net/smc/smc_core.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  *  Basic Transport Functions exploiting Infiniband API
   6  *
   7  *  Copyright IBM Corp. 2016
   8  *
   9  *  Author(s):  Ursula Braun <[email protected]>
  10  */
  11
  12 #include <linux/socket.h>
  13 #include <linux/if_vlan.h>
  14 #include <linux/random.h>
  15 #include <linux/workqueue.h>
  16 #include <linux/wait.h>
  17 #include <linux/reboot.h>
  18 #include <linux/mutex.h>
  19 #include <linux/list.h>
  20 #include <linux/smc.h>
  21 #include <net/tcp.h>
  22 #include <net/sock.h>
  23 #include <rdma/ib_verbs.h>
  24 #include <rdma/ib_cache.h>
  25
  26 #include "smc.h"
  27 #include "smc_clc.h"
  28 #include "smc_core.h"
  29 #include "smc_ib.h"
  30 #include "smc_wr.h"
  31 #include "smc_llc.h"
  32 #include "smc_cdc.h"
  33 #include "smc_close.h"
  34 #include "smc_ism.h"
  35 #include "smc_netlink.h"
  36 #include "smc_stats.h"
  37 #include "smc_tracepoint.h"
  38
  39 #define SMC_LGR_NUM_INCR                256
  40 #define SMC_LGR_FREE_DELAY_SERV         (600 * HZ)
  41 #define SMC_LGR_FREE_DELAY_CLNT         (SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
  42
  43 struct smc_lgr_list smc_lgr_list = {    /* established link groups */
  44         .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
  45         .list = LIST_HEAD_INIT(smc_lgr_list.list),
  46         .num = 0,
  47 };
  48
  49 static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */
  50 static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted);
  51
  52 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
  53                          struct smc_buf_desc *buf_desc);
  54 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft);
  55
  56 static void smc_link_down_work(struct work_struct *work);
  57
  58 /* return head of link group list and its lock for a given link group */
  59 static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr,
  60                                                   spinlock_t **lgr_lock)
  61 {
  62         if (lgr->is_smcd) {
  63                 *lgr_lock = &lgr->smcd->lgr_lock;
  64                 return &lgr->smcd->lgr_list;
  65         }
  66
  67         *lgr_lock = &smc_lgr_list.lock;
  68         return &smc_lgr_list.list;
  69 }
  70
  71 static void smc_ibdev_cnt_inc(struct smc_link *lnk)
  72 {
  73         atomic_inc(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
  74 }
  75
  76 static void smc_ibdev_cnt_dec(struct smc_link *lnk)
  77 {
  78         atomic_dec(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
  79 }
  80
  81 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
  82 {
  83         /* client link group creation always follows the server link group
  84          * creation. For client use a somewhat higher removal delay time,
  85          * otherwise there is a risk of out-of-sync link groups.
  86          */
  87         if (!lgr->freeing) {
  88                 mod_delayed_work(system_wq, &lgr->free_work,
  89                                  (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
  90                                                 SMC_LGR_FREE_DELAY_CLNT :
  91                                                 SMC_LGR_FREE_DELAY_SERV);
  92         }
  93 }
  94
  95 /* Register connection's alert token in our lookup structure.
  96  * To use rbtrees we have to implement our own insert core.
  97  * Requires @conns_lock
  98  * @smc         connection to register
  99  * Returns 0 on success, != otherwise.
 100  */
 101 static void smc_lgr_add_alert_token(struct smc_connection *conn)
 102 {
 103         struct rb_node **link, *parent = NULL;
 104         u32 token = conn->alert_token_local;
 105
 106         link = &conn->lgr->conns_all.rb_node;
 107         while (*link) {
 108                 struct smc_connection *cur = rb_entry(*link,
 109                                         struct smc_connection, alert_node);
 110
 111                 parent = *link;
 112                 if (cur->alert_token_local > token)
 113                         link = &parent->rb_left;
 114                 else
 115                         link = &parent->rb_right;
 116         }
 117         /* Put the new node there */
 118         rb_link_node(&conn->alert_node, parent, link);
 119         rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
 120 }
 121
 122 /* assign an SMC-R link to the connection */
 123 static int smcr_lgr_conn_assign_link(struct smc_connection *conn, bool first)
 124 {
 125         enum smc_link_state expected = first ? SMC_LNK_ACTIVATING :
 126                                        SMC_LNK_ACTIVE;
 127         int i, j;
 128
 129         /* do link balancing */
 130         conn->lnk = NULL;       /* reset conn->lnk first */
 131         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 132                 struct smc_link *lnk = &conn->lgr->lnk[i];
 133
 134                 if (lnk->state != expected || lnk->link_is_asym)
 135                         continue;
 136                 if (conn->lgr->role == SMC_CLNT) {
 137                         conn->lnk = lnk; /* temporary, SMC server assigns link*/
 138                         break;
 139                 }
 140                 if (conn->lgr->conns_num % 2) {
 141                         for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) {
 142                                 struct smc_link *lnk2;
 143
 144                                 lnk2 = &conn->lgr->lnk[j];
 145                                 if (lnk2->state == expected &&
 146                                     !lnk2->link_is_asym) {
 147                                         conn->lnk = lnk2;
 148                                         break;
 149                                 }
 150                         }
 151                 }
 152                 if (!conn->lnk)
 153                         conn->lnk = lnk;
 154                 break;
 155         }
 156         if (!conn->lnk)
 157                 return SMC_CLC_DECL_NOACTLINK;
 158         atomic_inc(&conn->lnk->conn_cnt);
 159         return 0;
 160 }
 161
 162 /* Register connection in link group by assigning an alert token
 163  * registered in a search tree.
 164  * Requires @conns_lock
 165  * Note that '0' is a reserved value and not assigned.
 166  */
 167 static int smc_lgr_register_conn(struct smc_connection *conn, bool first)
 168 {
 169         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 170         static atomic_t nexttoken = ATOMIC_INIT(0);
 171         int rc;
 172
 173         if (!conn->lgr->is_smcd) {
 174                 rc = smcr_lgr_conn_assign_link(conn, first);
 175                 if (rc) {
 176                         conn->lgr = NULL;
 177                         return rc;
 178                 }
 179         }
 180         /* find a new alert_token_local value not yet used by some connection
 181          * in this link group
 182          */
 183         sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
 184         while (!conn->alert_token_local) {
 185                 conn->alert_token_local = atomic_inc_return(&nexttoken);
 186                 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
 187                         conn->alert_token_local = 0;
 188         }
 189         smc_lgr_add_alert_token(conn);
 190         conn->lgr->conns_num++;
 191         return 0;
 192 }
 193
 194 /* Unregister connection and reset the alert token of the given connection<
 195  */
 196 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
 197 {
 198         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 199         struct smc_link_group *lgr = conn->lgr;
 200
 201         rb_erase(&conn->alert_node, &lgr->conns_all);
 202         if (conn->lnk)
 203                 atomic_dec(&conn->lnk->conn_cnt);
 204         lgr->conns_num--;
 205         conn->alert_token_local = 0;
 206         sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
 207 }
 208
 209 /* Unregister connection from lgr
 210  */
 211 static void smc_lgr_unregister_conn(struct smc_connection *conn)
 212 {
 213         struct smc_link_group *lgr = conn->lgr;
 214
 215         if (!smc_conn_lgr_valid(conn))
 216                 return;
 217         write_lock_bh(&lgr->conns_lock);
 218         if (conn->alert_token_local) {
 219                 __smc_lgr_unregister_conn(conn);
 220         }
 221         write_unlock_bh(&lgr->conns_lock);
 222 }
 223
 224 static void smc_lgr_buf_list_add(struct smc_link_group *lgr,
 225                                  bool is_rmb,
 226                                  struct list_head *buf_list,
 227                                  struct smc_buf_desc *buf_desc)
 228 {
 229         list_add(&buf_desc->list, buf_list);
 230         if (is_rmb) {
 231                 lgr->alloc_rmbs += buf_desc->len;
 232                 lgr->alloc_rmbs +=
 233                         lgr->is_smcd ? sizeof(struct smcd_cdc_msg) : 0;
 234         } else {
 235                 lgr->alloc_sndbufs += buf_desc->len;
 236         }
 237 }
 238
 239 static void smc_lgr_buf_list_del(struct smc_link_group *lgr,
 240                                  bool is_rmb,
 241                                  struct smc_buf_desc *buf_desc)
 242 {
 243         list_del(&buf_desc->list);
 244         if (is_rmb) {
 245                 lgr->alloc_rmbs -= buf_desc->len;
 246                 lgr->alloc_rmbs -=
 247                         lgr->is_smcd ? sizeof(struct smcd_cdc_msg) : 0;
 248         } else {
 249                 lgr->alloc_sndbufs -= buf_desc->len;
 250         }
 251 }
 252
 253 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb)
 254 {
 255         struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
 256         char hostname[SMC_MAX_HOSTNAME_LEN + 1];
 257         char smc_seid[SMC_MAX_EID_LEN + 1];
 258         struct nlattr *attrs;
 259         u8 *seid = NULL;
 260         u8 *host = NULL;
 261         void *nlh;
 262
 263         nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 264                           &smc_gen_nl_family, NLM_F_MULTI,
 265                           SMC_NETLINK_GET_SYS_INFO);
 266         if (!nlh)
 267                 goto errmsg;
 268         if (cb_ctx->pos[0])
 269                 goto errout;
 270         attrs = nla_nest_start(skb, SMC_GEN_SYS_INFO);
 271         if (!attrs)
 272                 goto errout;
 273         if (nla_put_u8(skb, SMC_NLA_SYS_VER, SMC_V2))
 274                 goto errattr;
 275         if (nla_put_u8(skb, SMC_NLA_SYS_REL, SMC_RELEASE))
 276                 goto errattr;
 277         if (nla_put_u8(skb, SMC_NLA_SYS_IS_ISM_V2, smc_ism_is_v2_capable()))
 278                 goto errattr;
 279         if (nla_put_u8(skb, SMC_NLA_SYS_IS_SMCR_V2, true))
 280                 goto errattr;
 281         smc_clc_get_hostname(&host);
 282         if (host) {
 283                 memcpy(hostname, host, SMC_MAX_HOSTNAME_LEN);
 284                 hostname[SMC_MAX_HOSTNAME_LEN] = 0;
 285                 if (nla_put_string(skb, SMC_NLA_SYS_LOCAL_HOST, hostname))
 286                         goto errattr;
 287         }
 288         if (smc_ism_is_v2_capable()) {
 289                 smc_ism_get_system_eid(&seid);
 290                 memcpy(smc_seid, seid, SMC_MAX_EID_LEN);
 291                 smc_seid[SMC_MAX_EID_LEN] = 0;
 292                 if (nla_put_string(skb, SMC_NLA_SYS_SEID, smc_seid))
 293                         goto errattr;
 294         }
 295         nla_nest_end(skb, attrs);
 296         genlmsg_end(skb, nlh);
 297         cb_ctx->pos[0] = 1;
 298         return skb->len;
 299
 300 errattr:
 301         nla_nest_cancel(skb, attrs);
 302 errout:
 303         genlmsg_cancel(skb, nlh);
 304 errmsg:
 305         return skb->len;
 306 }
 307
 308 /* Fill SMC_NLA_LGR_D_V2_COMMON/SMC_NLA_LGR_R_V2_COMMON nested attributes */
 309 static int smc_nl_fill_lgr_v2_common(struct smc_link_group *lgr,
 310                                      struct sk_buff *skb,
 311                                      struct netlink_callback *cb,
 312                                      struct nlattr *v2_attrs)
 313 {
 314         char smc_host[SMC_MAX_HOSTNAME_LEN + 1];
 315         char smc_eid[SMC_MAX_EID_LEN + 1];
 316
 317         if (nla_put_u8(skb, SMC_NLA_LGR_V2_VER, lgr->smc_version))
 318                 goto errv2attr;
 319         if (nla_put_u8(skb, SMC_NLA_LGR_V2_REL, lgr->peer_smc_release))
 320                 goto errv2attr;
 321         if (nla_put_u8(skb, SMC_NLA_LGR_V2_OS, lgr->peer_os))
 322                 goto errv2attr;
 323         memcpy(smc_host, lgr->peer_hostname, SMC_MAX_HOSTNAME_LEN);
 324         smc_host[SMC_MAX_HOSTNAME_LEN] = 0;
 325         if (nla_put_string(skb, SMC_NLA_LGR_V2_PEER_HOST, smc_host))
 326                 goto errv2attr;
 327         memcpy(smc_eid, lgr->negotiated_eid, SMC_MAX_EID_LEN);
 328         smc_eid[SMC_MAX_EID_LEN] = 0;
 329         if (nla_put_string(skb, SMC_NLA_LGR_V2_NEG_EID, smc_eid))
 330                 goto errv2attr;
 331
 332         nla_nest_end(skb, v2_attrs);
 333         return 0;
 334
 335 errv2attr:
 336         nla_nest_cancel(skb, v2_attrs);
 337         return -EMSGSIZE;
 338 }
 339
 340 static int smc_nl_fill_smcr_lgr_v2(struct smc_link_group *lgr,
 341                                    struct sk_buff *skb,
 342                                    struct netlink_callback *cb)
 343 {
 344         struct nlattr *v2_attrs;
 345
 346         v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2);
 347         if (!v2_attrs)
 348                 goto errattr;
 349         if (nla_put_u8(skb, SMC_NLA_LGR_R_V2_DIRECT, !lgr->uses_gateway))
 350                 goto errv2attr;
 351         if (nla_put_u8(skb, SMC_NLA_LGR_R_V2_MAX_CONNS, lgr->max_conns))
 352                 goto errv2attr;
