smc_core.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   4 *
   5 *  Basic Transport Functions exploiting Infiniband API
   6 *
   7 *  Copyright IBM Corp. 2016
   8 *
   9 *  Author(s):  Ursula Braun <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
  43struct 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
  49static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */
  50static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted);
  51
  52static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
  53			 struct smc_buf_desc *buf_desc);
  54static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft);
  55
  56static 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 */
  59static 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
  71static void smc_ibdev_cnt_inc(struct smc_link *lnk)
  72{
  73	atomic_inc(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
  74}
  75
  76static void smc_ibdev_cnt_dec(struct smc_link *lnk)
  77{
  78	atomic_dec(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
  79}
  80
  81static 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 */
 101static 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 */
 123static 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 */
 167static 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 */
 196static 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 */
 211static 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
 224static 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
 239static 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
 253int 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
 300errattr:
 301	nla_nest_cancel(skb, attrs);
 302errout:
 303	genlmsg_cancel(skb, nlh);
 304errmsg:
 305	return skb->len;
 306}
 307
 308/* Fill SMC_NLA_LGR_D_V2_COMMON/SMC_NLA_LGR_R_V2_COMMON nested attributes */
 309static 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
 335errv2attr:
 336	nla_nest_cancel(skb, v2_attrs);
 337	return -EMSGSIZE;
 338}
 339
 340static 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
 359errv2attr:
 360	nla_nest_cancel(skb, v2_attrs);
 361errattr:
 362	return -EMSGSIZE;
 363}
 364
 365static 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;
 411errattr:
 412	nla_nest_cancel(skb, attrs);
 413errout:
 414	return -EMSGSIZE;
 415}
 416
 417static 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;
 470errattr:
 471	nla_nest_cancel(skb, attrs);
 472errout:
 473	genlmsg_cancel(skb, nlh);
 474errmsg:
 475	return -EMSGSIZE;
 476}
 477
 478static 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	}
 503out:
 504	return 0;
 505
 506errout:
 507	genlmsg_cancel(skb, nlh);
 508errmsg:
 509	return -EMSGSIZE;
 510}
 511
 512static 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;
 528next:
 529		num++;
 530	}
 531errout:
 532	spin_unlock_bh(&smc_lgr->lock);
 533	cb_ctx->pos[0] = num;
 534}
 535
 536static 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
 598errattr:
 599	nla_nest_cancel(skb, attrs);
 600errout:
 601	genlmsg_cancel(skb, nlh);
 602errmsg:
 603	return -EMSGSIZE;
 604}
 605
 606static 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;
 624next:
 625		num++;
 626	}
 627errout:
 628	spin_unlock_bh(&dev->lgr_lock);
 629	cb_ctx->pos[1] = num;
 630	return rc;
 631}
 632
 633static 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;
 651next:
 652		num++;
 653	}
 654errout:
 655	mutex_unlock(&dev_list->mutex);
 656	cb_ctx->pos[0] = num;
 657	return rc;
 658}
 659
 660int 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
 668int 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
 676int 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
 682void 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
 698static 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
 712static void smc_lgr_free(struct smc_link_group *lgr);
 713
 714static 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
 750static 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 */
 759static u8 smcr_next_link_id(struct smc_link_group *lgr)
 760{
 761	u8 link_id;
 762	int i;
 763
 764	while (1) {
 765again:
 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
 779static 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
 788int 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
 868free_link_mem:
 869	smc_wr_free_link_mem(lnk);
 870destroy_qp:
 871	smc_ib_destroy_queue_pair(lnk);
 872dealloc_pd:
 873	smc_ib_dealloc_protection_domain(lnk);
 874clear_llc_lnk:
 875	smc_llc_link_clear(lnk, false);
 876out:
 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 */
 889static 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(sizeof(*lgr), GFP_KERNEL);
 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
1002free_wq:
1003	destroy_workqueue(lgr->tx_wq);
1004free_lgr:
1005	kfree(lgr);
1006ism_put_vlan:
1007	if (ini->is_smcd && ini->vlan_id)
1008		smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
1009out:
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
1019static 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
1033static 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			queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
1074			smc->sk.sk_data_ready(&smc->sk);
1075		}
1076	} else {
1077		smc_wr_tx_put_slot(conn->lnk,
1078				   (struct smc_wr_tx_pend_priv *)pend);
1079	}
1080	return rc;
1081}
1082
1083void smc_switch_link_and_count(struct smc_connection *conn,
1084			       struct smc_link *to_lnk)
1085{
1086	atomic_dec(&conn->lnk->conn_cnt);
1087	/* link_hold in smc_conn_create() */
1088	smcr_link_put(conn->lnk);
1089	conn->lnk = to_lnk;
1090	atomic_inc(&conn->lnk->conn_cnt);
1091	/* link_put in smc_conn_free() */
1092	smcr_link_hold(conn->lnk);
1093}
1094
1095struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
1096				  struct smc_link *from_lnk, bool is_dev_err)
1097{
1098	struct smc_link *to_lnk = NULL;
1099	struct smc_cdc_tx_pend *pend;
1100	struct smc_connection *conn;
1101	struct smc_wr_buf *wr_buf;
1102	struct smc_sock *smc;
1103	struct rb_node *node;
1104	int i, rc = 0;
1105
1106	/* link is inactive, wake up tx waiters */
1107	smc_wr_wakeup_tx_wait(from_lnk);
1108
1109	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1110		if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
1111			continue;
1112		if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
1113		    from_lnk->ibport == lgr->lnk[i].ibport) {
1114			continue;
1115		}
1116		to_lnk = &lgr->lnk[i];
1117		break;
1118	}
1119	if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
1120		smc_lgr_terminate_sched(lgr);
1121		return NULL;
1122	}
1123again:
1124	read_lock_bh(&lgr->conns_lock);
1125	for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
1126		conn = rb_entry(node, struct smc_connection, alert_node);
1127		if (conn->lnk != from_lnk)
1128			continue;
1129		smc = container_of(conn, struct smc_sock, conn);
1130		/* conn->lnk not yet set in SMC_INIT state */
1131		if (smc->sk.sk_state == SMC_INIT)
1132			continue;
1133		if (smc->sk.sk_state == SMC_CLOSED ||
1134		    smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
1135		    smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
1136		    smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
1137		    smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
1138		    smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
1139		    smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
1140		    smc->sk.sk_state == SMC_PEERABORTWAIT ||
1141		    smc->sk.sk_state == SMC_PROCESSABORT) {
1142			spin_lock_bh(&conn->send_lock);
1143			smc_switch_link_and_count(conn, to_lnk);
1144			spin_unlock_bh(&conn->send_lock);
1145			continue;
1146		}
1147		sock_hold(&smc->sk);
1148		read_unlock_bh(&lgr->conns_lock);
