1 // SPDX-License-Identifier: GPL-2.0
3  * Shared Memory Communications over RDMA (SMC-R) and RoCE
43 	struct socket *sock = sk->sk_socket;  in smc_tx_write_space()
48 	if (atomic_read(&smc->conn.sndbuf_space) && sock) {  in smc_tx_write_space()
49 		if (test_bit(SOCK_NOSPACE, &sock->flags))  in smc_tx_write_space()
50 			SMC_STAT_RMB_TX_FULL(smc, !smc->conn.lnk);  in smc_tx_write_space()
51 		clear_bit(SOCK_NOSPACE, &sock->flags);  in smc_tx_write_space()
53 		wq = rcu_dereference(sk->sk_wq);  in smc_tx_write_space()
55 			wake_up_interruptible_poll(&wq->wait,  in smc_tx_write_space()
58 		if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN))  in smc_tx_write_space()
69 	if (smc->sk.sk_socket &&  in smc_tx_sndbuf_nonfull()
70 	    test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags))  in smc_tx_sndbuf_nonfull()
71 		smc->sk.sk_write_space(&smc->sk);  in smc_tx_sndbuf_nonfull()
80 	struct smc_connection *conn = &smc->conn;  in smc_tx_wait()
81 	struct sock *sk = &smc->sk;  in smc_tx_wait()
90 		if (sk->sk_err ||  in smc_tx_wait()
91 		    (sk->sk_shutdown & SEND_SHUTDOWN) ||  in smc_tx_wait()
92 		    conn->killed ||  in smc_tx_wait()
93 		    conn->local_tx_ctrl.conn_state_flags.peer_done_writing) {  in smc_tx_wait()
94 			rc = -EPIPE;  in smc_tx_wait()
98 			rc = -ECONNRESET;  in smc_tx_wait()
103 			set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);  in smc_tx_wait()
104 			rc = -EAGAIN;  in smc_tx_wait()
112 		if (atomic_read(&conn->sndbuf_space) && !conn->urg_tx_pend)  in smc_tx_wait()
114 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);  in smc_tx_wait()
116 			      READ_ONCE(sk->sk_err) ||  in smc_tx_wait()
117 			      (READ_ONCE(sk->sk_shutdown) & SEND_SHUTDOWN) ||  in smc_tx_wait()
119 			      (atomic_read(&conn->sndbuf_space) &&  in smc_tx_wait()
120 			       !conn->urg_tx_pend),  in smc_tx_wait()
129 	struct tcp_sock *tp = tcp_sk(smc->clcsock->sk);  in smc_tx_is_corked()
131 	return (tp->nonagle & TCP_NAGLE_CORK) ? true : false;  in smc_tx_is_corked()
142  *  3. Don't cork to much data in a single RDMA Write to prevent burst
147 	struct smc_connection *conn = &smc->conn;  in smc_should_autocork()
150 	corking_size = min_t(unsigned int, conn->sndbuf_desc->len >> 1,  in smc_should_autocork()
151 			     sock_net(&smc->sk)->smc.sysctl_autocorking_size);  in smc_should_autocork()
153 	if (atomic_read(&conn->cdc_pend_tx_wr) == 0 ||  in smc_should_autocork()
161 	struct smc_connection *conn = &smc->conn;  in smc_tx_should_cork()
170 	if ((msg->msg_flags & MSG_MORE ||  in smc_tx_should_cork()
172 	    atomic_read(&conn->sndbuf_space))  in smc_tx_should_cork()
185 	struct smc_connection *conn = &smc->conn;  in smc_tx_sendmsg()
187 	struct sock *sk = &smc->sk;  in smc_tx_sendmsg()
196 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {  in smc_tx_sendmsg()
197 		rc = -EPIPE;  in smc_tx_sendmsg()
201 	if (sk->sk_state == SMC_INIT)  in smc_tx_sendmsg()
202 		return -ENOTCONN;  in smc_tx_sendmsg()
204 	if (len > conn->sndbuf_desc->len)  in smc_tx_sendmsg()
205 		SMC_STAT_RMB_TX_SIZE_SMALL(smc, !conn->lnk);  in smc_tx_sendmsg()
207 	if (len > conn->peer_rmbe_size)  in smc_tx_sendmsg()
208 		SMC_STAT_RMB_TX_PEER_SIZE_SMALL(smc, !conn->lnk);  in smc_tx_sendmsg()
210 	if (msg->msg_flags & MSG_OOB)  in smc_tx_sendmsg()
214 		if (smc->sk.sk_shutdown & SEND_SHUTDOWN ||  in smc_tx_sendmsg()
