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 /* Used to limit both RDMA and non-RDMA RDS message to 1MB */
54 #define RDS_MAX_MSG_SIZE ((unsigned int)(1 << 20))
56 #define RDS_CONG_MAP_BYTES (65536 / 8)
57 #define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
58 #define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
61 struct rb_node m_rb_node;
63 wait_queue_head_t m_waitq;
64 struct list_head m_conn_list;
65 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES];
70 * This is how we will track the connection state:
71 * A connection is always in one of the following
72 * states. Updates to the state are atomic and imply
78 RDS_CONN_DISCONNECTING,
84 /* Bits for c_flags */
85 #define RDS_LL_SEND_FULL 0
86 #define RDS_RECONNECT_PENDING 1
88 #define RDS_RECV_REFILL 3
90 /* Max number of multipaths per RDS connection. Must be a power of 2 */
91 #define RDS_MPATH_WORKERS 8
92 #define RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
93 (rs)->rs_hash_initval) & ((n) - 1))
95 /* Per mpath connection state */
96 struct rds_conn_path {
97 struct rds_connection *cp_conn;
98 struct rds_message *cp_xmit_rm;
99 unsigned long cp_xmit_sg;
100 unsigned int cp_xmit_hdr_off;
101 unsigned int cp_xmit_data_off;
102 unsigned int cp_xmit_atomic_sent;
103 unsigned int cp_xmit_rdma_sent;
104 unsigned int cp_xmit_data_sent;
106 spinlock_t cp_lock; /* protect msg queues */
108 struct list_head cp_send_queue;
109 struct list_head cp_retrans;
113 void *cp_transport_data;
116 unsigned long cp_send_gen;
117 unsigned long cp_flags;
118 unsigned long cp_reconnect_jiffies;
119 struct delayed_work cp_send_w;
120 struct delayed_work cp_recv_w;
121 struct delayed_work cp_conn_w;
122 struct work_struct cp_down_w;
123 struct mutex cp_cm_lock; /* protect cp_state & cm */
124 wait_queue_head_t cp_waitq;
126 unsigned int cp_unacked_packets;
127 unsigned int cp_unacked_bytes;
128 unsigned int cp_outgoing:1,
130 unsigned int cp_index;
133 /* One rds_connection per RDS address pair */
134 struct rds_connection {
135 struct hlist_node c_hash_node;
138 unsigned int c_loopback:1,
142 struct rds_connection *c_passive;
143 struct rds_transport *c_trans;
145 struct rds_cong_map *c_lcong;
146 struct rds_cong_map *c_fcong;
148 /* Protocol version */
149 unsigned int c_version;
150 possible_net_t c_net;
152 struct list_head c_map_item;
153 unsigned long c_map_queued;
155 struct rds_conn_path c_path[RDS_MPATH_WORKERS];
156 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 4
251 #define __RDS_EXTHDR_MAX 16 /* for now */
253 struct rds_incoming {
255 struct list_head i_item;
256 struct rds_connection *i_conn;
257 struct rds_conn_path *i_conn_path;
258 struct rds_header i_hdr;
259 unsigned long i_rx_jiffies;
262 rds_rdma_cookie_t i_rdma_cookie;
263 struct timeval i_rx_tstamp;
267 struct rb_node r_rb_node;
271 /* A copy of the creation flags */
272 unsigned int r_use_once:1;
273 unsigned int r_invalidate:1;
274 unsigned int r_write:1;
276 /* This is for RDS_MR_DEAD.
277 * It would be nice & consistent to make this part of the above
278 * bit field here, but we need to use test_and_set_bit.
280 unsigned long r_state;
281 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
282 struct rds_transport *r_trans;
283 void *r_trans_private;
286 /* Flags for mr->r_state */
287 #define RDS_MR_DEAD 0
289 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
291 return r_key | (((u64) offset) << 32);
294 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
299 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
304 /* atomic operation types */
305 #define RDS_ATOMIC_TYPE_CSWP 0
306 #define RDS_ATOMIC_TYPE_FADD 1
309 * m_sock_item and m_conn_item are on lists that are serialized under
310 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
311 * the message will not be put back on the retransmit list after being sent.
312 * messages that are canceled while being sent rely on this.
314 * m_inc is used by loopback so that it can pass an incoming message straight
315 * back up into the rx path. It embeds a wire header which is also used by
316 * the send path, which is kind of awkward.
