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 __printf(1, 2)
40 void 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_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
54 #define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
57 struct rb_node m_rb_node
;
59 wait_queue_head_t m_waitq
;
60 struct list_head m_conn_list
;
61 unsigned long m_page_addrs
[RDS_CONG_MAP_PAGES
];
66 * This is how we will track the connection state:
67 * A connection is always in one of the following
68 * states. Updates to the state are atomic and imply
74 RDS_CONN_DISCONNECTING
,
79 /* Bits for c_flags */
80 #define RDS_LL_SEND_FULL 0
81 #define RDS_RECONNECT_PENDING 1
84 struct rds_connection
{
85 struct hlist_node c_hash_node
;
88 unsigned int c_loopback
:1;
89 struct rds_connection
*c_passive
;
91 struct rds_cong_map
*c_lcong
;
92 struct rds_cong_map
*c_fcong
;
94 struct rds_message
*c_xmit_rm
;
95 unsigned long c_xmit_sg
;
96 unsigned int c_xmit_hdr_off
;
97 unsigned int c_xmit_data_off
;
98 unsigned int c_xmit_atomic_sent
;
99 unsigned int c_xmit_rdma_sent
;
100 unsigned int c_xmit_data_sent
;
102 spinlock_t c_lock
; /* protect msg queues */
104 struct list_head c_send_queue
;
105 struct list_head c_retrans
;
109 struct rds_transport
*c_trans
;
110 void *c_transport_data
;
113 unsigned long c_flags
;
114 unsigned long c_reconnect_jiffies
;
115 struct delayed_work c_send_w
;
116 struct delayed_work c_recv_w
;
117 struct delayed_work c_conn_w
;
118 struct work_struct c_down_w
;
119 struct mutex c_cm_lock
; /* protect conn state & cm */
120 wait_queue_head_t c_waitq
;
122 struct list_head c_map_item
;
123 unsigned long c_map_queued
;
125 unsigned int c_unacked_packets
;
126 unsigned int c_unacked_bytes
;
128 /* Protocol version */
129 unsigned int c_version
;
132 #define RDS_FLAG_CONG_BITMAP 0x01
133 #define RDS_FLAG_ACK_REQUIRED 0x02
134 #define RDS_FLAG_RETRANSMITTED 0x04
135 #define RDS_MAX_ADV_CREDIT 255
138 * Maximum space available for extension headers.
140 #define RDS_HEADER_EXT_SPACE 16
153 u8 h_exthdr
[RDS_HEADER_EXT_SPACE
];
157 * Reserved - indicates end of extensions
159 #define RDS_EXTHDR_NONE 0
162 * This extension header is included in the very
163 * first message that is sent on a new connection,
164 * and identifies the protocol level. This will help
165 * rolling updates if a future change requires breaking
167 * NB: This is no longer true for IB, where we do a version
168 * negotiation during the connection setup phase (protocol
169 * version information is included in the RDMA CM private data).
171 #define RDS_EXTHDR_VERSION 1
172 struct rds_ext_header_version
{
177 * This extension header is included in the RDS message
178 * chasing an RDMA operation.
180 #define RDS_EXTHDR_RDMA 2
181 struct rds_ext_header_rdma
{
186 * This extension header tells the peer about the
187 * destination <R_Key,offset> of the requested RDMA
190 #define RDS_EXTHDR_RDMA_DEST 3
191 struct rds_ext_header_rdma_dest
{
193 __be32 h_rdma_offset
;
196 #define __RDS_EXTHDR_MAX 16 /* for now */
198 struct rds_incoming
{
200 struct list_head i_item
;
201 struct rds_connection
*i_conn
;
202 struct rds_header i_hdr
;
203 unsigned long i_rx_jiffies
;
206 rds_rdma_cookie_t i_rdma_cookie
;
210 struct rb_node r_rb_node
;
214 /* A copy of the creation flags */
215 unsigned int r_use_once
:1;
216 unsigned int r_invalidate
:1;
217 unsigned int r_write
:1;
219 /* This is for RDS_MR_DEAD.
