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->max_send_wr = lgr->max_send_wr; 814 lnk->max_recv_wr = lgr->max_recv_wr; 815 lnk->lgr = lgr; 816 smc_lgr_hold(lgr); /* lgr_put in smcr_link_clear() */ 817 lnk->link_idx = link_idx; 818 lnk->wr_rx_id_compl = 0; 819 smc_ibdev_cnt_inc(lnk); 820 smcr_copy_dev_info_to_link(lnk); 821 atomic_set(&lnk->conn_cnt, 0); 822 smc_llc_link_set_uid(lnk); 823 INIT_WORK(&lnk->link_down_wrk, smc_link_down_work); 824 if (!lnk->smcibdev->initialized) { 825 rc = (int)smc_ib_setup_per_ibdev(lnk->smcibdev); 826 if (rc) 827 goto out; 828 } 829 get_random_bytes(rndvec, sizeof(rndvec)); 830 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) + 831 (rndvec[2] << 16); 832 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport, 833 ini->vlan_id, lnk->gid, &lnk->sgid_index, 834 lgr->smc_version == SMC_V2 ? 835 &ini->smcrv2 : NULL); 836 if (rc) 837 goto out; 838 rc = smc_llc_link_init(lnk); 839 if (rc) 840 goto out; 841 rc = smc_ib_create_protection_domain(lnk); 842 if (rc) 843 goto clear_llc_lnk; 844 do { 845 rc = smc_ib_create_queue_pair(lnk); 846 if (rc) 847 goto dealloc_pd; 848 rc = smc_wr_alloc_link_mem(lnk); 849 if (!rc) 850 break; 851 else if (rc != -ENOMEM) /* give up */ 852 goto destroy_qp; 853 /* retry with smaller ... */ 854 lnk->max_send_wr /= 2; 855 lnk->max_recv_wr /= 2; 856 /* ... unless droping below old SMC_WR_BUF_SIZE */ 857 if (lnk->max_send_wr < 16 || lnk->max_recv_wr < 48) 858 goto destroy_qp; 859 smc_ib_destroy_queue_pair(lnk); 860 } while (1); 861 862 rc = smc_wr_create_link(lnk); 863 if (rc) 864 goto free_link_mem; 865 lnk->state = SMC_LNK_ACTIVATING; 866 return 0; 867 868 free_link_mem: 869 smc_wr_free_link_mem(lnk); 870 destroy_qp: 871 smc_ib_destroy_queue_pair(lnk); 872 dealloc_pd: 873 smc_ib_dealloc_protection_domain(lnk); 874 clear_llc_lnk: 875 smc_llc_link_clear(lnk, false); 876 out: 877 smc_ibdev_cnt_dec(lnk); 878 put_device(&lnk->smcibdev->ibdev->dev); 879 smcibdev = lnk->smcibdev; 880 memset(lnk, 0, sizeof(struct smc_link)); 881 lnk->state = SMC_LNK_UNUSED; 882 if (!atomic_dec_return(&smcibdev->lnk_cnt)) 883 wake_up(&smcibdev->lnks_deleted); 884 smc_lgr_put(lgr); /* lgr_hold above */ 885 return rc; 886 } 887 888 /* create a new SMC link group */ 889 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini) 890 { 891 struct smc_link_group *lgr; 892 struct list_head *lgr_list; 893 struct smcd_dev *smcd; 894 struct smc_link *lnk; 895 spinlock_t *lgr_lock; 896 u8 link_idx; 897 int rc = 0; 898 int i; 899 900 if (ini->is_smcd && ini->vlan_id) { 901 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected], 902 ini->vlan_id)) { 903 rc = SMC_CLC_DECL_ISMVLANERR; 904 goto out; 905 } 906 } 907 908 lgr = kzalloc_obj(*lgr); 909 if (!lgr) { 910 rc = SMC_CLC_DECL_MEM; 911 goto ism_put_vlan; 912 } 913 lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", WQ_PERCPU, 0, 914 SMC_LGR_ID_SIZE, &lgr->id); 915 if (!lgr->tx_wq) { 916 rc = -ENOMEM; 917 goto free_lgr; 918 } 919 lgr->is_smcd = ini->is_smcd; 920 lgr->sync_err = 0; 921 lgr->terminating = 0; 922 lgr->freeing = 0; 923 lgr->vlan_id = ini->vlan_id; 924 refcount_set(&lgr->refcnt, 1); /* set lgr refcnt to 1 */ 925 init_rwsem(&lgr->sndbufs_lock); 926 init_rwsem(&lgr->rmbs_lock); 927 rwlock_init(&lgr->conns_lock); 928 for (i = 0; i < SMC_RMBE_SIZES; i++) { 929 INIT_LIST_HEAD(&lgr->sndbufs[i]); 930 INIT_LIST_HEAD(&lgr->rmbs[i]); 931 } 932 lgr->next_link_id = 0; 933 smc_lgr_list.num += SMC_LGR_NUM_INCR; 934 memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE); 935 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work); 936 INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work); 937 lgr->conns_all = RB_ROOT; 938 if (ini->is_smcd) { 939 /* SMC-D specific settings */ 940 smcd = ini->ism_dev[ini->ism_selected]; 941 get_device(&smcd->dibs->dev); 942 lgr->peer_gid.gid = 943 ini->ism_peer_gid[ini->ism_selected].gid; 944 lgr->peer_gid.gid_ext = 945 ini->ism_peer_gid[ini->ism_selected].gid_ext; 946 lgr->smcd = ini->ism_dev[ini->ism_selected]; 947 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list; 948 lgr_lock = &lgr->smcd->lgr_lock; 949 lgr->smc_version = ini->smcd_version; 950 lgr->peer_shutdown = 0; 951 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt); 952 } else { 953 /* SMC-R specific settings */ 954 struct smc_ib_device *ibdev; 955 int ibport; 956 957 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT; 958 lgr->smc_version = ini->smcr_version; 959 memcpy(lgr->peer_systemid, ini->peer_systemid, 960 SMC_SYSTEMID_LEN); 961 if (lgr->smc_version == SMC_V2) { 962 ibdev = ini->smcrv2.ib_dev_v2; 963 ibport = ini->smcrv2.ib_port_v2; 964 lgr->saddr = ini->smcrv2.saddr; 965 lgr->uses_gateway = ini->smcrv2.uses_gateway; 966 memcpy(lgr->nexthop_mac, ini->smcrv2.nexthop_mac, 967 ETH_ALEN); 968 lgr->max_conns = ini->max_conns; 969 lgr->max_links = ini->max_links; 970 } else { 971 ibdev = ini->ib_dev; 972 ibport = ini->ib_port; 973 lgr->max_conns = SMC_CONN_PER_LGR_MAX; 974 lgr->max_links = SMC_LINKS_ADD_LNK_MAX; 975 } 976 memcpy(lgr->pnet_id, ibdev->pnetid[ibport - 1], 977 SMC_MAX_PNETID_LEN); 978 rc = smc_wr_alloc_lgr_mem(lgr); 979 if (rc) 980 goto free_wq; 981 smc_llc_lgr_init(lgr, smc); 982 983 link_idx = SMC_SINGLE_LINK; 984 lnk = &lgr->lnk[link_idx]; 985 rc = smcr_link_init(lgr, lnk, link_idx, ini); 986 if (rc) { 987 smc_wr_free_lgr_mem(lgr); 988 goto free_wq; 989 } 990 lgr->net = smc_ib_net(lnk->smcibdev); 991 lgr_list = &smc_lgr_list.list; 992 lgr_lock = &smc_lgr_list.lock; 993 lgr->buf_type = lgr->net->smc.sysctl_smcr_buf_type; 994 atomic_inc(&lgr_cnt); 995 } 996 smc->conn.lgr = lgr; 997 spin_lock_bh(lgr_lock); 998 list_add_tail(&lgr->list, lgr_list); 999 spin_unlock_bh(lgr_lock); 1000 return 0; 1001 1002 free_wq: 1003 destroy_workqueue(lgr->tx_wq); 1004 free_lgr: 1005 kfree(lgr); 1006 ism_put_vlan: 1007 if (ini->is_smcd && ini->vlan_id) 1008 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id); 1009 out: 1010 if (rc < 0) { 1011 if (rc == -ENOMEM) 1012 rc = SMC_CLC_DECL_MEM; 1013 else 1014 rc = SMC_CLC_DECL_INTERR; 1015 } 1016 return rc; 1017 } 1018 1019 static int smc_write_space(struct smc_connection *conn) 1020 { 1021 int buffer_len = conn->peer_rmbe_size; 1022 union smc_host_cursor prod; 1023 union smc_host_cursor cons; 1024 int space; 1025 1026 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn); 1027 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn); 1028 /* determine rx_buf space */ 1029 space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod); 1030 return space; 1031 } 1032 1033 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend, 1034 struct smc_wr_buf *wr_buf) 1035 { 1036 struct smc_connection *conn = &smc->conn; 1037 union smc_host_cursor cons, fin; 1038 int rc = 0; 1039 int diff; 1040 1041 smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn); 1042 smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn); 1043 /* set prod cursor to old state, enforce tx_rdma_writes() */ 1044 smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn); 1045 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn); 1046 1047 if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) { 1048 /* cons cursor advanced more than fin, and prod was set 1049 * fin above, so now prod is smaller than cons. Fix that. 1050 */ 1051 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons); 1052 smc_curs_add(conn->sndbuf_desc->len, 1053 &conn->tx_curs_sent, diff); 1054 