xref: /openbmc/linux/net/smc/smc_ib.c (revision 227a0655)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
4  *
5  *  IB infrastructure:
6  *  Establish SMC-R as an Infiniband Client to be notified about added and
7  *  removed IB devices of type RDMA.
8  *  Determine device and port characteristics for these IB devices.
9  *
10  *  Copyright IBM Corp. 2016
11  *
12  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
13  */
14 
15 #include <linux/etherdevice.h>
16 #include <linux/if_vlan.h>
17 #include <linux/random.h>
18 #include <linux/workqueue.h>
19 #include <linux/scatterlist.h>
20 #include <linux/wait.h>
21 #include <linux/mutex.h>
22 #include <linux/inetdevice.h>
23 #include <rdma/ib_verbs.h>
24 #include <rdma/ib_cache.h>
25 
26 #include "smc_pnet.h"
27 #include "smc_ib.h"
28 #include "smc_core.h"
29 #include "smc_wr.h"
30 #include "smc.h"
31 #include "smc_netlink.h"
32 
33 #define SMC_MAX_CQE 32766	/* max. # of completion queue elements */
34 
35 #define SMC_QP_MIN_RNR_TIMER		5
36 #define SMC_QP_TIMEOUT			15 /* 4096 * 2 ** timeout usec */
37 #define SMC_QP_RETRY_CNT			7 /* 7: infinite */
38 #define SMC_QP_RNR_RETRY			7 /* 7: infinite */
39 
40 struct smc_ib_devices smc_ib_devices = {	/* smc-registered ib devices */
41 	.mutex = __MUTEX_INITIALIZER(smc_ib_devices.mutex),
42 	.list = LIST_HEAD_INIT(smc_ib_devices.list),
43 };
44 
45 u8 local_systemid[SMC_SYSTEMID_LEN];		/* unique system identifier */
46 
47 static int smc_ib_modify_qp_init(struct smc_link *lnk)
48 {
49 	struct ib_qp_attr qp_attr;
50 
51 	memset(&qp_attr, 0, sizeof(qp_attr));
52 	qp_attr.qp_state = IB_QPS_INIT;
53 	qp_attr.pkey_index = 0;
54 	qp_attr.port_num = lnk->ibport;
55 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
56 				| IB_ACCESS_REMOTE_WRITE;
57 	return ib_modify_qp(lnk->roce_qp, &qp_attr,
58 			    IB_QP_STATE | IB_QP_PKEY_INDEX |
59 			    IB_QP_ACCESS_FLAGS | IB_QP_PORT);
60 }
61 
62 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
63 {
64 	enum ib_qp_attr_mask qp_attr_mask =
65 		IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
66 		IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
67 	struct ib_qp_attr qp_attr;
68 	u8 hop_lim = 1;
69 
70 	memset(&qp_attr, 0, sizeof(qp_attr));
71 	qp_attr.qp_state = IB_QPS_RTR;
72 	qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
73 	qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
74 	rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
75 	if (lnk->lgr->smc_version == SMC_V2 && lnk->lgr->uses_gateway)
76 		hop_lim = IPV6_DEFAULT_HOPLIMIT;
77 	rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, lnk->sgid_index, hop_lim, 0);
78 	rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
79 	if (lnk->lgr->smc_version == SMC_V2 && lnk->lgr->uses_gateway)
80 		memcpy(&qp_attr.ah_attr.roce.dmac, lnk->lgr->nexthop_mac,
81 		       sizeof(lnk->lgr->nexthop_mac));
82 	else
83 		memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
84 		       sizeof(lnk->peer_mac));
85 	qp_attr.dest_qp_num = lnk->peer_qpn;
86 	qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
87 	qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
88 					 * requests
89 					 */
90 	qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
91 
92 	return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
93 }
94 
95 int smc_ib_modify_qp_rts(struct smc_link *lnk)
96 {
97 	struct ib_qp_attr qp_attr;
98 
99 	memset(&qp_attr, 0, sizeof(qp_attr));
100 	qp_attr.qp_state = IB_QPS_RTS;
101 	qp_attr.timeout = SMC_QP_TIMEOUT;	/* local ack timeout */
102 	qp_attr.retry_cnt = SMC_QP_RETRY_CNT;	/* retry count */
103 	qp_attr.rnr_retry = SMC_QP_RNR_RETRY;	/* RNR retries, 7=infinite */
104 	qp_attr.sq_psn = lnk->psn_initial;	/* starting send packet seq # */
105 	qp_attr.max_rd_atomic = 1;	/* # of outstanding RDMA reads and
106 					 * atomic ops allowed
107 					 */
