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 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);
  34
  35 static const struct nla_policy smc_pnet_policy[SMC_PNETID_MAX + 1] = {
  36         [SMC_PNETID_NAME] = {
  37                 .type = NLA_NUL_STRING,
  38                 .len = SMC_MAX_PNETID_LEN
  39         },
  40         [SMC_PNETID_ETHNAME] = {
  41                 .type = NLA_NUL_STRING,
  42                 .len = IFNAMSIZ - 1
  43         },
  44         [SMC_PNETID_IBNAME] = {
  45                 .type = NLA_NUL_STRING,
  46                 .len = IB_DEVICE_NAME_MAX - 1
  47         },
  48         [SMC_PNETID_IBPORT] = { .type = NLA_U8 }
  49 };
  50
  51 static struct genl_family smc_pnet_nl_family;
  52
  53 enum smc_pnet_nametype {
  54         SMC_PNET_ETH    = 1,
  55         SMC_PNET_IB     = 2,
  56 };
  57
  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;
  63         union {
  64                 struct {
  65                         char eth_name[IFNAMSIZ + 1];
  66                         struct net_device *ndev;
  67                 };
  68                 struct {
  69                         char ib_name[IB_DEVICE_NAME_MAX + 1];
  70                         u8 ib_port;
  71                 };
  72         };
  73 };
  74
  75 /* Check if the pnetid is set */
  76 bool smc_pnet_is_pnetid_set(u8 *pnetid)
  77 {
  78         if (pnetid[0] == 0 || pnetid[0] == _S)
  79                 return false;
  80         return true;
  81 }
  82
  83 /* Check if two given pnetids match */
  84 static bool smc_pnet_match(u8 *pnetid1, u8 *pnetid2)
  85 {
  86         int i;
  87
  88         for (i = 0; i < SMC_MAX_PNETID_LEN; i++) {
  89                 if ((pnetid1[i] == 0 || pnetid1[i] == _S) &&
  90                     (pnetid2[i] == 0 || pnetid2[i] == _S))
  91                         break;
  92                 if (pnetid1[i] != pnetid2[i])
  93                         return false;
  94         }
  95         return true;
  96 }
  97
  98 /* Remove a pnetid from the pnet table.
  99  */
 100 static int smc_pnet_remove_by_pnetid(struct net *net, char *pnet_name)
 101 {
 102         struct smc_pnetentry *pnetelem, *tmp_pe;
 103         struct smc_pnettable *pnettable;
 104         struct smc_ib_device *ibdev;
 105         struct smcd_dev *smcd_dev;
 106         struct smc_net *sn;
 107         int rc = -ENOENT;
 108         int ibport;
 109
 110         /* get pnettable for namespace */
 111         sn = net_generic(net, smc_net_id);
 112         pnettable = &sn->pnettable;
 113
 114         /* remove table entry */
 115         write_lock(&pnettable->lock);
 116         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist,
 117                                  list) {
 118                 if (!pnet_name ||
 119                     smc_pnet_match(pnetelem->pnet_name, pnet_name)) {
 120                         list_del(&pnetelem->list);
 121                         if (pnetelem->type == SMC_PNET_ETH && pnetelem->ndev) {
 122                                 dev_put(pnetelem->ndev);
 123                                 pr_warn_ratelimited("smc: net device %s "
 124                                                     "erased user defined "
 125                                                     "pnetid %.16s\n",
 126                                                     pnetelem->eth_name,
 127                                                     pnetelem->pnet_name);
 128                         }
 129                         kfree(pnetelem);
 130                         rc = 0;
 131                 }
 132         }
 133         write_unlock(&pnettable->lock);
 134
 135         /* if this is not the initial namespace, stop here */
 136         if (net != &init_net)
 137                 return rc;
 138
 139         /* remove ib devices */
 140         mutex_lock(&smc_ib_devices.mutex);
 141         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 142                 for (ibport = 0; ibport < SMC_MAX_PORTS; ibport++) {
 143                         if (ibdev->pnetid_by_user[ibport] &&
 144                             (!pnet_name ||
 145                              smc_pnet_match(pnet_name,
 146                                             ibdev->pnetid[ibport]))) {
 147                                 pr_warn_ratelimited("smc: ib device %s ibport "
 148                                                     "%d erased user defined "
 149                                                     "pnetid %.16s\n",
 150                                                     ibdev->ibdev->name,
 151                                                     ibport + 1,
 152                                                     ibdev->pnetid[ibport]);
 153                                 memset(ibdev->pnetid[ibport], 0,
 154                                        SMC_MAX_PNETID_LEN);
 155                                 ibdev->pnetid_by_user[ibport] = false;
 156                                 rc = 0;
 157                         }
 158                 }
 159         }
 160         mutex_unlock(&smc_ib_devices.mutex);
 161         /* remove smcd devices */
 162         mutex_lock(&smcd_dev_list.mutex);
 163         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 164                 if (smcd_dev->pnetid_by_user &&
 165                     (!pnet_name ||
 166                      smc_pnet_match(pnet_name, smcd_dev->pnetid))) {
 167                         pr_warn_ratelimited("smc: smcd device %s "
 168                                             "erased user defined pnetid "
 169                                             "%.16s\n", dev_name(&smcd_dev->dev),
 170                                             smcd_dev->pnetid);
 171                         memset(smcd_dev->pnetid, 0, SMC_MAX_PNETID_LEN);
 172                         smcd_dev->pnetid_by_user = false;
 173                         rc = 0;
 174                 }
 175         }
 176         mutex_unlock(&smcd_dev_list.mutex);
 177         return rc;
 178 }
 179
 180 /* Add the reference to a given network device to the pnet table.
