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(sk, SOCK_WAKE_WAITD, POLL_IN);
46 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
47 (sk->sk_state == SMC_CLOSED))
48 sk_wake_async(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
138 static const struct pipe_buf_operations smc_pipe_ops = {
139 .release = smc_rx_pipe_buf_release,
140 .get = generic_pipe_buf_get
141 };
142
smc_rx_spd_release(struct splice_pipe_desc * spd,unsigned int i)143 static void smc_rx_spd_release(struct splice_pipe_desc *spd,
144 unsigned int i)
145 {
146 put_page(spd->pages[i]);
147 }
148
smc_rx_splice(struct pipe_inode_info * pipe,char * src,size_t len,struct smc_sock * smc)149 static int smc_rx_splice(struct pipe_inode_info *pipe, char *src, size_t len,
150 struct smc_sock *smc)
151 {
152 struct smc_link_group *lgr = smc->conn.lgr;
153 int offset = offset_in_page(src);
154 struct partial_page *partial;
155 struct splice_pipe_desc spd;
156 struct smc_spd_priv **priv;
157 struct page **pages;
158 int bytes, nr_pages;
159 int i;
160
161 nr_pages = !lgr->is_smcd && smc->conn.rmb_desc->is_vm ?
162 PAGE_ALIGN(len + offset) / PAGE_SIZE : 1;
163
164 pages = kcalloc(nr_pages, sizeof(*pages), GFP_KERNEL);
165 if (!pages)
166 goto out;
167 partial = kcalloc(nr_pages, sizeof(*partial), GFP_KERNEL);
168 if (!partial)
169 goto out_page;
170 priv = kcalloc(nr_pages, sizeof(*priv), GFP_KERNEL);
171 if (!priv)
172 goto out_part;
173 for (i = 0; i < nr_pages; i++) {
174 priv[i] = kzalloc(sizeof(**priv), GFP_KERNEL);
175 if (!priv[i])
176 goto out_priv;
177 }
178
179 if (lgr->is_smcd ||
180 (!lgr->is_smcd && !smc->conn.rmb_desc->is_vm)) {
181 /* smcd or smcr that uses physically contiguous RMBs */
182 priv[0]->len = len;
183 priv[0]->smc = smc;
184 partial[0].offset = src - (char *)smc->conn.rmb_desc->cpu_addr;
185 partial[0].len = len;
186 partial[0].private = (unsigned long)priv[0];
187 pages[0] = smc->conn.rmb_desc->pages;
188 } else {
189 int size, left = len;
190 void *buf = src;
191 /* smcr that uses virtually contiguous RMBs*/
192 for (i = 0; i < nr_pages; i++) {
193 size = min_t(int, PAGE_SIZE - offset, left);
194 priv[i]->len = size;
195 priv[i]->smc = smc;
196 pages[i] = vmalloc_to_page(buf);
197 partial[i].offset = offset;
198 partial[i].len = size;
199 partial[i].private = (unsigned long)priv[i];
200 buf += size / sizeof(*buf);
201 left -= size;
202 offset = 0;
203 }
204 }
205 spd.nr_pages_max = nr_pages;
206 spd.nr_pages = nr_pages;
207 spd.pages = pages;
208 spd.partial = partial;
209 spd.ops = &smc_pipe_ops;
210 spd.spd_release = smc_rx_spd_release;
211
212 bytes = splice_to_pipe(pipe, &spd);
213 if (bytes > 0) {
214 sock_hold(&smc->sk);
215 if (!lgr->is_smcd && smc->conn.rmb_desc->is_vm) {
216 for (i = 0; i < PAGE_ALIGN(bytes + offset) / PAGE_SIZE; i++)
217 get_page(pages[i]);
218 } else {
219 get_page(smc->conn.rmb_desc->pages);
220 }
221 atomic_add(bytes, &smc->conn.splice_pending);
222 }
223 kfree(priv);
224 kfree(partial);
225 kfree(pages);
226
227 return bytes;
228
229 out_priv:
230 for (i = (i - 1); i >= 0; i--)
231 kfree(priv[i]);
232 kfree(priv);
233 out_part:
234 kfree(partial);
235 out_page:
236 kfree(pages);
237 out:
238 return -ENOMEM;
239 }
240
smc_rx_data_available_and_no_splice_pend(struct smc_connection * conn,size_t peeked)241 static int smc_rx_data_available_and_no_splice_pend(struct smc_connection *conn, size_t peeked)
242 {
243 return smc_rx_data_available(conn, peeked) &&
244 !atomic_read(&conn->splice_pending);
245 }
246
247 /* blocks rcvbuf consumer until >=len bytes available or timeout or interrupted
248 * @smc smc socket
249 * @timeo pointer to max seconds to wait, pointer to value 0 for no timeout
250 * @peeked number of bytes already peeked
251 * @fcrit add'l criterion to evaluate as function pointer
252 * Returns:
253 * 1 if at least 1 byte available in rcvbuf or if socket error/shutdown.
