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 38 #define SMC_LGR_NUM_INCR 256 39 #define SMC_LGR_FREE_DELAY_SERV (600 * HZ) 40 #define SMC_LGR_FREE_DELAY_CLNT (SMC_LGR_FREE_DELAY_SERV + 10 * HZ) 41 42 struct smc_lgr_list smc_lgr_list = { /* established link groups */ 43 .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock), 44 .list = LIST_HEAD_INIT(smc_lgr_list.list), 45 .num = 0, 46 }; 47 48 static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */ 49 static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted); 50 51 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb, 52 struct smc_buf_desc *buf_desc); 53 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft); 54 55 static void smc_link_down_work(struct work_struct *work); 56 57 /* return head of link group list and its lock for a given link group */ 58 static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr, 59 spinlock_t **lgr_lock) 60 { 61 if (lgr->is_smcd) { 62 *lgr_lock = &lgr->smcd->lgr_lock; 63 return &lgr->smcd->lgr_list; 64 } 65 66 *lgr_lock = &smc_lgr_list.lock; 67 return &smc_lgr_list.list; 68 } 69 70 static void smc_ibdev_cnt_inc(struct smc_link *lnk) 71 { 72 atomic_inc(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]); 73 } 74 75 static void smc_ibdev_cnt_dec(struct smc_link *lnk) 76 { 77 atomic_dec(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]); 78 } 79 80 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr) 81 { 82 /* client link group creation always follows the server link group 83 * creation. For client use a somewhat higher removal delay time, 84 * otherwise there is a risk of out-of-sync link groups. 85 */ 86 if (!lgr->freeing) { 87 mod_delayed_work(system_wq, &lgr->free_work, 88 (!lgr->is_smcd && lgr->role == SMC_CLNT) ? 89 SMC_LGR_FREE_DELAY_CLNT : 90 SMC_LGR_FREE_DELAY_SERV); 91 } 92 } 93 94 /* Register connection's alert token in our lookup structure. 95 * To use rbtrees we have to implement our own insert core. 96 * Requires @conns_lock 97 * @smc connection to register 98 * Returns 0 on success, != otherwise. 99 */ 100 static void smc_lgr_add_alert_token(struct smc_connection *conn) 101 { 102 struct rb_node **link, *parent = NULL; 103 u32 token = conn->alert_token_local; 104 105 link = &conn->lgr->conns_all.rb_node; 106 while (*link) { 107 struct smc_connection *cur = rb_entry(*link, 108 struct smc_connection, alert_node); 109 110 parent = *link; 111 if (cur->alert_token_local > token) 112 link = &parent->rb_left; 113 else 114 link = &parent->rb_right; 115 } 116 /* Put the new node there */ 117 rb_link_node(&conn->alert_node, parent, link); 118 rb_insert_color(&conn->alert_node, &conn->lgr->conns_all); 119 } 120 121 /* assign an SMC-R link to the connection */ 122 static int smcr_lgr_conn_assign_link(struct smc_connection *conn, bool first) 123 { 124 enum smc_link_state expected = first ? SMC_LNK_ACTIVATING : 125 SMC_LNK_ACTIVE; 126 int i, j; 127 128 /* do link balancing */ 129 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 130 struct smc_link *lnk = &conn->lgr->lnk[i]; 131 132 if (lnk->state != expected || lnk->link_is_asym) 133 continue; 134 if (conn->lgr->role == SMC_CLNT) { 135 conn->lnk = lnk; /* temporary, SMC server assigns link*/ 136 break; 137 } 138 if (conn->lgr->conns_num % 2) { 139 for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) { 140 struct smc_link *lnk2; 141 142 lnk2 = &conn->lgr->lnk[j]; 143 if (lnk2->state == expected && 144 !lnk2->link_is_asym) { 145 conn->lnk = lnk2; 146 break; 147 } 148 } 149 } 150 if (!conn->lnk) 151 conn->lnk = lnk; 152 break; 153 } 154 if (!conn->lnk) 155 return SMC_CLC_DECL_NOACTLINK; 156 atomic_inc(&conn->lnk->conn_cnt); 157 return 0; 158 } 159 160 /* Register connection in link group by assigning an alert token 161 * registered in a search tree. 162 * Requires @conns_lock 163 * Note that '0' is a reserved value and not assigned. 164 */ 165 static int smc_lgr_register_conn(struct smc_connection *conn, bool first) 166 { 167 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 168 static atomic_t nexttoken = ATOMIC_INIT(0); 169 int rc; 170 171 if (!conn->lgr->is_smcd) { 172 rc = smcr_lgr_conn_assign_link(conn, first); 173 if (rc) 174 return rc; 175 } 176 /* find a new alert_token_local value not yet used by some connection 177 * in this link group 178 */ 179 sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */ 180 while (!conn->alert_token_local) { 181 conn->alert_token_local = atomic_inc_return(&nexttoken); 182 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr)) 183 conn->alert_token_local = 0; 184 } 185 smc_lgr_add_alert_token(conn); 186 conn->lgr->conns_num++; 187 return 0; 188 } 189 190 /* Unregister connection and reset the alert token of the given connection< 191 */ 192 static void __smc_lgr_unregister_conn(struct smc_connection *conn) 193 { 194 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 195 struct smc_link_group *lgr = conn->lgr; 196 197 rb_erase(&conn->alert_node, &lgr->conns_all); 198 if (conn->lnk) 199 atomic_dec(&conn->lnk->conn_cnt); 200 lgr->conns_num--; 201 conn->alert_token_local = 0; 202 sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */ 203 } 204 205 /* Unregister connection from lgr 206 */ 207 static void smc_lgr_unregister_conn(struct smc_connection *conn) 208 { 209 struct smc_link_group *lgr = conn->lgr; 210 211 if (!lgr) 212 return; 213 write_lock_bh(&lgr->conns_lock); 214 if (conn->alert_token_local) { 215 __smc_lgr_unregister_conn(conn); 216 } 217 write_unlock_bh(&lgr->conns_lock); 218 conn->lgr = NULL; 219 } 220 221 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb) 222 { 223 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb); 224 char hostname[SMC_MAX_HOSTNAME_LEN + 1]; 225 char smc_seid[SMC_MAX_EID_LEN + 1]; 226 struct smcd_dev *smcd_dev; 227 struct nlattr *attrs; 228 u8 *seid = NULL; 229 u8 *host = NULL; 230 void *nlh; 231 232 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, 233 &smc_gen_nl_family, NLM_F_MULTI, 234 SMC_NETLINK_GET_SYS_INFO); 235 if (!nlh) 236 goto errmsg; 237 if (cb_ctx->pos[0]) 238 goto errout; 239 attrs = nla_nest_start(skb, SMC_GEN_SYS_INFO); 240 if (!attrs) 241 goto errout; 242 if (nla_put_u8(skb, SMC_NLA_SYS_VER, SMC_V2)) 243 goto errattr; 244 if (nla_put_u8(skb, SMC_NLA_SYS_REL, SMC_RELEASE)) 245 goto errattr; 246 if (nla_put_u8(skb, SMC_NLA_SYS_IS_ISM_V2, smc_ism_is_v2_capable())) 247 goto errattr; 248 smc_clc_get_hostname(&host); 249 if (host) { 250 memcpy(hostname, host, SMC_MAX_HOSTNAME_LEN); 251 hostname[SMC_MAX_HOSTNAME_LEN] = 0; 252 if (nla_put_string(skb, SMC_NLA_SYS_LOCAL_HOST, hostname)) 253 goto errattr; 254 } 255 mutex_lock(&smcd_dev_list.mutex); 256 smcd_dev = list_first_entry_or_null(&smcd_dev_list.list, 257 struct smcd_dev, list); 258 if (smcd_dev) 259 smc_ism_get_system_eid(smcd_dev, &seid); 260 mutex_unlock(&smcd_dev_list.mutex); 261 if (seid && smc_ism_is_v2_capable()) { 262 memcpy(smc_seid, seid, SMC_MAX_EID_LEN); 263 smc_seid[SMC_MAX_EID_LEN] = 0; 264 if (nla_put_string(skb, SMC_NLA_SYS_SEID, smc_seid)) 265 goto errattr; 266 } 267 nla_nest_end(skb, attrs); 268 genlmsg_end(skb, nlh); 269 cb_ctx->pos[0] = 1; 270 return skb->len; 271 272 errattr: 273 nla_nest_cancel(skb, attrs); 274 errout: 275 genlmsg_cancel(skb, nlh); 276 errmsg: 277 return skb->len; 278 } 279 280 static int smc_nl_fill_lgr(struct smc_link_group *lgr, 281 struct sk_buff *skb, 282 struct netlink_callback *cb) 283 { 284 char smc_target[SMC_MAX_PNETID_LEN + 1]; 285 struct nlattr *attrs; 286 287 attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCR); 288 if (!attrs) 289 goto errout; 290 291 if (nla_put_u32(skb, SMC_NLA_LGR_R_ID, *((u32 *)&lgr->id))) 292 goto errattr; 293 if (nla_put_u32(skb, SMC_NLA_LGR_R_CONNS_NUM, lgr->conns_num)) 294 goto errattr; 295 if (nla_put_u8(skb, SMC_NLA_LGR_R_ROLE, lgr->role)) 296 goto errattr; 297 if (nla_put_u8(skb, SMC_NLA_LGR_R_TYPE, lgr->type)) 298 goto errattr; 299 if (nla_put_u8(skb, SMC_NLA_LGR_R_VLAN_ID, lgr->vlan_id)) 300 goto errattr; 301 memcpy(smc_target, lgr->pnet_id, SMC_MAX_PNETID_LEN); 302 smc_target[SMC_MAX_PNETID_LEN] = 0; 303 if (nla_put_string(skb, SMC_NLA_LGR_R_PNETID, smc_target)) 304 goto errattr; 305 306 nla_nest_end(skb, attrs); 307 return 0; 308 errattr: 309 nla_nest_cancel(skb, attrs); 310 errout: 311 return -EMSGSIZE; 312 } 313 314 static int smc_nl_fill_lgr_link(struct smc_link_group *lgr, 315 struct smc_link *link, 316 struct sk_buff *skb, 317 struct netlink_callback *cb) 318 { 319 char smc_ibname[IB_DEVICE_NAME_MAX]; 320 u8 smc_gid_target[41]; 321 struct nlattr *attrs; 322 u32 link_uid = 0; 323 void *nlh; 324 