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 #include <linux/splice.h>
17
18 #include <net/sock.h>
19 #include <trace/events/sock.h>
20
21 #include "smc.h"
22 #include "smc_core.h"
23 #include "smc_cdc.h"
24 #include "smc_tx.h" /* smc_tx_consumer_update() */
25 #include "smc_rx.h"
26 #include "smc_stats.h"
27 #include "smc_tracepoint.h"
28
29 /* callback implementation to wakeup consumers blocked with smc_rx_wait().
30 * indirectly called by smc_cdc_msg_recv_action().
31 */
smc_rx_wake_up(struct sock * sk)32 static void smc_rx_wake_up(struct sock *sk)
33 {
34 struct socket_wq *wq;
35
36 trace_sk_data_ready(sk);
37
38 /* derived from sock_def_readable() */
39 /* called already in smc_listen_work() */
40 rcu_read_lock();
41 wq = rcu_dereference(sk->sk_wq);
42 if (skwq_has_sleeper(wq))
43 wake_up_interruptible_sync_poll(&wq->wait, EPOLLIN | EPOLLPRI |
44 EPOLLRDNORM | EPOLLRDBAND);
45 sk_wake_async_rcu(sk, SOCK_WAKE_WAITD, POLL_IN);
46 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
47 (sk->sk_state == SMC_CLOSED))
48 sk_wake_async_rcu(sk, SOCK_WAKE_WAITD, POLL_HUP);
49 rcu_read_unlock();
50 }
51
52 /* Update consumer cursor
53 * @conn connection to update
54 * @cons consumer cursor
55 * @len number of Bytes consumed
56 * Returns:
57 * 1 if we should end our receive, 0 otherwise
58 */
smc_rx_update_consumer(struct smc_sock * smc,union smc_host_cursor cons,size_t len)59 static int smc_rx_update_consumer(struct smc_sock *smc,
60 union smc_host_cursor cons, size_t len)
61 {
62 struct smc_connection *conn = &smc->conn;
63 struct sock *sk = &smc->sk;
64 bool force = false;
65 int diff, rc = 0;
66
67 smc_curs_add(conn->rmb_desc->len, &cons, len);
68
69 /* did we process urgent data? */
70 if (conn->urg_state == SMC_URG_VALID || conn->urg_rx_skip_pend) {
71 diff = smc_curs_comp(conn->rmb_desc->len, &cons,
72 &conn->urg_curs);
73 if (sock_flag(sk, SOCK_URGINLINE)) {
74 if (diff == 0) {
75 force = true;
76 rc = 1;
77 conn->urg_state = SMC_URG_READ;
78 }
79 } else {
80 if (diff == 1) {
81 /* skip urgent byte */
82 force = true;
83 smc_curs_add(conn->rmb_desc->len, &cons, 1);
84 conn->urg_rx_skip_pend = false;
85 } else if (diff < -1)
86 /* we read past urgent byte */
87 conn->urg_state = SMC_URG_READ;
88 }
89 }
90
91 smc_curs_copy(&conn->local_tx_ctrl.cons, &cons, conn);
92
93 /* send consumer cursor update if required */
94 /* similar to advertising new TCP rcv_wnd if required */
95 smc_tx_consumer_update(conn, force);
96
97 return rc;
98 }
99
smc_rx_update_cons(struct smc_sock * smc,size_t len)100 static void smc_rx_update_cons(struct smc_sock *smc, size_t len)
101 {
102 struct smc_connection *conn = &smc->conn;
103 union smc_host_cursor cons;
104
105 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
106 smc_rx_update_consumer(smc, cons, len);
107 }
108
109 struct smc_spd_priv {
110 struct smc_sock *smc;
111 size_t len;
112 };
113
smc_rx_pipe_buf_release(struct pipe_inode_info * pipe,struct pipe_buffer * buf)114 static void smc_rx_pipe_buf_release(struct pipe_inode_info *pipe,
115 struct pipe_buffer *buf)
116 {
117 struct smc_spd_priv *priv = (struct smc_spd_priv *)buf->private;
118 struct smc_sock *smc = priv->smc;
