cm.c

   1// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2/*
   3 * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
   4 * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   5 * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   6 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   7 * Copyright (c) 2019, Mellanox Technologies inc.  All rights reserved.
   8 */
   9
  10#include <linux/completion.h>
  11#include <linux/dma-mapping.h>
  12#include <linux/device.h>
  13#include <linux/module.h>
  14#include <linux/err.h>
  15#include <linux/idr.h>
  16#include <linux/interrupt.h>
  17#include <linux/random.h>
  18#include <linux/rbtree.h>
  19#include <linux/spinlock.h>
  20#include <linux/slab.h>
  21#include <linux/sysfs.h>
  22#include <linux/workqueue.h>
  23#include <linux/kdev_t.h>
  24#include <linux/etherdevice.h>
  25
  26#include <rdma/ib_cache.h>
  27#include <rdma/ib_cm.h>
  28#include <rdma/ib_sysfs.h>
  29#include "cm_msgs.h"
  30#include "core_priv.h"
  31#include "cm_trace.h"
  32
  33MODULE_AUTHOR("Sean Hefty");
  34MODULE_DESCRIPTION("InfiniBand CM");
  35MODULE_LICENSE("Dual BSD/GPL");
  36
  37#define CM_DIRECT_RETRY_CTX ((void *) 1UL)
  38#define CM_MRA_SETTING 24 /* 4.096us * 2^24 = ~68.7 seconds */
  39
  40static const char * const ibcm_rej_reason_strs[] = {
  41	[IB_CM_REJ_NO_QP]			= "no QP",
  42	[IB_CM_REJ_NO_EEC]			= "no EEC",
  43	[IB_CM_REJ_NO_RESOURCES]		= "no resources",
  44	[IB_CM_REJ_TIMEOUT]			= "timeout",
  45	[IB_CM_REJ_UNSUPPORTED]			= "unsupported",
  46	[IB_CM_REJ_INVALID_COMM_ID]		= "invalid comm ID",
  47	[IB_CM_REJ_INVALID_COMM_INSTANCE]	= "invalid comm instance",
  48	[IB_CM_REJ_INVALID_SERVICE_ID]		= "invalid service ID",
  49	[IB_CM_REJ_INVALID_TRANSPORT_TYPE]	= "invalid transport type",
  50	[IB_CM_REJ_STALE_CONN]			= "stale conn",
  51	[IB_CM_REJ_RDC_NOT_EXIST]		= "RDC not exist",
  52	[IB_CM_REJ_INVALID_GID]			= "invalid GID",
  53	[IB_CM_REJ_INVALID_LID]			= "invalid LID",
  54	[IB_CM_REJ_INVALID_SL]			= "invalid SL",
  55	[IB_CM_REJ_INVALID_TRAFFIC_CLASS]	= "invalid traffic class",
  56	[IB_CM_REJ_INVALID_HOP_LIMIT]		= "invalid hop limit",
  57	[IB_CM_REJ_INVALID_PACKET_RATE]		= "invalid packet rate",
  58	[IB_CM_REJ_INVALID_ALT_GID]		= "invalid alt GID",
  59	[IB_CM_REJ_INVALID_ALT_LID]		= "invalid alt LID",
  60	[IB_CM_REJ_INVALID_ALT_SL]		= "invalid alt SL",
  61	[IB_CM_REJ_INVALID_ALT_TRAFFIC_CLASS]	= "invalid alt traffic class",
  62	[IB_CM_REJ_INVALID_ALT_HOP_LIMIT]	= "invalid alt hop limit",
  63	[IB_CM_REJ_INVALID_ALT_PACKET_RATE]	= "invalid alt packet rate",
  64	[IB_CM_REJ_PORT_CM_REDIRECT]		= "port CM redirect",
  65	[IB_CM_REJ_PORT_REDIRECT]		= "port redirect",
  66	[IB_CM_REJ_INVALID_MTU]			= "invalid MTU",
  67	[IB_CM_REJ_INSUFFICIENT_RESP_RESOURCES]	= "insufficient resp resources",
  68	[IB_CM_REJ_CONSUMER_DEFINED]		= "consumer defined",
  69	[IB_CM_REJ_INVALID_RNR_RETRY]		= "invalid RNR retry",
  70	[IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID]	= "duplicate local comm ID",
