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 <net/netlink.h>
  16 #include <net/genetlink.h>
  17
  18 #include <uapi/linux/if.h>
  19 #include <uapi/linux/smc.h>
  20
  21 #include <rdma/ib_verbs.h>
  22
  23 #include <net/netns/generic.h>
  24 #include "smc_netns.h"
  25
  26 #include "smc_pnet.h"
  27 #include "smc_ib.h"
  28 #include "smc_ism.h"
  29 #include "smc_core.h"
  30
  31 #define SMC_ASCII_BLANK 32
  32
  33 static struct net_device *pnet_find_base_ndev(struct net_device *ndev);
  34
  35 static 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 /**
  54  * struct smc_user_pnetentry - pnet identifier name entry for/from user
  55  * @list: List node.
  56  * @pnet_name: Pnet identifier name
  57  * @ndev: pointer to network device.
  58  * @smcibdev: Pointer to IB device.
  59  * @ib_port: Port of IB device.
  60  * @smcd_dev: Pointer to smcd device.
  61  */
  62 struct smc_user_pnetentry {
  63         struct list_head list;
  64         char pnet_name[SMC_MAX_PNETID_LEN + 1];
  65         struct net_device *ndev;
  66         struct smc_ib_device *smcibdev;
  67         u8 ib_port;
  68         struct smcd_dev *smcd_dev;
  69 };
  70
  71 /* pnet entry stored in pnet table */
  72 struct smc_pnetentry {
  73         struct list_head list;
  74         char pnet_name[SMC_MAX_PNETID_LEN + 1];
  75         struct net_device *ndev;
  76 };
  77
  78 /* Check if two given pnetids match */
  79 static bool smc_pnet_match(u8 *pnetid1, u8 *pnetid2)
  80 {
  81         int i;
  82
  83         for (i = 0; i < SMC_MAX_PNETID_LEN; i++) {
  84                 if ((pnetid1[i] == 0 || pnetid1[i] == SMC_ASCII_BLANK) &&
  85                     (pnetid2[i] == 0 || pnetid2[i] == SMC_ASCII_BLANK))
  86                         break;
  87                 if (pnetid1[i] != pnetid2[i])
  88                         return false;
  89         }
  90         return true;
  91 }
  92
  93 /* Remove a pnetid from the pnet table.
  94  */
  95 static int smc_pnet_remove_by_pnetid(struct net *net, char *pnet_name)
  96 {
  97         struct smc_pnetentry *pnetelem, *tmp_pe;
  98         struct smc_pnettable *pnettable;
  99         struct smc_ib_device *ibdev;
 100         struct smcd_dev *smcd_dev;
 101         struct smc_net *sn;
 102         int rc = -ENOENT;
 103         int ibport;
 104
 105         /* get pnettable for namespace */
 106         sn = net_generic(net, smc_net_id);
 107         pnettable = &sn->pnettable;
 108
 109         /* remove netdevices */
 110         write_lock(&pnettable->lock);
 111         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist,
 112                                  list) {
 113                 if (!pnet_name ||
 114                     smc_pnet_match(pnetelem->pnet_name, pnet_name)) {
 115                         list_del(&pnetelem->list);
 116                         dev_put(pnetelem->ndev);
 117                         kfree(pnetelem);
 118                         rc = 0;
 119                 }
 120         }
 121         write_unlock(&pnettable->lock);
 122
 123         /* if this is not the initial namespace, stop here */
 124         if (net != &init_net)
 125                 return rc;
 126
 127         /* remove ib devices */
 128         spin_lock(&smc_ib_devices.lock);
 129         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 130                 for (ibport = 0; ibport < SMC_MAX_PORTS; ibport++) {
 131                         if (ibdev->pnetid_by_user[ibport] &&
 132                             (!pnet_name ||
 133                              smc_pnet_match(pnet_name,
 134                                             ibdev->pnetid[ibport]))) {
 135                                 memset(ibdev->pnetid[ibport], 0,
 136                                        SMC_MAX_PNETID_LEN);
 137                                 ibdev->pnetid_by_user[ibport] = false;
 138                                 rc = 0;
 139                         }
 140                 }
 141         }
 142         spin_unlock(&smc_ib_devices.lock);
 143         /* remove smcd devices */
 144         spin_lock(&smcd_dev_list.lock);
 145         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 146                 if (smcd_dev->pnetid_by_user &&
 147                     (!pnet_name ||
 148                      smc_pnet_match(pnet_name, smcd_dev->pnetid))) {
 149                         memset(smcd_dev->pnetid, 0, SMC_MAX_PNETID_LEN);
 150                         smcd_dev->pnetid_by_user = false;
 151                         rc = 0;
 152                 }
 153         }
 154         spin_unlock(&smcd_dev_list.lock);
 155         return rc;
 156 }
 157
 158 /* Remove a pnet entry mentioning a given network device from the pnet table.
