2 * Copyright (c) 2006, 2019 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/dmapool.h>
34 #include <linux/kernel.h>
36 #include <linux/slab.h>
37 #include <linux/vmalloc.h>
38 #include <linux/ratelimit.h>
39 #include <net/addrconf.h>
40 #include <rdma/ib_cm.h>
42 #include "rds_single_path.h"
48 * Set the selected protocol version
50 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
52 conn->c_version = version;
58 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
60 struct rds_ib_connection *ic = conn->c_transport_data;
62 if (rds_ib_sysctl_flow_control && credits != 0) {
63 /* We're doing flow control */
65 rds_ib_send_add_credits(conn, credits);
72 * Tune RNR behavior. Without flow control, we use a rather
73 * low timeout, but not the absolute minimum - this should
76 * We already set the RNR retry count to 7 (which is the
77 * smallest infinite number :-) above.
78 * If flow control is off, we want to change this back to 0
79 * so that we learn quickly when our credit accounting is
82 * Caller passes in a qp_attr pointer - don't waste stack spacv
83 * by allocation this twice.
86 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
90 attr->min_rnr_timer = IB_RNR_TIMER_000_32;
91 ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
93 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
97 * Connection established.
98 * We get here for both outgoing and incoming connection.
100 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
102 struct rds_ib_connection *ic = conn->c_transport_data;
103 const union rds_ib_conn_priv *dp = NULL;
104 struct ib_qp_attr qp_attr;
111 dp = event->param.conn.private_data;
113 if (event->param.conn.private_data_len >=
114 sizeof(struct rds6_ib_connect_private)) {
115 major = dp->ricp_v6.dp_protocol_major;
116 minor = dp->ricp_v6.dp_protocol_minor;
117 credit = dp->ricp_v6.dp_credit;
118 /* dp structure start is not guaranteed to be 8 bytes
119 * aligned. Since dp_ack_seq is 64-bit extended load
120 * operations can be used so go through get_unaligned
121 * to avoid unaligned errors.
123 ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq);
125 } else if (event->param.conn.private_data_len >=
126 sizeof(struct rds_ib_connect_private)) {
127 major = dp->ricp_v4.dp_protocol_major;
128 minor = dp->ricp_v4.dp_protocol_minor;
129 credit = dp->ricp_v4.dp_credit;
130 ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq);
133 /* make sure it isn't empty data */
135 rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor));
136 rds_ib_set_flow_control(conn, be32_to_cpu(credit));
139 if (conn->c_version < RDS_PROTOCOL_VERSION) {
140 if (conn->c_version != RDS_PROTOCOL_COMPAT_VERSION) {
141 pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n",
142 &conn->c_laddr, &conn->c_faddr,
143 RDS_PROTOCOL_MAJOR(conn->c_version),
144 RDS_PROTOCOL_MINOR(conn->c_version));
145 rds_conn_destroy(conn);
150 pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c,%d> version %u.%u%s\n",
151 ic->i_active_side ? "Active" : "Passive",
152 &conn->c_laddr, &conn->c_faddr, conn->c_tos,
153 RDS_PROTOCOL_MAJOR(conn->c_version),
154 RDS_PROTOCOL_MINOR(conn->c_version),
155 ic->i_flowctl ? ", flow control" : "");
157 /* receive sl from the peer */
158 ic->i_sl = ic->i_cm_id->route.path_rec->sl;
160 atomic_set(&ic->i_cq_quiesce, 0);
162 /* Init rings and fill recv. this needs to wait until protocol
163 * negotiation is complete, since ring layout is different
166 rds_ib_send_init_ring(ic);
167 rds_ib_recv_init_ring(ic);
