xref: /openbmc/linux/net/smc/smc_tx.c (revision 6db6b729)
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  */
41 static 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  */
67 void 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  */
77 static 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 
127 static 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  */
145 static 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 
159 static 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  */
181 int 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 
292 out_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 */
303 int 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 */
317 static 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 */
340 static 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() */
356 static 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() */
421 static 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  */
463 static 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  */
547 static 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 
599 out_unlock:
600 	spin_unlock_bh(&conn->send_lock);
601 	smc_wr_tx_link_put(link);
602 	return rc;
603 }
604 
605 static 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 
624 static int __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 
654 out:
655 	return rc;
656 }
657 
658 int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
659 {
660 	int rc;
661 
662 	/* This make sure only one can send simultaneously to prevent wasting
663 	 * of CPU and CDC slot.
664 	 * Record whether someone has tried to push while we are pushing.
665 	 */
666 	if (atomic_inc_return(&conn->tx_pushing) > 1)
667 		return 0;
668 
669 again:
670 	atomic_set(&conn->tx_pushing, 1);
671 	smp_wmb(); /* Make sure tx_pushing is 1 before real send */
672 	rc = __smc_tx_sndbuf_nonempty(conn);
673 
674 	/* We need to check whether someone else have added some data into
675 	 * the send queue and tried to push but failed after the atomic_set()
676 	 * when we are pushing.
677 	 * If so, we need to push again to prevent those data hang in the send
678 	 * queue.
679 	 */
680 	if (unlikely(!atomic_dec_and_test(&conn->tx_pushing)))
681 		goto again;
682 
683 	return rc;
684 }
685 
686 /* Wakeup sndbuf consumers from process context
687  * since there is more data to transmit. The caller
688  * must hold sock lock.
689  */
690 void smc_tx_pending(struct smc_connection *conn)
691 {
692 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
693 	int rc;
694 
695 	if (smc->sk.sk_err)
696 		return;
697 
698 	rc = smc_tx_sndbuf_nonempty(conn);
699 	if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked &&
700 	    !atomic_read(&conn->bytes_to_rcv))
701 		conn->local_rx_ctrl.prod_flags.write_blocked = 0;
702 }
703 
704 /* Wakeup sndbuf consumers from process context
705  * since there is more data to transmit in locked
706  * sock.
707  */
708 void smc_tx_work(struct work_struct *work)
709 {
710 	struct smc_connection *conn = container_of(to_delayed_work(work),
711 						   struct smc_connection,
712 						   tx_work);
713 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
714 
715 	lock_sock(&smc->sk);
716 	smc_tx_pending(conn);
717 	release_sock(&smc->sk);
718 }
719 
720 void smc_tx_consumer_update(struct smc_connection *conn, bool force)
721 {
722 	union smc_host_cursor cfed, cons, prod;
723 	int sender_free = conn->rmb_desc->len;
724 	int to_confirm;
725 
726 	smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
727 	smc_curs_copy(&cfed, &conn->rx_curs_confirmed, conn);
728 	to_confirm = smc_curs_diff(conn->rmb_desc->len, &cfed, &cons);
729 	if (to_confirm > conn->rmbe_update_limit) {
730 		smc_curs_copy(&prod, &conn->local_rx_ctrl.prod, conn);
731 		sender_free = conn->rmb_desc->len -
732 			      smc_curs_diff_large(conn->rmb_desc->len,
733 						  &cfed, &prod);
734 	}
735 
736 	if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
737 	    force ||
738 	    ((to_confirm > conn->rmbe_update_limit) &&
739 	     ((sender_free <= (conn->rmb_desc->len / 2)) ||
740 	      conn->local_rx_ctrl.prod_flags.write_blocked))) {
741 		if (conn->killed ||
742 		    conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
743 			return;
744 		if ((smc_cdc_get_slot_and_msg_send(conn) < 0) &&
745 		    !conn->killed) {
746 			queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
747 					   SMC_TX_WORK_DELAY);
748 			return;
749 		}
750 	}
751 	if (conn->local_rx_ctrl.prod_flags.write_blocked &&
752 	    !atomic_read(&conn->bytes_to_rcv))
753 		conn->local_rx_ctrl.prod_flags.write_blocked = 0;
754 }
755 
756 /***************************** send initialize *******************************/
757 
758 /* Initialize send properties on connection establishment. NB: not __init! */
759 void smc_tx_init(struct smc_sock *smc)
760 {
761 	smc->sk.sk_write_space = smc_tx_write_space;
762 }
763