1 // SPDX-License-Identifier: GPL-2.0
3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
5 * Generic netlink support functions to configure an SMC-R PNET table
7 * Copyright IBM Corp. 2016
12 #include <linux/module.h>
13 #include <linux/list.h>
14 #include <linux/ctype.h>
15 #include <linux/mutex.h>
16 #include <net/netlink.h>
17 #include <net/genetlink.h>
19 #include <uapi/linux/if.h>
20 #include <uapi/linux/smc.h>
22 #include <rdma/ib_verbs.h>
24 #include <net/netns/generic.h>
25 #include "smc_netns.h"
32 static struct net_device *__pnet_find_base_ndev(struct net_device *ndev);
33 static struct net_device *pnet_find_base_ndev(struct net_device *ndev);
35 static const struct nla_policy smc_pnet_policy[SMC_PNETID_MAX + 1] = {
37 .type = NLA_NUL_STRING,
38 .len = SMC_MAX_PNETID_LEN
40 [SMC_PNETID_ETHNAME] = {
41 .type = NLA_NUL_STRING,
44 [SMC_PNETID_IBNAME] = {
45 .type = NLA_NUL_STRING,
46 .len = IB_DEVICE_NAME_MAX - 1
48 [SMC_PNETID_IBPORT] = { .type = NLA_U8 }
51 static struct genl_family smc_pnet_nl_family;
53 enum smc_pnet_nametype {
58 /* pnet entry stored in pnet table */
59 struct smc_pnetentry {
60 struct list_head list;
61 char pnet_name[SMC_MAX_PNETID_LEN + 1];
62 enum smc_pnet_nametype type;
65 char eth_name[IFNAMSIZ + 1];
66 struct net_device *ndev;
67 netdevice_tracker dev_tracker;
70 char ib_name[IB_DEVICE_NAME_MAX + 1];
76 /* Check if the pnetid is set */
77 bool smc_pnet_is_pnetid_set(u8 *pnetid)
79 if (pnetid[0] == 0 || pnetid[0] == _S)
84 /* Check if two given pnetids match */
85 static bool smc_pnet_match(u8 *pnetid1, u8 *pnetid2)
89 for (i = 0; i < SMC_MAX_PNETID_LEN; i++) {
90 if ((pnetid1[i] == 0 || pnetid1[i] == _S) &&
91 (pnetid2[i] == 0 || pnetid2[i] == _S))
93 if (pnetid1[i] != pnetid2[i])
99 /* Remove a pnetid from the pnet table.
101 static int smc_pnet_remove_by_pnetid(struct net *net, char *pnet_name)
103 struct smc_pnetentry *pnetelem, *tmp_pe;
104 struct smc_pnettable *pnettable;
105 struct smc_ib_device *ibdev;
106 struct smcd_dev *smcd;
111 /* get pnettable for namespace */
112 sn = net_generic(net, smc_net_id);
113 pnettable = &sn->pnettable;
115 /* remove table entry */
116 mutex_lock(&pnettable->lock);
117 list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist,
120 smc_pnet_match(pnetelem->pnet_name, pnet_name)) {
121 list_del(&pnetelem->list);
122 if (pnetelem->type == SMC_PNET_ETH && pnetelem->ndev) {
123 netdev_put(pnetelem->ndev,
124 &pnetelem->dev_tracker);
125 pr_warn_ratelimited("smc: net device %s "
126 "erased user defined "
129 pnetelem->pnet_name);
135 mutex_unlock(&pnettable->lock);
137 /* if this is not the initial namespace, stop here */
138 if (net != &init_net)
141 /* remove ib devices */
142 mutex_lock(&smc_ib_devices.mutex);
143 list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
144 for (ibport = 0; ibport < SMC_MAX_PORTS; ibport++) {
145 if (ibdev->pnetid_by_user[ibport] &&
147 smc_pnet_match(pnet_name,
148 ibdev->pnetid[ibport]))) {
149 pr_warn_ratelimited("smc: ib device %s ibport "
150 "%d erased user defined "
154 ibdev->pnetid[ibport]);
155 memset(ibdev->pnetid[ibport], 0,
157 ibdev->pnetid_by_user[ibport] = false;
162 mutex_unlock(&smc_ib_devices.mutex);
163 /* remove smcd devices */
164 mutex_lock(&smcd_dev_list.mutex);
165 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
166 if (smcd->pnetid_by_user &&
168 smc_pnet_match(pnet_name, smcd->pnetid))) {
169 pr_warn_ratelimited("smc: smcd device %s "
170 "erased user defined pnetid "
172 dev_name(smcd->ops->get_dev(smcd)),
174 memset(smcd->pnetid, 0, SMC_MAX_PNETID_LEN);
175 smcd->pnetid_by_user = false;
179 mutex_unlock(&smcd_dev_list.mutex);
183 /* Add the reference to a given network device to the pnet table.