 181  */
 182 static int smc_pnet_add_by_ndev(struct net_device *ndev)
 183 {
 184         struct smc_pnetentry *pnetelem, *tmp_pe;
 185         struct smc_pnettable *pnettable;
 186         struct net *net = dev_net(ndev);
 187         struct smc_net *sn;
 188         int rc = -ENOENT;
 189
 190         /* get pnettable for namespace */
 191         sn = net_generic(net, smc_net_id);
 192         pnettable = &sn->pnettable;
 193
 194         write_lock(&pnettable->lock);
 195         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
 196                 if (pnetelem->type == SMC_PNET_ETH && !pnetelem->ndev &&
 197                     !strncmp(pnetelem->eth_name, ndev->name, IFNAMSIZ)) {
 198                         dev_hold(ndev);
 199                         pnetelem->ndev = ndev;
 200                         rc = 0;
 201                         pr_warn_ratelimited("smc: adding net device %s with "
 202                                             "user defined pnetid %.16s\n",
 203                                             pnetelem->eth_name,
 204                                             pnetelem->pnet_name);
 205                         break;
 206                 }
 207         }
 208         write_unlock(&pnettable->lock);
 209         return rc;
 210 }
 211
 212 /* Remove the reference to a given network device from the pnet table.
 213  */
 214 static int smc_pnet_remove_by_ndev(struct net_device *ndev)
 215 {
 216         struct smc_pnetentry *pnetelem, *tmp_pe;
 217         struct smc_pnettable *pnettable;
 218         struct net *net = dev_net(ndev);
 219         struct smc_net *sn;
 220         int rc = -ENOENT;
 221
 222         /* get pnettable for namespace */
 223         sn = net_generic(net, smc_net_id);
 224         pnettable = &sn->pnettable;
 225
 226         write_lock(&pnettable->lock);
 227         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
 228                 if (pnetelem->type == SMC_PNET_ETH && pnetelem->ndev == ndev) {
 229                         dev_put(pnetelem->ndev);
 230                         pnetelem->ndev = NULL;
 231                         rc = 0;
 232                         pr_warn_ratelimited("smc: removing net device %s with "
 233                                             "user defined pnetid %.16s\n",
 234                                             pnetelem->eth_name,
 235                                             pnetelem->pnet_name);
 236                         break;
 237                 }
 238         }
 239         write_unlock(&pnettable->lock);
 240         return rc;
 241 }
 242
 243 /* Apply pnetid to ib device when no pnetid is set.
 244  */
 245 static bool smc_pnet_apply_ib(struct smc_ib_device *ib_dev, u8 ib_port,
 246                               char *pnet_name)
 247 {
 248         bool applied = false;
 249
 250         mutex_lock(&smc_ib_devices.mutex);
 251         if (!smc_pnet_is_pnetid_set(ib_dev->pnetid[ib_port - 1])) {
 252                 memcpy(ib_dev->pnetid[ib_port - 1], pnet_name,
 253                        SMC_MAX_PNETID_LEN);
 254                 ib_dev->pnetid_by_user[ib_port - 1] = true;
 255                 applied = true;
 256         }
 257         mutex_unlock(&smc_ib_devices.mutex);
 258         return applied;
 259 }
 260
 261 /* Apply pnetid to smcd device when no pnetid is set.
 262  */
 263 static bool smc_pnet_apply_smcd(struct smcd_dev *smcd_dev, char *pnet_name)
 264 {
 265         bool applied = false;
 266
 267         mutex_lock(&smcd_dev_list.mutex);
 268         if (!smc_pnet_is_pnetid_set(smcd_dev->pnetid)) {
 269                 memcpy(smcd_dev->pnetid, pnet_name, SMC_MAX_PNETID_LEN);
 270                 smcd_dev->pnetid_by_user = true;
 271                 applied = true;
 272         }
 273         mutex_unlock(&smcd_dev_list.mutex);
 274         return applied;
 275 }
 276
 277 /* The limit for pnetid is 16 characters.
 278  * Valid characters should be (single-byte character set) a-z, A-Z, 0-9.
 279  * Lower case letters are converted to upper case.
 280  * Interior blanks should not be used.
