net/smc/smc_pnet.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  *  Generic netlink support functions to configure an SMC-R PNET table
   6  *
   7  *  Copyright IBM Corp. 2016
   8  *
   9  *  Author(s):  Thomas Richter <tmricht@linux.vnet.ibm.com>
  10  */
  11
  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>
  18
  19 #include <uapi/linux/if.h>
  20 #include <uapi/linux/smc.h>
  21
  22 #include <rdma/ib_verbs.h>
  23
  24 #include <net/netns/generic.h>
  25 #include "smc_netns.h"
  26
  27 #include "smc_pnet.h"
  28 #include "smc_ib.h"
  29 #include "smc_ism.h"
  30 #include "smc_core.h"
  31
  32 #define SMC_ASCII_BLANK 32
  33
  34 static struct net_device *pnet_find_base_ndev(struct net_device *ndev);
  35
  36 static const struct nla_policy smc_pnet_policy[SMC_PNETID_MAX + 1] = {
  37         [SMC_PNETID_NAME] = {
  38                 .type = NLA_NUL_STRING,
  39                 .len = SMC_MAX_PNETID_LEN
  40         },
  41         [SMC_PNETID_ETHNAME] = {
  42                 .type = NLA_NUL_STRING,
  43                 .len = IFNAMSIZ - 1
  44         },
  45         [SMC_PNETID_IBNAME] = {
  46                 .type = NLA_NUL_STRING,
  47                 .len = IB_DEVICE_NAME_MAX - 1
  48         },
  49         [SMC_PNETID_IBPORT] = { .type = NLA_U8 }
  50 };
  51
  52 static struct genl_family smc_pnet_nl_family;
  53
  54 enum smc_pnet_nametype {
  55         SMC_PNET_ETH    = 1,
  56         SMC_PNET_IB     = 2,
  57 };
  58
  59 /* pnet entry stored in pnet table */
  60 struct smc_pnetentry {
  61         struct list_head list;
  62         char pnet_name[SMC_MAX_PNETID_LEN + 1];
  63         enum smc_pnet_nametype type;
  64         union {
  65                 struct {
  66                         char eth_name[IFNAMSIZ + 1];
  67                         struct net_device *ndev;
  68                 };
  69                 struct {
  70                         char ib_name[IB_DEVICE_NAME_MAX + 1];
  71                         u8 ib_port;
  72                 };
  73         };
  74 };
  75
  76 /* Check if two given pnetids match */
  77 static bool smc_pnet_match(u8 *pnetid1, u8 *pnetid2)
  78 {
  79         int i;
  80
  81         for (i = 0; i < SMC_MAX_PNETID_LEN; i++) {
  82                 if ((pnetid1[i] == 0 || pnetid1[i] == SMC_ASCII_BLANK) &&
  83                     (pnetid2[i] == 0 || pnetid2[i] == SMC_ASCII_BLANK))
  84                         break;
  85                 if (pnetid1[i] != pnetid2[i])
  86                         return false;
  87         }
  88         return true;
  89 }
  90
  91 /* Remove a pnetid from the pnet table.
  92  */
  93 static int smc_pnet_remove_by_pnetid(struct net *net, char *pnet_name)
  94 {
  95         struct smc_pnetentry *pnetelem, *tmp_pe;
  96         struct smc_pnettable *pnettable;
  97         struct smc_ib_device *ibdev;
  98         struct smcd_dev *smcd_dev;
  99         struct smc_net *sn;
 100         int rc = -ENOENT;
 101         int ibport;
 102
 103         /* get pnettable for namespace */
 104         sn = net_generic(net, smc_net_id);
 105         pnettable = &sn->pnettable;
 106
 107         /* remove table entry */
 108         write_lock(&pnettable->lock);
 109         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist,
 110                                  list) {
 111                 if (!pnet_name ||
 112                     smc_pnet_match(pnetelem->pnet_name, pnet_name)) {
 113                         list_del(&pnetelem->list);
 114                         if (pnetelem->type == SMC_PNET_ETH && pnetelem->ndev) {
 115                                 dev_put(pnetelem->ndev);
 116                                 pr_warn_ratelimited("smc: net device %s "
 117                                                     "erased user defined "
 118                                                     "pnetid %.16s\n",
 119                                                     pnetelem->eth_name,
 120                                                     pnetelem->pnet_name);
 121                         }
 122                         kfree(pnetelem);
 123                         rc = 0;
 124                 }
 125         }
 126         write_unlock(&pnettable->lock);
 127
 128         /* if this is not the initial namespace, stop here */
 129         if (net != &init_net)
 130                 return rc;
 131
 132         /* remove ib devices */
 133         mutex_lock(&smc_ib_devices.mutex);
 134         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 135                 for (ibport = 0; ibport < SMC_MAX_PORTS; ibport++) {
 136                         if (ibdev->pnetid_by_user[ibport] &&
 137                             (!pnet_name ||
 138                              smc_pnet_match(pnet_name,
 139                                             ibdev->pnetid[ibport]))) {
 140                                 pr_warn_ratelimited("smc: ib device %s ibport "
 141                                                     "%d erased user defined "