185 static int smc_pnet_add_by_ndev(struct net_device *ndev)
187 struct smc_pnetentry *pnetelem, *tmp_pe;
188 struct smc_pnettable *pnettable;
189 struct net *net = dev_net(ndev);
193 /* get pnettable for namespace */
194 sn = net_generic(net, smc_net_id);
195 pnettable = &sn->pnettable;
197 mutex_lock(&pnettable->lock);
198 list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
199 if (pnetelem->type == SMC_PNET_ETH && !pnetelem->ndev &&
200 !strncmp(pnetelem->eth_name, ndev->name, IFNAMSIZ)) {
201 netdev_hold(ndev, &pnetelem->dev_tracker, GFP_ATOMIC);
202 pnetelem->ndev = ndev;
204 pr_warn_ratelimited("smc: adding net device %s with "
205 "user defined pnetid %.16s\n",
207 pnetelem->pnet_name);
211 mutex_unlock(&pnettable->lock);
215 /* Remove the reference to a given network device from the pnet table.
217 static int smc_pnet_remove_by_ndev(struct net_device *ndev)
219 struct smc_pnetentry *pnetelem, *tmp_pe;
220 struct smc_pnettable *pnettable;
221 struct net *net = dev_net(ndev);
225 /* get pnettable for namespace */
226 sn = net_generic(net, smc_net_id);
227 pnettable = &sn->pnettable;
229 mutex_lock(&pnettable->lock);
230 list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
231 if (pnetelem->type == SMC_PNET_ETH && pnetelem->ndev == ndev) {
232 netdev_put(pnetelem->ndev, &pnetelem->dev_tracker);
233 pnetelem->ndev = NULL;
235 pr_warn_ratelimited("smc: removing net device %s with "
236 "user defined pnetid %.16s\n",
238 pnetelem->pnet_name);
242 mutex_unlock(&pnettable->lock);
246 /* Apply pnetid to ib device when no pnetid is set.
248 static bool smc_pnet_apply_ib(struct smc_ib_device *ib_dev, u8 ib_port,
251 bool applied = false;
253 mutex_lock(&smc_ib_devices.mutex);
254 if (!smc_pnet_is_pnetid_set(ib_dev->pnetid[ib_port - 1])) {
255 memcpy(ib_dev->pnetid[ib_port - 1], pnet_name,
257 ib_dev->pnetid_by_user[ib_port - 1] = true;
260 mutex_unlock(&smc_ib_devices.mutex);
264 /* Apply pnetid to smcd device when no pnetid is set.
266 static bool smc_pnet_apply_smcd(struct smcd_dev *smcd_dev, char *pnet_name)
268 bool applied = false;
270 mutex_lock(&smcd_dev_list.mutex);
271 if (!smc_pnet_is_pnetid_set(smcd_dev->pnetid)) {
272 memcpy(smcd_dev->pnetid, pnet_name, SMC_MAX_PNETID_LEN);
273 smcd_dev->pnetid_by_user = true;
276 mutex_unlock(&smcd_dev_list.mutex);
280 /* The limit for pnetid is 16 characters.
281 * Valid characters should be (single-byte character set) a-z, A-Z, 0-9.
282 * Lower case letters are converted to upper case.
283 * Interior blanks should not be used.