 353         if (nla_put_u8(skb, SMC_NLA_LGR_R_V2_MAX_LINKS, lgr->max_links))
 354                 goto errv2attr;
 355
 356         nla_nest_end(skb, v2_attrs);
 357         return 0;
 358
 359 errv2attr:
 360         nla_nest_cancel(skb, v2_attrs);
 361 errattr:
 362         return -EMSGSIZE;
 363 }
 364
 365 static int smc_nl_fill_lgr(struct smc_link_group *lgr,
 366                            struct sk_buff *skb,
 367                            struct netlink_callback *cb)
 368 {
 369         char smc_target[SMC_MAX_PNETID_LEN + 1];
 370         struct nlattr *attrs, *v2_attrs;
 371
 372         attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCR);
 373         if (!attrs)
 374                 goto errout;
 375
 376         if (nla_put_u32(skb, SMC_NLA_LGR_R_ID, *((u32 *)&lgr->id)))
 377                 goto errattr;
 378         if (nla_put_u32(skb, SMC_NLA_LGR_R_CONNS_NUM, lgr->conns_num))
 379                 goto errattr;
 380         if (nla_put_u8(skb, SMC_NLA_LGR_R_ROLE, lgr->role))
 381                 goto errattr;
 382         if (nla_put_u8(skb, SMC_NLA_LGR_R_TYPE, lgr->type))
 383                 goto errattr;
 384         if (nla_put_u8(skb, SMC_NLA_LGR_R_BUF_TYPE, lgr->buf_type))
 385                 goto errattr;
 386         if (nla_put_u8(skb, SMC_NLA_LGR_R_VLAN_ID, lgr->vlan_id))
 387                 goto errattr;
 388         if (nla_put_u64_64bit(skb, SMC_NLA_LGR_R_NET_COOKIE,
 389                               lgr->net->net_cookie, SMC_NLA_LGR_R_PAD))
 390                 goto errattr;
 391         memcpy(smc_target, lgr->pnet_id, SMC_MAX_PNETID_LEN);
 392         smc_target[SMC_MAX_PNETID_LEN] = 0;
 393         if (nla_put_string(skb, SMC_NLA_LGR_R_PNETID, smc_target))
 394                 goto errattr;
 395         if (nla_put_uint(skb, SMC_NLA_LGR_R_SNDBUF_ALLOC, lgr->alloc_sndbufs))
 396                 goto errattr;
 397         if (nla_put_uint(skb, SMC_NLA_LGR_R_RMB_ALLOC, lgr->alloc_rmbs))
 398                 goto errattr;
 399         if (lgr->smc_version > SMC_V1) {
 400                 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2_COMMON);
 401                 if (!v2_attrs)
 402                         goto errattr;
 403                 if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
 404                         goto errattr;
 405                 if (smc_nl_fill_smcr_lgr_v2(lgr, skb, cb))
 406                         goto errattr;
 407         }
 408
 409         nla_nest_end(skb, attrs);
 410         return 0;
 411 errattr:
 412         nla_nest_cancel(skb, attrs);
 413 errout:
 414         return -EMSGSIZE;
 415 }
 416
 417 static int smc_nl_fill_lgr_link(struct smc_link_group *lgr,
 418                                 struct smc_link *link,
 419                                 struct sk_buff *skb,
 420                                 struct netlink_callback *cb)
 421 {
 422         char smc_ibname[IB_DEVICE_NAME_MAX];
 423         u8 smc_gid_target[41];
 424         struct nlattr *attrs;
 425         u32 link_uid = 0;
 426         void *nlh;
 427
 428         nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 429                           &smc_gen_nl_family, NLM_F_MULTI,
 430                           SMC_NETLINK_GET_LINK_SMCR);
 431         if (!nlh)
 432                 goto errmsg;
 433
 434         attrs = nla_nest_start(skb, SMC_GEN_LINK_SMCR);
 435         if (!attrs)
 436                 goto errout;
 437
 438         if (nla_put_u8(skb, SMC_NLA_LINK_ID, link->link_id))
 439                 goto errattr;
 440         if (nla_put_u32(skb, SMC_NLA_LINK_STATE, link->state))
 441                 goto errattr;
 442         if (nla_put_u32(skb, SMC_NLA_LINK_CONN_CNT,
 443                         atomic_read(&link->conn_cnt)))
 444                 goto errattr;
 445         if (nla_put_u8(skb, SMC_NLA_LINK_IB_PORT, link->ibport))
 446                 goto errattr;
 447         if (nla_put_u32(skb, SMC_NLA_LINK_NET_DEV, link->ndev_ifidx))
 448                 goto errattr;
 449         snprintf(smc_ibname, sizeof(smc_ibname), "%s", link->ibname);
 450         if (nla_put_string(skb, SMC_NLA_LINK_IB_DEV, smc_ibname))
 451                 goto errattr;
 452         memcpy(&link_uid, link->link_uid, sizeof(link_uid));
 453         if (nla_put_u32(skb, SMC_NLA_LINK_UID, link_uid))
 454                 goto errattr;
 455         memcpy(&link_uid, link->peer_link_uid, sizeof(link_uid));
 456         if (nla_put_u32(skb, SMC_NLA_LINK_PEER_UID, link_uid))
 457                 goto errattr;
 458         memset(smc_gid_target, 0, sizeof(smc_gid_target));
 459         smc_gid_be16_convert(smc_gid_target, link->gid);
 460         if (nla_put_string(skb, SMC_NLA_LINK_GID, smc_gid_target))
 461                 goto errattr;
 462         memset(smc_gid_target, 0, sizeof(smc_gid_target));
 463         smc_gid_be16_convert(smc_gid_target, link->peer_gid);
 464         if (nla_put_string(skb, SMC_NLA_LINK_PEER_GID, smc_gid_target))
 465                 goto errattr;
 466
 467         nla_nest_end(skb, attrs);
 468         genlmsg_end(skb, nlh);
 469         return 0;
 470 errattr:
 471         nla_nest_cancel(skb, attrs);
 472 errout:
 473         genlmsg_cancel(skb, nlh);
 474 errmsg:
 475         return -EMSGSIZE;
 476 }
 477
 478 static int smc_nl_handle_lgr(struct smc_link_group *lgr,
 479                              struct sk_buff *skb,
 480                              struct netlink_callback *cb,
 481                              bool list_links)
 482 {
 483         void *nlh;
 484         int i;
 485
 486         nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 487                           &smc_gen_nl_family, NLM_F_MULTI,
 488                           SMC_NETLINK_GET_LGR_SMCR);
 489         if (!nlh)
 490                 goto errmsg;
 491         if (smc_nl_fill_lgr(lgr, skb, cb))
 492                 goto errout;
 493
 494         genlmsg_end(skb, nlh);
 495         if (!list_links)
 496                 goto out;
 497         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 498                 if (!smc_link_usable(&lgr->lnk[i]))
 499                         continue;
 500                 if (smc_nl_fill_lgr_link(lgr, &lgr->lnk[i], skb, cb))
 501                         goto errout;
 502         }
 503 out:
 504         return 0;
 505
 506 errout:
 507         genlmsg_cancel(skb, nlh);
 508 errmsg:
 509         return -EMSGSIZE;
 510 }
 511
 512 static void smc_nl_fill_lgr_list(struct smc_lgr_list *smc_lgr,
 513                                  struct sk_buff *skb,
 514                                  struct netlink_callback *cb,
 515                                  bool list_links)
 516 {
 517         struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
 518         struct smc_link_group *lgr;
 519         int snum = cb_ctx->pos[0];
 520         int num = 0;
 521
 522         spin_lock_bh(&smc_lgr->lock);
 523         list_for_each_entry(lgr, &smc_lgr->list, list) {
 524                 if (num < snum)
 525                         goto next;
 526                 if (smc_nl_handle_lgr(lgr, skb, cb, list_links))
 527                         goto errout;
 528 next:
 529                 num++;
 530         }
 531 errout:
 532         spin_unlock_bh(&smc_lgr->lock);
 533         cb_ctx->pos[0] = num;
 534 }
 535
 536 static int smc_nl_fill_smcd_lgr(struct smc_link_group *lgr,
 537                                 struct sk_buff *skb,
 538                                 struct netlink_callback *cb)
 539 {
 540         char smc_pnet[SMC_MAX_PNETID_LEN + 1];
 541         struct smcd_dev *smcd = lgr->smcd;
 542         struct smcd_gid smcd_gid;
 543         struct nlattr *attrs;
 544         void *nlh;
 545
 546         nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 547                           &smc_gen_nl_family, NLM_F_MULTI,
 548                           SMC_NETLINK_GET_LGR_SMCD);
 549         if (!nlh)
 550                 goto errmsg;
 551
 552         attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD);
 553         if (!attrs)
 554                 goto errout;
 555
 556         if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id)))
 557                 goto errattr;
 558         smcd->ops->get_local_gid(smcd, &smcd_gid);
 559         if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID,
 560                               smcd_gid.gid, SMC_NLA_LGR_D_PAD))
 561                 goto errattr;
 562         if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_EXT_GID,
 563                               smcd_gid.gid_ext, SMC_NLA_LGR_D_PAD))
 564                 goto errattr;
 565         if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid.gid,
 566                               SMC_NLA_LGR_D_PAD))
 567                 goto errattr;
 568         if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_EXT_GID,
 569                               lgr->peer_gid.gid_ext, SMC_NLA_LGR_D_PAD))
 570                 goto errattr;
 571         if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id))
 572                 goto errattr;
 573         if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num))
 574                 goto errattr;
 575         if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd)))
 576                 goto errattr;
 577         if (nla_put_uint(skb, SMC_NLA_LGR_D_SNDBUF_ALLOC, lgr->alloc_sndbufs))
 578                 goto errattr;
 579         if (nla_put_uint(skb, SMC_NLA_LGR_D_DMB_ALLOC, lgr->alloc_rmbs))
 580                 goto errattr;
 581         memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN);
 582         smc_pnet[SMC_MAX_PNETID_LEN] = 0;
 583         if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet))
 584                 goto errattr;
 585         if (lgr->smc_version > SMC_V1) {
 586                 struct nlattr *v2_attrs;
 587
 588                 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_D_V2_COMMON);
 589                 if (!v2_attrs)
 590                         goto errattr;
 591                 if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
 592                         goto errattr;
 593         }
 594         nla_nest_end(skb, attrs);
 595         genlmsg_end(skb, nlh);
 596         return 0;
 597
 598 errattr:
 599         nla_nest_cancel(skb, attrs);
 600 errout:
 601         genlmsg_cancel(skb, nlh);
 602 errmsg:
 603         return -EMSGSIZE;
 604 }
 605
 606 static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev,
 607                                   struct sk_buff *skb,
 608                                   struct netlink_callback *cb)
 609 {
 610         struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
 611         struct smc_link_group *lgr;
 612         int snum = cb_ctx->pos[1];
 613         int rc = 0, num = 0;
 614
 615         spin_lock_bh(&dev->lgr_lock);
 616         list_for_each_entry(lgr, &dev->lgr_list, list) {
 617                 if (!lgr->is_smcd)
 618                         continue;
 619                 if (num < snum)
 620                         goto next;
 621                 rc = smc_nl_fill_smcd_lgr(lgr, skb, cb);
 622                 if (rc)
 623                         goto errout;
 624 next:
 625                 num++;
 626         }
 627 errout:
 628         spin_unlock_bh(&dev->lgr_lock);
 629         cb_ctx->pos[1] = num;
 630         return rc;
 631 }
 632
 633 static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list,
 634                                 struct sk_buff *skb,
 635                                 struct netlink_callback *cb)
 636 {
 637         struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
 638         struct smcd_dev *smcd_dev;
 639         int snum = cb_ctx->pos[0];
 640         int rc = 0, num = 0;
 641
 642         mutex_lock(&dev_list->mutex);
 643         list_for_each_entry(smcd_dev, &dev_list->list, list) {
 644                 if (list_empty(&smcd_dev->lgr_list))
 645                         continue;
 646                 if (num < snum)
 647                         goto next;
 648                 rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb);
 649                 if (rc)
 650                         goto errout;
 651 next:
 652                 num++;
 653         }
 654 errout:
 655         mutex_unlock(&dev_list->mutex);
 656         cb_ctx->pos[0] = num;
 657         return rc;
 658 }
 659
 660 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
 661 {
 662         bool list_links = false;
 663
 664         smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
 665         return skb->len;
 666 }
 667
 668 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb)
 669 {
 670         bool list_links = true;
 671
 672         smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
 673         return skb->len;
 674 }
 675
 676 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
 677 {
 678         smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb);
 679         return skb->len;
 680 }
 681
 682 void smc_lgr_cleanup_early(struct smc_link_group *lgr)
 683 {
 684         spinlock_t *lgr_lock;
 685
 686         if (!lgr)
 687                 return;
 688
 689         smc_lgr_list_head(lgr, &lgr_lock);
 690         spin_lock_bh(lgr_lock);
 691         /* do not use this link group for new connections */
 692         if (!list_empty(&lgr->list))
 693                 list_del_init(&lgr->list);
 694         spin_unlock_bh(lgr_lock);
 695         __smc_lgr_terminate(lgr, true);
 696 }
 697
 698 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
 699 {
 700         int i;
 701
 702         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 703                 struct smc_link *lnk = &lgr->lnk[i];
 704
 705                 if (smc_link_sendable(lnk))
 706                         lnk->state = SMC_LNK_INACTIVE;
