blob: 67ba67c058b1b6b8e02631d664c41cac4deaadcc [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>
santosh.shilimkar@oracle.com7b565432015-10-30 08:49:10 -070010#include <linux/rhashtable.h>
Andy Grover39de8282009-02-24 15:30:19 +000011
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
Andy Grover8cbd9602009-04-01 08:20:20 +000032#ifdef ATOMIC64_INIT
33#define KERNEL_HAS_ATOMIC64
34#endif
35
shamir rabinovitchff570872016-10-27 05:46:38 -040036#ifdef RDS_DEBUG
Andy Grover39de8282009-02-24 15:30:19 +000037#define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
38#else
39/* sigh, pr_debug() causes unused variable warnings */
Joe Perchesb9075fa2011-10-31 17:11:33 -070040static inline __printf(1, 2)
41void rdsdebug(char *fmt, ...)
Andy Grover39de8282009-02-24 15:30:19 +000042{
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)
Andy Grover39de8282009-02-24 15:30:19 +000054#define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
55#define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
56
57struct rds_cong_map {
58 struct rb_node m_rb_node;
59 __be32 m_addr;
60 wait_queue_head_t m_waitq;
61 struct list_head m_conn_list;
62 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES];
63};
64
65
66/*
67 * This is how we will track the connection state:
68 * A connection is always in one of the following
69 * states. Updates to the state are atomic and imply
70 * a memory barrier.
71 */
72enum {
73 RDS_CONN_DOWN = 0,
74 RDS_CONN_CONNECTING,
75 RDS_CONN_DISCONNECTING,
76 RDS_CONN_UP,
Sowmini Varadhan9c794402016-06-04 14:00:00 -070077 RDS_CONN_RESETTING,
Andy Grover39de8282009-02-24 15:30:19 +000078 RDS_CONN_ERROR,
79};
80
81/* Bits for c_flags */
82#define RDS_LL_SEND_FULL 0
83#define RDS_RECONNECT_PENDING 1
Zach Brown0f4b1c72010-06-04 14:41:41 -070084#define RDS_IN_XMIT 2
santosh.shilimkar@oracle.com73ce4312015-08-22 15:45:26 -070085#define RDS_RECV_REFILL 3
Andy Grover39de8282009-02-24 15:30:19 +000086
Sowmini Varadhan0cb43962016-06-13 09:44:26 -070087/* Max number of multipaths per RDS connection. Must be a power of 2 */
Sowmini Varadhan5916e2c2016-07-14 03:51:03 -070088#define RDS_MPATH_WORKERS 8
89#define RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
90 (rs)->rs_hash_initval) & ((n) - 1))
Sowmini Varadhan0cb43962016-06-13 09:44:26 -070091
92/* Per mpath connection state */
93struct rds_conn_path {
94 struct rds_connection *cp_conn;
95 struct rds_message *cp_xmit_rm;
96 unsigned long cp_xmit_sg;
97 unsigned int cp_xmit_hdr_off;
98 unsigned int cp_xmit_data_off;
99 unsigned int cp_xmit_atomic_sent;
100 unsigned int cp_xmit_rdma_sent;
101 unsigned int cp_xmit_data_sent;
102
103 spinlock_t cp_lock; /* protect msg queues */
104 u64 cp_next_tx_seq;
105 struct list_head cp_send_queue;
106 struct list_head cp_retrans;
107
108 u64 cp_next_rx_seq;
109
110 void *cp_transport_data;
111
112 atomic_t cp_state;
113 unsigned long cp_send_gen;
114 unsigned long cp_flags;
115 unsigned long cp_reconnect_jiffies;
116 struct delayed_work cp_send_w;
117 struct delayed_work cp_recv_w;
118 struct delayed_work cp_conn_w;
119 struct work_struct cp_down_w;
120 struct mutex cp_cm_lock; /* protect cp_state & cm */
121 wait_queue_head_t cp_waitq;
122
123 unsigned int cp_unacked_packets;
124 unsigned int cp_unacked_bytes;
125 unsigned int cp_outgoing:1,
126 cp_pad_to_32:31;
127 unsigned int cp_index;
128};
129
130/* One rds_connection per RDS address pair */
Andy Grover39de8282009-02-24 15:30:19 +0000131struct rds_connection {
132 struct hlist_node c_hash_node;
133 __be32 c_laddr;
134 __be32 c_faddr;
Sowmini Varadhan3b20fc32015-09-30 16:54:07 -0400135 unsigned int c_loopback:1,
Sowmini Varadhan5916e2c2016-07-14 03:51:03 -0700136 c_ping_triggered:1,
137 c_pad_to_32:30;
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700138 int c_npaths;
Andy Grover39de8282009-02-24 15:30:19 +0000139 struct rds_connection *c_passive;
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700140 struct rds_transport *c_trans;
Andy Grover39de8282009-02-24 15:30:19 +0000141
142 struct rds_cong_map *c_lcong;
143 struct rds_cong_map *c_fcong;
144
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700145 /* Protocol version */
146 unsigned int c_version;
147 possible_net_t c_net;
Andy Grover39de8282009-02-24 15:30:19 +0000148
