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
  30 #define SMC_ASCII_BLANK 32
  31
  32 static struct net_device *pnet_find_base_ndev(struct net_device *ndev);
  33
  34 static struct nla_policy smc_pnet_policy[SMC_PNETID_MAX + 1] = {
  35         [SMC_PNETID_NAME] = {
  36                 .type = NLA_NUL_STRING,
  37                 .len = SMC_MAX_PNETID_LEN
  38         },
  39         [SMC_PNETID_ETHNAME] = {
  40                 .type = NLA_NUL_STRING,
  41                 .len = IFNAMSIZ - 1
  42         },
  43         [SMC_PNETID_IBNAME] = {
  44                 .type = NLA_NUL_STRING,
  45                 .len = IB_DEVICE_NAME_MAX - 1
  46         },
  47         [SMC_PNETID_IBPORT] = { .type = NLA_U8 }
  48 };
  49
  50 static struct genl_family smc_pnet_nl_family;
  51
  52 /**
  53  * struct smc_user_pnetentry - pnet identifier name entry for/from user
  54  * @list: List node.
  55  * @pnet_name: Pnet identifier name
  56  * @ndev: pointer to network device.
  57  * @smcibdev: Pointer to IB device.
  58  * @ib_port: Port of IB device.
  59  * @smcd_dev: Pointer to smcd device.
  60  */
  61 struct smc_user_pnetentry {
  62         struct list_head list;
  63         char pnet_name[SMC_MAX_PNETID_LEN + 1];
  64         struct net_device *ndev;
  65         struct smc_ib_device *smcibdev;
  66         u8 ib_port;
  67         struct smcd_dev *smcd_dev;
  68 };
  69
  70 /* pnet entry stored in pnet table */
  71 struct smc_pnetentry {
  72         struct list_head list;
  73         char pnet_name[SMC_MAX_PNETID_LEN + 1];
  74         struct net_device *ndev;
  75 };
  76
  77 /* Check if two given pnetids match */
  78 static bool smc_pnet_match(u8 *pnetid1, u8 *pnetid2)
  79 {
  80         int i;
  81
  82         for (i = 0; i < SMC_MAX_PNETID_LEN; i++) {
  83                 if ((pnetid1[i] == 0 || pnetid1[i] == SMC_ASCII_BLANK) &&
  84                     (pnetid2[i] == 0 || pnetid2[i] == SMC_ASCII_BLANK))
  85                         break;
  86                 if (pnetid1[i] != pnetid2[i])
  87                         return false;
  88         }
  89         return true;
  90 }
  91
  92 /* Remove a pnetid from the pnet table.
  93  */
  94 static int smc_pnet_remove_by_pnetid(struct net *net, char *pnet_name)
  95 {
  96         struct smc_pnetentry *pnetelem, *tmp_pe;
  97         struct smc_pnettable *pnettable;
  98         struct smc_ib_device *ibdev;
  99         struct smcd_dev *smcd_dev;
 100         struct smc_net *sn;
 101         int rc = -ENOENT;
 102         int ibport;
 103
 104         /* get pnettable for namespace */
 105         sn = net_generic(net, smc_net_id);
 106         pnettable = &sn->pnettable;
 107
 108         /* remove netdevices */
 109         write_lock(&pnettable->lock);
 110         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist,
 111                                  list) {
 112                 if (!pnet_name ||
 113                     smc_pnet_match(pnetelem->pnet_name, pnet_name)) {
 114                         list_del(&pnetelem->list);
 115                         dev_put(pnetelem->ndev);
 116                         kfree(pnetelem);
 117                         rc = 0;
 118                 }
 119         }
 120         write_unlock(&pnettable->lock);
 121
 122         /* if this is not the initial namespace, stop here */
 123         if (net != &init_net)
 124                 return rc;
 125
 126         /* remove ib devices */
 127         spin_lock(&smc_ib_devices.lock);
 128         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 129                 for (ibport = 0; ibport < SMC_MAX_PORTS; ibport++) {
 130                         if (ibdev->pnetid_by_user[ibport] &&
 131                             (!pnet_name ||
 132                              smc_pnet_match(pnet_name,
 133                                             ibdev->pnetid[ibport]))) {
 134                                 memset(ibdev->pnetid[ibport], 0,
 135                                        SMC_MAX_PNETID_LEN);
 136                                 ibdev->pnetid_by_user[ibport] = false;
 137                                 rc = 0;
 138                         }
 139                 }
 140         }
 141         spin_unlock(&smc_ib_devices.lock);
 142         /* remove smcd devices */
 143         spin_lock(&smcd_dev_list.lock);
 144         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 145                 if (smcd_dev->pnetid_by_user &&
 146                     (!pnet_name ||
 147                      smc_pnet_match(pnet_name, smcd_dev->pnetid))) {
 148                         memset(smcd_dev->pnetid, 0, SMC_MAX_PNETID_LEN);
 149                         smcd_dev->pnetid_by_user = false;
 150                         rc = 0;
 151                 }
 152         }
 153         spin_unlock(&smcd_dev_list.lock);
 154         return rc;
 155 }
 156
 157 /* Remove a pnet entry mentioning a given network device from the pnet table.
