Back to home page

OSCL-LXR
 
 

     

0001 // SPDX-License-Identifier: GPL-2.0
0002 /*
0003  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
0004  *
0005  *  Generic netlink support functions to configure an SMC-R PNET table
0006  *
0007  *  Copyright IBM Corp. 2016
0008  *
0009  *  Author(s):  Thomas Richter <tmricht@linux.vnet.ibm.com>
0010  */
0011 
0012 #include <linux/module.h>
0013 #include <linux/list.h>
0014 #include <linux/ctype.h>
0015 #include <linux/mutex.h>
0016 #include <net/netlink.h>
0017 #include <net/genetlink.h>
0018 
0019 #include <uapi/linux/if.h>
0020 #include <uapi/linux/smc.h>
0021 
0022 #include <rdma/ib_verbs.h>
0023 
0024 #include <net/netns/generic.h>
0025 #include "smc_netns.h"
0026 
0027 #include "smc_pnet.h"
0028 #include "smc_ib.h"
0029 #include "smc_ism.h"
0030 #include "smc_core.h"
0031 
0032 static struct net_device *__pnet_find_base_ndev(struct net_device *ndev);
0033 static struct net_device *pnet_find_base_ndev(struct net_device *ndev);
0034 
0035 static const struct nla_policy smc_pnet_policy[SMC_PNETID_MAX + 1] = {
0036     [SMC_PNETID_NAME] = {
0037         .type = NLA_NUL_STRING,
0038         .len = SMC_MAX_PNETID_LEN
0039     },
0040     [SMC_PNETID_ETHNAME] = {
0041         .type = NLA_NUL_STRING,
0042         .len = IFNAMSIZ - 1
0043     },
0044     [SMC_PNETID_IBNAME] = {
0045         .type = NLA_NUL_STRING,
0046         .len = IB_DEVICE_NAME_MAX - 1
0047     },
0048     [SMC_PNETID_IBPORT] = { .type = NLA_U8 }
0049 };
0050 
0051 static struct genl_family smc_pnet_nl_family;
0052 
0053 enum smc_pnet_nametype {
0054     SMC_PNET_ETH    = 1,
0055     SMC_PNET_IB = 2,
0056 };
0057 
0058 /* pnet entry stored in pnet table */
0059 struct smc_pnetentry {
0060     struct list_head list;
0061     char pnet_name[SMC_MAX_PNETID_LEN + 1];
0062     enum smc_pnet_nametype type;
0063     union {
0064         struct {
0065             char eth_name[IFNAMSIZ + 1];
0066             struct net_device *ndev;
0067             netdevice_tracker dev_tracker;
0068         };
0069         struct {
0070             char ib_name[IB_DEVICE_NAME_MAX + 1];
0071             u8 ib_port;
0072         };
0073     };
0074 };
0075 
0076 /* Check if the pnetid is set */
0077 bool smc_pnet_is_pnetid_set(u8 *pnetid)
0078 {
0079     if (pnetid[0] == 0 || pnetid[0] == _S)
0080         return false;
0081     return true;
0082 }
0083 
0084 /* Check if two given pnetids match */
0085 static bool smc_pnet_match(u8 *pnetid1, u8 *pnetid2)
0086 {
0087     int i;
0088 
0089     for (i = 0; i < SMC_MAX_PNETID_LEN; i++) {
0090         if ((pnetid1[i] == 0 || pnetid1[i] == _S) &&
0091             (pnetid2[i] == 0 || pnetid2[i] == _S))
0092             break;
0093         if (pnetid1[i] != pnetid2[i])
0094             return false;
0095     }
0096     return true;
0097 }
0098 
0099 /* Remove a pnetid from the pnet table.
0100  */
0101 static int smc_pnet_remove_by_pnetid(struct net *net, char *pnet_name)
0102 {
0103     struct smc_pnetentry *pnetelem, *tmp_pe;
0104     struct smc_pnettable *pnettable;
0105     struct smc_ib_device *ibdev;
0106     struct smcd_dev *smcd_dev;
0107     struct smc_net *sn;
0108     int rc = -ENOENT;
0109     int ibport;
0110 
0111     /* get pnettable for namespace */
0112     sn = net_generic(net, smc_net_id);
0113     pnettable = &sn->pnettable;
0114 
0115     /* remove table entry */
0116     mutex_lock(&pnettable->lock);
0117     list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist,
0118                  list) {
0119         if (!pnet_name ||
0120             smc_pnet_match(pnetelem->pnet_name, pnet_name)) {
0121             list_del(&pnetelem->list);
0122             if (pnetelem->type == SMC_PNET_ETH && pnetelem->ndev) {
0123                 netdev_put(pnetelem->ndev,
0124                        &pnetelem->dev_tracker);
0125                 pr_warn_ratelimited("smc: net device %s "
0126                             "erased user defined "
0127                             "pnetid %.16s\n",
0128                             pnetelem->eth_name,
0129                             pnetelem->pnet_name);
0130             }
0131             kfree(pnetelem);
0132             rc = 0;
0133         }
0134     }
0135     mutex_unlock(&pnettable->lock);
0136 
0137     /* if this is not the initial namespace, stop here */
0138     if (net != &init_net)
0139         return rc;
0140 
0141     /* remove ib devices */
0142     mutex_lock(&smc_ib_devices.mutex);
0143     list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
0144         for (ibport = 0; ibport < SMC_MAX_PORTS; ibport++) {
0145             if (ibdev->pnetid_by_user[ibport] &&
0146                 (!pnet_name ||
0147                  smc_pnet_match(pnet_name,
0148                         ibdev->pnetid[ibport]))) {
0149                 pr_warn_ratelimited("smc: ib device %s ibport "
0150                             "%d erased user defined "
0151                             "pnetid %.16s\n",
0152                             ibdev->ibdev->name,
0153                             ibport + 1,
0154                             ibdev->pnetid[ibport]);
0155                 memset(ibdev->pnetid[ibport], 0,
0156                        SMC_MAX_PNETID_LEN);
0157                 ibdev->pnetid_by_user[ibport] = false;
0158                 rc = 0;
0159             }
0160         }
0161     }
0162     mutex_unlock(&smc_ib_devices.mutex);
0163     /* remove smcd devices */
0164     mutex_lock(&smcd_dev_list.mutex);
0165     list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
0166         if (smcd_dev->pnetid_by_user &&
0167             (!pnet_name ||
0168              smc_pnet_match(pnet_name, smcd_dev->pnetid))) {
0169             pr_warn_ratelimited("smc: smcd device %s "
0170                         "erased user defined pnetid "
0171                         "%.16s\n", dev_name(&smcd_dev->dev),
0172                         smcd_dev->pnetid);
0173             memset(smcd_dev->pnetid, 0, SMC_MAX_PNETID_LEN);
0174             smcd_dev->pnetid_by_user = false;
0175             rc = 0;
0176         }
0177     }
0178     mutex_unlock(&smcd_dev_list.mutex);
0179     return rc;
0180 }
0181 
0182 /* Add the reference to a given network device to the pnet table.
