1 // SPDX-License-Identifier: GPL-2.0
3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
7 * Copy user space data into send buffer, if send buffer space available.
9 * Trigger RDMA write into RMBE of peer and send CDC, if RMBE space available.
11 * Copyright IBM Corp. 2016
13 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
16 #include <linux/net.h>
17 #include <linux/rcupdate.h>
18 #include <linux/workqueue.h>
19 #include <linux/sched/signal.h>
29 #define SMC_TX_WORK_DELAY HZ
30 #define SMC_TX_CORK_DELAY (HZ >> 2) /* 250 ms */
32 /***************************** sndbuf producer *******************************/
34 /* callback implementation for sk.sk_write_space()
35 * to wakeup sndbuf producers that blocked with smc_tx_wait().
36 * called under sk_socket lock.
38 static void smc_tx_write_space(struct sock *sk)
40 struct socket *sock = sk->sk_socket;
41 struct smc_sock *smc = smc_sk(sk);
44 /* similar to sk_stream_write_space */
45 if (atomic_read(&smc->conn.sndbuf_space) && sock) {
46 clear_bit(SOCK_NOSPACE, &sock->flags);
48 wq = rcu_dereference(sk->sk_wq);
49 if (skwq_has_sleeper(wq))
50 wake_up_interruptible_poll(&wq->wait,
51 EPOLLOUT | EPOLLWRNORM |
53 if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN))
54 sock_wake_async(wq, SOCK_WAKE_SPACE, POLL_OUT);
59 /* Wakeup sndbuf producers that blocked with smc_tx_wait().
60 * Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space().
62 void smc_tx_sndbuf_nonfull(struct smc_sock *smc)
64 if (smc->sk.sk_socket &&
65 test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags))
66 smc->sk.sk_write_space(&smc->sk);
69 /* blocks sndbuf producer until at least one byte of free space available
70 * or urgent Byte was consumed
72 static int smc_tx_wait(struct smc_sock *smc, int flags)
74 DEFINE_WAIT_FUNC(wait, woken_wake_function);
75 struct smc_connection *conn = &smc->conn;
76 struct sock *sk = &smc->sk;
81 /* similar to sk_stream_wait_memory */
82 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
83 noblock = timeo ? false : true;
84 add_wait_queue(sk_sleep(sk), &wait);
86 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
88 (sk->sk_shutdown & SEND_SHUTDOWN) ||
89 conn->local_tx_ctrl.conn_state_flags.peer_done_writing) {
93 if (smc_cdc_rxed_any_close(conn)) {
99 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
103 if (signal_pending(current)) {
104 rc = sock_intr_errno(timeo);
107 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
108 if (atomic_read(&conn->sndbuf_space) && !conn->urg_tx_pend)
109 break; /* at least 1 byte of free & no urgent data */
110 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
111 sk_wait_event(sk, &timeo,
113 (sk->sk_shutdown & SEND_SHUTDOWN) ||
114 smc_cdc_rxed_any_close(conn) ||
115 (atomic_read(&conn->sndbuf_space) &&
119 remove_wait_queue(sk_sleep(sk), &wait);
123 static bool smc_tx_is_corked(struct smc_sock *smc)
125 struct tcp_sock *tp = tcp_sk(smc->clcsock->sk);
127 return (tp->nonagle & TCP_NAGLE_CORK) ? true : false;
130 /* sndbuf producer: main API called by socket layer.
131 * called under sock lock.