 707         }
 708         wake_up_all(&lgr->llc_msg_waiter);
 709         wake_up_all(&lgr->llc_flow_waiter);
 710 }
 711
 712 static void smc_lgr_free(struct smc_link_group *lgr);
 713
 714 static void smc_lgr_free_work(struct work_struct *work)
 715 {
 716         struct smc_link_group *lgr = container_of(to_delayed_work(work),
 717                                                   struct smc_link_group,
 718                                                   free_work);
 719         spinlock_t *lgr_lock;
 720         bool conns;
 721
 722         smc_lgr_list_head(lgr, &lgr_lock);
 723         spin_lock_bh(lgr_lock);
 724         if (lgr->freeing) {
 725                 spin_unlock_bh(lgr_lock);
 726                 return;
 727         }
 728         read_lock_bh(&lgr->conns_lock);
 729         conns = RB_EMPTY_ROOT(&lgr->conns_all);
 730         read_unlock_bh(&lgr->conns_lock);
 731         if (!conns) { /* number of lgr connections is no longer zero */
 732                 spin_unlock_bh(lgr_lock);
 733                 return;
 734         }
 735         list_del_init(&lgr->list); /* remove from smc_lgr_list */
 736         lgr->freeing = 1; /* this instance does the freeing, no new schedule */
 737         spin_unlock_bh(lgr_lock);
 738         cancel_delayed_work(&lgr->free_work);
 739
 740         if (!lgr->is_smcd && !lgr->terminating)
 741                 smc_llc_send_link_delete_all(lgr, true,
 742                                              SMC_LLC_DEL_PROG_INIT_TERM);
 743         if (lgr->is_smcd && !lgr->terminating)
 744                 smc_ism_signal_shutdown(lgr);
 745         if (!lgr->is_smcd)
 746                 smcr_lgr_link_deactivate_all(lgr);
 747         smc_lgr_free(lgr);
 748 }
 749
 750 static void smc_lgr_terminate_work(struct work_struct *work)
 751 {
 752         struct smc_link_group *lgr = container_of(work, struct smc_link_group,
 753                                                   terminate_work);
 754
 755         __smc_lgr_terminate(lgr, true);
 756 }
 757
 758 /* return next unique link id for the lgr */
 759 static u8 smcr_next_link_id(struct smc_link_group *lgr)
 760 {
 761         u8 link_id;
 762         int i;
 763
 764         while (1) {
 765 again:
 766                 link_id = ++lgr->next_link_id;
 767                 if (!link_id)   /* skip zero as link_id */
 768                         link_id = ++lgr->next_link_id;
 769                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 770                         if (smc_link_usable(&lgr->lnk[i]) &&
 771                             lgr->lnk[i].link_id == link_id)
 772                                 goto again;
 773                 }
 774                 break;
 775         }
 776         return link_id;
 777 }
 778
 779 static void smcr_copy_dev_info_to_link(struct smc_link *link)
 780 {
 781         struct smc_ib_device *smcibdev = link->smcibdev;
 782
 783         snprintf(link->ibname, sizeof(link->ibname), "%s",
 784                  smcibdev->ibdev->name);
 785         link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
 786 }
 787
 788 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
 789                    u8 link_idx, struct smc_init_info *ini)
 790 {
 791         struct smc_ib_device *smcibdev;
 792         u8 rndvec[3];
 793         int rc;
 794
 795         if (lgr->smc_version == SMC_V2) {
 796                 lnk->smcibdev = ini->smcrv2.ib_dev_v2;
 797                 lnk->ibport = ini->smcrv2.ib_port_v2;
 798                 lnk->wr_rx_sge_cnt = lnk->smcibdev->ibdev->attrs.max_recv_sge < 2 ? 1 : 2;
 799                 lnk->wr_rx_buflen = smc_link_shared_v2_rxbuf(lnk) ?
 800                         SMC_WR_BUF_SIZE : SMC_WR_BUF_V2_SIZE;
 801         } else {
 802                 lnk->smcibdev = ini->ib_dev;
 803                 lnk->ibport = ini->ib_port;
 804                 lnk->wr_rx_sge_cnt = 1;
 805                 lnk->wr_rx_buflen = SMC_WR_BUF_SIZE;
 806         }
 807         get_device(&lnk->smcibdev->ibdev->dev);
 808         atomic_inc(&lnk->smcibdev->lnk_cnt);
 809         refcount_set(&lnk->refcnt, 1); /* link refcnt is set to 1 */
 810         lnk->clearing = 0;
 811         lnk->path_mtu = lnk->smcibdev->pattr[lnk->ibport - 1].active_mtu;
 812         lnk->link_id = smcr_next_link_id(lgr);
 813         lnk->lgr = lgr;
 814         smc_lgr_hold(lgr); /* lgr_put in smcr_link_clear() */
 815         lnk->link_idx = link_idx;
 816         lnk->wr_rx_id_compl = 0;
 817         smc_ibdev_cnt_inc(lnk);
 818         smcr_copy_dev_info_to_link(lnk);
 819         atomic_set(&lnk->conn_cnt, 0);
 820         smc_llc_link_set_uid(lnk);
 821         INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
 822         if (!lnk->smcibdev->initialized) {
 823                 rc = (int)smc_ib_setup_per_ibdev(lnk->smcibdev);
 824                 if (rc)
 825                         goto out;
 826         }
 827         get_random_bytes(rndvec, sizeof(rndvec));
 828         lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
 829                 (rndvec[2] << 16);
 830         rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
 831                                   ini->vlan_id, lnk->gid, &lnk->sgid_index,
 832                                   lgr->smc_version == SMC_V2 ?
 833                                                   &ini->smcrv2 : NULL);
 834         if (rc)
 835                 goto out;
 836         rc = smc_llc_link_init(lnk);
 837         if (rc)
 838                 goto out;
 839         rc = smc_wr_alloc_link_mem(lnk);
 840         if (rc)
 841                 goto clear_llc_lnk;
 842         rc = smc_ib_create_protection_domain(lnk);
 843         if (rc)
 844                 goto free_link_mem;
 845         rc = smc_ib_create_queue_pair(lnk);
 846         if (rc)
 847                 goto dealloc_pd;
 848         rc = smc_wr_create_link(lnk);
 849         if (rc)
 850                 goto destroy_qp;
 851         lnk->state = SMC_LNK_ACTIVATING;
 852         return 0;
 853
 854 destroy_qp:
 855         smc_ib_destroy_queue_pair(lnk);
 856 dealloc_pd:
 857         smc_ib_dealloc_protection_domain(lnk);
 858 free_link_mem:
 859         smc_wr_free_link_mem(lnk);
 860 clear_llc_lnk:
 861         smc_llc_link_clear(lnk, false);
 862 out:
 863         smc_ibdev_cnt_dec(lnk);
 864         put_device(&lnk->smcibdev->ibdev->dev);
 865         smcibdev = lnk->smcibdev;
 866         memset(lnk, 0, sizeof(struct smc_link));
 867         lnk->state = SMC_LNK_UNUSED;
 868         if (!atomic_dec_return(&smcibdev->lnk_cnt))
 869                 wake_up(&smcibdev->lnks_deleted);
 870         smc_lgr_put(lgr); /* lgr_hold above */
 871         return rc;
 872 }
 873
 874 /* create a new SMC link group */
 875 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
 876 {
 877         struct smc_link_group *lgr;
 878         struct list_head *lgr_list;
 879         struct smcd_dev *smcd;
 880         struct smc_link *lnk;
 881         spinlock_t *lgr_lock;
 882         u8 link_idx;
 883         int rc = 0;
 884         int i;
 885
 886         if (ini->is_smcd && ini->vlan_id) {
 887                 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
 888                                      ini->vlan_id)) {
 889                         rc = SMC_CLC_DECL_ISMVLANERR;
 890                         goto out;
 891                 }
 892         }
 893
 894         lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
 895         if (!lgr) {
 896                 rc = SMC_CLC_DECL_MEM;
 897                 goto ism_put_vlan;
 898         }
 899         lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
 900                                      SMC_LGR_ID_SIZE, &lgr->id);
 901         if (!lgr->tx_wq) {
 902                 rc = -ENOMEM;
 903                 goto free_lgr;
 904         }
 905         lgr->is_smcd = ini->is_smcd;
 906         lgr->sync_err = 0;
 907         lgr->terminating = 0;
 908         lgr->freeing = 0;
 909         lgr->vlan_id = ini->vlan_id;
 910         refcount_set(&lgr->refcnt, 1); /* set lgr refcnt to 1 */
 911         init_rwsem(&lgr->sndbufs_lock);
 912         init_rwsem(&lgr->rmbs_lock);
 913         rwlock_init(&lgr->conns_lock);
 914         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 915                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 916                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 917         }
 918         lgr->next_link_id = 0;
 919         smc_lgr_list.num += SMC_LGR_NUM_INCR;
 920         memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
 921         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 922         INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
 923         lgr->conns_all = RB_ROOT;
 924         if (ini->is_smcd) {
 925                 /* SMC-D specific settings */
 926                 smcd = ini->ism_dev[ini->ism_selected];
 927                 get_device(smcd->ops->get_dev(smcd));
 928                 lgr->peer_gid.gid =
 929                         ini->ism_peer_gid[ini->ism_selected].gid;
 930                 lgr->peer_gid.gid_ext =
 931                         ini->ism_peer_gid[ini->ism_selected].gid_ext;
 932                 lgr->smcd = ini->ism_dev[ini->ism_selected];
 933                 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
 934                 lgr_lock = &lgr->smcd->lgr_lock;
 935                 lgr->smc_version = ini->smcd_version;
 936                 lgr->peer_shutdown = 0;
 937                 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
 938         } else {
 939                 /* SMC-R specific settings */
 940                 struct smc_ib_device *ibdev;
 941                 int ibport;
 942
 943                 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 944                 lgr->smc_version = ini->smcr_version;
 945                 memcpy(lgr->peer_systemid, ini->peer_systemid,
 946                        SMC_SYSTEMID_LEN);
 947                 if (lgr->smc_version == SMC_V2) {
 948                         ibdev = ini->smcrv2.ib_dev_v2;
 949                         ibport = ini->smcrv2.ib_port_v2;
 950                         lgr->saddr = ini->smcrv2.saddr;
 951                         lgr->uses_gateway = ini->smcrv2.uses_gateway;
 952                         memcpy(lgr->nexthop_mac, ini->smcrv2.nexthop_mac,
 953                                ETH_ALEN);
 954                         lgr->max_conns = ini->max_conns;
 955                         lgr->max_links = ini->max_links;
 956                 } else {
 957                         ibdev = ini->ib_dev;
 958                         ibport = ini->ib_port;
 959                         lgr->max_conns = SMC_CONN_PER_LGR_MAX;
 960                         lgr->max_links = SMC_LINKS_ADD_LNK_MAX;
 961                 }
 962                 memcpy(lgr->pnet_id, ibdev->pnetid[ibport - 1],
 963                        SMC_MAX_PNETID_LEN);
 964                 rc = smc_wr_alloc_lgr_mem(lgr);
 965                 if (rc)
 966                         goto free_wq;
 967                 smc_llc_lgr_init(lgr, smc);
 968
 969                 link_idx = SMC_SINGLE_LINK;
 970                 lnk = &lgr->lnk[link_idx];
 971                 rc = smcr_link_init(lgr, lnk, link_idx, ini);
 972                 if (rc) {
 973                         smc_wr_free_lgr_mem(lgr);
 974                         goto free_wq;
 975                 }
 976                 lgr->net = smc_ib_net(lnk->smcibdev);
 977                 lgr_list = &smc_lgr_list.list;
 978                 lgr_lock = &smc_lgr_list.lock;
 979                 lgr->buf_type = lgr->net->smc.sysctl_smcr_buf_type;
 980                 atomic_inc(&lgr_cnt);
 981         }
 982         smc->conn.lgr = lgr;
 983         spin_lock_bh(lgr_lock);
 984         list_add_tail(&lgr->list, lgr_list);
 985         spin_unlock_bh(lgr_lock);
 986         return 0;
 987
 988 free_wq:
 989         destroy_workqueue(lgr->tx_wq);
 990 free_lgr:
 991         kfree(lgr);
 992 ism_put_vlan:
 993         if (ini->is_smcd && ini->vlan_id)
 994                 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
 995 out:
 996         if (rc < 0) {
 997                 if (rc == -ENOMEM)
 998                         rc = SMC_CLC_DECL_MEM;
 999                 else
1000                         rc = SMC_CLC_DECL_INTERR;
1001         }
1002         return rc;
1003 }
1004
1005 static int smc_write_space(struct smc_connection *conn)
1006 {
1007         int buffer_len = conn->peer_rmbe_size;
1008         union smc_host_cursor prod;
1009         union smc_host_cursor cons;
1010         int space;
1011
1012         smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
1013         smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
1014         /* determine rx_buf space */
1015         space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
1016         return space;
1017 }
1018
1019 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
1020                              struct smc_wr_buf *wr_buf)
1021 {
1022         struct smc_connection *conn = &smc->conn;
1023         union smc_host_cursor cons, fin;
1024         int rc = 0;
1025         int diff;
1026
1027         smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
1028         smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
1029         /* set prod cursor to old state, enforce tx_rdma_writes() */
1030         smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
1031         smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
1032
1033         if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
1034                 /* cons cursor advanced more than fin, and prod was set
1035                  * fin above, so now prod is smaller than cons. Fix that.
