xref: /linux/net/smc/smc_pnet.c (revision 4246b92cf9fb32da8d8b060c92d8302797c6fbea) 
1 /*
2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
3  *
4  *  Generic netlink support functions to configure an SMC-R PNET table
5  *
6  *  Copyright IBM Corp. 2016
7  *
8  *  Author(s):  Thomas Richter <tmricht@linux.vnet.ibm.com>
9  */
10 
11 #include <linux/module.h>
12 #include <linux/list.h>
13 #include <linux/ctype.h>
14 #include <net/netlink.h>
15 #include <net/genetlink.h>
16 
17 #include <uapi/linux/if.h>
18 #include <uapi/linux/smc.h>
19 
20 #include <rdma/ib_verbs.h>
21 
22 #include "smc_pnet.h"
23 #include "smc_ib.h"
24 
25 #define SMC_MAX_PNET_ID_LEN	16	/* Max. length of PNET id */
26 
27 static struct nla_policy smc_pnet_policy[SMC_PNETID_MAX + 1] = {
28 	[SMC_PNETID_NAME] = {
29 		.type = NLA_NUL_STRING,
30 		.len = SMC_MAX_PNET_ID_LEN - 1
31 	},
32 	[SMC_PNETID_ETHNAME] = {
33 		.type = NLA_NUL_STRING,
34 		.len = IFNAMSIZ - 1
35 	},
36 	[SMC_PNETID_IBNAME] = {
37 		.type = NLA_NUL_STRING,
38 		.len = IB_DEVICE_NAME_MAX - 1
39 	},
40 	[SMC_PNETID_IBPORT] = { .type = NLA_U8 }
41 };
42 
43 static struct genl_family smc_pnet_nl_family;
44 
45 /**
46  * struct smc_pnettable - SMC PNET table anchor
47  * @lock: Lock for list action
48  * @pnetlist: List of PNETIDs
49  */
50 static struct smc_pnettable {
51 	rwlock_t lock;
52 	struct list_head pnetlist;
53 } smc_pnettable = {
54 	.pnetlist = LIST_HEAD_INIT(smc_pnettable.pnetlist),
55 	.lock = __RW_LOCK_UNLOCKED(smc_pnettable.lock)
56 };
57 
58 /**
59  * struct smc_pnetentry - pnet identifier name entry
60  * @list: List node.
61  * @pnet_name: Pnet identifier name
62  * @ndev: pointer to network device.
63  * @smcibdev: Pointer to IB device.
64  */
65 struct smc_pnetentry {
66 	struct list_head list;
67 	char pnet_name[SMC_MAX_PNET_ID_LEN + 1];
68 	struct net_device *ndev;
69 	struct smc_ib_device *smcibdev;
70 	u8 ib_port;
71 };
72 
73 /* Check if two RDMA device entries are identical. Use device name and port
74  * number for comparison.
75  */
76 static bool smc_pnet_same_ibname(struct smc_pnetentry *pnetelem, char *ibname,
77 				 u8 ibport)
78 {
79 	return pnetelem->ib_port == ibport &&
80 	       !strncmp(pnetelem->smcibdev->ibdev->name, ibname,
81 			sizeof(pnetelem->smcibdev->ibdev->name));
82 }
83 
84 /* Find a pnetid in the pnet table.
85  */
86 static struct smc_pnetentry *smc_pnet_find_pnetid(char *pnet_name)
87 {
88 	struct smc_pnetentry *pnetelem, *found_pnetelem = NULL;
89 
90 	read_lock(&smc_pnettable.lock);
91 	list_for_each_entry(pnetelem, &smc_pnettable.pnetlist, list) {
92 		if (!strncmp(pnetelem->pnet_name, pnet_name,
93 			     sizeof(pnetelem->pnet_name))) {
94 			found_pnetelem = pnetelem;
95 			break;
96 		}
97 	}
98 	read_unlock(&smc_pnettable.lock);
99 	return found_pnetelem;
100 }
101 
102 /* Remove a pnetid from the pnet table.
