1 /* SPDX-License-Identifier: GPL-2.0 */
6 #include <linux/scatterlist.h>
7 #include <linux/highmem.h>
8 #include <rdma/rdma_cm.h>
9 #include <linux/mutex.h>
10 #include <linux/rds.h>
11 #include <linux/rhashtable.h>
12 #include <linux/refcount.h>
13 #include <linux/in6.h>
18 * RDS Network protocol version
20 #define RDS_PROTOCOL_3_0 0x0300
21 #define RDS_PROTOCOL_3_1 0x0301
22 #define RDS_PROTOCOL_VERSION RDS_PROTOCOL_3_1
23 #define RDS_PROTOCOL_MAJOR(v) ((v) >> 8)
24 #define RDS_PROTOCOL_MINOR(v) ((v) & 255)
25 #define RDS_PROTOCOL(maj, min) (((maj) << 8) | min)
27 /* The following ports, 16385, 18634, 18635, are registered with IANA as
28 * the ports to be used for RDS over TCP and UDP. Currently, only RDS over
29 * TCP and RDS over IB/RDMA are implemented. 18634 is the historical value
30 * used for the RDMA_CM listener port. RDS/TCP uses port 16385. After
31 * IPv6 work, RDMA_CM also uses 16385 as the listener port. 18634 is kept
32 * to ensure compatibility with older RDS modules. Those ports are defined
33 * in each transport's header file.
35 #define RDS_PORT 18634
38 #define KERNEL_HAS_ATOMIC64
42 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
44 /* sigh, pr_debug() causes unused variable warnings */
45 static inline __printf(1, 2)
46 void rdsdebug(char *fmt, ...)
51 /* XXX is there one of these somewhere? */
53 ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
55 #define RDS_FRAG_SHIFT 12
56 #define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT))
58 /* Used to limit both RDMA and non-RDMA RDS message to 1MB */
59 #define RDS_MAX_MSG_SIZE ((unsigned int)(1 << 20))
61 #define RDS_CONG_MAP_BYTES (65536 / 8)
62 #define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
63 #define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
66 struct rb_node m_rb_node;
67 struct in6_addr m_addr;
68 wait_queue_head_t m_waitq;
69 struct list_head m_conn_list;
70 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES];
75 * This is how we will track the connection state:
76 * A connection is always in one of the following
77 * states. Updates to the state are atomic and imply
83 RDS_CONN_DISCONNECTING,
89 /* Bits for c_flags */
90 #define RDS_LL_SEND_FULL 0
91 #define RDS_RECONNECT_PENDING 1
93 #define RDS_RECV_REFILL 3
94 #define RDS_DESTROY_PENDING 4
96 /* Max number of multipaths per RDS connection. Must be a power of 2 */
97 #define RDS_MPATH_WORKERS 8
98 #define RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
99 (rs)->rs_hash_initval) & ((n) - 1))
101 #define IS_CANONICAL(laddr, faddr) (htonl(laddr) < htonl(faddr))
103 /* Per mpath connection state */
104 struct rds_conn_path {
105 struct rds_connection *cp_conn;
106 struct rds_message *cp_xmit_rm;
107 unsigned long cp_xmit_sg;
108 unsigned int cp_xmit_hdr_off;
109 unsigned int cp_xmit_data_off;
110 unsigned int cp_xmit_atomic_sent;
111 unsigned int cp_xmit_rdma_sent;
112 unsigned int cp_xmit_data_sent;
114 spinlock_t cp_lock; /* protect msg queues */
116 struct list_head cp_send_queue;
117 struct list_head cp_retrans;
121 void *cp_transport_data;
124 unsigned long cp_send_gen;
125 unsigned long cp_flags;
126 unsigned long cp_reconnect_jiffies;
127 struct delayed_work cp_send_w;
128 struct delayed_work cp_recv_w;
129 struct delayed_work cp_conn_w;
130 struct work_struct cp_down_w;
131 struct mutex cp_cm_lock; /* protect cp_state & cm */
132 wait_queue_head_t cp_waitq;
134 unsigned int cp_unacked_packets;
135 unsigned int cp_unacked_bytes;
136 unsigned int cp_index;
139 /* One rds_connection per RDS address pair */
140 struct rds_connection {