325 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, 326 &smc_gen_nl_family, NLM_F_MULTI, 327 SMC_NETLINK_GET_LINK_SMCR); 328 if (!nlh) 329 goto errmsg; 330 331 attrs = nla_nest_start(skb, SMC_GEN_LINK_SMCR); 332 if (!attrs) 333 goto errout; 334 335 if (nla_put_u8(skb, SMC_NLA_LINK_ID, link->link_id)) 336 goto errattr; 337 if (nla_put_u32(skb, SMC_NLA_LINK_STATE, link->state)) 338 goto errattr; 339 if (nla_put_u32(skb, SMC_NLA_LINK_CONN_CNT, 340 atomic_read(&link->conn_cnt))) 341 goto errattr; 342 if (nla_put_u8(skb, SMC_NLA_LINK_IB_PORT, link->ibport)) 343 goto errattr; 344 if (nla_put_u32(skb, SMC_NLA_LINK_NET_DEV, link->ndev_ifidx)) 345 goto errattr; 346 snprintf(smc_ibname, sizeof(smc_ibname), "%s", link->ibname); 347 if (nla_put_string(skb, SMC_NLA_LINK_IB_DEV, smc_ibname)) 348 goto errattr; 349 memcpy(&link_uid, link->link_uid, sizeof(link_uid)); 350 if (nla_put_u32(skb, SMC_NLA_LINK_UID, link_uid)) 351 goto errattr; 352 memcpy(&link_uid, link->peer_link_uid, sizeof(link_uid)); 353 if (nla_put_u32(skb, SMC_NLA_LINK_PEER_UID, link_uid)) 354 goto errattr; 355 memset(smc_gid_target, 0, sizeof(smc_gid_target)); 356 smc_gid_be16_convert(smc_gid_target, link->gid); 357 if (nla_put_string(skb, SMC_NLA_LINK_GID, smc_gid_target)) 358 goto errattr; 359 memset(smc_gid_target, 0, sizeof(smc_gid_target)); 360 smc_gid_be16_convert(smc_gid_target, link->peer_gid); 361 if (nla_put_string(skb, SMC_NLA_LINK_PEER_GID, smc_gid_target)) 362 goto errattr; 363 364 nla_nest_end(skb, attrs); 365 genlmsg_end(skb, nlh); 366 return 0; 367 errattr: 368 nla_nest_cancel(skb, attrs); 369 errout: 370 genlmsg_cancel(skb, nlh); 371 errmsg: 372 return -EMSGSIZE; 373 } 374 375 static int smc_nl_handle_lgr(struct smc_link_group *lgr, 376 struct sk_buff *skb, 377 struct netlink_callback *cb, 378 bool list_links) 379 { 380 void *nlh; 381 int i; 382 383 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, 384 &smc_gen_nl_family, NLM_F_MULTI, 385 SMC_NETLINK_GET_LGR_SMCR); 386 if (!nlh) 387 goto errmsg; 388 if (smc_nl_fill_lgr(lgr, skb, cb)) 389 goto errout; 390 391 genlmsg_end(skb, nlh); 392 if (!list_links) 393 goto out; 394 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 395 if (!smc_link_usable(&lgr->lnk[i])) 396 continue; 397 if (smc_nl_fill_lgr_link(lgr, &lgr->lnk[i], skb, cb)) 398 goto errout; 399 } 400 out: 401 return 0; 402 403 errout: 404 genlmsg_cancel(skb, nlh); 405 errmsg: 406 return -EMSGSIZE; 407 } 408 409 static void smc_nl_fill_lgr_list(struct smc_lgr_list *smc_lgr, 410 struct sk_buff *skb, 411 struct netlink_callback *cb, 412 bool list_links) 413 { 414 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb); 415 struct smc_link_group *lgr; 416 int snum = cb_ctx->pos[0]; 417 int num = 0; 418 419 spin_lock_bh(&smc_lgr->lock); 420 list_for_each_entry(lgr, &smc_lgr->list, list) { 421 if (num < snum) 422 goto next; 423 if (smc_nl_handle_lgr(lgr, skb, cb, list_links)) 424 goto errout; 425 next: 426 num++; 427 } 428 errout: 429 spin_unlock_bh(&smc_lgr->lock); 430 cb_ctx->pos[0] = num; 431 } 432 433 static int smc_nl_fill_smcd_lgr(struct smc_link_group *lgr, 434 struct sk_buff *skb, 435 struct netlink_callback *cb) 436 { 437 char smc_host[SMC_MAX_HOSTNAME_LEN + 1]; 438 char smc_pnet[SMC_MAX_PNETID_LEN + 1]; 439 char smc_eid[SMC_MAX_EID_LEN + 1]; 440 struct nlattr *v2_attrs; 441 struct nlattr *attrs; 442 void *nlh; 443 444 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, 445 &smc_gen_nl_family, NLM_F_MULTI, 446 SMC_NETLINK_GET_LGR_SMCD); 447 if (!nlh) 448 goto errmsg; 449 450 attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD); 451 if (!attrs) 452 goto errout; 453 454 if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id))) 455 goto errattr; 456 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID, lgr->smcd->local_gid, 457 SMC_NLA_LGR_D_PAD)) 458 goto errattr; 459 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid, 460 SMC_NLA_LGR_D_PAD)) 461 goto errattr; 462 if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id)) 463 goto errattr; 464 if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num)) 465 goto errattr; 466 if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd))) 467 goto errattr; 468 memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN); 469 smc_pnet[SMC_MAX_PNETID_LEN] = 0; 470 if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet)) 471 goto errattr; 472 473 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_V2); 474 if (!v2_attrs) 475 goto errattr; 476 if (nla_put_u8(skb, SMC_NLA_LGR_V2_VER, lgr->smc_version)) 477 goto errv2attr; 478 if (nla_put_u8(skb, SMC_NLA_LGR_V2_REL, lgr->peer_smc_release)) 479 goto errv2attr; 480 if (nla_put_u8(skb, SMC_NLA_LGR_V2_OS, lgr->peer_os)) 481 goto errv2attr; 482 memcpy(smc_host, lgr->peer_hostname, SMC_MAX_HOSTNAME_LEN); 483 smc_host[SMC_MAX_HOSTNAME_LEN] = 0; 484 if (nla_put_string(skb, SMC_NLA_LGR_V2_PEER_HOST, smc_host)) 485 goto errv2attr; 486 memcpy(smc_eid, lgr->negotiated_eid, SMC_MAX_EID_LEN); 487 smc_eid[SMC_MAX_EID_LEN] = 0; 488 if (nla_put_string(skb, SMC_NLA_LGR_V2_NEG_EID, smc_eid)) 489 goto errv2attr; 490 491 nla_nest_end(skb, v2_attrs); 492 nla_nest_end(skb, attrs); 493 genlmsg_end(skb, nlh); 494 return 0; 495 496 errv2attr: 497 nla_nest_cancel(skb, v2_attrs); 498 errattr: 499 nla_nest_cancel(skb, attrs); 500 errout: 501 genlmsg_cancel(skb, nlh); 502 errmsg: 503 return -EMSGSIZE; 504 } 505 506 static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev, 507 struct sk_buff *skb, 508 struct netlink_callback *cb) 509 { 510 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb); 511 struct smc_link_group *lgr; 512 int snum = cb_ctx->pos[1]; 513 int rc = 0, num = 0; 514 515 spin_lock_bh(&dev->lgr_lock); 516 list_for_each_entry(lgr, &dev->lgr_list, list) { 517 if (!lgr->is_smcd) 518 continue; 519 if (num < snum) 520 goto next; 521 rc = smc_nl_fill_smcd_lgr(lgr, skb, cb); 522 if (rc) 523 goto errout; 524 next: 525 num++; 526 } 527 errout: 528 spin_unlock_bh(&dev->lgr_lock); 529 cb_ctx->pos[1] = num; 530 return rc; 531 } 532 533 static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list, 534 struct sk_buff *skb, 535 struct netlink_callback *cb) 536 { 537 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb); 538 struct smcd_dev *smcd_dev; 539 int snum = cb_ctx->pos[0]; 540 int rc = 0, num = 0; 541 542 mutex_lock(&dev_list->mutex); 543 list_for_each_entry(smcd_dev, &dev_list->list, list) { 544 if (list_empty(&smcd_dev->lgr_list)) 545 continue; 546 if (num < snum) 547 goto next; 548 rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb); 549 if (rc) 550 goto errout; 551 next: 552 num++; 553 } 554 errout: 555 mutex_unlock(&dev_list->mutex); 556 cb_ctx->pos[0] = num; 557 return rc; 558 } 559 560 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb) 561 { 562 bool list_links = false; 563 564 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links); 565 return skb->len; 566 } 567 568 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb) 569 { 570 bool list_links = true; 571 572 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links); 573 return skb->len; 574 } 575 576 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb) 577 { 578 smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb); 579 return skb->len; 580 } 581 582 void smc_lgr_cleanup_early(struct smc_connection *conn) 583 { 584 struct smc_link_group *lgr = conn->lgr; 585 struct list_head *lgr_list; 586 spinlock_t *lgr_lock; 587 588 if (!lgr) 589 return; 590 591 smc_conn_free(conn); 592 lgr_list = smc_lgr_list_head(lgr, &lgr_lock); 593 spin_lock_bh(lgr_lock); 594 /* do not use this link group for new connections */ 595 if (!list_empty(lgr_list)) 596 list_del_init(lgr_list); 597 spin_unlock_bh(lgr_lock); 598 __smc_lgr_terminate(lgr, true); 599 } 600 601 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr) 602 { 603 int i; 604 605 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 606 struct smc_link *lnk = &lgr->lnk[i]; 607 608 if (smc_link_usable(lnk)) 609 lnk->state = SMC_LNK_INACTIVE; 610 } 611 wake_up_all(&lgr->llc_msg_waiter); 612 wake_up_all(&lgr->llc_flow_waiter); 613 } 614 615 static void smc_lgr_free(struct smc_link_group *lgr); 616 617 static void smc_lgr_free_work(struct work_struct *work) 618 { 619 struct smc_link_group *lgr = container_of(to_delayed_work(work), 620 struct smc_link_group, 621 free_work); 622 spinlock_t *lgr_lock; 623 bool conns; 624 625 smc_lgr_list_head(lgr, &lgr_lock); 626 spin_lock_bh(lgr_lock); 627 if (lgr->freeing) { 628 spin_unlock_bh(lgr_lock); 629 return; 630 } 631 read_lock_bh(&lgr->conns_lock); 632 conns = RB_EMPTY_ROOT(&lgr->conns_all); 633 read_unlock_bh(&lgr->conns_lock); 634 if (!conns) { /* number of lgr connections is no longer zero */ 635 spin_unlock_bh(lgr_lock); 636 return; 637 } 638 list_del_init(&lgr->list); /* remove from smc_lgr_list */ 639 lgr->freeing = 1; /* this instance does the freeing, no new schedule */ 640 spin_unlock_bh(lgr_lock); 641 cancel_delayed_work(&lgr->free_work); 642 643 if (!lgr->is_smcd && !lgr->terminating) 644 smc_llc_send_link_delete_all(lgr, true, 645 SMC_LLC_DEL_PROG_INIT_TERM); 646 if (lgr->is_smcd && !lgr->terminating) 647 smc_ism_signal_shutdown(lgr); 648 if (!lgr->is_smcd) 649 smcr_lgr_link_deactivate_all(lgr); 650 smc_lgr_free(lgr); 651 } 652 653 static void smc_lgr_terminate_work(struct work_struct *work) 654 { 655 struct smc_link_group *lgr = container_of(work, struct smc_link_group, 656 terminate_work); 657 658 __smc_lgr_terminate(lgr, true); 659 } 660 661 /* return next unique link id for the lgr */ 662 static u8 smcr_next_link_id(struct smc_link_group *lgr) 663 { 664 u8 link_id; 665 int i; 666 667 while (1) { 668 link_id = ++lgr->next_link_id; 669 if (!link_id) /* skip zero as link_id */ 670 link_id = ++lgr->next_link_id; 671 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 672 if (smc_link_usable(&lgr->lnk[i]) && 673 lgr->lnk[i].link_id == link_id) 674 continue; 675 } 676 break; 677 } 678 return link_id; 679 } 680 681 static void smcr_copy_dev_info_to_link(struct smc_link *link) 682 { 683 struct smc_ib_device *smcibdev = link->smcibdev; 684 685 snprintf(link->ibname, sizeof(link->ibname), "%s", 686 smcibdev->ibdev->name); 687 link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1]; 688 } 689 690 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk, 691 u8 link_idx, struct smc_init_info *ini) 692 { 693 u8 rndvec[3]; 694 int rc; 695 696 get_device(&ini->ib_dev->ibdev->dev); 697 atomic_inc(&ini->ib_dev->lnk_cnt); 698 lnk->link_id = smcr_next_link_id(lgr); 699 lnk->lgr = lgr; 700 lnk->link_idx = link_idx; 701 lnk->smcibdev = ini->ib_dev; 702 lnk->ibport = ini->ib_port; 703 smc_ibdev_cnt_inc(lnk); 704 smcr_copy_dev_info_to_link(lnk); 705 lnk->path_mtu = ini->ib_dev->pattr[ini->ib_port - 1].active_mtu; 706 atomic_set(&lnk->conn_cnt, 0); 707 smc_llc_link_set_uid(lnk); 708 INIT_WORK(&lnk->link_down_wrk, smc_link_down_work); 709 if (!ini->ib_dev->initialized) { 710 rc = (int)smc_ib_setup_per_ibdev(ini->ib_dev); 711 if (rc) 712 goto out; 713 } 714 get_random_bytes(rndvec, sizeof(rndvec)); 715 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) + 716 (rndvec[2] << 16); 717 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport, 718 ini->vlan_id, lnk->gid, &lnk->sgid_index); 719 if (rc) 720 goto out; 721 rc = smc_llc_link_init(lnk); 722 if (rc) 723 goto out; 724 rc = smc_wr_alloc_link_mem(lnk); 725 if (rc) 726 goto clear_llc_lnk; 727 rc = smc_ib_create_protection_domain(lnk); 728 if (rc) 729 goto free_link_mem; 730 rc = smc_ib_create_queue_pair(lnk); 731 if (rc) 732 goto dealloc_pd; 733 rc = smc_wr_create_link(lnk); 734 if (rc) 735 goto destroy_qp; 736 lnk->state = SMC_LNK_ACTIVATING; 737 return 0; 738 739 destroy_qp: 740 smc_ib_destroy_queue_pair(lnk); 741 dealloc_pd: 742 smc_ib_dealloc_protection_domain(lnk); 743 free_link_mem: 744 smc_wr_free_link_mem(lnk); 745 clear_llc_lnk: 746 smc_llc_link_clear(lnk, false); 747 out: 748 smc_ibdev_cnt_dec(lnk); 749 put_device(&ini->ib_dev->ibdev->dev); 750 memset(lnk, 0, sizeof(struct smc_link)); 751 lnk->state = SMC_LNK_UNUSED; 752 if (!atomic_dec_return(&ini->ib_dev->lnk_cnt)) 753 wake_up(&ini->ib_dev->lnks_deleted); 754 return rc; 755 } 756 757 /* create a new SMC link group */ 758 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini) 759 { 760 struct smc_link_group *lgr; 761 struct list_head *lgr_list; 762 struct smc_link *lnk; 763 spinlock_t *lgr_lock; 764 u8 link_idx; 765 int rc = 0; 766 int i; 767 768 if (ini->is_smcd && ini->vlan_id) { 769 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected], 770 ini->vlan_id)) { 771 rc = SMC_CLC_DECL_ISMVLANERR; 772 goto out; 773 } 774 } 775 776 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL); 777 if (!lgr) { 778 rc = SMC_CLC_DECL_MEM; 779 goto ism_put_vlan; 780 } 781 lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0, 782 SMC_LGR_ID_SIZE, &lgr->id); 783 if (!lgr->tx_wq) { 784 rc = -ENOMEM; 785 goto free_lgr; 786 } 787 lgr->is_smcd = ini->is_smcd; 788 lgr->sync_err = 0; 789 lgr->terminating = 0; 790 lgr->freeing = 0; 791 lgr->vlan_id = ini->vlan_id; 792 mutex_init(&lgr->sndbufs_lock); 793 mutex_init(&lgr->rmbs_lock); 794 rwlock_init(&lgr->conns_lock); 795 for (i = 0; i < SMC_RMBE_SIZES; i++) { 796 INIT_LIST_HEAD(&lgr->sndbufs[i]); 797 INIT_LIST_HEAD(&lgr->rmbs[i]); 798 } 799 lgr->next_link_id = 0; 800 smc_lgr_list.num += SMC_LGR_NUM_INCR; 801 memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE); 802 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work); 803 INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work); 804 lgr->conns_all = RB_ROOT; 805 if (ini->is_smcd) { 806 /* SMC-D specific settings */ 807 get_device(&ini->ism_dev[ini->ism_selected]->dev); 808 lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected]; 809 lgr->smcd = ini->ism_dev[ini->ism_selected]; 810 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list; 811 lgr_lock = &lgr->smcd->lgr_lock; 812 lgr->smc_version = ini->smcd_version; 813 lgr->peer_shutdown = 0; 814 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt); 815 } else { 816 /* SMC-R specific settings */ 817 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT; 818 memcpy(lgr->peer_systemid, ini->ib_lcl->id_for_peer, 819 SMC_SYSTEMID_LEN); 820 memcpy(lgr->pnet_id, ini->ib_dev->pnetid[ini->ib_port - 1], 821 SMC_MAX_PNETID_LEN); 822 smc_llc_lgr_init(lgr, smc); 823 824 link_idx = SMC_SINGLE_LINK; 825 lnk = &lgr->lnk[link_idx]; 826 rc = smcr_link_init(lgr, lnk, link_idx, ini); 827 if (rc) 828 goto free_wq; 829 lgr_list = &smc_lgr_list.list; 830 lgr_lock = &smc_lgr_list.lock; 831 atomic_inc(&lgr_cnt); 832 } 833 smc->conn.lgr = lgr; 834 spin_lock_bh(lgr_lock); 835 list_add_tail(&lgr->list, lgr_list); 836 spin_unlock_bh(lgr_lock); 837 return 0; 838 839 free_wq: 840 destroy_workqueue(lgr->tx_wq); 841 free_lgr: 842 kfree(lgr); 843 ism_put_vlan: 844 if (ini->is_smcd && ini->vlan_id) 845 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id); 846 out: 847 if (rc < 0) { 848 if (rc == -ENOMEM) 849 rc = SMC_CLC_DECL_MEM; 850 else 851 rc = SMC_CLC_DECL_INTERR; 852 } 853 return rc; 854 } 855 856 static int smc_write_space(struct smc_connection *conn) 857 { 858 int buffer_len = conn->peer_rmbe_size; 859 union smc_host_cursor prod; 860 union smc_host_cursor cons; 861 int space; 862 863 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn); 864 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn); 865 /* determine rx_buf space */ 866 space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod); 867 return space; 868 } 869 870 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend, 871 struct smc_wr_buf *wr_buf) 872 { 873 struct smc_connection *conn = &smc->conn; 874 union smc_host_cursor cons, fin; 875 int rc = 0; 876 int diff; 877 878 smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn); 879 smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn); 880 /* set prod cursor to old state, enforce tx_rdma_writes() */ 881 smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn); 882 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn); 883 884 if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) { 885 /* cons cursor advanced more than fin, and prod was set 886 * fin above, so now prod is smaller than cons. Fix that. 887 */ 888 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons); 889 smc_curs_add(conn->sndbuf_desc->len, 890 &conn->tx_curs_sent, diff); 891 smc_curs_add(conn->sndbuf_desc->len, 892 &conn->tx_curs_fin, diff); 893 894 smp_mb__before_atomic(); 895 atomic_add(diff, &conn->sndbuf_space); 896 smp_mb__after_atomic(); 897 898 smc_curs_add(conn->peer_rmbe_size, 899 &conn->local_tx_ctrl.prod, diff); 900 smc_curs_add(conn->peer_rmbe_size, 901 &conn->local_tx_ctrl_fin, diff); 902 } 903 /* recalculate, value is used by tx_rdma_writes() */ 904 atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn)); 905 906 if (smc->sk.sk_state != SMC_INIT && 907 smc->sk.sk_state != SMC_CLOSED) { 908 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf); 909 if (!rc) { 910 queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0); 911 smc->sk.sk_data_ready(&smc->sk); 912 } 913 } else { 914 smc_wr_tx_put_slot(conn->lnk, 915 (struct