Loading...
1/*
2 * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/in.h>
35#include <linux/slab.h>
36#include <linux/vmalloc.h>
37#include <linux/ratelimit.h>
38#include <net/addrconf.h>
39#include <rdma/ib_cm.h>
40
41#include "rds_single_path.h"
42#include "rds.h"
43#include "ib.h"
44#include "ib_mr.h"
45
46/*
47 * Set the selected protocol version
48 */
49static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
50{
51 conn->c_version = version;
52}
53
54/*
55 * Set up flow control
56 */
57static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
58{
59 struct rds_ib_connection *ic = conn->c_transport_data;
60
61 if (rds_ib_sysctl_flow_control && credits != 0) {
62 /* We're doing flow control */
63 ic->i_flowctl = 1;
64 rds_ib_send_add_credits(conn, credits);
65 } else {
66 ic->i_flowctl = 0;
67 }
68}
69
70/*
71 * Connection established.
72 * We get here for both outgoing and incoming connection.
73 */
74void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
75{
76 struct rds_ib_connection *ic = conn->c_transport_data;
77 const union rds_ib_conn_priv *dp = NULL;
78 __be64 ack_seq = 0;
79 __be32 credit = 0;
80 u8 major = 0;
81 u8 minor = 0;
82 int err;
83
84 dp = event->param.conn.private_data;
85 if (conn->c_isv6) {
86 if (event->param.conn.private_data_len >=
87 sizeof(struct rds6_ib_connect_private)) {
88 major = dp->ricp_v6.dp_protocol_major;
89 minor = dp->ricp_v6.dp_protocol_minor;
90 credit = dp->ricp_v6.dp_credit;
91 /* dp structure start is not guaranteed to be 8 bytes
92 * aligned. Since dp_ack_seq is 64-bit extended load
93 * operations can be used so go through get_unaligned
94 * to avoid unaligned errors.
95 */
96 ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq);
97 }
98 } else if (event->param.conn.private_data_len >=
99 sizeof(struct rds_ib_connect_private)) {
100 major = dp->ricp_v4.dp_protocol_major;
101 minor = dp->ricp_v4.dp_protocol_minor;
102 credit = dp->ricp_v4.dp_credit;
103 ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq);
104 }
105
106 /* make sure it isn't empty data */
107 if (major) {
108 rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor));
109 rds_ib_set_flow_control(conn, be32_to_cpu(credit));
110 }
111
112 if (conn->c_version < RDS_PROTOCOL_VERSION) {
113 if (conn->c_version != RDS_PROTOCOL_COMPAT_VERSION) {
114 pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n",
115 &conn->c_laddr, &conn->c_faddr,
116 RDS_PROTOCOL_MAJOR(conn->c_version),
117 RDS_PROTOCOL_MINOR(conn->c_version));
118 rds_conn_destroy(conn);
119 return;
120 }
121 }
122
123 pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c,%d> version %u.%u%s\n",
124 ic->i_active_side ? "Active" : "Passive",
125 &conn->c_laddr, &conn->c_faddr, conn->c_tos,
126 RDS_PROTOCOL_MAJOR(conn->c_version),
127 RDS_PROTOCOL_MINOR(conn->c_version),
128 ic->i_flowctl ? ", flow control" : "");
129
130 /* receive sl from the peer */
131 ic->i_sl = ic->i_cm_id->route.path_rec->sl;
132
133 atomic_set(&ic->i_cq_quiesce, 0);
134
135 /* Init rings and fill recv. this needs to wait until protocol
136 * negotiation is complete, since ring layout is different
137 * from 3.1 to 4.1.
138 */
139 rds_ib_send_init_ring(ic);
140 rds_ib_recv_init_ring(ic);
141 /* Post receive buffers - as a side effect, this will update
142 * the posted credit count. */
143 rds_ib_recv_refill(conn, 1, GFP_KERNEL);
144
145 /* update ib_device with this local ipaddr */
146 err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr);
147 if (err)
148 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
149 err);
150
151 /* If the peer gave us the last packet it saw, process this as if
152 * we had received a regular ACK. */
153 if (dp) {
154 if (ack_seq)
155 rds_send_drop_acked(conn, be64_to_cpu(ack_seq),
156 NULL);
157 }
158
159 conn->c_proposed_version = conn->c_version;
160 rds_connect_complete(conn);
161}
162
163static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
164 struct rdma_conn_param *conn_param,
165 union rds_ib_conn_priv *dp,
166 u32 protocol_version,
167 u32 max_responder_resources,
168 u32 max_initiator_depth,
169 bool isv6)
170{
171 struct rds_ib_connection *ic = conn->c_transport_data;
172 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
173
174 memset(conn_param, 0, sizeof(struct rdma_conn_param));
175
176 conn_param->responder_resources =
177 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
178 conn_param->initiator_depth =
179 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
180 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
181 conn_param->rnr_retry_count = 7;
182
183 if (dp) {
184 memset(dp, 0, sizeof(*dp));
185 if (isv6) {
186 dp->ricp_v6.dp_saddr = conn->c_laddr;
187 dp->ricp_v6.dp_daddr = conn->c_faddr;
188 dp->ricp_v6.dp_protocol_major =
189 RDS_PROTOCOL_MAJOR(protocol_version);
190 dp->ricp_v6.dp_protocol_minor =
191 RDS_PROTOCOL_MINOR(protocol_version);
192 dp->ricp_v6.dp_protocol_minor_mask =
193 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
194 dp->ricp_v6.dp_ack_seq =
195 cpu_to_be64(rds_ib_piggyb_ack(ic));
196 dp->ricp_v6.dp_cmn.ricpc_dp_toss = conn->c_tos;
197
198 conn_param->private_data = &dp->ricp_v6;
199 conn_param->private_data_len = sizeof(dp->ricp_v6);
200 } else {
201 dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3];
202 dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3];
203 dp->ricp_v4.dp_protocol_major =
204 RDS_PROTOCOL_MAJOR(protocol_version);
205 dp->ricp_v4.dp_protocol_minor =
206 RDS_PROTOCOL_MINOR(protocol_version);
207 dp->ricp_v4.dp_protocol_minor_mask =
208 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
209 dp->ricp_v4.dp_ack_seq =
210 cpu_to_be64(rds_ib_piggyb_ack(ic));
211 dp->ricp_v4.dp_cmn.ricpc_dp_toss = conn->c_tos;
212
213 conn_param->private_data = &dp->ricp_v4;
214 conn_param->private_data_len = sizeof(dp->ricp_v4);
215 }
216
217 /* Advertise flow control */
218 if (ic->i_flowctl) {
219 unsigned int credits;
220
221 credits = IB_GET_POST_CREDITS
222 (atomic_read(&ic->i_credits));
223 if (isv6)
224 dp->ricp_v6.dp_credit = cpu_to_be32(credits);
225 else
226 dp->ricp_v4.dp_credit = cpu_to_be32(credits);
227 atomic_sub(IB_SET_POST_CREDITS(credits),
228 &ic->i_credits);
229 }
230 }
231}
232
233static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
234{
235 rdsdebug("event %u (%s) data %p\n",
236 event->event, ib_event_msg(event->event), data);
237}
238
239/* Plucking the oldest entry from the ring can be done concurrently with
240 * the thread refilling the ring. Each ring operation is protected by
241 * spinlocks and the transient state of refilling doesn't change the
242 * recording of which entry is oldest.
243 *
244 * This relies on IB only calling one cq comp_handler for each cq so that
245 * there will only be one caller of rds_recv_incoming() per RDS connection.