108 	return ib_modify_qp(lnk->roce_qp, &qp_attr,
109 			    IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
110 			    IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
111 			    IB_QP_MAX_QP_RD_ATOMIC);
112 }
113 
114 int smc_ib_modify_qp_error(struct smc_link *lnk)
115 {
116 	struct ib_qp_attr qp_attr;
117 
118 	memset(&qp_attr, 0, sizeof(qp_attr));
119 	qp_attr.qp_state = IB_QPS_ERR;
120 	return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
121 }
122 
123 int smc_ib_ready_link(struct smc_link *lnk)
124 {
125 	struct smc_link_group *lgr = smc_get_lgr(lnk);
126 	int rc = 0;
127 
128 	rc = smc_ib_modify_qp_init(lnk);
129 	if (rc)
130 		goto out;
131 
132 	rc = smc_ib_modify_qp_rtr(lnk);
133 	if (rc)
134 		goto out;
135 	smc_wr_remember_qp_attr(lnk);
136 	rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
137 			      IB_CQ_SOLICITED_MASK);
138 	if (rc)
139 		goto out;
140 	rc = smc_wr_rx_post_init(lnk);
141 	if (rc)
142 		goto out;
143 	smc_wr_remember_qp_attr(lnk);
144 
145 	if (lgr->role == SMC_SERV) {
146 		rc = smc_ib_modify_qp_rts(lnk);
147 		if (rc)
148 			goto out;
149 		smc_wr_remember_qp_attr(lnk);
150 	}
151 out:
152 	return rc;
153 }
154 
155 static int smc_ib_fill_mac(struct smc_ib_device *smcibdev, u8 ibport)
156 {
157 	const struct ib_gid_attr *attr;
158 	int rc;
159 
160 	attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, 0);
161 	if (IS_ERR(attr))
162 		return -ENODEV;
163 
164 	rc = rdma_read_gid_l2_fields(attr, NULL, smcibdev->mac[ibport - 1]);
165 	rdma_put_gid_attr(attr);
166 	return rc;
167 }
168 
169 /* Create an identifier unique for this instance of SMC-R.
170  * The MAC-address of the first active registered IB device
171  * plus a random 2-byte number is used to create this identifier.
172  * This name is delivered to the peer during connection initialization.
173  */
174 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
175 						u8 ibport)
176 {
177 	memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
178 	       sizeof(smcibdev->mac[ibport - 1]));
179 }
180 
181 bool smc_ib_is_valid_local_systemid(void)
182 {
183 	return !is_zero_ether_addr(&local_systemid[2]);
184 }
185 
186 static void smc_ib_init_local_systemid(void)
187 {
188 	get_random_bytes(&local_systemid[0], 2);
189 }
190 
191 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
192 {
193 	return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
194 }
195 
196 int smc_ib_find_route(__be32 saddr, __be32 daddr,
197 		      u8 nexthop_mac[], u8 *uses_gateway)
198 {
199 	struct neighbour *neigh = NULL;
200 	struct rtable *rt = NULL;
201 	struct flowi4 fl4 = {
202 		.saddr = saddr,
203 		.daddr = daddr
204 	};
205 
206 	if (daddr == cpu_to_be32(INADDR_NONE))
207 		goto out;
208 	rt = ip_route_output_flow(&init_net, &fl4, NULL);
209 	if (IS_ERR(rt))
210 		goto out;
211 	if (rt->rt_uses_gateway && rt->rt_gw_family != AF_INET)
212 		goto out;
213 	neigh = rt->dst.ops->neigh_lookup(&rt->dst, NULL, &fl4.daddr);
214 	if (neigh) {
215 		memcpy(nexthop_mac, neigh->ha, ETH_ALEN);
216 		*uses_gateway = rt->rt_uses_gateway;
217 		return 0;
218 	}
219 out:
220 	return -ENOENT;
221 }
222 
223 static int smc_ib_determine_gid_rcu(const struct net_device *ndev,
224 				    const struct ib_gid_attr *attr,
225 				    u8 gid[], u8 *sgid_index,
226 				    struct smc_init_info_smcrv2 *smcrv2)
227 {
228 	if (!smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE) {
229 		if (gid)
230 			memcpy(gid, &attr->gid, SMC_GID_SIZE);
231 		if (sgid_index)
232 			*sgid_index = attr->index;
233 		return 0;
234 	}
235 	if (smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP &&
236 	    smc_ib_gid_to_ipv4((u8 *)&attr->gid) != cpu_to_be32(INADDR_NONE)) {
237 		struct in_device *in_dev = __in_dev_get_rcu(ndev);
238 		const struct in_ifaddr *ifa;
239 		bool subnet_match = false;
240 
241 		if (!in_dev)
242 			goto out;