0183  */
0184 static int smc_pnet_add_by_ndev(struct net_device *ndev)
0185 {
0186     struct smc_pnetentry *pnetelem, *tmp_pe;
0187     struct smc_pnettable *pnettable;
0188     struct net *net = dev_net(ndev);
0189     struct smc_net *sn;
0190     int rc = -ENOENT;
0191 
0192     /* get pnettable for namespace */
0193     sn = net_generic(net, smc_net_id);
0194     pnettable = &sn->pnettable;
0195 
0196     mutex_lock(&pnettable->lock);
0197     list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
0198         if (pnetelem->type == SMC_PNET_ETH && !pnetelem->ndev &&
0199             !strncmp(pnetelem->eth_name, ndev->name, IFNAMSIZ)) {
0200             netdev_hold(ndev, &pnetelem->dev_tracker, GFP_ATOMIC);
0201             pnetelem->ndev = ndev;
0202             rc = 0;
0203             pr_warn_ratelimited("smc: adding net device %s with "
0204                         "user defined pnetid %.16s\n",
0205                         pnetelem->eth_name,
0206                         pnetelem->pnet_name);
0207             break;
0208         }
0209     }
0210     mutex_unlock(&pnettable->lock);
0211     return rc;
0212 }
0213 
0214 /* Remove the reference to a given network device from the pnet table.
0215  */
0216 static int smc_pnet_remove_by_ndev(struct net_device *ndev)
0217 {
0218     struct smc_pnetentry *pnetelem, *tmp_pe;
0219     struct smc_pnettable *pnettable;
0220     struct net *net = dev_net(ndev);
0221     struct smc_net *sn;
0222     int rc = -ENOENT;
0223 
0224     /* get pnettable for namespace */
0225     sn = net_generic(net, smc_net_id);
0226     pnettable = &sn->pnettable;
0227 
0228     mutex_lock(&pnettable->lock);
0229     list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
0230         if (pnetelem->type == SMC_PNET_ETH && pnetelem->ndev == ndev) {
0231             netdev_put(pnetelem->ndev, &pnetelem->dev_tracker);
0232             pnetelem->ndev = NULL;
0233             rc = 0;
0234             pr_warn_ratelimited("smc: removing net device %s with "
0235                         "user defined pnetid %.16s\n",
0236                         pnetelem->eth_name,
0237                         pnetelem->pnet_name);
0238             break;
0239         }
0240     }
0241     mutex_unlock(&pnettable->lock);
0242     return rc;
0243 }
0244 
0245 /* Apply pnetid to ib device when no pnetid is set.
0246  */
0247 static bool smc_pnet_apply_ib(struct smc_ib_device *ib_dev, u8 ib_port,
0248                   char *pnet_name)
0249 {
0250     bool applied = false;
0251 
0252     mutex_lock(&smc_ib_devices.mutex);
0253     if (!smc_pnet_is_pnetid_set(ib_dev->pnetid[ib_port - 1])) {
0254         memcpy(ib_dev->pnetid[ib_port - 1], pnet_name,
0255                SMC_MAX_PNETID_LEN);
0256         ib_dev->pnetid_by_user[ib_port - 1] = true;
0257         applied = true;
0258     }
0259     mutex_unlock(&smc_ib_devices.mutex);
0260     return applied;
0261 }
0262 
0263 /* Apply pnetid to smcd device when no pnetid is set.
0264  */
0265 static bool smc_pnet_apply_smcd(struct smcd_dev *smcd_dev, char *pnet_name)
0266 {
0267     bool applied = false;
0268 
0269     mutex_lock(&smcd_dev_list.mutex);
0270     if (!smc_pnet_is_pnetid_set(smcd_dev->pnetid)) {
0271         memcpy(smcd_dev->pnetid, pnet_name, SMC_MAX_PNETID_LEN);
0272         smcd_dev->pnetid_by_user = true;
0273         applied = true;
0274     }
0275     mutex_unlock(&smcd_dev_list.mutex);
0276     return applied;
0277 }
0278 
0279 /* The limit for pnetid is 16 characters.
0280  * Valid characters should be (single-byte character set) a-z, A-Z, 0-9.
0281  * Lower case letters are converted to upper case.
0282  * Interior blanks should not be used.
