2 * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/kernel.h>
35 #include <linux/slab.h>
36 #include <linux/vmalloc.h>
37 #include <linux/ratelimit.h>
38 #include <net/addrconf.h>
40 #include "rds_single_path.h"
45 * Set the selected protocol version
47 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
49 conn->c_version = version;
55 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
57 struct rds_ib_connection *ic = conn->c_transport_data;
59 if (rds_ib_sysctl_flow_control && credits != 0) {
60 /* We're doing flow control */
62 rds_ib_send_add_credits(conn, credits);
69 * Tune RNR behavior. Without flow control, we use a rather
70 * low timeout, but not the absolute minimum - this should
73 * We already set the RNR retry count to 7 (which is the
74 * smallest infinite number :-) above.
75 * If flow control is off, we want to change this back to 0
76 * so that we learn quickly when our credit accounting is
79 * Caller passes in a qp_attr pointer - don't waste stack spacv
80 * by allocation this twice.
83 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
87 attr->min_rnr_timer = IB_RNR_TIMER_000_32;
88 ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
90 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
94 * Connection established.
95 * We get here for both outgoing and incoming connection.
97 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
99 struct rds_ib_connection *ic = conn->c_transport_data;
100 const union rds_ib_conn_priv *dp = NULL;
101 struct ib_qp_attr qp_attr;
108 dp = event->param.conn.private_data;
110 if (event->param.conn.private_data_len >=
111 sizeof(struct rds6_ib_connect_private)) {
112 major = dp->ricp_v6.dp_protocol_major;
113 minor = dp->ricp_v6.dp_protocol_minor;
114 credit = dp->ricp_v6.dp_credit;
115 /* dp structure start is not guaranteed to be 8 bytes
116 * aligned. Since dp_ack_seq is 64-bit extended load
117 * operations can be used so go through get_unaligned
118 * to avoid unaligned errors.
120 ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq);
122 } else if (event->param.conn.private_data_len >=
123 sizeof(struct rds_ib_connect_private)) {
124 major = dp->ricp_v4.dp_protocol_major;
125 minor = dp->ricp_v4.dp_protocol_minor;
126 credit = dp->ricp_v4.dp_credit;
127 ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq);
130 /* make sure it isn't empty data */
132 rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor));
133 rds_ib_set_flow_control(conn, be32_to_cpu(credit));
136 if (conn->c_version < RDS_PROTOCOL_VERSION) {
137 if (conn->c_version != RDS_PROTOCOL_COMPAT_VERSION) {
138 pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n",
139 &conn->c_laddr, &conn->c_faddr,
140 RDS_PROTOCOL_MAJOR(conn->c_version),
141 RDS_PROTOCOL_MINOR(conn->c_version));
142 rds_conn_destroy(conn);
147 pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c,%d> version %u.%u%s\n",
148 ic->i_active_side ? "Active" : "Passive",
149 &conn->c_laddr, &conn->c_faddr, conn->c_tos,
150 RDS_PROTOCOL_MAJOR(conn->c_version),
151 RDS_PROTOCOL_MINOR(conn->c_version),
152 ic->i_flowctl ? ", flow control" : "");
154 atomic_set(&ic->i_cq_quiesce, 0);
156 /* Init rings and fill recv. this needs to wait until protocol
157 * negotiation is complete, since ring layout is different
160 rds_ib_send_init_ring(ic);
161 rds_ib_recv_init_ring(ic);
162 /* Post receive buffers - as a side effect, this will update
163 * the posted credit count. */
164 rds_ib_recv_refill(conn, 1, GFP_KERNEL);
166 /* Tune RNR behavior */
167 rds_ib_tune_rnr(ic, &qp_attr);
169 qp_attr.qp_state = IB_QPS_RTS;
170 err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
172 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
174 /* update ib_device with this local ipaddr */
175 err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr);
177 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
180 /* If the peer gave us the last packet it saw, process this as if
181 * we had received a regular ACK. */
184 rds_send_drop_acked(conn, be64_to_cpu(ack_seq),
188 conn->c_proposed_version = conn->c_version;
189 rds_connect_complete(conn);
192 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
193 struct rdma_conn_param *conn_param,
194 union rds_ib_conn_priv *dp,
195 u32 protocol_version,
196 u32 max_responder_resources,
197 u32 max_initiator_depth,
200 struct rds_ib_connection *ic = conn->c_transport_data;
201 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
203 memset(conn_param, 0, sizeof(struct rdma_conn_param));
205 conn_param->responder_resources =
206 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
207 conn_param->initiator_depth =
208 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