243 		in_dev_for_each_ifa_rcu(ifa, in_dev) {
244 			if (!inet_ifa_match(smcrv2->saddr, ifa))
245 				continue;
246 			subnet_match = true;
247 			break;
248 		}
249 		if (!subnet_match)
250 			goto out;
251 		if (smcrv2->daddr && smc_ib_find_route(smcrv2->saddr,
252 						       smcrv2->daddr,
253 						       smcrv2->nexthop_mac,
254 						       &smcrv2->uses_gateway))
255 			goto out;
256 
257 		if (gid)
258 			memcpy(gid, &attr->gid, SMC_GID_SIZE);
259 		if (sgid_index)
260 			*sgid_index = attr->index;
261 		return 0;
262 	}
263 out:
264 	return -ENODEV;
265 }
266 
267 /* determine the gid for an ib-device port and vlan id */
268 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
269 			 unsigned short vlan_id, u8 gid[], u8 *sgid_index,
270 			 struct smc_init_info_smcrv2 *smcrv2)
271 {
272 	const struct ib_gid_attr *attr;
273 	const struct net_device *ndev;
274 	int i;
275 
276 	for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
277 		attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
278 		if (IS_ERR(attr))
279 			continue;
280 
281 		rcu_read_lock();
282 		ndev = rdma_read_gid_attr_ndev_rcu(attr);
283 		if (!IS_ERR(ndev) &&
284 		    ((!vlan_id && !is_vlan_dev(ndev)) ||
285 		     (vlan_id && is_vlan_dev(ndev) &&
286 		      vlan_dev_vlan_id(ndev) == vlan_id))) {
287 			if (!smc_ib_determine_gid_rcu(ndev, attr, gid,
288 						      sgid_index, smcrv2)) {
289 				rcu_read_unlock();
290 				rdma_put_gid_attr(attr);
291 				return 0;
292 			}
293 		}
294 		rcu_read_unlock();
295 		rdma_put_gid_attr(attr);
296 	}
297 	return -ENODEV;
298 }
299 
300 /* check if gid is still defined on smcibdev */
301 static bool smc_ib_check_link_gid(u8 gid[SMC_GID_SIZE], bool smcrv2,
302 				  struct smc_ib_device *smcibdev, u8 ibport)
303 {
304 	const struct ib_gid_attr *attr;
305 	bool rc = false;
306 	int i;
307 
308 	for (i = 0; !rc && i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
309 		attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
310 		if (IS_ERR(attr))
311 			continue;
312 
313 		rcu_read_lock();
314 		if ((!smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE) ||
315 		    (smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP &&
316 		     !(ipv6_addr_type((const struct in6_addr *)&attr->gid)
317 				     & IPV6_ADDR_LINKLOCAL)))
318 			if (!memcmp(gid, &attr->gid, SMC_GID_SIZE))
319 				rc = true;
320 		rcu_read_unlock();
321 		rdma_put_gid_attr(attr);
322 	}
323 	return rc;
324 }
325 
326 /* check all links if the gid is still defined on smcibdev */
327 static void smc_ib_gid_check(struct smc_ib_device *smcibdev, u8 ibport)
328 {
329 	struct smc_link_group *lgr;
330 	int i;
331 
332 	spin_lock_bh(&smc_lgr_list.lock);
333 	list_for_each_entry(lgr, &smc_lgr_list.list, list) {
334 		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
335 			    SMC_MAX_PNETID_LEN))
336 			continue; /* lgr is not affected */
337 		if (list_empty(&lgr->list))
338 			continue;
339 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
340 			if (lgr->lnk[i].state == SMC_LNK_UNUSED ||
341 			    lgr->lnk[i].smcibdev != smcibdev)
342 				continue;
343 			if (!smc_ib_check_link_gid(lgr->lnk[i].gid,
344 						   lgr->smc_version == SMC_V2,
345 						   smcibdev, ibport))
346 				smcr_port_err(smcibdev, ibport);
347 		}
348 	}
349 	spin_unlock_bh(&smc_lgr_list.lock);
350 }
351 
352 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
353 {
354 	int rc;
355 
356 	memset(&smcibdev->pattr[ibport - 1], 0,
357 	       sizeof(smcibdev->pattr[ibport - 1]));
358 	rc = ib_query_port(smcibdev->ibdev, ibport,
359 			   &smcibdev->pattr[ibport - 1]);
360 	if (rc)
361 		goto out;
362 	/* the SMC protocol requires specification of the RoCE MAC address */
363 	rc = smc_ib_fill_mac(smcibdev, ibport);
364 	if (rc)
365 		goto out;
366 	if (!smc_ib_is_valid_local_systemid() &&
367 	    smc_ib_port_active(smcibdev, ibport))
