]>
Commit | Line | Data |
---|---|---|
6812baab TR |
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 | goto found; | |
186 | } | |
187 | list_add_tail(&new_pnetelem->list, &smc_pnettable.pnetlist); | |
188 | rc = 0; | |
189 | found: | |
190 | write_unlock(&smc_pnettable.lock); | |
191 | return rc; | |
192 | } | |
193 | ||
194 | /* The limit for pnetid is 16 characters. | |
195 | * Valid characters should be (single-byte character set) a-z, A-Z, 0-9. | |
196 | * Lower case letters are converted to upper case. | |
197 | * Interior blanks should not be used. | |
198 | */ | |
199 | static bool smc_pnetid_valid(const char *pnet_name, char *pnetid) | |
200 | { | |
201 | char *bf = skip_spaces(pnet_name); | |
202 | size_t len = strlen(bf); | |
203 | char *end = bf + len; | |
204 | ||
205 | if (!len) | |
206 | return false; | |
207 | while (--end >= bf && isspace(*end)) | |
208 | ; | |
209 | if (end - bf >= SMC_MAX_PNET_ID_LEN) | |
210 | return false; | |
211 | while (bf <= end) { | |
212 | if (!isalnum(*bf)) | |
213 | return false; | |
214 | *pnetid++ = islower(*bf) ? toupper(*bf) : *bf; | |
215 | bf++; | |
216 | } | |
217 | *pnetid = '\0'; | |
218 | return true; | |
219 | } | |
220 | ||
221 | /* Find an infiniband device by a given name. The device might not exist. */ | |
249633a4 | 222 | static struct smc_ib_device *smc_pnet_find_ib(char *ib_name) |
6812baab TR |
223 | { |
224 | struct smc_ib_device *ibdev; | |
225 | ||
226 | spin_lock(&smc_ib_devices.lock); | |
227 | list_for_each_entry(ibdev, &smc_ib_devices.list, list) { | |
228 | if (!strncmp(ibdev->ibdev->name, ib_name, | |
229 | sizeof(ibdev->ibdev->name))) { | |
230 | goto out; | |
231 | } | |
232 | } | |
233 | ibdev = NULL; | |
234 | out: | |
235 | spin_unlock(&smc_ib_devices.lock); | |
236 | return ibdev; | |
237 | } | |
238 | ||
239 | /* Parse the supplied netlink attributes and fill a pnetentry structure. | |
240 | * For ethernet and infiniband device names verify that the devices exist. | |
241 | */ | |
242 | static int smc_pnet_fill_entry(struct net *net, struct smc_pnetentry *pnetelem, | |
243 | struct nlattr *tb[]) | |
244 | { | |
245 | char *string, *ibname = NULL; | |
246 | int rc = 0; | |
247 | ||
248 | memset(pnetelem, 0, sizeof(*pnetelem)); | |
249 | INIT_LIST_HEAD(&pnetelem->list); | |
250 | if (tb[SMC_PNETID_NAME]) { | |
251 | string = (char *)nla_data(tb[SMC_PNETID_NAME]); | |
252 | if (!smc_pnetid_valid(string, pnetelem->pnet_name)) { | |
253 | rc = -EINVAL; | |
254 | goto error; | |
255 | } | |
256 | } | |
257 | if (tb[SMC_PNETID_ETHNAME]) { | |
258 | string = (char *)nla_data(tb[SMC_PNETID_ETHNAME]); | |
259 | pnetelem->ndev = dev_get_by_name(net, string); | |
260 | if (!pnetelem->ndev) | |
261 | return -ENOENT; | |
262 | } | |
263 | if (tb[SMC_PNETID_IBNAME]) { | |
264 | ibname = (char *)nla_data(tb[SMC_PNETID_IBNAME]); | |