smc_wr_tx_pend_priv *)pend); 916 } 917 return rc; 918 } 919 920 void smc_switch_link_and_count(struct smc_connection *conn, 921 struct smc_link *to_lnk) 922 { 923 atomic_dec(&conn->lnk->conn_cnt); 924 conn->lnk = to_lnk; 925 atomic_inc(&conn->lnk->conn_cnt); 926 } 927 928 struct smc_link *smc_switch_conns(struct smc_link_group *lgr, 929 struct smc_link *from_lnk, bool is_dev_err) 930 { 931 struct smc_link *to_lnk = NULL; 932 struct smc_cdc_tx_pend *pend; 933 struct smc_connection *conn; 934 struct smc_wr_buf *wr_buf; 935 struct smc_sock *smc; 936 struct rb_node *node; 937 int i, rc = 0; 938 939 /* link is inactive, wake up tx waiters */ 940 smc_wr_wakeup_tx_wait(from_lnk); 941 942 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 943 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx) 944 continue; 945 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev && 946 from_lnk->ibport == lgr->lnk[i].ibport) { 947 continue; 948 } 949 to_lnk = &lgr->lnk[i]; 950 break; 951 } 952 if (!to_lnk) { 953 smc_lgr_terminate_sched(lgr); 954 return NULL; 955 } 956 again: 957 read_lock_bh(&lgr->conns_lock); 958 for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) { 959 conn = rb_entry(node, struct smc_connection, alert_node); 960 if (conn->lnk != from_lnk) 961 continue; 962 smc = container_of(conn, struct smc_sock, conn); 963 /* conn->lnk not yet set in SMC_INIT state */ 964 if (smc->sk.sk_state == SMC_INIT) 965 continue; 966 if (smc->sk.sk_state == SMC_CLOSED || 967 smc->sk.sk_state == SMC_PEERCLOSEWAIT1 || 968 smc->sk.sk_state == SMC_PEERCLOSEWAIT2 || 969 smc->sk.sk_state == SMC_APPFINCLOSEWAIT || 970 smc->sk.sk_state == SMC_APPCLOSEWAIT1 || 971 smc->sk.sk_state == SMC_APPCLOSEWAIT2 || 972 smc->sk.sk_state == SMC_PEERFINCLOSEWAIT || 973 smc->sk.sk_state == SMC_PEERABORTWAIT || 974 smc->sk.sk_state == SMC_PROCESSABORT) { 975 spin_lock_bh(&conn->send_lock); 976 smc_switch_link_and_count(conn, to_lnk); 977 spin_unlock_bh(&conn->send_lock); 978 continue; 979 } 980 sock_hold(&smc->sk); 981 read_unlock_bh(&lgr->conns_lock); 982 /* pre-fetch buffer outside of send_lock, might sleep */ 983 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend); 984 if (rc) { 985 smcr_link_down_cond_sched(to_lnk); 986 return NULL; 987 } 988 /* avoid race with smcr_tx_sndbuf_nonempty() */ 989 spin_lock_bh(&conn->send_lock); 990 smc_switch_link_and_count(conn, to_lnk); 991 rc = smc_switch_cursor(smc, pend, wr_buf); 992 spin_unlock_bh(&conn->send_lock); 993 sock_put(&smc->sk); 994 if (rc) { 995 smcr_link_down_cond_sched(to_lnk); 996 return NULL; 997 } 998 goto again; 999 } 1000 read_unlock_bh(&lgr->conns_lock); 1001 return to_lnk; 1002 } 1003 1004 static void smcr_buf_unuse(struct smc_buf_desc *rmb_desc, 1005 struct smc_link_group *lgr) 1006 { 1007 int rc; 1008 1009 if (rmb_desc->is_conf_rkey && !list_empty(&lgr->list)) { 1010 /* unregister rmb with peer */ 1011 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY); 1012 if (!rc) { 1013 /* protect against smc_llc_cli_rkey_exchange() */ 1014 mutex_lock(&lgr->llc_conf_mutex); 1015 smc_llc_do_delete_rkey(lgr, rmb_desc); 1016 rmb_desc->is_conf_rkey = false; 1017 mutex_unlock(&lgr->llc_conf_mutex); 1018 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl); 1019 } 1020 } 1021 1022 if (rmb_desc->is_reg_err) { 1023 /* buf registration failed, reuse not possible */ 1024 mutex_lock(&lgr->rmbs_lock); 1025 list_del(&rmb_desc->list); 1026 mutex_unlock(&lgr->rmbs_lock); 1027 1028 smc_buf_free(lgr, true, rmb_desc); 1029 } else { 1030 rmb_desc->used = 0; 1031 } 1032 } 1033 1034 static void smc_buf_unuse(struct smc_connection *conn, 1035 struct smc_link_group *lgr) 1036 { 1037 if (conn->sndbuf_desc) 1038 conn->sndbuf_desc->used = 0; 1039 if (conn->rmb_desc && lgr->is_smcd) 1040 conn->rmb_desc->used = 0; 1041 else if (conn->rmb_desc) 1042 smcr_buf_unuse(conn->rmb_desc, lgr); 1043 } 1044 1045 /* remove a finished connection from its link group */ 1046 void smc_conn_free(struct smc_connection *conn) 1047 { 1048 struct smc_link_group *lgr = conn->lgr; 1049 1050 if (!lgr) 1051 return; 1052 if (lgr->is_smcd) { 1053 if (!list_empty(&lgr->list)) 1054 smc_ism_unset_conn(conn); 1055 tasklet_kill(&conn->rx_tsklet); 1056 } else { 1057 smc_cdc_tx_dismiss_slots(conn); 1058 if (current_work() != &conn->abort_work) 1059 cancel_work_sync(&conn->abort_work); 1060 } 1061 if (!list_empty(&lgr->list)) { 1062 smc_lgr_unregister_conn(conn); 1063 smc_buf_unuse(conn, lgr); /* allow buffer reuse */ 1064 } 1065 1066 if (!lgr->conns_num) 1067 smc_lgr_schedule_free_work(lgr); 1068 } 1069 1070 /* unregister a link from a buf_desc */ 1071 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb, 1072 struct smc_link *lnk) 1073 { 1074 if (is_rmb) 1075 buf_desc->is_reg_mr[lnk->link_idx] = false; 1076 if (!buf_desc->is_map_ib[lnk->link_idx]) 1077 return; 1078 if (is_rmb) { 1079 if (buf_desc->mr_rx[lnk->link_idx]) { 1080 smc_ib_put_memory_region( 1081 buf_desc->mr_rx[lnk->link_idx]); 1082 buf_desc->mr_rx[lnk->link_idx] = NULL; 1083 } 1084 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE); 1085 } else { 1086 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE); 1087 } 1088 sg_free_table(&buf_desc->sgt[lnk->link_idx]); 1089 buf_desc->is_map_ib[lnk->link_idx] = false; 1090 } 1091 1092 /* unmap all buffers of lgr for a deleted link */ 1093 static void smcr_buf_unmap_lgr(struct smc_link *lnk) 1094 { 1095 struct smc_link_group *lgr = lnk->lgr; 1096 struct smc_buf_desc *buf_desc, *bf; 1097 int i; 1098 1099 for (i = 0; i < SMC_RMBE_SIZES; i++) { 1100 mutex_lock(&lgr->rmbs_lock); 1101 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) 1102 smcr_buf_unmap_link(buf_desc, true, lnk); 1103 mutex_unlock(&lgr->rmbs_lock); 1104 mutex_lock(&lgr->sndbufs_lock); 1105 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], 1106 list) 1107 smcr_buf_unmap_link(buf_desc, false, lnk); 1108 mutex_unlock(&lgr->sndbufs_lock); 1109 } 1110 } 1111 1112 static void smcr_rtoken_clear_link(struct smc_link *lnk) 1113 { 1114 struct smc_link_group *lgr = lnk->lgr; 1115 int i; 1116 1117 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) { 1118 lgr->rtokens[i][lnk->link_idx].rkey = 0; 1119 lgr->rtokens[i][lnk->link_idx].dma_addr = 0; 1120 } 1121 } 1122 1123 /* must be called under lgr->llc_conf_mutex lock */ 1124 void smcr_link_clear(struct smc_link *lnk, bool log) 1125 { 1126 struct smc_ib_device *smcibdev; 1127 1128 if (!lnk->lgr || lnk->state == SMC_LNK_UNUSED) 1129 return; 1130 lnk->peer_qpn = 0; 1131 smc_llc_link_clear(lnk, log); 1132 smcr_buf_unmap_lgr(lnk); 1133 smcr_rtoken_clear_link(lnk); 1134 smc_ib_modify_qp_reset(lnk); 1135 smc_wr_free_link(lnk); 1136 smc_ib_destroy_queue_pair(lnk); 1137 smc_ib_dealloc_protection_domain(lnk); 1138 smc_wr_free_link_mem(lnk); 1139 smc_ibdev_cnt_dec(lnk); 1140 put_device(&lnk->smcibdev->ibdev->dev); 1141 smcibdev = lnk->smcibdev; 1142 memset(lnk, 0, sizeof(struct smc_link)); 1143 lnk->state = SMC_LNK_UNUSED; 1144 if (!atomic_dec_return(&smcibdev->lnk_cnt)) 1145 wake_up(&smcibdev->lnks_deleted); 1146 } 1147 1148 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb, 1149 struct smc_buf_desc *buf_desc) 1150 { 1151 int i; 1152 1153 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) 1154 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]); 1155 1156 if (buf_desc->pages) 1157 __free_pages(buf_desc->pages, buf_desc->order); 1158 kfree(buf_desc); 1159 } 1160 1161 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb, 1162 struct smc_buf_desc *buf_desc) 1163 { 1164 if (is_dmb) { 1165 /* restore original buf len */ 1166 buf_desc->len += sizeof(struct smcd_cdc_msg); 1167 smc_ism_unregister_dmb(lgr->smcd, buf_desc); 1168 } else { 1169 kfree(buf_desc->cpu_addr); 1170 } 1171 kfree(buf_desc); 1172 } 1173 1174 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb, 1175 struct smc_buf_desc *buf_desc) 1176 { 1177 if (lgr->is_smcd) 1178 smcd_buf_free(lgr, is_rmb, buf_desc); 1179 else 1180 smcr_buf_free(lgr, is_rmb, buf_desc); 1181 } 1182 1183 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb) 1184 { 1185 struct smc_buf_desc *buf_desc, *bf_desc; 1186 struct list_head *buf_list; 1187 int i; 1188 1189 for (i = 0; i < SMC_RMBE_SIZES; i++) { 1190 if (is_rmb) 1191 buf_list = &lgr->rmbs[i]; 1192 else 1193 buf_list = &lgr->sndbufs[i]; 1194 list_for_each_entry_safe(buf_desc, bf_desc, buf_list, 1195 list) { 1196 list_del(&buf_desc->list); 1197 smc_buf_free(lgr, is_rmb, buf_desc); 1198 } 1199 } 1200 } 1201 1202 static void smc_lgr_free_bufs(struct smc_link_group *lgr) 1203 { 1204 /* free send buffers */ 1205 __smc_lgr_free_bufs(lgr, false); 1206 /* free rmbs */ 1207 __smc_lgr_free_bufs(lgr, true); 1208 } 1209 1210 /* remove a link group */ 1211 static void smc_lgr_free(struct smc_link_group *lgr) 1212 { 1213 int i; 1214 1215 if (!lgr->is_smcd) { 1216 mutex_lock(&lgr->llc_conf_mutex); 1217 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 1218 if (lgr->lnk[i].state != SMC_LNK_UNUSED) 1219 smcr_link_clear(&lgr->lnk[i], false); 1220 } 1221 mutex_unlock(&lgr->llc_conf_mutex); 1222 smc_llc_lgr_clear(lgr); 1223 } 1224 1225 smc_lgr_free_bufs(lgr); 1226 destroy_workqueue(lgr->tx_wq); 1227 if (lgr->is_smcd) { 1228 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id); 1229 put_device(&lgr->smcd->dev); 1230 if (!atomic_dec_return(&lgr->smcd->lgr_cnt)) 1231 wake_up(&lgr->smcd->lgrs_deleted); 1232 } else { 1233 if (!atomic_dec_return(&lgr_cnt)) 1234 wake_up(&lgrs_deleted); 1235 } 1236 kfree(lgr); 1237 } 1238 1239 static void smc_sk_wake_ups(struct smc_sock *smc) 1240 { 1241 smc->sk.sk_write_space(&smc->sk); 1242 smc->sk.sk_data_ready(&smc->sk); 1243 smc->sk.sk_state_change(&smc->sk); 1244 } 1245 1246 /* kill a connection */ 1247 static void smc_conn_kill(struct smc_connection *conn, bool soft) 1248 { 1249 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 1250 1251 if (conn->lgr->is_smcd && conn->lgr->peer_shutdown) 1252 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 1253 else 1254 smc_close_abort(conn); 1255 conn->killed = 1; 1256 smc->sk.sk_err = ECONNABORTED; 1257 smc_sk_wake_ups(smc); 1258 if (conn->lgr->is_smcd) { 1259 smc_ism_unset_conn(conn); 1260 if (soft) 1261 tasklet_kill(&conn->rx_tsklet); 1262 else 1263 tasklet_unlock_wait(&conn->rx_tsklet); 1264 } else { 1265 smc_cdc_tx_dismiss_slots(conn); 1266 } 1267 smc_lgr_unregister_conn(conn); 1268 smc_close_active_abort(smc); 1269 } 1270 1271 static void smc_lgr_cleanup(struct smc_link_group *lgr) 1272 { 1273 if (lgr->is_smcd) { 1274 smc_ism_signal_shutdown(lgr); 1275 } else { 1276 u32 rsn = lgr->llc_termination_rsn; 1277 1278 if (!rsn) 1279 rsn = SMC_LLC_DEL_PROG_INIT_TERM; 1280 smc_llc_send_link_delete_all(lgr, false, rsn); 1281 smcr_lgr_link_deactivate_all(lgr); 1282 } 1283 } 1284 1285 /* terminate link group 1286 * @soft: true if link group shutdown can take its time 1287 * false if immediate link group shutdown is required 1288 */ 1289 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft) 1290 { 1291 struct smc_connection *conn; 1292 struct smc_sock *smc; 1293 struct rb_node *node; 1294 1295 if (lgr->terminating) 1296 return; /* lgr already terminating */ 1297 /* cancel free_work sync, will terminate when lgr->freeing is set */ 1298 cancel_delayed_work_sync(&lgr->free_work); 1299 lgr->terminating = 1; 1300 1301 /* kill remaining link group connections */ 1302 read_lock_bh(&lgr->conns_lock); 1303 node = rb_first(&lgr->conns_all); 1304 while (node) { 1305 read_unlock_bh(&lgr->conns_lock); 1306 conn = rb_entry(node, struct smc_connection, alert_node); 1307 smc = container_of(conn, struct smc_sock, conn); 1308 sock_hold(&smc->sk); /* sock_put below */ 1309 lock_sock(&smc->sk); 1310 smc_conn_kill(conn, soft); 1311 release_sock(&smc->sk); 1312 sock_put(&smc->sk); /* sock_hold above */ 1313 read_lock_bh(&lgr->conns_lock); 1314 node = rb_first(&lgr->conns_all); 1315 } 1316 read_unlock_bh(&lgr->conns_lock); 1317 smc_lgr_cleanup(lgr); 1318 smc_lgr_free(lgr); 1319 } 1320 1321 /* unlink link group and schedule termination */ 1322 void smc_lgr_terminate_sched(struct smc_link_group *lgr) 1323 { 1324 spinlock_t *lgr_lock; 1325 1326 smc_lgr_list_head(lgr, &lgr_lock); 1327 spin_lock_bh(lgr_lock); 1328 if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) { 1329 spin_unlock_bh(lgr_lock); 1330 return; /* lgr already terminating */ 1331 } 1332 list_del_init(&lgr->list); 1333 lgr->freeing = 1; 1334 spin_unlock_bh(lgr_lock); 1335 schedule_work(&lgr->terminate_work); 1336 } 1337 1338 /* Called when peer lgr shutdown (regularly or abnormally) is received */ 1339 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan) 1340 { 1341 struct smc_link_group *lgr, *l; 1342 LIST_HEAD(lgr_free_list); 1343 1344 /* run common cleanup function and build free list */ 1345 spin_lock_bh(&dev->lgr_lock); 1346 list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) { 1347 if ((!peer_gid || lgr->peer_gid == peer_gid) && 1348 (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) { 1349 if (peer_gid) /* peer triggered termination */ 1350 lgr->peer_shutdown = 1; 1351 list_move(&lgr->list, &lgr_free_list); 1352 lgr->freeing = 1; 1353 } 1354 } 1355 spin_unlock_bh(&dev->lgr_lock); 1356 1357 /* cancel the regular free workers and actually free lgrs */ 1358 list_for_each_entry_safe(lgr, l, &lgr_free_list, list) { 1359 list_del_init(&lgr->list); 1360 schedule_work(&lgr->terminate_work); 1361 } 1362 } 1363 1364 /* Called when an SMCD device is removed or the smc module is unloaded */ 1365 void smc_smcd_terminate_all(struct smcd_dev *smcd) 1366 { 1367 struct smc_link_group *lgr, *lg; 1368 LIST_HEAD(lgr_free_list); 1369 1370 spin_lock_bh(&smcd->lgr_lock); 1371 list_splice_init(&smcd->lgr_list, &lgr_free_list); 1372 list_for_each_entry(lgr, &lgr_free_list, list) 1373 lgr->freeing = 1; 1374 spin_unlock_bh(&smcd->lgr_lock); 1375 1376 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) { 1377 list_del_init(&lgr->list); 1378 __smc_lgr_terminate(lgr, false); 1379 } 1380 1381 if (atomic_read(&smcd->lgr_cnt)) 1382 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt)); 1383 } 1384 1385 /* Called when an SMCR device is removed or the smc module is unloaded. 1386 * If smcibdev is given, all SMCR link groups using this device are terminated. 1387 * If smcibdev is NULL, all SMCR link groups are terminated. 1388 */ 1389 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev) 1390 { 1391 struct smc_link_group *lgr, *lg; 1392 LIST_HEAD(lgr_free_list); 1393 int i; 1394 1395 spin_lock_bh(&smc_lgr_list.lock); 1396 if (!smcibdev) { 1397 list_splice_init(&smc_lgr_list.list, &lgr_free_list); 1398 list_for_each_entry(lgr, &lgr_free_list, list) 1399 lgr->freeing = 1; 1400 } else { 1401 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) { 1402 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 1403 if (lgr->lnk[i].smcibdev == smcibdev) 1404 smcr_link_down_cond_sched(&lgr->lnk[i]); 1405 } 1406 } 1407 } 1408 spin_unlock_bh(&smc_lgr_list.lock); 1409 1410 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) { 1411 list_del_init(&lgr->list); 1412 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM); 1413 __smc_lgr_terminate(lgr, false); 1414 } 1415 1416 if (smcibdev) { 1417 if (atomic_read(&smcibdev->lnk_cnt)) 1418 wait_event(smcibdev->lnks_deleted, 1419 !atomic_read(&smcibdev->lnk_cnt)); 1420 } else { 1421 if (atomic_read(&lgr_cnt)) 1422 wait_event(lgrs_deleted, !atomic_read(&lgr_cnt)); 1423 } 1424 } 1425 1426 /* set new lgr type and clear all asymmetric link tagging */ 1427 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type) 1428 { 1429 char *lgr_type = ""; 1430 int i; 1431 1432 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) 1433 if (smc_link_usable(&lgr->lnk[i])) 1434 lgr->lnk[i].link_is_asym = false; 1435 if (lgr->type == new_type) 1436 return; 1437 lgr->type = new_type; 1438 1439 switch (lgr->type) { 1440 case SMC_LGR_NONE: 1441 lgr_type = "NONE"; 1442 break; 1443 case SMC_LGR_SINGLE: 1444 lgr_type = "SINGLE"; 1445 break; 1446 case SMC_LGR_SYMMETRIC: 1447 lgr_type = "SYMMETRIC"; 1448 break; 1449 case SMC_LGR_ASYMMETRIC_PEER: 1450 lgr_type = "ASYMMETRIC_PEER"; 1451 break; 1452 case SMC_LGR_ASYMMETRIC_LOCAL: 1453 lgr_type = "ASYMMETRIC_LOCAL"; 1454 break; 1455 } 1456 pr_warn_ratelimited("smc: SMC-R lg %*phN state changed: " 1457 "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id, 1458 lgr_type, lgr->pnet_id); 1459 } 1460 1461 /* set new lgr type and tag a link as asymmetric */ 1462 void smcr_lgr_set_type_asym(struct smc_link_group *lgr, 1463 enum smc_lgr_type new_type, int asym_lnk_idx) 1464 { 1465 smcr_lgr_set_type(lgr, new_type); 1466 lgr->lnk[asym_lnk_idx].link_is_asym = true; 1467 } 1468 1469 /* abort connection, abort_work scheduled from tasklet context */ 1470 static void smc_conn_abort_work(struct work_struct *work) 1471 { 1472 struct smc_connection *conn = container_of(work, 1473 struct smc_connection, 1474 abort_work); 1475 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 1476 1477 smc_conn_kill(conn, true); 1478 sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */ 1479 } 1480 1481 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport) 1482 { 1483 struct smc_link_group *lgr, *n; 1484 1485 