xref: /openbmc/linux/net/smc/smc_close.c (revision f7d84fa7)
1 /*
2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
3  *
4  *  Socket Closing - normal and abnormal
5  *
6  *  Copyright IBM Corp. 2016
7  *
8  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
9  */
10 
11 #include <linux/workqueue.h>
12 #include <linux/sched/signal.h>
13 
14 #include <net/sock.h>
15 
16 #include "smc.h"
17 #include "smc_tx.h"
18 #include "smc_cdc.h"
19 #include "smc_close.h"
20 
21 #define SMC_CLOSE_WAIT_TX_PENDS_TIME		(5 * HZ)
22 
23 static void smc_close_cleanup_listen(struct sock *parent)
24 {
25 	struct sock *sk;
26 
27 	/* Close non-accepted connections */
28 	while ((sk = smc_accept_dequeue(parent, NULL)))
29 		smc_close_non_accepted(sk);
30 }
31 
32 static void smc_close_wait_tx_pends(struct smc_sock *smc)
33 {
34 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
35 	struct sock *sk = &smc->sk;
36 	signed long timeout;
37 
38 	timeout = SMC_CLOSE_WAIT_TX_PENDS_TIME;
39 	add_wait_queue(sk_sleep(sk), &wait);
40 	while (!signal_pending(current) && timeout) {
41 		int rc;
42 
43 		rc = sk_wait_event(sk, &timeout,
44 				   !smc_cdc_tx_has_pending(&smc->conn),
45 				   &wait);
46 		if (rc)
47 			break;
48 	}
49 	remove_wait_queue(sk_sleep(sk), &wait);
50 }
51 
52 /* wait for sndbuf data being transmitted */
53 static void smc_close_stream_wait(struct smc_sock *smc, long timeout)
54 {
55 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
56 	struct sock *sk = &smc->sk;
57 
58 	if (!timeout)
59 		return;
60 
61 	if (!smc_tx_prepared_sends(&smc->conn))
62 		return;
63 
64 	smc->wait_close_tx_prepared = 1;
65 	add_wait_queue(sk_sleep(sk), &wait);
66 	while (!signal_pending(current) && timeout) {
67 		int rc;
68 
69 		rc = sk_wait_event(sk, &timeout,
70 				   !smc_tx_prepared_sends(&smc->conn) ||
71 				   (sk->sk_err == ECONNABORTED) ||
72 				   (sk->sk_err == ECONNRESET),
73 				   &wait);
74 		if (rc)
75 			break;
76 	}
77 	remove_wait_queue(sk_sleep(sk), &wait);
78 	smc->wait_close_tx_prepared = 0;
79 }
80 
81 void smc_close_wake_tx_prepared(struct smc_sock *smc)
82 {
83 	if (smc->wait_close_tx_prepared)
84 		/* wake up socket closing */
85 		smc->sk.sk_state_change(&smc->sk);
86 }
87 
88 static int smc_close_wr(struct smc_connection *conn)
89 {
90 	conn->local_tx_ctrl.conn_state_flags.peer_done_writing = 1;
91 
92 	return smc_cdc_get_slot_and_msg_send(conn);
93 }
94 
95 static int smc_close_final(struct smc_connection *conn)
96 {
97 	if (atomic_read(&conn->bytes_to_rcv))
98 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
99 	else
100 		conn->local_tx_ctrl.conn_state_flags.peer_conn_closed = 1;
101 
102 	return smc_cdc_get_slot_and_msg_send(conn);
103 }
104 
105 static int smc_close_abort(struct smc_connection *conn)
106 {
107 	conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
108 
109 	return smc_cdc_get_slot_and_msg_send(conn);
110 }
111 
112 /* terminate smc socket abnormally - active abort
113  * RDMA communication no longer possible
114  */
115 void smc_close_active_abort(struct smc_sock *smc)
116 {