265 | ibname = strim(ibname); | |
266 | pnetelem->smcibdev = smc_pnet_find_ib(ibname); | |
267 | if (!pnetelem->smcibdev) { | |
268 | rc = -ENOENT; | |
269 | goto error; | |
270 | } | |
271 | } | |
272 | if (tb[SMC_PNETID_IBPORT]) { | |
273 | pnetelem->ib_port = nla_get_u8(tb[SMC_PNETID_IBPORT]); | |
274 | if (pnetelem->ib_port > SMC_MAX_PORTS) { | |
275 | rc = -EINVAL; | |
276 | goto error; | |
277 | } | |
278 | } | |
279 | return 0; | |
280 | ||
281 | error: | |
282 | if (pnetelem->ndev) | |
283 | dev_put(pnetelem->ndev); | |
284 | return rc; | |
285 | } | |
286 | ||
287 | /* Convert an smc_pnetentry to a netlink attribute sequence */ | |
288 | static int smc_pnet_set_nla(struct sk_buff *msg, struct smc_pnetentry *pnetelem) | |
289 | { | |
290 | if (nla_put_string(msg, SMC_PNETID_NAME, pnetelem->pnet_name) || | |
291 | nla_put_string(msg, SMC_PNETID_ETHNAME, pnetelem->ndev->name) || | |
292 | nla_put_string(msg, SMC_PNETID_IBNAME, | |
293 | pnetelem->smcibdev->ibdev->name) || | |
294 | nla_put_u8(msg, SMC_PNETID_IBPORT, pnetelem->ib_port)) | |
295 | return -1; | |
296 | return 0; | |
297 | } | |
298 | ||
299 | /* Retrieve one PNETID entry */ | |
300 | static int smc_pnet_get(struct sk_buff *skb, struct genl_info *info) | |
301 | { | |
302 | struct smc_pnetentry *pnetelem; | |
303 | struct sk_buff *msg; | |
304 | void *hdr; | |
305 | int rc; | |
306 | ||
307 | pnetelem = smc_pnet_find_pnetid( | |
308 | (char *)nla_data(info->attrs[SMC_PNETID_NAME])); | |
309 | if (!pnetelem) | |
310 | return -ENOENT; | |
311 | msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL); | |
312 | if (!msg) | |
313 | return -ENOMEM; | |
314 | ||
315 | hdr = genlmsg_put(msg, info->snd_portid, info->snd_seq, | |
316 | &smc_pnet_nl_family, 0, SMC_PNETID_GET); | |
317 | if (!hdr) { | |
318 | rc = -EMSGSIZE; | |
319 | goto err_out; | |
320 | } | |
321 | ||
322 | if (smc_pnet_set_nla(msg, pnetelem)) { | |
323 | rc = -ENOBUFS; | |
324 | goto err_out; | |
325 | } | |
326 | ||
327 | genlmsg_end(msg, hdr); | |
328 | return genlmsg_reply(msg, info); | |
329 | ||
330 | err_out: | |
331 | nlmsg_free(msg); | |
332 | return rc; | |
333 | } | |
334 | ||
335 | static int smc_pnet_add(struct sk_buff *skb, struct genl_info *info) | |
336 | { | |
337 | struct net *net = genl_info_net(info); | |
338 | struct smc_pnetentry *pnetelem; | |
339 | int rc; | |
340 | ||
341 | pnetelem = kzalloc(sizeof(*pnetelem), GFP_KERNEL); | |
342 | if (!pnetelem) | |
343 | return -ENOMEM; | |
344 | rc = smc_pnet_fill_entry(net, pnetelem, info->attrs); | |
345 | if (!rc) | |
346 | rc = smc_pnet_enter(pnetelem); | |
347 | if (rc) { | |
348 | kfree(pnetelem); | |
349 | return rc; | |
350 | } | |
351 | rc = smc_ib_remember_port_attr(pnetelem->smcibdev, pnetelem->ib_port); | |
352 | if (rc) | |
353 | smc_pnet_remove_by_pnetid(pnetelem->pnet_name); | |
