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 struct smc_spd_priv *priv = (struct smc_spd_priv *)spd->partial[i].private;
154 struct sock *sk = &priv->smc->sk;
155
156 kfree(priv);
157 put_page(spd->pages[i]);
158 sock_put(sk);
159 }
160
smc_rx_splice(struct pipe_inode_info * pipe,char * src,size_t len,struct smc_sock * smc)161 static int smc_rx_splice(struct pipe_inode_info *pipe, char *src, size_t len,
162 struct smc_sock *smc)
163 {
164 struct smc_link_group *lgr = smc->conn.lgr;
165 int offset = offset_in_page(src);
166 struct partial_page *partial;
167 struct splice_pipe_desc spd;
168 struct smc_spd_priv **priv;
169 struct page **pages;
170 int bytes, nr_pages;
171 int i;
172
173 nr_pages = !lgr->is_smcd && smc->conn.rmb_desc->is_vm ?
174 PAGE_ALIGN(len + offset) / PAGE_SIZE : 1;
175
176 pages = kzalloc_objs(*pages, nr_pages);
177 if (!pages)
178 goto out;
179 partial = kzalloc_objs(*partial, nr_pages);
180 if (!partial)
181 goto out_page;
182 priv = kzalloc_objs(*priv, nr_pages);
183 if (!priv)
184 goto out_part;
185 for (i = 0; i < nr_pages; i++) {
186 priv[i] = kzalloc_obj(**priv);
187 if (!priv[i])
188 goto out_priv;
189 }
190
191 if (lgr->is_smcd ||
192 (!lgr->is_smcd && !smc->conn.rmb_desc->is_vm)) {
193 /* smcd or smcr that uses physically contiguous RMBs */
194 priv[0]->len = len;
195 priv[0]->smc = smc;
196 partial[0].offset = src - (char *)smc->conn.rmb_desc->cpu_addr;
197 partial[0].len = len;
198 partial[0].private = (unsigned long)priv[0];
199 pages[0] = smc->conn.rmb_desc->pages;
200 } else {
201 int size, left = len;
202 void *buf = src;
203 /* smcr that uses virtually contiguous RMBs*/
204 for (i = 0; i < nr_pages; i++) {
205 size = min_t(int, PAGE_SIZE - offset, left);
206 priv[i]->len = size;
207 priv[i]->smc = smc;
208 pages[i] = vmalloc_to_page(buf);
209 partial[i].offset = offset;
210 partial[i].len = size;
211 partial[i].private = (unsigned long)priv[i];
212 buf += size;
213 left -= size;
214 offset = 0;
215 }
216 }
217 for (i = 0; i < nr_pages; i++) {
218 get_page(pages[i]);
219 sock_hold(&smc->sk);
220 }
221 spd.nr_pages_max = nr_pages;
222 spd.nr_pages = nr_pages;
223 spd.pages = pages;
224 spd.partial = partial;
225 spd.ops = &smc_pipe_ops;
226 spd.spd_release = smc_rx_spd_release;
227
228 bytes = splice_to_pipe(pipe, &spd);
229 if (bytes > 0)
230 atomic_add(bytes, &smc->conn.splice_pending);
231 kfree(priv);
232 kfree(partial);
233 kfree(pages);
234
235 return bytes;
236
237 out_priv:
238 for (i = (i - 1); i >= 0; i--)
239 kfree(priv[i]);
240 kfree(priv);
241 out_part:
242 kfree(partial);
243 out_page:
244 kfree(pages);
245 out:
246 return -ENOMEM;
247 }
248
smc_rx_data_available_and_no_splice_pend(struct smc_connection * conn,size_t peeked)249 static int smc_rx_data_available_and_no_splice_pend(struct smc_connection *conn, size_t peeked)
250 {
251 return smc_rx_data_available(conn, peeked) &&
252 !atomic_read(&conn->splice_pending);
253 }
254
255 /* blocks rcvbuf consumer until >=len bytes available or timeout or interrupted
256 * @smc smc socket
257 * @timeo pointer to max seconds to wait, pointer to value 0 for no timeout
258 * @peeked number of bytes already peeked
259 * @fcrit add'l criterion to evaluate as function pointer
260 * Returns:
261 * 1 if at least 1 byte available in rcvbuf or if socket error/shutdown.
262 * 0 otherwise (nothing in rcvbuf nor timeout, e.g. interrupted).