285 static bool smc_pnetid_valid(const char *pnet_name, char *pnetid)
287 char *bf = skip_spaces(pnet_name);
288 size_t len = strlen(bf);
289 char *end = bf + len;
293 while (--end >= bf && isspace(*end))
295 if (end - bf >= SMC_MAX_PNETID_LEN)
300 *pnetid++ = islower(*bf) ? toupper(*bf) : *bf;
307 /* Find an infiniband device by a given name. The device might not exist. */
308 static struct smc_ib_device *smc_pnet_find_ib(char *ib_name)
310 struct smc_ib_device *ibdev;
312 mutex_lock(&smc_ib_devices.mutex);
313 list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
314 if (!strncmp(ibdev->ibdev->name, ib_name,
315 sizeof(ibdev->ibdev->name)) ||
316 (ibdev->ibdev->dev.parent &&
317 !strncmp(dev_name(ibdev->ibdev->dev.parent), ib_name,
318 IB_DEVICE_NAME_MAX - 1))) {
324 mutex_unlock(&smc_ib_devices.mutex);
328 /* Find an smcd device by a given name. The device might not exist. */
329 static struct smcd_dev *smc_pnet_find_smcd(char *smcd_name)
331 struct smcd_dev *smcd_dev;
333 mutex_lock(&smcd_dev_list.mutex);
334 list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
335 if (!strncmp(dev_name(smcd_dev->ops->get_dev(smcd_dev)),
336 smcd_name, IB_DEVICE_NAME_MAX - 1))
341 mutex_unlock(&smcd_dev_list.mutex);
345 static int smc_pnet_add_eth(struct smc_pnettable *pnettable, struct net *net,
346 char *eth_name, char *pnet_name)
348 struct smc_pnetentry *tmp_pe, *new_pe;
349 struct net_device *ndev, *base_ndev;
350 u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
354 /* check if (base) netdev already has a pnetid. If there is one, we do
355 * not want to add a pnet table entry
358 ndev = dev_get_by_name(net, eth_name); /* dev_hold() */
360 base_ndev = pnet_find_base_ndev(ndev);
361 if (!smc_pnetid_by_dev_port(base_ndev->dev.parent,
362 base_ndev->dev_port, ndev_pnetid))
366 /* add a new netdev entry to the pnet table if there isn't one */
368 new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
371 new_pe->type = SMC_PNET_ETH;
372 memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
373 strncpy(new_pe->eth_name, eth_name, IFNAMSIZ);
376 mutex_lock(&pnettable->lock);
377 list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
378 if (tmp_pe->type == SMC_PNET_ETH &&
379 !strncmp(tmp_pe->eth_name, eth_name, IFNAMSIZ)) {
387 netdev_tracker_alloc(ndev, &new_pe->dev_tracker,
390 list_add_tail(&new_pe->list, &pnettable->pnetlist);
391 mutex_unlock(&pnettable->lock);
393 mutex_unlock(&pnettable->lock);
398 pr_warn_ratelimited("smc: net device %s "
399 "applied user defined pnetid %.16s\n",
400 new_pe->eth_name, new_pe->pnet_name);
408 static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
409 u8 ib_port, char *pnet_name)
411 struct smc_pnetentry *tmp_pe, *new_pe;
412 struct smc_ib_device *ib_dev;
413 bool smcddev_applied = true;
414 bool ibdev_applied = true;
415 struct smcd_dev *smcd;
419 /* try to apply the pnetid to active devices */
420 ib_dev = smc_pnet_find_ib(ib_name);
422 ibdev_applied = smc_pnet_apply_ib(ib_dev, ib_port, pnet_name);
424 pr_warn_ratelimited("smc: ib device %s ibport %d "
425 "applied user defined pnetid "
426 "%.16s\n", ib_dev->ibdev->name,
428 ib_dev->pnetid[ib_port - 1]);
430 smcd = smc_pnet_find_smcd(ib_name);
432 smcddev_applied = smc_pnet_apply_smcd(smcd, pnet_name);
433 if (smcddev_applied) {
434 dev = smcd->ops->get_dev(smcd);
435 pr_warn_ratelimited("smc: smcd device %s "
436 "applied user defined pnetid "
437 "%.16s\n", dev_name(dev),
441 /* Apply fails when a device has a hardware-defined pnetid set, do not
442 * add a pnet table entry in that case.
444 if (!ibdev_applied || !smcddev_applied)
447 /* add a new ib entry to the pnet table if there isn't one */
448 new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
451 new_pe->type = SMC_PNET_IB;
452 memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
453 strncpy(new_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX);
454 new_pe->ib_port = ib_port;
457 mutex_lock(&pnettable->lock);
458 list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
459 if (tmp_pe->type == SMC_PNET_IB &&
460 !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
466 list_add_tail(&new_pe->list, &pnettable->pnetlist);
467 mutex_unlock(&pnettable->lock);
469 mutex_unlock(&pnettable->lock);
472 return (new_ibdev) ? 0 : -EEXIST;
475 /* Append a pnetid to the end of the pnet table if not already on this list.