1149		/* pre-fetch buffer outside of send_lock, might sleep */
1150		rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
1151		if (rc)
1152			goto err_out;
1153		/* avoid race with smcr_tx_sndbuf_nonempty() */
1154		spin_lock_bh(&conn->send_lock);
1155		smc_switch_link_and_count(conn, to_lnk);
1156		rc = smc_switch_cursor(smc, pend, wr_buf);
1157		spin_unlock_bh(&conn->send_lock);
1158		sock_put(&smc->sk);
1159		if (rc)
1160			goto err_out;
1161		goto again;
1162	}
1163	read_unlock_bh(&lgr->conns_lock);
1164	smc_wr_tx_link_put(to_lnk);
1165	return to_lnk;
1166
1167err_out:
1168	smcr_link_down_cond_sched(to_lnk);
1169	smc_wr_tx_link_put(to_lnk);
1170	return NULL;
1171}
1172
1173static void smcr_buf_unuse(struct smc_buf_desc *buf_desc, bool is_rmb,
1174			   struct smc_link_group *lgr)
1175{
1176	struct rw_semaphore *lock;	/* lock buffer list */
1177	int rc;
1178
1179	if (is_rmb && buf_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1180		/* unregister rmb with peer */
1181		rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1182		if (!rc) {
1183			/* protect against smc_llc_cli_rkey_exchange() */
1184			down_read(&lgr->llc_conf_mutex);
1185			smc_llc_do_delete_rkey(lgr, buf_desc);
1186			buf_desc->is_conf_rkey = false;
1187			up_read(&lgr->llc_conf_mutex);
1188			smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1189		}
1190	}
1191
1192	if (buf_desc->is_reg_err) {
1193		/* buf registration failed, reuse not possible */
1194		lock = is_rmb ? &lgr->rmbs_lock :
1195				&lgr->sndbufs_lock;
1196		down_write(lock);
1197		smc_lgr_buf_list_del(lgr, is_rmb, buf_desc);
1198		up_write(lock);
1199
1200		smc_buf_free(lgr, is_rmb, buf_desc);
1201	} else {
1202		/* memzero_explicit provides potential memory barrier semantics */
1203		memzero_explicit(buf_desc->cpu_addr, buf_desc->len);
1204		WRITE_ONCE(buf_desc->used, 0);
1205	}
1206}
1207
1208static void smcd_buf_detach(struct smc_connection *conn)
1209{
1210	struct smcd_dev *smcd = conn->lgr->smcd;
1211	u64 peer_token = conn->peer_token;
1212
1213	if (!conn->sndbuf_desc)
1214		return;
1215
1216	smc_ism_detach_dmb(smcd, peer_token);
1217
1218	kfree(conn->sndbuf_desc);
1219	conn->sndbuf_desc = NULL;
1220}
1221
1222static void smc_buf_unuse(struct smc_connection *conn,
1223			  struct smc_link_group *lgr)
1224{
1225	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1226	bool is_smcd = lgr->is_smcd;
1227	int bufsize;
1228
1229	if (conn->sndbuf_desc) {
1230		bufsize = conn->sndbuf_desc->len;
1231		if (!is_smcd && conn->sndbuf_desc->is_vm) {
1232			smcr_buf_unuse(conn->sndbuf_desc, false, lgr);
1233		} else {
1234			memzero_explicit(conn->sndbuf_desc->cpu_addr, bufsize);
1235			WRITE_ONCE(conn->sndbuf_desc->used, 0);
1236		}
1237		SMC_STAT_RMB_SIZE(smc, is_smcd, false, false, bufsize);
1238	}
1239	if (conn->rmb_desc) {
1240		bufsize = conn->rmb_desc->len;
1241		if (!is_smcd) {
1242			smcr_buf_unuse(conn->rmb_desc, true, lgr);
1243		} else {
1244			bufsize += sizeof(struct smcd_cdc_msg);
1245			memzero_explicit(conn->rmb_desc->cpu_addr, bufsize);
1246			WRITE_ONCE(conn->rmb_desc->used, 0);
1247		}
1248		SMC_STAT_RMB_SIZE(smc, is_smcd, true, false, bufsize);
1249	}
1250}
1251
1252/* remove a finished connection from its link group */
1253void smc_conn_free(struct smc_connection *conn)
1254{
1255	struct smc_link_group *lgr = conn->lgr;
1256
1257	if (!lgr || conn->freed)
1258		/* Connection has never been registered in a
1259		 * link group, or has already been freed.
1260		 */
1261		return;
1262
1263	conn->freed = 1;
1264	if (!smc_conn_lgr_valid(conn))
1265		/* Connection has already unregistered from
1266		 * link group.
1267		 */
1268		goto lgr_put;
1269
1270	if (lgr->is_smcd) {
1271		if (!list_empty(&lgr->list))
1272			smc_ism_unset_conn(conn);
1273		if (smc_ism_support_dmb_nocopy(lgr->smcd))
1274			smcd_buf_detach(conn);
1275		tasklet_kill(&conn->rx_tsklet);
1276	} else {
1277		smc_cdc_wait_pend_tx_wr(conn);
1278		if (current_work() != &conn->abort_work)
1279			cancel_work_sync(&conn->abort_work);
1280	}
1281	if (!list_empty(&lgr->list)) {
1282		smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1283		smc_lgr_unregister_conn(conn);
1284	}
1285
1286	if (!lgr->conns_num)
1287		smc_lgr_schedule_free_work(lgr);
1288lgr_put:
1289	if (!lgr->is_smcd)
1290		smcr_link_put(conn->lnk); /* link_hold in smc_conn_create() */
1291	smc_lgr_put(lgr); /* lgr_hold in smc_conn_create() */
1292}
1293
1294/* unregister a link from a buf_desc */
1295static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1296				struct smc_link *lnk)
1297{
1298	if (is_rmb || buf_desc->is_vm)
1299		buf_desc->is_reg_mr[lnk->link_idx] = false;
1300	if (!buf_desc->is_map_ib[lnk->link_idx])
1301		return;
1302
1303	if ((is_rmb || buf_desc->is_vm) &&
1304	    buf_desc->mr[lnk->link_idx]) {
1305		smc_ib_put_memory_region(buf_desc->mr[lnk->link_idx]);
1306		buf_desc->mr[lnk->link_idx] = NULL;
1307	}
1308	if (is_rmb)
1309		smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1310	else
1311		smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1312
1313	sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1314	buf_desc->is_map_ib[lnk->link_idx] = false;
1315}
1316
1317/* unmap all buffers of lgr for a deleted link */
1318static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1319{
1320	struct smc_link_group *lgr = lnk->lgr;
1321	struct smc_buf_desc *buf_desc, *bf;
1322	int i;
1323
1324	for (i = 0; i < SMC_RMBE_SIZES; i++) {
1325		down_write(&lgr->rmbs_lock);
1326		list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1327			smcr_buf_unmap_link(buf_desc, true, lnk);
1328		up_write(&lgr->rmbs_lock);
1329
1330		down_write(&lgr->sndbufs_lock);
1331		list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1332					 list)
1333			smcr_buf_unmap_link(buf_desc, false, lnk);
1334		up_write(&lgr->sndbufs_lock);
1335	}
1336}
1337
1338static void smcr_rtoken_clear_link(struct smc_link *lnk)
1339{
1340	struct smc_link_group *lgr = lnk->lgr;
1341	int i;
1342
1343	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1344		lgr->rtokens[i][lnk->link_idx].rkey = 0;
1345		lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1346	}
1347}
1348
1349static void __smcr_link_clear(struct smc_link *lnk)
1350{
1351	struct smc_link_group *lgr = lnk->lgr;
1352	struct smc_ib_device *smcibdev;
1353
1354	smc_wr_free_link_mem(lnk);
1355	smc_ibdev_cnt_dec(lnk);
1356	put_device(&lnk->smcibdev->ibdev->dev);
1357	smcibdev = lnk->smcibdev;
1358	memset(lnk, 0, sizeof(struct smc_link));
1359	lnk->state = SMC_LNK_UNUSED;
1360	if (!atomic_dec_return(&smcibdev->lnk_cnt))
1361		wake_up(&smcibdev->lnks_deleted);
1362	smc_lgr_put(lgr); /* lgr_hold in smcr_link_init() */
1363}
1364
1365/* must be called under lgr->llc_conf_mutex lock */
1366void smcr_link_clear(struct smc_link *lnk, bool log)
1367{
1368	if (!lnk->lgr || lnk->clearing ||
1369	    lnk->state == SMC_LNK_UNUSED)
1370		return;
1371	lnk->clearing = 1;
1372	lnk->peer_qpn = 0;
1373	smc_llc_link_clear(lnk, log);
1374	smcr_buf_unmap_lgr(lnk);
1375	smcr_rtoken_clear_link(lnk);
1376	smc_ib_modify_qp_error(lnk);
1377	smc_wr_free_link(lnk);
1378	smc_ib_destroy_queue_pair(lnk);
1379	smc_ib_dealloc_protection_domain(lnk);
1380	smcr_link_put(lnk); /* theoretically last link_put */
1381}
1382
1383void smcr_link_hold(struct smc_link *lnk)
1384{
1385	refcount_inc(&lnk->refcnt);
1386}
1387
1388void smcr_link_put(struct smc_link *lnk)
1389{
1390	if (refcount_dec_and_test(&lnk->refcnt))
1391		__smcr_link_clear(lnk);
1392}
1393
1394static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1395			  struct smc_buf_desc *buf_desc)
1396{
1397	int i;
1398
1399	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1400		smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1401
1402	if (!buf_desc->is_vm && buf_desc->pages)