209 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
210 conn_param->rnr_retry_count = 7;
213 memset(dp, 0, sizeof(*dp));
215 dp->ricp_v6.dp_saddr = conn->c_laddr;
216 dp->ricp_v6.dp_daddr = conn->c_faddr;
217 dp->ricp_v6.dp_protocol_major =
218 RDS_PROTOCOL_MAJOR(protocol_version);
219 dp->ricp_v6.dp_protocol_minor =
220 RDS_PROTOCOL_MINOR(protocol_version);
221 dp->ricp_v6.dp_protocol_minor_mask =
222 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
223 dp->ricp_v6.dp_ack_seq =
224 cpu_to_be64(rds_ib_piggyb_ack(ic));
225 dp->ricp_v6.dp_cmn.ricpc_dp_toss = conn->c_tos;
227 conn_param->private_data = &dp->ricp_v6;
228 conn_param->private_data_len = sizeof(dp->ricp_v6);
230 dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3];
231 dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3];
232 dp->ricp_v4.dp_protocol_major =
233 RDS_PROTOCOL_MAJOR(protocol_version);
234 dp->ricp_v4.dp_protocol_minor =
235 RDS_PROTOCOL_MINOR(protocol_version);
236 dp->ricp_v4.dp_protocol_minor_mask =
237 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
238 dp->ricp_v4.dp_ack_seq =
239 cpu_to_be64(rds_ib_piggyb_ack(ic));
240 dp->ricp_v4.dp_cmn.ricpc_dp_toss = conn->c_tos;
242 conn_param->private_data = &dp->ricp_v4;
243 conn_param->private_data_len = sizeof(dp->ricp_v4);
246 /* Advertise flow control */
248 unsigned int credits;
250 credits = IB_GET_POST_CREDITS
251 (atomic_read(&ic->i_credits));
253 dp->ricp_v6.dp_credit = cpu_to_be32(credits);
255 dp->ricp_v4.dp_credit = cpu_to_be32(credits);
256 atomic_sub(IB_SET_POST_CREDITS(credits),
262 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
264 rdsdebug("event %u (%s) data %p\n",
265 event->event, ib_event_msg(event->event), data);
268 /* Plucking the oldest entry from the ring can be done concurrently with
269 * the thread refilling the ring. Each ring operation is protected by
270 * spinlocks and the transient state of refilling doesn't change the
271 * recording of which entry is oldest.
273 * This relies on IB only calling one cq comp_handler for each cq so that
274 * there will only be one caller of rds_recv_incoming() per RDS connection.
276 static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
278 struct rds_connection *conn = context;
279 struct rds_ib_connection *ic = conn->c_transport_data;
281 rdsdebug("conn %p cq %p\n", conn, cq);
283 rds_ib_stats_inc(s_ib_evt_handler_call);
285 tasklet_schedule(&ic->i_recv_tasklet);
288 static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
294 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
295 for (i = 0; i < nr; i++) {
297 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
298 (unsigned long long)wc->wr_id, wc->status,
299 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
301 if (wc->wr_id <= ic->i_send_ring.w_nr ||
302 wc->wr_id == RDS_IB_ACK_WR_ID)
303 rds_ib_send_cqe_handler(ic, wc);
305 rds_ib_mr_cqe_handler(ic, wc);
311 static void rds_ib_tasklet_fn_send(unsigned long data)
313 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
314 struct rds_connection *conn = ic->conn;
316 rds_ib_stats_inc(s_ib_tasklet_call);
318 /* if cq has been already reaped, ignore incoming cq event */
319 if (atomic_read(&ic->i_cq_quiesce))
322 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
323 ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
324 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
326 if (rds_conn_up(conn) &&
327 (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
328 test_bit(0, &conn->c_map_queued)))
329 rds_send_xmit(&ic->conn->c_path[0]);
332 static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
334 struct rds_ib_ack_state *ack_state)
339 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
340 for (i = 0; i < nr; i++) {
342 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
343 (unsigned long long)wc->wr_id, wc->status,
344 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
346 rds_ib_recv_cqe_handler(ic, wc, ack_state);
351 static void rds_ib_tasklet_fn_recv(unsigned long data)
353 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
354 struct rds_connection *conn = ic->conn;
355 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
356 struct rds_ib_ack_state state;
361 rds_ib_stats_inc(s_ib_tasklet_call);
363 /* if cq has been already reaped, ignore incoming cq event */
364 if (atomic_read(&ic->i_cq_quiesce))
367 memset(&state, 0, sizeof(state));
368 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
369 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
370 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
372 if (state.ack_next_valid)
373 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
