5 #include <linux/scatterlist.h>
6 #include <linux/highmem.h>
7 #include <rdma/rdma_cm.h>
8 #include <linux/mutex.h>
10 #include <linux/rhashtable.h>
15 * RDS Network protocol version
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)
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.
30 #define RDS_PORT 18634
33 #define KERNEL_HAS_ATOMIC64
37 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
39 /* sigh, pr_debug() causes unused variable warnings */
40 static inline __printf(1, 2)
41 void rdsdebug(char *fmt, ...)
46 /* XXX is there one of these somewhere? */
48 ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
50 #define RDS_FRAG_SHIFT 12
51 #define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT))
53 #define RDS_CONG_MAP_BYTES (65536 / 8)
54 #define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
55 #define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
58 struct rb_node m_rb_node;
60 wait_queue_head_t m_waitq;
61 struct list_head m_conn_list;
62 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES];
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
75 RDS_CONN_DISCONNECTING,
81 /* Bits for c_flags */
82 #define RDS_LL_SEND_FULL 0
83 #define RDS_RECONNECT_PENDING 1
85 #define RDS_RECV_REFILL 3
87 /* Max number of multipaths per RDS connection. Must be a power of 2 */
88 #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))
92 /* Per mpath connection state */
93 struct 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;
103 spinlock_t cp_lock; /* protect msg queues */
105 struct list_head cp_send_queue;
106 struct list_head cp_retrans;
110 void *cp_transport_data;
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;
123 unsigned int cp_unacked_packets;
124 unsigned int cp_unacked_bytes;
125 unsigned int cp_outgoing:1,
127 unsigned int cp_index;
130 /* One rds_connection per RDS address pair */
131 struct rds_connection {
132 struct hlist_node c_hash_node;
135 unsigned int c_loopback:1,
139 struct rds_connection *c_passive;
140 struct rds_transport *c_trans;
142 struct rds_cong_map *c_lcong;
143 struct rds_cong_map *c_fcong;
145 /* Protocol version */
146 unsigned int c_version;
147 possible_net_t c_net;
149 struct list_head c_map_item;
150 unsigned long c_map_queued;
152 struct rds_conn_path c_path[RDS_MPATH_WORKERS];
153 wait_queue_head_t c_hs_waitq; /* handshake waitq */
160 struct net *rds_conn_net(struct rds_connection *conn)
162 return read_pnet(&conn->c_net);
166 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
168 write_pnet(&conn->c_net, net);
171 #define RDS_FLAG_CONG_BITMAP 0x01
172 #define RDS_FLAG_ACK_REQUIRED 0x02
173 #define RDS_FLAG_RETRANSMITTED 0x04
174 #define RDS_MAX_ADV_CREDIT 255
176 /* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
177 * probe to exchange control information before establishing a connection.
178 * Currently the control information that is exchanged is the number of
179 * supported paths. If the peer is a legacy (older kernel revision) peer,
180 * it would return a pong message without additional control information
181 * that would then alert the sender that the peer was an older rev.
183 #define RDS_FLAG_PROBE_PORT 1
184 #define RDS_HS_PROBE(sport, dport) \
185 ((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
186 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
188 * Maximum space available for extension headers.
190 #define RDS_HEADER_EXT_SPACE 16
203 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
207 * Reserved - indicates end of extensions
209 #define RDS_EXTHDR_NONE 0
212 * This extension header is included in the very
213 * first message that is sent on a new connection,
214 * and identifies the protocol level. This will help
215 * rolling updates if a future change requires breaking
217 * NB: This is no longer true for IB, where we do a version
218 * negotiation during the connection setup phase (protocol
219 * version information is included in the RDMA CM private data).
221 #define RDS_EXTHDR_VERSION 1
222 struct rds_ext_header_version {
227 * This extension header is included in the RDS message
228 * chasing an RDMA operation.
230 #define RDS_EXTHDR_RDMA 2
231 struct rds_ext_header_rdma {
236 * This extension header tells the peer about the
237 * destination <R_Key,offset> of the requested RDMA
240 #define RDS_EXTHDR_RDMA_DEST 3
241 struct rds_ext_header_rdma_dest {
243 __be32 h_rdma_offset;
246 /* Extension header announcing number of paths.
247 * Implicit length = 2 bytes.