318 * m_sock_item indicates the message's presence on a socket's send or receive
319 * queue. m_rs will point to that socket.
321 * m_daddr is used by cancellation to prune messages to a given destination.
323 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
324 * nesting. As paths iterate over messages on a sock, or conn, they must
325 * also lock the conn, or sock, to remove the message from those lists too.
326 * Testing the flag to determine if the message is still on the lists lets
327 * us avoid testing the list_head directly. That means each path can use
328 * the message's list_head to keep it on a local list while juggling locks
329 * without confusing the other path.
331 * m_ack_seq is an optional field set by transports who need a different
332 * sequence number range to invalidate. They can use this in a callback
333 * that they pass to rds_send_drop_acked() to see if each message has been
334 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
335 * had ack_seq set yet.
337 #define RDS_MSG_ON_SOCK 1
338 #define RDS_MSG_ON_CONN 2
339 #define RDS_MSG_HAS_ACK_SEQ 3
340 #define RDS_MSG_ACK_REQUIRED 4
341 #define RDS_MSG_RETRANSMITTED 5
342 #define RDS_MSG_MAPPED 6
343 #define RDS_MSG_PAGEVEC 7
347 struct list_head m_sock_item;
348 struct list_head m_conn_item;
349 struct rds_incoming m_inc;
352 unsigned long m_flags;
354 /* Never access m_rs without holding m_rs_lock.
359 spinlock_t m_rs_lock;
360 wait_queue_head_t m_flush_wait;
362 struct rds_sock *m_rs;
364 /* cookie to send to remote, in rds header */
365 rds_rdma_cookie_t m_rdma_cookie;
367 unsigned int m_used_sgs;
368 unsigned int m_total_sgs;
373 struct rm_atomic_op {
379 uint64_t compare_mask;
384 uint64_t nocarry_mask;
390 unsigned int op_notify:1;
391 unsigned int op_recverr:1;
392 unsigned int op_mapped:1;
393 unsigned int op_silent:1;
394 unsigned int op_active:1;
395 struct scatterlist *op_sg;
396 struct rds_notifier *op_notifier;
398 struct rds_mr *op_rdma_mr;
403 unsigned int op_write:1;
404 unsigned int op_fence:1;
405 unsigned int op_notify:1;
406 unsigned int op_recverr:1;
407 unsigned int op_mapped:1;
408 unsigned int op_silent:1;
409 unsigned int op_active:1;
410 unsigned int op_bytes;
411 unsigned int op_nents;
412 unsigned int op_count;
413 struct scatterlist *op_sg;
414 struct rds_notifier *op_notifier;
416 struct rds_mr *op_rdma_mr;
419 unsigned int op_active:1;
420 unsigned int op_notify:1;
421 unsigned int op_nents;
422 unsigned int op_count;
423 unsigned int op_dmasg;
424 unsigned int op_dmaoff;
425 struct scatterlist *op_sg;
431 * The RDS notifier is used (optionally) to tell the application about
432 * completed RDMA operations. Rather than keeping the whole rds message
433 * around on the queue, we allocate a small notifier that is put on the
434 * socket's notifier_list. Notifications are delivered to the application
435 * through control messages.
437 struct rds_notifier {
438 struct list_head n_list;
439 uint64_t n_user_token;
443 /* Available as part of RDS core, so doesn't need to participate
444 * in get_preferred transport etc
446 #define RDS_TRANS_LOOP 3
449 * struct rds_transport - transport specific behavioural hooks
451 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
452 * part of a message. The caller serializes on the send_sem so this
453 * doesn't need to be reentrant for a given conn. The header must be
454 * sent before the data payload. .xmit must be prepared to send a
455 * message with no data payload. .xmit should return the number of
456 * bytes that were sent down the connection, including header bytes.
457 * Returning 0 tells the caller that it doesn't need to perform any
458 * additional work now. This is usually the case when the transport has
459 * filled the sending queue for its connection and will handle
460 * triggering the rds thread to continue the send when space becomes
461 * available. Returning -EAGAIN tells the caller to retry the send
462 * immediately. Returning -ENOMEM tells the caller to retry the send at
463 * some point in the future.
465 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
466 * it returns the connection can not call rds_recv_incoming().