215 		    (smc->sk.sk_err == ECONNABORTED) ||  in smc_tx_sendmsg()
216 		    conn->killed)  in smc_tx_sendmsg()
217 			return -EPIPE;  in smc_tx_sendmsg()
219 			return send_done ?: -ECONNRESET;  in smc_tx_sendmsg()
221 		if (msg->msg_flags & MSG_OOB)  in smc_tx_sendmsg()
222 			conn->local_tx_ctrl.prod_flags.urg_data_pending = 1;  in smc_tx_sendmsg()
224 		if (!atomic_read(&conn->sndbuf_space) || conn->urg_tx_pend) {  in smc_tx_sendmsg()
227 			rc = smc_tx_wait(smc, msg->msg_flags);  in smc_tx_sendmsg()
235 		writespace = atomic_read(&conn->sndbuf_space);  in smc_tx_sendmsg()
239 		sndbuf_base = conn->sndbuf_desc->cpu_addr;  in smc_tx_sendmsg()
240 		smc_curs_copy(&prep, &conn->tx_curs_prep, conn);  in smc_tx_sendmsg()
244 		chunk_len = min_t(size_t, copylen, conn->sndbuf_desc->len -  in smc_tx_sendmsg()
258 			send_remaining -= chunk_len;  in smc_tx_sendmsg()
263 			chunk_len = copylen - chunk_len; /* remainder */  in smc_tx_sendmsg()
269 		smc_curs_add(conn->sndbuf_desc->len, &prep, copylen);  in smc_tx_sendmsg()
270 		smc_curs_copy(&conn->tx_curs_prep, &prep, conn);  in smc_tx_sendmsg()
273 		atomic_sub(copylen, &conn->sndbuf_space);  in smc_tx_sendmsg()
274 		/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */  in smc_tx_sendmsg()
279 		if ((msg->msg_flags & MSG_OOB) && !send_remaining)  in smc_tx_sendmsg()
280 			conn->urg_tx_pend = true;  in smc_tx_sendmsg()
293 	rc = sk_stream_error(sk, msg->msg_flags, rc);  in smc_tx_sendmsg()
295 	if (unlikely(rc == -EAGAIN))  in smc_tx_sendmsg()
296 		sk->sk_write_space(sk);  in smc_tx_sendmsg()
308 	rc = smc_ism_write(conn->lgr->smcd, conn->peer_token,  in smcd_tx_ism_write()
309 			   conn->peer_rmbe_idx, signal, conn->tx_off + offset,  in smcd_tx_ism_write()
312 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;  in smcd_tx_ism_write()
320 	struct smc_link_group *lgr = conn->lgr;  in smc_tx_rdma_write()
321 	struct smc_link *link = conn->lnk;  in smc_tx_rdma_write()
324 	rdma_wr->wr.wr_id = smc_wr_tx_get_next_wr_id(link);  in smc_tx_rdma_write()
325 	rdma_wr->wr.num_sge = num_sges;  in smc_tx_rdma_write()
326 	rdma_wr->remote_addr =  in smc_tx_rdma_write()
327 		lgr->rtokens[conn->rtoken_idx][link->link_idx].dma_addr +  in smc_tx_rdma_write()
329 		conn->tx_off +  in smc_tx_rdma_write()
332 	rdma_wr->rkey = lgr->rtokens[conn->rtoken_idx][link->link_idx].rkey;  in smc_tx_rdma_write()
333 	rc = ib_post_send(link->roce_qp, &rdma_wr->wr, NULL);  in smc_tx_rdma_write()
345 	smc_curs_add(conn->peer_rmbe_size, prod, len);  in smc_tx_advance_cursors()
349 	atomic_sub(len, &conn->peer_rmbe_space);  in smc_tx_advance_cursors()
352 	smc_curs_add(conn->sndbuf_desc->len, sent, len);  in smc_tx_advance_cursors()
355 /* SMC-R helper for smc_tx_rdma_writes() */
361 	struct smc_link *link = conn->lnk;  in smcr_tx_rdma_writes()
364 		sg_dma_address(conn->sndbuf_desc->sgt[link->link_idx].sgl);  in smcr_tx_rdma_writes()
365 	u64 virt_addr = (uintptr_t)conn->sndbuf_desc->cpu_addr;  in smcr_tx_rdma_writes()
373 		struct ib_rdma_wr *wr = &wr_rdma_buf->wr_tx_rdma[dstchunk];  in smcr_tx_rdma_writes()
374 		struct ib_sge *sge = wr->wr.sg_list;  in smcr_tx_rdma_writes()
377 		if (dst_len < link->qp_attr.cap.max_inline_data) {  in smcr_tx_rdma_writes()
379 			wr->wr.send_flags |= IB_SEND_INLINE;  in smcr_tx_rdma_writes()
381 			wr->wr.send_flags &= ~IB_SEND_INLINE;  in smcr_tx_rdma_writes()
