Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
4 *
5 * Manage send buffer.
6 * Producer:
7 * Copy user space data into send buffer, if send buffer space available.
8 * Consumer:
9 * Trigger RDMA write into RMBE of peer and send CDC, if RMBE space available.
10 *
11 * Copyright IBM Corp. 2016
12 *
13 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
14 */
15
16#include <linux/net.h>
17#include <linux/rcupdate.h>
18#include <linux/workqueue.h>
19#include <linux/sched/signal.h>
20
21#include <net/sock.h>
22#include <net/tcp.h>
23
24#include "smc.h"
25#include "smc_wr.h"
26#include "smc_cdc.h"
27#include "smc_close.h"
28#include "smc_ism.h"
29#include "smc_tx.h"
30#include "smc_stats.h"
31#include "smc_tracepoint.h"
32
33#define SMC_TX_WORK_DELAY 0
34
35/***************************** sndbuf producer *******************************/
36
37/* callback implementation for sk.sk_write_space()
38 * to wakeup sndbuf producers that blocked with smc_tx_wait().
39 * called under sk_socket lock.
40 */
41static void smc_tx_write_space(struct sock *sk)
42{
43 struct socket *sock = sk->sk_socket;
44 struct smc_sock *smc = smc_sk(sk);
45 struct socket_wq *wq;
46
47 /* similar to sk_stream_write_space */
48 if (atomic_read(&smc->conn.sndbuf_space) && sock) {
49 if (test_bit(SOCK_NOSPACE, &sock->flags))
50 SMC_STAT_RMB_TX_FULL(smc, !smc->conn.lnk);
51 clear_bit(SOCK_NOSPACE, &sock->flags);
52 rcu_read_lock();
53 wq = rcu_dereference(sk->sk_wq);
54 if (skwq_has_sleeper(wq))
55 wake_up_interruptible_poll(&wq->wait,
56 EPOLLOUT | EPOLLWRNORM |
57 EPOLLWRBAND);
58 if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN))
59 sock_wake_async(wq, SOCK_WAKE_SPACE, POLL_OUT);
60 rcu_read_unlock();
61 }
62}
63
64/* Wakeup sndbuf producers that blocked with smc_tx_wait().
65 * Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space().
66 */
67void smc_tx_sndbuf_nonfull(struct smc_sock *smc)
68{
69 if (smc->sk.sk_socket &&
70 test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags))
71 smc->sk.sk_write_space(&smc->sk);
72}
73
74/* blocks sndbuf producer until at least one byte of free space available
75 * or urgent Byte was consumed
76 */
77static int smc_tx_wait(struct smc_sock *smc, int flags)
78{
79 DEFINE_WAIT_FUNC(wait, woken_wake_function);
80 struct smc_connection *conn = &smc->conn;
81 struct sock *sk = &smc->sk;
82 long timeo;
83 int rc = 0;
84
85 /* similar to sk_stream_wait_memory */
86 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
87 add_wait_queue(sk_sleep(sk), &wait);
88 while (1) {
89 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
90 if (sk->sk_err ||
91 (sk->sk_shutdown & SEND_SHUTDOWN) ||
92 conn->killed ||
93 conn->local_tx_ctrl.conn_state_flags.peer_done_writing) {
94 rc = -EPIPE;
95 break;
96 }
97 if (smc_cdc_rxed_any_close(conn)) {
98 rc = -ECONNRESET;
99 break;
100 }
101 if (!timeo) {
102 /* ensure EPOLLOUT is subsequently generated */
103 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
104 rc = -EAGAIN;
105 break;
106 }
107 if (signal_pending(current)) {
108 rc = sock_intr_errno(timeo);
109 break;
110 }
111 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
112 if (atomic_read(&conn->sndbuf_space) && !conn->urg_tx_pend)
113 break; /* at least 1 byte of free & no urgent data */
114 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
115 sk_wait_event(sk, &timeo,
116 READ_ONCE(sk->sk_err) ||
117 (READ_ONCE(sk->sk_shutdown) & SEND_SHUTDOWN) ||
118 smc_cdc_rxed_any_close(conn) ||
119 (atomic_read(&conn->sndbuf_space) &&
120 !conn->urg_tx_pend),
121 &wait);
122 }
123 remove_wait_queue(sk_sleep(sk), &wait);
124 return rc;
125}
126
127static bool smc_tx_is_corked(struct smc_sock *smc)
128{
129 struct tcp_sock *tp = tcp_sk(smc->clcsock->sk);
130
131 return (tp->nonagle & TCP_NAGLE_CORK) ? true : false;
132}
133
134/* If we have pending CDC messages, do not send:
135 * Because CQE of this CDC message will happen shortly, it gives
136 * a chance to coalesce future sendmsg() payload in to one RDMA Write,
137 * without need for a timer, and with no latency trade off.
138 * Algorithm here:
139 * 1. First message should never cork
140 * 2. If we have pending Tx CDC messages, wait for the first CDC
141 * message's completion
142 * 3. Don't cork to much data in a single RDMA Write to prevent burst
143 * traffic, total corked message should not exceed sendbuf/2
144 */
145static bool smc_should_autocork(struct smc_sock *smc)
146{
147 struct smc_connection *conn = &smc->conn;
148 int corking_size;
149
150 corking_size = min_t(unsigned int, conn->sndbuf_desc->len >> 1,
151 sock_net(&smc->sk)->smc.sysctl_autocorking_size);
152
153 if (atomic_read(&conn->cdc_pend_tx_wr) == 0 ||
154 smc_tx_prepared_sends(conn) > corking_size)
155 return false;
156 return true;
157}
158
159static bool smc_tx_should_cork(struct smc_sock *smc, struct msghdr *msg)
160{
161 struct smc_connection *conn = &smc->conn;
162
163 if (smc_should_autocork(smc))
164 return true;
165
166 /* for a corked socket defer the RDMA writes if
167 * sndbuf_space is still available. The applications
168 * should known how/when to uncork it.
169 */
170 if ((msg->msg_flags & MSG_MORE ||
171 smc_tx_is_corked(smc)) &&
172 atomic_read(&conn->sndbuf_space))
173 return true;
174
175 return false;
176}
177
178/* sndbuf producer: main API called by socket layer.
179 * called under sock lock.