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) { 1486 struct smc_link *link; 1487 1488 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id, 1489 SMC_MAX_PNETID_LEN) || 1490 lgr->type == SMC_LGR_SYMMETRIC || 1491 lgr->type == SMC_LGR_ASYMMETRIC_PEER) 1492 continue; 1493 1494 /* trigger local add link processing */ 1495 link = smc_llc_usable_link(lgr); 1496 if (link) 1497 smc_llc_add_link_local(link); 1498 } 1499 } 1500 1501 /* link is down - switch connections to alternate link, 1502 * must be called under lgr->llc_conf_mutex lock 1503 */ 1504 static void smcr_link_down(struct smc_link *lnk) 1505 { 1506 struct smc_link_group *lgr = lnk->lgr; 1507 struct smc_link *to_lnk; 1508 int del_link_id; 1509 1510 if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list)) 1511 return; 1512 1513 smc_ib_modify_qp_reset(lnk); 1514 to_lnk = smc_switch_conns(lgr, lnk, true); 1515 if (!to_lnk) { /* no backup link available */ 1516 smcr_link_clear(lnk, true); 1517 return; 1518 } 1519 smcr_lgr_set_type(lgr, SMC_LGR_SINGLE); 1520 del_link_id = lnk->link_id; 1521 1522 if (lgr->role == SMC_SERV) { 1523 /* trigger local delete link processing */ 1524 smc_llc_srv_delete_link_local(to_lnk, del_link_id); 1525 } else { 1526 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) { 1527 /* another llc task is ongoing */ 1528 mutex_unlock(&lgr->llc_conf_mutex); 1529 wait_event_timeout(lgr->llc_flow_waiter, 1530 (list_empty(&lgr->list) || 1531 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE), 1532 SMC_LLC_WAIT_TIME); 1533 mutex_lock(&lgr->llc_conf_mutex); 1534 } 1535 if (!list_empty(&lgr->list)) { 1536 smc_llc_send_delete_link(to_lnk, del_link_id, 1537 SMC_LLC_REQ, true, 1538 SMC_LLC_DEL_LOST_PATH); 1539 smcr_link_clear(lnk, true); 1540 } 1541 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */ 1542 } 1543 } 1544 1545 /* must be called under lgr->llc_conf_mutex lock */ 1546 void smcr_link_down_cond(struct smc_link *lnk) 1547 { 1548 if (smc_link_downing(&lnk->state)) 1549 smcr_link_down(lnk); 1550 } 1551 1552 /* will get the lgr->llc_conf_mutex lock */ 1553 void smcr_link_down_cond_sched(struct smc_link *lnk) 1554 { 1555 if (smc_link_downing(&lnk->state)) 1556 schedule_work(&lnk->link_down_wrk); 1557 } 1558 1559 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport) 1560 { 1561 struct smc_link_group *lgr, *n; 1562 int i; 1563 1564 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) { 1565 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id, 1566 SMC_MAX_PNETID_LEN)) 1567 continue; /* lgr is not affected */ 1568 if (list_empty(&lgr->list)) 1569 continue; 1570 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 1571 struct smc_link *lnk = &lgr->lnk[i]; 1572 1573 if (smc_link_usable(lnk) && 1574 lnk->smcibdev == smcibdev && lnk->ibport == ibport) 1575 smcr_link_down_cond_sched(lnk); 1576 } 1577 } 1578 } 1579 1580 static void smc_link_down_work(struct work_struct *work) 1581 { 1582 struct smc_link *link = container_of(work, struct smc_link, 1583 link_down_wrk); 1584 struct smc_link_group *lgr = link->lgr; 1585 1586 if (list_empty(&lgr->list)) 1587 return; 1588 wake_up_all(&lgr->llc_msg_waiter); 1589 mutex_lock(&lgr->llc_conf_mutex); 1590 smcr_link_down(link); 1591 mutex_unlock(&lgr->llc_conf_mutex); 1592 } 1593 1594 /* Determine vlan of internal TCP socket. 1595 * @vlan_id: address to store the determined vlan id into 1596 */ 1597 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini) 1598 { 1599 struct dst_entry *dst = sk_dst_get(clcsock->sk); 1600 struct net_device *ndev; 1601 int i, nest_lvl, rc = 0; 1602 1603 ini->vlan_id = 0; 1604 if (!dst) { 1605 rc = -ENOTCONN; 1606 goto out; 1607 } 1608 if (!dst->dev) { 1609 rc = -ENODEV; 1610 goto out_rel; 1611 } 1612 1613 ndev = dst->dev; 1614 if (is_vlan_dev(ndev)) { 1615 ini->vlan_id = vlan_dev_vlan_id(ndev); 1616 goto out_rel; 1617 } 1618 1619 rtnl_lock(); 1620 nest_lvl = ndev->lower_level; 1621 for (i = 0; i < nest_lvl; i++) { 1622 struct list_head *lower = &ndev->adj_list.lower; 1623 1624 if (list_empty(lower)) 1625 break; 1626 lower = lower->next; 1627 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower); 1628 if (is_vlan_dev(ndev)) { 1629 ini->vlan_id = vlan_dev_vlan_id(ndev); 1630 break; 1631 } 1632 } 1633 rtnl_unlock(); 1634 1635 out_rel: 1636 dst_release(dst); 1637 out: 1638 return rc; 1639 } 1640 1641 static bool smcr_lgr_match(struct smc_link_group *lgr, 1642 struct smc_clc_msg_local *lcl, 1643 enum smc_lgr_role role, u32 clcqpn) 1644 { 1645 int i; 1646 1647 if (memcmp(lgr->peer_systemid, lcl->id_for_peer, SMC_SYSTEMID_LEN) || 1648 lgr->role != role) 1649 return false; 1650 1651 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 1652 if (!smc_link_active(&lgr->lnk[i])) 1653 continue; 1654 if ((lgr->role == SMC_SERV || lgr->lnk[i].peer_qpn == clcqpn) && 1655 !memcmp(lgr->lnk[i].peer_gid, &lcl->gid, SMC_GID_SIZE) && 1656 !memcmp(lgr->lnk[i].peer_mac, lcl->mac, sizeof(lcl->mac))) 1657 return true; 1658 } 1659 return false; 1660 } 1661 1662 static bool smcd_lgr_match(struct smc_link_group *lgr, 1663 struct smcd_dev *smcismdev, u64 peer_gid) 1664 { 1665 return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev; 1666 } 1667 1668 /* create a new SMC connection (and a new link group if necessary) */ 1669 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini) 1670 { 1671 struct smc_connection *conn = &smc->conn; 1672 struct list_head *lgr_list; 1673 struct smc_link_group *lgr; 1674 enum smc_lgr_role role; 1675 spinlock_t *lgr_lock; 1676 int rc = 0; 1677 1678 lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list : 1679 &smc_lgr_list.list; 1680 lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock : 1681 &smc_lgr_list.lock; 1682 ini->first_contact_local = 1; 1683 role = smc->listen_smc ? SMC_SERV : SMC_CLNT; 1684 if (role == SMC_CLNT && ini->first_contact_peer) 1685 /* create new link group as well */ 1686 goto create; 1687 1688 /* determine if an existing link group can be reused */ 1689 spin_lock_bh(lgr_lock); 1690 list_for_each_entry(lgr, lgr_list, list) { 1691 write_lock_bh(&lgr->conns_lock); 1692 if ((ini->is_smcd ? 1693 smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected], 1694 ini->ism_peer_gid[ini->ism_selected]) : 1695 smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) && 1696 !lgr->sync_err && 1697 (ini->smcd_version == SMC_V2 || 1698 lgr->vlan_id == ini->vlan_id) && 1699 (role == SMC_CLNT || ini->is_smcd || 1700 lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) { 1701 /* link group found */ 1702 ini->first_contact_local = 0; 1703 conn->lgr = lgr; 1704 rc = smc_lgr_register_conn(conn, false); 1705 write_unlock_bh(&lgr->conns_lock); 1706 if (!rc && delayed_work_pending(&lgr->free_work)) 1707 cancel_delayed_work(&lgr->free_work); 1708 break; 1709 } 1710 write_unlock_bh(&lgr->conns_lock); 1711 } 1712 spin_unlock_bh(lgr_lock); 1713 if (rc) 1714 return rc; 1715 1716 if (role == SMC_CLNT && !ini->first_contact_peer && 1717 ini->first_contact_local) { 1718 /* Server reuses a link group, but Client wants to start 1719 * a new one 1720 * send out_of_sync decline, reason synchr. error 1721 */ 1722 return SMC_CLC_DECL_SYNCERR; 1723 } 1724 1725 create: 1726 if (ini->first_contact_local) { 1727 rc = smc_lgr_create(smc, ini); 1728 if (rc) 1729 goto out; 1730 lgr = conn->lgr; 1731 write_lock_bh(&lgr->conns_lock); 1732 rc = smc_lgr_register_conn(conn, true); 1733 write_unlock_bh(&lgr->conns_lock); 1734 if (rc) 1735 goto out; 1736 } 1737 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE; 1738 conn->local_tx_ctrl.len = SMC_WR_TX_SIZE; 1739 conn->urg_state = SMC_URG_READ; 1740 INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work); 1741 if (ini->is_smcd) { 1742 conn->rx_off = sizeof(struct smcd_cdc_msg); 1743 smcd_cdc_rx_init(conn); /* init tasklet for this conn */ 1744 } else { 1745 conn->rx_off = 0; 1746 } 1747 #ifndef KERNEL_HAS_ATOMIC64 1748 spin_lock_init(&conn->acurs_lock); 1749 #endif 1750 1751 out: 1752 return rc; 1753 } 1754 1755 #define SMCD_DMBE_SIZES 6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */ 1756 #define SMCR_RMBE_SIZES 5 /* 0 -> 16KB, 1 -> 32KB, .. 5 -> 512KB */ 1757 1758 /* convert the RMB size into the compressed notation (minimum 16K, see 1759 * SMCD/R_DMBE_SIZES. 