 158  */
 159 static int smc_pnet_remove_by_ndev(struct net_device *ndev)
 160 {
 161         struct smc_pnetentry *pnetelem, *tmp_pe;
 162         struct smc_pnettable *pnettable;
 163         struct net *net = dev_net(ndev);
 164         struct smc_net *sn;
 165         int rc = -ENOENT;
 166
 167         /* get pnettable for namespace */
 168         sn = net_generic(net, smc_net_id);
 169         pnettable = &sn->pnettable;
 170
 171         write_lock(&pnettable->lock);
 172         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
 173                 if (pnetelem->ndev == ndev) {
 174                         list_del(&pnetelem->list);
 175                         dev_put(pnetelem->ndev);
 176                         kfree(pnetelem);
 177                         rc = 0;
 178                         break;
 179                 }
 180         }
 181         write_unlock(&pnettable->lock);
 182         return rc;
 183 }
 184
 185 /* Append a pnetid to the end of the pnet table if not already on this list.
 186  */
 187 static int smc_pnet_enter(struct smc_pnettable *pnettable,
 188                           struct smc_user_pnetentry *new_pnetelem)
 189 {
 190         u8 pnet_null[SMC_MAX_PNETID_LEN] = {0};
 191         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 192         struct smc_pnetentry *tmp_pnetelem;
 193         struct smc_pnetentry *pnetelem;
 194         bool new_smcddev = false;
 195         struct net_device *ndev;
 196         bool new_netdev = true;
 197         bool new_ibdev = false;
 198
 199         if (new_pnetelem->smcibdev) {
 200                 struct smc_ib_device *ib_dev = new_pnetelem->smcibdev;
 201                 int ib_port = new_pnetelem->ib_port;
 202
 203                 spin_lock(&smc_ib_devices.lock);
 204                 if (smc_pnet_match(ib_dev->pnetid[ib_port - 1], pnet_null)) {
 205                         memcpy(ib_dev->pnetid[ib_port - 1],
 206                                new_pnetelem->pnet_name, SMC_MAX_PNETID_LEN);
 207                         ib_dev->pnetid_by_user[ib_port - 1] = true;
 208                         new_ibdev = true;
 209                 }
 210                 spin_unlock(&smc_ib_devices.lock);
 211         }
 212         if (new_pnetelem->smcd_dev) {
 213                 struct smcd_dev *smcd_dev = new_pnetelem->smcd_dev;
 214
 215                 spin_lock(&smcd_dev_list.lock);
 216                 if (smc_pnet_match(smcd_dev->pnetid, pnet_null)) {
 217                         memcpy(smcd_dev->pnetid, new_pnetelem->pnet_name,
 218                                SMC_MAX_PNETID_LEN);
 219                         smcd_dev->pnetid_by_user = true;
 220                         new_smcddev = true;
 221                 }
 222                 spin_unlock(&smcd_dev_list.lock);
 223         }
 224
 225         if (!new_pnetelem->ndev)
 226                 return (new_ibdev || new_smcddev) ? 0 : -EEXIST;
 227
 228         /* check if (base) netdev already has a pnetid. If there is one, we do
 229          * not want to add a pnet table entry
 230          */
 231         ndev = pnet_find_base_ndev(new_pnetelem->ndev);
 232         if (!smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
 233                                     ndev_pnetid))
 234                 return (new_ibdev || new_smcddev) ? 0 : -EEXIST;
 235
 236         /* add a new netdev entry to the pnet table if there isn't one */