374 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
375 rds_send_drop_acked(conn, state.ack_recv, NULL);
376 ic->i_ack_recv = state.ack_recv;
379 if (rds_conn_up(conn))
380 rds_ib_attempt_ack(ic);
383 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
385 struct rds_connection *conn = data;
386 struct rds_ib_connection *ic = conn->c_transport_data;
388 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
389 ib_event_msg(event->event));
391 switch (event->event) {
392 case IB_EVENT_COMM_EST:
393 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
396 rdsdebug("Fatal QP Event %u (%s) - connection %pI6c->%pI6c, reconnecting\n",
397 event->event, ib_event_msg(event->event),
398 &conn->c_laddr, &conn->c_faddr);
404 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
406 struct rds_connection *conn = context;
407 struct rds_ib_connection *ic = conn->c_transport_data;
409 rdsdebug("conn %p cq %p\n", conn, cq);
411 rds_ib_stats_inc(s_ib_evt_handler_call);
413 tasklet_schedule(&ic->i_send_tasklet);
416 static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
418 int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
419 int index = rds_ibdev->dev->num_comp_vectors - 1;
422 for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
423 if (rds_ibdev->vector_load[i] < min) {
425 min = rds_ibdev->vector_load[i];
429 rds_ibdev->vector_load[index]++;
433 static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
435 rds_ibdev->vector_load[index]--;
439 * This needs to be very careful to not leave IS_ERR pointers around for
440 * cleanup to trip over.
442 static int rds_ib_setup_qp(struct rds_connection *conn)
444 struct rds_ib_connection *ic = conn->c_transport_data;
445 struct ib_device *dev = ic->i_cm_id->device;
446 struct ib_qp_init_attr attr;
447 struct ib_cq_init_attr cq_attr = {};
448 struct rds_ib_device *rds_ibdev;
449 int ret, fr_queue_space;
452 * It's normal to see a null device if an incoming connection races
453 * with device removal, so we don't print a warning.
455 rds_ibdev = rds_ib_get_client_data(dev);
459 /* The fr_queue_space is currently set to 512, to add extra space on
460 * completion queue and send queue. This extra space is used for FRMR
461 * registration and invalidation work requests
463 fr_queue_space = (rds_ibdev->use_fastreg ? RDS_IB_DEFAULT_FR_WR : 0);
465 /* add the conn now so that connection establishment has the dev */
466 rds_ib_add_conn(rds_ibdev, conn);
468 if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
469 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
470 if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
471 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
473 /* Protection domain and memory range */
474 ic->i_pd = rds_ibdev->pd;
476 ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
477 cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
478 cq_attr.comp_vector = ic->i_scq_vector;
479 ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
480 rds_ib_cq_event_handler, conn,
482 if (IS_ERR(ic->i_send_cq)) {
483 ret = PTR_ERR(ic->i_send_cq);
484 ic->i_send_cq = NULL;
485 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
486 rdsdebug("ib_create_cq send failed: %d\n", ret);
490 ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
491 cq_attr.cqe = ic->i_recv_ring.w_nr;
492 cq_attr.comp_vector = ic->i_rcq_vector;
493 ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
494 rds_ib_cq_event_handler, conn,
496 if (IS_ERR(ic->i_recv_cq)) {
497 ret = PTR_ERR(ic->i_recv_cq);
498 ic->i_recv_cq = NULL;
499 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
500 rdsdebug("ib_create_cq recv failed: %d\n", ret);
504 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
506 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
510 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
512 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
516 /* XXX negotiate max send/recv with remote? */
517 memset(&attr, 0, sizeof(attr));
518 attr.event_handler = rds_ib_qp_event_handler;
519 attr.qp_context = conn;
520 /* + 1 to allow for the single ack message */
521 attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
522 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
523 attr.cap.max_send_sge = rds_ibdev->max_sge;
524 attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
525 attr.sq_sig_type = IB_SIGNAL_REQ_WR;
526 attr.qp_type = IB_QPT_RC;
527 attr.send_cq = ic->i_send_cq;
528 attr.recv_cq = ic->i_recv_cq;
529 atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
532 * XXX this can fail if max_*_wr is too large? Are we supposed
533 * to back off until we get a value that the hardware can support?
