xref: /linux/net/smc/smc_ib.c (revision 306ec721d043bbe5e818d59fbb37c28d999b5d8b)
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;
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 		struct net *net = dev_net(ndev);
239 		const struct in_ifaddr *ifa;
240 		bool subnet_match = false;
241 
242 		if (!in_dev)
243 			goto out;
244 		in_dev_for_each_ifa_rcu(ifa, in_dev) {
245 			if (!inet_ifa_match(smcrv2->saddr, ifa))
246 				continue;
247 			subnet_match = true;
248 			break;
249 		}
250 		if (!subnet_match)
251 			goto out;
252 		if (smcrv2->daddr && smc_ib_find_route(net, smcrv2->saddr,
253 						       smcrv2->daddr,
254 						       smcrv2->nexthop_mac,
255 						       &smcrv2->uses_gateway))
256 			goto out;
257 
258 		if (gid)
259 			memcpy(gid, &attr->gid, SMC_GID_SIZE);
260 		if (sgid_index)
261 			*sgid_index = attr->index;
262 		return 0;
263 	}
264 out:
265 	return -ENODEV;
266 }
267 
268 /* determine the gid for an ib-device port and vlan id */
269 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
270 			 unsigned short vlan_id, u8 gid[], u8 *sgid_index,
271 			 struct smc_init_info_smcrv2 *smcrv2)
272 {
273 	const struct ib_gid_attr *attr;
274 	const struct net_device *ndev;
275 	int i;
276 
277 	for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
278 		attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
279 		if (IS_ERR(attr))
280 			continue;
281 
282 		rcu_read_lock();
283 		ndev = rdma_read_gid_attr_ndev_rcu(attr);
284 		if (!IS_ERR(ndev) &&
285 		    ((!vlan_id && !is_vlan_dev(ndev)) ||
286 		     (vlan_id && is_vlan_dev(ndev) &&
287 		      vlan_dev_vlan_id(ndev) == vlan_id))) {
288 			if (!smc_ib_determine_gid_rcu(ndev, attr, gid,
289 						      sgid_index, smcrv2)) {
290 				rcu_read_unlock();
291 				rdma_put_gid_attr(attr);
292 				return 0;
293 			}
294 		}
295 		rcu_read_unlock();
296 		rdma_put_gid_attr(attr);
297 	}
298 	return -ENODEV;
299 }
300 
301 /* check if gid is still defined on smcibdev */
302 static bool smc_ib_check_link_gid(u8 gid[SMC_GID_SIZE], bool smcrv2,
303 				  struct smc_ib_device *smcibdev, u8 ibport)
304 {
305 	const struct ib_gid_attr *attr;
306 	bool rc = false;
307 	int i;
308 
309 	for (i = 0; !rc && i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
310 		attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
311 		if (IS_ERR(attr))
312 			continue;
313 
314 		rcu_read_lock();
315 		if ((!smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE) ||
316 		    (smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP &&
317 		     !(ipv6_addr_type((const struct in6_addr *)&attr->gid)
318 				     & IPV6_ADDR_LINKLOCAL)))
319 			if (!memcmp(gid, &attr->gid, SMC_GID_SIZE))
320 				rc = true;
321 		rcu_read_unlock();
322 		rdma_put_gid_attr(attr);
323 	}
324 	return rc;
325 }
326 
327 /* check all links if the gid is still defined on smcibdev */
328 static void smc_ib_gid_check(struct smc_ib_device *smcibdev, u8 ibport)
329 {
330 	struct smc_link_group *lgr;
331 	int i;
332 
333 	spin_lock_bh(&smc_lgr_list.lock);
334 	list_for_each_entry(lgr, &smc_lgr_list.list, list) {
335 		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
336 			    SMC_MAX_PNETID_LEN))
337 			continue; /* lgr is not affected */
338 		if (list_empty(&lgr->list))
339 			continue;
340 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
341 			if (lgr->lnk[i].state == SMC_LNK_UNUSED ||
342 			    lgr->lnk[i].smcibdev != smcibdev)
343 				continue;
344 			if (!smc_ib_check_link_gid(lgr->lnk[i].gid,
