xref: /openbmc/linux/net/smc/smc_ib.c (revision 50dc9a85)
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/random.h>
16 #include <linux/workqueue.h>
17 #include <linux/scatterlist.h>
18 #include <linux/wait.h>
19 #include <linux/mutex.h>
20 #include <linux/inetdevice.h>
21 #include <rdma/ib_verbs.h>
22 #include <rdma/ib_cache.h>
23 
24 #include "smc_pnet.h"
25 #include "smc_ib.h"
26 #include "smc_core.h"
27 #include "smc_wr.h"
28 #include "smc.h"
29 #include "smc_netlink.h"
30 
31 #define SMC_MAX_CQE 32766	/* max. # of completion queue elements */
32 
33 #define SMC_QP_MIN_RNR_TIMER		5
34 #define SMC_QP_TIMEOUT			15 /* 4096 * 2 ** timeout usec */
35 #define SMC_QP_RETRY_CNT			7 /* 7: infinite */
36 #define SMC_QP_RNR_RETRY			7 /* 7: infinite */
37 
38 struct smc_ib_devices smc_ib_devices = {	/* smc-registered ib devices */
39 	.mutex = __MUTEX_INITIALIZER(smc_ib_devices.mutex),
40 	.list = LIST_HEAD_INIT(smc_ib_devices.list),
41 };
42 
43 u8 local_systemid[SMC_SYSTEMID_LEN];		/* unique system identifier */
44 
45 static int smc_ib_modify_qp_init(struct smc_link *lnk)
46 {
47 	struct ib_qp_attr qp_attr;
48 
49 	memset(&qp_attr, 0, sizeof(qp_attr));
50 	qp_attr.qp_state = IB_QPS_INIT;
51 	qp_attr.pkey_index = 0;
52 	qp_attr.port_num = lnk->ibport;
53 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
54 				| IB_ACCESS_REMOTE_WRITE;
55 	return ib_modify_qp(lnk->roce_qp, &qp_attr,
56 			    IB_QP_STATE | IB_QP_PKEY_INDEX |
57 			    IB_QP_ACCESS_FLAGS | IB_QP_PORT);
58 }
59 
60 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
61 {
62 	enum ib_qp_attr_mask qp_attr_mask =
63 		IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
64 		IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
65 	struct ib_qp_attr qp_attr;
66 	u8 hop_lim = 1;
67 
68 	memset(&qp_attr, 0, sizeof(qp_attr));
69 	qp_attr.qp_state = IB_QPS_RTR;
70 	qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
71 	qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
72 	rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
73 	if (lnk->lgr->smc_version == SMC_V2 && lnk->lgr->uses_gateway)
74 		hop_lim = IPV6_DEFAULT_HOPLIMIT;
75 	rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, lnk->sgid_index, hop_lim, 0);
76 	rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
77 	if (lnk->lgr->smc_version == SMC_V2 && lnk->lgr->uses_gateway)
78 		memcpy(&qp_attr.ah_attr.roce.dmac, lnk->lgr->nexthop_mac,
79 		       sizeof(lnk->lgr->nexthop_mac));
80 	else
81 		memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
82 		       sizeof(lnk->peer_mac));
83 	qp_attr.dest_qp_num = lnk->peer_qpn;
84 	qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
85 	qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
86 					 * requests
87 					 */
88 	qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
89 
90 	return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
91 }
92 
93 int smc_ib_modify_qp_rts(struct smc_link *lnk)
94 {
95 	struct ib_qp_attr qp_attr;
96 
97 	memset(&qp_attr, 0, sizeof(qp_attr));
98 	qp_attr.qp_state = IB_QPS_RTS;
99 	qp_attr.timeout = SMC_QP_TIMEOUT;	/* local ack timeout */
100 	qp_attr.retry_cnt = SMC_QP_RETRY_CNT;	/* retry count */
101 	qp_attr.rnr_retry = SMC_QP_RNR_RETRY;	/* RNR retries, 7=infinite */
102 	qp_attr.sq_psn = lnk->psn_initial;	/* starting send packet seq # */
103 	qp_attr.max_rd_atomic = 1;	/* # of outstanding RDMA reads and
104 					 * atomic ops allowed
105 					 */
106 	return ib_modify_qp(lnk->roce_qp, &qp_attr,
107 			    IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
108 			    IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
109 			    IB_QP_MAX_QP_RD_ATOMIC);
110 }
111 
112 int smc_ib_modify_qp_reset(struct smc_link *lnk)
113 {
114 	struct ib_qp_attr qp_attr;
115 
116 	memset(&qp_attr, 0, sizeof(qp_attr));