168 /* Post receive buffers - as a side effect, this will update
169 * the posted credit count. */
170 rds_ib_recv_refill(conn, 1, GFP_KERNEL);
172 /* Tune RNR behavior */
173 rds_ib_tune_rnr(ic, &qp_attr);
175 qp_attr.qp_state = IB_QPS_RTS;
176 err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
178 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
180 /* update ib_device with this local ipaddr */
181 err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr);
183 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
186 /* If the peer gave us the last packet it saw, process this as if
187 * we had received a regular ACK. */
190 rds_send_drop_acked(conn, be64_to_cpu(ack_seq),
194 conn->c_proposed_version = conn->c_version;
195 rds_connect_complete(conn);
198 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
199 struct rdma_conn_param *conn_param,
200 union rds_ib_conn_priv *dp,
201 u32 protocol_version,
202 u32 max_responder_resources,
203 u32 max_initiator_depth,
206 struct rds_ib_connection *ic = conn->c_transport_data;
207 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
209 memset(conn_param, 0, sizeof(struct rdma_conn_param));
211 conn_param->responder_resources =
212 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
213 conn_param->initiator_depth =
214 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
215 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
216 conn_param->rnr_retry_count = 7;
219 memset(dp, 0, sizeof(*dp));
221 dp->ricp_v6.dp_saddr = conn->c_laddr;
222 dp->ricp_v6.dp_daddr = conn->c_faddr;
223 dp->ricp_v6.dp_protocol_major =
224 RDS_PROTOCOL_MAJOR(protocol_version);
225 dp->ricp_v6.dp_protocol_minor =
226 RDS_PROTOCOL_MINOR(protocol_version);
227 dp->ricp_v6.dp_protocol_minor_mask =
228 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
229 dp->ricp_v6.dp_ack_seq =
230 cpu_to_be64(rds_ib_piggyb_ack(ic));
231 dp->ricp_v6.dp_cmn.ricpc_dp_toss = conn->c_tos;
233 conn_param->private_data = &dp->ricp_v6;
234 conn_param->private_data_len = sizeof(dp->ricp_v6);
236 dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3];
237 dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3];
238 dp->ricp_v4.dp_protocol_major =
239 RDS_PROTOCOL_MAJOR(protocol_version);
240 dp->ricp_v4.dp_protocol_minor =
241 RDS_PROTOCOL_MINOR(protocol_version);
242 dp->ricp_v4.dp_protocol_minor_mask =
243 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
244 dp->ricp_v4.dp_ack_seq =
245 cpu_to_be64(rds_ib_piggyb_ack(ic));
246 dp->ricp_v4.dp_cmn.ricpc_dp_toss = conn->c_tos;
248 conn_param->private_data = &dp->ricp_v4;
249 conn_param->private_data_len = sizeof(dp->ricp_v4);
252 /* Advertise flow control */
254 unsigned int credits;
256 credits = IB_GET_POST_CREDITS
257 (atomic_read(&ic->i_credits));
259 dp->ricp_v6.dp_credit = cpu_to_be32(credits);
261 dp->ricp_v4.dp_credit = cpu_to_be32(credits);
262 atomic_sub(IB_SET_POST_CREDITS(credits),
268 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
270 rdsdebug("event %u (%s) data %p\n",
271 event->event, ib_event_msg(event->event), data);
274 /* Plucking the oldest entry from the ring can be done concurrently with
275 * the thread refilling the ring. Each ring operation is protected by
276 * spinlocks and the transient state of refilling doesn't change the
277 * recording of which entry is oldest.
279 * This relies on IB only calling one cq comp_handler for each cq so that
280 * there will only be one caller of rds_recv_incoming() per RDS connection.