 281  */
 282 static bool smc_pnetid_valid(const char *pnet_name, char *pnetid)
 283 {
 284         char *bf = skip_spaces(pnet_name);
 285         size_t len = strlen(bf);
 286         char *end = bf + len;
 287
 288         if (!len)
 289                 return false;
 290         while (--end >= bf && isspace(*end))
 291                 ;
 292         if (end - bf >= SMC_MAX_PNETID_LEN)
 293                 return false;
 294         while (bf <= end) {
 295                 if (!isalnum(*bf))
 296                         return false;
 297                 *pnetid++ = islower(*bf) ? toupper(*bf) : *bf;
 298                 bf++;
 299         }
 300         *pnetid = '\0';
 301         return true;
 302 }
 303
 304 /* Find an infiniband device by a given name. The device might not exist. */
 305 static struct smc_ib_device *smc_pnet_find_ib(char *ib_name)
 306 {
 307         struct smc_ib_device *ibdev;
 308
 309         mutex_lock(&smc_ib_devices.mutex);
 310         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 311                 if (!strncmp(ibdev->ibdev->name, ib_name,
 312                              sizeof(ibdev->ibdev->name)) ||
 313                     !strncmp(dev_name(ibdev->ibdev->dev.parent), ib_name,
 314                              IB_DEVICE_NAME_MAX - 1)) {
 315                         goto out;
 316                 }
 317         }
 318         ibdev = NULL;
 319 out:
 320         mutex_unlock(&smc_ib_devices.mutex);
 321         return ibdev;
 322 }
 323
 324 /* Find an smcd device by a given name. The device might not exist. */
 325 static struct smcd_dev *smc_pnet_find_smcd(char *smcd_name)
 326 {
 327         struct smcd_dev *smcd_dev;
 328
 329         mutex_lock(&smcd_dev_list.mutex);
 330         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 331                 if (!strncmp(dev_name(&smcd_dev->dev), smcd_name,
 332                              IB_DEVICE_NAME_MAX - 1))
 333                         goto out;
 334         }
 335         smcd_dev = NULL;
 336 out:
 337         mutex_unlock(&smcd_dev_list.mutex);
 338         return smcd_dev;
 339 }
 340
 341 static int smc_pnet_add_eth(struct smc_pnettable *pnettable, struct net *net,
 342                             char *eth_name, char *pnet_name)
 343 {
 344         struct smc_pnetentry *tmp_pe, *new_pe;
 345         struct net_device *ndev, *base_ndev;
 346         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 347         bool new_netdev;
 348         int rc;
 349
 350         /* check if (base) netdev already has a pnetid. If there is one, we do
 351          * not want to add a pnet table entry
 352          */
 353         rc = -EEXIST;
 354         ndev = dev_get_by_name(net, eth_name);  /* dev_hold() */
 355         if (ndev) {
 356                 base_ndev = pnet_find_base_ndev(ndev);
 357                 if (!smc_pnetid_by_dev_port(base_ndev->dev.parent,
 358                                             base_ndev->dev_port, ndev_pnetid))
 359                         goto out_put;
 360         }
 361
 362         /* add a new netdev entry to the pnet table if there isn't one */
 363         rc = -ENOMEM;
 364         new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
 365         if (!new_pe)
 366                 goto out_put;
 367         new_pe->type = SMC_PNET_ETH;
 368         memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
 369         strncpy(new_pe->eth_name, eth_name, IFNAMSIZ);
 370         new_pe->ndev = ndev;
 371
 372         rc = -EEXIST;
 373         new_netdev = true;
 374         write_lock(&pnettable->lock);
 375         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
 376                 if (tmp_pe->type == SMC_PNET_ETH &&
 377                     !strncmp(tmp_pe->eth_name, eth_name, IFNAMSIZ)) {
 378                         new_netdev = false;
 379                         break;
 380                 }
 381         }
 382         if (new_netdev) {
 383                 list_add_tail(&new_pe->list, &pnettable->pnetlist);
 384                 write_unlock(&pnettable->lock);
 385         } else {
 386                 write_unlock(&pnettable->lock);
 387                 kfree(new_pe);
 388                 goto out_put;
 389         }
 390         if (ndev)
 391                 pr_warn_ratelimited("smc: net device %s "
 392                                     "applied user defined pnetid %.16s\n",
 393                                     new_pe->eth_name, new_pe->pnet_name);
 394         return 0;
 395
 396 out_put:
 397         if (ndev)
 398                 dev_put(ndev);
 399         return rc;
 400 }
 401
 402 static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
 403                            u8 ib_port, char *pnet_name)
 404 {
 405         struct smc_pnetentry *tmp_pe, *new_pe;
 406         struct smc_ib_device *ib_dev;
 407         bool smcddev_applied = true;
 408         bool ibdev_applied = true;
 409         struct smcd_dev *smcd_dev;
 410         bool new_ibdev;
 411
 412         /* try to apply the pnetid to active devices */
 413         ib_dev = smc_pnet_find_ib(ib_name);
 414         if (ib_dev) {
 415                 ibdev_applied = smc_pnet_apply_ib(ib_dev, ib_port, pnet_name);
 416                 if (ibdev_applied)
 417                         pr_warn_ratelimited("smc: ib device %s ibport %d "
 418                                             "applied user defined pnetid "
 419                                             "%.16s\n", ib_dev->ibdev->name,
 420                                             ib_port,
 421                                             ib_dev->pnetid[ib_port - 1]);
 422         }
 423         smcd_dev = smc_pnet_find_smcd(ib_name);
 424         if (smcd_dev) {
 425                 smcddev_applied = smc_pnet_apply_smcd(smcd_dev, pnet_name);
 426                 if (smcddev_applied)
 427                         pr_warn_ratelimited("smc: smcd device %s "
 428                                             "applied user defined pnetid "
 429                                             "%.16s\n", dev_name(&smcd_dev->dev),
 430                                             smcd_dev->pnetid);
 431         }
 432         /* Apply fails when a device has a hardware-defined pnetid set, do not
 433          * add a pnet table entry in that case.
 434          */
 435         if (!ibdev_applied || !smcddev_applied)
 436                 return -EEXIST;
 437
 438         /* add a new ib entry to the pnet table if there isn't one */
 439         new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
 440         if (!new_pe)
 441                 return -ENOMEM;
 442         new_pe->type = SMC_PNET_IB;
 443         memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
 444         strncpy(new_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX);
 445         new_pe->ib_port = ib_port;
 446
 447         new_ibdev = true;
 448         write_lock(&pnettable->lock);
 449         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
 450                 if (tmp_pe->type == SMC_PNET_IB &&
 451                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
 452                         new_ibdev = false;
 453                         break;
 454                 }
 455         }
 456         if (new_ibdev) {
 457                 list_add_tail(&new_pe->list, &pnettable->pnetlist);
 458                 write_unlock(&pnettable->lock);
 459         } else {
 460                 write_unlock(&pnettable->lock);
 461                 kfree(new_pe);
 462         }
 463         return (new_ibdev) ? 0 : -EEXIST;
 464 }
 465
 466 /* Append a pnetid to the end of the pnet table if not already on this list.