1760 * In contrast to plain ilog2, this rounds towards the next power of 2, 1761 * so the socket application gets at least its desired sndbuf / rcvbuf size. 1762 */ 1763 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb) 1764 { 1765 const unsigned int max_scat = SG_MAX_SINGLE_ALLOC * PAGE_SIZE; 1766 u8 compressed; 1767 1768 if (size <= SMC_BUF_MIN_SIZE) 1769 return 0; 1770 1771 size = (size - 1) >> 14; /* convert to 16K multiple */ 1772 compressed = min_t(u8, ilog2(size) + 1, 1773 is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES); 1774 1775 if (!is_smcd && is_rmb) 1776 /* RMBs are backed by & limited to max size of scatterlists */ 1777 compressed = min_t(u8, compressed, ilog2(max_scat >> 14)); 1778 1779 return compressed; 1780 } 1781 1782 /* convert the RMB size from compressed notation into integer */ 1783 int smc_uncompress_bufsize(u8 compressed) 1784 { 1785 u32 size; 1786 1787 size = 0x00000001 << (((int)compressed) + 14); 1788 return (int)size; 1789 } 1790 1791 /* try to reuse a sndbuf or rmb description slot for a certain 1792 * buffer size; if not available, return NULL 1793 */ 1794 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize, 1795 struct mutex *lock, 1796 struct list_head *buf_list) 1797 { 1798 struct smc_buf_desc *buf_slot; 1799 1800 mutex_lock(lock); 1801 list_for_each_entry(buf_slot, buf_list, list) { 1802 if (cmpxchg(&buf_slot->used, 0, 1) == 0) { 1803 mutex_unlock(lock); 1804 return buf_slot; 1805 } 1806 } 1807 mutex_unlock(lock); 1808 return NULL; 1809 } 1810 1811 /* one of the conditions for announcing a receiver's current window size is 1812 * that it "results in a minimum increase in the window size of 10% of the 1813 * receive buffer space" [RFC7609] 1814 */ 1815 static inline int smc_rmb_wnd_update_limit(int rmbe_size) 1816 { 1817 return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2); 1818 } 1819 1820 /* map an rmb buf to a link */ 1821 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb, 1822 struct smc_link *lnk) 1823 { 1824 int rc; 1825 1826 if (buf_desc->is_map_ib[lnk->link_idx]) 1827 return 0; 1828 1829 rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], 1, GFP_KERNEL); 1830 if (rc) 1831 return rc; 1832 sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl, 1833 buf_desc->cpu_addr, buf_desc->len); 1834 1835 /* map sg table to DMA address */ 1836 rc = smc_ib_buf_map_sg(lnk, buf_desc, 1837 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE); 1838 /* SMC protocol depends on mapping to one DMA address only */ 1839 if (rc != 1) { 1840 rc = -EAGAIN; 1841 goto free_table; 1842 } 1843 1844 /* create a new memory region for the RMB */ 1845 if (is_rmb) { 1846 rc = smc_ib_get_memory_region(lnk->roce_pd, 1847 IB_ACCESS_REMOTE_WRITE | 1848 IB_ACCESS_LOCAL_WRITE, 1849 buf_desc, lnk->link_idx); 1850 if (rc) 1851 goto buf_unmap; 1852 smc_ib_sync_sg_for_device(lnk, buf_desc, DMA_FROM_DEVICE); 1853 } 1854 buf_desc->is_map_ib[lnk->link_idx] = true; 1855 return 0; 1856 1857 buf_unmap: 1858 smc_ib_buf_unmap_sg(lnk, buf_desc, 1859 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE); 1860 free_table: 1861 sg_free_table(&buf_desc->sgt[lnk->link_idx]); 1862 return rc; 1863 } 1864 1865 /* register a new rmb on IB device, 1866 * must be called under lgr->llc_conf_mutex lock 1867 */ 1868 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc) 1869 { 1870 if (list_empty(&link->lgr->list)) 1871 return -ENOLINK; 1872 if (!rmb_desc->is_reg_mr[link->link_idx]) { 1873 /* register memory region for new rmb */ 1874 if (smc_wr_reg_send(link, rmb_desc->mr_rx[link->link_idx])) { 1875 rmb_desc->is_reg_err = true; 1876 return -EFAULT; 1877 } 1878 rmb_desc->is_reg_mr[link->link_idx] = true; 1879 } 1880 return 0; 1881 } 1882 1883 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock, 1884 struct list_head *lst, bool is_rmb) 1885 { 1886 struct smc_buf_desc *buf_desc, *bf; 1887 int rc = 0; 1888 1889 mutex_lock(lock); 1890 list_for_each_entry_safe(buf_desc, bf, lst, list) { 1891 if (!buf_desc->used) 1892 continue; 1893 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk); 1894 if (rc) 1895 goto out; 1896 } 1897 out: 1898 mutex_unlock(lock); 1899 return rc; 1900 } 1901 1902 /* map all used buffers of lgr for a new link */ 1903 int smcr_buf_map_lgr(struct smc_link *lnk) 1904 { 1905 struct smc_link_group *lgr = lnk->lgr; 1906 int i, rc = 0; 1907 1908 for (i = 0; i < SMC_RMBE_SIZES; i++) { 1909 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock, 1910 &lgr->rmbs[i], true); 1911 if (rc) 1912 return rc; 1913 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock, 1914 &lgr->sndbufs[i], false); 1915 if (rc) 1916 return rc; 1917 } 1918 return 0; 1919 } 1920 1921 /* register all used buffers of lgr for a new link, 1922 * must be called under lgr->llc_conf_mutex lock 1923 */ 1924 int smcr_buf_reg_lgr(struct smc_link *lnk) 1925 { 1926 struct smc_link_group *lgr = lnk->lgr; 1927 struct smc_buf_desc *buf_desc, *bf; 1928 int i, rc = 0; 1929 1930 mutex_lock(&lgr->rmbs_lock); 1931 for (i = 0; i < SMC_RMBE_SIZES; i++) { 1932 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) { 1933 if (!buf_desc->used) 1934 continue; 1935 rc = smcr_link_reg_rmb(lnk, buf_desc); 1936 if (rc) 1937 goto out; 1938 } 1939 } 1940 out: 1941 mutex_unlock(&lgr->rmbs_lock); 1942 return rc; 1943 } 1944 1945 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr, 1946 bool is_rmb, int bufsize) 1947 { 1948 struct smc_buf_desc *buf_desc; 1949 1950 /* try to alloc a new buffer */ 1951 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL); 1952 if (!buf_desc) 1953 return ERR_PTR(-ENOMEM); 1954 1955 buf_desc->order = get_order(bufsize); 1956 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN | 1957 __GFP_NOMEMALLOC | __GFP_COMP | 1958 __GFP_NORETRY | __GFP_ZERO, 1959 buf_desc->order); 1960 if (!buf_desc->pages) { 1961 kfree(buf_desc); 1962 return ERR_PTR(-EAGAIN); 1963 } 1964 buf_desc->cpu_addr = (void *)page_address(buf_desc->pages); 1965 buf_desc->len = bufsize; 1966 return buf_desc; 1967 } 1968 1969 /* map buf_desc on all usable links, 1970 * unused buffers stay mapped as long as the link is up 1971 */ 1972 static int smcr_buf_map_usable_links(struct smc_link_group *lgr, 1973 struct smc_buf_desc *buf_desc, bool is_rmb) 1974 { 1975 int i, rc = 0; 1976 1977 /* protect against parallel link reconfiguration */ 1978 mutex_lock(&lgr->llc_conf_mutex); 1979 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 1980 struct smc_link *lnk = &lgr->lnk[i]; 1981 1982 if (!smc_link_usable(lnk)) 1983 continue; 1984 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) { 1985 rc = -ENOMEM; 1986 goto out; 1987 } 1988 } 1989 out: 1990 mutex_unlock(&lgr->llc_conf_mutex); 1991 return rc; 1992 } 1993 1994 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr, 1995 bool is_dmb, int bufsize) 1996 { 1997 struct smc_buf_desc *buf_desc; 1998 int rc; 1999 2000 /* try to alloc a new DMB */ 2001 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL); 2002 if (!buf_desc) 2003 return ERR_PTR(-ENOMEM); 2004 if (is_dmb) { 2005 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc); 2006 if (rc) { 2007 kfree(buf_desc); 2008 if (rc == -ENOMEM) 2009 return ERR_PTR(-EAGAIN); 2010 if (rc == -ENOSPC) 2011 return ERR_PTR(-ENOSPC); 2012 return ERR_PTR(-EIO); 2013 } 2014 buf_desc->pages = virt_to_page(buf_desc->cpu_addr); 2015 /* CDC header stored in buf. So, pretend it was smaller */ 2016 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg); 2017 } else { 2018 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL | 2019 __GFP_NOWARN | __GFP_NORETRY | 2020 __GFP_NOMEMALLOC); 2021 if (!buf_desc->cpu_addr) { 2022 kfree(buf_desc); 2023 return ERR_PTR(-EAGAIN); 2024 } 2025 buf_desc->len = bufsize; 2026 } 2027 return buf_desc; 2028 } 2029 2030 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb) 2031 { 2032 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM); 2033 struct smc_connection *conn = &smc->conn; 2034 struct smc_link_group *lgr = conn->lgr; 2035 struct list_head *buf_list; 2036 int bufsize, bufsize_short; 2037 bool is_dgraded = false; 2038 struct mutex *lock; /* lock buffer list */ 2039 int sk_buf_size; 2040 2041 if (is_rmb) 2042 /* use socket recv buffer size (w/o overhead) as start value */ 2043 sk_buf_size = smc->sk.sk_rcvbuf / 2; 2044 else 2045 /* use socket send buffer size (w/o overhead) as start value */ 2046 sk_buf_size = smc->sk.sk_sndbuf / 2; 2047 2048 for (bufsize_short = smc_compress_bufsize(sk_buf_size, is_smcd, is_rmb); 2049 bufsize_short >= 0; bufsize_short--) { 2050 if (is_rmb) { 2051 lock = &lgr->rmbs_lock; 2052 buf_list = &lgr->rmbs[bufsize_short]; 2053 } else { 2054 lock = &lgr->sndbufs_lock; 2055 buf_list = &lgr->sndbufs[bufsize_short]; 2056 } 2057 bufsize = smc_uncompress_bufsize(bufsize_short); 2058 2059 /* check for reusable slot in the link group */ 2060 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list); 2061 if (buf_desc) { 2062 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize); 