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
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 mutex c_send_lock; /* protect send ring */
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_rdma_sent;
101 spinlock_t c_lock; /* protect msg queues */
103 struct list_head c_send_queue;
104 struct list_head c_retrans;
108 struct rds_transport *c_trans;
109 void *c_transport_data;
112 unsigned long c_flags;
113 unsigned long c_reconnect_jiffies;
114 struct delayed_work c_send_w;
115 struct delayed_work c_recv_w;
116 struct delayed_work c_conn_w;
117 struct work_struct c_down_w;
118 struct mutex c_cm_lock; /* protect conn state & cm */
120 struct list_head c_map_item;
121 unsigned long c_map_queued;
122 unsigned long c_map_offset;
123 unsigned long c_map_bytes;
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 * m_sock_item and m_conn_item are on lists that are serialized under
211 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
212 * the message will not be put back on the retransmit list after being sent.
213 * messages that are canceled while being sent rely on this.
215 * m_inc is used by loopback so that it can pass an incoming message straight
216 * back up into the rx path. It embeds a wire header which is also used by
217 * the send path, which is kind of awkward.
219 * m_sock_item indicates the message's presence on a socket's send or receive
220 * queue. m_rs will point to that socket.
222 * m_daddr is used by cancellation to prune messages to a given destination.
224 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
225 * nesting. As paths iterate over messages on a sock, or conn, they must
226 * also lock the conn, or sock, to remove the message from those lists too.
227 * Testing the flag to determine if the message is still on the lists lets
228 * us avoid testing the list_head directly. That means each path can use
229 * the message's list_head to keep it on a local list while juggling locks
230 * without confusing the other path.
232 * m_ack_seq is an optional field set by transports who need a different
233 * sequence number range to invalidate. They can use this in a callback
234 * that they pass to rds_send_drop_acked() to see if each message has been
235 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
236 * had ack_seq set yet.
238 #define RDS_MSG_ON_SOCK 1
239 #define RDS_MSG_ON_CONN 2
240 #define RDS_MSG_HAS_ACK_SEQ 3
241 #define RDS_MSG_ACK_REQUIRED 4
242 #define RDS_MSG_RETRANSMITTED 5
243 #define RDS_MSG_MAPPED 6
244 #define RDS_MSG_PAGEVEC 7
248 struct list_head m_sock_item;
249 struct list_head m_conn_item;
250 struct rds_incoming m_inc;
253 unsigned long m_flags;
255 /* Never access m_rs without holding m_rs_lock.
260 spinlock_t m_rs_lock;
261 struct rds_sock *m_rs;
262 rds_rdma_cookie_t m_rdma_cookie;
265 struct rds_rdma_op *m_rdma_op;
266 struct rds_mr *m_rdma_mr;
269 unsigned int m_nents;
270 unsigned int m_count;
271 struct scatterlist m_sg[0];
277 * The RDS notifier is used (optionally) to tell the application about
278 * completed RDMA operations. Rather than keeping the whole rds message
279 * around on the queue, we allocate a small notifier that is put on the
280 * socket's notifier_list. Notifications are delivered to the application
281 * through control messages.
283 struct rds_notifier {
284 struct list_head n_list;
285 uint64_t n_user_token;
290 * struct rds_transport - transport specific behavioural hooks
292 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
293 * part of a message. The caller serializes on the send_sem so this
294 * doesn't need to be reentrant for a given conn. The header must be
295 * sent before the data payload. .xmit must be prepared to send a
296 * message with no data payload. .xmit should return the number of
297 * bytes that were sent down the connection, including header bytes.
298 * Returning 0 tells the caller that it doesn't need to perform any
299 * additional work now. This is usually the case when the transport has
300 * filled the sending queue for its connection and will handle
301 * triggering the rds thread to continue the send when space becomes
302 * available. Returning -EAGAIN tells the caller to retry the send
303 * immediately. Returning -ENOMEM tells the caller to retry the send at
304 * some point in the future.
306 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
307 * it returns the connection can not call rds_recv_incoming().
