1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/ulnds/socklnd/usocklnd.h
38 * Author: Maxim Patlasov <maxim@clusterfs.com>
45 #include <lnet/lib-lnet.h>
46 #include <lnet/socklnd.h>
49 cfs_list_t tx_list; /* neccessary to form tx list */
50 lnet_msg_t *tx_lnetmsg; /* lnet message for lnet_finalize() */
51 ksock_msg_t tx_msg; /* buffer for wire header of ksock msg */
52 int tx_resid; /* # of residual bytes */
53 int tx_nob; /* # of packet bytes */
54 int tx_size; /* size of this descriptor */
55 struct iovec *tx_iov; /* points to tx_iova[i] */
56 int tx_niov; /* # of packet iovec frags */
57 struct iovec tx_iova[1]; /* iov for header */
63 cfs_socket_t *uc_sock; /* socket */
64 int uc_type; /* conn type */
65 int uc_activeflag; /* active side of connection? */
66 int uc_flip; /* is peer other endian? */
67 int uc_state; /* connection state */
68 struct usock_peer_s *uc_peer; /* owning peer */
69 lnet_process_id_t uc_peerid; /* id of remote peer */
70 int uc_pt_idx; /* index in ud_pollthreads[] of
71 * owning poll thread */
72 lnet_ni_t *uc_ni; /* parent NI while accepting */
73 struct usock_preq_s *uc_preq; /* preallocated request */
74 __u32 uc_peer_ip; /* IP address of the peer */
75 __u16 uc_peer_port; /* port of the peer */
76 cfs_list_t uc_stale_list; /* orphaned connections */
79 int uc_rx_state; /* message or hello state */
80 ksock_hello_msg_t *uc_rx_hello; /* hello buffer */
81 struct iovec *uc_rx_iov; /* points to uc_rx_iova[i] */
82 struct iovec uc_rx_iova[LNET_MAX_IOV]; /* message frags */
83 int uc_rx_niov; /* # frags */
84 int uc_rx_nob_left; /* # bytes to next hdr/body */
85 int uc_rx_nob_wanted; /* # of bytes actually wanted */
86 void *uc_rx_lnetmsg; /* LNET message being received */
87 cfs_time_t uc_rx_deadline; /* when to time out */
88 int uc_rx_flag; /* deadline valid? */
89 ksock_msg_t uc_rx_msg; /* message buffer */
92 cfs_list_t uc_tx_list; /* pending txs */
93 cfs_list_t uc_zcack_list; /* pending zc_acks */
94 cfs_time_t uc_tx_deadline; /* when to time out */
95 int uc_tx_flag; /* deadline valid? */
96 int uc_sending; /* send op is in progress */
97 usock_tx_t *uc_tx_hello; /* fake tx with hello */
99 cfs_mt_atomic_t uc_refcount; /* # of users */
100 pthread_mutex_t uc_lock; /* serialize */
101 int uc_errored; /* a flag for lnet_notify() */
104 /* Allowable conn states are: */
105 #define UC_CONNECTING 1
106 #define UC_SENDING_HELLO 2
107 #define UC_RECEIVING_HELLO 3
111 /* Allowable RX states are: */
112 #define UC_RX_HELLO_MAGIC 1
113 #define UC_RX_HELLO_VERSION 2
114 #define UC_RX_HELLO_BODY 3
115 #define UC_RX_HELLO_IPS 4
116 #define UC_RX_KSM_HEADER 5
117 #define UC_RX_LNET_HEADER 6
118 #define UC_RX_PARSE 7
119 #define UC_RX_PARSE_WAIT 8
120 #define UC_RX_LNET_PAYLOAD 9
121 #define UC_RX_SKIPPING 10
123 #define N_CONN_TYPES 3 /* CONTROL, BULK_IN and BULK_OUT */
125 typedef struct usock_peer_s {
126 cfs_list_t up_list; /* neccessary to form peer list */
127 lnet_process_id_t up_peerid; /* id of remote peer */
128 usock_conn_t *up_conns[N_CONN_TYPES]; /* conns that connect us