282 static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
284 struct rds_connection *conn = context;
285 struct rds_ib_connection *ic = conn->c_transport_data;
287 rdsdebug("conn %p cq %p\n", conn, cq);
289 rds_ib_stats_inc(s_ib_evt_handler_call);
291 tasklet_schedule(&ic->i_recv_tasklet);
294 static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
300 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
301 for (i = 0; i < nr; i++) {
303 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
304 (unsigned long long)wc->wr_id, wc->status,
305 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
307 if (wc->wr_id <= ic->i_send_ring.w_nr ||
308 wc->wr_id == RDS_IB_ACK_WR_ID)
309 rds_ib_send_cqe_handler(ic, wc);
311 rds_ib_mr_cqe_handler(ic, wc);
317 static void rds_ib_tasklet_fn_send(unsigned long data)
319 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
320 struct rds_connection *conn = ic->conn;
322 rds_ib_stats_inc(s_ib_tasklet_call);
324 /* if cq has been already reaped, ignore incoming cq event */
325 if (atomic_read(&ic->i_cq_quiesce))
328 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
329 ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
330 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
332 if (rds_conn_up(conn) &&
333 (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
334 test_bit(0, &conn->c_map_queued)))
335 rds_send_xmit(&ic->conn->c_path[0]);
338 static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
340 struct rds_ib_ack_state *ack_state)
345 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
346 for (i = 0; i < nr; i++) {
348 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
349 (unsigned long long)wc->wr_id, wc->status,
350 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
352 rds_ib_recv_cqe_handler(ic, wc, ack_state);
357 static void rds_ib_tasklet_fn_recv(unsigned long data)
359 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
360 struct rds_connection *conn = ic->conn;
361 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
362 struct rds_ib_ack_state state;
367 rds_ib_stats_inc(s_ib_tasklet_call);
369 /* if cq has been already reaped, ignore incoming cq event */
370 if (atomic_read(&ic->i_cq_quiesce))
373 memset(&state, 0, sizeof(state));
374 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
375 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
376 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
378 if (state.ack_next_valid)
379 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
380 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
381 rds_send_drop_acked(conn, state.ack_recv, NULL);
382 ic->i_ack_recv = state.ack_recv;
385 if (rds_conn_up(conn))
386 rds_ib_attempt_ack(ic);
389 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
391 struct rds_connection *conn = data;
392 struct rds_ib_connection *ic = conn->c_transport_data;
394 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
395 ib_event_msg(event->event));
397 switch (event->event) {
398 case IB_EVENT_COMM_EST:
399 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
402 rdsdebug("Fatal QP Event %u (%s) - connection %pI6c->%pI6c, reconnecting\n",
403 event->event, ib_event_msg(event->event),
404 &conn->c_laddr, &conn->c_faddr);
410 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
412 struct rds_connection *conn = context;
413 struct rds_ib_connection *ic = conn->c_transport_data;
415 rdsdebug("conn %p cq %p\n", conn, cq);
417 rds_ib_stats_inc(s_ib_evt_handler_call);
419 tasklet_schedule(&ic->i_send_tasklet);
422 static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
424 int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
425 int index = rds_ibdev->dev->num_comp_vectors - 1;
428 for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
429 if (rds_ibdev->vector_load[i] < min) {
431 min = rds_ibdev->vector_load[i];
435 rds_ibdev->vector_load[index]++;
439 static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
441 rds_ibdev->vector_load[index]--;
444 /* Allocate DMA coherent memory to be used to store struct rds_header for
445 * sending/receiving packets. The pointers to the DMA memory and the
446 * associated DMA addresses are stored in two arrays.
448 * @ibdev: the IB device
449 * @pool: the DMA memory pool
450 * @dma_addrs: pointer to the array for storing DMA addresses
451 * @num_hdrs: number of headers to allocate
453 * It returns the pointer to the array storing the DMA memory pointers. On
454 * error, NULL pointer is returned.
456 struct rds_header **rds_dma_hdrs_alloc(struct ib_device *ibdev,
457 struct dma_pool *pool,
458 dma_addr_t **dma_addrs, u32 num_hdrs)
460 struct rds_header **hdrs;
461 dma_addr_t *hdr_daddrs;
464 hdrs = kvmalloc_node(sizeof(*hdrs) * num_hdrs, GFP_KERNEL,
465 ibdev_to_node(ibdev));
469 hdr_daddrs = kvmalloc_node(sizeof(*hdr_daddrs) * num_hdrs, GFP_KERNEL,
470 ibdev_to_node(ibdev));
476 for (i = 0; i < num_hdrs; i++) {
477 hdrs[i] = dma_pool_zalloc(pool, GFP_KERNEL, &hdr_daddrs[i]);
479 rds_dma_hdrs_free(pool, hdrs, hdr_daddrs, i);
484 *dma_addrs = hdr_daddrs;
488 /* Free the DMA memory used to store struct rds_header.
490 * @pool: the DMA memory pool
491 * @hdrs: pointer to the array storing DMA memory pointers
492 * @dma_addrs: pointer to the array storing DMA addresses
493 * @num_hdars: number of headers to free.
495 void rds_dma_hdrs_free(struct dma_pool *pool, struct rds_header **hdrs,
496 dma_addr_t *dma_addrs, u32 num_hdrs)
500 for (i = 0; i < num_hdrs; i++)
501 dma_pool_free(pool, hdrs[i], dma_addrs[i]);
507 * This needs to be very careful to not leave IS_ERR pointers around for
508 * cleanup to trip over.