141 struct hlist_node c_hash_node;
142 struct in6_addr c_laddr;
143 struct in6_addr c_faddr;
144 int c_dev_if; /* ifindex used for this conn */
145 int c_bound_if; /* ifindex of c_laddr */
146 unsigned int c_loopback:1,
151 struct rds_connection *c_passive;
152 struct rds_transport *c_trans;
154 struct rds_cong_map *c_lcong;
155 struct rds_cong_map *c_fcong;
157 /* Protocol version */
158 unsigned int c_version;
159 possible_net_t c_net;
161 struct list_head c_map_item;
162 unsigned long c_map_queued;
164 struct rds_conn_path *c_path;
165 wait_queue_head_t c_hs_waitq; /* handshake waitq */
172 struct net *rds_conn_net(struct rds_connection *conn)
174 return read_pnet(&conn->c_net);
178 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
180 write_pnet(&conn->c_net, net);
183 #define RDS_FLAG_CONG_BITMAP 0x01
184 #define RDS_FLAG_ACK_REQUIRED 0x02
185 #define RDS_FLAG_RETRANSMITTED 0x04
186 #define RDS_MAX_ADV_CREDIT 255
188 /* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
189 * probe to exchange control information before establishing a connection.
190 * Currently the control information that is exchanged is the number of
191 * supported paths. If the peer is a legacy (older kernel revision) peer,
192 * it would return a pong message without additional control information
193 * that would then alert the sender that the peer was an older rev.
195 #define RDS_FLAG_PROBE_PORT 1
196 #define RDS_HS_PROBE(sport, dport) \
197 ((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
198 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
200 * Maximum space available for extension headers.
202 #define RDS_HEADER_EXT_SPACE 16
215 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
219 * Reserved - indicates end of extensions
221 #define RDS_EXTHDR_NONE 0
224 * This extension header is included in the very
225 * first message that is sent on a new connection,
226 * and identifies the protocol level. This will help
227 * rolling updates if a future change requires breaking
229 * NB: This is no longer true for IB, where we do a version
230 * negotiation during the connection setup phase (protocol
231 * version information is included in the RDMA CM private data).
233 #define RDS_EXTHDR_VERSION 1
234 struct rds_ext_header_version {
239 * This extension header is included in the RDS message
240 * chasing an RDMA operation.
242 #define RDS_EXTHDR_RDMA 2
243 struct rds_ext_header_rdma {
248 * This extension header tells the peer about the
249 * destination <R_Key,offset> of the requested RDMA
252 #define RDS_EXTHDR_RDMA_DEST 3
253 struct rds_ext_header_rdma_dest {
255 __be32 h_rdma_offset;
258 /* Extension header announcing number of paths.
259 * Implicit length = 2 bytes.
261 #define RDS_EXTHDR_NPATHS 5
262 #define RDS_EXTHDR_GEN_NUM 6
264 #define __RDS_EXTHDR_MAX 16 /* for now */
265 #define RDS_RX_MAX_TRACES (RDS_MSG_RX_DGRAM_TRACE_MAX + 1)
266 #define RDS_MSG_RX_HDR 0
267 #define RDS_MSG_RX_START 1
268 #define RDS_MSG_RX_END 2
269 #define RDS_MSG_RX_CMSG 3
271 struct rds_incoming {
272 refcount_t i_refcount;
273 struct list_head i_item;
274 struct rds_connection *i_conn;
275 struct rds_conn_path *i_conn_path;
276 struct rds_header i_hdr;
277 unsigned long i_rx_jiffies;
278 struct in6_addr i_saddr;
280 rds_rdma_cookie_t i_rdma_cookie;
281 struct timeval i_rx_tstamp;
282 u64 i_rx_lat_trace[RDS_RX_MAX_TRACES];
286 struct rb_node r_rb_node;
287 refcount_t r_refcount;
290 /* A copy of the creation flags */
291 unsigned int r_use_once:1;
292 unsigned int r_invalidate:1;
293 unsigned int r_write:1;
295 /* This is for RDS_MR_DEAD.