1403		__free_pages(buf_desc->pages, buf_desc->order);
1404	else if (buf_desc->is_vm && buf_desc->cpu_addr)
1405		vfree(buf_desc->cpu_addr);
1406	kfree(buf_desc);
1407}
1408
1409static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1410			  struct smc_buf_desc *buf_desc)
1411{
1412	if (is_dmb) {
1413		/* restore original buf len */
1414		buf_desc->len += sizeof(struct smcd_cdc_msg);
1415		smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1416	} else {
1417		kfree(buf_desc->cpu_addr);
1418	}
1419	kfree(buf_desc);
1420}
1421
1422static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1423			 struct smc_buf_desc *buf_desc)
1424{
1425	if (lgr->is_smcd)
1426		smcd_buf_free(lgr, is_rmb, buf_desc);
1427	else
1428		smcr_buf_free(lgr, is_rmb, buf_desc);
1429}
1430
1431static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1432{
1433	struct smc_buf_desc *buf_desc, *bf_desc;
1434	struct list_head *buf_list;
1435	int i;
1436
1437	for (i = 0; i < SMC_RMBE_SIZES; i++) {
1438		if (is_rmb)
1439			buf_list = &lgr->rmbs[i];
1440		else
1441			buf_list = &lgr->sndbufs[i];
1442		list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1443					 list) {
1444			smc_lgr_buf_list_del(lgr, is_rmb, buf_desc);
1445			smc_buf_free(lgr, is_rmb, buf_desc);
1446		}
1447	}
1448}
1449
1450static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1451{
1452	/* free send buffers */
1453	__smc_lgr_free_bufs(lgr, false);
1454	/* free rmbs */
1455	__smc_lgr_free_bufs(lgr, true);
1456}
1457
1458/* won't be freed until no one accesses to lgr anymore */
1459static void __smc_lgr_free(struct smc_link_group *lgr)
1460{
1461	smc_lgr_free_bufs(lgr);
1462	if (lgr->is_smcd) {
1463		if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1464			wake_up(&lgr->smcd->lgrs_deleted);
1465	} else {
1466		smc_wr_free_lgr_mem(lgr);
1467		if (!atomic_dec_return(&lgr_cnt))
1468			wake_up(&lgrs_deleted);
1469	}
1470	kfree(lgr);
1471}
1472
1473/* remove a link group */
1474static void smc_lgr_free(struct smc_link_group *lgr)
1475{
1476	int i;
1477
1478	if (!lgr->is_smcd) {
1479		down_write(&lgr->llc_conf_mutex);
1480		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1481			if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1482				smcr_link_clear(&lgr->lnk[i], false);
1483		}
1484		up_write(&lgr->llc_conf_mutex);
1485		smc_llc_lgr_clear(lgr);
1486	}
1487
1488	destroy_workqueue(lgr->tx_wq);
1489	if (lgr->is_smcd) {
1490		smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1491		put_device(&lgr->smcd->dibs->dev);
1492	}
1493	smc_lgr_put(lgr); /* theoretically last lgr_put */
1494}
1495
1496void smc_lgr_hold(struct smc_link_group *lgr)
1497{
1498	refcount_inc(&lgr->refcnt);
1499}
1500
1501void smc_lgr_put(struct smc_link_group *lgr)
1502{
1503	if (refcount_dec_and_test(&lgr->refcnt))
1504		__smc_lgr_free(lgr);
1505}
1506
1507static void smc_sk_wake_ups(struct smc_sock *smc)
1508{
1509	smc->sk.sk_write_space(&smc->sk);
1510	smc->sk.sk_data_ready(&smc->sk);
1511	smc->sk.sk_state_change(&smc->sk);
1512}
1513
1514/* kill a connection */
1515static void smc_conn_kill(struct smc_connection *conn, bool soft)
1516{
1517	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1518
1519	if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1520		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1521	else
1522		smc_close_abort(conn);
1523	conn->killed = 1;
1524	smc->sk.sk_err = ECONNABORTED;
1525	smc_sk_wake_ups(smc);
1526	if (conn->lgr->is_smcd) {
1527		smc_ism_unset_conn(conn);
1528		if (smc_ism_support_dmb_nocopy(conn->lgr->smcd))
1529			smcd_buf_detach(conn);
1530		if (soft)
1531			tasklet_kill(&conn->rx_tsklet);
1532		else
1533			tasklet_unlock_wait(&conn->rx_tsklet);
1534	} else {
1535		smc_cdc_wait_pend_tx_wr(conn);
1536	}
1537	smc_lgr_unregister_conn(conn);
1538	smc_close_active_abort(smc);
1539}
1540
1541static void smc_lgr_cleanup(struct smc_link_group *lgr)
1542{
1543	if (lgr->is_smcd) {
1544		smc_ism_signal_shutdown(lgr);
1545	} else {
1546		u32 rsn = lgr->llc_termination_rsn;
1547
1548		if (!rsn)
1549			rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1550		smc_llc_send_link_delete_all(lgr, false, rsn);
1551		smcr_lgr_link_deactivate_all(lgr);
1552	}
1553}
1554
1555/* terminate link group
1556 * @soft: true if link group shutdown can take its time
1557 *	  false if immediate link group shutdown is required
1558 */
1559static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1560{
1561	struct smc_connection *conn;
1562	struct smc_sock *smc;
1563	struct rb_node *node;
1564
1565	if (lgr->terminating)
1566		return;	/* lgr already terminating */
1567	/* cancel free_work sync, will terminate when lgr->freeing is set */
1568	cancel_delayed_work(&lgr->free_work);
1569	lgr->terminating = 1;
1570
1571	/* kill remaining link group connections */
1572	read_lock_bh(&lgr->conns_lock);
1573	node = rb_first(&lgr->conns_all);
1574	while (node) {
1575		read_unlock_bh(&lgr->conns_lock);
1576		conn = rb_entry(node, struct smc_connection, alert_node);
1577		smc = container_of(conn, struct smc_sock, conn);
1578		sock_hold(&smc->sk); /* sock_put below */
1579		lock_sock(&smc->sk);
1580		smc_conn_kill(conn, soft);
1581		release_sock(&smc->sk);
1582		sock_put(&smc->sk); /* sock_hold above */
1583		read_lock_bh(&lgr->conns_lock);
1584		node = rb_first(&lgr->conns_all);
1585	}
1586	read_unlock_bh(&lgr->conns_lock);
1587	smc_lgr_cleanup(lgr);
1588	smc_lgr_free(lgr);
1589}
1590
1591/* unlink link group and schedule termination */
1592void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1593{
1594	spinlock_t *lgr_lock;
1595
1596	smc_lgr_list_head(lgr, &lgr_lock);
1597	spin_lock_bh(lgr_lock);
1598	if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1599		spin_unlock_bh(lgr_lock);
1600		return;	/* lgr already terminating */
1601	}
1602	list_del_init(&lgr->list);
1603	lgr->freeing = 1;
1604	spin_unlock_bh(lgr_lock);
1605	schedule_work(&lgr->terminate_work);
1606}
1607
1608/* Called when peer lgr shutdown (regularly or abnormally) is received */
1609void smc_smcd_terminate(struct smcd_dev *dev, struct smcd_gid *peer_gid,
1610			unsigned short vlan)
1611{
1612	struct smc_link_group *lgr, *l;
1613	LIST_HEAD(lgr_free_list);
1614
1615	/* run common cleanup function and build free list */
1616	spin_lock_bh(&dev->lgr_lock);
1617	list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1618		if ((!peer_gid->gid ||
1619		     (lgr->peer_gid.gid == peer_gid->gid &&
1620		      !smc_ism_is_emulated(dev) ? 1 :
1621		      lgr->peer_gid.gid_ext == peer_gid->gid_ext)) &&
1622		    (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1623			if (peer_gid->gid) /* peer triggered termination */
1624				lgr->peer_shutdown = 1;
1625			list_move(&lgr->list, &lgr_free_list);
1626			lgr->freeing = 1;
1627		}
1628	}
1629	spin_unlock_bh(&dev->lgr_lock);
1630
1631	/* cancel the regular free workers and actually free lgrs */
1632	list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1633		list_del_init(&lgr->list);
1634		schedule_work(&lgr->terminate_work);
1635	}
1636}
1637
1638/* Called when an SMCD device is removed or the smc module is unloaded */
1639void smc_smcd_terminate_all(struct smcd_dev *smcd)
1640{
1641	struct smc_link_group *lgr, *lg;
1642	LIST_HEAD(lgr_free_list);
1643
1644	spin_lock_bh(&smcd->lgr_lock);
1645	list_splice_init(&smcd->lgr_list, &lgr_free_list);
1646	list_for_each_entry(lgr, &lgr_free_list, list)
1647		lgr->freeing = 1;
1648	spin_unlock_bh(&smcd->lgr_lock);
1649
1650	list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1651		list_del_init(&lgr->list);
1652		__smc_lgr_terminate(lgr, false);
1653	}
1654
1655	if (atomic_read(&smcd->lgr_cnt))
1656		wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1657}
1658
1659/* Called when an SMCR device is removed or the smc module is unloaded.