368 		/* create unique system identifier */
369 		smc_ib_define_local_systemid(smcibdev, ibport);
370 out:
371 	return rc;
372 }
373 
374 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
375 static void smc_ib_port_event_work(struct work_struct *work)
376 {
377 	struct smc_ib_device *smcibdev = container_of(
378 		work, struct smc_ib_device, port_event_work);
379 	u8 port_idx;
380 
381 	for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
382 		smc_ib_remember_port_attr(smcibdev, port_idx + 1);
383 		clear_bit(port_idx, &smcibdev->port_event_mask);
384 		if (!smc_ib_port_active(smcibdev, port_idx + 1)) {
385 			set_bit(port_idx, smcibdev->ports_going_away);
386 			smcr_port_err(smcibdev, port_idx + 1);
387 		} else {
388 			clear_bit(port_idx, smcibdev->ports_going_away);
389 			smcr_port_add(smcibdev, port_idx + 1);
390 			smc_ib_gid_check(smcibdev, port_idx + 1);
391 		}
392 	}
393 }
394 
395 /* can be called in IRQ context */
396 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
397 					struct ib_event *ibevent)
398 {
399 	struct smc_ib_device *smcibdev;
400 	bool schedule = false;
401 	u8 port_idx;
402 
403 	smcibdev = container_of(handler, struct smc_ib_device, event_handler);
404 
405 	switch (ibevent->event) {
406 	case IB_EVENT_DEVICE_FATAL:
407 		/* terminate all ports on device */
408 		for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) {
409 			set_bit(port_idx, &smcibdev->port_event_mask);
410 			if (!test_and_set_bit(port_idx,
411 					      smcibdev->ports_going_away))
412 				schedule = true;
413 		}
414 		if (schedule)
415 			schedule_work(&smcibdev->port_event_work);
416 		break;
417 	case IB_EVENT_PORT_ACTIVE:
418 		port_idx = ibevent->element.port_num - 1;
419 		if (port_idx >= SMC_MAX_PORTS)
420 			break;
421 		set_bit(port_idx, &smcibdev->port_event_mask);
422 		if (test_and_clear_bit(port_idx, smcibdev->ports_going_away))
423 			schedule_work(&smcibdev->port_event_work);
424 		break;
425 	case IB_EVENT_PORT_ERR:
426 		port_idx = ibevent->element.port_num - 1;
427 		if (port_idx >= SMC_MAX_PORTS)
428 			break;
429 		set_bit(port_idx, &smcibdev->port_event_mask);
430 		if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
431 			schedule_work(&smcibdev->port_event_work);
432 		break;
433 	case IB_EVENT_GID_CHANGE:
434 		port_idx = ibevent->element.port_num - 1;
435 		if (port_idx >= SMC_MAX_PORTS)
436 			break;
437 		set_bit(port_idx, &smcibdev->port_event_mask);
438 		schedule_work(&smcibdev->port_event_work);
439 		break;
440 	default:
441 		break;
442 	}
443 }
444 
445 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
446 {
447 	if (lnk->roce_pd)
448 		ib_dealloc_pd(lnk->roce_pd);
449 	lnk->roce_pd = NULL;
450 }
451 
452 int smc_ib_create_protection_domain(struct smc_link *lnk)
453 {
454 	int rc;
455 
456 	lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
457 	rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
458 	if (IS_ERR(lnk->roce_pd))
459 		lnk->roce_pd = NULL;
460 	return rc;
461 }
462 
463 static bool smcr_diag_is_dev_critical(struct smc_lgr_list *smc_lgr,
464 				      struct smc_ib_device *smcibdev)
465 {
466 	struct smc_link_group *lgr;
467 	bool rc = false;
468 	int i;
469 
470 	spin_lock_bh(&smc_lgr->lock);
471 	list_for_each_entry(lgr, &smc_lgr->list, list) {
472 		if (lgr->is_smcd)
473 			continue;
474 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
475 			if (lgr->lnk[i].state == SMC_LNK_UNUSED ||
476 			    lgr->lnk[i].smcibdev != smcibdev)
477 				continue;
478 			if (lgr->type == SMC_LGR_SINGLE ||
479 			    lgr->type == SMC_LGR_ASYMMETRIC_LOCAL) {
480 				rc = true;
481 				goto out;
482 			}
483 		}
484 	}
485 out:
486 	spin_unlock_bh(&smc_lgr->lock);
487 	return rc;
488 }
489 
490 static int smc_nl_handle_dev_port(struct sk_buff *skb,
491 				  struct ib_device *ibdev,
492 				  struct smc_ib_device *smcibdev,