 142                                                     "pnetid %.16s\n",
 143                                                     ibdev->ibdev->name,
 144                                                     ibport + 1,
 145                                                     ibdev->pnetid[ibport]);
 146                                 memset(ibdev->pnetid[ibport], 0,
 147                                        SMC_MAX_PNETID_LEN);
 148                                 ibdev->pnetid_by_user[ibport] = false;
 149                                 rc = 0;
 150                         }
 151                 }
 152         }
 153         mutex_unlock(&smc_ib_devices.mutex);
 154         /* remove smcd devices */
 155         mutex_lock(&smcd_dev_list.mutex);
 156         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 157                 if (smcd_dev->pnetid_by_user &&
 158                     (!pnet_name ||
 159                      smc_pnet_match(pnet_name, smcd_dev->pnetid))) {
 160                         pr_warn_ratelimited("smc: smcd device %s "
 161                                             "erased user defined pnetid "
 162                                             "%.16s\n", dev_name(&smcd_dev->dev),
 163                                             smcd_dev->pnetid);
 164                         memset(smcd_dev->pnetid, 0, SMC_MAX_PNETID_LEN);
 165                         smcd_dev->pnetid_by_user = false;
 166                         rc = 0;
 167                 }
 168         }
 169         mutex_unlock(&smcd_dev_list.mutex);
 170         return rc;
 171 }
 172
 173 /* Add the reference to a given network device to the pnet table.
 174  */
 175 static int smc_pnet_add_by_ndev(struct net_device *ndev)
 176 {
 177         struct smc_pnetentry *pnetelem, *tmp_pe;
 178         struct smc_pnettable *pnettable;
 179         struct net *net = dev_net(ndev);
 180         struct smc_net *sn;
 181         int rc = -ENOENT;
 182
 183         /* get pnettable for namespace */
 184         sn = net_generic(net, smc_net_id);
 185         pnettable = &sn->pnettable;
 186
 187         write_lock(&pnettable->lock);
 188         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
 189                 if (pnetelem->type == SMC_PNET_ETH && !pnetelem->ndev &&
 190                     !strncmp(pnetelem->eth_name, ndev->name, IFNAMSIZ)) {
 191                         dev_hold(ndev);
 192                         pnetelem->ndev = ndev;
 193                         rc = 0;
 194                         pr_warn_ratelimited("smc: adding net device %s with "
 195                                             "user defined pnetid %.16s\n",
 196                                             pnetelem->eth_name,
 197                                             pnetelem->pnet_name);
 198                         break;
 199                 }
 200         }
 201         write_unlock(&pnettable->lock);
 202         return rc;
 203 }
 204
 205 /* Remove the reference to a given network device from the pnet table.
 206  */
 207 static int smc_pnet_remove_by_ndev(struct net_device *ndev)
 208 {
 209         struct smc_pnetentry *pnetelem, *tmp_pe;
 210         struct smc_pnettable *pnettable;
 211         struct net *net = dev_net(ndev);
 212         struct smc_net *sn;
 213         int rc = -ENOENT;
 214
 215         /* get pnettable for namespace */
 216         sn = net_generic(net, smc_net_id);
 217         pnettable = &sn->pnettable;
 218
 219         write_lock(&pnettable->lock);
 220         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
 221                 if (pnetelem->type == SMC_PNET_ETH && pnetelem->ndev == ndev) {
 222                         dev_put(pnetelem->ndev);
 223                         pnetelem->ndev = NULL;
 224                         rc = 0;
 225                         pr_warn_ratelimited("smc: removing net device %s with "
 226                                             "user defined pnetid %.16s\n",
 227                                             pnetelem->eth_name,
 228                                             pnetelem->pnet_name);
 229                         break;
 230                 }
 231         }
 232         write_unlock(&pnettable->lock);
 233         return rc;
 234 }
 235
 236 /* Apply pnetid to ib device when no pnetid is set.
 237  */
 238 static bool smc_pnet_apply_ib(struct smc_ib_device *ib_dev, u8 ib_port,
 239                               char *pnet_name)
 240 {
 241         u8 pnet_null[SMC_MAX_PNETID_LEN] = {0};
 242         bool applied = false;
 243
 244         mutex_lock(&smc_ib_devices.mutex);
 245         if (smc_pnet_match(ib_dev->pnetid[ib_port - 1], pnet_null)) {
 246                 memcpy(ib_dev->pnetid[ib_port - 1], pnet_name,
 247                        SMC_MAX_PNETID_LEN);
 248                 ib_dev->pnetid_by_user[ib_port - 1] = true;
 249                 applied = true;
 250         }
 251         mutex_unlock(&smc_ib_devices.mutex);
 252         return applied;
 253 }
 254
 255 /* Apply pnetid to smcd device when no pnetid is set.