535 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
537 rdsdebug("rdma_create_qp failed: %d\n", ret);
541 ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
542 ic->i_send_ring.w_nr *
543 sizeof(struct rds_header),
544 &ic->i_send_hdrs_dma, GFP_KERNEL);
545 if (!ic->i_send_hdrs) {
547 rdsdebug("ib_dma_alloc_coherent send failed\n");
551 ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
552 ic->i_recv_ring.w_nr *
553 sizeof(struct rds_header),
554 &ic->i_recv_hdrs_dma, GFP_KERNEL);
555 if (!ic->i_recv_hdrs) {
557 rdsdebug("ib_dma_alloc_coherent recv failed\n");
558 goto send_hdrs_dma_out;
561 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
562 &ic->i_ack_dma, GFP_KERNEL);
565 rdsdebug("ib_dma_alloc_coherent ack failed\n");
566 goto recv_hdrs_dma_out;
569 ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
570 ic->i_send_ring.w_nr),
574 rdsdebug("send allocation failed\n");
578 ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
579 ic->i_recv_ring.w_nr),
583 rdsdebug("recv allocation failed\n");
587 rds_ib_recv_init_ack(ic);
589 rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
590 ic->i_send_cq, ic->i_recv_cq);
597 ib_dma_free_coherent(dev, sizeof(struct rds_header),
598 ic->i_ack, ic->i_ack_dma);
600 ib_dma_free_coherent(dev, ic->i_recv_ring.w_nr *
601 sizeof(struct rds_header),
602 ic->i_recv_hdrs, ic->i_recv_hdrs_dma);
604 ib_dma_free_coherent(dev, ic->i_send_ring.w_nr *
605 sizeof(struct rds_header),
606 ic->i_send_hdrs, ic->i_send_hdrs_dma);
608 rdma_destroy_qp(ic->i_cm_id);
610 ib_destroy_cq(ic->i_recv_cq);
611 ic->i_recv_cq = NULL;
613 ib_destroy_cq(ic->i_send_cq);
614 ic->i_send_cq = NULL;
616 rds_ib_remove_conn(rds_ibdev, conn);
618 rds_ib_dev_put(rds_ibdev);
623 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
625 const union rds_ib_conn_priv *dp = event->param.conn.private_data;
626 u8 data_len, major, minor;
632 * rdma_cm private data is odd - when there is any private data in the
633 * request, we will be given a pretty large buffer without telling us the
634 * original size. The only way to tell the difference is by looking at
635 * the contents, which are initialized to zero.
636 * If the protocol version fields aren't set, this is a connection attempt
637 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
638 * We really should have changed this for OFED 1.3 :-(
641 /* Be paranoid. RDS always has privdata */
642 if (!event->param.conn.private_data_len) {
643 printk(KERN_NOTICE "RDS incoming connection has no private data, "
649 data_len = sizeof(struct rds6_ib_connect_private);
650 major = dp->ricp_v6.dp_protocol_major;
651 minor = dp->ricp_v6.dp_protocol_minor;
652 mask = dp->ricp_v6.dp_protocol_minor_mask;
654 data_len = sizeof(struct rds_ib_connect_private);
655 major = dp->ricp_v4.dp_protocol_major;
656 minor = dp->ricp_v4.dp_protocol_minor;
657 mask = dp->ricp_v4.dp_protocol_minor_mask;
660 /* Even if len is crap *now* I still want to check it. -ASG */
661 if (event->param.conn.private_data_len < data_len || major == 0)
662 return RDS_PROTOCOL_4_0;
664 common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS;
665 if (major == 4 && common) {
666 version = RDS_PROTOCOL_4_0;
667 while ((common >>= 1) != 0)
669 } else if (RDS_PROTOCOL_COMPAT_VERSION ==
670 RDS_PROTOCOL(major, minor)) {
671 version = RDS_PROTOCOL_COMPAT_VERSION;
674 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
675 &dp->ricp_v6.dp_saddr, major, minor);
677 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
678 &dp->ricp_v4.dp_saddr, major, minor);
683 #if IS_ENABLED(CONFIG_IPV6)
684 /* Given an IPv6 address, find the net_device which hosts that address and
685 * return its index. This is used by the rds_ib_cm_handle_connect() code to
686 * find the interface index of where an incoming request comes from when
687 * the request is using a link local address.