133 int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len)
135 size_t copylen, send_done = 0, send_remaining = len;
136 size_t chunk_len, chunk_off, chunk_len_sum;
137 struct smc_connection *conn = &smc->conn;
138 union smc_host_cursor prep;
139 struct sock *sk = &smc->sk;
145 /* This should be in poll */
146 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
148 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
153 while (msg_data_left(msg)) {
154 if (sk->sk_state == SMC_INIT)
156 if (smc->sk.sk_shutdown & SEND_SHUTDOWN ||
157 (smc->sk.sk_err == ECONNABORTED) ||
158 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort)
160 if (smc_cdc_rxed_any_close(conn))
161 return send_done ?: -ECONNRESET;
163 if (msg->msg_flags & MSG_OOB)
164 conn->local_tx_ctrl.prod_flags.urg_data_pending = 1;
166 if (!atomic_read(&conn->sndbuf_space) || conn->urg_tx_pend) {
167 rc = smc_tx_wait(smc, msg->msg_flags);
176 /* initialize variables for 1st iteration of subsequent loop */
177 /* could be just 1 byte, even after smc_tx_wait above */
178 writespace = atomic_read(&conn->sndbuf_space);
179 /* not more than what user space asked for */
180 copylen = min_t(size_t, send_remaining, writespace);
181 /* determine start of sndbuf */
182 sndbuf_base = conn->sndbuf_desc->cpu_addr;
183 smc_curs_write(&prep,
184 smc_curs_read(&conn->tx_curs_prep, conn),
186 tx_cnt_prep = prep.count;
187 /* determine chunks where to write into sndbuf */
188 /* either unwrapped case, or 1st chunk of wrapped case */
189 chunk_len = min_t(size_t, copylen, conn->sndbuf_desc->len -
191 chunk_len_sum = chunk_len;
192 chunk_off = tx_cnt_prep;
193 smc_sndbuf_sync_sg_for_cpu(conn);
194 for (chunk = 0; chunk < 2; chunk++) {
195 rc = memcpy_from_msg(sndbuf_base + chunk_off,
198 smc_sndbuf_sync_sg_for_device(conn);
203 send_done += chunk_len;
204 send_remaining -= chunk_len;
206 if (chunk_len_sum == copylen)
207 break; /* either on 1st or 2nd iteration */
208 /* prepare next (== 2nd) iteration */
209 chunk_len = copylen - chunk_len; /* remainder */
210 chunk_len_sum += chunk_len;
211 chunk_off = 0; /* modulo offset in send ring buffer */
213 smc_sndbuf_sync_sg_for_device(conn);
215 smc_curs_add(conn->sndbuf_desc->len, &prep, copylen);
216 smc_curs_write(&conn->tx_curs_prep,
217 smc_curs_read(&prep, conn),
219 /* increased in send tasklet smc_cdc_tx_handler() */
220 smp_mb__before_atomic();
221 atomic_sub(copylen, &conn->sndbuf_space);
222 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
223 smp_mb__after_atomic();
224 /* since we just produced more new data into sndbuf,
225 * trigger sndbuf consumer: RDMA write into peer RMBE and CDC
227 if ((msg->msg_flags & MSG_OOB) && !send_remaining)
228 conn->urg_tx_pend = true;
229 if ((msg->msg_flags & MSG_MORE || smc_tx_is_corked(smc)) &&
230 (atomic_read(&conn->sndbuf_space) >
231 (conn->sndbuf_desc->len >> 1)))
232 /* for a corked socket defer the RDMA writes if there
233 * is still sufficient sndbuf_space available
235 schedule_delayed_work(&conn->tx_work,
238 smc_tx_sndbuf_nonempty(conn);
239 } /* while (msg_data_left(msg)) */
244 rc = sk_stream_error(sk, msg->msg_flags, rc);
245 /* make sure we wake any epoll edge trigger waiter */
246 if (unlikely(rc == -EAGAIN))
247 sk->sk_write_space(sk);
251 /***************************** sndbuf consumer *******************************/
253 /* sndbuf consumer: actual data transfer of one target chunk with RDMA write */
254 static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset,
255 int num_sges, struct ib_sge sges[])
257 struct smc_link_group *lgr = conn->lgr;
258 struct ib_rdma_wr rdma_wr;
259 struct smc_link *link;
262 memset(&rdma_wr, 0, sizeof(rdma_wr));
263 link = &lgr->lnk[SMC_SINGLE_LINK];
264 rdma_wr.wr.wr_id = smc_wr_tx_get_next_wr_id(link);
265 rdma_wr.wr.sg_list = sges;
266 rdma_wr.wr.num_sge = num_sges;
267 rdma_wr.wr.opcode = IB_WR_RDMA_WRITE;
268 rdma_wr.remote_addr =
