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>
41 #include "rds_single_path.h"
47 * Set the selected protocol version
49 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
51 conn->c_version = version;
57 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
59 struct rds_ib_connection *ic = conn->c_transport_data;
61 if (rds_ib_sysctl_flow_control && credits != 0) {
62 /* We're doing flow control */
64 rds_ib_send_add_credits(conn, credits);
71 * Tune RNR behavior. Without flow control, we use a rather
72 * low timeout, but not the absolute minimum - this should
75 * We already set the RNR retry count to 7 (which is the
76 * smallest infinite number :-) above.
77 * If flow control is off, we want to change this back to 0
78 * so that we learn quickly when our credit accounting is
81 * Caller passes in a qp_attr pointer - don't waste stack spacv
82 * by allocation this twice.
85 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
89 attr->min_rnr_timer = IB_RNR_TIMER_000_32;
90 ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
92 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
96 * Connection established.
97 * We get here for both outgoing and incoming connection.
99 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
101 struct rds_ib_connection *ic = conn->c_transport_data;
102 const union rds_ib_conn_priv *dp = NULL;
103 struct ib_qp_attr qp_attr;
110 dp = event->param.conn.private_data;
112 if (event->param.conn.private_data_len >=
113 sizeof(struct rds6_ib_connect_private)) {
114 major = dp->ricp_v6.dp_protocol_major;
115 minor = dp->ricp_v6.dp_protocol_minor;
116 credit = dp->ricp_v6.dp_credit;
117 /* dp structure start is not guaranteed to be 8 bytes
118 * aligned. Since dp_ack_seq is 64-bit extended load
119 * operations can be used so go through get_unaligned
120 * to avoid unaligned errors.
122 ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq);
124 } else if (event->param.conn.private_data_len >=
125 sizeof(struct rds_ib_connect_private)) {
126 major = dp->ricp_v4.dp_protocol_major;
127 minor = dp->ricp_v4.dp_protocol_minor;
128 credit = dp->ricp_v4.dp_credit;
129 ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq);
132 /* make sure it isn't empty data */
134 rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor));
135 rds_ib_set_flow_control(conn, be32_to_cpu(credit));
138 if (conn->c_version < RDS_PROTOCOL_VERSION) {
139 if (conn->c_version != RDS_PROTOCOL_COMPAT_VERSION) {
140 pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n",
141 &conn->c_laddr, &conn->c_faddr,
142 RDS_PROTOCOL_MAJOR(conn->c_version),
143 RDS_PROTOCOL_MINOR(conn->c_version));
144 rds_conn_destroy(conn);
149 pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c,%d> version %u.%u%s\n",
150 ic->i_active_side ? "Active" : "Passive",
151 &conn->c_laddr, &conn->c_faddr, conn->c_tos,
152 RDS_PROTOCOL_MAJOR(conn->c_version),
153 RDS_PROTOCOL_MINOR(conn->c_version),
154 ic->i_flowctl ? ", flow control" : "");
156 /* receive sl from the peer */
157 ic->i_sl = ic->i_cm_id->route.path_rec->sl;
159 atomic_set(&ic->i_cq_quiesce, 0);
161 /* Init rings and fill recv. this needs to wait until protocol
162 * negotiation is complete, since ring layout is different
165 rds_ib_send_init_ring(ic);
166 rds_ib_recv_init_ring(ic);
167 /* Post receive buffers - as a side effect, this will update
168 * the posted credit count. */