  71	[IB_CM_REJ_INVALID_CLASS_VERSION]	= "invalid class version",
  72	[IB_CM_REJ_INVALID_FLOW_LABEL]		= "invalid flow label",
  73	[IB_CM_REJ_INVALID_ALT_FLOW_LABEL]	= "invalid alt flow label",
  74	[IB_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED] =
  75		"vendor option is not supported",
  76};
  77
  78const char *__attribute_const__ ibcm_reject_msg(int reason)
  79{
  80	size_t index = reason;
  81
  82	if (index < ARRAY_SIZE(ibcm_rej_reason_strs) &&
  83	    ibcm_rej_reason_strs[index])
  84		return ibcm_rej_reason_strs[index];
  85	else
  86		return "unrecognized reason";
  87}
  88EXPORT_SYMBOL(ibcm_reject_msg);
  89
  90struct cm_id_private;
  91struct cm_work;
  92static int cm_add_one(struct ib_device *device);
  93static void cm_remove_one(struct ib_device *device, void *client_data);
  94static void cm_process_work(struct cm_id_private *cm_id_priv,
  95			    struct cm_work *work);
  96static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv,
  97				   struct ib_cm_sidr_rep_param *param);
  98static void cm_issue_dreq(struct cm_id_private *cm_id_priv);
  99static int cm_send_drep_locked(struct cm_id_private *cm_id_priv,
 100			       void *private_data, u8 private_data_len);
 101static int cm_send_rej_locked(struct cm_id_private *cm_id_priv,
 102			      enum ib_cm_rej_reason reason, void *ari,
 103			      u8 ari_length, const void *private_data,
 104			      u8 private_data_len);
 105
 106static struct ib_client cm_client = {
 107	.name   = "cm",
 108	.add    = cm_add_one,
 109	.remove = cm_remove_one
 110};
 111
 112static struct ib_cm {
 113	spinlock_t lock;
 114	struct list_head device_list;
 115	rwlock_t device_lock;
 116	struct rb_root listen_service_table;
 117	u64 listen_service_id;
 118	/* struct rb_root peer_service_table; todo: fix peer to peer */
 119	struct rb_root remote_qp_table;
 120	struct rb_root remote_id_table;
 121	struct rb_root remote_sidr_table;
 122	struct xarray local_id_table;
 123	u32 local_id_next;
 124	__be32 random_id_operand;
 125	struct list_head timewait_list;
 126	struct workqueue_struct *wq;
 127} cm;
 128
 129/* Counter indexes ordered by attribute ID */
 130enum {
 131	CM_REQ_COUNTER,
 132	CM_MRA_COUNTER,
 133	CM_REJ_COUNTER,
 134	CM_REP_COUNTER,
 135	CM_RTU_COUNTER,
 136	CM_DREQ_COUNTER,
 137	CM_DREP_COUNTER,
 138	CM_SIDR_REQ_COUNTER,
 139	CM_SIDR_REP_COUNTER,
 140	CM_LAP_COUNTER,
 141	CM_APR_COUNTER,
 142	CM_ATTR_COUNT,
 143	CM_ATTR_ID_OFFSET = 0x0010,
 144};
 145
 146enum {
 147	CM_XMIT,
 148	CM_XMIT_RETRIES,
 149	CM_RECV,
 150	CM_RECV_DUPLICATES,
 151	CM_COUNTER_GROUPS
 152};
 153
 154struct cm_counter_attribute {
 155	struct ib_port_attribute attr;
 156	unsigned short group;
 157	unsigned short index;
 158};
 159
 160struct cm_port {
 161	struct cm_device *cm_dev;
 162	struct ib_mad_agent *mad_agent;
 163	struct ib_mad_agent *rep_agent;
 164	u32 port_num;
 165	atomic_long_t counters[CM_COUNTER_GROUPS][CM_ATTR_COUNT];
 166};
 167
 168struct cm_device {