1660 * If smcibdev is given, all SMCR link groups using this device are terminated.
1661 * If smcibdev is NULL, all SMCR link groups are terminated.
1662 */
1663void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1664{
1665	struct smc_link_group *lgr, *lg;
1666	LIST_HEAD(lgr_free_list);
1667	int i;
1668
1669	spin_lock_bh(&smc_lgr_list.lock);
1670	if (!smcibdev) {
1671		list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1672		list_for_each_entry(lgr, &lgr_free_list, list)
1673			lgr->freeing = 1;
1674	} else {
1675		list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1676			for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1677				if (lgr->lnk[i].smcibdev == smcibdev)
1678					smcr_link_down_cond_sched(&lgr->lnk[i]);
1679			}
1680		}
1681	}
1682	spin_unlock_bh(&smc_lgr_list.lock);
1683
1684	list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1685		list_del_init(&lgr->list);
1686		smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1687		__smc_lgr_terminate(lgr, false);
1688	}
1689
1690	if (smcibdev) {
1691		if (atomic_read(&smcibdev->lnk_cnt))
1692			wait_event(smcibdev->lnks_deleted,
1693				   !atomic_read(&smcibdev->lnk_cnt));
1694	} else {
1695		if (atomic_read(&lgr_cnt))
1696			wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1697	}
1698}
1699
1700/* set new lgr type and clear all asymmetric link tagging */
1701void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1702{
1703	char *lgr_type = "";
1704	int i;
1705
1706	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1707		if (smc_link_usable(&lgr->lnk[i]))
1708			lgr->lnk[i].link_is_asym = false;
1709	if (lgr->type == new_type)
1710		return;
1711	lgr->type = new_type;
1712
1713	switch (lgr->type) {
1714	case SMC_LGR_NONE:
1715		lgr_type = "NONE";
1716		break;
1717	case SMC_LGR_SINGLE:
1718		lgr_type = "SINGLE";
1719		break;
1720	case SMC_LGR_SYMMETRIC:
1721		lgr_type = "SYMMETRIC";
1722		break;
1723	case SMC_LGR_ASYMMETRIC_PEER:
1724		lgr_type = "ASYMMETRIC_PEER";
1725		break;
1726	case SMC_LGR_ASYMMETRIC_LOCAL:
1727		lgr_type = "ASYMMETRIC_LOCAL";
1728		break;
1729	}
1730	pr_warn_ratelimited("smc: SMC-R lg %*phN net %llu state changed: "
1731			    "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1732			    lgr->net->net_cookie, lgr_type, lgr->pnet_id);
1733}
1734
1735/* set new lgr type and tag a link as asymmetric */
1736void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1737			    enum smc_lgr_type new_type, int asym_lnk_idx)
1738{
1739	smcr_lgr_set_type(lgr, new_type);
1740	lgr->lnk[asym_lnk_idx].link_is_asym = true;
1741}
1742
1743/* abort connection, abort_work scheduled from tasklet context */
1744static void smc_conn_abort_work(struct work_struct *work)
1745{
1746	struct smc_connection *conn = container_of(work,
1747						   struct smc_connection,
1748						   abort_work);
1749	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1750
1751	lock_sock(&smc->sk);
1752	smc_conn_kill(conn, true);
1753	release_sock(&smc->sk);
1754	sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1755}
1756
1757void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1758{
1759	struct smc_link_group *lgr, *n;
1760
1761	spin_lock_bh(&smc_lgr_list.lock);
1762	list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1763		struct smc_link *link;
1764
1765		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1766			    SMC_MAX_PNETID_LEN) ||
1767		    lgr->type == SMC_LGR_SYMMETRIC ||
1768		    lgr->type == SMC_LGR_ASYMMETRIC_PEER ||
1769		    !rdma_dev_access_netns(smcibdev->ibdev, lgr->net))
1770			continue;
1771
1772		if (lgr->type == SMC_LGR_SINGLE && lgr->max_links <= 1)
1773			continue;
1774
1775		/* trigger local add link processing */
1776		link = smc_llc_usable_link(lgr);
1777		if (link)
1778			smc_llc_add_link_local(link);
1779	}
1780	spin_unlock_bh(&smc_lgr_list.lock);
1781}
1782
1783/* link is down - switch connections to alternate link,
1784 * must be called under lgr->llc_conf_mutex lock
1785 */
1786static void smcr_link_down(struct smc_link *lnk)
1787{
1788	struct smc_link_group *lgr = lnk->lgr;
1789	struct smc_link *to_lnk;
1790	int del_link_id;
1791
1792	if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1793		return;
1794
1795	to_lnk = smc_switch_conns(lgr, lnk, true);
1796	if (!to_lnk) { /* no backup link available */
1797		smcr_link_clear(lnk, true);
1798		return;
1799	}
1800	smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1801	del_link_id = lnk->link_id;
1802
1803	if (lgr->role == SMC_SERV) {
1804		/* trigger local delete link processing */
1805		smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1806	} else {
1807		if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1808			/* another llc task is ongoing */
1809			up_write(&lgr->llc_conf_mutex);
1810			wait_event_timeout(lgr->llc_flow_waiter,
1811				(list_empty(&lgr->list) ||
1812				 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1813				SMC_LLC_WAIT_TIME);
1814			down_write(&lgr->llc_conf_mutex);
1815		}
1816		if (!list_empty(&lgr->list)) {
1817			smc_llc_send_delete_link(to_lnk, del_link_id,
1818						 SMC_LLC_REQ, true,
1819						 SMC_LLC_DEL_LOST_PATH);
1820			smcr_link_clear(lnk, true);
1821		}
1822		wake_up(&lgr->llc_flow_waiter);	/* wake up next waiter */
1823	}
1824}
1825
1826/* must be called under lgr->llc_conf_mutex lock */
1827void smcr_link_down_cond(struct smc_link *lnk)
1828{
1829	if (smc_link_downing(&lnk->state)) {
1830		trace_smcr_link_down(lnk, __builtin_return_address(0));
1831		smcr_link_down(lnk);
1832	}
1833}
1834
1835/* will get the lgr->llc_conf_mutex lock */
1836void smcr_link_down_cond_sched(struct smc_link *lnk)
1837{
1838	if (smc_link_downing(&lnk->state)) {
1839		trace_smcr_link_down(lnk, __builtin_return_address(0));
1840		smcr_link_hold(lnk); /* smcr_link_put in link_down_wrk */
1841		if (!schedule_work(&lnk->link_down_wrk))
1842			smcr_link_put(lnk);
1843	}
1844}
1845
1846void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1847{
1848	struct smc_link_group *lgr, *n;
1849	int i;
1850
1851	list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1852		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1853			    SMC_MAX_PNETID_LEN))
1854			continue; /* lgr is not affected */
1855		if (list_empty(&lgr->list))
1856			continue;
1857		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1858			struct smc_link *lnk = &lgr->lnk[i];
1859
1860			if (smc_link_usable(lnk) &&
1861			    lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1862				smcr_link_down_cond_sched(lnk);
1863		}
1864	}