1036                  */
1037                 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
1038                 smc_curs_add(conn->sndbuf_desc->len,
1039                              &conn->tx_curs_sent, diff);
1040                 smc_curs_add(conn->sndbuf_desc->len,
1041                              &conn->tx_curs_fin, diff);
1042
1043                 smp_mb__before_atomic();
1044                 atomic_add(diff, &conn->sndbuf_space);
1045                 smp_mb__after_atomic();
1046
1047                 smc_curs_add(conn->peer_rmbe_size,
1048                              &conn->local_tx_ctrl.prod, diff);
1049                 smc_curs_add(conn->peer_rmbe_size,
1050                              &conn->local_tx_ctrl_fin, diff);
1051         }
1052         /* recalculate, value is used by tx_rdma_writes() */
1053         atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
1054
1055         if (smc->sk.sk_state != SMC_INIT &&
1056             smc->sk.sk_state != SMC_CLOSED) {
1057                 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
1058                 if (!rc) {
1059                         queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
1060                         smc->sk.sk_data_ready(&smc->sk);
1061                 }
1062         } else {
1063                 smc_wr_tx_put_slot(conn->lnk,
1064                                    (struct smc_wr_tx_pend_priv *)pend);
1065         }
1066         return rc;
1067 }
1068
1069 void smc_switch_link_and_count(struct smc_connection *conn,
1070                                struct smc_link *to_lnk)
1071 {
1072         atomic_dec(&conn->lnk->conn_cnt);
1073         /* link_hold in smc_conn_create() */
1074         smcr_link_put(conn->lnk);
1075         conn->lnk = to_lnk;
1076         atomic_inc(&conn->lnk->conn_cnt);
1077         /* link_put in smc_conn_free() */
1078         smcr_link_hold(conn->lnk);
1079 }
1080
1081 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
1082                                   struct smc_link *from_lnk, bool is_dev_err)
1083 {
1084         struct smc_link *to_lnk = NULL;
1085         struct smc_cdc_tx_pend *pend;
1086         struct smc_connection *conn;
1087         struct smc_wr_buf *wr_buf;
1088         struct smc_sock *smc;
1089         struct rb_node *node;
1090         int i, rc = 0;
1091
1092         /* link is inactive, wake up tx waiters */
1093         smc_wr_wakeup_tx_wait(from_lnk);
1094
1095         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1096                 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
1097                         continue;
1098                 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
1099                     from_lnk->ibport == lgr->lnk[i].ibport) {
1100                         continue;
1101                 }
1102                 to_lnk = &lgr->lnk[i];
1103                 break;
1104         }
1105         if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
1106                 smc_lgr_terminate_sched(lgr);
1107                 return NULL;
1108         }
1109 again:
1110         read_lock_bh(&lgr->conns_lock);
1111         for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
1112                 conn = rb_entry(node, struct smc_connection, alert_node);
1113                 if (conn->lnk != from_lnk)
1114                         continue;
1115                 smc = container_of(conn, struct smc_sock, conn);
1116                 /* conn->lnk not yet set in SMC_INIT state */
1117                 if (smc->sk.sk_state == SMC_INIT)
1118                         continue;
1119                 if (smc->sk.sk_state == SMC_CLOSED ||
1120                     smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
1121                     smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
1122                     smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
1123                     smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
1124                     smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
1125                     smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
1126                     smc->sk.sk_state == SMC_PEERABORTWAIT ||
1127                     smc->sk.sk_state == SMC_PROCESSABORT) {
1128                         spin_lock_bh(&conn->send_lock);
1129                         smc_switch_link_and_count(conn, to_lnk);
1130                         spin_unlock_bh(&conn->send_lock);
1131                         continue;
1132                 }
1133                 sock_hold(&smc->sk);
1134                 read_unlock_bh(&lgr->conns_lock);
1135                 /* pre-fetch buffer outside of send_lock, might sleep */
1136                 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
1137                 if (rc)
1138                         goto err_out;
1139                 /* avoid race with smcr_tx_sndbuf_nonempty() */
1140                 spin_lock_bh(&conn->send_lock);
1141                 smc_switch_link_and_count(conn, to_lnk);
1142                 rc = smc_switch_cursor(smc, pend, wr_buf);
1143                 spin_unlock_bh(&conn->send_lock);
1144                 sock_put(&smc->sk);
1145                 if (rc)
1146                         goto err_out;
1147                 goto again;
1148         }
1149         read_unlock_bh(&lgr->conns_lock);
1150         smc_wr_tx_link_put(to_lnk);
1151         return to_lnk;
1152
1153 err_out:
1154         smcr_link_down_cond_sched(to_lnk);
1155         smc_wr_tx_link_put(to_lnk);
1156         return NULL;
1157 }
1158
1159 static void smcr_buf_unuse(struct smc_buf_desc *buf_desc, bool is_rmb,
1160                            struct smc_link_group *lgr)
1161 {
1162         struct rw_semaphore *lock;      /* lock buffer list */
1163         int rc;
1164
1165         if (is_rmb && buf_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1166                 /* unregister rmb with peer */
1167                 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1168                 if (!rc) {
1169                         /* protect against smc_llc_cli_rkey_exchange() */
1170                         down_read(&lgr->llc_conf_mutex);
1171                         smc_llc_do_delete_rkey(lgr, buf_desc);
1172                         buf_desc->is_conf_rkey = false;
1173                         up_read(&lgr->llc_conf_mutex);
1174                         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1175                 }
1176         }
1177
1178         if (buf_desc->is_reg_err) {
1179                 /* buf registration failed, reuse not possible */
1180                 lock = is_rmb ? &lgr->rmbs_lock :
1181                                 &lgr->sndbufs_lock;
1182                 down_write(lock);
1183                 smc_lgr_buf_list_del(lgr, is_rmb, buf_desc);
1184                 up_write(lock);
1185
1186                 smc_buf_free(lgr, is_rmb, buf_desc);
1187         } else {
1188                 /* memzero_explicit provides potential memory barrier semantics */
1189                 memzero_explicit(buf_desc->cpu_addr, buf_desc->len);
1190                 WRITE_ONCE(buf_desc->used, 0);
1191         }
1192 }
1193
1194 static void smcd_buf_detach(struct smc_connection *conn)
1195 {
1196         struct smcd_dev *smcd = conn->lgr->smcd;
1197         u64 peer_token = conn->peer_token;
1198
1199         if (!conn->sndbuf_desc)
1200                 return;
1201
1202         smc_ism_detach_dmb(smcd, peer_token);
1203
1204         kfree(conn->sndbuf_desc);
1205         conn->sndbuf_desc = NULL;
1206 }
1207
1208 static void smc_buf_unuse(struct smc_connection *conn,
1209                           struct smc_link_group *lgr)
1210 {
1211         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1212         bool is_smcd = lgr->is_smcd;
1213         int bufsize;
1214
1215         if (conn->sndbuf_desc) {
1216                 bufsize = conn->sndbuf_desc->len;
1217                 if (!is_smcd && conn->sndbuf_desc->is_vm) {
1218                         smcr_buf_unuse(conn->sndbuf_desc, false, lgr);
1219                 } else {
1220                         memzero_explicit(conn->sndbuf_desc->cpu_addr, bufsize);
1221                         WRITE_ONCE(conn->sndbuf_desc->used, 0);
1222                 }
1223                 SMC_STAT_RMB_SIZE(smc, is_smcd, false, false, bufsize);
1224         }
1225         if (conn->rmb_desc) {
1226                 bufsize = conn->rmb_desc->len;
1227                 if (!is_smcd) {
1228                         smcr_buf_unuse(conn->rmb_desc, true, lgr);
1229                 } else {
1230                         bufsize += sizeof(struct smcd_cdc_msg);
1231                         memzero_explicit(conn->rmb_desc->cpu_addr, bufsize);
1232                         WRITE_ONCE(conn->rmb_desc->used, 0);
1233                 }
1234                 SMC_STAT_RMB_SIZE(smc, is_smcd, true, false, bufsize);
1235         }
1236 }
1237
1238 /* remove a finished connection from its link group */
1239 void smc_conn_free(struct smc_connection *conn)
1240 {
1241         struct smc_link_group *lgr = conn->lgr;
1242
1243         if (!lgr || conn->freed)
1244                 /* Connection has never been registered in a
1245                  * link group, or has already been freed.
1246                  */
1247                 return;
1248
1249         conn->freed = 1;
1250         if (!smc_conn_lgr_valid(conn))
1251                 /* Connection has already unregistered from
1252                  * link group.