510 static int rds_ib_setup_qp(struct rds_connection *conn)
512 struct rds_ib_connection *ic = conn->c_transport_data;
513 struct ib_device *dev = ic->i_cm_id->device;
514 struct ib_qp_init_attr attr;
515 struct ib_cq_init_attr cq_attr = {};
516 struct rds_ib_device *rds_ibdev;
517 unsigned long max_wrs;
518 int ret, fr_queue_space;
519 struct dma_pool *pool;
522 * It's normal to see a null device if an incoming connection races
523 * with device removal, so we don't print a warning.
525 rds_ibdev = rds_ib_get_client_data(dev);
529 /* The fr_queue_space is currently set to 512, to add extra space on
530 * completion queue and send queue. This extra space is used for FRWR
531 * registration and invalidation work requests
533 fr_queue_space = RDS_IB_DEFAULT_FR_WR;
535 /* add the conn now so that connection establishment has the dev */
536 rds_ib_add_conn(rds_ibdev, conn);
538 max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_send_wr + 1 ?
539 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_send_wr;
540 if (ic->i_send_ring.w_nr != max_wrs)
541 rds_ib_ring_resize(&ic->i_send_ring, max_wrs);
543 max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_recv_wr + 1 ?
544 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_recv_wr;
545 if (ic->i_recv_ring.w_nr != max_wrs)
546 rds_ib_ring_resize(&ic->i_recv_ring, max_wrs);
548 /* Protection domain and memory range */
549 ic->i_pd = rds_ibdev->pd;
551 ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
552 cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
553 cq_attr.comp_vector = ic->i_scq_vector;
554 ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
555 rds_ib_cq_event_handler, conn,
557 if (IS_ERR(ic->i_send_cq)) {
558 ret = PTR_ERR(ic->i_send_cq);
559 ic->i_send_cq = NULL;
560 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
561 rdsdebug("ib_create_cq send failed: %d\n", ret);
565 ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
566 cq_attr.cqe = ic->i_recv_ring.w_nr;
567 cq_attr.comp_vector = ic->i_rcq_vector;
568 ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
569 rds_ib_cq_event_handler, conn,
571 if (IS_ERR(ic->i_recv_cq)) {
572 ret = PTR_ERR(ic->i_recv_cq);
573 ic->i_recv_cq = NULL;
574 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
575 rdsdebug("ib_create_cq recv failed: %d\n", ret);
579 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
581 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
585 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
587 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
591 /* XXX negotiate max send/recv with remote? */
592 memset(&attr, 0, sizeof(attr));
593 attr.event_handler = rds_ib_qp_event_handler;
594 attr.qp_context = conn;
595 /* + 1 to allow for the single ack message */
596 attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
597 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
598 attr.cap.max_send_sge = rds_ibdev->max_sge;
599 attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
600 attr.sq_sig_type = IB_SIGNAL_REQ_WR;
601 attr.qp_type = IB_QPT_RC;
602 attr.send_cq = ic->i_send_cq;
603 attr.recv_cq = ic->i_recv_cq;
606 * XXX this can fail if max_*_wr is too large? Are we supposed
607 * to back off until we get a value that the hardware can support?