129 * us with the peer */
130 lnet_ni_t *up_ni; /* pointer to parent NI */
131 __u64 up_incarnation; /* peer's incarnation */
132 int up_incrn_is_set;/* 0 if peer's incarnation
133 * hasn't been set so far */
134 cfs_mt_atomic_t up_refcount; /* # of users */
135 pthread_mutex_t up_lock; /* serialize */
136 int up_errored; /* a flag for lnet_notify() */
137 cfs_time_t up_last_alive; /* when the peer was last alive */
141 cfs_socket_t *upt_notifier[2]; /* notifier sockets: 1st for
142 * writing, 2nd for reading */
143 struct pollfd *upt_pollfd; /* poll fds */
144 int upt_nfds; /* active poll fds */
145 int upt_npollfd; /* allocated poll fds */
146 usock_conn_t **upt_idx2conn; /* conns corresponding to
148 int *upt_skip; /* skip chain */
149 int *upt_fd2idx; /* index into upt_pollfd[]
151 int upt_nfd2idx; /* # of allocated elements
153 cfs_list_t upt_stale_list; /* list of orphaned conns */
154 cfs_list_t upt_pollrequests; /* list of poll requests */
155 pthread_mutex_t upt_pollrequests_lock; /* serialize */
156 int upt_errno; /* non-zero if errored */
157 cfs_mt_completion_t upt_completion; /* wait/signal facility for
158 * syncronizing shutdown */
159 } usock_pollthread_t;
161 /* Number of elements in upt_pollfd[], upt_idx2conn[] and upt_fd2idx[]
162 * at initialization time. Will be resized on demand */
163 #define UPT_START_SIZ 32
166 #define UD_PEER_HASH_SIZE 101
169 int ud_state; /* initialization state */
170 int ud_npollthreads; /* # of poll threads */
171 usock_pollthread_t *ud_pollthreads; /* their state */
172 int ud_shutdown; /* shutdown flag */
173 int ud_nets_count; /* # of instances */
174 cfs_list_t ud_peers[UD_PEER_HASH_SIZE]; /* peer hash table */
175 pthread_rwlock_t ud_peers_lock; /* serialize */
178 extern usock_data_t usock_data;
180 /* ud_state allowed values */
181 #define UD_STATE_INIT_NOTHING 0
182 #define UD_STATE_INITIALIZED 1
185 int un_peercount; /* # of peers */
186 int un_shutdown; /* shutdown flag */
187 __u64 un_incarnation; /* my epoch */
188 pthread_cond_t un_cond; /* condvar to wait for notifications */
189 pthread_mutex_t un_lock; /* a lock to protect un_cond */
193 int ut_poll_timeout; /* the third arg for poll(2) (seconds) */
194 int ut_timeout; /* "stuck" socket timeout (seconds) */
195 int ut_npollthreads; /* number of poll thread to spawn */
196 int ut_fair_limit; /* how many packets can we receive or transmit
197 * without calling poll(2) */
198 int ut_min_bulk; /* smallest "large" message */
199 int ut_txcredits; /* # concurrent sends */
200 int ut_peertxcredits; /* # concurrent sends to 1 peer */
201 int ut_socknagle; /* Is Nagle alg on ? */
202 int ut_sockbufsiz; /* size of socket buffers */
205 extern usock_tunables_t usock_tuns;
207 typedef struct usock_preq_s {
208 int upr_type; /* type of requested action */
209 short upr_value; /* bitmask of POLLIN and POLLOUT bits */
210 usock_conn_t * upr_conn; /* a conn for the sake of which
211 * action will be performed */
212 cfs_list_t upr_list; /* neccessary to form list */
213 } usock_pollrequest_t;