296 * It would be nice & consistent to make this part of the above
297 * bit field here, but we need to use test_and_set_bit.
299 unsigned long r_state;
300 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
301 struct rds_transport *r_trans;
302 void *r_trans_private;
305 /* Flags for mr->r_state */
306 #define RDS_MR_DEAD 0
308 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
310 return r_key | (((u64) offset) << 32);
313 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
318 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
323 /* atomic operation types */
324 #define RDS_ATOMIC_TYPE_CSWP 0
325 #define RDS_ATOMIC_TYPE_FADD 1
328 * m_sock_item and m_conn_item are on lists that are serialized under
329 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
330 * the message will not be put back on the retransmit list after being sent.
331 * messages that are canceled while being sent rely on this.
333 * m_inc is used by loopback so that it can pass an incoming message straight
334 * back up into the rx path. It embeds a wire header which is also used by
335 * the send path, which is kind of awkward.
337 * m_sock_item indicates the message's presence on a socket's send or receive
338 * queue. m_rs will point to that socket.
340 * m_daddr is used by cancellation to prune messages to a given destination.
342 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
343 * nesting. As paths iterate over messages on a sock, or conn, they must
344 * also lock the conn, or sock, to remove the message from those lists too.
345 * Testing the flag to determine if the message is still on the lists lets
346 * us avoid testing the list_head directly. That means each path can use
347 * the message's list_head to keep it on a local list while juggling locks
348 * without confusing the other path.
350 * m_ack_seq is an optional field set by transports who need a different
351 * sequence number range to invalidate. They can use this in a callback
352 * that they pass to rds_send_drop_acked() to see if each message has been
353 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
354 * had ack_seq set yet.
356 #define RDS_MSG_ON_SOCK 1
357 #define RDS_MSG_ON_CONN 2
358 #define RDS_MSG_HAS_ACK_SEQ 3
359 #define RDS_MSG_ACK_REQUIRED 4
360 #define RDS_MSG_RETRANSMITTED 5
361 #define RDS_MSG_MAPPED 6
362 #define RDS_MSG_PAGEVEC 7
363 #define RDS_MSG_FLUSH 8
365 struct rds_znotifier {
370 struct rds_msg_zcopy_info {
371 struct list_head rs_zcookie_next;
373 struct rds_znotifier znotif;
374 struct rds_zcopy_cookies zcookies;
378 struct rds_msg_zcopy_queue {
379 struct list_head zcookie_head;
380 spinlock_t lock; /* protects zcookie_head queue */
383 static inline void rds_message_zcopy_queue_init(struct rds_msg_zcopy_queue *q)
385 spin_lock_init(&q->lock);
386 INIT_LIST_HEAD(&q->zcookie_head);
389 struct rds_iov_vector {
390 struct rds_iovec *iov;
394 struct rds_iov_vector_arr {
395 struct rds_iov_vector *vec;
402 refcount_t m_refcount;
403 struct list_head m_sock_item;
404 struct list_head m_conn_item;
405 struct rds_incoming m_inc;
407 struct in6_addr m_daddr;
408 unsigned long m_flags;
410 /* Never access m_rs without holding m_rs_lock.
415 spinlock_t m_rs_lock;
416 wait_queue_head_t m_flush_wait;
418 struct rds_sock *m_rs;
420 /* cookie to send to remote, in rds header */
421 rds_rdma_cookie_t m_rdma_cookie;
423 unsigned int m_used_sgs;
424 unsigned int m_total_sgs;
429 struct rm_atomic_op {
435 uint64_t compare_mask;
440 uint64_t nocarry_mask;
446 unsigned int op_notify:1;
447 unsigned int op_recverr:1;
448 unsigned int op_mapped:1;
449 unsigned int op_silent:1;
450 unsigned int op_active:1;
451 struct scatterlist *op_sg;
452 struct rds_notifier *op_notifier;
454 struct rds_mr *op_rdma_mr;
459 unsigned int op_write:1;
460 unsigned int op_fence:1;
461 unsigned int op_notify:1;
462 unsigned int op_recverr:1;
463 unsigned int op_mapped:1;
464 unsigned int op_silent:1;
465 unsigned int op_active:1;
466 unsigned int op_bytes;
467 unsigned int op_nents;
468 unsigned int op_count;
469 struct scatterlist *op_sg;
470 struct rds_notifier *op_notifier;
472 struct rds_mr *op_rdma_mr;
475 unsigned int op_active:1;
476 unsigned int op_notify:1;
477 unsigned int op_nents;
478 unsigned int op_count;
479 unsigned int op_dmasg;
480 unsigned int op_dmaoff;
481 struct rds_znotifier *op_mmp_znotifier;
482 struct scatterlist *op_sg;
486 struct rds_conn_path *m_conn_path;
490 * The RDS notifier is used (optionally) to tell the application about
491 * completed RDMA operations. Rather than keeping the whole rds message
492 * around on the queue, we allocate a small notifier that is put on the
493 * socket's notifier_list. Notifications are delivered to the application
494 * through control messages.
