xref: /linux/net/smc/smc_ib.c (revision 662fa3d6099374c4615bf64d06895e3573b935b2)
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 <rdma/ib_verbs.h>
21 #include <rdma/ib_cache.h>
22 
23 #include "smc_pnet.h"
24 #include "smc_ib.h"
25 #include "smc_core.h"
26 #include "smc_wr.h"
27 #include "smc.h"
28 #include "smc_netlink.h"
29 
30 #define SMC_MAX_CQE 32766	/* max. # of completion queue elements */
31 
32 #define SMC_QP_MIN_RNR_TIMER		5
33 #define SMC_QP_TIMEOUT			15 /* 4096 * 2 ** timeout usec */
34 #define SMC_QP_RETRY_CNT			7 /* 7: infinite */
35 #define SMC_QP_RNR_RETRY			7 /* 7: infinite */
36 
37 struct smc_ib_devices smc_ib_devices = {	/* smc-registered ib devices */
38 	.mutex = __MUTEX_INITIALIZER(smc_ib_devices.mutex),
39 	.list = LIST_HEAD_INIT(smc_ib_devices.list),
40 };
41 
42 u8 local_systemid[SMC_SYSTEMID_LEN];		/* unique system identifier */
43 
44 static int smc_ib_modify_qp_init(struct smc_link *lnk)
45 {
46 	struct ib_qp_attr qp_attr;
47 
48 	memset(&qp_attr, 0, sizeof(qp_attr));
49 	qp_attr.qp_state = IB_QPS_INIT;
50 	qp_attr.pkey_index = 0;
51 	qp_attr.port_num = lnk->ibport;
52 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
53 				| IB_ACCESS_REMOTE_WRITE;
54 	return ib_modify_qp(lnk->roce_qp, &qp_attr,
55 			    IB_QP_STATE | IB_QP_PKEY_INDEX |
56 			    IB_QP_ACCESS_FLAGS | IB_QP_PORT);
57 }
58 
59 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
60 {
61 	enum ib_qp_attr_mask qp_attr_mask =
62 		IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
63 		IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
64 	struct ib_qp_attr qp_attr;
65 
66 	memset(&qp_attr, 0, sizeof(qp_attr));
67 	qp_attr.qp_state = IB_QPS_RTR;
68 	qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
69 	qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
70 	rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
71 	rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, lnk->sgid_index, 1, 0);
72 	rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
73 	memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
74 	       sizeof(lnk->peer_mac));
75 	qp_attr.dest_qp_num = lnk->peer_qpn;
76 	qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
77 	qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
78 					 * requests
79 					 */
80 	qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
81 
82 	return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
83 }
84 
85 int smc_ib_modify_qp_rts(struct smc_link *lnk)
86 {
87 	struct ib_qp_attr qp_attr;
88 
89 	memset(&qp_attr, 0, sizeof(qp_attr));
90 	qp_attr.qp_state = IB_QPS_RTS;
91 	qp_attr.timeout = SMC_QP_TIMEOUT;	/* local ack timeout */
92 	qp_attr.retry_cnt = SMC_QP_RETRY_CNT;	/* retry count */
93 	qp_attr.rnr_retry = SMC_QP_RNR_RETRY;	/* RNR retries, 7=infinite */
94 	qp_attr.sq_psn = lnk->psn_initial;	/* starting send packet seq # */
95 	qp_attr.max_rd_atomic = 1;	/* # of outstanding RDMA reads and
96 					 * atomic ops allowed
97 					 */
98 	return ib_modify_qp(lnk->roce_qp, &qp_attr,
99 			    IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
100 			    IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
101 			    IB_QP_MAX_QP_RD_ATOMIC);
102 }
103 
104 int smc_ib_modify_qp_reset(struct smc_link *lnk)
105 {
106 	struct ib_qp_attr qp_attr;