149 struct list_head c_map_item;
150 unsigned long c_map_queued;
Andy Grover39de8282009-02-24 15:30:19 +0000151
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700152 struct rds_conn_path c_path[RDS_MPATH_WORKERS];
Sowmini Varadhan5916e2c2016-07-14 03:51:03 -0700153 wait_queue_head_t c_hs_waitq; /* handshake waitq */
Andy Grover39de8282009-02-24 15:30:19 +0000154};
155
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400156static inline
157struct net *rds_conn_net(struct rds_connection *conn)
158{
159 return read_pnet(&conn->c_net);
160}
161
162static inline
163void rds_conn_net_set(struct rds_connection *conn, struct net *net)
164{
165 write_pnet(&conn->c_net, net);
166}
167
Andy Grover39de8282009-02-24 15:30:19 +0000168#define RDS_FLAG_CONG_BITMAP 0x01
169#define RDS_FLAG_ACK_REQUIRED 0x02
170#define RDS_FLAG_RETRANSMITTED 0x04
Steve Wise7b70d032009-04-09 14:09:39 +0000171#define RDS_MAX_ADV_CREDIT 255
Andy Grover39de8282009-02-24 15:30:19 +0000172
Sowmini Varadhan5916e2c2016-07-14 03:51:03 -0700173/* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
174 * probe to exchange control information before establishing a connection.
175 * Currently the control information that is exchanged is the number of
176 * supported paths. If the peer is a legacy (older kernel revision) peer,
177 * it would return a pong message without additional control information
178 * that would then alert the sender that the peer was an older rev.
179 */
180#define RDS_FLAG_PROBE_PORT 1
181#define RDS_HS_PROBE(sport, dport) \
182 ((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
183 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
Andy Grover39de8282009-02-24 15:30:19 +0000184/*
185 * Maximum space available for extension headers.
186 */
187#define RDS_HEADER_EXT_SPACE 16
188
189struct rds_header {
190 __be64 h_sequence;
191 __be64 h_ack;
192 __be32 h_len;
193 __be16 h_sport;
194 __be16 h_dport;
195 u8 h_flags;
196 u8 h_credit;
197 u8 h_padding[4];
198 __sum16 h_csum;
199
200 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
201};
202
203/*
204 * Reserved - indicates end of extensions
205 */
206#define RDS_EXTHDR_NONE 0
207
208/*
209 * This extension header is included in the very
210 * first message that is sent on a new connection,
211 * and identifies the protocol level. This will help
212 * rolling updates if a future change requires breaking
213 * the protocol.
214 * NB: This is no longer true for IB, where we do a version
215 * negotiation during the connection setup phase (protocol
216 * version information is included in the RDMA CM private data).
217 */
218#define RDS_EXTHDR_VERSION 1
219struct rds_ext_header_version {
220 __be32 h_version;
221};
222
223/*
224 * This extension header is included in the RDS message
225 * chasing an RDMA operation.
226 */
227#define RDS_EXTHDR_RDMA 2
228struct rds_ext_header_rdma {
229 __be32 h_rdma_rkey;
230};
231
232/*
233 * This extension header tells the peer about the
234 * destination <R_Key,offset> of the requested RDMA
235 * operation.
236 */
237#define RDS_EXTHDR_RDMA_DEST 3
238struct rds_ext_header_rdma_dest {
239 __be32 h_rdma_rkey;
240 __be32 h_rdma_offset;
241};
242
Sowmini Varadhan5916e2c2016-07-14 03:51:03 -0700243/* Extension header announcing number of paths.
244 * Implicit length = 2 bytes.
245 */
246#define RDS_EXTHDR_NPATHS 4
247
Andy Grover39de8282009-02-24 15:30:19 +0000248#define __RDS_EXTHDR_MAX 16 /* for now */
249
250struct rds_incoming {
251 atomic_t i_refcount;
252 struct list_head i_item;
253 struct rds_connection *i_conn;
Sowmini Varadhanef9e62c2016-06-13 09:44:28 -0700254 struct rds_conn_path *i_conn_path;
Andy Grover39de8282009-02-24 15:30:19 +0000255 struct rds_header i_hdr;
256 unsigned long i_rx_jiffies;
257 __be32 i_saddr;
258
259 rds_rdma_cookie_t i_rdma_cookie;
santosh.shilimkar@oracle.com5711f8b2016-03-01 15:20:43 -0800260 struct timeval i_rx_tstamp;
Andy Grover39de8282009-02-24 15:30:19 +0000261};
262
Andy Grover21f79af2010-01-12 12:57:27 -0800263struct rds_mr {
264 struct rb_node r_rb_node;
265 atomic_t r_refcount;
266 u32 r_key;
267
268 /* A copy of the creation flags */
269 unsigned int r_use_once:1;
270 unsigned int r_invalidate:1;
271 unsigned int r_write:1;
272
273 /* This is for RDS_MR_DEAD.