493 				  int port)
494 {
495 	char smc_pnet[SMC_MAX_PNETID_LEN + 1];
496 	struct nlattr *port_attrs;
497 	unsigned char port_state;
498 	int lnk_count = 0;
499 
500 	port_attrs = nla_nest_start(skb, SMC_NLA_DEV_PORT + port);
501 	if (!port_attrs)
502 		goto errout;
503 
504 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_PNET_USR,
505 		       smcibdev->pnetid_by_user[port]))
506 		goto errattr;
507 	memcpy(smc_pnet, &smcibdev->pnetid[port], SMC_MAX_PNETID_LEN);
508 	smc_pnet[SMC_MAX_PNETID_LEN] = 0;
509 	if (nla_put_string(skb, SMC_NLA_DEV_PORT_PNETID, smc_pnet))
510 		goto errattr;
511 	if (nla_put_u32(skb, SMC_NLA_DEV_PORT_NETDEV,
512 			smcibdev->ndev_ifidx[port]))
513 		goto errattr;
514 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_VALID, 1))
515 		goto errattr;
516 	port_state = smc_ib_port_active(smcibdev, port + 1);
517 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_STATE, port_state))
518 		goto errattr;
519 	lnk_count = atomic_read(&smcibdev->lnk_cnt_by_port[port]);
520 	if (nla_put_u32(skb, SMC_NLA_DEV_PORT_LNK_CNT, lnk_count))
521 		goto errattr;
522 	nla_nest_end(skb, port_attrs);
523 	return 0;
524 errattr:
525 	nla_nest_cancel(skb, port_attrs);
526 errout:
527 	return -EMSGSIZE;
528 }
529 
530 static bool smc_nl_handle_pci_values(const struct smc_pci_dev *smc_pci_dev,
531 				     struct sk_buff *skb)
532 {
533 	if (nla_put_u32(skb, SMC_NLA_DEV_PCI_FID, smc_pci_dev->pci_fid))
534 		return false;
535 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_CHID, smc_pci_dev->pci_pchid))
536 		return false;
537 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_VENDOR, smc_pci_dev->pci_vendor))
538 		return false;
539 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_DEVICE, smc_pci_dev->pci_device))
540 		return false;
541 	if (nla_put_string(skb, SMC_NLA_DEV_PCI_ID, smc_pci_dev->pci_id))
542 		return false;
543 	return true;
544 }
545 
546 static int smc_nl_handle_smcr_dev(struct smc_ib_device *smcibdev,
547 				  struct sk_buff *skb,
548 				  struct netlink_callback *cb)
549 {
550 	char smc_ibname[IB_DEVICE_NAME_MAX];
551 	struct smc_pci_dev smc_pci_dev;
552 	struct pci_dev *pci_dev;
553 	unsigned char is_crit;
554 	struct nlattr *attrs;
555 	void *nlh;
556 	int i;
557 
558 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
559 			  &smc_gen_nl_family, NLM_F_MULTI,
560 			  SMC_NETLINK_GET_DEV_SMCR);
561 	if (!nlh)
562 		goto errmsg;
563 	attrs = nla_nest_start(skb, SMC_GEN_DEV_SMCR);
564 	if (!attrs)
565 		goto errout;
566 	is_crit = smcr_diag_is_dev_critical(&smc_lgr_list, smcibdev);
567 	if (nla_put_u8(skb, SMC_NLA_DEV_IS_CRIT, is_crit))
568 		goto errattr;
569 	if (smcibdev->ibdev->dev.parent) {
570 		memset(&smc_pci_dev, 0, sizeof(smc_pci_dev));
571 		pci_dev = to_pci_dev(smcibdev->ibdev->dev.parent);
572 		smc_set_pci_values(pci_dev, &smc_pci_dev);
573 		if (!smc_nl_handle_pci_values(&smc_pci_dev, skb))
574 			goto errattr;
575 	}
576 	snprintf(smc_ibname, sizeof(smc_ibname), "%s", smcibdev->ibdev->name);
577 	if (nla_put_string(skb, SMC_NLA_DEV_IB_NAME, smc_ibname))
578 		goto errattr;
579 	for (i = 1; i <= SMC_MAX_PORTS; i++) {
580 		if (!rdma_is_port_valid(smcibdev->ibdev, i))
581 			continue;
582 		if (smc_nl_handle_dev_port(skb, smcibdev->ibdev,
583 					   smcibdev, i - 1))
584 			goto errattr;
585 	}
586 
587 	nla_nest_end(skb, attrs);
588 	genlmsg_end(skb, nlh);
589 	return 0;
590 
591 errattr:
592 	nla_nest_cancel(skb, attrs);
593 errout:
594 	genlmsg_cancel(skb, nlh);
595 errmsg:
596 	return -EMSGSIZE;
597 }
598 
599 static void smc_nl_prep_smcr_dev(struct smc_ib_devices *dev_list,
600 				 struct sk_buff *skb,
601 				 struct netlink_callback *cb)
602 {
603 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
604 	struct smc_ib_device *smcibdev;
605 	int snum = cb_ctx->pos[0];
606 	int num = 0;
607 
608 	mutex_lock(&dev_list->mutex);