0283  */
0284 static bool smc_pnetid_valid(const char *pnet_name, char *pnetid)
0285 {
0286     char *bf = skip_spaces(pnet_name);
0287     size_t len = strlen(bf);
0288     char *end = bf + len;
0289 
0290     if (!len)
0291         return false;
0292     while (--end >= bf && isspace(*end))
0293         ;
0294     if (end - bf >= SMC_MAX_PNETID_LEN)
0295         return false;
0296     while (bf <= end) {
0297         if (!isalnum(*bf))
0298             return false;
0299         *pnetid++ = islower(*bf) ? toupper(*bf) : *bf;
0300         bf++;
0301     }
0302     *pnetid = '\0';
0303     return true;
0304 }
0305 
0306 /* Find an infiniband device by a given name. The device might not exist. */
0307 static struct smc_ib_device *smc_pnet_find_ib(char *ib_name)
0308 {
0309     struct smc_ib_device *ibdev;
0310 
0311     mutex_lock(&smc_ib_devices.mutex);
0312     list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
0313         if (!strncmp(ibdev->ibdev->name, ib_name,
0314                  sizeof(ibdev->ibdev->name)) ||
0315             (ibdev->ibdev->dev.parent &&
0316              !strncmp(dev_name(ibdev->ibdev->dev.parent), ib_name,
0317                  IB_DEVICE_NAME_MAX - 1))) {
0318             goto out;
0319         }
0320     }
0321     ibdev = NULL;
0322 out:
0323     mutex_unlock(&smc_ib_devices.mutex);
0324     return ibdev;
0325 }
0326 
0327 /* Find an smcd device by a given name. The device might not exist. */
0328 static struct smcd_dev *smc_pnet_find_smcd(char *smcd_name)
0329 {
0330     struct smcd_dev *smcd_dev;
0331 
0332     mutex_lock(&smcd_dev_list.mutex);
0333     list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
0334         if (!strncmp(dev_name(&smcd_dev->dev), smcd_name,
0335                  IB_DEVICE_NAME_MAX - 1))
0336             goto out;
0337     }
0338     smcd_dev = NULL;
0339 out:
0340     mutex_unlock(&smcd_dev_list.mutex);
0341     return smcd_dev;
0342 }
0343 
0344 static int smc_pnet_add_eth(struct smc_pnettable *pnettable, struct net *net,
0345                 char *eth_name, char *pnet_name)
0346 {
0347     struct smc_pnetentry *tmp_pe, *new_pe;
0348     struct net_device *ndev, *base_ndev;
0349     u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
0350     bool new_netdev;
0351     int rc;
0352 
0353     /* check if (base) netdev already has a pnetid. If there is one, we do
0354      * not want to add a pnet table entry
0355      */
0356     rc = -EEXIST;
0357     ndev = dev_get_by_name(net, eth_name);  /* dev_hold() */
0358     if (ndev) {
0359         base_ndev = pnet_find_base_ndev(ndev);
0360         if (!smc_pnetid_by_dev_port(base_ndev->dev.parent,
0361                         base_ndev->dev_port, ndev_pnetid))
0362             goto out_put;
0363     }
0364 
0365     /* add a new netdev entry to the pnet table if there isn't one */
0366     rc = -ENOMEM;
0367     new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
0368     if (!new_pe)
0369         goto out_put;
0370     new_pe->type = SMC_PNET_ETH;
0371     memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
0372     strncpy(new_pe->eth_name, eth_name, IFNAMSIZ);
0373     rc = -EEXIST;
0374     new_netdev = true;
0375     mutex_lock(&pnettable->lock);
0376     list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
0377         if (tmp_pe->type == SMC_PNET_ETH &&
0378             !strncmp(tmp_pe->eth_name, eth_name, IFNAMSIZ)) {
0379             new_netdev = false;
0380             break;
0381         }
0382     }
0383     if (new_netdev) {
0384         if (ndev) {
0385             new_pe->ndev = ndev;
0386             netdev_tracker_alloc(ndev, &new_pe->dev_tracker,
0387                          GFP_ATOMIC);
0388         }
0389         list_add_tail(&new_pe->list, &pnettable->pnetlist);
0390         mutex_unlock(&pnettable->lock);
0391     } else {
0392         mutex_unlock(&pnettable->lock);
0393         kfree(new_pe);
0394         goto out_put;
0395     }
0396     if (ndev)
0397         pr_warn_ratelimited("smc: net device %s "
0398                     "applied user defined pnetid %.16s\n",
0399                     new_pe->eth_name, new_pe->pnet_name);
0400     return 0;
0401 
0402 out_put:
0403     dev_put(ndev);
0404     return rc;
0405 }
0406 
0407 static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
0408                u8 ib_port, char *pnet_name)
0409 {
0410     struct smc_pnetentry *tmp_pe, *new_pe;
0411     struct smc_ib_device *ib_dev;
0412     bool smcddev_applied = true;
0413     bool ibdev_applied = true;
0414     struct smcd_dev *smcd_dev;
0415     bool new_ibdev;
0416 
0417     /* try to apply the pnetid to active devices */
0418     ib_dev = smc_pnet_find_ib(ib_name);
0419     if (ib_dev) {
0420         ibdev_applied = smc_pnet_apply_ib(ib_dev, ib_port, pnet_name);
0421         if (ibdev_applied)
0422             pr_warn_ratelimited("smc: ib device %s ibport %d "
0423                         "applied user defined pnetid "
0424                         "%.16s\n", ib_dev->ibdev->name,
0425                         ib_port,
0426                         ib_dev->pnetid[ib_port - 1]);
0427     }
0428     smcd_dev = smc_pnet_find_smcd(ib_name);
0429     if (smcd_dev) {
0430         smcddev_applied = smc_pnet_apply_smcd(smcd_dev, pnet_name);
0431         if (smcddev_applied)
0432             pr_warn_ratelimited("smc: smcd device %s "
0433                         "applied user defined pnetid "
0434                         "%.16s\n", dev_name(&smcd_dev->dev),
0435                         smcd_dev->pnetid);
0436     }
0437     /* Apply fails when a device has a hardware-defined pnetid set, do not
0438      * add a pnet table entry in that case.