274 * It would be nice & consistent to make this part of the above
275 * bit field here, but we need to use test_and_set_bit.
276 */
277 unsigned long r_state;
278 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
279 struct rds_transport *r_trans;
280 void *r_trans_private;
281};
282
283/* Flags for mr->r_state */
284#define RDS_MR_DEAD 0
285
Andy Grover21f79af2010-01-12 12:57:27 -0800286static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
287{
288 return r_key | (((u64) offset) << 32);
289}
290
291static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
292{
293 return cookie;
294}
295
296static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
297{
298 return cookie >> 32;
299}
300
Andy Grover15133f62010-01-12 14:33:38 -0800301/* atomic operation types */
302#define RDS_ATOMIC_TYPE_CSWP 0
303#define RDS_ATOMIC_TYPE_FADD 1
304
Andy Grover39de8282009-02-24 15:30:19 +0000305/*
306 * m_sock_item and m_conn_item are on lists that are serialized under
307 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
308 * the message will not be put back on the retransmit list after being sent.
309 * messages that are canceled while being sent rely on this.
310 *
311 * m_inc is used by loopback so that it can pass an incoming message straight
312 * back up into the rx path. It embeds a wire header which is also used by
313 * the send path, which is kind of awkward.
314 *
315 * m_sock_item indicates the message's presence on a socket's send or receive
316 * queue. m_rs will point to that socket.
317 *
318 * m_daddr is used by cancellation to prune messages to a given destination.
319 *
320 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
321 * nesting. As paths iterate over messages on a sock, or conn, they must
322 * also lock the conn, or sock, to remove the message from those lists too.
323 * Testing the flag to determine if the message is still on the lists lets
324 * us avoid testing the list_head directly. That means each path can use
325 * the message's list_head to keep it on a local list while juggling locks
326 * without confusing the other path.
327 *
328 * m_ack_seq is an optional field set by transports who need a different
329 * sequence number range to invalidate. They can use this in a callback
330 * that they pass to rds_send_drop_acked() to see if each message has been
331 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
332 * had ack_seq set yet.
333 */
334#define RDS_MSG_ON_SOCK 1
335#define RDS_MSG_ON_CONN 2
336#define RDS_MSG_HAS_ACK_SEQ 3
337#define RDS_MSG_ACK_REQUIRED 4
338#define RDS_MSG_RETRANSMITTED 5
339#define RDS_MSG_MAPPED 6
340#define RDS_MSG_PAGEVEC 7
341
342struct rds_message {
343 atomic_t m_refcount;
344 struct list_head m_sock_item;
345 struct list_head m_conn_item;
346 struct rds_incoming m_inc;
347 u64 m_ack_seq;
348 __be32 m_daddr;
349 unsigned long m_flags;
350
351 /* Never access m_rs without holding m_rs_lock.
352 * Lock nesting is
353 * rm->m_rs_lock
354 * -> rs->rs_lock
355 */
356 spinlock_t m_rs_lock;
Chris Masonc83188d2010-04-21 13:09:28 -0700357 wait_queue_head_t m_flush_wait;
358
Andy Grover39de8282009-02-24 15:30:19 +0000359 struct rds_sock *m_rs;
Andy Grover7e3bd652010-03-01 16:04:59 -0800360
361 /* cookie to send to remote, in rds header */
Andy Grover39de8282009-02-24 15:30:19 +0000362 rds_rdma_cookie_t m_rdma_cookie;
Andy Grover7e3bd652010-03-01 16:04:59 -0800363
364 unsigned int m_used_sgs;
365 unsigned int m_total_sgs;
366
Andy Groverff3d7d32010-03-01 14:03:09 -0800367 void *m_final_op;
368
Andy Grovere7791372010-01-12 12:15:02 -0800369 struct {
Andy Grover15133f62010-01-12 14:33:38 -0800370 struct rm_atomic_op {
371 int op_type;
Andy Grover20c72bd2010-08-25 05:51:28 -0700372 union {
373 struct {
374 uint64_t compare;
375 uint64_t swap;
376 uint64_t compare_mask;
377 uint64_t swap_mask;
378 } op_m_cswp;
379 struct {
380 uint64_t add;
381 uint64_t nocarry_mask;
382 } op_m_fadd;
383 };
Andy Grover15133f62010-01-12 14:33:38 -0800384
385 u32 op_rkey;
386 u64 op_remote_addr;
387 unsigned int op_notify:1;
388 unsigned int op_recverr:1;
389 unsigned int op_mapped:1;
Andy Grover2c3a5f92010-03-01 16:10:40 -0800390 unsigned int op_silent:1;
Andy Grover15133f62010-01-12 14:33:38 -0800391 unsigned int op_active:1;
Andy Grover15133f62010-01-12 14:33:38 -0800392 struct scatterlist *op_sg;
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800393 struct rds_notifier *op_notifier;
Andy Grover15133f62010-01-12 14:33:38 -0800394
395 struct rds_mr *op_rdma_mr;
396 } atomic;
397 struct rm_rdma_op {
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800398 u32 op_rkey;
399 u64 op_remote_addr;
400 unsigned int op_write:1;
401 unsigned int op_fence:1;
402 unsigned int op_notify:1;
403 unsigned int op_recverr:1;
404 unsigned int op_mapped:1;
Andy Grover2c3a5f92010-03-01 16:10:40 -0800405 unsigned int op_silent:1;
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800406 unsigned int op_active:1;
407 unsigned int op_bytes;
408 unsigned int op_nents;
409 unsigned int op_count;
410 struct scatterlist *op_sg;
