1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
5 * Author: Maxim Patlasov <maxim@clusterfs.com>
7 * This file is part of the Lustre file system, http://www.lustre.org
8 * Lustre is a trademark of Cluster File Systems, Inc.
14 #include <lnet/lib-lnet.h>
15 #include <lnet/socklnd.h>
18 struct list_head tx_list; /* neccessary to form tx list */
19 lnet_msg_t *tx_lnetmsg; /* lnet message for lnet_finalize() */
20 ksock_msg_t tx_msg; /* buffer for wire header of ksock msg */
21 int tx_resid; /* # of residual bytes */
22 int tx_nob; /* # of packet bytes */
23 int tx_size; /* size of this descriptor */
24 struct iovec *tx_iov; /* points to tx_iova[i] */
25 int tx_niov; /* # of packet iovec frags */
26 struct iovec tx_iova[1]; /* iov for header */
32 int uc_fd; /* socket */
33 int uc_type; /* conn type */
34 int uc_activeflag; /* active side of connection? */
35 int uc_flip; /* is peer other endian? */
36 int uc_state; /* connection state */
37 struct usock_peer_s *uc_peer; /* owning peer */
38 lnet_process_id_t uc_peerid; /* id of remote peer */
39 int uc_pt_idx; /* index in ud_pollthreads[] of
40 * owning poll thread */
41 lnet_ni_t *uc_ni; /* parent NI while accepting */
42 struct usock_preq_s *uc_preq; /* preallocated request */
43 __u32 uc_peer_ip; /* IP address of the peer */
44 __u16 uc_peer_port; /* port of the peer */
45 struct list_head uc_stale_list; /* orphaned connections */
48 int uc_rx_state; /* message or hello state */
49 ksock_hello_msg_t *uc_rx_hello; /* hello buffer */
50 struct iovec *uc_rx_iov; /* points to uc_rx_iova[i] */
51 struct iovec uc_rx_iova[LNET_MAX_IOV]; /* message frags */
52 int uc_rx_niov; /* # frags */
53 int uc_rx_nob_left; /* # bytes to next hdr/body */
54 int uc_rx_nob_wanted; /* # of bytes actually wanted */
55 void *uc_rx_lnetmsg; /* LNET message being received */
56 cfs_time_t uc_rx_deadline; /* when to time out */
57 int uc_rx_flag; /* deadline valid? */
58 ksock_msg_t uc_rx_msg; /* message buffer */
61 struct list_head uc_tx_list; /* pending txs */
62 struct list_head uc_zcack_list; /* pending zc_acks */
63 cfs_time_t uc_tx_deadline; /* when to time out */
64 int uc_tx_flag; /* deadline valid? */
65 int uc_sending; /* send op is in progress */
66 usock_tx_t *uc_tx_hello; /* fake tx with hello */
68 cfs_atomic_t uc_refcount; /* # of users */
69 pthread_mutex_t uc_lock; /* serialize */
70 int uc_errored; /* a flag for lnet_notify() */
73 /* Allowable conn states are: */
74 #define UC_CONNECTING 1
75 #define UC_SENDING_HELLO 2
76 #define UC_RECEIVING_HELLO 3
80 /* Allowable RX states are: */
81 #define UC_RX_HELLO_MAGIC 1
82 #define UC_RX_HELLO_VERSION 2
83 #define UC_RX_HELLO_BODY 3
84 #define UC_RX_HELLO_IPS 4
85 #define UC_RX_KSM_HEADER 5
86 #define UC_RX_LNET_HEADER 6
88 #define UC_RX_PARSE_WAIT 8
89 #define UC_RX_LNET_PAYLOAD 9
90 #define UC_RX_SKIPPING 10
92 #define N_CONN_TYPES 3 /* CONTROL, BULK_IN and BULK_OUT */
94 typedef struct usock_peer_s {
95 struct list_head up_list; /* neccessary to form peer list */
96 lnet_process_id_t up_peerid; /* id of remote peer */