254 * 0 otherwise (nothing in rcvbuf nor timeout, e.g. interrupted).
255 */
smc_rx_wait(struct smc_sock * smc,long * timeo,size_t peeked,int (* fcrit)(struct smc_connection * conn,size_t baseline))256 int smc_rx_wait(struct smc_sock *smc, long *timeo, size_t peeked,
257 int (*fcrit)(struct smc_connection *conn, size_t baseline))
258 {
259 DEFINE_WAIT_FUNC(wait, woken_wake_function);
260 struct smc_connection *conn = &smc->conn;
261 struct smc_cdc_conn_state_flags *cflags =
262 &conn->local_tx_ctrl.conn_state_flags;
263 struct sock *sk = &smc->sk;
264 int rc;
265
266 if (fcrit(conn, peeked))
267 return 1;
268 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
269 add_wait_queue(sk_sleep(sk), &wait);
270 rc = sk_wait_event(sk, timeo,
271 READ_ONCE(sk->sk_err) ||
272 cflags->peer_conn_abort ||
273 READ_ONCE(sk->sk_shutdown) & RCV_SHUTDOWN ||
274 conn->killed ||
275 fcrit(conn, peeked),
276 &wait);
277 remove_wait_queue(sk_sleep(sk), &wait);
278 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
279 return rc;
280 }
281
smc_rx_recv_urg(struct smc_sock * smc,struct msghdr * msg,int len,int flags)282 static int smc_rx_recv_urg(struct smc_sock *smc, struct msghdr *msg, int len,
283 int flags)
284 {
285 struct smc_connection *conn = &smc->conn;
286 union smc_host_cursor cons;
287 struct sock *sk = &smc->sk;
288 int rc = 0;
289
290 if (sock_flag(sk, SOCK_URGINLINE) ||
291 !(conn->urg_state == SMC_URG_VALID) ||
292 conn->urg_state == SMC_URG_READ)
293 return -EINVAL;
294
295 SMC_STAT_INC(smc, urg_data_cnt);
296 if (conn->urg_state == SMC_URG_VALID) {
297 if (!(flags & MSG_PEEK))
298 smc->conn.urg_state = SMC_URG_READ;
299 msg->msg_flags |= MSG_OOB;
300 if (len > 0) {
301 if (!(flags & MSG_TRUNC))
302 rc = memcpy_to_msg(msg, &conn->urg_rx_byte, 1);
303 len = 1;
304 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
305 if (smc_curs_diff(conn->rmb_desc->len, &cons,
306 &conn->urg_curs) > 1)
307 conn->urg_rx_skip_pend = true;
308 /* Urgent Byte was already accounted for, but trigger
309 * skipping the urgent byte in non-inline case
310 */
311 if (!(flags & MSG_PEEK))
312 smc_rx_update_consumer(smc, cons, 0);
313 } else {
314 msg->msg_flags |= MSG_TRUNC;
315 }
316
317 return rc ? -EFAULT : len;
318 }
319
320 if (sk->sk_state == SMC_CLOSED || sk->sk_shutdown & RCV_SHUTDOWN)
321 return 0;
322
323 return -EAGAIN;
324 }
325
smc_rx_recvmsg_data_available(struct smc_sock * smc,size_t peeked)326 static bool smc_rx_recvmsg_data_available(struct smc_sock *smc, size_t peeked)
327 {
328 struct smc_connection *conn = &smc->conn;
329
330 if (smc_rx_data_available(conn, peeked))
331 return true;
332 else if (conn->urg_state == SMC_URG_VALID)
333 /* we received a single urgent Byte - skip */
334 smc_rx_update_cons(smc, 0);
335 return false;
336 }
337
338 /* smc_rx_recvmsg - receive data from RMBE
339 * @msg: copy data to receive buffer
340 * @pipe: copy data to pipe if set - indicates splice() call
341 *
342 * rcvbuf consumer: main API called by socket layer.