 256  */
 257 static bool smc_pnet_apply_smcd(struct smcd_dev *smcd_dev, char *pnet_name)
 258 {
 259         u8 pnet_null[SMC_MAX_PNETID_LEN] = {0};
 260         bool applied = false;
 261
 262         mutex_lock(&smcd_dev_list.mutex);
 263         if (smc_pnet_match(smcd_dev->pnetid, pnet_null)) {
 264                 memcpy(smcd_dev->pnetid, pnet_name, SMC_MAX_PNETID_LEN);
 265                 smcd_dev->pnetid_by_user = true;
 266                 applied = true;
 267         }
 268         mutex_unlock(&smcd_dev_list.mutex);
 269         return applied;
 270 }
 271
 272 /* The limit for pnetid is 16 characters.
 273  * Valid characters should be (single-byte character set) a-z, A-Z, 0-9.
 274  * Lower case letters are converted to upper case.
 275  * Interior blanks should not be used.
 276  */
 277 static bool smc_pnetid_valid(const char *pnet_name, char *pnetid)
 278 {
 279         char *bf = skip_spaces(pnet_name);
 280         size_t len = strlen(bf);
 281         char *end = bf + len;
 282
 283         if (!len)
 284                 return false;
 285         while (--end >= bf && isspace(*end))
 286                 ;
 287         if (end - bf >= SMC_MAX_PNETID_LEN)
 288                 return false;
 289         while (bf <= end) {
 290                 if (!isalnum(*bf))
 291                         return false;
 292                 *pnetid++ = islower(*bf) ? toupper(*bf) : *bf;
 293                 bf++;
 294         }
 295         *pnetid = '\0';
 296         return true;
 297 }
 298
 299 /* Find an infiniband device by a given name. The device might not exist. */
 300 static struct smc_ib_device *smc_pnet_find_ib(char *ib_name)
 301 {
 302         struct smc_ib_device *ibdev;
 303
 304         mutex_lock(&smc_ib_devices.mutex);
 305         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 306                 if (!strncmp(ibdev->ibdev->name, ib_name,
 307                              sizeof(ibdev->ibdev->name)) ||
 308                     !strncmp(dev_name(ibdev->ibdev->dev.parent), ib_name,
 309                              IB_DEVICE_NAME_MAX - 1)) {
 310                         goto out;
 311                 }
 312         }
 313         ibdev = NULL;
 314 out:
 315         mutex_unlock(&smc_ib_devices.mutex);
 316         return ibdev;
 317 }
 318
 319 /* Find an smcd device by a given name. The device might not exist. */
 320 static struct smcd_dev *smc_pnet_find_smcd(char *smcd_name)
 321 {
 322         struct smcd_dev *smcd_dev;
 323
 324         mutex_lock(&smcd_dev_list.mutex);
 325         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 326                 if (!strncmp(dev_name(&smcd_dev->dev), smcd_name,
 327                              IB_DEVICE_NAME_MAX - 1))
 328                         goto out;
 329         }
 330         smcd_dev = NULL;
 331 out:
 332         mutex_unlock(&smcd_dev_list.mutex);
 333         return smcd_dev;
 334 }
 335
 336 static int smc_pnet_add_eth(struct smc_pnettable *pnettable, struct net *net,
 337                             char *eth_name, char *pnet_name)
 338 {
 339         struct smc_pnetentry *tmp_pe, *new_pe;
 340         struct net_device *ndev, *base_ndev;
 341         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 342         bool new_netdev;
 343         int rc;
 344
 345         /* check if (base) netdev already has a pnetid. If there is one, we do
 346          * not want to add a pnet table entry
 347          */
 348         rc = -EEXIST;
 349         ndev = dev_get_by_name(net, eth_name);  /* dev_hold() */
 350         if (ndev) {
 351                 base_ndev = pnet_find_base_ndev(ndev);
 352                 if (!smc_pnetid_by_dev_port(base_ndev->dev.parent,
 353                                             base_ndev->dev_port, ndev_pnetid))
 354                         goto out_put;
 355         }
 356
 357         /* add a new netdev entry to the pnet table if there isn't one */
 358         rc = -ENOMEM;
 359         new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
 360         if (!new_pe)
 361                 goto out_put;
 362         new_pe->type = SMC_PNET_ETH;
 363         memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
 364         strncpy(new_pe->eth_name, eth_name, IFNAMSIZ);
 365         new_pe->ndev = ndev;
 366
 367         rc = -EEXIST;
 368         new_netdev = true;
 369         write_lock(&pnettable->lock);
 370         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
 371                 if (tmp_pe->type == SMC_PNET_ETH &&
 372                     !strncmp(tmp_pe->eth_name, eth_name, IFNAMSIZ)) {
 373                         new_netdev = false;
 374                         break;
 375                 }
 376         }
 377         if (new_netdev) {
 378                 list_add_tail(&new_pe->list, &pnettable->pnetlist);
 379                 write_unlock(&pnettable->lock);
 380         } else {
 381                 write_unlock(&pnettable->lock);
 382                 kfree(new_pe);
 383                 goto out_put;
 384         }
 385         if (ndev)
 386                 pr_warn_ratelimited("smc: net device %s "
 387                                     "applied user defined pnetid %.16s\n",
 388                                     new_pe->eth_name, new_pe->pnet_name);