249 #define RDS_EXTHDR_NPATHS 5
250 #define RDS_EXTHDR_GEN_NUM 6
252 #define __RDS_EXTHDR_MAX 16 /* for now */
254 struct rds_incoming {
256 struct list_head i_item;
257 struct rds_connection *i_conn;
258 struct rds_conn_path *i_conn_path;
259 struct rds_header i_hdr;
260 unsigned long i_rx_jiffies;
263 rds_rdma_cookie_t i_rdma_cookie;
264 struct timeval i_rx_tstamp;
268 struct rb_node r_rb_node;
272 /* A copy of the creation flags */
273 unsigned int r_use_once:1;
274 unsigned int r_invalidate:1;
275 unsigned int r_write:1;
277 /* This is for RDS_MR_DEAD.
278 * It would be nice & consistent to make this part of the above
279 * bit field here, but we need to use test_and_set_bit.
281 unsigned long r_state;
282 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
283 struct rds_transport *r_trans;
284 void *r_trans_private;
287 /* Flags for mr->r_state */
288 #define RDS_MR_DEAD 0
290 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
292 return r_key | (((u64) offset) << 32);
295 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
300 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
305 /* atomic operation types */
306 #define RDS_ATOMIC_TYPE_CSWP 0
307 #define RDS_ATOMIC_TYPE_FADD 1
310 * m_sock_item and m_conn_item are on lists that are serialized under
311 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
312 * the message will not be put back on the retransmit list after being sent.
313 * messages that are canceled while being sent rely on this.
315 * m_inc is used by loopback so that it can pass an incoming message straight
316 * back up into the rx path. It embeds a wire header which is also used by
317 * the send path, which is kind of awkward.
319 * m_sock_item indicates the message's presence on a socket's send or receive
320 * queue. m_rs will point to that socket.
322 * m_daddr is used by cancellation to prune messages to a given destination.
324 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
325 * nesting. As paths iterate over messages on a sock, or conn, they must
326 * also lock the conn, or sock, to remove the message from those lists too.
327 * Testing the flag to determine if the message is still on the lists lets
328 * us avoid testing the list_head directly. That means each path can use
329 * the message's list_head to keep it on a local list while juggling locks
330 * without confusing the other path.
332 * m_ack_seq is an optional field set by transports who need a different
333 * sequence number range to invalidate. They can use this in a callback
334 * that they pass to rds_send_drop_acked() to see if each message has been
335 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
336 * had ack_seq set yet.
338 #define RDS_MSG_ON_SOCK 1
339 #define RDS_MSG_ON_CONN 2
340 #define RDS_MSG_HAS_ACK_SEQ 3
341 #define RDS_MSG_ACK_REQUIRED 4
342 #define RDS_MSG_RETRANSMITTED 5
343 #define RDS_MSG_MAPPED 6
344 #define RDS_MSG_PAGEVEC 7
345 #define RDS_MSG_FLUSH 8
349 struct list_head m_sock_item;
350 struct list_head m_conn_item;
351 struct rds_incoming m_inc;
354 unsigned long m_flags;
356 /* Never access m_rs without holding m_rs_lock.
361 spinlock_t m_rs_lock;
362 wait_queue_head_t m_flush_wait;
364 struct rds_sock *m_rs;
366 /* cookie to send to remote, in rds header */
367 rds_rdma_cookie_t m_rdma_cookie;
369 unsigned int m_used_sgs;
370 unsigned int m_total_sgs;
375 struct rm_atomic_op {
381 uint64_t compare_mask;
386 uint64_t nocarry_mask;
392 unsigned int op_notify:1;
393 unsigned int op_recverr:1;
394 unsigned int op_mapped:1;
395 unsigned int op_silent:1;
396 unsigned int op_active:1;
397 struct scatterlist *op_sg;
398 struct rds_notifier *op_notifier;
400 struct rds_mr *op_rdma_mr;
405 unsigned int op_write:1;
406 unsigned int op_fence:1;
407 unsigned int op_notify:1;
408 unsigned int op_recverr:1;
409 unsigned int op_mapped:1;
410 unsigned int op_silent:1;
411 unsigned int op_active:1;
412 unsigned int op_bytes;
413 unsigned int op_nents;
414 unsigned int op_count;
415 struct scatterlist *op_sg;
416 struct rds_notifier *op_notifier;
418 struct rds_mr *op_rdma_mr;
421 unsigned int op_active:1;
422 unsigned int op_notify:1;
423 unsigned int op_nents;
424 unsigned int op_count;
425 unsigned int op_dmasg;
426 unsigned int op_dmaoff;
427 struct scatterlist *op_sg;
433 * The RDS notifier is used (optionally) to tell the application about
434 * completed RDMA operations. Rather than keeping the whole rds message
435 * around on the queue, we allocate a small notifier that is put on the
436 * socket's notifier_list. Notifications are delivered to the application
437 * through control messages.