117 	qp_attr.qp_state = IB_QPS_RESET;
118 	return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
119 }
120 
121 int smc_ib_ready_link(struct smc_link *lnk)
122 {
123 	struct smc_link_group *lgr = smc_get_lgr(lnk);
124 	int rc = 0;
125 
126 	rc = smc_ib_modify_qp_init(lnk);
127 	if (rc)
128 		goto out;
129 
130 	rc = smc_ib_modify_qp_rtr(lnk);
131 	if (rc)
132 		goto out;
133 	smc_wr_remember_qp_attr(lnk);
134 	rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
135 			      IB_CQ_SOLICITED_MASK);
136 	if (rc)
137 		goto out;
138 	rc = smc_wr_rx_post_init(lnk);
139 	if (rc)
140 		goto out;
141 	smc_wr_remember_qp_attr(lnk);
142 
143 	if (lgr->role == SMC_SERV) {
144 		rc = smc_ib_modify_qp_rts(lnk);
145 		if (rc)
146 			goto out;
147 		smc_wr_remember_qp_attr(lnk);
148 	}
149 out:
150 	return rc;
151 }
152 
153 static int smc_ib_fill_mac(struct smc_ib_device *smcibdev, u8 ibport)
154 {
155 	const struct ib_gid_attr *attr;
156 	int rc;
157 
158 	attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, 0);
159 	if (IS_ERR(attr))
160 		return -ENODEV;
161 
162 	rc = rdma_read_gid_l2_fields(attr, NULL, smcibdev->mac[ibport - 1]);
163 	rdma_put_gid_attr(attr);
164 	return rc;
165 }
166 
167 /* Create an identifier unique for this instance of SMC-R.
168  * The MAC-address of the first active registered IB device
169  * plus a random 2-byte number is used to create this identifier.
170  * This name is delivered to the peer during connection initialization.
171  */
172 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
173 						u8 ibport)
174 {
175 	memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
176 	       sizeof(smcibdev->mac[ibport - 1]));
177 }
178 
179 bool smc_ib_is_valid_local_systemid(void)
180 {
181 	return !is_zero_ether_addr(&local_systemid[2]);
182 }
183 
184 static void smc_ib_init_local_systemid(void)
185 {
186 	get_random_bytes(&local_systemid[0], 2);
187 }
188 
189 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
190 {
191 	return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
192 }
193 
194 int smc_ib_find_route(__be32 saddr, __be32 daddr,
195 		      u8 nexthop_mac[], u8 *uses_gateway)
196 {
197 	struct neighbour *neigh = NULL;
198 	struct rtable *rt = NULL;
199 	struct flowi4 fl4 = {
200 		.saddr = saddr,
201 		.daddr = daddr
202 	};
203 
204 	if (daddr == cpu_to_be32(INADDR_NONE))
205 		goto out;
206 	rt = ip_route_output_flow(&init_net, &fl4, NULL);
207 	if (IS_ERR(rt))
208 		goto out;
209 	if (rt->rt_uses_gateway && rt->rt_gw_family != AF_INET)
210 		goto out;
211 	neigh = rt->dst.ops->neigh_lookup(&rt->dst, NULL, &fl4.daddr);
212 	if (neigh) {
213 		memcpy(nexthop_mac, neigh->ha, ETH_ALEN);
214 		*uses_gateway = rt->rt_uses_gateway;
215 		return 0;
216 	}
217 out:
218 	return -ENOENT;
219 }
220 
221 static int smc_ib_determine_gid_rcu(const struct net_device *ndev,
222 				    const struct ib_gid_attr *attr,
223 				    u8 gid[], u8 *sgid_index,
224 				    struct smc_init_info_smcrv2 *smcrv2)
225 {
226 	if (!smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE) {
227 		if (gid)
228 			memcpy(gid, &attr->gid, SMC_GID_SIZE);
229 		if (sgid_index)
230 			*sgid_index = attr->index;
231 		return 0;
232 	}
233 	if (smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP &&
234 	    smc_ib_gid_to_ipv4((u8 *)&attr->gid) != cpu_to_be32(INADDR_NONE)) {
235 		struct in_device *in_dev = __in_dev_get_rcu(ndev);
236 		const struct in_ifaddr *ifa;
237 		bool subnet_match = false;
238 
239 		if (!in_dev)
240 			goto out;
241 		in_dev_for_each_ifa_rcu(ifa, in_dev) {
242 			if (!inet_ifa_match(smcrv2->saddr, ifa))
243 				continue;
244 			subnet_match = true;
245 			break;
246 		}
247 		if (!subnet_match)
248 			goto out;
249 		if (smcrv2->daddr && smc_ib_find_route(smcrv2->saddr,
250 						       smcrv2->daddr,