220 * It would be nice & consistent to make this part of the above
221 * bit field here, but we need to use test_and_set_bit.
223 unsigned long r_state
;
224 struct rds_sock
*r_sock
; /* back pointer to the socket that owns us */
225 struct rds_transport
*r_trans
;
226 void *r_trans_private
;
229 /* Flags for mr->r_state */
230 #define RDS_MR_DEAD 0
232 static inline rds_rdma_cookie_t
rds_rdma_make_cookie(u32 r_key
, u32 offset
)
234 return r_key
| (((u64
) offset
) << 32);
237 static inline u32
rds_rdma_cookie_key(rds_rdma_cookie_t cookie
)
242 static inline u32
rds_rdma_cookie_offset(rds_rdma_cookie_t cookie
)
247 /* atomic operation types */
248 #define RDS_ATOMIC_TYPE_CSWP 0
249 #define RDS_ATOMIC_TYPE_FADD 1
252 * m_sock_item and m_conn_item are on lists that are serialized under
253 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
254 * the message will not be put back on the retransmit list after being sent.
255 * messages that are canceled while being sent rely on this.
257 * m_inc is used by loopback so that it can pass an incoming message straight
258 * back up into the rx path. It embeds a wire header which is also used by
259 * the send path, which is kind of awkward.
261 * m_sock_item indicates the message's presence on a socket's send or receive
262 * queue. m_rs will point to that socket.
264 * m_daddr is used by cancellation to prune messages to a given destination.
266 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
267 * nesting. As paths iterate over messages on a sock, or conn, they must
268 * also lock the conn, or sock, to remove the message from those lists too.
269 * Testing the flag to determine if the message is still on the lists lets
270 * us avoid testing the list_head directly. That means each path can use
271 * the message's list_head to keep it on a local list while juggling locks
272 * without confusing the other path.
274 * m_ack_seq is an optional field set by transports who need a different
275 * sequence number range to invalidate. They can use this in a callback
276 * that they pass to rds_send_drop_acked() to see if each message has been
277 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
278 * had ack_seq set yet.
280 #define RDS_MSG_ON_SOCK 1
281 #define RDS_MSG_ON_CONN 2
282 #define RDS_MSG_HAS_ACK_SEQ 3
283 #define RDS_MSG_ACK_REQUIRED 4
284 #define RDS_MSG_RETRANSMITTED 5
285 #define RDS_MSG_MAPPED 6
286 #define RDS_MSG_PAGEVEC 7
290 struct list_head m_sock_item
;
291 struct list_head m_conn_item
;
292 struct rds_incoming m_inc
;
295 unsigned long m_flags
;
297 /* Never access m_rs without holding m_rs_lock.
302 spinlock_t m_rs_lock
;
303 wait_queue_head_t m_flush_wait
;
305 struct rds_sock
*m_rs
;
307 /* cookie to send to remote, in rds header */
308 rds_rdma_cookie_t m_rdma_cookie
;
310 unsigned int m_used_sgs
;
311 unsigned int m_total_sgs
;
316 struct rm_atomic_op
{
322 uint64_t compare_mask
;
327 uint64_t nocarry_mask
;
333 unsigned int op_notify
:1;
334 unsigned int op_recverr
:1;
335 unsigned int op_mapped
:1;
336 unsigned int op_silent
:1;
337 unsigned int op_active
:1;
338 struct scatterlist
*op_sg
;
339 struct rds_notifier
*op_notifier
;
341 struct rds_mr
*op_rdma_mr
;
346 unsigned int op_write
:1;
347 unsigned int op_fence
:1;
348 unsigned int op_notify
:1;
349 unsigned int op_recverr
:1;
350 unsigned int op_mapped
:1;
351 unsigned int op_silent
:1;
352 unsigned int op_active
:1;
353 unsigned int op_bytes
;
354 unsigned int op_nents
;
355 unsigned int op_count
;
356 struct scatterlist
*op_sg
;
357 struct rds_notifier
*op_notifier
;
359 struct rds_mr
*op_rdma_mr
;
362 unsigned int op_active
:1;
363 unsigned int op_nents
;
364 unsigned int op_count
;
365 struct scatterlist
*op_sg
;
371 * The RDS notifier is used (optionally) to tell the application about
372 * completed RDMA operations. Rather than keeping the whole rds message
373 * around on the queue, we allocate a small notifier that is put on the
374 * socket's notifier_list. Notifications are delivered to the application
375 * through control messages.