269 lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr +
270 /* RMBE within RMB */
272 /* offset within RMBE */
274 rdma_wr.rkey = lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey;
275 rc = ib_post_send(link->roce_qp, &rdma_wr.wr, NULL);
277 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
278 smc_lgr_terminate(lgr);
283 /* sndbuf consumer */
284 static inline void smc_tx_advance_cursors(struct smc_connection *conn,
285 union smc_host_cursor *prod,
286 union smc_host_cursor *sent,
289 smc_curs_add(conn->peer_rmbe_size, prod, len);
290 /* increased in recv tasklet smc_cdc_msg_rcv() */
291 smp_mb__before_atomic();
292 /* data in flight reduces usable snd_wnd */
293 atomic_sub(len, &conn->peer_rmbe_space);
294 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
295 smp_mb__after_atomic();
296 smc_curs_add(conn->sndbuf_desc->len, sent, len);
299 /* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit;
300 * usable snd_wnd as max transmit
302 static int smc_tx_rdma_writes(struct smc_connection *conn)
304 size_t src_off, src_len, dst_off, dst_len; /* current chunk values */
305 size_t len, dst_len_sum, src_len_sum, dstchunk, srcchunk;
306 union smc_host_cursor sent, prep, prod, cons;
307 struct ib_sge sges[SMC_IB_MAX_SEND_SGE];
308 struct smc_link_group *lgr = conn->lgr;
309 struct smc_cdc_producer_flags *pflags;
310 int to_send, rmbespace;
311 struct smc_link *link;
317 smc_curs_write(&sent, smc_curs_read(&conn->tx_curs_sent, conn), conn);
318 smc_curs_write(&prep, smc_curs_read(&conn->tx_curs_prep, conn), conn);
319 /* cf. wmem_alloc - (snd_max - snd_una) */
320 to_send = smc_curs_diff(conn->sndbuf_desc->len, &sent, &prep);
324 /* destination: RMBE */
326 rmbespace = atomic_read(&conn->peer_rmbe_space);
329 smc_curs_write(&prod,
330 smc_curs_read(&conn->local_tx_ctrl.prod, conn),
332 smc_curs_write(&cons,
333 smc_curs_read(&conn->local_rx_ctrl.cons, conn),
336 /* if usable snd_wnd closes ask peer to advertise once it opens again */
337 pflags = &conn->local_tx_ctrl.prod_flags;
338 pflags->write_blocked = (to_send >= rmbespace);
339 /* cf. usable snd_wnd */
340 len = min(to_send, rmbespace);
342 /* initialize variables for first iteration of subsequent nested loop */
343 link = &lgr->lnk[SMC_SINGLE_LINK];
344 dst_off = prod.count;
345 if (prod.wrap == cons.wrap) {
346 /* the filled destination area is unwrapped,
347 * hence the available free destination space is wrapped
348 * and we need 2 destination chunks of sum len; start with 1st
349 * which is limited by what's available in sndbuf
351 dst_len = min_t(size_t,
352 conn->peer_rmbe_size - prod.count, len);
354 /* the filled destination area is wrapped,
355 * hence the available free destination space is unwrapped
356 * and we need a single destination chunk of entire len
360 dst_len_sum = dst_len;
361 src_off = sent.count;
362 /* dst_len determines the maximum src_len */
363 if (sent.count + dst_len <= conn->sndbuf_desc->len) {
364 /* unwrapped src case: single chunk of entire dst_len */
367 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */
368 src_len = conn->sndbuf_desc->len - sent.count;
370 src_len_sum = src_len;
371 dma_addr = sg_dma_address(conn->sndbuf_desc->sgt[SMC_SINGLE_LINK].sgl);
372 for (dstchunk = 0; dstchunk < 2; dstchunk++) {
374 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
375 sges[srcchunk].addr = dma_addr + src_off;
376 sges[srcchunk].length = src_len;
377 sges[srcchunk].lkey = link->roce_pd->local_dma_lkey;
380 if (src_off >= conn->sndbuf_desc->len)
381 src_off -= conn->sndbuf_desc->len;
382 /* modulo in send ring */
383 if (src_len_sum == dst_len)
384 break; /* either on 1st or 2nd iteration */
385 /* prepare next (== 2nd) iteration */
386 src_len = dst_len - src_len; /* remainder */
387 src_len_sum += src_len;
389 rc = smc_tx_rdma_write(conn, dst_off, num_sges, sges);
392 if (dst_len_sum == len)
393 break; /* either on 1st or 2nd iteration */