345 						   lgr->smc_version == SMC_V2,
346 						   smcibdev, ibport))
347 				smcr_port_err(smcibdev, ibport);
348 		}
349 	}
350 	spin_unlock_bh(&smc_lgr_list.lock);
351 }
352 
353 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
354 {
355 	int rc;
356 
357 	memset(&smcibdev->pattr[ibport - 1], 0,
358 	       sizeof(smcibdev->pattr[ibport - 1]));
359 	rc = ib_query_port(smcibdev->ibdev, ibport,
360 			   &smcibdev->pattr[ibport - 1]);
361 	if (rc)
362 		goto out;
363 	/* the SMC protocol requires specification of the RoCE MAC address */
364 	rc = smc_ib_fill_mac(smcibdev, ibport);
365 	if (rc)
366 		goto out;
367 	if (!smc_ib_is_valid_local_systemid() &&
368 	    smc_ib_port_active(smcibdev, ibport))
369 		/* create unique system identifier */
370 		smc_ib_define_local_systemid(smcibdev, ibport);
371 out:
372 	return rc;
373 }
374 
375 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
376 static void smc_ib_port_event_work(struct work_struct *work)
377 {
378 	struct smc_ib_device *smcibdev = container_of(
379 		work, struct smc_ib_device, port_event_work);
380 	u8 port_idx;
381 
382 	for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
383 		smc_ib_remember_port_attr(smcibdev, port_idx + 1);
384 		clear_bit(port_idx, &smcibdev->port_event_mask);
385 		if (!smc_ib_port_active(smcibdev, port_idx + 1)) {
386 			set_bit(port_idx, smcibdev->ports_going_away);
387 			smcr_port_err(smcibdev, port_idx + 1);
388 		} else {
389 			clear_bit(port_idx, smcibdev->ports_going_away);
390 			smcr_port_add(smcibdev, port_idx + 1);
391 			smc_ib_gid_check(smcibdev, port_idx + 1);
392 		}
393 	}
394 }
395 
396 /* can be called in IRQ context */
397 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
398 					struct ib_event *ibevent)
399 {
400 	struct smc_ib_device *smcibdev;
401 	bool schedule = false;
402 	u8 port_idx;
403 
404 	smcibdev = container_of(handler, struct smc_ib_device, event_handler);
405 
406 	switch (ibevent->event) {
407 	case IB_EVENT_DEVICE_FATAL:
408 		/* terminate all ports on device */
409 		for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) {
410 			set_bit(port_idx, &smcibdev->port_event_mask);
411 			if (!test_and_set_bit(port_idx,
412 					      smcibdev->ports_going_away))
413 				schedule = true;
414 		}
415 		if (schedule)
416 			schedule_work(&smcibdev->port_event_work);
417 		break;
418 	case IB_EVENT_PORT_ACTIVE:
419 		port_idx = ibevent->element.port_num - 1;
420 		if (port_idx >= SMC_MAX_PORTS)
421 			break;
422 		set_bit(port_idx, &smcibdev->port_event_mask);
423 		if (test_and_clear_bit(port_idx, smcibdev->ports_going_away))
424 			schedule_work(&smcibdev->port_event_work);
425 		break;
426 	case IB_EVENT_PORT_ERR:
427 		port_idx = ibevent->element.port_num - 1;
428 		if (port_idx >= SMC_MAX_PORTS)
429 			break;
430 		set_bit(port_idx, &smcibdev->port_event_mask);
431 		if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
432 			schedule_work(&smcibdev->port_event_work);
433 		break;
434 	case IB_EVENT_GID_CHANGE:
435 		port_idx = ibevent->element.port_num - 1;
436 		if (port_idx >= SMC_MAX_PORTS)
437 			break;
438 		set_bit(port_idx, &smcibdev->port_event_mask);
439 		schedule_work(&smcibdev->port_event_work);
440 		break;
441 	default:
442 		break;
443 	}
444 }
445 
446 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
447 {
448 	if (lnk->roce_pd)
449 		ib_dealloc_pd(lnk->roce_pd);
450 	lnk->roce_pd = NULL;
451 }
452 