 467  */
 468 static int smc_pnet_enter(struct net *net, struct nlattr *tb[])
 469 {
 470         char pnet_name[SMC_MAX_PNETID_LEN + 1];
 471         struct smc_pnettable *pnettable;
 472         bool new_netdev = false;
 473         bool new_ibdev = false;
 474         struct smc_net *sn;
 475         u8 ibport = 1;
 476         char *string;
 477         int rc;
 478
 479         /* get pnettable for namespace */
 480         sn = net_generic(net, smc_net_id);
 481         pnettable = &sn->pnettable;
 482
 483         rc = -EINVAL;
 484         if (!tb[SMC_PNETID_NAME])
 485                 goto error;
 486         string = (char *)nla_data(tb[SMC_PNETID_NAME]);
 487         if (!smc_pnetid_valid(string, pnet_name))
 488                 goto error;
 489
 490         if (tb[SMC_PNETID_ETHNAME]) {
 491                 string = (char *)nla_data(tb[SMC_PNETID_ETHNAME]);
 492                 rc = smc_pnet_add_eth(pnettable, net, string, pnet_name);
 493                 if (!rc)
 494                         new_netdev = true;
 495                 else if (rc != -EEXIST)
 496                         goto error;
 497         }
 498
 499         /* if this is not the initial namespace, stop here */
 500         if (net != &init_net)
 501                 return new_netdev ? 0 : -EEXIST;
 502
 503         rc = -EINVAL;
 504         if (tb[SMC_PNETID_IBNAME]) {
 505                 string = (char *)nla_data(tb[SMC_PNETID_IBNAME]);
 506                 string = strim(string);
 507                 if (tb[SMC_PNETID_IBPORT]) {
 508                         ibport = nla_get_u8(tb[SMC_PNETID_IBPORT]);
 509                         if (ibport < 1 || ibport > SMC_MAX_PORTS)
 510                                 goto error;
 511                 }
 512                 rc = smc_pnet_add_ib(pnettable, string, ibport, pnet_name);
 513                 if (!rc)
 514                         new_ibdev = true;
 515                 else if (rc != -EEXIST)
 516                         goto error;
 517         }
 518         return (new_netdev || new_ibdev) ? 0 : -EEXIST;
 519
 520 error:
 521         return rc;
 522 }
 523
 524 /* Convert an smc_pnetentry to a netlink attribute sequence */
 525 static int smc_pnet_set_nla(struct sk_buff *msg,
 526                             struct smc_pnetentry *pnetelem)
 527 {
 528         if (nla_put_string(msg, SMC_PNETID_NAME, pnetelem->pnet_name))
 529                 return -1;
 530         if (pnetelem->type == SMC_PNET_ETH) {
 531                 if (nla_put_string(msg, SMC_PNETID_ETHNAME,
 532                                    pnetelem->eth_name))
 533                         return -1;
 534         } else {
 535                 if (nla_put_string(msg, SMC_PNETID_ETHNAME, "n/a"))
 536                         return -1;
 537         }
 538         if (pnetelem->type == SMC_PNET_IB) {
 539                 if (nla_put_string(msg, SMC_PNETID_IBNAME, pnetelem->ib_name) ||
 540                     nla_put_u8(msg, SMC_PNETID_IBPORT, pnetelem->ib_port))
 541                         return -1;
 542         } else {
 543                 if (nla_put_string(msg, SMC_PNETID_IBNAME, "n/a") ||
 544                     nla_put_u8(msg, SMC_PNETID_IBPORT, 0xff))
 545                         return -1;
 546         }
 547
 548         return 0;
 549 }
 550
 551 static int smc_pnet_add(struct sk_buff *skb, struct genl_info *info)
 552 {
 553         struct net *net = genl_info_net(info);
 554
 555         return smc_pnet_enter(net, info->attrs);
 556 }
 557
 558 static int smc_pnet_del(struct sk_buff *skb, struct genl_info *info)
 559 {
 560         struct net *net = genl_info_net(info);
 561
 562         if (!info->attrs[SMC_PNETID_NAME])
 563                 return -EINVAL;
 564         return smc_pnet_remove_by_pnetid(net,
 565                                 (char *)nla_data(info->attrs[SMC_PNETID_NAME]));
 566 }
 567
 568 static int smc_pnet_dump_start(struct netlink_callback *cb)
 569 {
 570         cb->args[0] = 0;
 571         return 0;
 572 }
 573
 574 static int smc_pnet_dumpinfo(struct sk_buff *skb,
 575                              u32 portid, u32 seq, u32 flags,
 576                              struct smc_pnetentry *pnetelem)
 577 {
 578         void *hdr;
 579
 580         hdr = genlmsg_put(skb, portid, seq, &smc_pnet_nl_family,
 581                           flags, SMC_PNETID_GET);
 582         if (!hdr)
 583                 return -ENOMEM;
 584         if (smc_pnet_set_nla(skb, pnetelem) < 0) {
 585                 genlmsg_cancel(skb, hdr);
 586                 return -EMSGSIZE;
 587         }
 588         genlmsg_end(skb, hdr);
 589         return 0;
 590 }
 591
 592 static int _smc_pnet_dump(struct net *net, struct sk_buff *skb, u32 portid,
 593                           u32 seq, u8 *pnetid, int start_idx)
 594 {
 595         struct smc_pnettable *pnettable;
 596         struct smc_pnetentry *pnetelem;
 597         struct smc_net *sn;
 598         int idx = 0;
 599
 600         /* get pnettable for namespace */
 601         sn = net_generic(net, smc_net_id);
 602         pnettable = &sn->pnettable;
 603
 604         /* dump pnettable entries */
 605         read_lock(&pnettable->lock);
 606         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 607                 if (pnetid && !smc_pnet_match(pnetelem->pnet_name, pnetid))
 608                         continue;
 609                 if (idx++ < start_idx)
 610                         continue;
 611                 /* if this is not the initial namespace, dump only netdev */