107 
108 	memset(&qp_attr, 0, sizeof(qp_attr));
109 	qp_attr.qp_state = IB_QPS_RESET;
110 	return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
111 }
112 
113 int smc_ib_ready_link(struct smc_link *lnk)
114 {
115 	struct smc_link_group *lgr = smc_get_lgr(lnk);
116 	int rc = 0;
117 
118 	rc = smc_ib_modify_qp_init(lnk);
119 	if (rc)
120 		goto out;
121 
122 	rc = smc_ib_modify_qp_rtr(lnk);
123 	if (rc)
124 		goto out;
125 	smc_wr_remember_qp_attr(lnk);
126 	rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
127 			      IB_CQ_SOLICITED_MASK);
128 	if (rc)
129 		goto out;
130 	rc = smc_wr_rx_post_init(lnk);
131 	if (rc)
132 		goto out;
133 	smc_wr_remember_qp_attr(lnk);
134 
135 	if (lgr->role == SMC_SERV) {
136 		rc = smc_ib_modify_qp_rts(lnk);
137 		if (rc)
138 			goto out;
139 		smc_wr_remember_qp_attr(lnk);
140 	}
141 out:
142 	return rc;
143 }
144 
145 static int smc_ib_fill_mac(struct smc_ib_device *smcibdev, u8 ibport)
146 {
147 	const struct ib_gid_attr *attr;
148 	int rc;
149 
150 	attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, 0);
151 	if (IS_ERR(attr))
152 		return -ENODEV;
153 
154 	rc = rdma_read_gid_l2_fields(attr, NULL, smcibdev->mac[ibport - 1]);
155 	rdma_put_gid_attr(attr);
156 	return rc;
157 }
158 
159 /* Create an identifier unique for this instance of SMC-R.
160  * The MAC-address of the first active registered IB device
161  * plus a random 2-byte number is used to create this identifier.
162  * This name is delivered to the peer during connection initialization.
163  */
164 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
165 						u8 ibport)
166 {
167 	memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
168 	       sizeof(smcibdev->mac[ibport - 1]));
169 }
170 
171 bool smc_ib_is_valid_local_systemid(void)
172 {
173 	return !is_zero_ether_addr(&local_systemid[2]);
174 }
175 
176 static void smc_ib_init_local_systemid(void)
177 {
178 	get_random_bytes(&local_systemid[0], 2);
179 }
180 
181 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
182 {
183 	return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
184 }
185 
186 /* determine the gid for an ib-device port and vlan id */
187 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
188 			 unsigned short vlan_id, u8 gid[], u8 *sgid_index)
189 {
190 	const struct ib_gid_attr *attr;
191 	const struct net_device *ndev;
192 	int i;
193 
194 	for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
195 		attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
196 		if (IS_ERR(attr))
197 			continue;
198 
199 		rcu_read_lock();
200 		ndev = rdma_read_gid_attr_ndev_rcu(attr);
201 		if (!IS_ERR(ndev) &&
202 		    ((!vlan_id && !is_vlan_dev(ndev)) ||
203 		     (vlan_id && is_vlan_dev(ndev) &&
204 		      vlan_dev_vlan_id(ndev) == vlan_id)) &&
205 		    attr->gid_type == IB_GID_TYPE_ROCE) {
206 			rcu_read_unlock();
207 			if (gid)
208 				memcpy(gid, &attr->gid, SMC_GID_SIZE);
209 			if (sgid_index)
210 				*sgid_index = attr->index;
211 			rdma_put_gid_attr(attr);
212 			return 0;
213 		}
214 		rcu_read_unlock();
215 		rdma_put_gid_attr(attr);
216 	}
217 	return -ENODEV;
218 }
219 
220 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
221 {
222 	int rc;
223 
224 	memset(&smcibdev->pattr[ibport - 1], 0,