97 usock_conn_t *up_conns[N_CONN_TYPES]; /* conns that connect us
99 lnet_ni_t *up_ni; /* pointer to parent NI */
100 __u64 up_incarnation; /* peer's incarnation */
101 int up_incrn_is_set; /* 0 if peer's incarnation
102 * hasn't been set so far */
103 cfs_atomic_t up_refcount; /* # of users */
104 pthread_mutex_t up_lock; /* serialize */
105 int up_errored; /* a flag for lnet_notify() */
106 cfs_time_t up_last_alive; /* when the peer was last alive */
110 int upt_notifier_fd; /* notifier fd for writing */
111 struct pollfd *upt_pollfd; /* poll fds */
112 int upt_nfds; /* active poll fds */
113 int upt_npollfd; /* allocated poll fds */
114 usock_conn_t **upt_idx2conn; /* conns corresponding to
116 int *upt_skip; /* skip chain */
117 int *upt_fd2idx; /* index into upt_pollfd[]
119 int upt_nfd2idx; /* # of allocated elements
121 struct list_head upt_stale_list; /* list of orphaned conns */
122 struct list_head upt_pollrequests; /* list of poll requests */
123 pthread_mutex_t upt_pollrequests_lock; /* serialize */
124 int upt_errno; /* non-zero if errored */
125 struct cfs_completion upt_completion; /* wait/signal facility for
126 * syncronizing shutdown */
127 } usock_pollthread_t;
129 /* Number of elements in upt_pollfd[], upt_idx2conn[] and upt_fd2idx[]
130 * at initialization time. Will be resized on demand */
131 #define UPT_START_SIZ 32
134 #define UD_PEER_HASH_SIZE 101
137 int ud_state; /* initialization state */
138 int ud_npollthreads; /* # of poll threads */
139 usock_pollthread_t *ud_pollthreads; /* their state */
140 int ud_shutdown; /* shutdown flag */
141 int ud_nets_count; /* # of instances */
142 struct list_head ud_peers[UD_PEER_HASH_SIZE]; /* peer hash table */
143 pthread_rwlock_t ud_peers_lock; /* serialize */
146 extern usock_data_t usock_data;
148 /* ud_state allowed values */
149 #define UD_STATE_INIT_NOTHING 0
150 #define UD_STATE_INITIALIZED 1
153 int un_peercount; /* # of peers */
154 int un_shutdown; /* shutdown flag */
155 __u64 un_incarnation; /* my epoch */
156 pthread_cond_t un_cond; /* condvar to wait for notifications */
157 pthread_mutex_t un_lock; /* a lock to protect un_cond */
161 int ut_poll_timeout; /* the third arg for poll(2) (seconds) */
162 int ut_timeout; /* "stuck" socket timeout (seconds) */
163 int ut_npollthreads; /* number of poll thread to spawn */
164 int ut_fair_limit; /* how many packets can we receive or transmit
165 * without calling poll(2) */
166 int ut_min_bulk; /* smallest "large" message */
167 int ut_txcredits; /* # concurrent sends */
168 int ut_peertxcredits; /* # concurrent sends to 1 peer */
169 int ut_socknagle; /* Is Nagle alg on ? */
170 int ut_sockbufsiz; /* size of socket buffers */
173 extern usock_tunables_t usock_tuns;
175 typedef struct usock_preq_s {
176 int upr_type; /* type of requested action */
177 short upr_value; /* bitmask of POLLIN and POLLOUT bits */
178 usock_conn_t * upr_conn; /* a conn for the sake of which
179 * action will be performed */
180 struct list_head upr_list; /* neccessary to form list */
181 } usock_pollrequest_t;