 159  */
 160 static int smc_pnet_remove_by_ndev(struct net_device *ndev)
 161 {
 162         struct smc_pnetentry *pnetelem, *tmp_pe;
 163         struct smc_pnettable *pnettable;
 164         struct net *net = dev_net(ndev);
 165         struct smc_net *sn;
 166         int rc = -ENOENT;
 167
 168         /* get pnettable for namespace */
 169         sn = net_generic(net, smc_net_id);
 170         pnettable = &sn->pnettable;
 171
 172         write_lock(&pnettable->lock);
 173         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
 174                 if (pnetelem->ndev == ndev) {
 175                         list_del(&pnetelem->list);
 176                         dev_put(pnetelem->ndev);
 177                         kfree(pnetelem);
 178                         rc = 0;
 179                         break;
 180                 }
 181         }
 182         write_unlock(&pnettable->lock);
 183         return rc;
 184 }
 185
 186 /* Append a pnetid to the end of the pnet table if not already on this list.
 187  */
 188 static int smc_pnet_enter(struct smc_pnettable *pnettable,
 189                           struct smc_user_pnetentry *new_pnetelem)
 190 {
 191         u8 pnet_null[SMC_MAX_PNETID_LEN] = {0};
 192         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 193         struct smc_pnetentry *tmp_pnetelem;
 194         struct smc_pnetentry *pnetelem;
 195         bool new_smcddev = false;
 196         struct net_device *ndev;
 197         bool new_netdev = true;
 198         bool new_ibdev = false;
 199
 200         if (new_pnetelem->smcibdev) {
 201                 struct smc_ib_device *ib_dev = new_pnetelem->smcibdev;
 202                 int ib_port = new_pnetelem->ib_port;
 203
 204                 spin_lock(&smc_ib_devices.lock);
 205                 if (smc_pnet_match(ib_dev->pnetid[ib_port - 1], pnet_null)) {
 206                         memcpy(ib_dev->pnetid[ib_port - 1],
 207                                new_pnetelem->pnet_name, SMC_MAX_PNETID_LEN);
 208                         ib_dev->pnetid_by_user[ib_port - 1] = true;
 209                         new_ibdev = true;
 210                 }
 211                 spin_unlock(&smc_ib_devices.lock);
 212         }
 213         if (new_pnetelem->smcd_dev) {
 214                 struct smcd_dev *smcd_dev = new_pnetelem->smcd_dev;
 215
 216                 spin_lock(&smcd_dev_list.lock);
 217                 if (smc_pnet_match(smcd_dev->pnetid, pnet_null)) {
 218                         memcpy(smcd_dev->pnetid, new_pnetelem->pnet_name,
 219                                SMC_MAX_PNETID_LEN);
 220                         smcd_dev->pnetid_by_user = true;
 221                         new_smcddev = true;
 222                 }
 223                 spin_unlock(&smcd_dev_list.lock);
 224         }
 225
 226         if (!new_pnetelem->ndev)
 227                 return (new_ibdev || new_smcddev) ? 0 : -EEXIST;
 228
 229         /* check if (base) netdev already has a pnetid. If there is one, we do
 230          * not want to add a pnet table entry
 231          */
 232         ndev = pnet_find_base_ndev(new_pnetelem->ndev);
 233         if (!smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
 234                                     ndev_pnetid))
 235                 return (new_ibdev || new_smcddev) ? 0 : -EEXIST;
 236
 237         /* add a new netdev entry to the pnet table if there isn't one */
 238         tmp_pnetelem = kzalloc(sizeof(*pnetelem), GFP_KERNEL);
 239         if (!tmp_pnetelem)
 240                 return -ENOMEM;
 241         memcpy(tmp_pnetelem->pnet_name, new_pnetelem->pnet_name,
 242                SMC_MAX_PNETID_LEN);
 243         tmp_pnetelem->ndev = new_pnetelem->ndev;
 244
 245         write_lock(&pnettable->lock);
 246         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 247                 if (pnetelem->ndev == new_pnetelem->ndev)
 248                         new_netdev = false;
 249         }
 250         if (new_netdev) {
 251                 dev_hold(tmp_pnetelem->ndev);
 252                 list_add_tail(&tmp_pnetelem->list, &pnettable->pnetlist);
 253                 write_unlock(&pnettable->lock);
 254         } else {
 255                 write_unlock(&pnettable->lock);
 256                 kfree(tmp_pnetelem);
 257         }
 258
 259         return (new_netdev || new_ibdev || new_smcddev) ? 0 : -EEXIST;
 260 }
 261
 262 /* The limit for pnetid is 16 characters.