453 int smc_ib_create_protection_domain(struct smc_link *lnk)
454 {
455 	int rc;
456 
457 	lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
458 	rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
459 	if (IS_ERR(lnk->roce_pd))
460 		lnk->roce_pd = NULL;
461 	return rc;
462 }
463 
464 static bool smcr_diag_is_dev_critical(struct smc_lgr_list *smc_lgr,
465 				      struct smc_ib_device *smcibdev)
466 {
467 	struct smc_link_group *lgr;
468 	bool rc = false;
469 	int i;
470 
471 	spin_lock_bh(&smc_lgr->lock);
472 	list_for_each_entry(lgr, &smc_lgr->list, list) {
473 		if (lgr->is_smcd)
474 			continue;
475 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
476 			if (lgr->lnk[i].state == SMC_LNK_UNUSED ||
477 			    lgr->lnk[i].smcibdev != smcibdev)
478 				continue;
479 			if (lgr->type == SMC_LGR_SINGLE ||
480 			    lgr->type == SMC_LGR_ASYMMETRIC_LOCAL) {
481 				rc = true;
482 				goto out;
483 			}
484 		}
485 	}
486 out:
487 	spin_unlock_bh(&smc_lgr->lock);
488 	return rc;
489 }
490 
491 static int smc_nl_handle_dev_port(struct sk_buff *skb,
492 				  struct ib_device *ibdev,
493 				  struct smc_ib_device *smcibdev,
494 				  int port)
495 {
496 	char smc_pnet[SMC_MAX_PNETID_LEN + 1];
497 	struct nlattr *port_attrs;
498 	unsigned char port_state;
499 	int lnk_count = 0;
500 
501 	port_attrs = nla_nest_start(skb, SMC_NLA_DEV_PORT + port);
502 	if (!port_attrs)
503 		goto errout;
504 
505 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_PNET_USR,
506 		       smcibdev->pnetid_by_user[port]))
507 		goto errattr;
508 	memcpy(smc_pnet, &smcibdev->pnetid[port], SMC_MAX_PNETID_LEN);
509 	smc_pnet[SMC_MAX_PNETID_LEN] = 0;
510 	if (nla_put_string(skb, SMC_NLA_DEV_PORT_PNETID, smc_pnet))
511 		goto errattr;
512 	if (nla_put_u32(skb, SMC_NLA_DEV_PORT_NETDEV,
513 			smcibdev->ndev_ifidx[port]))
514 		goto errattr;
515 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_VALID, 1))
516 		goto errattr;
517 	port_state = smc_ib_port_active(smcibdev, port + 1);
518 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_STATE, port_state))
519 		goto errattr;
520 	lnk_count = atomic_read(&smcibdev->lnk_cnt_by_port[port]);
521 	if (nla_put_u32(skb, SMC_NLA_DEV_PORT_LNK_CNT, lnk_count))
522 		goto errattr;
523 	nla_nest_end(skb, port_attrs);
524 	return 0;
525 errattr:
526 	nla_nest_cancel(skb, port_attrs);
527 errout:
528 	return -EMSGSIZE;
529 }
530 
531 static bool smc_nl_handle_pci_values(const struct smc_pci_dev *smc_pci_dev,
532 				     struct sk_buff *skb)
533 {
534 	if (nla_put_u32(skb, SMC_NLA_DEV_PCI_FID, smc_pci_dev->pci_fid))
535 		return false;
536 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_CHID, smc_pci_dev->pci_pchid))
537 		return false;
538 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_VENDOR, smc_pci_dev->pci_vendor))
539 		return false;
540 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_DEVICE, smc_pci_dev->pci_device))
541 		return false;
542 	if (nla_put_string(skb, SMC_NLA_DEV_PCI_ID, smc_pci_dev->pci_id))
543 		return false;
544 	return true;
545 }
546 
547 static int smc_nl_handle_smcr_dev(struct smc_ib_device *smcibdev,
548 				  struct sk_buff *skb,
549 				  struct netlink_callback *cb)
550 {
551 	char smc_ibname[IB_DEVICE_NAME_MAX];
552 	struct smc_pci_dev smc_pci_dev;
553 	struct pci_dev *pci_dev;
554 	unsigned char is_crit;
555 	struct nlattr *attrs;
556 	void *nlh;
557 	int i;
558 
559 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
560 			  &smc_gen_nl_family, NLM_F_MULTI,
