xref: /openbmc/linux/net/smc/smc_ib.c (revision 93696d8f)
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(struct net *net, __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(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_rt;
213 	neigh = dst_neigh_lookup(&rt->dst, &fl4.daddr);
214 	if (!neigh)
215 		goto out_rt;
216 	memcpy(nexthop_mac, neigh->ha, ETH_ALEN);
217 	*uses_gateway = rt->rt_uses_gateway;
218 	neigh_release(neigh);
219 	ip_rt_put(rt);
220 	return 0;
221 
222 out_rt:
223 	ip_rt_put(rt);
224 out:
225 	return -ENOENT;
226 }
227 
228 static int smc_ib_determine_gid_rcu(const struct net_device *ndev,
229 				    const struct ib_gid_attr *attr,
230 				    u8 gid[], u8 *sgid_index,
231 				    struct smc_init_info_smcrv2 *smcrv2)
232 {
233 	if (!smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE) {
234 		if (gid)
235 			memcpy(gid, &attr->gid, SMC_GID_SIZE);
236 		if (sgid_index)
237 			*sgid_index = attr->index;
238 		return 0;
239 	}
240 	if (smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP &&
241 	    smc_ib_gid_to_ipv4((u8 *)&attr->gid) != cpu_to_be32(INADDR_NONE)) {
242 		struct in_device *in_dev = __in_dev_get_rcu(ndev);
243 		struct net *net = dev_net(ndev);
244 		const struct in_ifaddr *ifa;
245 		bool subnet_match = false;
246 
247 		if (!in_dev)
248 			goto out;
249 		in_dev_for_each_ifa_rcu(ifa, in_dev) {
250 			if (!inet_ifa_match(smcrv2->saddr, ifa))
251 				continue;
252 			subnet_match = true;
253 			break;
254 		}
255 		if (!subnet_match)
256 			goto out;
257 		if (smcrv2->daddr && smc_ib_find_route(net, smcrv2->saddr,
258 						       smcrv2->daddr,
259 						       smcrv2->nexthop_mac,
260 						       &smcrv2->uses_gateway))
261 			goto out;
262 
263 		if (gid)
264 			memcpy(gid, &attr->gid, SMC_GID_SIZE);
265 		if (sgid_index)
266 			*sgid_index = attr->index;
267 		return 0;
268 	}
269 out:
270 	return -ENODEV;
271 }
272 
273 /* determine the gid for an ib-device port and vlan id */
274 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
275 			 unsigned short vlan_id, u8 gid[], u8 *sgid_index,
276 			 struct smc_init_info_smcrv2 *smcrv2)
277 {
278 	const struct ib_gid_attr *attr;
279 	const struct net_device *ndev;
280 	int i;
281 
282 	for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
283 		attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
284 		if (IS_ERR(attr))
285 			continue;
286 
287 		rcu_read_lock();
288 		ndev = rdma_read_gid_attr_ndev_rcu(attr);
289 		if (!IS_ERR(ndev) &&
290 		    ((!vlan_id && !is_vlan_dev(ndev)) ||
291 		     (vlan_id && is_vlan_dev(ndev) &&
292 		      vlan_dev_vlan_id(ndev) == vlan_id))) {
293 			if (!smc_ib_determine_gid_rcu(ndev, attr, gid,
294 						      sgid_index, smcrv2)) {
295 				rcu_read_unlock();
296 				rdma_put_gid_attr(attr);
297 				return 0;
298 			}
299 		}
300 		rcu_read_unlock();
301 		rdma_put_gid_attr(attr);
302 	}
303 	return -ENODEV;
304 }
305 
306 /* check if gid is still defined on smcibdev */
307 static bool smc_ib_check_link_gid(u8 gid[SMC_GID_SIZE], bool smcrv2,
308 				  struct smc_ib_device *smcibdev, u8 ibport)
309 {
310 	const struct ib_gid_attr *attr;