183 /* Allowable poll request types are: */
184 #define POLL_ADD_REQUEST 1
185 #define POLL_DEL_REQUEST 2
186 #define POLL_RX_SET_REQUEST 3
187 #define POLL_TX_SET_REQUEST 4
188 #define POLL_SET_REQUEST 5
191 struct list_head zc_list; /* neccessary to form zc_ack list */
192 __u64 zc_cookie; /* zero-copy cookie */
196 usocklnd_conn_addref(usock_conn_t *conn)
198 LASSERT (cfs_atomic_read(&conn->uc_refcount) > 0);
199 cfs_atomic_inc(&conn->uc_refcount);
202 void usocklnd_destroy_conn(usock_conn_t *conn);
205 usocklnd_conn_decref(usock_conn_t *conn)
207 LASSERT (cfs_atomic_read(&conn->uc_refcount) > 0);
208 if (cfs_atomic_dec_and_test(&conn->uc_refcount))
209 usocklnd_destroy_conn(conn);
213 usocklnd_peer_addref(usock_peer_t *peer)
215 LASSERT (cfs_atomic_read(&peer->up_refcount) > 0);
216 cfs_atomic_inc(&peer->up_refcount);
219 void usocklnd_destroy_peer(usock_peer_t *peer);
222 usocklnd_peer_decref(usock_peer_t *peer)
224 LASSERT (cfs_atomic_read(&peer->up_refcount) > 0);
225 if (cfs_atomic_dec_and_test(&peer->up_refcount))
226 usocklnd_destroy_peer(peer);
230 usocklnd_ip2pt_idx(__u32 ip) {
231 return ip % usock_data.ud_npollthreads;
234 static inline struct list_head *
235 usocklnd_nid2peerlist(lnet_nid_t nid)
237 unsigned int hash = ((unsigned int)nid) % UD_PEER_HASH_SIZE;
239 return &usock_data.ud_peers[hash];
242 int usocklnd_startup(lnet_ni_t *ni);
243 void usocklnd_shutdown(lnet_ni_t *ni);
244 int usocklnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
245 int usocklnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
246 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
247 unsigned int offset, unsigned int mlen, unsigned int rlen);
248 int usocklnd_accept(lnet_ni_t *ni, int sock_fd);
250 int usocklnd_poll_thread(void *arg);
251 int usocklnd_add_pollrequest(usock_conn_t *conn, int type, short value);
252 void usocklnd_add_killrequest(usock_conn_t *conn);
253 int usocklnd_process_pollrequest(usock_pollrequest_t *pr,
254 usock_pollthread_t *pt_data);
255 void usocklnd_execute_handlers(usock_pollthread_t *pt_data);
256 int usocklnd_calculate_chunk_size(int num);
257 void usocklnd_wakeup_pollthread(int i);
259 int usocklnd_notifier_handler(int fd);
260 void usocklnd_exception_handler(usock_conn_t *conn);
261 int usocklnd_read_handler(usock_conn_t *conn);
262 int usocklnd_read_msg(usock_conn_t *conn, int *cont_flag);
263 int usocklnd_handle_zc_req(usock_peer_t *peer, __u64 cookie);
264 int usocklnd_read_hello(usock_conn_t *conn, int *cont_flag);
265 int usocklnd_activeconn_hellorecv(usock_conn_t *conn);
266 int usocklnd_passiveconn_hellorecv(usock_conn_t *conn);
267 int usocklnd_write_handler(usock_conn_t *conn);
268 usock_tx_t * usocklnd_try_piggyback(struct list_head *tx_list_p,
269 struct list_head *zcack_list_p);
270 int usocklnd_activeconn_hellosent(usock_conn_t *conn);
271 int usocklnd_passiveconn_hellosent(usock_conn_t *conn);
272 int usocklnd_send_tx(usock_conn_t *conn, usock_tx_t *tx);
273 int usocklnd_read_data(usock_conn_t *conn);
275 void usocklnd_release_poll_states(int n);