496 struct rds_notifier {
497 struct list_head n_list;
498 uint64_t n_user_token;
502 /* Available as part of RDS core, so doesn't need to participate
503 * in get_preferred transport etc
505 #define RDS_TRANS_LOOP 3
508 * struct rds_transport - transport specific behavioural hooks
510 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
511 * part of a message. The caller serializes on the send_sem so this
512 * doesn't need to be reentrant for a given conn. The header must be
513 * sent before the data payload. .xmit must be prepared to send a
514 * message with no data payload. .xmit should return the number of
515 * bytes that were sent down the connection, including header bytes.
516 * Returning 0 tells the caller that it doesn't need to perform any
517 * additional work now. This is usually the case when the transport has
518 * filled the sending queue for its connection and will handle
519 * triggering the rds thread to continue the send when space becomes
520 * available. Returning -EAGAIN tells the caller to retry the send
521 * immediately. Returning -ENOMEM tells the caller to retry the send at
522 * some point in the future.
524 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
525 * it returns the connection can not call rds_recv_incoming().
526 * This will only be called once after conn_connect returns
527 * non-zero success and will The caller serializes this with
528 * the send and connecting paths (xmit_* and conn_*). The
529 * transport is responsible for other serialization, including
530 * rds_recv_incoming(). This is called in process context but
531 * should try hard not to block.
534 struct rds_transport {
535 char t_name[TRANSNAMSIZ];
536 struct list_head t_item;
537 struct module *t_owner;
538 unsigned int t_prefer_loopback:1,
542 int (*laddr_check)(struct net *net, const struct in6_addr *addr,
544 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
545 void (*conn_free)(void *data);
546 int (*conn_path_connect)(struct rds_conn_path *cp);
547 void (*conn_path_shutdown)(struct rds_conn_path *conn);
548 void (*xmit_path_prepare)(struct rds_conn_path *cp);
549 void (*xmit_path_complete)(struct rds_conn_path *cp);
550 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
551 unsigned int hdr_off, unsigned int sg, unsigned int off);
552 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
553 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
554 int (*recv_path)(struct rds_conn_path *cp);
555 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
556 void (*inc_free)(struct rds_incoming *inc);
558 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
559 struct rdma_cm_event *event, bool isv6);
560 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id, bool isv6);
561 void (*cm_connect_complete)(struct rds_connection *conn,
562 struct rdma_cm_event *event);
564 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
567 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
568 struct rds_sock *rs, u32 *key_ret,
569 struct rds_connection *conn);
570 void (*sync_mr)(void *trans_private, int direction);
571 void (*free_mr)(void *trans_private, int invalidate);
572 void (*flush_mrs)(void);
573 bool (*t_unloading)(struct rds_connection *conn);
576 /* Bind hash table key length. It is the sum of the size of a struct
577 * in6_addr, a scope_id and a port.