467 * This will only be called once after conn_connect returns
468 * non-zero success and will The caller serializes this with
469 * the send and connecting paths (xmit_* and conn_*). The
470 * transport is responsible for other serialization, including
471 * rds_recv_incoming(). This is called in process context but
472 * should try hard not to block.
475 struct rds_transport {
476 char t_name[TRANSNAMSIZ];
477 struct list_head t_item;
478 struct module *t_owner;
479 unsigned int t_prefer_loopback:1,
483 int (*laddr_check)(struct net *net, __be32 addr);
484 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
485 void (*conn_free)(void *data);
486 int (*conn_path_connect)(struct rds_conn_path *cp);
487 void (*conn_path_shutdown)(struct rds_conn_path *conn);
488 void (*xmit_path_prepare)(struct rds_conn_path *cp);
489 void (*xmit_path_complete)(struct rds_conn_path *cp);
490 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
491 unsigned int hdr_off, unsigned int sg, unsigned int off);
492 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
493 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
494 int (*recv_path)(struct rds_conn_path *cp);
495 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
496 void (*inc_free)(struct rds_incoming *inc);
498 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
499 struct rdma_cm_event *event);
500 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
501 void (*cm_connect_complete)(struct rds_connection *conn,
502 struct rdma_cm_event *event);
504 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
507 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
508 struct rds_sock *rs, u32 *key_ret);
509 void (*sync_mr)(void *trans_private, int direction);
510 void (*free_mr)(void *trans_private, int invalidate);
511 void (*flush_mrs)(void);
521 * bound_addr used for both incoming and outgoing, no INADDR_ANY
524 struct rhash_head rs_bound_node;
526 __be32 rs_bound_addr;
528 __be16 rs_bound_port;
530 struct rds_transport *rs_transport;
533 * rds_sendmsg caches the conn it used the last time around.
534 * This helps avoid costly lookups.
536 struct rds_connection *rs_conn;
538 /* flag indicating we were congested or not */
540 /* seen congestion (ENOBUFS) when sending? */
541 int rs_seen_congestion;
543 /* rs_lock protects all these adjacent members before the newline */
545 struct list_head rs_send_queue;
548 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
550 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
551 * to decide whether the application should be woken up.
552 * If not set, we use rs_cong_track to find out whether a cong map
555 uint64_t rs_cong_mask;
556 uint64_t rs_cong_notify;
557 struct list_head rs_cong_list;
558 unsigned long rs_cong_track;
561 * rs_recv_lock protects the receive queue, and is
562 * used to serialize with rds_release.
564 rwlock_t rs_recv_lock;
565 struct list_head rs_recv_queue;
567 /* just for stats reporting */
568 struct list_head rs_item;
570 /* these have their own lock */
571 spinlock_t rs_rdma_lock;
572 struct rb_root rs_rdma_keys;
574 /* Socket options - in case there will be more */
575 unsigned char rs_recverr,
580 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
582 return container_of(sk, struct rds_sock, rs_sk);
584 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
590 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
591 * to account for overhead. We don't account for overhead, we just apply
592 * the number of payload bytes to the specified value.