0439      */
0440     if (!ibdev_applied || !smcddev_applied)
0441         return -EEXIST;
0442 
0443     /* add a new ib entry to the pnet table if there isn't one */
0444     new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
0445     if (!new_pe)
0446         return -ENOMEM;
0447     new_pe->type = SMC_PNET_IB;
0448     memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
0449     strncpy(new_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX);
0450     new_pe->ib_port = ib_port;
0451 
0452     new_ibdev = true;
0453     mutex_lock(&pnettable->lock);
0454     list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
0455         if (tmp_pe->type == SMC_PNET_IB &&
0456             !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
0457             new_ibdev = false;
0458             break;
0459         }
0460     }
0461     if (new_ibdev) {
0462         list_add_tail(&new_pe->list, &pnettable->pnetlist);
0463         mutex_unlock(&pnettable->lock);
0464     } else {
0465         mutex_unlock(&pnettable->lock);
0466         kfree(new_pe);
0467     }
0468     return (new_ibdev) ? 0 : -EEXIST;
0469 }
0470 
0471 /* Append a pnetid to the end of the pnet table if not already on this list.
0472  */
0473 static int smc_pnet_enter(struct net *net, struct nlattr *tb[])
0474 {
0475     char pnet_name[SMC_MAX_PNETID_LEN + 1];
0476     struct smc_pnettable *pnettable;
0477     bool new_netdev = false;
0478     bool new_ibdev = false;
0479     struct smc_net *sn;
0480     u8 ibport = 1;
0481     char *string;
0482     int rc;
0483 
0484     /* get pnettable for namespace */
0485     sn = net_generic(net, smc_net_id);
0486     pnettable = &sn->pnettable;
0487 
0488     rc = -EINVAL;
0489     if (!tb[SMC_PNETID_NAME])
0490         goto error;
0491     string = (char *)nla_data(tb[SMC_PNETID_NAME]);
0492     if (!smc_pnetid_valid(string, pnet_name))
0493         goto error;
0494 
0495     if (tb[SMC_PNETID_ETHNAME]) {
0496         string = (char *)nla_data(tb[SMC_PNETID_ETHNAME]);
0497         rc = smc_pnet_add_eth(pnettable, net, string, pnet_name);
0498         if (!rc)
0499             new_netdev = true;
0500         else if (rc != -EEXIST)
0501             goto error;
0502     }
0503 
0504     /* if this is not the initial namespace, stop here */
0505     if (net != &init_net)
0506         return new_netdev ? 0 : -EEXIST;
0507 
0508     rc = -EINVAL;
0509     if (tb[SMC_PNETID_IBNAME]) {
0510         string = (char *)nla_data(tb[SMC_PNETID_IBNAME]);
0511         string = strim(string);
0512         if (tb[SMC_PNETID_IBPORT]) {
0513             ibport = nla_get_u8(tb[SMC_PNETID_IBPORT]);
0514             if (ibport < 1 || ibport > SMC_MAX_PORTS)
0515                 goto error;
0516         }
0517         rc = smc_pnet_add_ib(pnettable, string, ibport, pnet_name);
0518         if (!rc)
0519             new_ibdev = true;
0520         else if (rc != -EEXIST)
0521             goto error;
0522     }
0523     return (new_netdev || new_ibdev) ? 0 : -EEXIST;
0524 
0525 error:
0526     return rc;
0527 }
0528 
0529 /* Convert an smc_pnetentry to a netlink attribute sequence */
0530 static int smc_pnet_set_nla(struct sk_buff *msg,
0531                 struct smc_pnetentry *pnetelem)
0532 {
0533     if (nla_put_string(msg, SMC_PNETID_NAME, pnetelem->pnet_name))
0534         return -1;
0535     if (pnetelem->type == SMC_PNET_ETH) {
0536         if (nla_put_string(msg, SMC_PNETID_ETHNAME,
0537                    pnetelem->eth_name))
0538             return -1;
0539     } else {
0540         if (nla_put_string(msg, SMC_PNETID_ETHNAME, "n/a"))
0541             return -1;
0542     }
0543     if (pnetelem->type == SMC_PNET_IB) {
0544         if (nla_put_string(msg, SMC_PNETID_IBNAME, pnetelem->ib_name) ||
0545             nla_put_u8(msg, SMC_PNETID_IBPORT, pnetelem->ib_port))
0546             return -1;
0547     } else {
0548         if (nla_put_string(msg, SMC_PNETID_IBNAME, "n/a") ||
0549             nla_put_u8(msg, SMC_PNETID_IBPORT, 0xff))
0550             return -1;
0551     }
0552 
0553     return 0;
0554 }
0555 
0556 static int smc_pnet_add(struct sk_buff *skb, struct genl_info *info)
0557 {
0558     struct net *net = genl_info_net(info);
0559 
0560     return smc_pnet_enter(net, info->attrs);
0561 }
0562 
0563 static int smc_pnet_del(struct sk_buff *skb, struct genl_info *info)
0564 {
0565     struct net *net = genl_info_net(info);
0566 
0567     if (!info->attrs[SMC_PNETID_NAME])
0568         return -EINVAL;
0569     return smc_pnet_remove_by_pnetid(net,
0570                 (char *)nla_data(info->attrs[SMC_PNETID_NAME]));
0571 }
0572 
0573 static int smc_pnet_dump_start(struct netlink_callback *cb)
0574 {
0575     cb->args[0] = 0;
0576     return 0;
0577 }
0578 
0579 static int smc_pnet_dumpinfo(struct sk_buff *skb,
0580                  u32 portid, u32 seq, u32 flags,
0581                  struct smc_pnetentry *pnetelem)
0582 {
0583     void *hdr;
0584 
0585     hdr = genlmsg_put(skb, portid, seq, &smc_pnet_nl_family,
0586               flags, SMC_PNETID_GET);
0587     if (!hdr)
0588         return -ENOMEM;
0589     if (smc_pnet_set_nla(skb, pnetelem) < 0) {
0590         genlmsg_cancel(skb, hdr);
0591         return -EMSGSIZE;
0592     }
0593     genlmsg_end(skb, hdr);
0594     return 0;
0595 }
0596 
0597 static int _smc_pnet_dump(struct net *net, struct sk_buff *skb, u32 portid,