579 #define RDS_BOUND_KEY_LEN \
580 (sizeof(struct in6_addr) + sizeof(__u32) + sizeof(__be16))
589 * bound_addr used for both incoming and outgoing, no INADDR_ANY
592 struct rhash_head rs_bound_node;
593 u8 rs_bound_key[RDS_BOUND_KEY_LEN];
594 struct sockaddr_in6 rs_bound_sin6;
595 #define rs_bound_addr rs_bound_sin6.sin6_addr
596 #define rs_bound_addr_v4 rs_bound_sin6.sin6_addr.s6_addr32[3]
597 #define rs_bound_port rs_bound_sin6.sin6_port
598 #define rs_bound_scope_id rs_bound_sin6.sin6_scope_id
599 struct in6_addr rs_conn_addr;
600 #define rs_conn_addr_v4 rs_conn_addr.s6_addr32[3]
602 struct rds_transport *rs_transport;
605 * rds_sendmsg caches the conn it used the last time around.
606 * This helps avoid costly lookups.
608 struct rds_connection *rs_conn;
610 /* flag indicating we were congested or not */
612 /* seen congestion (ENOBUFS) when sending? */
613 int rs_seen_congestion;
615 /* rs_lock protects all these adjacent members before the newline */
617 struct list_head rs_send_queue;
620 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
622 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
623 * to decide whether the application should be woken up.
624 * If not set, we use rs_cong_track to find out whether a cong map
627 uint64_t rs_cong_mask;
628 uint64_t rs_cong_notify;
629 struct list_head rs_cong_list;
630 unsigned long rs_cong_track;
633 * rs_recv_lock protects the receive queue, and is
634 * used to serialize with rds_release.
636 rwlock_t rs_recv_lock;
637 struct list_head rs_recv_queue;
639 /* just for stats reporting */
640 struct list_head rs_item;
642 /* these have their own lock */
643 spinlock_t rs_rdma_lock;
644 struct rb_root rs_rdma_keys;
646 /* Socket options - in case there will be more */
647 unsigned char rs_recverr,
651 /* Socket receive path trace points*/
653 u8 rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
654 struct rds_msg_zcopy_queue rs_zcookie_queue;
657 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
659 return container_of(sk, struct rds_sock, rs_sk);
661 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
667 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
668 * to account for overhead. We don't account for overhead, we just apply
669 * the number of payload bytes to the specified value.
671 static inline int rds_sk_sndbuf(struct rds_sock *rs)
673 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
675 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
677 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
680 struct rds_statistics {
681 uint64_t s_conn_reset;
682 uint64_t s_recv_drop_bad_checksum;
683 uint64_t s_recv_drop_old_seq;
684 uint64_t s_recv_drop_no_sock;
685 uint64_t s_recv_drop_dead_sock;
686 uint64_t s_recv_deliver_raced;
687 uint64_t s_recv_delivered;
688 uint64_t s_recv_queued;
689 uint64_t s_recv_immediate_retry;
690 uint64_t s_recv_delayed_retry;
691 uint64_t s_recv_ack_required;
692 uint64_t s_recv_rdma_bytes;
693 uint64_t s_recv_ping;
694 uint64_t s_send_queue_empty;
695 uint64_t s_send_queue_full;
696 uint64_t s_send_lock_contention;
697 uint64_t s_send_lock_queue_raced;
698 uint64_t s_send_immediate_retry;
699 uint64_t s_send_delayed_retry;
700 uint64_t s_send_drop_acked;
701 uint64_t s_send_ack_required;
702 uint64_t s_send_queued;
703 uint64_t s_send_rdma;
704 uint64_t s_send_rdma_bytes;
705 uint64_t s_send_pong;
706 uint64_t s_page_remainder_hit;
707 uint64_t s_page_remainder_miss;
708 uint64_t s_copy_to_user;
709 uint64_t s_copy_from_user;
710 uint64_t s_cong_update_queued;
711 uint64_t s_cong_update_received;
712 uint64_t s_cong_send_error;
713 uint64_t s_cong_send_blocked;
714 uint64_t s_recv_bytes_added_to_socket;
715 uint64_t s_recv_bytes_removed_from_socket;
720 void rds_sock_addref(struct rds_sock *rs);
721 void rds_sock_put(struct rds_sock *rs);
722 void rds_wake_sk_sleep(struct rds_sock *rs);
723 static inline void __rds_wake_sk_sleep(struct sock *sk)
725 wait_queue_head_t *waitq = sk_sleep(sk);
727 if (!sock_flag(sk, SOCK_DEAD) && waitq)
730 extern wait_queue_head_t rds_poll_waitq;
734 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
735 void rds_remove_bound(struct rds_sock *rs);
736 struct rds_sock *rds_find_bound(const struct in6_addr *addr, __be16 port,
738 int rds_bind_lock_init(void);
739 void rds_bind_lock_destroy(void);
742 int rds_cong_get_maps(struct rds_connection *conn);
743 void rds_cong_add_conn(struct rds_connection *conn);
744 void rds_cong_remove_conn(struct rds_connection *conn);