311 	bool rc = false;
312 	int i;
313 
314 	for (i = 0; !rc && i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
315 		attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
316 		if (IS_ERR(attr))
317 			continue;
318 
319 		rcu_read_lock();
320 		if ((!smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE) ||
321 		    (smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP &&
322 		     !(ipv6_addr_type((const struct in6_addr *)&attr->gid)
323 				     & IPV6_ADDR_LINKLOCAL)))
324 			if (!memcmp(gid, &attr->gid, SMC_GID_SIZE))
325 				rc = true;
326 		rcu_read_unlock();
327 		rdma_put_gid_attr(attr);
328 	}
329 	return rc;
330 }
331 
332 /* check all links if the gid is still defined on smcibdev */
333 static void smc_ib_gid_check(struct smc_ib_device *smcibdev, u8 ibport)
334 {
335 	struct smc_link_group *lgr;
336 	int i;
337 
338 	spin_lock_bh(&smc_lgr_list.lock);
339 	list_for_each_entry(lgr, &smc_lgr_list.list, list) {
340 		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
341 			    SMC_MAX_PNETID_LEN))
342 			continue; /* lgr is not affected */
343 		if (list_empty(&lgr->list))
344 			continue;
345 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
346 			if (lgr->lnk[i].state == SMC_LNK_UNUSED ||
347 			    lgr->lnk[i].smcibdev != smcibdev)
348 				continue;
349 			if (!smc_ib_check_link_gid(lgr->lnk[i].gid,
350 						   lgr->smc_version == SMC_V2,
351 						   smcibdev, ibport))
352 				smcr_port_err(smcibdev, ibport);
353 		}
354 	}
355 	spin_unlock_bh(&smc_lgr_list.lock);
356 }
357 
358 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
359 {
360 	int rc;
361 
362 	memset(&smcibdev->pattr[ibport - 1], 0,
363 	       sizeof(smcibdev->pattr[ibport - 1]));
364 	rc = ib_query_port(smcibdev->ibdev, ibport,
365 			   &smcibdev->pattr[ibport - 1]);
366 	if (rc)
367 		goto out;
368 	/* the SMC protocol requires specification of the RoCE MAC address */
369 	rc = smc_ib_fill_mac(smcibdev, ibport);
370 	if (rc)
371 		goto out;
372 	if (!smc_ib_is_valid_local_systemid() &&
373 	    smc_ib_port_active(smcibdev, ibport))
374 		/* create unique system identifier */
375 		smc_ib_define_local_systemid(smcibdev, ibport);
376 out:
377 	return rc;
378 }
379 
380 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
381 static void smc_ib_port_event_work(struct work_struct *work)
382 {
383 	struct smc_ib_device *smcibdev = container_of(
384 		work, struct smc_ib_device, port_event_work);
385 	u8 port_idx;
386 
387 	for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
388 		smc_ib_remember_port_attr(smcibdev, port_idx + 1);
389 		clear_bit(port_idx, &smcibdev->port_event_mask);
390 		if (!smc_ib_port_active(smcibdev, port_idx + 1)) {
391 			set_bit(port_idx, smcibdev->ports_going_away);
392 			smcr_port_err(smcibdev, port_idx + 1);
393 		} else {
394 			clear_bit(port_idx, smcibdev->ports_going_away);
395 			smcr_port_add(smcibdev, port_idx + 1);
396 			smc_ib_gid_check(smcibdev, port_idx + 1);
397 		}
398 	}
399 }
400 
401 /* can be called in IRQ context */
402 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
403 					struct ib_event *ibevent)
404 {