smc_curs_add(conn->sndbuf_desc->len, 1055 &conn->tx_curs_fin, diff); 1056 1057 smp_mb__before_atomic(); 1058 atomic_add(diff, &conn->sndbuf_space); 1059 smp_mb__after_atomic(); 1060 1061 smc_curs_add(conn->peer_rmbe_size, 1062 &conn->local_tx_ctrl.prod, diff); 1063 smc_curs_add(conn->peer_rmbe_size, 1064 &conn->local_tx_ctrl_fin, diff); 1065 } 1066 /* recalculate, value is used by tx_rdma_writes() */ 1067 atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn)); 1068 1069 if (smc->sk.sk_state != SMC_INIT && 1070 smc->sk.sk_state != SMC_CLOSED) { 1071 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf); 1072 if (!rc) { 1073 sock_hold(&smc->sk); 1074 if (!queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0)) 1075 sock_put(&smc->sk); 1076 smc->sk.sk_data_ready(&smc->sk); 1077 } 1078 } else { 1079 smc_wr_tx_put_slot(conn->lnk, 1080 (struct smc_wr_tx_pend_priv *)pend); 1081 } 1082 return rc; 1083 } 1084 1085 void smc_switch_link_and_count(struct smc_connection *conn, 1086 struct smc_link *to_lnk) 1087 { 1088 atomic_dec(&conn->lnk->conn_cnt); 1089 /* link_hold in smc_conn_create() */ 1090 smcr_link_put(conn->lnk); 1091 conn->lnk = to_lnk; 1092 atomic_inc(&conn->lnk->conn_cnt); 1093 /* link_put in smc_conn_free() */ 1094 smcr_link_hold(conn->lnk); 1095 } 1096 1097 struct smc_link *smc_switch_conns(struct smc_link_group *lgr, 1098 struct smc_link *from_lnk, bool is_dev_err) 1099 { 1100 struct smc_link *to_lnk = NULL; 1101 struct smc_cdc_tx_pend *pend; 1102 struct smc_connection *conn; 1103 struct smc_wr_buf *wr_buf; 1104 struct smc_sock *smc; 1105 struct rb_node *node; 1106 int i, rc = 0; 1107 1108 /* link is inactive, wake up tx waiters */ 1109 smc_wr_wakeup_tx_wait(from_lnk); 1110 1111 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 1112 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx) 1113 continue; 1114 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev && 1115 from_lnk->ibport == lgr->lnk[i].ibport) { 1116 continue; 1117 } 1118 to_lnk = &lgr->lnk[i]; 1119 break; 1120 } 1121 if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) { 1122 smc_lgr_terminate_sched(lgr); 1123 return NULL; 1124 } 1125 again: 1126 read_lock_bh(&lgr->conns_lock); 1127 for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) { 1128 conn = rb_entry(node, struct smc_connection, alert_node); 1129 if (conn->lnk != from_lnk) 1130 continue; 1131 smc = container_of(conn, struct smc_sock, conn); 1132 /* conn->lnk not yet set in SMC_INIT state */ 1133 if (smc->sk.sk_state == SMC_INIT) 1134 continue; 1135 if (smc->sk.sk_state == SMC_CLOSED || 1136 smc->sk.sk_state == SMC_PEERCLOSEWAIT1 || 1137 smc->sk.sk_state == SMC_PEERCLOSEWAIT2 || 1138 smc->sk.sk_state == SMC_APPFINCLOSEWAIT || 1139 smc->sk.sk_state == SMC_APPCLOSEWAIT1 || 1140 smc->sk.sk_state == SMC_APPCLOSEWAIT2 || 1141 smc->sk.sk_state == SMC_PEERFINCLOSEWAIT || 1142 smc->sk.sk_state == SMC_PEERABORTWAIT || 1143 smc->sk.sk_state == SMC_PROCESSABORT) { 1144 spin_lock_bh(&conn->send_lock); 1145 smc_switch_link_and_count(conn, to_lnk); 1146 spin_unlock_bh(&conn->send_lock); 1147 continue; 1148 } 1149 sock_hold(&smc->sk); 1150 read_unlock_bh(&lgr->conns_lock); 1151 /* pre-fetch buffer outside of send_lock, might sleep */ 1152 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend); 1153 if (!rc) { 1154 /* avoid race with smcr_tx_sndbuf_nonempty() */ 1155 spin_lock_bh(&conn->send_lock); 1156 smc_switch_link_and_count(conn, to_lnk); 1157 rc = smc_switch_cursor(smc, pend, wr_buf); 1158 spin_unlock_bh(&conn->send_lock); 1159 } 1160 sock_put(&smc->sk); 1161 if (rc) 1162 goto err_out; 1163 goto again; 1164 } 1165 read_unlock_bh(&lgr->conns_lock); 1166 smc_wr_tx_link_put(to_lnk); 1167 return to_lnk; 1168 1169 err_out: 1170 smcr_link_down_cond_sched(to_lnk); 1171 smc_wr_tx_link_put(to_lnk); 1172 return NULL; 1173 } 1174 1175 static void smcr_buf_unuse(struct smc_buf_desc *buf_desc, bool is_rmb, 1176 struct smc_link_group *lgr) 1177 { 1178 struct rw_semaphore *lock; /* lock buffer list */ 1179 int rc; 1180 1181 if (is_rmb && buf_desc->is_conf_rkey && !list_empty(&lgr->list)) { 1182 /* unregister rmb with peer */ 1183 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY); 1184 if (!rc) { 1185 /* protect against smc_llc_cli_rkey_exchange() */ 1186 down_read(&lgr->llc_conf_mutex); 1187 smc_llc_do_delete_rkey(lgr, buf_desc); 1188 buf_desc->is_conf_rkey = false; 1189 up_read(&lgr->llc_conf_mutex); 1190 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl); 1191 } 1192 } 1193 1194 if (buf_desc->is_reg_err) { 1195 /* buf registration failed, reuse not possible */ 1196 lock = is_rmb ? &lgr->rmbs_lock : 1197 &lgr->sndbufs_lock; 1198 down_write(lock); 1199 smc_lgr_buf_list_del(lgr, is_rmb, buf_desc); 1200 up_write(lock); 1201 1202 smc_buf_free(lgr, is_rmb, buf_desc); 1203 } else { 1204 /* memzero_explicit provides potential memory barrier semantics */ 1205 memzero_explicit(buf_desc->cpu_addr, buf_desc->len); 1206 WRITE_ONCE(buf_desc->used, 0); 1207 } 1208 } 1209 1210 static void smcd_buf_detach(struct smc_connection *conn) 1211 { 1212 struct smcd_dev *smcd = conn->lgr->smcd; 1213 u64 peer_token = conn->peer_token; 1214 struct smc_buf_desc *buf_desc; 1215 1216 if (!conn->sndbuf_desc) 1217 return; 1218 1219 smc_ism_detach_dmb(smcd, peer_token); 1220 1221 buf_desc = conn->sndbuf_desc; 1222 conn->sndbuf_desc = NULL; 1223 kfree(buf_desc); 1224 } 1225 1226 static void smc_buf_unuse(struct smc_connection *conn, 1227 struct smc_link_group *lgr) 1228 { 1229 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 1230 bool is_smcd = lgr->is_smcd; 1231 int bufsize; 1232 1233 if (conn->sndbuf_desc) { 1234 bufsize = conn->sndbuf_desc->len; 1235 if (!is_smcd && conn->sndbuf_desc->is_vm) { 1236 smcr_buf_unuse(conn->sndbuf_desc, false, lgr); 1237 } else { 1238 memzero_explicit(conn->sndbuf_desc->cpu_addr, bufsize); 1239 WRITE_ONCE(conn->sndbuf_desc->used, 0); 1240 } 1241 SMC_STAT_RMB_SIZE(smc, is_smcd, false, false, bufsize); 1242 } 1243 if (conn->rmb_desc) { 1244 bufsize = conn->rmb_desc->len; 1245 if (!is_smcd) { 1246 smcr_buf_unuse(conn->rmb_desc, true, lgr); 1247 } else { 1248 bufsize += sizeof(struct smcd_cdc_msg); 1249 memzero_explicit(conn->rmb_desc->cpu_addr, bufsize); 1250 WRITE_ONCE(conn->rmb_desc->used, 0); 1251 } 1252 SMC_STAT_RMB_SIZE(smc, is_smcd, true, false, bufsize); 1253 } 1254 } 1255 1256 /* remove a finished connection from its link group */ 1257 void smc_conn_free(struct smc_connection *conn) 1258 { 1259 struct smc_link_group *lgr = conn->lgr; 1260 1261 if (!lgr || conn->freed) 1262 /* Connection has never been registered in a 1263 * link group, or has already been freed. 1264 */ 1265 return; 1266 1267 conn->freed = 1; 1268 if (!smc_conn_lgr_valid(conn)) 1269 /* Connection has already unregistered from 1270 * link group. 