745 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
746 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
747 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
748 void rds_cong_queue_updates(struct rds_cong_map *map);
749 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
750 int rds_cong_updated_since(unsigned long *recent);
751 void rds_cong_add_socket(struct rds_sock *);
752 void rds_cong_remove_socket(struct rds_sock *);
753 void rds_cong_exit(void);
754 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
757 extern u32 rds_gen_num;
758 int rds_conn_init(void);
759 void rds_conn_exit(void);
760 struct rds_connection *rds_conn_create(struct net *net,
761 const struct in6_addr *laddr,
762 const struct in6_addr *faddr,
763 struct rds_transport *trans, gfp_t gfp,
765 struct rds_connection *rds_conn_create_outgoing(struct net *net,
766 const struct in6_addr *laddr,
767 const struct in6_addr *faddr,
768 struct rds_transport *trans,
769 gfp_t gfp, int dev_if);
770 void rds_conn_shutdown(struct rds_conn_path *cpath);
771 void rds_conn_destroy(struct rds_connection *conn);
772 void rds_conn_drop(struct rds_connection *conn);
773 void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy);
774 void rds_conn_connect_if_down(struct rds_connection *conn);
775 void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
776 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
777 struct rds_info_iterator *iter,
778 struct rds_info_lengths *lens,
779 int (*visitor)(struct rds_connection *, void *),
784 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
785 #define rds_conn_path_error(cp, fmt...) \
786 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
789 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
791 return atomic_cmpxchg(&cp->cp_state, old, new) == old;
795 rds_conn_transition(struct rds_connection *conn, int old, int new)
797 WARN_ON(conn->c_trans->t_mp_capable);
798 return rds_conn_path_transition(&conn->c_path[0], old, new);
802 rds_conn_path_state(struct rds_conn_path *cp)
804 return atomic_read(&cp->cp_state);
808 rds_conn_state(struct rds_connection *conn)
810 WARN_ON(conn->c_trans->t_mp_capable);
811 return rds_conn_path_state(&conn->c_path[0]);
815 rds_conn_path_up(struct rds_conn_path *cp)
817 return atomic_read(&cp->cp_state) == RDS_CONN_UP;
821 rds_conn_up(struct rds_connection *conn)
823 WARN_ON(conn->c_trans->t_mp_capable);
824 return rds_conn_path_up(&conn->c_path[0]);
828 rds_conn_path_connecting(struct rds_conn_path *cp)
830 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
834 rds_conn_connecting(struct rds_connection *conn)
836 WARN_ON(conn->c_trans->t_mp_capable);
837 return rds_conn_path_connecting(&conn->c_path[0]);
841 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
842 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
843 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
845 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
846 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
847 __be16 dport, u64 seq);
848 int rds_message_add_extension(struct rds_header *hdr,
849 unsigned int type, const void *data, unsigned int len);
850 int rds_message_next_extension(struct rds_header *hdr,
851 unsigned int *pos, void *buf, unsigned int *buflen);
852 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
853 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
854 void rds_message_inc_free(struct rds_incoming *inc);
855 void rds_message_addref(struct rds_message *rm);
856 void rds_message_put(struct rds_message *rm);
857 void rds_message_wait(struct rds_message *rm);
858 void rds_message_unmapped(struct rds_message *rm);
859 void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *info);
861 static inline void rds_message_make_checksum(struct rds_header *hdr)
864 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
867 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
869 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
874 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
876 void rds_page_exit(void);
879 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
880 struct in6_addr *saddr);
881 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
882 struct in6_addr *saddr);
883 void rds_inc_put(struct rds_incoming *inc);
884 void rds_recv_incoming(struct rds_connection *conn, struct in6_addr *saddr,
885 struct in6_addr *daddr,
886 struct rds_incoming *inc, gfp_t gfp);
887 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