 263  * Valid characters should be (single-byte character set) a-z, A-Z, 0-9.
 264  * Lower case letters are converted to upper case.
 265  * Interior blanks should not be used.
 266  */
 267 static bool smc_pnetid_valid(const char *pnet_name, char *pnetid)
 268 {
 269         char *bf = skip_spaces(pnet_name);
 270         size_t len = strlen(bf);
 271         char *end = bf + len;
 272
 273         if (!len)
 274                 return false;
 275         while (--end >= bf && isspace(*end))
 276                 ;
 277         if (end - bf >= SMC_MAX_PNETID_LEN)
 278                 return false;
 279         while (bf <= end) {
 280                 if (!isalnum(*bf))
 281                         return false;
 282                 *pnetid++ = islower(*bf) ? toupper(*bf) : *bf;
 283                 bf++;
 284         }
 285         *pnetid = '\0';
 286         return true;
 287 }
 288
 289 /* Find an infiniband device by a given name. The device might not exist. */
 290 static struct smc_ib_device *smc_pnet_find_ib(char *ib_name)
 291 {
 292         struct smc_ib_device *ibdev;
 293
 294         spin_lock(&smc_ib_devices.lock);
 295         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 296                 if (!strncmp(ibdev->ibdev->name, ib_name,
 297                              sizeof(ibdev->ibdev->name)) ||
 298                     !strncmp(dev_name(ibdev->ibdev->dev.parent), ib_name,
 299                              IB_DEVICE_NAME_MAX - 1)) {
 300                         goto out;
 301                 }
 302         }
 303         ibdev = NULL;
 304 out:
 305         spin_unlock(&smc_ib_devices.lock);
 306         return ibdev;
 307 }
 308
 309 /* Find an smcd device by a given name. The device might not exist. */
 310 static struct smcd_dev *smc_pnet_find_smcd(char *smcd_name)
 311 {
 312         struct smcd_dev *smcd_dev;
 313
 314         spin_lock(&smcd_dev_list.lock);
 315         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 316                 if (!strncmp(dev_name(&smcd_dev->dev), smcd_name,
 317                              IB_DEVICE_NAME_MAX - 1))
 318                         goto out;
 319         }
 320         smcd_dev = NULL;
 321 out:
 322         spin_unlock(&smcd_dev_list.lock);
 323         return smcd_dev;
 324 }
 325
 326 /* Parse the supplied netlink attributes and fill a pnetentry structure.
 327  * For ethernet and infiniband device names verify that the devices exist.