1271 */ 1272 goto lgr_put; 1273 1274 if (lgr->is_smcd) { 1275 smc_ism_unset_conn(conn); 1276 tasklet_kill(&conn->rx_tsklet); 1277 if (smc_ism_support_dmb_nocopy(lgr->smcd)) 1278 smcd_buf_detach(conn); 1279 } else { 1280 smc_cdc_wait_pend_tx_wr(conn); 1281 if (current_work() != &conn->abort_work) 1282 cancel_work_sync(&conn->abort_work); 1283 } 1284 if (!list_empty(&lgr->list)) { 1285 smc_buf_unuse(conn, lgr); /* allow buffer reuse */ 1286 smc_lgr_unregister_conn(conn); 1287 } 1288 1289 if (!lgr->conns_num) 1290 smc_lgr_schedule_free_work(lgr); 1291 lgr_put: 1292 if (!lgr->is_smcd) 1293 smcr_link_put(conn->lnk); /* link_hold in smc_conn_create() */ 1294 smc_lgr_put(lgr); /* lgr_hold in smc_conn_create() */ 1295 } 1296 1297 /* unregister a link from a buf_desc */ 1298 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb, 1299 struct smc_link *lnk) 1300 { 1301 if (is_rmb || buf_desc->is_vm) 1302 buf_desc->is_reg_mr[lnk->link_idx] = false; 1303 if (!buf_desc->is_map_ib[lnk->link_idx]) 1304 return; 1305 1306 if ((is_rmb || buf_desc->is_vm) && 1307 buf_desc->mr[lnk->link_idx]) { 1308 smc_ib_put_memory_region(buf_desc->mr[lnk->link_idx]); 1309 buf_desc->mr[lnk->link_idx] = NULL; 1310 } 1311 if (is_rmb) 1312 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE); 1313 else 1314 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE); 1315 1316 sg_free_table(&buf_desc->sgt[lnk->link_idx]); 1317 buf_desc->is_map_ib[lnk->link_idx] = false; 1318 } 1319 1320 /* unmap all buffers of lgr for a deleted link */ 1321 static void smcr_buf_unmap_lgr(struct smc_link *lnk) 1322 { 1323 struct smc_link_group *lgr = lnk->lgr; 1324 struct smc_buf_desc *buf_desc, *bf; 1325 int i; 1326 1327 for (i = 0; i < SMC_RMBE_SIZES; i++) { 1328 down_write(&lgr->rmbs_lock); 1329 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) 1330 smcr_buf_unmap_link(buf_desc, true, lnk); 1331 up_write(&lgr->rmbs_lock); 1332 1333 down_write(&lgr->sndbufs_lock); 1334 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], 1335 list) 1336 smcr_buf_unmap_link(buf_desc, false, lnk); 1337 up_write(&lgr->sndbufs_lock); 1338 } 1339 } 1340 1341 static void smcr_rtoken_clear_link(struct smc_link *lnk) 1342 { 1343 struct smc_link_group *lgr = lnk->lgr; 1344 int i; 1345 1346 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) { 1347 lgr->rtokens[i][lnk->link_idx].rkey = 0; 1348 lgr->rtokens[i][lnk->link_idx].dma_addr = 0; 1349 } 1350 } 1351 1352 static void __smcr_link_clear(struct smc_link *lnk) 1353 { 1354 struct smc_link_group *lgr = lnk->lgr; 1355 struct smc_ib_device *smcibdev; 1356 1357 smc_wr_free_link_mem(lnk); 1358 smc_ibdev_cnt_dec(lnk); 1359 put_device(&lnk->smcibdev->ibdev->dev); 1360 smcibdev = lnk->smcibdev; 1361 memset(lnk, 0, sizeof(struct smc_link)); 1362 lnk->state = SMC_LNK_UNUSED; 1363 if (!atomic_dec_return(&smcibdev->lnk_cnt)) 1364 wake_up(&smcibdev->lnks_deleted); 1365 smc_lgr_put(lgr); /* lgr_hold in smcr_link_init() */ 1366 } 1367 1368 /* must be called under lgr->llc_conf_mutex lock */ 1369 void smcr_link_clear(struct smc_link *lnk, bool log) 1370 { 1371 if (!lnk->lgr || lnk->clearing || 1372 lnk->state == SMC_LNK_UNUSED) 1373 return; 1374 lnk->clearing = 1; 1375 lnk->peer_qpn = 0; 1376 smc_llc_link_clear(lnk, log); 1377 smcr_buf_unmap_lgr(lnk); 1378 smcr_rtoken_clear_link(lnk); 1379 smc_ib_modify_qp_error(lnk); 1380 smc_wr_free_link(lnk); 1381 smc_ib_destroy_queue_pair(lnk); 1382 smc_ib_dealloc_protection_domain(lnk); 1383 smcr_link_put(lnk); /* theoretically last link_put */ 1384 } 1385 1386 void smcr_link_hold(struct smc_link *lnk) 1387 { 1388 refcount_inc(&lnk->refcnt); 1389 } 1390 1391 void smcr_link_put(struct smc_link *lnk) 1392 { 1393 if (refcount_dec_and_test(&lnk->refcnt)) 1394 __smcr_link_clear(lnk); 1395 } 1396 1397 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb, 1398 struct smc_buf_desc *buf_desc) 1399 { 1400 int i; 1401 1402 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) 1403 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]); 1404 1405 if (!buf_desc->is_vm && buf_desc->pages) 1406 __free_pages(buf_desc->pages, buf_desc->order); 1407 else if (buf_desc->is_vm && buf_desc->cpu_addr) 1408 vfree(buf_desc->cpu_addr); 1409 kfree(buf_desc); 1410 } 1411 1412 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb, 1413 struct smc_buf_desc *buf_desc) 1414 { 1415 if (is_dmb) { 1416 /* restore original buf len */ 1417 buf_desc->len += sizeof(struct smcd_cdc_msg); 1418 smc_ism_unregister_dmb(lgr->smcd, buf_desc); 1419 } else { 1420 kfree(buf_desc->cpu_addr); 1421 } 1422 kfree(buf_desc); 1423 } 1424 1425 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb, 1426 struct smc_buf_desc *buf_desc) 1427 { 1428 if (lgr->is_smcd) 1429 smcd_buf_free(lgr, is_rmb, buf_desc); 1430 else 1431 smcr_buf_free(lgr, is_rmb, buf_desc); 1432 } 1433 1434 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb) 1435 { 1436 struct smc_buf_desc *buf_desc, *bf_desc; 1437 struct list_head *buf_list; 1438 int i; 1439 1440 for (i = 0; i < SMC_RMBE_SIZES; i++) { 1441 if (is_rmb) 1442 buf_list = &lgr->rmbs[i]; 1443 else 1444 buf_list = &lgr->sndbufs[i]; 1445 list_for_each_entry_safe(buf_desc, bf_desc, buf_list, 1446 list) { 1447 smc_lgr_buf_list_del(lgr, is_rmb, buf_desc); 1448 smc_buf_free(lgr, is_rmb, buf_desc); 1449 } 1450 } 1451 } 1452 1453 static void smc_lgr_free_bufs(struct smc_link_group *lgr) 1454 { 1455 /* free send buffers */ 1456 __smc_lgr_free_bufs(lgr, false); 1457 /* free rmbs */ 1458 __smc_lgr_free_bufs(lgr, true); 1459 } 1460 1461 /* won't be freed until no one accesses to lgr anymore */ 1462 static void __smc_lgr_free(struct smc_link_group *lgr) 1463 { 1464 smc_lgr_free_bufs(lgr); 1465 if (lgr->is_smcd) { 1466 if (!atomic_dec_return(&lgr->smcd->lgr_cnt)) 1467 wake_up(&lgr->smcd->lgrs_deleted); 1468 } else { 1469 smc_wr_free_lgr_mem(lgr); 1470 if (!atomic_dec_return(&lgr_cnt)) 1471 wake_up(&lgrs_deleted); 1472 } 1473 kfree(lgr); 1474 } 1475 1476 /* remove a link group */ 1477 static void smc_lgr_free(struct smc_link_group *lgr) 1478 { 1479 int i; 1480 1481 if (!lgr->is_smcd) { 1482 down_write(&lgr->llc_conf_mutex); 1483 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 1484 if (lgr->lnk[i].state != SMC_LNK_UNUSED) 1485 smcr_link_clear(&lgr->lnk[i], false); 1486 } 1487 up_write(&lgr->llc_conf_mutex); 1488 smc_llc_lgr_clear(lgr); 1489 } 1490 1491 destroy_workqueue(lgr->tx_wq); 1492 if (lgr->is_smcd) { 1493 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id); 1494 put_device(&lgr->smcd->dibs->dev); 1495 } 1496 smc_lgr_put(lgr); /* theoretically last lgr_put */ 1497 } 1498 1499 void smc_lgr_hold(struct smc_link_group *lgr) 1500 { 1501 refcount_inc(&lgr->refcnt); 1502 } 1503 1504 void smc_lgr_put(struct smc_link_group *lgr) 1505 { 1506 if (refcount_dec_and_test(&lgr->refcnt)) 1507 __smc_lgr_free(lgr); 1508 } 1509 1510 static void smc_sk_wake_ups(struct smc_sock *smc) 1511 { 1512 smc->sk.sk_write_space(&smc->sk); 1513 smc->sk.sk_data_ready(&smc->sk); 1514 smc->sk.sk_state_change(&smc->sk); 1515 } 1516 1517 /* kill a connection */ 1518 static void smc_conn_kill(struct smc_connection *conn, bool soft) 1519 { 1520 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 1521 1522 if (conn->lgr->is_smcd && conn->lgr->peer_shutdown) 1523 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 1524 else 1525 smc_close_abort(conn); 1526 conn->killed = 1; 1527 smc->sk.sk_err = ECONNABORTED; 1528 smc_sk_wake_ups(smc); 1529 if (conn->lgr->is_smcd) { 1530 smc_ism_unset_conn(conn); 1531 if (soft) 1532 tasklet_kill(&conn->rx_tsklet); 1533 else 1534 tasklet_unlock_wait(&conn->rx_tsklet); 1535 if (smc_ism_support_dmb_nocopy(conn->lgr->smcd)) 1536 smcd_buf_detach(conn); 1537 } else { 1538 smc_cdc_wait_pend_tx_wr(conn); 1539 } 1540 smc_lgr_unregister_conn(conn); 1541 smc_close_active_abort(smc); 1542 } 1543 1544 static void smc_lgr_cleanup(struct smc_link_group *lgr) 1545 { 1546 if (lgr->is_smcd) { 1547 smc_ism_signal_shutdown(lgr); 1548 } else { 1549 u32 rsn = lgr->llc_termination_rsn; 1550 1551 if (!rsn) 1552 rsn = SMC_LLC_DEL_PROG_INIT_TERM; 1553 