263 */
smc_rx_wait(struct smc_sock * smc,long * timeo,size_t peeked,int (* fcrit)(struct smc_connection * conn,size_t baseline))264 int smc_rx_wait(struct smc_sock *smc, long *timeo, size_t peeked,
265 int (*fcrit)(struct smc_connection *conn, size_t baseline))
266 {
267 DEFINE_WAIT_FUNC(wait, woken_wake_function);
268 struct smc_connection *conn = &smc->conn;
269 struct smc_cdc_conn_state_flags *cflags =
270 &conn->local_tx_ctrl.conn_state_flags;
271 struct sock *sk = &smc->sk;
272 int rc;
273
274 if (fcrit(conn, peeked))
275 return 1;
276 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
277 add_wait_queue(sk_sleep(sk), &wait);
278 rc = sk_wait_event(sk, timeo,
279 READ_ONCE(sk->sk_err) ||
280 cflags->peer_conn_abort ||
281 READ_ONCE(sk->sk_shutdown) & RCV_SHUTDOWN ||
282 conn->killed ||
283 fcrit(conn, peeked),
284 &wait);
285 remove_wait_queue(sk_sleep(sk), &wait);
286 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
287 return rc;
288 }
289
smc_rx_recv_urg(struct smc_sock * smc,struct msghdr * msg,int len,int flags)290 static int smc_rx_recv_urg(struct smc_sock *smc, struct msghdr *msg, int len,
291 int flags)
292 {
293 struct smc_connection *conn = &smc->conn;
294 union smc_host_cursor cons;
295 struct sock *sk = &smc->sk;
296 int rc = 0;
297
298 if (sock_flag(sk, SOCK_URGINLINE) ||
299 !(conn->urg_state == SMC_URG_VALID) ||
300 conn->urg_state == SMC_URG_READ)
301 return -EINVAL;
302
303 SMC_STAT_INC(smc, urg_data_cnt);
304 if (conn->urg_state == SMC_URG_VALID) {
305 if (!(flags & MSG_PEEK))
306 smc->conn.urg_state = SMC_URG_READ;
307 msg->msg_flags |= MSG_OOB;
308 if (len > 0) {
309 if (!(flags & MSG_TRUNC))
310 rc = memcpy_to_msg(msg, &conn->urg_rx_byte, 1);
311 len = 1;
312 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
313 if (smc_curs_diff(conn->rmb_desc->len, &cons,
314 &conn->urg_curs) > 1)
315 conn->urg_rx_skip_pend = true;
316 /* Urgent Byte was already accounted for, but trigger
317 * skipping the urgent byte in non-inline case
318 */
319 if (!(flags & MSG_PEEK))
320 smc_rx_update_consumer(smc, cons, 0);
321 } else {
322 msg->msg_flags |= MSG_TRUNC;
323 }
324
325 return rc ? -EFAULT : len;
326 }
327
328 if (sk->sk_state == SMC_CLOSED || sk->sk_shutdown & RCV_SHUTDOWN)
329 return 0;
330
331 return -EAGAIN;
332 }
333
smc_rx_recvmsg_data_available(struct smc_sock * smc,size_t peeked)334 static bool smc_rx_recvmsg_data_available(struct smc_sock *smc, size_t peeked)
335 {
336 struct smc_connection *conn = &smc->conn;
337
338 if (smc_rx_data_available(conn, peeked))
339 return true;
340 else if (conn->urg_state == SMC_URG_VALID)
341 /* we received a single urgent Byte - skip */
342 smc_rx_update_cons(smc, 0);
343 return false;
344 }
345
346 /* smc_rx_recvmsg - receive data from RMBE
347 * @msg: copy data to receive buffer
348 * @pipe: copy data to pipe if set - indicates splice() call
349 *
350 * rcvbuf consumer: main API called by socket layer.
351 * Called under sk lock.
352 */
smc_rx_recvmsg(struct smc_sock * smc,struct msghdr * msg,struct pipe_inode_info * pipe,size_t len,int flags)353 int smc_rx_recvmsg(struct smc_sock *smc, struct msghdr *msg,
354 struct pipe_inode_info *pipe, size_t len, int flags)
355 {
356 size_t copylen, read_done = 0, read_remaining = len, peeked_bytes = 0;
357 size_t chunk_len, chunk_off, chunk_len_sum;
358 struct smc_connection *conn = &smc->conn;
359 int (*func)(struct smc_connection *conn, size_t baseline);
360 union smc_host_cursor cons;
361 int readable, chunk;
362 char *rcvbuf_base;
363 struct sock *sk;
364 int splbytes;
365 long timeo;
366 int target; /* Read at least these many bytes */
367 int rc;
368
369 if (unlikely(flags & MSG_ERRQUEUE))
370 return -EINVAL; /* future work for sk.sk_family == AF_SMC */
371
372 sk = &smc->sk;
373 if (sk->sk_state == SMC_LISTEN)
374 return -ENOTCONN;
375 if (flags & MSG_OOB)
376 return smc_rx_recv_urg(smc, msg, len, flags);
377 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
378 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
379
380 readable = atomic_read(&conn->bytes_to_rcv);
381 if (readable >= conn->rmb_desc->len)
382 SMC_STAT_RMB_RX_FULL(smc, !conn->lnk);
383
384 if (len < readable)
385 SMC_STAT_RMB_RX_SIZE_SMALL(smc, !conn->lnk);
386 /* we currently use 1 RMBE per RMB, so RMBE == RMB base addr */
387 rcvbuf_base = conn->rx_off + conn->rmb_desc->cpu_addr;
388
389 do { /* while (read_remaining) */
390 if (read_done >= target || (pipe && read_done))
391 break;
392
393 if (conn->killed)
394 break;
395
396 if (smc_rx_recvmsg_data_available(smc, peeked_bytes))
397 goto copy;
398
399 if (sk->sk_shutdown & RCV_SHUTDOWN) {
400 /* smc_cdc_msg_recv_action() could have run after
401 * above smc_rx_recvmsg_data_available()
402 */
403 if (smc_rx_recvmsg_data_available(smc, peeked_bytes))
404 goto copy;
405 break;
406 }
407
408 if (read_done) {
409 if (sk->sk_err ||
410 sk->sk_state == SMC_CLOSED ||
411 !timeo ||
412 signal_pending(current))
413 break;
414 } else {
415 if (sk->sk_err) {
416 read_done = sock_error(sk);
417 break;
418 }
419 if (sk->sk_state == SMC_CLOSED) {
420 if (!sock_flag(sk, SOCK_DONE)) {
421 /* This occurs when user tries to read
422 * from never connected socket.
