5 #include <linux/scatterlist.h>
6 #include <linux/highmem.h>
7 #include <rdma/rdma_cm.h>
8 #include <linux/mutex.h>
14 * RDS Network protocol version
16 #define RDS_PROTOCOL_3_0 0x0300
17 #define RDS_PROTOCOL_3_1 0x0301
18 #define RDS_PROTOCOL_VERSION RDS_PROTOCOL_3_1
19 #define RDS_PROTOCOL_MAJOR(v) ((v) >> 8)
20 #define RDS_PROTOCOL_MINOR(v) ((v) & 255)
21 #define RDS_PROTOCOL(maj, min) (((maj) << 8) | min)
24 * XXX randomly chosen, but at least seems to be unused:
25 * # 18464-18768 Unassigned
26 * We should do better. We want a reserved port to discourage unpriv'ed
27 * userspace from listening.
29 #define RDS_PORT 18634
32 #define KERNEL_HAS_ATOMIC64
36 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
38 /* sigh, pr_debug() causes unused variable warnings */
39 static inline void __attribute__ ((format (printf, 1, 2)))
40 rdsdebug(char *fmt, ...)
45 /* XXX is there one of these somewhere? */
47 ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
49 #define RDS_FRAG_SHIFT 12
50 #define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT))
52 #define RDS_CONG_MAP_BYTES (65536 / 8)
53 #define RDS_CONG_MAP_LONGS (RDS_CONG_MAP_BYTES / sizeof(unsigned long))
54 #define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
55 #define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
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,
80 /* Bits for c_flags */
81 #define RDS_LL_SEND_FULL 0
82 #define RDS_RECONNECT_PENDING 1
85 struct rds_connection {
86 struct hlist_node c_hash_node;
89 unsigned int c_loopback:1;
90 struct rds_connection *c_passive;
92 struct rds_cong_map *c_lcong;
93 struct rds_cong_map *c_fcong;
95 struct rds_message *c_xmit_rm;
96 unsigned long c_xmit_sg;
97 unsigned int c_xmit_hdr_off;
98 unsigned int c_xmit_data_off;
99 unsigned int c_xmit_atomic_sent;
100 unsigned int c_xmit_rdma_sent;
101 unsigned int c_xmit_data_sent;
103 spinlock_t c_lock; /* protect msg queues */
105 struct list_head c_send_queue;
106 struct list_head c_retrans;
110 struct rds_transport *c_trans;
111 void *c_transport_data;
114 unsigned long c_flags;
115 unsigned long c_reconnect_jiffies;
116 struct delayed_work c_send_w;
117 struct delayed_work c_recv_w;
118 struct delayed_work c_conn_w;
119 struct work_struct c_down_w;
120 struct mutex c_cm_lock; /* protect conn state & cm */
121 wait_queue_head_t c_waitq;
123 struct list_head c_map_item;
124 unsigned long c_map_queued;
126 unsigned int c_unacked_packets;
127 unsigned int c_unacked_bytes;
129 /* Protocol version */
130 unsigned int c_version;
133 #define RDS_FLAG_CONG_BITMAP 0x01
134 #define RDS_FLAG_ACK_REQUIRED 0x02
135 #define RDS_FLAG_RETRANSMITTED 0x04
136 #define RDS_MAX_ADV_CREDIT 255
139 * Maximum space available for extension headers.
141 #define RDS_HEADER_EXT_SPACE 16
154 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
158 * Reserved - indicates end of extensions
160 #define RDS_EXTHDR_NONE 0
163 * This extension header is included in the very
164 * first message that is sent on a new connection,
165 * and identifies the protocol level. This will help
166 * rolling updates if a future change requires breaking
168 * NB: This is no longer true for IB, where we do a version
169 * negotiation during the connection setup phase (protocol
170 * version information is included in the RDMA CM private data).
172 #define RDS_EXTHDR_VERSION 1
173 struct rds_ext_header_version {
178 * This extension header is included in the RDS message
179 * chasing an RDMA operation.
