | int op_type; | |
20c72bd5 AG |
338 | union { |
339 | struct { | |
340 | uint64_t compare; | |
341 | uint64_t swap; | |
342 | uint64_t compare_mask; | |
343 | uint64_t swap_mask; | |
344 | } op_m_cswp; | |
345 | struct { | |
346 | uint64_t add; | |
347 | uint64_t nocarry_mask; | |
348 | } op_m_fadd; | |
349 | }; | |
15133f6e AG |
350 | |
351 | u32 op_rkey; | |
352 | u64 op_remote_addr; | |
353 | unsigned int op_notify:1; | |
354 | unsigned int op_recverr:1; | |
355 | unsigned int op_mapped:1; | |
2c3a5f9a | 356 | unsigned int op_silent:1; |
15133f6e | 357 | unsigned int op_active:1; |
15133f6e | 358 | struct scatterlist *op_sg; |
f8b3aaf2 | 359 | struct rds_notifier *op_notifier; |
15133f6e AG |
360 | |
361 | struct rds_mr *op_rdma_mr; | |
362 | } atomic; | |
363 | struct rm_rdma_op { | |
f8b3aaf2 AG |
364 | u32 op_rkey; |
365 | u64 op_remote_addr; | |
366 | unsigned int op_write:1; | |
367 | unsigned int op_fence:1; | |
368 | unsigned int op_notify:1; | |
369 | unsigned int op_recverr:1; | |
370 | unsigned int op_mapped:1; | |
2c3a5f9a | 371 | unsigned int op_silent:1; |
f8b3aaf2 AG |
372 | unsigned int op_active:1; |
373 | unsigned int op_bytes; | |
374 | unsigned int op_nents; | |
375 | unsigned int op_count; | |
376 | struct scatterlist *op_sg; | |
377 | struct rds_notifier *op_notifier; | |
378 | ||
379 | struct rds_mr *op_rdma_mr; | |
e779137a | 380 | } rdma; |
15133f6e | 381 | struct rm_data_op { |
241eef3e | 382 | unsigned int op_active:1; |
6c7cc6e4 AG |
383 | unsigned int op_nents; |
384 | unsigned int op_count; | |
d655a9fb WW |
385 | unsigned int op_dmasg; |
386 | unsigned int op_dmaoff; | |
6c7cc6e4 | 387 | struct scatterlist *op_sg; |
e779137a AG |
388 | } data; |
389 | }; | |
39de8281 AG |
390 | }; |
391 | ||
392 | /* | |
393 | * The RDS notifier is used (optionally) to tell the application about | |
394 | * completed RDMA operations. Rather than keeping the whole rds message | |
395 | * around on the queue, we allocate a small notifier that is put on the | |
396 | * socket's notifier_list. Notifications are delivered to the application | |
397 | * through control messages. | |
398 | */ | |
399 | struct rds_notifier { | |
400 | struct list_head n_list; | |
401 | uint64_t n_user_token; | |
402 | int n_status; | |
403 | }; | |
404 | ||
405 | /** | |
406 | * struct rds_transport - transport specific behavioural hooks | |
407 | * | |
408 | * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send | |
409 | * part of a message. The caller serializes on the send_sem so this | |
410 | * doesn't need to be reentrant for a given conn. The header must be | |
411 | * sent before the data payload. .xmit must be prepared to send a | |
412 | * message with no data payload. .xmit should return the number of | |
413 | * bytes that were sent down the connection, including header bytes. | |
414 | * Returning 0 tells the caller that it doesn't need to perform any | |
415 | * additional work now. This is usually the case when the transport has | |
416 | * filled the sending queue for its connection and will handle | |
417 | * triggering the rds thread to continue the send when space becomes | |
418 | * available. Returning -EAGAIN tells the caller to retry the send | |
419 | * immediately. Returning -ENOMEM tells the caller to retry the send at | |
420 | * some point in the future. | |
421 | * | |
422 | * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once | |
423 | * it returns the connection can not call rds_recv_incoming(). | |
424 | * This will only be called once after conn_connect returns | |
425 | * non-zero success and will The caller serializes this with | |
