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         dev_put(ndev);
 398         return rc;
 399 }
 400
 401 static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
 402                            u8 ib_port, char *pnet_name)
 403 {
 404         struct smc_pnetentry *tmp_pe, *new_pe;
 405         struct smc_ib_device *ib_dev;
 406         bool smcddev_applied = true;
 407         bool ibdev_applied = true;
 408         struct smcd_dev *smcd_dev;
 409         bool new_ibdev;
 410
 411         /* try to apply the pnetid to active devices */
 412         ib_dev = smc_pnet_find_ib(ib_name);
 413         if (ib_dev) {
 414                 ibdev_applied = smc_pnet_apply_ib(ib_dev, ib_port, pnet_name);
 415                 if (ibdev_applied)
 416                         pr_warn_ratelimited("smc: ib device %s ibport %d "
 417                                             "applied user defined pnetid "
 418                                             "%.16s\n", ib_dev->ibdev->name,
 419                                             ib_port,
 420                                             ib_dev->pnetid[ib_port - 1]);
 421         }
 422         smcd_dev = smc_pnet_find_smcd(ib_name);
 423         if (smcd_dev) {
 424                 smcddev_applied = smc_pnet_apply_smcd(smcd_dev, pnet_name);
 425                 if (smcddev_applied)
 426                         pr_warn_ratelimited("smc: smcd device %s "
 427                                             "applied user defined pnetid "
 428                                             "%.16s\n", dev_name(&smcd_dev->dev),
 429                                             smcd_dev->pnetid);
 430         }
 431         /* Apply fails when a device has a hardware-defined pnetid set, do not
 432          * add a pnet table entry in that case.
 433          */
 434         if (!ibdev_applied || !smcddev_applied)
 435                 return -EEXIST;
 436
 437         /* add a new ib entry to the pnet table if there isn't one */
 438         new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
 439         if (!new_pe)
 440                 return -ENOMEM;
 441         new_pe->type = SMC_PNET_IB;
 442         memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
 443         strncpy(new_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX);
 444         new_pe->ib_port = ib_port;
 445
 446         new_ibdev = true;
 447         write_lock(&pnettable->lock);
 448         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
 449                 if (tmp_pe->type == SMC_PNET_IB &&
 450                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
 451                         new_ibdev = false;
 452                         break;
 453                 }
 454         }
 455         if (new_ibdev) {
 456                 list_add_tail(&new_pe->list, &pnettable->pnetlist);
 457                 write_unlock(&pnettable->lock);
 458         } else {
 459                 write_unlock(&pnettable->lock);
 460                 kfree(new_pe);
 461         }
 462         return (new_ibdev) ? 0 : -EEXIST;
 463 }
 464
 465 /* Append a pnetid to the end of the pnet table if not already on this list.
 466  */
 467 static int smc_pnet_enter(struct net *net, struct nlattr *tb[])
 468 {
 469         char pnet_name[SMC_MAX_PNETID_LEN + 1];
 470         struct smc_pnettable *pnettable;
 471         bool new_netdev = false;
 472         bool new_ibdev = false;
 473         struct smc_net *sn;
 474         u8 ibport = 1;
 475         char *string;
 476         int rc;
 477
 478         /* get pnettable for namespace */
 479         sn = net_generic(net, smc_net_id);
 480         pnettable = &sn->pnettable;
 481
 482         rc = -EINVAL;
 483         if (!tb[SMC_PNETID_NAME])
 484                 goto error;
 485         string = (char *)nla_data(tb[SMC_PNETID_NAME]);
 486         if (!smc_pnetid_valid(string, pnet_name))
 487                 goto error;
 488
 489         if (tb[SMC_PNETID_ETHNAME]) {