477 static int smc_pnet_enter(struct net *net, struct nlattr *tb[])
479 char pnet_name[SMC_MAX_PNETID_LEN + 1];
480 struct smc_pnettable *pnettable;
481 bool new_netdev = false;
482 bool new_ibdev = false;
488 /* get pnettable for namespace */
489 sn = net_generic(net, smc_net_id);
490 pnettable = &sn->pnettable;
493 if (!tb[SMC_PNETID_NAME])
495 string = (char *)nla_data(tb[SMC_PNETID_NAME]);
496 if (!smc_pnetid_valid(string, pnet_name))
499 if (tb[SMC_PNETID_ETHNAME]) {
500 string = (char *)nla_data(tb[SMC_PNETID_ETHNAME]);
501 rc = smc_pnet_add_eth(pnettable, net, string, pnet_name);
504 else if (rc != -EEXIST)
508 /* if this is not the initial namespace, stop here */
509 if (net != &init_net)
510 return new_netdev ? 0 : -EEXIST;
513 if (tb[SMC_PNETID_IBNAME]) {
514 string = (char *)nla_data(tb[SMC_PNETID_IBNAME]);
515 string = strim(string);
516 if (tb[SMC_PNETID_IBPORT]) {
517 ibport = nla_get_u8(tb[SMC_PNETID_IBPORT]);
518 if (ibport < 1 || ibport > SMC_MAX_PORTS)
521 rc = smc_pnet_add_ib(pnettable, string, ibport, pnet_name);
524 else if (rc != -EEXIST)
527 return (new_netdev || new_ibdev) ? 0 : -EEXIST;
533 /* Convert an smc_pnetentry to a netlink attribute sequence */
534 static int smc_pnet_set_nla(struct sk_buff *msg,
535 struct smc_pnetentry *pnetelem)
537 if (nla_put_string(msg, SMC_PNETID_NAME, pnetelem->pnet_name))
539 if (pnetelem->type == SMC_PNET_ETH) {
540 if (nla_put_string(msg, SMC_PNETID_ETHNAME,
544 if (nla_put_string(msg, SMC_PNETID_ETHNAME, "n/a"))
547 if (pnetelem->type == SMC_PNET_IB) {
548 if (nla_put_string(msg, SMC_PNETID_IBNAME, pnetelem->ib_name) ||
549 nla_put_u8(msg, SMC_PNETID_IBPORT, pnetelem->ib_port))
552 if (nla_put_string(msg, SMC_PNETID_IBNAME, "n/a") ||
553 nla_put_u8(msg, SMC_PNETID_IBPORT, 0xff))
560 static int smc_pnet_add(struct sk_buff *skb, struct genl_info *info)
562 struct net *net = genl_info_net(info);
564 return smc_pnet_enter(net, info->attrs);
567 static int smc_pnet_del(struct sk_buff *skb, struct genl_info *info)
569 struct net *net = genl_info_net(info);
571 if (!info->attrs[SMC_PNETID_NAME])
573 return smc_pnet_remove_by_pnetid(net,
574 (char *)nla_data(info->attrs[SMC_PNETID_NAME]));
577 static int smc_pnet_dump_start(struct netlink_callback *cb)
583 static int smc_pnet_dumpinfo(struct sk_buff *skb,
584 u32 portid, u32 seq, u32 flags,
585 struct smc_pnetentry *pnetelem)
589 hdr = genlmsg_put(skb, portid, seq, &smc_pnet_nl_family,
590 flags, SMC_PNETID_GET);
593 if (smc_pnet_set_nla(skb, pnetelem) < 0) {
594 genlmsg_cancel(skb, hdr);
597 genlmsg_end(skb, hdr);
601 static int _smc_pnet_dump(struct net *net, struct sk_buff *skb, u32 portid,
602 u32 seq, u8 *pnetid, int start_idx)
604 struct smc_pnettable *pnettable;
605 struct smc_pnetentry *pnetelem;
609 /* get pnettable for namespace */
610 sn = net_generic(net, smc_net_id);
611 pnettable = &sn->pnettable;
613 /* dump pnettable entries */
614 mutex_lock(&pnettable->lock);
615 list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
616 if (pnetid && !smc_pnet_match(pnetelem->pnet_name, pnetid))
618 if (idx++ < start_idx)
620 /* if this is not the initial namespace, dump only netdev */
621 if (net != &init_net && pnetelem->type != SMC_PNET_ETH)
623 if (smc_pnet_dumpinfo(skb, portid, seq, NLM_F_MULTI,
629 mutex_unlock(&pnettable->lock);
633 static int smc_pnet_dump(struct sk_buff *skb, struct netlink_callback *cb)
635 struct net *net = sock_net(skb->sk);
638 idx = _smc_pnet_dump(net, skb, NETLINK_CB(cb->skb).portid,
639 cb->nlh->nlmsg_seq, NULL, cb->args[0]);
645 /* Retrieve one PNETID entry */
646 static int smc_pnet_get(struct sk_buff *skb, struct genl_info *info)