426 | * the send and connecting paths (xmit_* and conn_*). The | |
427 | * transport is responsible for other serialization, including | |
428 | * rds_recv_incoming(). This is called in process context but | |
429 | * should try hard not to block. | |
39de8281 AG |
430 | */ |
431 | ||
432 | struct rds_transport { | |
433 | char t_name[TRANSNAMSIZ]; | |
434 | struct list_head t_item; | |
435 | struct module *t_owner; | |
436 | unsigned int t_prefer_loopback:1; | |
335776bd | 437 | unsigned int t_type; |
39de8281 | 438 | |
d5a8ac28 | 439 | int (*laddr_check)(struct net *net, __be32 addr); |
39de8281 AG |
440 | int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp); |
441 | void (*conn_free)(void *data); | |
442 | int (*conn_connect)(struct rds_connection *conn); | |
443 | void (*conn_shutdown)(struct rds_connection *conn); | |
444 | void (*xmit_prepare)(struct rds_connection *conn); | |
445 | void (*xmit_complete)(struct rds_connection *conn); | |
446 | int (*xmit)(struct rds_connection *conn, struct rds_message *rm, | |
447 | unsigned int hdr_off, unsigned int sg, unsigned int off); | |
f8b3aaf2 | 448 | int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op); |
ff3d7d36 | 449 | int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op); |
39de8281 | 450 | int (*recv)(struct rds_connection *conn); |
c310e72c | 451 | int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to); |
39de8281 AG |
452 | void (*inc_free)(struct rds_incoming *inc); |
453 | ||
454 | int (*cm_handle_connect)(struct rdma_cm_id *cm_id, | |
455 | struct rdma_cm_event *event); | |
456 | int (*cm_initiate_connect)(struct rdma_cm_id *cm_id); | |
457 | void (*cm_connect_complete)(struct rds_connection *conn, | |
458 | struct rdma_cm_event *event); | |
459 | ||
460 | unsigned int (*stats_info_copy)(struct rds_info_iterator *iter, | |
461 | unsigned int avail); | |
462 | void (*exit)(void); | |
463 | void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg, | |
464 | struct rds_sock *rs, u32 *key_ret); | |
465 | void (*sync_mr)(void *trans_private, int direction); | |
466 | void (*free_mr)(void *trans_private, int invalidate); | |
467 | void (*flush_mrs)(void); | |
468 | }; | |
469 | ||
470 | struct rds_sock { | |
471 | struct sock rs_sk; | |
472 | ||
473 | u64 rs_user_addr; | |
474 | u64 rs_user_bytes; | |
475 | ||
476 | /* | |
477 | * bound_addr used for both incoming and outgoing, no INADDR_ANY | |
478 | * support. | |
479 | */ | |
7b565434 | 480 | struct rhash_head rs_bound_node; |
481 | u64 rs_bound_key; | |
39de8281 AG |
482 | __be32 rs_bound_addr; |
483 | __be32 rs_conn_addr; | |
484 | __be16 rs_bound_port; | |
485 | __be16 rs_conn_port; | |
39de8281 AG |
486 | struct rds_transport *rs_transport; |
487 | ||
488 | /* | |
489 | * rds_sendmsg caches the conn it used the last time around. | |
490 | * This helps avoid costly lookups. | |
491 | */ | |
492 | struct rds_connection *rs_conn; | |
493 | ||
494 | /* flag indicating we were congested or not */ | |
495 | int rs_congested; | |
b98ba52f AG |
496 | /* seen congestion (ENOBUFS) when sending? */ |
497 | int rs_seen_congestion; | |
39de8281 AG |
498 | |
499 | /* rs_lock protects all these adjacent members before the newline */ | |
500 | spinlock_t rs_lock; | |
501 | struct list_head rs_send_queue; | |
502 | u32 rs_snd_bytes; | |
503 | int rs_rcv_bytes; | |
504 | struct list_head rs_notify_queue; /* currently used for failed RDMAs */ | |
505 | ||
506 | /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask | |
507 | * to decide whether the application should be woken up. | |
508 | * If not set, we use rs_cong_track to find out whether a cong map | |
509 | * update arrived. | |