246 */
247static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
248{
249 struct rds_connection *conn = context;
250 struct rds_ib_connection *ic = conn->c_transport_data;
251
252 rdsdebug("conn %p cq %p\n", conn, cq);
253
254 rds_ib_stats_inc(s_ib_evt_handler_call);
255
256 tasklet_schedule(&ic->i_recv_tasklet);
257}
258
259static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
260 struct ib_wc *wcs)
261{
262 int nr, i;
263 struct ib_wc *wc;
264
265 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
266 for (i = 0; i < nr; i++) {
267 wc = wcs + i;
268 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
269 (unsigned long long)wc->wr_id, wc->status,
270 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
271
272 if (wc->wr_id <= ic->i_send_ring.w_nr ||
273 wc->wr_id == RDS_IB_ACK_WR_ID)
274 rds_ib_send_cqe_handler(ic, wc);
275 else
276 rds_ib_mr_cqe_handler(ic, wc);
277
278 }
279 }
280}
281
282static void rds_ib_tasklet_fn_send(unsigned long data)
283{
284 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
285 struct rds_connection *conn = ic->conn;
286
287 rds_ib_stats_inc(s_ib_tasklet_call);
288
289 /* if cq has been already reaped, ignore incoming cq event */
290 if (atomic_read(&ic->i_cq_quiesce))
291 return;
292
293 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
294 ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
295 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
296
297 if (rds_conn_up(conn) &&
298 (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
299 test_bit(0, &conn->c_map_queued)))
300 rds_send_xmit(&ic->conn->c_path[0]);
301}
302
303static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
304 struct ib_wc *wcs,
305 struct rds_ib_ack_state *ack_state)
306{
307 int nr, i;
308 struct ib_wc *wc;
309
310 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
311 for (i = 0; i < nr; i++) {
312 wc = wcs + i;
313 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
314 (unsigned long long)wc->wr_id, wc->status,
315 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
316
317 rds_ib_recv_cqe_handler(ic, wc, ack_state);
318 }
319 }
320}
321
322static void rds_ib_tasklet_fn_recv(unsigned long data)
323{
324 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
325 struct rds_connection *conn = ic->conn;
326 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
327 struct rds_ib_ack_state state;
328
329 if (!rds_ibdev)
330 rds_conn_drop(conn);
331
332 rds_ib_stats_inc(s_ib_tasklet_call);
333
334 /* if cq has been already reaped, ignore incoming cq event */
335 if (atomic_read(&ic->i_cq_quiesce))
336 return;
337
338 memset(&state, 0, sizeof(state));
339 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
340 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
341 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
342
343 if (state.ack_next_valid)
344 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
345 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
346 rds_send_drop_acked(conn, state.ack_recv, NULL);
347 ic->i_ack_recv = state.ack_recv;
348 }
349
350 if (rds_conn_up(conn))
351 rds_ib_attempt_ack(ic);
352}
353
354static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
355{
356 struct rds_connection *conn = data;
357 struct rds_ib_connection *ic = conn->c_transport_data;
358
359 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
360 ib_event_msg(event->event));
361
362 switch (event->event) {
363 case IB_EVENT_COMM_EST:
364 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
365 break;
366 default:
367 rdsdebug("Fatal QP Event %u (%s) - connection %pI6c->%pI6c, reconnecting\n",
368 event->event, ib_event_msg(event->event),
369 &conn->c_laddr, &conn->c_faddr);
370 rds_conn_drop(conn);
371 break;
372 }
373}
374
375static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
376{
377 struct rds_connection *conn = context;
378 struct rds_ib_connection *ic = conn->c_transport_data;
379
380 rdsdebug("conn %p cq %p\n", conn, cq);
381
382 rds_ib_stats_inc(s_ib_evt_handler_call);
383
384 tasklet_schedule(&ic->i_send_tasklet);
385}
386
387static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
388{
389 int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
390 int index = rds_ibdev->dev->num_comp_vectors - 1;
391 int i;
392
393 for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
394 if (rds_ibdev->vector_load[i] < min) {
395 index = i;
396 min = rds_ibdev->vector_load[i];
397 }
398 }
399
400 rds_ibdev->vector_load[index]++;
401 return index;
402}
403
404static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
405{
406 rds_ibdev->vector_load[index]--;
407}
408
409static void rds_dma_hdr_free(struct ib_device *dev, struct rds_header *hdr,
410 dma_addr_t dma_addr, enum dma_data_direction dir)
411{
412 ib_dma_unmap_single(dev, dma_addr, sizeof(*hdr), dir);
413 kfree(hdr);
414}
415
416static struct rds_header *rds_dma_hdr_alloc(struct ib_device *dev,
417 dma_addr_t *dma_addr, enum dma_data_direction dir)
418{
419 struct rds_header *hdr;
420
421 hdr = kzalloc_node(sizeof(*hdr), GFP_KERNEL, ibdev_to_node(dev));
422 if (!hdr)
423 return NULL;
424
425 *dma_addr = ib_dma_map_single(dev, hdr, sizeof(*hdr),
426 DMA_BIDIRECTIONAL);
427 if (ib_dma_mapping_error(dev, *dma_addr)) {
428 kfree(hdr);
429 return NULL;
430 }
431
432 return hdr;
433}
434
435/* Free the DMA memory used to store struct rds_header.
436 *
437 * @dev: the RDS IB device
438 * @hdrs: pointer to the array storing DMA memory pointers
439 * @dma_addrs: pointer to the array storing DMA addresses
440 * @num_hdars: number of headers to free.
441 */
442static void rds_dma_hdrs_free(struct rds_ib_device *dev,
443 struct rds_header **hdrs, dma_addr_t *dma_addrs, u32 num_hdrs,
444 enum dma_data_direction dir)
445{
446 u32 i;
447
448 for (i = 0; i < num_hdrs; i++)
449 rds_dma_hdr_free(dev->dev, hdrs[i], dma_addrs[i], dir);
450 kvfree(hdrs);
451 kvfree(dma_addrs);
452}
453
454
455/* Allocate DMA coherent memory to be used to store struct rds_header for
456 * sending/receiving packets. The pointers to the DMA memory and the
457 * associated DMA addresses are stored in two arrays.
458 *
459 * @dev: the RDS IB device
460 * @dma_addrs: pointer to the array for storing DMA addresses
461 * @num_hdrs: number of headers to allocate
462 *
463 * It returns the pointer to the array storing the DMA memory pointers. On
464 * error, NULL pointer is returned.
465 */
466static struct rds_header **rds_dma_hdrs_alloc(struct rds_ib_device *dev,
467 dma_addr_t **dma_addrs, u32 num_hdrs,
468 enum dma_data_direction dir)
469{
470 struct rds_header **hdrs;
471 dma_addr_t *hdr_daddrs;
472 u32 i;
473
474 hdrs = kvmalloc_node(sizeof(*hdrs) * num_hdrs, GFP_KERNEL,
475 ibdev_to_node(dev->dev));
476 if (!hdrs)
477 return NULL;
478
479 hdr_daddrs = kvmalloc_node(sizeof(*hdr_daddrs) * num_hdrs, GFP_KERNEL,
480 ibdev_to_node(dev->dev));
481 if (!hdr_daddrs) {
482 kvfree(hdrs);
483 return NULL;
484 }
485
486 for (i = 0; i < num_hdrs; i++) {
487 hdrs[i] = rds_dma_hdr_alloc(dev->dev, &hdr_daddrs[i], dir);
488 if (!hdrs[i]) {
489 rds_dma_hdrs_free(dev, hdrs, hdr_daddrs, i, dir);
490 return NULL;
491 }
492 }
493
494 *dma_addrs = hdr_daddrs;
495 return hdrs;
496}
497
498/*
499 * This needs to be very careful to not leave IS_ERR pointers around for
500 * cleanup to trip over.
501 */
502static int rds_ib_setup_qp(struct rds_connection *conn)
503{
504 struct rds_ib_connection *ic = conn->c_transport_data;
505 struct ib_device *dev = ic->i_cm_id->device;
506 struct ib_qp_init_attr attr;
507 struct ib_cq_init_attr cq_attr = {};
508 struct rds_ib_device *rds_ibdev;
509 unsigned long max_wrs;
510 int ret, fr_queue_space;
511
512 /*
513 * It's normal to see a null device if an incoming connection races
514 * with device removal, so we don't print a warning.
515 */
516 rds_ibdev = rds_ib_get_client_data(dev);
517 if (!rds_ibdev)
518 return -EOPNOTSUPP;
519
520 /* The fr_queue_space is currently set to 512, to add extra space on
521 * completion queue and send queue. This extra space is used for FRWR
522 * registration and invalidation work requests
523 */
524 fr_queue_space = RDS_IB_DEFAULT_FR_WR;
525
526 /* add the conn now so that connection establishment has the dev */
527 rds_ib_add_conn(rds_ibdev, conn);
528
529 max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_send_wr + 1 ?