609 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
611 rdsdebug("rdma_create_qp failed: %d\n", ret);
615 pool = rds_ibdev->rid_hdrs_pool;
616 ic->i_send_hdrs = rds_dma_hdrs_alloc(dev, pool, &ic->i_send_hdrs_dma,
617 ic->i_send_ring.w_nr);
618 if (!ic->i_send_hdrs) {
620 rdsdebug("DMA send hdrs alloc failed\n");
624 ic->i_recv_hdrs = rds_dma_hdrs_alloc(dev, pool, &ic->i_recv_hdrs_dma,
625 ic->i_recv_ring.w_nr);
626 if (!ic->i_recv_hdrs) {
628 rdsdebug("DMA recv hdrs alloc failed\n");
629 goto send_hdrs_dma_out;
632 ic->i_ack = dma_pool_zalloc(pool, GFP_KERNEL,
636 rdsdebug("DMA ack header alloc failed\n");
637 goto recv_hdrs_dma_out;
640 ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
641 ic->i_send_ring.w_nr),
645 rdsdebug("send allocation failed\n");
649 ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
650 ic->i_recv_ring.w_nr),
654 rdsdebug("recv allocation failed\n");
658 rds_ib_recv_init_ack(ic);
660 rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
661 ic->i_send_cq, ic->i_recv_cq);
669 dma_pool_free(pool, ic->i_ack, ic->i_ack_dma);
673 rds_dma_hdrs_free(pool, ic->i_recv_hdrs, ic->i_recv_hdrs_dma,
674 ic->i_recv_ring.w_nr);
675 ic->i_recv_hdrs = NULL;
676 ic->i_recv_hdrs_dma = NULL;
679 rds_dma_hdrs_free(pool, ic->i_send_hdrs, ic->i_send_hdrs_dma,
680 ic->i_send_ring.w_nr);
681 ic->i_send_hdrs = NULL;
682 ic->i_send_hdrs_dma = NULL;
685 rdma_destroy_qp(ic->i_cm_id);
687 ib_destroy_cq(ic->i_recv_cq);
688 ic->i_recv_cq = NULL;
690 ib_destroy_cq(ic->i_send_cq);
691 ic->i_send_cq = NULL;
693 rds_ib_remove_conn(rds_ibdev, conn);
695 rds_ib_dev_put(rds_ibdev);
700 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
702 const union rds_ib_conn_priv *dp = event->param.conn.private_data;
703 u8 data_len, major, minor;
709 * rdma_cm private data is odd - when there is any private data in the
710 * request, we will be given a pretty large buffer without telling us the
711 * original size. The only way to tell the difference is by looking at
712 * the contents, which are initialized to zero.
713 * If the protocol version fields aren't set, this is a connection attempt
714 * from an older version. This could be 3.0 or 2.0 - we can't tell.
715 * We really should have changed this for OFED 1.3 :-(
718 /* Be paranoid. RDS always has privdata */
719 if (!event->param.conn.private_data_len) {
720 printk(KERN_NOTICE "RDS incoming connection has no private data, "
726 data_len = sizeof(struct rds6_ib_connect_private);
727 major = dp->ricp_v6.dp_protocol_major;
728 minor = dp->ricp_v6.dp_protocol_minor;
729 mask = dp->ricp_v6.dp_protocol_minor_mask;
731 data_len = sizeof(struct rds_ib_connect_private);
732 major = dp->ricp_v4.dp_protocol_major;
733 minor = dp->ricp_v4.dp_protocol_minor;
734 mask = dp->ricp_v4.dp_protocol_minor_mask;
737 /* Even if len is crap *now* I still want to check it. -ASG */
738 if (event->param.conn.private_data_len < data_len || major == 0)
739 return RDS_PROTOCOL_4_0;
741 common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS;
742 if (major == 4 && common) {
743 version = RDS_PROTOCOL_4_0;
744 while ((common >>= 1) != 0)
746 } else if (RDS_PROTOCOL_COMPAT_VERSION ==
747 RDS_PROTOCOL(major, minor)) {
748 version = RDS_PROTOCOL_COMPAT_VERSION;
751 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
752 &dp->ricp_v6.dp_saddr, major, minor);
754 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
755 &dp->ricp_v4.dp_saddr, major, minor);
760 #if IS_ENABLED(CONFIG_IPV6)
761 /* Given an IPv6 address, find the net_device which hosts that address and
762 * return its index. This is used by the rds_ib_cm_handle_connect() code to
763 * find the interface index of where an incoming request comes from when
764 * the request is using a link local address.
766 * Note one problem in this search. It is possible that two interfaces have
767 * the same link local address. Unfortunately, this cannot be solved unless
768 * the underlying layer gives us the interface which an incoming RDMA connect
769 * request comes from.
