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>
27 #include "smc_close.h"
31 #define SMC_TX_WORK_DELAY 0
32 #define SMC_TX_CORK_DELAY (HZ >> 2) /* 250 ms */
34 /***************************** sndbuf producer *******************************/
36 /* callback implementation for sk.sk_write_space()
37 * to wakeup sndbuf producers that blocked with smc_tx_wait().
38 * called under sk_socket lock.
40 static void smc_tx_write_space(struct sock *sk)
42 struct socket *sock = sk->sk_socket;
43 struct smc_sock *smc = smc_sk(sk);
46 /* similar to sk_stream_write_space */
47 if (atomic_read(&smc->conn.sndbuf_space) && sock) {
48 clear_bit(SOCK_NOSPACE, &sock->flags);
50 wq = rcu_dereference(sk->sk_wq);
51 if (skwq_has_sleeper(wq))
52 wake_up_interruptible_poll(&wq->wait,
53 EPOLLOUT | EPOLLWRNORM |
55 if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN))
56 sock_wake_async(wq, SOCK_WAKE_SPACE, POLL_OUT);
61 /* Wakeup sndbuf producers that blocked with smc_tx_wait().
62 * Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space().
64 void smc_tx_sndbuf_nonfull(struct smc_sock *smc)
66 if (smc->sk.sk_socket &&
67 test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags))
68 smc->sk.sk_write_space(&smc->sk);
71 /* blocks sndbuf producer until at least one byte of free space available
72 * or urgent Byte was consumed
74 static int smc_tx_wait(struct smc_sock *smc, int flags)
76 DEFINE_WAIT_FUNC(wait, woken_wake_function);
77 struct smc_connection *conn = &smc->conn;
78 struct sock *sk = &smc->sk;
82 /* similar to sk_stream_wait_memory */
83 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
84 add_wait_queue(sk_sleep(sk), &wait);
86 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
88 (sk->sk_shutdown & SEND_SHUTDOWN) ||
90 conn->local_tx_ctrl.conn_state_flags.peer_done_writing) {
94 if (smc_cdc_rxed_any_close(conn)) {
99 /* ensure EPOLLOUT is subsequently generated */
100 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
104 if (signal_pending(current)) {
105 rc = sock_intr_errno(timeo);
108 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
109 if (atomic_read(&conn->sndbuf_space) && !conn->urg_tx_pend)
110 break; /* at least 1 byte of free & no urgent data */
111 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
112 sk_wait_event(sk, &timeo,
114 (sk->sk_shutdown & SEND_SHUTDOWN) ||
115 smc_cdc_rxed_any_close(conn) ||
116 (atomic_read(&conn->sndbuf_space) &&
120 remove_wait_queue(sk_sleep(sk), &wait);
124 static bool smc_tx_is_corked(struct smc_sock *smc)
126 struct tcp_sock *tp = tcp_sk(smc->clcsock->sk);
128 return (tp->nonagle & TCP_NAGLE_CORK) ? true : false;
131 /* sndbuf producer: main API called by socket layer.
132 * called under sock lock.
134 int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len)
136 size_t copylen, send_done = 0, send_remaining = len;
137 size_t chunk_len, chunk_off, chunk_len_sum;
138 struct smc_connection *conn = &smc->conn;
139 union smc_host_cursor prep;
140 struct sock *sk = &smc->sk;
146 /* This should be in poll */
147 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
149 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
154 while (msg_data_left(msg)) {
155 if (sk->sk_state == SMC_INIT)
157 if (smc->sk.sk_shutdown & SEND_SHUTDOWN ||
158 (smc->sk.sk_err == ECONNABORTED) ||
161 if (smc_cdc_rxed_any_close(conn))
162 return send_done ?: -ECONNRESET;
164 if (msg->msg_flags & MSG_OOB)
165 conn->local_tx_ctrl.prod_flags.urg_data_pending = 1;
167 if (!atomic_read(&conn->sndbuf_space) || conn->urg_tx_pend) {
170 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_copy(&prep, &conn->tx_curs_prep, conn);
184 tx_cnt_prep = prep.count;
185 /* determine chunks where to write into sndbuf */
186 /* either unwrapped case, or 1st chunk of wrapped case */
187 chunk_len = min_t(size_t, copylen, conn->sndbuf_desc->len -
189 chunk_len_sum = chunk_len;
190 chunk_off = tx_cnt_prep;
191 smc_sndbuf_sync_sg_for_cpu(conn);
192 for (chunk = 0; chunk < 2; chunk++) {
193 rc = memcpy_from_msg(sndbuf_base + chunk_off,
196 smc_sndbuf_sync_sg_for_device(conn);
201 send_done += chunk_len;