530 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_send_wr;
531 if (ic->i_send_ring.w_nr != max_wrs)
532 rds_ib_ring_resize(&ic->i_send_ring, max_wrs);
533
534 max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_recv_wr + 1 ?
535 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_recv_wr;
536 if (ic->i_recv_ring.w_nr != max_wrs)
537 rds_ib_ring_resize(&ic->i_recv_ring, max_wrs);
538
539 /* Protection domain and memory range */
540 ic->i_pd = rds_ibdev->pd;
541
542 ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
543 cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
544 cq_attr.comp_vector = ic->i_scq_vector;
545 ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
546 rds_ib_cq_event_handler, conn,
547 &cq_attr);
548 if (IS_ERR(ic->i_send_cq)) {
549 ret = PTR_ERR(ic->i_send_cq);
550 ic->i_send_cq = NULL;
551 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
552 rdsdebug("ib_create_cq send failed: %d\n", ret);
553 goto rds_ibdev_out;
554 }
555
556 ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
557 cq_attr.cqe = ic->i_recv_ring.w_nr;
558 cq_attr.comp_vector = ic->i_rcq_vector;
559 ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
560 rds_ib_cq_event_handler, conn,
561 &cq_attr);
562 if (IS_ERR(ic->i_recv_cq)) {
563 ret = PTR_ERR(ic->i_recv_cq);
564 ic->i_recv_cq = NULL;
565 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
566 rdsdebug("ib_create_cq recv failed: %d\n", ret);
567 goto send_cq_out;
568 }
569
570 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
571 if (ret) {
572 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
573 goto recv_cq_out;
574 }
575
576 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
577 if (ret) {
578 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
579 goto recv_cq_out;
580 }
581
582 /* XXX negotiate max send/recv with remote? */
583 memset(&attr, 0, sizeof(attr));
584 attr.event_handler = rds_ib_qp_event_handler;
585 attr.qp_context = conn;
586 /* + 1 to allow for the single ack message */
587 attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
588 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
589 attr.cap.max_send_sge = rds_ibdev->max_sge;
590 attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
591 attr.sq_sig_type = IB_SIGNAL_REQ_WR;
592 attr.qp_type = IB_QPT_RC;
593 attr.send_cq = ic->i_send_cq;
594 attr.recv_cq = ic->i_recv_cq;
595
596 /*
597 * XXX this can fail if max_*_wr is too large? Are we supposed
598 * to back off until we get a value that the hardware can support?
599 */
600 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
601 if (ret) {
602 rdsdebug("rdma_create_qp failed: %d\n", ret);
603 goto recv_cq_out;
604 }
605
606 ic->i_send_hdrs = rds_dma_hdrs_alloc(rds_ibdev, &ic->i_send_hdrs_dma,
607 ic->i_send_ring.w_nr,
608 DMA_TO_DEVICE);
609 if (!ic->i_send_hdrs) {
610 ret = -ENOMEM;
611 rdsdebug("DMA send hdrs alloc failed\n");
612 goto qp_out;
613 }
614
615 ic->i_recv_hdrs = rds_dma_hdrs_alloc(rds_ibdev, &ic->i_recv_hdrs_dma,
616 ic->i_recv_ring.w_nr,
617 DMA_FROM_DEVICE);
618 if (!ic->i_recv_hdrs) {
619 ret = -ENOMEM;
620 rdsdebug("DMA recv hdrs alloc failed\n");
621 goto send_hdrs_dma_out;
622 }
623
624 ic->i_ack = rds_dma_hdr_alloc(rds_ibdev->dev, &ic->i_ack_dma,
625 DMA_TO_DEVICE);
626 if (!ic->i_ack) {
627 ret = -ENOMEM;
628 rdsdebug("DMA ack header alloc failed\n");
629 goto recv_hdrs_dma_out;
630 }
631
632 ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
633 ic->i_send_ring.w_nr),
634 ibdev_to_node(dev));
635 if (!ic->i_sends) {
636 ret = -ENOMEM;
637 rdsdebug("send allocation failed\n");
638 goto ack_dma_out;
639 }
640
641 ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
642 ic->i_recv_ring.w_nr),
643 ibdev_to_node(dev));
644 if (!ic->i_recvs) {
645 ret = -ENOMEM;
646 rdsdebug("recv allocation failed\n");
647 goto sends_out;
648 }
649
650 rds_ib_recv_init_ack(ic);
651
652 rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
653 ic->i_send_cq, ic->i_recv_cq);
654
655 goto out;
656
657sends_out:
658 vfree(ic->i_sends);
659
660ack_dma_out:
661 rds_dma_hdr_free(rds_ibdev->dev, ic->i_ack, ic->i_ack_dma,
662 DMA_TO_DEVICE);
663 ic->i_ack = NULL;
664
665recv_hdrs_dma_out:
666 rds_dma_hdrs_free(rds_ibdev, ic->i_recv_hdrs, ic->i_recv_hdrs_dma,
667 ic->i_recv_ring.w_nr, DMA_FROM_DEVICE);
668 ic->i_recv_hdrs = NULL;
669 ic->i_recv_hdrs_dma = NULL;
670
671send_hdrs_dma_out:
672 rds_dma_hdrs_free(rds_ibdev, ic->i_send_hdrs, ic->i_send_hdrs_dma,
673 ic->i_send_ring.w_nr, DMA_TO_DEVICE);
674 ic->i_send_hdrs = NULL;
675 ic->i_send_hdrs_dma = NULL;
676
677qp_out:
678 rdma_destroy_qp(ic->i_cm_id);
679recv_cq_out:
680 ib_destroy_cq(ic->i_recv_cq);
681 ic->i_recv_cq = NULL;
682send_cq_out:
683 ib_destroy_cq(ic->i_send_cq);
684 ic->i_send_cq = NULL;
685rds_ibdev_out:
686 rds_ib_remove_conn(rds_ibdev, conn);
687out:
688 rds_ib_dev_put(rds_ibdev);
689
690 return ret;
691}
692
693static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
694{
695 const union rds_ib_conn_priv *dp = event->param.conn.private_data;
696 u8 data_len, major, minor;
697 u32 version = 0;
698 __be16 mask;
699 u16 common;
700
701 /*
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 be 3.0 or 2.0 - we can't tell.
708 * We really should have changed this for OFED 1.3 :-(
709 */
710
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, "
714 "rejecting\n");
715 return 0;
716 }
717
718 if (isv6) {
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;
723 } else {
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;
728 }
729
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;
733
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)
738 version++;
739 } else if (RDS_PROTOCOL_COMPAT_VERSION ==
740 RDS_PROTOCOL(major, minor)) {
741 version = RDS_PROTOCOL_COMPAT_VERSION;
742 } else {
743 if (isv6)
744 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
745 &dp->ricp_v6.dp_saddr, major, minor);
746 else
747 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
748 &dp->ricp_v4.dp_saddr, major, minor);
749 }
750 return version;
751}
752
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.
758 *
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.
763 */
764static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
765{
766 struct net_device *dev;
767 int idx = 0;
768
769 rcu_read_lock();
770 for_each_netdev_rcu(net, dev) {
771 if (ipv6_chk_addr(net, addr, dev, 1)) {
772 idx = dev->ifindex;
773 break;
774 }
775 }
776 rcu_read_unlock();
777
778 return idx;
779}
780#endif
781
782int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
783 struct rdma_cm_event *event, bool isv6)
784{
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;
797 int destroy = 1;
798 u32 ifindex = 0;
799 u32 version;
800 int err = 1;
801
802 /* Check whether the remote protocol version matches ours. */
803 version = rds_ib_protocol_compatible(event, isv6);
804 if (!version) {
805 err = RDS_RDMA_REJ_INCOMPAT;
806 goto out;
807 }
808
809 dp = event->param.conn.private_data;
810 if (isv6) {
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
817 * connection.