561 			  SMC_NETLINK_GET_DEV_SMCR);
562 	if (!nlh)
563 		goto errmsg;
564 	attrs = nla_nest_start(skb, SMC_GEN_DEV_SMCR);
565 	if (!attrs)
566 		goto errout;
567 	is_crit = smcr_diag_is_dev_critical(&smc_lgr_list, smcibdev);
568 	if (nla_put_u8(skb, SMC_NLA_DEV_IS_CRIT, is_crit))
569 		goto errattr;
570 	if (smcibdev->ibdev->dev.parent) {
571 		memset(&smc_pci_dev, 0, sizeof(smc_pci_dev));
572 		pci_dev = to_pci_dev(smcibdev->ibdev->dev.parent);
573 		smc_set_pci_values(pci_dev, &smc_pci_dev);
574 		if (!smc_nl_handle_pci_values(&smc_pci_dev, skb))
575 			goto errattr;
576 	}
577 	snprintf(smc_ibname, sizeof(smc_ibname), "%s", smcibdev->ibdev->name);
578 	if (nla_put_string(skb, SMC_NLA_DEV_IB_NAME, smc_ibname))
579 		goto errattr;
580 	for (i = 1; i <= SMC_MAX_PORTS; i++) {
581 		if (!rdma_is_port_valid(smcibdev->ibdev, i))
582 			continue;
583 		if (smc_nl_handle_dev_port(skb, smcibdev->ibdev,
584 					   smcibdev, i - 1))
585 			goto errattr;
586 	}
587 
588 	nla_nest_end(skb, attrs);
589 	genlmsg_end(skb, nlh);
590 	return 0;
591 
592 errattr:
593 	nla_nest_cancel(skb, attrs);
594 errout:
595 	genlmsg_cancel(skb, nlh);
596 errmsg:
597 	return -EMSGSIZE;
598 }
599 
600 static void smc_nl_prep_smcr_dev(struct smc_ib_devices *dev_list,
601 				 struct sk_buff *skb,
602 				 struct netlink_callback *cb)
603 {
604 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
605 	struct smc_ib_device *smcibdev;
606 	int snum = cb_ctx->pos[0];
607 	int num = 0;
608 
609 	mutex_lock(&dev_list->mutex);
610 	list_for_each_entry(smcibdev, &dev_list->list, list) {
611 		if (num < snum)
612 			goto next;
613 		if (smc_nl_handle_smcr_dev(smcibdev, skb, cb))
614 			goto errout;
615 next:
616 		num++;
617 	}
618 errout:
619 	mutex_unlock(&dev_list->mutex);
620 	cb_ctx->pos[0] = num;
621 }
622 
623 int smcr_nl_get_device(struct sk_buff *skb, struct netlink_callback *cb)
624 {
625 	smc_nl_prep_smcr_dev(&smc_ib_devices, skb, cb);
626 	return skb->len;
627 }
628 
629 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
630 {
631 	struct smc_link *lnk = (struct smc_link *)priv;
632 	struct smc_ib_device *smcibdev = lnk->smcibdev;
633 	u8 port_idx;
634 
635 	switch (ibevent->event) {
636 	case IB_EVENT_QP_FATAL:
637 	case IB_EVENT_QP_ACCESS_ERR:
638 		port_idx = ibevent->element.qp->port - 1;
639 		if (port_idx >= SMC_MAX_PORTS)
640 			break;
641 		set_bit(port_idx, &smcibdev->port_event_mask);
642 		if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
643 			schedule_work(&smcibdev->port_event_work);
644 		break;
645 	default:
646 		break;
647 	}
648 }
649 
650 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
651 {
652 	if (lnk->roce_qp)
653 		ib_destroy_qp(lnk->roce_qp);
654 	lnk->roce_qp = NULL;
655 }
656 
657 /* create a queue pair within the protection domain for a link */
658 int smc_ib_create_queue_pair(struct smc_link *lnk)
659 {
660 	int sges_per_buf = (lnk->lgr->smc_version == SMC_V2) ? 2 : 1;
661 	struct ib_qp_init_attr qp_attr = {
662 		.event_handler = smc_ib_qp_event_handler,
663 		.qp_context = lnk,
664 		.send_cq = lnk->smcibdev->roce_cq_send,
665 		.recv_cq = lnk->smcibdev->roce_cq_recv,
666 		.srq = NULL,
667 		.cap = {
668 				/* include unsolicited rdma_writes as well,
669 				 * there are max. 2 RDMA_WRITE per 1 WR_SEND
670 				 */
671 			.max_send_wr = SMC_WR_BUF_CNT * 3,
672 			.max_recv_wr = SMC_WR_BUF_CNT * 3,
673 			.max_send_sge = SMC_IB_MAX_SEND_SGE,
674 			.max_recv_sge = sges_per_buf,
675 			.max_inline_data = 0,