169 rds_ib_recv_refill(conn, 1, GFP_KERNEL);
171 /* Tune RNR behavior */
172 rds_ib_tune_rnr(ic, &qp_attr);
174 qp_attr.qp_state = IB_QPS_RTS;
175 err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
177 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
179 /* update ib_device with this local ipaddr */
180 err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr);
182 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
185 /* If the peer gave us the last packet it saw, process this as if
186 * we had received a regular ACK. */
189 rds_send_drop_acked(conn, be64_to_cpu(ack_seq),
193 conn->c_proposed_version = conn->c_version;
194 rds_connect_complete(conn);
197 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
198 struct rdma_conn_param *conn_param,
199 union rds_ib_conn_priv *dp,
200 u32 protocol_version,
201 u32 max_responder_resources,
202 u32 max_initiator_depth,
205 struct rds_ib_connection *ic = conn->c_transport_data;
206 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
208 memset(conn_param, 0, sizeof(struct rdma_conn_param));
210 conn_param->responder_resources =
211 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
212 conn_param->initiator_depth =
213 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
214 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
215 conn_param->rnr_retry_count = 7;
218 memset(dp, 0, sizeof(*dp));
220 dp->ricp_v6.dp_saddr = conn->c_laddr;
221 dp->ricp_v6.dp_daddr = conn->c_faddr;
222 dp->ricp_v6.dp_protocol_major =
223 RDS_PROTOCOL_MAJOR(protocol_version);
224 dp->ricp_v6.dp_protocol_minor =
225 RDS_PROTOCOL_MINOR(protocol_version);
226 dp->ricp_v6.dp_protocol_minor_mask =
227 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
228 dp->ricp_v6.dp_ack_seq =
229 cpu_to_be64(rds_ib_piggyb_ack(ic));
230 dp->ricp_v6.dp_cmn.ricpc_dp_toss = conn->c_tos;
232 conn_param->private_data = &dp->ricp_v6;
233 conn_param->private_data_len = sizeof(dp->ricp_v6);
235 dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3];
236 dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3];
237 dp->ricp_v4.dp_protocol_major =
238 RDS_PROTOCOL_MAJOR(protocol_version);
239 dp->ricp_v4.dp_protocol_minor =
240 RDS_PROTOCOL_MINOR(protocol_version);
241 dp->ricp_v4.dp_protocol_minor_mask =
242 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
243 dp->ricp_v4.dp_ack_seq =
244 cpu_to_be64(rds_ib_piggyb_ack(ic));
245 dp->ricp_v4.dp_cmn.ricpc_dp_toss = conn->c_tos;
247 conn_param->private_data = &dp->ricp_v4;
248 conn_param->private_data_len = sizeof(dp->ricp_v4);
251 /* Advertise flow control */
253 unsigned int credits;
255 credits = IB_GET_POST_CREDITS
256 (atomic_read(&ic->i_credits));
258 dp->ricp_v6.dp_credit = cpu_to_be32(credits);
260 dp->ricp_v4.dp_credit = cpu_to_be32(credits);
261 atomic_sub(IB_SET_POST_CREDITS(credits),
267 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
269 rdsdebug("event %u (%s) data %p\n",
270 event->event, ib_event_msg(event->event), data);
273 /* Plucking the oldest entry from the ring can be done concurrently with
274 * the thread refilling the ring. Each ring operation is protected by
275 * spinlocks and the transient state of refilling doesn't change the
276 * recording of which entry is oldest.
278 * This relies on IB only calling one cq comp_handler for each cq so that
279 * there will only be one caller of rds_recv_incoming() per RDS connection.