 237         tmp_pnetelem = kzalloc(sizeof(*pnetelem), GFP_KERNEL);
 238         if (!tmp_pnetelem)
 239                 return -ENOMEM;
 240         memcpy(tmp_pnetelem->pnet_name, new_pnetelem->pnet_name,
 241                SMC_MAX_PNETID_LEN);
 242         tmp_pnetelem->ndev = new_pnetelem->ndev;
 243
 244         write_lock(&pnettable->lock);
 245         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 246                 if (pnetelem->ndev == new_pnetelem->ndev)
 247                         new_netdev = false;
 248         }
 249         if (new_netdev) {
 250                 dev_hold(tmp_pnetelem->ndev);
 251                 list_add_tail(&tmp_pnetelem->list, &pnettable->pnetlist);
 252                 write_unlock(&pnettable->lock);
 253         } else {
 254                 write_unlock(&pnettable->lock);
 255                 kfree(tmp_pnetelem);
 256         }
 257
 258         return (new_netdev || new_ibdev || new_smcddev) ? 0 : -EEXIST;
 259 }
 260
 261 /* The limit for pnetid is 16 characters.
 262  * Valid characters should be (single-byte character set) a-z, A-Z, 0-9.
 263  * Lower case letters are converted to upper case.
 264  * Interior blanks should not be used.
 265  */
 266 static bool smc_pnetid_valid(const char *pnet_name, char *pnetid)
 267 {
 268         char *bf = skip_spaces(pnet_name);
 269         size_t len = strlen(bf);
 270         char *end = bf + len;
 271
 272         if (!len)
 273                 return false;
 274         while (--end >= bf && isspace(*end))
 275                 ;
 276         if (end - bf >= SMC_MAX_PNETID_LEN)
 277                 return false;
 278         while (bf <= end) {
 279                 if (!isalnum(*bf))
 280                         return false;
 281                 *pnetid++ = islower(*bf) ? toupper(*bf) : *bf;
 282                 bf++;
 283         }
 284         *pnetid = '\0';
 285         return true;
 286 }
 287
 288 /* Find an infiniband device by a given name. The device might not exist. */
 289 static struct smc_ib_device *smc_pnet_find_ib(char *ib_name)
 290 {
 291         struct smc_ib_device *ibdev;
 292
 293         spin_lock(&smc_ib_devices.lock);
 294         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 295                 if (!strncmp(ibdev->ibdev->name, ib_name,
 296                              sizeof(ibdev->ibdev->name)) ||
 297                     !strncmp(dev_name(ibdev->ibdev->dev.parent), ib_name,
 298                              IB_DEVICE_NAME_MAX - 1)) {
 299                         goto out;
 300                 }
 301         }
 302         ibdev = NULL;
 303 out:
 304         spin_unlock(&smc_ib_devices.lock);
 305         return ibdev;
 306 }
 307
 308 /* Find an smcd device by a given name. The device might not exist. */
 309 static struct smcd_dev *smc_pnet_find_smcd(char *smcd_name)
 310 {
 311         struct smcd_dev *smcd_dev;
 312
 313         spin_lock(&smcd_dev_list.lock);
 314         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 315                 if (!strncmp(dev_name(&smcd_dev->dev), smcd_name,
 316                              IB_DEVICE_NAME_MAX - 1))
 317                         goto out;
 318         }
 319         smcd_dev = NULL;
 320 out:
 321         spin_unlock(&smcd_dev_list.lock);
 322         return smcd_dev;
 323 }
 324
 325 /* Parse the supplied netlink attributes and fill a pnetentry structure.
 326  * For ethernet and infiniband device names verify that the devices exist.