771 static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
773 struct net_device *dev;
777 for_each_netdev_rcu(net, dev) {
778 if (ipv6_chk_addr(net, addr, dev, 1)) {
789 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
790 struct rdma_cm_event *event, bool isv6)
792 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
793 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
794 const struct rds_ib_conn_priv_cmn *dp_cmn;
795 struct rds_connection *conn = NULL;
796 struct rds_ib_connection *ic = NULL;
797 struct rdma_conn_param conn_param;
798 const union rds_ib_conn_priv *dp;
799 union rds_ib_conn_priv dp_rep;
800 struct in6_addr s_mapped_addr;
801 struct in6_addr d_mapped_addr;
802 const struct in6_addr *saddr6;
803 const struct in6_addr *daddr6;
809 /* Check whether the remote protocol version matches ours. */
810 version = rds_ib_protocol_compatible(event, isv6);
812 err = RDS_RDMA_REJ_INCOMPAT;
816 dp = event->param.conn.private_data;
818 #if IS_ENABLED(CONFIG_IPV6)
819 dp_cmn = &dp->ricp_v6.dp_cmn;
820 saddr6 = &dp->ricp_v6.dp_saddr;
821 daddr6 = &dp->ricp_v6.dp_daddr;
822 /* If either address is link local, need to find the
823 * interface index in order to create a proper RDS
826 if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) {
827 /* Using init_net for now .. */
828 ifindex = __rds_find_ifindex(&init_net, daddr6);
829 /* No index found... Need to bail out. */
834 } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) {
835 /* Use our address to find the correct index. */
836 ifindex = __rds_find_ifindex(&init_net, daddr6);
837 /* No index found... Need to bail out. */
848 dp_cmn = &dp->ricp_v4.dp_cmn;
849 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr);
850 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
851 saddr6 = &s_mapped_addr;
852 daddr6 = &d_mapped_addr;
855 rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid 0x%llx, tos:%d\n",
856 saddr6, daddr6, RDS_PROTOCOL_MAJOR(version),
857 RDS_PROTOCOL_MINOR(version),
858 (unsigned long long)be64_to_cpu(lguid),
859 (unsigned long long)be64_to_cpu(fguid), dp_cmn->ricpc_dp_toss);
861 /* RDS/IB is not currently netns aware, thus init_net */
862 conn = rds_conn_create(&init_net, daddr6, saddr6,
863 &rds_ib_transport, dp_cmn->ricpc_dp_toss,
864 GFP_KERNEL, ifindex);
866 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
872 * The connection request may occur while the
873 * previous connection exist, e.g. in case of failover.
874 * But as connections may be initiated simultaneously
875 * by both hosts, we have a random backoff mechanism -
876 * see the comment above rds_queue_reconnect()
878 mutex_lock(&conn->c_cm_lock);
879 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
880 if (rds_conn_state(conn) == RDS_CONN_UP) {
881 rdsdebug("incoming connect while connecting\n");
883 rds_ib_stats_inc(s_ib_listen_closed_stale);
885 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
886 /* Wait and see - our connect may still be succeeding */
887 rds_ib_stats_inc(s_ib_connect_raced);
892 ic = conn->c_transport_data;
894 rds_ib_set_protocol(conn, version);
895 rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
897 /* If the peer gave us the last packet it saw, process this as if
898 * we had received a regular ACK. */
899 if (dp_cmn->ricpc_ack_seq)
900 rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq),
903 BUG_ON(cm_id->context);
907 cm_id->context = conn;
909 /* We got halfway through setting up the ib_connection, if we
910 * fail now, we have to take the long route out of this mess. */
913 err = rds_ib_setup_qp(conn);
915 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
919 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
920 event->param.conn.responder_resources,
921 event->param.conn.initiator_depth, isv6);
923 /* rdma_accept() calls rdma_reject() internally if it fails */
924 if (rdma_accept(cm_id, &conn_param))
925 rds_ib_conn_error(conn, "rdma_accept failed\n");
929 mutex_unlock(&conn->c_cm_lock);
931 rdma_reject(cm_id, &err, sizeof(int),
932 IB_CM_REJ_CONSUMER_DEFINED);