181 #define RDS_EXTHDR_RDMA 2
182 struct rds_ext_header_rdma {
187 * This extension header tells the peer about the
188 * destination <R_Key,offset> of the requested RDMA
191 #define RDS_EXTHDR_RDMA_DEST 3
192 struct rds_ext_header_rdma_dest {
194 __be32 h_rdma_offset;
197 #define __RDS_EXTHDR_MAX 16 /* for now */
199 struct rds_incoming {
201 struct list_head i_item;
202 struct rds_connection *i_conn;
203 struct rds_header i_hdr;
204 unsigned long i_rx_jiffies;
207 rds_rdma_cookie_t i_rdma_cookie;
211 struct rb_node r_rb_node;
215 /* A copy of the creation flags */
216 unsigned int r_use_once:1;
217 unsigned int r_invalidate:1;
218 unsigned int r_write:1;
220 /* This is for RDS_MR_DEAD.
221 * It would be nice & consistent to make this part of the above
222 * bit field here, but we need to use test_and_set_bit.
224 unsigned long r_state;
225 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
226 struct rds_transport *r_trans;
227 void *r_trans_private;
230 /* Flags for mr->r_state */
231 #define RDS_MR_DEAD 0
233 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
235 return r_key | (((u64) offset) << 32);
238 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
243 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
248 /* atomic operation types */
249 #define RDS_ATOMIC_TYPE_CSWP 0
250 #define RDS_ATOMIC_TYPE_FADD 1
253 * m_sock_item and m_conn_item are on lists that are serialized under
254 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
255 * the message will not be put back on the retransmit list after being sent.
256 * messages that are canceled while being sent rely on this.
258 * m_inc is used by loopback so that it can pass an incoming message straight
259 * back up into the rx path. It embeds a wire header which is also used by
260 * the send path, which is kind of awkward.
262 * m_sock_item indicates the message's presence on a socket's send or receive
263 * queue. m_rs will point to that socket.
265 * m_daddr is used by cancellation to prune messages to a given destination.
267 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
268 * nesting. As paths iterate over messages on a sock, or conn, they must
269 * also lock the conn, or sock, to remove the message from those lists too.
270 * Testing the flag to determine if the message is still on the lists lets
271 * us avoid testing the list_head directly. That means each path can use
272 * the message's list_head to keep it on a local list while juggling locks
273 * without confusing the other path.
275 * m_ack_seq is an optional field set by transports who need a different
276 * sequence number range to invalidate. They can use this in a callback
277 * that they pass to rds_send_drop_acked() to see if each message has been
278 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
279 * had ack_seq set yet.
281 #define RDS_MSG_ON_SOCK 1
282 #define RDS_MSG_ON_CONN 2
283 #define RDS_MSG_HAS_ACK_SEQ 3
284 #define RDS_MSG_ACK_REQUIRED 4
285 #define RDS_MSG_RETRANSMITTED 5
286 #define RDS_MSG_MAPPED 6
287 #define RDS_MSG_PAGEVEC 7
291 struct list_head m_sock_item;
292 struct list_head m_conn_item;
293 struct rds_incoming m_inc;
296 unsigned long m_flags;
298 /* Never access m_rs without holding m_rs_lock.
303 spinlock_t m_rs_lock;
304 wait_queue_head_t m_flush_wait;
306 struct rds_sock *m_rs;
308 /* cookie to send to remote, in rds header */
309 rds_rdma_cookie_t m_rdma_cookie;
311 unsigned int m_used_sgs;
312 unsigned int m_total_sgs;
317 struct rm_atomic_op {
319 uint64_t op_swap_add;
324 unsigned int op_notify:1;
325 unsigned int op_recverr:1;
326 unsigned int op_mapped:1;
327 unsigned int op_silent:1;
328 unsigned int op_active:1;
329 struct scatterlist *op_sg;
330 struct rds_notifier *op_notifier;
332 struct rds_mr *op_rdma_mr;
337 unsigned int op_write:1;
338 unsigned int op_fence:1;
339 unsigned int op_notify:1;
340 unsigned int op_recverr:1;
341 unsigned int op_mapped:1;
342 unsigned int op_silent:1;
343 unsigned int op_active:1;
344 unsigned int op_bytes;
345 unsigned int op_nents;
346 unsigned int op_count;
347 struct scatterlist *op_sg;
348 struct rds_notifier *op_notifier;
350 struct rds_mr *op_rdma_mr;
353 unsigned int op_active:1;
354 unsigned int op_nents;
355 unsigned int op_count;
356 struct scatterlist *op_sg;
362 * The RDS notifier is used (optionally) to tell the application about
363 * completed RDMA operations. Rather than keeping the whole rds message
364 * around on the queue, we allocate a small notifier that is put on the
365 * socket's notifier_list. Notifications are delivered to the application
366 * through control messages.