225 	       sizeof(smcibdev->pattr[ibport - 1]));
226 	rc = ib_query_port(smcibdev->ibdev, ibport,
227 			   &smcibdev->pattr[ibport - 1]);
228 	if (rc)
229 		goto out;
230 	/* the SMC protocol requires specification of the RoCE MAC address */
231 	rc = smc_ib_fill_mac(smcibdev, ibport);
232 	if (rc)
233 		goto out;
234 	if (!smc_ib_is_valid_local_systemid() &&
235 	    smc_ib_port_active(smcibdev, ibport))
236 		/* create unique system identifier */
237 		smc_ib_define_local_systemid(smcibdev, ibport);
238 out:
239 	return rc;
240 }
241 
242 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
243 static void smc_ib_port_event_work(struct work_struct *work)
244 {
245 	struct smc_ib_device *smcibdev = container_of(
246 		work, struct smc_ib_device, port_event_work);
247 	u8 port_idx;
248 
249 	for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
250 		smc_ib_remember_port_attr(smcibdev, port_idx + 1);
251 		clear_bit(port_idx, &smcibdev->port_event_mask);
252 		if (!smc_ib_port_active(smcibdev, port_idx + 1)) {
253 			set_bit(port_idx, smcibdev->ports_going_away);
254 			smcr_port_err(smcibdev, port_idx + 1);
255 		} else {
256 			clear_bit(port_idx, smcibdev->ports_going_away);
257 			smcr_port_add(smcibdev, port_idx + 1);
258 		}
259 	}
260 }
261 
262 /* can be called in IRQ context */
263 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
264 					struct ib_event *ibevent)
265 {
266 	struct smc_ib_device *smcibdev;
267 	bool schedule = false;
268 	u8 port_idx;
269 
270 	smcibdev = container_of(handler, struct smc_ib_device, event_handler);
271 
272 	switch (ibevent->event) {
273 	case IB_EVENT_DEVICE_FATAL:
274 		/* terminate all ports on device */
275 		for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) {
276 			set_bit(port_idx, &smcibdev->port_event_mask);
277 			if (!test_and_set_bit(port_idx,
278 					      smcibdev->ports_going_away))
279 				schedule = true;
280 		}
281 		if (schedule)
282 			schedule_work(&smcibdev->port_event_work);
283 		break;
284 	case IB_EVENT_PORT_ACTIVE:
285 		port_idx = ibevent->element.port_num - 1;
286 		if (port_idx >= SMC_MAX_PORTS)
287 			break;
288 		set_bit(port_idx, &smcibdev->port_event_mask);
289 		if (test_and_clear_bit(port_idx, smcibdev->ports_going_away))
290 			schedule_work(&smcibdev->port_event_work);
291 		break;
292 	case IB_EVENT_PORT_ERR:
293 		port_idx = ibevent->element.port_num - 1;
294 		if (port_idx >= SMC_MAX_PORTS)
295 			break;
296 		set_bit(port_idx, &smcibdev->port_event_mask);
297 		if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
298 			schedule_work(&smcibdev->port_event_work);
299 		break;
300 	case IB_EVENT_GID_CHANGE:
301 		port_idx = ibevent->element.port_num - 1;
302 		if (port_idx >= SMC_MAX_PORTS)
303 			break;
304 		set_bit(port_idx, &smcibdev->port_event_mask);
305 		schedule_work(&smcibdev->port_event_work);
306 		break;
307 	default:
308 		break;
309 	}
310 }
311 
312 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
313 {
314 	if (lnk->roce_pd)
315 		ib_dealloc_pd(lnk->roce_pd);
316 	lnk->roce_pd = NULL;
317 }
318 
319 int smc_ib_create_protection_domain(struct smc_link *lnk)
320 {
321 	int rc;
322 
323 	lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
324 	rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