215 /* Allowable poll request types are: */
216 #define POLL_ADD_REQUEST 1
217 #define POLL_DEL_REQUEST 2
218 #define POLL_RX_SET_REQUEST 3
219 #define POLL_TX_SET_REQUEST 4
220 #define POLL_SET_REQUEST 5
223 cfs_list_t zc_list; /* neccessary to form zc_ack list */
224 __u64 zc_cookie; /* zero-copy cookie */
228 usocklnd_conn_addref(usock_conn_t *conn)
230 LASSERT (cfs_mt_atomic_read(&conn->uc_refcount) > 0);
231 cfs_mt_atomic_inc(&conn->uc_refcount);
234 void usocklnd_destroy_conn(usock_conn_t *conn);
237 usocklnd_conn_decref(usock_conn_t *conn)
239 LASSERT (cfs_mt_atomic_read(&conn->uc_refcount) > 0);
240 if (cfs_mt_atomic_dec_and_test(&conn->uc_refcount))
241 usocklnd_destroy_conn(conn);
245 usocklnd_peer_addref(usock_peer_t *peer)
247 LASSERT (cfs_mt_atomic_read(&peer->up_refcount) > 0);
248 cfs_mt_atomic_inc(&peer->up_refcount);
251 void usocklnd_destroy_peer(usock_peer_t *peer);
254 usocklnd_peer_decref(usock_peer_t *peer)
256 LASSERT (cfs_mt_atomic_read(&peer->up_refcount) > 0);
257 if (cfs_mt_atomic_dec_and_test(&peer->up_refcount))
258 usocklnd_destroy_peer(peer);
262 usocklnd_ip2pt_idx(__u32 ip) {
263 return ip % usock_data.ud_npollthreads;
266 static inline cfs_list_t *
267 usocklnd_nid2peerlist(lnet_nid_t nid)
269 unsigned int hash = ((unsigned int)nid) % UD_PEER_HASH_SIZE;
271 return &usock_data.ud_peers[hash];
274 int usocklnd_startup(lnet_ni_t *ni);
275 void usocklnd_shutdown(lnet_ni_t *ni);
276 int usocklnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
277 int usocklnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
278 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
279 unsigned int offset, unsigned int mlen, unsigned int rlen);
280 int usocklnd_accept(lnet_ni_t *ni, cfs_socket_t *sock);
282 int usocklnd_poll_thread(void *arg);
283 int usocklnd_add_pollrequest(usock_conn_t *conn, int type, short value);
284 void usocklnd_add_killrequest(usock_conn_t *conn);
285 int usocklnd_process_pollrequest(usock_pollrequest_t *pr,
286 usock_pollthread_t *pt_data);
287 void usocklnd_execute_handlers(usock_pollthread_t *pt_data);
288 int usocklnd_calculate_chunk_size(int num);
289 void usocklnd_wakeup_pollthread(int i);
291 int usocklnd_notifier_handler(int fd);
292 void usocklnd_exception_handler(usock_conn_t *conn);
293 int usocklnd_read_handler(usock_conn_t *conn);
294 int usocklnd_read_msg(usock_conn_t *conn, int *cont_flag);
295 int usocklnd_handle_zc_req(usock_peer_t *peer, __u64 cookie);
296 int usocklnd_read_hello(usock_conn_t *conn, int *cont_flag);
297 int usocklnd_activeconn_hellorecv(usock_conn_t *conn);
298 int usocklnd_passiveconn_hellorecv(usock_conn_t *conn);
299 int usocklnd_write_handler(usock_conn_t *conn);
300 usock_tx_t * usocklnd_try_piggyback(cfs_list_t *tx_list_p,
301 cfs_list_t *zcack_list_p);
302 int usocklnd_activeconn_hellosent(usock_conn_t *conn);
303 int usocklnd_passiveconn_hellosent(usock_conn_t *conn);
304 int usocklnd_send_tx(usock_conn_t *conn, usock_tx_t *tx);
305 int usocklnd_read_data(usock_conn_t *conn);
307 void usocklnd_release_poll_states(int n);