251 						       smcrv2->nexthop_mac,
252 						       &smcrv2->uses_gateway))
253 			goto out;
254 
255 		if (gid)
256 			memcpy(gid, &attr->gid, SMC_GID_SIZE);
257 		if (sgid_index)
258 			*sgid_index = attr->index;
259 		return 0;
260 	}
261 out:
262 	return -ENODEV;
263 }
264 
265 /* determine the gid for an ib-device port and vlan id */
266 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
267 			 unsigned short vlan_id, u8 gid[], u8 *sgid_index,
268 			 struct smc_init_info_smcrv2 *smcrv2)
269 {
270 	const struct ib_gid_attr *attr;
271 	const struct net_device *ndev;
272 	int i;
273 
274 	for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
275 		attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
276 		if (IS_ERR(attr))
277 			continue;
278 
279 		rcu_read_lock();
280 		ndev = rdma_read_gid_attr_ndev_rcu(attr);
281 		if (!IS_ERR(ndev) &&
282 		    ((!vlan_id && !is_vlan_dev(ndev)) ||
283 		     (vlan_id && is_vlan_dev(ndev) &&
284 		      vlan_dev_vlan_id(ndev) == vlan_id))) {
285 			if (!smc_ib_determine_gid_rcu(ndev, attr, gid,
286 						      sgid_index, smcrv2)) {
287 				rcu_read_unlock();
288 				rdma_put_gid_attr(attr);
289 				return 0;
290 			}
291 		}
292 		rcu_read_unlock();
293 		rdma_put_gid_attr(attr);
294 	}
295 	return -ENODEV;
296 }
297 
298 /* check if gid is still defined on smcibdev */
299 static bool smc_ib_check_link_gid(u8 gid[SMC_GID_SIZE], bool smcrv2,
300 				  struct smc_ib_device *smcibdev, u8 ibport)
301 {
302 	const struct ib_gid_attr *attr;
303 	bool rc = false;
304 	int i;
305 
306 	for (i = 0; !rc && i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
307 		attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
308 		if (IS_ERR(attr))
309 			continue;
310 
311 		rcu_read_lock();
312 		if ((!smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE) ||
313 		    (smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP &&
314 		     !(ipv6_addr_type((const struct in6_addr *)&attr->gid)
315 				     & IPV6_ADDR_LINKLOCAL)))
316 			if (!memcmp(gid, &attr->gid, SMC_GID_SIZE))
317 				rc = true;
318 		rcu_read_unlock();
319 		rdma_put_gid_attr(attr);
320 	}
321 	return rc;
322 }
323 
324 /* check all links if the gid is still defined on smcibdev */
325 static void smc_ib_gid_check(struct smc_ib_device *smcibdev, u8 ibport)
326 {
327 	struct smc_link_group *lgr;
328 	int i;
329 
330 	spin_lock_bh(&smc_lgr_list.lock);
331 	list_for_each_entry(lgr, &smc_lgr_list.list, list) {
332 		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
333 			    SMC_MAX_PNETID_LEN))
334 			continue; /* lgr is not affected */
335 		if (list_empty(&lgr->list))
336 			continue;
337 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
338 			if (lgr->lnk[i].state == SMC_LNK_UNUSED ||
339 			    lgr->lnk[i].smcibdev != smcibdev)
340 				continue;
341 			if (!smc_ib_check_link_gid(lgr->lnk[i].gid,
342 						   lgr->smc_version == SMC_V2,
343 						   smcibdev, ibport))
344 				smcr_port_err(smcibdev, ibport);
345 		}
346 	}
347 	spin_unlock_bh(&smc_lgr_list.lock);
348 }
349 
350 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
351 {
352 	int rc;
353 
354 	memset(&smcibdev->pattr[ibport - 1], 0,
355 	       sizeof(smcibdev->pattr[ibport - 1]));
356 	rc = ib_query_port(smcibdev->ibdev, ibport,
357 			   &smcibdev->pattr[ibport - 1]);
358 	if (rc)
359 		goto out;
360 	/* the SMC protocol requires specification of the RoCE MAC address */
361 	rc = smc_ib_fill_mac(smcibdev, ibport);
362 	if (rc)
363 		goto out;
364 	if (!smc_ib_is_valid_local_systemid() &&
365 	    smc_ib_port_active(smcibdev, ibport))
366 		/* create unique system identifier */
367 		smc_ib_define_local_systemid(smcibdev, ibport);
368 out:
369 	return rc;
370 }
371 
372 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
373 static void smc_ib_port_event_work(struct work_struct *work)