smc_llc_send_link_delete_all(lgr, false, rsn); 1554 smcr_lgr_link_deactivate_all(lgr); 1555 } 1556 } 1557 1558 /* terminate link group 1559 * @soft: true if link group shutdown can take its time 1560 * false if immediate link group shutdown is required 1561 */ 1562 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft) 1563 { 1564 struct smc_connection *conn; 1565 struct smc_sock *smc; 1566 struct rb_node *node; 1567 1568 if (lgr->terminating) 1569 return; /* lgr already terminating */ 1570 /* cancel free_work sync, will terminate when lgr->freeing is set */ 1571 cancel_delayed_work(&lgr->free_work); 1572 lgr->terminating = 1; 1573 1574 /* kill remaining link group connections */ 1575 read_lock_bh(&lgr->conns_lock); 1576 node = rb_first(&lgr->conns_all); 1577 while (node) { 1578 conn = rb_entry(node, struct smc_connection, alert_node); 1579 smc = container_of(conn, struct smc_sock, conn); 1580 sock_hold(&smc->sk); /* sock_put below */ 1581 read_unlock_bh(&lgr->conns_lock); 1582 lock_sock(&smc->sk); 1583 smc_conn_kill(conn, soft); 1584 release_sock(&smc->sk); 1585 sock_put(&smc->sk); /* sock_hold above */ 1586 read_lock_bh(&lgr->conns_lock); 1587 node = rb_first(&lgr->conns_all); 1588 } 1589 read_unlock_bh(&lgr->conns_lock); 1590 smc_lgr_cleanup(lgr); 1591 smc_lgr_free(lgr); 1592 } 1593 1594 /* unlink link group and schedule termination */ 1595 void smc_lgr_terminate_sched(struct smc_link_group *lgr) 1596 { 1597 spinlock_t *lgr_lock; 1598 1599 smc_lgr_list_head(lgr, &lgr_lock); 1600 spin_lock_bh(lgr_lock); 1601 if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) { 1602 spin_unlock_bh(lgr_lock); 1603 return; /* lgr already terminating */ 1604 } 1605 list_del_init(&lgr->list); 1606 lgr->freeing = 1; 1607 spin_unlock_bh(lgr_lock); 1608 schedule_work(&lgr->terminate_work); 1609 } 1610 1611 /* Called when peer lgr shutdown (regularly or abnormally) is received */ 1612 void smc_smcd_terminate(struct smcd_dev *dev, struct smcd_gid *peer_gid, 1613 unsigned short vlan) 1614 { 1615 struct smc_link_group *lgr, *l; 1616 LIST_HEAD(lgr_free_list); 1617 1618 /* run common cleanup function and build free list */ 1619 spin_lock_bh(&dev->lgr_lock); 1620 list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) { 1621 if ((!peer_gid->gid || 1622 (lgr->peer_gid.gid == peer_gid->gid && 1623 !smc_ism_is_emulated(dev) ? 1 : 1624 lgr->peer_gid.gid_ext == peer_gid->gid_ext)) && 1625 (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) { 1626 if (peer_gid->gid) /* peer triggered termination */ 1627 lgr->peer_shutdown = 1; 1628 list_move(&lgr->list, &lgr_free_list); 1629 lgr->freeing = 1; 1630 } 1631 } 1632 spin_unlock_bh(&dev->lgr_lock); 1633 1634 /* cancel the regular free workers and actually free lgrs */ 1635 list_for_each_entry_safe(lgr, l, &lgr_free_list, list) { 1636 list_del_init(&lgr->list); 1637 schedule_work(&lgr->terminate_work); 1638 } 1639 } 1640 1641 /* Called when an SMCD device is removed or the smc module is unloaded */ 1642 void smc_smcd_terminate_all(struct smcd_dev *smcd) 1643 { 1644 struct smc_link_group *lgr, *lg; 1645 LIST_HEAD(lgr_free_list); 1646 1647 spin_lock_bh(&smcd->lgr_lock); 1648 list_splice_init(&smcd->lgr_list, &lgr_free_list); 1649 list_for_each_entry(lgr, &lgr_free_list, list) 1650 lgr->freeing = 1; 1651 spin_unlock_bh(&smcd->lgr_lock); 1652 1653 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) { 1654 list_del_init(&lgr->list); 1655 __smc_lgr_terminate(lgr, false); 1656 } 1657 1658 if (atomic_read(&smcd->lgr_cnt)) 1659 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt)); 1660 } 1661 1662 /* Called when an SMCR device is removed or the smc module is unloaded. 1663 * If smcibdev is given, all SMCR link groups using this device are terminated. 1664 * If smcibdev is NULL, all SMCR link groups are terminated. 1665 */ 1666 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev) 1667 { 1668 struct smc_link_group *lgr, *lg; 1669 LIST_HEAD(lgr_free_list); 1670 int i; 1671 1672 spin_lock_bh(&smc_lgr_list.lock); 1673 if (!smcibdev) { 1674 list_splice_init(&smc_lgr_list.list, &lgr_free_list); 1675 list_for_each_entry(lgr, &lgr_free_list, list) 1676 lgr->freeing = 1; 1677 } else { 1678 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) { 1679 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 1680 if (lgr->lnk[i].smcibdev == smcibdev) 1681 smcr_link_down_cond_sched(&lgr->lnk[i]); 1682 } 1683 } 1684 } 1685 spin_unlock_bh(&smc_lgr_list.lock); 1686 1687 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) { 1688 list_del_init(&lgr->list); 1689 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM); 1690 __smc_lgr_terminate(lgr, false); 1691 } 1692 1693 if (smcibdev) { 1694 if (atomic_read(&smcibdev->lnk_cnt)) 1695 wait_event(smcibdev->lnks_deleted, 1696 !atomic_read(&smcibdev->lnk_cnt)); 1697 } else { 1698 if (atomic_read(&lgr_cnt)) 1699 wait_event(lgrs_deleted, !atomic_read(&lgr_cnt)); 1700 } 1701 } 1702 1703 /* set new lgr type and clear all asymmetric link tagging */ 1704 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type) 1705 { 1706 char *lgr_type = ""; 1707 int i; 1708 1709 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) 1710 if (smc_link_usable(&lgr->lnk[i])) 1711 lgr->lnk[i].link_is_asym = false; 1712 if (lgr->type == new_type) 1713 return; 1714 lgr->type = new_type; 1715 1716 switch (lgr->type) { 1717 case SMC_LGR_NONE: 1718 lgr_type = "NONE"; 1719 break; 1720 case SMC_LGR_SINGLE: 1721 lgr_type = "SINGLE"; 1722 break; 1723 case SMC_LGR_SYMMETRIC: 1724 lgr_type = "SYMMETRIC"; 1725 break; 1726 case SMC_LGR_ASYMMETRIC_PEER: 1727 lgr_type = "ASYMMETRIC_PEER"; 1728 break; 1729 case SMC_LGR_ASYMMETRIC_LOCAL: 1730 lgr_type = "ASYMMETRIC_LOCAL"; 1731 break; 1732 } 1733 pr_warn_ratelimited("smc: SMC-R lg %*phN net %llu state changed: " 1734 "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id, 1735 lgr->net->net_cookie, lgr_type, lgr->pnet_id); 1736 } 1737 1738 /* set new lgr type and tag a link as asymmetric */ 1739 void smcr_lgr_set_type_asym(struct smc_link_group *lgr, 1740 enum smc_lgr_type new_type, int asym_lnk_idx) 1741 { 1742 smcr_lgr_set_type(lgr, new_type); 1743 lgr->lnk[asym_lnk_idx].link_is_asym = true; 1744 } 1745 1746 /* abort connection, abort_work scheduled from tasklet context */ 1747 static void smc_conn_abort_work(struct work_struct *work) 1748 { 1749 struct smc_connection *conn = container_of(work, 1750 struct smc_connection, 1751 abort_work); 1752 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 1753 1754 lock_sock(&smc->sk); 1755 smc_conn_kill(conn, true); 1756 release_sock(&smc->sk); 1757 sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */ 1758 } 1759 1760 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport) 1761 { 1762 struct smc_link_group *lgr, *n; 1763 1764 spin_lock_bh(&smc_lgr_list.lock); 1765 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) { 1766 struct smc_link *link; 1767 1768 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id, 1769 SMC_MAX_PNETID_LEN) || 1770 lgr->type == SMC_LGR_SYMMETRIC || 1771 lgr->type == SMC_LGR_ASYMMETRIC_PEER || 1772 !rdma_dev_access_netns(smcibdev->ibdev, lgr->net)) 1773 continue; 1774 1775 if (lgr->type == SMC_LGR_SINGLE && lgr->max_links <= 1) 1776 continue; 1777 1778 /* trigger local add link processing */ 1779 link = smc_llc_usable_link(lgr); 1780 if (link) 1781 smc_llc_add_link_local(link); 1782 } 1783 spin_unlock_bh(&smc_lgr_list.lock); 1784 } 1785 1786 /* link is down - switch connections to alternate link, 1787 * must be called under lgr->llc_conf_mutex lock 1788 */ 1789 static void smcr_link_down(struct smc_link *lnk) 1790 { 1791 struct smc_link_group *lgr = lnk->lgr; 1792 struct smc_link *to_lnk; 1793 int del_link_id; 1794 1795 if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list)) 1796 return; 1797 1798 to_lnk = smc_switch_conns(lgr, lnk, true); 1799 if (!to_lnk) { /* no backup link available */ 1800 smcr_link_clear(lnk, true); 1801 return; 1802 } 1803 smcr_lgr_set_type(lgr, SMC_LGR_SINGLE); 1804 del_link_id = lnk->link_id; 1805 1806 if (lgr->role == SMC_SERV) { 1807 /* trigger local delete link processing */ 1808 smc_llc_srv_delete_link_local(to_lnk, del_link_id); 1809 } else { 1810 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) { 1811 /* another llc task is ongoing */ 1812 up_write(&lgr->llc_conf_mutex); 1813 wait_event_timeout(lgr->llc_flow_waiter, 1814 (list_empty(&lgr->list) || 1815 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE), 1816 SMC_LLC_WAIT_TIME); 1817 down_write(&lgr->llc_conf_mutex); 1818 } 1819 if (!list_empty(&lgr->list)) { 1820 smc_llc_send_delete_link(to_lnk, del_link_id, 1821 SMC_LLC_REQ, true, 1822 SMC_LLC_DEL_LOST_PATH); 1823 smcr_link_clear(lnk, true); 1824 } 1825 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */ 1826 } 1827 } 1828 1829 /* must be called under lgr->llc_conf_mutex lock */ 1830 void smcr_link_down_cond(struct smc_link *lnk) 1831 { 1832 if (smc_link_downing(&lnk->state)) { 1833 trace_smcr_link_down(lnk, __builtin_return_address(0)); 1834 smcr_link_down(lnk); 1835 } 1836 } 1837 1838 /* will get the lgr->llc_conf_mutex lock */ 1839 void smcr_link_down_cond_sched(struct smc_link *lnk) 1840 { 1841 if (smc_link_downing(&lnk->state)) { 1842 trace_smcr_link_down(lnk, __builtin_return_address(0)); 1843 smcr_link_hold(lnk); /* smcr_link_put in link_down_wrk */ 1844 if (!schedule_work(&lnk->link_down_wrk)) 1845 smcr_link_put(lnk); 1846 } 1847 } 1848 1849 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport) 1850 { 1851 struct smc_link_group *lgr, *n; 1852 int i; 1853 1854 spin_lock_bh(&smc_lgr_list.lock); 1855 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) { 1856 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id, 1857 SMC_MAX_PNETID_LEN)) 1858 continue; /* lgr is not affected */ 1859 if (list_empty(&lgr->list)) 1860 continue; 1861 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 1862 struct smc_link *lnk = &lgr->lnk[i]; 1863 1864 if (smc_link_usable(lnk) && 1865 lnk->smcibdev == smcibdev && lnk->ibport == ibport) 1866 smcr_link_down_cond_sched(lnk); 1867 } 1868 } 1869 spin_unlock_bh(&smc_lgr_list.lock); 1870 } 1871 1872 static void smc_link_down_work(struct work_struct *work) 1873 { 1874 struct smc_link *link = container_of(work, struct smc_link, 1875 link_down_wrk); 1876 struct smc_link_group *lgr = link->lgr; 1877 1878 if (list_empty(&lgr->list)) 1879 goto out; 1880 wake_up_all(&lgr->llc_msg_waiter); 1881 down_write(&lgr->llc_conf_mutex); 1882 smcr_link_down(link); 1883 up_write(&lgr->llc_conf_mutex); 1884 1885 out: 1886 smcr_link_put(link); /* smcr_link_hold by schedulers of link_down_work */ 1887 } 1888 1889 static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev, 1890 struct netdev_nested_priv *priv) 1891 { 1892 unsigned short *vlan_id = (unsigned short *)priv->data; 1893 1894 if (is_vlan_dev(lower_dev)) { 1895 *vlan_id = vlan_dev_vlan_id(lower_dev); 1896 return 1; 1897 } 1898 1899 return 0; 1900 } 1901 1902 /* Determine vlan of internal TCP socket. */ 1903 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini) 1904 { 1905 struct netdev_nested_priv priv; 1906 struct net_device *ndev; 1907 struct dst_entry *dst; 1908 int rc = 0; 1909 1910 ini->vlan_id = 0; 1911 1912 rcu_read_lock(); 1913 1914 dst = __sk_dst_get(clcsock->sk); 1915 ndev = dst ? dst_dev_rcu(dst) : NULL; 1916 if (!ndev) { 1917 rc = -ENODEV; 1918 goto out; 1919 } 1920 1921 if (is_vlan_dev(ndev)) { 1922 ini->vlan_id = vlan_dev_vlan_id(ndev); 1923 goto out; 1924 } 1925 1926 priv.data = (void *)&ini->vlan_id; 1927 netdev_walk_all_lower_dev_rcu(ndev, smc_vlan_by_tcpsk_walk, &priv); 1928 out: 1929 rcu_read_unlock(); 1930 1931 return rc; 1932 } 1933 1934 static bool smcr_lgr_match(struct smc_link_group *lgr, u8 smcr_version, 1935 u8 peer_systemid[], 1936 u8 peer_gid[], 1937 u8 peer_mac_v1[], 1938 enum smc_lgr_role role, u32 clcqpn, 1939 struct net *net) 1940 { 1941 struct smc_link *lnk; 1942 int i; 1943 1944 if (memcmp(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN) || 1945 lgr->role != role) 1946 return false; 1947 1948 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 1949 lnk = &lgr->lnk[i]; 1950 1951 if (!smc_link_active(lnk)) 1952 continue; 1953 /* use verbs API to check netns, instead of lgr->net */ 1954 if (!rdma_dev_access_netns(lnk->smcibdev->ibdev, net)) 1955 return false; 1956 if ((lgr->role == SMC_SERV || lnk->peer_qpn == clcqpn) && 1957 !memcmp(lnk->peer_gid, peer_gid, SMC_GID_SIZE) && 1958 (smcr_version == SMC_V2 || 1959 !memcmp(lnk->peer_mac, peer_mac_v1, ETH_ALEN))) 1960 return true; 1961 } 1962 return false; 1963 } 1964 1965 static bool smcd_lgr_match(struct smc_link_group *lgr, 1966 struct smcd_dev *smcismdev, 1967 struct smcd_gid *peer_gid) 1968 { 1969 if (lgr->peer_gid.gid != peer_gid->gid || 1970 lgr->smcd != smcismdev) 1971 return false; 1972 1973 if (smc_ism_is_emulated(smcismdev) && 1974 lgr->peer_gid.gid_ext != peer_gid->gid_ext) 1975 return false; 1976 1977 return true; 1978 } 1979 1980 /* create a new SMC connection (and a new link group if necessary) */ 1981 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini) 1982 { 1983 struct smc_connection *conn = &smc->conn; 1984 struct net *net = sock_net(&smc->sk); 1985 struct list_head *lgr_list; 1986 struct smc_link_group *lgr; 1987 enum smc_lgr_role role; 1988 spinlock_t *lgr_lock; 1989 int rc = 0; 1990 1991 lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list : 1992 &smc_lgr_list.list; 1993 lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock : 1994 &smc_lgr_list.lock; 1995 ini->first_contact_local = 1; 1996 role = smc->listen_smc ? SMC_SERV : SMC_CLNT; 1997 if (role == SMC_CLNT && ini->first_contact_peer) 1998 /* create new link group as well */ 1999 goto create; 2000 2001 /* determine if an existing link group can be reused */ 2002 spin_lock_bh(lgr_lock); 2003 list_for_each_entry(lgr, lgr_list, list) { 2004 write_lock_bh(&lgr->conns_lock); 2005 if ((ini->is_smcd ? 2006 smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected], 2007 &ini->ism_peer_gid[ini->ism_selected]) : 2008 smcr_lgr_match(lgr, ini->smcr_version, 2009 ini->peer_systemid, 2010 ini->peer_gid, ini->peer_mac, role, 2011 ini->ib_clcqpn, net)) && 2012 !lgr->sync_err && 2013 (ini->smcd_version == SMC_V2 || 2014 lgr->vlan_id == ini->vlan_id) && 2015 (role == SMC_CLNT || ini->is_smcd || 2016 (lgr->conns_num < lgr->max_conns && 2017 !