510 | */ | |
511 | uint64_t rs_cong_mask; | |
512 | uint64_t rs_cong_notify; | |
513 | struct list_head rs_cong_list; | |
514 | unsigned long rs_cong_track; | |
515 | ||
516 | /* | |
517 | * rs_recv_lock protects the receive queue, and is | |
518 | * used to serialize with rds_release. | |
519 | */ | |
520 | rwlock_t rs_recv_lock; | |
521 | struct list_head rs_recv_queue; | |
522 | ||
523 | /* just for stats reporting */ | |
524 | struct list_head rs_item; | |
525 | ||
526 | /* these have their own lock */ | |
527 | spinlock_t rs_rdma_lock; | |
528 | struct rb_root rs_rdma_keys; | |
529 | ||
530 | /* Socket options - in case there will be more */ | |
531 | unsigned char rs_recverr, | |
532 | rs_cong_monitor; | |
533 | }; | |
534 | ||
535 | static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk) | |
536 | { | |
537 | return container_of(sk, struct rds_sock, rs_sk); | |
538 | } | |
539 | static inline struct sock *rds_rs_to_sk(struct rds_sock *rs) | |
540 | { | |
541 | return &rs->rs_sk; | |
542 | } | |
543 | ||
544 | /* | |
545 | * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value | |
546 | * to account for overhead. We don't account for overhead, we just apply | |
547 | * the number of payload bytes to the specified value. | |
548 | */ | |
549 | static inline int rds_sk_sndbuf(struct rds_sock *rs) | |
550 | { | |
551 | return rds_rs_to_sk(rs)->sk_sndbuf / 2; | |
552 | } | |
553 | static inline int rds_sk_rcvbuf(struct rds_sock *rs) | |
554 | { | |
555 | return rds_rs_to_sk(rs)->sk_rcvbuf / 2; | |
556 | } | |
557 | ||
558 | struct rds_statistics { | |
559 | uint64_t s_conn_reset; | |
560 | uint64_t s_recv_drop_bad_checksum; | |
561 | uint64_t s_recv_drop_old_seq; | |
562 | uint64_t s_recv_drop_no_sock; | |
563 | uint64_t s_recv_drop_dead_sock; | |
564 | uint64_t s_recv_deliver_raced; | |
565 | uint64_t s_recv_delivered; | |
566 | uint64_t s_recv_queued; | |
567 | uint64_t s_recv_immediate_retry; | |
568 | uint64_t s_recv_delayed_retry; | |
569 | uint64_t s_recv_ack_required; | |
570 | uint64_t s_recv_rdma_bytes; | |
571 | uint64_t s_recv_ping; | |
572 | uint64_t s_send_queue_empty; | |
573 | uint64_t s_send_queue_full; | |
049ee3f5 AG |
574 | uint64_t s_send_lock_contention; |
575 | uint64_t s_send_lock_queue_raced; | |
39de8281 AG |
576 | uint64_t s_send_immediate_retry; |
577 | uint64_t s_send_delayed_retry; | |
578 | uint64_t s_send_drop_acked; | |
579 | uint64_t s_send_ack_required; | |
580 | uint64_t s_send_queued; | |
581 | uint64_t s_send_rdma; | |
582 | uint64_t s_send_rdma_bytes; | |
583 | uint64_t s_send_pong; | |
584 | uint64_t s_page_remainder_hit; | |
585 | uint64_t s_page_remainder_miss; | |
586 | uint64_t s_copy_to_user; | |
587 | uint64_t s_copy_from_user; | |
588 | uint64_t s_cong_update_queued; | |
589 | uint64_t s_cong_update_received; | |
590 | uint64_t s_cong_send_error; | |
591 | uint64_t s_cong_send_blocked; | |
592 | }; | |
593 | ||
594 | /* af_rds.c */ | |
595 | void rds_sock_addref(struct rds_sock *rs); | |
596 | void rds_sock_put(struct rds_sock *rs); | |
597 | void rds_wake_sk_sleep(struct rds_sock *rs); | |
598 | static inline void __rds_wake_sk_sleep(struct sock *sk) | |
599 | { | |
aa395145 | 600 | wait_queue_head_t *waitq = sk_sleep(sk); |
39de8281 AG |
601 | |
602 | if (!sock_flag(sk, SOCK_DEAD) && waitq) | |
603 | wake_up(waitq); | |
604 | } | |
605 | extern wait_queue_head_t rds_poll_waitq; | |
606 | ||
607 | ||
608 | /* bind.c */ | |
609 | int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len); | |
610 | void rds_remove_bound(struct rds_sock *rs); | |
611 | struct rds_sock *rds_find_bound(__be32 addr, __be16 port); | |