103  */
104 static int smc_pnet_remove_by_pnetid(char *pnet_name)
105 {
106 	struct smc_pnetentry *pnetelem, *tmp_pe;
107 	int rc = -ENOENT;
108 
109 	write_lock(&smc_pnettable.lock);
110 	list_for_each_entry_safe(pnetelem, tmp_pe, &smc_pnettable.pnetlist,
111 				 list) {
112 		if (!strncmp(pnetelem->pnet_name, pnet_name,
113 			     sizeof(pnetelem->pnet_name))) {
114 			list_del(&pnetelem->list);
115 			dev_put(pnetelem->ndev);
116 			kfree(pnetelem);
117 			rc = 0;
118 			break;
119 		}
120 	}
121 	write_unlock(&smc_pnettable.lock);
122 	return rc;
123 }
124 
125 /* Remove a pnet entry mentioning a given network device from the pnet table.
126  */
127 static int smc_pnet_remove_by_ndev(struct net_device *ndev)
128 {
129 	struct smc_pnetentry *pnetelem, *tmp_pe;
130 	int rc = -ENOENT;
131 
132 	write_lock(&smc_pnettable.lock);
133 	list_for_each_entry_safe(pnetelem, tmp_pe, &smc_pnettable.pnetlist,
134 				 list) {
135 		if (pnetelem->ndev == ndev) {
136 			list_del(&pnetelem->list);
137 			dev_put(pnetelem->ndev);
138 			kfree(pnetelem);
139 			rc = 0;
140 			break;
141 		}
142 	}
143 	write_unlock(&smc_pnettable.lock);
144 	return rc;
145 }
146 
147 /* Remove a pnet entry mentioning a given ib device from the pnet table.
148  */
149 int smc_pnet_remove_by_ibdev(struct smc_ib_device *ibdev)
150 {
151 	struct smc_pnetentry *pnetelem, *tmp_pe;
152 	int rc = -ENOENT;
153 
154 	write_lock(&smc_pnettable.lock);
155 	list_for_each_entry_safe(pnetelem, tmp_pe, &smc_pnettable.pnetlist,
156 				 list) {
157 		if (pnetelem->smcibdev == ibdev) {
158 			list_del(&pnetelem->list);
159 			dev_put(pnetelem->ndev);
160 			kfree(pnetelem);
161 			rc = 0;
162 			break;
163 		}
164 	}
165 	write_unlock(&smc_pnettable.lock);
166 	return rc;
167 }
168 
169 /* Append a pnetid to the end of the pnet table if not already on this list.
170  */
171 static int smc_pnet_enter(struct smc_pnetentry *new_pnetelem)
172 {
173 	struct smc_pnetentry *pnetelem;
174 	int rc = -EEXIST;
175 
176 	write_lock(&smc_pnettable.lock);
177 	list_for_each_entry(pnetelem, &smc_pnettable.pnetlist, list) {
178 		if (!strncmp(pnetelem->pnet_name, new_pnetelem->pnet_name,
179 			     sizeof(new_pnetelem->pnet_name)) ||
180 		    !strncmp(pnetelem->ndev->name, new_pnetelem->ndev->name,
181 			     sizeof(new_pnetelem->ndev->name)) ||
182 		    smc_pnet_same_ibname(pnetelem,
183 					 new_pnetelem->smcibdev->ibdev->name,
184 					 new_pnetelem->ib_port)) {
185 			dev_put(pnetelem->ndev);
186 			goto found;
187 		}
188 	}
189 	list_add_tail(&new_pnetelem->list, &smc_pnettable.pnetlist);
190 	rc = 0;
191 found:
192 	write_unlock(&smc_pnettable.lock);
193 	return rc;
194 }
195 
196 /* The limit for pnetid is 16 characters.
197  * Valid characters should be (single-byte character set) a-z, A-Z, 0-9.
198  * Lower case letters are converted to upper case.
199  * Interior blanks should not be used.