676 		},
677 		.sq_sig_type = IB_SIGNAL_REQ_WR,
678 		.qp_type = IB_QPT_RC,
679 	};
680 	int rc;
681 
682 	lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
683 	rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
684 	if (IS_ERR(lnk->roce_qp))
685 		lnk->roce_qp = NULL;
686 	else
687 		smc_wr_remember_qp_attr(lnk);
688 	return rc;
689 }
690 
691 void smc_ib_put_memory_region(struct ib_mr *mr)
692 {
693 	ib_dereg_mr(mr);
694 }
695 
696 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot, u8 link_idx)
697 {
698 	unsigned int offset = 0;
699 	int sg_num;
700 
701 	/* map the largest prefix of a dma mapped SG list */
702 	sg_num = ib_map_mr_sg(buf_slot->mr[link_idx],
703 			      buf_slot->sgt[link_idx].sgl,
704 			      buf_slot->sgt[link_idx].orig_nents,
705 			      &offset, PAGE_SIZE);
706 
707 	return sg_num;
708 }
709 
710 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
711 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
712 			     struct smc_buf_desc *buf_slot, u8 link_idx)
713 {
714 	if (buf_slot->mr[link_idx])
715 		return 0; /* already done */
716 
717 	buf_slot->mr[link_idx] =
718 		ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
719 	if (IS_ERR(buf_slot->mr[link_idx])) {
720 		int rc;
721 
722 		rc = PTR_ERR(buf_slot->mr[link_idx]);
723 		buf_slot->mr[link_idx] = NULL;
724 		return rc;
725 	}
726 
727 	if (smc_ib_map_mr_sg(buf_slot, link_idx) !=
728 			     buf_slot->sgt[link_idx].orig_nents)
729 		return -EINVAL;
730 
731 	return 0;
732 }
733 
734 bool smc_ib_is_sg_need_sync(struct smc_link *lnk,
735 			    struct smc_buf_desc *buf_slot)
736 {
737 	struct scatterlist *sg;
738 	unsigned int i;
739 	bool ret = false;
740 
741 	/* for now there is just one DMA address */
742 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
743 		    buf_slot->sgt[lnk->link_idx].nents, i) {
744 		if (!sg_dma_len(sg))
745 			break;
746 		if (dma_need_sync(lnk->smcibdev->ibdev->dma_device,
747 				  sg_dma_address(sg))) {
748 			ret = true;
749 			goto out;
750 		}
751 	}
752 
753 out:
754 	return ret;
755 }
756 
757 /* synchronize buffer usage for cpu access */
758 void smc_ib_sync_sg_for_cpu(struct smc_link *lnk,
759 			    struct smc_buf_desc *buf_slot,
760 			    enum dma_data_direction data_direction)
761 {
762 	struct scatterlist *sg;
763 	unsigned int i;
764 
765 	if (!(buf_slot->is_dma_need_sync & (1U << lnk->link_idx)))
766 		return;
767 
768 	/* for now there is just one DMA address */
769 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
770 		    buf_slot->sgt[lnk->link_idx].nents, i) {
771 		if (!sg_dma_len(sg))
772 			break;
773 		ib_dma_sync_single_for_cpu(lnk->smcibdev->ibdev,
774 					   sg_dma_address(sg),
775 					   sg_dma_len(sg),
776 					   data_direction);
777 	}
778 }
779 
780 /* synchronize buffer usage for device access */
781 void smc_ib_sync_sg_for_device(struct smc_link *lnk,
782 			       struct smc_buf_desc *buf_slot,
783 			       enum dma_data_direction data_direction)
784 {
785 	struct scatterlist *sg;
786 	unsigned int i;
787 
788 	if (!(buf_slot->is_dma_need_sync & (1U << lnk->link_idx)))
789 		return;
790 
791 	/* for now there is just one DMA address */
792 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
793 		    buf_slot->sgt[lnk->link_idx].nents, i) {
794 		if (!sg_dma_len(sg))
795 			break;
796 		ib_dma_sync_single_for_device(lnk->smcibdev->ibdev,
797 					      sg_dma_address(sg),
798 					      sg_dma_len(sg),
799 					      data_direction);
800 	}