119 struct smc_connection *conn;
120 struct sock *sk = &smc->sk;
121
122 if (sk->sk_state == SMC_CLOSED ||
123 sk->sk_state == SMC_PEERFINCLOSEWAIT ||
124 sk->sk_state == SMC_APPFINCLOSEWAIT)
125 goto out;
126 conn = &smc->conn;
127 lock_sock(sk);
128 smc_rx_update_cons(smc, priv->len);
129 release_sock(sk);
130 if (atomic_sub_and_test(priv->len, &conn->splice_pending))
131 smc_rx_wake_up(sk);
132 out:
133 kfree(priv);
134 put_page(buf->page);
135 sock_put(sk);
136 }
137
smc_rx_pipe_buf_get(struct pipe_inode_info * pipe,struct pipe_buffer * buf)138 static bool smc_rx_pipe_buf_get(struct pipe_inode_info *pipe,
139 struct pipe_buffer *buf)
140 {
141 /* smc_spd_priv in buf->private is not shareable; disallow cloning. */
142 return false;
143 }
144
145 static const struct pipe_buf_operations smc_pipe_ops = {
146 .release = smc_rx_pipe_buf_release,
147 .get = smc_rx_pipe_buf_get,
148 };
149
smc_rx_spd_release(struct splice_pipe_desc * spd,unsigned int i)150 static void smc_rx_spd_release(struct splice_pipe_desc *spd,
151 unsigned int i)
152 {
153 put_page(spd->pages[i]);
154 }
155
smc_rx_splice(struct pipe_inode_info * pipe,char * src,size_t len,struct smc_sock * smc)156 static int smc_rx_splice(struct pipe_inode_info *pipe, char *src, size_t len,
157 struct smc_sock *smc)
158 {
159 struct smc_link_group *lgr = smc->conn.lgr;
160 int offset = offset_in_page(src);
161 struct partial_page *partial;
162 struct splice_pipe_desc spd;
163 struct smc_spd_priv **priv;
164 struct page **pages;
165 int bytes, nr_pages;
166 int i;
167
168 nr_pages = !lgr->is_smcd && smc->conn.rmb_desc->is_vm ?
169 PAGE_ALIGN(len + offset) / PAGE_SIZE : 1;
170
171 pages = kzalloc_objs(*pages, nr_pages);
172 if (!pages)
173 goto out;
174 partial = kzalloc_objs(*partial, nr_pages);
175 if (!partial)
176 goto out_page;
177 priv = kzalloc_objs(*priv, nr_pages);
178 if (!priv)
179 goto out_part;
180 for (i = 0; i < nr_pages; i++) {
181 priv[i] = kzalloc_obj(**priv);
182 if (!priv[i])
183 goto out_priv;
184 }
185
186 if (lgr->is_smcd ||
187 (!lgr->is_smcd && !smc->conn.rmb_desc->is_vm)) {
188 /* smcd or smcr that uses physically contiguous RMBs */
189 priv[0]->len = len;
190 priv[0]->smc = smc;
191 partial[0].offset = src - (char *)smc->conn.rmb_desc->cpu_addr;
192 partial[0].len = len;
193 partial[0].private = (unsigned long)priv[0];
194 pages[0] = smc->conn.rmb_desc->pages;
195 } else {
196 int size, left = len;
197 void *buf = src;
198 /* smcr that uses virtually contiguous RMBs*/
199 for (i = 0; i < nr_pages; i++) {
200 size = min_t(int, PAGE_SIZE - offset, left);
201 priv[i]->len = size;
202 priv[i]->smc = smc;
203 pages[i] = vmalloc_to_page(buf);
204 partial[i].offset = offset;
205 partial[i].len = size;
206 partial[i].private = (unsigned long)priv[i];
207 buf += size;
208 left -= size;
209 offset = 0;
210 }
211 }
212 spd.nr_pages_max = nr_pages;
213 spd.nr_pages = nr_pages;
214 spd.pages = pages;
215 spd.partial = partial;
216 spd.ops = &smc_pipe_ops;
217 spd.spd_release = smc_rx_spd_release;
218
219 bytes = splice_to_pipe(pipe, &spd);
220 if (bytes > 0) {
221 sock_hold(&smc->sk);
222 if (!lgr->is_smcd && smc->conn.rmb_desc->is_vm) {