325 	if (IS_ERR(lnk->roce_pd))
326 		lnk->roce_pd = NULL;
327 	return rc;
328 }
329 
330 static bool smcr_diag_is_dev_critical(struct smc_lgr_list *smc_lgr,
331 				      struct smc_ib_device *smcibdev)
332 {
333 	struct smc_link_group *lgr;
334 	bool rc = false;
335 	int i;
336 
337 	spin_lock_bh(&smc_lgr->lock);
338 	list_for_each_entry(lgr, &smc_lgr->list, list) {
339 		if (lgr->is_smcd)
340 			continue;
341 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
342 			if (lgr->lnk[i].state == SMC_LNK_UNUSED ||
343 			    lgr->lnk[i].smcibdev != smcibdev)
344 				continue;
345 			if (lgr->type == SMC_LGR_SINGLE ||
346 			    lgr->type == SMC_LGR_ASYMMETRIC_LOCAL) {
347 				rc = true;
348 				goto out;
349 			}
350 		}
351 	}
352 out:
353 	spin_unlock_bh(&smc_lgr->lock);
354 	return rc;
355 }
356 
357 static int smc_nl_handle_dev_port(struct sk_buff *skb,
358 				  struct ib_device *ibdev,
359 				  struct smc_ib_device *smcibdev,
360 				  int port)
361 {
362 	char smc_pnet[SMC_MAX_PNETID_LEN + 1];
363 	struct nlattr *port_attrs;
364 	unsigned char port_state;
365 	int lnk_count = 0;
366 
367 	port_attrs = nla_nest_start(skb, SMC_NLA_DEV_PORT + port);
368 	if (!port_attrs)
369 		goto errout;
370 
371 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_PNET_USR,
372 		       smcibdev->pnetid_by_user[port]))
373 		goto errattr;
374 	memcpy(smc_pnet, &smcibdev->pnetid[port], SMC_MAX_PNETID_LEN);
375 	smc_pnet[SMC_MAX_PNETID_LEN] = 0;
376 	if (nla_put_string(skb, SMC_NLA_DEV_PORT_PNETID, smc_pnet))
377 		goto errattr;
378 	if (nla_put_u32(skb, SMC_NLA_DEV_PORT_NETDEV,
379 			smcibdev->ndev_ifidx[port]))
380 		goto errattr;
381 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_VALID, 1))
382 		goto errattr;
383 	port_state = smc_ib_port_active(smcibdev, port + 1);
384 	if (nla_put_u8(skb, SMC_NLA_DEV_PORT_STATE, port_state))
385 		goto errattr;
386 	lnk_count = atomic_read(&smcibdev->lnk_cnt_by_port[port]);
387 	if (nla_put_u32(skb, SMC_NLA_DEV_PORT_LNK_CNT, lnk_count))
388 		goto errattr;
389 	nla_nest_end(skb, port_attrs);
390 	return 0;
391 errattr:
392 	nla_nest_cancel(skb, port_attrs);
393 errout:
394 	return -EMSGSIZE;
395 }
396 
397 static bool smc_nl_handle_pci_values(const struct smc_pci_dev *smc_pci_dev,
398 				     struct sk_buff *skb)
399 {
400 	if (nla_put_u32(skb, SMC_NLA_DEV_PCI_FID, smc_pci_dev->pci_fid))
401 		return false;
402 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_CHID, smc_pci_dev->pci_pchid))
403 		return false;
404 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_VENDOR, smc_pci_dev->pci_vendor))
405 		return false;
406 	if (nla_put_u16(skb, SMC_NLA_DEV_PCI_DEVICE, smc_pci_dev->pci_device))
407 		return false;
408 	if (nla_put_string(skb, SMC_NLA_DEV_PCI_ID, smc_pci_dev->pci_id))
409 		return false;
410 	return true;
411 }
412 
413 static int smc_nl_handle_smcr_dev(struct smc_ib_device *smcibdev,
414 				  struct sk_buff *skb,
415 				  struct netlink_callback *cb)
416 {
417 	char smc_ibname[IB_DEVICE_NAME_MAX];
418 	struct smc_pci_dev smc_pci_dev;
419 	struct pci_dev *pci_dev;
420 	unsigned char is_crit;
421 	struct nlattr *attrs;
422 	void *nlh;
423 	int i;
424 
425 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
426 			  &smc_gen_nl_family, NLM_F_MULTI,