368 struct rds_notifier {
369 struct list_head n_list;
370 uint64_t n_user_token;
375 * struct rds_transport - transport specific behavioural hooks
377 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
378 * part of a message. The caller serializes on the send_sem so this
379 * doesn't need to be reentrant for a given conn. The header must be
380 * sent before the data payload. .xmit must be prepared to send a
381 * message with no data payload. .xmit should return the number of
382 * bytes that were sent down the connection, including header bytes.
383 * Returning 0 tells the caller that it doesn't need to perform any
384 * additional work now. This is usually the case when the transport has
385 * filled the sending queue for its connection and will handle
386 * triggering the rds thread to continue the send when space becomes
387 * available. Returning -EAGAIN tells the caller to retry the send
388 * immediately. Returning -ENOMEM tells the caller to retry the send at
389 * some point in the future.
391 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
392 * it returns the connection can not call rds_recv_incoming().
393 * This will only be called once after conn_connect returns
394 * non-zero success and will The caller serializes this with
395 * the send and connecting paths (xmit_* and conn_*). The
396 * transport is responsible for other serialization, including
397 * rds_recv_incoming(). This is called in process context but
398 * should try hard not to block.
401 #define RDS_TRANS_IB 0
402 #define RDS_TRANS_IWARP 1
403 #define RDS_TRANS_TCP 2
404 #define RDS_TRANS_COUNT 3
406 struct rds_transport {
407 char t_name[TRANSNAMSIZ];
408 struct list_head t_item;
409 struct module *t_owner;
410 unsigned int t_prefer_loopback:1;
413 int (*laddr_check)(__be32 addr);
414 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
415 void (*conn_free)(void *data);
416 int (*conn_connect)(struct rds_connection *conn);
417 void (*conn_shutdown)(struct rds_connection *conn);
418 void (*xmit_prepare)(struct rds_connection *conn);
419 void (*xmit_complete)(struct rds_connection *conn);
420 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
421 unsigned int hdr_off, unsigned int sg, unsigned int off);
422 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
423 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
424 int (*recv)(struct rds_connection *conn);
425 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
427 void (*inc_free)(struct rds_incoming *inc);
429 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
430 struct rdma_cm_event *event);
431 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
432 void (*cm_connect_complete)(struct rds_connection *conn,
433 struct rdma_cm_event *event);
435 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
438 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
439 struct rds_sock *rs, u32 *key_ret);
440 void (*sync_mr)(void *trans_private, int direction);
441 void (*free_mr)(void *trans_private, int invalidate);
442 void (*flush_mrs)(void);
452 * bound_addr used for both incoming and outgoing, no INADDR_ANY
455 struct hlist_node rs_bound_node;
456 __be32 rs_bound_addr;
458 __be16 rs_bound_port;
460 struct rds_transport *rs_transport;
463 * rds_sendmsg caches the conn it used the last time around.
464 * This helps avoid costly lookups.
466 struct rds_connection *rs_conn;
468 /* flag indicating we were congested or not */
470 /* seen congestion (ENOBUFS) when sending? */
471 int rs_seen_congestion;
473 /* rs_lock protects all these adjacent members before the newline */
475 struct list_head rs_send_queue;
478 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
480 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
481 * to decide whether the application should be woken up.
482 * If not set, we use rs_cong_track to find out whether a cong map
485 uint64_t rs_cong_mask;
486 uint64_t rs_cong_notify;
487 struct list_head rs_cong_list;
488 unsigned long rs_cong_track;
491 * rs_recv_lock protects the receive queue, and is
492 * used to serialize with rds_release.
494 rwlock_t rs_recv_lock;
495 struct list_head rs_recv_queue;
497 /* just for stats reporting */
498 struct list_head rs_item;
500 /* these have their own lock */
501 spinlock_t rs_rdma_lock;
502 struct rb_root rs_rdma_keys;
504 /* Socket options - in case there will be more */
505 unsigned char rs_recverr,
509 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
511 return container_of(sk, struct rds_sock, rs_sk);
513 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
519 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
520 * to account for overhead. We don't account for overhead, we just apply
521 * the number of payload bytes to the specified value.