937 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
939 struct rds_connection *conn = cm_id->context;
940 struct rds_ib_connection *ic = conn->c_transport_data;
941 struct rdma_conn_param conn_param;
942 union rds_ib_conn_priv dp;
945 /* If the peer doesn't do protocol negotiation, we must
946 * default to RDSv3.0 */
947 rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1);
948 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
950 ret = rds_ib_setup_qp(conn);
952 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
956 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
957 conn->c_proposed_version,
958 UINT_MAX, UINT_MAX, isv6);
959 ret = rdma_connect(cm_id, &conn_param);
961 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
964 /* Beware - returning non-zero tells the rdma_cm to destroy
965 * the cm_id. We should certainly not do it as long as we still
966 * "own" the cm_id. */
968 if (ic->i_cm_id == cm_id)
971 ic->i_active_side = true;
975 int rds_ib_conn_path_connect(struct rds_conn_path *cp)
977 struct rds_connection *conn = cp->cp_conn;
978 struct sockaddr_storage src, dest;
979 rdma_cm_event_handler handler;
980 struct rds_ib_connection *ic;
983 ic = conn->c_transport_data;
985 /* XXX I wonder what affect the port space has */
986 /* delegate cm event handler to rdma_transport */
987 #if IS_ENABLED(CONFIG_IPV6)
989 handler = rds6_rdma_cm_event_handler;
992 handler = rds_rdma_cm_event_handler;
993 ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
994 RDMA_PS_TCP, IB_QPT_RC);
995 if (IS_ERR(ic->i_cm_id)) {
996 ret = PTR_ERR(ic->i_cm_id);
998 rdsdebug("rdma_create_id() failed: %d\n", ret);
1002 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
1004 if (ipv6_addr_v4mapped(&conn->c_faddr)) {
1005 struct sockaddr_in *sin;
1007 sin = (struct sockaddr_in *)&src;
1008 sin->sin_family = AF_INET;
1009 sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
1012 sin = (struct sockaddr_in *)&dest;
1013 sin->sin_family = AF_INET;
1014 sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
1015 sin->sin_port = htons(RDS_PORT);
1017 struct sockaddr_in6 *sin6;
1019 sin6 = (struct sockaddr_in6 *)&src;
1020 sin6->sin6_family = AF_INET6;
1021 sin6->sin6_addr = conn->c_laddr;
1022 sin6->sin6_port = 0;
1023 sin6->sin6_scope_id = conn->c_dev_if;
1025 sin6 = (struct sockaddr_in6 *)&dest;
1026 sin6->sin6_family = AF_INET6;
1027 sin6->sin6_addr = conn->c_faddr;
1028 sin6->sin6_port = htons(RDS_CM_PORT);
1029 sin6->sin6_scope_id = conn->c_dev_if;
1032 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
1033 (struct sockaddr *)&dest,
1034 RDS_RDMA_RESOLVE_TIMEOUT_MS);
1036 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
1038 rdma_destroy_id(ic->i_cm_id);
1047 * This is so careful about only cleaning up resources that were built up
1048 * so that it can be called at any point during startup. In fact it
1049 * can be called multiple times for a given connection.
1051 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
1053 struct rds_connection *conn = cp->cp_conn;
1054 struct rds_ib_connection *ic = conn->c_transport_data;
1057 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
1058 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
1059 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
1062 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
1063 err = rdma_disconnect(ic->i_cm_id);
1065 /* Actually this may happen quite frequently, when
1066 * an outgoing connect raced with an incoming connect.
1068 rdsdebug("failed to disconnect, cm: %p err %d\n",
1072 /* kick off "flush_worker" for all pools in order to reap
1073 * all FRMR registrations that are still marked "FRMR_IS_INUSE"
1078 * We want to wait for tx and rx completion to finish
1079 * before we tear down the connection, but we have to be
1080 * careful not to get stuck waiting on a send ring that
1081 * only has unsignaled sends in it. We've shutdown new
1082 * sends before getting here so by waiting for signaled
1083 * sends to complete we're ensured that there will be no
1084 * more tx processing.