 612                 if (net != &init_net && pnetelem->type != SMC_PNET_ETH)
 613                         continue;
 614                 if (smc_pnet_dumpinfo(skb, portid, seq, NLM_F_MULTI,
 615                                       pnetelem)) {
 616                         --idx;
 617                         break;
 618                 }
 619         }
 620         read_unlock(&pnettable->lock);
 621         return idx;
 622 }
 623
 624 static int smc_pnet_dump(struct sk_buff *skb, struct netlink_callback *cb)
 625 {
 626         struct net *net = sock_net(skb->sk);
 627         int idx;
 628
 629         idx = _smc_pnet_dump(net, skb, NETLINK_CB(cb->skb).portid,
 630                              cb->nlh->nlmsg_seq, NULL, cb->args[0]);
 631
 632         cb->args[0] = idx;
 633         return skb->len;
 634 }
 635
 636 /* Retrieve one PNETID entry */
 637 static int smc_pnet_get(struct sk_buff *skb, struct genl_info *info)
 638 {
 639         struct net *net = genl_info_net(info);
 640         struct sk_buff *msg;
 641         void *hdr;
 642
 643         if (!info->attrs[SMC_PNETID_NAME])
 644                 return -EINVAL;
 645
 646         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 647         if (!msg)
 648                 return -ENOMEM;
 649
 650         _smc_pnet_dump(net, msg, info->snd_portid, info->snd_seq,
 651                        nla_data(info->attrs[SMC_PNETID_NAME]), 0);
 652
 653         /* finish multi part message and send it */
 654         hdr = nlmsg_put(msg, info->snd_portid, info->snd_seq, NLMSG_DONE, 0,
 655                         NLM_F_MULTI);
 656         if (!hdr) {
 657                 nlmsg_free(msg);
 658                 return -EMSGSIZE;
 659         }
 660         return genlmsg_reply(msg, info);
 661 }
 662
 663 /* Remove and delete all pnetids from pnet table.
 664  */
 665 static int smc_pnet_flush(struct sk_buff *skb, struct genl_info *info)
 666 {
 667         struct net *net = genl_info_net(info);
 668
 669         smc_pnet_remove_by_pnetid(net, NULL);
 670         return 0;
 671 }
 672
 673 /* SMC_PNETID generic netlink operation definition */
 674 static const struct genl_ops smc_pnet_ops[] = {
 675         {
 676                 .cmd = SMC_PNETID_GET,
 677                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 678                 /* can be retrieved by unprivileged users */
 679                 .doit = smc_pnet_get,
 680                 .dumpit = smc_pnet_dump,
 681                 .start = smc_pnet_dump_start
 682         },
 683         {
 684                 .cmd = SMC_PNETID_ADD,
 685                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 686                 .flags = GENL_ADMIN_PERM,
 687                 .doit = smc_pnet_add
 688         },
 689         {
 690                 .cmd = SMC_PNETID_DEL,
 691                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 692                 .flags = GENL_ADMIN_PERM,
 693                 .doit = smc_pnet_del
 694         },
 695         {
 696                 .cmd = SMC_PNETID_FLUSH,
 697                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 698                 .flags = GENL_ADMIN_PERM,
 699                 .doit = smc_pnet_flush
 700         }
 701 };
 702
 703 /* SMC_PNETID family definition */
 704 static struct genl_family smc_pnet_nl_family __ro_after_init = {
 705         .hdrsize = 0,
 706         .name = SMCR_GENL_FAMILY_NAME,
 707         .version = SMCR_GENL_FAMILY_VERSION,
 708         .maxattr = SMC_PNETID_MAX,
 709         .policy = smc_pnet_policy,
 710         .netnsok = true,
 711         .module = THIS_MODULE,
 712         .ops = smc_pnet_ops,
 713         .n_ops =  ARRAY_SIZE(smc_pnet_ops)
 714 };
 715
 716 bool smc_pnet_is_ndev_pnetid(struct net *net, u8 *pnetid)
 717 {
 718         struct smc_net *sn = net_generic(net, smc_net_id);
 719         struct smc_pnetids_ndev_entry *pe;
 720         bool rc = false;
 721
 722         read_lock(&sn->pnetids_ndev.lock);
 723         list_for_each_entry(pe, &sn->pnetids_ndev.list, list) {
 724                 if (smc_pnet_match(pnetid, pe->pnetid)) {
 725                         rc = true;
 726                         goto unlock;
 727                 }
 728         }
 729
 730 unlock:
 731         read_unlock(&sn->pnetids_ndev.lock);
 732         return rc;
 733 }
 734
 735 static int smc_pnet_add_pnetid(struct net *net, u8 *pnetid)
 736 {
 737         struct smc_net *sn = net_generic(net, smc_net_id);
 738         struct smc_pnetids_ndev_entry *pe, *pi;
 739
 740         pe = kzalloc(sizeof(*pe), GFP_KERNEL);
 741         if (!pe)
 742                 return -ENOMEM;
 743
 744         write_lock(&sn->pnetids_ndev.lock);
 745         list_for_each_entry(pi, &sn->pnetids_ndev.list, list) {
 746                 if (smc_pnet_match(pnetid, pe->pnetid)) {
 747                         refcount_inc(&pi->refcnt);
 748                         kfree(pe);
 749                         goto unlock;
 750                 }
 751         }
 752         refcount_set(&pe->refcnt, 1);
 753         memcpy(pe->pnetid, pnetid, SMC_MAX_PNETID_LEN);
 754         list_add_tail(&pe->list, &sn->pnetids_ndev.list);
 755
 756 unlock:
 757         write_unlock(&sn->pnetids_ndev.lock);
 758         return 0;
 759 }
 760
 761 static void smc_pnet_remove_pnetid(struct net *net, u8 *pnetid)
 762 {
 763         struct smc_net *sn = net_generic(net, smc_net_id);
 764         struct smc_pnetids_ndev_entry *pe, *pe2;
 765
 766         write_lock(&sn->pnetids_ndev.lock);