180 */
181int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len)
182{
183 size_t copylen, send_done = 0, send_remaining = len;
184 size_t chunk_len, chunk_off, chunk_len_sum;
185 struct smc_connection *conn = &smc->conn;
186 union smc_host_cursor prep;
187 struct sock *sk = &smc->sk;
188 char *sndbuf_base;
189 int tx_cnt_prep;
190 int writespace;
191 int rc, chunk;
192
193 /* This should be in poll */
194 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
195
196 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
197 rc = -EPIPE;
198 goto out_err;
199 }
200
201 if (sk->sk_state == SMC_INIT)
202 return -ENOTCONN;
203
204 if (len > conn->sndbuf_desc->len)
205 SMC_STAT_RMB_TX_SIZE_SMALL(smc, !conn->lnk);
206
207 if (len > conn->peer_rmbe_size)
208 SMC_STAT_RMB_TX_PEER_SIZE_SMALL(smc, !conn->lnk);
209
210 if (msg->msg_flags & MSG_OOB)
211 SMC_STAT_INC(smc, urg_data_cnt);
212
213 while (msg_data_left(msg)) {
214 if (smc->sk.sk_shutdown & SEND_SHUTDOWN ||
215 (smc->sk.sk_err == ECONNABORTED) ||
216 conn->killed)
217 return -EPIPE;
218 if (smc_cdc_rxed_any_close(conn))
219 return send_done ?: -ECONNRESET;
220
221 if (msg->msg_flags & MSG_OOB)
222 conn->local_tx_ctrl.prod_flags.urg_data_pending = 1;
223
224 if (!atomic_read(&conn->sndbuf_space) || conn->urg_tx_pend) {
225 if (send_done)
226 return send_done;
227 rc = smc_tx_wait(smc, msg->msg_flags);
228 if (rc)
229 goto out_err;
230 continue;
231 }
232
233 /* initialize variables for 1st iteration of subsequent loop */
234 /* could be just 1 byte, even after smc_tx_wait above */
235 writespace = atomic_read(&conn->sndbuf_space);
236 /* not more than what user space asked for */
237 copylen = min_t(size_t, send_remaining, writespace);
238 /* determine start of sndbuf */
239 sndbuf_base = conn->sndbuf_desc->cpu_addr;
240 smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
241 tx_cnt_prep = prep.count;
242 /* determine chunks where to write into sndbuf */
243 /* either unwrapped case, or 1st chunk of wrapped case */
244 chunk_len = min_t(size_t, copylen, conn->sndbuf_desc->len -
245 tx_cnt_prep);
246 chunk_len_sum = chunk_len;
247 chunk_off = tx_cnt_prep;
248 for (chunk = 0; chunk < 2; chunk++) {
249 rc = memcpy_from_msg(sndbuf_base + chunk_off,
250 msg, chunk_len);
251 if (rc) {
252 smc_sndbuf_sync_sg_for_device(conn);
253 if (send_done)
254 return send_done;
255 goto out_err;
256 }
257 send_done += chunk_len;
258 send_remaining -= chunk_len;
259
260 if (chunk_len_sum == copylen)
261 break; /* either on 1st or 2nd iteration */
262 /* prepare next (== 2nd) iteration */
263 chunk_len = copylen - chunk_len; /* remainder */
264 chunk_len_sum += chunk_len;
265 chunk_off = 0; /* modulo offset in send ring buffer */
266 }
267 smc_sndbuf_sync_sg_for_device(conn);
268 /* update cursors */
269 smc_curs_add(conn->sndbuf_desc->len, &prep, copylen);
270 smc_curs_copy(&conn->tx_curs_prep, &prep, conn);
271 /* increased in send tasklet smc_cdc_tx_handler() */
272 smp_mb__before_atomic();
273 atomic_sub(copylen, &conn->sndbuf_space);
274 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
275 smp_mb__after_atomic();
276 /* since we just produced more new data into sndbuf,
277 * trigger sndbuf consumer: RDMA write into peer RMBE and CDC
278 */
279 if ((msg->msg_flags & MSG_OOB) && !send_remaining)
280 conn->urg_tx_pend = true;
281 /* If we need to cork, do nothing and wait for the next
282 * sendmsg() call or push on tx completion
283 */
284 if (!smc_tx_should_cork(smc, msg))
285 smc_tx_sndbuf_nonempty(conn);
286
287 trace_smc_tx_sendmsg(smc, copylen);
288 } /* while (msg_data_left(msg)) */
289
290 return send_done;
291
292out_err:
293 rc = sk_stream_error(sk, msg->msg_flags, rc);
294 /* make sure we wake any epoll edge trigger waiter */
295 if (unlikely(rc == -EAGAIN))
296 sk->sk_write_space(sk);
297 return rc;
298}
299
300/***************************** sndbuf consumer *******************************/
301
302/* sndbuf consumer: actual data transfer of one target chunk with ISM write */
303int smcd_tx_ism_write(struct smc_connection *conn, void *data, size_t len,
304 u32 offset, int signal)
305{
306 int rc;
307
308 rc = smc_ism_write(conn->lgr->smcd, conn->peer_token,
309 conn->peer_rmbe_idx, signal, conn->tx_off + offset,
310 data, len);
311 if (rc)
312 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
313 return rc;
314}
315
316/* sndbuf consumer: actual data transfer of one target chunk with RDMA write */
317static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset,
318 int num_sges, struct ib_rdma_wr *rdma_wr)
319{
320 struct smc_link_group *lgr = conn->lgr;
321 struct smc_link *link = conn->lnk;
322 int rc;
323
324 rdma_wr->wr.wr_id = smc_wr_tx_get_next_wr_id(link);
325 rdma_wr->wr.num_sge = num_sges;
326 rdma_wr->remote_addr =
327 lgr->rtokens[conn->rtoken_idx][link->link_idx].dma_addr +
328 /* RMBE within RMB */
329 conn->tx_off +
330 /* offset within RMBE */
331 peer_rmbe_offset;
332 rdma_wr->rkey = lgr->rtokens[conn->rtoken_idx][link->link_idx].rkey;
333 rc = ib_post_send(link->roce_qp, &rdma_wr->wr, NULL);
334 if (rc)
335 smcr_link_down_cond_sched(link);
336 return rc;
337}
338
339/* sndbuf consumer */
340static inline void smc_tx_advance_cursors(struct smc_connection *conn,
341 union smc_host_cursor *prod,
342 union smc_host_cursor *sent,
343 size_t len)
344{
345 smc_curs_add(conn->peer_rmbe_size, prod, len);
346 /* increased in recv tasklet smc_cdc_msg_rcv() */
347 smp_mb__before_atomic();
348 /* data in flight reduces usable snd_wnd */
349 atomic_sub(len, &conn->peer_rmbe_space);
350 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
351 smp_mb__after_atomic();
352 smc_curs_add(conn->sndbuf_desc->len, sent, len);
353}
354
355/* SMC-R helper for smc_tx_rdma_writes() */
356static int smcr_tx_rdma_writes(struct smc_connection *conn, size_t len,
357 size_t src_off, size_t src_len,
358 size_t dst_off, size_t dst_len,
359 struct smc_rdma_wr *wr_rdma_buf)
360{
361 struct smc_link *link = conn->lnk;
362
363 dma_addr_t dma_addr =
364 sg_dma_address(conn->sndbuf_desc->sgt[link->link_idx].sgl);
365 u64 virt_addr = (uintptr_t)conn->sndbuf_desc->cpu_addr;
366 int src_len_sum = src_len, dst_len_sum = dst_len;
367 int sent_count = src_off;
368 int srcchunk, dstchunk;
369 int num_sges;
370 int rc;
371
372 for (dstchunk = 0; dstchunk < 2; dstchunk++) {
373 struct ib_rdma_wr *wr = &wr_rdma_buf->wr_tx_rdma[dstchunk];
374 struct ib_sge *sge = wr->wr.sg_list;
375 u64 base_addr = dma_addr;
376
377 if (dst_len < link->qp_attr.cap.max_inline_data) {
378 base_addr = virt_addr;
379 wr->wr.send_flags |= IB_SEND_INLINE;
380 } else {
381 wr->wr.send_flags &= ~IB_SEND_INLINE;
382 }
383
384 num_sges = 0;
385 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
386 sge[srcchunk].addr = conn->sndbuf_desc->is_vm ?