648 struct net *net = genl_info_net(info);
652 if (!info->attrs[SMC_PNETID_NAME])
655 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
659 _smc_pnet_dump(net, msg, info->snd_portid, info->snd_seq,
660 nla_data(info->attrs[SMC_PNETID_NAME]), 0);
662 /* finish multi part message and send it */
663 hdr = nlmsg_put(msg, info->snd_portid, info->snd_seq, NLMSG_DONE, 0,
669 return genlmsg_reply(msg, info);
672 /* Remove and delete all pnetids from pnet table.
674 static int smc_pnet_flush(struct sk_buff *skb, struct genl_info *info)
676 struct net *net = genl_info_net(info);
678 smc_pnet_remove_by_pnetid(net, NULL);
682 /* SMC_PNETID generic netlink operation definition */
683 static const struct genl_ops smc_pnet_ops[] = {
685 .cmd = SMC_PNETID_GET,
686 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
687 /* can be retrieved by unprivileged users */
688 .doit = smc_pnet_get,
689 .dumpit = smc_pnet_dump,
690 .start = smc_pnet_dump_start
693 .cmd = SMC_PNETID_ADD,
694 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
695 .flags = GENL_ADMIN_PERM,
699 .cmd = SMC_PNETID_DEL,
700 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
701 .flags = GENL_ADMIN_PERM,
705 .cmd = SMC_PNETID_FLUSH,
706 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
707 .flags = GENL_ADMIN_PERM,
708 .doit = smc_pnet_flush
712 /* SMC_PNETID family definition */
713 static struct genl_family smc_pnet_nl_family __ro_after_init = {
715 .name = SMCR_GENL_FAMILY_NAME,
716 .version = SMCR_GENL_FAMILY_VERSION,
717 .maxattr = SMC_PNETID_MAX,
718 .policy = smc_pnet_policy,
720 .module = THIS_MODULE,
722 .n_ops = ARRAY_SIZE(smc_pnet_ops),
723 .resv_start_op = SMC_PNETID_FLUSH + 1,
726 bool smc_pnet_is_ndev_pnetid(struct net *net, u8 *pnetid)
728 struct smc_net *sn = net_generic(net, smc_net_id);
729 struct smc_pnetids_ndev_entry *pe;
732 read_lock(&sn->pnetids_ndev.lock);
733 list_for_each_entry(pe, &sn->pnetids_ndev.list, list) {
734 if (smc_pnet_match(pnetid, pe->pnetid)) {
741 read_unlock(&sn->pnetids_ndev.lock);
745 static int smc_pnet_add_pnetid(struct net *net, u8 *pnetid)
747 struct smc_net *sn = net_generic(net, smc_net_id);
748 struct smc_pnetids_ndev_entry *pe, *pi;
750 pe = kzalloc(sizeof(*pe), GFP_KERNEL);
754 write_lock(&sn->pnetids_ndev.lock);
755 list_for_each_entry(pi, &sn->pnetids_ndev.list, list) {
756 if (smc_pnet_match(pnetid, pe->pnetid)) {
757 refcount_inc(&pi->refcnt);
762 refcount_set(&pe->refcnt, 1);
763 memcpy(pe->pnetid, pnetid, SMC_MAX_PNETID_LEN);
764 list_add_tail(&pe->list, &sn->pnetids_ndev.list);
767 write_unlock(&sn->pnetids_ndev.lock);
771 static void smc_pnet_remove_pnetid(struct net *net, u8 *pnetid)
773 struct smc_net *sn = net_generic(net, smc_net_id);
774 struct smc_pnetids_ndev_entry *pe, *pe2;
776 write_lock(&sn->pnetids_ndev.lock);
777 list_for_each_entry_safe(pe, pe2, &sn->pnetids_ndev.list, list) {
778 if (smc_pnet_match(pnetid, pe->pnetid)) {
779 if (refcount_dec_and_test(&pe->refcnt)) {
786 write_unlock(&sn->pnetids_ndev.lock);
789 static void smc_pnet_add_base_pnetid(struct net *net, struct net_device *dev,
792 struct net_device *base_dev;
794 base_dev = __pnet_find_base_ndev(dev);
795 if (base_dev->flags & IFF_UP &&
796 !smc_pnetid_by_dev_port(base_dev->dev.parent, base_dev->dev_port,
798 /* add to PNETIDs list */
799 smc_pnet_add_pnetid(net, ndev_pnetid);
803 /* create initial list of netdevice pnetids */
804 static void smc_pnet_create_pnetids_list(struct net *net)
806 u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
807 struct net_device *dev;
809 /* Newly created netns do not have devices.