0598               u32 seq, u8 *pnetid, int start_idx)
0599 {
0600     struct smc_pnettable *pnettable;
0601     struct smc_pnetentry *pnetelem;
0602     struct smc_net *sn;
0603     int idx = 0;
0604 
0605     /* get pnettable for namespace */
0606     sn = net_generic(net, smc_net_id);
0607     pnettable = &sn->pnettable;
0608 
0609     /* dump pnettable entries */
0610     mutex_lock(&pnettable->lock);
0611     list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
0612         if (pnetid && !smc_pnet_match(pnetelem->pnet_name, pnetid))
0613             continue;
0614         if (idx++ < start_idx)
0615             continue;
0616         /* if this is not the initial namespace, dump only netdev */
0617         if (net != &init_net && pnetelem->type != SMC_PNET_ETH)
0618             continue;
0619         if (smc_pnet_dumpinfo(skb, portid, seq, NLM_F_MULTI,
0620                       pnetelem)) {
0621             --idx;
0622             break;
0623         }
0624     }
0625     mutex_unlock(&pnettable->lock);
0626     return idx;
0627 }
0628 
0629 static int smc_pnet_dump(struct sk_buff *skb, struct netlink_callback *cb)
0630 {
0631     struct net *net = sock_net(skb->sk);
0632     int idx;
0633 
0634     idx = _smc_pnet_dump(net, skb, NETLINK_CB(cb->skb).portid,
0635                  cb->nlh->nlmsg_seq, NULL, cb->args[0]);
0636 
0637     cb->args[0] = idx;
0638     return skb->len;
0639 }
0640 
0641 /* Retrieve one PNETID entry */
0642 static int smc_pnet_get(struct sk_buff *skb, struct genl_info *info)
0643 {
0644     struct net *net = genl_info_net(info);
0645     struct sk_buff *msg;
0646     void *hdr;
0647 
0648     if (!info->attrs[SMC_PNETID_NAME])
0649         return -EINVAL;
0650 
0651     msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
0652     if (!msg)
0653         return -ENOMEM;
0654 
0655     _smc_pnet_dump(net, msg, info->snd_portid, info->snd_seq,
0656                nla_data(info->attrs[SMC_PNETID_NAME]), 0);
0657 
0658     /* finish multi part message and send it */
0659     hdr = nlmsg_put(msg, info->snd_portid, info->snd_seq, NLMSG_DONE, 0,
0660             NLM_F_MULTI);
0661     if (!hdr) {
0662         nlmsg_free(msg);
0663         return -EMSGSIZE;
0664     }
0665     return genlmsg_reply(msg, info);
0666 }
0667 
0668 /* Remove and delete all pnetids from pnet table.
0669  */
0670 static int smc_pnet_flush(struct sk_buff *skb, struct genl_info *info)
0671 {
0672     struct net *net = genl_info_net(info);
0673 
0674     smc_pnet_remove_by_pnetid(net, NULL);
0675     return 0;
0676 }
0677 
0678 /* SMC_PNETID generic netlink operation definition */
0679 static const struct genl_ops smc_pnet_ops[] = {
0680     {
0681         .cmd = SMC_PNETID_GET,
0682         .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
0683         /* can be retrieved by unprivileged users */
0684         .doit = smc_pnet_get,
0685         .dumpit = smc_pnet_dump,
0686         .start = smc_pnet_dump_start
0687     },
0688     {
0689         .cmd = SMC_PNETID_ADD,
0690         .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
0691         .flags = GENL_ADMIN_PERM,
0692         .doit = smc_pnet_add
0693     },
0694     {
0695         .cmd = SMC_PNETID_DEL,
0696         .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
0697         .flags = GENL_ADMIN_PERM,
0698         .doit = smc_pnet_del
0699     },
0700     {
0701         .cmd = SMC_PNETID_FLUSH,
0702         .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
0703         .flags = GENL_ADMIN_PERM,
0704         .doit = smc_pnet_flush
0705     }
0706 };
0707 
0708 /* SMC_PNETID family definition */
0709 static struct genl_family smc_pnet_nl_family __ro_after_init = {
0710     .hdrsize = 0,
0711     .name = SMCR_GENL_FAMILY_NAME,
0712     .version = SMCR_GENL_FAMILY_VERSION,
0713     .maxattr = SMC_PNETID_MAX,
0714     .policy = smc_pnet_policy,
0715     .netnsok = true,
0716     .module = THIS_MODULE,
0717     .ops = smc_pnet_ops,
0718     .n_ops =  ARRAY_SIZE(smc_pnet_ops)
0719 };
0720 
0721 bool smc_pnet_is_ndev_pnetid(struct net *net, u8 *pnetid)
0722 {
0723     struct smc_net *sn = net_generic(net, smc_net_id);
0724     struct smc_pnetids_ndev_entry *pe;
0725     bool rc = false;
0726 
0727     read_lock(&sn->pnetids_ndev.lock);
0728     list_for_each_entry(pe, &sn->pnetids_ndev.list, list) {
0729         if (smc_pnet_match(pnetid, pe->pnetid)) {
0730             rc = true;
0731             goto unlock;
0732         }
0733     }
0734 
0735 unlock:
0736     read_unlock(&sn->pnetids_ndev.lock);
0737     return rc;
0738 }
0739 
0740 static int smc_pnet_add_pnetid(struct net *net, u8 *pnetid)
0741 {
0742     struct smc_net *sn = net_generic(net, smc_net_id);
0743     struct smc_pnetids_ndev_entry *pe, *pi;
0744 
0745     pe = kzalloc(sizeof(*pe), GFP_KERNEL);
0746     if (!pe)
0747         return -ENOMEM;
0748 
0749     write_lock(&sn->pnetids_ndev.lock);
0750     list_for_each_entry(pi, &sn->pnetids_ndev.list, list) {
0751         if (smc_pnet_match(pnetid, pe->pnetid)) {
0752             refcount_inc(&pi->refcnt);
0753             kfree(pe);
0754             goto unlock;
0755         }
0756     }
0757     refcount_set(&pe->refcnt, 1);
0758     memcpy(pe->pnetid, pnetid, SMC_MAX_PNETID_LEN);
0759     list_add_tail(&pe->list, &sn->pnetids_ndev.list);
0760 
0761 unlock:
0762     write_unlock(&sn->pnetids_ndev.lock);
0763     return 0;
0764 }
0765 
0766 static void smc_pnet_remove_pnetid(struct net *net, u8 *pnetid)
0767 {
0768     struct smc_net *sn = net_generic(net, smc_net_id);
0769     struct smc_pnetids_ndev_entry *pe, *pe2;
0770 
0771     write_lock(&sn->pnetids_ndev.lock);
0772     list_for_each_entry_safe(pe, pe2, &sn->pnetids_ndev.list, list) {
0773         if (smc_pnet_match(pnetid, pe->pnetid)) {
0774             if (refcount_dec_and_test(&pe->refcnt)) {
0775                 list_del(&pe->list);
0776                 kfree(pe);
0777             }
0778             break;
0779         }
0780     }
0781     write_unlock(&sn->pnetids_ndev.lock);
0782 }
0783 
0784 static void smc_pnet_add_base_pnetid(struct net *net, struct net_device *dev,
0785                      u8 *ndev_pnetid)
0786 {
0787     struct net_device *base_dev;
0788 
0789     base_dev = __pnet_find_base_ndev(dev);
0790     if (base_dev->flags & IFF_UP &&
0791         !smc_pnetid_by_dev_port(base_dev->dev.parent, base_dev->dev_port,
0792                     ndev_pnetid)) {
0793         /* add to PNETIDs list */
0794         smc_pnet_add_pnetid(net, ndev_pnetid);
0795     }
0796 }
0797 
0798 /* create initial list of netdevice pnetids */
0799 static void smc_pnet_create_pnetids_list(struct net *net)
0800 {
0801     u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
0802     struct net_device *dev;
0803 
0804     rtnl_lock();
0805     for_each_netdev(net, dev)
0806         smc_pnet_add_base_pnetid(net, dev, ndev_pnetid);
0807     rtnl_unlock();
0808 }
0809 
0810 /* clean up list of netdevice pnetids */
0811 static void smc_pnet_destroy_pnetids_list(struct net *net)
0812 {
0813     struct smc_net *sn = net_generic(net, smc_net_id);
0814     struct smc_pnetids_ndev_entry *pe, *temp_pe;
0815 
0816     write_lock(&sn->pnetids_ndev.lock);
0817     list_for_each_entry_safe(pe, temp_pe, &sn->pnetids_ndev.list, list) {
0818         list_del(&pe->list);
0819         kfree(pe);
0820     }
0821     write_unlock(&sn->pnetids_ndev.lock);
0822 }
0823 
0824 static int smc_pnet_netdev_event(struct notifier_block *this,
0825                  unsigned long event, void *ptr)
0826 {
0827     struct net_device *event_dev = netdev_notifier_info_to_dev(ptr);
0828     struct net *net = dev_net(event_dev);
0829     u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
0830 
0831     switch (event) {
0832     case NETDEV_REBOOT:
0833     case NETDEV_UNREGISTER:
0834         smc_pnet_remove_by_ndev(event_dev);
0835         smc_ib_ndev_change(event_dev, event);
0836         return NOTIFY_OK;
0837     case NETDEV_REGISTER:
0838         smc_pnet_add_by_ndev(event_dev);
0839         smc_ib_ndev_change(event_dev, event);
0840         return NOTIFY_OK;
0841     case NETDEV_UP:
0842         smc_pnet_add_base_pnetid(net, event_dev, ndev_pnetid);
0843         return NOTIFY_OK;
0844     case NETDEV_DOWN:
0845         event_dev = __pnet_find_base_ndev(event_dev);
0846         if (!smc_pnetid_by_dev_port(event_dev->dev.parent,
0847                         event_dev->dev_port, ndev_pnetid)) {
0848             /* remove from PNETIDs list */
0849             smc_pnet_remove_pnetid(net, ndev_pnetid);
0850         }
0851         return NOTIFY_OK;
0852     default:
0853         return NOTIFY_DONE;
0854     }
0855 }
0856 
0857 static struct notifier_block smc_netdev_notifier = {
0858     .notifier_call = smc_pnet_netdev_event
0859 };
0860 
0861 /* init network namespace */
0862 int smc_pnet_net_init(struct net *net)
0863 {
0864     struct smc_net *sn = net_generic(net, smc_net_id);
0865     struct smc_pnettable *pnettable = &sn->pnettable;
0866     struct smc_pnetids_ndev *pnetids_ndev = &sn->pnetids_ndev;
0867 
0868     INIT_LIST_HEAD(&pnettable->pnetlist);
0869     mutex_init(&pnettable->lock);
0870     INIT_LIST_HEAD(&pnetids_ndev->list);
0871     rwlock_init(&pnetids_ndev->lock);
0872 
0873     smc_pnet_create_pnetids_list(net);
0874 
0875     /* disable handshake limitation by default */
0876     net->smc.limit_smc_hs = 0;
0877 
0878     return 0;
0879 }
0880 
0881 int __init smc_pnet_init(void)
0882 {
0883     int rc;
0884 
0885     rc = genl_register_family(&smc_pnet_nl_family);
0886     if (rc)
0887         return rc;
0888     rc = register_netdevice_notifier(&smc_netdev_notifier);
0889     if (rc)
0890         genl_unregister_family(&smc_pnet_nl_family);
0891 
0892     return rc;
0893 }
0894 
0895 /* exit network namespace */
0896 void smc_pnet_net_exit(struct net *net)
0897 {
0898     /* flush pnet table */
0899     smc_pnet_remove_by_pnetid(net, NULL);
0900     smc_pnet_destroy_pnetids_list(net);
0901 }
0902 
0903 void smc_pnet_exit(void)
0904 {
0905     unregister_netdevice_notifier(&smc_netdev_notifier);
0906     genl_unregister_family(&smc_pnet_nl_family);
0907 }
0908 
0909 static struct net_device *__pnet_find_base_ndev(struct net_device *ndev)
0910 {
0911     int i, nest_lvl;
0912 
0913     ASSERT_RTNL();
0914     nest_lvl = ndev->lower_level;
0915     for (i = 0; i < nest_lvl; i++) {
0916         struct list_head *lower = &ndev->adj_list.lower;
0917 
0918         if (list_empty(lower))
0919             break;
0920         lower = lower->next;
0921         ndev = netdev_lower_get_next(ndev, &lower);
0922     }
0923     return ndev;
0924 }
0925 
0926 /* Determine one base device for stacked net devices.