374 {
375 	struct smc_ib_device *smcibdev = container_of(
376 		work, struct smc_ib_device, port_event_work);
377 	u8 port_idx;
378 
379 	for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
380 		smc_ib_remember_port_attr(smcibdev, port_idx + 1);
381 		clear_bit(port_idx, &smcibdev->port_event_mask);
382 		if (!smc_ib_port_active(smcibdev, port_idx + 1)) {
383 			set_bit(port_idx, smcibdev->ports_going_away);
384 			smcr_port_err(smcibdev, port_idx + 1);
385 		} else {
386 			clear_bit(port_idx, smcibdev->ports_going_away);
387 			smcr_port_add(smcibdev, port_idx + 1);
388 			smc_ib_gid_check(smcibdev, port_idx + 1);
389 		}
390 	}
391 }
392 
393 /* can be called in IRQ context */
394 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
395 					struct ib_event *ibevent)
396 {
397 	struct smc_ib_device *smcibdev;
398 	bool schedule = false;
399 	u8 port_idx;
400 
401 	smcibdev = container_of(handler, struct smc_ib_device, event_handler);
402 
403 	switch (ibevent->event) {
404 	case IB_EVENT_DEVICE_FATAL:
405 		/* terminate all ports on device */
406 		for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) {
407 			set_bit(port_idx, &smcibdev->port_event_mask);
408 			if (!test_and_set_bit(port_idx,
409 					      smcibdev->ports_going_away))
410 				schedule = true;
411 		}
412 		if (schedule)
413 			schedule_work(&smcibdev->port_event_work);
414 		break;
415 	case IB_EVENT_PORT_ACTIVE:
416 		port_idx = ibevent->element.port_num - 1;
417 		if (port_idx >= SMC_MAX_PORTS)
418 			break;
419 		set_bit(port_idx, &smcibdev->port_event_mask);
420 		if (test_and_clear_bit(port_idx, smcibdev->ports_going_away))
421 			schedule_work(&smcibdev->port_event_work);
422 		break;
423 	case IB_EVENT_PORT_ERR:
424 		port_idx = ibevent->element.port_num - 1;
425 		if (port_idx >= SMC_MAX_PORTS)
426 			break;
427 		set_bit(port_idx, &smcibdev->port_event_mask);
428 		if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
429 			schedule_work(&smcibdev->port_event_work);
430 		break;
431 	case IB_EVENT_GID_CHANGE:
432 		port_idx = ibevent->element.port_num - 1;
433 		if (port_idx >= SMC_MAX_PORTS)
434 			break;
435 		set_bit(port_idx, &smcibdev->port_event_mask);
436 		schedule_work(&smcibdev->port_event_work);
437 		break;
438 	default:
439 		break;
440 	}
441 }
442 
443 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
444 {
445 	if (lnk->roce_pd)
446 		ib_dealloc_pd(lnk->roce_pd);
447 	lnk->roce_pd = NULL;
448 }
449 
450 int smc_ib_create_protection_domain(struct smc_link *lnk)
451 {
452 	int rc;
453 
454 	lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
455 	rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
456 	if (IS_ERR(lnk->roce_pd))
457 		lnk->roce_pd = NULL;
458 	return rc;
459 }
460 
461 static bool smcr_diag_is_dev_critical(struct smc_lgr_list *smc_lgr,
462 				      struct smc_ib_device *smcibdev)
463 {
464 	struct smc_link_group *lgr;
465 	bool rc = false;
466 	int i;
467 
468 	spin_lock_bh(&smc_lgr->lock);
469 	list_for_each_entry(lgr, &smc_lgr->list, list) {
470 		if (lgr->is_smcd)
471 			continue;
472 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
473 			if (lgr->lnk[i].state == SMC_LNK_UNUSED ||
474 			    lgr->lnk[i].smcibdev != smcibdev)
475 				continue;
476 			if (lgr->type == SMC_LGR_SINGLE ||
477 			    lgr->type == SMC_LGR_ASYMMETRIC_LOCAL) {
478 				rc = true;
479 				goto out;
480 			}
481 		}
482 	}
483 out:
484 	spin_unlock_bh(&smc_lgr->lock);
485 	return rc;
486 }
487 
488 static int smc_nl_handle_dev_port(struct sk_buff *skb,
489 				  struct ib_device *ibdev,
490 				  struct smc_ib_device *smcibdev,
491 				  int port)
492 {
493 	char smc_pnet[SMC_MAX_PNETID_LEN + 1];
494 	struct nlattr *port_attrs;
495 	unsigned char port_state;
496 	int lnk_count = 0;
497 
498 	port_attrs = nla_nest_start(skb, SMC_NLA_DEV_PORT + port);