689 * Note one problem in this search. It is possible that two interfaces have
690 * the same link local address. Unfortunately, this cannot be solved unless
691 * the underlying layer gives us the interface which an incoming RDMA connect
692 * request comes from.
694 static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
696 struct net_device *dev;
700 for_each_netdev_rcu(net, dev) {
701 if (ipv6_chk_addr(net, addr, dev, 1)) {
712 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
713 struct rdma_cm_event *event, bool isv6)
715 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
716 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
717 const struct rds_ib_conn_priv_cmn *dp_cmn;
718 struct rds_connection *conn = NULL;
719 struct rds_ib_connection *ic = NULL;
720 struct rdma_conn_param conn_param;
721 const union rds_ib_conn_priv *dp;
722 union rds_ib_conn_priv dp_rep;
723 struct in6_addr s_mapped_addr;
724 struct in6_addr d_mapped_addr;
725 const struct in6_addr *saddr6;
726 const struct in6_addr *daddr6;
732 /* Check whether the remote protocol version matches ours. */
733 version = rds_ib_protocol_compatible(event, isv6);
735 err = RDS_RDMA_REJ_INCOMPAT;
739 dp = event->param.conn.private_data;
741 #if IS_ENABLED(CONFIG_IPV6)
742 dp_cmn = &dp->ricp_v6.dp_cmn;
743 saddr6 = &dp->ricp_v6.dp_saddr;
744 daddr6 = &dp->ricp_v6.dp_daddr;
745 /* If either address is link local, need to find the
746 * interface index in order to create a proper RDS
749 if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) {
750 /* Using init_net for now .. */
751 ifindex = __rds_find_ifindex(&init_net, daddr6);
752 /* No index found... Need to bail out. */
757 } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) {
758 /* Use our address to find the correct index. */
759 ifindex = __rds_find_ifindex(&init_net, daddr6);
760 /* No index found... Need to bail out. */
771 dp_cmn = &dp->ricp_v4.dp_cmn;
772 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr);
773 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
774 saddr6 = &s_mapped_addr;
775 daddr6 = &d_mapped_addr;
778 rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid 0x%llx, tos:%d\n",
779 saddr6, daddr6, RDS_PROTOCOL_MAJOR(version),
780 RDS_PROTOCOL_MINOR(version),
781 (unsigned long long)be64_to_cpu(lguid),
782 (unsigned long long)be64_to_cpu(fguid), dp_cmn->ricpc_dp_toss);
784 /* RDS/IB is not currently netns aware, thus init_net */
785 conn = rds_conn_create(&init_net, daddr6, saddr6,
786 &rds_ib_transport, dp_cmn->ricpc_dp_toss,
787 GFP_KERNEL, ifindex);
789 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
795 * The connection request may occur while the
796 * previous connection exist, e.g. in case of failover.