200  */
201 static bool smc_pnetid_valid(const char *pnet_name, char *pnetid)
202 {
203 	char *bf = skip_spaces(pnet_name);
204 	size_t len = strlen(bf);
205 	char *end = bf + len;
206 
207 	if (!len)
208 		return false;
209 	while (--end >= bf && isspace(*end))
210 		;
211 	if (end - bf >= SMC_MAX_PNET_ID_LEN)
212 		return false;
213 	while (bf <= end) {
214 		if (!isalnum(*bf))
215 			return false;
216 		*pnetid++ = islower(*bf) ? toupper(*bf) : *bf;
217 		bf++;
218 	}
219 	*pnetid = '\0';
220 	return true;
221 }
222 
223 /* Find an infiniband device by a given name. The device might not exist. */
224 static struct smc_ib_device *smc_pnet_find_ib(char *ib_name)
225 {
226 	struct smc_ib_device *ibdev;
227 
228 	spin_lock(&smc_ib_devices.lock);
229 	list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
230 		if (!strncmp(ibdev->ibdev->name, ib_name,
231 			     sizeof(ibdev->ibdev->name))) {
232 			goto out;
233 		}
234 	}
235 	ibdev = NULL;
236 out:
237 	spin_unlock(&smc_ib_devices.lock);
238 	return ibdev;
239 }
240 
241 /* Parse the supplied netlink attributes and fill a pnetentry structure.
242  * For ethernet and infiniband device names verify that the devices exist.
243  */
244 static int smc_pnet_fill_entry(struct net *net, struct smc_pnetentry *pnetelem,
245 			       struct nlattr *tb[])
246 {
247 	char *string, *ibname = NULL;
248 	int rc = 0;
249 
250 	memset(pnetelem, 0, sizeof(*pnetelem));
251 	INIT_LIST_HEAD(&pnetelem->list);
252 	if (tb[SMC_PNETID_NAME]) {
253 		string = (char *)nla_data(tb[SMC_PNETID_NAME]);
254 		if (!smc_pnetid_valid(string, pnetelem->pnet_name)) {
255 			rc = -EINVAL;
256 			goto error;
257 		}
258 	}
259 	if (tb[SMC_PNETID_ETHNAME]) {
260 		string = (char *)nla_data(tb[SMC_PNETID_ETHNAME]);
261 		pnetelem->ndev = dev_get_by_name(net, string);
262 		if (!pnetelem->ndev)
263 			return -ENOENT;
264 	}
265 	if (tb[SMC_PNETID_IBNAME]) {
266 		ibname = (char *)nla_data(tb[SMC_PNETID_IBNAME]);
267 		ibname = strim(ibname);
268 		pnetelem->smcibdev = smc_pnet_find_ib(ibname);
269 		if (!pnetelem->smcibdev) {
270 			rc = -ENOENT;
271 			goto error;
272 		}
273 	}
274 	if (tb[SMC_PNETID_IBPORT]) {
275 		pnetelem->ib_port = nla_get_u8(tb[SMC_PNETID_IBPORT]);
276 		if (pnetelem->ib_port > SMC_MAX_PORTS) {
277 			rc = -EINVAL;
278 			goto error;
279 		}
280 	}
281 	return 0;
282 
283 error:
284 	if (pnetelem->ndev)
285 		dev_put(pnetelem->ndev);
286 	return rc;
287 }
288 
289 /* Convert an smc_pnetentry to a netlink attribute sequence */
290 static int smc_pnet_set_nla(struct sk_buff *msg, struct smc_pnetentry *pnetelem)
291 {
292 	if (nla_put_string(msg, SMC_PNETID_NAME, pnetelem->pnet_name) ||
293 	    nla_put_string(msg, SMC_PNETID_ETHNAME, pnetelem->ndev->name) ||
294 	    nla_put_string(msg, SMC_PNETID_IBNAME,
295 			   pnetelem->smcibdev->ibdev->name) ||
296 	    nla_put_u8(msg, SMC_PNETID_IBPORT, pnetelem->ib_port))