 327  */
 328 static int smc_pnet_fill_entry(struct net *net,
 329                                struct smc_user_pnetentry *pnetelem,
 330                                struct nlattr *tb[])
 331 {
 332         char *string, *ibname;
 333         int rc;
 334
 335         memset(pnetelem, 0, sizeof(*pnetelem));
 336         INIT_LIST_HEAD(&pnetelem->list);
 337
 338         rc = -EINVAL;
 339         if (!tb[SMC_PNETID_NAME])
 340                 goto error;
 341         string = (char *)nla_data(tb[SMC_PNETID_NAME]);
 342         if (!smc_pnetid_valid(string, pnetelem->pnet_name))
 343                 goto error;
 344
 345         rc = -EINVAL;
 346         if (tb[SMC_PNETID_ETHNAME]) {
 347                 string = (char *)nla_data(tb[SMC_PNETID_ETHNAME]);
 348                 pnetelem->ndev = dev_get_by_name(net, string);
 349                 if (!pnetelem->ndev)
 350                         goto error;
 351         }
 352
 353         /* if this is not the initial namespace, stop here */
 354         if (net != &init_net)
 355                 return 0;
 356
 357         rc = -EINVAL;
 358         if (tb[SMC_PNETID_IBNAME]) {
 359                 ibname = (char *)nla_data(tb[SMC_PNETID_IBNAME]);
 360                 ibname = strim(ibname);
 361                 pnetelem->smcibdev = smc_pnet_find_ib(ibname);
 362                 pnetelem->smcd_dev = smc_pnet_find_smcd(ibname);
 363                 if (!pnetelem->smcibdev && !pnetelem->smcd_dev)
 364                         goto error;
 365                 if (pnetelem->smcibdev) {
 366                         if (!tb[SMC_PNETID_IBPORT])
 367                                 goto error;
 368                         pnetelem->ib_port = nla_get_u8(tb[SMC_PNETID_IBPORT]);
 369                         if (pnetelem->ib_port < 1 ||
 370                             pnetelem->ib_port > SMC_MAX_PORTS)
 371                                 goto error;
 372                 }
 373         }
 374
 375         return 0;
 376
 377 error:
 378         if (pnetelem->ndev)
 379                 dev_put(pnetelem->ndev);
 380         return rc;
 381 }
 382
 383 /* Convert an smc_pnetentry to a netlink attribute sequence */
 384 static int smc_pnet_set_nla(struct sk_buff *msg,
 385                             struct smc_user_pnetentry *pnetelem)
 386 {
 387         if (nla_put_string(msg, SMC_PNETID_NAME, pnetelem->pnet_name))
 388                 return -1;
 389         if (pnetelem->ndev) {
 390                 if (nla_put_string(msg, SMC_PNETID_ETHNAME,
 391                                    pnetelem->ndev->name))
 392                         return -1;
 393         } else {
 394                 if (nla_put_string(msg, SMC_PNETID_ETHNAME, "n/a"))
 395                         return -1;
 396         }
 397         if (pnetelem->smcibdev) {
 398                 if (nla_put_string(msg, SMC_PNETID_IBNAME,
 399                         dev_name(pnetelem->smcibdev->ibdev->dev.parent)) ||
 400                     nla_put_u8(msg, SMC_PNETID_IBPORT, pnetelem->ib_port))
 401                         return -1;
 402         } else if (pnetelem->smcd_dev) {
 403                 if (nla_put_string(msg, SMC_PNETID_IBNAME,
 404                                    dev_name(&pnetelem->smcd_dev->dev)) ||
 405                     nla_put_u8(msg, SMC_PNETID_IBPORT, 1))
 406                         return -1;
 407         } else {
 408                 if (nla_put_string(msg, SMC_PNETID_IBNAME, "n/a") ||
 409                     nla_put_u8(msg, SMC_PNETID_IBPORT, 0xff))
 410                         return -1;
 411         }
 412
 413         return 0;
 414 }
 415
 416 static int smc_pnet_add(struct sk_buff *skb, struct genl_info *info)
 417 {
 418         struct net *net = genl_info_net(info);
 419         struct smc_user_pnetentry pnetelem;
 420         struct smc_pnettable *pnettable;
 421         struct smc_net *sn;
 422         int rc;
 423
 424         /* get pnettable for namespace */
 425         sn = net_generic(net, smc_net_id);
 426         pnettable = &sn->pnettable;
 427
 428         rc = smc_pnet_fill_entry(net, &pnetelem, info->attrs);
 429         if (!rc)
 430                 rc = smc_pnet_enter(pnettable, &pnetelem);
 431         if (pnetelem.ndev)
 432                 dev_put(pnetelem.ndev);
 433         return rc;
 434 }
 435
 436 static int smc_pnet_del(struct sk_buff *skb, struct genl_info *info)
 437 {
