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