386 			sge[srcchunk].addr = conn->sndbuf_desc->is_vm ?  in smcr_tx_rdma_writes()
389 			if (conn->sndbuf_desc->is_vm)  in smcr_tx_rdma_writes()
391 					conn->sndbuf_desc->mr[link->link_idx]->lkey;  in smcr_tx_rdma_writes()
395 			if (src_off >= conn->sndbuf_desc->len)  in smcr_tx_rdma_writes()
396 				src_off -= conn->sndbuf_desc->len;  in smcr_tx_rdma_writes()
401 			src_len = dst_len - src_len; /* remainder */  in smcr_tx_rdma_writes()
411 		dst_len = len - dst_len; /* remainder */  in smcr_tx_rdma_writes()
413 		src_len = min_t(int, dst_len, conn->sndbuf_desc->len -  in smcr_tx_rdma_writes()
420 /* SMC-D helper for smc_tx_rdma_writes() */
431 			void *data = conn->sndbuf_desc->cpu_addr + src_off;  in smcd_tx_rdma_writes()
439 			if (src_off >= conn->sndbuf_desc->len)  in smcd_tx_rdma_writes()
440 				src_off -= conn->sndbuf_desc->len;  in smcd_tx_rdma_writes()
445 			src_len = dst_len - src_len; /* remainder */  in smcd_tx_rdma_writes()
452 		dst_len = len - dst_len; /* remainder */  in smcd_tx_rdma_writes()
454 		src_len = min_t(int, dst_len, conn->sndbuf_desc->len - src_off);  in smcd_tx_rdma_writes()
473 	smc_curs_copy(&sent, &conn->tx_curs_sent, conn);  in smc_tx_rdma_writes()
474 	smc_curs_copy(&prep, &conn->tx_curs_prep, conn);  in smc_tx_rdma_writes()
475 	/* cf. wmem_alloc - (snd_max - snd_una) */  in smc_tx_rdma_writes()
476 	to_send = smc_curs_diff(conn->sndbuf_desc->len, &sent, &prep);  in smc_tx_rdma_writes()
482 	rmbespace = atomic_read(&conn->peer_rmbe_space);  in smc_tx_rdma_writes()
486 		SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);  in smc_tx_rdma_writes()
489 	smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);  in smc_tx_rdma_writes()
490 	smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);  in smc_tx_rdma_writes()
493 	pflags = &conn->local_tx_ctrl.prod_flags;  in smc_tx_rdma_writes()
494 	pflags->write_blocked = (to_send >= rmbespace);  in smc_tx_rdma_writes()
500 	if (prod.wrap == cons.wrap) {  in smc_tx_rdma_writes()
507 				conn->peer_rmbe_size - prod.count, len);  in smc_tx_rdma_writes()
516 	if (sent.count + dst_len <= conn->sndbuf_desc->len) {  in smc_tx_rdma_writes()
521 		src_len = conn->sndbuf_desc->len - sent.count;  in smc_tx_rdma_writes()
524 	if (conn->lgr->is_smcd)  in smc_tx_rdma_writes()
533 	if (conn->urg_tx_pend && len == to_send)  in smc_tx_rdma_writes()
534 		pflags->urg_data_present = 1;  in smc_tx_rdma_writes()
537 	smc_curs_copy(&conn->local_tx_ctrl.prod, &prod, conn);  in smc_tx_rdma_writes()
539 	smc_curs_copy(&conn->tx_curs_sent, &sent, conn);/* src: local sndbuf */  in smc_tx_rdma_writes()
549 	struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;  in smcr_tx_sndbuf_nonempty()
550 	struct smc_link *link = conn->lnk;  in smcr_tx_sndbuf_nonempty()
557 		return -ENOLINK;  in smcr_tx_sndbuf_nonempty()
561 		if (rc == -EBUSY) {  in smcr_tx_sndbuf_nonempty()
565 			if (smc->sk.sk_err == ECONNABORTED)  in smcr_tx_sndbuf_nonempty()
566 				return sock_error(&smc->sk);  in smcr_tx_sndbuf_nonempty()
567 			if (conn->killed)  in smcr_tx_sndbuf_nonempty()
568 				return -EPIPE;  in smcr_tx_sndbuf_nonempty()
570 			mod_delayed_work(conn->lgr->tx_wq, &conn->tx_work,  in smcr_tx_sndbuf_nonempty()
576 	spin_lock_bh(&conn->send_lock);  in smcr_tx_sndbuf_nonempty()
577 	if (link != conn->lnk) {  in smcr_tx_sndbuf_nonempty()