387 (virt_addr + src_off) : (base_addr + src_off);
388 sge[srcchunk].length = src_len;
389 if (conn->sndbuf_desc->is_vm)
390 sge[srcchunk].lkey =
391 conn->sndbuf_desc->mr[link->link_idx]->lkey;
392 num_sges++;
393
394 src_off += src_len;
395 if (src_off >= conn->sndbuf_desc->len)
396 src_off -= conn->sndbuf_desc->len;
397 /* modulo in send ring */
398 if (src_len_sum == dst_len)
399 break; /* either on 1st or 2nd iteration */
400 /* prepare next (== 2nd) iteration */
401 src_len = dst_len - src_len; /* remainder */
402 src_len_sum += src_len;
403 }
404 rc = smc_tx_rdma_write(conn, dst_off, num_sges, wr);
405 if (rc)
406 return rc;
407 if (dst_len_sum == len)
408 break; /* either on 1st or 2nd iteration */
409 /* prepare next (== 2nd) iteration */
410 dst_off = 0; /* modulo offset in RMBE ring buffer */
411 dst_len = len - dst_len; /* remainder */
412 dst_len_sum += dst_len;
413 src_len = min_t(int, dst_len, conn->sndbuf_desc->len -
414 sent_count);
415 src_len_sum = src_len;
416 }
417 return 0;
418}
419
420/* SMC-D helper for smc_tx_rdma_writes() */
421static int smcd_tx_rdma_writes(struct smc_connection *conn, size_t len,
422 size_t src_off, size_t src_len,
423 size_t dst_off, size_t dst_len)
424{
425 int src_len_sum = src_len, dst_len_sum = dst_len;
426 int srcchunk, dstchunk;
427 int rc;
428
429 for (dstchunk = 0; dstchunk < 2; dstchunk++) {
430 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
431 void *data = conn->sndbuf_desc->cpu_addr + src_off;
432
433 rc = smcd_tx_ism_write(conn, data, src_len, dst_off +
434 sizeof(struct smcd_cdc_msg), 0);
435 if (rc)
436 return rc;
437 dst_off += src_len;
438 src_off += src_len;
439 if (src_off >= conn->sndbuf_desc->len)
440 src_off -= conn->sndbuf_desc->len;
441 /* modulo in send ring */
442 if (src_len_sum == dst_len)
443 break; /* either on 1st or 2nd iteration */
444 /* prepare next (== 2nd) iteration */
445 src_len = dst_len - src_len; /* remainder */
446 src_len_sum += src_len;
447 }
448 if (dst_len_sum == len)
449 break; /* either on 1st or 2nd iteration */
450 /* prepare next (== 2nd) iteration */
451 dst_off = 0; /* modulo offset in RMBE ring buffer */
452 dst_len = len - dst_len; /* remainder */
453 dst_len_sum += dst_len;
454 src_len = min_t(int, dst_len, conn->sndbuf_desc->len - src_off);
455 src_len_sum = src_len;
456 }
457 return 0;
458}
459
460/* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit;
461 * usable snd_wnd as max transmit
462 */
463static int smc_tx_rdma_writes(struct smc_connection *conn,
464 struct smc_rdma_wr *wr_rdma_buf)
465{
466 size_t len, src_len, dst_off, dst_len; /* current chunk values */
467 union smc_host_cursor sent, prep, prod, cons;
468 struct smc_cdc_producer_flags *pflags;
469 int to_send, rmbespace;
470 int rc;
471
472 /* source: sndbuf */
473 smc_curs_copy(&sent, &conn->tx_curs_sent, conn);
474 smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
475 /* cf. wmem_alloc - (snd_max - snd_una) */
476 to_send = smc_curs_diff(conn->sndbuf_desc->len, &sent, &prep);
477 if (to_send <= 0)
478 return 0;
479
480 /* destination: RMBE */
481 /* cf. snd_wnd */
482 rmbespace = atomic_read(&conn->peer_rmbe_space);
483 if (rmbespace <= 0) {
484 struct smc_sock *smc = container_of(conn, struct smc_sock,
485 conn);
486 SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);
487 return 0;
488 }
489 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
490 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
491
492 /* if usable snd_wnd closes ask peer to advertise once it opens again */
493 pflags = &conn->local_tx_ctrl.prod_flags;
494 pflags->write_blocked = (to_send >= rmbespace);
495 /* cf. usable snd_wnd */
496 len = min(to_send, rmbespace);
497
498 /* initialize variables for first iteration of subsequent nested loop */
499 dst_off = prod.count;
500 if (prod.wrap == cons.wrap) {
501 /* the filled destination area is unwrapped,
502 * hence the available free destination space is wrapped
503 * and we need 2 destination chunks of sum len; start with 1st
504 * which is limited by what's available in sndbuf
505 */
506 dst_len = min_t(size_t,
507 conn->peer_rmbe_size - prod.count, len);
508 } else {
509 /* the filled destination area is wrapped,
510 * hence the available free destination space is unwrapped
511 * and we need a single destination chunk of entire len
512 */
513 dst_len = len;
514 }
515 /* dst_len determines the maximum src_len */
516 if (sent.count + dst_len <= conn->sndbuf_desc->len) {
517 /* unwrapped src case: single chunk of entire dst_len */
518 src_len = dst_len;
519 } else {
520 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */
521 src_len = conn->sndbuf_desc->len - sent.count;
522 }
523
524 if (conn->lgr->is_smcd)
525 rc = smcd_tx_rdma_writes(conn, len, sent.count, src_len,
526 dst_off, dst_len);
527 else
528 rc = smcr_tx_rdma_writes(conn, len, sent.count, src_len,
529 dst_off, dst_len, wr_rdma_buf);
530 if (rc)
531 return rc;
532
533 if (conn->urg_tx_pend && len == to_send)
534 pflags->urg_data_present = 1;
535 smc_tx_advance_cursors(conn, &prod, &sent, len);
536 /* update connection's cursors with advanced local cursors */
537 smc_curs_copy(&conn->local_tx_ctrl.prod, &prod, conn);
538 /* dst: peer RMBE */
539 smc_curs_copy(&conn->tx_curs_sent, &sent, conn);/* src: local sndbuf */
540
541 return 0;
542}
543
544/* Wakeup sndbuf consumers from any context (IRQ or process)
545 * since there is more data to transmit; usable snd_wnd as max transmit
546 */
547static int smcr_tx_sndbuf_nonempty(struct smc_connection *conn)
548{
549 struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
550 struct smc_link *link = conn->lnk;
551 struct smc_rdma_wr *wr_rdma_buf;
552 struct smc_cdc_tx_pend *pend;
553 struct smc_wr_buf *wr_buf;
554 int rc;
555
556 if (!link || !smc_wr_tx_link_hold(link))
557 return -ENOLINK;
558 rc = smc_cdc_get_free_slot(conn, link, &wr_buf, &wr_rdma_buf, &pend);
559 if (rc < 0) {