1253                  */
1254                 goto lgr_put;
1255
1256         if (lgr->is_smcd) {
1257                 if (!list_empty(&lgr->list))
1258                         smc_ism_unset_conn(conn);
1259                 if (smc_ism_support_dmb_nocopy(lgr->smcd))
1260                         smcd_buf_detach(conn);
1261                 tasklet_kill(&conn->rx_tsklet);
1262         } else {
1263                 smc_cdc_wait_pend_tx_wr(conn);
1264                 if (current_work() != &conn->abort_work)
1265                         cancel_work_sync(&conn->abort_work);
1266         }
1267         if (!list_empty(&lgr->list)) {
1268                 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1269                 smc_lgr_unregister_conn(conn);
1270         }
1271
1272         if (!lgr->conns_num)
1273                 smc_lgr_schedule_free_work(lgr);
1274 lgr_put:
1275         if (!lgr->is_smcd)
1276                 smcr_link_put(conn->lnk); /* link_hold in smc_conn_create() */
1277         smc_lgr_put(lgr); /* lgr_hold in smc_conn_create() */
1278 }
1279
1280 /* unregister a link from a buf_desc */
1281 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1282                                 struct smc_link *lnk)
1283 {
1284         if (is_rmb || buf_desc->is_vm)
1285                 buf_desc->is_reg_mr[lnk->link_idx] = false;
1286         if (!buf_desc->is_map_ib[lnk->link_idx])
1287                 return;
1288
1289         if ((is_rmb || buf_desc->is_vm) &&
1290             buf_desc->mr[lnk->link_idx]) {
1291                 smc_ib_put_memory_region(buf_desc->mr[lnk->link_idx]);
1292                 buf_desc->mr[lnk->link_idx] = NULL;
1293         }
1294         if (is_rmb)
1295                 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1296         else
1297                 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1298
1299         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1300         buf_desc->is_map_ib[lnk->link_idx] = false;
1301 }
1302
1303 /* unmap all buffers of lgr for a deleted link */
1304 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1305 {
1306         struct smc_link_group *lgr = lnk->lgr;
1307         struct smc_buf_desc *buf_desc, *bf;
1308         int i;
1309
1310         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1311                 down_write(&lgr->rmbs_lock);
1312                 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1313                         smcr_buf_unmap_link(buf_desc, true, lnk);
1314                 up_write(&lgr->rmbs_lock);
1315
1316                 down_write(&lgr->sndbufs_lock);
1317                 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1318                                          list)
1319                         smcr_buf_unmap_link(buf_desc, false, lnk);
1320                 up_write(&lgr->sndbufs_lock);
1321         }
1322 }
1323
1324 static void smcr_rtoken_clear_link(struct smc_link *lnk)
1325 {
1326         struct smc_link_group *lgr = lnk->lgr;
1327         int i;
1328
1329         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1330                 lgr->rtokens[i][lnk->link_idx].rkey = 0;
1331                 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1332         }
1333 }
1334
1335 static void __smcr_link_clear(struct smc_link *lnk)
1336 {
1337         struct smc_link_group *lgr = lnk->lgr;
1338         struct smc_ib_device *smcibdev;
1339
1340         smc_wr_free_link_mem(lnk);
1341         smc_ibdev_cnt_dec(lnk);
1342         put_device(&lnk->smcibdev->ibdev->dev);
1343         smcibdev = lnk->smcibdev;
1344         memset(lnk, 0, sizeof(struct smc_link));
1345         lnk->state = SMC_LNK_UNUSED;
1346         if (!atomic_dec_return(&smcibdev->lnk_cnt))
1347                 wake_up(&smcibdev->lnks_deleted);
1348         smc_lgr_put(lgr); /* lgr_hold in smcr_link_init() */
1349 }
1350
1351 /* must be called under lgr->llc_conf_mutex lock */
1352 void smcr_link_clear(struct smc_link *lnk, bool log)
1353 {
1354         if (!lnk->lgr || lnk->clearing ||
1355             lnk->state == SMC_LNK_UNUSED)
1356                 return;
1357         lnk->clearing = 1;
1358         lnk->peer_qpn = 0;
1359         smc_llc_link_clear(lnk, log);
1360         smcr_buf_unmap_lgr(lnk);
1361         smcr_rtoken_clear_link(lnk);
1362         smc_ib_modify_qp_error(lnk);
1363         smc_wr_free_link(lnk);
1364         smc_ib_destroy_queue_pair(lnk);
1365         smc_ib_dealloc_protection_domain(lnk);
1366         smcr_link_put(lnk); /* theoretically last link_put */
1367 }
1368
1369 void smcr_link_hold(struct smc_link *lnk)
1370 {
1371         refcount_inc(&lnk->refcnt);
1372 }
1373
1374 void smcr_link_put(struct smc_link *lnk)
1375 {
1376         if (refcount_dec_and_test(&lnk->refcnt))
1377                 __smcr_link_clear(lnk);
1378 }
1379
1380 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1381                           struct smc_buf_desc *buf_desc)
1382 {
1383         int i;
1384
1385         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1386                 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1387
1388         if (!buf_desc->is_vm && buf_desc->pages)
1389                 __free_pages(buf_desc->pages, buf_desc->order);
1390         else if (buf_desc->is_vm && buf_desc->cpu_addr)
1391                 vfree(buf_desc->cpu_addr);
1392         kfree(buf_desc);
1393 }
1394
1395 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1396                           struct smc_buf_desc *buf_desc)
1397 {
1398         if (is_dmb) {
1399                 /* restore original buf len */
1400                 buf_desc->len += sizeof(struct smcd_cdc_msg);
1401                 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1402         } else {
1403                 kfree(buf_desc->cpu_addr);
1404         }
1405         kfree(buf_desc);
1406 }
1407
1408 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1409                          struct smc_buf_desc *buf_desc)
1410 {
1411         if (lgr->is_smcd)
1412                 smcd_buf_free(lgr, is_rmb, buf_desc);
1413         else
1414                 smcr_buf_free(lgr, is_rmb, buf_desc);
1415 }
1416
1417 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1418 {
1419         struct smc_buf_desc *buf_desc, *bf_desc;
1420         struct list_head *buf_list;
1421         int i;
1422
1423         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1424                 if (is_rmb)
1425                         buf_list = &lgr->rmbs[i];
1426                 else
1427                         buf_list = &lgr->sndbufs[i];
1428                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1429                                          list) {
1430                         smc_lgr_buf_list_del(lgr, is_rmb, buf_desc);
1431                         smc_buf_free(lgr, is_rmb, buf_desc);
1432                 }
1433         }
1434 }
1435
1436 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1437 {
1438         /* free send buffers */
1439         __smc_lgr_free_bufs(lgr, false);
1440         /* free rmbs */
1441         __smc_lgr_free_bufs(lgr, true);
1442 }
1443
1444 /* won't be freed until no one accesses to lgr anymore */
1445 static void __smc_lgr_free(struct smc_link_group *lgr)
1446 {
1447         smc_lgr_free_bufs(lgr);
1448         if (lgr->is_smcd) {
1449                 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1450                         wake_up(&lgr->smcd->lgrs_deleted);
1451         } else {
1452                 smc_wr_free_lgr_mem(lgr);
1453                 if (!atomic_dec_return(&lgr_cnt))
1454                         wake_up(&lgrs_deleted);
1455         }
1456         kfree(lgr);
1457 }
1458
1459 /* remove a link group */
1460 static void smc_lgr_free(struct smc_link_group *lgr)
1461 {
1462         int i;
1463
1464         if (!lgr->is_smcd) {
1465                 down_write(&lgr->llc_conf_mutex);
1466                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1467                         if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1468                                 smcr_link_clear(&lgr->lnk[i], false);
1469                 }
1470                 up_write(&lgr->llc_conf_mutex);
1471                 smc_llc_lgr_clear(lgr);
1472         }
1473
1474         destroy_workqueue(lgr->tx_wq);
1475         if (lgr->is_smcd) {
1476                 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1477                 put_device(lgr->smcd->ops->get_dev(lgr->smcd));
1478         }
1479         smc_lgr_put(lgr); /* theoretically last lgr_put */
1480 }
1481
1482 void smc_lgr_hold(struct smc_link_group *lgr)
1483 {
1484         refcount_inc(&lgr->refcnt);
1485 }
1486
1487 void smc_lgr_put(struct smc_link_group *lgr)
1488 {
1489         if (refcount_dec_and_test(&lgr->refcnt))
1490                 __smc_lgr_free(lgr);
1491 }
1492
1493 static void smc_sk_wake_ups(struct smc_sock *smc)
1494 {
1495         smc->sk.sk_write_space(&smc->sk);
1496         smc->sk.sk_data_ready(&smc->sk);
1497         smc->sk.sk_state_change(&smc->sk);
1498 }
1499
1500 /* kill a connection */
1501 static void smc_conn_kill(struct smc_connection *conn, bool soft)
1502 {
1503         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1504
1505         if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1506                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1507         else
1508                 smc_close_abort(conn);
1509         conn->killed = 1;
1510         smc->sk.sk_err = ECONNABORTED;
1511         smc_sk_wake_ups(smc);
1512         if (conn->lgr->is_smcd) {
1513                 smc_ism_unset_conn(conn);
1514                 if (smc_ism_support_dmb_nocopy(conn->lgr->smcd))
1515                         smcd_buf_detach(conn);
1516                 if (soft)
1517                         tasklet_kill(&conn->rx_tsklet);
1518                 else
1519                         tasklet_unlock_wait(&conn->rx_tsklet);
1520         } else {
1521                 smc_cdc_wait_pend_tx_wr(conn);
1522         }
1523         smc_lgr_unregister_conn(conn);
1524         smc_close_active_abort(smc);
1525 }
1526
1527 static void smc_lgr_cleanup(struct smc_link_group *lgr)
1528 {
1529         if (lgr->is_smcd) {
1530                 smc_ism_signal_shutdown(lgr);
1531         } else {
1532                 u32 rsn = lgr->llc_termination_rsn;
1533
1534                 if (!rsn)
1535                         rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1536                 smc_llc_send_link_delete_all(lgr, false, rsn);
1537                 smcr_lgr_link_deactivate_all(lgr);
1538         }
1539 }
1540
1541 /* terminate link group
1542  * @soft: true if link group shutdown can take its time
1543  *        false if immediate link group shutdown is required
1544  */
1545 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1546 {
1547         struct smc_connection *conn;
1548         struct smc_sock *smc;
1549         struct rb_node *node;
1550
1551         if (lgr->terminating)
1552                 return; /* lgr already terminating */
1553         /* cancel free_work sync, will terminate when lgr->freeing is set */
1554         cancel_delayed_work(&lgr->free_work);
1555         lgr->terminating = 1;
1556
1557         /* kill remaining link group connections */
1558         read_lock_bh(&lgr->conns_lock);
1559         node = rb_first(&lgr->conns_all);
1560         while (node) {
1561                 read_unlock_bh(&lgr->conns_lock);
1562                 conn = rb_entry(node, struct smc_connection, alert_node);
1563                 smc = container_of(conn, struct smc_sock, conn);
1564                 sock_hold(&smc->sk); /* sock_put below */
1565                 lock_sock(&smc->sk);
1566                 smc_conn_kill(conn, soft);
1567                 release_sock(&smc->sk);
1568                 sock_put(&smc->sk); /* sock_hold above */
1569                 read_lock_bh(&lgr->conns_lock);
1570                 node = rb_first(&lgr->conns_all);
1571         }
1572         read_unlock_bh(&lgr->conns_lock);
1573         smc_lgr_cleanup(lgr);
1574         smc_lgr_free(lgr);
1575 }
1576
1577 /* unlink link group and schedule termination */
1578 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1579 {
1580         spinlock_t *lgr_lock;
1581
1582         smc_lgr_list_head(lgr, &lgr_lock);
1583         spin_lock_bh(lgr_lock);
1584         if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1585                 spin_unlock_bh(lgr_lock);
1586                 return; /* lgr already terminating */
1587         }
1588         list_del_init(&lgr->list);
1589         lgr->freeing = 1;
1590         spin_unlock_bh(lgr_lock);
1591         schedule_work(&lgr->terminate_work);
1592 }
1593
1594 /* Called when peer lgr shutdown (regularly or abnormally) is received */
1595 void smc_smcd_terminate(struct smcd_dev *dev, struct smcd_gid *peer_gid,
1596                         unsigned short vlan)
1597 {
1598         struct smc_link_group *lgr, *l;
1599         LIST_HEAD(lgr_free_list);
1600
1601         /* run common cleanup function and build free list */
1602         spin_lock_bh(&dev->lgr_lock);
1603         list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1604                 if ((!peer_gid->gid ||
1605                      (lgr->peer_gid.gid == peer_gid->gid &&
1606                       !smc_ism_is_emulated(dev) ? 1 :
1607                       lgr->peer_gid.gid_ext == peer_gid->gid_ext)) &&
1608                     (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1609                         if (peer_gid->gid) /* peer triggered termination */
1610                                 lgr->peer_shutdown = 1;
1611                         list_move(&lgr->list, &lgr_free_list);
1612                         lgr->freeing = 1;
1613                 }
1614         }
1615         spin_unlock_bh(&dev->lgr_lock);
1616
1617         /* cancel the regular free workers and actually free lgrs */
1618         list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1619                 list_del_init(&lgr->list);
1620                 schedule_work(&lgr->terminate_work);
1621         }
1622 }
1623
1624 /* Called when an SMCD device is removed or the smc module is unloaded */
1625 void smc_smcd_terminate_all(struct smcd_dev *smcd)
1626 {
1627         struct smc_link_group *lgr, *lg;
1628         LIST_HEAD(lgr_free_list);
1629
1630         spin_lock_bh(&smcd->lgr_lock);
1631         list_splice_init(&smcd->lgr_list, &lgr_free_list);
1632         list_for_each_entry(lgr, &lgr_free_list, list)
1633                 lgr->freeing = 1;
1634         spin_unlock_bh(&smcd->lgr_lock);
1635
1636         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1637                 list_del_init(&lgr->list);
1638                 __smc_lgr_terminate(lgr, false);
1639         }
1640
1641         if (atomic_read(&smcd->lgr_cnt))
1642                 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1643 }
1644
1645 /* Called when an SMCR device is removed or the smc module is unloaded.
1646  * If smcibdev is given, all SMCR link groups using this device are terminated.
1647  * If smcibdev is NULL, all SMCR link groups are terminated.