 490                 string = (char *)nla_data(tb[SMC_PNETID_ETHNAME]);
 491                 rc = smc_pnet_add_eth(pnettable, net, string, pnet_name);
 492                 if (!rc)
 493                         new_netdev = true;
 494                 else if (rc != -EEXIST)
 495                         goto error;
 496         }
 497
 498         /* if this is not the initial namespace, stop here */
 499         if (net != &init_net)
 500                 return new_netdev ? 0 : -EEXIST;
 501
 502         rc = -EINVAL;
 503         if (tb[SMC_PNETID_IBNAME]) {
 504                 string = (char *)nla_data(tb[SMC_PNETID_IBNAME]);
 505                 string = strim(string);
 506                 if (tb[SMC_PNETID_IBPORT]) {
 507                         ibport = nla_get_u8(tb[SMC_PNETID_IBPORT]);
 508                         if (ibport < 1 || ibport > SMC_MAX_PORTS)
 509                                 goto error;
 510                 }
 511                 rc = smc_pnet_add_ib(pnettable, string, ibport, pnet_name);
 512                 if (!rc)
 513                         new_ibdev = true;
 514                 else if (rc != -EEXIST)
 515                         goto error;
 516         }
 517         return (new_netdev || new_ibdev) ? 0 : -EEXIST;
 518
 519 error:
 520         return rc;
 521 }
 522
 523 /* Convert an smc_pnetentry to a netlink attribute sequence */
 524 static int smc_pnet_set_nla(struct sk_buff *msg,
 525                             struct smc_pnetentry *pnetelem)
 526 {
 527         if (nla_put_string(msg, SMC_PNETID_NAME, pnetelem->pnet_name))
 528                 return -1;
 529         if (pnetelem->type == SMC_PNET_ETH) {
 530                 if (nla_put_string(msg, SMC_PNETID_ETHNAME,
 531                                    pnetelem->eth_name))
 532                         return -1;
 533         } else {
 534                 if (nla_put_string(msg, SMC_PNETID_ETHNAME, "n/a"))
 535                         return -1;
 536         }
 537         if (pnetelem->type == SMC_PNET_IB) {
 538                 if (nla_put_string(msg, SMC_PNETID_IBNAME, pnetelem->ib_name) ||
 539                     nla_put_u8(msg, SMC_PNETID_IBPORT, pnetelem->ib_port))
 540                         return -1;
 541         } else {
 542                 if (nla_put_string(msg, SMC_PNETID_IBNAME, "n/a") ||
 543                     nla_put_u8(msg, SMC_PNETID_IBPORT, 0xff))
 544                         return -1;
 545         }
 546
 547         return 0;
 548 }
 549
 550 static int smc_pnet_add(struct sk_buff *skb, struct genl_info *info)
 551 {
 552         struct net *net = genl_info_net(info);
 553
 554         return smc_pnet_enter(net, info->attrs);
 555 }
 556
 557 static int smc_pnet_del(struct sk_buff *skb, struct genl_info *info)
 558 {
 559         struct net *net = genl_info_net(info);
 560
 561         if (!info->attrs[SMC_PNETID_NAME])
 562                 return -EINVAL;
 563         return smc_pnet_remove_by_pnetid(net,
 564                                 (char *)nla_data(info->attrs[SMC_PNETID_NAME]));
 565 }
 566
 567 static int smc_pnet_dump_start(struct netlink_callback *cb)
 568 {
 569         cb->args[0] = 0;
 570         return 0;
 571 }
 572
 573 static int smc_pnet_dumpinfo(struct sk_buff *skb,
 574                              u32 portid, u32 seq, u32 flags,
 575                              struct smc_pnetentry *pnetelem)
 576 {
 577         void *hdr;
 578
 579         hdr = genlmsg_put(skb, portid, seq, &smc_pnet_nl_family,
 580                           flags, SMC_PNETID_GET);
 581         if (!hdr)
 582                 return -ENOMEM;
 583         if (smc_pnet_set_nla(skb, pnetelem) < 0) {
 584                 genlmsg_cancel(skb, hdr);
 585                 return -EMSGSIZE;
 586         }
 587         genlmsg_end(skb, hdr);
 588         return 0;
 589 }
 590
 591 static int _smc_pnet_dump(struct net *net, struct sk_buff *skb, u32 portid,
 592                           u32 seq, u8 *pnetid, int start_idx)
 593 {