411 struct rds_notifier *op_notifier;
412
413 struct rds_mr *op_rdma_mr;
Andy Grovere7791372010-01-12 12:15:02 -0800414 } rdma;
Andy Grover15133f62010-01-12 14:33:38 -0800415 struct rm_data_op {
Andy Grover241eef32010-01-19 21:25:26 -0800416 unsigned int op_active:1;
Andy Grover6c7cc6e2010-01-27 18:04:18 -0800417 unsigned int op_nents;
418 unsigned int op_count;
Wengang Wangd655a9f2015-05-21 13:11:40 +0800419 unsigned int op_dmasg;
420 unsigned int op_dmaoff;
Andy Grover6c7cc6e2010-01-27 18:04:18 -0800421 struct scatterlist *op_sg;
Andy Grovere7791372010-01-12 12:15:02 -0800422 } data;
423 };
Andy Grover39de8282009-02-24 15:30:19 +0000424};
425
426/*
427 * The RDS notifier is used (optionally) to tell the application about
428 * completed RDMA operations. Rather than keeping the whole rds message
429 * around on the queue, we allocate a small notifier that is put on the
430 * socket's notifier_list. Notifications are delivered to the application
431 * through control messages.
432 */
433struct rds_notifier {
434 struct list_head n_list;
435 uint64_t n_user_token;
436 int n_status;
437};
438
439/**
440 * struct rds_transport - transport specific behavioural hooks
441 *
442 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
443 * part of a message. The caller serializes on the send_sem so this
444 * doesn't need to be reentrant for a given conn. The header must be
445 * sent before the data payload. .xmit must be prepared to send a
446 * message with no data payload. .xmit should return the number of
447 * bytes that were sent down the connection, including header bytes.
448 * Returning 0 tells the caller that it doesn't need to perform any
449 * additional work now. This is usually the case when the transport has
450 * filled the sending queue for its connection and will handle
451 * triggering the rds thread to continue the send when space becomes
452 * available. Returning -EAGAIN tells the caller to retry the send
453 * immediately. Returning -ENOMEM tells the caller to retry the send at
454 * some point in the future.
455 *
456 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
457 * it returns the connection can not call rds_recv_incoming().
458 * This will only be called once after conn_connect returns
459 * non-zero success and will The caller serializes this with
460 * the send and connecting paths (xmit_* and conn_*). The
461 * transport is responsible for other serialization, including
462 * rds_recv_incoming(). This is called in process context but
463 * should try hard not to block.
Andy Grover39de8282009-02-24 15:30:19 +0000464 */
465
466struct rds_transport {
467 char t_name[TRANSNAMSIZ];
468 struct list_head t_item;
469 struct module *t_owner;
Sowmini Varadhan7e8f4412016-06-13 09:44:27 -0700470 unsigned int t_prefer_loopback:1,
471 t_mp_capable:1;
Andy Grover335776b2009-08-21 12:28:34 +0000472 unsigned int t_type;
Andy Grover39de8282009-02-24 15:30:19 +0000473
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400474 int (*laddr_check)(struct net *net, __be32 addr);
Andy Grover39de8282009-02-24 15:30:19 +0000475 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
476 void (*conn_free)(void *data);
Sowmini Varadhanb04e8552016-06-30 16:11:16 -0700477 int (*conn_path_connect)(struct rds_conn_path *cp);
Sowmini Varadhand769ef82016-06-13 09:44:41 -0700478 void (*conn_path_shutdown)(struct rds_conn_path *conn);
Sowmini Varadhan1f9ecd72016-06-13 09:44:34 -0700479 void (*xmit_path_prepare)(struct rds_conn_path *cp);
Sowmini Varadhan1f9ecd72016-06-13 09:44:34 -0700480 void (*xmit_path_complete)(struct rds_conn_path *cp);
Andy Grover39de8282009-02-24 15:30:19 +0000481 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
482 unsigned int hdr_off, unsigned int sg, unsigned int off);
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800483 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
Andy Groverff3d7d32010-03-01 14:03:09 -0800484 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
Sowmini Varadhan2da43c42016-06-30 16:11:15 -0700485 int (*recv_path)(struct rds_conn_path *cp);
Al Viroc310e722014-11-20 09:21:14 -0500486 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
Andy Grover39de8282009-02-24 15:30:19 +0000487 void (*inc_free)(struct rds_incoming *inc);
488
489 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
490 struct rdma_cm_event *event);
491 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
492 void (*cm_connect_complete)(struct rds_connection *conn,
493 struct rdma_cm_event *event);
494
495 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
496 unsigned int avail);
497 void (*exit)(void);
498 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
499 struct rds_sock *rs, u32 *key_ret);
500 void (*sync_mr)(void *trans_private, int direction);
501 void (*free_mr)(void *trans_private, int invalidate);
502 void (*flush_mrs)(void);
503};
504
505struct rds_sock {
506 struct sock rs_sk;
507
508 u64 rs_user_addr;
509 u64 rs_user_bytes;
510
511 /*
512 * bound_addr used for both incoming and outgoing, no INADDR_ANY
513 * support.