308 * This will only be called once after conn_connect returns
309 * non-zero success and will The caller serializes this with
310 * the send and connecting paths (xmit_* and conn_*). The
311 * transport is responsible for other serialization, including
312 * rds_recv_incoming(). This is called in process context but
313 * should try hard not to block.
315 * @xmit_cong_map: This asks the transport to send the local bitmap down the
316 * given connection. XXX get a better story about the bitmap
320 #define RDS_TRANS_IB 0
321 #define RDS_TRANS_IWARP 1
322 #define RDS_TRANS_TCP 2
323 #define RDS_TRANS_COUNT 3
325 struct rds_transport {
326 char t_name[TRANSNAMSIZ];
327 struct list_head t_item;
328 struct module *t_owner;
329 unsigned int t_prefer_loopback:1;
332 int (*laddr_check)(__be32 addr);
333 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
334 void (*conn_free)(void *data);
335 int (*conn_connect)(struct rds_connection *conn);
336 void (*conn_shutdown)(struct rds_connection *conn);
337 void (*xmit_prepare)(struct rds_connection *conn);
338 void (*xmit_complete)(struct rds_connection *conn);
339 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
340 unsigned int hdr_off, unsigned int sg, unsigned int off);
341 int (*xmit_cong_map)(struct rds_connection *conn,
342 struct rds_cong_map *map, unsigned long offset);
343 int (*xmit_rdma)(struct rds_connection *conn, struct rds_rdma_op *op);
344 int (*recv)(struct rds_connection *conn);
345 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
347 void (*inc_purge)(struct rds_incoming *inc);
348 void (*inc_free)(struct rds_incoming *inc);
350 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
351 struct rdma_cm_event *event);
352 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
353 void (*cm_connect_complete)(struct rds_connection *conn,
354 struct rdma_cm_event *event);
356 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
359 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
360 struct rds_sock *rs, u32 *key_ret);
361 void (*sync_mr)(void *trans_private, int direction);
362 void (*free_mr)(void *trans_private, int invalidate);
363 void (*flush_mrs)(void);
373 * bound_addr used for both incoming and outgoing, no INADDR_ANY
376 struct rb_node rs_bound_node;
377 __be32 rs_bound_addr;
379 __be16 rs_bound_port;
383 * This is only used to communicate the transport between bind and
384 * initiating connections. All other trans use is referenced through
387 struct rds_transport *rs_transport;
390 * rds_sendmsg caches the conn it used the last time around.
391 * This helps avoid costly lookups.
393 struct rds_connection *rs_conn;
395 /* flag indicating we were congested or not */
397 /* seen congestion (ENOBUFS) when sending? */
398 int rs_seen_congestion;
400 /* rs_lock protects all these adjacent members before the newline */
402 struct list_head rs_send_queue;
405 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
407 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
408 * to decide whether the application should be woken up.
409 * If not set, we use rs_cong_track to find out whether a cong map
412 uint64_t rs_cong_mask;
413 uint64_t rs_cong_notify;
414 struct list_head rs_cong_list;
415 unsigned long rs_cong_track;
418 * rs_recv_lock protects the receive queue, and is
419 * used to serialize with rds_release.
421 rwlock_t rs_recv_lock;
422 struct list_head rs_recv_queue;
424 /* just for stats reporting */
425 struct list_head rs_item;
427 /* these have their own lock */
428 spinlock_t rs_rdma_lock;
429 struct rb_root rs_rdma_keys;
431 /* Socket options - in case there will be more */
432 unsigned char rs_recverr,
436 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
438 return container_of(sk, struct rds_sock, rs_sk);
440 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
446 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
447 * to account for overhead. We don't account for overhead, we just apply
448 * the number of payload bytes to the specified value.