 328  */
 329 static int smc_pnet_fill_entry(struct net *net,
 330                                struct smc_user_pnetentry *pnetelem,
 331                                struct nlattr *tb[])
 332 {
 333         char *string, *ibname;
 334         int rc;
 335
 336         memset(pnetelem, 0, sizeof(*pnetelem));
 337         INIT_LIST_HEAD(&pnetelem->list);
 338
 339         rc = -EINVAL;
 340         if (!tb[SMC_PNETID_NAME])
 341                 goto error;
 342         string = (char *)nla_data(tb[SMC_PNETID_NAME]);
 343         if (!smc_pnetid_valid(string, pnetelem->pnet_name))
 344                 goto error;
 345
 346         rc = -EINVAL;
 347         if (tb[SMC_PNETID_ETHNAME]) {
 348                 string = (char *)nla_data(tb[SMC_PNETID_ETHNAME]);
 349                 pnetelem->ndev = dev_get_by_name(net, string);
 350                 if (!pnetelem->ndev)
 351                         goto error;
 352         }
 353
 354         /* if this is not the initial namespace, stop here */
 355         if (net != &init_net)
 356                 return 0;
 357
 358         rc = -EINVAL;
 359         if (tb[SMC_PNETID_IBNAME]) {
 360                 ibname = (char *)nla_data(tb[SMC_PNETID_IBNAME]);
 361                 ibname = strim(ibname);
 362                 pnetelem->smcibdev = smc_pnet_find_ib(ibname);
 363                 pnetelem->smcd_dev = smc_pnet_find_smcd(ibname);
 364                 if (!pnetelem->smcibdev && !pnetelem->smcd_dev)
 365                         goto error;
 366                 if (pnetelem->smcibdev) {
 367                         if (!tb[SMC_PNETID_IBPORT])
 368                                 goto error;
 369                         pnetelem->ib_port = nla_get_u8(tb[SMC_PNETID_IBPORT]);
 370                         if (pnetelem->ib_port < 1 ||
 371                             pnetelem->ib_port > SMC_MAX_PORTS)
 372                                 goto error;
 373                 }
 374         }
 375
 376         return 0;
 377
 378 error:
 379         return rc;
 380 }
 381
 382 /* Convert an smc_pnetentry to a netlink attribute sequence */
 383 static int smc_pnet_set_nla(struct sk_buff *msg,
 384                             struct smc_user_pnetentry *pnetelem)
 385 {
 386         if (nla_put_string(msg, SMC_PNETID_NAME, pnetelem->pnet_name))
 387                 return -1;
 388         if (pnetelem->ndev) {
 389                 if (nla_put_string(msg, SMC_PNETID_ETHNAME,
 390                                    pnetelem->ndev->name))
 391                         return -1;
 392         } else {
 393                 if (nla_put_string(msg, SMC_PNETID_ETHNAME, "n/a"))
 394                         return -1;
 395         }
 396         if (pnetelem->smcibdev) {
 397                 if (nla_put_string(msg, SMC_PNETID_IBNAME,
 398                         dev_name(pnetelem->smcibdev->ibdev->dev.parent)) ||
 399                     nla_put_u8(msg, SMC_PNETID_IBPORT, pnetelem->ib_port))
 400                         return -1;
 401         } else if (pnetelem->smcd_dev) {
 402                 if (nla_put_string(msg, SMC_PNETID_IBNAME,
 403                                    dev_name(&pnetelem->smcd_dev->dev)) ||
 404                     nla_put_u8(msg, SMC_PNETID_IBPORT, 1))
 405                         return -1;
 406         } else {
 407                 if (nla_put_string(msg, SMC_PNETID_IBNAME, "n/a") ||
 408                     nla_put_u8(msg, SMC_PNETID_IBPORT, 0xff))
 409                         return -1;
 410         }
 411
 412         return 0;
 413 }
 414
 415 static int smc_pnet_add(struct sk_buff *skb, struct genl_info *info)
 416 {
 417         struct net *net = genl_info_net(info);
 418         struct smc_user_pnetentry pnetelem;
 419         struct smc_pnettable *pnettable;
 420         struct smc_net *sn;
 421         int rc;
 422
 423         /* get pnettable for namespace */
 424         sn = net_generic(net, smc_net_id);
 425         pnettable = &sn->pnettable;
 426
 427         rc = smc_pnet_fill_entry(net, &pnetelem, info->attrs);
 428         if (!rc)
 429                 rc = smc_pnet_enter(pnettable, &pnetelem);
 430         if (pnetelem.ndev)
 431                 dev_put(pnetelem.ndev);
 432         return rc;
 433 }
 434
 435 static int smc_pnet_del(struct sk_buff *skb, struct genl_info *info)
 436 {
 437         struct net *net = genl_info_net(info);
 438
 439         if (!info->attrs[SMC_PNETID_NAME])
 440                 return -EINVAL;
 441         return smc_pnet_remove_by_pnetid(net,