 389         return 0;
 390
 391 out_put:
 392         if (ndev)
 393                 dev_put(ndev);
 394         return rc;
 395 }
 396
 397 static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
 398                            u8 ib_port, char *pnet_name)
 399 {
 400         struct smc_pnetentry *tmp_pe, *new_pe;
 401         struct smc_ib_device *ib_dev;
 402         bool smcddev_applied = true;
 403         bool ibdev_applied = true;
 404         struct smcd_dev *smcd_dev;
 405         bool new_ibdev;
 406
 407         /* try to apply the pnetid to active devices */
 408         ib_dev = smc_pnet_find_ib(ib_name);
 409         if (ib_dev) {
 410                 ibdev_applied = smc_pnet_apply_ib(ib_dev, ib_port, pnet_name);
 411                 if (ibdev_applied)
 412                         pr_warn_ratelimited("smc: ib device %s ibport %d "
 413                                             "applied user defined pnetid "
 414                                             "%.16s\n", ib_dev->ibdev->name,
 415                                             ib_port,
 416                                             ib_dev->pnetid[ib_port - 1]);
 417         }
 418         smcd_dev = smc_pnet_find_smcd(ib_name);
 419         if (smcd_dev) {
 420                 smcddev_applied = smc_pnet_apply_smcd(smcd_dev, pnet_name);
 421                 if (smcddev_applied)
 422                         pr_warn_ratelimited("smc: smcd device %s "
 423                                             "applied user defined pnetid "
 424                                             "%.16s\n", dev_name(&smcd_dev->dev),
 425                                             smcd_dev->pnetid);
 426         }
 427         /* Apply fails when a device has a hardware-defined pnetid set, do not
 428          * add a pnet table entry in that case.
 429          */
 430         if (!ibdev_applied || !smcddev_applied)
 431                 return -EEXIST;
 432
 433         /* add a new ib entry to the pnet table if there isn't one */
 434         new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
 435         if (!new_pe)
 436                 return -ENOMEM;
 437         new_pe->type = SMC_PNET_IB;
 438         memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
 439         strncpy(new_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX);
 440         new_pe->ib_port = ib_port;
 441
 442         new_ibdev = true;
 443         write_lock(&pnettable->lock);
 444         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
 445                 if (tmp_pe->type == SMC_PNET_IB &&
 446                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
 447                         new_ibdev = false;
 448                         break;
 449                 }
 450         }
 451         if (new_ibdev) {
 452                 list_add_tail(&new_pe->list, &pnettable->pnetlist);
 453                 write_unlock(&pnettable->lock);
 454         } else {
 455                 write_unlock(&pnettable->lock);
 456                 kfree(new_pe);
 457         }
 458         return (new_ibdev) ? 0 : -EEXIST;
 459 }
 460
 461 /* Append a pnetid to the end of the pnet table if not already on this list.
 462  */
 463 static int smc_pnet_enter(struct net *net, struct nlattr *tb[])
 464 {
 465         char pnet_name[SMC_MAX_PNETID_LEN + 1];
 466         struct smc_pnettable *pnettable;
 467         bool new_netdev = false;
 468         bool new_ibdev = false;
 469         struct smc_net *sn;
 470         u8 ibport = 1;
 471         char *string;
 472         int rc;
 473
 474         /* get pnettable for namespace */
 475         sn = net_generic(net, smc_net_id);
 476         pnettable = &sn->pnettable;
 477
 478         rc = -EINVAL;
 479         if (!tb[SMC_PNETID_NAME])
 480                 goto error;
 481         string = (char *)nla_data(tb[SMC_PNETID_NAME]);
 482         if (!smc_pnetid_valid(string, pnet_name))
 483                 goto error;
 484
 485         if (tb[SMC_PNETID_ETHNAME]) {
 486                 string = (char *)nla_data(tb[SMC_PNETID_ETHNAME]);
 487                 rc = smc_pnet_add_eth(pnettable, net, string, pnet_name);
 488                 if (!rc)
 489                         new_netdev = true;
 490                 else if (rc != -EEXIST)
 491                         goto error;
 492         }
 493
 494         /* if this is not the initial namespace, stop here */
 495         if (net != &init_net)
 496                 return new_netdev ? 0 : -EEXIST;
 497
 498         rc = -EINVAL;
 499         if (tb[SMC_PNETID_IBNAME]) {
 500                 string = (char *)nla_data(tb[SMC_PNETID_IBNAME]);
 501                 string = strim(string);
 502                 if (tb[SMC_PNETID_IBPORT]) {
 503                         ibport = nla_get_u8(tb[SMC_PNETID_IBPORT]);
 504                         if (ibport < 1 || ibport > SMC_MAX_PORTS)
 505                                 goto error;
 506                 }
 507                 rc = smc_pnet_add_ib(pnettable, string, ibport, pnet_name);
 508                 if (!rc)
 509                         new_ibdev = true;
 510                 else if (rc != -EEXIST)
 511                         goto error;
 512         }
 513         return (new_netdev || new_ibdev) ? 0 : -EEXIST;
 514
 515 error:
 516         return rc;
 517 }
 518
 519 /* Convert an smc_pnetentry to a netlink attribute sequence */
 520 static int smc_pnet_set_nla(struct sk_buff *msg,
 521                             struct smc_pnetentry *pnetelem)
 522 {
 523         if (nla_put_string(msg, SMC_PNETID_NAME, pnetelem->pnet_name))
 524                 return -1;
 525         if (pnetelem->type == SMC_PNET_ETH) {
 526                 if (nla_put_string(msg, SMC_PNETID_ETHNAME,
 527                                    pnetelem->eth_name))
 528                         return -1;
 529         } else {
 530                 if (nla_put_string(msg, SMC_PNETID_ETHNAME, "n/a"))
 531                         return -1;
 532         }
 533         if (pnetelem->type == SMC_PNET_IB) {
 534                 if (nla_put_string(msg, SMC_PNETID_IBNAME, pnetelem->ib_name) ||
 535                     nla_put_u8(msg, SMC_PNETID_IBPORT, pnetelem->ib_port))
 536                         return -1;
 537         } else {
 538                 if (nla_put_string(msg, SMC_PNETID_IBNAME, "n/a") ||
 539                     nla_put_u8(msg, SMC_PNETID_IBPORT, 0xff))
 540                         return -1;
 541         }
 542
 543         return 0;
 544 }
 545
 546 static int smc_pnet_add(struct sk_buff *skb, struct genl_info *info)
 547 {
 548         struct net *net = genl_info_net(info);
 549
 550         return smc_pnet_enter(net, info->attrs);
 551 }
 552
 553 static int smc_pnet_del(struct sk_buff *skb, struct genl_info *info)
 554 {
 555         struct net *net = genl_info_net(info);
 556
 557         if (!info->attrs[SMC_PNETID_NAME])
 558                 return -EINVAL;
 559         return smc_pnet_remove_by_pnetid(net,
 560                                 (char *)nla_data(info->attrs[SMC_PNETID_NAME]));
 561 }
 562
 563 static int smc_pnet_dump_start(struct netlink_callback *cb)
 564 {
 565         cb->args[0] = 0;
 566         return 0;
 567 }
 568
 569 static int smc_pnet_dumpinfo(struct sk_buff *skb,
 570                              u32 portid, u32 seq, u32 flags,
 571                              struct smc_pnetentry *pnetelem)
 572 {
 573         void *hdr;
 574
 575         hdr = genlmsg_put(skb, portid, seq, &smc_pnet_nl_family,
 576                           flags, SMC_PNETID_GET);
 577         if (!hdr)
 578                 return -ENOMEM;
 579         if (smc_pnet_set_nla(skb, pnetelem) < 0) {
 580                 genlmsg_cancel(skb, hdr);
 581                 return -EMSGSIZE;
 582         }
 583         genlmsg_end(skb, hdr);
 584         return 0;
 585 }
 586
 587 static int _smc_pnet_dump(struct net *net, struct sk_buff *skb, u32 portid,
 588                           u32 seq, u8 *pnetid, int start_idx)
 589 {
 590         struct smc_pnettable *pnettable;
 591         struct smc_pnetentry *pnetelem;
 592         struct smc_net *sn;
 593         int idx = 0;
 594
 595         /* get pnettable for namespace */
 596         sn = net_generic(net, smc_net_id);
 597         pnettable = &sn->pnettable;
 598
 599         /* dump pnettable entries */
 600         read_lock(&pnettable->lock);
 601         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 602                 if (pnetid && !smc_pnet_match(pnetelem->pnet_name, pnetid))
 603                         continue;
 604                 if (idx++ < start_idx)
 605                         continue;
 606                 /* if this is not the initial namespace, dump only netdev */
 607                 if (net != &init_net && pnetelem->type != SMC_PNET_ETH)
 608                         continue;
 609                 if (smc_pnet_dumpinfo(skb, portid, seq, NLM_F_MULTI,
 610                                       pnetelem)) {
 611                         --idx;
 612                         break;
 613                 }
 614         }
 615         read_unlock(&pnettable->lock);
 616         return idx;
 617 }
 618
 619 static int smc_pnet_dump(struct sk_buff *skb, struct netlink_callback *cb)
 620 {
 621         struct net *net = sock_net(skb->sk);
 622         int idx;
 623
 624         idx = _smc_pnet_dump(net, skb, NETLINK_CB(cb->skb).portid,
 625                              cb->nlh->nlmsg_seq, NULL, cb->args[0]);
 626
 627         cb->args[0] = idx;
 628         return skb->len;
 629 }
 630
 631 /* Retrieve one PNETID entry */
 632 static int smc_pnet_get(struct sk_buff *skb, struct genl_info *info)
 633 {
 634         struct net *net = genl_info_net(info);
 635         struct sk_buff *msg;
 636         void *hdr;
 637
 638         if (!info->attrs[SMC_PNETID_NAME])
 639                 return -EINVAL;
 640
 641         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 642         if (!msg)
 643                 return -ENOMEM;
 644
 645         _smc_pnet_dump(net, msg, info->snd_portid, info->snd_seq,
 646                        nla_data(info->attrs[SMC_PNETID_NAME]), 0);
 647
 648         /* finish multi part message and send it */
 649         hdr = nlmsg_put(msg, info->snd_portid, info->snd_seq, NLMSG_DONE, 0,
 650                         NLM_F_MULTI);
 651         if (!hdr) {
 652                 nlmsg_free(msg);
 653                 return -EMSGSIZE;
 654         }
 655         return genlmsg_reply(msg, info);
 656 }
 657
 658 /* Remove and delete all pnetids from pnet table.