499 	if (!port_attrs)
500 		goto errout;
501 
502 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_PNET_USR,
503 		       smcibdev->pnetid_by_user[port]))
504 		goto errattr;
505 	memcpy(smc_pnet, &smcibdev->pnetid[port], SMC_MAX_PNETID_LEN);
506 	smc_pnet[SMC_MAX_PNETID_LEN] = 0;
507 	if (nla_put_string(skb, SMC_NLA_DEV_PORT_PNETID, smc_pnet))
508 		goto errattr;
509 	if (nla_put_u32(skb, SMC_NLA_DEV_PORT_NETDEV,
510 			smcibdev->ndev_ifidx[port]))
511 		goto errattr;
512 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_VALID, 1))
513 		goto errattr;
514 	port_state = smc_ib_port_active(smcibdev, port + 1);
515 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_STATE, port_state))
516 		goto errattr;
517 	lnk_count = atomic_read(&smcibdev->lnk_cnt_by_port[port]);
518 	if (nla_put_u32(skb, SMC_NLA_DEV_PORT_LNK_CNT, lnk_count))
519 		goto errattr;
520 	nla_nest_end(skb, port_attrs);
521 	return 0;
522 errattr:
523 	nla_nest_cancel(skb, port_attrs);
524 errout:
525 	return -EMSGSIZE;
526 }
527 
528 static bool smc_nl_handle_pci_values(const struct smc_pci_dev *smc_pci_dev,
529 				     struct sk_buff *skb)
530 {
531 	if (nla_put_u32(skb, SMC_NLA_DEV_PCI_FID, smc_pci_dev->pci_fid))
532 		return false;
533 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_CHID, smc_pci_dev->pci_pchid))
534 		return false;
535 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_VENDOR, smc_pci_dev->pci_vendor))
536 		return false;
537 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_DEVICE, smc_pci_dev->pci_device))
538 		return false;
539 	if (nla_put_string(skb, SMC_NLA_DEV_PCI_ID, smc_pci_dev->pci_id))
540 		return false;
541 	return true;
542 }
543 
544 static int smc_nl_handle_smcr_dev(struct smc_ib_device *smcibdev,
545 				  struct sk_buff *skb,
546 				  struct netlink_callback *cb)
547 {
548 	char smc_ibname[IB_DEVICE_NAME_MAX];
549 	struct smc_pci_dev smc_pci_dev;
550 	struct pci_dev *pci_dev;
551 	unsigned char is_crit;
552 	struct nlattr *attrs;
553 	void *nlh;
554 	int i;
555 
556 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
557 			  &smc_gen_nl_family, NLM_F_MULTI,
558 			  SMC_NETLINK_GET_DEV_SMCR);
559 	if (!nlh)
560 		goto errmsg;
561 	attrs = nla_nest_start(skb, SMC_GEN_DEV_SMCR);
562 	if (!attrs)
563 		goto errout;
564 	is_crit = smcr_diag_is_dev_critical(&smc_lgr_list, smcibdev);
565 	if (nla_put_u8(skb, SMC_NLA_DEV_IS_CRIT, is_crit))
566 		goto errattr;
567 	if (smcibdev->ibdev->dev.parent) {
568 		memset(&smc_pci_dev, 0, sizeof(smc_pci_dev));
569 		pci_dev = to_pci_dev(smcibdev->ibdev->dev.parent);
570 		smc_set_pci_values(pci_dev, &smc_pci_dev);
571 		if (!smc_nl_handle_pci_values(&smc_pci_dev, skb))
572 			goto errattr;
573 	}
574 	snprintf(smc_ibname, sizeof(smc_ibname), "%s", smcibdev->ibdev->name);
575 	if (nla_put_string(skb, SMC_NLA_DEV_IB_NAME, smc_ibname))
576 		goto errattr;
577 	for (i = 1; i <= SMC_MAX_PORTS; i++) {
578 		if (!rdma_is_port_valid(smcibdev->ibdev, i))
579 			continue;
580 		if (smc_nl_handle_dev_port(skb, smcibdev->ibdev,
581 					   smcibdev, i - 1))
582 			goto errattr;
583 	}
584 
585 	nla_nest_end(skb, attrs);
586 	genlmsg_end(skb, nlh);
587 	return 0;
588 
589 errattr:
590 	nla_nest_cancel(skb, attrs);
591 errout:
592 	genlmsg_cancel(skb, nlh);
593 errmsg:
594 	return -EMSGSIZE;
595 }
596 
597 static void smc_nl_prep_smcr_dev(struct smc_ib_devices *dev_list,
598 				 struct sk_buff *skb,
599 				 struct netlink_callback *cb)
600 {
601 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
602 	struct smc_ib_device *smcibdev;
603 	int snum = cb_ctx->pos[0];
604 	int num = 0;
605 
606 	mutex_lock(&dev_list->mutex);
607 	list_for_each_entry(smcibdev, &dev_list->list, list) {
608 		if (num < snum)
609 			goto next;
610 		if (smc_nl_handle_smcr_dev(smcibdev, skb, cb))