 442                                 (char *)nla_data(info->attrs[SMC_PNETID_NAME]));
 443 }
 444
 445 static int smc_pnet_dump_start(struct netlink_callback *cb)
 446 {
 447         cb->args[0] = 0;
 448         return 0;
 449 }
 450
 451 static int smc_pnet_dumpinfo(struct sk_buff *skb,
 452                              u32 portid, u32 seq, u32 flags,
 453                              struct smc_user_pnetentry *pnetelem)
 454 {
 455         void *hdr;
 456
 457         hdr = genlmsg_put(skb, portid, seq, &smc_pnet_nl_family,
 458                           flags, SMC_PNETID_GET);
 459         if (!hdr)
 460                 return -ENOMEM;
 461         if (smc_pnet_set_nla(skb, pnetelem) < 0) {
 462                 genlmsg_cancel(skb, hdr);
 463                 return -EMSGSIZE;
 464         }
 465         genlmsg_end(skb, hdr);
 466         return 0;
 467 }
 468
 469 static int _smc_pnet_dump(struct net *net, struct sk_buff *skb, u32 portid,
 470                           u32 seq, u8 *pnetid, int start_idx)
 471 {
 472         struct smc_user_pnetentry tmp_entry;
 473         struct smc_pnettable *pnettable;
 474         struct smc_pnetentry *pnetelem;
 475         struct smc_ib_device *ibdev;
 476         struct smcd_dev *smcd_dev;
 477         struct smc_net *sn;
 478         int idx = 0;
 479         int ibport;
 480
 481         /* get pnettable for namespace */
 482         sn = net_generic(net, smc_net_id);
 483         pnettable = &sn->pnettable;
 484
 485         /* dump netdevices */
 486         read_lock(&pnettable->lock);
 487         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 488                 if (pnetid && !smc_pnet_match(pnetelem->pnet_name, pnetid))
 489                         continue;
 490                 if (idx++ < start_idx)
 491                         continue;
 492                 memset(&tmp_entry, 0, sizeof(tmp_entry));
 493                 memcpy(&tmp_entry.pnet_name, pnetelem->pnet_name,
 494                        SMC_MAX_PNETID_LEN);
 495                 tmp_entry.ndev = pnetelem->ndev;
 496                 if (smc_pnet_dumpinfo(skb, portid, seq, NLM_F_MULTI,
 497                                       &tmp_entry)) {
 498                         --idx;
 499                         break;
 500                 }
 501         }
 502         read_unlock(&pnettable->lock);
 503
 504         /* if this is not the initial namespace, stop here */
 505         if (net != &init_net)
 506                 return idx;
 507
 508         /* dump ib devices */
 509         spin_lock(&smc_ib_devices.lock);
 510         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 511                 for (ibport = 0; ibport < SMC_MAX_PORTS; ibport++) {
 512                         if (ibdev->pnetid_by_user[ibport]) {
 513                                 if (pnetid &&
 514                                     !smc_pnet_match(ibdev->pnetid[ibport],
 515                                                     pnetid))
 516                                         continue;
 517                                 if (idx++ < start_idx)
 518                                         continue;
 519                                 memset(&tmp_entry, 0, sizeof(tmp_entry));
 520                                 memcpy(&tmp_entry.pnet_name,
 521                                        ibdev->pnetid[ibport],
 522                                        SMC_MAX_PNETID_LEN);
 523                                 tmp_entry.smcibdev = ibdev;
 524                                 tmp_entry.ib_port = ibport + 1;
 525                                 if (smc_pnet_dumpinfo(skb, portid, seq,
 526                                                       NLM_F_MULTI,
 527                                                       &tmp_entry)) {
 528                                         --idx;
 529                                         break;
 530                                 }
 531                         }
 532                 }
 533         }
 534         spin_unlock(&smc_ib_devices.lock);
 535
 536         /* dump smcd devices */
 537         spin_lock(&smcd_dev_list.lock);
 538         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 539                 if (smcd_dev->pnetid_by_user) {
 540                         if (pnetid && !smc_pnet_match(smcd_dev->pnetid, pnetid))
 541                                 continue;
 542                         if (idx++ < start_idx)
 543                                 continue;
 544                         memset(&tmp_entry, 0, sizeof(tmp_entry));