 767         list_for_each_entry_safe(pe, pe2, &sn->pnetids_ndev.list, list) {
 768                 if (smc_pnet_match(pnetid, pe->pnetid)) {
 769                         if (refcount_dec_and_test(&pe->refcnt)) {
 770                                 list_del(&pe->list);
 771                                 kfree(pe);
 772                         }
 773                         break;
 774                 }
 775         }
 776         write_unlock(&sn->pnetids_ndev.lock);
 777 }
 778
 779 static void smc_pnet_add_base_pnetid(struct net *net, struct net_device *dev,
 780                                      u8 *ndev_pnetid)
 781 {
 782         struct net_device *base_dev;
 783
 784         base_dev = __pnet_find_base_ndev(dev);
 785         if (base_dev->flags & IFF_UP &&
 786             !smc_pnetid_by_dev_port(base_dev->dev.parent, base_dev->dev_port,
 787                                     ndev_pnetid)) {
 788                 /* add to PNETIDs list */
 789                 smc_pnet_add_pnetid(net, ndev_pnetid);
 790         }
 791 }
 792
 793 /* create initial list of netdevice pnetids */
 794 static void smc_pnet_create_pnetids_list(struct net *net)
 795 {
 796         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 797         struct net_device *dev;
 798
 799         rtnl_lock();
 800         for_each_netdev(net, dev)
 801                 smc_pnet_add_base_pnetid(net, dev, ndev_pnetid);
 802         rtnl_unlock();
 803 }
 804
 805 /* clean up list of netdevice pnetids */
 806 static void smc_pnet_destroy_pnetids_list(struct net *net)
 807 {
 808         struct smc_net *sn = net_generic(net, smc_net_id);
 809         struct smc_pnetids_ndev_entry *pe, *temp_pe;
 810
 811         write_lock(&sn->pnetids_ndev.lock);
 812         list_for_each_entry_safe(pe, temp_pe, &sn->pnetids_ndev.list, list) {
 813                 list_del(&pe->list);
 814                 kfree(pe);
 815         }
 816         write_unlock(&sn->pnetids_ndev.lock);
 817 }
 818
 819 static int smc_pnet_netdev_event(struct notifier_block *this,
 820                                  unsigned long event, void *ptr)
 821 {
 822         struct net_device *event_dev = netdev_notifier_info_to_dev(ptr);
 823         struct net *net = dev_net(event_dev);
 824         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 825
 826         switch (event) {
 827         case NETDEV_REBOOT:
 828         case NETDEV_UNREGISTER:
 829                 smc_pnet_remove_by_ndev(event_dev);
 830                 smc_ib_ndev_change(event_dev, event);
 831                 return NOTIFY_OK;
 832         case NETDEV_REGISTER:
 833                 smc_pnet_add_by_ndev(event_dev);
 834                 smc_ib_ndev_change(event_dev, event);
 835                 return NOTIFY_OK;
 836         case NETDEV_UP:
 837                 smc_pnet_add_base_pnetid(net, event_dev, ndev_pnetid);
 838                 return NOTIFY_OK;
 839         case NETDEV_DOWN:
 840                 event_dev = __pnet_find_base_ndev(event_dev);
 841                 if (!smc_pnetid_by_dev_port(event_dev->dev.parent,
 842                                             event_dev->dev_port, ndev_pnetid)) {
 843                         /* remove from PNETIDs list */
 844                         smc_pnet_remove_pnetid(net, ndev_pnetid);
 845                 }
 846                 return NOTIFY_OK;
 847         default:
 848                 return NOTIFY_DONE;
 849         }
 850 }
 851
 852 static struct notifier_block smc_netdev_notifier = {
 853         .notifier_call = smc_pnet_netdev_event
 854 };
 855
 856 /* init network namespace */
 857 int smc_pnet_net_init(struct net *net)
 858 {
 859         struct smc_net *sn = net_generic(net, smc_net_id);
 860         struct smc_pnettable *pnettable = &sn->pnettable;
 861         struct smc_pnetids_ndev *pnetids_ndev = &sn->pnetids_ndev;
 862
 863         INIT_LIST_HEAD(&pnettable->pnetlist);
 864         rwlock_init(&pnettable->lock);
 865         INIT_LIST_HEAD(&pnetids_ndev->list);
 866         rwlock_init(&pnetids_ndev->lock);
 867
 868         smc_pnet_create_pnetids_list(net);
 869
 870         return 0;
 871 }
 872
 873 int __init smc_pnet_init(void)
 874 {
 875         int rc;
 876
 877         rc = genl_register_family(&smc_pnet_nl_family);
 878         if (rc)
 879                 return rc;
 880         rc = register_netdevice_notifier(&smc_netdev_notifier);
 881         if (rc)
 882                 genl_unregister_family(&smc_pnet_nl_family);
 883
 884         return rc;
 885 }
 886
 887 /* exit network namespace */
 888 void smc_pnet_net_exit(struct net *net)
 889 {
 890         /* flush pnet table */
 891         smc_pnet_remove_by_pnetid(net, NULL);
 892         smc_pnet_destroy_pnetids_list(net);
 893 }
 894
 895 void smc_pnet_exit(void)
 896 {
 897         unregister_netdevice_notifier(&smc_netdev_notifier);
 898         genl_unregister_family(&smc_pnet_nl_family);
 899 }
 900
 901 static struct net_device *__pnet_find_base_ndev(struct net_device *ndev)
 902 {
 903         int i, nest_lvl;
 904
 905         ASSERT_RTNL();
 906         nest_lvl = ndev->lower_level;
 907         for (i = 0; i < nest_lvl; i++) {
 908                 struct list_head *lower = &ndev->adj_list.lower;
 909
 910                 if (list_empty(lower))
 911                         break;
 912                 lower = lower->next;
 913                 ndev = netdev_lower_get_next(ndev, &lower);
 914         }
 915         return ndev;
 916 }
 917
 918 /* Determine one base device for stacked net devices.