609 	list_for_each_entry(smcibdev, &dev_list->list, list) {
610 		if (num < snum)
611 			goto next;
612 		if (smc_nl_handle_smcr_dev(smcibdev, skb, cb))
613 			goto errout;
614 next:
615 		num++;
616 	}
617 errout:
618 	mutex_unlock(&dev_list->mutex);
619 	cb_ctx->pos[0] = num;
620 }
621 
622 int smcr_nl_get_device(struct sk_buff *skb, struct netlink_callback *cb)
623 {
624 	smc_nl_prep_smcr_dev(&smc_ib_devices, skb, cb);
625 	return skb->len;
626 }
627 
628 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
629 {
630 	struct smc_link *lnk = (struct smc_link *)priv;
631 	struct smc_ib_device *smcibdev = lnk->smcibdev;
632 	u8 port_idx;
633 
634 	switch (ibevent->event) {
635 	case IB_EVENT_QP_FATAL:
636 	case IB_EVENT_QP_ACCESS_ERR:
637 		port_idx = ibevent->element.qp->port - 1;
638 		if (port_idx >= SMC_MAX_PORTS)
639 			break;
640 		set_bit(port_idx, &smcibdev->port_event_mask);
641 		if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
642 			schedule_work(&smcibdev->port_event_work);
643 		break;
644 	default:
645 		break;
646 	}
647 }
648 
649 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
650 {
651 	if (lnk->roce_qp)
652 		ib_destroy_qp(lnk->roce_qp);
653 	lnk->roce_qp = NULL;
654 }
655 
656 /* create a queue pair within the protection domain for a link */
657 int smc_ib_create_queue_pair(struct smc_link *lnk)
658 {
659 	int sges_per_buf = (lnk->lgr->smc_version == SMC_V2) ? 2 : 1;
660 	struct ib_qp_init_attr qp_attr = {
661 		.event_handler = smc_ib_qp_event_handler,
662 		.qp_context = lnk,
663 		.send_cq = lnk->smcibdev->roce_cq_send,
664 		.recv_cq = lnk->smcibdev->roce_cq_recv,
665 		.srq = NULL,
666 		.cap = {
667 				/* include unsolicited rdma_writes as well,
668 				 * there are max. 2 RDMA_WRITE per 1 WR_SEND
669 				 */
670 			.max_send_wr = SMC_WR_BUF_CNT * 3,
671 			.max_recv_wr = SMC_WR_BUF_CNT * 3,
672 			.max_send_sge = SMC_IB_MAX_SEND_SGE,
673 			.max_recv_sge = sges_per_buf,
674 		},
675 		.sq_sig_type = IB_SIGNAL_REQ_WR,
676 		.qp_type = IB_QPT_RC,
677 	};
678 	int rc;
679 
680 	lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
681 	rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
682 	if (IS_ERR(lnk->roce_qp))
683 		lnk->roce_qp = NULL;
684 	else
685 		smc_wr_remember_qp_attr(lnk);
686 	return rc;
687 }
688 
689 void smc_ib_put_memory_region(struct ib_mr *mr)
690 {
691 	ib_dereg_mr(mr);
692 }
693 
694 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot, u8 link_idx)
695 {
696 	unsigned int offset = 0;
697 	int sg_num;
698 
699 	/* map the largest prefix of a dma mapped SG list */
700 	sg_num = ib_map_mr_sg(buf_slot->mr_rx[link_idx],
701 			      buf_slot->sgt[link_idx].sgl,
702 			      buf_slot->sgt[link_idx].orig_nents,
703 			      &offset, PAGE_SIZE);
704 
705 	return sg_num;
706 }
707 
708 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
709 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
710 			     struct smc_buf_desc *buf_slot, u8 link_idx)
711 {
712 	if (buf_slot->mr_rx[link_idx])
713 		return 0; /* already done */
714 
715 	buf_slot->mr_rx[link_idx] =
716 		ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
717 	if (IS_ERR(buf_slot->mr_rx[link_idx])) {
718 		int rc;
719 
720 		rc = PTR_ERR(buf_slot->mr_rx[link_idx]);
721 		buf_slot->mr_rx[link_idx] = NULL;
722 		return rc;
723 	}
724 
725 	if (smc_ib_map_mr_sg(buf_slot, link_idx) != 1)
726 		return -EINVAL;
727 
728 	return 0;
729 }
730 
731 /* synchronize buffer usage for cpu access */
732 void smc_ib_sync_sg_for_cpu(struct smc_link *lnk,
733 			    struct smc_buf_desc *buf_slot,
734 			    enum dma_data_direction data_direction)
735 {
736 	struct scatterlist *sg;
737 	unsigned int i;
738 
739 	/* for now there is just one DMA address */
740 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
741 		    buf_slot->sgt[lnk->link_idx].nents, i) {
742 		if (!sg_dma_len(sg))
743 			break;