405 	struct smc_ib_device *smcibdev;
406 	bool schedule = false;
407 	u8 port_idx;
408 
409 	smcibdev = container_of(handler, struct smc_ib_device, event_handler);
410 
411 	switch (ibevent->event) {
412 	case IB_EVENT_DEVICE_FATAL:
413 		/* terminate all ports on device */
414 		for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) {
415 			set_bit(port_idx, &smcibdev->port_event_mask);
416 			if (!test_and_set_bit(port_idx,
417 					      smcibdev->ports_going_away))
418 				schedule = true;
419 		}
420 		if (schedule)
421 			schedule_work(&smcibdev->port_event_work);
422 		break;
423 	case IB_EVENT_PORT_ACTIVE:
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_clear_bit(port_idx, smcibdev->ports_going_away))
429 			schedule_work(&smcibdev->port_event_work);
430 		break;
431 	case IB_EVENT_PORT_ERR:
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 		if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
437 			schedule_work(&smcibdev->port_event_work);
438 		break;
439 	case IB_EVENT_GID_CHANGE:
440 		port_idx = ibevent->element.port_num - 1;
441 		if (port_idx >= SMC_MAX_PORTS)
442 			break;
443 		set_bit(port_idx, &smcibdev->port_event_mask);
444 		schedule_work(&smcibdev->port_event_work);
445 		break;
446 	default:
447 		break;
448 	}
449 }
450 
451 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
452 {
453 	if (lnk->roce_pd)
454 		ib_dealloc_pd(lnk->roce_pd);
455 	lnk->roce_pd = NULL;
456 }
457 
458 int smc_ib_create_protection_domain(struct smc_link *lnk)
459 {
460 	int rc;
461 
462 	lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
463 	rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
464 	if (IS_ERR(lnk->roce_pd))
465 		lnk->roce_pd = NULL;
466 	return rc;
467 }
468 
469 static bool smcr_diag_is_dev_critical(struct smc_lgr_list *smc_lgr,
470 				      struct smc_ib_device *smcibdev)
471 {
472 	struct smc_link_group *lgr;
473 	bool rc = false;
474 	int i;
475 
476 	spin_lock_bh(&smc_lgr->lock);
477 	list_for_each_entry(lgr, &smc_lgr->list, list) {
478 		if (lgr->is_smcd)
479 			continue;
480 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
481 			if (lgr->lnk[i].state == SMC_LNK_UNUSED ||
482 			    lgr->lnk[i].smcibdev != smcibdev)
483 				continue;
484 			if (lgr->type == SMC_LGR_SINGLE ||
485 			    lgr->type == SMC_LGR_ASYMMETRIC_LOCAL) {
486 				rc = true;
487 				goto out;
488 			}
489 		}
490 	}
491 out:
492 	spin_unlock_bh(&smc_lgr->lock);
493 	return rc;
494 }
495 
496 static int smc_nl_handle_dev_port(struct sk_buff *skb,
497 				  struct ib_device *ibdev,
498 				  struct smc_ib_device *smcibdev,
499 				  int port)
500 {
501 	char smc_pnet[SMC_MAX_PNETID_LEN + 1];
502 	struct nlattr *port_attrs;
503 	unsigned char port_state;
504 	int lnk_count = 0;
505 
506 	port_attrs = nla_nest_start(skb, SMC_NLA_DEV_PORT + port);
507 	if (!port_attrs)
508 		goto errout;
509 
510 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_PNET_USR,
511 		       smcibdev->pnetid_by_user[port]))
512 		goto errattr;
513 	memcpy(smc_pnet, &smcibdev->pnetid[port], SMC_MAX_PNETID_LEN);
514 	smc_pnet[SMC_MAX_PNETID_LEN] = 0;
515 	if (nla_put_string(skb, SMC_NLA_DEV_PORT_PNETID, smc_pnet))