560 smc_wr_tx_link_put(link);
561 if (rc == -EBUSY) {
562 struct smc_sock *smc =
563 container_of(conn, struct smc_sock, conn);
564
565 if (smc->sk.sk_err == ECONNABORTED)
566 return sock_error(&smc->sk);
567 if (conn->killed)
568 return -EPIPE;
569 rc = 0;
570 mod_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
571 SMC_TX_WORK_DELAY);
572 }
573 return rc;
574 }
575
576 spin_lock_bh(&conn->send_lock);
577 if (link != conn->lnk) {
578 /* link of connection changed, tx_work will restart */
579 smc_wr_tx_put_slot(link,
580 (struct smc_wr_tx_pend_priv *)pend);
581 rc = -ENOLINK;
582 goto out_unlock;
583 }
584 if (!pflags->urg_data_present) {
585 rc = smc_tx_rdma_writes(conn, wr_rdma_buf);
586 if (rc) {
587 smc_wr_tx_put_slot(link,
588 (struct smc_wr_tx_pend_priv *)pend);
589 goto out_unlock;
590 }
591 }
592
593 rc = smc_cdc_msg_send(conn, wr_buf, pend);
594 if (!rc && pflags->urg_data_present) {
595 pflags->urg_data_pending = 0;
596 pflags->urg_data_present = 0;
597 }
598
599out_unlock:
600 spin_unlock_bh(&conn->send_lock);
601 smc_wr_tx_link_put(link);
602 return rc;
603}
604
605static int smcd_tx_sndbuf_nonempty(struct smc_connection *conn)
606{
607 struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
608 int rc = 0;
609
610 spin_lock_bh(&conn->send_lock);
611 if (!pflags->urg_data_present)
612 rc = smc_tx_rdma_writes(conn, NULL);
613 if (!rc)
614 rc = smcd_cdc_msg_send(conn);
615
616 if (!rc && pflags->urg_data_present) {
617 pflags->urg_data_pending = 0;
618 pflags->urg_data_present = 0;
619 }
620 spin_unlock_bh(&conn->send_lock);
621 return rc;
622}
623
624int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
625{
626 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
627 int rc = 0;
628
629 /* No data in the send queue */
630 if (unlikely(smc_tx_prepared_sends(conn) <= 0))
631 goto out;
632
633 /* Peer don't have RMBE space */
634 if (unlikely(atomic_read(&conn->peer_rmbe_space) <= 0)) {
635 SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);
636 goto out;
637 }
638
639 if (conn->killed ||
640 conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
641 rc = -EPIPE; /* connection being aborted */
642 goto out;
643 }
644 if (conn->lgr->is_smcd)
645 rc = smcd_tx_sndbuf_nonempty(conn);
646 else
647 rc = smcr_tx_sndbuf_nonempty(conn);
648
649 if (!rc) {
650 /* trigger socket release if connection is closing */
651 smc_close_wake_tx_prepared(smc);
652 }
653
654out:
655 return rc;
656}
657
658/* Wakeup sndbuf consumers from process context
659 * since there is more data to transmit. The caller
660 * must hold sock lock.
661 */
662void smc_tx_pending(struct smc_connection *conn)
663{
664 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
665 int rc;
666
667 if (smc->sk.sk_err)
668 return;
669
670 rc = smc_tx_sndbuf_nonempty(conn);
671 if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked &&
672 !atomic_read(&conn->bytes_to_rcv))
673 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
674}
675
676/* Wakeup sndbuf consumers from process context
677 * since there is more data to transmit in locked
678 * sock.
679 */
680void smc_tx_work(struct work_struct *work)
681{
682 struct smc_connection *conn = container_of(to_delayed_work(work),
683 struct smc_connection,
684 tx_work);
685 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
686
687 lock_sock(&smc->sk);
688 smc_tx_pending(conn);
689 release_sock(&smc->sk);
690}
691
692void smc_tx_consumer_update(struct smc_connection *conn, bool force)
693{
694 union smc_host_cursor cfed, cons, prod;
695 int sender_free = conn->rmb_desc->len;
696 int to_confirm;
697
698 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
699 smc_curs_copy(&cfed, &conn->rx_curs_confirmed, conn);
700 to_confirm = smc_curs_diff(conn->rmb_desc->len, &cfed, &cons);
701 if (to_confirm > conn->rmbe_update_limit) {
702 smc_curs_copy(&prod, &conn->local_rx_ctrl.prod, conn);
703 sender_free = conn->rmb_desc->len -
704 smc_curs_diff_large(conn->rmb_desc->len,
705 &cfed, &prod);
706 }
707
708 if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
709 force ||
710 ((to_confirm > conn->rmbe_update_limit) &&
711 ((sender_free <= (conn->rmb_desc->len / 2)) ||
712 conn->local_rx_ctrl.prod_flags.write_blocked))) {
713 if (conn->killed ||
714 conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
715 return;
716 if ((smc_cdc_get_slot_and_msg_send(conn) < 0) &&
717 !conn->killed) {
718 queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
719 SMC_TX_WORK_DELAY);
720 return;
721 }
722 }
723 if (conn->local_rx_ctrl.prod_flags.write_blocked &&
724 !atomic_read(&conn->bytes_to_rcv))
725 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
726}
727
728/***************************** send initialize *******************************/
729
730/* Initialize send properties on connection establishment. NB: not __init! */
731void smc_tx_init(struct smc_sock *smc)
732{
733 smc->sk.sk_write_space = smc_tx_write_space;
734}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
4 *
5 * Manage send buffer.
6 * Producer:
7 * Copy user space data into send buffer, if send buffer space available.
8 * Consumer:
9 * Trigger RDMA write into RMBE of peer and send CDC, if RMBE space available.
10 *
11 * Copyright IBM Corp. 2016
12 *
13 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
14 */
15
16#include <linux/net.h>
17#include <linux/rcupdate.h>
18#include <linux/workqueue.h>
19#include <linux/sched/signal.h>
20
21#include <net/sock.h>
22#include <net/tcp.h>
23
24#include "smc.h"
25#include "smc_wr.h"
26#include "smc_cdc.h"
27#include "smc_close.h"
28#include "smc_ism.h"
29#include "smc_tx.h"
30#include "smc_stats.h"
31#include "smc_tracepoint.h"
32
33#define SMC_TX_WORK_DELAY 0
34
35/***************************** sndbuf producer *******************************/
36
37/* callback implementation for sk.sk_write_space()
38 * to wakeup sndbuf producers that blocked with smc_tx_wait().