2063 SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb); 2064 memset(buf_desc->cpu_addr, 0, bufsize); 2065 break; /* found reusable slot */ 2066 } 2067 2068 if (is_smcd) 2069 buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize); 2070 else 2071 buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize); 2072 2073 if (PTR_ERR(buf_desc) == -ENOMEM) 2074 break; 2075 if (IS_ERR(buf_desc)) { 2076 if (!is_dgraded) { 2077 is_dgraded = true; 2078 SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb); 2079 } 2080 continue; 2081 } 2082 2083 SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb); 2084 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize); 2085 buf_desc->used = 1; 2086 mutex_lock(lock); 2087 list_add(&buf_desc->list, buf_list); 2088 mutex_unlock(lock); 2089 break; /* found */ 2090 } 2091 2092 if (IS_ERR(buf_desc)) 2093 return PTR_ERR(buf_desc); 2094 2095 if (!is_smcd) { 2096 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) { 2097 smcr_buf_unuse(buf_desc, lgr); 2098 return -ENOMEM; 2099 } 2100 } 2101 2102 if (is_rmb) { 2103 conn->rmb_desc = buf_desc; 2104 conn->rmbe_size_short = bufsize_short; 2105 smc->sk.sk_rcvbuf = bufsize * 2; 2106 atomic_set(&conn->bytes_to_rcv, 0); 2107 conn->rmbe_update_limit = 2108 smc_rmb_wnd_update_limit(buf_desc->len); 2109 if (is_smcd) 2110 smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */ 2111 } else { 2112 conn->sndbuf_desc = buf_desc; 2113 smc->sk.sk_sndbuf = bufsize * 2; 2114 atomic_set(&conn->sndbuf_space, bufsize); 2115 } 2116 return 0; 2117 } 2118 2119 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn) 2120 { 2121 if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk)) 2122 return; 2123 smc_ib_sync_sg_for_cpu(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE); 2124 } 2125 2126 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn) 2127 { 2128 if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk)) 2129 return; 2130 smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE); 2131 } 2132 2133 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn) 2134 { 2135 int i; 2136 2137 if (!conn->lgr || conn->lgr->is_smcd) 2138 return; 2139 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 2140 if (!smc_link_active(&conn->lgr->lnk[i])) 2141 continue; 2142 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc, 2143 DMA_FROM_DEVICE); 2144 } 2145 } 2146 2147 void smc_rmb_sync_sg_for_device(struct smc_connection *conn) 2148 { 2149 int i; 2150 2151 if (!conn->lgr || conn->lgr->is_smcd) 2152 return; 2153 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 2154 if (!smc_link_active(&conn->lgr->lnk[i])) 2155 continue; 2156 smc_ib_sync_sg_for_device(&conn->lgr->lnk[i], conn->rmb_desc, 2157 DMA_FROM_DEVICE); 2158 } 2159 } 2160 2161 /* create the send and receive buffer for an SMC socket; 2162 * receive buffers are called RMBs; 2163 * (even though the SMC protocol allows more than one RMB-element per RMB, 2164 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an 2165 * extra RMB for every connection in a link group 2166 */ 2167 int smc_buf_create(struct smc_sock *smc, bool is_smcd) 2168 { 2169 int rc; 2170 2171 /* create send buffer */ 2172 rc = __smc_buf_create(smc, is_smcd, false); 2173 if (rc) 2174 return rc; 2175 /* create rmb */ 2176 rc = __smc_buf_create(smc, is_smcd, true); 2177 if (rc) { 2178 mutex_lock(&smc->conn.lgr->sndbufs_lock); 2179 list_del(&smc->conn.sndbuf_desc->list); 2180 mutex_unlock(&smc->conn.lgr->sndbufs_lock); 2181 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc); 2182 smc->conn.sndbuf_desc = NULL; 2183 } 2184 return rc; 2185 } 2186 2187 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr) 2188 { 2189 int i; 2190 2191 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) { 2192 if (!test_and_set_bit(i, lgr->rtokens_used_mask)) 2193 return i; 2194 } 2195 return -ENOSPC; 2196 } 2197 2198 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx, 2199 u32 rkey) 2200 { 2201 int i; 2202 2203 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) { 2204 if (test_bit(i, lgr->rtokens_used_mask) && 2205 lgr->rtokens[i][lnk_idx].rkey == rkey) 2206 return i; 2207 } 2208 return -ENOENT; 2209 } 2210 2211 /* set rtoken for a new link to an existing rmb */ 2212 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new, 2213 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey) 2214 { 2215 int rtok_idx; 2216 2217 rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known)); 2218 if (rtok_idx == -ENOENT) 2219 return; 2220 lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey); 2221 lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr); 2222 } 2223 2224 /* set rtoken for a new link whose link_id is given */ 2225 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id, 2226 __be64 nw_vaddr, __be32 nw_rkey) 2227 { 2228 u64 dma_addr = be64_to_cpu(nw_vaddr); 2229 u32 rkey = ntohl(nw_rkey); 2230 bool found = false; 2231 int link_idx; 2232 2233 for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) { 2234 if (lgr->lnk[link_idx].link_id == link_id) { 2235 found = true; 2236 break; 2237 } 2238 } 2239 if (!found) 2240 return; 2241 lgr->rtokens[rtok_idx][link_idx].rkey = rkey; 2242 lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr; 2243 } 2244 2245 /* add a new rtoken from peer */ 2246 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey) 2247 { 2248 struct smc_link_group *lgr = smc_get_lgr(lnk); 2249 u64 dma_addr = be64_to_cpu(nw_vaddr); 2250 u32 rkey = ntohl(nw_rkey); 2251 int i; 2252 2253 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) { 2254 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey && 2255 lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr && 2256 test_bit(i, lgr->rtokens_used_mask)) { 2257 /* already in list */ 2258 return i; 2259 } 2260 } 2261 i = smc_rmb_reserve_rtoken_idx(lgr); 2262 if (i < 0) 2263 return i; 2264 lgr->rtokens[i][lnk->link_idx].rkey = rkey; 2265 lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr; 2266 return i; 2267 } 2268 2269 /* delete an rtoken from all links */ 2270 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey) 2271 { 2272 struct smc_link_group *lgr = smc_get_lgr(lnk); 2273 u32 rkey = ntohl(nw_rkey); 2274 int i, j; 2275 2276 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) { 2277 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey && 2278 test_bit(i, lgr->rtokens_used_mask)) { 2279 for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) { 2280 lgr->rtokens[i][j].rkey = 0; 2281 lgr->rtokens[i][j].dma_addr = 0; 2282 } 2283 clear_bit(i, lgr->rtokens_used_mask); 2284 return 0; 2285 } 2286 } 2287 return -ENOENT; 2288 } 2289 2290 /* save rkey and dma_addr received from peer during clc handshake */ 2291 int smc_rmb_rtoken_handling(struct smc_connection *conn, 2292 struct smc_link *lnk, 2293 struct smc_clc_msg_accept_confirm *clc) 2294 { 2295 conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr, 2296 clc->r0.rmb_rkey); 2297 if (conn->rtoken_idx < 0) 2298 return conn->rtoken_idx; 2299 return 0; 2300 } 2301 2302 static void smc_core_going_away(void) 2303 { 2304 struct smc_ib_device *smcibdev; 2305 struct smcd_dev *smcd; 2306 2307 mutex_lock(&smc_ib_devices.mutex); 2308 list_for_each_entry(smcibdev, &smc_ib_devices.list, list) { 2309 int i; 2310 2311 for (i = 0; i < SMC_MAX_PORTS; i++) 2312 set_bit(i, smcibdev->ports_going_away); 2313 } 2314 mutex_unlock(&smc_ib_devices.mutex); 2315 2316 mutex_lock(&smcd_dev_list.mutex); 2317 list_for_each_entry(smcd, &smcd_dev_list.list, list) { 2318 smcd->going_away = 1; 2319 } 2320 mutex_unlock(&smcd_dev_list.mutex); 2321 } 2322 2323 /* Clean up all SMC link groups */ 2324 static void smc_lgrs_shutdown(void) 2325 { 2326 struct smcd_dev *smcd; 2327 2328 smc_core_going_away(); 2329 2330 smc_smcr_terminate_all(NULL); 2331 2332 mutex_lock(&smcd_dev_list.mutex); 2333 list_for_each_entry(smcd, &smcd_dev_list.list, list) 2334 smc_smcd_terminate_all(smcd); 2335 mutex_unlock(&smcd_dev_list.mutex); 2336 } 2337 2338 static int smc_core_reboot_event(struct notifier_block *this, 2339 unsigned long event, void *ptr) 2340 { 2341 smc_lgrs_shutdown(); 2342 smc_ib_unregister_client(); 2343 return 0; 2344 } 2345 2346 static struct notifier_block smc_reboot_notifier = { 2347 .notifier_call = smc_core_reboot_event, 2348 }; 2349 2350 int __init smc_core_init(void) 2351 { 2352 return register_reboot_notifier(&smc_reboot_notifier); 2353 } 2354 2355 /* Called (from smc_exit) when module is removed */ 2356 void smc_core_exit(void) 2357 { 2358 unregister_reboot_notifier(&smc_reboot_notifier); 2359 smc_lgrs_shutdown(); 2360 } 2361