818 */
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. */
823 if (ifindex == 0) {
824 err = -EOPNOTSUPP;
825 goto out;
826 }
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. */
831 if (ifindex == 0) {
832 err = -EOPNOTSUPP;
833 goto out;
834 }
835 }
836#else
837 err = -EOPNOTSUPP;
838 goto out;
839#endif
840 } else {
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;
846 }
847
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);
853
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);
858 if (IS_ERR(conn)) {
859 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
860 conn = NULL;
861 goto out;
862 }
863
864 /*
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()
870 */
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");
875 rds_conn_drop(conn);
876 rds_ib_stats_inc(s_ib_listen_closed_stale);
877 } else
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);
881 }
882 goto out;
883 }
884
885 ic = conn->c_transport_data;
886
887 rds_ib_set_protocol(conn, version);
888 rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
889
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),
894 NULL);
895
896 BUG_ON(cm_id->context);
897 BUG_ON(ic->i_cm_id);
898
899 ic->i_cm_id = cm_id;
900 cm_id->context = conn;
901
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. */
904 destroy = 0;
905
906 err = rds_ib_setup_qp(conn);
907 if (err) {
908 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
909 goto out;
910 }
911
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);
915
916 rdma_set_min_rnr_timer(cm_id, IB_RNR_TIMER_000_32);
917 /* rdma_accept() calls rdma_reject() internally if it fails */
918 if (rdma_accept(cm_id, &conn_param))
919 rds_ib_conn_error(conn, "rdma_accept failed\n");
920
921out:
922 if (conn)
923 mutex_unlock(&conn->c_cm_lock);
924 if (err)
925 rdma_reject(cm_id, &err, sizeof(int),
926 IB_CM_REJ_CONSUMER_DEFINED);
927 return destroy;
928}
929
930
931int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
932{
933 struct rds_connection *conn = cm_id->context;
934 struct rds_ib_connection *ic = conn->c_transport_data;
935 struct rdma_conn_param conn_param;
936 union rds_ib_conn_priv dp;
937 int ret;
938
939 /* If the peer doesn't do protocol negotiation, we must
940 * default to RDSv3.0 */
941 rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1);
942 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
943
944 ret = rds_ib_setup_qp(conn);
945 if (ret) {
946 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
947 goto out;
948 }
949
950 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
951 conn->c_proposed_version,
952 UINT_MAX, UINT_MAX, isv6);
953 ret = rdma_connect_locked(cm_id, &conn_param);
954 if (ret)
955 rds_ib_conn_error(conn, "rdma_connect_locked failed (%d)\n",
956 ret);
957
958out:
959 /* Beware - returning non-zero tells the rdma_cm to destroy
960 * the cm_id. We should certainly not do it as long as we still
961 * "own" the cm_id. */
962 if (ret) {
963 if (ic->i_cm_id == cm_id)
964 ret = 0;
965 }
966 ic->i_active_side = true;
967 return ret;
968}
969
970int rds_ib_conn_path_connect(struct rds_conn_path *cp)
971{
972 struct rds_connection *conn = cp->cp_conn;
973 struct sockaddr_storage src, dest;
974 rdma_cm_event_handler handler;
975 struct rds_ib_connection *ic;
976 int ret;
977
978 ic = conn->c_transport_data;
979
980 /* XXX I wonder what affect the port space has */
981 /* delegate cm event handler to rdma_transport */
982#if IS_ENABLED(CONFIG_IPV6)
983 if (conn->c_isv6)
984 handler = rds6_rdma_cm_event_handler;
985 else
986#endif
987 handler = rds_rdma_cm_event_handler;
988 ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
989 RDMA_PS_TCP, IB_QPT_RC);
990 if (IS_ERR(ic->i_cm_id)) {
991 ret = PTR_ERR(ic->i_cm_id);
992 ic->i_cm_id = NULL;
993 rdsdebug("rdma_create_id() failed: %d\n", ret);
994 goto out;
995 }
996
997 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
998
999 if (ipv6_addr_v4mapped(&conn->c_faddr)) {
1000 struct sockaddr_in *sin;
1001
1002 sin = (struct sockaddr_in *)&src;
1003 sin->sin_family = AF_INET;
1004 sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
1005 sin->sin_port = 0;
1006
1007 sin = (struct sockaddr_in *)&dest;
1008 sin->sin_family = AF_INET;
1009 sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
1010 sin->sin_port = htons(RDS_PORT);
1011 } else {
1012 struct sockaddr_in6 *sin6;
1013
1014 sin6 = (struct sockaddr_in6 *)&src;
1015 sin6->sin6_family = AF_INET6;
1016 sin6->sin6_addr = conn->c_laddr;
1017 sin6->sin6_port = 0;
1018 sin6->sin6_scope_id = conn->c_dev_if;
1019
1020 sin6 = (struct sockaddr_in6 *)&dest;
1021 sin6->sin6_family = AF_INET6;
1022 sin6->sin6_addr = conn->c_faddr;
1023 sin6->sin6_port = htons(RDS_CM_PORT);
1024 sin6->sin6_scope_id = conn->c_dev_if;
1025 }
1026
1027 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
1028 (struct sockaddr *)&dest,
1029 RDS_RDMA_RESOLVE_TIMEOUT_MS);
1030 if (ret) {
1031 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
1032 ret);
1033 rdma_destroy_id(ic->i_cm_id);
1034 ic->i_cm_id = NULL;
1035 }
1036
1037out:
1038 return ret;
1039}
1040
1041/*
1042 * This is so careful about only cleaning up resources that were built up
1043 * so that it can be called at any point during startup. In fact it
1044 * can be called multiple times for a given connection.
1045 */
1046void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
1047{
1048 struct rds_connection *conn = cp->cp_conn;
1049 struct rds_ib_connection *ic = conn->c_transport_data;
1050 int err = 0;
1051
1052 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
1053 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
1054 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
1055
1056 if (ic->i_cm_id) {
1057 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
1058 err = rdma_disconnect(ic->i_cm_id);
1059 if (err) {
1060 /* Actually this may happen quite frequently, when
1061 * an outgoing connect raced with an incoming connect.
1062 */
1063 rdsdebug("failed to disconnect, cm: %p err %d\n",
1064 ic->i_cm_id, err);
1065 }
1066
1067 /* kick off "flush_worker" for all pools in order to reap
1068 * all FRMR registrations that are still marked "FRMR_IS_INUSE"
1069 */
1070 rds_ib_flush_mrs();
1071
1072 /*
1073 * We want to wait for tx and rx completion to finish
1074 * before we tear down the connection, but we have to be
1075 * careful not to get stuck waiting on a send ring that
1076 * only has unsignaled sends in it. We've shutdown new
1077 * sends before getting here so by waiting for signaled
1078 * sends to complete we're ensured that there will be no
1079 * more tx processing.
1080 */
1081 wait_event(rds_ib_ring_empty_wait,
1082 rds_ib_ring_empty(&ic->i_recv_ring) &&
1083 (atomic_read(&ic->i_signaled_sends) == 0) &&
1084 (atomic_read(&ic->i_fastreg_inuse_count) == 0) &&
1085 (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
1086 tasklet_kill(&ic->i_send_tasklet);
1087 tasklet_kill(&ic->i_recv_tasklet);
1088
1089 atomic_set(&ic->i_cq_quiesce, 1);
1090
1091 /* first destroy the ib state that generates callbacks */
1092 if (ic->i_cm_id->qp)
1093 rdma_destroy_qp(ic->i_cm_id);
1094 if (ic->i_send_cq) {
1095 if (ic->rds_ibdev)
1096 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1097 ib_destroy_cq(ic->i_send_cq);
1098 }
1099
1100 if (ic->i_recv_cq) {
1101 if (ic->rds_ibdev)
1102 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1103 ib_destroy_cq(ic->i_recv_cq);
1104 }
1105
1106 if (ic->rds_ibdev) {
1107 /* then free the resources that ib callbacks use */
1108 if (ic->i_send_hdrs) {
1109 rds_dma_hdrs_free(ic->rds_ibdev,
1110 ic->i_send_hdrs,
1111 ic->i_send_hdrs_dma,
1112 ic->i_send_ring.w_nr,
1113 DMA_TO_DEVICE);
1114 ic->i_send_hdrs = NULL;
1115 ic->i_send_hdrs_dma = NULL;
1116 }
1117
1118 if (ic->i_recv_hdrs) {
1119 rds_dma_hdrs_free(ic->rds_ibdev,
1120 ic->i_recv_hdrs,
1121 ic->i_recv_hdrs_dma,
1122 ic->i_recv_ring.w_nr,
1123 DMA_FROM_DEVICE);
1124 ic->i_recv_hdrs = NULL;
1125 ic->i_recv_hdrs_dma = NULL;
1126 }
1127
1128 if (ic->i_ack) {
1129 rds_dma_hdr_free(ic->rds_ibdev->dev, ic->i_ack,
1130 ic->i_ack_dma, DMA_TO_DEVICE);
1131 ic->i_ack = NULL;
1132 }
1133 } else {
1134 WARN_ON(ic->i_send_hdrs);
1135 WARN_ON(ic->i_send_hdrs_dma);
1136 WARN_ON(ic->i_recv_hdrs);
1137 WARN_ON(ic->i_recv_hdrs_dma);
1138 WARN_ON(ic->i_ack);
1139 }
1140
1141 if (ic->i_sends)
1142 rds_ib_send_clear_ring(ic);
1143 if (ic->i_recvs)
1144 rds_ib_recv_clear_ring(ic);
1145
1146 rdma_destroy_id(ic->i_cm_id);
1147
1148 /*
1149 * Move connection back to the nodev list.