516 		goto errattr;
517 	if (nla_put_u32(skb, SMC_NLA_DEV_PORT_NETDEV,
518 			smcibdev->ndev_ifidx[port]))
519 		goto errattr;
520 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_VALID, 1))
521 		goto errattr;
522 	port_state = smc_ib_port_active(smcibdev, port + 1);
523 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_STATE, port_state))
524 		goto errattr;
525 	lnk_count = atomic_read(&smcibdev->lnk_cnt_by_port[port]);
526 	if (nla_put_u32(skb, SMC_NLA_DEV_PORT_LNK_CNT, lnk_count))
527 		goto errattr;
528 	nla_nest_end(skb, port_attrs);
529 	return 0;
530 errattr:
531 	nla_nest_cancel(skb, port_attrs);
532 errout:
533 	return -EMSGSIZE;
534 }
535 
536 static bool smc_nl_handle_pci_values(const struct smc_pci_dev *smc_pci_dev,
537 				     struct sk_buff *skb)
538 {
539 	if (nla_put_u32(skb, SMC_NLA_DEV_PCI_FID, smc_pci_dev->pci_fid))
540 		return false;
541 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_CHID, smc_pci_dev->pci_pchid))
542 		return false;
543 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_VENDOR, smc_pci_dev->pci_vendor))
544 		return false;
545 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_DEVICE, smc_pci_dev->pci_device))
546 		return false;
547 	if (nla_put_string(skb, SMC_NLA_DEV_PCI_ID, smc_pci_dev->pci_id))
548 		return false;
549 	return true;
550 }
551 
552 static int smc_nl_handle_smcr_dev(struct smc_ib_device *smcibdev,
553 				  struct sk_buff *skb,
554 				  struct netlink_callback *cb)
555 {
556 	char smc_ibname[IB_DEVICE_NAME_MAX];
557 	struct smc_pci_dev smc_pci_dev;
558 	struct pci_dev *pci_dev;
559 	unsigned char is_crit;
560 	struct nlattr *attrs;
561 	void *nlh;
562 	int i;
563 
564 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
565 			  &smc_gen_nl_family, NLM_F_MULTI,
566 			  SMC_NETLINK_GET_DEV_SMCR);
567 	if (!nlh)
568 		goto errmsg;
569 	attrs = nla_nest_start(skb, SMC_GEN_DEV_SMCR);
570 	if (!attrs)
571 		goto errout;
572 	is_crit = smcr_diag_is_dev_critical(&smc_lgr_list, smcibdev);
573 	if (nla_put_u8(skb, SMC_NLA_DEV_IS_CRIT, is_crit))
574 		goto errattr;
575 	if (smcibdev->ibdev->dev.parent) {
576 		memset(&smc_pci_dev, 0, sizeof(smc_pci_dev));
577 		pci_dev = to_pci_dev(smcibdev->ibdev->dev.parent);
578 		smc_set_pci_values(pci_dev, &smc_pci_dev);
579 		if (!smc_nl_handle_pci_values(&smc_pci_dev, skb))
580 			goto errattr;
581 	}
582 	snprintf(smc_ibname, sizeof(smc_ibname), "%s", smcibdev->ibdev->name);
583 	if (nla_put_string(skb, SMC_NLA_DEV_IB_NAME, smc_ibname))
584 		goto errattr;
585 	for (i = 1; i <= SMC_MAX_PORTS; i++) {
586 		if (!rdma_is_port_valid(smcibdev->ibdev, i))
587 			continue;
588 		if (smc_nl_handle_dev_port(skb, smcibdev->ibdev,
589 					   smcibdev, i - 1))
590 			goto errattr;
591 	}
592 
593 	nla_nest_end(skb, attrs);
594 	genlmsg_end(skb, nlh);
595 	return 0;
596 
597 errattr:
598 	nla_nest_cancel(skb, attrs);
599 errout:
600 	genlmsg_cancel(skb, nlh);
601 errmsg:
602 	return -EMSGSIZE;
603 }
604 
605 static void smc_nl_prep_smcr_dev(struct smc_ib_devices *dev_list,
606 				 struct sk_buff *skb,
607 				 struct netlink_callback *cb)
608 {
609 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