427 			  SMC_NETLINK_GET_DEV_SMCR);
428 	if (!nlh)
429 		goto errmsg;
430 	attrs = nla_nest_start(skb, SMC_GEN_DEV_SMCR);
431 	if (!attrs)
432 		goto errout;
433 	is_crit = smcr_diag_is_dev_critical(&smc_lgr_list, smcibdev);
434 	if (nla_put_u8(skb, SMC_NLA_DEV_IS_CRIT, is_crit))
435 		goto errattr;
436 	if (smcibdev->ibdev->dev.parent) {
437 		memset(&smc_pci_dev, 0, sizeof(smc_pci_dev));
438 		pci_dev = to_pci_dev(smcibdev->ibdev->dev.parent);
439 		smc_set_pci_values(pci_dev, &smc_pci_dev);
440 		if (!smc_nl_handle_pci_values(&smc_pci_dev, skb))
441 			goto errattr;
442 	}
443 	snprintf(smc_ibname, sizeof(smc_ibname), "%s", smcibdev->ibdev->name);
444 	if (nla_put_string(skb, SMC_NLA_DEV_IB_NAME, smc_ibname))
445 		goto errattr;
446 	for (i = 1; i <= SMC_MAX_PORTS; i++) {
447 		if (!rdma_is_port_valid(smcibdev->ibdev, i))
448 			continue;
449 		if (smc_nl_handle_dev_port(skb, smcibdev->ibdev,
450 					   smcibdev, i - 1))
451 			goto errattr;
452 	}
453 
454 	nla_nest_end(skb, attrs);
455 	genlmsg_end(skb, nlh);
456 	return 0;
457 
458 errattr:
459 	nla_nest_cancel(skb, attrs);
460 errout:
461 	genlmsg_cancel(skb, nlh);
462 errmsg:
463 	return -EMSGSIZE;
464 }
465 
466 static void smc_nl_prep_smcr_dev(struct smc_ib_devices *dev_list,
467 				 struct sk_buff *skb,
468 				 struct netlink_callback *cb)
469 {
470 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
471 	struct smc_ib_device *smcibdev;
472 	int snum = cb_ctx->pos[0];
473 	int num = 0;
474 
475 	mutex_lock(&dev_list->mutex);
476 	list_for_each_entry(smcibdev, &dev_list->list, list) {
477 		if (num < snum)
478 			goto next;
479 		if (smc_nl_handle_smcr_dev(smcibdev, skb, cb))
480 			goto errout;
481 next:
482 		num++;
483 	}
484 errout:
485 	mutex_unlock(&dev_list->mutex);
486 	cb_ctx->pos[0] = num;
487 }
488 
489 int smcr_nl_get_device(struct sk_buff *skb, struct netlink_callback *cb)
490 {
491 	smc_nl_prep_smcr_dev(&smc_ib_devices, skb, cb);
492 	return skb->len;
493 }
494 
495 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
496 {
497 	struct smc_link *lnk = (struct smc_link *)priv;
498 	struct smc_ib_device *smcibdev = lnk->smcibdev;
499 	u8 port_idx;
500 
501 	switch (ibevent->event) {
502 	case IB_EVENT_QP_FATAL:
503 	case IB_EVENT_QP_ACCESS_ERR:
504 		port_idx = ibevent->element.qp->port - 1;
505 		if (port_idx >= SMC_MAX_PORTS)
506 			break;
507 		set_bit(port_idx, &smcibdev->port_event_mask);
508 		if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
509 			schedule_work(&smcibdev->port_event_work);
510 		break;
511 	default:
512 		break;
513 	}
514 }
515 
516 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
517 {
518 	if (lnk->roce_qp)
519 		ib_destroy_qp(lnk->roce_qp);
520 	lnk->roce_qp = NULL;
521 }
522 
523 /* create a queue pair within the protection domain for a link */
524 int smc_ib_create_queue_pair(struct smc_link *lnk)
525 {
526 	struct ib_qp_init_attr qp_attr = {
527 		.event_handler = smc_ib_qp_event_handler,
528 		.qp_context = lnk,
529 		.send_cq = lnk->smcibdev->roce_cq_send,
530 		.recv_cq = lnk->smcibdev->roce_cq_recv,
531 		.srq = NULL,
532 		.cap = {
533 				/* include unsolicited rdma_writes as well,
534 				 * there are max. 2 RDMA_WRITE per 1 WR_SEND
535 				 */
536 			.max_send_wr = SMC_WR_BUF_CNT * 3,
