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