0927  * If the lower device level contains more than one devices
0928  * (for instance with bonding slaves), just the first device
0929  * is used to reach a base device.
0930  */
0931 static struct net_device *pnet_find_base_ndev(struct net_device *ndev)
0932 {
0933     rtnl_lock();
0934     ndev = __pnet_find_base_ndev(ndev);
0935     rtnl_unlock();
0936     return ndev;
0937 }
0938 
0939 static int smc_pnet_find_ndev_pnetid_by_table(struct net_device *ndev,
0940                           u8 *pnetid)
0941 {
0942     struct smc_pnettable *pnettable;
0943     struct net *net = dev_net(ndev);
0944     struct smc_pnetentry *pnetelem;
0945     struct smc_net *sn;
0946     int rc = -ENOENT;
0947 
0948     /* get pnettable for namespace */
0949     sn = net_generic(net, smc_net_id);
0950     pnettable = &sn->pnettable;
0951 
0952     mutex_lock(&pnettable->lock);
0953     list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
0954         if (pnetelem->type == SMC_PNET_ETH && ndev == pnetelem->ndev) {
0955             /* get pnetid of netdev device */
0956             memcpy(pnetid, pnetelem->pnet_name, SMC_MAX_PNETID_LEN);
0957             rc = 0;
0958             break;
0959         }
0960     }
0961     mutex_unlock(&pnettable->lock);
0962     return rc;
0963 }
0964 
0965 static int smc_pnet_determine_gid(struct smc_ib_device *ibdev, int i,
0966                   struct smc_init_info *ini)
0967 {
0968     if (!ini->check_smcrv2 &&
0969         !smc_ib_determine_gid(ibdev, i, ini->vlan_id, ini->ib_gid, NULL,
0970                   NULL)) {
0971         ini->ib_dev = ibdev;
0972         ini->ib_port = i;
0973         return 0;
0974     }
0975     if (ini->check_smcrv2 &&
0976         !smc_ib_determine_gid(ibdev, i, ini->vlan_id, ini->smcrv2.ib_gid_v2,
0977                   NULL, &ini->smcrv2)) {
0978         ini->smcrv2.ib_dev_v2 = ibdev;
0979         ini->smcrv2.ib_port_v2 = i;
0980         return 0;
0981     }
0982     return -ENODEV;
0983 }
0984 
0985 /* find a roce device for the given pnetid */
0986 static void _smc_pnet_find_roce_by_pnetid(u8 *pnet_id,
0987                       struct smc_init_info *ini,
0988                       struct smc_ib_device *known_dev,
0989                       struct net *net)
0990 {
0991     struct smc_ib_device *ibdev;
0992     int i;
0993 
0994     mutex_lock(&smc_ib_devices.mutex);
0995     list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
0996         if (ibdev == known_dev ||
0997             !rdma_dev_access_netns(ibdev->ibdev, net))
0998             continue;
0999         for (i = 1; i <= SMC_MAX_PORTS; i++) {
1000             if (!rdma_is_port_valid(ibdev->ibdev, i))
1001                 continue;
1002             if (smc_pnet_match(ibdev->pnetid[i - 1], pnet_id) &&
1003                 smc_ib_port_active(ibdev, i) &&
1004                 !test_bit(i - 1, ibdev->ports_going_away)) {
1005                 if (!smc_pnet_determine_gid(ibdev, i, ini))
1006                     goto out;
1007             }
1008         }
1009     }
1010 out:
1011     mutex_unlock(&smc_ib_devices.mutex);
1012 }
1013 
1014 /* find alternate roce device with same pnet_id, vlan_id and net namespace */
1015 void smc_pnet_find_alt_roce(struct smc_link_group *lgr,
1016                 struct smc_init_info *ini,
1017                 struct smc_ib_device *known_dev)
1018 {
1019     struct net *net = lgr->net;
1020 
1021     _smc_pnet_find_roce_by_pnetid(lgr->pnet_id, ini, known_dev, net);
1022 }
1023 
1024 /* if handshake network device belongs to a roce device, return its
1025  * IB device and port
1026  */
1027 static void smc_pnet_find_rdma_dev(struct net_device *netdev,
1028                    struct smc_init_info *ini)
1029 {
1030     struct net *net = dev_net(netdev);
1031     struct smc_ib_device *ibdev;
1032 
1033     mutex_lock(&smc_ib_devices.mutex);
1034     list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
1035         struct net_device *ndev;
1036         int i;
1037 
1038         /* check rdma net namespace */
1039         if (!rdma_dev_access_netns(ibdev->ibdev, net))
1040             continue;
1041 
1042         for (i = 1; i <= SMC_MAX_PORTS; i++) {
1043             if (!rdma_is_port_valid(ibdev->ibdev, i))
1044                 continue;
1045             if (!ibdev->ibdev->ops.get_netdev)
1046                 continue;
1047             ndev = ibdev->ibdev->ops.get_netdev(ibdev->ibdev, i);
1048             if (!ndev)
1049                 continue;
1050             dev_put(ndev);
1051             if (netdev == ndev &&
1052                 smc_ib_port_active(ibdev, i) &&
1053                 !test_bit(i - 1, ibdev->ports_going_away)) {
1054                 if (!smc_pnet_determine_gid(ibdev, i, ini))
1055                     break;
1056             }
1057         }
1058     }
1059     mutex_unlock(&smc_ib_devices.mutex);
1060 }
1061 
1062 /* Determine the corresponding IB device port based on the hardware PNETID.