810 * Do not even acquire rtnl.
812 if (list_empty(&net->dev_base_head))
815 /* Note: This might not be needed, because smc_pnet_netdev_event()
816 * is also calling smc_pnet_add_base_pnetid() when handling
820 for_each_netdev(net, dev)
821 smc_pnet_add_base_pnetid(net, dev, ndev_pnetid);
825 /* clean up list of netdevice pnetids */
826 static void smc_pnet_destroy_pnetids_list(struct net *net)
828 struct smc_net *sn = net_generic(net, smc_net_id);
829 struct smc_pnetids_ndev_entry *pe, *temp_pe;
831 write_lock(&sn->pnetids_ndev.lock);
832 list_for_each_entry_safe(pe, temp_pe, &sn->pnetids_ndev.list, list) {
836 write_unlock(&sn->pnetids_ndev.lock);
839 static int smc_pnet_netdev_event(struct notifier_block *this,
840 unsigned long event, void *ptr)
842 struct net_device *event_dev = netdev_notifier_info_to_dev(ptr);
843 struct net *net = dev_net(event_dev);
844 u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
848 case NETDEV_UNREGISTER:
849 smc_pnet_remove_by_ndev(event_dev);
850 smc_ib_ndev_change(event_dev, event);
852 case NETDEV_REGISTER:
853 smc_pnet_add_by_ndev(event_dev);
854 smc_ib_ndev_change(event_dev, event);
857 smc_pnet_add_base_pnetid(net, event_dev, ndev_pnetid);
860 event_dev = __pnet_find_base_ndev(event_dev);
861 if (!smc_pnetid_by_dev_port(event_dev->dev.parent,
862 event_dev->dev_port, ndev_pnetid)) {
863 /* remove from PNETIDs list */
864 smc_pnet_remove_pnetid(net, ndev_pnetid);
872 static struct notifier_block smc_netdev_notifier = {
873 .notifier_call = smc_pnet_netdev_event
876 /* init network namespace */
877 int smc_pnet_net_init(struct net *net)
879 struct smc_net *sn = net_generic(net, smc_net_id);
880 struct smc_pnettable *pnettable = &sn->pnettable;
881 struct smc_pnetids_ndev *pnetids_ndev = &sn->pnetids_ndev;
883 INIT_LIST_HEAD(&pnettable->pnetlist);
884 mutex_init(&pnettable->lock);
885 INIT_LIST_HEAD(&pnetids_ndev->list);
886 rwlock_init(&pnetids_ndev->lock);
888 smc_pnet_create_pnetids_list(net);
890 /* disable handshake limitation by default */
891 net->smc.limit_smc_hs = 0;
896 int __init smc_pnet_init(void)
900 rc = genl_register_family(&smc_pnet_nl_family);
903 rc = register_netdevice_notifier(&smc_netdev_notifier);
905 genl_unregister_family(&smc_pnet_nl_family);
910 /* exit network namespace */
911 void smc_pnet_net_exit(struct net *net)
913 /* flush pnet table */
914 smc_pnet_remove_by_pnetid(net, NULL);
915 smc_pnet_destroy_pnetids_list(net);
918 void smc_pnet_exit(void)
920 unregister_netdevice_notifier(&smc_netdev_notifier);
921 genl_unregister_family(&smc_pnet_nl_family);
924 static struct net_device *__pnet_find_base_ndev(struct net_device *ndev)
929 nest_lvl = ndev->lower_level;
930 for (i = 0; i < nest_lvl; i++) {
931 struct list_head *lower = &ndev->adj_list.lower;
933 if (list_empty(lower))
936 ndev = netdev_lower_get_next(ndev, &lower);
941 /* Determine one base device for stacked net devices.