514 */
santosh.shilimkar@oracle.com7b565432015-10-30 08:49:10 -0700515 struct rhash_head rs_bound_node;
516 u64 rs_bound_key;
Andy Grover39de8282009-02-24 15:30:19 +0000517 __be32 rs_bound_addr;
518 __be32 rs_conn_addr;
519 __be16 rs_bound_port;
520 __be16 rs_conn_port;
Andy Grover39de8282009-02-24 15:30:19 +0000521 struct rds_transport *rs_transport;
522
523 /*
524 * rds_sendmsg caches the conn it used the last time around.
525 * This helps avoid costly lookups.
526 */
527 struct rds_connection *rs_conn;
528
529 /* flag indicating we were congested or not */
530 int rs_congested;
Andy Groverb98ba522010-03-11 13:50:04 +0000531 /* seen congestion (ENOBUFS) when sending? */
532 int rs_seen_congestion;
Andy Grover39de8282009-02-24 15:30:19 +0000533
534 /* rs_lock protects all these adjacent members before the newline */
535 spinlock_t rs_lock;
536 struct list_head rs_send_queue;
537 u32 rs_snd_bytes;
538 int rs_rcv_bytes;
539 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
540
541 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
542 * to decide whether the application should be woken up.
543 * If not set, we use rs_cong_track to find out whether a cong map
544 * update arrived.
545 */
546 uint64_t rs_cong_mask;
547 uint64_t rs_cong_notify;
548 struct list_head rs_cong_list;
549 unsigned long rs_cong_track;
550
551 /*
552 * rs_recv_lock protects the receive queue, and is
553 * used to serialize with rds_release.
554 */
555 rwlock_t rs_recv_lock;
556 struct list_head rs_recv_queue;
557
558 /* just for stats reporting */
559 struct list_head rs_item;
560
561 /* these have their own lock */
562 spinlock_t rs_rdma_lock;
563 struct rb_root rs_rdma_keys;
564
565 /* Socket options - in case there will be more */
566 unsigned char rs_recverr,
567 rs_cong_monitor;
Sowmini Varadhan5916e2c2016-07-14 03:51:03 -0700568 u32 rs_hash_initval;
Andy Grover39de8282009-02-24 15:30:19 +0000569};
570
571static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
572{
573 return container_of(sk, struct rds_sock, rs_sk);
574}
575static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
576{
577 return &rs->rs_sk;
578}
579
580/*
581 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
582 * to account for overhead. We don't account for overhead, we just apply
583 * the number of payload bytes to the specified value.
