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 #include "smc_stats.h"
25
26 /* callback implementation to wakeup consumers blocked with smc_rx_wait().
27 * indirectly called by smc_cdc_msg_recv_action().
28 */
smc_rx_wake_up(struct sock * sk)29 static void smc_rx_wake_up(struct sock *sk)
30 {
31 struct socket_wq *wq;
32
33 /* derived from sock_def_readable() */
34 /* called already in smc_listen_work() */
35 rcu_read_lock();
36 wq = rcu_dereference(sk->sk_wq);
37 if (skwq_has_sleeper(wq))
38 wake_up_interruptible_sync_poll(&wq->wait, EPOLLIN | EPOLLPRI |
39 EPOLLRDNORM | EPOLLRDBAND);
40 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
41 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
42 (sk->sk_state == SMC_CLOSED))
43 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
44 rcu_read_unlock();
45 }
46
47 /* Update consumer cursor
48 * @conn connection to update
49 * @cons consumer cursor
50 * @len number of Bytes consumed
51 * Returns:
52 * 1 if we should end our receive, 0 otherwise
53 */
smc_rx_update_consumer(struct smc_sock * smc,union smc_host_cursor cons,size_t len)54 static int smc_rx_update_consumer(struct smc_sock *smc,
55 union smc_host_cursor cons, size_t len)
56 {
57 struct smc_connection *conn = &smc->conn;
58 struct sock *sk = &smc->sk;
59 bool force = false;
60 int diff, rc = 0;
61
62 smc_curs_add(conn->rmb_desc->len, &cons, len);
63
64 /* did we process urgent data? */
65 if (conn->urg_state == SMC_URG_VALID || conn->urg_rx_skip_pend) {
66 diff = smc_curs_comp(conn->rmb_desc->len, &cons,
67 &conn->urg_curs);
68 if (sock_flag(sk, SOCK_URGINLINE)) {
69 if (diff == 0) {
70 force = true;
71 rc = 1;
72 conn->urg_state = SMC_URG_READ;
73 }
74 } else {
75 if (diff == 1) {
76 /* skip urgent byte */
77 force = true;
78 smc_curs_add(conn->rmb_desc->len, &cons, 1);
79 conn->urg_rx_skip_pend = false;
80 } else if (diff < -1)
81 /* we read past urgent byte */
82 conn->urg_state = SMC_URG_READ;
83 }
84 }
85
86 smc_curs_copy(&conn->local_tx_ctrl.cons, &cons, conn);
87
88 /* send consumer cursor update if required */
89 /* similar to advertising new TCP rcv_wnd if required */
90 smc_tx_consumer_update(conn, force);
91
92 return rc;
93 }
94
smc_rx_update_cons(struct smc_sock * smc,size_t len)95 static void smc_rx_update_cons(struct smc_sock *smc, size_t len)
96 {
97 struct smc_connection *conn = &smc->conn;
98 union smc_host_cursor cons;
99
100 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
101 smc_rx_update_consumer(smc, cons, len);
102 }
103
104 struct smc_spd_priv {
105 struct smc_sock *smc;
106 size_t len;
107 };
108
smc_rx_pipe_buf_release(struct pipe_inode_info * pipe,struct pipe_buffer * buf)109 static void smc_rx_pipe_buf_release(struct pipe_inode_info *pipe,
110 struct pipe_buffer *buf)
111 {
112 struct smc_spd_priv *priv = (struct smc_spd_priv *)buf->private;
113 struct smc_sock *smc = priv->smc;
114 struct smc_connection *conn;
115 struct sock *sk = &smc->sk;
116
117 if (sk->sk_state == SMC_CLOSED ||
118 sk->sk_state == SMC_PEERFINCLOSEWAIT ||
119 sk->sk_state == SMC_APPFINCLOSEWAIT)
120 goto out;
121 conn = &smc->conn;
122 lock_sock(sk);
123 smc_rx_update_cons(smc, priv->len);
124 release_sock(sk);
125 if (atomic_sub_and_test(priv->len, &conn->splice_pending))
126 smc_rx_wake_up(sk);
127 out:
128 kfree(priv);
129 put_page(buf->page);
130 sock_put(sk);
131 }
132
133 static const struct pipe_buf_operations smc_pipe_ops = {
134 .release = smc_rx_pipe_buf_release,