7b565434 | 612 | int rds_bind_lock_init(void); |
613 | void rds_bind_lock_destroy(void); | |
39de8281 AG |
614 | |
615 | /* cong.c */ | |
616 | int rds_cong_get_maps(struct rds_connection *conn); | |
617 | void rds_cong_add_conn(struct rds_connection *conn); | |
618 | void rds_cong_remove_conn(struct rds_connection *conn); | |
619 | void rds_cong_set_bit(struct rds_cong_map *map, __be16 port); | |
620 | void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port); | |
621 | int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs); | |
622 | void rds_cong_queue_updates(struct rds_cong_map *map); | |
623 | void rds_cong_map_updated(struct rds_cong_map *map, uint64_t); | |
624 | int rds_cong_updated_since(unsigned long *recent); | |
625 | void rds_cong_add_socket(struct rds_sock *); | |
626 | void rds_cong_remove_socket(struct rds_sock *); | |
627 | void rds_cong_exit(void); | |
628 | struct rds_message *rds_cong_update_alloc(struct rds_connection *conn); | |
629 | ||
630 | /* conn.c */ | |
ef87b7ea | 631 | int rds_conn_init(void); |
39de8281 | 632 | void rds_conn_exit(void); |
d5a8ac28 SV |
633 | struct rds_connection *rds_conn_create(struct net *net, |
634 | __be32 laddr, __be32 faddr, | |
39de8281 | 635 | struct rds_transport *trans, gfp_t gfp); |
d5a8ac28 SV |
636 | struct rds_connection *rds_conn_create_outgoing(struct net *net, |
637 | __be32 laddr, __be32 faddr, | |
39de8281 | 638 | struct rds_transport *trans, gfp_t gfp); |
2dc39357 | 639 | void rds_conn_shutdown(struct rds_connection *conn); |
39de8281 | 640 | void rds_conn_destroy(struct rds_connection *conn); |
39de8281 | 641 | void rds_conn_drop(struct rds_connection *conn); |
f3c6808d | 642 | void rds_conn_connect_if_down(struct rds_connection *conn); |
39de8281 AG |
643 | void rds_for_each_conn_info(struct socket *sock, unsigned int len, |
644 | struct rds_info_iterator *iter, | |
645 | struct rds_info_lengths *lens, | |
646 | int (*visitor)(struct rds_connection *, void *), | |
647 | size_t item_len); | |
b9075fa9 JP |
648 | __printf(2, 3) |
649 | void __rds_conn_error(struct rds_connection *conn, const char *, ...); | |
39de8281 AG |
650 | #define rds_conn_error(conn, fmt...) \ |
651 | __rds_conn_error(conn, KERN_WARNING "RDS: " fmt) | |
652 | ||
653 | static inline int | |
654 | rds_conn_transition(struct rds_connection *conn, int old, int new) | |
655 | { | |
656 | return atomic_cmpxchg(&conn->c_state, old, new) == old; | |
657 | } | |
658 | ||
659 | static inline int | |
660 | rds_conn_state(struct rds_connection *conn) | |
661 | { | |
662 | return atomic_read(&conn->c_state); | |
663 | } | |
664 | ||
665 | static inline int | |
666 | rds_conn_up(struct rds_connection *conn) | |
667 | { | |
668 | return atomic_read(&conn->c_state) == RDS_CONN_UP; | |
669 | } | |
670 | ||
671 | static inline int | |
672 | rds_conn_connecting(struct rds_connection *conn) | |
673 | { | |
674 | return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING; | |
675 | } | |
676 | ||
677 | /* message.c */ | |
678 | struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp); | |
fc445084 | 679 | struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents); |
083735f4 | 680 | int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from); |
39de8281 AG |
681 | struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len); |
682 | void rds_message_populate_header(struct rds_header *hdr, __be16 sport, | |
683 | __be16 dport, u64 seq); | |
684 | int rds_message_add_extension(struct rds_header *hdr, | |
685 | unsigned int type, const void *data, unsigned int len); | |
686 | int rds_message_next_extension(struct rds_header *hdr, | |