611 			goto errout;
612 next:
613 		num++;
614 	}
615 errout:
616 	mutex_unlock(&dev_list->mutex);
617 	cb_ctx->pos[0] = num;
618 }
619 
620 int smcr_nl_get_device(struct sk_buff *skb, struct netlink_callback *cb)
621 {
622 	smc_nl_prep_smcr_dev(&smc_ib_devices, skb, cb);
623 	return skb->len;
624 }
625 
626 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
627 {
628 	struct smc_link *lnk = (struct smc_link *)priv;
629 	struct smc_ib_device *smcibdev = lnk->smcibdev;
630 	u8 port_idx;
631 
632 	switch (ibevent->event) {
633 	case IB_EVENT_QP_FATAL:
634 	case IB_EVENT_QP_ACCESS_ERR:
635 		port_idx = ibevent->element.qp->port - 1;
636 		if (port_idx >= SMC_MAX_PORTS)
637 			break;
638 		set_bit(port_idx, &smcibdev->port_event_mask);
639 		if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
640 			schedule_work(&smcibdev->port_event_work);
641 		break;
642 	default:
643 		break;
644 	}
645 }
646 
647 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
648 {
649 	if (lnk->roce_qp)
650 		ib_destroy_qp(lnk->roce_qp);
651 	lnk->roce_qp = NULL;
652 }
653 
654 /* create a queue pair within the protection domain for a link */
655 int smc_ib_create_queue_pair(struct smc_link *lnk)
656 {
657 	int sges_per_buf = (lnk->lgr->smc_version == SMC_V2) ? 2 : 1;
658 	struct ib_qp_init_attr qp_attr = {
659 		.event_handler = smc_ib_qp_event_handler,
660 		.qp_context = lnk,
661 		.send_cq = lnk->smcibdev->roce_cq_send,
662 		.recv_cq = lnk->smcibdev->roce_cq_recv,
663 		.srq = NULL,
664 		.cap = {
665 				/* include unsolicited rdma_writes as well,
666 				 * there are max. 2 RDMA_WRITE per 1 WR_SEND
667 				 */
668 			.max_send_wr = SMC_WR_BUF_CNT * 3,
669 			.max_recv_wr = SMC_WR_BUF_CNT * 3,
670 			.max_send_sge = SMC_IB_MAX_SEND_SGE,
671 			.max_recv_sge = sges_per_buf,
672 		},
673 		.sq_sig_type = IB_SIGNAL_REQ_WR,
674 		.qp_type = IB_QPT_RC,
675 	};
676 	int rc;
677 
678 	lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
679 	rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
680 	if (IS_ERR(lnk->roce_qp))
681 		lnk->roce_qp = NULL;
682 	else
683 		smc_wr_remember_qp_attr(lnk);
684 	return rc;
685 }
686 
687 void smc_ib_put_memory_region(struct ib_mr *mr)
688 {
689 	ib_dereg_mr(mr);
690 }
691 
692 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot, u8 link_idx)
693 {
694 	unsigned int offset = 0;
695 	int sg_num;
696 
697 	/* map the largest prefix of a dma mapped SG list */
698 	sg_num = ib_map_mr_sg(buf_slot->mr_rx[link_idx],
699 			      buf_slot->sgt[link_idx].sgl,
700 			      buf_slot->sgt[link_idx].orig_nents,
701 			      &offset, PAGE_SIZE);
702 
703 	return sg_num;
704 }
705 
706 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
707 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
708 			     struct smc_buf_desc *buf_slot, u8 link_idx)
709 {
710 	if (buf_slot->mr_rx[link_idx])
711 		return 0; /* already done */
712 
713 	buf_slot->mr_rx[link_idx] =
714 		ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
715 	if (IS_ERR(buf_slot->mr_rx[link_idx])) {
716 		int rc;
717 
718 		rc = PTR_ERR(buf_slot->mr_rx[link_idx]);
719 		buf_slot->mr_rx[link_idx] = NULL;
720 		return rc;
721 	}
722 
723 	if (smc_ib_map_mr_sg(buf_slot, link_idx) != 1)
724 		return -EINVAL;
725 
726 	return 0;
727 }
728 
729 /* synchronize buffer usage for cpu access */
730 void smc_ib_sync_sg_for_cpu(struct smc_link *lnk,
731 			    struct smc_buf_desc *buf_slot,
732 			    enum dma_data_direction data_direction)
733 {
734 	struct scatterlist *sg;
735 	unsigned int i;
736 
737 	/* for now there is just one DMA address */
738 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
739 		    buf_slot->sgt[lnk->link_idx].nents, i) {
740 		if (!sg_dma_len(sg))
741 			break;
742 		ib_dma_sync_single_for_cpu(lnk->smcibdev->ibdev,
743 					   sg_dma_address(sg),