801 }
802 
803 /* Map a new TX or RX buffer SG-table to DMA */
804 int smc_ib_buf_map_sg(struct smc_link *lnk,
805 		      struct smc_buf_desc *buf_slot,
806 		      enum dma_data_direction data_direction)
807 {
808 	int mapped_nents;
809 
810 	mapped_nents = ib_dma_map_sg(lnk->smcibdev->ibdev,
811 				     buf_slot->sgt[lnk->link_idx].sgl,
812 				     buf_slot->sgt[lnk->link_idx].orig_nents,
813 				     data_direction);
814 	if (!mapped_nents)
815 		return -ENOMEM;
816 
817 	return mapped_nents;
818 }
819 
820 void smc_ib_buf_unmap_sg(struct smc_link *lnk,
821 			 struct smc_buf_desc *buf_slot,
822 			 enum dma_data_direction data_direction)
823 {
824 	if (!buf_slot->sgt[lnk->link_idx].sgl->dma_address)
825 		return; /* already unmapped */
826 
827 	ib_dma_unmap_sg(lnk->smcibdev->ibdev,
828 			buf_slot->sgt[lnk->link_idx].sgl,
829 			buf_slot->sgt[lnk->link_idx].orig_nents,
830 			data_direction);
831 	buf_slot->sgt[lnk->link_idx].sgl->dma_address = 0;
832 }
833 
834 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
835 {
836 	struct ib_cq_init_attr cqattr =	{
837 		.cqe = SMC_MAX_CQE, .comp_vector = 0 };
838 	int cqe_size_order, smc_order;
839 	long rc;
840 
841 	mutex_lock(&smcibdev->mutex);
842 	rc = 0;
843 	if (smcibdev->initialized)
844 		goto out;
845 	/* the calculated number of cq entries fits to mlx5 cq allocation */
846 	cqe_size_order = cache_line_size() == 128 ? 7 : 6;
847 	smc_order = MAX_PAGE_ORDER - cqe_size_order;
848 	if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
849 		cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
850 	smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
851 					      smc_wr_tx_cq_handler, NULL,
852 					      smcibdev, &cqattr);
853 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
854 	if (IS_ERR(smcibdev->roce_cq_send)) {
855 		smcibdev->roce_cq_send = NULL;
856 		goto out;
857 	}
858 	smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
859 					      smc_wr_rx_cq_handler, NULL,
860 					      smcibdev, &cqattr);
861 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
862 	if (IS_ERR(smcibdev->roce_cq_recv)) {
863 		smcibdev->roce_cq_recv = NULL;
864 		goto err;
865 	}
866 	smc_wr_add_dev(smcibdev);
867 	smcibdev->initialized = 1;
868 	goto out;
869 
870 err:
871 	ib_destroy_cq(smcibdev->roce_cq_send);
872 out:
873 	mutex_unlock(&smcibdev->mutex);
874 	return rc;
875 }
876 
877 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
878 {
879 	mutex_lock(&smcibdev->mutex);
880 	if (!smcibdev->initialized)
881 		goto out;
882 	smcibdev->initialized = 0;
883 	ib_destroy_cq(smcibdev->roce_cq_recv);
884 	ib_destroy_cq(smcibdev->roce_cq_send);
885 	smc_wr_remove_dev(smcibdev);
886 out:
887 	mutex_unlock(&smcibdev->mutex);
888 }
889 
890 static struct ib_client smc_ib_client;
891 
892 static void smc_copy_netdev_ifindex(struct smc_ib_device *smcibdev, int port)
893 {
894 	struct ib_device *ibdev = smcibdev->ibdev;
895 	struct net_device *ndev;
896 
897 	if (!ibdev->ops.get_netdev)
898 		return;
899 	ndev = ibdev->ops.get_netdev(ibdev, port + 1);
900 	if (ndev) {
901 		smcibdev->ndev_ifidx[port] = ndev->ifindex;
902 		dev_put(ndev);
903 	}
904 }
905 
906 void smc_ib_ndev_change(struct net_device *ndev, unsigned long event)
907 {
908 	struct smc_ib_device *smcibdev;
909 	struct ib_device *libdev;
910 	struct net_device *lndev;
911 	u8 port_cnt;
912 	int i;
913 
914 	mutex_lock(&smc_ib_devices.mutex);
915 	list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