594 static inline int rds_sk_sndbuf(struct rds_sock *rs)
596 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
598 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
600 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
603 struct rds_statistics {
604 uint64_t s_conn_reset;
605 uint64_t s_recv_drop_bad_checksum;
606 uint64_t s_recv_drop_old_seq;
607 uint64_t s_recv_drop_no_sock;
608 uint64_t s_recv_drop_dead_sock;
609 uint64_t s_recv_deliver_raced;
610 uint64_t s_recv_delivered;
611 uint64_t s_recv_queued;
612 uint64_t s_recv_immediate_retry;
613 uint64_t s_recv_delayed_retry;
614 uint64_t s_recv_ack_required;
615 uint64_t s_recv_rdma_bytes;
616 uint64_t s_recv_ping;
617 uint64_t s_send_queue_empty;
618 uint64_t s_send_queue_full;
619 uint64_t s_send_lock_contention;
620 uint64_t s_send_lock_queue_raced;
621 uint64_t s_send_immediate_retry;
622 uint64_t s_send_delayed_retry;
623 uint64_t s_send_drop_acked;
624 uint64_t s_send_ack_required;
625 uint64_t s_send_queued;
626 uint64_t s_send_rdma;
627 uint64_t s_send_rdma_bytes;
628 uint64_t s_send_pong;
629 uint64_t s_page_remainder_hit;
630 uint64_t s_page_remainder_miss;
631 uint64_t s_copy_to_user;
632 uint64_t s_copy_from_user;
633 uint64_t s_cong_update_queued;
634 uint64_t s_cong_update_received;
635 uint64_t s_cong_send_error;
636 uint64_t s_cong_send_blocked;
640 void rds_sock_addref(struct rds_sock *rs);
641 void rds_sock_put(struct rds_sock *rs);
642 void rds_wake_sk_sleep(struct rds_sock *rs);
643 static inline void __rds_wake_sk_sleep(struct sock *sk)
645 wait_queue_head_t *waitq = sk_sleep(sk);
647 if (!sock_flag(sk, SOCK_DEAD) && waitq)
650 extern wait_queue_head_t rds_poll_waitq;
654 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
655 void rds_remove_bound(struct rds_sock *rs);
656 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
657 int rds_bind_lock_init(void);
658 void rds_bind_lock_destroy(void);
661 int rds_cong_get_maps(struct rds_connection *conn);
662 void rds_cong_add_conn(struct rds_connection *conn);
663 void rds_cong_remove_conn(struct rds_connection *conn);
664 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
665 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
666 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
667 void rds_cong_queue_updates(struct rds_cong_map *map);
668 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
669 int rds_cong_updated_since(unsigned long *recent);
670 void rds_cong_add_socket(struct rds_sock *);
671 void rds_cong_remove_socket(struct rds_sock *);
672 void rds_cong_exit(void);
673 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
676 int rds_conn_init(void);
677 void rds_conn_exit(void);
678 struct rds_connection *rds_conn_create(struct net *net,
679 __be32 laddr, __be32 faddr,
680 struct rds_transport *trans, gfp_t gfp);
681 struct rds_connection *rds_conn_create_outgoing(struct net *net,
682 __be32 laddr, __be32 faddr,
683 struct rds_transport *trans, gfp_t gfp);
684 void rds_conn_shutdown(struct rds_conn_path *cpath);
685 void rds_conn_destroy(struct rds_connection *conn);
686 void rds_conn_drop(struct rds_connection *conn);
687 void rds_conn_path_drop(struct rds_conn_path *cpath);
688 void rds_conn_connect_if_down(struct rds_connection *conn);
689 void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
690 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
691 struct rds_info_iterator *iter,
692 struct rds_info_lengths *lens,
693 int (*visitor)(struct rds_connection *, void *),
696 void __rds_conn_error(struct rds_connection *conn, const char *, ...);
697 #define rds_conn_error(conn, fmt...) \
698 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
701 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
702 #define rds_conn_path_error(cp, fmt...) \
703 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
706 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
708 return atomic_cmpxchg(&cp->cp_state, old, new) == old;
712 rds_conn_transition(struct rds_connection *conn, int old, int new)
714 WARN_ON(conn->c_trans->t_mp_capable);
715 return rds_conn_path_transition(&conn->c_path[0], old, new);
719 rds_conn_path_state(struct rds_conn_path *cp)
721 return atomic_read(&cp->cp_state);
725 rds_conn_state(struct rds_connection *conn)
727 WARN_ON(conn->c_trans->t_mp_capable);
728 return rds_conn_path_state(&conn->c_path[0]);
732 rds_conn_path_up(struct rds_conn_path *cp)
734 return atomic_read(&cp->cp_state) == RDS_CONN_UP;
738 rds_conn_up(struct rds_connection *conn)
740 WARN_ON(conn->c_trans->t_mp_capable);
741 return rds_conn_path_up(&conn->c_path[0]);
745 rds_conn_path_connecting(struct rds_conn_path *cp)
747 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
751 rds_conn_connecting(struct rds_connection *conn)
753 WARN_ON(conn->c_trans->t_mp_capable);
754 return rds_conn_path_connecting(&conn->c_path[0]);
758 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
759 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
760 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
761 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
762 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
763 __be16 dport, u64 seq);
764 int rds_message_add_extension(struct rds_header *hdr,
765 unsigned int type, const void *data, unsigned int len);
766 int rds_message_next_extension(struct rds_header *hdr,