537 			.max_recv_wr = SMC_WR_BUF_CNT * 3,
538 			.max_send_sge = SMC_IB_MAX_SEND_SGE,
539 			.max_recv_sge = 1,
540 		},
541 		.sq_sig_type = IB_SIGNAL_REQ_WR,
542 		.qp_type = IB_QPT_RC,
543 	};
544 	int rc;
545 
546 	lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
547 	rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
548 	if (IS_ERR(lnk->roce_qp))
549 		lnk->roce_qp = NULL;
550 	else
551 		smc_wr_remember_qp_attr(lnk);
552 	return rc;
553 }
554 
555 void smc_ib_put_memory_region(struct ib_mr *mr)
556 {
557 	ib_dereg_mr(mr);
558 }
559 
560 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot, u8 link_idx)
561 {
562 	unsigned int offset = 0;
563 	int sg_num;
564 
565 	/* map the largest prefix of a dma mapped SG list */
566 	sg_num = ib_map_mr_sg(buf_slot->mr_rx[link_idx],
567 			      buf_slot->sgt[link_idx].sgl,
568 			      buf_slot->sgt[link_idx].orig_nents,
569 			      &offset, PAGE_SIZE);
570 
571 	return sg_num;
572 }
573 
574 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
575 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
576 			     struct smc_buf_desc *buf_slot, u8 link_idx)
577 {
578 	if (buf_slot->mr_rx[link_idx])
579 		return 0; /* already done */
580 
581 	buf_slot->mr_rx[link_idx] =
582 		ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
583 	if (IS_ERR(buf_slot->mr_rx[link_idx])) {
584 		int rc;
585 
586 		rc = PTR_ERR(buf_slot->mr_rx[link_idx]);
587 		buf_slot->mr_rx[link_idx] = NULL;
588 		return rc;
589 	}
590 
591 	if (smc_ib_map_mr_sg(buf_slot, link_idx) != 1)
592 		return -EINVAL;
593 
594 	return 0;
595 }
596 
597 /* synchronize buffer usage for cpu access */
598 void smc_ib_sync_sg_for_cpu(struct smc_link *lnk,
599 			    struct smc_buf_desc *buf_slot,
600 			    enum dma_data_direction data_direction)
601 {
602 	struct scatterlist *sg;
603 	unsigned int i;
604 
605 	/* for now there is just one DMA address */
606 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
607 		    buf_slot->sgt[lnk->link_idx].nents, i) {
608 		if (!sg_dma_len(sg))
609 			break;
610 		ib_dma_sync_single_for_cpu(lnk->smcibdev->ibdev,
611 					   sg_dma_address(sg),
612 					   sg_dma_len(sg),
613 					   data_direction);
614 	}
615 }
616 
617 /* synchronize buffer usage for device access */
618 void smc_ib_sync_sg_for_device(struct smc_link *lnk,
619 			       struct smc_buf_desc *buf_slot,
620 			       enum dma_data_direction data_direction)
621 {
622 	struct scatterlist *sg;
623 	unsigned int i;
624 
625 	/* for now there is just one DMA address */
626 	for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
627 		    buf_slot->sgt[lnk->link_idx].nents, i) {
628 		if (!sg_dma_len(sg))
629 			break;
630 		ib_dma_sync_single_for_device(lnk->smcibdev->ibdev,
631 					      sg_dma_address(sg),
632 					      sg_dma_len(sg),
633 					      data_direction);
634 	}
635 }
636 
637 /* Map a new TX or RX buffer SG-table to DMA */
638 int smc_ib_buf_map_sg(struct smc_link *lnk,
639 		      struct smc_buf_desc *buf_slot,
640 		      enum dma_data_direction data_direction)
641 {
642 	int mapped_nents;
643 
644 	mapped_nents = ib_dma_map_sg(lnk->smcibdev->ibdev,
645 				     buf_slot->sgt[lnk->link_idx].sgl,
646 				     buf_slot->sgt[lnk->link_idx].orig_nents,
647 				     data_direction);