916 		port_cnt = smcibdev->ibdev->phys_port_cnt;
917 		for (i = 0; i < min_t(size_t, port_cnt, SMC_MAX_PORTS); i++) {
918 			libdev = smcibdev->ibdev;
919 			if (!libdev->ops.get_netdev)
920 				continue;
921 			lndev = libdev->ops.get_netdev(libdev, i + 1);
922 			dev_put(lndev);
923 			if (lndev != ndev)
924 				continue;
925 			if (event == NETDEV_REGISTER)
926 				smcibdev->ndev_ifidx[i] = ndev->ifindex;
927 			if (event == NETDEV_UNREGISTER)
928 				smcibdev->ndev_ifidx[i] = 0;
929 		}
930 	}
931 	mutex_unlock(&smc_ib_devices.mutex);
932 }
933 
934 /* callback function for ib_register_client() */
935 static int smc_ib_add_dev(struct ib_device *ibdev)
936 {
937 	struct smc_ib_device *smcibdev;
938 	u8 port_cnt;
939 	int i;
940 
941 	if (ibdev->node_type != RDMA_NODE_IB_CA)
942 		return -EOPNOTSUPP;
943 
944 	smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
945 	if (!smcibdev)
946 		return -ENOMEM;
947 
948 	smcibdev->ibdev = ibdev;
949 	INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
950 	atomic_set(&smcibdev->lnk_cnt, 0);
951 	init_waitqueue_head(&smcibdev->lnks_deleted);
952 	mutex_init(&smcibdev->mutex);
953 	mutex_lock(&smc_ib_devices.mutex);
954 	list_add_tail(&smcibdev->list, &smc_ib_devices.list);
955 	mutex_unlock(&smc_ib_devices.mutex);
956 	ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
957 	INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
958 			      smc_ib_global_event_handler);
959 	ib_register_event_handler(&smcibdev->event_handler);
960 
961 	/* trigger reading of the port attributes */
962 	port_cnt = smcibdev->ibdev->phys_port_cnt;
963 	pr_warn_ratelimited("smc: adding ib device %s with port count %d\n",
964 			    smcibdev->ibdev->name, port_cnt);
965 	for (i = 0;
966 	     i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
967 	     i++) {
968 		set_bit(i, &smcibdev->port_event_mask);
969 		/* determine pnetids of the port */
970 		if (smc_pnetid_by_dev_port(ibdev->dev.parent, i,
971 					   smcibdev->pnetid[i]))
972 			smc_pnetid_by_table_ib(smcibdev, i + 1);
973 		smc_copy_netdev_ifindex(smcibdev, i);
974 		pr_warn_ratelimited("smc:    ib device %s port %d has pnetid "
975 				    "%.16s%s\n",
976 				    smcibdev->ibdev->name, i + 1,
977 				    smcibdev->pnetid[i],
978 				    smcibdev->pnetid_by_user[i] ?
979 				     " (user defined)" :
980 				     "");
981 	}
982 	schedule_work(&smcibdev->port_event_work);
983 	return 0;
984 }
985 
986 /* callback function for ib_unregister_client() */
987 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
988 {
989 	struct smc_ib_device *smcibdev = client_data;
990 
991 	mutex_lock(&smc_ib_devices.mutex);
992 	list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
993 	mutex_unlock(&smc_ib_devices.mutex);
994 	pr_warn_ratelimited("smc: removing ib device %s\n",
995 			    smcibdev->ibdev->name);
996 	smc_smcr_terminate_all(smcibdev);
997 	smc_ib_cleanup_per_ibdev(smcibdev);
998 	ib_unregister_event_handler(&smcibdev->event_handler);
999 	cancel_work_sync(&smcibdev->port_event_work);
1000 	kfree(smcibdev);
1001 }
1002 
1003 static struct ib_client smc_ib_client = {
1004 	.name	= "smc_ib",
1005 	.add	= smc_ib_add_dev,
1006 	.remove = smc_ib_remove_dev,
1007 };
1008 
1009 int __init smc_ib_register_client(void)
1010 {
1011 	smc_ib_init_local_systemid();
1012 	return ib_register_client(&smc_ib_client);
1013 }
1014 
1015 void smc_ib_unregister_client(void)
1016 {
1017 	ib_unregister_client(&smc_ib_client);
1018 }
1019