377 struct rds_notifier
{
378 struct list_head n_list
;
379 uint64_t n_user_token
;
384 * struct rds_transport - transport specific behavioural hooks
386 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
387 * part of a message. The caller serializes on the send_sem so this
388 * doesn't need to be reentrant for a given conn. The header must be
389 * sent before the data payload. .xmit must be prepared to send a
390 * message with no data payload. .xmit should return the number of
391 * bytes that were sent down the connection, including header bytes.
392 * Returning 0 tells the caller that it doesn't need to perform any
393 * additional work now. This is usually the case when the transport has
394 * filled the sending queue for its connection and will handle
395 * triggering the rds thread to continue the send when space becomes
396 * available. Returning -EAGAIN tells the caller to retry the send
397 * immediately. Returning -ENOMEM tells the caller to retry the send at
398 * some point in the future.
400 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
401 * it returns the connection can not call rds_recv_incoming().
402 * This will only be called once after conn_connect returns
403 * non-zero success and will The caller serializes this with
404 * the send and connecting paths (xmit_* and conn_*). The
405 * transport is responsible for other serialization, including
406 * rds_recv_incoming(). This is called in process context but
407 * should try hard not to block.
410 #define RDS_TRANS_IB 0
411 #define RDS_TRANS_IWARP 1
412 #define RDS_TRANS_TCP 2
413 #define RDS_TRANS_COUNT 3
415 struct rds_transport
{
416 char t_name
[TRANSNAMSIZ
];
417 struct list_head t_item
;
418 struct module
*t_owner
;
419 unsigned int t_prefer_loopback
:1;
422 int (*laddr_check
)(__be32 addr
);
423 int (*conn_alloc
)(struct rds_connection
*conn
, gfp_t gfp
);
424 void (*conn_free
)(void *data
);
425 int (*conn_connect
)(struct rds_connection
*conn
);
426 void (*conn_shutdown
)(struct rds_connection
*conn
);
427 void (*xmit_prepare
)(struct rds_connection
*conn
);
428 void (*xmit_complete
)(struct rds_connection
*conn
);
429 int (*xmit
)(struct rds_connection
*conn
, struct rds_message
*rm
,
430 unsigned int hdr_off
, unsigned int sg
, unsigned int off
);
431 int (*xmit_rdma
)(struct rds_connection
*conn
, struct rm_rdma_op
*op
);
432 int (*xmit_atomic
)(struct rds_connection
*conn
, struct rm_atomic_op
*op
);
433 int (*recv
)(struct rds_connection
*conn
);
434 int (*inc_copy_to_user
)(struct rds_incoming
*inc
, struct iov_iter
*to
);
435 void (*inc_free
)(struct rds_incoming
*inc
);
437 int (*cm_handle_connect
)(struct rdma_cm_id
*cm_id
,
438 struct rdma_cm_event
*event
);
439 int (*cm_initiate_connect
)(struct rdma_cm_id
*cm_id
);
440 void (*cm_connect_complete
)(struct rds_connection
*conn
,
441 struct rdma_cm_event
*event
);
443 unsigned int (*stats_info_copy
)(struct rds_info_iterator
*iter
,
446 void *(*get_mr
)(struct scatterlist
*sg
, unsigned long nr_sg
,
447 struct rds_sock
*rs
, u32
*key_ret
);
448 void (*sync_mr
)(void *trans_private
, int direction
);
449 void (*free_mr
)(void *trans_private
, int invalidate
);
450 void (*flush_mrs
)(void);
460 * bound_addr used for both incoming and outgoing, no INADDR_ANY
463 struct hlist_node rs_bound_node
;
464 __be32 rs_bound_addr
;
466 __be16 rs_bound_port
;
468 struct rds_transport
*rs_transport
;
471 * rds_sendmsg caches the conn it used the last time around.