117 	struct smc_cdc_conn_state_flags *txflags =
118 		&smc->conn.local_tx_ctrl.conn_state_flags;
119 
120 	smc->sk.sk_err = ECONNABORTED;
121 	if (smc->clcsock && smc->clcsock->sk) {
122 		smc->clcsock->sk->sk_err = ECONNABORTED;
123 		smc->clcsock->sk->sk_state_change(smc->clcsock->sk);
124 	}
125 	switch (smc->sk.sk_state) {
126 	case SMC_INIT:
127 	case SMC_ACTIVE:
128 		smc->sk.sk_state = SMC_PEERABORTWAIT;
129 		break;
130 	case SMC_APPCLOSEWAIT1:
131 	case SMC_APPCLOSEWAIT2:
132 		txflags->peer_conn_abort = 1;
133 		sock_release(smc->clcsock);
134 		if (!smc_cdc_rxed_any_close(&smc->conn))
135 			smc->sk.sk_state = SMC_PEERABORTWAIT;
136 		else
137 			smc->sk.sk_state = SMC_CLOSED;
138 		break;
139 	case SMC_PEERCLOSEWAIT1:
140 	case SMC_PEERCLOSEWAIT2:
141 		if (!txflags->peer_conn_closed) {
142 			smc->sk.sk_state = SMC_PEERABORTWAIT;
143 			txflags->peer_conn_abort = 1;
144 			sock_release(smc->clcsock);
145 		} else {
146 			smc->sk.sk_state = SMC_CLOSED;
147 		}
148 		break;
149 	case SMC_PROCESSABORT:
150 	case SMC_APPFINCLOSEWAIT:
151 		if (!txflags->peer_conn_closed) {
152 			txflags->peer_conn_abort = 1;
153 			sock_release(smc->clcsock);
154 		}
155 		smc->sk.sk_state = SMC_CLOSED;
156 		break;
157 	case SMC_PEERFINCLOSEWAIT:
158 	case SMC_PEERABORTWAIT:
159 	case SMC_CLOSED:
160 		break;
161 	}
162 
163 	sock_set_flag(&smc->sk, SOCK_DEAD);
164 	smc->sk.sk_state_change(&smc->sk);
165 }
166 
167 static inline bool smc_close_sent_any_close(struct smc_connection *conn)
168 {
169 	return conn->local_tx_ctrl.conn_state_flags.peer_conn_abort ||
170 	       conn->local_tx_ctrl.conn_state_flags.peer_conn_closed;
171 }
172 
173 int smc_close_active(struct smc_sock *smc)
174 {
175 	struct smc_cdc_conn_state_flags *txflags =
176 		&smc->conn.local_tx_ctrl.conn_state_flags;
177 	long timeout = SMC_MAX_STREAM_WAIT_TIMEOUT;
178 	struct smc_connection *conn = &smc->conn;
179 	struct sock *sk = &smc->sk;
180 	int old_state;
181 	int rc = 0;
182 
183 	if (sock_flag(sk, SOCK_LINGER) &&
184 	    !(current->flags & PF_EXITING))
185 		timeout = sk->sk_lingertime;
186 
187 again:
188 	old_state = sk->sk_state;
189 	switch (old_state) {
190 	case SMC_INIT:
191 		sk->sk_state = SMC_CLOSED;
192 		if (smc->smc_listen_work.func)
193 			cancel_work_sync(&smc->smc_listen_work);
194 		break;
195 	case SMC_LISTEN:
196 		sk->sk_state = SMC_CLOSED;
197 		sk->sk_state_change(sk); /* wake up accept */
198 		if (smc->clcsock && smc->clcsock->sk) {
199 			rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
200 			/* wake up kernel_accept of smc_tcp_listen_worker */
201 			smc->clcsock->sk->sk_data_ready(smc->clcsock->sk);
202 		}
203 		release_sock(sk);
204 		smc_close_cleanup_listen(sk);
205 		cancel_work_sync(&smc->smc_listen_work);
206 		lock_sock(sk);
207 		break;
208 	case SMC_ACTIVE:
209 		smc_close_stream_wait(smc, timeout);
210 		release_sock(sk);
211 		cancel_work_sync(&conn->tx_work);
212 		lock_sock(sk);
213 		if (sk->sk_state == SMC_ACTIVE) {
214 			/* send close request */
215 			rc = smc_close_final(conn);
216 			sk->sk_state = SMC_PEERCLOSEWAIT1;
217 		} else {