450 static inline int rds_sk_sndbuf(struct rds_sock *rs)
452 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
454 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
456 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
459 struct rds_statistics {
460 uint64_t s_conn_reset;
461 uint64_t s_recv_drop_bad_checksum;
462 uint64_t s_recv_drop_old_seq;
463 uint64_t s_recv_drop_no_sock;
464 uint64_t s_recv_drop_dead_sock;
465 uint64_t s_recv_deliver_raced;
466 uint64_t s_recv_delivered;
467 uint64_t s_recv_queued;
468 uint64_t s_recv_immediate_retry;
469 uint64_t s_recv_delayed_retry;
470 uint64_t s_recv_ack_required;
471 uint64_t s_recv_rdma_bytes;
472 uint64_t s_recv_ping;
473 uint64_t s_send_queue_empty;
474 uint64_t s_send_queue_full;
475 uint64_t s_send_sem_contention;
476 uint64_t s_send_sem_queue_raced;
477 uint64_t s_send_immediate_retry;
478 uint64_t s_send_delayed_retry;
479 uint64_t s_send_drop_acked;
480 uint64_t s_send_ack_required;
481 uint64_t s_send_queued;
482 uint64_t s_send_rdma;
483 uint64_t s_send_rdma_bytes;
484 uint64_t s_send_pong;
485 uint64_t s_page_remainder_hit;
486 uint64_t s_page_remainder_miss;
487 uint64_t s_copy_to_user;
488 uint64_t s_copy_from_user;
489 uint64_t s_cong_update_queued;
490 uint64_t s_cong_update_received;
491 uint64_t s_cong_send_error;
492 uint64_t s_cong_send_blocked;
496 void rds_sock_addref(struct rds_sock *rs);
497 void rds_sock_put(struct rds_sock *rs);
498 void rds_wake_sk_sleep(struct rds_sock *rs);
499 static inline void __rds_wake_sk_sleep(struct sock *sk)
501 wait_queue_head_t *waitq = sk_sleep(sk);
503 if (!sock_flag(sk, SOCK_DEAD) && waitq)
506 extern wait_queue_head_t rds_poll_waitq;
510 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
511 void rds_remove_bound(struct rds_sock *rs);
512 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
515 int rds_cong_get_maps(struct rds_connection *conn);
516 void rds_cong_add_conn(struct rds_connection *conn);
517 void rds_cong_remove_conn(struct rds_connection *conn);
518 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
519 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
520 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
521 void rds_cong_queue_updates(struct rds_cong_map *map);
522 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
523 int rds_cong_updated_since(unsigned long *recent);
524 void rds_cong_add_socket(struct rds_sock *);
525 void rds_cong_remove_socket(struct rds_sock *);
526 void rds_cong_exit(void);
527 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
530 int __init rds_conn_init(void);
531 void rds_conn_exit(void);
532 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
533 struct rds_transport *trans, gfp_t gfp);
534 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
535 struct rds_transport *trans, gfp_t gfp);
536 void rds_conn_shutdown(struct rds_connection *conn);
537 void rds_conn_destroy(struct rds_connection *conn);
538 void rds_conn_reset(struct rds_connection *conn);
539 void rds_conn_drop(struct rds_connection *conn);
540 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
541 struct rds_info_iterator *iter,
542 struct rds_info_lengths *lens,
543 int (*visitor)(struct rds_connection *, void *),
545 void __rds_conn_error(struct rds_connection *conn, const char *, ...)