1648  */
1649 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1650 {
1651         struct smc_link_group *lgr, *lg;
1652         LIST_HEAD(lgr_free_list);
1653         int i;
1654
1655         spin_lock_bh(&smc_lgr_list.lock);
1656         if (!smcibdev) {
1657                 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1658                 list_for_each_entry(lgr, &lgr_free_list, list)
1659                         lgr->freeing = 1;
1660         } else {
1661                 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1662                         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1663                                 if (lgr->lnk[i].smcibdev == smcibdev)
1664                                         smcr_link_down_cond_sched(&lgr->lnk[i]);
1665                         }
1666                 }
1667         }
1668         spin_unlock_bh(&smc_lgr_list.lock);
1669
1670         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1671                 list_del_init(&lgr->list);
1672                 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1673                 __smc_lgr_terminate(lgr, false);
1674         }
1675
1676         if (smcibdev) {
1677                 if (atomic_read(&smcibdev->lnk_cnt))
1678                         wait_event(smcibdev->lnks_deleted,
1679                                    !atomic_read(&smcibdev->lnk_cnt));
1680         } else {
1681                 if (atomic_read(&lgr_cnt))
1682                         wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1683         }
1684 }
1685
1686 /* set new lgr type and clear all asymmetric link tagging */
1687 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1688 {
1689         char *lgr_type = "";
1690         int i;
1691
1692         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1693                 if (smc_link_usable(&lgr->lnk[i]))
1694                         lgr->lnk[i].link_is_asym = false;
1695         if (lgr->type == new_type)
1696                 return;
1697         lgr->type = new_type;
1698
1699         switch (lgr->type) {
1700         case SMC_LGR_NONE:
1701                 lgr_type = "NONE";
1702                 break;
1703         case SMC_LGR_SINGLE:
1704                 lgr_type = "SINGLE";
1705                 break;
1706         case SMC_LGR_SYMMETRIC:
1707                 lgr_type = "SYMMETRIC";
1708                 break;
1709         case SMC_LGR_ASYMMETRIC_PEER:
1710                 lgr_type = "ASYMMETRIC_PEER";
1711                 break;
1712         case SMC_LGR_ASYMMETRIC_LOCAL:
1713                 lgr_type = "ASYMMETRIC_LOCAL";
1714                 break;
1715         }
1716         pr_warn_ratelimited("smc: SMC-R lg %*phN net %llu state changed: "
1717                             "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1718                             lgr->net->net_cookie, lgr_type, lgr->pnet_id);
1719 }
1720
1721 /* set new lgr type and tag a link as asymmetric */
1722 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1723                             enum smc_lgr_type new_type, int asym_lnk_idx)
1724 {
1725         smcr_lgr_set_type(lgr, new_type);
1726         lgr->lnk[asym_lnk_idx].link_is_asym = true;
1727 }
1728
1729 /* abort connection, abort_work scheduled from tasklet context */
1730 static void smc_conn_abort_work(struct work_struct *work)
1731 {
1732         struct smc_connection *conn = container_of(work,
1733                                                    struct smc_connection,
1734                                                    abort_work);
1735         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1736
1737         lock_sock(&smc->sk);
1738         smc_conn_kill(conn, true);
1739         release_sock(&smc->sk);
1740         sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1741 }
1742
1743 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1744 {
1745         struct smc_link_group *lgr, *n;
1746
1747         spin_lock_bh(&smc_lgr_list.lock);
1748         list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1749                 struct smc_link *link;
1750
1751                 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1752                             SMC_MAX_PNETID_LEN) ||
1753                     lgr->type == SMC_LGR_SYMMETRIC ||
1754                     lgr->type == SMC_LGR_ASYMMETRIC_PEER ||
1755                     !rdma_dev_access_netns(smcibdev->ibdev, lgr->net))
1756                         continue;
1757
1758                 if (lgr->type == SMC_LGR_SINGLE && lgr->max_links <= 1)
1759                         continue;
1760
1761                 /* trigger local add link processing */
1762                 link = smc_llc_usable_link(lgr);
1763                 if (link)
1764                         smc_llc_add_link_local(link);
1765         }
1766         spin_unlock_bh(&smc_lgr_list.lock);
1767 }
1768
1769 /* link is down - switch connections to alternate link,
1770  * must be called under lgr->llc_conf_mutex lock
1771  */
1772 static void smcr_link_down(struct smc_link *lnk)
1773 {
1774         struct smc_link_group *lgr = lnk->lgr;
1775         struct smc_link *to_lnk;
1776         int del_link_id;
1777
1778         if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1779                 return;
1780
1781         to_lnk = smc_switch_conns(lgr, lnk, true);
1782         if (!to_lnk) { /* no backup link available */
1783                 smcr_link_clear(lnk, true);
1784                 return;
1785         }
1786         smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1787         del_link_id = lnk->link_id;
1788
1789         if (lgr->role == SMC_SERV) {
1790                 /* trigger local delete link processing */
1791                 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1792         } else {
1793                 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1794                         /* another llc task is ongoing */
1795                         up_write(&lgr->llc_conf_mutex);
1796                         wait_event_timeout(lgr->llc_flow_waiter,
1797                                 (list_empty(&lgr->list) ||
1798                                  lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1799                                 SMC_LLC_WAIT_TIME);
1800                         down_write(&lgr->llc_conf_mutex);
1801                 }
1802                 if (!list_empty(&lgr->list)) {
1803                         smc_llc_send_delete_link(to_lnk, del_link_id,
1804                                                  SMC_LLC_REQ, true,
1805                                                  SMC_LLC_DEL_LOST_PATH);
1806                         smcr_link_clear(lnk, true);
1807                 }
1808                 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1809         }
1810 }
1811
1812 /* must be called under lgr->llc_conf_mutex lock */
1813 void smcr_link_down_cond(struct smc_link *lnk)
1814 {
1815         if (smc_link_downing(&lnk->state)) {
1816                 trace_smcr_link_down(lnk, __builtin_return_address(0));
1817                 smcr_link_down(lnk);
1818         }
1819 }
1820
1821 /* will get the lgr->llc_conf_mutex lock */
1822 void smcr_link_down_cond_sched(struct smc_link *lnk)
1823 {
1824         if (smc_link_downing(&lnk->state)) {
1825                 trace_smcr_link_down(lnk, __builtin_return_address(0));
1826                 smcr_link_hold(lnk); /* smcr_link_put in link_down_wrk */
1827                 if (!schedule_work(&lnk->link_down_wrk))
1828                         smcr_link_put(lnk);
1829         }
1830 }
1831
1832 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1833 {
1834         struct smc_link_group *lgr, *n;
1835         int i;
1836
1837         list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1838                 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1839                             SMC_MAX_PNETID_LEN))
1840                         continue; /* lgr is not affected */
1841                 if (list_empty(&lgr->list))
1842                         continue;
1843                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1844                         struct smc_link *lnk = &lgr->lnk[i];
1845
1846                         if (smc_link_usable(lnk) &&
1847                             lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1848                                 smcr_link_down_cond_sched(lnk);
1849                 }
1850         }
1851 }
1852
1853 static void smc_link_down_work(struct work_struct *work)
1854 {
1855         struct smc_link *link = container_of(work, struct smc_link,
1856                                              link_down_wrk);
1857         struct smc_link_group *lgr = link->lgr;
1858
1859         if (list_empty(&lgr->list))
1860                 goto out;
1861         wake_up_all(&lgr->llc_msg_waiter);
1862         down_write(&lgr->llc_conf_mutex);
1863         smcr_link_down(link);
1864         up_write(&lgr->llc_conf_mutex);
1865
1866 out:
1867         smcr_link_put(link); /* smcr_link_hold by schedulers of link_down_work */
1868 }
1869
1870 static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
1871                                   struct netdev_nested_priv *priv)
1872 {
1873         unsigned short *vlan_id = (unsigned short *)priv->data;
1874
1875         if (is_vlan_dev(lower_dev)) {
1876                 *vlan_id = vlan_dev_vlan_id(lower_dev);
1877                 return 1;
1878         }
1879
1880         return 0;
1881 }
1882
1883 /* Determine vlan of internal TCP socket. */
1884 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1885 {
1886         struct dst_entry *dst = sk_dst_get(clcsock->sk);
1887         struct netdev_nested_priv priv;
1888         struct net_device *ndev;
1889         int rc = 0;
1890
1891         ini->vlan_id = 0;
1892         if (!dst) {
1893                 rc = -ENOTCONN;
1894                 goto out;
1895         }
1896         if (!dst->dev) {
1897                 rc = -ENODEV;
1898                 goto out_rel;
1899         }
1900
1901         ndev = dst->dev;
1902         if (is_vlan_dev(ndev)) {
1903                 ini->vlan_id = vlan_dev_vlan_id(ndev);
1904                 goto out_rel;
1905         }
1906
1907         priv.data = (void *)&ini->vlan_id;
1908         rtnl_lock();
1909         netdev_walk_all_lower_dev(ndev, smc_vlan_by_tcpsk_walk, &priv);
1910         rtnl_unlock();
1911
1912 out_rel:
1913         dst_release(dst);
1914 out:
1915         return rc;
1916 }
1917
1918 static bool smcr_lgr_match(struct smc_link_group *lgr, u8 smcr_version,
1919                            u8 peer_systemid[],
1920                            u8 peer_gid[],
1921                            u8 peer_mac_v1[],
1922                            enum smc_lgr_role role, u32 clcqpn,
1923                            struct net *net)
1924 {
1925         struct smc_link *lnk;
1926         int i;
1927
1928         if (memcmp(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN) ||
1929             lgr->role != role)
1930                 return false;
1931
1932         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1933                 lnk = &lgr->lnk[i];
1934
1935                 if (!smc_link_active(lnk))
1936                         continue;
1937                 /* use verbs API to check netns, instead of lgr->net */
1938                 if (!rdma_dev_access_netns(lnk->smcibdev->ibdev, net))
1939                         return false;
1940                 if ((lgr->role == SMC_SERV || lnk->peer_qpn == clcqpn) &&
1941                     !memcmp(lnk->peer_gid, peer_gid, SMC_GID_SIZE) &&
1942                     (smcr_version == SMC_V2 ||
1943                      !memcmp(lnk->peer_mac, peer_mac_v1, ETH_ALEN)))
1944                         return true;
1945         }
1946         return false;
1947 }
1948
1949 static bool smcd_lgr_match(struct smc_link_group *lgr,
1950                            struct smcd_dev *smcismdev,
1951                            struct smcd_gid *peer_gid)
1952 {
1953         if (lgr->peer_gid.gid != peer_gid->gid ||
1954             lgr->smcd != smcismdev)
1955                 return false;
1956
1957         if (smc_ism_is_emulated(smcismdev) &&
1958             lgr->peer_gid.gid_ext != peer_gid->gid_ext)
1959                 return false;
1960
1961         return true;
1962 }
1963
1964 /* create a new SMC connection (and a new link group if necessary) */
1965 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1966 {
1967         struct smc_connection *conn = &smc->conn;
1968         struct net *net = sock_net(&smc->sk);
1969         struct list_head *lgr_list;
1970         struct smc_link_group *lgr;
1971         enum smc_lgr_role role;
1972         spinlock_t *lgr_lock;
1973         int rc = 0;
1974
1975         lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1976                                   &smc_lgr_list.list;
1977         lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1978                                   &smc_lgr_list.lock;
1979         ini->first_contact_local = 1;
1980         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1981         if (role == SMC_CLNT && ini->first_contact_peer)
1982                 /* create new link group as well */
1983                 goto create;
1984
1985         /* determine if an existing link group can be reused */
1986         spin_lock_bh(lgr_lock);
1987         list_for_each_entry(lgr, lgr_list, list) {
1988                 write_lock_bh(&lgr->conns_lock);
1989                 if ((ini->is_smcd ?
