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 struct rds_rdma_op *m_rdma_op;
263 rds_rdma_cookie_t m_rdma_cookie;
264 struct rds_mr *m_rdma_mr;
265 unsigned int m_nents;
266 unsigned int m_count;
267 struct scatterlist m_sg[0];
271 * The RDS notifier is used (optionally) to tell the application about
272 * completed RDMA operations. Rather than keeping the whole rds message
273 * around on the queue, we allocate a small notifier that is put on the
274 * socket's notifier_list. Notifications are delivered to the application
275 * through control messages.
277 struct rds_notifier {
278 struct list_head n_list;
279 uint64_t n_user_token;
284 * struct rds_transport - transport specific behavioural hooks
286 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
287 * part of a message. The caller serializes on the send_sem so this
288 * doesn't need to be reentrant for a given conn. The header must be
289 * sent before the data payload. .xmit must be prepared to send a
290 * message with no data payload. .xmit should return the number of
291 * bytes that were sent down the connection, including header bytes.
292 * Returning 0 tells the caller that it doesn't need to perform any
293 * additional work now. This is usually the case when the transport has
294 * filled the sending queue for its connection and will handle
295 * triggering the rds thread to continue the send when space becomes
296 * available. Returning -EAGAIN tells the caller to retry the send
297 * immediately. Returning -ENOMEM tells the caller to retry the send at
298 * some point in the future.
300 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
301 * it returns the connection can not call rds_recv_incoming().
302 * This will only be called once after conn_connect returns
303 * non-zero success and will The caller serializes this with
304 * the send and connecting paths (xmit_* and conn_*). The
305 * transport is responsible for other serialization, including
306 * rds_recv_incoming(). This is called in process context but
307 * should try hard not to block.
309 * @xmit_cong_map: This asks the transport to send the local bitmap down the
310 * given connection. XXX get a better story about the bitmap
314 struct rds_transport {
315 char t_name[TRANSNAMSIZ];
316 struct list_head t_item;
317 struct module *t_owner;
318 unsigned int t_prefer_loopback:1;
320 int (*laddr_check)(__be32 addr);
321 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
322 void (*conn_free)(void *data);
323 int (*conn_connect)(struct rds_connection *conn);
324 void (*conn_shutdown)(struct rds_connection *conn);
325 void (*xmit_prepare)(struct rds_connection *conn);
326 void (*xmit_complete)(struct rds_connection *conn);
327 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
328 unsigned int hdr_off, unsigned int sg, unsigned int off);
329 int (*xmit_cong_map)(struct rds_connection *conn,
330 struct rds_cong_map *map, unsigned long offset);
331 int (*xmit_rdma)(struct rds_connection *conn, struct rds_rdma_op *op);
332 int (*recv)(struct rds_connection *conn);
333 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
335 void (*inc_purge)(struct rds_incoming *inc);
336 void (*inc_free)(struct rds_incoming *inc);
338 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
339 struct rdma_cm_event *event);
340 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
341 void (*cm_connect_complete)(struct rds_connection *conn,
342 struct rdma_cm_event *event);
344 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
347 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
348 struct rds_sock *rs, u32 *key_ret);
349 void (*sync_mr)(void *trans_private, int direction);
350 void (*free_mr)(void *trans_private, int invalidate);
351 void (*flush_mrs)(void);
361 * bound_addr used for both incoming and outgoing, no INADDR_ANY
364 struct rb_node rs_bound_node;
365 __be32 rs_bound_addr;
367 __be16 rs_bound_port;
371 * This is only used to communicate the transport between bind and
372 * initiating connections. All other trans use is referenced through
375 struct rds_transport *rs_transport;
378 * rds_sendmsg caches the conn it used the last time around.
379 * This helps avoid costly lookups.
381 struct rds_connection *rs_conn;
383 /* flag indicating we were congested or not */
386 /* rs_lock protects all these adjacent members before the newline */
388 struct list_head rs_send_queue;
391 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
393 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
394 * to decide whether the application should be woken up.
395 * If not set, we use rs_cong_track to find out whether a cong map
398 uint64_t rs_cong_mask;
399 uint64_t rs_cong_notify;
400 struct list_head rs_cong_list;
401 unsigned long rs_cong_track;
404 * rs_recv_lock protects the receive queue, and is
405 * used to serialize with rds_release.