223 for (i = 0; i < PAGE_ALIGN(bytes + offset) / PAGE_SIZE; i++)
224 get_page(pages[i]);
225 } else {
226 get_page(smc->conn.rmb_desc->pages);
227 }
228 atomic_add(bytes, &smc->conn.splice_pending);
229 }
230 kfree(priv);
231 kfree(partial);
232 kfree(pages);
233
234 return bytes;
235
236 out_priv:
237 for (i = (i - 1); i >= 0; i--)
238 kfree(priv[i]);
239 kfree(priv);
240 out_part:
241 kfree(partial);
242 out_page:
243 kfree(pages);
244 out:
245 return -ENOMEM;
246 }
247
smc_rx_data_available_and_no_splice_pend(struct smc_connection * conn,size_t peeked)248 static int smc_rx_data_available_and_no_splice_pend(struct smc_connection *conn, size_t peeked)
249 {
250 return smc_rx_data_available(conn, peeked) &&
251 !atomic_read(&conn->splice_pending);
252 }
253
254 /* blocks rcvbuf consumer until >=len bytes available or timeout or interrupted
255 * @smc smc socket
256 * @timeo pointer to max seconds to wait, pointer to value 0 for no timeout
257 * @peeked number of bytes already peeked
258 * @fcrit add'l criterion to evaluate as function pointer
259 * Returns:
260 * 1 if at least 1 byte available in rcvbuf or if socket error/shutdown.
261 * 0 otherwise (nothing in rcvbuf nor timeout, e.g. interrupted).
262 */
smc_rx_wait(struct smc_sock * smc,long * timeo,size_t peeked,int (* fcrit)(struct smc_connection * conn,size_t baseline))263 int smc_rx_wait(struct smc_sock *smc, long *timeo, size_t peeked,
264 int (*fcrit)(struct smc_connection *conn, size_t baseline))
265 {
266 DEFINE_WAIT_FUNC(wait, woken_wake_function);
267 struct smc_connection *conn = &smc->conn;
268 struct smc_cdc_conn_state_flags *cflags =
269 &conn->local_tx_ctrl.conn_state_flags;
270 struct sock *sk = &smc->sk;
271 int rc;
272
273 if (fcrit(conn, peeked))
274 return 1;
275 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
276 add_wait_queue(sk_sleep(sk), &wait);
277 rc = sk_wait_event(sk, timeo,
278 READ_ONCE(sk->sk_err) ||
279 cflags->peer_conn_abort ||
280 READ_ONCE(sk->sk_shutdown) & RCV_SHUTDOWN ||
281 conn->killed ||
282 fcrit(conn, peeked),
283 &wait);
284 remove_wait_queue(sk_sleep(sk), &wait);
285 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
286 return rc;
287 }
288
smc_rx_recv_urg(struct smc_sock * smc,struct msghdr * msg,int len,int flags)289 static int smc_rx_recv_urg(struct smc_sock *smc, struct msghdr *msg, int len,
290 int flags)
291 {
292 struct smc_connection *conn = &smc->conn;
293 union smc_host_cursor cons;
294 struct sock *sk = &smc->sk;
295 int rc = 0;
296
297 if (sock_flag(sk, SOCK_URGINLINE) ||
298 !(conn->urg_state == SMC_URG_VALID) ||
299 conn->urg_state == SMC_URG_READ)
300 return -EINVAL;
301
302 SMC_STAT_INC(smc, urg_data_cnt);
303 if (conn->urg_state == SMC_URG_VALID) {
304 if (!(flags & MSG_PEEK))
305 smc->conn.urg_state = SMC_URG_READ;
306 msg->msg_flags |= MSG_OOB;
307 if (len > 0) {
308 if (!(flags & MSG_TRUNC))
309 rc = memcpy_to_msg(msg, &conn->urg_rx_byte, 1);
310 len = 1;
311 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
312 if (smc_curs_diff(conn->rmb_desc->len, &cons,
313 &conn->urg_curs) > 1)
314 conn->urg_rx_skip_pend = true;
315 /* Urgent Byte was already accounted for, but trigger
316 * skipping the urgent byte in non-inline case