 659  */
 660 static int smc_pnet_flush(struct sk_buff *skb, struct genl_info *info)
 661 {
 662         struct net *net = genl_info_net(info);
 663
 664         smc_pnet_remove_by_pnetid(net, NULL);
 665         return 0;
 666 }
 667
 668 /* SMC_PNETID generic netlink operation definition */
 669 static const struct genl_ops smc_pnet_ops[] = {
 670         {
 671                 .cmd = SMC_PNETID_GET,
 672                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 673                 /* can be retrieved by unprivileged users */
 674                 .doit = smc_pnet_get,
 675                 .dumpit = smc_pnet_dump,
 676                 .start = smc_pnet_dump_start
 677         },
 678         {
 679                 .cmd = SMC_PNETID_ADD,
 680                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 681                 .flags = GENL_ADMIN_PERM,
 682                 .doit = smc_pnet_add
 683         },
 684         {
 685                 .cmd = SMC_PNETID_DEL,
 686                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 687                 .flags = GENL_ADMIN_PERM,
 688                 .doit = smc_pnet_del
 689         },
 690         {
 691                 .cmd = SMC_PNETID_FLUSH,
 692                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 693                 .flags = GENL_ADMIN_PERM,
 694                 .doit = smc_pnet_flush
 695         }
 696 };
 697
 698 /* SMC_PNETID family definition */
 699 static struct genl_family smc_pnet_nl_family __ro_after_init = {
 700         .hdrsize = 0,
 701         .name = SMCR_GENL_FAMILY_NAME,
 702         .version = SMCR_GENL_FAMILY_VERSION,
 703         .maxattr = SMC_PNETID_MAX,
 704         .policy = smc_pnet_policy,
 705         .netnsok = true,
 706         .module = THIS_MODULE,
 707         .ops = smc_pnet_ops,
 708         .n_ops =  ARRAY_SIZE(smc_pnet_ops)
 709 };
 710
 711 static int smc_pnet_netdev_event(struct notifier_block *this,
 712                                  unsigned long event, void *ptr)
 713 {
 714         struct net_device *event_dev = netdev_notifier_info_to_dev(ptr);
 715
 716         switch (event) {
 717         case NETDEV_REBOOT:
 718         case NETDEV_UNREGISTER:
 719                 smc_pnet_remove_by_ndev(event_dev);
 720                 return NOTIFY_OK;
 721         case NETDEV_REGISTER:
 722                 smc_pnet_add_by_ndev(event_dev);
 723                 return NOTIFY_OK;
 724         default:
 725                 return NOTIFY_DONE;
 726         }
 727 }
 728
 729 static struct notifier_block smc_netdev_notifier = {
 730         .notifier_call = smc_pnet_netdev_event
 731 };
 732
 733 /* init network namespace */
 734 int smc_pnet_net_init(struct net *net)
 735 {
 736         struct smc_net *sn = net_generic(net, smc_net_id);
 737         struct smc_pnettable *pnettable = &sn->pnettable;
 738
 739         INIT_LIST_HEAD(&pnettable->pnetlist);
 740         rwlock_init(&pnettable->lock);
 741
 742         return 0;
 743 }
 744
 745 int __init smc_pnet_init(void)
 746 {
 747         int rc;
 748
 749         rc = genl_register_family(&smc_pnet_nl_family);
 750         if (rc)
 751                 return rc;
 752         rc = register_netdevice_notifier(&smc_netdev_notifier);
 753         if (rc)
 754                 genl_unregister_family(&smc_pnet_nl_family);
 755         return rc;
 756 }
 757
 758 /* exit network namespace */
 759 void smc_pnet_net_exit(struct net *net)
 760 {
 761         /* flush pnet table */
 762         smc_pnet_remove_by_pnetid(net, NULL);
 763 }
 764
 765 void smc_pnet_exit(void)
 766 {
 767         unregister_netdevice_notifier(&smc_netdev_notifier);
 768         genl_unregister_family(&smc_pnet_nl_family);
 769 }
 770
 771 /* Determine one base device for stacked net devices.