942 * If the lower device level contains more than one devices
943 * (for instance with bonding slaves), just the first device
944 * is used to reach a base device.
946 static struct net_device *pnet_find_base_ndev(struct net_device *ndev)
949 ndev = __pnet_find_base_ndev(ndev);
954 static int smc_pnet_find_ndev_pnetid_by_table(struct net_device *ndev,
957 struct smc_pnettable *pnettable;
958 struct net *net = dev_net(ndev);
959 struct smc_pnetentry *pnetelem;
963 /* get pnettable for namespace */
964 sn = net_generic(net, smc_net_id);
965 pnettable = &sn->pnettable;
967 mutex_lock(&pnettable->lock);
968 list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
969 if (pnetelem->type == SMC_PNET_ETH && ndev == pnetelem->ndev) {
970 /* get pnetid of netdev device */
971 memcpy(pnetid, pnetelem->pnet_name, SMC_MAX_PNETID_LEN);
976 mutex_unlock(&pnettable->lock);
980 static int smc_pnet_determine_gid(struct smc_ib_device *ibdev, int i,
981 struct smc_init_info *ini)
983 if (!ini->check_smcrv2 &&
984 !smc_ib_determine_gid(ibdev, i, ini->vlan_id, ini->ib_gid, NULL,
990 if (ini->check_smcrv2 &&
991 !smc_ib_determine_gid(ibdev, i, ini->vlan_id, ini->smcrv2.ib_gid_v2,
992 NULL, &ini->smcrv2)) {
993 ini->smcrv2.ib_dev_v2 = ibdev;
994 ini->smcrv2.ib_port_v2 = i;
1000 /* find a roce device for the given pnetid */
1001 static void _smc_pnet_find_roce_by_pnetid(u8 *pnet_id,
1002 struct smc_init_info *ini,
1003 struct smc_ib_device *known_dev,
1006 struct smc_ib_device *ibdev;
1009 mutex_lock(&smc_ib_devices.mutex);
1010 list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
1011 if (ibdev == known_dev ||
1012 !rdma_dev_access_netns(ibdev->ibdev, net))
1014 for (i = 1; i <= SMC_MAX_PORTS; i++) {
1015 if (!rdma_is_port_valid(ibdev->ibdev, i))
1017 if (smc_pnet_match(ibdev->pnetid[i - 1], pnet_id) &&
1018 smc_ib_port_active(ibdev, i) &&
1019 !test_bit(i - 1, ibdev->ports_going_away)) {
1020 if (!smc_pnet_determine_gid(ibdev, i, ini))
1026 mutex_unlock(&smc_ib_devices.mutex);
1029 /* find alternate roce device with same pnet_id, vlan_id and net namespace */
1030 void smc_pnet_find_alt_roce(struct smc_link_group *lgr,
1031 struct smc_init_info *ini,
1032 struct smc_ib_device *known_dev)
1034 struct net *net = lgr->net;
1036 _smc_pnet_find_roce_by_pnetid(lgr->pnet_id, ini, known_dev, net);
1039 /* if handshake network device belongs to a roce device, return its
1040 * IB device and port
1042 static void smc_pnet_find_rdma_dev(struct net_device *netdev,
1043 struct smc_init_info *ini)
1045 struct net *net = dev_net(netdev);
1046 struct smc_ib_device *ibdev;
1048 mutex_lock(&smc_ib_devices.mutex);
1049 list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
1050 struct net_device *ndev;
1053 /* check rdma net namespace */
1054 if (!rdma_dev_access_netns(ibdev->ibdev, net))
1057 for (i = 1; i <= SMC_MAX_PORTS; i++) {
1058 if (!rdma_is_port_valid(ibdev->ibdev, i))
1060 if (!ibdev->ibdev->ops.get_netdev)
1062 ndev = ibdev->ibdev->ops.get_netdev(ibdev->ibdev, i);
1066 if (netdev == ndev &&
1067 smc_ib_port_active(ibdev, i) &&
1068 !test_bit(i - 1, ibdev->ports_going_away)) {
1069 if (!smc_pnet_determine_gid(ibdev, i, ini))
1074 mutex_unlock(&smc_ib_devices.mutex);
1077 /* Determine the corresponding IB device port based on the hardware PNETID.
