• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
4  *
5  *  Socket Closing - normal and abnormal
6  *
7  *  Copyright IBM Corp. 2016
8  *
9  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
10  */
11 
12 #include <linux/workqueue.h>
13 #include <linux/sched/signal.h>
14 
15 #include <net/sock.h>
16 #include <net/tcp.h>
17 
18 #include "smc.h"
19 #include "smc_tx.h"
20 #include "smc_cdc.h"
21 #include "smc_close.h"
22 
23 /* release the clcsock that is assigned to the smc_sock */
smc_clcsock_release(struct smc_sock * smc)24 void smc_clcsock_release(struct smc_sock *smc)
25 {
26 	struct socket *tcp;
27 
28 	if (smc->listen_smc && current_work() != &smc->smc_listen_work)
29 		cancel_work_sync(&smc->smc_listen_work);
30 	mutex_lock(&smc->clcsock_release_lock);
31 	if (smc->clcsock) {
32 		tcp = smc->clcsock;
33 		smc->clcsock = NULL;
34 		sock_release(tcp);
35 	}
36 	mutex_unlock(&smc->clcsock_release_lock);
37 }
38 
smc_close_cleanup_listen(struct sock * parent)39 static void smc_close_cleanup_listen(struct sock *parent)
40 {
41 	struct sock *sk;
42 
43 	/* Close non-accepted connections */
44 	while ((sk = smc_accept_dequeue(parent, NULL)))
45 		smc_close_non_accepted(sk);
46 }
47 
48 /* wait for sndbuf data being transmitted */
smc_close_stream_wait(struct smc_sock * smc,long timeout)49 static void smc_close_stream_wait(struct smc_sock *smc, long timeout)
50 {
51 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
52 	struct sock *sk = &smc->sk;
53 
54 	if (!timeout)
55 		return;
56 
57 	if (!smc_tx_prepared_sends(&smc->conn))
58 		return;
59 
60 	smc->wait_close_tx_prepared = 1;
61 	add_wait_queue(sk_sleep(sk), &wait);
62 	while (!signal_pending(current) && timeout) {
63 		int rc;
64 
65 		rc = sk_wait_event(sk, &timeout,
66 				   !smc_tx_prepared_sends(&smc->conn) ||
67 				   READ_ONCE(sk->sk_err) == ECONNABORTED ||
68 				   READ_ONCE(sk->sk_err) == ECONNRESET ||
69 				   smc->conn.killed,
70 				   &wait);
71 		if (rc)
72 			break;
73 	}
74 	remove_wait_queue(sk_sleep(sk), &wait);
75 	smc->wait_close_tx_prepared = 0;
76 }
77 
smc_close_wake_tx_prepared(struct smc_sock * smc)78 void smc_close_wake_tx_prepared(struct smc_sock *smc)
79 {
80 	if (smc->wait_close_tx_prepared)
81 		/* wake up socket closing */
82 		smc->sk.sk_state_change(&smc->sk);
83 }
84 
smc_close_wr(struct smc_connection * conn)85 static int smc_close_wr(struct smc_connection *conn)
86 {
87 	conn->local_tx_ctrl.conn_state_flags.peer_done_writing = 1;
88 
89 	return smc_cdc_get_slot_and_msg_send(conn);
90 }
91 
smc_close_final(struct smc_connection * conn)92 static int smc_close_final(struct smc_connection *conn)
93 {
94 	if (atomic_read(&conn->bytes_to_rcv))
95 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
96 	else
97 		conn->local_tx_ctrl.conn_state_flags.peer_conn_closed = 1;
98 	if (conn->killed)
99 		return -EPIPE;
100 
101 	return smc_cdc_get_slot_and_msg_send(conn);
102 }
103 
smc_close_abort(struct smc_connection * conn)104 int smc_close_abort(struct smc_connection *conn)