 594         struct smc_pnettable *pnettable;
 595         struct smc_pnetentry *pnetelem;
 596         struct smc_net *sn;
 597         int idx = 0;
 598
 599         /* get pnettable for namespace */
 600         sn = net_generic(net, smc_net_id);
 601         pnettable = &sn->pnettable;
 602
 603         /* dump pnettable entries */
 604         read_lock(&pnettable->lock);
 605         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 606                 if (pnetid && !smc_pnet_match(pnetelem->pnet_name, pnetid))
 607                         continue;
 608                 if (idx++ < start_idx)
 609                         continue;
 610                 /* if this is not the initial namespace, dump only netdev */
 611                 if (net != &init_net && pnetelem->type != SMC_PNET_ETH)
 612                         continue;
 613                 if (smc_pnet_dumpinfo(skb, portid, seq, NLM_F_MULTI,
 614                                       pnetelem)) {
 615                         --idx;
 616                         break;
 617                 }
 618         }
 619         read_unlock(&pnettable->lock);
 620         return idx;
 621 }
 622
 623 static int smc_pnet_dump(struct sk_buff *skb, struct netlink_callback *cb)
 624 {
 625         struct net *net = sock_net(skb->sk);
 626         int idx;
 627
 628         idx = _smc_pnet_dump(net, skb, NETLINK_CB(cb->skb).portid,
 629                              cb->nlh->nlmsg_seq, NULL, cb->args[0]);
 630
 631         cb->args[0] = idx;
 632         return skb->len;
 633 }
 634
 635 /* Retrieve one PNETID entry */
 636 static int smc_pnet_get(struct sk_buff *skb, struct genl_info *info)
 637 {
 638         struct net *net = genl_info_net(info);
 639         struct sk_buff *msg;
 640         void *hdr;
 641
 642         if (!info->attrs[SMC_PNETID_NAME])
 643                 return -EINVAL;
 644
 645         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 646         if (!msg)
 647                 return -ENOMEM;
 648
 649         _smc_pnet_dump(net, msg, info->snd_portid, info->snd_seq,
 650                        nla_data(info->attrs[SMC_PNETID_NAME]), 0);
 651
 652         /* finish multi part message and send it */
 653         hdr = nlmsg_put(msg, info->snd_portid, info->snd_seq, NLMSG_DONE, 0,
 654                         NLM_F_MULTI);
 655         if (!hdr) {
 656                 nlmsg_free(msg);
 657                 return -EMSGSIZE;
 658         }
 659         return genlmsg_reply(msg, info);
 660 }
 661
 662 /* Remove and delete all pnetids from pnet table.
 663  */
 664 static int smc_pnet_flush(struct sk_buff *skb, struct genl_info *info)
 665 {
 666         struct net *net = genl_info_net(info);
 667
 668         smc_pnet_remove_by_pnetid(net, NULL);
 669         return 0;
 670 }
 671
 672 /* SMC_PNETID generic netlink operation definition */
 673 static const struct genl_ops smc_pnet_ops[] = {
 674         {
 675                 .cmd = SMC_PNETID_GET,
 676                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 677                 /* can be retrieved by unprivileged users */
 678                 .doit = smc_pnet_get,
 679                 .dumpit = smc_pnet_dump,
 680                 .start = smc_pnet_dump_start
 681         },
 682         {
 683                 .cmd = SMC_PNETID_ADD,
 684                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 685                 .flags = GENL_ADMIN_PERM,
 686                 .doit = smc_pnet_add
 687         },
 688         {
 689                 .cmd = SMC_PNETID_DEL,
 690                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 691                 .flags = GENL_ADMIN_PERM,
 692                 .doit = smc_pnet_del
 693         },
 694         {
 695                 .cmd = SMC_PNETID_FLUSH,
 696                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 697                 .flags = GENL_ADMIN_PERM,
 698                 .doit = smc_pnet_flush
 699         }
 700 };
 701