 545                         memcpy(&tmp_entry.pnet_name, smcd_dev->pnetid,
 546                                SMC_MAX_PNETID_LEN);
 547                         tmp_entry.smcd_dev = smcd_dev;
 548                         if (smc_pnet_dumpinfo(skb, portid, seq, NLM_F_MULTI,
 549                                               &tmp_entry)) {
 550                                 --idx;
 551                                 break;
 552                         }
 553                 }
 554         }
 555         spin_unlock(&smcd_dev_list.lock);
 556
 557         return idx;
 558 }
 559
 560 static int smc_pnet_dump(struct sk_buff *skb, struct netlink_callback *cb)
 561 {
 562         struct net *net = sock_net(skb->sk);
 563         int idx;
 564
 565         idx = _smc_pnet_dump(net, skb, NETLINK_CB(cb->skb).portid,
 566                              cb->nlh->nlmsg_seq, NULL, cb->args[0]);
 567
 568         cb->args[0] = idx;
 569         return skb->len;
 570 }
 571
 572 /* Retrieve one PNETID entry */
 573 static int smc_pnet_get(struct sk_buff *skb, struct genl_info *info)
 574 {
 575         struct net *net = genl_info_net(info);
 576         struct sk_buff *msg;
 577         void *hdr;
 578
 579         if (!info->attrs[SMC_PNETID_NAME])
 580                 return -EINVAL;
 581
 582         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 583         if (!msg)
 584                 return -ENOMEM;
 585
 586         _smc_pnet_dump(net, msg, info->snd_portid, info->snd_seq,
 587                        nla_data(info->attrs[SMC_PNETID_NAME]), 0);
 588
 589         /* finish multi part message and send it */
 590         hdr = nlmsg_put(msg, info->snd_portid, info->snd_seq, NLMSG_DONE, 0,
 591                         NLM_F_MULTI);
 592         if (!hdr) {
 593                 nlmsg_free(msg);
 594                 return -EMSGSIZE;
 595         }
 596         return genlmsg_reply(msg, info);
 597 }
 598
 599 /* Remove and delete all pnetids from pnet table.
 600  */
 601 static int smc_pnet_flush(struct sk_buff *skb, struct genl_info *info)
 602 {
 603         struct net *net = genl_info_net(info);
 604
 605         smc_pnet_remove_by_pnetid(net, NULL);
 606         return 0;
 607 }
 608
 609 /* SMC_PNETID generic netlink operation definition */
 610 static const struct genl_ops smc_pnet_ops[] = {
 611         {
 612                 .cmd = SMC_PNETID_GET,
 613                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 614                 /* can be retrieved by unprivileged users */
 615                 .doit = smc_pnet_get,
 616                 .dumpit = smc_pnet_dump,
 617                 .start = smc_pnet_dump_start
 618         },
 619         {
 620                 .cmd = SMC_PNETID_ADD,
 621                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 622                 .flags = GENL_ADMIN_PERM,
 623                 .doit = smc_pnet_add
 624         },
 625         {
 626                 .cmd = SMC_PNETID_DEL,
 627                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 628                 .flags = GENL_ADMIN_PERM,
 629                 .doit = smc_pnet_del
 630         },
 631         {
 632                 .cmd = SMC_PNETID_FLUSH,
 633                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 634                 .flags = GENL_ADMIN_PERM,
 635                 .doit = smc_pnet_flush
 636         }
 637 };
 638
 639 /* SMC_PNETID family definition */
 640 static struct genl_family smc_pnet_nl_family __ro_after_init = {
 641         .hdrsize = 0,
 642         .name = SMCR_GENL_FAMILY_NAME,
 643         .version = SMCR_GENL_FAMILY_VERSION,
 644         .maxattr = SMC_PNETID_MAX,
 645         .policy = smc_pnet_policy,
 646         .netnsok = true,
 647         .module = THIS_MODULE,
 648         .ops = smc_pnet_ops,
 649         .n_ops =  ARRAY_SIZE(smc_pnet_ops)
 650 };
 651
 652 static int smc_pnet_netdev_event(struct notifier_block *this,
 653                                  unsigned long event, void *ptr)
 654 {
 655         struct net_device *event_dev = netdev_notifier_info_to_dev(ptr);
 656
 657         switch (event) {
 658         case NETDEV_REBOOT:
 659         case NETDEV_UNREGISTER:
 660                 smc_pnet_remove_by_ndev(event_dev);
 661                 return NOTIFY_OK;
 662         default:
 663                 return NOTIFY_DONE;
 664         }
 665 }
 666
 667 static struct notifier_block smc_netdev_notifier = {
 668         .notifier_call = smc_pnet_netdev_event
 669 };
 670
 671 /* init network namespace */
 672 int smc_pnet_net_init(struct net *net)