105 {
106 	conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
107 
108 	return smc_cdc_get_slot_and_msg_send(conn);
109 }
110 
smc_close_cancel_work(struct smc_sock * smc)111 static void smc_close_cancel_work(struct smc_sock *smc)
112 {
113 	struct sock *sk = &smc->sk;
114 
115 	release_sock(sk);
116 	if (cancel_work_sync(&smc->conn.close_work))
117 		sock_put(sk);
118 	cancel_delayed_work_sync(&smc->conn.tx_work);
119 	lock_sock(sk);
120 }
121 
122 /* terminate smc socket abnormally - active abort
123  * link group is terminated, i.e. RDMA communication no longer possible
124  */
smc_close_active_abort(struct smc_sock * smc)125 void smc_close_active_abort(struct smc_sock *smc)
126 {
127 	struct sock *sk = &smc->sk;
128 	bool release_clcsock = false;
129 
130 	if (sk->sk_state != SMC_INIT && smc->clcsock && smc->clcsock->sk) {
131 		sk->sk_err = ECONNABORTED;
132 		if (smc->clcsock && smc->clcsock->sk)
133 			tcp_abort(smc->clcsock->sk, ECONNABORTED);
134 	}
135 	switch (sk->sk_state) {
136 	case SMC_ACTIVE:
137 	case SMC_APPCLOSEWAIT1:
138 	case SMC_APPCLOSEWAIT2:
139 		sk->sk_state = SMC_PEERABORTWAIT;
140 		smc_close_cancel_work(smc);
141 		if (sk->sk_state != SMC_PEERABORTWAIT)
142 			break;
143 		sk->sk_state = SMC_CLOSED;
144 		sock_put(sk); /* (postponed) passive closing */
145 		break;
146 	case SMC_PEERCLOSEWAIT1:
147 	case SMC_PEERCLOSEWAIT2:
148 	case SMC_PEERFINCLOSEWAIT:
149 		sk->sk_state = SMC_PEERABORTWAIT;
150 		smc_close_cancel_work(smc);
151 		if (sk->sk_state != SMC_PEERABORTWAIT)
152 			break;
153 		sk->sk_state = SMC_CLOSED;
154 		smc_conn_free(&smc->conn);
155 		release_clcsock = true;
156 		sock_put(sk); /* passive closing */
157 		break;
158 	case SMC_PROCESSABORT:
159 	case SMC_APPFINCLOSEWAIT:
160 		sk->sk_state = SMC_PEERABORTWAIT;
161 		smc_close_cancel_work(smc);
162 		if (sk->sk_state != SMC_PEERABORTWAIT)
163 			break;
164 		sk->sk_state = SMC_CLOSED;
165 		smc_conn_free(&smc->conn);
166 		release_clcsock = true;
167 		break;
168 	case SMC_INIT:
169 	case SMC_PEERABORTWAIT:
170 	case SMC_CLOSED:
171 		break;
172 	}
173 
174 	smc_sock_set_flag(sk, SOCK_DEAD);
175 	sk->sk_state_change(sk);
176 
177 	if (release_clcsock) {
178 		release_sock(sk);
179 		smc_clcsock_release(smc);
180 		lock_sock(sk);
181 	}
182 }
183 
smc_close_sent_any_close(struct smc_connection * conn)184 static inline bool smc_close_sent_any_close(struct smc_connection *conn)
185 {
186 	return conn->local_tx_ctrl.conn_state_flags.peer_conn_abort ||
187 	       conn->local_tx_ctrl.conn_state_flags.peer_conn_closed;
188 }
189 
smc_close_active(struct smc_sock * smc)190 int smc_close_active(struct smc_sock *smc)
191 {
192 	struct smc_cdc_conn_state_flags *txflags =
193 		&smc->conn.local_tx_ctrl.conn_state_flags;
194 	struct smc_connection *conn = &smc->conn;
195 	struct sock *sk = &smc->sk;
196 	int old_state;
197 	long timeout;
198 	int rc = 0;
199 	int rc1 = 0;
200 
201 	timeout = current->flags & PF_EXITING ?
202 		  0 : sock_flag(sk, SOCK_LINGER) ?