744 		ib_dma_sync_single_for_cpu(lnk->smcibdev->ibdev,
745 					   sg_dma_address(sg),
746 					   sg_dma_len(sg),
747 					   data_direction);
748 	}
749 }
750 
751 /* synchronize buffer usage for device access */
752 void smc_ib_sync_sg_for_device(struct smc_link *lnk,
753 			       struct smc_buf_desc *buf_slot,
754 			       enum dma_data_direction data_direction)
755 {
756 	struct scatterlist *sg;
757 	unsigned int i;
758 
759 	/* for now there is just one DMA address */
760 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
761 		    buf_slot->sgt[lnk->link_idx].nents, i) {
762 		if (!sg_dma_len(sg))
763 			break;
764 		ib_dma_sync_single_for_device(lnk->smcibdev->ibdev,
765 					      sg_dma_address(sg),
766 					      sg_dma_len(sg),
767 					      data_direction);
768 	}
769 }
770 
771 /* Map a new TX or RX buffer SG-table to DMA */
772 int smc_ib_buf_map_sg(struct smc_link *lnk,
773 		      struct smc_buf_desc *buf_slot,
774 		      enum dma_data_direction data_direction)
775 {
776 	int mapped_nents;
777 
778 	mapped_nents = ib_dma_map_sg(lnk->smcibdev->ibdev,
779 				     buf_slot->sgt[lnk->link_idx].sgl,
780 				     buf_slot->sgt[lnk->link_idx].orig_nents,
781 				     data_direction);
782 	if (!mapped_nents)
783 		return -ENOMEM;
784 
785 	return mapped_nents;
786 }
787 
788 void smc_ib_buf_unmap_sg(struct smc_link *lnk,
789 			 struct smc_buf_desc *buf_slot,
790 			 enum dma_data_direction data_direction)
791 {
792 	if (!buf_slot->sgt[lnk->link_idx].sgl->dma_address)
793 		return; /* already unmapped */
794 
795 	ib_dma_unmap_sg(lnk->smcibdev->ibdev,
796 			buf_slot->sgt[lnk->link_idx].sgl,
797 			buf_slot->sgt[lnk->link_idx].orig_nents,
798 			data_direction);
799 	buf_slot->sgt[lnk->link_idx].sgl->dma_address = 0;
800 }
801 
802 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
803 {
804 	struct ib_cq_init_attr cqattr =	{
805 		.cqe = SMC_MAX_CQE, .comp_vector = 0 };
806 	int cqe_size_order, smc_order;
807 	long rc;
808 
809 	mutex_lock(&smcibdev->mutex);
810 	rc = 0;
811 	if (smcibdev->initialized)
812 		goto out;
813 	/* the calculated number of cq entries fits to mlx5 cq allocation */
814 	cqe_size_order = cache_line_size() == 128 ? 7 : 6;
815 	smc_order = MAX_ORDER - cqe_size_order - 1;
816 	if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
817 		cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
818 	smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
819 					      smc_wr_tx_cq_handler, NULL,
820 					      smcibdev, &cqattr);
821 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
822 	if (IS_ERR(smcibdev->roce_cq_send)) {
823 		smcibdev->roce_cq_send = NULL;
824 		goto out;
825 	}
826 	smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
827 					      smc_wr_rx_cq_handler, NULL,
828 					      smcibdev, &cqattr);
829 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
830 	if (IS_ERR(smcibdev->roce_cq_recv)) {
831 		smcibdev->roce_cq_recv = NULL;
832 		goto err;
833 	}
834 	smc_wr_add_dev(smcibdev);
835 	smcibdev->initialized = 1;
836 	goto out;
837 
838 err:
839 	ib_destroy_cq(smcibdev->roce_cq_send);
840 out:
841 	mutex_unlock(&smcibdev->mutex);
842 	return rc;
843 }
844 
845 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
846 {
847 	mutex_lock(&smcibdev->mutex);
848 	if (!smcibdev->initialized)
849 		goto out;
850 	smcibdev->initialized = 0;
851 	ib_destroy_cq(smcibdev->roce_cq_recv);
852 	ib_destroy_cq(smcibdev->roce_cq_send);
853 	smc_wr_remove_dev(smcibdev);
854 out:
855 	mutex_unlock(&smcibdev->mutex);
856 }
857 
858 static struct ib_client smc_ib_client;
859 
860 static void smc_copy_netdev_ifindex(struct smc_ib_device *smcibdev, int port)
861 {
862 	struct ib_device *ibdev = smcibdev->ibdev;
863 	struct net_device *ndev;
864 
865 	if (!ibdev->ops.get_netdev)
866 		return;
867 	ndev = ibdev->ops.get_netdev(ibdev, port + 1);