 702 /* SMC_PNETID family definition */
 703 static struct genl_family smc_pnet_nl_family __ro_after_init = {
 704         .hdrsize = 0,
 705         .name = SMCR_GENL_FAMILY_NAME,
 706         .version = SMCR_GENL_FAMILY_VERSION,
 707         .maxattr = SMC_PNETID_MAX,
 708         .policy = smc_pnet_policy,
 709         .netnsok = true,
 710         .module = THIS_MODULE,
 711         .ops = smc_pnet_ops,
 712         .n_ops =  ARRAY_SIZE(smc_pnet_ops)
 713 };
 714
 715 bool smc_pnet_is_ndev_pnetid(struct net *net, u8 *pnetid)
 716 {
 717         struct smc_net *sn = net_generic(net, smc_net_id);
 718         struct smc_pnetids_ndev_entry *pe;
 719         bool rc = false;
 720
 721         read_lock(&sn->pnetids_ndev.lock);
 722         list_for_each_entry(pe, &sn->pnetids_ndev.list, list) {
 723                 if (smc_pnet_match(pnetid, pe->pnetid)) {
 724                         rc = true;
 725                         goto unlock;
 726                 }
 727         }
 728
 729 unlock:
 730         read_unlock(&sn->pnetids_ndev.lock);
 731         return rc;
 732 }
 733
 734 static int smc_pnet_add_pnetid(struct net *net, u8 *pnetid)
 735 {
 736         struct smc_net *sn = net_generic(net, smc_net_id);
 737         struct smc_pnetids_ndev_entry *pe, *pi;
 738
 739         pe = kzalloc(sizeof(*pe), GFP_KERNEL);
 740         if (!pe)
 741                 return -ENOMEM;
 742
 743         write_lock(&sn->pnetids_ndev.lock);
 744         list_for_each_entry(pi, &sn->pnetids_ndev.list, list) {
 745                 if (smc_pnet_match(pnetid, pe->pnetid)) {
 746                         refcount_inc(&pi->refcnt);
 747                         kfree(pe);
 748                         goto unlock;
 749                 }
 750         }
 751         refcount_set(&pe->refcnt, 1);
 752         memcpy(pe->pnetid, pnetid, SMC_MAX_PNETID_LEN);
 753         list_add_tail(&pe->list, &sn->pnetids_ndev.list);
 754
 755 unlock:
 756         write_unlock(&sn->pnetids_ndev.lock);
 757         return 0;
 758 }
 759
 760 static void smc_pnet_remove_pnetid(struct net *net, u8 *pnetid)
 761 {
 762         struct smc_net *sn = net_generic(net, smc_net_id);
 763         struct smc_pnetids_ndev_entry *pe, *pe2;
 764
 765         write_lock(&sn->pnetids_ndev.lock);
 766         list_for_each_entry_safe(pe, pe2, &sn->pnetids_ndev.list, list) {
 767                 if (smc_pnet_match(pnetid, pe->pnetid)) {
 768                         if (refcount_dec_and_test(&pe->refcnt)) {
 769                                 list_del(&pe->list);
 770                                 kfree(pe);
 771                         }
 772                         break;
 773                 }
 774         }
 775         write_unlock(&sn->pnetids_ndev.lock);
 776 }
 777
 778 static void smc_pnet_add_base_pnetid(struct net *net, struct net_device *dev,
 779                                      u8 *ndev_pnetid)
 780 {
 781         struct net_device *base_dev;
 782
 783         base_dev = __pnet_find_base_ndev(dev);
 784         if (base_dev->flags & IFF_UP &&
 785             !smc_pnetid_by_dev_port(base_dev->dev.parent, base_dev->dev_port,
 786                                     ndev_pnetid)) {
 787                 /* add to PNETIDs list */
 788                 smc_pnet_add_pnetid(net, ndev_pnetid);
 789         }
 790 }
 791
 792 /* create initial list of netdevice pnetids */
 793 static void smc_pnet_create_pnetids_list(struct net *net)
 794 {
 795         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 796         struct net_device *dev;
 797
 798         rtnl_lock();
 799         for_each_netdev(net, dev)
 800                 smc_pnet_add_base_pnetid(net, dev, ndev_pnetid);
 801         rtnl_unlock();
 802 }
 803
 804 /* clean up list of netdevice pnetids */
 805 static void smc_pnet_destroy_pnetids_list(struct net *net)
 806 {
 807         struct smc_net *sn = net_generic(net, smc_net_id);
 808         struct smc_pnetids_ndev_entry *pe, *temp_pe;
 809