648 	if (!mapped_nents)
649 		return -ENOMEM;
650 
651 	return mapped_nents;
652 }
653 
654 void smc_ib_buf_unmap_sg(struct smc_link *lnk,
655 			 struct smc_buf_desc *buf_slot,
656 			 enum dma_data_direction data_direction)
657 {
658 	if (!buf_slot->sgt[lnk->link_idx].sgl->dma_address)
659 		return; /* already unmapped */
660 
661 	ib_dma_unmap_sg(lnk->smcibdev->ibdev,
662 			buf_slot->sgt[lnk->link_idx].sgl,
663 			buf_slot->sgt[lnk->link_idx].orig_nents,
664 			data_direction);
665 	buf_slot->sgt[lnk->link_idx].sgl->dma_address = 0;
666 }
667 
668 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
669 {
670 	struct ib_cq_init_attr cqattr =	{
671 		.cqe = SMC_MAX_CQE, .comp_vector = 0 };
672 	int cqe_size_order, smc_order;
673 	long rc;
674 
675 	mutex_lock(&smcibdev->mutex);
676 	rc = 0;
677 	if (smcibdev->initialized)
678 		goto out;
679 	/* the calculated number of cq entries fits to mlx5 cq allocation */
680 	cqe_size_order = cache_line_size() == 128 ? 7 : 6;
681 	smc_order = MAX_ORDER - cqe_size_order - 1;
682 	if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
683 		cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
684 	smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
685 					      smc_wr_tx_cq_handler, NULL,
686 					      smcibdev, &cqattr);
687 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
688 	if (IS_ERR(smcibdev->roce_cq_send)) {
689 		smcibdev->roce_cq_send = NULL;
690 		goto out;
691 	}
692 	smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
693 					      smc_wr_rx_cq_handler, NULL,
694 					      smcibdev, &cqattr);
695 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
696 	if (IS_ERR(smcibdev->roce_cq_recv)) {
697 		smcibdev->roce_cq_recv = NULL;
698 		goto err;
699 	}
700 	smc_wr_add_dev(smcibdev);
701 	smcibdev->initialized = 1;
702 	goto out;
703 
704 err:
705 	ib_destroy_cq(smcibdev->roce_cq_send);
706 out:
707 	mutex_unlock(&smcibdev->mutex);
708 	return rc;
709 }
710 
711 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
712 {
713 	mutex_lock(&smcibdev->mutex);
714 	if (!smcibdev->initialized)
715 		goto out;
716 	smcibdev->initialized = 0;
717 	ib_destroy_cq(smcibdev->roce_cq_recv);
718 	ib_destroy_cq(smcibdev->roce_cq_send);
719 	smc_wr_remove_dev(smcibdev);
720 out:
721 	mutex_unlock(&smcibdev->mutex);
722 }
723 
724 static struct ib_client smc_ib_client;
725 
726 static void smc_copy_netdev_ifindex(struct smc_ib_device *smcibdev, int port)
727 {
728 	struct ib_device *ibdev = smcibdev->ibdev;
729 	struct net_device *ndev;
730 
731 	if (!ibdev->ops.get_netdev)
732 		return;
733 	ndev = ibdev->ops.get_netdev(ibdev, port + 1);
734 	if (ndev) {
735 		smcibdev->ndev_ifidx[port] = ndev->ifindex;
736 		dev_put(ndev);
737 	}
738 }
739 
740 void smc_ib_ndev_change(struct net_device *ndev, unsigned long event)
741 {
742 	struct smc_ib_device *smcibdev;
743 	struct ib_device *libdev;
744 	struct net_device *lndev;
745 	u8 port_cnt;
746 	int i;
747 
748 	mutex_lock(&smc_ib_devices.mutex);
749 	list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
750 		port_cnt = smcibdev->ibdev->phys_port_cnt;
751 		for (i = 0; i < min_t(size_t, port_cnt, SMC_MAX_PORTS); i++) {
752 			libdev = smcibdev->ibdev;
753 			if (!libdev->ops.get_netdev)
754 				continue;