1865}
1866
1867static void smc_link_down_work(struct work_struct *work)
1868{
1869	struct smc_link *link = container_of(work, struct smc_link,
1870					     link_down_wrk);
1871	struct smc_link_group *lgr = link->lgr;
1872
1873	if (list_empty(&lgr->list))
1874		goto out;
1875	wake_up_all(&lgr->llc_msg_waiter);
1876	down_write(&lgr->llc_conf_mutex);
1877	smcr_link_down(link);
1878	up_write(&lgr->llc_conf_mutex);
1879
1880out:
1881	smcr_link_put(link); /* smcr_link_hold by schedulers of link_down_work */
1882}
1883
1884static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
1885				  struct netdev_nested_priv *priv)
1886{
1887	unsigned short *vlan_id = (unsigned short *)priv->data;
1888
1889	if (is_vlan_dev(lower_dev)) {
1890		*vlan_id = vlan_dev_vlan_id(lower_dev);
1891		return 1;
1892	}
1893
1894	return 0;
1895}
1896
1897/* Determine vlan of internal TCP socket. */
1898int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1899{
1900	struct netdev_nested_priv priv;
1901	struct net_device *ndev;
1902	struct dst_entry *dst;
1903	int rc = 0;
1904
1905	ini->vlan_id = 0;
1906
1907	rcu_read_lock();
1908
1909	dst = __sk_dst_get(clcsock->sk);
1910	ndev = dst ? dst_dev_rcu(dst) : NULL;
1911	if (!ndev) {
1912		rc = -ENODEV;
1913		goto out;
1914	}
1915
1916	if (is_vlan_dev(ndev)) {
1917		ini->vlan_id = vlan_dev_vlan_id(ndev);
1918		goto out;
1919	}
1920
1921	priv.data = (void *)&ini->vlan_id;
1922	netdev_walk_all_lower_dev_rcu(ndev, smc_vlan_by_tcpsk_walk, &priv);
1923out:
1924	rcu_read_unlock();
1925
1926	return rc;
1927}
1928
1929static bool smcr_lgr_match(struct smc_link_group *lgr, u8 smcr_version,
1930			   u8 peer_systemid[],
1931			   u8 peer_gid[],
1932			   u8 peer_mac_v1[],
1933			   enum smc_lgr_role role, u32 clcqpn,
1934			   struct net *net)
1935{
1936	struct smc_link *lnk;
1937	int i;
1938
1939	if (memcmp(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN) ||
1940	    lgr->role != role)
1941		return false;
1942
1943	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1944		lnk = &lgr->lnk[i];
1945
1946		if (!smc_link_active(lnk))
1947			continue;
1948		/* use verbs API to check netns, instead of lgr->net */
1949		if (!rdma_dev_access_netns(lnk->smcibdev->ibdev, net))
1950			return false;
1951		if ((lgr->role == SMC_SERV || lnk->peer_qpn == clcqpn) &&
1952		    !memcmp(lnk->peer_gid, peer_gid, SMC_GID_SIZE) &&
1953		    (smcr_version == SMC_V2 ||
1954		     !memcmp(lnk->peer_mac, peer_mac_v1, ETH_ALEN)))
1955			return true;
1956	}
1957	return false;
1958}
1959
1960static bool smcd_lgr_match(struct smc_link_group *lgr,
1961			   struct smcd_dev *smcismdev,
1962			   struct smcd_gid *peer_gid)
1963{
1964	if (lgr->peer_gid.gid != peer_gid->gid ||
1965	    lgr->smcd != smcismdev)
1966		return false;
1967
1968	if (smc_ism_is_emulated(smcismdev) &&
1969	    lgr->peer_gid.gid_ext != peer_gid->gid_ext)
1970		return false;
1971
1972	return true;
1973}
1974
1975/* create a new SMC connection (and a new link group if necessary) */
1976int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1977{
1978	struct smc_connection *conn = &smc->conn;
1979	struct net *net = sock_net(&smc->sk);
1980	struct list_head *lgr_list;
1981	struct smc_link_group *lgr;
1982	enum smc_lgr_role role;
1983	spinlock_t *lgr_lock;
1984	int rc = 0;
1985
1986	lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1987				  &smc_lgr_list.list;
1988	lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1989				  &smc_lgr_list.lock;
1990	ini->first_contact_local = 1;
1991	role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1992	if (role == SMC_CLNT && ini->first_contact_peer)
1993		/* create new link group as well */
1994		goto create;
1995
1996	/* determine if an existing link group can be reused */
1997	spin_lock_bh(lgr_lock);
1998	list_for_each_entry(lgr, lgr_list, list) {
1999		write_lock_bh(&lgr->conns_lock);
2000		if ((ini->is_smcd ?
2001		     smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
2002				    &ini->ism_peer_gid[ini->ism_selected]) :
2003		     smcr_lgr_match(lgr, ini->smcr_version,
2004				    ini->peer_systemid,
2005				    ini->peer_gid, ini->peer_mac, role,
2006				    ini->ib_clcqpn, net)) &&
2007		    !lgr->sync_err &&
2008		    (ini->smcd_version == SMC_V2 ||
2009		     lgr->vlan_id == ini->vlan_id) &&
2010		    (role == SMC_CLNT || ini->is_smcd ||
2011		    (lgr->conns_num < lgr->max_conns &&
2012		      !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
2013			/* link group found */
2014			ini->first_contact_local = 0;
2015			conn->lgr = lgr;
2016			rc = smc_lgr_register_conn(conn, false);
2017			write_unlock_bh(&lgr->conns_lock);
2018			if (!rc && delayed_work_pending(&lgr->free_work))
2019				cancel_delayed_work(&lgr->free_work);
2020			break;
2021		}
2022		write_unlock_bh(&lgr->conns_lock);
2023	}
2024	spin_unlock_bh(lgr_lock);
2025	if (rc)
2026		return rc;
2027
2028	if (role == SMC_CLNT && !ini->first_contact_peer &&
2029	    ini->first_contact_local) {
2030		/* Server reuses a link group, but Client wants to start
2031		 * a new one
2032		 * send out_of_sync decline, reason synchr. error
2033		 */
2034		return SMC_CLC_DECL_SYNCERR;
2035	}
2036
2037create:
2038	if (ini->first_contact_local) {
2039		rc = smc_lgr_create(smc, ini);
2040		if (rc)
2041			goto out;
2042		lgr = conn->lgr;
2043		write_lock_bh(&lgr->conns_lock);
2044		rc = smc_lgr_register_conn(conn, true);
2045		write_unlock_bh(&lgr->conns_lock);
2046		if (rc) {
2047			smc_lgr_cleanup_early(lgr);
2048			goto out;
2049		}
2050	}
2051	smc_lgr_hold(conn->lgr); /* lgr_put in smc_conn_free() */
2052	if (!conn->lgr->is_smcd)
2053		smcr_link_hold(conn->lnk); /* link_put in smc_conn_free() */
2054	conn->freed = 0;
2055	conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
2056	conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
2057	conn->urg_state = SMC_URG_READ;
2058	init_waitqueue_head(&conn->cdc_pend_tx_wq);
2059	INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
2060	if (ini->is_smcd) {
2061		conn->rx_off = sizeof(struct smcd_cdc_msg);
2062		smcd_cdc_rx_init(conn); /* init tasklet for this conn */
2063	} else {
2064		conn->rx_off = 0;
2065	}
2066#ifndef KERNEL_HAS_ATOMIC64
2067	spin_lock_init(&conn->acurs_lock);
2068#endif
2069
2070out:
2071	return rc;
2072}
2073
2074#define SMCD_DMBE_SIZES		6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
2075#define SMCR_RMBE_SIZES		15 /* 0 -> 16KB, 1 -> 32KB, .. 15 -> 512MB */
2076
2077/* convert the RMB size into the compressed notation (minimum 16K, see
2078 * SMCD/R_DMBE_SIZES.
