xref: /linux/net/smc/smc_rx.c (revision 17cfcb68af3bc7d5e8ae08779b1853310a2949f3)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Shared Memory Communications over RDMA (SMC-R) and RoCE
4  *
5  * Manage RMBE
6  * copy new RMBE data into user space
7  *
8  * Copyright IBM Corp. 2016
9  *
10  * Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
11  */
12 
13 #include <linux/net.h>
14 #include <linux/rcupdate.h>
15 #include <linux/sched/signal.h>
16 
17 #include <net/sock.h>
18 
19 #include "smc.h"
20 #include "smc_core.h"
21 #include "smc_cdc.h"
22 #include "smc_tx.h" /* smc_tx_consumer_update() */
23 #include "smc_rx.h"
24 
25 /* callback implementation to wakeup consumers blocked with smc_rx_wait().
26  * indirectly called by smc_cdc_msg_recv_action().
27  */
28 static void smc_rx_wake_up(struct sock *sk)
29 {
30 	struct socket_wq *wq;
31 
32 	/* derived from sock_def_readable() */
33 	/* called already in smc_listen_work() */
34 	rcu_read_lock();
35 	wq = rcu_dereference(sk->sk_wq);
36 	if (skwq_has_sleeper(wq))
37 		wake_up_interruptible_sync_poll(&wq->wait, EPOLLIN | EPOLLPRI |
38 						EPOLLRDNORM | EPOLLRDBAND);
39 	sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
40 	if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
41 	    (sk->sk_state == SMC_CLOSED))
42 		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
43 	rcu_read_unlock();
44 }
45 
46 /* Update consumer cursor
47  *   @conn   connection to update
48  *   @cons   consumer cursor
49  *   @len    number of Bytes consumed
50  *   Returns:
51  *   1 if we should end our receive, 0 otherwise
52  */
53 static int smc_rx_update_consumer(struct smc_sock *smc,
54 				  union smc_host_cursor cons, size_t len)
55 {
56 	struct smc_connection *conn = &smc->conn;
57 	struct sock *sk = &smc->sk;
58 	bool force = false;
59 	int diff, rc = 0;
60 
61 	smc_curs_add(conn->rmb_desc->len, &cons, len);
62 
63 	/* did we process urgent data? */
64 	if (conn->urg_state == SMC_URG_VALID || conn->urg_rx_skip_pend) {
65 		diff = smc_curs_comp(conn->rmb_desc->len, &cons,
66 				     &conn->urg_curs);
67 		if (sock_flag(sk, SOCK_URGINLINE)) {
68 			if (diff == 0) {
69 				force = true;
70 				rc = 1;
71 				conn->urg_state = SMC_URG_READ;
72 			}
73 		} else {
74 			if (diff == 1) {
75 				/* skip urgent byte */
76 				force = true;
77 				smc_curs_add(conn->rmb_desc->len, &cons, 1);
78 				conn->urg_rx_skip_pend = false;
79 			} else if (diff < -1)
80 				/* we read past urgent byte */
81 				conn->urg_state = SMC_URG_READ;
82 		}
83 	}
84 
85 	smc_curs_copy(&conn->local_tx_ctrl.cons, &cons, conn);
86 
87 	/* send consumer cursor update if required */
88 	/* similar to advertising new TCP rcv_wnd if required */
89 	smc_tx_consumer_update(conn, force);
90 
91 	return rc;
92 }
93 
94 static void smc_rx_update_cons(struct smc_sock *smc, size_t len)
95 {
96 	struct smc_connection *conn = &smc->conn;
97 	union smc_host_cursor cons;
98 
99 	smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