610 	struct smc_ib_device *smcibdev;
611 	int snum = cb_ctx->pos[0];
612 	int num = 0;
613 
614 	mutex_lock(&dev_list->mutex);
615 	list_for_each_entry(smcibdev, &dev_list->list, list) {
616 		if (num < snum)
617 			goto next;
618 		if (smc_nl_handle_smcr_dev(smcibdev, skb, cb))
619 			goto errout;
620 next:
621 		num++;
622 	}
623 errout:
624 	mutex_unlock(&dev_list->mutex);
625 	cb_ctx->pos[0] = num;
626 }
627 
628 int smcr_nl_get_device(struct sk_buff *skb, struct netlink_callback *cb)
629 {
630 	smc_nl_prep_smcr_dev(&smc_ib_devices, skb, cb);
631 	return skb->len;
632 }
633 
634 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
635 {
636 	struct smc_link *lnk = (struct smc_link *)priv;
637 	struct smc_ib_device *smcibdev = lnk->smcibdev;
638 	u8 port_idx;
639 
640 	switch (ibevent->event) {
641 	case IB_EVENT_QP_FATAL:
642 	case IB_EVENT_QP_ACCESS_ERR:
643 		port_idx = ibevent->element.qp->port - 1;
644 		if (port_idx >= SMC_MAX_PORTS)
645 			break;
646 		set_bit(port_idx, &smcibdev->port_event_mask);
647 		if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
648 			schedule_work(&smcibdev->port_event_work);
649 		break;
650 	default:
651 		break;
652 	}
653 }
654 
655 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
656 {
657 	if (lnk->roce_qp)
658 		ib_destroy_qp(lnk->roce_qp);
659 	lnk->roce_qp = NULL;
660 }
661 
662 /* create a queue pair within the protection domain for a link */
663 int smc_ib_create_queue_pair(struct smc_link *lnk)
664 {
665 	int sges_per_buf = (lnk->lgr->smc_version == SMC_V2) ? 2 : 1;
666 	struct ib_qp_init_attr qp_attr = {
667 		.event_handler = smc_ib_qp_event_handler,
668 		.qp_context = lnk,
669 		.send_cq = lnk->smcibdev->roce_cq_send,
670 		.recv_cq = lnk->smcibdev->roce_cq_recv,
671 		.srq = NULL,
672 		.cap = {
673 				/* include unsolicited rdma_writes as well,
674 				 * there are max. 2 RDMA_WRITE per 1 WR_SEND
675 				 */
676 			.max_send_wr = SMC_WR_BUF_CNT * 3,
677 			.max_recv_wr = SMC_WR_BUF_CNT * 3,
678 			.max_send_sge = SMC_IB_MAX_SEND_SGE,
679 			.max_recv_sge = sges_per_buf,
680 			.max_inline_data = 0,
681 		},
682 		.sq_sig_type = IB_SIGNAL_REQ_WR,
683 		.qp_type = IB_QPT_RC,
684 	};
685 	int rc;
686 
687 	lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
688 	rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
689 	if (IS_ERR(lnk->roce_qp))
690 		lnk->roce_qp = NULL;
691 	else
692 		smc_wr_remember_qp_attr(lnk);
693 	return rc;
694 }
695 
696 void smc_ib_put_memory_region(struct ib_mr *mr)
697 {
698 	ib_dereg_mr(mr);
699 }
700 
701 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot, u8 link_idx)
702 {
703 	unsigned int offset = 0;
704 	int sg_num;
705 
706 	/* map the largest prefix of a dma mapped SG list */
707 	sg_num = ib_map_mr_sg(buf_slot->mr[link_idx],
708 			      buf_slot->sgt[link_idx].sgl,
709 			      buf_slot->sgt[link_idx].orig_nents,
710 			      &offset, PAGE_SIZE);
711 
712 	return sg_num;
713 }
714 
715 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
716 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
717 			     struct smc_buf_desc *buf_slot, u8 link_idx)