203 		      sk->sk_lingertime : SMC_MAX_STREAM_WAIT_TIMEOUT;
204 
205 	old_state = sk->sk_state;
206 again:
207 	switch (sk->sk_state) {
208 	case SMC_INIT:
209 		sk->sk_state = SMC_CLOSED;
210 		break;
211 	case SMC_LISTEN:
212 		sk->sk_state = SMC_CLOSED;
213 		sk->sk_state_change(sk); /* wake up accept */
214 		if (smc->clcsock && smc->clcsock->sk) {
215 			smc->clcsock->sk->sk_data_ready = smc->clcsk_data_ready;
216 			smc->clcsock->sk->sk_user_data = NULL;
217 			rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
218 		}
219 		smc_close_cleanup_listen(sk);
220 		release_sock(sk);
221 		flush_work(&smc->tcp_listen_work);
222 		lock_sock(sk);
223 		break;
224 	case SMC_ACTIVE:
225 		smc_close_stream_wait(smc, timeout);
226 		release_sock(sk);
227 		cancel_delayed_work_sync(&conn->tx_work);
228 		lock_sock(sk);
229 		if (sk->sk_state == SMC_ACTIVE) {
230 			/* send close request */
231 			rc = smc_close_final(conn);
232 			sk->sk_state = SMC_PEERCLOSEWAIT1;
233 
234 			/* actively shutdown clcsock before peer close it,
235 			 * prevent peer from entering TIME_WAIT state.
236 			 */
237 			if (smc->clcsock && smc->clcsock->sk) {
238 				rc1 = kernel_sock_shutdown(smc->clcsock,
239 							   SHUT_RDWR);
240 				rc = rc ? rc : rc1;
241 			}
242 		} else {
243 			/* peer event has changed the state */
244 			goto again;
245 		}
246 		break;
247 	case SMC_APPFINCLOSEWAIT:
248 		/* socket already shutdown wr or both (active close) */
249 		if (txflags->peer_done_writing &&
250 		    !smc_close_sent_any_close(conn)) {
251 			/* just shutdown wr done, send close request */
252 			rc = smc_close_final(conn);
253 		}
254 		sk->sk_state = SMC_CLOSED;
255 		break;
256 	case SMC_APPCLOSEWAIT1:
257 	case SMC_APPCLOSEWAIT2:
258 		if (!smc_cdc_rxed_any_close(conn))
259 			smc_close_stream_wait(smc, timeout);
260 		release_sock(sk);
261 		cancel_delayed_work_sync(&conn->tx_work);
262 		lock_sock(sk);
263 		if (sk->sk_state != SMC_APPCLOSEWAIT1 &&
264 		    sk->sk_state != SMC_APPCLOSEWAIT2)
265 			goto again;
266 		/* confirm close from peer */
267 		rc = smc_close_final(conn);
268 		if (smc_cdc_rxed_any_close(conn)) {
269 			/* peer has closed the socket already */
270 			sk->sk_state = SMC_CLOSED;
271 			sock_put(sk); /* postponed passive closing */
272 		} else {
273 			/* peer has just issued a shutdown write */
274 			sk->sk_state = SMC_PEERFINCLOSEWAIT;
275 		}
276 		break;
277 	case SMC_PEERCLOSEWAIT1:
278 	case SMC_PEERCLOSEWAIT2:
279 		if (txflags->peer_done_writing &&
280 		    !smc_close_sent_any_close(conn)) {
281 			/* just shutdown wr done, send close request */
282 			rc = smc_close_final(conn);
283 		}
284 		/* peer sending PeerConnectionClosed will cause transition */
285 		break;
286 	case SMC_PEERFINCLOSEWAIT:
287 		/* peer sending PeerConnectionClosed will cause transition */
288 		break;
289 	case SMC_PROCESSABORT:
290 		rc = smc_close_abort(conn);
291 		sk->sk_state = SMC_CLOSED;
292 		break;
293 	case SMC_PEERABORTWAIT:
294 		sk->sk_state = SMC_CLOSED;
295 		break;
296 	case SMC_CLOSED:
297 		/* nothing to do, add tracing in future patch */