439 struct rds_notifier {
440 struct list_head n_list;
441 uint64_t n_user_token;
446 * struct rds_transport - transport specific behavioural hooks
448 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
449 * part of a message. The caller serializes on the send_sem so this
450 * doesn't need to be reentrant for a given conn. The header must be
451 * sent before the data payload. .xmit must be prepared to send a
452 * message with no data payload. .xmit should return the number of
453 * bytes that were sent down the connection, including header bytes.
454 * Returning 0 tells the caller that it doesn't need to perform any
455 * additional work now. This is usually the case when the transport has
456 * filled the sending queue for its connection and will handle
457 * triggering the rds thread to continue the send when space becomes
458 * available. Returning -EAGAIN tells the caller to retry the send
459 * immediately. Returning -ENOMEM tells the caller to retry the send at
460 * some point in the future.
462 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
463 * it returns the connection can not call rds_recv_incoming().
464 * This will only be called once after conn_connect returns
465 * non-zero success and will The caller serializes this with
466 * the send and connecting paths (xmit_* and conn_*). The
467 * transport is responsible for other serialization, including
468 * rds_recv_incoming(). This is called in process context but
469 * should try hard not to block.
472 struct rds_transport {
473 char t_name[TRANSNAMSIZ];
474 struct list_head t_item;
475 struct module *t_owner;
476 unsigned int t_prefer_loopback:1,
480 int (*laddr_check)(struct net *net, __be32 addr);
481 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
482 void (*conn_free)(void *data);
483 int (*conn_path_connect)(struct rds_conn_path *cp);
484 void (*conn_path_shutdown)(struct rds_conn_path *conn);
485 void (*xmit_path_prepare)(struct rds_conn_path *cp);
486 void (*xmit_path_complete)(struct rds_conn_path *cp);
487 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
488 unsigned int hdr_off, unsigned int sg, unsigned int off);
489 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
490 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
491 int (*recv_path)(struct rds_conn_path *cp);
492 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
493 void (*inc_free)(struct rds_incoming *inc);
495 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
496 struct rdma_cm_event *event);
497 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
498 void (*cm_connect_complete)(struct rds_connection *conn,
499 struct rdma_cm_event *event);
501 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
504 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
505 struct rds_sock *rs, u32 *key_ret);
506 void (*sync_mr)(void *trans_private, int direction);
507 void (*free_mr)(void *trans_private, int invalidate);
508 void (*flush_mrs)(void);
518 * bound_addr used for both incoming and outgoing, no INADDR_ANY
521 struct rhash_head rs_bound_node;
523 __be32 rs_bound_addr;
525 __be16 rs_bound_port;
527 struct rds_transport *rs_transport;
530 * rds_sendmsg caches the conn it used the last time around.
531 * This helps avoid costly lookups.
533 struct rds_connection *rs_conn;
535 /* flag indicating we were congested or not */
537 /* seen congestion (ENOBUFS) when sending? */
538 int rs_seen_congestion;
540 /* rs_lock protects all these adjacent members before the newline */
542 struct list_head rs_send_queue;
545 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
547 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
548 * to decide whether the application should be woken up.
549 * If not set, we use rs_cong_track to find out whether a cong map
552 uint64_t rs_cong_mask;
553 uint64_t rs_cong_notify;
554 struct list_head rs_cong_list;
555 unsigned long rs_cong_track;
558 * rs_recv_lock protects the receive queue, and is
559 * used to serialize with rds_release.
561 rwlock_t rs_recv_lock;
562 struct list_head rs_recv_queue;
564 /* just for stats reporting */
565 struct list_head rs_item;
567 /* these have their own lock */
568 spinlock_t rs_rdma_lock;
569 struct rb_root rs_rdma_keys;
571 /* Socket options - in case there will be more */
572 unsigned char rs_recverr,
577 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
579 return container_of(sk, struct rds_sock, rs_sk);
581 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
587 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
588 * to account for overhead. We don't account for overhead, we just apply
589 * the number of payload bytes to the specified value.