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) { 2018 /* link group found */ 2019 ini->first_contact_local = 0; 2020 conn->lgr = lgr; 2021 rc = smc_lgr_register_conn(conn, false); 2022 write_unlock_bh(&lgr->conns_lock); 2023 if (!rc && delayed_work_pending(&lgr->free_work)) 2024 cancel_delayed_work(&lgr->free_work); 2025 break; 2026 } 2027 write_unlock_bh(&lgr->conns_lock); 2028 } 2029 spin_unlock_bh(lgr_lock); 2030 if (rc) 2031 return rc; 2032 2033 if (role == SMC_CLNT && !ini->first_contact_peer && 2034 ini->first_contact_local) { 2035 /* Server reuses a link group, but Client wants to start 2036 * a new one 2037 * send out_of_sync decline, reason synchr. error 2038 */ 2039 return SMC_CLC_DECL_SYNCERR; 2040 } 2041 2042 create: 2043 if (ini->first_contact_local) { 2044 rc = smc_lgr_create(smc, ini); 2045 if (rc) 2046 goto out; 2047 lgr = conn->lgr; 2048 write_lock_bh(&lgr->conns_lock); 2049 rc = smc_lgr_register_conn(conn, true); 2050 write_unlock_bh(&lgr->conns_lock); 2051 if (rc) { 2052 smc_lgr_cleanup_early(lgr); 2053 goto out; 2054 } 2055 } 2056 smc_lgr_hold(conn->lgr); /* lgr_put in smc_conn_free() */ 2057 if (!conn->lgr->is_smcd) 2058 smcr_link_hold(conn->lnk); /* link_put in smc_conn_free() */ 2059 conn->freed = 0; 2060 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE; 2061 conn->local_tx_ctrl.len = SMC_WR_TX_SIZE; 2062 conn->urg_state = SMC_URG_READ; 2063 init_waitqueue_head(&conn->cdc_pend_tx_wq); 2064 INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work); 2065 if (ini->is_smcd) { 2066 conn->rx_off = sizeof(struct smcd_cdc_msg); 2067 smcd_cdc_rx_init(conn); /* init tasklet for this conn */ 2068 } else { 2069 conn->rx_off = 0; 2070 } 2071 #ifndef KERNEL_HAS_ATOMIC64 2072 spin_lock_init(&conn->acurs_lock); 2073 #endif 2074 2075 out: 2076 return rc; 2077 } 2078 2079 #define SMCD_DMBE_SIZES 6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */ 2080 #define SMCR_RMBE_SIZES 15 /* 0 -> 16KB, 1 -> 32KB, .. 15 -> 512MB */ 2081 2082 /* convert the RMB size into the compressed notation (minimum 16K, see 2083 * SMCD/R_DMBE_SIZES. 2084 * In contrast to plain ilog2, this rounds towards the next power of 2, 2085 * so the socket application gets at least its desired sndbuf / rcvbuf size. 2086 */ 2087 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb) 2088 { 2089 u8 compressed; 2090 2091 if (size <= SMC_BUF_MIN_SIZE) 2092 return 0; 2093 2094 size = (size - 1) >> 14; /* convert to 16K multiple */ 2095 compressed = min_t(u8, ilog2(size) + 1, 2096 is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES); 2097 2098 #ifdef CONFIG_ARCH_NO_SG_CHAIN 2099 if (!is_smcd && is_rmb) 2100 /* RMBs are backed by & limited to max size of scatterlists */ 2101 compressed = min_t(u8, compressed, ilog2((SG_MAX_SINGLE_ALLOC * PAGE_SIZE) >> 14)); 2102 #endif 2103 2104 return compressed; 2105 } 2106 2107 /* convert the RMB size from compressed notation into integer */ 2108 int smc_uncompress_bufsize(u8 compressed) 2109 { 2110 u32 size; 2111 2112 size = 0x00000001 << (((int)compressed) + 14); 2113 return (int)size; 2114 } 2115 2116 /* try to reuse a sndbuf or rmb description slot for a certain 2117 * buffer size; if not available, return NULL 2118 */ 2119 static struct smc_buf_desc *smc_buf_get_slot(struct rw_semaphore *lock, 2120 struct list_head *buf_list) 2121 { 2122 struct smc_buf_desc *buf_slot; 2123 2124 down_read(lock); 2125 list_for_each_entry(buf_slot, buf_list, list) { 2126 if (cmpxchg(&buf_slot->used, 0, 1) == 0) { 2127 up_read(lock); 2128 return buf_slot; 2129 } 2130 } 2131 up_read(lock); 2132 return NULL; 2133 } 2134 2135 /* one of the conditions for announcing a receiver's current window size is 2136 * that it "results in a minimum increase in the window size of 10% of the 2137 * receive buffer space" [RFC7609] 2138 */ 2139 static inline int smc_rmb_wnd_update_limit(int rmbe_size) 2140 { 2141 return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2); 2142 } 2143 2144 /* map an buf to a link */ 2145 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb, 2146 struct smc_link *lnk) 2147 { 2148 int rc, i, nents, offset, buf_size, size, access_flags; 2149 struct scatterlist *sg; 2150 void *buf; 2151 2152 if (buf_desc->is_map_ib[lnk->link_idx]) 2153 return 0; 2154 2155 if (buf_desc->is_vm) { 2156 buf = buf_desc->cpu_addr; 2157 buf_size = buf_desc->len; 2158 offset = offset_in_page(buf_desc->cpu_addr); 2159 nents = PAGE_ALIGN(buf_size + offset) / PAGE_SIZE; 2160 } else { 2161 nents = 1; 2162 } 2163 2164 rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], nents, GFP_KERNEL); 2165 if (rc) 2166 return rc; 2167 2168 if (buf_desc->is_vm) { 2169 /* virtually contiguous buffer */ 2170 for_each_sg(buf_desc->sgt[lnk->link_idx].sgl, sg, nents, i) { 2171 size = min_t(int, PAGE_SIZE - offset, buf_size); 2172 sg_set_page(sg, vmalloc_to_page(buf), size, offset); 2173 buf += size; 2174 buf_size -= size; 2175 offset = 0; 2176 } 2177 } else { 2178 /* physically contiguous buffer */ 2179 sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl, 2180 buf_desc->cpu_addr, buf_desc->len); 2181 } 2182 2183 /* map sg table to DMA address */ 2184 rc = smc_ib_buf_map_sg(lnk, buf_desc, 2185 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE); 2186 /* SMC protocol depends on mapping to one DMA address only */ 2187 if (rc != nents) { 2188 rc = -EAGAIN; 2189 goto free_table; 2190 } 2191 2192 buf_desc->is_dma_need_sync |= 2193 smc_ib_is_sg_need_sync(lnk, buf_desc) << lnk->link_idx; 2194 2195 if (is_rmb || buf_desc->is_vm) { 2196 /* create a new memory region for the RMB or vzalloced sndbuf */ 2197 access_flags = is_rmb ? 2198 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE : 2199 IB_ACCESS_LOCAL_WRITE; 2200 2201 rc = smc_ib_get_memory_region(lnk->roce_pd, access_flags, 2202 buf_desc, lnk->link_idx); 2203 if (rc) 2204 goto buf_unmap; 2205 smc_ib_sync_sg_for_device(lnk, buf_desc, 2206 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE); 2207 } 2208 buf_desc->is_map_ib[lnk->link_idx] = true; 2209 return 0; 2210 2211 buf_unmap: 2212 smc_ib_buf_unmap_sg(lnk, buf_desc, 2213 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE); 2214 free_table: 2215 sg_free_table(&buf_desc->sgt[lnk->link_idx]); 2216 return rc; 2217 } 2218 2219 /* register a new buf on IB device, rmb or vzalloced sndbuf 2220 * must be called under lgr->llc_conf_mutex lock 2221 */ 2222 int smcr_link_reg_buf(struct smc_link *link, struct smc_buf_desc *buf_desc) 2223 { 2224 if (list_empty(&link->lgr->list)) 2225 return -ENOLINK; 2226 if (!buf_desc->is_reg_mr[link->link_idx]) { 2227 /* register memory region for new buf */ 2228 if (buf_desc->is_vm) 2229 buf_desc->mr[link->link_idx]->iova = 2230 (uintptr_t)buf_desc->cpu_addr; 2231 if (smc_wr_reg_send(link, buf_desc->mr[link->link_idx])) { 2232 buf_desc->is_reg_err = true; 2233 return -EFAULT; 2234 } 2235 buf_desc->is_reg_mr[link->link_idx] = true; 2236 } 2237 return 0; 2238 } 2239 2240 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct rw_semaphore *lock, 2241 struct list_head *lst, bool is_rmb) 2242 { 2243 struct smc_buf_desc *buf_desc, *bf; 2244 int rc = 0; 2245 2246 down_write(lock); 2247 list_for_each_entry_safe(buf_desc, bf, lst, list) { 2248 if (!buf_desc->used) 2249 continue; 2250 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk); 2251 if (rc) 2252 goto out; 2253 } 2254 out: 2255 up_write(lock); 2256 return rc; 2257 } 2258 2259 /* map all used buffers of lgr for a new link */ 2260 int smcr_buf_map_lgr(struct smc_link *lnk) 2261 { 2262 struct smc_link_group *lgr = lnk->lgr; 2263 int i, rc = 0; 2264 2265 for (i = 0; i < SMC_RMBE_SIZES; i++) { 2266 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock, 2267 &lgr->rmbs[i], true); 2268 if (rc) 2269 return rc; 2270 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock, 2271 &lgr->sndbufs[i], false); 2272 if (rc) 2273 return rc; 2274 } 2275 return 0; 2276 } 2277 2278 /* register all used buffers of lgr for a new link, 2279 * must