298 		break;
299 	}
300 
301 	if (old_state != sk->sk_state)
302 		sk->sk_state_change(sk);
303 	return rc;
304 }
305 
smc_close_passive_abort_received(struct smc_sock * smc)306 static void smc_close_passive_abort_received(struct smc_sock *smc)
307 {
308 	struct smc_cdc_conn_state_flags *txflags =
309 		&smc->conn.local_tx_ctrl.conn_state_flags;
310 	struct sock *sk = &smc->sk;
311 
312 	switch (sk->sk_state) {
313 	case SMC_INIT:
314 	case SMC_ACTIVE:
315 	case SMC_APPCLOSEWAIT1:
316 		sk->sk_state = SMC_PROCESSABORT;
317 		sock_put(sk); /* passive closing */
318 		break;
319 	case SMC_APPFINCLOSEWAIT:
320 		sk->sk_state = SMC_PROCESSABORT;
321 		break;
322 	case SMC_PEERCLOSEWAIT1:
323 	case SMC_PEERCLOSEWAIT2:
324 		if (txflags->peer_done_writing &&
325 		    !smc_close_sent_any_close(&smc->conn))
326 			/* just shutdown, but not yet closed locally */
327 			sk->sk_state = SMC_PROCESSABORT;
328 		else
329 			sk->sk_state = SMC_CLOSED;
330 		sock_put(sk); /* passive closing */
331 		break;
332 	case SMC_APPCLOSEWAIT2:
333 	case SMC_PEERFINCLOSEWAIT:
334 		sk->sk_state = SMC_CLOSED;
335 		sock_put(sk); /* passive closing */
336 		break;
337 	case SMC_PEERABORTWAIT:
338 		sk->sk_state = SMC_CLOSED;
339 		break;
340 	case SMC_PROCESSABORT:
341 	/* nothing to do, add tracing in future patch */
342 		break;
343 	}
344 }
345 
346 /* Either some kind of closing has been received: peer_conn_closed,
347  * peer_conn_abort, or peer_done_writing
348  * or the link group of the connection terminates abnormally.
349  */
smc_close_passive_work(struct work_struct * work)350 static void smc_close_passive_work(struct work_struct *work)
351 {
352 	struct smc_connection *conn = container_of(work,
353 						   struct smc_connection,
354 						   close_work);
355 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
356 	struct smc_cdc_conn_state_flags *rxflags;
357 	bool release_clcsock = false;
358 	struct sock *sk = &smc->sk;
359 	int old_state;
360 
361 	lock_sock(sk);
362 	old_state = sk->sk_state;
363 
364 	rxflags = &conn->local_rx_ctrl.conn_state_flags;
365 	if (rxflags->peer_conn_abort) {
366 		/* peer has not received all data */
367 		smc_close_passive_abort_received(smc);
368 		release_sock(&smc->sk);
369 		cancel_delayed_work_sync(&conn->tx_work);
370 		lock_sock(&smc->sk);
371 		goto wakeup;
372 	}
373 
374 	switch (sk->sk_state) {
375 	case SMC_INIT:
376 		sk->sk_state = SMC_APPCLOSEWAIT1;
377 		break;
378 	case SMC_ACTIVE:
379 		sk->sk_state = SMC_APPCLOSEWAIT1;
380 		/* postpone sock_put() for passive closing to cover
381 		 * received SEND_SHUTDOWN as well
382 		 */
383 		break;
384 	case SMC_PEERCLOSEWAIT1:
385 		if (rxflags->peer_done_writing)
386 			sk->sk_state = SMC_PEERCLOSEWAIT2;
387 		fallthrough;
388 		/* to check for closing */
389 	case SMC_PEERCLOSEWAIT2:
390 		if (!smc_cdc_rxed_any_close(conn))
391 			break;
392 		if (sock_flag(sk, SOCK_DEAD) &&
393 		    smc_close_sent_any_close(conn)) {
394 			/* smc_release has already been called locally */
395 			sk->sk_state = SMC_CLOSED;
396 		} else {