2079 * In contrast to plain ilog2, this rounds towards the next power of 2,
2080 * so the socket application gets at least its desired sndbuf / rcvbuf size.
2081 */
2082static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
2083{
2084	u8 compressed;
2085
2086	if (size <= SMC_BUF_MIN_SIZE)
2087		return 0;
2088
2089	size = (size - 1) >> 14;  /* convert to 16K multiple */
2090	compressed = min_t(u8, ilog2(size) + 1,
2091			   is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
2092
2093#ifdef CONFIG_ARCH_NO_SG_CHAIN
2094	if (!is_smcd && is_rmb)
2095		/* RMBs are backed by & limited to max size of scatterlists */
2096		compressed = min_t(u8, compressed, ilog2((SG_MAX_SINGLE_ALLOC * PAGE_SIZE) >> 14));
2097#endif
2098
2099	return compressed;
2100}
2101
2102/* convert the RMB size from compressed notation into integer */
2103int smc_uncompress_bufsize(u8 compressed)
2104{
2105	u32 size;
2106
2107	size = 0x00000001 << (((int)compressed) + 14);
2108	return (int)size;
2109}
2110
2111/* try to reuse a sndbuf or rmb description slot for a certain
2112 * buffer size; if not available, return NULL
2113 */
2114static struct smc_buf_desc *smc_buf_get_slot(struct rw_semaphore *lock,
2115					     struct list_head *buf_list)
2116{
2117	struct smc_buf_desc *buf_slot;
2118
2119	down_read(lock);
2120	list_for_each_entry(buf_slot, buf_list, list) {
2121		if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
2122			up_read(lock);
2123			return buf_slot;
2124		}
2125	}
2126	up_read(lock);
2127	return NULL;
2128}
2129
2130/* one of the conditions for announcing a receiver's current window size is
2131 * that it "results in a minimum increase in the window size of 10% of the
2132 * receive buffer space" [RFC7609]
2133 */
2134static inline int smc_rmb_wnd_update_limit(int rmbe_size)
2135{
2136	return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
2137}
2138
2139/* map an buf to a link */
2140static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
2141			     struct smc_link *lnk)
2142{
2143	int rc, i, nents, offset, buf_size, size, access_flags;
2144	struct scatterlist *sg;
2145	void *buf;
2146
2147	if (buf_desc->is_map_ib[lnk->link_idx])
2148		return 0;
2149
2150	if (buf_desc->is_vm) {
2151		buf = buf_desc->cpu_addr;
2152		buf_size = buf_desc->len;
2153		offset = offset_in_page(buf_desc->cpu_addr);
2154		nents = PAGE_ALIGN(buf_size + offset) / PAGE_SIZE;
2155	} else {
2156		nents = 1;
2157	}
2158
2159	rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], nents, GFP_KERNEL);
2160	if (rc)
2161		return rc;
2162
2163	if (buf_desc->is_vm) {
2164		/* virtually contiguous buffer */
2165		for_each_sg(buf_desc->sgt[lnk->link_idx].sgl, sg, nents, i) {
2166			size = min_t(int, PAGE_SIZE - offset, buf_size);
2167			sg_set_page(sg, vmalloc_to_page(buf), size, offset);
2168			buf += size;
2169			buf_size -= size;
2170			offset = 0;
2171		}
2172	} else {
2173		/* physically contiguous buffer */
2174		sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
2175			   buf_desc->cpu_addr, buf_desc->len);
2176	}
2177
2178	/* map sg table to DMA address */
2179	rc = smc_ib_buf_map_sg(lnk, buf_desc,
2180			       is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2181	/* SMC protocol depends on mapping to one DMA address only */
2182	if (rc != nents) {
2183		rc = -EAGAIN;
2184		goto free_table;
2185	}
2186
2187	buf_desc->is_dma_need_sync |=
2188		smc_ib_is_sg_need_sync(lnk, buf_desc) << lnk->link_idx;
2189
2190	if (is_rmb || buf_desc->is_vm) {
2191		/* create a new memory region for the RMB or vzalloced sndbuf */
2192		access_flags = is_rmb ?
2193			       IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2194			       IB_ACCESS_LOCAL_WRITE;
2195
2196		rc = smc_ib_get_memory_region(lnk->roce_pd, access_flags,
2197					      buf_desc, lnk->link_idx);
2198		if (rc)
2199			goto buf_unmap;
2200		smc_ib_sync_sg_for_device(lnk, buf_desc,
2201					  is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2202	}
2203	buf_desc->is_map_ib[lnk->link_idx] = true;
2204	return 0;
2205
2206buf_unmap:
2207	smc_ib_buf_unmap_sg(lnk, buf_desc,
2208			    is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2209free_table:
2210	sg_free_table(&buf_desc->sgt[lnk->link_idx]);
2211	return rc;
2212}
2213
2214/* register a new buf on IB device, rmb or vzalloced sndbuf
2215 * must be called under lgr->llc_conf_mutex lock
2216 */
2217int smcr_link_reg_buf(struct smc_link *link, struct smc_buf_desc *buf_desc)
2218{
2219	if (list_empty(&link->lgr->list))
2220		return -ENOLINK;
2221	if (!buf_desc->is_reg_mr[link->link_idx]) {
2222		/* register memory region for new buf */
2223		if (buf_desc->is_vm)
2224			buf_desc->mr[link->link_idx]->iova =
2225				(uintptr_t)buf_desc->cpu_addr;
2226		if (smc_wr_reg_send(link, buf_desc->mr[link->link_idx])) {
2227			buf_desc->is_reg_err = true;
2228			return -EFAULT;
2229		}
2230		buf_desc->is_reg_mr[link->link_idx] = true;
2231	}
2232	return 0;
2233}
2234
2235static int _smcr_buf_map_lgr(struct smc_link *lnk, struct rw_semaphore *lock,
2236			     struct list_head *lst, bool is_rmb)
2237{
2238	struct smc_buf_desc *buf_desc, *bf;
2239	int rc = 0;
2240
2241	down_write(lock);
2242	list_for_each_entry_safe(buf_desc, bf, lst, list) {
2243		if (!buf_desc->used)
2244			continue;
2245		rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
2246		if (rc)
2247			goto out;
2248	}
2249out:
2250	up_write(lock);
2251	return rc;
2252}
2253
2254/* map all used buffers of lgr for a new link */
2255int smcr_buf_map_lgr(struct smc_link *lnk)
2256{
2257	struct smc_link_group *lgr = lnk->lgr;
2258	int i, rc = 0;
2259
2260	for (i = 0; i < SMC_RMBE_SIZES; i++) {
2261		rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
2262				       &lgr->rmbs[i], true);
2263		if (rc)
2264			return rc;
2265		rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
2266				       &lgr->sndbufs[i], false);
2267		if (rc)
2268			return rc;
2269	}
2270	return 0;
2271}
2272
2273/* register all used buffers of lgr for a new link,
2274 * must be called under lgr->llc_conf_mutex lock
2275 */
2276int smcr_buf_reg_lgr(struct smc_link *lnk)
2277{
2278	struct smc_link_group *lgr = lnk->lgr;
2279	struct smc_buf_desc *buf_desc, *bf;
2280	int i, rc = 0;
2281
2282	/* reg all RMBs for a new link */
2283	down_write(&lgr->rmbs_lock);
2284	for (i = 0; i < SMC_RMBE_SIZES; i++) {
2285		list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
2286			if (!buf_desc->used)
2287				continue;
2288			rc = smcr_link_reg_buf(lnk, buf_desc);
2289			if (rc) {
2290				up_write(&lgr->rmbs_lock);
2291				return rc;
2292			}
2293		}
2294	}
2295	up_write(&lgr->rmbs_lock);
2296
2297	if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2298		return rc;
2299
2300	/* reg all vzalloced sndbufs for a new link */
2301	down_write(&lgr->sndbufs_lock);
2302	for (i = 0; i < SMC_RMBE_SIZES; i++) {
2303		list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], list) {
2304			if (!buf_desc->used || !buf_desc->is_vm)
2305				continue;
2306			rc = smcr_link_reg_buf(lnk, buf_desc);
2307			if (rc) {
2308				up_write(&lgr->sndbufs_lock);
2309				return rc;
2310			}
2311		}
2312	}
2313	up_write(&lgr->sndbufs_lock);
2314	return rc;
2315}
2316
2317static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
2318						int bufsize)
2319{
2320	struct smc_buf_desc *buf_desc;
2321
2322	/* try to alloc a new buffer */
2323	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2324	if (!buf_desc)
2325		return ERR_PTR(-ENOMEM);
2326
2327	switch (lgr->buf_type) {
2328	case SMCR_PHYS_CONT_BUFS:
2329	case SMCR_MIXED_BUFS:
2330		buf_desc->order = get_order(bufsize);
2331		buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
2332					      __GFP_NOMEMALLOC | __GFP_COMP |
2333					      __GFP_NORETRY | __GFP_ZERO,