584 */
585static inline int rds_sk_sndbuf(struct rds_sock *rs)
586{
587 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
588}
589static inline int rds_sk_rcvbuf(struct rds_sock *rs)
590{
591 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
592}
593
594struct rds_statistics {
595 uint64_t s_conn_reset;
596 uint64_t s_recv_drop_bad_checksum;
597 uint64_t s_recv_drop_old_seq;
598 uint64_t s_recv_drop_no_sock;
599 uint64_t s_recv_drop_dead_sock;
600 uint64_t s_recv_deliver_raced;
601 uint64_t s_recv_delivered;
602 uint64_t s_recv_queued;
603 uint64_t s_recv_immediate_retry;
604 uint64_t s_recv_delayed_retry;
605 uint64_t s_recv_ack_required;
606 uint64_t s_recv_rdma_bytes;
607 uint64_t s_recv_ping;
608 uint64_t s_send_queue_empty;
609 uint64_t s_send_queue_full;
Andy Grover049ee3f2010-03-23 17:39:07 -0700610 uint64_t s_send_lock_contention;
611 uint64_t s_send_lock_queue_raced;
Andy Grover39de8282009-02-24 15:30:19 +0000612 uint64_t s_send_immediate_retry;
613 uint64_t s_send_delayed_retry;
614 uint64_t s_send_drop_acked;
615 uint64_t s_send_ack_required;
616 uint64_t s_send_queued;
617 uint64_t s_send_rdma;
618 uint64_t s_send_rdma_bytes;
619 uint64_t s_send_pong;
620 uint64_t s_page_remainder_hit;
621 uint64_t s_page_remainder_miss;
622 uint64_t s_copy_to_user;
623 uint64_t s_copy_from_user;
624 uint64_t s_cong_update_queued;
625 uint64_t s_cong_update_received;
626 uint64_t s_cong_send_error;
627 uint64_t s_cong_send_blocked;
628};
629
630/* af_rds.c */
631void rds_sock_addref(struct rds_sock *rs);
632void rds_sock_put(struct rds_sock *rs);
633void rds_wake_sk_sleep(struct rds_sock *rs);
634static inline void __rds_wake_sk_sleep(struct sock *sk)
635{
Eric Dumazetaa395142010-04-20 13:03:51 +0000636 wait_queue_head_t *waitq = sk_sleep(sk);
Andy Grover39de8282009-02-24 15:30:19 +0000637
638 if (!sock_flag(sk, SOCK_DEAD) && waitq)
639 wake_up(waitq);
640}
641extern wait_queue_head_t rds_poll_waitq;
642
643
644/* bind.c */
645int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
646void rds_remove_bound(struct rds_sock *rs);
647struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
santosh.shilimkar@oracle.com7b565432015-10-30 08:49:10 -0700648int rds_bind_lock_init(void);
649void rds_bind_lock_destroy(void);
Andy Grover39de8282009-02-24 15:30:19 +0000650
651/* cong.c */
652int rds_cong_get_maps(struct rds_connection *conn);
653void rds_cong_add_conn(struct rds_connection *conn);
654void rds_cong_remove_conn(struct rds_connection *conn);
655void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
656void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
657int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
658void rds_cong_queue_updates(struct rds_cong_map *map);
659void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
660int rds_cong_updated_since(unsigned long *recent);
661void rds_cong_add_socket(struct rds_sock *);
662void rds_cong_remove_socket(struct rds_sock *);
663void rds_cong_exit(void);
664struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
665
666/* conn.c */
Zach Brownef87b7e2010-07-09 12:26:20 -0700667int rds_conn_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000668void rds_conn_exit(void);
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400669struct rds_connection *rds_conn_create(struct net *net,
670 __be32 laddr, __be32 faddr,
Andy Grover39de8282009-02-24 15:30:19 +0000671 struct rds_transport *trans, gfp_t gfp);
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400672struct rds_connection *rds_conn_create_outgoing(struct net *net,
673 __be32 laddr, __be32 faddr,
Andy Grover39de8282009-02-24 15:30:19 +0000674 struct rds_transport *trans, gfp_t gfp);
Sowmini Varadhand769ef82016-06-13 09:44:41 -0700675void rds_conn_shutdown(struct rds_conn_path *cpath);
Andy Grover39de8282009-02-24 15:30:19 +0000676void rds_conn_destroy(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000677void rds_conn_drop(struct rds_connection *conn);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700678void rds_conn_path_drop(struct rds_conn_path *cpath);
Zach Brownf3c68082010-05-24 13:14:36 -0700679void rds_conn_connect_if_down(struct rds_connection *conn);
Sowmini Varadhan3c0a5902016-06-13 09:44:37 -0700680void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
Andy Grover39de8282009-02-24 15:30:19 +0000681void rds_for_each_conn_info(struct socket *sock, unsigned int len,
682 struct rds_info_iterator *iter,
683 struct rds_info_lengths *lens,
684 int (*visitor)(struct rds_connection *, void *),
685 size_t item_len);
Joe Perchesb9075fa2011-10-31 17:11:33 -0700686__printf(2, 3)
687void __rds_conn_error(struct rds_connection *conn, const char *, ...);
Andy Grover39de8282009-02-24 15:30:19 +0000688#define rds_conn_error(conn, fmt...) \
689 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
690
Nicolas Iooss6cdaf032016-08-05 22:11:12 +0200691__printf(2, 3)
Sowmini Varadhanfb1b3dc2016-06-13 09:44:39 -0700692void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
693#define rds_conn_path_error(cp, fmt...) \
694 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
695
Andy Grover39de8282009-02-24 15:30:19 +0000696static inline int