797 * But as connections may be initiated simultaneously
798 * by both hosts, we have a random backoff mechanism -
799 * see the comment above rds_queue_reconnect()
801 mutex_lock(&conn->c_cm_lock);
802 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
803 if (rds_conn_state(conn) == RDS_CONN_UP) {
804 rdsdebug("incoming connect while connecting\n");
806 rds_ib_stats_inc(s_ib_listen_closed_stale);
808 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
809 /* Wait and see - our connect may still be succeeding */
810 rds_ib_stats_inc(s_ib_connect_raced);
815 ic = conn->c_transport_data;
817 rds_ib_set_protocol(conn, version);
818 rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
820 /* If the peer gave us the last packet it saw, process this as if
821 * we had received a regular ACK. */
822 if (dp_cmn->ricpc_ack_seq)
823 rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq),
826 BUG_ON(cm_id->context);
830 cm_id->context = conn;
832 /* We got halfway through setting up the ib_connection, if we
833 * fail now, we have to take the long route out of this mess. */
836 err = rds_ib_setup_qp(conn);
838 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
842 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
843 event->param.conn.responder_resources,
844 event->param.conn.initiator_depth, isv6);
846 /* rdma_accept() calls rdma_reject() internally if it fails */
847 if (rdma_accept(cm_id, &conn_param))
848 rds_ib_conn_error(conn, "rdma_accept failed\n");
852 mutex_unlock(&conn->c_cm_lock);
854 rdma_reject(cm_id, &err, sizeof(int));
859 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
861 struct rds_connection *conn = cm_id->context;
862 struct rds_ib_connection *ic = conn->c_transport_data;
863 struct rdma_conn_param conn_param;
864 union rds_ib_conn_priv dp;
867 /* If the peer doesn't do protocol negotiation, we must
868 * default to RDSv3.0 */
869 rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1);
870 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
872 ret = rds_ib_setup_qp(conn);
874 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
878 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
879 conn->c_proposed_version,
880 UINT_MAX, UINT_MAX, isv6);
881 ret = rdma_connect(cm_id, &conn_param);
883 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
886 /* Beware - returning non-zero tells the rdma_cm to destroy
887 * the cm_id. We should certainly not do it as long as we still
888 * "own" the cm_id. */
890 if (ic->i_cm_id == cm_id)
893 ic->i_active_side = true;
897 int rds_ib_conn_path_connect(struct rds_conn_path *cp)
899 struct rds_connection *conn = cp->cp_conn;
900 struct sockaddr_storage src, dest;
901 rdma_cm_event_handler handler;
902 struct rds_ib_connection *ic;
905 ic = conn->c_transport_data;
907 /* XXX I wonder what affect the port space has */
908 /* delegate cm event handler to rdma_transport */
909 #if IS_ENABLED(CONFIG_IPV6)
911 handler = rds6_rdma_cm_event_handler;
914 handler = rds_rdma_cm_event_handler;
915 ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
916 RDMA_PS_TCP, IB_QPT_RC);
917 if (IS_ERR(ic->i_cm_id)) {
918 ret = PTR_ERR(ic->i_cm_id);
920 rdsdebug("rdma_create_id() failed: %d\n", ret);
924 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
926 if (ipv6_addr_v4mapped(&conn->c_faddr)) {
927 struct sockaddr_in *sin;
929 sin = (struct sockaddr_in *)&src;
930 sin->sin_family = AF_INET;
931 sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
934 sin = (struct sockaddr_in *)&dest;
935 sin->sin_family = AF_INET;
936 sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
937 sin->sin_port = htons(RDS_PORT);
939 struct sockaddr_in6 *sin6;
941 sin6 = (struct sockaddr_in6 *)&src;
942 sin6->sin6_family = AF_INET6;
943 sin6->sin6_addr = conn->c_laddr;
945 sin6->sin6_scope_id = conn->c_dev_if;
947 sin6 = (struct sockaddr_in6 *)&dest;
948 sin6->sin6_family = AF_INET6;
949 sin6->sin6_addr = conn->c_faddr;
950 sin6->sin6_port = htons(RDS_CM_PORT);
951 sin6->sin6_scope_id = conn->c_dev_if;
954 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
955 (struct sockaddr *)&dest,
956 RDS_RDMA_RESOLVE_TIMEOUT_MS);
958 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
960 rdma_destroy_id(ic->i_cm_id);
969 * This is so careful about only cleaning up resources that were built up
970 * so that it can be called at any point during startup. In fact it
971 * can be called multiple times for a given connection.
973 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
975 struct rds_connection *conn = cp->cp_conn;
976 struct rds_ib_connection *ic = conn->c_transport_data;
979 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
980 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
981 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
984 struct ib_device *dev = ic->i_cm_id->device;
986 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
987 err = rdma_disconnect(ic->i_cm_id);
989 /* Actually this may happen quite frequently, when
990 * an outgoing connect raced with an incoming connect.
992 rdsdebug("failed to disconnect, cm: %p err %d\n",
997 * We want to wait for tx and rx completion to finish
998 * before we tear down the connection, but we have to be
999 * careful not to get stuck waiting on a send ring that
1000 * only has unsignaled sends in it. We've shutdown new
1001 * sends before getting here so by waiting for signaled
1002 * sends to complete we're ensured that there will be no
1003 * more tx processing.