135 .get = generic_pipe_buf_get
136 };
137
smc_rx_spd_release(struct splice_pipe_desc * spd,unsigned int i)138 static void smc_rx_spd_release(struct splice_pipe_desc *spd,
139 unsigned int i)
140 {
141 put_page(spd->pages[i]);
142 }
143
smc_rx_splice(struct pipe_inode_info * pipe,char * src,size_t len,struct smc_sock * smc)144 static int smc_rx_splice(struct pipe_inode_info *pipe, char *src, size_t len,
145 struct smc_sock *smc)
146 {
147 struct splice_pipe_desc spd;
148 struct partial_page partial;
149 struct smc_spd_priv *priv;
150 int bytes;
151
152 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
153 if (!priv)
154 return -ENOMEM;
155 priv->len = len;
156 priv->smc = smc;
157 partial.offset = src - (char *)smc->conn.rmb_desc->cpu_addr;
158 partial.len = len;
159 partial.private = (unsigned long)priv;
160
161 spd.nr_pages_max = 1;
162 spd.nr_pages = 1;
163 spd.pages = &smc->conn.rmb_desc->pages;
164 spd.partial = &partial;
165 spd.ops = &smc_pipe_ops;
166 spd.spd_release = smc_rx_spd_release;
167
168 bytes = splice_to_pipe(pipe, &spd);
169 if (bytes > 0) {
170 sock_hold(&smc->sk);
171 get_page(smc->conn.rmb_desc->pages);
172 atomic_add(bytes, &smc->conn.splice_pending);
173 }
174
175 return bytes;
176 }
177
smc_rx_data_available_and_no_splice_pend(struct smc_connection * conn)178 static int smc_rx_data_available_and_no_splice_pend(struct smc_connection *conn)
179 {
180 return atomic_read(&conn->bytes_to_rcv) &&
181 !atomic_read(&conn->splice_pending);
182 }
183
184 /* blocks rcvbuf consumer until >=len bytes available or timeout or interrupted
185 * @smc smc socket
186 * @timeo pointer to max seconds to wait, pointer to value 0 for no timeout
187 * @fcrit add'l criterion to evaluate as function pointer
188 * Returns:
189 * 1 if at least 1 byte available in rcvbuf or if socket error/shutdown.
190 * 0 otherwise (nothing in rcvbuf nor timeout, e.g. interrupted).
191 */
smc_rx_wait(struct smc_sock * smc,long * timeo,int (* fcrit)(struct smc_connection * conn))192 int smc_rx_wait(struct smc_sock *smc, long *timeo,
193 int (*fcrit)(struct smc_connection *conn))
194 {
195 DEFINE_WAIT_FUNC(wait, woken_wake_function);
196 struct smc_connection *conn = &smc->conn;
197 struct smc_cdc_conn_state_flags *cflags =
198 &conn->local_tx_ctrl.conn_state_flags;
199 struct sock *sk = &smc->sk;
200 int rc;
201
202 if (fcrit(conn))
203 return 1;
204 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
205 add_wait_queue(sk_sleep(sk), &wait);
206 rc = sk_wait_event(sk, timeo,
207 READ_ONCE(sk->sk_err) ||
208 cflags->peer_conn_abort ||
209 READ_ONCE(sk->sk_shutdown) & RCV_SHUTDOWN ||
210 conn->killed ||
211 fcrit(conn),
212 &wait);
213 remove_wait_queue(sk_sleep(sk), &wait);
214 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
215 return rc;
216 }
217
smc_rx_recv_urg(struct smc_sock * smc,struct msghdr * msg,int len,int flags)218 static int smc_rx_recv_urg(struct smc_sock *smc, struct msghdr *msg, int len,
219 int flags)
220 {
221 struct smc_connection *conn = &smc->conn;
222 union smc_host_cursor cons;
223 struct sock *sk = &smc->sk;
224 int rc = 0;
225
226 if (sock_flag(sk, SOCK_URGINLINE) ||
227 !(conn->urg_state == SMC_URG_VALID) ||
228 conn->urg_state == SMC_URG_READ)
229 return -EINVAL;
230
231 SMC_STAT_INC(smc, urg_data_cnt);
232 if (conn->urg_state == SMC_URG_VALID) {
233 if (!(flags & MSG_PEEK))
234 smc->conn.urg_state = SMC_URG_READ;
235 msg->msg_flags |= MSG_OOB;
236 if (len > 0) {