354 | return rc; | |
355 | } | |
356 | ||
357 | static int smc_pnet_del(struct sk_buff *skb, struct genl_info *info) | |
358 | { | |
359 | return smc_pnet_remove_by_pnetid( | |
360 | (char *)nla_data(info->attrs[SMC_PNETID_NAME])); | |
361 | } | |
362 | ||
363 | static int smc_pnet_dump_start(struct netlink_callback *cb) | |
364 | { | |
365 | cb->args[0] = 0; | |
366 | return 0; | |
367 | } | |
368 | ||
369 | static int smc_pnet_dumpinfo(struct sk_buff *skb, | |
370 | u32 portid, u32 seq, u32 flags, | |
371 | struct smc_pnetentry *pnetelem) | |
372 | { | |
373 | void *hdr; | |
374 | ||
375 | hdr = genlmsg_put(skb, portid, seq, &smc_pnet_nl_family, | |
376 | flags, SMC_PNETID_GET); | |
377 | if (!hdr) | |
378 | return -ENOMEM; | |
379 | if (smc_pnet_set_nla(skb, pnetelem) < 0) { | |
380 | genlmsg_cancel(skb, hdr); | |
381 | return -EMSGSIZE; | |
382 | } | |
383 | genlmsg_end(skb, hdr); | |
384 | return 0; | |
385 | } | |
386 | ||
387 | static int smc_pnet_dump(struct sk_buff *skb, struct netlink_callback *cb) | |
388 | { | |
389 | struct smc_pnetentry *pnetelem; | |
390 | int idx = 0; | |
391 | ||
392 | read_lock(&smc_pnettable.lock); | |
393 | list_for_each_entry(pnetelem, &smc_pnettable.pnetlist, list) { | |
394 | if (idx++ < cb->args[0]) | |
395 | continue; | |
396 | if (smc_pnet_dumpinfo(skb, NETLINK_CB(cb->skb).portid, | |
397 | cb->nlh->nlmsg_seq, NLM_F_MULTI, | |
398 | pnetelem)) { | |
399 | --idx; | |
400 | break; | |
401 | } | |
402 | } | |
403 | cb->args[0] = idx; | |
404 | read_unlock(&smc_pnettable.lock); | |
405 | return skb->len; | |
406 | } | |
407 | ||
408 | /* Remove and delete all pnetids from pnet table. | |
409 | */ | |
410 | static int smc_pnet_flush(struct sk_buff *skb, struct genl_info *info) | |
411 | { | |
412 | struct smc_pnetentry *pnetelem, *tmp_pe; | |
413 | ||
414 | write_lock(&smc_pnettable.lock); | |
415 | list_for_each_entry_safe(pnetelem, tmp_pe, &smc_pnettable.pnetlist, | |
416 | list) { | |
417 | list_del(&pnetelem->list); | |
418 | dev_put(pnetelem->ndev); | |
419 | kfree(pnetelem); | |
420 | } | |
421 | write_unlock(&smc_pnettable.lock); | |
422 | return 0; | |
423 | } | |
424 | ||
425 | /* SMC_PNETID generic netlink operation definition */ | |
426 | static const struct genl_ops smc_pnet_ops[] = { | |
427 | { | |
428 | .cmd = SMC_PNETID_GET, | |
429 | .flags = GENL_ADMIN_PERM, | |
430 | .policy = smc_pnet_policy, | |
431 | .doit = smc_pnet_get, | |
432 | .dumpit = smc_pnet_dump, | |
433 | .start = smc_pnet_dump_start | |
434 | }, | |
435 | { | |
436 | .cmd = SMC_PNETID_ADD, | |
437 | .flags = GENL_ADMIN_PERM, | |
438 | .policy = smc_pnet_policy, | |
439 | .doit = smc_pnet_add | |
440 | }, | |
441 | { | |
442 | .cmd = SMC_PNETID_DEL, | |
443 | .flags = GENL_ADMIN_PERM, | |
444 | .policy = smc_pnet_policy, | |
445 | .doit = smc_pnet_del | |