 438         struct net *net = genl_info_net(info);
 439
 440         if (!info->attrs[SMC_PNETID_NAME])
 441                 return -EINVAL;
 442         return smc_pnet_remove_by_pnetid(net,
 443                                 (char *)nla_data(info->attrs[SMC_PNETID_NAME]));
 444 }
 445
 446 static int smc_pnet_dump_start(struct netlink_callback *cb)
 447 {
 448         cb->args[0] = 0;
 449         return 0;
 450 }
 451
 452 static int smc_pnet_dumpinfo(struct sk_buff *skb,
 453                              u32 portid, u32 seq, u32 flags,
 454                              struct smc_user_pnetentry *pnetelem)
 455 {
 456         void *hdr;
 457
 458         hdr = genlmsg_put(skb, portid, seq, &smc_pnet_nl_family,
 459                           flags, SMC_PNETID_GET);
 460         if (!hdr)
 461                 return -ENOMEM;
 462         if (smc_pnet_set_nla(skb, pnetelem) < 0) {
 463                 genlmsg_cancel(skb, hdr);
 464                 return -EMSGSIZE;
 465         }
 466         genlmsg_end(skb, hdr);
 467         return 0;
 468 }
 469
 470 static int _smc_pnet_dump(struct net *net, struct sk_buff *skb, u32 portid,
 471                           u32 seq, u8 *pnetid, int start_idx)
 472 {
 473         struct smc_user_pnetentry tmp_entry;
 474         struct smc_pnettable *pnettable;
 475         struct smc_pnetentry *pnetelem;
 476         struct smc_ib_device *ibdev;
 477         struct smcd_dev *smcd_dev;
 478         struct smc_net *sn;
 479         int idx = 0;
 480         int ibport;
 481
 482         /* get pnettable for namespace */
 483         sn = net_generic(net, smc_net_id);
 484         pnettable = &sn->pnettable;
 485
 486         /* dump netdevices */
 487         read_lock(&pnettable->lock);
 488         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 489                 if (pnetid && !smc_pnet_match(pnetelem->pnet_name, pnetid))
 490                         continue;
 491                 if (idx++ < start_idx)
 492                         continue;
 493                 memset(&tmp_entry, 0, sizeof(tmp_entry));
 494                 memcpy(&tmp_entry.pnet_name, pnetelem->pnet_name,
 495                        SMC_MAX_PNETID_LEN);
 496                 tmp_entry.ndev = pnetelem->ndev;
 497                 if (smc_pnet_dumpinfo(skb, portid, seq, NLM_F_MULTI,
 498                                       &tmp_entry)) {
 499                         --idx;
 500                         break;
 501                 }
 502         }
 503         read_unlock(&pnettable->lock);
 504
 505         /* if this is not the initial namespace, stop here */
 506         if (net != &init_net)
 507                 return idx;
 508
 509         /* dump ib devices */
 510         spin_lock(&smc_ib_devices.lock);
 511         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 512                 for (ibport = 0; ibport < SMC_MAX_PORTS; ibport++) {
 513                         if (ibdev->pnetid_by_user[ibport]) {
 514                                 if (pnetid &&
 515                                     !smc_pnet_match(ibdev->pnetid[ibport],
 516                                                     pnetid))
 517                                         continue;
 518                                 if (idx++ < start_idx)
 519                                         continue;
 520                                 memset(&tmp_entry, 0, sizeof(tmp_entry));
 521                                 memcpy(&tmp_entry.pnet_name,
 522                                        ibdev->pnetid[ibport],
 523                                        SMC_MAX_PNETID_LEN);
 524                                 tmp_entry.smcibdev = ibdev;
 525                                 tmp_entry.ib_port = ibport + 1;
 526                                 if (smc_pnet_dumpinfo(skb, portid, seq,
 527                                                       NLM_F_MULTI,
 528                                                       &tmp_entry)) {
 529                                         --idx;
 530                                         break;
 531                                 }
 532                         }
 533                 }
 534         }
 535         spin_unlock(&smc_ib_devices.lock);
 536
 537         /* dump smcd devices */
 538         spin_lock(&smcd_dev_list.lock);
 539         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 540                 if (smcd_dev->pnetid_by_user) {