1150 */
1151 if (ic->rds_ibdev)
1152 rds_ib_remove_conn(ic->rds_ibdev, conn);
1153
1154 ic->i_cm_id = NULL;
1155 ic->i_pd = NULL;
1156 ic->i_send_cq = NULL;
1157 ic->i_recv_cq = NULL;
1158 }
1159 BUG_ON(ic->rds_ibdev);
1160
1161 /* Clear pending transmit */
1162 if (ic->i_data_op) {
1163 struct rds_message *rm;
1164
1165 rm = container_of(ic->i_data_op, struct rds_message, data);
1166 rds_message_put(rm);
1167 ic->i_data_op = NULL;
1168 }
1169
1170 /* Clear the ACK state */
1171 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1172#ifdef KERNEL_HAS_ATOMIC64
1173 atomic64_set(&ic->i_ack_next, 0);
1174#else
1175 ic->i_ack_next = 0;
1176#endif
1177 ic->i_ack_recv = 0;
1178
1179 /* Clear flow control state */
1180 ic->i_flowctl = 0;
1181 atomic_set(&ic->i_credits, 0);
1182
1183 /* Re-init rings, but retain sizes. */
1184 rds_ib_ring_init(&ic->i_send_ring, ic->i_send_ring.w_nr);
1185 rds_ib_ring_init(&ic->i_recv_ring, ic->i_recv_ring.w_nr);
1186
1187 if (ic->i_ibinc) {
1188 rds_inc_put(&ic->i_ibinc->ii_inc);
1189 ic->i_ibinc = NULL;
1190 }
1191
1192 vfree(ic->i_sends);
1193 ic->i_sends = NULL;
1194 vfree(ic->i_recvs);
1195 ic->i_recvs = NULL;
1196 ic->i_active_side = false;
1197}
1198
1199int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1200{
1201 struct rds_ib_connection *ic;
1202 unsigned long flags;
1203 int ret;
1204
1205 /* XXX too lazy? */
1206 ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1207 if (!ic)
1208 return -ENOMEM;
1209
1210 ret = rds_ib_recv_alloc_caches(ic, gfp);
1211 if (ret) {
1212 kfree(ic);
1213 return ret;
1214 }
1215
1216 INIT_LIST_HEAD(&ic->ib_node);
1217 tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1218 (unsigned long)ic);
1219 tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1220 (unsigned long)ic);
1221 mutex_init(&ic->i_recv_mutex);
1222#ifndef KERNEL_HAS_ATOMIC64
1223 spin_lock_init(&ic->i_ack_lock);
1224#endif
1225 atomic_set(&ic->i_signaled_sends, 0);
1226 atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
1227
1228 /*
1229 * rds_ib_conn_shutdown() waits for these to be emptied so they
1230 * must be initialized before it can be called.
1231 */
1232 rds_ib_ring_init(&ic->i_send_ring, 0);
1233 rds_ib_ring_init(&ic->i_recv_ring, 0);
1234
1235 ic->conn = conn;
1236 conn->c_transport_data = ic;
1237
1238 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1239 list_add_tail(&ic->ib_node, &ib_nodev_conns);
1240 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1241
1242
1243 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1244 return 0;
1245}
1246
1247/*
1248 * Free a connection. Connection must be shut down and not set for reconnect.
1249 */
1250void rds_ib_conn_free(void *arg)
1251{
1252 struct rds_ib_connection *ic = arg;
1253 spinlock_t *lock_ptr;
1254
1255 rdsdebug("ic %p\n", ic);
1256
1257 /*
1258 * Conn is either on a dev's list or on the nodev list.
1259 * A race with shutdown() or connect() would cause problems
1260 * (since rds_ibdev would change) but that should never happen.
1261 */
1262 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1263
1264 spin_lock_irq(lock_ptr);
1265 list_del(&ic->ib_node);
1266 spin_unlock_irq(lock_ptr);
1267
1268 rds_ib_recv_free_caches(ic);
1269
1270 kfree(ic);
1271}
1272
1273
1274/*
1275 * An error occurred on the connection
1276 */
1277void
1278__rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1279{
1280 va_list ap;
1281
1282 rds_conn_drop(conn);
1283
1284 va_start(ap, fmt);
1285 vprintk(fmt, ap);
1286 va_end(ap);
1287}
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/in.h>
35#include <linux/slab.h>
36#include <linux/vmalloc.h>
37#include <linux/ratelimit.h>
38
39#include "rds_single_path.h"
40#include "rds.h"
41#include "ib.h"
42
43/*
44 * Set the selected protocol version
45 */
46static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
47{
48 conn->c_version = version;
49}
50
51/*
52 * Set up flow control
53 */
54static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
55{
56 struct rds_ib_connection *ic = conn->c_transport_data;
57
58 if (rds_ib_sysctl_flow_control && credits != 0) {
59 /* We're doing flow control */
60 ic->i_flowctl = 1;
61 rds_ib_send_add_credits(conn, credits);
62 } else {
63 ic->i_flowctl = 0;
64 }
65}
66
67/*
68 * Tune RNR behavior. Without flow control, we use a rather
69 * low timeout, but not the absolute minimum - this should
70 * be tunable.
71 *
72 * We already set the RNR retry count to 7 (which is the
73 * smallest infinite number :-) above.
74 * If flow control is off, we want to change this back to 0
75 * so that we learn quickly when our credit accounting is
76 * buggy.
77 *
78 * Caller passes in a qp_attr pointer - don't waste stack spacv
79 * by allocation this twice.
80 */
81static void
82rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
83{
84 int ret;
85
86 attr->min_rnr_timer = IB_RNR_TIMER_000_32;
87 ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
88 if (ret)
89 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
90}
91
92/*
93 * Connection established.
94 * We get here for both outgoing and incoming connection.
95 */
96void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
97{
98 const struct rds_ib_connect_private *dp = NULL;
99 struct rds_ib_connection *ic = conn->c_transport_data;
100 struct ib_qp_attr qp_attr;
101 int err;
102
103 if (event->param.conn.private_data_len >= sizeof(*dp)) {
104 dp = event->param.conn.private_data;
105
106 /* make sure it isn't empty data */
107 if (dp->dp_protocol_major) {
108 rds_ib_set_protocol(conn,
109 RDS_PROTOCOL(dp->dp_protocol_major,
110 dp->dp_protocol_minor));
111 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
112 }
113 }
114
115 if (conn->c_version < RDS_PROTOCOL(3, 1)) {
116 printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed,"
117 " no longer supported\n",
118 &conn->c_faddr,
119 RDS_PROTOCOL_MAJOR(conn->c_version),
120 RDS_PROTOCOL_MINOR(conn->c_version));
121 rds_conn_destroy(conn);
122 return;
123 } else {
124 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
125 &conn->c_faddr,
126 RDS_PROTOCOL_MAJOR(conn->c_version),
127 RDS_PROTOCOL_MINOR(conn->c_version),
128 ic->i_flowctl ? ", flow control" : "");
129 }
130
131 /*
132 * Init rings and fill recv. this needs to wait until protocol negotiation
133 * is complete, since ring layout is different from 3.0 to 3.1.