39 * called under sk_socket lock.
40 */
41static void smc_tx_write_space(struct sock *sk)
42{
43 struct socket *sock = sk->sk_socket;
44 struct smc_sock *smc = smc_sk(sk);
45 struct socket_wq *wq;
46
47 /* similar to sk_stream_write_space */
48 if (atomic_read(&smc->conn.sndbuf_space) && sock) {
49 if (test_bit(SOCK_NOSPACE, &sock->flags))
50 SMC_STAT_RMB_TX_FULL(smc, !smc->conn.lnk);
51 clear_bit(SOCK_NOSPACE, &sock->flags);
52 rcu_read_lock();
53 wq = rcu_dereference(sk->sk_wq);
54 if (skwq_has_sleeper(wq))
55 wake_up_interruptible_poll(&wq->wait,
56 EPOLLOUT | EPOLLWRNORM |
57 EPOLLWRBAND);
58 if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN))
59 sock_wake_async(wq, SOCK_WAKE_SPACE, POLL_OUT);
60 rcu_read_unlock();
61 }
62}
63
64/* Wakeup sndbuf producers that blocked with smc_tx_wait().
65 * Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space().
66 */
67void smc_tx_sndbuf_nonfull(struct smc_sock *smc)
68{
69 if (smc->sk.sk_socket &&
70 test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags))
71 smc->sk.sk_write_space(&smc->sk);
72}
73
74/* blocks sndbuf producer until at least one byte of free space available
75 * or urgent Byte was consumed
76 */
77static int smc_tx_wait(struct smc_sock *smc, int flags)
78{
79 DEFINE_WAIT_FUNC(wait, woken_wake_function);
80 struct smc_connection *conn = &smc->conn;
81 struct sock *sk = &smc->sk;
82 long timeo;
83 int rc = 0;
84
85 /* similar to sk_stream_wait_memory */
86 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
87 add_wait_queue(sk_sleep(sk), &wait);
88 while (1) {
89 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
90 if (sk->sk_err ||
91 (sk->sk_shutdown & SEND_SHUTDOWN) ||
92 conn->killed ||
93 conn->local_tx_ctrl.conn_state_flags.peer_done_writing) {
94 rc = -EPIPE;
95 break;
96 }
97 if (smc_cdc_rxed_any_close(conn)) {
98 rc = -ECONNRESET;
99 break;
100 }
101 if (!timeo) {
102 /* ensure EPOLLOUT is subsequently generated */
103 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
104 rc = -EAGAIN;
105 break;
106 }
107 if (signal_pending(current)) {
108 rc = sock_intr_errno(timeo);
109 break;
110 }
111 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
112 if (atomic_read(&conn->sndbuf_space) && !conn->urg_tx_pend)
113 break; /* at least 1 byte of free & no urgent data */
114 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
115 sk_wait_event(sk, &timeo,
116 sk->sk_err ||
117 (sk->sk_shutdown & SEND_SHUTDOWN) ||
118 smc_cdc_rxed_any_close(conn) ||
119 (atomic_read(&conn->sndbuf_space) &&
120 !conn->urg_tx_pend),
121 &wait);
122 }
123 remove_wait_queue(sk_sleep(sk), &wait);
124 return rc;
125}
126
127static bool smc_tx_is_corked(struct smc_sock *smc)
128{
129 struct tcp_sock *tp = tcp_sk(smc->clcsock->sk);
130
131 return (tp->nonagle & TCP_NAGLE_CORK) ? true : false;
132}
133
134/* If we have pending CDC messages, do not send:
135 * Because CQE of this CDC message will happen shortly, it gives
136 * a chance to coalesce future sendmsg() payload in to one RDMA Write,
137 * without need for a timer, and with no latency trade off.
138 * Algorithm here:
139 * 1. First message should never cork
140 * 2. If we have pending Tx CDC messages, wait for the first CDC
141 * message's completion
142 * 3. Don't cork to much data in a single RDMA Write to prevent burst
143 * traffic, total corked message should not exceed sendbuf/2
144 */
145static bool smc_should_autocork(struct smc_sock *smc)
146{
147 struct smc_connection *conn = &smc->conn;
148 int corking_size;
149
150 corking_size = min_t(unsigned int, conn->sndbuf_desc->len >> 1,
151 sock_net(&smc->sk)->smc.sysctl_autocorking_size);
152
153 if (atomic_read(&conn->cdc_pend_tx_wr) == 0 ||
154 smc_tx_prepared_sends(conn) > corking_size)
155 return false;
156 return true;
157}
158
159static bool smc_tx_should_cork(struct smc_sock *smc, struct msghdr *msg)
160{
161 struct smc_connection *conn = &smc->conn;
162
163 if (smc_should_autocork(smc))
164 return true;
165
166 /* for a corked socket defer the RDMA writes if
167 * sndbuf_space is still available. The applications
168 * should known how/when to uncork it.
169 */
170 if ((msg->msg_flags & MSG_MORE ||
171 smc_tx_is_corked(smc) ||
172 msg->msg_flags & MSG_SENDPAGE_NOTLAST) &&
173 atomic_read(&conn->sndbuf_space))
174 return true;
175
176 return false;
177}
178
179/* sndbuf producer: main API called by socket layer.
180 * called under sock lock.