889 void rds_clear_recv_queue(struct rds_sock *rs);
890 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
891 void rds_inc_info_copy(struct rds_incoming *inc,
892 struct rds_info_iterator *iter,
893 __be32 saddr, __be32 daddr, int flip);
894 void rds6_inc_info_copy(struct rds_incoming *inc,
895 struct rds_info_iterator *iter,
896 struct in6_addr *saddr, struct in6_addr *daddr,
900 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
901 void rds_send_path_reset(struct rds_conn_path *conn);
902 int rds_send_xmit(struct rds_conn_path *cp);
904 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest);
905 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
906 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
907 is_acked_func is_acked);
908 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
909 is_acked_func is_acked);
910 void rds_send_ping(struct rds_connection *conn, int cp_index);
911 int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
914 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
915 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
916 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
917 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
918 void rds_rdma_drop_keys(struct rds_sock *rs);
919 int rds_rdma_extra_size(struct rds_rdma_args *args,
920 struct rds_iov_vector *iov);
921 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
922 struct cmsghdr *cmsg);
923 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
924 struct cmsghdr *cmsg,
925 struct rds_iov_vector *vec);
926 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
927 struct cmsghdr *cmsg);
928 void rds_rdma_free_op(struct rm_rdma_op *ro);
929 void rds_atomic_free_op(struct rm_atomic_op *ao);
930 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
931 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
932 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
933 struct cmsghdr *cmsg);
935 void __rds_put_mr_final(struct rds_mr *mr);
936 static inline void rds_mr_put(struct rds_mr *mr)
938 if (refcount_dec_and_test(&mr->r_refcount))
939 __rds_put_mr_final(mr);
942 static inline bool rds_destroy_pending(struct rds_connection *conn)
944 return !check_net(rds_conn_net(conn)) ||
945 (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn));
949 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
950 #define rds_stats_inc_which(which, member) do { \
951 per_cpu(which, get_cpu()).member++; \
954 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
955 #define rds_stats_add_which(which, member, count) do { \
956 per_cpu(which, get_cpu()).member += count; \
959 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
960 int rds_stats_init(void);
961 void rds_stats_exit(void);
962 void rds_stats_info_copy(struct rds_info_iterator *iter,
963 uint64_t *values, const char *const *names,
967 int rds_sysctl_init(void);
968 void rds_sysctl_exit(void);
969 extern unsigned long rds_sysctl_sndbuf_min;
970 extern unsigned long rds_sysctl_sndbuf_default;
971 extern unsigned long rds_sysctl_sndbuf_max;
972 extern unsigned long rds_sysctl_reconnect_min_jiffies;
973 extern unsigned long rds_sysctl_reconnect_max_jiffies;
974 extern unsigned int rds_sysctl_max_unacked_packets;
975 extern unsigned int rds_sysctl_max_unacked_bytes;
976 extern unsigned int rds_sysctl_ping_enable;
977 extern unsigned long rds_sysctl_trace_flags;
978 extern unsigned int rds_sysctl_trace_level;
981 int rds_threads_init(void);
982 void rds_threads_exit(void);
983 extern struct workqueue_struct *rds_wq;
984 void rds_queue_reconnect(struct rds_conn_path *cp);
985 void rds_connect_worker(struct work_struct *);
986 void rds_shutdown_worker(struct work_struct *);
987 void rds_send_worker(struct work_struct *);
988 void rds_recv_worker(struct work_struct *);
989 void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
990 void rds_connect_complete(struct rds_connection *conn);
991 int rds_addr_cmp(const struct in6_addr *a1, const struct in6_addr *a2);
994 void rds_trans_register(struct rds_transport *trans);
995 void rds_trans_unregister(struct rds_transport *trans);
996 struct rds_transport *rds_trans_get_preferred(struct net *net,
997 const struct in6_addr *addr,
999 void rds_trans_put(struct rds_transport *trans);
1000 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
1001 unsigned int avail);
1002 struct rds_transport *rds_trans_get(int t_type);
1003 int rds_trans_init(void);
1004 void rds_trans_exit(void);