134 */
135 rds_ib_send_init_ring(ic);
136 rds_ib_recv_init_ring(ic);
137 /* Post receive buffers - as a side effect, this will update
138 * the posted credit count. */
139 rds_ib_recv_refill(conn, 1, GFP_KERNEL);
140
141 /* Tune RNR behavior */
142 rds_ib_tune_rnr(ic, &qp_attr);
143
144 qp_attr.qp_state = IB_QPS_RTS;
145 err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
146 if (err)
147 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
148
149 /* update ib_device with this local ipaddr */
150 err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
151 if (err)
152 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
153 err);
154
155 /* If the peer gave us the last packet it saw, process this as if
156 * we had received a regular ACK. */
157 if (dp) {
158 /* dp structure start is not guaranteed to be 8 bytes aligned.
159 * Since dp_ack_seq is 64-bit extended load operations can be
160 * used so go through get_unaligned to avoid unaligned errors.
161 */
162 __be64 dp_ack_seq = get_unaligned(&dp->dp_ack_seq);
163
164 if (dp_ack_seq)
165 rds_send_drop_acked(conn, be64_to_cpu(dp_ack_seq),
166 NULL);
167 }
168
169 rds_connect_complete(conn);
170}
171
172static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
173 struct rdma_conn_param *conn_param,
174 struct rds_ib_connect_private *dp,
175 u32 protocol_version,
176 u32 max_responder_resources,
177 u32 max_initiator_depth)
178{
179 struct rds_ib_connection *ic = conn->c_transport_data;
180 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
181
182 memset(conn_param, 0, sizeof(struct rdma_conn_param));
183
184 conn_param->responder_resources =
185 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
186 conn_param->initiator_depth =
187 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
188 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
189 conn_param->rnr_retry_count = 7;
190
191 if (dp) {
192 memset(dp, 0, sizeof(*dp));
193 dp->dp_saddr = conn->c_laddr;
194 dp->dp_daddr = conn->c_faddr;
195 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
196 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
197 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
198 dp->dp_ack_seq = cpu_to_be64(rds_ib_piggyb_ack(ic));
199
200 /* Advertise flow control */
201 if (ic->i_flowctl) {
202 unsigned int credits;
203
204 credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
205 dp->dp_credit = cpu_to_be32(credits);
206 atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
207 }
208
209 conn_param->private_data = dp;
210 conn_param->private_data_len = sizeof(*dp);
211 }
212}
213
214static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
215{
216 rdsdebug("event %u (%s) data %p\n",
217 event->event, ib_event_msg(event->event), data);
218}
219
220/* Plucking the oldest entry from the ring can be done concurrently with
221 * the thread refilling the ring. Each ring operation is protected by
222 * spinlocks and the transient state of refilling doesn't change the
223 * recording of which entry is oldest.
224 *
225 * This relies on IB only calling one cq comp_handler for each cq so that
226 * there will only be one caller of rds_recv_incoming() per RDS connection.
227 */
228static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
229{
230 struct rds_connection *conn = context;
231 struct rds_ib_connection *ic = conn->c_transport_data;
232
233 rdsdebug("conn %p cq %p\n", conn, cq);
234
235 rds_ib_stats_inc(s_ib_evt_handler_call);
236
237 tasklet_schedule(&ic->i_recv_tasklet);
238}
239
240static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
241 struct ib_wc *wcs)
242{
243 int nr, i;
244 struct ib_wc *wc;
245
246 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
247 for (i = 0; i < nr; i++) {
248 wc = wcs + i;
249 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
250 (unsigned long long)wc->wr_id, wc->status,
251 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
252
253 if (wc->wr_id <= ic->i_send_ring.w_nr ||
254 wc->wr_id == RDS_IB_ACK_WR_ID)
255 rds_ib_send_cqe_handler(ic, wc);
256 else
257 rds_ib_mr_cqe_handler(ic, wc);
258
259 }
260 }
261}
262
263static void rds_ib_tasklet_fn_send(unsigned long data)
264{
265 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
266 struct rds_connection *conn = ic->conn;
267
268 rds_ib_stats_inc(s_ib_tasklet_call);
269
270 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
271 ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
272 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
273
274 if (rds_conn_up(conn) &&
275 (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
276 test_bit(0, &conn->c_map_queued)))
277 rds_send_xmit(&ic->conn->c_path[0]);
278}
279
280static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
281 struct ib_wc *wcs,
282 struct rds_ib_ack_state *ack_state)
283{
284 int nr, i;
285 struct ib_wc *wc;
286
287 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
288 for (i = 0; i < nr; i++) {
289 wc = wcs + i;
290 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
291 (unsigned long long)wc->wr_id, wc->status,
292 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
293
294 rds_ib_recv_cqe_handler(ic, wc, ack_state);
295 }
296 }
297}
298
299static void rds_ib_tasklet_fn_recv(unsigned long data)
300{
301 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
302 struct rds_connection *conn = ic->conn;
303 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
304 struct rds_ib_ack_state state;
305
306 if (!rds_ibdev)
307 rds_conn_drop(conn);
308
309 rds_ib_stats_inc(s_ib_tasklet_call);
310
311 memset(&state, 0, sizeof(state));
312 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
313 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
314 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
315
316 if (state.ack_next_valid)
317 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
318 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
319 rds_send_drop_acked(conn, state.ack_recv, NULL);
320 ic->i_ack_recv = state.ack_recv;
321 }
322
323 if (rds_conn_up(conn))
324 rds_ib_attempt_ack(ic);
325}
326
327static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
328{
329 struct rds_connection *conn = data;
330 struct rds_ib_connection *ic = conn->c_transport_data;
331
332 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
333 ib_event_msg(event->event));
334
335 switch (event->event) {
336 case IB_EVENT_COMM_EST:
337 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
338 break;
339 default:
340 rdsdebug("Fatal QP Event %u (%s) "
341 "- connection %pI4->%pI4, reconnecting\n",
342 event->event, ib_event_msg(event->event),
343 &conn->c_laddr, &conn->c_faddr);
344 rds_conn_drop(conn);
345 break;
346 }
347}
348
349static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
350{
351 struct rds_connection *conn = context;
352 struct rds_ib_connection *ic = conn->c_transport_data;
353
354 rdsdebug("conn %p cq %p\n", conn, cq);
355
356 rds_ib_stats_inc(s_ib_evt_handler_call);
357
358 tasklet_schedule(&ic->i_send_tasklet);
359}
360
361/*
362 * This needs to be very careful to not leave IS_ERR pointers around for
363 * cleanup to trip over.
364 */
365static int rds_ib_setup_qp(struct rds_connection *conn)
366{
367 struct rds_ib_connection *ic = conn->c_transport_data;
368 struct ib_device *dev = ic->i_cm_id->device;
369 struct ib_qp_init_attr attr;
370 struct ib_cq_init_attr cq_attr = {};
371 struct rds_ib_device *rds_ibdev;
372 int ret, fr_queue_space;
373
374 /*
375 * It's normal to see a null device if an incoming connection races
376 * with device removal, so we don't print a warning.