1086 wait_event(rds_ib_ring_empty_wait,
1087 rds_ib_ring_empty(&ic->i_recv_ring) &&
1088 (atomic_read(&ic->i_signaled_sends) == 0) &&
1089 (atomic_read(&ic->i_fastreg_inuse_count) == 0) &&
1090 (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
1091 tasklet_kill(&ic->i_send_tasklet);
1092 tasklet_kill(&ic->i_recv_tasklet);
1094 atomic_set(&ic->i_cq_quiesce, 1);
1096 /* first destroy the ib state that generates callbacks */
1097 if (ic->i_cm_id->qp)
1098 rdma_destroy_qp(ic->i_cm_id);
1099 if (ic->i_send_cq) {
1101 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1102 ib_destroy_cq(ic->i_send_cq);
1105 if (ic->i_recv_cq) {
1107 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1108 ib_destroy_cq(ic->i_recv_cq);
1111 if (ic->rds_ibdev) {
1112 struct dma_pool *pool;
1114 pool = ic->rds_ibdev->rid_hdrs_pool;
1116 /* then free the resources that ib callbacks use */
1117 if (ic->i_send_hdrs) {
1118 rds_dma_hdrs_free(pool, ic->i_send_hdrs,
1119 ic->i_send_hdrs_dma,
1120 ic->i_send_ring.w_nr);
1121 ic->i_send_hdrs = NULL;
1122 ic->i_send_hdrs_dma = NULL;
1125 if (ic->i_recv_hdrs) {
1126 rds_dma_hdrs_free(pool, ic->i_recv_hdrs,
1127 ic->i_recv_hdrs_dma,
1128 ic->i_recv_ring.w_nr);
1129 ic->i_recv_hdrs = NULL;
1130 ic->i_recv_hdrs_dma = NULL;
1134 dma_pool_free(pool, ic->i_ack, ic->i_ack_dma);
1138 WARN_ON(ic->i_send_hdrs);
1139 WARN_ON(ic->i_send_hdrs_dma);
1140 WARN_ON(ic->i_recv_hdrs);
1141 WARN_ON(ic->i_recv_hdrs_dma);
1146 rds_ib_send_clear_ring(ic);
1148 rds_ib_recv_clear_ring(ic);
1150 rdma_destroy_id(ic->i_cm_id);
1153 * Move connection back to the nodev list.
1156 rds_ib_remove_conn(ic->rds_ibdev, conn);
1160 ic->i_send_cq = NULL;
1161 ic->i_recv_cq = NULL;
1163 BUG_ON(ic->rds_ibdev);
1165 /* Clear pending transmit */
1166 if (ic->i_data_op) {
1167 struct rds_message *rm;
1169 rm = container_of(ic->i_data_op, struct rds_message, data);
1170 rds_message_put(rm);
1171 ic->i_data_op = NULL;
1174 /* Clear the ACK state */
1175 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1176 #ifdef KERNEL_HAS_ATOMIC64
1177 atomic64_set(&ic->i_ack_next, 0);
1183 /* Clear flow control state */
1185 atomic_set(&ic->i_credits, 0);
1187 /* Re-init rings, but retain sizes. */
1188 rds_ib_ring_init(&ic->i_send_ring, ic->i_send_ring.w_nr);
1189 rds_ib_ring_init(&ic->i_recv_ring, ic->i_recv_ring.w_nr);
1192 rds_inc_put(&ic->i_ibinc->ii_inc);
1200 ic->i_active_side = false;
1203 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1205 struct rds_ib_connection *ic;
1206 unsigned long flags;
1210 ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1214 ret = rds_ib_recv_alloc_caches(ic, gfp);
1220 INIT_LIST_HEAD(&ic->ib_node);
1221 tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1223 tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1225 mutex_init(&ic->i_recv_mutex);
1226 #ifndef KERNEL_HAS_ATOMIC64
1227 spin_lock_init(&ic->i_ack_lock);
1229 atomic_set(&ic->i_signaled_sends, 0);
1230 atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
1233 * rds_ib_conn_shutdown() waits for these to be emptied so they
1234 * must be initialized before it can be called.
1236 rds_ib_ring_init(&ic->i_send_ring, 0);
1237 rds_ib_ring_init(&ic->i_recv_ring, 0);
1240 conn->c_transport_data = ic;
1242 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1243 list_add_tail(&ic->ib_node, &ib_nodev_conns);
1244 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1247 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1252 * Free a connection. Connection must be shut down and not set for reconnect.
1254 void rds_ib_conn_free(void *arg)
1256 struct rds_ib_connection *ic = arg;
1257 spinlock_t *lock_ptr;
1259 rdsdebug("ic %p\n", ic);
1262 * Conn is either on a dev's list or on the nodev list.
1263 * A race with shutdown() or connect() would cause problems
1264 * (since rds_ibdev would change) but that should never happen.
1266 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1268 spin_lock_irq(lock_ptr);
1269 list_del(&ic->ib_node);
1270 spin_unlock_irq(lock_ptr);
1272 rds_ib_recv_free_caches(ic);
1279 * An error occurred on the connection
1282 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1286 rds_conn_drop(conn);