687 | unsigned int *pos, void *buf, unsigned int *buflen); | |
39de8281 | 688 | int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset); |
c310e72c | 689 | int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); |
39de8281 AG |
690 | void rds_message_inc_free(struct rds_incoming *inc); |
691 | void rds_message_addref(struct rds_message *rm); | |
692 | void rds_message_put(struct rds_message *rm); | |
693 | void rds_message_wait(struct rds_message *rm); | |
694 | void rds_message_unmapped(struct rds_message *rm); | |
695 | ||
696 | static inline void rds_message_make_checksum(struct rds_header *hdr) | |
697 | { | |
698 | hdr->h_csum = 0; | |
699 | hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2); | |
700 | } | |
701 | ||
702 | static inline int rds_message_verify_checksum(const struct rds_header *hdr) | |
703 | { | |
704 | return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0; | |
705 | } | |
706 | ||
707 | ||
708 | /* page.c */ | |
709 | int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes, | |
710 | gfp_t gfp); | |
711 | int rds_page_copy_user(struct page *page, unsigned long offset, | |
712 | void __user *ptr, unsigned long bytes, | |
713 | int to_user); | |
714 | #define rds_page_copy_to_user(page, offset, ptr, bytes) \ | |
715 | rds_page_copy_user(page, offset, ptr, bytes, 1) | |
716 | #define rds_page_copy_from_user(page, offset, ptr, bytes) \ | |
717 | rds_page_copy_user(page, offset, ptr, bytes, 0) | |
718 | void rds_page_exit(void); | |
719 | ||
720 | /* recv.c */ | |
721 | void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn, | |
722 | __be32 saddr); | |
39de8281 AG |
723 | void rds_inc_put(struct rds_incoming *inc); |
724 | void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr, | |
6114eab5 | 725 | struct rds_incoming *inc, gfp_t gfp); |
1b784140 YX |
726 | int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size, |
727 | int msg_flags); | |
39de8281 AG |
728 | void rds_clear_recv_queue(struct rds_sock *rs); |
729 | int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg); | |
730 | void rds_inc_info_copy(struct rds_incoming *inc, | |
731 | struct rds_info_iterator *iter, | |
732 | __be32 saddr, __be32 daddr, int flip); | |
733 | ||
734 | /* send.c */ | |
1b784140 | 735 | int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len); |
39de8281 AG |
736 | void rds_send_reset(struct rds_connection *conn); |
737 | int rds_send_xmit(struct rds_connection *conn); | |
738 | struct sockaddr_in; | |
739 | void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest); | |
740 | typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack); | |
741 | void rds_send_drop_acked(struct rds_connection *conn, u64 ack, | |
742 | is_acked_func is_acked); | |
39de8281 AG |
743 | int rds_send_pong(struct rds_connection *conn, __be16 dport); |
744 | struct rds_message *rds_send_get_message(struct rds_connection *, | |
f8b3aaf2 | 745 | struct rm_rdma_op *); |
39de8281 AG |
746 | |
747 | /* rdma.c */ | |
748 | void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force); | |
21f79afa AG |
749 | int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen); |
750 | int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen); | |
751 | int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen); | |
752 | void rds_rdma_drop_keys(struct rds_sock *rs); | |
753 | int rds_rdma_extra_size(struct rds_rdma_args *args); | |
754 | int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, | |
755 | struct cmsghdr *cmsg); | |
756 | int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm, | |
757 | struct cmsghdr *cmsg); | |
758 | int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, | |