394 /* prepare next (== 2nd) iteration */
395 dst_off = 0; /* modulo offset in RMBE ring buffer */
396 dst_len = len - dst_len; /* remainder */
397 dst_len_sum += dst_len;
399 dst_len, conn->sndbuf_desc->len - sent.count);
400 src_len_sum = src_len;
403 if (conn->urg_tx_pend && len == to_send)
404 pflags->urg_data_present = 1;
405 smc_tx_advance_cursors(conn, &prod, &sent, len);
406 /* update connection's cursors with advanced local cursors */
407 smc_curs_write(&conn->local_tx_ctrl.prod,
408 smc_curs_read(&prod, conn),
411 smc_curs_write(&conn->tx_curs_sent,
412 smc_curs_read(&sent, conn),
414 /* src: local sndbuf */
419 /* Wakeup sndbuf consumers from any context (IRQ or process)
420 * since there is more data to transmit; usable snd_wnd as max transmit
422 int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
424 struct smc_cdc_producer_flags *pflags;
425 struct smc_cdc_tx_pend *pend;
426 struct smc_wr_buf *wr_buf;
429 spin_lock_bh(&conn->send_lock);
430 rc = smc_cdc_get_free_slot(conn, &wr_buf, &pend);
433 struct smc_sock *smc =
434 container_of(conn, struct smc_sock, conn);
436 if (smc->sk.sk_err == ECONNABORTED) {
437 rc = sock_error(&smc->sk);
441 if (conn->alert_token_local) /* connection healthy */
442 mod_delayed_work(system_wq, &conn->tx_work,
448 if (!conn->local_tx_ctrl.prod_flags.urg_data_present) {
449 rc = smc_tx_rdma_writes(conn);
451 smc_wr_tx_put_slot(&conn->lgr->lnk[SMC_SINGLE_LINK],
452 (struct smc_wr_tx_pend_priv *)pend);
457 rc = smc_cdc_msg_send(conn, wr_buf, pend);
458 pflags = &conn->local_tx_ctrl.prod_flags;
459 if (!rc && pflags->urg_data_present) {
460 pflags->urg_data_pending = 0;
461 pflags->urg_data_present = 0;
465 spin_unlock_bh(&conn->send_lock);
469 /* Wakeup sndbuf consumers from process context
470 * since there is more data to transmit
472 void smc_tx_work(struct work_struct *work)
474 struct smc_connection *conn = container_of(to_delayed_work(work),
475 struct smc_connection,
477 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
481 if (smc->sk.sk_err ||
482 !conn->alert_token_local ||
483 conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
486 rc = smc_tx_sndbuf_nonempty(conn);
487 if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked &&
488 !atomic_read(&conn->bytes_to_rcv))
489 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
492 release_sock(&smc->sk);
495 void smc_tx_consumer_update(struct smc_connection *conn, bool force)
497 union smc_host_cursor cfed, cons, prod;
498 int sender_free = conn->rmb_desc->len;
501 smc_curs_write(&cons,
502 smc_curs_read(&conn->local_tx_ctrl.cons, conn),
504 smc_curs_write(&cfed,
505 smc_curs_read(&conn->rx_curs_confirmed, conn),
507 to_confirm = smc_curs_diff(conn->rmb_desc->len, &cfed, &cons);
508 if (to_confirm > conn->rmbe_update_limit) {
509 smc_curs_write(&prod,
510 smc_curs_read(&conn->local_rx_ctrl.prod, conn),
512 sender_free = conn->rmb_desc->len -
513 smc_curs_diff(conn->rmb_desc->len, &prod, &cfed);
516 if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
518 ((to_confirm > conn->rmbe_update_limit) &&
519 ((sender_free <= (conn->rmb_desc->len / 2)) ||
520 conn->local_rx_ctrl.prod_flags.write_blocked))) {
521 if ((smc_cdc_get_slot_and_msg_send(conn) < 0) &&
522 conn->alert_token_local) { /* connection healthy */
523 schedule_delayed_work(&conn->tx_work,
527 smc_curs_write(&conn->rx_curs_confirmed,
528 smc_curs_read(&conn->local_tx_ctrl.cons, conn),
530 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
532 if (conn->local_rx_ctrl.prod_flags.write_blocked &&
533 !atomic_read(&conn->bytes_to_rcv))
534 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
537 /***************************** send initialize *******************************/
539 /* Initialize send properties on connection establishment. NB: not __init! */
540 void smc_tx_init(struct smc_sock *smc)
542 smc->sk.sk_write_space = smc_tx_write_space;