472 * This helps avoid costly lookups.
474 struct rds_connection
*rs_conn
;
476 /* flag indicating we were congested or not */
478 /* seen congestion (ENOBUFS) when sending? */
479 int rs_seen_congestion
;
481 /* rs_lock protects all these adjacent members before the newline */
483 struct list_head rs_send_queue
;
486 struct list_head rs_notify_queue
; /* currently used for failed RDMAs */
488 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
489 * to decide whether the application should be woken up.
490 * If not set, we use rs_cong_track to find out whether a cong map
493 uint64_t rs_cong_mask
;
494 uint64_t rs_cong_notify
;
495 struct list_head rs_cong_list
;
496 unsigned long rs_cong_track
;
499 * rs_recv_lock protects the receive queue, and is
500 * used to serialize with rds_release.
502 rwlock_t rs_recv_lock
;
503 struct list_head rs_recv_queue
;
505 /* just for stats reporting */
506 struct list_head rs_item
;
508 /* these have their own lock */
509 spinlock_t rs_rdma_lock
;
510 struct rb_root rs_rdma_keys
;
512 /* Socket options - in case there will be more */
513 unsigned char rs_recverr
,
517 static inline struct rds_sock
*rds_sk_to_rs(const struct sock
*sk
)
519 return container_of(sk
, struct rds_sock
, rs_sk
);
521 static inline struct sock
*rds_rs_to_sk(struct rds_sock
*rs
)
527 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
528 * to account for overhead. We don't account for overhead, we just apply
529 * the number of payload bytes to the specified value.
531 static inline int rds_sk_sndbuf(struct rds_sock
*rs
)
533 return rds_rs_to_sk(rs
)->sk_sndbuf
/ 2;
535 static inline int rds_sk_rcvbuf(struct rds_sock
*rs
)
537 return rds_rs_to_sk(rs
)->sk_rcvbuf
/ 2;
540 struct rds_statistics
{
541 uint64_t s_conn_reset
;
542 uint64_t s_recv_drop_bad_checksum
;
543 uint64_t s_recv_drop_old_seq
;
544 uint64_t s_recv_drop_no_sock
;
545 uint64_t s_recv_drop_dead_sock
;
546 uint64_t s_recv_deliver_raced
;
547 uint64_t s_recv_delivered
;
548 uint64_t s_recv_queued
;
549 uint64_t s_recv_immediate_retry
;
550 uint64_t s_recv_delayed_retry
;
551 uint64_t s_recv_ack_required
;
552 uint64_t s_recv_rdma_bytes
;
553 uint64_t s_recv_ping
;
554 uint64_t s_send_queue_empty
;
555 uint64_t s_send_queue_full
;
556 uint64_t s_send_lock_contention
;
557 uint64_t s_send_lock_queue_raced
;
558 uint64_t s_send_immediate_retry
;
559 uint64_t s_send_delayed_retry
;
560 uint64_t s_send_drop_acked
;
561 uint64_t s_send_ack_required
;
562 uint64_t s_send_queued
;
563 uint64_t s_send_rdma
;
564 uint64_t s_send_rdma_bytes
;
565 uint64_t s_send_pong
;
566 uint64_t s_page_remainder_hit
;
567 uint64_t s_page_remainder_miss
;
568 uint64_t s_copy_to_user
;
569 uint64_t s_copy_from_user
;
570 uint64_t s_cong_update_queued
;
571 uint64_t s_cong_update_received
;
572 uint64_t s_cong_send_error
;
573 uint64_t s_cong_send_blocked
;
577 char *rds_str_array(char **array
, size_t elements
, size_t index
);
578 void rds_sock_addref(struct rds_sock
*rs
);
579 void rds_sock_put(struct rds_sock
*rs
);