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700697rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
698{
699 return atomic_cmpxchg(&cp->cp_state, old, new) == old;
700}
701
702static inline int
Andy Grover39de8282009-02-24 15:30:19 +0000703rds_conn_transition(struct rds_connection *conn, int old, int new)
704{
Sowmini Varadhan7e8f4412016-06-13 09:44:27 -0700705 WARN_ON(conn->c_trans->t_mp_capable);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700706 return rds_conn_path_transition(&conn->c_path[0], old, new);
707}
708
709static inline int
710rds_conn_path_state(struct rds_conn_path *cp)
711{
712 return atomic_read(&cp->cp_state);
Andy Grover39de8282009-02-24 15:30:19 +0000713}
714
715static inline int
716rds_conn_state(struct rds_connection *conn)
717{
Sowmini Varadhan7e8f4412016-06-13 09:44:27 -0700718 WARN_ON(conn->c_trans->t_mp_capable);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700719 return rds_conn_path_state(&conn->c_path[0]);
720}
721
722static inline int
723rds_conn_path_up(struct rds_conn_path *cp)
724{
725 return atomic_read(&cp->cp_state) == RDS_CONN_UP;
Andy Grover39de8282009-02-24 15:30:19 +0000726}
727
728static inline int
729rds_conn_up(struct rds_connection *conn)
730{
Sowmini Varadhan7e8f4412016-06-13 09:44:27 -0700731 WARN_ON(conn->c_trans->t_mp_capable);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700732 return rds_conn_path_up(&conn->c_path[0]);
733}
734
735static inline int
736rds_conn_path_connecting(struct rds_conn_path *cp)
737{
738 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
Andy Grover39de8282009-02-24 15:30:19 +0000739}
740
741static inline int
742rds_conn_connecting(struct rds_connection *conn)
743{
Sowmini Varadhan7e8f4412016-06-13 09:44:27 -0700744 WARN_ON(conn->c_trans->t_mp_capable);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700745 return rds_conn_path_connecting(&conn->c_path[0]);
Andy Grover39de8282009-02-24 15:30:19 +0000746}
747
748/* message.c */
749struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
Andy Groverfc445082010-01-12 12:56:06 -0800750struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
Al Viro083735f2014-11-20 09:31:08 -0500751int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
Andy Grover39de8282009-02-24 15:30:19 +0000752struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
753void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
754 __be16 dport, u64 seq);
755int rds_message_add_extension(struct rds_header *hdr,
756 unsigned int type, const void *data, unsigned int len);
757int rds_message_next_extension(struct rds_header *hdr,
758 unsigned int *pos, void *buf, unsigned int *buflen);
Andy Grover39de8282009-02-24 15:30:19 +0000759int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
Al Viroc310e722014-11-20 09:21:14 -0500760int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
Andy Grover39de8282009-02-24 15:30:19 +0000761void rds_message_inc_free(struct rds_incoming *inc);
762void rds_message_addref(struct rds_message *rm);
763void rds_message_put(struct rds_message *rm);
764void rds_message_wait(struct rds_message *rm);
765void rds_message_unmapped(struct rds_message *rm);
766
767static inline void rds_message_make_checksum(struct rds_header *hdr)
768{
769 hdr->h_csum = 0;
770 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
771}
772
773static inline int rds_message_verify_checksum(const struct rds_header *hdr)
774{
775 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
776}
777
778
779/* page.c */
780int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
781 gfp_t gfp);
782int rds_page_copy_user(struct page *page, unsigned long offset,
783 void __user *ptr, unsigned long bytes,
784 int to_user);
785#define rds_page_copy_to_user(page, offset, ptr, bytes) \
786 rds_page_copy_user(page, offset, ptr, bytes, 1)
787#define rds_page_copy_from_user(page, offset, ptr, bytes) \
788 rds_page_copy_user(page, offset, ptr, bytes, 0)
789void rds_page_exit(void);
790
791/* recv.c */
792void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
793 __be32 saddr);
Sowmini Varadhan5e833e02016-06-13 09:44:29 -0700794void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
795 __be32 saddr);
Andy Grover39de8282009-02-24 15:30:19 +0000796void rds_inc_put(struct rds_incoming *inc);
797void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
Cong Wang6114eab2011-11-25 23:14:40 +0800798 struct rds_incoming *inc, gfp_t gfp);
Ying Xue1b784142015-03-02 15:37:48 +0800799int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
800 int msg_flags);
Andy Grover39de8282009-02-24 15:30:19 +0000801void rds_clear_recv_queue(struct rds_sock *rs);
802int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
803void rds_inc_info_copy(struct rds_incoming *inc,
804 struct rds_info_iterator *iter,
805 __be32 saddr, __be32 daddr, int flip);
806
807/* send.c */
Ying Xue1b784142015-03-02 15:37:48 +0800808int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
Sowmini Varadhand769ef82016-06-13 09:44:41 -0700809void rds_send_path_reset(struct rds_conn_path *conn);
Sowmini Varadhan1f9ecd72016-06-13 09:44:34 -0700810int rds_send_xmit(struct rds_conn_path *cp);
Andy Grover39de8282009-02-24 15:30:19 +0000811struct sockaddr_in;