2334					      buf_desc->order);
2335		if (buf_desc->pages) {
2336			buf_desc->cpu_addr =
2337				(void *)page_address(buf_desc->pages);
2338			buf_desc->len = bufsize;
2339			buf_desc->is_vm = false;
2340			break;
2341		}
2342		if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2343			goto out;
2344		fallthrough;	// try virtually contiguous buf
2345	case SMCR_VIRT_CONT_BUFS:
2346		buf_desc->order = get_order(bufsize);
2347		buf_desc->cpu_addr = vzalloc(PAGE_SIZE << buf_desc->order);
2348		if (!buf_desc->cpu_addr)
2349			goto out;
2350		buf_desc->pages = NULL;
2351		buf_desc->len = bufsize;
2352		buf_desc->is_vm = true;
2353		break;
2354	}
2355	return buf_desc;
2356
2357out:
2358	kfree(buf_desc);
2359	return ERR_PTR(-EAGAIN);
2360}
2361
2362/* map buf_desc on all usable links,
2363 * unused buffers stay mapped as long as the link is up
2364 */
2365static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
2366				     struct smc_buf_desc *buf_desc, bool is_rmb)
2367{
2368	int i, rc = 0, cnt = 0;
2369
2370	/* protect against parallel link reconfiguration */
2371	down_read(&lgr->llc_conf_mutex);
2372	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2373		struct smc_link *lnk = &lgr->lnk[i];
2374
2375		if (!smc_link_usable(lnk))
2376			continue;
2377		if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
2378			rc = -ENOMEM;
2379			goto out;
2380		}
2381		cnt++;
2382	}
2383out:
2384	up_read(&lgr->llc_conf_mutex);
2385	if (!rc && !cnt)
2386		rc = -EINVAL;
2387	return rc;
2388}
2389
2390static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
2391						bool is_dmb, int bufsize)
2392{
2393	struct smc_buf_desc *buf_desc;
2394	int rc;
2395
2396	/* try to alloc a new DMB */
2397	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2398	if (!buf_desc)
2399		return ERR_PTR(-ENOMEM);
2400	if (is_dmb) {
2401		rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2402		if (rc) {
2403			kfree(buf_desc);
2404			if (rc == -ENOMEM)
2405				return ERR_PTR(-EAGAIN);
2406			if (rc == -ENOSPC)
2407				return ERR_PTR(-ENOSPC);
2408			return ERR_PTR(-EIO);
2409		}
2410		buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2411		/* CDC header stored in buf. So, pretend it was smaller */
2412		buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2413	} else {
2414		buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2415					     __GFP_NOWARN | __GFP_NORETRY |
2416					     __GFP_NOMEMALLOC);
2417		if (!buf_desc->cpu_addr) {
2418			kfree(buf_desc);
2419			return ERR_PTR(-EAGAIN);
2420		}
2421		buf_desc->len = bufsize;
2422	}
2423	return buf_desc;
2424}
2425
2426static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2427{
2428	struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2429	struct smc_connection *conn = &smc->conn;
2430	struct smc_link_group *lgr = conn->lgr;
2431	struct list_head *buf_list;
2432	int bufsize, bufsize_comp;
2433	struct rw_semaphore *lock;	/* lock buffer list */
2434	bool is_dgraded = false;
2435
2436	if (is_rmb)
2437		/* use socket recv buffer size (w/o overhead) as start value */
2438		bufsize = smc->sk.sk_rcvbuf / 2;
2439	else
2440		/* use socket send buffer size (w/o overhead) as start value */
2441		bufsize = smc->sk.sk_sndbuf / 2;
2442
2443	for (bufsize_comp = smc_compress_bufsize(bufsize, is_smcd, is_rmb);
2444	     bufsize_comp >= 0; bufsize_comp--) {
2445		if (is_rmb) {
2446			lock = &lgr->rmbs_lock;
2447			buf_list = &lgr->rmbs[bufsize_comp];
2448		} else {
2449			lock = &lgr->sndbufs_lock;
2450			buf_list = &lgr->sndbufs[bufsize_comp];
2451		}
2452		bufsize = smc_uncompress_bufsize(bufsize_comp);
2453
2454		/* check for reusable slot in the link group */
2455		buf_desc = smc_buf_get_slot(lock, buf_list);
2456		if (buf_desc) {
2457			buf_desc->is_dma_need_sync = 0;
2458			SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, true, bufsize);
2459			SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2460			break; /* found reusable slot */
2461		}
2462
2463		if (is_smcd)
2464			buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2465		else
2466			buf_desc = smcr_new_buf_create(lgr, bufsize);
2467
2468		if (PTR_ERR(buf_desc) == -ENOMEM)
2469			break;
2470		if (IS_ERR(buf_desc)) {
2471			if (!is_dgraded) {
2472				is_dgraded = true;
2473				SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2474			}
2475			continue;
2476		}
2477
2478		SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2479		SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, true, bufsize);
2480		buf_desc->used = 1;
2481		down_write(lock);
2482		smc_lgr_buf_list_add(lgr, is_rmb, buf_list, buf_desc);
2483		up_write(lock);
2484		break; /* found */
2485	}
2486
2487	if (IS_ERR(buf_desc))
2488		return PTR_ERR(buf_desc);
2489
2490	if (!is_smcd) {
2491		if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2492			smcr_buf_unuse(buf_desc, is_rmb, lgr);
2493			return -ENOMEM;
2494		}
2495	}
2496
2497	if (is_rmb) {
2498		conn->rmb_desc = buf_desc;
2499		conn->rmbe_size_comp = bufsize_comp;
2500		smc->sk.sk_rcvbuf = bufsize * 2;
2501		atomic_set(&conn->bytes_to_rcv, 0);
2502		conn->rmbe_update_limit =
2503			smc_rmb_wnd_update_limit(buf_desc->len);
2504		if (is_smcd)
2505			smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2506	} else {
2507		conn->sndbuf_desc = buf_desc;
2508		smc->sk.sk_sndbuf = bufsize * 2;
2509		atomic_set(&conn->sndbuf_space, bufsize);
2510	}
2511	return 0;
2512}
2513
2514void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2515{
2516	if (!conn->sndbuf_desc->is_dma_need_sync)
2517		return;
2518	if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd ||
2519	    !smc_link_active(conn->lnk))
2520		return;
2521	smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2522}
2523
2524void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2525{
2526	int i;
2527
2528	if (!conn->rmb_desc->is_dma_need_sync)
2529		return;
2530	if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd)
2531		return;
2532	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2533		if (!smc_link_active(&conn->lgr->lnk[i]))
2534			continue;
2535		smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2536				       DMA_FROM_DEVICE);
2537	}
2538}
2539
2540/* create the send and receive buffer for an SMC socket;
2541 * receive buffers are called RMBs;
2542 * (even though the SMC protocol allows more than one RMB-element per RMB,
2543 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2544 * extra RMB for every connection in a link group
2545 */
2546int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2547{
2548	int rc;
2549
2550	/* create send buffer */
2551	if (is_smcd &&
2552	    smc_ism_support_dmb_nocopy(smc->conn.lgr->smcd))
2553		goto create_rmb;
2554
2555	rc = __smc_buf_create(smc, is_smcd, false);
2556	if (rc)
2557		return rc;
2558
2559create_rmb:
2560	/* create rmb */
2561	rc = __smc_buf_create(smc, is_smcd, true);
2562	if (rc && smc->conn.sndbuf_desc) {
2563		down_write(&smc->conn.lgr->sndbufs_lock);
2564		smc_lgr_buf_list_del(smc->conn.lgr, false,
2565				     smc->conn.sndbuf_desc);
2566		up_write(&smc->conn.lgr->sndbufs_lock);
2567		smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2568		smc->conn.sndbuf_desc = NULL;
2569	}
2570	return rc;
2571}
2572
2573int smcd_buf_attach(struct smc_sock *smc)
2574{
2575	struct smc_connection *conn = &smc->conn;
2576	struct smcd_dev *smcd = conn->lgr->smcd;
2577	u64 peer_token = conn->peer_token;
2578	struct smc_buf_desc *buf_desc;
2579	int rc;
2580
2581	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2582	if (!buf_desc)
2583		return -ENOMEM;
2584
2585	/* The ghost sndbuf_desc describes the same memory region as
2586	 * peer RMB. Its lifecycle is consistent with the connection's
2587	 * and it will be freed with the connections instead of the
2588	 * link group.