407 rwlock_t rs_recv_lock;
408 struct list_head rs_recv_queue;
410 /* just for stats reporting */
411 struct list_head rs_item;
413 /* these have their own lock */
414 spinlock_t rs_rdma_lock;
415 struct rb_root rs_rdma_keys;
417 /* Socket options - in case there will be more */
418 unsigned char rs_recverr,
422 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
424 return container_of(sk, struct rds_sock, rs_sk);
426 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
432 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
433 * to account for overhead. We don't account for overhead, we just apply
434 * the number of payload bytes to the specified value.
436 static inline int rds_sk_sndbuf(struct rds_sock *rs)
438 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
440 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
442 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
445 struct rds_statistics {
446 uint64_t s_conn_reset;
447 uint64_t s_recv_drop_bad_checksum;
448 uint64_t s_recv_drop_old_seq;
449 uint64_t s_recv_drop_no_sock;
450 uint64_t s_recv_drop_dead_sock;
451 uint64_t s_recv_deliver_raced;
452 uint64_t s_recv_delivered;
453 uint64_t s_recv_queued;
454 uint64_t s_recv_immediate_retry;
455 uint64_t s_recv_delayed_retry;
456 uint64_t s_recv_ack_required;
457 uint64_t s_recv_rdma_bytes;
458 uint64_t s_recv_ping;
459 uint64_t s_send_queue_empty;
460 uint64_t s_send_queue_full;
461 uint64_t s_send_sem_contention;
462 uint64_t s_send_sem_queue_raced;
463 uint64_t s_send_immediate_retry;
464 uint64_t s_send_delayed_retry;
465 uint64_t s_send_drop_acked;
466 uint64_t s_send_ack_required;
467 uint64_t s_send_queued;
468 uint64_t s_send_rdma;
469 uint64_t s_send_rdma_bytes;
470 uint64_t s_send_pong;
471 uint64_t s_page_remainder_hit;
472 uint64_t s_page_remainder_miss;
473 uint64_t s_copy_to_user;
474 uint64_t s_copy_from_user;
475 uint64_t s_cong_update_queued;
476 uint64_t s_cong_update_received;
477 uint64_t s_cong_send_error;
478 uint64_t s_cong_send_blocked;
482 void rds_sock_addref(struct rds_sock *rs);
483 void rds_sock_put(struct rds_sock *rs);
484 void rds_wake_sk_sleep(struct rds_sock *rs);
485 static inline void __rds_wake_sk_sleep(struct sock *sk)
487 wait_queue_head_t *waitq = sk->sk_sleep;
489 if (!sock_flag(sk, SOCK_DEAD) && waitq)
492 extern wait_queue_head_t rds_poll_waitq;
496 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
497 void rds_remove_bound(struct rds_sock *rs);
498 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
501 int rds_cong_get_maps(struct rds_connection *conn);
502 void rds_cong_add_conn(struct rds_connection *conn);
503 void rds_cong_remove_conn(struct rds_connection *conn);
504 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
505 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
506 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
507 void rds_cong_queue_updates(struct rds_cong_map *map);
508 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
509 int rds_cong_updated_since(unsigned long *recent);
510 void rds_cong_add_socket(struct rds_sock *);
511 void rds_cong_remove_socket(struct rds_sock *);
512 void rds_cong_exit(void);
513 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
516 int __init rds_conn_init(void);
517 void rds_conn_exit(void);
518 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
519 struct rds_transport *trans, gfp_t gfp);
520 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
521 struct rds_transport *trans, gfp_t gfp);
522 void rds_conn_destroy(struct rds_connection *conn);
523 void rds_conn_reset(struct rds_connection *conn);
524 void rds_conn_drop(struct rds_connection *conn);
525 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
526 struct rds_info_iterator *iter,
527 struct rds_info_lengths *lens,
528 int (*visitor)(struct rds_connection *, void *),
530 void __rds_conn_error(struct rds_connection *conn, const char *, ...)