100 	smc_rx_update_consumer(smc, cons, len);
101 }
102 
103 struct smc_spd_priv {
104 	struct smc_sock *smc;
105 	size_t		 len;
106 };
107 
108 static void smc_rx_pipe_buf_release(struct pipe_inode_info *pipe,
109 				    struct pipe_buffer *buf)
110 {
111 	struct smc_spd_priv *priv = (struct smc_spd_priv *)buf->private;
112 	struct smc_sock *smc = priv->smc;
113 	struct smc_connection *conn;
114 	struct sock *sk = &smc->sk;
115 
116 	if (sk->sk_state == SMC_CLOSED ||
117 	    sk->sk_state == SMC_PEERFINCLOSEWAIT ||
118 	    sk->sk_state == SMC_APPFINCLOSEWAIT)
119 		goto out;
120 	conn = &smc->conn;
121 	lock_sock(sk);
122 	smc_rx_update_cons(smc, priv->len);
123 	release_sock(sk);
124 	if (atomic_sub_and_test(priv->len, &conn->splice_pending))
125 		smc_rx_wake_up(sk);
126 out:
127 	kfree(priv);
128 	put_page(buf->page);
129 	sock_put(sk);
130 }
131 
132 static int smc_rx_pipe_buf_nosteal(struct pipe_inode_info *pipe,
133 				   struct pipe_buffer *buf)
134 {
135 	return 1;
136 }
137 
138 static const struct pipe_buf_operations smc_pipe_ops = {
139 	.confirm = generic_pipe_buf_confirm,
140 	.release = smc_rx_pipe_buf_release,
141 	.steal = smc_rx_pipe_buf_nosteal,
142 	.get = generic_pipe_buf_get
143 };
144 
145 static void smc_rx_spd_release(struct splice_pipe_desc *spd,
146 			       unsigned int i)
147 {
148 	put_page(spd->pages[i]);
149 }
150 
151 static int smc_rx_splice(struct pipe_inode_info *pipe, char *src, size_t len,
152 			 struct smc_sock *smc)
153 {
154 	struct splice_pipe_desc spd;
155 	struct partial_page partial;
156 	struct smc_spd_priv *priv;
157 	int bytes;
158 
159 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
160 	if (!priv)
161 		return -ENOMEM;
162 	priv->len = len;
163 	priv->smc = smc;
164 	partial.offset = src - (char *)smc->conn.rmb_desc->cpu_addr;
165 	partial.len = len;
166 	partial.private = (unsigned long)priv;
167 
168 	spd.nr_pages_max = 1;
169 	spd.nr_pages = 1;
170 	spd.pages = &smc->conn.rmb_desc->pages;
171 	spd.partial = &partial;
172 	spd.ops = &smc_pipe_ops;
173 	spd.spd_release = smc_rx_spd_release;
174 
175 	bytes = splice_to_pipe(pipe, &spd);
176 	if (bytes > 0) {
177 		sock_hold(&smc->sk);
178 		get_page(smc->conn.rmb_desc->pages);
179 		atomic_add(bytes, &smc->conn.splice_pending);
180 	}
181 
182 	return bytes;
183 }
184 
185 static int smc_rx_data_available_and_no_splice_pend(struct smc_connection *conn)
186 {
187 	return atomic_read(&conn->bytes_to_rcv) &&
188 	       !atomic_read(&conn->splice_pending);
189 }
190 
191 /* blocks rcvbuf consumer until >=len bytes available or timeout or interrupted
192  *   @smc    smc socket
193  *   @timeo  pointer to max seconds to wait, pointer to value 0 for no timeout
194  *   @fcrit  add'l criterion to evaluate as function pointer
195  * Returns:
196  * 1 if at least 1 byte available in rcvbuf or if socket error/shutdown.
197  * 0 otherwise (nothing in rcvbuf nor timeout, e.g. interrupted).