1078 * Searching stops at the first matching active IB device port with vlan_id
1080 * If nothing found, check pnetid table.
1081 * If nothing found, try to use handshake device
1083 static void smc_pnet_find_roce_by_pnetid(struct net_device *ndev,
1084 struct smc_init_info *ini)
1086 u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
1089 ndev = pnet_find_base_ndev(ndev);
1090 net = dev_net(ndev);
1091 if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
1093 smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid)) {
1094 smc_pnet_find_rdma_dev(ndev, ini);
1095 return; /* pnetid could not be determined */
1097 _smc_pnet_find_roce_by_pnetid(ndev_pnetid, ini, NULL, net);
1100 static void smc_pnet_find_ism_by_pnetid(struct net_device *ndev,
1101 struct smc_init_info *ini)
1103 u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
1104 struct smcd_dev *ismdev;
1106 ndev = pnet_find_base_ndev(ndev);
1107 if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
1109 smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid))
1110 return; /* pnetid could not be determined */
1112 mutex_lock(&smcd_dev_list.mutex);
1113 list_for_each_entry(ismdev, &smcd_dev_list.list, list) {
1114 if (smc_pnet_match(ismdev->pnetid, ndev_pnetid) &&
1115 !ismdev->going_away &&
1116 (!ini->ism_peer_gid[0].gid ||
1117 !smc_ism_cantalk(&ini->ism_peer_gid[0], ini->vlan_id,
1119 ini->ism_dev[0] = ismdev;
1123 mutex_unlock(&smcd_dev_list.mutex);
1126 /* PNET table analysis for a given sock:
1127 * determine ib_device and port belonging to used internal TCP socket
1128 * ethernet interface.
1130 void smc_pnet_find_roce_resource(struct sock *sk, struct smc_init_info *ini)
1132 struct dst_entry *dst = sk_dst_get(sk);
1139 smc_pnet_find_roce_by_pnetid(dst->dev, ini);
1147 void smc_pnet_find_ism_resource(struct sock *sk, struct smc_init_info *ini)
1149 struct dst_entry *dst = sk_dst_get(sk);
1151 ini->ism_dev[0] = NULL;
1157 smc_pnet_find_ism_by_pnetid(dst->dev, ini);
1165 /* Lookup and apply a pnet table entry to the given ib device.
1167 int smc_pnetid_by_table_ib(struct smc_ib_device *smcibdev, u8 ib_port)
1169 char *ib_name = smcibdev->ibdev->name;
1170 struct smc_pnettable *pnettable;
1171 struct smc_pnetentry *tmp_pe;
1175 /* get pnettable for init namespace */
1176 sn = net_generic(&init_net, smc_net_id);
1177 pnettable = &sn->pnettable;
1179 mutex_lock(&pnettable->lock);
1180 list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
1181 if (tmp_pe->type == SMC_PNET_IB &&
1182 !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX) &&
1183 tmp_pe->ib_port == ib_port) {
1184 smc_pnet_apply_ib(smcibdev, ib_port, tmp_pe->pnet_name);
1189 mutex_unlock(&pnettable->lock);
1194 /* Lookup and apply a pnet table entry to the given smcd device.
1196 int smc_pnetid_by_table_smcd(struct smcd_dev *smcddev)
1198 const char *ib_name = dev_name(smcddev->ops->get_dev(smcddev));
1199 struct smc_pnettable *pnettable;
1200 struct smc_pnetentry *tmp_pe;
1204 /* get pnettable for init namespace */
1205 sn = net_generic(&init_net, smc_net_id);
1206 pnettable = &sn->pnettable;
1208 mutex_lock(&pnettable->lock);
1209 list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
1210 if (tmp_pe->type == SMC_PNET_IB &&
1211 !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
1212 smc_pnet_apply_smcd(smcddev, tmp_pe->pnet_name);
1217 mutex_unlock(&pnettable->lock);