446 | }, | |
447 | { | |
448 | .cmd = SMC_PNETID_FLUSH, | |
449 | .flags = GENL_ADMIN_PERM, | |
450 | .policy = smc_pnet_policy, | |
451 | .doit = smc_pnet_flush | |
452 | } | |
453 | }; | |
454 | ||
455 | /* SMC_PNETID family definition */ | |
456 | static struct genl_family smc_pnet_nl_family = { | |
457 | .hdrsize = 0, | |
458 | .name = SMCR_GENL_FAMILY_NAME, | |
459 | .version = SMCR_GENL_FAMILY_VERSION, | |
460 | .maxattr = SMC_PNETID_MAX, | |
461 | .netnsok = true, | |
462 | .module = THIS_MODULE, | |
463 | .ops = smc_pnet_ops, | |
464 | .n_ops = ARRAY_SIZE(smc_pnet_ops) | |
465 | }; | |
466 | ||
467 | static int smc_pnet_netdev_event(struct notifier_block *this, | |
468 | unsigned long event, void *ptr) | |
469 | { | |
470 | struct net_device *event_dev = netdev_notifier_info_to_dev(ptr); | |
471 | ||
472 | switch (event) { | |
473 | case NETDEV_REBOOT: | |
474 | case NETDEV_UNREGISTER: | |
475 | smc_pnet_remove_by_ndev(event_dev); | |
476 | default: | |
477 | break; | |
478 | } | |
479 | return NOTIFY_DONE; | |
480 | } | |
481 | ||
482 | static struct notifier_block smc_netdev_notifier = { | |
483 | .notifier_call = smc_pnet_netdev_event | |
484 | }; | |
485 | ||
486 | int __init smc_pnet_init(void) | |
487 | { | |
488 | int rc; | |
489 | ||
490 | rc = genl_register_family(&smc_pnet_nl_family); | |
491 | if (rc) | |
492 | return rc; | |
493 | rc = register_netdevice_notifier(&smc_netdev_notifier); | |
494 | if (rc) | |
495 | genl_unregister_family(&smc_pnet_nl_family); | |
496 | return rc; | |
497 | } | |
498 | ||
499 | void smc_pnet_exit(void) | |
500 | { | |
501 | smc_pnet_flush(NULL, NULL); | |
502 | unregister_netdevice_notifier(&smc_netdev_notifier); | |
503 | genl_unregister_family(&smc_pnet_nl_family); | |
504 | } | |
505 | ||
506 | /* PNET table analysis for a given sock: | |
507 | * determine ib_device and port belonging to used internal TCP socket | |
508 | * ethernet interface. | |
509 | */ | |
510 | void smc_pnet_find_roce_resource(struct sock *sk, | |
511 | struct smc_ib_device **smcibdev, u8 *ibport) | |
512 | { | |
513 | struct dst_entry *dst = sk_dst_get(sk); | |
514 | struct smc_pnetentry *pnetelem; | |
515 | ||
516 | *smcibdev = NULL; | |
517 | *ibport = 0; | |
518 | ||
519 | if (!dst) | |
520 | return; | |
521 | if (!dst->dev) | |
522 | goto out_rel; | |
523 | read_lock(&smc_pnettable.lock); | |
524 | list_for_each_entry(pnetelem, &smc_pnettable.pnetlist, list) { | |
525 | if (dst->dev == pnetelem->ndev) { | |
5da7e4d3 UB |
526 | if (smc_ib_port_active(pnetelem->smcibdev, |
527 | pnetelem->ib_port)) { | |
528 | *smcibdev = pnetelem->smcibdev; | |
529 | *ibport = pnetelem->ib_port; | |
530 | } | |
6812baab TR |
531 | break; |
532 | } | |
533 | } | |
534 | read_unlock(&smc_pnettable.lock); | |
535 | out_rel: | |
536 | dst_release(dst); | |
537 | } |