202 send_remaining -= chunk_len;
204 if (chunk_len_sum == copylen)
205 break; /* either on 1st or 2nd iteration */
206 /* prepare next (== 2nd) iteration */
207 chunk_len = copylen - chunk_len; /* remainder */
208 chunk_len_sum += chunk_len;
209 chunk_off = 0; /* modulo offset in send ring buffer */
211 smc_sndbuf_sync_sg_for_device(conn);
213 smc_curs_add(conn->sndbuf_desc->len, &prep, copylen);
214 smc_curs_copy(&conn->tx_curs_prep, &prep, conn);
215 /* increased in send tasklet smc_cdc_tx_handler() */
216 smp_mb__before_atomic();
217 atomic_sub(copylen, &conn->sndbuf_space);
218 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
219 smp_mb__after_atomic();
220 /* since we just produced more new data into sndbuf,
221 * trigger sndbuf consumer: RDMA write into peer RMBE and CDC
223 if ((msg->msg_flags & MSG_OOB) && !send_remaining)
224 conn->urg_tx_pend = true;
225 if ((msg->msg_flags & MSG_MORE || smc_tx_is_corked(smc)) &&
226 (atomic_read(&conn->sndbuf_space) >
227 (conn->sndbuf_desc->len >> 1)))
228 /* for a corked socket defer the RDMA writes if there
229 * is still sufficient sndbuf_space available
231 queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
234 smc_tx_sndbuf_nonempty(conn);
235 } /* while (msg_data_left(msg)) */
240 rc = sk_stream_error(sk, msg->msg_flags, rc);
241 /* make sure we wake any epoll edge trigger waiter */
242 if (unlikely(rc == -EAGAIN))
243 sk->sk_write_space(sk);
247 /***************************** sndbuf consumer *******************************/
249 /* sndbuf consumer: actual data transfer of one target chunk with ISM write */
250 int smcd_tx_ism_write(struct smc_connection *conn, void *data, size_t len,
251 u32 offset, int signal)
253 struct smc_ism_position pos;
256 memset(&pos, 0, sizeof(pos));
257 pos.token = conn->peer_token;
258 pos.index = conn->peer_rmbe_idx;
259 pos.offset = conn->tx_off + offset;
261 rc = smc_ism_write(conn->lgr->smcd, &pos, data, len);
263 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
267 /* sndbuf consumer: actual data transfer of one target chunk with RDMA write */
268 static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset,
269 int num_sges, struct ib_rdma_wr *rdma_wr)
271 struct smc_link_group *lgr = conn->lgr;
272 struct smc_link *link = conn->lnk;
275 rdma_wr->wr.wr_id = smc_wr_tx_get_next_wr_id(link);
276 rdma_wr->wr.num_sge = num_sges;
277 rdma_wr->remote_addr =
278 lgr->rtokens[conn->rtoken_idx][link->link_idx].dma_addr +
279 /* RMBE within RMB */
281 /* offset within RMBE */
283 rdma_wr->rkey = lgr->rtokens[conn->rtoken_idx][link->link_idx].rkey;
284 rc = ib_post_send(link->roce_qp, &rdma_wr->wr, NULL);
286 smcr_link_down_cond_sched(link);
290 /* sndbuf consumer */
291 static inline void smc_tx_advance_cursors(struct smc_connection *conn,
292 union smc_host_cursor *prod,
293 union smc_host_cursor *sent,
296 smc_curs_add(conn->peer_rmbe_size, prod, len);
297 /* increased in recv tasklet smc_cdc_msg_rcv() */
298 smp_mb__before_atomic();
299 /* data in flight reduces usable snd_wnd */
300 atomic_sub(len, &conn->peer_rmbe_space);
301 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
302 smp_mb__after_atomic();
303 smc_curs_add(conn->sndbuf_desc->len, sent, len);
306 /* SMC-R helper for smc_tx_rdma_writes() */
307 static int smcr_tx_rdma_writes(struct smc_connection *conn, size_t len,
308 size_t src_off, size_t src_len,
309 size_t dst_off, size_t dst_len,
310 struct smc_rdma_wr *wr_rdma_buf)
312 struct smc_link *link = conn->lnk;
314 dma_addr_t dma_addr =
315 sg_dma_address(conn->sndbuf_desc->sgt[link->link_idx].sgl);
316 int src_len_sum = src_len, dst_len_sum = dst_len;
317 int sent_count = src_off;
318 int srcchunk, dstchunk;
322 for (dstchunk = 0; dstchunk < 2; dstchunk++) {
324 wr_rdma_buf->wr_tx_rdma[dstchunk].wr.sg_list;
327 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
328 sge[srcchunk].addr = dma_addr + src_off;
329 sge[srcchunk].length = src_len;
333 if (src_off >= conn->sndbuf_desc->len)
334 src_off -= conn->sndbuf_desc->len;
335 /* modulo in send ring */
336 if (src_len_sum == dst_len)