1005 wait_event(rds_ib_ring_empty_wait,
1006 rds_ib_ring_empty(&ic->i_recv_ring) &&
1007 (atomic_read(&ic->i_signaled_sends) == 0) &&
1008 (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
1009 tasklet_kill(&ic->i_send_tasklet);
1010 tasklet_kill(&ic->i_recv_tasklet);
1012 atomic_set(&ic->i_cq_quiesce, 1);
1014 /* first destroy the ib state that generates callbacks */
1015 if (ic->i_cm_id->qp)
1016 rdma_destroy_qp(ic->i_cm_id);
1017 if (ic->i_send_cq) {
1019 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1020 ib_destroy_cq(ic->i_send_cq);
1023 if (ic->i_recv_cq) {
1025 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1026 ib_destroy_cq(ic->i_recv_cq);
1029 /* then free the resources that ib callbacks use */
1030 if (ic->i_send_hdrs)
1031 ib_dma_free_coherent(dev,
1032 ic->i_send_ring.w_nr *
1033 sizeof(struct rds_header),
1035 ic->i_send_hdrs_dma);
1037 if (ic->i_recv_hdrs)
1038 ib_dma_free_coherent(dev,
1039 ic->i_recv_ring.w_nr *
1040 sizeof(struct rds_header),
1042 ic->i_recv_hdrs_dma);
1045 ib_dma_free_coherent(dev, sizeof(struct rds_header),
1046 ic->i_ack, ic->i_ack_dma);
1049 rds_ib_send_clear_ring(ic);
1051 rds_ib_recv_clear_ring(ic);
1053 rdma_destroy_id(ic->i_cm_id);
1056 * Move connection back to the nodev list.
1059 rds_ib_remove_conn(ic->rds_ibdev, conn);
1063 ic->i_send_cq = NULL;
1064 ic->i_recv_cq = NULL;
1065 ic->i_send_hdrs = NULL;
1066 ic->i_recv_hdrs = NULL;
1069 BUG_ON(ic->rds_ibdev);
1071 /* Clear pending transmit */
1072 if (ic->i_data_op) {
1073 struct rds_message *rm;
1075 rm = container_of(ic->i_data_op, struct rds_message, data);
1076 rds_message_put(rm);
1077 ic->i_data_op = NULL;
1080 /* Clear the ACK state */
1081 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1082 #ifdef KERNEL_HAS_ATOMIC64
1083 atomic64_set(&ic->i_ack_next, 0);
1089 /* Clear flow control state */
1091 atomic_set(&ic->i_credits, 0);
1093 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
1094 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
1097 rds_inc_put(&ic->i_ibinc->ii_inc);
1105 ic->i_active_side = false;
1108 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1110 struct rds_ib_connection *ic;
1111 unsigned long flags;
1115 ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1119 ret = rds_ib_recv_alloc_caches(ic, gfp);
1125 INIT_LIST_HEAD(&ic->ib_node);
1126 tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1128 tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1130 mutex_init(&ic->i_recv_mutex);
1131 #ifndef KERNEL_HAS_ATOMIC64
1132 spin_lock_init(&ic->i_ack_lock);
1134 atomic_set(&ic->i_signaled_sends, 0);
1137 * rds_ib_conn_shutdown() waits for these to be emptied so they
1138 * must be initialized before it can be called.
1140 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
1141 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
1144 conn->c_transport_data = ic;
1146 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1147 list_add_tail(&ic->ib_node, &ib_nodev_conns);
1148 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1151 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1156 * Free a connection. Connection must be shut down and not set for reconnect.
1158 void rds_ib_conn_free(void *arg)
1160 struct rds_ib_connection *ic = arg;
1161 spinlock_t *lock_ptr;
1163 rdsdebug("ic %p\n", ic);
1166 * Conn is either on a dev's list or on the nodev list.
1167 * A race with shutdown() or connect() would cause problems
1168 * (since rds_ibdev would change) but that should never happen.
1170 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1172 spin_lock_irq(lock_ptr);
1173 list_del(&ic->ib_node);
1174 spin_unlock_irq(lock_ptr);
1176 rds_ib_recv_free_caches(ic);
1183 * An error occurred on the connection
1186 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1190 rds_conn_drop(conn);