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