580 void rds_wake_sk_sleep(struct rds_sock
*rs
);
581 static inline void __rds_wake_sk_sleep(struct sock
*sk
)
583 wait_queue_head_t
*waitq
= sk_sleep(sk
);
585 if (!sock_flag(sk
, SOCK_DEAD
) && waitq
)
588 extern wait_queue_head_t rds_poll_waitq
;
592 int rds_bind(struct socket
*sock
, struct sockaddr
*uaddr
, int addr_len
);
593 void rds_remove_bound(struct rds_sock
*rs
);
594 struct rds_sock
*rds_find_bound(__be32 addr
, __be16 port
);
597 int rds_cong_get_maps(struct rds_connection
*conn
);
598 void rds_cong_add_conn(struct rds_connection
*conn
);
599 void rds_cong_remove_conn(struct rds_connection
*conn
);
600 void rds_cong_set_bit(struct rds_cong_map
*map
, __be16 port
);
601 void rds_cong_clear_bit(struct rds_cong_map
*map
, __be16 port
);
602 int rds_cong_wait(struct rds_cong_map
*map
, __be16 port
, int nonblock
, struct rds_sock
*rs
);
603 void rds_cong_queue_updates(struct rds_cong_map
*map
);
604 void rds_cong_map_updated(struct rds_cong_map
*map
, uint64_t);
605 int rds_cong_updated_since(unsigned long *recent
);
606 void rds_cong_add_socket(struct rds_sock
*);
607 void rds_cong_remove_socket(struct rds_sock
*);
608 void rds_cong_exit(void);
609 struct rds_message
*rds_cong_update_alloc(struct rds_connection
*conn
);
612 int rds_conn_init(void);
613 void rds_conn_exit(void);
614 struct rds_connection
*rds_conn_create(__be32 laddr
, __be32 faddr
,
615 struct rds_transport
*trans
, gfp_t gfp
);
616 struct rds_connection
*rds_conn_create_outgoing(__be32 laddr
, __be32 faddr
,
617 struct rds_transport
*trans
, gfp_t gfp
);
618 void rds_conn_shutdown(struct rds_connection
*conn
);
619 void rds_conn_destroy(struct rds_connection
*conn
);
620 void rds_conn_drop(struct rds_connection
*conn
);
621 void rds_conn_connect_if_down(struct rds_connection
*conn
);
622 void rds_for_each_conn_info(struct socket
*sock
, unsigned int len
,
623 struct rds_info_iterator
*iter
,
624 struct rds_info_lengths
*lens
,
625 int (*visitor
)(struct rds_connection
*, void *),
628 void __rds_conn_error(struct rds_connection
*conn
, const char *, ...);
629 #define rds_conn_error(conn, fmt...) \
630 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
633 rds_conn_transition(struct rds_connection
*conn
, int old
, int new)
635 return atomic_cmpxchg(&conn
->c_state
, old
, new) == old
;
639 rds_conn_state(struct rds_connection
*conn
)
641 return atomic_read(&conn
->c_state
);
645 rds_conn_up(struct rds_connection
*conn
)
647 return atomic_read(&conn
->c_state
) == RDS_CONN_UP
;
651 rds_conn_connecting(struct rds_connection
*conn
)
653 return atomic_read(&conn
->c_state
) == RDS_CONN_CONNECTING
;
657 struct rds_message
*rds_message_alloc(unsigned int nents
, gfp_t gfp
);
658 struct scatterlist
*rds_message_alloc_sgs(struct rds_message
*rm
, int nents
);
659 int rds_message_copy_from_user(struct rds_message
*rm
, struct iov_iter
*from
);
660 struct rds_message
*rds_message_map_pages(unsigned long *page_addrs
, unsigned int total_len
);
661 void rds_message_populate_header(struct rds_header
*hdr