181 */
182int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len)
183{
184 size_t copylen, send_done = 0, send_remaining = len;
185 size_t chunk_len, chunk_off, chunk_len_sum;
186 struct smc_connection *conn = &smc->conn;
187 union smc_host_cursor prep;
188 struct sock *sk = &smc->sk;
189 char *sndbuf_base;
190 int tx_cnt_prep;
191 int writespace;
192 int rc, chunk;
193
194 /* This should be in poll */
195 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
196
197 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
198 rc = -EPIPE;
199 goto out_err;
200 }
201
202 if (sk->sk_state == SMC_INIT)
203 return -ENOTCONN;
204
205 if (len > conn->sndbuf_desc->len)
206 SMC_STAT_RMB_TX_SIZE_SMALL(smc, !conn->lnk);
207
208 if (len > conn->peer_rmbe_size)
209 SMC_STAT_RMB_TX_PEER_SIZE_SMALL(smc, !conn->lnk);
210
211 if (msg->msg_flags & MSG_OOB)
212 SMC_STAT_INC(smc, urg_data_cnt);
213
214 while (msg_data_left(msg)) {
215 if (smc->sk.sk_shutdown & SEND_SHUTDOWN ||
216 (smc->sk.sk_err == ECONNABORTED) ||
217 conn->killed)
218 return -EPIPE;
219 if (smc_cdc_rxed_any_close(conn))
220 return send_done ?: -ECONNRESET;
221
222 if (msg->msg_flags & MSG_OOB)
223 conn->local_tx_ctrl.prod_flags.urg_data_pending = 1;
224
225 if (!atomic_read(&conn->sndbuf_space) || conn->urg_tx_pend) {
226 if (send_done)
227 return send_done;
228 rc = smc_tx_wait(smc, msg->msg_flags);
229 if (rc)
230 goto out_err;
231 continue;
232 }
233
234 /* initialize variables for 1st iteration of subsequent loop */
235 /* could be just 1 byte, even after smc_tx_wait above */
236 writespace = atomic_read(&conn->sndbuf_space);
237 /* not more than what user space asked for */
238 copylen = min_t(size_t, send_remaining, writespace);
239 /* determine start of sndbuf */
240 sndbuf_base = conn->sndbuf_desc->cpu_addr;
241 smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
242 tx_cnt_prep = prep.count;
243 /* determine chunks where to write into sndbuf */
244 /* either unwrapped case, or 1st chunk of wrapped case */
245 chunk_len = min_t(size_t, copylen, conn->sndbuf_desc->len -
246 tx_cnt_prep);
247 chunk_len_sum = chunk_len;
248 chunk_off = tx_cnt_prep;
249 for (chunk = 0; chunk < 2; chunk++) {
250 rc = memcpy_from_msg(sndbuf_base + chunk_off,
251 msg, chunk_len);
252 if (rc) {
253 smc_sndbuf_sync_sg_for_device(conn);
254 if (send_done)
255 return send_done;
256 goto out_err;
257 }
258 send_done += chunk_len;
259 send_remaining -= chunk_len;
260
261 if (chunk_len_sum == copylen)
262 break; /* either on 1st or 2nd iteration */
263 /* prepare next (== 2nd) iteration */
264 chunk_len = copylen - chunk_len; /* remainder */
265 chunk_len_sum += chunk_len;
266 chunk_off = 0; /* modulo offset in send ring buffer */
267 }
268 smc_sndbuf_sync_sg_for_device(conn);
269 /* update cursors */
270 smc_curs_add(conn->sndbuf_desc->len, &prep, copylen);
271 smc_curs_copy(&conn->tx_curs_prep, &prep, conn);
272 /* increased in send tasklet smc_cdc_tx_handler() */
273 smp_mb__before_atomic();
274 atomic_sub(copylen, &conn->sndbuf_space);
275 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
276 smp_mb__after_atomic();
277 /* since we just produced more new data into sndbuf,
278 * trigger sndbuf consumer: RDMA write into peer RMBE and CDC
279 */
280 if ((msg->msg_flags & MSG_OOB) && !send_remaining)
281 conn->urg_tx_pend = true;
282 /* If we need to cork, do nothing and wait for the next
283 * sendmsg() call or push on tx completion
284 */
285 if (!smc_tx_should_cork(smc, msg))
286 smc_tx_sndbuf_nonempty(conn);
287
288 trace_smc_tx_sendmsg(smc, copylen);
289 } /* while (msg_data_left(msg)) */
290
291 return send_done;
292
293out_err:
294 rc = sk_stream_error(sk, msg->msg_flags, rc);
295 /* make sure we wake any epoll edge trigger waiter */
296 if (unlikely(rc == -EAGAIN))
297 sk->sk_write_space(sk);
298 return rc;
299}
300
301int smc_tx_sendpage(struct smc_sock *smc, struct page *page, int offset,
302 size_t size, int flags)
303{
304 struct msghdr msg = {.msg_flags = flags};
305 char *kaddr = kmap(page);
306 struct kvec iov;
307 int rc;
308
309 iov.iov_base = kaddr + offset;
310 iov.iov_len = size;
311 iov_iter_kvec(&msg.msg_iter, ITER_SOURCE, &iov, 1, size);
312 rc = smc_tx_sendmsg(smc, &msg, size);
313 kunmap(page);
314 return rc;
315}
316
317/***************************** sndbuf consumer *******************************/
318
319/* sndbuf consumer: actual data transfer of one target chunk with ISM write */
320int smcd_tx_ism_write(struct smc_connection *conn, void *data, size_t len,
321 u32 offset, int signal)
322{
323 int rc;
324
325 rc = smc_ism_write(conn->lgr->smcd, conn->peer_token,
326 conn->peer_rmbe_idx, signal, conn->tx_off + offset,
327 data, len);
328 if (rc)
329 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
330 return rc;
331}
332
333/* sndbuf consumer: actual data transfer of one target chunk with RDMA write */
334static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset,
335 int num_sges, struct ib_rdma_wr *rdma_wr)
336{
337 struct smc_link_group *lgr = conn->lgr;
338 struct smc_link *link = conn->lnk;
339 int rc;
340
341 rdma_wr->wr.wr_id = smc_wr_tx_get_next_wr_id(link);
342 rdma_wr->wr.num_sge = num_sges;
343 rdma_wr->remote_addr =
344 lgr->rtokens[conn->rtoken_idx][link->link_idx].dma_addr +
345 /* RMBE within RMB */
346 conn->tx_off +
347 /* offset within RMBE */
348 peer_rmbe_offset;
349 rdma_wr->rkey = lgr->rtokens[conn->rtoken_idx][link->link_idx].rkey;
350 rc = ib_post_send(link->roce_qp, &rdma_wr->wr, NULL);
351 if (rc)
352 smcr_link_down_cond_sched(link);
353 return rc;
354}
355
356/* sndbuf consumer */
357static inline void smc_tx_advance_cursors(struct smc_connection *conn,
358 union smc_host_cursor *prod,
359 union smc_host_cursor *sent,
360 size_t len)
361{
362 smc_curs_add(conn->peer_rmbe_size, prod, len);
363 /* increased in recv tasklet smc_cdc_msg_rcv() */
364 smp_mb__before_atomic();
365 /* data in flight reduces usable snd_wnd */
366 atomic_sub(len, &conn->peer_rmbe_space);
367 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
368 smp_mb__after_atomic();
369 smc_curs_add(conn->sndbuf_desc->len, sent, len);
370}
371
372/* SMC-R helper for smc_tx_rdma_writes() */
373static int smcr_tx_rdma_writes(struct smc_connection *conn, size_t len,
374 size_t src_off, size_t src_len,
375 size_t dst_off, size_t dst_len,
376 struct smc_rdma_wr *wr_rdma_buf)
377{
378 struct smc_link *link = conn->lnk;
379
380 dma_addr_t dma_addr =
381 sg_dma_address(conn->sndbuf_desc->sgt[link->link_idx].sgl);
382 u64 virt_addr = (uintptr_t)conn->sndbuf_desc->cpu_addr;
383 int src_len_sum = src_len, dst_len_sum = dst_len;
384 int sent_count = src_off;
385 int srcchunk, dstchunk;
386 int num_sges;
387 int rc;
388
389 for (dstchunk = 0; dstchunk < 2; dstchunk++) {
390 struct ib_rdma_wr *wr = &wr_rdma_buf->wr_tx_rdma[dstchunk];
391 struct ib_sge *sge = wr->wr.sg_list;
392 u64 base_addr = dma_addr;
393
394 if (dst_len < link->qp_attr.cap.max_inline_data) {
395 base_addr = virt_addr;
396 wr->wr.send_flags |= IB_SEND_INLINE;
397 } else {
398 wr->wr.send_flags &= ~IB_SEND_INLINE;
399 }
400
401 num_sges = 0;
402 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
403 sge[srcchunk].addr = conn->sndbuf_desc->is_vm ?