281 static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
283 struct rds_connection *conn = context;
284 struct rds_ib_connection *ic = conn->c_transport_data;
286 rdsdebug("conn %p cq %p\n", conn, cq);
288 rds_ib_stats_inc(s_ib_evt_handler_call);
290 tasklet_schedule(&ic->i_recv_tasklet);
293 static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
299 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
300 for (i = 0; i < nr; i++) {
302 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
303 (unsigned long long)wc->wr_id, wc->status,
304 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
306 if (wc->wr_id <= ic->i_send_ring.w_nr ||
307 wc->wr_id == RDS_IB_ACK_WR_ID)
308 rds_ib_send_cqe_handler(ic, wc);
310 rds_ib_mr_cqe_handler(ic, wc);
316 static void rds_ib_tasklet_fn_send(unsigned long data)
318 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
319 struct rds_connection *conn = ic->conn;
321 rds_ib_stats_inc(s_ib_tasklet_call);
323 /* if cq has been already reaped, ignore incoming cq event */
324 if (atomic_read(&ic->i_cq_quiesce))
327 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
328 ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
329 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
331 if (rds_conn_up(conn) &&
332 (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
333 test_bit(0, &conn->c_map_queued)))
334 rds_send_xmit(&ic->conn->c_path[0]);
337 static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
339 struct rds_ib_ack_state *ack_state)
344 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
345 for (i = 0; i < nr; i++) {
347 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
348 (unsigned long long)wc->wr_id, wc->status,
349 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
351 rds_ib_recv_cqe_handler(ic, wc, ack_state);
356 static void rds_ib_tasklet_fn_recv(unsigned long data)
358 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
359 struct rds_connection *conn = ic->conn;
360 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
361 struct rds_ib_ack_state state;
366 rds_ib_stats_inc(s_ib_tasklet_call);
368 /* if cq has been already reaped, ignore incoming cq event */
369 if (atomic_read(&ic->i_cq_quiesce))
372 memset(&state, 0, sizeof(state));
373 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
374 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
375 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
377 if (state.ack_next_valid)
378 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
379 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
380 rds_send_drop_acked(conn, state.ack_recv, NULL);
381 ic->i_ack_recv = state.ack_recv;
384 if (rds_conn_up(conn))
385 rds_ib_attempt_ack(ic);
388 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
390 struct rds_connection *conn = data;
391 struct rds_ib_connection *ic = conn->c_transport_data;
393 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
394 ib_event_msg(event->event));
396 switch (event->event) {
397 case IB_EVENT_COMM_EST:
398 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
401 rdsdebug("Fatal QP Event %u (%s) - connection %pI6c->%pI6c, reconnecting\n",
402 event->event, ib_event_msg(event->event),
403 &conn->c_laddr, &conn->c_faddr);
409 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
411 struct rds_connection *conn = context;
412 struct rds_ib_connection *ic = conn->c_transport_data;
414 rdsdebug("conn %p cq %p\n", conn, cq);
416 rds_ib_stats_inc(s_ib_evt_handler_call);
418 tasklet_schedule(&ic->i_send_tasklet);
421 static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
423 int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
424 int index = rds_ibdev->dev->num_comp_vectors - 1;
427 for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
428 if (rds_ibdev->vector_load[i] < min) {
430 min = rds_ibdev->vector_load[i];
434 rds_ibdev->vector_load[index]++;
438 static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
440 rds_ibdev->vector_load[index]--;
443 /* Allocate DMA coherent memory to be used to store struct rds_header for
444 * sending/receiving packets. The pointers to the DMA memory and the
445 * associated DMA addresses are stored in two arrays.
447 * @ibdev: the IB device
448 * @pool: the DMA memory pool
449 * @dma_addrs: pointer to the array for storing DMA addresses
450 * @num_hdrs: number of headers to allocate
452 * It returns the pointer to the array storing the DMA memory pointers. On
453 * error, NULL pointer is returned.
455 struct rds_header **rds_dma_hdrs_alloc(struct ib_device *ibdev,
456 struct dma_pool *pool,
457 dma_addr_t **dma_addrs, u32 num_hdrs)
459 struct rds_header **hdrs;
460 dma_addr_t *hdr_daddrs;
463 hdrs = kvmalloc_node(sizeof(*hdrs) * num_hdrs, GFP_KERNEL,
464 ibdev_to_node(ibdev));
468 hdr_daddrs = kvmalloc_node(sizeof(*hdr_daddrs) * num_hdrs, GFP_KERNEL,
469 ibdev_to_node(ibdev));
475 for (i = 0; i < num_hdrs; i++) {
476 hdrs[i] = dma_pool_zalloc(pool, GFP_KERNEL, &hdr_daddrs[i]);
478 rds_dma_hdrs_free(pool, hdrs, hdr_daddrs, i);
483 *dma_addrs = hdr_daddrs;
487 /* Free the DMA memory used to store struct rds_header.
489 * @pool: the DMA memory pool
490 * @hdrs: pointer to the array storing DMA memory pointers
491 * @dma_addrs: pointer to the array storing DMA addresses
492 * @num_hdars: number of headers to free.
494 void rds_dma_hdrs_free(struct dma_pool *pool, struct rds_header **hdrs,
495 dma_addr_t *dma_addrs, u32 num_hdrs)
499 for (i = 0; i < num_hdrs; i++)
500 dma_pool_free(pool, hdrs[i], dma_addrs[i]);
506 * This needs to be very careful to not leave IS_ERR pointers around for
507 * cleanup to trip over.