523 static inline int rds_sk_sndbuf(struct rds_sock *rs)
525 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
527 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
529 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
532 struct rds_statistics {
533 uint64_t s_conn_reset;
534 uint64_t s_recv_drop_bad_checksum;
535 uint64_t s_recv_drop_old_seq;
536 uint64_t s_recv_drop_no_sock;
537 uint64_t s_recv_drop_dead_sock;
538 uint64_t s_recv_deliver_raced;
539 uint64_t s_recv_delivered;
540 uint64_t s_recv_queued;
541 uint64_t s_recv_immediate_retry;
542 uint64_t s_recv_delayed_retry;
543 uint64_t s_recv_ack_required;
544 uint64_t s_recv_rdma_bytes;
545 uint64_t s_recv_ping;
546 uint64_t s_send_queue_empty;
547 uint64_t s_send_queue_full;
548 uint64_t s_send_lock_contention;
549 uint64_t s_send_lock_queue_raced;
550 uint64_t s_send_immediate_retry;
551 uint64_t s_send_delayed_retry;
552 uint64_t s_send_drop_acked;
553 uint64_t s_send_ack_required;
554 uint64_t s_send_queued;
555 uint64_t s_send_rdma;
556 uint64_t s_send_rdma_bytes;
557 uint64_t s_send_pong;
558 uint64_t s_page_remainder_hit;
559 uint64_t s_page_remainder_miss;
560 uint64_t s_copy_to_user;
561 uint64_t s_copy_from_user;
562 uint64_t s_cong_update_queued;
563 uint64_t s_cong_update_received;
564 uint64_t s_cong_send_error;
565 uint64_t s_cong_send_blocked;
569 char *rds_str_array(char **array, size_t elements, size_t index);
570 void rds_sock_addref(struct rds_sock *rs);
571 void rds_sock_put(struct rds_sock *rs);
572 void rds_wake_sk_sleep(struct rds_sock *rs);
573 static inline void __rds_wake_sk_sleep(struct sock *sk)
575 wait_queue_head_t *waitq = sk_sleep(sk);
577 if (!sock_flag(sk, SOCK_DEAD) && waitq)
580 extern wait_queue_head_t rds_poll_waitq;
584 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
585 void rds_remove_bound(struct rds_sock *rs);
586 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
589 int rds_cong_get_maps(struct rds_connection *conn);
590 void rds_cong_add_conn(struct rds_connection *conn);
591 void rds_cong_remove_conn(struct rds_connection *conn);
592 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
593 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
594 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
595 void rds_cong_queue_updates(struct rds_cong_map *map);
596 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
597 int rds_cong_updated_since(unsigned long *recent);
598 void rds_cong_add_socket(struct rds_sock *);
599 void rds_cong_remove_socket(struct rds_sock *);
600 void rds_cong_exit(void);
601 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
604 int rds_conn_init(void);
605 void rds_conn_exit(void);
606 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
607 struct rds_transport *trans, gfp_t gfp);
608 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
609 struct rds_transport *trans, gfp_t gfp);
610 void rds_conn_shutdown(struct rds_connection *conn);
611 void rds_conn_destroy(struct rds_connection *conn);
612 void rds_conn_reset(struct rds_connection *conn);
613 void rds_conn_drop(struct rds_connection *conn);
614 void rds_conn_connect_if_down(struct rds_connection *conn);
615 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
616 struct rds_info_iterator *iter,
617 struct rds_info_lengths *lens,
618 int (*visitor)(struct rds_connection *, void *),
620 void __rds_conn_error(struct rds_connection *conn, const char *, ...)