317 */
318 if (!(flags & MSG_PEEK))
319 smc_rx_update_consumer(smc, cons, 0);
320 } else {
321 msg->msg_flags |= MSG_TRUNC;
322 }
323
324 return rc ? -EFAULT : len;
325 }
326
327 if (sk->sk_state == SMC_CLOSED || sk->sk_shutdown & RCV_SHUTDOWN)
328 return 0;
329
330 return -EAGAIN;
331 }
332
smc_rx_recvmsg_data_available(struct smc_sock * smc,size_t peeked)333 static bool smc_rx_recvmsg_data_available(struct smc_sock *smc, size_t peeked)
334 {
335 struct smc_connection *conn = &smc->conn;
336
337 if (smc_rx_data_available(conn, peeked))
338 return true;
339 else if (conn->urg_state == SMC_URG_VALID)
340 /* we received a single urgent Byte - skip */
341 smc_rx_update_cons(smc, 0);
342 return false;
343 }
344
345 /* smc_rx_recvmsg - receive data from RMBE
346 * @msg: copy data to receive buffer
347 * @pipe: copy data to pipe if set - indicates splice() call
348 *
349 * rcvbuf consumer: main API called by socket layer.
350 * Called under sk lock.
351 */
smc_rx_recvmsg(struct smc_sock * smc,struct msghdr * msg,struct pipe_inode_info * pipe,size_t len,int flags)352 int smc_rx_recvmsg(struct smc_sock *smc, struct msghdr *msg,
353 struct pipe_inode_info *pipe, size_t len, int flags)
354 {
355 size_t copylen, read_done = 0, read_remaining = len, peeked_bytes = 0;
356 size_t chunk_len, chunk_off, chunk_len_sum;
357 struct smc_connection *conn = &smc->conn;
358 int (*func)(struct smc_connection *conn, size_t baseline);
359 union smc_host_cursor cons;
360 int readable, chunk;
361 char *rcvbuf_base;
362 struct sock *sk;
363 int splbytes;
364 long timeo;
365 int target; /* Read at least these many bytes */
366 int rc;
367
368 if (unlikely(flags & MSG_ERRQUEUE))
369 return -EINVAL; /* future work for sk.sk_family == AF_SMC */
370
371 sk = &smc->sk;
372 if (sk->sk_state == SMC_LISTEN)
373 return -ENOTCONN;
374 if (flags & MSG_OOB)
375 return smc_rx_recv_urg(smc, msg, len, flags);
376 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
377 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
378
379 readable = atomic_read(&conn->bytes_to_rcv);
380 if (readable >= conn->rmb_desc->len)
381 SMC_STAT_RMB_RX_FULL(smc, !conn->lnk);
382
383 if (len < readable)
384 SMC_STAT_RMB_RX_SIZE_SMALL(smc, !conn->lnk);
385 /* we currently use 1 RMBE per RMB, so RMBE == RMB base addr */
386 rcvbuf_base = conn->rx_off + conn->rmb_desc->cpu_addr;
387
388 do { /* while (read_remaining) */
389 if (read_done >= target || (pipe && read_done))
390 break;
391
392 if (conn->killed)
393 break;
394
395 if (smc_rx_recvmsg_data_available(smc, peeked_bytes))
396 goto copy;
397
398 if (sk->sk_shutdown & RCV_SHUTDOWN) {
399 /* smc_cdc_msg_recv_action() could have run after
400 * above smc_rx_recvmsg_data_available()
401 */
402 if (smc_rx_recvmsg_data_available(smc, peeked_bytes))
403 goto copy;
404 break;
405 }
406
407 if (read_done) {
408 if (sk->sk_err ||
409 sk->sk_state == SMC_CLOSED ||
410 !timeo ||
411 signal_pending(current))
412 break;
413 } else {
414 if (sk->sk_err) {
415 read_done = sock_error(sk);
416 break;
417 }
418 if (sk->sk_state == SMC_CLOSED) {
419 if (!sock_flag(sk, SOCK_DONE)) {
420 /* This occurs when user tries to read
421 * from never connected socket.