546 __attribute__ ((format (printf, 2, 3)));
547 #define rds_conn_error(conn, fmt...) \
548 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
551 rds_conn_transition(struct rds_connection *conn, int old, int new)
553 return atomic_cmpxchg(&conn->c_state, old, new) == old;
557 rds_conn_state(struct rds_connection *conn)
559 return atomic_read(&conn->c_state);
563 rds_conn_up(struct rds_connection *conn)
565 return atomic_read(&conn->c_state) == RDS_CONN_UP;
569 rds_conn_connecting(struct rds_connection *conn)
571 return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
575 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
576 struct rds_message *rds_message_copy_from_user(struct iovec *first_iov,
578 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
579 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
580 __be16 dport, u64 seq);
581 int rds_message_add_extension(struct rds_header *hdr,
582 unsigned int type, const void *data, unsigned int len);
583 int rds_message_next_extension(struct rds_header *hdr,
584 unsigned int *pos, void *buf, unsigned int *buflen);
585 int rds_message_add_version_extension(struct rds_header *hdr, unsigned int version);
586 int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *version);
587 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
588 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
589 struct iovec *first_iov, size_t size);
590 void rds_message_inc_purge(struct rds_incoming *inc);
591 void rds_message_inc_free(struct rds_incoming *inc);
592 void rds_message_addref(struct rds_message *rm);
593 void rds_message_put(struct rds_message *rm);
594 void rds_message_wait(struct rds_message *rm);
595 void rds_message_unmapped(struct rds_message *rm);
597 static inline void rds_message_make_checksum(struct rds_header *hdr)
600 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
603 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
605 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
610 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
612 int rds_page_copy_user(struct page *page, unsigned long offset,
613 void __user *ptr, unsigned long bytes,
615 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
616 rds_page_copy_user(page, offset, ptr, bytes, 1)
617 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
618 rds_page_copy_user(page, offset, ptr, bytes, 0)
619 void rds_page_exit(void);
622 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
624 void rds_inc_addref(struct rds_incoming *inc);
625 void rds_inc_put(struct rds_incoming *inc);
626 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
627 struct rds_incoming *inc, gfp_t gfp, enum km_type km);
628 int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
629 size_t size, int msg_flags);
630 void rds_clear_recv_queue(struct rds_sock *rs);
631 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
632 void rds_inc_info_copy(struct rds_incoming *inc,
633 struct rds_info_iterator *iter,
634 __be32 saddr, __be32 daddr, int flip);
637 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
639 void rds_send_reset(struct rds_connection *conn);
640 int rds_send_xmit(struct rds_connection *conn);
642 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
643 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
644 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
645 is_acked_func is_acked);
646 int rds_send_acked_before(struct rds_connection *conn, u64 seq);
647 void rds_send_remove_from_sock(struct list_head *messages, int status);
648 int rds_send_pong(struct rds_connection *conn, __be16 dport);
649 struct rds_message *rds_send_get_message(struct rds_connection *,
650 struct rds_rdma_op *);
653 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
656 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
657 #define rds_stats_inc_which(which, member) do { \
658 per_cpu(which, get_cpu()).member++; \
661 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
662 #define rds_stats_add_which(which, member, count) do { \
663 per_cpu(which, get_cpu()).member += count; \
666 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
667 int __init rds_stats_init(void);
668 void rds_stats_exit(void);
669 void rds_stats_info_copy(struct rds_info_iterator *iter,
670 uint64_t *values, const char *const *names,
674 int __init rds_sysctl_init(void);
675 void rds_sysctl_exit(void);
676 extern unsigned long rds_sysctl_sndbuf_min;
677 extern unsigned long rds_sysctl_sndbuf_default;
678 extern unsigned long rds_sysctl_sndbuf_max;
679 extern unsigned long rds_sysctl_reconnect_min_jiffies;
680 extern unsigned long rds_sysctl_reconnect_max_jiffies;
681 extern unsigned int rds_sysctl_max_unacked_packets;
682 extern unsigned int rds_sysctl_max_unacked_bytes;
683 extern unsigned int rds_sysctl_ping_enable;
684 extern unsigned long rds_sysctl_trace_flags;
685 extern unsigned int rds_sysctl_trace_level;
688 int __init rds_threads_init(void);
689 void rds_threads_exit(void);
690 extern struct workqueue_struct *rds_wq;
691 void rds_queue_reconnect(struct rds_connection *conn);
692 void rds_connect_worker(struct work_struct *);
693 void rds_shutdown_worker(struct work_struct *);
694 void rds_send_worker(struct work_struct *);
695 void rds_recv_worker(struct work_struct *);
696 void rds_connect_complete(struct rds_connection *conn);
699 int rds_trans_register(struct rds_transport *trans);
700 void rds_trans_unregister(struct rds_transport *trans);
701 struct rds_transport *rds_trans_get_preferred(__be32 addr);
702 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
704 int __init rds_trans_init(void);
705 void rds_trans_exit(void);