 673 {
 674         struct smc_net *sn = net_generic(net, smc_net_id);
 675         struct smc_pnettable *pnettable = &sn->pnettable;
 676
 677         INIT_LIST_HEAD(&pnettable->pnetlist);
 678         rwlock_init(&pnettable->lock);
 679
 680         return 0;
 681 }
 682
 683 int __init smc_pnet_init(void)
 684 {
 685         int rc;
 686
 687         rc = genl_register_family(&smc_pnet_nl_family);
 688         if (rc)
 689                 return rc;
 690         rc = register_netdevice_notifier(&smc_netdev_notifier);
 691         if (rc)
 692                 genl_unregister_family(&smc_pnet_nl_family);
 693         return rc;
 694 }
 695
 696 /* exit network namespace */
 697 void smc_pnet_net_exit(struct net *net)
 698 {
 699         /* flush pnet table */
 700         smc_pnet_remove_by_pnetid(net, NULL);
 701 }
 702
 703 void smc_pnet_exit(void)
 704 {
 705         unregister_netdevice_notifier(&smc_netdev_notifier);
 706         genl_unregister_family(&smc_pnet_nl_family);
 707 }
 708
 709 /* Determine one base device for stacked net devices.
 710  * If the lower device level contains more than one devices
 711  * (for instance with bonding slaves), just the first device
 712  * is used to reach a base device.
 713  */
 714 static struct net_device *pnet_find_base_ndev(struct net_device *ndev)
 715 {
 716         int i, nest_lvl;
 717
 718         rtnl_lock();
 719         nest_lvl = ndev->lower_level;
 720         for (i = 0; i < nest_lvl; i++) {
 721                 struct list_head *lower = &ndev->adj_list.lower;
 722
 723                 if (list_empty(lower))
 724                         break;
 725                 lower = lower->next;
 726                 ndev = netdev_lower_get_next(ndev, &lower);
 727         }
 728         rtnl_unlock();
 729         return ndev;
 730 }
 731
 732 static int smc_pnet_find_ndev_pnetid_by_table(struct net_device *ndev,
 733                                               u8 *pnetid)
 734 {
 735         struct smc_pnettable *pnettable;
 736         struct net *net = dev_net(ndev);
 737         struct smc_pnetentry *pnetelem;
 738         struct smc_net *sn;
 739         int rc = -ENOENT;
 740
 741         /* get pnettable for namespace */
 742         sn = net_generic(net, smc_net_id);
 743         pnettable = &sn->pnettable;
 744
 745         read_lock(&pnettable->lock);
 746         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 747                 if (ndev == pnetelem->ndev) {
 748                         /* get pnetid of netdev device */
 749                         memcpy(pnetid, pnetelem->pnet_name, SMC_MAX_PNETID_LEN);
 750                         rc = 0;
 751                         break;
 752                 }
 753         }
 754         read_unlock(&pnettable->lock);
 755         return rc;
 756 }
 757
 758 /* if handshake network device belongs to a roce device, return its
 759  * IB device and port
 760  */
 761 static void smc_pnet_find_rdma_dev(struct net_device *netdev,
 762                                    struct smc_init_info *ini)
 763 {
 764         struct smc_ib_device *ibdev;
 765
 766         spin_lock(&smc_ib_devices.lock);
 767         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 768                 struct net_device *ndev;
 769                 int i;
 770
 771                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
 772                         if (!rdma_is_port_valid(ibdev->ibdev, i))
 773                                 continue;
 774                         if (!ibdev->ibdev->ops.get_netdev)
 775                                 continue;
 776                         ndev = ibdev->ibdev->ops.get_netdev(ibdev->ibdev, i);
 777                         if (!ndev)
 778                                 continue;
 779                         dev_put(ndev);
 780                         if (netdev == ndev &&
 781                             smc_ib_port_active(ibdev, i) &&
 782                             !test_bit(i - 1, ibdev->ports_going_away) &&
 783                             !smc_ib_determine_gid(ibdev, i, ini->vlan_id,
 784                                                   ini->ib_gid, NULL)) {
 785                                 ini->ib_dev = ibdev;
 786                                 ini->ib_port = i;
 787                                 break;
 788                         }
 789                 }
 790         }
 791         spin_unlock(&smc_ib_devices.lock);
 792 }
 793
 794 /* Determine the corresponding IB device port based on the hardware PNETID.