 772  * If the lower device level contains more than one devices
 773  * (for instance with bonding slaves), just the first device
 774  * is used to reach a base device.
 775  */
 776 static struct net_device *pnet_find_base_ndev(struct net_device *ndev)
 777 {
 778         int i, nest_lvl;
 779
 780         rtnl_lock();
 781         nest_lvl = ndev->lower_level;
 782         for (i = 0; i < nest_lvl; i++) {
 783                 struct list_head *lower = &ndev->adj_list.lower;
 784
 785                 if (list_empty(lower))
 786                         break;
 787                 lower = lower->next;
 788                 ndev = netdev_lower_get_next(ndev, &lower);
 789         }
 790         rtnl_unlock();
 791         return ndev;
 792 }
 793
 794 static int smc_pnet_find_ndev_pnetid_by_table(struct net_device *ndev,
 795                                               u8 *pnetid)
 796 {
 797         struct smc_pnettable *pnettable;
 798         struct net *net = dev_net(ndev);
 799         struct smc_pnetentry *pnetelem;
 800         struct smc_net *sn;
 801         int rc = -ENOENT;
 802
 803         /* get pnettable for namespace */
 804         sn = net_generic(net, smc_net_id);
 805         pnettable = &sn->pnettable;
 806
 807         read_lock(&pnettable->lock);
 808         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 809                 if (pnetelem->type == SMC_PNET_ETH && ndev == pnetelem->ndev) {
 810                         /* get pnetid of netdev device */
 811                         memcpy(pnetid, pnetelem->pnet_name, SMC_MAX_PNETID_LEN);
 812                         rc = 0;
 813                         break;
 814                 }
 815         }
 816         read_unlock(&pnettable->lock);
 817         return rc;
 818 }
 819
 820 /* find a roce device for the given pnetid */
 821 static void _smc_pnet_find_roce_by_pnetid(u8 *pnet_id,
 822                                           struct smc_init_info *ini,
 823                                           struct smc_ib_device *known_dev)
 824 {
 825         struct smc_ib_device *ibdev;
 826         int i;
 827
 828         ini->ib_dev = NULL;
 829         mutex_lock(&smc_ib_devices.mutex);
 830         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 831                 if (ibdev == known_dev)
 832                         continue;
 833                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
 834                         if (!rdma_is_port_valid(ibdev->ibdev, i))
 835                                 continue;
 836                         if (smc_pnet_match(ibdev->pnetid[i - 1], pnet_id) &&
 837                             smc_ib_port_active(ibdev, i) &&
 838                             !test_bit(i - 1, ibdev->ports_going_away) &&
 839                             !smc_ib_determine_gid(ibdev, i, ini->vlan_id,
 840                                                   ini->ib_gid, NULL)) {
 841                                 ini->ib_dev = ibdev;
 842                                 ini->ib_port = i;
 843                                 goto out;
 844                         }
 845                 }
 846         }
 847 out:
 848         mutex_unlock(&smc_ib_devices.mutex);
 849 }
 850
 851 /* find alternate roce device with same pnet_id and vlan_id */
 852 void smc_pnet_find_alt_roce(struct smc_link_group *lgr,
 853                             struct smc_init_info *ini,
 854                             struct smc_ib_device *known_dev)
 855 {
 856         _smc_pnet_find_roce_by_pnetid(lgr->pnet_id, ini, known_dev);
 857 }
 858
 859 /* if handshake network device belongs to a roce device, return its
 860  * IB device and port
 861  */
 862 static void smc_pnet_find_rdma_dev(struct net_device *netdev,
 863                                    struct smc_init_info *ini)
 864 {
 865         struct smc_ib_device *ibdev;
 866
 867         mutex_lock(&smc_ib_devices.mutex);
 868         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 869                 struct net_device *ndev;
 870                 int i;
 871
 872                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
 873                         if (!rdma_is_port_valid(ibdev->ibdev, i))
 874                                 continue;
 875                         if (!ibdev->ibdev->ops.get_netdev)
 876                                 continue;
 877                         ndev = ibdev->ibdev->ops.get_netdev(ibdev->ibdev, i);
 878                         if (!ndev)
 879                                 continue;
 880                         dev_put(ndev);
 881                         if (netdev == ndev &&
 882                             smc_ib_port_active(ibdev, i) &&
 883                             !test_bit(i - 1, ibdev->ports_going_away) &&
 884                             !smc_ib_determine_gid(ibdev, i, ini->vlan_id,
 885                                                   ini->ib_gid, NULL)) {
 886                                 ini->ib_dev = ibdev;
 887                                 ini->ib_port = i;
 888                                 break;
 889                         }
 890                 }
 891         }
 892         mutex_unlock(&smc_ib_devices.mutex);
 893 }
 894
 895 /* Determine the corresponding IB device port based on the hardware PNETID.