377 */
378 rds_ibdev = rds_ib_get_client_data(dev);
379 if (!rds_ibdev)
380 return -EOPNOTSUPP;
381
382 /* The fr_queue_space is currently set to 512, to add extra space on
383 * completion queue and send queue. This extra space is used for FRMR
384 * registration and invalidation work requests
385 */
386 fr_queue_space = (rds_ibdev->use_fastreg ? RDS_IB_DEFAULT_FR_WR : 0);
387
388 /* add the conn now so that connection establishment has the dev */
389 rds_ib_add_conn(rds_ibdev, conn);
390
391 if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
392 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
393 if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
394 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
395
396 /* Protection domain and memory range */
397 ic->i_pd = rds_ibdev->pd;
398
399 cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
400
401 ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
402 rds_ib_cq_event_handler, conn,
403 &cq_attr);
404 if (IS_ERR(ic->i_send_cq)) {
405 ret = PTR_ERR(ic->i_send_cq);
406 ic->i_send_cq = NULL;
407 rdsdebug("ib_create_cq send failed: %d\n", ret);
408 goto out;
409 }
410
411 cq_attr.cqe = ic->i_recv_ring.w_nr;
412 ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
413 rds_ib_cq_event_handler, conn,
414 &cq_attr);
415 if (IS_ERR(ic->i_recv_cq)) {
416 ret = PTR_ERR(ic->i_recv_cq);
417 ic->i_recv_cq = NULL;
418 rdsdebug("ib_create_cq recv failed: %d\n", ret);
419 goto out;
420 }
421
422 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
423 if (ret) {
424 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
425 goto out;
426 }
427
428 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
429 if (ret) {
430 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
431 goto out;
432 }
433
434 /* XXX negotiate max send/recv with remote? */
435 memset(&attr, 0, sizeof(attr));
436 attr.event_handler = rds_ib_qp_event_handler;
437 attr.qp_context = conn;
438 /* + 1 to allow for the single ack message */
439 attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
440 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
441 attr.cap.max_send_sge = rds_ibdev->max_sge;
442 attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
443 attr.sq_sig_type = IB_SIGNAL_REQ_WR;
444 attr.qp_type = IB_QPT_RC;
445 attr.send_cq = ic->i_send_cq;
446 attr.recv_cq = ic->i_recv_cq;
447 atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
448
449 /*
450 * XXX this can fail if max_*_wr is too large? Are we supposed
451 * to back off until we get a value that the hardware can support?
452 */
453 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
454 if (ret) {
455 rdsdebug("rdma_create_qp failed: %d\n", ret);
456 goto out;
457 }
458
459 ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
460 ic->i_send_ring.w_nr *
461 sizeof(struct rds_header),
462 &ic->i_send_hdrs_dma, GFP_KERNEL);
463 if (!ic->i_send_hdrs) {
464 ret = -ENOMEM;
465 rdsdebug("ib_dma_alloc_coherent send failed\n");
466 goto out;
467 }
468
469 ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
470 ic->i_recv_ring.w_nr *
471 sizeof(struct rds_header),
472 &ic->i_recv_hdrs_dma, GFP_KERNEL);
473 if (!ic->i_recv_hdrs) {
474 ret = -ENOMEM;
475 rdsdebug("ib_dma_alloc_coherent recv failed\n");
476 goto out;
477 }
478
479 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
480 &ic->i_ack_dma, GFP_KERNEL);
481 if (!ic->i_ack) {
482 ret = -ENOMEM;
483 rdsdebug("ib_dma_alloc_coherent ack failed\n");
484 goto out;
485 }
486
487 ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
488 ibdev_to_node(dev));
489 if (!ic->i_sends) {
490 ret = -ENOMEM;
491 rdsdebug("send allocation failed\n");
492 goto out;
493 }
494
495 ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
496 ibdev_to_node(dev));
497 if (!ic->i_recvs) {
498 ret = -ENOMEM;
499 rdsdebug("recv allocation failed\n");
500 goto out;
501 }
502
503 rds_ib_recv_init_ack(ic);
504
505 rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
506 ic->i_send_cq, ic->i_recv_cq);
507
508out:
509 rds_ib_dev_put(rds_ibdev);
510 return ret;
511}
512
513static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
514{
515 const struct rds_ib_connect_private *dp = event->param.conn.private_data;
516 u16 common;
517 u32 version = 0;
518
519 /*
520 * rdma_cm private data is odd - when there is any private data in the
521 * request, we will be given a pretty large buffer without telling us the
522 * original size. The only way to tell the difference is by looking at
523 * the contents, which are initialized to zero.
524 * If the protocol version fields aren't set, this is a connection attempt
525 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
526 * We really should have changed this for OFED 1.3 :-(
527 */
528
529 /* Be paranoid. RDS always has privdata */
530 if (!event->param.conn.private_data_len) {
531 printk(KERN_NOTICE "RDS incoming connection has no private data, "
532 "rejecting\n");
533 return 0;
534 }
535
536 /* Even if len is crap *now* I still want to check it. -ASG */
537 if (event->param.conn.private_data_len < sizeof (*dp) ||
538 dp->dp_protocol_major == 0)
539 return RDS_PROTOCOL_3_0;
540
541 common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
542 if (dp->dp_protocol_major == 3 && common) {
543 version = RDS_PROTOCOL_3_0;
544 while ((common >>= 1) != 0)
545 version++;
546 } else
547 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
548 &dp->dp_saddr,
549 dp->dp_protocol_major,
550 dp->dp_protocol_minor);
551 return version;
552}
553
554int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
555 struct rdma_cm_event *event)
556{
557 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
558 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
559 const struct rds_ib_connect_private *dp = event->param.conn.private_data;
560 struct rds_ib_connect_private dp_rep;
561 struct rds_connection *conn = NULL;
562 struct rds_ib_connection *ic = NULL;
563 struct rdma_conn_param conn_param;
564 u32 version;
565 int err = 1, destroy = 1;
566
567 /* Check whether the remote protocol version matches ours. */
568 version = rds_ib_protocol_compatible(event);
569 if (!version)
570 goto out;
571
572 rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
573 "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
574 RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
575 (unsigned long long)be64_to_cpu(lguid),
576 (unsigned long long)be64_to_cpu(fguid));
577
578 /* RDS/IB is not currently netns aware, thus init_net */
579 conn = rds_conn_create(&init_net, dp->dp_daddr, dp->dp_saddr,
580 &rds_ib_transport, GFP_KERNEL);
581 if (IS_ERR(conn)) {
582 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
583 conn = NULL;
584 goto out;
585 }
586
587 /*
588 * The connection request may occur while the
589 * previous connection exist, e.g. in case of failover.
590 * But as connections may be initiated simultaneously
591 * by both hosts, we have a random backoff mechanism -
592 * see the comment above rds_queue_reconnect()
593 */
594 mutex_lock(&conn->c_cm_lock);
595 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
596 if (rds_conn_state(conn) == RDS_CONN_UP) {
597 rdsdebug("incoming connect while connecting\n");
598 rds_conn_drop(conn);
599 rds_ib_stats_inc(s_ib_listen_closed_stale);
600 } else
601 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
602 /* Wait and see - our connect may still be succeeding */
603 rds_ib_stats_inc(s_ib_connect_raced);
604 }
605 goto out;
606 }
607
608 ic = conn->c_transport_data;
609
610 rds_ib_set_protocol(conn, version);
611 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
612
613 /* If the peer gave us the last packet it saw, process this as if
614 * we had received a regular ACK. */
615 if (dp->dp_ack_seq)
616 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
617
618 BUG_ON(cm_id->context);
619 BUG_ON(ic->i_cm_id);
620
621 ic->i_cm_id = cm_id;
622 cm_id->context = conn;
623
624 /* We got halfway through setting up the ib_connection, if we
625 * fail now, we have to take the long route out of this mess. */
626 destroy = 0;
627
628 err = rds_ib_setup_qp(conn);
629 if (err) {
630 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
631 goto out;
632 }
633
634 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
635 event->param.conn.responder_resources,
636 event->param.conn.initiator_depth);
637
638 /* rdma_accept() calls rdma_reject() internally if it fails */
639 err = rdma_accept(cm_id, &conn_param);
640 if (err)
641 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
642
643out:
644 if (conn)
645 mutex_unlock(&conn->c_cm_lock);
646 if (err)
647 rdma_reject(cm_id, NULL, 0);
648 return destroy;
649}
650
651
652int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
653{
654 struct rds_connection *conn = cm_id->context;
655 struct rds_ib_connection *ic = conn->c_transport_data;
656 struct rdma_conn_param conn_param;
657 struct rds_ib_connect_private dp;
658 int ret;
659
660 /* If the peer doesn't do protocol negotiation, we must
661 * default to RDSv3.0 */
662 rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
663 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
664
665 ret = rds_ib_setup_qp(conn);
666 if (ret) {
667 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
668 goto out;
669 }
670
671 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
672 UINT_MAX, UINT_MAX);
673 ret = rdma_connect(cm_id, &conn_param);
674 if (ret)
675 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
676
677out:
678 /* Beware - returning non-zero tells the rdma_cm to destroy
679 * the cm_id. We should certainly not do it as long as we still
680 * "own" the cm_id. */
681 if (ret) {
682 if (ic->i_cm_id == cm_id)
683 ret = 0;
684 }
685 return ret;
686}
687
688int rds_ib_conn_path_connect(struct rds_conn_path *cp)
689{
690 struct rds_connection *conn = cp->cp_conn;
691 struct rds_ib_connection *ic = conn->c_transport_data;
692 struct sockaddr_in src, dest;
693 int ret;
694
695 /* XXX I wonder what affect the port space has */
696 /* delegate cm event handler to rdma_transport */
697 ic->i_cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler, conn,
698 RDMA_PS_TCP, IB_QPT_RC);
699 if (IS_ERR(ic->i_cm_id)) {
700 ret = PTR_ERR(ic->i_cm_id);
701 ic->i_cm_id = NULL;
702 rdsdebug("rdma_create_id() failed: %d\n", ret);
703 goto out;
704 }
705
706 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
707
708 src.sin_family = AF_INET;
709 src.sin_addr.s_addr = (__force u32)conn->c_laddr;
710 src.sin_port = (__force u16)htons(0);
711
712 dest.sin_family = AF_INET;
713 dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
714 dest.sin_port = (__force u16)htons(RDS_PORT);
715
716 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
717 (struct sockaddr *)&dest,
718 RDS_RDMA_RESOLVE_TIMEOUT_MS);
719 if (ret) {
720 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
721 ret);
722 rdma_destroy_id(ic->i_cm_id);
723 ic->i_cm_id = NULL;
724 }
725
726out:
727 return ret;
728}
729
730/*
731 * This is so careful about only cleaning up resources that were built up
732 * so that it can be called at any point during startup. In fact it
733 * can be called multiple times for a given connection.