218 			/* peer event has changed the state */
219 			goto again;
220 		}
221 		break;
222 	case SMC_APPFINCLOSEWAIT:
223 		/* socket already shutdown wr or both (active close) */
224 		if (txflags->peer_done_writing &&
225 		    !smc_close_sent_any_close(conn)) {
226 			/* just shutdown wr done, send close request */
227 			rc = smc_close_final(conn);
228 		}
229 		sk->sk_state = SMC_CLOSED;
230 		smc_close_wait_tx_pends(smc);
231 		break;
232 	case SMC_APPCLOSEWAIT1:
233 	case SMC_APPCLOSEWAIT2:
234 		if (!smc_cdc_rxed_any_close(conn))
235 			smc_close_stream_wait(smc, timeout);
236 		release_sock(sk);
237 		cancel_work_sync(&conn->tx_work);
238 		lock_sock(sk);
239 		if (sk->sk_err != ECONNABORTED) {
240 			/* confirm close from peer */
241 			rc = smc_close_final(conn);
242 			if (rc)
243 				break;
244 		}
245 		if (smc_cdc_rxed_any_close(conn))
246 			/* peer has closed the socket already */
247 			sk->sk_state = SMC_CLOSED;
248 		else
249 			/* peer has just issued a shutdown write */
250 			sk->sk_state = SMC_PEERFINCLOSEWAIT;
251 		smc_close_wait_tx_pends(smc);
252 		break;
253 	case SMC_PEERCLOSEWAIT1:
254 	case SMC_PEERCLOSEWAIT2:
255 		if (txflags->peer_done_writing &&
256 		    !smc_close_sent_any_close(conn)) {
257 			/* just shutdown wr done, send close request */
258 			rc = smc_close_final(conn);
259 		}
260 		/* peer sending PeerConnectionClosed will cause transition */
261 		break;
262 	case SMC_PEERFINCLOSEWAIT:
263 		/* peer sending PeerConnectionClosed will cause transition */
264 		break;
265 	case SMC_PROCESSABORT:
266 		cancel_work_sync(&conn->tx_work);
267 		smc_close_abort(conn);
268 		sk->sk_state = SMC_CLOSED;
269 		smc_close_wait_tx_pends(smc);
270 		break;
271 	case SMC_PEERABORTWAIT:
272 	case SMC_CLOSED:
273 		/* nothing to do, add tracing in future patch */
274 		break;
275 	}
276 
277 	if (old_state != sk->sk_state)
278 		sk->sk_state_change(&smc->sk);
279 	return rc;
280 }
281 
282 static void smc_close_passive_abort_received(struct smc_sock *smc)
283 {
284 	struct smc_cdc_conn_state_flags *txflags =
285 		&smc->conn.local_tx_ctrl.conn_state_flags;
286 	struct sock *sk = &smc->sk;
287 
288 	switch (sk->sk_state) {
289 	case SMC_ACTIVE:
290 	case SMC_APPFINCLOSEWAIT:
291 	case SMC_APPCLOSEWAIT1:
292 	case SMC_APPCLOSEWAIT2:
293 		smc_close_abort(&smc->conn);
294 		sk->sk_state = SMC_PROCESSABORT;
295 		break;
296 	case SMC_PEERCLOSEWAIT1:
297 	case SMC_PEERCLOSEWAIT2:
298 		if (txflags->peer_done_writing &&
299 		    !smc_close_sent_any_close(&smc->conn)) {
300 			/* just shutdown, but not yet closed locally */
301 			smc_close_abort(&smc->conn);
302 			sk->sk_state = SMC_PROCESSABORT;
303 		} else {
304 			sk->sk_state = SMC_CLOSED;
305 		}
306 		break;
307 	case SMC_PEERFINCLOSEWAIT:
308 	case SMC_PEERABORTWAIT:
309 		sk->sk_state = SMC_CLOSED;
310 		break;
311 	case SMC_INIT:
312 	case SMC_PROCESSABORT:
313 	/* nothing to do, add tracing in future patch */
314 		break;
315 	}
316 }
317 
318 /* Some kind of closing has been received: peer_conn_closed, peer_conn_abort,
319  * or peer_done_writing.