718 {
719 	if (buf_slot->mr[link_idx])
720 		return 0; /* already done */
721 
722 	buf_slot->mr[link_idx] =
723 		ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
724 	if (IS_ERR(buf_slot->mr[link_idx])) {
725 		int rc;
726 
727 		rc = PTR_ERR(buf_slot->mr[link_idx]);
728 		buf_slot->mr[link_idx] = NULL;
729 		return rc;
730 	}
731 
732 	if (smc_ib_map_mr_sg(buf_slot, link_idx) !=
733 			     buf_slot->sgt[link_idx].orig_nents)
734 		return -EINVAL;
735 
736 	return 0;
737 }
738 
739 bool smc_ib_is_sg_need_sync(struct smc_link *lnk,
740 			    struct smc_buf_desc *buf_slot)
741 {
742 	struct scatterlist *sg;
743 	unsigned int i;
744 	bool ret = false;
745 
746 	/* for now there is just one DMA address */
747 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
748 		    buf_slot->sgt[lnk->link_idx].nents, i) {
749 		if (!sg_dma_len(sg))
750 			break;
751 		if (dma_need_sync(lnk->smcibdev->ibdev->dma_device,
752 				  sg_dma_address(sg))) {
753 			ret = true;
754 			goto out;
755 		}
756 	}
757 
758 out:
759 	return ret;
760 }
761 
762 /* synchronize buffer usage for cpu access */
763 void smc_ib_sync_sg_for_cpu(struct smc_link *lnk,
764 			    struct smc_buf_desc *buf_slot,
765 			    enum dma_data_direction data_direction)
766 {
767 	struct scatterlist *sg;
768 	unsigned int i;
769 
770 	if (!(buf_slot->is_dma_need_sync & (1U << lnk->link_idx)))
771 		return;
772 
773 	/* for now there is just one DMA address */
774 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
775 		    buf_slot->sgt[lnk->link_idx].nents, i) {
776 		if (!sg_dma_len(sg))
777 			break;
778 		ib_dma_sync_single_for_cpu(lnk->smcibdev->ibdev,
779 					   sg_dma_address(sg),
780 					   sg_dma_len(sg),
781 					   data_direction);
782 	}
783 }
784 
785 /* synchronize buffer usage for device access */
786 void smc_ib_sync_sg_for_device(struct smc_link *lnk,
787 			       struct smc_buf_desc *buf_slot,
788 			       enum dma_data_direction data_direction)
789 {
790 	struct scatterlist *sg;
791 	unsigned int i;
792 
793 	if (!(buf_slot->is_dma_need_sync & (1U << lnk->link_idx)))
794 		return;
795 
796 	/* for now there is just one DMA address */
797 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
798 		    buf_slot->sgt[lnk->link_idx].nents, i) {
799 		if (!sg_dma_len(sg))
800 			break;
801 		ib_dma_sync_single_for_device(lnk->smcibdev->ibdev,
802 					      sg_dma_address(sg),
803 					      sg_dma_len(sg),
804 					      data_direction);
805 	}
806 }
807 
808 /* Map a new TX or RX buffer SG-table to DMA */
809 int smc_ib_buf_map_sg(struct smc_link *lnk,
810 		      struct smc_buf_desc *buf_slot,
811 		      enum dma_data_direction data_direction)
812 {
813 	int mapped_nents;
814 
815 	mapped_nents = ib_dma_map_sg(lnk->smcibdev->ibdev,
816 				     buf_slot->sgt[lnk->link_idx].sgl,
817 				     buf_slot->sgt[lnk->link_idx].orig_nents,
818 				     data_direction);
819 	if (!mapped_nents)
820 		return -ENOMEM;
821 
822 	return mapped_nents;
823 }
824 
825 void smc_ib_buf_unmap_sg(struct smc_link *lnk,
826 			 struct smc_buf_desc *buf_slot,
827 			 enum dma_data_direction data_direction)
828 {