868 	if (ndev) {
869 		smcibdev->ndev_ifidx[port] = ndev->ifindex;
870 		dev_put(ndev);
871 	}
872 }
873 
874 void smc_ib_ndev_change(struct net_device *ndev, unsigned long event)
875 {
876 	struct smc_ib_device *smcibdev;
877 	struct ib_device *libdev;
878 	struct net_device *lndev;
879 	u8 port_cnt;
880 	int i;
881 
882 	mutex_lock(&smc_ib_devices.mutex);
883 	list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
884 		port_cnt = smcibdev->ibdev->phys_port_cnt;
885 		for (i = 0; i < min_t(size_t, port_cnt, SMC_MAX_PORTS); i++) {
886 			libdev = smcibdev->ibdev;
887 			if (!libdev->ops.get_netdev)
888 				continue;
889 			lndev = libdev->ops.get_netdev(libdev, i + 1);
890 			dev_put(lndev);
891 			if (lndev != ndev)
892 				continue;
893 			if (event == NETDEV_REGISTER)
894 				smcibdev->ndev_ifidx[i] = ndev->ifindex;
895 			if (event == NETDEV_UNREGISTER)
896 				smcibdev->ndev_ifidx[i] = 0;
897 		}
898 	}
899 	mutex_unlock(&smc_ib_devices.mutex);
900 }
901 
902 /* callback function for ib_register_client() */
903 static int smc_ib_add_dev(struct ib_device *ibdev)
904 {
905 	struct smc_ib_device *smcibdev;
906 	u8 port_cnt;
907 	int i;
908 
909 	if (ibdev->node_type != RDMA_NODE_IB_CA)
910 		return -EOPNOTSUPP;
911 
912 	smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
913 	if (!smcibdev)
914 		return -ENOMEM;
915 
916 	smcibdev->ibdev = ibdev;
917 	INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
918 	atomic_set(&smcibdev->lnk_cnt, 0);
919 	init_waitqueue_head(&smcibdev->lnks_deleted);
920 	mutex_init(&smcibdev->mutex);
921 	mutex_lock(&smc_ib_devices.mutex);
922 	list_add_tail(&smcibdev->list, &smc_ib_devices.list);
923 	mutex_unlock(&smc_ib_devices.mutex);
924 	ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
925 	INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
926 			      smc_ib_global_event_handler);
927 	ib_register_event_handler(&smcibdev->event_handler);
928 
929 	/* trigger reading of the port attributes */
930 	port_cnt = smcibdev->ibdev->phys_port_cnt;
931 	pr_warn_ratelimited("smc: adding ib device %s with port count %d\n",
932 			    smcibdev->ibdev->name, port_cnt);
933 	for (i = 0;
934 	     i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
935 	     i++) {
936 		set_bit(i, &smcibdev->port_event_mask);
937 		/* determine pnetids of the port */
938 		if (smc_pnetid_by_dev_port(ibdev->dev.parent, i,
939 					   smcibdev->pnetid[i]))
940 			smc_pnetid_by_table_ib(smcibdev, i + 1);
941 		smc_copy_netdev_ifindex(smcibdev, i);
942 		pr_warn_ratelimited("smc:    ib device %s port %d has pnetid "
943 				    "%.16s%s\n",
944 				    smcibdev->ibdev->name, i + 1,
945 				    smcibdev->pnetid[i],
946 				    smcibdev->pnetid_by_user[i] ?
947 				     " (user defined)" :
948 				     "");
949 	}
950 	schedule_work(&smcibdev->port_event_work);
951 	return 0;
952 }
953 
954 /* callback function for ib_unregister_client() */
955 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
956 {
957 	struct smc_ib_device *smcibdev = client_data;
958 
959 	mutex_lock(&smc_ib_devices.mutex);
960 	list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
961 	mutex_unlock(&smc_ib_devices.mutex);
962 	pr_warn_ratelimited("smc: removing ib device %s\n",
963 			    smcibdev->ibdev->name);
964 	smc_smcr_terminate_all(smcibdev);
965 	smc_ib_cleanup_per_ibdev(smcibdev);
966 	ib_unregister_event_handler(&smcibdev->event_handler);
967 	cancel_work_sync(&smcibdev->port_event_work);
968 	kfree(smcibdev);
969 }
970 
971 static struct ib_client smc_ib_client = {
972 	.name	= "smc_ib",
973 	.add	= smc_ib_add_dev,
974 	.remove = smc_ib_remove_dev,
975 };
976 
977 int __init smc_ib_register_client(void)
978 {
979 	smc_ib_init_local_systemid();
980 	return ib_register_client(&smc_ib_client);
981 }
982 
983 void smc_ib_unregister_client(void)
984 {
985 	ib_unregister_client(&smc_ib_client);
986 }
987