509 static int rds_ib_setup_qp(struct rds_connection *conn)
511 struct rds_ib_connection *ic = conn->c_transport_data;
512 struct ib_device *dev = ic->i_cm_id->device;
513 struct ib_qp_init_attr attr;
514 struct ib_cq_init_attr cq_attr = {};
515 struct rds_ib_device *rds_ibdev;
516 int ret, fr_queue_space;
517 struct dma_pool *pool;
520 * It's normal to see a null device if an incoming connection races
521 * with device removal, so we don't print a warning.
523 rds_ibdev = rds_ib_get_client_data(dev);
527 /* The fr_queue_space is currently set to 512, to add extra space on
528 * completion queue and send queue. This extra space is used for FRMR
529 * registration and invalidation work requests
531 fr_queue_space = (rds_ibdev->use_fastreg ? RDS_IB_DEFAULT_FR_WR : 0);
533 /* add the conn now so that connection establishment has the dev */
534 rds_ib_add_conn(rds_ibdev, conn);
536 if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
537 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
538 if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
539 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
541 /* Protection domain and memory range */
542 ic->i_pd = rds_ibdev->pd;
544 ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
545 cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
546 cq_attr.comp_vector = ic->i_scq_vector;
547 ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
548 rds_ib_cq_event_handler, conn,
550 if (IS_ERR(ic->i_send_cq)) {
551 ret = PTR_ERR(ic->i_send_cq);
552 ic->i_send_cq = NULL;
553 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
554 rdsdebug("ib_create_cq send failed: %d\n", ret);
558 ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
559 cq_attr.cqe = ic->i_recv_ring.w_nr;
560 cq_attr.comp_vector = ic->i_rcq_vector;
561 ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
562 rds_ib_cq_event_handler, conn,
564 if (IS_ERR(ic->i_recv_cq)) {
565 ret = PTR_ERR(ic->i_recv_cq);
566 ic->i_recv_cq = NULL;
567 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
568 rdsdebug("ib_create_cq recv failed: %d\n", ret);
572 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
574 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
578 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
580 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
584 /* XXX negotiate max send/recv with remote? */
585 memset(&attr, 0, sizeof(attr));
586 attr.event_handler = rds_ib_qp_event_handler;
587 attr.qp_context = conn;
588 /* + 1 to allow for the single ack message */
589 attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
590 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
591 attr.cap.max_send_sge = rds_ibdev->max_sge;
592 attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
593 attr.sq_sig_type = IB_SIGNAL_REQ_WR;
594 attr.qp_type = IB_QPT_RC;
595 attr.send_cq = ic->i_send_cq;
596 attr.recv_cq = ic->i_recv_cq;
599 * XXX this can fail if max_*_wr is too large? Are we supposed
600 * to back off until we get a value that the hardware can support?