198  */
199 int smc_rx_wait(struct smc_sock *smc, long *timeo,
200 		int (*fcrit)(struct smc_connection *conn))
201 {
202 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
203 	struct smc_connection *conn = &smc->conn;
204 	struct sock *sk = &smc->sk;
205 	int rc;
206 
207 	if (fcrit(conn))
208 		return 1;
209 	sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
210 	add_wait_queue(sk_sleep(sk), &wait);
211 	rc = sk_wait_event(sk, timeo,
212 			   sk->sk_err ||
213 			   sk->sk_shutdown & RCV_SHUTDOWN ||
214 			   fcrit(conn),
215 			   &wait);
216 	remove_wait_queue(sk_sleep(sk), &wait);
217 	sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
218 	return rc;
219 }
220 
221 static int smc_rx_recv_urg(struct smc_sock *smc, struct msghdr *msg, int len,
222 			   int flags)
223 {
224 	struct smc_connection *conn = &smc->conn;
225 	union smc_host_cursor cons;
226 	struct sock *sk = &smc->sk;
227 	int rc = 0;
228 
229 	if (sock_flag(sk, SOCK_URGINLINE) ||
230 	    !(conn->urg_state == SMC_URG_VALID) ||
231 	    conn->urg_state == SMC_URG_READ)
232 		return -EINVAL;
233 
234 	if (conn->urg_state == SMC_URG_VALID) {
235 		if (!(flags & MSG_PEEK))
236 			smc->conn.urg_state = SMC_URG_READ;
237 		msg->msg_flags |= MSG_OOB;
238 		if (len > 0) {
239 			if (!(flags & MSG_TRUNC))
240 				rc = memcpy_to_msg(msg, &conn->urg_rx_byte, 1);
241 			len = 1;
242 			smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
243 			if (smc_curs_diff(conn->rmb_desc->len, &cons,
244 					  &conn->urg_curs) > 1)
245 				conn->urg_rx_skip_pend = true;
246 			/* Urgent Byte was already accounted for, but trigger
247 			 * skipping the urgent byte in non-inline case
248 			 */
249 			if (!(flags & MSG_PEEK))
250 				smc_rx_update_consumer(smc, cons, 0);
251 		} else {
252 			msg->msg_flags |= MSG_TRUNC;
253 		}
254 
255 		return rc ? -EFAULT : len;
256 	}
257 
258 	if (sk->sk_state == SMC_CLOSED || sk->sk_shutdown & RCV_SHUTDOWN)
259 		return 0;
260 
261 	return -EAGAIN;
262 }
263 
264 static bool smc_rx_recvmsg_data_available(struct smc_sock *smc)
265 {
266 	struct smc_connection *conn = &smc->conn;
267 
268 	if (smc_rx_data_available(conn))
269 		return true;
270 	else if (conn->urg_state == SMC_URG_VALID)
271 		/* we received a single urgent Byte - skip */
272 		smc_rx_update_cons(smc, 0);
273 	return false;
274 }
275 
276 /* smc_rx_recvmsg - receive data from RMBE
277  * @msg:	copy data to receive buffer
278  * @pipe:	copy data to pipe if set - indicates splice() call
279  *
280  * rcvbuf consumer: main API called by socket layer.
281  * Called under sk lock.
282  */
283 int smc_rx_recvmsg(struct smc_sock *smc, struct msghdr *msg,
284 		   struct pipe_inode_info *pipe, size_t len, int flags)
285 {
286 	size_t copylen, read_done = 0, read_remaining = len;
287 	size_t chunk_len, chunk_off, chunk_len_sum;
288 	struct smc_connection *conn = &smc->conn;
289 	int (*func)(struct smc_connection *conn);
290 	union smc_host_cursor cons;
291 	int readable, chunk;
292 	char *rcvbuf_base;
293 	struct sock *sk;
294 	int splbytes;
295 	long timeo;
296 	int target;		/* Read at least these many bytes */
297 	int rc;
298 
299 	if (unlikely(flags & MSG_ERRQUEUE))
300 		return -EINVAL; /* future work for sk.sk_family == AF_SMC */
301 
302 	sk = &smc->sk;
303 	if (sk->sk_state == SMC_LISTEN)
304 		return -ENOTCONN;
305 	if (flags & MSG_OOB)
306 		return smc_rx_recv_urg(smc, msg, len, flags);
307 	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
308 	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
309 
310 	/* we currently use 1 RMBE per RMB, so RMBE == RMB base addr */
311 	rcvbuf_base = conn->rx_off + conn->rmb_desc->cpu_addr;
312 
313 	do { /* while (read_remaining) */
314 		if (read_done >= target || (pipe && read_done))
315 			break;
316 
317 		if (smc_rx_recvmsg_data_available(smc))
318 			goto copy;
319 
320 		if (sk->sk_shutdown & RCV_SHUTDOWN ||
321 		    conn->local_tx_ctrl.conn_state_flags.peer_conn_abort) {
322 			/* smc_cdc_msg_recv_action() could have run after
323 			 * above smc_rx_recvmsg_data_available()
324 			 */
325 			if (smc_rx_recvmsg_data_available(smc))
326 				goto copy;
327 			break;
328 		}
329 
330 		if (read_done) {
331 			if (sk->sk_err ||
332 			    sk->sk_state == SMC_CLOSED ||
333 			    !timeo ||
334 			    signal_pending(current))
335 				break;
336 		} else {
337 			if (sk->sk_err) {
338 				read_done = sock_error(sk);
339 				break;
340 			}
341 			if (sk->sk_state == SMC_CLOSED) {
342 				if (!sock_flag(sk, SOCK_DONE)) {
343 					/* This occurs when user tries to read
344 					 * from never connected socket.