734 */
735void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
736{
737 struct rds_connection *conn = cp->cp_conn;
738 struct rds_ib_connection *ic = conn->c_transport_data;
739 int err = 0;
740
741 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
742 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
743 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
744
745 if (ic->i_cm_id) {
746 struct ib_device *dev = ic->i_cm_id->device;
747
748 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
749 err = rdma_disconnect(ic->i_cm_id);
750 if (err) {
751 /* Actually this may happen quite frequently, when
752 * an outgoing connect raced with an incoming connect.
753 */
754 rdsdebug("failed to disconnect, cm: %p err %d\n",
755 ic->i_cm_id, err);
756 }
757
758 /*
759 * We want to wait for tx and rx completion to finish
760 * before we tear down the connection, but we have to be
761 * careful not to get stuck waiting on a send ring that
762 * only has unsignaled sends in it. We've shutdown new
763 * sends before getting here so by waiting for signaled
764 * sends to complete we're ensured that there will be no
765 * more tx processing.
766 */
767 wait_event(rds_ib_ring_empty_wait,
768 rds_ib_ring_empty(&ic->i_recv_ring) &&
769 (atomic_read(&ic->i_signaled_sends) == 0) &&
770 (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
771 tasklet_kill(&ic->i_send_tasklet);
772 tasklet_kill(&ic->i_recv_tasklet);
773
774 /* first destroy the ib state that generates callbacks */
775 if (ic->i_cm_id->qp)
776 rdma_destroy_qp(ic->i_cm_id);
777 if (ic->i_send_cq)
778 ib_destroy_cq(ic->i_send_cq);
779 if (ic->i_recv_cq)
780 ib_destroy_cq(ic->i_recv_cq);
781
782 /* then free the resources that ib callbacks use */
783 if (ic->i_send_hdrs)
784 ib_dma_free_coherent(dev,
785 ic->i_send_ring.w_nr *
786 sizeof(struct rds_header),
787 ic->i_send_hdrs,
788 ic->i_send_hdrs_dma);
789
790 if (ic->i_recv_hdrs)
791 ib_dma_free_coherent(dev,
792 ic->i_recv_ring.w_nr *
793 sizeof(struct rds_header),
794 ic->i_recv_hdrs,
795 ic->i_recv_hdrs_dma);
796
797 if (ic->i_ack)
798 ib_dma_free_coherent(dev, sizeof(struct rds_header),
799 ic->i_ack, ic->i_ack_dma);
800
801 if (ic->i_sends)
802 rds_ib_send_clear_ring(ic);
803 if (ic->i_recvs)
804 rds_ib_recv_clear_ring(ic);
805
806 rdma_destroy_id(ic->i_cm_id);
807
808 /*
809 * Move connection back to the nodev list.
810 */
811 if (ic->rds_ibdev)
812 rds_ib_remove_conn(ic->rds_ibdev, conn);
813
814 ic->i_cm_id = NULL;
815 ic->i_pd = NULL;
816 ic->i_send_cq = NULL;
817 ic->i_recv_cq = NULL;
818 ic->i_send_hdrs = NULL;
819 ic->i_recv_hdrs = NULL;
820 ic->i_ack = NULL;
821 }
822 BUG_ON(ic->rds_ibdev);
823
824 /* Clear pending transmit */
825 if (ic->i_data_op) {
826 struct rds_message *rm;
827
828 rm = container_of(ic->i_data_op, struct rds_message, data);
829 rds_message_put(rm);
830 ic->i_data_op = NULL;
831 }
832
833 /* Clear the ACK state */
834 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
835#ifdef KERNEL_HAS_ATOMIC64
836 atomic64_set(&ic->i_ack_next, 0);
837#else
838 ic->i_ack_next = 0;
839#endif
840 ic->i_ack_recv = 0;
841
842 /* Clear flow control state */
843 ic->i_flowctl = 0;
844 atomic_set(&ic->i_credits, 0);
845
846 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
847 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
848
849 if (ic->i_ibinc) {
850 rds_inc_put(&ic->i_ibinc->ii_inc);
851 ic->i_ibinc = NULL;
852 }
853
854 vfree(ic->i_sends);
855 ic->i_sends = NULL;
856 vfree(ic->i_recvs);
857 ic->i_recvs = NULL;
858}
859
860int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
861{
862 struct rds_ib_connection *ic;
863 unsigned long flags;
864 int ret;
865
866 /* XXX too lazy? */
867 ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
868 if (!ic)
869 return -ENOMEM;
870
871 ret = rds_ib_recv_alloc_caches(ic);
872 if (ret) {
873 kfree(ic);
874 return ret;
875 }
876
877 INIT_LIST_HEAD(&ic->ib_node);
878 tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
879 (unsigned long)ic);
880 tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
881 (unsigned long)ic);
882 mutex_init(&ic->i_recv_mutex);
883#ifndef KERNEL_HAS_ATOMIC64
884 spin_lock_init(&ic->i_ack_lock);
885#endif
886 atomic_set(&ic->i_signaled_sends, 0);
887
888 /*
889 * rds_ib_conn_shutdown() waits for these to be emptied so they
890 * must be initialized before it can be called.
891 */
892 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
893 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
894
895 ic->conn = conn;
896 conn->c_transport_data = ic;
897
898 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
899 list_add_tail(&ic->ib_node, &ib_nodev_conns);
900 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
901
902
903 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
904 return 0;
905}
906
907/*
908 * Free a connection. Connection must be shut down and not set for reconnect.
909 */
910void rds_ib_conn_free(void *arg)
911{
912 struct rds_ib_connection *ic = arg;
913 spinlock_t *lock_ptr;
914
915 rdsdebug("ic %p\n", ic);
916
917 /*
918 * Conn is either on a dev's list or on the nodev list.
919 * A race with shutdown() or connect() would cause problems
920 * (since rds_ibdev would change) but that should never happen.
921 */
922 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
923
924 spin_lock_irq(lock_ptr);
925 list_del(&ic->ib_node);
926 spin_unlock_irq(lock_ptr);
927
928 rds_ib_recv_free_caches(ic);
929
930 kfree(ic);
931}
932
933
934/*
935 * An error occurred on the connection
936 */
937void
938__rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
939{
940 va_list ap;
941
942 rds_conn_drop(conn);
943
944 va_start(ap, fmt);
945 vprintk(fmt, ap);
946 va_end(ap);
947}