602 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
604 rdsdebug("rdma_create_qp failed: %d\n", ret);
608 pool = rds_ibdev->rid_hdrs_pool;
609 ic->i_send_hdrs = rds_dma_hdrs_alloc(dev, pool, &ic->i_send_hdrs_dma,
610 ic->i_send_ring.w_nr);
611 if (!ic->i_send_hdrs) {
613 rdsdebug("DMA send hdrs alloc failed\n");
617 ic->i_recv_hdrs = rds_dma_hdrs_alloc(dev, pool, &ic->i_recv_hdrs_dma,
618 ic->i_recv_ring.w_nr);
619 if (!ic->i_recv_hdrs) {
621 rdsdebug("DMA recv hdrs alloc failed\n");
622 goto send_hdrs_dma_out;
625 ic->i_ack = dma_pool_zalloc(pool, GFP_KERNEL,
629 rdsdebug("DMA ack header alloc failed\n");
630 goto recv_hdrs_dma_out;
633 ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
634 ic->i_send_ring.w_nr),
638 rdsdebug("send allocation failed\n");
642 ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
643 ic->i_recv_ring.w_nr),
647 rdsdebug("recv allocation failed\n");
651 rds_ib_recv_init_ack(ic);
653 rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
654 ic->i_send_cq, ic->i_recv_cq);
662 dma_pool_free(pool, ic->i_ack, ic->i_ack_dma);
666 rds_dma_hdrs_free(pool, ic->i_recv_hdrs, ic->i_recv_hdrs_dma,
667 ic->i_recv_ring.w_nr);
668 ic->i_recv_hdrs = NULL;
669 ic->i_recv_hdrs_dma = NULL;
672 rds_dma_hdrs_free(pool, ic->i_send_hdrs, ic->i_send_hdrs_dma,
673 ic->i_send_ring.w_nr);
674 ic->i_send_hdrs = NULL;
675 ic->i_send_hdrs_dma = NULL;
678 rdma_destroy_qp(ic->i_cm_id);
680 ib_destroy_cq(ic->i_recv_cq);
681 ic->i_recv_cq = NULL;
683 ib_destroy_cq(ic->i_send_cq);
684 ic->i_send_cq = NULL;
686 rds_ib_remove_conn(rds_ibdev, conn);
688 rds_ib_dev_put(rds_ibdev);
693 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
695 const union rds_ib_conn_priv *dp = event->param.conn.private_data;
696 u8 data_len, major, minor;
702 * rdma_cm private data is odd - when there is any private data in the
703 * request, we will be given a pretty large buffer without telling us the
704 * original size. The only way to tell the difference is by looking at
705 * the contents, which are initialized to zero.
706 * If the protocol version fields aren't set, this is a connection attempt
707 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
708 * We really should have changed this for OFED 1.3 :-(
711 /* Be paranoid. RDS always has privdata */
712 if (!event->param.conn.private_data_len) {
713 printk(KERN_NOTICE "RDS incoming connection has no private data, "
719 data_len = sizeof(struct rds6_ib_connect_private);
720 major = dp->ricp_v6.dp_protocol_major;
721 minor = dp->ricp_v6.dp_protocol_minor;
722 mask = dp->ricp_v6.dp_protocol_minor_mask;
724 data_len = sizeof(struct rds_ib_connect_private);
725 major = dp->ricp_v4.dp_protocol_major;
726 minor = dp->ricp_v4.dp_protocol_minor;
727 mask = dp->ricp_v4.dp_protocol_minor_mask;
730 /* Even if len is crap *now* I still want to check it. -ASG */
731 if (event->param.conn.private_data_len < data_len || major == 0)
732 return RDS_PROTOCOL_4_0;
734 common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS;
735 if (major == 4 && common) {
736 version = RDS_PROTOCOL_4_0;
737 while ((common >>= 1) != 0)
739 } else if (RDS_PROTOCOL_COMPAT_VERSION ==
740 RDS_PROTOCOL(major, minor)) {
741 version = RDS_PROTOCOL_COMPAT_VERSION;
744 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
745 &dp->ricp_v6.dp_saddr, major, minor);
747 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
748 &dp->ricp_v4.dp_saddr, major, minor);
753 #if IS_ENABLED(CONFIG_IPV6)
754 /* Given an IPv6 address, find the net_device which hosts that address and
755 * return its index. This is used by the rds_ib_cm_handle_connect() code to
756 * find the interface index of where an incoming request comes from when
757 * the request is using a link local address.
759 * Note one problem in this search. It is possible that two interfaces have
760 * the same link local address. Unfortunately, this cannot be solved unless
761 * the underlying layer gives us the interface which an incoming RDMA connect
762 * request comes from.