404 (virt_addr + src_off) : (base_addr + src_off);
405 sge[srcchunk].length = src_len;
406 if (conn->sndbuf_desc->is_vm)
407 sge[srcchunk].lkey =
408 conn->sndbuf_desc->mr[link->link_idx]->lkey;
409 num_sges++;
410
411 src_off += src_len;
412 if (src_off >= conn->sndbuf_desc->len)
413 src_off -= conn->sndbuf_desc->len;
414 /* modulo in send ring */
415 if (src_len_sum == dst_len)
416 break; /* either on 1st or 2nd iteration */
417 /* prepare next (== 2nd) iteration */
418 src_len = dst_len - src_len; /* remainder */
419 src_len_sum += src_len;
420 }
421 rc = smc_tx_rdma_write(conn, dst_off, num_sges, wr);
422 if (rc)
423 return rc;
424 if (dst_len_sum == len)
425 break; /* either on 1st or 2nd iteration */
426 /* prepare next (== 2nd) iteration */
427 dst_off = 0; /* modulo offset in RMBE ring buffer */
428 dst_len = len - dst_len; /* remainder */
429 dst_len_sum += dst_len;
430 src_len = min_t(int, dst_len, conn->sndbuf_desc->len -
431 sent_count);
432 src_len_sum = src_len;
433 }
434 return 0;
435}
436
437/* SMC-D helper for smc_tx_rdma_writes() */
438static int smcd_tx_rdma_writes(struct smc_connection *conn, size_t len,
439 size_t src_off, size_t src_len,
440 size_t dst_off, size_t dst_len)
441{
442 int src_len_sum = src_len, dst_len_sum = dst_len;
443 int srcchunk, dstchunk;
444 int rc;
445
446 for (dstchunk = 0; dstchunk < 2; dstchunk++) {
447 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
448 void *data = conn->sndbuf_desc->cpu_addr + src_off;
449
450 rc = smcd_tx_ism_write(conn, data, src_len, dst_off +
451 sizeof(struct smcd_cdc_msg), 0);
452 if (rc)
453 return rc;
454 dst_off += src_len;
455 src_off += src_len;
456 if (src_off >= conn->sndbuf_desc->len)
457 src_off -= conn->sndbuf_desc->len;
458 /* modulo in send ring */
459 if (src_len_sum == dst_len)
460 break; /* either on 1st or 2nd iteration */
461 /* prepare next (== 2nd) iteration */
462 src_len = dst_len - src_len; /* remainder */
463 src_len_sum += src_len;
464 }
465 if (dst_len_sum == len)
466 break; /* either on 1st or 2nd iteration */
467 /* prepare next (== 2nd) iteration */
468 dst_off = 0; /* modulo offset in RMBE ring buffer */
469 dst_len = len - dst_len; /* remainder */
470 dst_len_sum += dst_len;
471 src_len = min_t(int, dst_len, conn->sndbuf_desc->len - src_off);
472 src_len_sum = src_len;
473 }
474 return 0;
475}
476
477/* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit;
478 * usable snd_wnd as max transmit
479 */
480static int smc_tx_rdma_writes(struct smc_connection *conn,
481 struct smc_rdma_wr *wr_rdma_buf)
482{
483 size_t len, src_len, dst_off, dst_len; /* current chunk values */
484 union smc_host_cursor sent, prep, prod, cons;
485 struct smc_cdc_producer_flags *pflags;
486 int to_send, rmbespace;
487 int rc;
488
489 /* source: sndbuf */
490 smc_curs_copy(&sent, &conn->tx_curs_sent, conn);
491 smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
492 /* cf. wmem_alloc - (snd_max - snd_una) */
493 to_send = smc_curs_diff(conn->sndbuf_desc->len, &sent, &prep);
494 if (to_send <= 0)
495 return 0;
496
497 /* destination: RMBE */
498 /* cf. snd_wnd */
499 rmbespace = atomic_read(&conn->peer_rmbe_space);
500 if (rmbespace <= 0) {
501 struct smc_sock *smc = container_of(conn, struct smc_sock,
502 conn);
503 SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);
504 return 0;
505 }
506 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
507 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
508
509 /* if usable snd_wnd closes ask peer to advertise once it opens again */
510 pflags = &conn->local_tx_ctrl.prod_flags;
511 pflags->write_blocked = (to_send >= rmbespace);
512 /* cf. usable snd_wnd */
513 len = min(to_send, rmbespace);
514
515 /* initialize variables for first iteration of subsequent nested loop */
516 dst_off = prod.count;
517 if (prod.wrap == cons.wrap) {
518 /* the filled destination area is unwrapped,
519 * hence the available free destination space is wrapped
520 * and we need 2 destination chunks of sum len; start with 1st
521 * which is limited by what's available in sndbuf
522 */
523 dst_len = min_t(size_t,
524 conn->peer_rmbe_size - prod.count, len);
525 } else {
526 /* the filled destination area is wrapped,
527 * hence the available free destination space is unwrapped
528 * and we need a single destination chunk of entire len
529 */
530 dst_len = len;
531 }
532 /* dst_len determines the maximum src_len */
533 if (sent.count + dst_len <= conn->sndbuf_desc->len) {
534 /* unwrapped src case: single chunk of entire dst_len */
535 src_len = dst_len;
536 } else {
537 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */
538 src_len = conn->sndbuf_desc->len - sent.count;
539 }
540
541 if (conn->lgr->is_smcd)
542 rc = smcd_tx_rdma_writes(conn, len, sent.count, src_len,
543 dst_off, dst_len);
544 else
545 rc = smcr_tx_rdma_writes(conn, len, sent.count, src_len,
546 dst_off, dst_len, wr_rdma_buf);
547 if (rc)
548 return rc;
549
550 if (conn->urg_tx_pend && len == to_send)
551 pflags->urg_data_present = 1;
552 smc_tx_advance_cursors(conn, &prod, &sent, len);
553 /* update connection's cursors with advanced local cursors */
554 smc_curs_copy(&conn->local_tx_ctrl.prod, &prod, conn);
555 /* dst: peer RMBE */
556 smc_curs_copy(&conn->tx_curs_sent, &sent, conn);/* src: local sndbuf */
557
558 return 0;
559}
560
561/* Wakeup sndbuf consumers from any context (IRQ or process)
562 * since there is more data to transmit; usable snd_wnd as max transmit
563 */
564static int smcr_tx_sndbuf_nonempty(struct smc_connection *conn)
565{
566 struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
567 struct smc_link *link = conn->lnk;
568 struct smc_rdma_wr *wr_rdma_buf;
569 struct smc_cdc_tx_pend *pend;
570 struct smc_wr_buf *wr_buf;
571 int rc;
572
573 if (!link || !smc_wr_tx_link_hold(link))
574 return -ENOLINK;