 896  * Searching stops at the first matching active IB device port with vlan_id
 897  * configured.
 898  * If nothing found, check pnetid table.
 899  * If nothing found, try to use handshake device
 900  */
 901 static void smc_pnet_find_roce_by_pnetid(struct net_device *ndev,
 902                                          struct smc_init_info *ini)
 903 {
 904         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 905
 906         ndev = pnet_find_base_ndev(ndev);
 907         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
 908                                    ndev_pnetid) &&
 909             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid)) {
 910                 smc_pnet_find_rdma_dev(ndev, ini);
 911                 return; /* pnetid could not be determined */
 912         }
 913         _smc_pnet_find_roce_by_pnetid(ndev_pnetid, ini, NULL);
 914 }
 915
 916 static void smc_pnet_find_ism_by_pnetid(struct net_device *ndev,
 917                                         struct smc_init_info *ini)
 918 {
 919         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 920         struct smcd_dev *ismdev;
 921
 922         ndev = pnet_find_base_ndev(ndev);
 923         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
 924                                    ndev_pnetid) &&
 925             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid))
 926                 return; /* pnetid could not be determined */
 927
 928         mutex_lock(&smcd_dev_list.mutex);
 929         list_for_each_entry(ismdev, &smcd_dev_list.list, list) {
 930                 if (smc_pnet_match(ismdev->pnetid, ndev_pnetid) &&
 931                     !ismdev->going_away) {
 932                         ini->ism_dev = ismdev;
 933                         break;
 934                 }
 935         }
 936         mutex_unlock(&smcd_dev_list.mutex);
 937 }
 938
 939 /* PNET table analysis for a given sock:
 940  * determine ib_device and port belonging to used internal TCP socket
 941  * ethernet interface.
 942  */
 943 void smc_pnet_find_roce_resource(struct sock *sk, struct smc_init_info *ini)
 944 {
 945         struct dst_entry *dst = sk_dst_get(sk);
 946
 947         ini->ib_dev = NULL;
 948         ini->ib_port = 0;
 949         if (!dst)
 950                 goto out;
 951         if (!dst->dev)
 952                 goto out_rel;
 953
 954         smc_pnet_find_roce_by_pnetid(dst->dev, ini);
 955
 956 out_rel:
 957         dst_release(dst);
 958 out:
 959         return;
 960 }
 961
 962 void smc_pnet_find_ism_resource(struct sock *sk, struct smc_init_info *ini)
 963 {
 964         struct dst_entry *dst = sk_dst_get(sk);
 965
 966         ini->ism_dev = NULL;
 967         if (!dst)
 968                 goto out;
 969         if (!dst->dev)
 970                 goto out_rel;
 971
 972         smc_pnet_find_ism_by_pnetid(dst->dev, ini);
 973
 974 out_rel:
 975         dst_release(dst);
 976 out:
 977         return;
 978 }
 979
 980 /* Lookup and apply a pnet table entry to the given ib device.
 981  */
 982 int smc_pnetid_by_table_ib(struct smc_ib_device *smcibdev, u8 ib_port)
 983 {
 984         char *ib_name = smcibdev->ibdev->name;
 985         struct smc_pnettable *pnettable;
 986         struct smc_pnetentry *tmp_pe;
 987         struct smc_net *sn;
 988         int rc = -ENOENT;
 989
 990         /* get pnettable for init namespace */
 991         sn = net_generic(&init_net, smc_net_id);
 992         pnettable = &sn->pnettable;
 993
 994         read_lock(&pnettable->lock);
 995         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
 996                 if (tmp_pe->type == SMC_PNET_IB &&
 997                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX) &&
 998                     tmp_pe->ib_port == ib_port) {
 999                         smc_pnet_apply_ib(smcibdev, ib_port, tmp_pe->pnet_name);
1000                         rc = 0;
1001                         break;
1002                 }
1003         }
1004         read_unlock(&pnettable->lock);
1005
1006         return rc;
1007 }
1008
1009 /* Lookup and apply a pnet table entry to the given smcd device.
1010  */
1011 int smc_pnetid_by_table_smcd(struct smcd_dev *smcddev)
1012 {
1013         const char *ib_name = dev_name(&smcddev->dev);
1014         struct smc_pnettable *pnettable;
1015         struct smc_pnetentry *tmp_pe;
1016         struct smc_net *sn;
1017         int rc = -ENOENT;
1018
1019         /* get pnettable for init namespace */
1020         sn = net_generic(&init_net, smc_net_id);
1021         pnettable = &sn->pnettable;
1022
1023         read_lock(&pnettable->lock);
1024         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
1025                 if (tmp_pe->type == SMC_PNET_IB &&
1026                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
1027                         smc_pnet_apply_smcd(smcddev, tmp_pe->pnet_name);
1028                         rc = 0;
1029                         break;
1030                 }
1031         }
1032         read_unlock(&pnettable->lock);
1033
1034         return rc;
1035 }