764 static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
766 struct net_device *dev;
770 for_each_netdev_rcu(net, dev) {
771 if (ipv6_chk_addr(net, addr, dev, 1)) {
782 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
783 struct rdma_cm_event *event, bool isv6)
785 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
786 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
787 const struct rds_ib_conn_priv_cmn *dp_cmn;
788 struct rds_connection *conn = NULL;
789 struct rds_ib_connection *ic = NULL;
790 struct rdma_conn_param conn_param;
791 const union rds_ib_conn_priv *dp;
792 union rds_ib_conn_priv dp_rep;
793 struct in6_addr s_mapped_addr;
794 struct in6_addr d_mapped_addr;
795 const struct in6_addr *saddr6;
796 const struct in6_addr *daddr6;
802 /* Check whether the remote protocol version matches ours. */
803 version = rds_ib_protocol_compatible(event, isv6);
805 err = RDS_RDMA_REJ_INCOMPAT;
809 dp = event->param.conn.private_data;
811 #if IS_ENABLED(CONFIG_IPV6)
812 dp_cmn = &dp->ricp_v6.dp_cmn;
813 saddr6 = &dp->ricp_v6.dp_saddr;
814 daddr6 = &dp->ricp_v6.dp_daddr;
815 /* If either address is link local, need to find the
816 * interface index in order to create a proper RDS
819 if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) {
820 /* Using init_net for now .. */
821 ifindex = __rds_find_ifindex(&init_net, daddr6);
822 /* No index found... Need to bail out. */
827 } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) {
828 /* Use our address to find the correct index. */
829 ifindex = __rds_find_ifindex(&init_net, daddr6);
830 /* No index found... Need to bail out. */
841 dp_cmn = &dp->ricp_v4.dp_cmn;
842 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr);
843 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
844 saddr6 = &s_mapped_addr;
845 daddr6 = &d_mapped_addr;
848 rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid 0x%llx, tos:%d\n",
849 saddr6, daddr6, RDS_PROTOCOL_MAJOR(version),
850 RDS_PROTOCOL_MINOR(version),
851 (unsigned long long)be64_to_cpu(lguid),
852 (unsigned long long)be64_to_cpu(fguid), dp_cmn->ricpc_dp_toss);
854 /* RDS/IB is not currently netns aware, thus init_net */
855 conn = rds_conn_create(&init_net, daddr6, saddr6,
856 &rds_ib_transport, dp_cmn->ricpc_dp_toss,
857 GFP_KERNEL, ifindex);
859 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
865 * The connection request may occur while the
866 * previous connection exist, e.g. in case of failover.
867 * But as connections may be initiated simultaneously
868 * by both hosts, we have a random backoff mechanism -
869 * see the comment above rds_queue_reconnect()
871 mutex_lock(&conn->c_cm_lock);
872 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
873 if (rds_conn_state(conn) == RDS_CONN_UP) {
874 rdsdebug("incoming connect while connecting\n");
876 rds_ib_stats_inc(s_ib_listen_closed_stale);
878 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
879 /* Wait and see - our connect may still be succeeding */
880 rds_ib_stats_inc(s_ib_connect_raced);
885 ic = conn->c_transport_data;
887 rds_ib_set_protocol(conn, version);
888 rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
890 /* If the peer gave us the last packet it saw, process this as if
891 * we had received a regular ACK. */
892 if (dp_cmn->ricpc_ack_seq)
893 rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq),
896 BUG_ON(cm_id->context);
900 cm_id->context = conn;
902 /* We got halfway through setting up the ib_connection, if we
903 * fail now, we have to take the long route out of this mess. */
906 err = rds_ib_setup_qp(conn);
908 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
912 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
913 event->param.conn.responder_resources,
914 event->param.conn.initiator_depth, isv6);
916 /* rdma_accept() calls rdma_reject() internally if it fails */
917 if (rdma_accept(cm_id, &conn_param))
918 rds_ib_conn_error(conn, "rdma_accept failed\n");
922 mutex_unlock(&conn->c_cm_lock);
924 rdma_reject(cm_id, &err, sizeof(int));