297 		return -1;
298 	return 0;
299 }
300 
301 /* Retrieve one PNETID entry */
302 static int smc_pnet_get(struct sk_buff *skb, struct genl_info *info)
303 {
304 	struct smc_pnetentry *pnetelem;
305 	struct sk_buff *msg;
306 	void *hdr;
307 	int rc;
308 
309 	pnetelem = smc_pnet_find_pnetid(
310 				(char *)nla_data(info->attrs[SMC_PNETID_NAME]));
311 	if (!pnetelem)
312 		return -ENOENT;
313 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
314 	if (!msg)
315 		return -ENOMEM;
316 
317 	hdr = genlmsg_put(msg, info->snd_portid, info->snd_seq,
318 			  &smc_pnet_nl_family, 0, SMC_PNETID_GET);
319 	if (!hdr) {
320 		rc = -EMSGSIZE;
321 		goto err_out;
322 	}
323 
324 	if (smc_pnet_set_nla(msg, pnetelem)) {
325 		rc = -ENOBUFS;
326 		goto err_out;
327 	}
328 
329 	genlmsg_end(msg, hdr);
330 	return genlmsg_reply(msg, info);
331 
332 err_out:
333 	nlmsg_free(msg);
334 	return rc;
335 }
336 
337 static int smc_pnet_add(struct sk_buff *skb, struct genl_info *info)
338 {
339 	struct net *net = genl_info_net(info);
340 	struct smc_pnetentry *pnetelem;
341 	int rc;
342 
343 	pnetelem = kzalloc(sizeof(*pnetelem), GFP_KERNEL);
344 	if (!pnetelem)
345 		return -ENOMEM;
346 	rc = smc_pnet_fill_entry(net, pnetelem, info->attrs);
347 	if (!rc)
348 		rc = smc_pnet_enter(pnetelem);
349 	if (rc) {
350 		kfree(pnetelem);
351 		return rc;
352 	}
353 	rc = smc_ib_remember_port_attr(pnetelem->smcibdev, pnetelem->ib_port);
354 	if (rc)
355 		smc_pnet_remove_by_pnetid(pnetelem->pnet_name);
356 	return rc;
357 }
358 
359 static int smc_pnet_del(struct sk_buff *skb, struct genl_info *info)
360 {
361 	return smc_pnet_remove_by_pnetid(
362 				(char *)nla_data(info->attrs[SMC_PNETID_NAME]));
363 }
364 
365 static int smc_pnet_dump_start(struct netlink_callback *cb)
366 {
367 	cb->args[0] = 0;
368 	return 0;
369 }
370 
371 static int smc_pnet_dumpinfo(struct sk_buff *skb,
372 			     u32 portid, u32 seq, u32 flags,
373 			     struct smc_pnetentry *pnetelem)
374 {
375 	void *hdr;
376 
377 	hdr = genlmsg_put(skb, portid, seq, &smc_pnet_nl_family,
378 			  flags, SMC_PNETID_GET);
379 	if (!hdr)
380 		return -ENOMEM;
381 	if (smc_pnet_set_nla(skb, pnetelem) < 0) {
382 		genlmsg_cancel(skb, hdr);
383 		return -EMSGSIZE;
384 	}
385 	genlmsg_end(skb, hdr);
386 	return 0;
387 }
388 
389 static int smc_pnet_dump(struct sk_buff *skb, struct netlink_callback *cb)
390 {
391 	struct smc_pnetentry *pnetelem;
392 	int idx = 0;
393 
394 	read_lock(&smc_pnettable.lock);
395 	list_for_each_entry(pnetelem, &smc_pnettable.pnetlist, list) {
396 		if (idx++ < cb->args[0])
397 			continue;
398 		if (smc_pnet_dumpinfo(skb, NETLINK_CB(cb->skb).portid,
399 				      cb->nlh->nlmsg_seq, NLM_F_MULTI,
400 				      pnetelem)) {
401 			--idx;
402 			break;
403 		}
404 	}
405 	cb->args[0] = idx;
406 	read_unlock(&smc_pnettable.lock);
407 	return skb->len;
408 }
409 
410 /* Remove and delete all pnetids from pnet table.