1063  * Searching stops at the first matching active IB device port with vlan_id
1064  * configured.
1065  * If nothing found, check pnetid table.
1066  * If nothing found, try to use handshake device
1067  */
1068 static void smc_pnet_find_roce_by_pnetid(struct net_device *ndev,
1069                      struct smc_init_info *ini)
1070 {
1071     u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
1072     struct net *net;
1073 
1074     ndev = pnet_find_base_ndev(ndev);
1075     net = dev_net(ndev);
1076     if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
1077                    ndev_pnetid) &&
1078         smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid)) {
1079         smc_pnet_find_rdma_dev(ndev, ini);
1080         return; /* pnetid could not be determined */
1081     }
1082     _smc_pnet_find_roce_by_pnetid(ndev_pnetid, ini, NULL, net);
1083 }
1084 
1085 static void smc_pnet_find_ism_by_pnetid(struct net_device *ndev,
1086                     struct smc_init_info *ini)
1087 {
1088     u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
1089     struct smcd_dev *ismdev;
1090 
1091     ndev = pnet_find_base_ndev(ndev);
1092     if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
1093                    ndev_pnetid) &&
1094         smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid))
1095         return; /* pnetid could not be determined */
1096 
1097     mutex_lock(&smcd_dev_list.mutex);
1098     list_for_each_entry(ismdev, &smcd_dev_list.list, list) {
1099         if (smc_pnet_match(ismdev->pnetid, ndev_pnetid) &&
1100             !ismdev->going_away &&
1101             (!ini->ism_peer_gid[0] ||
1102              !smc_ism_cantalk(ini->ism_peer_gid[0], ini->vlan_id,
1103                       ismdev))) {
1104             ini->ism_dev[0] = ismdev;
1105             break;
1106         }
1107     }
1108     mutex_unlock(&smcd_dev_list.mutex);
1109 }
1110 
1111 /* PNET table analysis for a given sock:
1112  * determine ib_device and port belonging to used internal TCP socket
1113  * ethernet interface.
1114  */
1115 void smc_pnet_find_roce_resource(struct sock *sk, struct smc_init_info *ini)
1116 {
1117     struct dst_entry *dst = sk_dst_get(sk);
1118 
1119     if (!dst)
1120         goto out;
1121     if (!dst->dev)
1122         goto out_rel;
1123 
1124     smc_pnet_find_roce_by_pnetid(dst->dev, ini);
1125 
1126 out_rel:
1127     dst_release(dst);
1128 out:
1129     return;
1130 }
1131 
1132 void smc_pnet_find_ism_resource(struct sock *sk, struct smc_init_info *ini)
1133 {
1134     struct dst_entry *dst = sk_dst_get(sk);
1135 
1136     ini->ism_dev[0] = NULL;
1137     if (!dst)
1138         goto out;
1139     if (!dst->dev)
1140         goto out_rel;
1141 
1142     smc_pnet_find_ism_by_pnetid(dst->dev, ini);
1143 
1144 out_rel:
1145     dst_release(dst);
1146 out:
1147     return;
1148 }
1149 
1150 /* Lookup and apply a pnet table entry to the given ib device.
1151  */
1152 int smc_pnetid_by_table_ib(struct smc_ib_device *smcibdev, u8 ib_port)
1153 {
1154     char *ib_name = smcibdev->ibdev->name;
1155     struct smc_pnettable *pnettable;
1156     struct smc_pnetentry *tmp_pe;
1157     struct smc_net *sn;
1158     int rc = -ENOENT;
1159 
1160     /* get pnettable for init namespace */
1161     sn = net_generic(&init_net, smc_net_id);
1162     pnettable = &sn->pnettable;
1163 
1164     mutex_lock(&pnettable->lock);
1165     list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
1166         if (tmp_pe->type == SMC_PNET_IB &&
1167             !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX) &&
1168             tmp_pe->ib_port == ib_port) {
1169             smc_pnet_apply_ib(smcibdev, ib_port, tmp_pe->pnet_name);
1170             rc = 0;
1171             break;
1172         }
1173     }
1174     mutex_unlock(&pnettable->lock);
1175 
1176     return rc;
1177 }
1178 
1179 /* Lookup and apply a pnet table entry to the given smcd device.
1180  */
1181 int smc_pnetid_by_table_smcd(struct smcd_dev *smcddev)
1182 {
1183     const char *ib_name = dev_name(&smcddev->dev);
1184     struct smc_pnettable *pnettable;
1185     struct smc_pnetentry *tmp_pe;
1186     struct smc_net *sn;
1187     int rc = -ENOENT;
1188 
1189     /* get pnettable for init namespace */
1190     sn = net_generic(&init_net, smc_net_id);
1191     pnettable = &sn->pnettable;
1192 
1193     mutex_lock(&pnettable->lock);
1194     list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
1195         if (tmp_pe->type == SMC_PNET_IB &&
1196             !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
1197             smc_pnet_apply_smcd(smcddev, tmp_pe->pnet_name);
1198             rc = 0;
1199             break;
1200         }
1201     }
1202     mutex_unlock(&pnettable->lock);
1203 
1204     return rc;
1205 }
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!