929 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
931 struct rds_connection *conn = cm_id->context;
932 struct rds_ib_connection *ic = conn->c_transport_data;
933 struct rdma_conn_param conn_param;
934 union rds_ib_conn_priv dp;
937 /* If the peer doesn't do protocol negotiation, we must
938 * default to RDSv3.0 */
939 rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1);
940 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
942 ret = rds_ib_setup_qp(conn);
944 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
948 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
949 conn->c_proposed_version,
950 UINT_MAX, UINT_MAX, isv6);
951 ret = rdma_connect(cm_id, &conn_param);
953 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
956 /* Beware - returning non-zero tells the rdma_cm to destroy
957 * the cm_id. We should certainly not do it as long as we still
958 * "own" the cm_id. */
960 if (ic->i_cm_id == cm_id)
963 ic->i_active_side = true;
967 int rds_ib_conn_path_connect(struct rds_conn_path *cp)
969 struct rds_connection *conn = cp->cp_conn;
970 struct sockaddr_storage src, dest;
971 rdma_cm_event_handler handler;
972 struct rds_ib_connection *ic;
975 ic = conn->c_transport_data;
977 /* XXX I wonder what affect the port space has */
978 /* delegate cm event handler to rdma_transport */
979 #if IS_ENABLED(CONFIG_IPV6)
981 handler = rds6_rdma_cm_event_handler;
984 handler = rds_rdma_cm_event_handler;
985 ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
986 RDMA_PS_TCP, IB_QPT_RC);
987 if (IS_ERR(ic->i_cm_id)) {
988 ret = PTR_ERR(ic->i_cm_id);
990 rdsdebug("rdma_create_id() failed: %d\n", ret);
994 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
996 if (ipv6_addr_v4mapped(&conn->c_faddr)) {
997 struct sockaddr_in *sin;
999 sin = (struct sockaddr_in *)&src;
1000 sin->sin_family = AF_INET;
1001 sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
1004 sin = (struct sockaddr_in *)&dest;
1005 sin->sin_family = AF_INET;
1006 sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
1007 sin->sin_port = htons(RDS_PORT);
1009 struct sockaddr_in6 *sin6;
1011 sin6 = (struct sockaddr_in6 *)&src;
1012 sin6->sin6_family = AF_INET6;
1013 sin6->sin6_addr = conn->c_laddr;
1014 sin6->sin6_port = 0;
1015 sin6->sin6_scope_id = conn->c_dev_if;
1017 sin6 = (struct sockaddr_in6 *)&dest;
1018 sin6->sin6_family = AF_INET6;
1019 sin6->sin6_addr = conn->c_faddr;
1020 sin6->sin6_port = htons(RDS_CM_PORT);
1021 sin6->sin6_scope_id = conn->c_dev_if;
1024 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
1025 (struct sockaddr *)&dest,
1026 RDS_RDMA_RESOLVE_TIMEOUT_MS);
1028 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
1030 rdma_destroy_id(ic->i_cm_id);
1039 * This is so careful about only cleaning up resources that were built up
1040 * so that it can be called at any point during startup. In fact it
1041 * can be called multiple times for a given connection.
1043 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
1045 struct rds_connection *conn = cp->cp_conn;
1046 struct rds_ib_connection *ic = conn->c_transport_data;
1049 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
1050 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
1051 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
1054 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
1055 err = rdma_disconnect(ic->i_cm_id);
1057 /* Actually this may happen quite frequently, when
1058 * an outgoing connect raced with an incoming connect.
1060 rdsdebug("failed to disconnect, cm: %p err %d\n",
1064 /* kick off "flush_worker" for all pools in order to reap
1065 * all FRMR registrations that are still marked "FRMR_IS_INUSE"
1070 * We want to wait for tx and rx completion to finish
1071 * before we tear down the connection, but we have to be
1072 * careful not to get stuck waiting on a send ring that
1073 * only has unsignaled sends in it. We've shutdown new
1074 * sends before getting here so by waiting for signaled
1075 * sends to complete we're ensured that there will be no
1076 * more tx processing.