 810         write_lock(&sn->pnetids_ndev.lock);
 811         list_for_each_entry_safe(pe, temp_pe, &sn->pnetids_ndev.list, list) {
 812                 list_del(&pe->list);
 813                 kfree(pe);
 814         }
 815         write_unlock(&sn->pnetids_ndev.lock);
 816 }
 817
 818 static int smc_pnet_netdev_event(struct notifier_block *this,
 819                                  unsigned long event, void *ptr)
 820 {
 821         struct net_device *event_dev = netdev_notifier_info_to_dev(ptr);
 822         struct net *net = dev_net(event_dev);
 823         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 824
 825         switch (event) {
 826         case NETDEV_REBOOT:
 827         case NETDEV_UNREGISTER:
 828                 smc_pnet_remove_by_ndev(event_dev);
 829                 smc_ib_ndev_change(event_dev, event);
 830                 return NOTIFY_OK;
 831         case NETDEV_REGISTER:
 832                 smc_pnet_add_by_ndev(event_dev);
 833                 smc_ib_ndev_change(event_dev, event);
 834                 return NOTIFY_OK;
 835         case NETDEV_UP:
 836                 smc_pnet_add_base_pnetid(net, event_dev, ndev_pnetid);
 837                 return NOTIFY_OK;
 838         case NETDEV_DOWN:
 839                 event_dev = __pnet_find_base_ndev(event_dev);
 840                 if (!smc_pnetid_by_dev_port(event_dev->dev.parent,
 841                                             event_dev->dev_port, ndev_pnetid)) {
 842                         /* remove from PNETIDs list */
 843                         smc_pnet_remove_pnetid(net, ndev_pnetid);
 844                 }
 845                 return NOTIFY_OK;
 846         default:
 847                 return NOTIFY_DONE;
 848         }
 849 }
 850
 851 static struct notifier_block smc_netdev_notifier = {
 852         .notifier_call = smc_pnet_netdev_event
 853 };
 854
 855 /* init network namespace */
 856 int smc_pnet_net_init(struct net *net)
 857 {
 858         struct smc_net *sn = net_generic(net, smc_net_id);
 859         struct smc_pnettable *pnettable = &sn->pnettable;
 860         struct smc_pnetids_ndev *pnetids_ndev = &sn->pnetids_ndev;
 861
 862         INIT_LIST_HEAD(&pnettable->pnetlist);
 863         rwlock_init(&pnettable->lock);
 864         INIT_LIST_HEAD(&pnetids_ndev->list);
 865         rwlock_init(&pnetids_ndev->lock);
 866
 867         smc_pnet_create_pnetids_list(net);
 868
 869         return 0;
 870 }
 871
 872 int __init smc_pnet_init(void)
 873 {
 874         int rc;
 875
 876         rc = genl_register_family(&smc_pnet_nl_family);
 877         if (rc)
 878                 return rc;
 879         rc = register_netdevice_notifier(&smc_netdev_notifier);
 880         if (rc)
 881                 genl_unregister_family(&smc_pnet_nl_family);
 882
 883         return rc;
 884 }
 885
 886 /* exit network namespace */
 887 void smc_pnet_net_exit(struct net *net)
 888 {
 889         /* flush pnet table */
 890         smc_pnet_remove_by_pnetid(net, NULL);
 891         smc_pnet_destroy_pnetids_list(net);
 892 }
 893
 894 void smc_pnet_exit(void)
 895 {
 896         unregister_netdevice_notifier(&smc_netdev_notifier);
 897         genl_unregister_family(&smc_pnet_nl_family);
 898 }
 899
 900 static struct net_device *__pnet_find_base_ndev(struct net_device *ndev)
 901 {
 902         int i, nest_lvl;
 903
 904         ASSERT_RTNL();
 905         nest_lvl = ndev->lower_level;
 906         for (i = 0; i < nest_lvl; i++) {
 907                 struct list_head *lower = &ndev->adj_list.lower;
 908
 909                 if (list_empty(lower))
 910                         break;
 911                 lower = lower->next;
 912                 ndev = netdev_lower_get_next(ndev, &lower);
 913         }
 914         return ndev;
 915 }
 916
 917 /* Determine one base device for stacked net devices.
 918  * If the lower device level contains more than one devices
 919  * (for instance with bonding slaves), just the first device
 920  * is used to reach a base device.