423 */
424 read_done = -ENOTCONN;
425 break;
426 }
427 break;
428 }
429 if (!timeo)
430 return -EAGAIN;
431 if (signal_pending(current)) {
432 read_done = sock_intr_errno(timeo);
433 break;
434 }
435 }
436
437 if (!smc_rx_data_available(conn, peeked_bytes)) {
438 smc_rx_wait(smc, &timeo, peeked_bytes, smc_rx_data_available);
439 continue;
440 }
441
442 copy:
443 /* initialize variables for 1st iteration of subsequent loop */
444 /* could be just 1 byte, even after waiting on data above */
445 readable = smc_rx_data_available(conn, peeked_bytes);
446 splbytes = atomic_read(&conn->splice_pending);
447 if (!readable || (msg && splbytes)) {
448 if (splbytes)
449 func = smc_rx_data_available_and_no_splice_pend;
450 else
451 func = smc_rx_data_available;
452 smc_rx_wait(smc, &timeo, peeked_bytes, func);
453 continue;
454 }
455
456 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
457 if ((flags & MSG_PEEK) && peeked_bytes)
458 smc_curs_add(conn->rmb_desc->len, &cons, peeked_bytes);
459 /* subsequent splice() calls pick up where previous left */
460 if (splbytes)
461 smc_curs_add(conn->rmb_desc->len, &cons, splbytes);
462 if (conn->urg_state == SMC_URG_VALID &&
463 sock_flag(&smc->sk, SOCK_URGINLINE) &&
464 readable > 1)
465 readable--; /* always stop at urgent Byte */
466 /* not more than what user space asked for */
467 copylen = min_t(size_t, read_remaining, readable);
468 /* determine chunks where to read from rcvbuf */
469 /* either unwrapped case, or 1st chunk of wrapped case */
470 chunk_len = min_t(size_t, copylen, conn->rmb_desc->len -
471 cons.count);
472 chunk_len_sum = chunk_len;
473 chunk_off = cons.count;
474 smc_rmb_sync_sg_for_cpu(conn);
475 for (chunk = 0; chunk < 2; chunk++) {
476 if (!(flags & MSG_TRUNC)) {
477 if (msg) {
478 rc = memcpy_to_msg(msg, rcvbuf_base +
479 chunk_off,
480 chunk_len);
481 } else {
482 rc = smc_rx_splice(pipe, rcvbuf_base +
483 chunk_off, chunk_len,
484 smc);
485 }
486 if (rc < 0) {
487 if (!read_done)
488 read_done = -EFAULT;
489 goto out;
490 }
491 }
492 read_remaining -= chunk_len;
493 read_done += chunk_len;
494 if (flags & MSG_PEEK)
495 peeked_bytes += chunk_len;
496
497 if (chunk_len_sum == copylen)
498 break; /* either on 1st or 2nd iteration */
499 /* prepare next (== 2nd) iteration */
500 chunk_len = copylen - chunk_len; /* remainder */
501 chunk_len_sum += chunk_len;
502 chunk_off = 0; /* modulo offset in recv ring buffer */
503 }
504
505 /* update cursors */
506 if (!(flags & MSG_PEEK)) {
507 /* increased in recv tasklet smc_cdc_msg_rcv() */
508 smp_mb__before_atomic();
509 atomic_sub(copylen, &conn->bytes_to_rcv);
510 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
511 smp_mb__after_atomic();
512 if (msg && smc_rx_update_consumer(smc, cons, copylen))
513 goto out;
514 }
515
516 trace_smc_rx_recvmsg(smc, copylen);
517 } while (read_remaining);
518 out:
519 return read_done;
520 }
521
522 /* Initialize receive properties on connection establishment. NB: not __init! */
smc_rx_init(struct smc_sock * smc)523 void smc_rx_init(struct smc_sock *smc)
524 {
525 smc->sk.sk_data_ready = smc_rx_wake_up;
526 atomic_set(&smc->conn.splice_pending, 0);
527 smc->conn.urg_state = SMC_URG_READ;
528 }
529