591 static inline int rds_sk_sndbuf(struct rds_sock *rs)
593 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
595 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
597 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
600 struct rds_statistics {
601 uint64_t s_conn_reset;
602 uint64_t s_recv_drop_bad_checksum;
603 uint64_t s_recv_drop_old_seq;
604 uint64_t s_recv_drop_no_sock;
605 uint64_t s_recv_drop_dead_sock;
606 uint64_t s_recv_deliver_raced;
607 uint64_t s_recv_delivered;
608 uint64_t s_recv_queued;
609 uint64_t s_recv_immediate_retry;
610 uint64_t s_recv_delayed_retry;
611 uint64_t s_recv_ack_required;
612 uint64_t s_recv_rdma_bytes;
613 uint64_t s_recv_ping;
614 uint64_t s_send_queue_empty;
615 uint64_t s_send_queue_full;
616 uint64_t s_send_lock_contention;
617 uint64_t s_send_lock_queue_raced;
618 uint64_t s_send_immediate_retry;
619 uint64_t s_send_delayed_retry;
620 uint64_t s_send_drop_acked;
621 uint64_t s_send_ack_required;
622 uint64_t s_send_queued;
623 uint64_t s_send_rdma;
624 uint64_t s_send_rdma_bytes;
625 uint64_t s_send_pong;
626 uint64_t s_page_remainder_hit;
627 uint64_t s_page_remainder_miss;
628 uint64_t s_copy_to_user;
629 uint64_t s_copy_from_user;
630 uint64_t s_cong_update_queued;
631 uint64_t s_cong_update_received;
632 uint64_t s_cong_send_error;
633 uint64_t s_cong_send_blocked;
637 void rds_sock_addref(struct rds_sock *rs);
638 void rds_sock_put(struct rds_sock *rs);
639 void rds_wake_sk_sleep(struct rds_sock *rs);
640 static inline void __rds_wake_sk_sleep(struct sock *sk)
642 wait_queue_head_t *waitq = sk_sleep(sk);
644 if (!sock_flag(sk, SOCK_DEAD) && waitq)
647 extern wait_queue_head_t rds_poll_waitq;
651 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
652 void rds_remove_bound(struct rds_sock *rs);
653 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
654 int rds_bind_lock_init(void);
655 void rds_bind_lock_destroy(void);
658 int rds_cong_get_maps(struct rds_connection *conn);
659 void rds_cong_add_conn(struct rds_connection *conn);
660 void rds_cong_remove_conn(struct rds_connection *conn);
661 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
662 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
663 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
664 void rds_cong_queue_updates(struct rds_cong_map *map);
665 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
666 int rds_cong_updated_since(unsigned long *recent);
667 void rds_cong_add_socket(struct rds_sock *);
668 void rds_cong_remove_socket(struct rds_sock *);
669 void rds_cong_exit(void);
670 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
673 extern u32 rds_gen_num;
674 int rds_conn_init(void);
675 void rds_conn_exit(void);
676 struct rds_connection *rds_conn_create(struct net *net,
677 __be32 laddr, __be32 faddr,
678 struct rds_transport *trans, gfp_t gfp);
679 struct rds_connection *rds_conn_create_outgoing(struct net *net,
680 __be32 laddr, __be32 faddr,
681 struct rds_transport *trans, gfp_t gfp);
682 void rds_conn_shutdown(struct rds_conn_path *cpath);
683 void rds_conn_destroy(struct rds_connection *conn);
684 void rds_conn_drop(struct rds_connection *conn);
685 void rds_conn_path_drop(struct rds_conn_path *cpath);
686 void rds_conn_connect_if_down(struct rds_connection *conn);
687 void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
688 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
689 struct rds_info_iterator *iter,
690 struct rds_info_lengths *lens,
691 int (*visitor)(struct rds_connection *, void *),
695 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
696 #define rds_conn_path_error(cp, fmt...) \
697 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
700 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
702 return atomic_cmpxchg(&cp->cp_state, old, new) == old;
706 rds_conn_transition(struct rds_connection *conn, int old, int new)
708 WARN_ON(conn->c_trans->t_mp_capable);
709 return rds_conn_path_transition(&conn->c_path[0], old, new);
713 rds_conn_path_state(struct rds_conn_path *cp)
715 return atomic_read(&cp->cp_state);
719 rds_conn_state(struct rds_connection *conn)
721 WARN_ON(conn->c_trans->t_mp_capable);
722 return rds_conn_path_state(&conn->c_path[0]);
726 rds_conn_path_up(struct rds_conn_path *cp)
728 return atomic_read(&cp->cp_state) == RDS_CONN_UP;
732 rds_conn_up(struct rds_connection *conn)
734 WARN_ON(conn->c_trans->t_mp_capable);
735 return rds_conn_path_up(&conn->c_path[0]);
739 rds_conn_path_connecting(struct rds_conn_path *cp)
741 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
745 rds_conn_connecting(struct rds_connection *conn)
747 WARN_ON(conn->c_trans->t_mp_capable);
748 return rds_conn_path_connecting(&conn->c_path[0]);
752 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
753 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
754 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
755 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
756 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
757 __be16 dport, u64 seq);
758 int rds_message_add_extension(struct rds_header *hdr,
759 unsigned int type, const void *data, unsigned int len);
760 int rds_message_next_extension(struct rds_header *hdr,
761 unsigned int *pos, void *buf, unsigned int *buflen);