744 					   sg_dma_len(sg),
745 					   data_direction);
746 	}
747 }
748 
749 /* synchronize buffer usage for device access */
750 void smc_ib_sync_sg_for_device(struct smc_link *lnk,
751 			       struct smc_buf_desc *buf_slot,
752 			       enum dma_data_direction data_direction)
753 {
754 	struct scatterlist *sg;
755 	unsigned int i;
756 
757 	/* for now there is just one DMA address */
758 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
759 		    buf_slot->sgt[lnk->link_idx].nents, i) {
760 		if (!sg_dma_len(sg))
761 			break;
762 		ib_dma_sync_single_for_device(lnk->smcibdev->ibdev,
763 					      sg_dma_address(sg),
764 					      sg_dma_len(sg),
765 					      data_direction);
766 	}
767 }
768 
769 /* Map a new TX or RX buffer SG-table to DMA */
770 int smc_ib_buf_map_sg(struct smc_link *lnk,
771 		      struct smc_buf_desc *buf_slot,
772 		      enum dma_data_direction data_direction)
773 {
774 	int mapped_nents;
775 
776 	mapped_nents = ib_dma_map_sg(lnk->smcibdev->ibdev,
777 				     buf_slot->sgt[lnk->link_idx].sgl,
778 				     buf_slot->sgt[lnk->link_idx].orig_nents,
779 				     data_direction);
780 	if (!mapped_nents)
781 		return -ENOMEM;
782 
783 	return mapped_nents;
784 }
785 
786 void smc_ib_buf_unmap_sg(struct smc_link *lnk,
787 			 struct smc_buf_desc *buf_slot,
788 			 enum dma_data_direction data_direction)
789 {
790 	if (!buf_slot->sgt[lnk->link_idx].sgl->dma_address)
791 		return; /* already unmapped */
792 
793 	ib_dma_unmap_sg(lnk->smcibdev->ibdev,
794 			buf_slot->sgt[lnk->link_idx].sgl,
795 			buf_slot->sgt[lnk->link_idx].orig_nents,
796 			data_direction);
797 	buf_slot->sgt[lnk->link_idx].sgl->dma_address = 0;
798 }
799 
800 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
801 {
802 	struct ib_cq_init_attr cqattr =	{
803 		.cqe = SMC_MAX_CQE, .comp_vector = 0 };
804 	int cqe_size_order, smc_order;
805 	long rc;
806 
807 	mutex_lock(&smcibdev->mutex);
808 	rc = 0;
809 	if (smcibdev->initialized)
810 		goto out;
811 	/* the calculated number of cq entries fits to mlx5 cq allocation */
812 	cqe_size_order = cache_line_size() == 128 ? 7 : 6;
813 	smc_order = MAX_ORDER - cqe_size_order - 1;
814 	if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
815 		cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
816 	smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
817 					      smc_wr_tx_cq_handler, NULL,
818 					      smcibdev, &cqattr);
819 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
820 	if (IS_ERR(smcibdev->roce_cq_send)) {
821 		smcibdev->roce_cq_send = NULL;
822 		goto out;
823 	}
824 	smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
825 					      smc_wr_rx_cq_handler, NULL,
826 					      smcibdev, &cqattr);
827 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
828 	if (IS_ERR(smcibdev->roce_cq_recv)) {
829 		smcibdev->roce_cq_recv = NULL;
830 		goto err;
831 	}
832 	smc_wr_add_dev(smcibdev);
833 	smcibdev->initialized = 1;
834 	goto out;
835 
836 err:
837 	ib_destroy_cq(smcibdev->roce_cq_send);
838 out:
839 	mutex_unlock(&smcibdev->mutex);
840 	return rc;
841 }
842 
843 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
844 {
845 	mutex_lock(&smcibdev->mutex);
846 	if (!smcibdev->initialized)
847 		goto out;
848 	smcibdev->initialized = 0;
849 	ib_destroy_cq(smcibdev->roce_cq_recv);
850 	ib_destroy_cq(smcibdev->roce_cq_send);
851 	smc_wr_remove_dev(smcibdev);
852 out:
853 	mutex_unlock(&smcibdev->mutex);
854 }
855 
856 static struct ib_client smc_ib_client;
857 
858 static void smc_copy_netdev_ifindex(struct smc_ib_device *smcibdev, int port)
859 {
860 	struct ib_device *ibdev = smcibdev->ibdev;
861 	struct net_device *ndev;
862 
863 	if (!ibdev->ops.get_netdev)
864 		return;
865 	ndev = ibdev->ops.get_netdev(ibdev, port + 1);
866 	if (ndev) {
867 		smcibdev->ndev_ifidx[port] = ndev->ifindex;