320  */
321 static void smc_close_passive_work(struct work_struct *work)
322 {
323 	struct smc_connection *conn = container_of(work,
324 						   struct smc_connection,
325 						   close_work);
326 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
327 	struct smc_cdc_conn_state_flags *rxflags;
328 	struct sock *sk = &smc->sk;
329 	int old_state;
330 
331 	lock_sock(&smc->sk);
332 	old_state = sk->sk_state;
333 
334 	if (!conn->alert_token_local) {
335 		/* abnormal termination */
336 		smc_close_active_abort(smc);
337 		goto wakeup;
338 	}
339 
340 	rxflags = &smc->conn.local_rx_ctrl.conn_state_flags;
341 	if (rxflags->peer_conn_abort) {
342 		smc_close_passive_abort_received(smc);
343 		goto wakeup;
344 	}
345 
346 	switch (sk->sk_state) {
347 	case SMC_INIT:
348 		if (atomic_read(&smc->conn.bytes_to_rcv) ||
349 		    (rxflags->peer_done_writing &&
350 		     !smc_cdc_rxed_any_close(conn)))
351 			sk->sk_state = SMC_APPCLOSEWAIT1;
352 		else
353 			sk->sk_state = SMC_CLOSED;
354 		break;
355 	case SMC_ACTIVE:
356 		sk->sk_state = SMC_APPCLOSEWAIT1;
357 		break;
358 	case SMC_PEERCLOSEWAIT1:
359 		if (rxflags->peer_done_writing)
360 			sk->sk_state = SMC_PEERCLOSEWAIT2;
361 		/* fall through to check for closing */
362 	case SMC_PEERCLOSEWAIT2:
363 	case SMC_PEERFINCLOSEWAIT:
364 		if (!smc_cdc_rxed_any_close(&smc->conn))
365 			break;
366 		if (sock_flag(sk, SOCK_DEAD) &&
367 		    smc_close_sent_any_close(conn)) {
368 			/* smc_release has already been called locally */
369 			sk->sk_state = SMC_CLOSED;
370 		} else {
371 			/* just shutdown, but not yet closed locally */
372 			sk->sk_state = SMC_APPFINCLOSEWAIT;
373 		}
374 		break;
375 	case SMC_APPCLOSEWAIT1:
376 	case SMC_APPCLOSEWAIT2:
377 	case SMC_APPFINCLOSEWAIT:
378 	case SMC_PEERABORTWAIT:
379 	case SMC_PROCESSABORT:
380 	case SMC_CLOSED:
381 		/* nothing to do, add tracing in future patch */
382 		break;
383 	}
384 
385 wakeup:
386 	sk->sk_data_ready(sk); /* wakeup blocked rcvbuf consumers */
387 	sk->sk_write_space(sk); /* wakeup blocked sndbuf producers */
388 
389 	if (old_state != sk->sk_state) {
390 		sk->sk_state_change(sk);
391 		if ((sk->sk_state == SMC_CLOSED) &&
392 		    (sock_flag(sk, SOCK_DEAD) || !sk->sk_socket)) {
393 			smc_conn_free(&smc->conn);
394 			schedule_delayed_work(&smc->sock_put_work,
395 					      SMC_CLOSE_SOCK_PUT_DELAY);
396 		}
397 	}
398 	release_sock(&smc->sk);
399 }
400 
401 void smc_close_sock_put_work(struct work_struct *work)
402 {
403 	struct smc_sock *smc = container_of(to_delayed_work(work),
404 					    struct smc_sock,
405 					    sock_put_work);
406 
407 	smc->sk.sk_prot->unhash(&smc->sk);
408 	sock_put(&smc->sk);
409 }
410 
411 int smc_close_shutdown_write(struct smc_sock *smc)
412 {
413 	struct smc_connection *conn = &smc->conn;
414 	long timeout = SMC_MAX_STREAM_WAIT_TIMEOUT;
415 	struct sock *sk = &smc->sk;
416 	int old_state;
417 	int rc = 0;
418 
419 	if (sock_flag(sk, SOCK_LINGER))
420 		timeout = sk->sk_lingertime;
421 
422 again:
423 	old_state = sk->sk_state;
424 	switch (old_state) {
425 	case SMC_ACTIVE:
426 		smc_close_stream_wait(smc, timeout);
427 		release_sock(sk);
428 		cancel_work_sync(&conn->tx_work);
429 		lock_sock(sk);
430 		/* send close wr request */
431 		rc = smc_close_wr(conn);
432 		if (sk->sk_state == SMC_ACTIVE)
433 			sk->sk_state = SMC_PEERCLOSEWAIT1;
434 		else
435 			goto again;
436 		break;
437 	case SMC_APPCLOSEWAIT1:
438 		/* passive close */
439 		if (!smc_cdc_rxed_any_close(conn))
440 			smc_close_stream_wait(smc, timeout);
441 		release_sock(sk);
442 		cancel_work_sync(&conn->tx_work);
443 		lock_sock(sk);
444 		/* confirm close from peer */
445 		rc = smc_close_wr(conn);
446 		sk->sk_state = SMC_APPCLOSEWAIT2;
447 		break;
448 	case SMC_APPCLOSEWAIT2:
449 	case SMC_PEERFINCLOSEWAIT:
450 	case SMC_PEERCLOSEWAIT1:
451 	case SMC_PEERCLOSEWAIT2:
452 	case SMC_APPFINCLOSEWAIT:
453 	case SMC_PROCESSABORT:
454 	case SMC_PEERABORTWAIT:
455 		/* nothing to do, add tracing in future patch */
456 		break;
457 	}
458 
459 	if (old_state != sk->sk_state)
460 		sk->sk_state_change(&smc->sk);
461 	return rc;
462 }
463 
464 /* Initialize close properties on connection establishment. */
465 void smc_close_init(struct smc_sock *smc)
466 {
467 	INIT_WORK(&smc->conn.close_work, smc_close_passive_work);
468 }
469