, __be16 sport
,
662 __be16 dport
, u64 seq
);
663 int rds_message_add_extension(struct rds_header
*hdr
,
664 unsigned int type
, const void *data
, unsigned int len
);
665 int rds_message_next_extension(struct rds_header
*hdr
,
666 unsigned int *pos
, void *buf
, unsigned int *buflen
);
667 int rds_message_add_rdma_dest_extension(struct rds_header
*hdr
, u32 r_key
, u32 offset
);
668 int rds_message_inc_copy_to_user(struct rds_incoming
*inc
, struct iov_iter
*to
);
669 void rds_message_inc_free(struct rds_incoming
*inc
);
670 void rds_message_addref(struct rds_message
*rm
);
671 void rds_message_put(struct rds_message
*rm
);
672 void rds_message_wait(struct rds_message
*rm
);
673 void rds_message_unmapped(struct rds_message
*rm
);
675 static inline void rds_message_make_checksum(struct rds_header
*hdr
)
678 hdr
->h_csum
= ip_fast_csum((void *) hdr
, sizeof(*hdr
) >> 2);
681 static inline int rds_message_verify_checksum(const struct rds_header
*hdr
)
683 return !hdr
->h_csum
|| ip_fast_csum((void *) hdr
, sizeof(*hdr
) >> 2) == 0;
688 int rds_page_remainder_alloc(struct scatterlist
*scat
, unsigned long bytes
,
690 int rds_page_copy_user(struct page
*page
, unsigned long offset
,
691 void __user
*ptr
, unsigned long bytes
,
693 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
694 rds_page_copy_user(page, offset, ptr, bytes, 1)
695 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
696 rds_page_copy_user(page, offset, ptr, bytes, 0)
697 void rds_page_exit(void);
700 void rds_inc_init(struct rds_incoming
*inc
, struct rds_connection
*conn
,
702 void rds_inc_put(struct rds_incoming
*inc
);
703 void rds_recv_incoming(struct rds_connection
*conn
, __be32 saddr
, __be32 daddr
,
704 struct rds_incoming
*inc
, gfp_t gfp
);
705 int rds_recvmsg(struct kiocb
*iocb
, struct socket
*sock
, struct msghdr
*msg
,
706 size_t size
, int msg_flags
);
707 void rds_clear_recv_queue(struct rds_sock
*rs
);
708 int rds_notify_queue_get(struct rds_sock
*rs
, struct msghdr
*msg
);
709 void rds_inc_info_copy(struct rds_incoming
*inc
,
710 struct rds_info_iterator
*iter
,
711 __be32 saddr
, __be32 daddr
, int flip
);
714 int rds_sendmsg(struct kiocb
*iocb
, struct socket
*sock
, struct msghdr
*msg
,
716 void rds_send_reset(struct rds_connection
*conn
);
717 int rds_send_xmit(struct rds_connection
*conn
);
719 void rds_send_drop_to(struct rds_sock
*rs
, struct sockaddr_in
*dest
);
720 typedef int (*is_acked_func
)(struct rds_message
*rm
, uint64_t ack
);
721 void rds_send_drop_acked(struct rds_connection
*conn
, u64 ack
,
722 is_acked_func is_acked
);
723 int rds_send_pong(struct rds_connection
*conn
, __be16 dport
);
724 struct rds_message
*rds_send_get_message(struct rds_connection
*,
725 struct rm_rdma_op
*);
728 void rds_rdma_unuse(struct rds_sock
*rs
, u32 r_key
, int force
);
729 int rds_get_mr(struct rds_sock
*rs
, char __user
*optval
, int optlen
);
730 int rds_get_mr_for_dest(struct rds_sock
*rs
, char __user
*optval
, int optlen
);
731 int rds_free_mr(struct rds_sock
*rs
, char __user
*optval