237 if (!(flags & MSG_TRUNC))
238 rc = memcpy_to_msg(msg, &conn->urg_rx_byte, 1);
239 len = 1;
240 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
241 if (smc_curs_diff(conn->rmb_desc->len, &cons,
242 &conn->urg_curs) > 1)
243 conn->urg_rx_skip_pend = true;
244 /* Urgent Byte was already accounted for, but trigger
245 * skipping the urgent byte in non-inline case
246 */
247 if (!(flags & MSG_PEEK))
248 smc_rx_update_consumer(smc, cons, 0);
249 } else {
250 msg->msg_flags |= MSG_TRUNC;
251 }
252
253 return rc ? -EFAULT : len;
254 }
255
256 if (sk->sk_state == SMC_CLOSED || sk->sk_shutdown & RCV_SHUTDOWN)
257 return 0;
258
259 return -EAGAIN;
260 }
261
smc_rx_recvmsg_data_available(struct smc_sock * smc)262 static bool smc_rx_recvmsg_data_available(struct smc_sock *smc)
263 {
264 struct smc_connection *conn = &smc->conn;
265
266 if (smc_rx_data_available(conn))
267 return true;
268 else if (conn->urg_state == SMC_URG_VALID)
269 /* we received a single urgent Byte - skip */
270 smc_rx_update_cons(smc, 0);
271 return false;
272 }
273
274 /* smc_rx_recvmsg - receive data from RMBE
275 * @msg: copy data to receive buffer
276 * @pipe: copy data to pipe if set - indicates splice() call
277 *
278 * rcvbuf consumer: main API called by socket layer.
279 * Called under sk lock.
280 */
smc_rx_recvmsg(struct smc_sock * smc,struct msghdr * msg,struct pipe_inode_info * pipe,size_t len,int flags)281 int smc_rx_recvmsg(struct smc_sock *smc, struct msghdr *msg,
282 struct pipe_inode_info *pipe, size_t len, int flags)
283 {
284 size_t copylen, read_done = 0, read_remaining = len;
285 size_t chunk_len, chunk_off, chunk_len_sum;
286 struct smc_connection *conn = &smc->conn;
287 int (*func)(struct smc_connection *conn);
288 union smc_host_cursor cons;
289 int readable, chunk;
290 char *rcvbuf_base;
291 struct sock *sk;
292 int splbytes;
293 long timeo;
294 int target; /* Read at least these many bytes */
295 int rc;
296
297 if (unlikely(flags & MSG_ERRQUEUE))
298 return -EINVAL; /* future work for sk.sk_family == AF_SMC */
299
300 sk = &smc->sk;
301 if (sk->sk_state == SMC_LISTEN)
302 return -ENOTCONN;
303 if (flags & MSG_OOB)
304 return smc_rx_recv_urg(smc, msg, len, flags);
305 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
306 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
307
308 readable = atomic_read(&conn->bytes_to_rcv);
309 if (readable >= conn->rmb_desc->len)
310 SMC_STAT_RMB_RX_FULL(smc, !conn->lnk);
311
312 if (len < readable)
313 SMC_STAT_RMB_RX_SIZE_SMALL(smc, !conn->lnk);
314 /* we currently use 1 RMBE per RMB, so RMBE == RMB base addr */
315 rcvbuf_base = conn->rx_off + conn->rmb_desc->cpu_addr;
316
317 do { /* while (read_remaining) */
318 if (read_done >= target || (pipe && read_done))
319 break;
320
321 if (conn->killed)
322 break;
323
324 if (smc_rx_recvmsg_data_available(smc))
325 goto copy;
326
327 if (sk->sk_shutdown & RCV_SHUTDOWN) {
328 /* smc_cdc_msg_recv_action() could have run after
329 * above smc_rx_recvmsg_data_available()
330 */
331 if (smc_rx_recvmsg_data_available(smc))
332 goto copy;
333 break;
334 }
335
336 if (read_done) {
337 if (sk->sk_err ||
338 sk->sk_state == SMC_CLOSED ||
339 !timeo ||
340 signal_pending(current))
341 break;
342 } else {
343 if (sk->sk_err) {
344 read_done = sock_error(sk);
345 break;
346 }
347 if (sk->sk_state == SMC_CLOSED) {
348 if (!sock_flag(sk, SOCK_DONE)) {
349 /* This occurs when user tries to read
350 * from never connected socket.