755 			lndev = libdev->ops.get_netdev(libdev, i + 1);
756 			if (lndev)
757 				dev_put(lndev);
758 			if (lndev != ndev)
759 				continue;
760 			if (event == NETDEV_REGISTER)
761 				smcibdev->ndev_ifidx[i] = ndev->ifindex;
762 			if (event == NETDEV_UNREGISTER)
763 				smcibdev->ndev_ifidx[i] = 0;
764 		}
765 	}
766 	mutex_unlock(&smc_ib_devices.mutex);
767 }
768 
769 /* callback function for ib_register_client() */
770 static int smc_ib_add_dev(struct ib_device *ibdev)
771 {
772 	struct smc_ib_device *smcibdev;
773 	u8 port_cnt;
774 	int i;
775 
776 	if (ibdev->node_type != RDMA_NODE_IB_CA)
777 		return -EOPNOTSUPP;
778 
779 	smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
780 	if (!smcibdev)
781 		return -ENOMEM;
782 
783 	smcibdev->ibdev = ibdev;
784 	INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
785 	atomic_set(&smcibdev->lnk_cnt, 0);
786 	init_waitqueue_head(&smcibdev->lnks_deleted);
787 	mutex_init(&smcibdev->mutex);
788 	mutex_lock(&smc_ib_devices.mutex);
789 	list_add_tail(&smcibdev->list, &smc_ib_devices.list);
790 	mutex_unlock(&smc_ib_devices.mutex);
791 	ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
792 	INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
793 			      smc_ib_global_event_handler);
794 	ib_register_event_handler(&smcibdev->event_handler);
795 
796 	/* trigger reading of the port attributes */
797 	port_cnt = smcibdev->ibdev->phys_port_cnt;
798 	pr_warn_ratelimited("smc: adding ib device %s with port count %d\n",
799 			    smcibdev->ibdev->name, port_cnt);
800 	for (i = 0;
801 	     i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
802 	     i++) {
803 		set_bit(i, &smcibdev->port_event_mask);
804 		/* determine pnetids of the port */
805 		if (smc_pnetid_by_dev_port(ibdev->dev.parent, i,
806 					   smcibdev->pnetid[i]))
807 			smc_pnetid_by_table_ib(smcibdev, i + 1);
808 		smc_copy_netdev_ifindex(smcibdev, i);
809 		pr_warn_ratelimited("smc:    ib device %s port %d has pnetid "
810 				    "%.16s%s\n",
811 				    smcibdev->ibdev->name, i + 1,
812 				    smcibdev->pnetid[i],
813 				    smcibdev->pnetid_by_user[i] ?
814 				     " (user defined)" :
815 				     "");
816 	}
817 	schedule_work(&smcibdev->port_event_work);
818 	return 0;
819 }
820 
821 /* callback function for ib_unregister_client() */
822 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
823 {
824 	struct smc_ib_device *smcibdev = client_data;
825 
826 	mutex_lock(&smc_ib_devices.mutex);
827 	list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
828 	mutex_unlock(&smc_ib_devices.mutex);
829 	pr_warn_ratelimited("smc: removing ib device %s\n",
830 			    smcibdev->ibdev->name);
831 	smc_smcr_terminate_all(smcibdev);
832 	smc_ib_cleanup_per_ibdev(smcibdev);
833 	ib_unregister_event_handler(&smcibdev->event_handler);
834 	cancel_work_sync(&smcibdev->port_event_work);
835 	kfree(smcibdev);
836 }
837 
838 static struct ib_client smc_ib_client = {
839 	.name	= "smc_ib",
840 	.add	= smc_ib_add_dev,
841 	.remove = smc_ib_remove_dev,
842 };
843 
844 int __init smc_ib_register_client(void)
845 {
846 	smc_ib_init_local_systemid();
847 	return ib_register_client(&smc_ib_client);
848 }
849 
850 void smc_ib_unregister_client(void)
851 {
852 	ib_unregister_client(&smc_ib_client);
853 }
854