, int optlen
);
732 void rds_rdma_drop_keys(struct rds_sock
*rs
);
733 int rds_rdma_extra_size(struct rds_rdma_args
*args
);
734 int rds_cmsg_rdma_args(struct rds_sock
*rs
, struct rds_message
*rm
,
735 struct cmsghdr
*cmsg
);
736 int rds_cmsg_rdma_dest(struct rds_sock
*rs
, struct rds_message
*rm
,
737 struct cmsghdr
*cmsg
);
738 int rds_cmsg_rdma_args(struct rds_sock
*rs
, struct rds_message
*rm
,
739 struct cmsghdr
*cmsg
);
740 int rds_cmsg_rdma_map(struct rds_sock
*rs
, struct rds_message
*rm
,
741 struct cmsghdr
*cmsg
);
742 void rds_rdma_free_op(struct rm_rdma_op
*ro
);
743 void rds_atomic_free_op(struct rm_atomic_op
*ao
);
744 void rds_rdma_send_complete(struct rds_message
*rm
, int wc_status
);
745 void rds_atomic_send_complete(struct rds_message
*rm
, int wc_status
);
746 int rds_cmsg_atomic(struct rds_sock
*rs
, struct rds_message
*rm
,
747 struct cmsghdr
*cmsg
);
749 void __rds_put_mr_final(struct rds_mr
*mr
);
750 static inline void rds_mr_put(struct rds_mr
*mr
)
752 if (atomic_dec_and_test(&mr
->r_refcount
))
753 __rds_put_mr_final(mr
);
757 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics
, rds_stats
);
758 #define rds_stats_inc_which(which, member) do { \
759 per_cpu(which, get_cpu()).member++; \
762 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
763 #define rds_stats_add_which(which, member, count) do { \
764 per_cpu(which, get_cpu()).member += count; \
767 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
768 int rds_stats_init(void);
769 void rds_stats_exit(void);
770 void rds_stats_info_copy(struct rds_info_iterator
*iter
,
771 uint64_t *values
, const char *const *names
,
775 int rds_sysctl_init(void);
776 void rds_sysctl_exit(void);
777 extern unsigned long rds_sysctl_sndbuf_min
;
778 extern unsigned long rds_sysctl_sndbuf_default
;
779 extern unsigned long rds_sysctl_sndbuf_max
;
780 extern unsigned long rds_sysctl_reconnect_min_jiffies
;
781 extern unsigned long rds_sysctl_reconnect_max_jiffies
;
782 extern unsigned int rds_sysctl_max_unacked_packets
;
783 extern unsigned int rds_sysctl_max_unacked_bytes
;
784 extern unsigned int rds_sysctl_ping_enable
;
785 extern unsigned long rds_sysctl_trace_flags
;
786 extern unsigned int rds_sysctl_trace_level
;
789 int rds_threads_init(void);
790 void rds_threads_exit(void);
791 extern struct workqueue_struct
*rds_wq
;
792 void rds_queue_reconnect(struct rds_connection
*conn
);
793 void rds_connect_worker(struct work_struct
*);
794 void rds_shutdown_worker(struct work_struct
*);
795 void rds_send_worker(struct work_struct
*);
796 void rds_recv_worker(struct work_struct
*);
797 void rds_connect_complete(struct rds_connection
*conn
);
800 int rds_trans_register(struct rds_transport
*trans
);
801 void rds_trans_unregister(struct rds_transport
*trans
);
802 struct rds_transport
*rds_trans_get_preferred(__be32 addr
);
803 void rds_trans_put(struct rds_transport
*trans
);
804 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator
*iter
,
806 int rds_trans_init(void);
807 void rds_trans_exit(void);