1078 wait_event(rds_ib_ring_empty_wait,
1079 rds_ib_ring_empty(&ic->i_recv_ring) &&
1080 (atomic_read(&ic->i_signaled_sends) == 0) &&
1081 (atomic_read(&ic->i_fastreg_inuse_count) == 0) &&
1082 (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
1083 tasklet_kill(&ic->i_send_tasklet);
1084 tasklet_kill(&ic->i_recv_tasklet);
1086 atomic_set(&ic->i_cq_quiesce, 1);
1088 /* first destroy the ib state that generates callbacks */
1089 if (ic->i_cm_id->qp)
1090 rdma_destroy_qp(ic->i_cm_id);
1091 if (ic->i_send_cq) {
1093 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1094 ib_destroy_cq(ic->i_send_cq);
1097 if (ic->i_recv_cq) {
1099 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1100 ib_destroy_cq(ic->i_recv_cq);
1103 if (ic->rds_ibdev) {
1104 struct dma_pool *pool;
1106 pool = ic->rds_ibdev->rid_hdrs_pool;
1108 /* then free the resources that ib callbacks use */
1109 if (ic->i_send_hdrs) {
1110 rds_dma_hdrs_free(pool, ic->i_send_hdrs,
1111 ic->i_send_hdrs_dma,
1112 ic->i_send_ring.w_nr);
1113 ic->i_send_hdrs = NULL;
1114 ic->i_send_hdrs_dma = NULL;
1117 if (ic->i_recv_hdrs) {
1118 rds_dma_hdrs_free(pool, ic->i_recv_hdrs,
1119 ic->i_recv_hdrs_dma,
1120 ic->i_recv_ring.w_nr);
1121 ic->i_recv_hdrs = NULL;
1122 ic->i_recv_hdrs_dma = NULL;
1126 dma_pool_free(pool, ic->i_ack, ic->i_ack_dma);
1130 WARN_ON(ic->i_send_hdrs);
1131 WARN_ON(ic->i_send_hdrs_dma);
1132 WARN_ON(ic->i_recv_hdrs);
1133 WARN_ON(ic->i_recv_hdrs_dma);
1138 rds_ib_send_clear_ring(ic);
1140 rds_ib_recv_clear_ring(ic);
1142 rdma_destroy_id(ic->i_cm_id);
1145 * Move connection back to the nodev list.
1148 rds_ib_remove_conn(ic->rds_ibdev, conn);
1152 ic->i_send_cq = NULL;
1153 ic->i_recv_cq = NULL;
1155 BUG_ON(ic->rds_ibdev);
1157 /* Clear pending transmit */
1158 if (ic->i_data_op) {
1159 struct rds_message *rm;
1161 rm = container_of(ic->i_data_op, struct rds_message, data);
1162 rds_message_put(rm);
1163 ic->i_data_op = NULL;
1166 /* Clear the ACK state */
1167 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1168 #ifdef KERNEL_HAS_ATOMIC64
1169 atomic64_set(&ic->i_ack_next, 0);
1175 /* Clear flow control state */
1177 atomic_set(&ic->i_credits, 0);
1179 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
1180 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
1183 rds_inc_put(&ic->i_ibinc->ii_inc);
1191 ic->i_active_side = false;
1194 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1196 struct rds_ib_connection *ic;
1197 unsigned long flags;
1201 ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1205 ret = rds_ib_recv_alloc_caches(ic, gfp);
1211 INIT_LIST_HEAD(&ic->ib_node);
1212 tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1214 tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1216 mutex_init(&ic->i_recv_mutex);
1217 #ifndef KERNEL_HAS_ATOMIC64
1218 spin_lock_init(&ic->i_ack_lock);
1220 atomic_set(&ic->i_signaled_sends, 0);
1221 atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
1224 * rds_ib_conn_shutdown() waits for these to be emptied so they
1225 * must be initialized before it can be called.
1227 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
1228 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
1231 conn->c_transport_data = ic;
1233 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1234 list_add_tail(&ic->ib_node, &ib_nodev_conns);
1235 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1238 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1243 * Free a connection. Connection must be shut down and not set for reconnect.
1245 void rds_ib_conn_free(void *arg)
1247 struct rds_ib_connection *ic = arg;
1248 spinlock_t *lock_ptr;
1250 rdsdebug("ic %p\n", ic);
1253 * Conn is either on a dev's list or on the nodev list.
1254 * A race with shutdown() or connect() would cause problems
1255 * (since rds_ibdev would change) but that should never happen.
1257 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1259 spin_lock_irq(lock_ptr);
1260 list_del(&ic->ib_node);
1261 spin_unlock_irq(lock_ptr);
1263 rds_ib_recv_free_caches(ic);
1270 * An error occurred on the connection
1273 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1277 rds_conn_drop(conn);