759 | struct cmsghdr *cmsg); | |
760 | int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm, | |
761 | struct cmsghdr *cmsg); | |
f8b3aaf2 | 762 | void rds_rdma_free_op(struct rm_rdma_op *ro); |
d0ab25a8 | 763 | void rds_atomic_free_op(struct rm_atomic_op *ao); |
15133f6e AG |
764 | void rds_rdma_send_complete(struct rds_message *rm, int wc_status); |
765 | void rds_atomic_send_complete(struct rds_message *rm, int wc_status); | |
766 | int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm, | |
767 | struct cmsghdr *cmsg); | |
21f79afa | 768 | |
c1b1203d | 769 | void __rds_put_mr_final(struct rds_mr *mr); |
21f79afa AG |
770 | static inline void rds_mr_put(struct rds_mr *mr) |
771 | { | |
772 | if (atomic_dec_and_test(&mr->r_refcount)) | |
773 | __rds_put_mr_final(mr); | |
774 | } | |
39de8281 AG |
775 | |
776 | /* stats.c */ | |
9b8de747 | 777 | DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats); |
39de8281 AG |
778 | #define rds_stats_inc_which(which, member) do { \ |
779 | per_cpu(which, get_cpu()).member++; \ | |
780 | put_cpu(); \ | |
781 | } while (0) | |
782 | #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member) | |
783 | #define rds_stats_add_which(which, member, count) do { \ | |
784 | per_cpu(which, get_cpu()).member += count; \ | |
785 | put_cpu(); \ | |
786 | } while (0) | |
787 | #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count) | |
ef87b7ea | 788 | int rds_stats_init(void); |
39de8281 AG |
789 | void rds_stats_exit(void); |
790 | void rds_stats_info_copy(struct rds_info_iterator *iter, | |
36cbd3dc JE |
791 | uint64_t *values, const char *const *names, |
792 | size_t nr); | |
39de8281 AG |
793 | |
794 | /* sysctl.c */ | |
ef87b7ea | 795 | int rds_sysctl_init(void); |
39de8281 AG |
796 | void rds_sysctl_exit(void); |
797 | extern unsigned long rds_sysctl_sndbuf_min; | |
798 | extern unsigned long rds_sysctl_sndbuf_default; | |
799 | extern unsigned long rds_sysctl_sndbuf_max; | |
800 | extern unsigned long rds_sysctl_reconnect_min_jiffies; | |
801 | extern unsigned long rds_sysctl_reconnect_max_jiffies; | |
802 | extern unsigned int rds_sysctl_max_unacked_packets; | |
803 | extern unsigned int rds_sysctl_max_unacked_bytes; | |
804 | extern unsigned int rds_sysctl_ping_enable; | |
805 | extern unsigned long rds_sysctl_trace_flags; | |
806 | extern unsigned int rds_sysctl_trace_level; | |
807 | ||
808 | /* threads.c */ | |
ef87b7ea | 809 | int rds_threads_init(void); |
39de8281 AG |
810 | void rds_threads_exit(void); |
811 | extern struct workqueue_struct *rds_wq; | |
2dc39357 | 812 | void rds_queue_reconnect(struct rds_connection *conn); |
39de8281 AG |
813 | void rds_connect_worker(struct work_struct *); |
814 | void rds_shutdown_worker(struct work_struct *); | |
815 | void rds_send_worker(struct work_struct *); | |
816 | void rds_recv_worker(struct work_struct *); | |
9c79440e | 817 | void rds_connect_path_complete(struct rds_connection *conn, int curr); |
39de8281 AG |
818 | void rds_connect_complete(struct rds_connection *conn); |
819 | ||
820 | /* transport.c */ | |
821 | int rds_trans_register(struct rds_transport *trans); | |
822 | void rds_trans_unregister(struct rds_transport *trans); | |
d5a8ac28 | 823 | struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr); |
5adb5bc6 | 824 | void rds_trans_put(struct rds_transport *trans); |
39de8281 AG |
825 | unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter, |
826 | unsigned int avail); | |
d97dac54 | 827 | struct rds_transport *rds_trans_get(int t_type); |
ef87b7ea | 828 | int rds_trans_init(void); |
39de8281 AG |
829 | void rds_trans_exit(void); |
830 | ||
831 | #endif |