 541                         if (pnetid && !smc_pnet_match(smcd_dev->pnetid, pnetid))
 542                                 continue;
 543                         if (idx++ < start_idx)
 544                                 continue;
 545                         memset(&tmp_entry, 0, sizeof(tmp_entry));
 546                         memcpy(&tmp_entry.pnet_name, smcd_dev->pnetid,
 547                                SMC_MAX_PNETID_LEN);
 548                         tmp_entry.smcd_dev = smcd_dev;
 549                         if (smc_pnet_dumpinfo(skb, portid, seq, NLM_F_MULTI,
 550                                               &tmp_entry)) {
 551                                 --idx;
 552                                 break;
 553                         }
 554                 }
 555         }
 556         spin_unlock(&smcd_dev_list.lock);
 557
 558         return idx;
 559 }
 560
 561 static int smc_pnet_dump(struct sk_buff *skb, struct netlink_callback *cb)
 562 {
 563         struct net *net = sock_net(skb->sk);
 564         int idx;
 565
 566         idx = _smc_pnet_dump(net, skb, NETLINK_CB(cb->skb).portid,
 567                              cb->nlh->nlmsg_seq, NULL, cb->args[0]);
 568
 569         cb->args[0] = idx;
 570         return skb->len;
 571 }
 572
 573 /* Retrieve one PNETID entry */
 574 static int smc_pnet_get(struct sk_buff *skb, struct genl_info *info)
 575 {
 576         struct net *net = genl_info_net(info);
 577         struct sk_buff *msg;
 578         void *hdr;
 579
 580         if (!info->attrs[SMC_PNETID_NAME])
 581                 return -EINVAL;
 582
 583         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 584         if (!msg)
 585                 return -ENOMEM;
 586
 587         _smc_pnet_dump(net, msg, info->snd_portid, info->snd_seq,
 588                        nla_data(info->attrs[SMC_PNETID_NAME]), 0);
 589
 590         /* finish multi part message and send it */
 591         hdr = nlmsg_put(msg, info->snd_portid, info->snd_seq, NLMSG_DONE, 0,
 592                         NLM_F_MULTI);
 593         if (!hdr) {
 594                 nlmsg_free(msg);
 595                 return -EMSGSIZE;
 596         }
 597         return genlmsg_reply(msg, info);
 598 }
 599
 600 /* Remove and delete all pnetids from pnet table.
 601  */
 602 static int smc_pnet_flush(struct sk_buff *skb, struct genl_info *info)
 603 {
 604         struct net *net = genl_info_net(info);
 605
 606         smc_pnet_remove_by_pnetid(net, NULL);
 607         return 0;
 608 }
 609
 610 /* SMC_PNETID generic netlink operation definition */
 611 static const struct genl_ops smc_pnet_ops[] = {
 612         {
 613                 .cmd = SMC_PNETID_GET,
 614                 .flags = GENL_ADMIN_PERM,
 615                 .policy = smc_pnet_policy,
 616                 .doit = smc_pnet_get,
 617                 .dumpit = smc_pnet_dump,
 618                 .start = smc_pnet_dump_start
 619         },
 620         {
 621                 .cmd = SMC_PNETID_ADD,
 622                 .flags = GENL_ADMIN_PERM,
 623                 .policy = smc_pnet_policy,
 624                 .doit = smc_pnet_add
 625         },
 626         {
 627                 .cmd = SMC_PNETID_DEL,
 628                 .flags = GENL_ADMIN_PERM,
 629                 .policy = smc_pnet_policy,
 630                 .doit = smc_pnet_del
 631         },
 632         {
 633                 .cmd = SMC_PNETID_FLUSH,
 634                 .flags = GENL_ADMIN_PERM,
 635                 .policy = smc_pnet_policy,
 636                 .doit = smc_pnet_flush
 637         }
 638 };
 639
 640 /* SMC_PNETID family definition */
 641 static struct genl_family smc_pnet_nl_family __ro_after_init = {
 642         .hdrsize = 0,
 643         .name = SMCR_GENL_FAMILY_NAME,
 644         .version = SMCR_GENL_FAMILY_VERSION,
 645         .maxattr = SMC_PNETID_MAX,
 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 = dev_get_nest_level(ndev);
 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_ib_device **smcibdev,
 763                                    u8 *ibport, unsigned short vlan_id, u8 gid[])
 764 {
 765         struct smc_ib_device *ibdev;
 766
 767         spin_lock(&smc_ib_devices.lock);
 768         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 769                 struct net_device *ndev;
 770                 int i;
 771