337 break; /* either on 1st or 2nd iteration */
338 /* prepare next (== 2nd) iteration */
339 src_len = dst_len - src_len; /* remainder */
340 src_len_sum += src_len;
342 rc = smc_tx_rdma_write(conn, dst_off, num_sges,
343 &wr_rdma_buf->wr_tx_rdma[dstchunk]);
346 if (dst_len_sum == len)
347 break; /* either on 1st or 2nd iteration */
348 /* prepare next (== 2nd) iteration */
349 dst_off = 0; /* modulo offset in RMBE ring buffer */
350 dst_len = len - dst_len; /* remainder */
351 dst_len_sum += dst_len;
352 src_len = min_t(int, dst_len, conn->sndbuf_desc->len -
354 src_len_sum = src_len;
359 /* SMC-D helper for smc_tx_rdma_writes() */
360 static int smcd_tx_rdma_writes(struct smc_connection *conn, size_t len,
361 size_t src_off, size_t src_len,
362 size_t dst_off, size_t dst_len)
364 int src_len_sum = src_len, dst_len_sum = dst_len;
365 int srcchunk, dstchunk;
368 for (dstchunk = 0; dstchunk < 2; dstchunk++) {
369 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
370 void *data = conn->sndbuf_desc->cpu_addr + src_off;
372 rc = smcd_tx_ism_write(conn, data, src_len, dst_off +
373 sizeof(struct smcd_cdc_msg), 0);
378 if (src_off >= conn->sndbuf_desc->len)
379 src_off -= conn->sndbuf_desc->len;
380 /* modulo in send ring */
381 if (src_len_sum == dst_len)
382 break; /* either on 1st or 2nd iteration */
383 /* prepare next (== 2nd) iteration */
384 src_len = dst_len - src_len; /* remainder */
385 src_len_sum += src_len;
387 if (dst_len_sum == len)
388 break; /* either on 1st or 2nd iteration */
389 /* prepare next (== 2nd) iteration */
390 dst_off = 0; /* modulo offset in RMBE ring buffer */
391 dst_len = len - dst_len; /* remainder */
392 dst_len_sum += dst_len;
393 src_len = min_t(int, dst_len, conn->sndbuf_desc->len - src_off);
394 src_len_sum = src_len;
399 /* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit;
400 * usable snd_wnd as max transmit
402 static int smc_tx_rdma_writes(struct smc_connection *conn,
403 struct smc_rdma_wr *wr_rdma_buf)
405 size_t len, src_len, dst_off, dst_len; /* current chunk values */
406 union smc_host_cursor sent, prep, prod, cons;
407 struct smc_cdc_producer_flags *pflags;
408 int to_send, rmbespace;
412 smc_curs_copy(&sent, &conn->tx_curs_sent, conn);
413 smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
414 /* cf. wmem_alloc - (snd_max - snd_una) */
415 to_send = smc_curs_diff(conn->sndbuf_desc->len, &sent, &prep);
419 /* destination: RMBE */
421 rmbespace = atomic_read(&conn->peer_rmbe_space);
424 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
425 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
427 /* if usable snd_wnd closes ask peer to advertise once it opens again */
428 pflags = &conn->local_tx_ctrl.prod_flags;
429 pflags->write_blocked = (to_send >= rmbespace);
430 /* cf. usable snd_wnd */
431 len = min(to_send, rmbespace);
433 /* initialize variables for first iteration of subsequent nested loop */
434 dst_off = prod.count;
435 if (prod.wrap == cons.wrap) {
436 /* the filled destination area is unwrapped,
437 * hence the available free destination space is wrapped
438 * and we need 2 destination chunks of sum len; start with 1st
439 * which is limited by what's available in sndbuf
441 dst_len = min_t(size_t,
442 conn->peer_rmbe_size - prod.count, len);
444 /* the filled destination area is wrapped,
445 * hence the available free destination space is unwrapped
446 * and we need a single destination chunk of entire len
450 /* dst_len determines the maximum src_len */
451 if (sent.count + dst_len <= conn->sndbuf_desc->len) {
452 /* unwrapped src case: single chunk of entire dst_len */
455 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */
456 src_len = conn->sndbuf_desc->len - sent.count;
459 if (conn->lgr->is_smcd)
460 rc = smcd_tx_rdma_writes(conn, len, sent.count, src_len,
463 rc = smcr_tx_rdma_writes(conn, len, sent.count, src_len,
464 dst_off, dst_len, wr_rdma_buf);
468 if (conn->urg_tx_pend && len == to_send)
469 pflags->urg_data_present = 1;
470 smc_tx_advance_cursors(conn, &prod, &sent, len);
471 /* update connection's cursors with advanced local cursors */