308 int usocklnd_base_startup();
309 void usocklnd_base_shutdown(int n);
310 __u64 usocklnd_new_incarnation();
311 void usocklnd_del_all_peers(lnet_ni_t *ni);
312 void usocklnd_del_peer_and_conns(usock_peer_t *peer);
313 void usocklnd_del_conns_locked(usock_peer_t *peer);
315 int usocklnd_conn_timed_out(usock_conn_t *conn, cfs_time_t current_time);
316 void usocklnd_conn_kill(usock_conn_t *conn);
317 void usocklnd_conn_kill_locked(usock_conn_t *conn);
318 usock_conn_t *usocklnd_conn_allocate();
319 void usocklnd_conn_free(usock_conn_t *conn);
320 void usocklnd_tear_peer_conn(usock_conn_t *conn);
321 void usocklnd_check_peer_stale(lnet_ni_t *ni, lnet_process_id_t id);
322 int usocklnd_create_passive_conn(lnet_ni_t *ni,
323 cfs_socket_t *sock, usock_conn_t **connp);
324 int usocklnd_create_active_conn(usock_peer_t *peer, int type,
325 usock_conn_t **connp);
326 int usocklnd_connect_srv_mode(cfs_socket_t **sockp,
327 __u32 dst_ip, __u16 dst_port);
328 int usocklnd_connect_cli_mode(cfs_socket_t **sockp,
329 __u32 dst_ip, __u16 dst_port);
330 int usocklnd_set_sock_options(cfs_socket_t *sock);
331 usock_tx_t *usocklnd_create_noop_tx(__u64 cookie);
332 usock_tx_t *usocklnd_create_tx(lnet_msg_t *lntmsg);
333 void usocklnd_init_hello_msg(ksock_hello_msg_t *hello,
334 lnet_ni_t *ni, int type, lnet_nid_t peer_nid);
335 usock_tx_t *usocklnd_create_hello_tx(lnet_ni_t *ni,
336 int type, lnet_nid_t peer_nid);
337 usock_tx_t *usocklnd_create_cr_hello_tx(lnet_ni_t *ni,
338 int type, lnet_nid_t peer_nid);
339 void usocklnd_destroy_tx(lnet_ni_t *ni, usock_tx_t *tx);
340 void usocklnd_destroy_txlist(lnet_ni_t *ni, cfs_list_t *txlist);
341 void usocklnd_destroy_zcack_list(cfs_list_t *zcack_list);
342 void usocklnd_destroy_peer (usock_peer_t *peer);
343 int usocklnd_get_conn_type(lnet_msg_t *lntmsg);
344 int usocklnd_type2idx(int type);
345 usock_peer_t *usocklnd_find_peer_locked(lnet_ni_t *ni, lnet_process_id_t id);
346 int usocklnd_create_peer(lnet_ni_t *ni, lnet_process_id_t id,
347 usock_peer_t **peerp);
348 int usocklnd_find_or_create_peer(lnet_ni_t *ni, lnet_process_id_t id,
349 usock_peer_t **peerp);
350 int usocklnd_find_or_create_conn(usock_peer_t *peer, int type,
351 usock_conn_t **connp,
352 usock_tx_t *tx, usock_zc_ack_t *zc_ack,
353 int *send_immediately_flag);
354 void usocklnd_link_conn_to_peer(usock_conn_t *conn, usock_peer_t *peer, int idx);
355 int usocklnd_invert_type(int type);
356 void usocklnd_conn_new_state(usock_conn_t *conn, int new_state);
357 void usocklnd_cleanup_stale_conns(usock_peer_t *peer, __u64 incrn,
358 usock_conn_t *skip_conn);
360 void usocklnd_rx_hellomagic_state_transition(usock_conn_t *conn);
361 void usocklnd_rx_helloversion_state_transition(usock_conn_t *conn);
362 void usocklnd_rx_hellobody_state_transition(usock_conn_t *conn);
363 void usocklnd_rx_helloIPs_state_transition(usock_conn_t *conn);
364 void usocklnd_rx_lnethdr_state_transition(usock_conn_t *conn);
365 void usocklnd_rx_ksmhdr_state_transition(usock_conn_t *conn);
366 void usocklnd_rx_skipping_state_transition(usock_conn_t *conn);