829 	if (!buf_slot->sgt[lnk->link_idx].sgl->dma_address)
830 		return; /* already unmapped */
831 
832 	ib_dma_unmap_sg(lnk->smcibdev->ibdev,
833 			buf_slot->sgt[lnk->link_idx].sgl,
834 			buf_slot->sgt[lnk->link_idx].orig_nents,
835 			data_direction);
836 	buf_slot->sgt[lnk->link_idx].sgl->dma_address = 0;
837 }
838 
839 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
840 {
841 	struct ib_cq_init_attr cqattr =	{
842 		.cqe = SMC_MAX_CQE, .comp_vector = 0 };
843 	int cqe_size_order, smc_order;
844 	long rc;
845 
846 	mutex_lock(&smcibdev->mutex);
847 	rc = 0;
848 	if (smcibdev->initialized)
849 		goto out;
850 	/* the calculated number of cq entries fits to mlx5 cq allocation */
851 	cqe_size_order = cache_line_size() == 128 ? 7 : 6;
852 	smc_order = MAX_ORDER - cqe_size_order;
853 	if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
854 		cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
855 	smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
856 					      smc_wr_tx_cq_handler, NULL,
857 					      smcibdev, &cqattr);
858 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
859 	if (IS_ERR(smcibdev->roce_cq_send)) {
860 		smcibdev->roce_cq_send = NULL;
861 		goto out;
862 	}
863 	smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
864 					      smc_wr_rx_cq_handler, NULL,
865 					      smcibdev, &cqattr);
866 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
867 	if (IS_ERR(smcibdev->roce_cq_recv)) {
868 		smcibdev->roce_cq_recv = NULL;
869 		goto err;
870 	}
871 	smc_wr_add_dev(smcibdev);
872 	smcibdev->initialized = 1;
873 	goto out;
874 
875 err:
876 	ib_destroy_cq(smcibdev->roce_cq_send);
877 out:
878 	mutex_unlock(&smcibdev->mutex);
879 	return rc;
880 }
881 
882 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
883 {
884 	mutex_lock(&smcibdev->mutex);
885 	if (!smcibdev->initialized)
886 		goto out;
887 	smcibdev->initialized = 0;
888 	ib_destroy_cq(smcibdev->roce_cq_recv);
889 	ib_destroy_cq(smcibdev->roce_cq_send);
890 	smc_wr_remove_dev(smcibdev);
891 out:
892 	mutex_unlock(&smcibdev->mutex);
893 }
894 
895 static struct ib_client smc_ib_client;
896 
897 static void smc_copy_netdev_ifindex(struct smc_ib_device *smcibdev, int port)
898 {
899 	struct ib_device *ibdev = smcibdev->ibdev;
900 	struct net_device *ndev;
901 
902 	if (!ibdev->ops.get_netdev)
903 		return;
904 	ndev = ibdev->ops.get_netdev(ibdev, port + 1);
905 	if (ndev) {
906 		smcibdev->ndev_ifidx[port] = ndev->ifindex;
907 		dev_put(ndev);
908 	}
909 }
910 
911 void smc_ib_ndev_change(struct net_device *ndev, unsigned long event)
912 {
913 	struct smc_ib_device *smcibdev;
914 	struct ib_device *libdev;
915 	struct net_device *lndev;
916 	u8 port_cnt;
917 	int i;
918 
919 	mutex_lock(&smc_ib_devices.mutex);
920 	list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
921 		port_cnt = smcibdev->ibdev->phys_port_cnt;
922 		for (i = 0; i < min_t(size_t, port_cnt, SMC_MAX_PORTS); i++) {
923 			libdev = smcibdev->ibdev;
924 			if (!libdev->ops.get_netdev)
925 				continue;
926 			lndev = libdev->ops.get_netdev(libdev, i + 1);
927 			dev_put(lndev);
928 			if (lndev != ndev)
929 				continue;