345 					 */
346 					read_done = -ENOTCONN;
347 					break;
348 				}
349 				break;
350 			}
351 			if (signal_pending(current)) {
352 				read_done = sock_intr_errno(timeo);
353 				break;
354 			}
355 			if (!timeo)
356 				return -EAGAIN;
357 		}
358 
359 		if (!smc_rx_data_available(conn)) {
360 			smc_rx_wait(smc, &timeo, smc_rx_data_available);
361 			continue;
362 		}
363 
364 copy:
365 		/* initialize variables for 1st iteration of subsequent loop */
366 		/* could be just 1 byte, even after waiting on data above */
367 		readable = atomic_read(&conn->bytes_to_rcv);
368 		splbytes = atomic_read(&conn->splice_pending);
369 		if (!readable || (msg && splbytes)) {
370 			if (splbytes)
371 				func = smc_rx_data_available_and_no_splice_pend;
372 			else
373 				func = smc_rx_data_available;
374 			smc_rx_wait(smc, &timeo, func);
375 			continue;
376 		}
377 
378 		smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
379 		/* subsequent splice() calls pick up where previous left */
380 		if (splbytes)
381 			smc_curs_add(conn->rmb_desc->len, &cons, splbytes);
382 		if (conn->urg_state == SMC_URG_VALID &&
383 		    sock_flag(&smc->sk, SOCK_URGINLINE) &&
384 		    readable > 1)
385 			readable--;	/* always stop at urgent Byte */
386 		/* not more than what user space asked for */
387 		copylen = min_t(size_t, read_remaining, readable);
388 		/* determine chunks where to read from rcvbuf */
389 		/* either unwrapped case, or 1st chunk of wrapped case */
390 		chunk_len = min_t(size_t, copylen, conn->rmb_desc->len -
391 				  cons.count);
392 		chunk_len_sum = chunk_len;
393 		chunk_off = cons.count;
394 		smc_rmb_sync_sg_for_cpu(conn);
395 		for (chunk = 0; chunk < 2; chunk++) {
396 			if (!(flags & MSG_TRUNC)) {
397 				if (msg) {
398 					rc = memcpy_to_msg(msg, rcvbuf_base +
399 							   chunk_off,
400 							   chunk_len);
401 				} else {
402 					rc = smc_rx_splice(pipe, rcvbuf_base +
403 							chunk_off, chunk_len,
404 							smc);
405 				}
406 				if (rc < 0) {
407 					if (!read_done)
408 						read_done = -EFAULT;
409 					smc_rmb_sync_sg_for_device(conn);
410 					goto out;
411 				}
412 			}
413 			read_remaining -= chunk_len;
414 			read_done += chunk_len;
415 
416 			if (chunk_len_sum == copylen)
417 				break; /* either on 1st or 2nd iteration */
418 			/* prepare next (== 2nd) iteration */
419 			chunk_len = copylen - chunk_len; /* remainder */
420 			chunk_len_sum += chunk_len;
421 			chunk_off = 0; /* modulo offset in recv ring buffer */
422 		}
423 		smc_rmb_sync_sg_for_device(conn);
424 
425 		/* update cursors */
426 		if (!(flags & MSG_PEEK)) {
427 			/* increased in recv tasklet smc_cdc_msg_rcv() */
428 			smp_mb__before_atomic();
429 			atomic_sub(copylen, &conn->bytes_to_rcv);
430 			/* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
431 			smp_mb__after_atomic();
432 			if (msg && smc_rx_update_consumer(smc, cons, copylen))
433 				goto out;
434 		}
435 	} while (read_remaining);
436 out:
437 	return read_done;
438 }
439 
440 /* Initialize receive properties on connection establishment. NB: not __init! */
441 void smc_rx_init(struct smc_sock *smc)
442 {
443 	smc->sk.sk_data_ready = smc_rx_wake_up;
444 	atomic_set(&smc->conn.splice_pending, 0);
445 	smc->conn.urg_state = SMC_URG_READ;
446 }
447