276 int usocklnd_base_startup();
277 void usocklnd_base_shutdown(int n);
278 __u64 usocklnd_new_incarnation();
279 void usocklnd_del_all_peers(lnet_ni_t *ni);
280 void usocklnd_del_peer_and_conns(usock_peer_t *peer);
281 void usocklnd_del_conns_locked(usock_peer_t *peer);
283 int usocklnd_conn_timed_out(usock_conn_t *conn, cfs_time_t current_time);
284 void usocklnd_conn_kill(usock_conn_t *conn);
285 void usocklnd_conn_kill_locked(usock_conn_t *conn);
286 usock_conn_t *usocklnd_conn_allocate();
287 void usocklnd_conn_free(usock_conn_t *conn);
288 void usocklnd_tear_peer_conn(usock_conn_t *conn);
289 void usocklnd_check_peer_stale(lnet_ni_t *ni, lnet_process_id_t id);
290 int usocklnd_create_passive_conn(lnet_ni_t *ni, int fd, usock_conn_t **connp);
291 int usocklnd_create_active_conn(usock_peer_t *peer, int type,
292 usock_conn_t **connp);
293 int usocklnd_connect_srv_mode(int *fdp, __u32 dst_ip, __u16 dst_port);
294 int usocklnd_connect_cli_mode(int *fdp, __u32 dst_ip, __u16 dst_port);
295 int usocklnd_set_sock_options(int fd);
296 void usocklnd_init_msg(ksock_msg_t *msg, int type);
297 usock_tx_t *usocklnd_create_noop_tx(__u64 cookie);
298 usock_tx_t *usocklnd_create_tx(lnet_msg_t *lntmsg);
299 void usocklnd_init_hello_msg(ksock_hello_msg_t *hello,
300 lnet_ni_t *ni, int type, lnet_nid_t peer_nid);
301 usock_tx_t *usocklnd_create_hello_tx(lnet_ni_t *ni,
302 int type, lnet_nid_t peer_nid);
303 usock_tx_t *usocklnd_create_cr_hello_tx(lnet_ni_t *ni,
304 int type, lnet_nid_t peer_nid);
305 void usocklnd_destroy_tx(lnet_ni_t *ni, usock_tx_t *tx);
306 void usocklnd_destroy_txlist(lnet_ni_t *ni, struct list_head *txlist);
307 void usocklnd_destroy_zcack_list(struct list_head *zcack_list);
308 void usocklnd_destroy_peer (usock_peer_t *peer);
309 int usocklnd_get_conn_type(lnet_msg_t *lntmsg);
310 int usocklnd_type2idx(int type);
311 usock_peer_t *usocklnd_find_peer_locked(lnet_ni_t *ni, lnet_process_id_t id);
312 int usocklnd_create_peer(lnet_ni_t *ni, lnet_process_id_t id,
313 usock_peer_t **peerp);
314 int usocklnd_find_or_create_peer(lnet_ni_t *ni, lnet_process_id_t id,
315 usock_peer_t **peerp);
316 int usocklnd_find_or_create_conn(usock_peer_t *peer, int type,
317 usock_conn_t **connp,
318 usock_tx_t *tx, usock_zc_ack_t *zc_ack,
319 int *send_immediately_flag);
320 void usocklnd_link_conn_to_peer(usock_conn_t *conn, usock_peer_t *peer, int idx);
321 int usocklnd_invert_type(int type);
322 void usocklnd_conn_new_state(usock_conn_t *conn, int new_state);
323 void usocklnd_cleanup_stale_conns(usock_peer_t *peer, __u64 incrn,
324 usock_conn_t *skip_conn);
326 void usocklnd_rx_hellomagic_state_transition(usock_conn_t *conn);
327 void usocklnd_rx_helloversion_state_transition(usock_conn_t *conn);
328 void usocklnd_rx_hellobody_state_transition(usock_conn_t *conn);
329 void usocklnd_rx_helloIPs_state_transition(usock_conn_t *conn);
330 void usocklnd_rx_lnethdr_state_transition(usock_conn_t *conn);
331 void usocklnd_rx_ksmhdr_state_transition(usock_conn_t *conn);
332 void usocklnd_rx_skipping_state_transition(usock_conn_t *conn);