531 __attribute__ ((format (printf, 2, 3)));
532 #define rds_conn_error(conn, fmt...) \
533 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
536 rds_conn_transition(struct rds_connection *conn, int old, int new)
538 return atomic_cmpxchg(&conn->c_state, old, new) == old;
542 rds_conn_state(struct rds_connection *conn)
544 return atomic_read(&conn->c_state);
548 rds_conn_up(struct rds_connection *conn)
550 return atomic_read(&conn->c_state) == RDS_CONN_UP;
554 rds_conn_connecting(struct rds_connection *conn)
556 return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
560 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
561 struct rds_message *rds_message_copy_from_user(struct iovec *first_iov,
563 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
564 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
565 __be16 dport, u64 seq);
566 int rds_message_add_extension(struct rds_header *hdr,
567 unsigned int type, const void *data, unsigned int len);
568 int rds_message_next_extension(struct rds_header *hdr,
569 unsigned int *pos, void *buf, unsigned int *buflen);
570 int rds_message_add_version_extension(struct rds_header *hdr, unsigned int version);
571 int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *version);
572 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
573 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
574 struct iovec *first_iov, size_t size);
575 void rds_message_inc_purge(struct rds_incoming *inc);
576 void rds_message_inc_free(struct rds_incoming *inc);
577 void rds_message_addref(struct rds_message *rm);
578 void rds_message_put(struct rds_message *rm);
579 void rds_message_wait(struct rds_message *rm);
580 void rds_message_unmapped(struct rds_message *rm);
582 static inline void rds_message_make_checksum(struct rds_header *hdr)
585 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
588 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
590 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
595 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
597 int rds_page_copy_user(struct page *page, unsigned long offset,
598 void __user *ptr, unsigned long bytes,
600 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
601 rds_page_copy_user(page, offset, ptr, bytes, 1)
602 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
603 rds_page_copy_user(page, offset, ptr, bytes, 0)
604 void rds_page_exit(void);
607 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
609 void rds_inc_addref(struct rds_incoming *inc);
610 void rds_inc_put(struct rds_incoming *inc);
611 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
612 struct rds_incoming *inc, gfp_t gfp, enum km_type km);
613 int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
614 size_t size, int msg_flags);
615 void rds_clear_recv_queue(struct rds_sock *rs);
616 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
617 void rds_inc_info_copy(struct rds_incoming *inc,
618 struct rds_info_iterator *iter,
619 __be32 saddr, __be32 daddr, int flip);
622 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
624 void rds_send_reset(struct rds_connection *conn);
625 int rds_send_xmit(struct rds_connection *conn);
627 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
628 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
629 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
630 is_acked_func is_acked);
631 int rds_send_acked_before(struct rds_connection *conn, u64 seq);
632 void rds_send_remove_from_sock(struct list_head *messages, int status);
633 int rds_send_pong(struct rds_connection *conn, __be16 dport);
634 struct rds_message *rds_send_get_message(struct rds_connection *,
635 struct rds_rdma_op *);
638 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
641 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
642 #define rds_stats_inc_which(which, member) do { \
643 per_cpu(which, get_cpu()).member++; \
646 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
647 #define rds_stats_add_which(which, member, count) do { \
648 per_cpu(which, get_cpu()).member += count; \
651 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
652 int __init rds_stats_init(void);
653 void rds_stats_exit(void);
654 void rds_stats_info_copy(struct rds_info_iterator *iter,
655 uint64_t *values, char **names, size_t nr);
658 int __init rds_sysctl_init(void);
659 void rds_sysctl_exit(void);
660 extern unsigned long rds_sysctl_sndbuf_min;
661 extern unsigned long rds_sysctl_sndbuf_default;
662 extern unsigned long rds_sysctl_sndbuf_max;
663 extern unsigned long rds_sysctl_reconnect_min_jiffies;
664 extern unsigned long rds_sysctl_reconnect_max_jiffies;
665 extern unsigned int rds_sysctl_max_unacked_packets;
666 extern unsigned int rds_sysctl_max_unacked_bytes;
667 extern unsigned int rds_sysctl_ping_enable;
668 extern unsigned long rds_sysctl_trace_flags;
669 extern unsigned int rds_sysctl_trace_level;
672 int __init rds_threads_init(void);
673 void rds_threads_exit(void);
674 extern struct workqueue_struct *rds_wq;
675 void rds_connect_worker(struct work_struct *);
676 void rds_shutdown_worker(struct work_struct *);
677 void rds_send_worker(struct work_struct *);
678 void rds_recv_worker(struct work_struct *);
679 void rds_connect_complete(struct rds_connection *conn);
682 int rds_trans_register(struct rds_transport *trans);
683 void rds_trans_unregister(struct rds_transport *trans);
684 struct rds_transport *rds_trans_get_preferred(__be32 addr);
685 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
687 int __init rds_trans_init(void);
688 void rds_trans_exit(void);