621 __attribute__ ((format (printf, 2, 3)));
622 #define rds_conn_error(conn, fmt...) \
623 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
626 rds_conn_transition(struct rds_connection *conn, int old, int new)
628 return atomic_cmpxchg(&conn->c_state, old, new) == old;
632 rds_conn_state(struct rds_connection *conn)
634 return atomic_read(&conn->c_state);
638 rds_conn_up(struct rds_connection *conn)
640 return atomic_read(&conn->c_state) == RDS_CONN_UP;
644 rds_conn_connecting(struct rds_connection *conn)
646 return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
650 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
651 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
652 int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
654 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
655 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
656 __be16 dport, u64 seq);
657 int rds_message_add_extension(struct rds_header *hdr,
658 unsigned int type, const void *data, unsigned int len);
659 int rds_message_next_extension(struct rds_header *hdr,
660 unsigned int *pos, void *buf, unsigned int *buflen);
661 int rds_message_add_version_extension(struct rds_header *hdr, unsigned int version);
662 int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *version);
663 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
664 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
665 struct iovec *first_iov, size_t size);
666 void rds_message_inc_free(struct rds_incoming *inc);
667 void rds_message_addref(struct rds_message *rm);
668 void rds_message_put(struct rds_message *rm);
669 void rds_message_wait(struct rds_message *rm);
670 void rds_message_unmapped(struct rds_message *rm);
672 static inline void rds_message_make_checksum(struct rds_header *hdr)
675 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
678 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
680 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
685 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
687 int rds_page_copy_user(struct page *page, unsigned long offset,
688 void __user *ptr, unsigned long bytes,
690 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
691 rds_page_copy_user(page, offset, ptr, bytes, 1)
692 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
693 rds_page_copy_user(page, offset, ptr, bytes, 0)
694 void rds_page_exit(void);
697 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
699 void rds_inc_addref(struct rds_incoming *inc);
700 void rds_inc_put(struct rds_incoming *inc);
701 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
702 struct rds_incoming *inc, gfp_t gfp, enum km_type km);
703 int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
704 size_t size, int msg_flags);
705 void rds_clear_recv_queue(struct rds_sock *rs);
706 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
707 void rds_inc_info_copy(struct rds_incoming *inc,
708 struct rds_info_iterator *iter,
709 __be32 saddr, __be32 daddr, int flip);
712 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
714 void rds_send_reset(struct rds_connection *conn);
715 int rds_send_xmit(struct rds_connection *conn);
717 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
718 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
719 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
720 is_acked_func is_acked);
721 void rds_send_remove_from_sock(struct list_head *messages, int status);
722 int rds_send_pong(struct rds_connection *conn, __be16 dport);
723 struct rds_message *rds_send_get_message(struct rds_connection *,
724 struct rm_rdma_op *);
727 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
728 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
729 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
730 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
731 void rds_rdma_drop_keys(struct rds_sock *rs);
732 int rds_rdma_extra_size(struct rds_rdma_args *args);
733 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
734 struct cmsghdr *cmsg);
735 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
736 struct cmsghdr *cmsg);
737 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
738 struct cmsghdr *cmsg);
739 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
740 struct cmsghdr *cmsg);
741 void rds_rdma_free_op(struct rm_rdma_op *ro);
742 void rds_atomic_free_op(struct rm_atomic_op *ao);
743 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
744 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
745 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
746 struct cmsghdr *cmsg);
748 extern void __rds_put_mr_final(struct rds_mr *mr);
749 static inline void rds_mr_put(struct rds_mr *mr)
751 if (atomic_dec_and_test(&mr->r_refcount))
752 __rds_put_mr_final(mr);
756 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
757 #define rds_stats_inc_which(which, member) do { \
758 per_cpu(which, get_cpu()).member++; \
761 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
762 #define rds_stats_add_which(which, member, count) do { \
763 per_cpu(which, get_cpu()).member += count; \
766 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
767 int rds_stats_init(void);
768 void rds_stats_exit(void);
769 void rds_stats_info_copy(struct rds_info_iterator *iter,
770 uint64_t *values, const char *const *names,
774 int rds_sysctl_init(void);
775 void rds_sysctl_exit(void);
776 extern unsigned long rds_sysctl_sndbuf_min;
777 extern unsigned long rds_sysctl_sndbuf_default;
778 extern unsigned long rds_sysctl_sndbuf_max;
779 extern unsigned long rds_sysctl_reconnect_min_jiffies;
780 extern unsigned long rds_sysctl_reconnect_max_jiffies;
781 extern unsigned int rds_sysctl_max_unacked_packets;
782 extern unsigned int rds_sysctl_max_unacked_bytes;
783 extern unsigned int rds_sysctl_ping_enable;
784 extern unsigned long rds_sysctl_trace_flags;
785 extern unsigned int rds_sysctl_trace_level;
788 int rds_threads_init(void);
789 void rds_threads_exit(void);
790 extern struct workqueue_struct *rds_wq;
791 void rds_queue_reconnect(struct rds_connection *conn);
792 void rds_connect_worker(struct work_struct *);
793 void rds_shutdown_worker(struct work_struct *);
794 void rds_send_worker(struct work_struct *);
795 void rds_recv_worker(struct work_struct *);
796 void rds_connect_complete(struct rds_connection *conn);
799 int rds_trans_register(struct rds_transport *trans);
800 void rds_trans_unregister(struct rds_transport *trans);
801 struct rds_transport *rds_trans_get_preferred(__be32 addr);
802 void rds_trans_put(struct rds_transport *trans);
803 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
805 int rds_trans_init(void);
806 void rds_trans_exit(void);