343 * Called under sk lock.
344 */
smc_rx_recvmsg(struct smc_sock * smc,struct msghdr * msg,struct pipe_inode_info * pipe,size_t len,int flags)345 int smc_rx_recvmsg(struct smc_sock *smc, struct msghdr *msg,
346 struct pipe_inode_info *pipe, size_t len, int flags)
347 {
348 size_t copylen, read_done = 0, read_remaining = len, peeked_bytes = 0;
349 size_t chunk_len, chunk_off, chunk_len_sum;
350 struct smc_connection *conn = &smc->conn;
351 int (*func)(struct smc_connection *conn, size_t baseline);
352 union smc_host_cursor cons;
353 int readable, chunk;
354 char *rcvbuf_base;
355 struct sock *sk;
356 int splbytes;
357 long timeo;
358 int target; /* Read at least these many bytes */
359 int rc;
360
361 if (unlikely(flags & MSG_ERRQUEUE))
362 return -EINVAL; /* future work for sk.sk_family == AF_SMC */
363
364 sk = &smc->sk;
365 if (sk->sk_state == SMC_LISTEN)
366 return -ENOTCONN;
367 if (flags & MSG_OOB)
368 return smc_rx_recv_urg(smc, msg, len, flags);
369 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
370 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
371
372 readable = atomic_read(&conn->bytes_to_rcv);
373 if (readable >= conn->rmb_desc->len)
374 SMC_STAT_RMB_RX_FULL(smc, !conn->lnk);
375
376 if (len < readable)
377 SMC_STAT_RMB_RX_SIZE_SMALL(smc, !conn->lnk);
378 /* we currently use 1 RMBE per RMB, so RMBE == RMB base addr */
379 rcvbuf_base = conn->rx_off + conn->rmb_desc->cpu_addr;
380
381 do { /* while (read_remaining) */
382 if (read_done >= target || (pipe && read_done))
383 break;
384
385 if (conn->killed)
386 break;
387
388 if (smc_rx_recvmsg_data_available(smc, peeked_bytes))
389 goto copy;
390
391 if (sk->sk_shutdown & RCV_SHUTDOWN) {
392 /* smc_cdc_msg_recv_action() could have run after
393 * above smc_rx_recvmsg_data_available()
394 */
395 if (smc_rx_recvmsg_data_available(smc, peeked_bytes))
396 goto copy;
397 break;
398 }
399
400 if (read_done) {
401 if (sk->sk_err ||
402 sk->sk_state == SMC_CLOSED ||
403 !timeo ||
404 signal_pending(current))
405 break;
406 } else {
407 if (sk->sk_err) {
408 read_done = sock_error(sk);
409 break;
410 }
411 if (sk->sk_state == SMC_CLOSED) {
412 if (!sock_flag(sk, SOCK_DONE)) {
413 /* This occurs when user tries to read
414 * from never connected socket.