2589	 */
2590	rc = smc_ism_attach_dmb(smcd, peer_token, buf_desc);
2591	if (rc)
2592		goto free;
2593
2594	smc->sk.sk_sndbuf = buf_desc->len;
2595	buf_desc->cpu_addr =
2596		(u8 *)buf_desc->cpu_addr + sizeof(struct smcd_cdc_msg);
2597	buf_desc->len -= sizeof(struct smcd_cdc_msg);
2598	conn->sndbuf_desc = buf_desc;
2599	conn->sndbuf_desc->used = 1;
2600	atomic_set(&conn->sndbuf_space, conn->sndbuf_desc->len);
2601	return 0;
2602
2603free:
2604	kfree(buf_desc);
2605	return rc;
2606}
2607
2608static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2609{
2610	int i;
2611
2612	for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2613		if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2614			return i;
2615	}
2616	return -ENOSPC;
2617}
2618
2619static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2620				   u32 rkey)
2621{
2622	int i;
2623
2624	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2625		if (test_bit(i, lgr->rtokens_used_mask) &&
2626		    lgr->rtokens[i][lnk_idx].rkey == rkey)
2627			return i;
2628	}
2629	return -ENOENT;
2630}
2631
2632/* set rtoken for a new link to an existing rmb */
2633void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2634		    __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2635{
2636	int rtok_idx;
2637
2638	rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2639	if (rtok_idx == -ENOENT)
2640		return;
2641	lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2642	lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2643}
2644
2645/* set rtoken for a new link whose link_id is given */
2646void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2647		     __be64 nw_vaddr, __be32 nw_rkey)
2648{
2649	u64 dma_addr = be64_to_cpu(nw_vaddr);
2650	u32 rkey = ntohl(nw_rkey);
2651	bool found = false;
2652	int link_idx;
2653
2654	for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2655		if (lgr->lnk[link_idx].link_id == link_id) {
2656			found = true;
2657			break;
2658		}
2659	}
2660	if (!found)
2661		return;
2662	lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2663	lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2664}
2665
2666/* add a new rtoken from peer */
2667int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2668{
2669	struct smc_link_group *lgr = smc_get_lgr(lnk);
2670	u64 dma_addr = be64_to_cpu(nw_vaddr);
2671	u32 rkey = ntohl(nw_rkey);
2672	int i;
2673
2674	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2675		if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2676		    lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2677		    test_bit(i, lgr->rtokens_used_mask)) {
2678			/* already in list */
2679			return i;
2680		}
2681	}
2682	i = smc_rmb_reserve_rtoken_idx(lgr);
2683	if (i < 0)
2684		return i;
2685	lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2686	lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2687	return i;
2688}
2689
2690/* delete an rtoken from all links */
2691int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2692{
2693	struct smc_link_group *lgr = smc_get_lgr(lnk);
2694	u32 rkey = ntohl(nw_rkey);
2695	int i, j;
2696
2697	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2698		if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2699		    test_bit(i, lgr->rtokens_used_mask)) {
2700			for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2701				lgr->rtokens[i][j].rkey = 0;
2702				lgr->rtokens[i][j].dma_addr = 0;
2703			}
2704			clear_bit(i, lgr->rtokens_used_mask);
2705			return 0;
2706		}
2707	}
2708	return -ENOENT;
2709}
2710
2711/* save rkey and dma_addr received from peer during clc handshake */
2712int smc_rmb_rtoken_handling(struct smc_connection *conn,
2713			    struct smc_link *lnk,
2714			    struct smc_clc_msg_accept_confirm *clc)
2715{
2716	conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2717					  clc->r0.rmb_rkey);
2718	if (conn->rtoken_idx < 0)
2719		return conn->rtoken_idx;
2720	return 0;
2721}
2722
2723static void smc_core_going_away(void)
2724{
2725	struct smc_ib_device *smcibdev;
2726	struct smcd_dev *smcd;
2727
2728	mutex_lock(&smc_ib_devices.mutex);
2729	list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2730		int i;
2731
2732		for (i = 0; i < SMC_MAX_PORTS; i++)
2733			set_bit(i, smcibdev->ports_going_away);
2734	}
2735	mutex_unlock(&smc_ib_devices.mutex);
2736
2737	mutex_lock(&smcd_dev_list.mutex);
2738	list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2739		smcd->going_away = 1;
2740	}
2741	mutex_unlock(&smcd_dev_list.mutex);
2742}
2743
2744/* Clean up all SMC link groups */
2745static void smc_lgrs_shutdown(void)
2746{
2747	struct smcd_dev *smcd;
2748
2749	smc_core_going_away();
2750
2751	smc_smcr_terminate_all(NULL);
2752
2753	mutex_lock(&smcd_dev_list.mutex);
2754	list_for_each_entry(smcd, &smcd_dev_list.list, list)
2755		smc_smcd_terminate_all(smcd);
2756	mutex_unlock(&smcd_dev_list.mutex);
2757}
2758
2759static int smc_core_reboot_event(struct notifier_block *this,
2760				 unsigned long event, void *ptr)
2761{
2762	smc_lgrs_shutdown();
2763	smc_ib_unregister_client();
2764	smc_ism_exit();
2765	return 0;
2766}
2767
2768static struct notifier_block smc_reboot_notifier = {
2769	.notifier_call = smc_core_reboot_event,
2770};
2771
2772int __init smc_core_init(void)
2773{
2774	return register_reboot_notifier(&smc_reboot_notifier);
2775}
2776
2777/* Called (from smc_exit) when module is removed */
2778void smc_core_exit(void)
2779{
2780	unregister_reboot_notifier(&smc_reboot_notifier);
2781	smc_lgrs_shutdown();
2782}