 919  * If the lower device level contains more than one devices
 920  * (for instance with bonding slaves), just the first device
 921  * is used to reach a base device.
 922  */
 923 static struct net_device *pnet_find_base_ndev(struct net_device *ndev)
 924 {
 925         rtnl_lock();
 926         ndev = __pnet_find_base_ndev(ndev);
 927         rtnl_unlock();
 928         return ndev;
 929 }
 930
 931 static int smc_pnet_find_ndev_pnetid_by_table(struct net_device *ndev,
 932                                               u8 *pnetid)
 933 {
 934         struct smc_pnettable *pnettable;
 935         struct net *net = dev_net(ndev);
 936         struct smc_pnetentry *pnetelem;
 937         struct smc_net *sn;
 938         int rc = -ENOENT;
 939
 940         /* get pnettable for namespace */
 941         sn = net_generic(net, smc_net_id);
 942         pnettable = &sn->pnettable;
 943
 944         read_lock(&pnettable->lock);
 945         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 946                 if (pnetelem->type == SMC_PNET_ETH && ndev == pnetelem->ndev) {
 947                         /* get pnetid of netdev device */
 948                         memcpy(pnetid, pnetelem->pnet_name, SMC_MAX_PNETID_LEN);
 949                         rc = 0;
 950                         break;
 951                 }
 952         }
 953         read_unlock(&pnettable->lock);
 954         return rc;
 955 }
 956
 957 /* find a roce device for the given pnetid */
 958 static void _smc_pnet_find_roce_by_pnetid(u8 *pnet_id,
 959                                           struct smc_init_info *ini,
 960                                           struct smc_ib_device *known_dev)
 961 {
 962         struct smc_ib_device *ibdev;
 963         int i;
 964
 965         ini->ib_dev = NULL;
 966         mutex_lock(&smc_ib_devices.mutex);
 967         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 968                 if (ibdev == known_dev)
 969                         continue;
 970                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
 971                         if (!rdma_is_port_valid(ibdev->ibdev, i))
 972                                 continue;
 973                         if (smc_pnet_match(ibdev->pnetid[i - 1], pnet_id) &&
 974                             smc_ib_port_active(ibdev, i) &&
 975                             !test_bit(i - 1, ibdev->ports_going_away) &&
 976                             !smc_ib_determine_gid(ibdev, i, ini->vlan_id,
 977                                                   ini->ib_gid, NULL)) {
 978                                 ini->ib_dev = ibdev;
 979                                 ini->ib_port = i;
 980                                 goto out;
 981                         }
 982                 }
 983         }
 984 out:
 985         mutex_unlock(&smc_ib_devices.mutex);
 986 }
 987
 988 /* find alternate roce device with same pnet_id and vlan_id */
 989 void smc_pnet_find_alt_roce(struct smc_link_group *lgr,
 990                             struct smc_init_info *ini,
 991                             struct smc_ib_device *known_dev)
 992 {
 993         _smc_pnet_find_roce_by_pnetid(lgr->pnet_id, ini, known_dev);
 994 }
 995
 996 /* if handshake network device belongs to a roce device, return its
 997  * IB device and port
 998  */
 999 static void smc_pnet_find_rdma_dev(struct net_device *netdev,
1000                                    struct smc_init_info *ini)
1001 {
1002         struct smc_ib_device *ibdev;
1003
1004         mutex_lock(&smc_ib_devices.mutex);
1005         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
1006                 struct net_device *ndev;
1007                 int i;
1008
1009                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
1010                         if (!rdma_is_port_valid(ibdev->ibdev, i))
1011                                 continue;
1012                         if (!ibdev->ibdev->ops.get_netdev)
1013                                 continue;
1014                         ndev = ibdev->ibdev->ops.get_netdev(ibdev->ibdev, i);
1015                         if (!ndev)
1016                                 continue;
1017                         dev_put(ndev);
1018                         if (netdev == ndev &&
1019                             smc_ib_port_active(ibdev, i) &&
1020                             !test_bit(i - 1, ibdev->ports_going_away) &&
1021                             !smc_ib_determine_gid(ibdev, i, ini->vlan_id,
1022                                                   ini->ib_gid, NULL)) {
1023                                 ini->ib_dev = ibdev;
1024                                 ini->ib_port = i;
1025                                 break;
1026                         }
1027                 }
1028         }
1029         mutex_unlock(&smc_ib_devices.mutex);
1030 }
1031
1032 /* Determine the corresponding IB device port based on the hardware PNETID.