415 */
416 read_done = -ENOTCONN;
417 break;
418 }
419 break;
420 }
421 if (!timeo)
422 return -EAGAIN;
423 if (signal_pending(current)) {
424 read_done = sock_intr_errno(timeo);
425 break;
426 }
427 }
428
429 if (!smc_rx_data_available(conn, peeked_bytes)) {
430 smc_rx_wait(smc, &timeo, peeked_bytes, smc_rx_data_available);
431 continue;
432 }
433
434 copy:
435 /* initialize variables for 1st iteration of subsequent loop */
436 /* could be just 1 byte, even after waiting on data above */
437 readable = smc_rx_data_available(conn, peeked_bytes);
438 splbytes = atomic_read(&conn->splice_pending);
439 if (!readable || (msg && splbytes)) {
440 if (splbytes)
441 func = smc_rx_data_available_and_no_splice_pend;
442 else
443 func = smc_rx_data_available;
444 smc_rx_wait(smc, &timeo, peeked_bytes, func);
445 continue;
446 }
447
448 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
449 if ((flags & MSG_PEEK) && peeked_bytes)
450 smc_curs_add(conn->rmb_desc->len, &cons, peeked_bytes);
451 /* subsequent splice() calls pick up where previous left */
452 if (splbytes)
453 smc_curs_add(conn->rmb_desc->len, &cons, splbytes);
454 if (conn->urg_state == SMC_URG_VALID &&
455 sock_flag(&smc->sk, SOCK_URGINLINE) &&
456 readable > 1)
457 readable--; /* always stop at urgent Byte */
458 /* not more than what user space asked for */
459 copylen = min_t(size_t, read_remaining, readable);
460 /* determine chunks where to read from rcvbuf */
461 /* either unwrapped case, or 1st chunk of wrapped case */
462 chunk_len = min_t(size_t, copylen, conn->rmb_desc->len -
463 cons.count);
464 chunk_len_sum = chunk_len;
465 chunk_off = cons.count;
466 smc_rmb_sync_sg_for_cpu(conn);
467 for (chunk = 0; chunk < 2; chunk++) {
468 if (!(flags & MSG_TRUNC)) {
469 if (msg) {
470 rc = memcpy_to_msg(msg, rcvbuf_base +
471 chunk_off,
472 chunk_len);
473 } else {
474 rc = smc_rx_splice(pipe, rcvbuf_base +
475 chunk_off, chunk_len,
476 smc);
477 }
478 if (rc < 0) {
479 if (!read_done)
480 read_done = -EFAULT;
481 goto out;
482 }
483 }
484 read_remaining -= chunk_len;
485 read_done += chunk_len;
486 if (flags & MSG_PEEK)
487 peeked_bytes += chunk_len;
488
489 if (chunk_len_sum == copylen)
490 break; /* either on 1st or 2nd iteration */
491 /* prepare next (== 2nd) iteration */
492 chunk_len = copylen - chunk_len; /* remainder */
493 chunk_len_sum += chunk_len;
494 chunk_off = 0; /* modulo offset in recv ring buffer */
495 }
496
497 /* update cursors */
498 if (!(flags & MSG_PEEK)) {
499 /* increased in recv tasklet smc_cdc_msg_rcv() */
500 smp_mb__before_atomic();
501 atomic_sub(copylen, &conn->bytes_to_rcv);
502 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
503 smp_mb__after_atomic();
504 if (msg && smc_rx_update_consumer(smc, cons, copylen))
505 goto out;
506 }
507
508 trace_smc_rx_recvmsg(smc, copylen);
509 } while (read_remaining);
510 out:
511 return read_done;
512 }
513
514 /* Initialize receive properties on connection establishment. NB: not __init! */
smc_rx_init(struct smc_sock * smc)515 void smc_rx_init(struct smc_sock *smc)
516 {
517 smc->sk.sk_data_ready = smc_rx_wake_up;
518 atomic_set(&smc->conn.splice_pending, 0);
519 smc->conn.urg_state = SMC_URG_READ;
520 }
521