411  */
412 static int smc_pnet_flush(struct sk_buff *skb, struct genl_info *info)
413 {
414 	struct smc_pnetentry *pnetelem, *tmp_pe;
415 
416 	write_lock(&smc_pnettable.lock);
417 	list_for_each_entry_safe(pnetelem, tmp_pe, &smc_pnettable.pnetlist,
418 				 list) {
419 		list_del(&pnetelem->list);
420 		dev_put(pnetelem->ndev);
421 		kfree(pnetelem);
422 	}
423 	write_unlock(&smc_pnettable.lock);
424 	return 0;
425 }
426 
427 /* SMC_PNETID generic netlink operation definition */
428 static const struct genl_ops smc_pnet_ops[] = {
429 	{
430 		.cmd = SMC_PNETID_GET,
431 		.flags = GENL_ADMIN_PERM,
432 		.policy = smc_pnet_policy,
433 		.doit = smc_pnet_get,
434 		.dumpit = smc_pnet_dump,
435 		.start = smc_pnet_dump_start
436 	},
437 	{
438 		.cmd = SMC_PNETID_ADD,
439 		.flags = GENL_ADMIN_PERM,
440 		.policy = smc_pnet_policy,
441 		.doit = smc_pnet_add
442 	},
443 	{
444 		.cmd = SMC_PNETID_DEL,
445 		.flags = GENL_ADMIN_PERM,
446 		.policy = smc_pnet_policy,
447 		.doit = smc_pnet_del
448 	},
449 	{
450 		.cmd = SMC_PNETID_FLUSH,
451 		.flags = GENL_ADMIN_PERM,
452 		.policy = smc_pnet_policy,
453 		.doit = smc_pnet_flush
454 	}
455 };
456 
457 /* SMC_PNETID family definition */
458 static struct genl_family smc_pnet_nl_family = {
459 	.hdrsize = 0,
460 	.name = SMCR_GENL_FAMILY_NAME,
461 	.version = SMCR_GENL_FAMILY_VERSION,
462 	.maxattr = SMC_PNETID_MAX,
463 	.netnsok = true,
464 	.module = THIS_MODULE,
465 	.ops = smc_pnet_ops,
466 	.n_ops =  ARRAY_SIZE(smc_pnet_ops)
467 };
468 
469 static int smc_pnet_netdev_event(struct notifier_block *this,
470 				 unsigned long event, void *ptr)
471 {
472 	struct net_device *event_dev = netdev_notifier_info_to_dev(ptr);
473 
474 	switch (event) {
475 	case NETDEV_REBOOT:
476 	case NETDEV_UNREGISTER:
477 		smc_pnet_remove_by_ndev(event_dev);
478 	default:
479 		break;
480 	}
481 	return NOTIFY_DONE;
482 }
483 
484 static struct notifier_block smc_netdev_notifier = {
485 	.notifier_call = smc_pnet_netdev_event
486 };
487 
488 int __init smc_pnet_init(void)
489 {
490 	int rc;
491 
492 	rc = genl_register_family(&smc_pnet_nl_family);
493 	if (rc)
494 		return rc;
495 	rc = register_netdevice_notifier(&smc_netdev_notifier);
496 	if (rc)
497 		genl_unregister_family(&smc_pnet_nl_family);
498 	return rc;
499 }
500 
501 void smc_pnet_exit(void)
502 {
503 	smc_pnet_flush(NULL, NULL);
504 	unregister_netdevice_notifier(&smc_netdev_notifier);
505 	genl_unregister_family(&smc_pnet_nl_family);
506 }
507 
508 /* PNET table analysis for a given sock:
509  * determine ib_device and port belonging to used internal TCP socket
510  * ethernet interface.
511  */
512 void smc_pnet_find_roce_resource(struct sock *sk,
513 				 struct smc_ib_device **smcibdev, u8 *ibport)
514 {
515 	struct dst_entry *dst = sk_dst_get(sk);
516 	struct smc_pnetentry *pnetelem;
517 
518 	*smcibdev = NULL;
519 	*ibport = 0;
520 
521 	if (!dst)
522 		return;
523 	if (!dst->dev)
524 		goto out_rel;
525 	read_lock(&smc_pnettable.lock);
526 	list_for_each_entry(pnetelem, &smc_pnettable.pnetlist, list) {
527 		if (dst->dev == pnetelem->ndev) {
528 			if (smc_ib_port_active(pnetelem->smcibdev,
529 					       pnetelem->ib_port)) {
530 				*smcibdev = pnetelem->smcibdev;
531 				*ibport = pnetelem->ib_port;
532 			}
533 			break;
534 		}
535 	}
536 	read_unlock(&smc_pnettable.lock);
537 out_rel:
538 	dst_release(dst);
539 }
540