be called under lgr->llc_conf_mutex lock 2280 */ 2281 int smcr_buf_reg_lgr(struct smc_link *lnk) 2282 { 2283 struct smc_link_group *lgr = lnk->lgr; 2284 struct smc_buf_desc *buf_desc, *bf; 2285 int i, rc = 0; 2286 2287 /* reg all RMBs for a new link */ 2288 down_write(&lgr->rmbs_lock); 2289 for (i = 0; i < SMC_RMBE_SIZES; i++) { 2290 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) { 2291 if (!buf_desc->used) 2292 continue; 2293 rc = smcr_link_reg_buf(lnk, buf_desc); 2294 if (rc) { 2295 up_write(&lgr->rmbs_lock); 2296 return rc; 2297 } 2298 } 2299 } 2300 up_write(&lgr->rmbs_lock); 2301 2302 if (lgr->buf_type == SMCR_PHYS_CONT_BUFS) 2303 return rc; 2304 2305 /* reg all vzalloced sndbufs for a new link */ 2306 down_write(&lgr->sndbufs_lock); 2307 for (i = 0; i < SMC_RMBE_SIZES; i++) { 2308 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], list) { 2309 if (!buf_desc->used || !buf_desc->is_vm) 2310 continue; 2311 rc = smcr_link_reg_buf(lnk, buf_desc); 2312 if (rc) { 2313 up_write(&lgr->sndbufs_lock); 2314 return rc; 2315 } 2316 } 2317 } 2318 up_write(&lgr->sndbufs_lock); 2319 return rc; 2320 } 2321 2322 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr, 2323 int bufsize) 2324 { 2325 struct smc_buf_desc *buf_desc; 2326 2327 /* try to alloc a new buffer */ 2328 buf_desc = kzalloc_obj(*buf_desc); 2329 if (!buf_desc) 2330 return ERR_PTR(-ENOMEM); 2331 2332 switch (lgr->buf_type) { 2333 case SMCR_PHYS_CONT_BUFS: 2334 case SMCR_MIXED_BUFS: 2335 buf_desc->order = get_order(bufsize); 2336 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN | 2337 __GFP_NOMEMALLOC | __GFP_COMP | 2338 __GFP_NORETRY | __GFP_ZERO, 2339 buf_desc->order); 2340 if (buf_desc->pages) { 2341 buf_desc->cpu_addr = 2342 (void *)page_address(buf_desc->pages); 2343 buf_desc->len = bufsize; 2344 buf_desc->is_vm = false; 2345 break; 2346 } 2347 if (lgr->buf_type == SMCR_PHYS_CONT_BUFS) 2348 goto out; 2349 fallthrough; // try virtually contiguous buf 2350 case SMCR_VIRT_CONT_BUFS: 2351 buf_desc->order = get_order(bufsize); 2352 buf_desc->cpu_addr = vzalloc(PAGE_SIZE << buf_desc->order); 2353 if (!buf_desc->cpu_addr) 2354 goto out; 2355 buf_desc->pages = NULL; 2356 buf_desc->len = bufsize; 2357 buf_desc->is_vm = true; 2358 break; 2359 } 2360 return buf_desc; 2361 2362 out: 2363 kfree(buf_desc); 2364 return ERR_PTR(-EAGAIN); 2365 } 2366 2367 /* map buf_desc on all usable links, 2368 * unused buffers stay mapped as long as the link is up 2369 */ 2370 static int smcr_buf_map_usable_links(struct smc_link_group *lgr, 2371 struct smc_buf_desc *buf_desc, bool is_rmb) 2372 { 2373 int i, rc = 0, cnt = 0; 2374 2375 /* protect against parallel link reconfiguration */ 2376 down_read(&lgr->llc_conf_mutex); 2377 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 2378 struct smc_link *lnk = &lgr->lnk[i]; 2379 2380 if (!smc_link_usable(lnk)) 2381 continue; 2382 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) { 2383 rc = -ENOMEM; 2384 goto out; 2385 } 2386 cnt++; 2387 } 2388 out: 2389 up_read(&lgr->llc_conf_mutex); 2390 if (!rc && !cnt) 2391 rc = -EINVAL; 2392 return rc; 2393 } 2394 2395 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr, 2396 bool is_dmb, int bufsize) 2397 { 2398 struct smc_buf_desc *buf_desc; 2399 int rc; 2400 2401 /* try to alloc a new DMB */ 2402 buf_desc = kzalloc_obj(*buf_desc); 2403 if (!buf_desc) 2404 return ERR_PTR(-ENOMEM); 2405 if (is_dmb) { 2406 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc); 2407 if (rc) { 2408 kfree(buf_desc); 2409 if (rc == -ENOMEM) 2410 return ERR_PTR(-EAGAIN); 2411 if (rc == -ENOSPC) 2412 return ERR_PTR(-ENOSPC); 2413 return ERR_PTR(-EIO); 2414 } 2415 buf_desc->pages = virt_to_page(buf_desc->cpu_addr); 2416 /* CDC header stored in buf. So, pretend it was smaller */ 2417 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg); 2418 } else { 2419 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL | 2420 __GFP_NOWARN | __GFP_NORETRY | 2421 __GFP_NOMEMALLOC); 2422 if (!buf_desc->cpu_addr) { 2423 kfree(buf_desc); 2424 return ERR_PTR(-EAGAIN); 2425 } 2426 buf_desc->len = bufsize; 2427 } 2428 return buf_desc; 2429 } 2430 2431 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb) 2432 { 2433 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM); 2434 struct smc_connection *conn = &smc->conn; 2435 struct smc_link_group *lgr = conn->lgr; 2436 struct list_head *buf_list; 2437 int bufsize, bufsize_comp; 2438 struct rw_semaphore *lock; /* lock buffer list */ 2439 bool is_dgraded = false; 2440 2441 if (is_rmb) 2442 /* use socket recv buffer size (w/o overhead) as start value */ 2443 bufsize = smc->sk.sk_rcvbuf / 2; 2444 else 2445 /* use socket send buffer size (w/o overhead) as start value */ 2446 bufsize = smc->sk.sk_sndbuf / 2; 2447 2448 /* limit bufsize for physically contiguous buffers */ 2449 if (!is_smcd && lgr->buf_type == SMCR_PHYS_CONT_BUFS) 2450 bufsize = min_t(int, bufsize, PAGE_SIZE << MAX_PAGE_ORDER); 2451 2452 for (bufsize_comp = smc_compress_bufsize(bufsize, is_smcd, is_rmb); 2453 bufsize_comp >= 0; bufsize_comp--) { 2454 if (is_rmb) { 2455 lock = &lgr->rmbs_lock; 2456 buf_list = &lgr->rmbs[bufsize_comp]; 2457 } else { 2458 lock = &lgr->sndbufs_lock; 2459 buf_list = &lgr->sndbufs[bufsize_comp]; 2460 } 2461 bufsize = smc_uncompress_bufsize(bufsize_comp); 2462 2463 /* check for reusable slot in the link group */ 2464 buf_desc = smc_buf_get_slot(lock, buf_list); 2465 if (buf_desc) { 2466 buf_desc->is_dma_need_sync = 0; 2467 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, true, bufsize); 2468 SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb); 2469 break; /* found reusable slot */ 2470 } 2471 2472 if (is_smcd) 2473 buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize); 2474 else 2475 buf_desc = smcr_new_buf_create(lgr, bufsize); 2476 2477 if (PTR_ERR(buf_desc) == -ENOMEM) 2478 break; 2479 if (IS_ERR(buf_desc)) { 2480 if (!is_dgraded) { 2481 is_dgraded = true; 2482 SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb); 2483 } 2484 continue; 2485 } 2486 2487 SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb); 2488 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, true, bufsize); 2489 buf_desc->used = 1; 2490 down_write(lock); 2491 smc_lgr_buf_list_add(lgr, is_rmb, buf_list, buf_desc); 2492 up_write(lock); 2493 break; /* found */ 2494 } 2495 2496 if (IS_ERR(buf_desc)) 2497 return PTR_ERR(buf_desc); 2498 2499 if (!is_smcd) { 2500 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) { 2501 smcr_buf_unuse(buf_desc, is_rmb, lgr); 2502 return -ENOMEM; 2503 } 2504 } 2505 2506 if (is_rmb) { 2507 conn->rmb_desc = buf_desc; 2508 conn->rmbe_size_comp = bufsize_comp; 2509 smc->sk.sk_rcvbuf = bufsize * 2; 2510 atomic_set(&conn->bytes_to_rcv, 0); 2511 conn->rmbe_update_limit = 2512 smc_rmb_wnd_update_limit(buf_desc->len); 2513 if (is_smcd) 2514 smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */ 2515 } else { 2516 conn->sndbuf_desc = buf_desc; 2517 smc->sk.sk_sndbuf = bufsize * 2; 2518 atomic_set(&conn->sndbuf_space, bufsize); 2519 } 2520 return 0; 2521 } 2522 2523 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn) 2524 { 2525 if (!conn->sndbuf_desc->is_dma_need_sync) 2526 return; 2527 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd || 2528 !smc_link_active(conn->lnk)) 2529 return; 2530 smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE); 2531 } 2532 2533 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn) 2534 { 2535 int i; 2536 2537 if (!conn->rmb_desc->is_dma_need_sync) 2538 return; 2539 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd) 2540 return; 2541 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 2542 if (!smc_link_active(&conn->lgr->lnk[i])) 2543 continue; 2544 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc, 2545 DMA_FROM_DEVICE); 2546 } 2547 } 2548 2549 /* create the send and receive buffer for an SMC socket; 2550 * receive buffers are called RMBs; 2551 * (even though the SMC protocol allows more than one RMB-element per RMB, 2552 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an 2553 * extra RMB for every connection in a link group 2554 */ 2555 int smc_buf_create(struct smc_sock *smc, bool is_smcd) 2556 { 2557 int rc; 2558 2559 /* create send buffer */ 2560 if (is_smcd && 2561 smc_ism_support_dmb_nocopy(smc->conn.lgr->smcd)) 2562 goto create_rmb; 2563 2564 rc = __smc_buf_create(smc, is_smcd, false); 2565 if (rc) 2566 return rc; 2567 2568 create_rmb: 2569 /* create rmb */ 2570 rc = __smc_buf_create(smc, is_smcd, true); 2571 if (rc && smc->conn.sndbuf_desc) { 2572 down_write(&smc->conn.lgr->sndbufs_lock); 2573 smc_lgr_buf_list_del(smc->conn.lgr, false, 2574 smc->conn.sndbuf_desc); 2575 up_write(&smc->conn.lgr->sndbufs_lock); 2576 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc); 2577 smc->conn.sndbuf_desc = NULL; 2578 } 2579 return rc; 2580 } 2581 2582 int smcd_buf_attach(struct smc_sock *smc) 2583 { 2584 struct smc_connection *conn = &smc->conn; 2585 struct smcd_dev *smcd = conn->lgr->smcd; 2586 u64 peer_token = conn->peer_token; 2587 struct smc_buf_desc *buf_desc; 2588 int rc; 2589 2590 buf_desc = kzalloc_obj(*buf_desc); 2591 if (!buf_desc) 2592 return -ENOMEM; 2593 2594 /* The ghost sndbuf_desc describes the same memory region as 2595 * peer RMB. Its lifecycle is consistent with the connection's 2596 * and it will be freed with the connections instead of the 2597 * link group. 2598 */ 2599 rc = smc_ism_attach_dmb(smcd, peer_token, buf_desc); 2600 if (rc) 2601 goto free; 2602 2603 smc->sk.sk_sndbuf = buf_desc->len; 2604 buf_desc->cpu_addr = 2605 (u8 *)buf_desc->cpu_addr + sizeof(struct smcd_cdc_msg); 2606 buf_desc->len -= sizeof(struct smcd_cdc_msg); 2607 conn->sndbuf_desc = buf_desc; 2608 conn->sndbuf_desc->used = 1; 2609 atomic_set(&conn->sndbuf_space, conn->sndbuf_desc->len); 2610 return 0; 2611 2612 free: 2613 kfree(buf_desc); 2614 return rc; 2615 } 2616 2617 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr) 2618 { 2619 int i; 2620 2621 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) { 2622 if (!test_and_set_bit(i, lgr->rtokens_used_mask)) 2623 return i; 2624 } 2625 return -ENOSPC; 2626 } 2627 2628 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx, 2629 u32 rkey) 2630 { 2631 int i; 2632 2633 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) { 2634 if (test_bit(i, lgr->rtokens_used_mask) && 2635 lgr->rtokens[i][lnk_idx].rkey == rkey) 2636 return i; 2637 } 2638 return -ENOENT; 2639 } 2640 2641 /* set rtoken for a new link to an existing rmb */ 2642 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new, 2643 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey) 2644 { 2645 int rtok_idx; 2646 2647 rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known)); 2648 if (rtok_idx == -ENOENT) 2649 return; 2650 lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey); 2651 lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr); 2652 } 2653 2654 /* set rtoken for a new link whose link_id is given */ 2655 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id, 2656 __be64 nw_vaddr, __be32 nw_rkey) 2657 { 2658 u64 dma_addr = be64_to_cpu(nw_vaddr); 2659 u32 rkey = ntohl(nw_rkey); 2660 bool found = false; 2661 int link_idx; 2662 2663 for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) { 2664 if (lgr->lnk[link_idx].link_id == link_id) { 2665 found = true; 2666 break; 2667 } 2668 } 2669 if (!found) 2670 return; 2671 lgr->rtokens[rtok_idx][link_idx].rkey = rkey; 2672 lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr; 2673 } 2674 2675 /* add a new rtoken from peer */ 2676 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey) 2677 { 2678 struct smc_link_group *lgr = smc_get_lgr(lnk); 2679 u64 dma_addr = be64_to_cpu(nw_vaddr); 2680 u32 rkey = ntohl(nw_rkey); 2681 int i; 2682 2683 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) { 2684 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey && 2685 lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr && 2686 test_bit(i, lgr->rtokens_used_mask)) { 2687 /* already in list */ 2688 return i; 2689 } 2690 } 2691 i = smc_rmb_reserve_rtoken_idx(lgr); 2692 if (i < 0) 2693 return i; 2694 lgr->rtokens[i][lnk->link_idx].rkey = rkey; 2695 lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr; 2696 return i; 2697 } 2698 2699 /* delete an rtoken from all links */ 2700 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey) 2701 { 2702 struct smc_link_group *lgr = smc_get_lgr(lnk); 2703 u32 rkey = ntohl(nw_rkey); 2704 int i, j; 2705 2706 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) { 2707 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey && 2708 test_bit(i, lgr->rtokens_used_mask)) { 2709 for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) { 2710 lgr->rtokens[i][j].rkey = 0; 2711 lgr->rtokens[i][j].dma_addr = 0; 2712 } 2713 clear_bit(i, lgr->rtokens_used_mask); 2714 return 0; 2715 } 2716 } 2717 return -ENOENT; 2718 } 2719 2720 /* save rkey and dma_addr received from peer during clc handshake */ 2721 int smc_rmb_rtoken_handling(struct smc_connection *conn, 2722 struct smc_link *lnk, 2723 struct smc_clc_msg_accept_confirm *clc) 2724 { 2725 conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr, 2726 clc->r0.rmb_rkey); 2727 if (conn->rtoken_idx < 0) 2728 return conn->rtoken_idx; 2729 return 0; 2730 } 2731 2732 static void smc_core_going_away(void) 2733 { 2734 struct smc_ib_device *smcibdev; 2735 struct smcd_dev *smcd; 2736 2737 mutex_lock(&smc_ib_devices.mutex); 2738 list_for_each_entry(smcibdev, &smc_ib_devices.list, list) { 2739 int i; 2740 2741 for (i = 0; i < SMC_MAX_PORTS; i++) 2742 set_bit(i, smcibdev->ports_going_away); 2743 } 2744 mutex_unlock(&smc_ib_devices.mutex); 2745 2746 mutex_lock(&smcd_dev_list.mutex); 2747 list_for_each_entry(smcd, &smcd_dev_list.list, list) { 2748 smcd->going_away = 1; 2749 } 2750 mutex_unlock(&smcd_dev_list.mutex); 2751 } 2752 2753 /* Clean up all SMC link groups */ 2754 static void smc_lgrs_shutdown(void) 2755 { 2756 struct smcd_dev *smcd; 2757 2758 smc_core_going_away(); 2759 2760 smc_smcr_terminate_all(NULL); 2761 2762 mutex_lock(&smcd_dev_list.mutex); 2763 list_for_each_entry(smcd, &smcd_dev_list.list, list) 2764 smc_smcd_terminate_all(smcd); 2765 mutex_unlock(&smcd_dev_list.mutex); 2766 } 2767 2768 static int smc_core_reboot_event(struct notifier_block *this, 2769 unsigned long event, void *ptr) 2770 { 2771 smc_lgrs_shutdown(); 2772 smc_ib_unregister_client(); 2773 smc_ism_exit(); 2774 return 0; 2775 } 2776 2777 static struct notifier_block smc_reboot_notifier = { 2778 .notifier_call = smc_core_reboot_event, 2779 }; 2780 2781 int __init smc_core_init(void) 2782 { 2783 return register_reboot_notifier(&smc_reboot_notifier); 2784 } 2785 2786 /* Called (from smc_exit) when module is removed */ 2787 void smc_core_exit(void) 2788 { 2789 unregister_reboot_notifier(&smc_reboot_notifier); 2790 smc_lgrs_shutdown(); 2791 } 2792