 921  */
 922 static struct net_device *pnet_find_base_ndev(struct net_device *ndev)
 923 {
 924         rtnl_lock();
 925         ndev = __pnet_find_base_ndev(ndev);
 926         rtnl_unlock();
 927         return ndev;
 928 }
 929
 930 static int smc_pnet_find_ndev_pnetid_by_table(struct net_device *ndev,
 931                                               u8 *pnetid)
 932 {
 933         struct smc_pnettable *pnettable;
 934         struct net *net = dev_net(ndev);
 935         struct smc_pnetentry *pnetelem;
 936         struct smc_net *sn;
 937         int rc = -ENOENT;
 938
 939         /* get pnettable for namespace */
 940         sn = net_generic(net, smc_net_id);
 941         pnettable = &sn->pnettable;
 942
 943         read_lock(&pnettable->lock);
 944         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 945                 if (pnetelem->type == SMC_PNET_ETH && ndev == pnetelem->ndev) {
 946                         /* get pnetid of netdev device */
 947                         memcpy(pnetid, pnetelem->pnet_name, SMC_MAX_PNETID_LEN);
 948                         rc = 0;
 949                         break;
 950                 }
 951         }
 952         read_unlock(&pnettable->lock);
 953         return rc;
 954 }
 955
 956 static int smc_pnet_determine_gid(struct smc_ib_device *ibdev, int i,
 957                                   struct smc_init_info *ini)
 958 {
 959         if (!ini->check_smcrv2 &&
 960             !smc_ib_determine_gid(ibdev, i, ini->vlan_id, ini->ib_gid, NULL,
 961                                   NULL)) {
 962                 ini->ib_dev = ibdev;
 963                 ini->ib_port = i;
 964                 return 0;
 965         }
 966         if (ini->check_smcrv2 &&
 967             !smc_ib_determine_gid(ibdev, i, ini->vlan_id, ini->smcrv2.ib_gid_v2,
 968                                   NULL, &ini->smcrv2)) {
 969                 ini->smcrv2.ib_dev_v2 = ibdev;
 970                 ini->smcrv2.ib_port_v2 = i;
 971                 return 0;
 972         }
 973         return -ENODEV;
 974 }
 975
 976 /* find a roce device for the given pnetid */
 977 static void _smc_pnet_find_roce_by_pnetid(u8 *pnet_id,
 978                                           struct smc_init_info *ini,
 979                                           struct smc_ib_device *known_dev)
 980 {
 981         struct smc_ib_device *ibdev;
 982         int i;
 983
 984         mutex_lock(&smc_ib_devices.mutex);
 985         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 986                 if (ibdev == known_dev)
 987                         continue;
 988                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
 989                         if (!rdma_is_port_valid(ibdev->ibdev, i))
 990                                 continue;
 991                         if (smc_pnet_match(ibdev->pnetid[i - 1], pnet_id) &&
 992                             smc_ib_port_active(ibdev, i) &&
 993                             !test_bit(i - 1, ibdev->ports_going_away)) {
 994                                 if (!smc_pnet_determine_gid(ibdev, i, ini))
 995                                         goto out;
 996                         }
 997                 }
 998         }
 999 out:
1000         mutex_unlock(&smc_ib_devices.mutex);
1001 }
1002
1003 /* find alternate roce device with same pnet_id and vlan_id */
1004 void smc_pnet_find_alt_roce(struct smc_link_group *lgr,
1005                             struct smc_init_info *ini,
1006                             struct smc_ib_device *known_dev)
1007 {
1008         _smc_pnet_find_roce_by_pnetid(lgr->pnet_id, ini, known_dev);
1009 }
1010
1011 /* if handshake network device belongs to a roce device, return its
1012  * IB device and port
1013  */
1014 static void smc_pnet_find_rdma_dev(struct net_device *netdev,
1015                                    struct smc_init_info *ini)
1016 {
1017         struct smc_ib_device *ibdev;
1018
1019         mutex_lock(&smc_ib_devices.mutex);
1020         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
1021                 struct net_device *ndev;
1022                 int i;
1023
1024                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
1025                         if (!rdma_is_port_valid(ibdev->ibdev, i))
1026                                 continue;
1027                         if (!ibdev->ibdev->ops.get_netdev)
1028                                 continue;
1029                         ndev = ibdev->ibdev->ops.get_netdev(ibdev->ibdev, i);
1030                         if (!ndev)
1031                                 continue;
1032                         dev_put(ndev);
1033                         if (netdev == ndev &&
1034                             smc_ib_port_active(ibdev, i) &&
1035                             !test_bit(i - 1, ibdev->ports_going_away)) {
1036                                 if (!smc_pnet_determine_gid(ibdev, i, ini))
1037                                         break;
1038                         }
1039                 }
1040         }
1041         mutex_unlock(&smc_ib_devices.mutex);
1042 }
1043
1044 /* Determine the corresponding IB device port based on the hardware PNETID.
1045  * Searching stops at the first matching active IB device port with vlan_id
1046  * configured.
1047  * If nothing found, check pnetid table.