1990                      smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1991                                     &ini->ism_peer_gid[ini->ism_selected]) :
1992                      smcr_lgr_match(lgr, ini->smcr_version,
1993                                     ini->peer_systemid,
1994                                     ini->peer_gid, ini->peer_mac, role,
1995                                     ini->ib_clcqpn, net)) &&
1996                     !lgr->sync_err &&
1997                     (ini->smcd_version == SMC_V2 ||
1998                      lgr->vlan_id == ini->vlan_id) &&
1999                     (role == SMC_CLNT || ini->is_smcd ||
2000                     (lgr->conns_num < lgr->max_conns &&
2001                       !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
2002                         /* link group found */
2003                         ini->first_contact_local = 0;
2004                         conn->lgr = lgr;
2005                         rc = smc_lgr_register_conn(conn, false);
2006                         write_unlock_bh(&lgr->conns_lock);
2007                         if (!rc && delayed_work_pending(&lgr->free_work))
2008                                 cancel_delayed_work(&lgr->free_work);
2009                         break;
2010                 }
2011                 write_unlock_bh(&lgr->conns_lock);
2012         }
2013         spin_unlock_bh(lgr_lock);
2014         if (rc)
2015                 return rc;
2016
2017         if (role == SMC_CLNT && !ini->first_contact_peer &&
2018             ini->first_contact_local) {
2019                 /* Server reuses a link group, but Client wants to start
2020                  * a new one
2021                  * send out_of_sync decline, reason synchr. error
2022                  */
2023                 return SMC_CLC_DECL_SYNCERR;
2024         }
2025
2026 create:
2027         if (ini->first_contact_local) {
2028                 rc = smc_lgr_create(smc, ini);
2029                 if (rc)
2030                         goto out;
2031                 lgr = conn->lgr;
2032                 write_lock_bh(&lgr->conns_lock);
2033                 rc = smc_lgr_register_conn(conn, true);
2034                 write_unlock_bh(&lgr->conns_lock);
2035                 if (rc) {
2036                         smc_lgr_cleanup_early(lgr);
2037                         goto out;
2038                 }
2039         }
2040         smc_lgr_hold(conn->lgr); /* lgr_put in smc_conn_free() */
2041         if (!conn->lgr->is_smcd)
2042                 smcr_link_hold(conn->lnk); /* link_put in smc_conn_free() */
2043         conn->freed = 0;
2044         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
2045         conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
2046         conn->urg_state = SMC_URG_READ;
2047         init_waitqueue_head(&conn->cdc_pend_tx_wq);
2048         INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
2049         if (ini->is_smcd) {
2050                 conn->rx_off = sizeof(struct smcd_cdc_msg);
2051                 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
2052         } else {
2053                 conn->rx_off = 0;
2054         }
2055 #ifndef KERNEL_HAS_ATOMIC64
2056         spin_lock_init(&conn->acurs_lock);
2057 #endif
2058
2059 out:
2060         return rc;
2061 }
2062
2063 #define SMCD_DMBE_SIZES         6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
2064 #define SMCR_RMBE_SIZES         15 /* 0 -> 16KB, 1 -> 32KB, .. 15 -> 512MB */
2065
2066 /* convert the RMB size into the compressed notation (minimum 16K, see
2067  * SMCD/R_DMBE_SIZES.
2068  * In contrast to plain ilog2, this rounds towards the next power of 2,
2069  * so the socket application gets at least its desired sndbuf / rcvbuf size.
2070  */
2071 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
2072 {
2073         u8 compressed;
2074
2075         if (size <= SMC_BUF_MIN_SIZE)
2076                 return 0;
2077
2078         size = (size - 1) >> 14;  /* convert to 16K multiple */
2079         compressed = min_t(u8, ilog2(size) + 1,
2080                            is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
2081
2082 #ifdef CONFIG_ARCH_NO_SG_CHAIN
2083         if (!is_smcd && is_rmb)
2084                 /* RMBs are backed by & limited to max size of scatterlists */
2085                 compressed = min_t(u8, compressed, ilog2((SG_MAX_SINGLE_ALLOC * PAGE_SIZE) >> 14));
2086 #endif
2087
2088         return compressed;
2089 }
2090
2091 /* convert the RMB size from compressed notation into integer */
2092 int smc_uncompress_bufsize(u8 compressed)
2093 {
2094         u32 size;
2095
2096         size = 0x00000001 << (((int)compressed) + 14);
2097         return (int)size;
2098 }
2099
2100 /* try to reuse a sndbuf or rmb description slot for a certain
2101  * buffer size; if not available, return NULL
2102  */
2103 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
2104                                              struct rw_semaphore *lock,
2105                                              struct list_head *buf_list)
2106 {
2107         struct smc_buf_desc *buf_slot;
2108
2109         down_read(lock);
2110         list_for_each_entry(buf_slot, buf_list, list) {
2111                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
2112                         up_read(lock);
2113                         return buf_slot;
2114                 }
2115         }
2116         up_read(lock);
2117         return NULL;
2118 }
2119
2120 /* one of the conditions for announcing a receiver's current window size is
2121  * that it "results in a minimum increase in the window size of 10% of the
2122  * receive buffer space" [RFC7609]
2123  */
2124 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
2125 {
2126         return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
2127 }
2128
2129 /* map an buf to a link */
2130 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
2131                              struct smc_link *lnk)
2132 {
2133         int rc, i, nents, offset, buf_size, size, access_flags;
2134         struct scatterlist *sg;
2135         void *buf;
2136
2137         if (buf_desc->is_map_ib[lnk->link_idx])
2138                 return 0;
2139
2140         if (buf_desc->is_vm) {
2141                 buf = buf_desc->cpu_addr;
2142                 buf_size = buf_desc->len;
2143                 offset = offset_in_page(buf_desc->cpu_addr);
2144                 nents = PAGE_ALIGN(buf_size + offset) / PAGE_SIZE;
2145         } else {
2146                 nents = 1;
2147         }
2148
2149         rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], nents, GFP_KERNEL);
2150         if (rc)
2151                 return rc;
2152
2153         if (buf_desc->is_vm) {
2154                 /* virtually contiguous buffer */
2155                 for_each_sg(buf_desc->sgt[lnk->link_idx].sgl, sg, nents, i) {
2156                         size = min_t(int, PAGE_SIZE - offset, buf_size);
2157                         sg_set_page(sg, vmalloc_to_page(buf), size, offset);
2158                         buf += size;
2159                         buf_size -= size;
2160                         offset = 0;
2161                 }
2162         } else {
2163                 /* physically contiguous buffer */
2164                 sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
2165                            buf_desc->cpu_addr, buf_desc->len);
2166         }
2167
2168         /* map sg table to DMA address */
2169         rc = smc_ib_buf_map_sg(lnk, buf_desc,
2170                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2171         /* SMC protocol depends on mapping to one DMA address only */
2172         if (rc != nents) {
2173                 rc = -EAGAIN;
2174                 goto free_table;
2175         }
2176
2177         buf_desc->is_dma_need_sync |=
2178                 smc_ib_is_sg_need_sync(lnk, buf_desc) << lnk->link_idx;
2179
2180         if (is_rmb || buf_desc->is_vm) {
2181                 /* create a new memory region for the RMB or vzalloced sndbuf */
2182                 access_flags = is_rmb ?
2183                                IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2184                                IB_ACCESS_LOCAL_WRITE;
2185
2186                 rc = smc_ib_get_memory_region(lnk->roce_pd, access_flags,
2187                                               buf_desc, lnk->link_idx);
2188                 if (rc)
2189                         goto buf_unmap;
2190                 smc_ib_sync_sg_for_device(lnk, buf_desc,
2191                                           is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2192         }
2193         buf_desc->is_map_ib[lnk->link_idx] = true;
2194         return 0;
2195
2196 buf_unmap:
2197         smc_ib_buf_unmap_sg(lnk, buf_desc,
2198                             is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2199 free_table:
2200         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
2201         return rc;
2202 }
2203
2204 /* register a new buf on IB device, rmb or vzalloced sndbuf
2205  * must be called under lgr->llc_conf_mutex lock
2206  */
2207 int smcr_link_reg_buf(struct smc_link *link, struct smc_buf_desc *buf_desc)
2208 {
2209         if (list_empty(&link->lgr->list))
2210                 return -ENOLINK;
2211         if (!buf_desc->is_reg_mr[link->link_idx]) {
2212                 /* register memory region for new buf */
2213                 if (buf_desc->is_vm)
2214                         buf_desc->mr[link->link_idx]->iova =
2215                                 (uintptr_t)buf_desc->cpu_addr;
2216                 if (smc_wr_reg_send(link, buf_desc->mr[link->link_idx])) {
2217                         buf_desc->is_reg_err = true;
2218                         return -EFAULT;
2219                 }
2220                 buf_desc->is_reg_mr[link->link_idx] = true;
2221         }
2222         return 0;
2223 }
2224
2225 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct rw_semaphore *lock,
2226                              struct list_head *lst, bool is_rmb)
2227 {
2228         struct smc_buf_desc *buf_desc, *bf;
2229         int rc = 0;
2230
2231         down_write(lock);
2232         list_for_each_entry_safe(buf_desc, bf, lst, list) {
2233                 if (!buf_desc->used)
2234                         continue;
2235                 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
2236                 if (rc)
2237                         goto out;
2238         }
2239 out:
2240         up_write(lock);
2241         return rc;
2242 }
2243
2244 /* map all used buffers of lgr for a new link */
2245 int smcr_buf_map_lgr(struct smc_link *lnk)
2246 {
2247         struct smc_link_group *lgr = lnk->lgr;
2248         int i, rc = 0;
2249
2250         for (i = 0; i < SMC_RMBE_SIZES; i++) {
2251                 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
2252                                        &lgr->rmbs[i], true);
2253                 if (rc)
2254                         return rc;
2255                 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
2256                                        &lgr->sndbufs[i], false);
2257                 if (rc)
2258                         return rc;
2259         }
2260         return 0;
2261 }
2262
2263 /* register all used buffers of lgr for a new link,
2264  * must be called under lgr->llc_conf_mutex lock
2265  */
2266 int smcr_buf_reg_lgr(struct smc_link *lnk)
2267 {
2268         struct smc_link_group *lgr = lnk->lgr;
2269         struct smc_buf_desc *buf_desc, *bf;
2270         int i, rc = 0;
2271
2272         /* reg all RMBs for a new link */
2273         down_write(&lgr->rmbs_lock);
2274         for (i = 0; i < SMC_RMBE_SIZES; i++) {
2275                 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
2276                         if (!buf_desc->used)
2277                                 continue;
2278                         rc = smcr_link_reg_buf(lnk, buf_desc);
2279                         if (rc) {
2280                                 up_write(&lgr->rmbs_lock);
2281                                 return rc;
2282                         }
2283                 }
2284         }
2285         up_write(&lgr->rmbs_lock);
2286
2287         if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2288                 return rc;
2289
2290         /* reg all vzalloced sndbufs for a new link */
2291         down_write(&lgr->sndbufs_lock);
2292         for (i = 0; i < SMC_RMBE_SIZES; i++) {
2293                 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], list) {
2294                         if (!buf_desc->used || !buf_desc->is_vm)
2295                                 continue;
2296                         rc = smcr_link_reg_buf(lnk, buf_desc);
2297                         if (rc) {
2298                                 up_write(&lgr->sndbufs_lock);
2299                                 return rc;
2300                         }
2301                 }
2302         }
2303         up_write(&lgr->sndbufs_lock);
2304         return rc;
2305 }
2306
2307 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
2308                                                 int bufsize)
2309 {
2310         struct smc_buf_desc *buf_desc;
2311
2312         /* try to alloc a new buffer */
2313         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2314         if (!buf_desc)
2315                 return ERR_PTR(-ENOMEM);
2316
2317         switch (lgr->buf_type) {
2318         case SMCR_PHYS_CONT_BUFS:
2319         case SMCR_MIXED_BUFS:
2320                 buf_desc->order = get_order(bufsize);
2321                 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
2322                                               __GFP_NOMEMALLOC | __GFP_COMP |
2323                                               __GFP_NORETRY | __GFP_ZERO,
2324                                               buf_desc->order);
2325                 if (buf_desc->pages) {
2326                         buf_desc->cpu_addr =
2327                                 (void *)page_address(buf_desc->pages);
2328                         buf_desc->len = bufsize;
2329                         buf_desc->is_vm = false;
2330                         break;
2331                 }
2332                 if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2333                         goto out;
2334                 fallthrough;    // try virtually contiguous buf
2335         case SMCR_VIRT_CONT_BUFS:
2336                 buf_desc->order = get_order(bufsize);
2337                 buf_desc->cpu_addr = vzalloc(PAGE_SIZE << buf_desc->order);
2338                 if (!buf_desc->cpu_addr)
2339                         goto out;
2340                 buf_desc->pages = NULL;
2341                 buf_desc->len = bufsize;
2342                 buf_desc->is_vm = true;
2343                 break;
2344         }
2345         return buf_desc;
2346
2347 out:
2348         kfree(buf_desc);
2349         return ERR_PTR(-EAGAIN);
2350 }
2351
2352 /* map buf_desc on all usable links,
2353  * unused buffers stay mapped as long as the link is up
2354  */
2355 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
2356                                      struct smc_buf_desc *buf_desc, bool is_rmb)
2357 {
2358         int i, rc = 0, cnt = 0;
2359
2360         /* protect against parallel link reconfiguration */
2361         down_read(&lgr->llc_conf_mutex);
2362         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2363                 struct smc_link *lnk = &lgr->lnk[i];
2364
2365                 if (!smc_link_usable(lnk))
2366                         continue;
2367                 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