930 			if (event == NETDEV_REGISTER)
931 				smcibdev->ndev_ifidx[i] = ndev->ifindex;
932 			if (event == NETDEV_UNREGISTER)
933 				smcibdev->ndev_ifidx[i] = 0;
934 		}
935 	}
936 	mutex_unlock(&smc_ib_devices.mutex);
937 }
938 
939 /* callback function for ib_register_client() */
940 static int smc_ib_add_dev(struct ib_device *ibdev)
941 {
942 	struct smc_ib_device *smcibdev;
943 	u8 port_cnt;
944 	int i;
945 
946 	if (ibdev->node_type != RDMA_NODE_IB_CA)
947 		return -EOPNOTSUPP;
948 
949 	smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
950 	if (!smcibdev)
951 		return -ENOMEM;
952 
953 	smcibdev->ibdev = ibdev;
954 	INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
955 	atomic_set(&smcibdev->lnk_cnt, 0);
956 	init_waitqueue_head(&smcibdev->lnks_deleted);
957 	mutex_init(&smcibdev->mutex);
958 	mutex_lock(&smc_ib_devices.mutex);
959 	list_add_tail(&smcibdev->list, &smc_ib_devices.list);
960 	mutex_unlock(&smc_ib_devices.mutex);
961 	ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
962 	INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
963 			      smc_ib_global_event_handler);
964 	ib_register_event_handler(&smcibdev->event_handler);
965 
966 	/* trigger reading of the port attributes */
967 	port_cnt = smcibdev->ibdev->phys_port_cnt;
968 	pr_warn_ratelimited("smc: adding ib device %s with port count %d\n",
969 			    smcibdev->ibdev->name, port_cnt);
970 	for (i = 0;
971 	     i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
972 	     i++) {
973 		set_bit(i, &smcibdev->port_event_mask);
974 		/* determine pnetids of the port */
975 		if (smc_pnetid_by_dev_port(ibdev->dev.parent, i,
976 					   smcibdev->pnetid[i]))
977 			smc_pnetid_by_table_ib(smcibdev, i + 1);
978 		smc_copy_netdev_ifindex(smcibdev, i);
979 		pr_warn_ratelimited("smc:    ib device %s port %d has pnetid "
980 				    "%.16s%s\n",
981 				    smcibdev->ibdev->name, i + 1,
982 				    smcibdev->pnetid[i],
983 				    smcibdev->pnetid_by_user[i] ?
984 				     " (user defined)" :
985 				     "");
986 	}
987 	schedule_work(&smcibdev->port_event_work);
988 	return 0;
989 }
990 
991 /* callback function for ib_unregister_client() */
992 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
993 {
994 	struct smc_ib_device *smcibdev = client_data;
995 
996 	mutex_lock(&smc_ib_devices.mutex);
997 	list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
998 	mutex_unlock(&smc_ib_devices.mutex);
999 	pr_warn_ratelimited("smc: removing ib device %s\n",
1000 			    smcibdev->ibdev->name);
1001 	smc_smcr_terminate_all(smcibdev);
1002 	smc_ib_cleanup_per_ibdev(smcibdev);
1003 	ib_unregister_event_handler(&smcibdev->event_handler);
1004 	cancel_work_sync(&smcibdev->port_event_work);
1005 	kfree(smcibdev);
1006 }
1007 
1008 static struct ib_client smc_ib_client = {
1009 	.name	= "smc_ib",
1010 	.add	= smc_ib_add_dev,
1011 	.remove = smc_ib_remove_dev,
1012 };
1013 
1014 int __init smc_ib_register_client(void)
1015 {
1016 	smc_ib_init_local_systemid();
1017 	return ib_register_client(&smc_ib_client);
1018 }
1019 
1020 void smc_ib_unregister_client(void)
1021 {
1022 	ib_unregister_client(&smc_ib_client);
1023 }
1024