351 */
352 read_done = -ENOTCONN;
353 break;
354 }
355 break;
356 }
357 if (!timeo)
358 return -EAGAIN;
359 if (signal_pending(current)) {
360 read_done = sock_intr_errno(timeo);
361 break;
362 }
363 }
364
365 if (!smc_rx_data_available(conn)) {
366 smc_rx_wait(smc, &timeo, smc_rx_data_available);
367 continue;
368 }
369
370 copy:
371 /* initialize variables for 1st iteration of subsequent loop */
372 /* could be just 1 byte, even after waiting on data above */
373 readable = atomic_read(&conn->bytes_to_rcv);
374 splbytes = atomic_read(&conn->splice_pending);
375 if (!readable || (msg && splbytes)) {
376 if (splbytes)
377 func = smc_rx_data_available_and_no_splice_pend;
378 else
379 func = smc_rx_data_available;
380 smc_rx_wait(smc, &timeo, func);
381 continue;
382 }
383
384 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
385 /* subsequent splice() calls pick up where previous left */
386 if (splbytes)
387 smc_curs_add(conn->rmb_desc->len, &cons, splbytes);
388 if (conn->urg_state == SMC_URG_VALID &&
389 sock_flag(&smc->sk, SOCK_URGINLINE) &&
390 readable > 1)
391 readable--; /* always stop at urgent Byte */
392 /* not more than what user space asked for */
393 copylen = min_t(size_t, read_remaining, readable);
394 /* determine chunks where to read from rcvbuf */
395 /* either unwrapped case, or 1st chunk of wrapped case */
396 chunk_len = min_t(size_t, copylen, conn->rmb_desc->len -
397 cons.count);
398 chunk_len_sum = chunk_len;
399 chunk_off = cons.count;
400 smc_rmb_sync_sg_for_cpu(conn);
401 for (chunk = 0; chunk < 2; chunk++) {
402 if (!(flags & MSG_TRUNC)) {
403 if (msg) {
404 rc = memcpy_to_msg(msg, rcvbuf_base +
405 chunk_off,
406 chunk_len);
407 } else {
408 rc = smc_rx_splice(pipe, rcvbuf_base +
409 chunk_off, chunk_len,
410 smc);
411 }
412 if (rc < 0) {
413 if (!read_done)
414 read_done = -EFAULT;
415 smc_rmb_sync_sg_for_device(conn);
416 goto out;
417 }
418 }
419 read_remaining -= chunk_len;
420 read_done += chunk_len;
421
422 if (chunk_len_sum == copylen)
423 break; /* either on 1st or 2nd iteration */
424 /* prepare next (== 2nd) iteration */
425 chunk_len = copylen - chunk_len; /* remainder */
426 chunk_len_sum += chunk_len;
427 chunk_off = 0; /* modulo offset in recv ring buffer */
428 }
429 smc_rmb_sync_sg_for_device(conn);
430
431 /* update cursors */
432 if (!(flags & MSG_PEEK)) {
433 /* increased in recv tasklet smc_cdc_msg_rcv() */
434 smp_mb__before_atomic();
435 atomic_sub(copylen, &conn->bytes_to_rcv);
436 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
437 smp_mb__after_atomic();
438 if (msg && smc_rx_update_consumer(smc, cons, copylen))
439 goto out;
440 }
441 } while (read_remaining);
442 out:
443 return read_done;
444 }
445
446 /* Initialize receive properties on connection establishment. NB: not __init! */
smc_rx_init(struct smc_sock * smc)447 void smc_rx_init(struct smc_sock *smc)
448 {
449 smc->sk.sk_data_ready = smc_rx_wake_up;
450 atomic_set(&smc->conn.splice_pending, 0);
451 smc->conn.urg_state = SMC_URG_READ;
452 }
453