767 unsigned int *pos, void *buf, unsigned int *buflen);
768 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
769 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
770 void rds_message_inc_free(struct rds_incoming *inc);
771 void rds_message_addref(struct rds_message *rm);
772 void rds_message_put(struct rds_message *rm);
773 void rds_message_wait(struct rds_message *rm);
774 void rds_message_unmapped(struct rds_message *rm);
776 static inline void rds_message_make_checksum(struct rds_header *hdr)
779 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
782 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
784 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
789 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
791 int rds_page_copy_user(struct page *page, unsigned long offset,
792 void __user *ptr, unsigned long bytes,
794 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
795 rds_page_copy_user(page, offset, ptr, bytes, 1)
796 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
797 rds_page_copy_user(page, offset, ptr, bytes, 0)
798 void rds_page_exit(void);
801 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
803 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
805 void rds_inc_put(struct rds_incoming *inc);
806 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
807 struct rds_incoming *inc, gfp_t gfp);
808 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
810 void rds_clear_recv_queue(struct rds_sock *rs);
811 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
812 void rds_inc_info_copy(struct rds_incoming *inc,
813 struct rds_info_iterator *iter,
814 __be32 saddr, __be32 daddr, int flip);
817 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
818 void rds_send_path_reset(struct rds_conn_path *conn);
819 int rds_send_xmit(struct rds_conn_path *cp);
821 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
822 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
823 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
824 is_acked_func is_acked);
825 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
826 is_acked_func is_acked);
827 int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
830 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
831 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
832 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
833 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
834 void rds_rdma_drop_keys(struct rds_sock *rs);
835 int rds_rdma_extra_size(struct rds_rdma_args *args);
836 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
837 struct cmsghdr *cmsg);
838 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
839 struct cmsghdr *cmsg);
840 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
841 struct cmsghdr *cmsg);
842 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
843 struct cmsghdr *cmsg);
844 void rds_rdma_free_op(struct rm_rdma_op *ro);
845 void rds_atomic_free_op(struct rm_atomic_op *ao);
846 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
847 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
848 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
849 struct cmsghdr *cmsg);
851 void __rds_put_mr_final(struct rds_mr *mr);
852 static inline void rds_mr_put(struct rds_mr *mr)
854 if (atomic_dec_and_test(&mr->r_refcount))
855 __rds_put_mr_final(mr);
859 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
860 #define rds_stats_inc_which(which, member) do { \
861 per_cpu(which, get_cpu()).member++; \
864 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
865 #define rds_stats_add_which(which, member, count) do { \
866 per_cpu(which, get_cpu()).member += count; \
869 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
870 int rds_stats_init(void);
871 void rds_stats_exit(void);
872 void rds_stats_info_copy(struct rds_info_iterator *iter,
873 uint64_t *values, const char *const *names,
877 int rds_sysctl_init(void);
878 void rds_sysctl_exit(void);
879 extern unsigned long rds_sysctl_sndbuf_min;
880 extern unsigned long rds_sysctl_sndbuf_default;
881 extern unsigned long rds_sysctl_sndbuf_max;
882 extern unsigned long rds_sysctl_reconnect_min_jiffies;
883 extern unsigned long rds_sysctl_reconnect_max_jiffies;
884 extern unsigned int rds_sysctl_max_unacked_packets;
885 extern unsigned int rds_sysctl_max_unacked_bytes;
886 extern unsigned int rds_sysctl_ping_enable;
887 extern unsigned long rds_sysctl_trace_flags;
888 extern unsigned int rds_sysctl_trace_level;
891 int rds_threads_init(void);
892 void rds_threads_exit(void);
893 extern struct workqueue_struct *rds_wq;
894 void rds_queue_reconnect(struct rds_conn_path *cp);
895 void rds_connect_worker(struct work_struct *);
896 void rds_shutdown_worker(struct work_struct *);
897 void rds_send_worker(struct work_struct *);
898 void rds_recv_worker(struct work_struct *);
899 void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
900 void rds_connect_complete(struct rds_connection *conn);
903 int rds_trans_register(struct rds_transport *trans);
904 void rds_trans_unregister(struct rds_transport *trans);
905 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
906 void rds_trans_put(struct rds_transport *trans);
907 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
909 struct rds_transport *rds_trans_get(int t_type);
910 int rds_trans_init(void);
911 void rds_trans_exit(void);