422 */
423 read_done = -ENOTCONN;
424 break;
425 }
426 break;
427 }
428 if (!timeo)
429 return -EAGAIN;
430 if (signal_pending(current)) {
431 read_done = sock_intr_errno(timeo);
432 break;
433 }
434 }
435
436 if (!smc_rx_data_available(conn, peeked_bytes)) {
437 smc_rx_wait(smc, &timeo, peeked_bytes, smc_rx_data_available);
438 continue;
439 }
440
441 copy:
442 /* initialize variables for 1st iteration of subsequent loop */
443 /* could be just 1 byte, even after waiting on data above */
444 readable = smc_rx_data_available(conn, peeked_bytes);
445 splbytes = atomic_read(&conn->splice_pending);
446 if (!readable || (msg && splbytes)) {
447 if (splbytes)
448 func = smc_rx_data_available_and_no_splice_pend;
449 else
450 func = smc_rx_data_available;
451 smc_rx_wait(smc, &timeo, peeked_bytes, func);
452 continue;
453 }
454
455 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
456 if ((flags & MSG_PEEK) && peeked_bytes)
457 smc_curs_add(conn->rmb_desc->len, &cons, peeked_bytes);
458 /* subsequent splice() calls pick up where previous left */
459 if (splbytes)
460 smc_curs_add(conn->rmb_desc->len, &cons, splbytes);
461 if (conn->urg_state == SMC_URG_VALID &&
462 sock_flag(&smc->sk, SOCK_URGINLINE) &&
463 readable > 1)
464 readable--; /* always stop at urgent Byte */
465 /* not more than what user space asked for */
466 copylen = min_t(size_t, read_remaining, readable);
467 /* determine chunks where to read from rcvbuf */
468 /* either unwrapped case, or 1st chunk of wrapped case */
469 chunk_len = min_t(size_t, copylen, conn->rmb_desc->len -
470 cons.count);
471 chunk_len_sum = chunk_len;
472 chunk_off = cons.count;
473 smc_rmb_sync_sg_for_cpu(conn);
474 for (chunk = 0; chunk < 2; chunk++) {
475 if (!(flags & MSG_TRUNC)) {
476 if (msg) {
477 rc = memcpy_to_msg(msg, rcvbuf_base +
478 chunk_off,
479 chunk_len);
480 } else {
481 rc = smc_rx_splice(pipe, rcvbuf_base +
482 chunk_off, chunk_len,
483 smc);
484 }
485 if (rc < 0) {
486 if (!read_done)
487 read_done = -EFAULT;
488 goto out;
489 }
490 }
491 read_remaining -= chunk_len;
492 read_done += chunk_len;
493 if (flags & MSG_PEEK)
494 peeked_bytes += chunk_len;
495
496 if (chunk_len_sum == copylen)
497 break; /* either on 1st or 2nd iteration */
498 /* prepare next (== 2nd) iteration */
499 chunk_len = copylen - chunk_len; /* remainder */
500 chunk_len_sum += chunk_len;
501 chunk_off = 0; /* modulo offset in recv ring buffer */
502 }
503
504 /* update cursors */
505 if (!(flags & MSG_PEEK)) {
506 /* increased in recv tasklet smc_cdc_msg_rcv() */
507 smp_mb__before_atomic();
508 atomic_sub(copylen, &conn->bytes_to_rcv);
509 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
510 smp_mb__after_atomic();
511 if (msg && smc_rx_update_consumer(smc, cons, copylen))
512 goto out;
513 }
514
515 trace_smc_rx_recvmsg(smc, copylen);
516 } while (read_remaining);
517 out:
518 return read_done;
519 }
520
521 /* Initialize receive properties on connection establishment. NB: not __init! */
smc_rx_init(struct smc_sock * smc)522 void smc_rx_init(struct smc_sock *smc)
523 {
524 smc->sk.sk_data_ready = smc_rx_wake_up;
525 atomic_set(&smc->conn.splice_pending, 0);
526 smc->conn.urg_state = SMC_URG_READ;
527 }
528