397 			/* just shutdown, but not yet closed locally */
398 			sk->sk_state = SMC_APPFINCLOSEWAIT;
399 		}
400 		sock_put(sk); /* passive closing */
401 		break;
402 	case SMC_PEERFINCLOSEWAIT:
403 		if (smc_cdc_rxed_any_close(conn)) {
404 			sk->sk_state = SMC_CLOSED;
405 			sock_put(sk); /* passive closing */
406 		}
407 		break;
408 	case SMC_APPCLOSEWAIT1:
409 	case SMC_APPCLOSEWAIT2:
410 		/* postpone sock_put() for passive closing to cover
411 		 * received SEND_SHUTDOWN as well
412 		 */
413 		break;
414 	case SMC_APPFINCLOSEWAIT:
415 	case SMC_PEERABORTWAIT:
416 	case SMC_PROCESSABORT:
417 	case SMC_CLOSED:
418 		/* nothing to do, add tracing in future patch */
419 		break;
420 	}
421 
422 wakeup:
423 	sk->sk_data_ready(sk); /* wakeup blocked rcvbuf consumers */
424 	sk->sk_write_space(sk); /* wakeup blocked sndbuf producers */
425 
426 	if (old_state != sk->sk_state) {
427 		sk->sk_state_change(sk);
428 		if ((sk->sk_state == SMC_CLOSED) &&
429 		    (sock_flag(sk, SOCK_DEAD) || !sk->sk_socket)) {
430 			smc_conn_free(conn);
431 			if (smc->clcsock)
432 				release_clcsock = true;
433 		}
434 	}
435 	release_sock(sk);
436 	if (release_clcsock)
437 		smc_clcsock_release(smc);
438 	sock_put(sk); /* sock_hold done by schedulers of close_work */
439 }
440 
smc_close_shutdown_write(struct smc_sock * smc)441 int smc_close_shutdown_write(struct smc_sock *smc)
442 {
443 	struct smc_connection *conn = &smc->conn;
444 	struct sock *sk = &smc->sk;
445 	int old_state;
446 	long timeout;
447 	int rc = 0;
448 
449 	timeout = current->flags & PF_EXITING ?
450 		  0 : sock_flag(sk, SOCK_LINGER) ?
451 		      sk->sk_lingertime : SMC_MAX_STREAM_WAIT_TIMEOUT;
452 
453 	old_state = sk->sk_state;
454 again:
455 	switch (sk->sk_state) {
456 	case SMC_ACTIVE:
457 		smc_close_stream_wait(smc, timeout);
458 		release_sock(sk);
459 		cancel_delayed_work_sync(&conn->tx_work);
460 		lock_sock(sk);
461 		if (sk->sk_state != SMC_ACTIVE)
462 			goto again;
463 		/* send close wr request */
464 		rc = smc_close_wr(conn);
465 		sk->sk_state = SMC_PEERCLOSEWAIT1;
466 		break;
467 	case SMC_APPCLOSEWAIT1:
468 		/* passive close */
469 		if (!smc_cdc_rxed_any_close(conn))
470 			smc_close_stream_wait(smc, timeout);
471 		release_sock(sk);
472 		cancel_delayed_work_sync(&conn->tx_work);
473 		lock_sock(sk);
474 		if (sk->sk_state != SMC_APPCLOSEWAIT1)
475 			goto again;
476 		/* confirm close from peer */
477 		rc = smc_close_wr(conn);
478 		sk->sk_state = SMC_APPCLOSEWAIT2;
479 		break;
480 	case SMC_APPCLOSEWAIT2:
481 	case SMC_PEERFINCLOSEWAIT:
482 	case SMC_PEERCLOSEWAIT1:
483 	case SMC_PEERCLOSEWAIT2:
484 	case SMC_APPFINCLOSEWAIT:
485 	case SMC_PROCESSABORT:
486 	case SMC_PEERABORTWAIT:
487 		/* nothing to do, add tracing in future patch */
488 		break;
489 	}
490 
491 	if (old_state != sk->sk_state)
492 		sk->sk_state_change(sk);
493 	return rc;
494 }
495 
496 /* Initialize close properties on connection establishment. */
smc_close_init(struct smc_sock * smc)497 void smc_close_init(struct smc_sock *smc)
498 {
499 	INIT_WORK(&smc->conn.close_work, smc_close_passive_work);
500 }
501