581 		rc = -ENOLINK;  in smcr_tx_sndbuf_nonempty()
584 	if (!pflags->urg_data_present) {  in smcr_tx_sndbuf_nonempty()
594 	if (!rc && pflags->urg_data_present) {  in smcr_tx_sndbuf_nonempty()
595 		pflags->urg_data_pending = 0;  in smcr_tx_sndbuf_nonempty()
596 		pflags->urg_data_present = 0;  in smcr_tx_sndbuf_nonempty()
600 	spin_unlock_bh(&conn->send_lock);  in smcr_tx_sndbuf_nonempty()
607 	struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;  in smcd_tx_sndbuf_nonempty()
610 	spin_lock_bh(&conn->send_lock);  in smcd_tx_sndbuf_nonempty()
611 	if (!pflags->urg_data_present)  in smcd_tx_sndbuf_nonempty()
616 	if (!rc && pflags->urg_data_present) {  in smcd_tx_sndbuf_nonempty()
617 		pflags->urg_data_pending = 0;  in smcd_tx_sndbuf_nonempty()
618 		pflags->urg_data_present = 0;  in smcd_tx_sndbuf_nonempty()
620 	spin_unlock_bh(&conn->send_lock);  in smcd_tx_sndbuf_nonempty()
634 	if (unlikely(atomic_read(&conn->peer_rmbe_space) <= 0)) {  in smc_tx_sndbuf_nonempty()
635 		SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);  in smc_tx_sndbuf_nonempty()
639 	if (conn->killed ||  in smc_tx_sndbuf_nonempty()
640 	    conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {  in smc_tx_sndbuf_nonempty()
641 		rc = -EPIPE;    /* connection being aborted */  in smc_tx_sndbuf_nonempty()
644 	if (conn->lgr->is_smcd)  in smc_tx_sndbuf_nonempty()
667 	if (smc->sk.sk_err)  in smc_tx_pending()
671 	if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked &&  in smc_tx_pending()
672 	    !atomic_read(&conn->bytes_to_rcv))  in smc_tx_pending()
673 		conn->local_rx_ctrl.prod_flags.write_blocked = 0;  in smc_tx_pending()
687 	lock_sock(&smc->sk);  in smc_tx_work()
689 	release_sock(&smc->sk);  in smc_tx_work()
695 	int sender_free = conn->rmb_desc->len;  in smc_tx_consumer_update()
698 	smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);  in smc_tx_consumer_update()
699 	smc_curs_copy(&cfed, &conn->rx_curs_confirmed, conn);  in smc_tx_consumer_update()
700 	to_confirm = smc_curs_diff(conn->rmb_desc->len, &cfed, &cons);  in smc_tx_consumer_update()
701 	if (to_confirm > conn->rmbe_update_limit) {  in smc_tx_consumer_update()
702 		smc_curs_copy(&prod, &conn->local_rx_ctrl.prod, conn);  in smc_tx_consumer_update()
703 		sender_free = conn->rmb_desc->len -  in smc_tx_consumer_update()
704 			      smc_curs_diff_large(conn->rmb_desc->len,  in smc_tx_consumer_update()
708 	if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||  in smc_tx_consumer_update()
710 	    ((to_confirm > conn->rmbe_update_limit) &&  in smc_tx_consumer_update()
711 	     ((sender_free <= (conn->rmb_desc->len / 2)) ||  in smc_tx_consumer_update()
712 	      conn->local_rx_ctrl.prod_flags.write_blocked))) {  in smc_tx_consumer_update()
713 		if (conn->killed ||  in smc_tx_consumer_update()
714 		    conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)  in smc_tx_consumer_update()
717 		    !conn->killed) {  in smc_tx_consumer_update()
718 			queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work,  in smc_tx_consumer_update()
723 	if (conn->local_rx_ctrl.prod_flags.write_blocked &&  in smc_tx_consumer_update()
724 	    !atomic_read(&conn->bytes_to_rcv))  in smc_tx_consumer_update()
725 		conn->local_rx_ctrl.prod_flags.write_blocked = 0;  in smc_tx_consumer_update()
733 	smc->sk.sk_write_space = smc_tx_write_space;  in smc_tx_init()