 795  * Searching stops at the first matching active IB device port with vlan_id
 796  * configured.
 797  * If nothing found, check pnetid table.
 798  * If nothing found, try to use handshake device
 799  */
 800 static void smc_pnet_find_roce_by_pnetid(struct net_device *ndev,
 801                                          struct smc_init_info *ini)
 802 {
 803         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 804         struct smc_ib_device *ibdev;
 805         int i;
 806
 807         ndev = pnet_find_base_ndev(ndev);
 808         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
 809                                    ndev_pnetid) &&
 810             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid)) {
 811                 smc_pnet_find_rdma_dev(ndev, ini);
 812                 return; /* pnetid could not be determined */
 813         }
 814
 815         spin_lock(&smc_ib_devices.lock);
 816         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 817                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
 818                         if (!rdma_is_port_valid(ibdev->ibdev, i))
 819                                 continue;
 820                         if (smc_pnet_match(ibdev->pnetid[i - 1], ndev_pnetid) &&
 821                             smc_ib_port_active(ibdev, i) &&
 822                             !test_bit(i - 1, ibdev->ports_going_away) &&
 823                             !smc_ib_determine_gid(ibdev, i, ini->vlan_id,
 824                                                   ini->ib_gid, NULL)) {
 825                                 ini->ib_dev = ibdev;
 826                                 ini->ib_port = i;
 827                                 goto out;
 828                         }
 829                 }
 830         }
 831 out:
 832         spin_unlock(&smc_ib_devices.lock);
 833 }
 834
 835 static void smc_pnet_find_ism_by_pnetid(struct net_device *ndev,
 836                                         struct smc_init_info *ini)
 837 {
 838         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 839         struct smcd_dev *ismdev;
 840
 841         ndev = pnet_find_base_ndev(ndev);
 842         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
 843                                    ndev_pnetid) &&
 844             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid))
 845                 return; /* pnetid could not be determined */
 846
 847         spin_lock(&smcd_dev_list.lock);
 848         list_for_each_entry(ismdev, &smcd_dev_list.list, list) {
 849                 if (smc_pnet_match(ismdev->pnetid, ndev_pnetid) &&
 850                     !ismdev->going_away) {
 851                         ini->ism_dev = ismdev;
 852                         break;
 853                 }
 854         }
 855         spin_unlock(&smcd_dev_list.lock);
 856 }
 857
 858 /* PNET table analysis for a given sock:
 859  * determine ib_device and port belonging to used internal TCP socket
 860  * ethernet interface.
 861  */
 862 void smc_pnet_find_roce_resource(struct sock *sk, struct smc_init_info *ini)
 863 {
 864         struct dst_entry *dst = sk_dst_get(sk);
 865
 866         ini->ib_dev = NULL;
 867         ini->ib_port = 0;
 868         if (!dst)
 869                 goto out;
 870         if (!dst->dev)
 871                 goto out_rel;
 872
 873         smc_pnet_find_roce_by_pnetid(dst->dev, ini);
 874
 875 out_rel:
 876         dst_release(dst);
 877 out:
 878         return;
 879 }
 880
 881 void smc_pnet_find_ism_resource(struct sock *sk, struct smc_init_info *ini)
 882 {
 883         struct dst_entry *dst = sk_dst_get(sk);
 884
 885         ini->ism_dev = NULL;
 886         if (!dst)
 887                 goto out;
 888         if (!dst->dev)
 889                 goto out_rel;
 890
 891         smc_pnet_find_ism_by_pnetid(dst->dev, ini);
 892
 893 out_rel:
 894         dst_release(dst);
 895 out:
 896         return;
 897 }