868 		dev_put(ndev);
869 	}
870 }
871 
872 void smc_ib_ndev_change(struct net_device *ndev, unsigned long event)
873 {
874 	struct smc_ib_device *smcibdev;
875 	struct ib_device *libdev;
876 	struct net_device *lndev;
877 	u8 port_cnt;
878 	int i;
879 
880 	mutex_lock(&smc_ib_devices.mutex);
881 	list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
882 		port_cnt = smcibdev->ibdev->phys_port_cnt;
883 		for (i = 0; i < min_t(size_t, port_cnt, SMC_MAX_PORTS); i++) {
884 			libdev = smcibdev->ibdev;
885 			if (!libdev->ops.get_netdev)
886 				continue;
887 			lndev = libdev->ops.get_netdev(libdev, i + 1);
888 			dev_put(lndev);
889 			if (lndev != ndev)
890 				continue;
891 			if (event == NETDEV_REGISTER)
892 				smcibdev->ndev_ifidx[i] = ndev->ifindex;
893 			if (event == NETDEV_UNREGISTER)
894 				smcibdev->ndev_ifidx[i] = 0;
895 		}
896 	}
897 	mutex_unlock(&smc_ib_devices.mutex);
898 }
899 
900 /* callback function for ib_register_client() */
901 static int smc_ib_add_dev(struct ib_device *ibdev)
902 {
903 	struct smc_ib_device *smcibdev;
904 	u8 port_cnt;
905 	int i;
906 
907 	if (ibdev->node_type != RDMA_NODE_IB_CA)
908 		return -EOPNOTSUPP;
909 
910 	smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
911 	if (!smcibdev)
912 		return -ENOMEM;
913 
914 	smcibdev->ibdev = ibdev;
915 	INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
916 	atomic_set(&smcibdev->lnk_cnt, 0);
917 	init_waitqueue_head(&smcibdev->lnks_deleted);
918 	mutex_init(&smcibdev->mutex);
919 	mutex_lock(&smc_ib_devices.mutex);
920 	list_add_tail(&smcibdev->list, &smc_ib_devices.list);
921 	mutex_unlock(&smc_ib_devices.mutex);
922 	ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
923 	INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
924 			      smc_ib_global_event_handler);
925 	ib_register_event_handler(&smcibdev->event_handler);
926 
927 	/* trigger reading of the port attributes */
928 	port_cnt = smcibdev->ibdev->phys_port_cnt;
929 	pr_warn_ratelimited("smc: adding ib device %s with port count %d\n",
930 			    smcibdev->ibdev->name, port_cnt);
931 	for (i = 0;
932 	     i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
933 	     i++) {
934 		set_bit(i, &smcibdev->port_event_mask);
935 		/* determine pnetids of the port */
936 		if (smc_pnetid_by_dev_port(ibdev->dev.parent, i,
937 					   smcibdev->pnetid[i]))
938 			smc_pnetid_by_table_ib(smcibdev, i + 1);
939 		smc_copy_netdev_ifindex(smcibdev, i);
940 		pr_warn_ratelimited("smc:    ib device %s port %d has pnetid "
941 				    "%.16s%s\n",
942 				    smcibdev->ibdev->name, i + 1,
943 				    smcibdev->pnetid[i],
944 				    smcibdev->pnetid_by_user[i] ?
945 				     " (user defined)" :
946 				     "");
947 	}
948 	schedule_work(&smcibdev->port_event_work);
949 	return 0;
950 }
951 
952 /* callback function for ib_unregister_client() */
953 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
954 {
955 	struct smc_ib_device *smcibdev = client_data;
956 
957 	mutex_lock(&smc_ib_devices.mutex);
958 	list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
959 	mutex_unlock(&smc_ib_devices.mutex);
960 	pr_warn_ratelimited("smc: removing ib device %s\n",
961 			    smcibdev->ibdev->name);
962 	smc_smcr_terminate_all(smcibdev);
963 	smc_ib_cleanup_per_ibdev(smcibdev);
964 	ib_unregister_event_handler(&smcibdev->event_handler);
965 	cancel_work_sync(&smcibdev->port_event_work);
966 	kfree(smcibdev);
967 }
968 
969 static struct ib_client smc_ib_client = {
970 	.name	= "smc_ib",
971 	.add	= smc_ib_add_dev,
972 	.remove = smc_ib_remove_dev,
973 };
974 
975 int __init smc_ib_register_client(void)
976 {
977 	smc_ib_init_local_systemid();
978 	return ib_register_client(&smc_ib_client);
979 }
980 
981 void smc_ib_unregister_client(void)
982 {
983 	ib_unregister_client(&smc_ib_client);
984 }
985