1048  * If nothing found, try to use handshake device
1049  */
1050 static void smc_pnet_find_roce_by_pnetid(struct net_device *ndev,
1051                                          struct smc_init_info *ini)
1052 {
1053         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
1054
1055         ndev = pnet_find_base_ndev(ndev);
1056         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
1057                                    ndev_pnetid) &&
1058             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid)) {
1059                 smc_pnet_find_rdma_dev(ndev, ini);
1060                 return; /* pnetid could not be determined */
1061         }
1062         _smc_pnet_find_roce_by_pnetid(ndev_pnetid, ini, NULL);
1063 }
1064
1065 static void smc_pnet_find_ism_by_pnetid(struct net_device *ndev,
1066                                         struct smc_init_info *ini)
1067 {
1068         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
1069         struct smcd_dev *ismdev;
1070
1071         ndev = pnet_find_base_ndev(ndev);
1072         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
1073                                    ndev_pnetid) &&
1074             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid))
1075                 return; /* pnetid could not be determined */
1076
1077         mutex_lock(&smcd_dev_list.mutex);
1078         list_for_each_entry(ismdev, &smcd_dev_list.list, list) {
1079                 if (smc_pnet_match(ismdev->pnetid, ndev_pnetid) &&
1080                     !ismdev->going_away &&
1081                     (!ini->ism_peer_gid[0] ||
1082                      !smc_ism_cantalk(ini->ism_peer_gid[0], ini->vlan_id,
1083                                       ismdev))) {
1084                         ini->ism_dev[0] = ismdev;
1085                         break;
1086                 }
1087         }
1088         mutex_unlock(&smcd_dev_list.mutex);
1089 }
1090
1091 /* PNET table analysis for a given sock:
1092  * determine ib_device and port belonging to used internal TCP socket
1093  * ethernet interface.
1094  */
1095 void smc_pnet_find_roce_resource(struct sock *sk, struct smc_init_info *ini)
1096 {
1097         struct dst_entry *dst = sk_dst_get(sk);
1098
1099         if (!dst)
1100                 goto out;
1101         if (!dst->dev)
1102                 goto out_rel;
1103
1104         smc_pnet_find_roce_by_pnetid(dst->dev, ini);
1105
1106 out_rel:
1107         dst_release(dst);
1108 out:
1109         return;
1110 }
1111
1112 void smc_pnet_find_ism_resource(struct sock *sk, struct smc_init_info *ini)
1113 {
1114         struct dst_entry *dst = sk_dst_get(sk);
1115
1116         ini->ism_dev[0] = NULL;
1117         if (!dst)
1118                 goto out;
1119         if (!dst->dev)
1120                 goto out_rel;
1121
1122         smc_pnet_find_ism_by_pnetid(dst->dev, ini);
1123
1124 out_rel:
1125         dst_release(dst);
1126 out:
1127         return;
1128 }
1129
1130 /* Lookup and apply a pnet table entry to the given ib device.
1131  */
1132 int smc_pnetid_by_table_ib(struct smc_ib_device *smcibdev, u8 ib_port)
1133 {
1134         char *ib_name = smcibdev->ibdev->name;
1135         struct smc_pnettable *pnettable;
1136         struct smc_pnetentry *tmp_pe;
1137         struct smc_net *sn;
1138         int rc = -ENOENT;
1139
1140         /* get pnettable for init namespace */
1141         sn = net_generic(&init_net, smc_net_id);
1142         pnettable = &sn->pnettable;
1143
1144         read_lock(&pnettable->lock);
1145         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
1146                 if (tmp_pe->type == SMC_PNET_IB &&
1147                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX) &&
1148                     tmp_pe->ib_port == ib_port) {
1149                         smc_pnet_apply_ib(smcibdev, ib_port, tmp_pe->pnet_name);
1150                         rc = 0;
1151                         break;
1152                 }
1153         }
1154         read_unlock(&pnettable->lock);
1155
1156         return rc;
1157 }
1158
1159 /* Lookup and apply a pnet table entry to the given smcd device.
1160  */
1161 int smc_pnetid_by_table_smcd(struct smcd_dev *smcddev)
1162 {
1163         const char *ib_name = dev_name(&smcddev->dev);
1164         struct smc_pnettable *pnettable;
1165         struct smc_pnetentry *tmp_pe;
1166         struct smc_net *sn;
1167         int rc = -ENOENT;
1168
1169         /* get pnettable for init namespace */
1170         sn = net_generic(&init_net, smc_net_id);
1171         pnettable = &sn->pnettable;
1172
1173         read_lock(&pnettable->lock);
1174         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
1175                 if (tmp_pe->type == SMC_PNET_IB &&
1176                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
1177                         smc_pnet_apply_smcd(smcddev, tmp_pe->pnet_name);
1178                         rc = 0;
1179                         break;
1180                 }
1181         }
1182         read_unlock(&pnettable->lock);
1183
1184         return rc;
1185 }