2368                         rc = -ENOMEM;
2369                         goto out;
2370                 }
2371                 cnt++;
2372         }
2373 out:
2374         up_read(&lgr->llc_conf_mutex);
2375         if (!rc && !cnt)
2376                 rc = -EINVAL;
2377         return rc;
2378 }
2379
2380 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
2381                                                 bool is_dmb, int bufsize)
2382 {
2383         struct smc_buf_desc *buf_desc;
2384         int rc;
2385
2386         /* try to alloc a new DMB */
2387         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2388         if (!buf_desc)
2389                 return ERR_PTR(-ENOMEM);
2390         if (is_dmb) {
2391                 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2392                 if (rc) {
2393                         kfree(buf_desc);
2394                         if (rc == -ENOMEM)
2395                                 return ERR_PTR(-EAGAIN);
2396                         if (rc == -ENOSPC)
2397                                 return ERR_PTR(-ENOSPC);
2398                         return ERR_PTR(-EIO);
2399                 }
2400                 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2401                 /* CDC header stored in buf. So, pretend it was smaller */
2402                 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2403         } else {
2404                 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2405                                              __GFP_NOWARN | __GFP_NORETRY |
2406                                              __GFP_NOMEMALLOC);
2407                 if (!buf_desc->cpu_addr) {
2408                         kfree(buf_desc);
2409                         return ERR_PTR(-EAGAIN);
2410                 }
2411                 buf_desc->len = bufsize;
2412         }
2413         return buf_desc;
2414 }
2415
2416 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2417 {
2418         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2419         struct smc_connection *conn = &smc->conn;
2420         struct smc_link_group *lgr = conn->lgr;
2421         struct list_head *buf_list;
2422         int bufsize, bufsize_comp;
2423         struct rw_semaphore *lock;      /* lock buffer list */
2424         bool is_dgraded = false;
2425
2426         if (is_rmb)
2427                 /* use socket recv buffer size (w/o overhead) as start value */
2428                 bufsize = smc->sk.sk_rcvbuf / 2;
2429         else
2430                 /* use socket send buffer size (w/o overhead) as start value */
2431                 bufsize = smc->sk.sk_sndbuf / 2;
2432
2433         for (bufsize_comp = smc_compress_bufsize(bufsize, is_smcd, is_rmb);
2434              bufsize_comp >= 0; bufsize_comp--) {
2435                 if (is_rmb) {
2436                         lock = &lgr->rmbs_lock;
2437                         buf_list = &lgr->rmbs[bufsize_comp];
2438                 } else {
2439                         lock = &lgr->sndbufs_lock;
2440                         buf_list = &lgr->sndbufs[bufsize_comp];
2441                 }
2442                 bufsize = smc_uncompress_bufsize(bufsize_comp);
2443
2444                 /* check for reusable slot in the link group */
2445                 buf_desc = smc_buf_get_slot(bufsize_comp, lock, buf_list);
2446                 if (buf_desc) {
2447                         buf_desc->is_dma_need_sync = 0;
2448                         SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, true, bufsize);
2449                         SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2450                         break; /* found reusable slot */
2451                 }
2452
2453                 if (is_smcd)
2454                         buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2455                 else
2456                         buf_desc = smcr_new_buf_create(lgr, bufsize);
2457
2458                 if (PTR_ERR(buf_desc) == -ENOMEM)
2459                         break;
2460                 if (IS_ERR(buf_desc)) {
2461                         if (!is_dgraded) {
2462                                 is_dgraded = true;
2463                                 SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2464                         }
2465                         continue;
2466                 }
2467
2468                 SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2469                 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, true, bufsize);
2470                 buf_desc->used = 1;
2471                 down_write(lock);
2472                 smc_lgr_buf_list_add(lgr, is_rmb, buf_list, buf_desc);
2473                 up_write(lock);
2474                 break; /* found */
2475         }
2476
2477         if (IS_ERR(buf_desc))
2478                 return PTR_ERR(buf_desc);
2479
2480         if (!is_smcd) {
2481                 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2482                         smcr_buf_unuse(buf_desc, is_rmb, lgr);
2483                         return -ENOMEM;
2484                 }
2485         }
2486
2487         if (is_rmb) {
2488                 conn->rmb_desc = buf_desc;
2489                 conn->rmbe_size_comp = bufsize_comp;
2490                 smc->sk.sk_rcvbuf = bufsize * 2;
2491                 atomic_set(&conn->bytes_to_rcv, 0);
2492                 conn->rmbe_update_limit =
2493                         smc_rmb_wnd_update_limit(buf_desc->len);
2494                 if (is_smcd)
2495                         smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2496         } else {
2497                 conn->sndbuf_desc = buf_desc;
2498                 smc->sk.sk_sndbuf = bufsize * 2;
2499                 atomic_set(&conn->sndbuf_space, bufsize);
2500         }
2501         return 0;
2502 }
2503
2504 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2505 {
2506         if (!conn->sndbuf_desc->is_dma_need_sync)
2507                 return;
2508         if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd ||
2509             !smc_link_active(conn->lnk))
2510                 return;
2511         smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2512 }
2513
2514 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2515 {
2516         int i;
2517
2518         if (!conn->rmb_desc->is_dma_need_sync)
2519                 return;
2520         if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd)
2521                 return;
2522         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2523                 if (!smc_link_active(&conn->lgr->lnk[i]))
2524                         continue;
2525                 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2526                                        DMA_FROM_DEVICE);
2527         }
2528 }
2529
2530 /* create the send and receive buffer for an SMC socket;
2531  * receive buffers are called RMBs;
2532  * (even though the SMC protocol allows more than one RMB-element per RMB,
2533  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2534  * extra RMB for every connection in a link group
2535  */
2536 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2537 {
2538         int rc;
2539
2540         /* create send buffer */
2541         if (is_smcd &&
2542             smc_ism_support_dmb_nocopy(smc->conn.lgr->smcd))
2543                 goto create_rmb;
2544
2545         rc = __smc_buf_create(smc, is_smcd, false);
2546         if (rc)
2547                 return rc;
2548
2549 create_rmb:
2550         /* create rmb */
2551         rc = __smc_buf_create(smc, is_smcd, true);
2552         if (rc && smc->conn.sndbuf_desc) {
2553                 down_write(&smc->conn.lgr->sndbufs_lock);
2554                 smc_lgr_buf_list_del(smc->conn.lgr, false,
2555                                      smc->conn.sndbuf_desc);
2556                 up_write(&smc->conn.lgr->sndbufs_lock);
2557                 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2558                 smc->conn.sndbuf_desc = NULL;
2559         }
2560         return rc;
2561 }
2562
2563 int smcd_buf_attach(struct smc_sock *smc)
2564 {
2565         struct smc_connection *conn = &smc->conn;
2566         struct smcd_dev *smcd = conn->lgr->smcd;
2567         u64 peer_token = conn->peer_token;
2568         struct smc_buf_desc *buf_desc;
2569         int rc;
2570
2571         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2572         if (!buf_desc)
2573                 return -ENOMEM;
2574
2575         /* The ghost sndbuf_desc describes the same memory region as
2576          * peer RMB. Its lifecycle is consistent with the connection's
2577          * and it will be freed with the connections instead of the
2578          * link group.
2579          */
2580         rc = smc_ism_attach_dmb(smcd, peer_token, buf_desc);
2581         if (rc)
2582                 goto free;
2583
2584         smc->sk.sk_sndbuf = buf_desc->len;
2585         buf_desc->cpu_addr =
2586                 (u8 *)buf_desc->cpu_addr + sizeof(struct smcd_cdc_msg);
2587         buf_desc->len -= sizeof(struct smcd_cdc_msg);
2588         conn->sndbuf_desc = buf_desc;
2589         conn->sndbuf_desc->used = 1;
2590         atomic_set(&conn->sndbuf_space, conn->sndbuf_desc->len);
2591         return 0;
2592
2593 free:
2594         kfree(buf_desc);
2595         return rc;
2596 }
2597
2598 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2599 {
2600         int i;
2601
2602         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2603                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2604                         return i;
2605         }
2606         return -ENOSPC;
2607 }
2608
2609 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2610                                    u32 rkey)
2611 {
2612         int i;
2613
2614         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2615                 if (test_bit(i, lgr->rtokens_used_mask) &&
2616                     lgr->rtokens[i][lnk_idx].rkey == rkey)
2617                         return i;
2618         }
2619         return -ENOENT;
2620 }
2621
2622 /* set rtoken for a new link to an existing rmb */
2623 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2624                     __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2625 {
2626         int rtok_idx;
2627
2628         rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2629         if (rtok_idx == -ENOENT)
2630                 return;
2631         lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2632         lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2633 }
2634
2635 /* set rtoken for a new link whose link_id is given */
2636 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2637                      __be64 nw_vaddr, __be32 nw_rkey)
2638 {
2639         u64 dma_addr = be64_to_cpu(nw_vaddr);
2640         u32 rkey = ntohl(nw_rkey);
2641         bool found = false;
2642         int link_idx;
2643
2644         for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2645                 if (lgr->lnk[link_idx].link_id == link_id) {
2646                         found = true;
2647                         break;
2648                 }
2649         }
2650         if (!found)
2651                 return;
2652         lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2653         lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2654 }
2655
2656 /* add a new rtoken from peer */
2657 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2658 {
2659         struct smc_link_group *lgr = smc_get_lgr(lnk);
2660         u64 dma_addr = be64_to_cpu(nw_vaddr);
2661         u32 rkey = ntohl(nw_rkey);
2662         int i;
2663
2664         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2665                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2666                     lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2667                     test_bit(i, lgr->rtokens_used_mask)) {
2668                         /* already in list */
2669                         return i;
2670                 }
2671         }
2672         i = smc_rmb_reserve_rtoken_idx(lgr);
2673         if (i < 0)
2674                 return i;
2675         lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2676         lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2677         return i;
2678 }
2679
2680 /* delete an rtoken from all links */
2681 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2682 {
2683         struct smc_link_group *lgr = smc_get_lgr(lnk);
2684         u32 rkey = ntohl(nw_rkey);
2685         int i, j;
2686
2687         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2688                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2689                     test_bit(i, lgr->rtokens_used_mask)) {
2690                         for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2691                                 lgr->rtokens[i][j].rkey = 0;
2692                                 lgr->rtokens[i][j].dma_addr = 0;
2693                         }
2694                         clear_bit(i, lgr->rtokens_used_mask);
2695                         return 0;
2696                 }
2697         }
2698         return -ENOENT;
2699 }
2700
2701 /* save rkey and dma_addr received from peer during clc handshake */
2702 int smc_rmb_rtoken_handling(struct smc_connection *conn,
2703                             struct smc_link *lnk,
2704                             struct smc_clc_msg_accept_confirm *clc)
2705 {
2706         conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2707                                           clc->r0.rmb_rkey);
2708         if (conn->rtoken_idx < 0)
2709                 return conn->rtoken_idx;
2710         return 0;
2711 }
2712
2713 static void smc_core_going_away(void)
2714 {
2715         struct smc_ib_device *smcibdev;
2716         struct smcd_dev *smcd;
2717
2718         mutex_lock(&smc_ib_devices.mutex);
2719         list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2720                 int i;
2721
2722                 for (i = 0; i < SMC_MAX_PORTS; i++)
2723                         set_bit(i, smcibdev->ports_going_away);
2724         }
2725         mutex_unlock(&smc_ib_devices.mutex);
2726
2727         mutex_lock(&smcd_dev_list.mutex);
2728         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2729                 smcd->going_away = 1;
2730         }
2731         mutex_unlock(&smcd_dev_list.mutex);
2732 }
2733
2734 /* Clean up all SMC link groups */
2735 static void smc_lgrs_shutdown(void)
2736 {
2737         struct smcd_dev *smcd;
2738
2739         smc_core_going_away();
2740
2741         smc_smcr_terminate_all(NULL);
2742
2743         mutex_lock(&smcd_dev_list.mutex);
2744         list_for_each_entry(smcd, &smcd_dev_list.list, list)
2745                 smc_smcd_terminate_all(smcd);
2746         mutex_unlock(&smcd_dev_list.mutex);
2747 }
2748
2749 static int smc_core_reboot_event(struct notifier_block *this,
2750                                  unsigned long event, void *ptr)
2751 {
2752         smc_lgrs_shutdown();
2753         smc_ib_unregister_client();
2754         smc_ism_exit();
2755         return 0;
2756 }
2757
2758 static struct notifier_block smc_reboot_notifier = {
2759         .notifier_call = smc_core_reboot_event,
2760 };
2761
2762 int __init smc_core_init(void)
2763 {
2764         return register_reboot_notifier(&smc_reboot_notifier);
2765 }
2766
2767 /* Called (from smc_exit) when module is removed */
2768 void smc_core_exit(void)
2769 {
2770         unregister_reboot_notifier(&smc_reboot_notifier);
2771         smc_lgrs_shutdown();
2772 }