812void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
813typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
814void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
815 is_acked_func is_acked);
Sowmini Varadhan5c3d2742016-06-13 09:44:31 -0700816void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
817 is_acked_func is_acked);
Sowmini Varadhan45997e92016-06-13 09:44:36 -0700818int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
Andy Grover39de8282009-02-24 15:30:19 +0000819
820/* rdma.c */
821void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
Andy Grover21f79af2010-01-12 12:57:27 -0800822int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
823int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
824int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
825void rds_rdma_drop_keys(struct rds_sock *rs);
826int rds_rdma_extra_size(struct rds_rdma_args *args);
827int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
828 struct cmsghdr *cmsg);
829int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
830 struct cmsghdr *cmsg);
831int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
832 struct cmsghdr *cmsg);
833int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
834 struct cmsghdr *cmsg);
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800835void rds_rdma_free_op(struct rm_rdma_op *ro);
Andy Groverd0ab25a2010-01-27 16:15:48 -0800836void rds_atomic_free_op(struct rm_atomic_op *ao);
Andy Grover15133f62010-01-12 14:33:38 -0800837void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
838void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
839int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
840 struct cmsghdr *cmsg);
Andy Grover21f79af2010-01-12 12:57:27 -0800841
Joe Perchesc1b12032013-10-18 13:48:25 -0700842void __rds_put_mr_final(struct rds_mr *mr);
Andy Grover21f79af2010-01-12 12:57:27 -0800843static inline void rds_mr_put(struct rds_mr *mr)
844{
845 if (atomic_dec_and_test(&mr->r_refcount))
846 __rds_put_mr_final(mr);
847}
Andy Grover39de8282009-02-24 15:30:19 +0000848
849/* stats.c */
David Howells9b8de742009-04-21 23:00:24 +0100850DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
Andy Grover39de8282009-02-24 15:30:19 +0000851#define rds_stats_inc_which(which, member) do { \
852 per_cpu(which, get_cpu()).member++; \
853 put_cpu(); \
854} while (0)
855#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
856#define rds_stats_add_which(which, member, count) do { \
857 per_cpu(which, get_cpu()).member += count; \
858 put_cpu(); \
859} while (0)
860#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
Zach Brownef87b7e2010-07-09 12:26:20 -0700861int rds_stats_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000862void rds_stats_exit(void);
863void rds_stats_info_copy(struct rds_info_iterator *iter,
Jan Engelhardt36cbd3d2009-08-05 10:42:58 -0700864 uint64_t *values, const char *const *names,
865 size_t nr);
Andy Grover39de8282009-02-24 15:30:19 +0000866
867/* sysctl.c */
Zach Brownef87b7e2010-07-09 12:26:20 -0700868int rds_sysctl_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000869void rds_sysctl_exit(void);
870extern unsigned long rds_sysctl_sndbuf_min;
871extern unsigned long rds_sysctl_sndbuf_default;
872extern unsigned long rds_sysctl_sndbuf_max;
873extern unsigned long rds_sysctl_reconnect_min_jiffies;
874extern unsigned long rds_sysctl_reconnect_max_jiffies;
875extern unsigned int rds_sysctl_max_unacked_packets;
876extern unsigned int rds_sysctl_max_unacked_bytes;
877extern unsigned int rds_sysctl_ping_enable;
878extern unsigned long rds_sysctl_trace_flags;
879extern unsigned int rds_sysctl_trace_level;
880
881/* threads.c */
Zach Brownef87b7e2010-07-09 12:26:20 -0700882int rds_threads_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000883void rds_threads_exit(void);
884extern struct workqueue_struct *rds_wq;
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700885void rds_queue_reconnect(struct rds_conn_path *cp);
Andy Grover39de8282009-02-24 15:30:19 +0000886void rds_connect_worker(struct work_struct *);
887void rds_shutdown_worker(struct work_struct *);
888void rds_send_worker(struct work_struct *);
889void rds_recv_worker(struct work_struct *);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700890void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
Andy Grover39de8282009-02-24 15:30:19 +0000891void rds_connect_complete(struct rds_connection *conn);
892
893/* transport.c */
894int rds_trans_register(struct rds_transport *trans);
895void rds_trans_unregister(struct rds_transport *trans);
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400896struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
Zach Brown5adb5bc2010-07-23 10:32:31 -0700897void rds_trans_put(struct rds_transport *trans);
Andy Grover39de8282009-02-24 15:30:19 +0000898unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
899 unsigned int avail);
Sowmini Varadhand97dac52015-05-29 17:28:08 -0400900struct rds_transport *rds_trans_get(int t_type);
Zach Brownef87b7e2010-07-09 12:26:20 -0700901int rds_trans_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000902void rds_trans_exit(void);
903
904#endif