1033  * Searching stops at the first matching active IB device port with vlan_id
1034  * configured.
1035  * If nothing found, check pnetid table.
1036  * If nothing found, try to use handshake device
1037  */
1038 static void smc_pnet_find_roce_by_pnetid(struct net_device *ndev,
1039                                          struct smc_init_info *ini)
1040 {
1041         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
1042
1043         ndev = pnet_find_base_ndev(ndev);
1044         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
1045                                    ndev_pnetid) &&
1046             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid)) {
1047                 smc_pnet_find_rdma_dev(ndev, ini);
1048                 return; /* pnetid could not be determined */
1049         }
1050         _smc_pnet_find_roce_by_pnetid(ndev_pnetid, ini, NULL);
1051 }
1052
1053 static void smc_pnet_find_ism_by_pnetid(struct net_device *ndev,
1054                                         struct smc_init_info *ini)
1055 {
1056         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
1057         struct smcd_dev *ismdev;
1058
1059         ndev = pnet_find_base_ndev(ndev);
1060         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
1061                                    ndev_pnetid) &&
1062             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid))
1063                 return; /* pnetid could not be determined */
1064
1065         mutex_lock(&smcd_dev_list.mutex);
1066         list_for_each_entry(ismdev, &smcd_dev_list.list, list) {
1067                 if (smc_pnet_match(ismdev->pnetid, ndev_pnetid) &&
1068                     !ismdev->going_away &&
1069                     (!ini->ism_peer_gid[0] ||
1070                      !smc_ism_cantalk(ini->ism_peer_gid[0], ini->vlan_id,
1071                                       ismdev))) {
1072                         ini->ism_dev[0] = ismdev;
1073                         break;
1074                 }
1075         }
1076         mutex_unlock(&smcd_dev_list.mutex);
1077 }
1078
1079 /* PNET table analysis for a given sock:
1080  * determine ib_device and port belonging to used internal TCP socket
1081  * ethernet interface.
1082  */
1083 void smc_pnet_find_roce_resource(struct sock *sk, struct smc_init_info *ini)
1084 {
1085         struct dst_entry *dst = sk_dst_get(sk);
1086
1087         ini->ib_dev = NULL;
1088         ini->ib_port = 0;
1089         if (!dst)
1090                 goto out;
1091         if (!dst->dev)
1092                 goto out_rel;
1093
1094         smc_pnet_find_roce_by_pnetid(dst->dev, ini);
1095
1096 out_rel:
1097         dst_release(dst);
1098 out:
1099         return;
1100 }
1101
1102 void smc_pnet_find_ism_resource(struct sock *sk, struct smc_init_info *ini)
1103 {
1104         struct dst_entry *dst = sk_dst_get(sk);
1105
1106         ini->ism_dev[0] = NULL;
1107         if (!dst)
1108                 goto out;
1109         if (!dst->dev)
1110                 goto out_rel;
1111
1112         smc_pnet_find_ism_by_pnetid(dst->dev, ini);
1113
1114 out_rel:
1115         dst_release(dst);
1116 out:
1117         return;
1118 }
1119
1120 /* Lookup and apply a pnet table entry to the given ib device.
1121  */
1122 int smc_pnetid_by_table_ib(struct smc_ib_device *smcibdev, u8 ib_port)
1123 {
1124         char *ib_name = smcibdev->ibdev->name;
1125         struct smc_pnettable *pnettable;
1126         struct smc_pnetentry *tmp_pe;
1127         struct smc_net *sn;
1128         int rc = -ENOENT;
1129
1130         /* get pnettable for init namespace */
1131         sn = net_generic(&init_net, smc_net_id);
1132         pnettable = &sn->pnettable;
1133
1134         read_lock(&pnettable->lock);
1135         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
1136                 if (tmp_pe->type == SMC_PNET_IB &&
1137                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX) &&
1138                     tmp_pe->ib_port == ib_port) {
1139                         smc_pnet_apply_ib(smcibdev, ib_port, tmp_pe->pnet_name);
1140                         rc = 0;
1141                         break;
1142                 }
1143         }
1144         read_unlock(&pnettable->lock);
1145
1146         return rc;
1147 }
1148
1149 /* Lookup and apply a pnet table entry to the given smcd device.
1150  */
1151 int smc_pnetid_by_table_smcd(struct smcd_dev *smcddev)
1152 {
1153         const char *ib_name = dev_name(&smcddev->dev);
1154         struct smc_pnettable *pnettable;
1155         struct smc_pnetentry *tmp_pe;
1156         struct smc_net *sn;
1157         int rc = -ENOENT;
1158
1159         /* get pnettable for init namespace */
1160         sn = net_generic(&init_net, smc_net_id);
1161         pnettable = &sn->pnettable;
1162
1163         read_lock(&pnettable->lock);
1164         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
1165                 if (tmp_pe->type == SMC_PNET_IB &&
1166                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
1167                         smc_pnet_apply_smcd(smcddev, tmp_pe->pnet_name);
1168                         rc = 0;
1169                         break;
1170                 }
1171         }
1172         read_unlock(&pnettable->lock);
1173
1174         return rc;
1175 }