 772                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
 773                         if (!rdma_is_port_valid(ibdev->ibdev, i))
 774                                 continue;
 775                         if (!ibdev->ibdev->ops.get_netdev)
 776                                 continue;
 777                         ndev = ibdev->ibdev->ops.get_netdev(ibdev->ibdev, i);
 778                         if (!ndev)
 779                                 continue;
 780                         dev_put(ndev);
 781                         if (netdev == ndev &&
 782                             smc_ib_port_active(ibdev, i) &&
 783                             !smc_ib_determine_gid(ibdev, i, vlan_id, gid,
 784                                                   NULL)) {
 785                                 *smcibdev = ibdev;
 786                                 *ibport = 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_ib_device **smcibdev,
 802                                          u8 *ibport, unsigned short vlan_id,
 803                                          u8 gid[])
 804 {
 805         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 806         struct smc_ib_device *ibdev;
 807         int i;
 808
 809         ndev = pnet_find_base_ndev(ndev);
 810         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
 811                                    ndev_pnetid) &&
 812             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid)) {
 813                 smc_pnet_find_rdma_dev(ndev, smcibdev, ibport, vlan_id, gid);
 814                 return; /* pnetid could not be determined */
 815         }
 816
 817         spin_lock(&smc_ib_devices.lock);
 818         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 819                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
 820                         if (!rdma_is_port_valid(ibdev->ibdev, i))
 821                                 continue;
 822                         if (smc_pnet_match(ibdev->pnetid[i - 1], ndev_pnetid) &&
 823                             smc_ib_port_active(ibdev, i) &&
 824                             !smc_ib_determine_gid(ibdev, i, vlan_id, gid,
 825                                                   NULL))  {
 826                                 *smcibdev = ibdev;
 827                                 *ibport = i;
 828                                 goto out;
 829                         }
 830                 }
 831         }
 832 out:
 833         spin_unlock(&smc_ib_devices.lock);
 834 }
 835
 836 static void smc_pnet_find_ism_by_pnetid(struct net_device *ndev,
 837                                         struct smcd_dev **smcismdev)
 838 {
 839         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 840         struct smcd_dev *ismdev;
 841
 842         ndev = pnet_find_base_ndev(ndev);
 843         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
 844                                    ndev_pnetid) &&
 845             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid))
 846                 return; /* pnetid could not be determined */
 847
 848         spin_lock(&smcd_dev_list.lock);
 849         list_for_each_entry(ismdev, &smcd_dev_list.list, list) {
 850                 if (smc_pnet_match(ismdev->pnetid, ndev_pnetid)) {
 851                         *smcismdev = 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,
 863                                  struct smc_ib_device **smcibdev, u8 *ibport,
 864                                  unsigned short vlan_id, u8 gid[])
 865 {
 866         struct dst_entry *dst = sk_dst_get(sk);
 867
 868         *smcibdev = NULL;
 869         *ibport = 0;
 870
 871         if (!dst)
 872                 goto out;
 873         if (!dst->dev)
 874                 goto out_rel;
 875
 876         smc_pnet_find_roce_by_pnetid(dst->dev, smcibdev, ibport, vlan_id, gid);
 877
 878 out_rel:
 879         dst_release(dst);
 880 out:
 881         return;
 882 }
 883
 884 void smc_pnet_find_ism_resource(struct sock *sk, struct smcd_dev **smcismdev)
 885 {
 886         struct dst_entry *dst = sk_dst_get(sk);
 887
 888         *smcismdev = NULL;
 889         if (!dst)
 890                 goto out;
 891         if (!dst->dev)
 892                 goto out_rel;
 893
 894         smc_pnet_find_ism_by_pnetid(dst->dev, smcismdev);
 895
 896 out_rel:
 897         dst_release(dst);
 898 out:
 899         return;
 900 }