762 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
763 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
764 void rds_message_inc_free(struct rds_incoming *inc);
765 void rds_message_addref(struct rds_message *rm);
766 void rds_message_put(struct rds_message *rm);
767 void rds_message_wait(struct rds_message *rm);
768 void rds_message_unmapped(struct rds_message *rm);
770 static inline void rds_message_make_checksum(struct rds_header *hdr)
773 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
776 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
778 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
783 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
785 int rds_page_copy_user(struct page *page, unsigned long offset,
786 void __user *ptr, unsigned long bytes,
788 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
789 rds_page_copy_user(page, offset, ptr, bytes, 1)
790 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
791 rds_page_copy_user(page, offset, ptr, bytes, 0)
792 void rds_page_exit(void);
795 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
797 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
799 void rds_inc_put(struct rds_incoming *inc);
800 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
801 struct rds_incoming *inc, gfp_t gfp);
802 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
804 void rds_clear_recv_queue(struct rds_sock *rs);
805 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
806 void rds_inc_info_copy(struct rds_incoming *inc,
807 struct rds_info_iterator *iter,
808 __be32 saddr, __be32 daddr, int flip);
811 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
812 void rds_send_path_reset(struct rds_conn_path *conn);
813 int rds_send_xmit(struct rds_conn_path *cp);
815 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
816 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
817 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
818 is_acked_func is_acked);
819 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
820 is_acked_func is_acked);
821 int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
824 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
825 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
826 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
827 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
828 void rds_rdma_drop_keys(struct rds_sock *rs);
829 int rds_rdma_extra_size(struct rds_rdma_args *args);
830 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
831 struct cmsghdr *cmsg);
832 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
833 struct cmsghdr *cmsg);
834 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
835 struct cmsghdr *cmsg);
836 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
837 struct cmsghdr *cmsg);
838 void rds_rdma_free_op(struct rm_rdma_op *ro);
839 void rds_atomic_free_op(struct rm_atomic_op *ao);
840 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
841 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
842 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
843 struct cmsghdr *cmsg);
845 void __rds_put_mr_final(struct rds_mr *mr);
846 static inline void rds_mr_put(struct rds_mr *mr)
848 if (atomic_dec_and_test(&mr->r_refcount))
849 __rds_put_mr_final(mr);
853 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
854 #define rds_stats_inc_which(which, member) do { \
855 per_cpu(which, get_cpu()).member++; \
858 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
859 #define rds_stats_add_which(which, member, count) do { \
860 per_cpu(which, get_cpu()).member += count; \
863 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
864 int rds_stats_init(void);
865 void rds_stats_exit(void);
866 void rds_stats_info_copy(struct rds_info_iterator *iter,
867 uint64_t *values, const char *const *names,
871 int rds_sysctl_init(void);
872 void rds_sysctl_exit(void);
873 extern unsigned long rds_sysctl_sndbuf_min;
874 extern unsigned long rds_sysctl_sndbuf_default;
875 extern unsigned long rds_sysctl_sndbuf_max;
876 extern unsigned long rds_sysctl_reconnect_min_jiffies;
877 extern unsigned long rds_sysctl_reconnect_max_jiffies;
878 extern unsigned int rds_sysctl_max_unacked_packets;
879 extern unsigned int rds_sysctl_max_unacked_bytes;
880 extern unsigned int rds_sysctl_ping_enable;
881 extern unsigned long rds_sysctl_trace_flags;
882 extern unsigned int rds_sysctl_trace_level;
885 int rds_threads_init(void);
886 void rds_threads_exit(void);
887 extern struct workqueue_struct *rds_wq;
888 void rds_queue_reconnect(struct rds_conn_path *cp);
889 void rds_connect_worker(struct work_struct *);
890 void rds_shutdown_worker(struct work_struct *);
891 void rds_send_worker(struct work_struct *);
892 void rds_recv_worker(struct work_struct *);
893 void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
894 void rds_connect_complete(struct rds_connection *conn);
897 int rds_trans_register(struct rds_transport *trans);
898 void rds_trans_unregister(struct rds_transport *trans);
899 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
900 void rds_trans_put(struct rds_transport *trans);
901 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
903 struct rds_transport *rds_trans_get(int t_type);
904 int rds_trans_init(void);
905 void rds_trans_exit(void);