472 smc_curs_copy(&conn->local_tx_ctrl.prod, &prod, conn);
474 smc_curs_copy(&conn->tx_curs_sent, &sent, conn);/* src: local sndbuf */
479 /* Wakeup sndbuf consumers from any context (IRQ or process)
480 * since there is more data to transmit; usable snd_wnd as max transmit
482 static int smcr_tx_sndbuf_nonempty(struct smc_connection *conn)
484 struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
485 struct smc_link *link = conn->lnk;
486 struct smc_rdma_wr *wr_rdma_buf;
487 struct smc_cdc_tx_pend *pend;
488 struct smc_wr_buf *wr_buf;
491 rc = smc_cdc_get_free_slot(conn, link, &wr_buf, &wr_rdma_buf, &pend);
494 struct smc_sock *smc =
495 container_of(conn, struct smc_sock, conn);
497 if (smc->sk.sk_err == ECONNABORTED)
498 return sock_error(&smc->sk);
502 mod_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
508 spin_lock_bh(&conn->send_lock);
509 if (link != conn->lnk) {
510 /* link of connection changed, tx_work will restart */
511 smc_wr_tx_put_slot(link,
512 (struct smc_wr_tx_pend_priv *)pend);
516 if (!pflags->urg_data_present) {
517 rc = smc_tx_rdma_writes(conn, wr_rdma_buf);
519 smc_wr_tx_put_slot(link,
520 (struct smc_wr_tx_pend_priv *)pend);
525 rc = smc_cdc_msg_send(conn, wr_buf, pend);
526 if (!rc && pflags->urg_data_present) {
527 pflags->urg_data_pending = 0;
528 pflags->urg_data_present = 0;
532 spin_unlock_bh(&conn->send_lock);
536 static int smcd_tx_sndbuf_nonempty(struct smc_connection *conn)
538 struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
541 spin_lock_bh(&conn->send_lock);
542 if (!pflags->urg_data_present)
543 rc = smc_tx_rdma_writes(conn, NULL);
545 rc = smcd_cdc_msg_send(conn);
547 if (!rc && pflags->urg_data_present) {
548 pflags->urg_data_pending = 0;
549 pflags->urg_data_present = 0;
551 spin_unlock_bh(&conn->send_lock);
555 int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
560 conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
561 return -EPIPE; /* connection being aborted */
562 if (conn->lgr->is_smcd)
563 rc = smcd_tx_sndbuf_nonempty(conn);
565 rc = smcr_tx_sndbuf_nonempty(conn);
568 /* trigger socket release if connection is closing */
569 struct smc_sock *smc = container_of(conn, struct smc_sock,
571 smc_close_wake_tx_prepared(smc);
576 /* Wakeup sndbuf consumers from process context
577 * since there is more data to transmit
579 void smc_tx_work(struct work_struct *work)
581 struct smc_connection *conn = container_of(to_delayed_work(work),
582 struct smc_connection,
584 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
591 rc = smc_tx_sndbuf_nonempty(conn);
592 if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked &&
593 !atomic_read(&conn->bytes_to_rcv))
594 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
597 release_sock(&smc->sk);
600 void smc_tx_consumer_update(struct smc_connection *conn, bool force)
602 union smc_host_cursor cfed, cons, prod;
603 int sender_free = conn->rmb_desc->len;
606 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
607 smc_curs_copy(&cfed, &conn->rx_curs_confirmed, conn);
608 to_confirm = smc_curs_diff(conn->rmb_desc->len, &cfed, &cons);
609 if (to_confirm > conn->rmbe_update_limit) {
610 smc_curs_copy(&prod, &conn->local_rx_ctrl.prod, conn);
611 sender_free = conn->rmb_desc->len -
612 smc_curs_diff_large(conn->rmb_desc->len,
616 if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
618 ((to_confirm > conn->rmbe_update_limit) &&
619 ((sender_free <= (conn->rmb_desc->len / 2)) ||
620 conn->local_rx_ctrl.prod_flags.write_blocked))) {
622 conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
624 if ((smc_cdc_get_slot_and_msg_send(conn) < 0) &&
626 queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
631 if (conn->local_rx_ctrl.prod_flags.write_blocked &&
632 !atomic_read(&conn->bytes_to_rcv))
633 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
636 /***************************** send initialize *******************************/
638 /* Initialize send properties on connection establishment. NB: not __init! */
639 void smc_tx_init(struct smc_sock *smc)
641 smc->sk.sk_write_space = smc_tx_write_space;