575 rc = smc_cdc_get_free_slot(conn, link, &wr_buf, &wr_rdma_buf, &pend);
576 if (rc < 0) {
577 smc_wr_tx_link_put(link);
578 if (rc == -EBUSY) {
579 struct smc_sock *smc =
580 container_of(conn, struct smc_sock, conn);
581
582 if (smc->sk.sk_err == ECONNABORTED)
583 return sock_error(&smc->sk);
584 if (conn->killed)
585 return -EPIPE;
586 rc = 0;
587 mod_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
588 SMC_TX_WORK_DELAY);
589 }
590 return rc;
591 }
592
593 spin_lock_bh(&conn->send_lock);
594 if (link != conn->lnk) {
595 /* link of connection changed, tx_work will restart */
596 smc_wr_tx_put_slot(link,
597 (struct smc_wr_tx_pend_priv *)pend);
598 rc = -ENOLINK;
599 goto out_unlock;
600 }
601 if (!pflags->urg_data_present) {
602 rc = smc_tx_rdma_writes(conn, wr_rdma_buf);
603 if (rc) {
604 smc_wr_tx_put_slot(link,
605 (struct smc_wr_tx_pend_priv *)pend);
606 goto out_unlock;
607 }
608 }
609
610 rc = smc_cdc_msg_send(conn, wr_buf, pend);
611 if (!rc && pflags->urg_data_present) {
612 pflags->urg_data_pending = 0;
613 pflags->urg_data_present = 0;
614 }
615
616out_unlock:
617 spin_unlock_bh(&conn->send_lock);
618 smc_wr_tx_link_put(link);
619 return rc;
620}
621
622static int smcd_tx_sndbuf_nonempty(struct smc_connection *conn)
623{
624 struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
625 int rc = 0;
626
627 spin_lock_bh(&conn->send_lock);
628 if (!pflags->urg_data_present)
629 rc = smc_tx_rdma_writes(conn, NULL);
630 if (!rc)
631 rc = smcd_cdc_msg_send(conn);
632
633 if (!rc && pflags->urg_data_present) {
634 pflags->urg_data_pending = 0;
635 pflags->urg_data_present = 0;
636 }
637 spin_unlock_bh(&conn->send_lock);
638 return rc;
639}
640
641static int __smc_tx_sndbuf_nonempty(struct smc_connection *conn)
642{
643 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
644 int rc = 0;
645
646 /* No data in the send queue */
647 if (unlikely(smc_tx_prepared_sends(conn) <= 0))
648 goto out;
649
650 /* Peer don't have RMBE space */
651 if (unlikely(atomic_read(&conn->peer_rmbe_space) <= 0)) {
652 SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);
653 goto out;
654 }
655
656 if (conn->killed ||
657 conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
658 rc = -EPIPE; /* connection being aborted */
659 goto out;
660 }
661 if (conn->lgr->is_smcd)
662 rc = smcd_tx_sndbuf_nonempty(conn);
663 else
664 rc = smcr_tx_sndbuf_nonempty(conn);
665
666 if (!rc) {
667 /* trigger socket release if connection is closing */
668 smc_close_wake_tx_prepared(smc);
669 }
670
671out:
672 return rc;
673}
674
675int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
676{
677 int rc;
678
679 /* This make sure only one can send simultaneously to prevent wasting
680 * of CPU and CDC slot.
681 * Record whether someone has tried to push while we are pushing.
682 */
683 if (atomic_inc_return(&conn->tx_pushing) > 1)
684 return 0;
685
686again:
687 atomic_set(&conn->tx_pushing, 1);
688 smp_wmb(); /* Make sure tx_pushing is 1 before real send */
689 rc = __smc_tx_sndbuf_nonempty(conn);
690
691 /* We need to check whether someone else have added some data into
692 * the send queue and tried to push but failed after the atomic_set()
693 * when we are pushing.
694 * If so, we need to push again to prevent those data hang in the send
695 * queue.
696 */
697 if (unlikely(!atomic_dec_and_test(&conn->tx_pushing)))
698 goto again;
699
700 return rc;
701}
702
703/* Wakeup sndbuf consumers from process context
704 * since there is more data to transmit. The caller
705 * must hold sock lock.
706 */
707void smc_tx_pending(struct smc_connection *conn)
708{
709 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
710 int rc;
711
712 if (smc->sk.sk_err)
713 return;
714
715 rc = smc_tx_sndbuf_nonempty(conn);
716 if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked &&
717 !atomic_read(&conn->bytes_to_rcv))
718 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
719}
720
721/* Wakeup sndbuf consumers from process context
722 * since there is more data to transmit in locked
723 * sock.
724 */
725void smc_tx_work(struct work_struct *work)
726{
727 struct smc_connection *conn = container_of(to_delayed_work(work),
728 struct smc_connection,
729 tx_work);
730 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
731
732 lock_sock(&smc->sk);
733 smc_tx_pending(conn);
734 release_sock(&smc->sk);
735}
736
737void smc_tx_consumer_update(struct smc_connection *conn, bool force)
738{
739 union smc_host_cursor cfed, cons, prod;
740 int sender_free = conn->rmb_desc->len;
741 int to_confirm;
742
743 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
744 smc_curs_copy(&cfed, &conn->rx_curs_confirmed, conn);
745 to_confirm = smc_curs_diff(conn->rmb_desc->len, &cfed, &cons);
746 if (to_confirm > conn->rmbe_update_limit) {
747 smc_curs_copy(&prod, &conn->local_rx_ctrl.prod, conn);
748 sender_free = conn->rmb_desc->len -
749 smc_curs_diff_large(conn->rmb_desc->len,
750 &cfed, &prod);
751 }
752
753 if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
754 force ||
755 ((to_confirm > conn->rmbe_update_limit) &&
756 ((sender_free <= (conn->rmb_desc->len / 2)) ||
757 conn->local_rx_ctrl.prod_flags.write_blocked))) {
758 if (conn->killed ||
759 conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
760 return;
761 if ((smc_cdc_get_slot_and_msg_send(conn) < 0) &&
762 !conn->killed) {
763 queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
764 SMC_TX_WORK_DELAY);
765 return;
766 }
767 }
768 if (conn->local_rx_ctrl.prod_flags.write_blocked &&
769 !atomic_read(&conn->bytes_to_rcv))
770 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
771}
772
773/***************************** send initialize *******************************/
774
775/* Initialize send properties on connection establishment. NB: not __init! */
776void smc_tx_init(struct smc_sock *smc)
777{
778 smc->sk.sk_write_space = smc_tx_write_space;
779}