 169	struct kref kref;
 170	struct list_head list;
 171	rwlock_t mad_agent_lock;
 172	struct ib_device *ib_device;
 173	u8 ack_delay;
 174	int going_down;
 175	struct cm_port *port[];
 176};
 177
 178struct cm_av {
 179	struct cm_port *port;
 180	struct rdma_ah_attr ah_attr;
 181	u16 dlid_datapath;
 182	u16 pkey_index;
 183	u8 timeout;
 184};
 185
 186struct cm_work {
 187	struct delayed_work work;
 188	struct list_head list;
 189	struct cm_port *port;
 190	struct ib_mad_recv_wc *mad_recv_wc;	/* Received MADs */
 191	__be32 local_id;			/* Established / timewait */
 192	__be32 remote_id;
 193	struct ib_cm_event cm_event;
 194	struct sa_path_rec path[];
 195};
 196
 197struct cm_timewait_info {
 198	struct cm_work work;
 199	struct list_head list;
 200	struct rb_node remote_qp_node;
 201	struct rb_node remote_id_node;
 202	__be64 remote_ca_guid;
 203	__be32 remote_qpn;
 204	u8 inserted_remote_qp;
 205	u8 inserted_remote_id;
 206};
 207
 208struct cm_id_private {
 209	struct ib_cm_id	id;
 210
 211	struct rb_node service_node;
 212	struct rb_node sidr_id_node;
 213	u32 sidr_slid;
 214	spinlock_t lock;	/* Do not acquire inside cm.lock */
 215	struct completion comp;
 216	refcount_t refcount;
 217	/* Number of clients sharing this ib_cm_id. Only valid for listeners.
 218	 * Protected by the cm.lock spinlock.
 219	 */
 220	int listen_sharecount;
 221	struct rcu_head rcu;
 222
 223	struct ib_mad_send_buf *msg;
 224	struct cm_timewait_info *timewait_info;
 225	/* todo: use alternate port on send failure */
 226	struct cm_av av;
 227	struct cm_av alt_av;
 228
 229	void *private_data;
 230	__be64 tid;
 231	__be32 local_qpn;
 232	__be32 remote_qpn;
 233	enum ib_qp_type qp_type;
 234	__be32 sq_psn;
 235	__be32 rq_psn;
 236	int timeout_ms;
 237	enum ib_mtu path_mtu;
 238	__be16 pkey;
 239	u8 private_data_len;
 240	u8 max_cm_retries;
 241	u8 responder_resources;
 242	u8 initiator_depth;
 243	u8 retry_count;
 244	u8 rnr_retry_count;
 245	u8 target_ack_delay;
 246
 247	struct list_head work_list;
 248	atomic_t work_count;
 249
 250	struct rdma_ucm_ece ece;
 251};
 252
 253static void cm_dev_release(struct kref *kref)
 254{
 255	struct cm_device *cm_dev = container_of(kref, struct cm_device, kref);
 256	u32 i;
 257
 258	rdma_for_each_port(cm_dev->ib_device, i)
 259		kfree(cm_dev->port[i - 1]);
 260
 261	kfree(cm_dev);
 262}
 263
 264static void cm_device_put(struct cm_device *cm_dev)
 265{
 266	kref_put(&cm_dev->kref, cm_dev_release);
 267}
 268
 269static void cm_work_handler(struct work_struct *work);
 270
 271static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 272{
 273	if (refcount_dec_and_test(&cm_id_priv->refcount))
 274		complete(&cm_id_priv->comp);
 275}
 276
 277static struct ib_mad_send_buf *
 278cm_alloc_msg_agent(struct cm_id_private *cm_id_priv, bool rep_agent)
 279{
 280	struct ib_mad_agent *mad_agent;
 281	struct ib_mad_send_buf *m;
 282	struct ib_ah *ah;
 283
 284	lockdep_assert_held(&cm_id_priv->lock);
 285
 286	if (!cm_id_priv->av.port)
 287		return ERR_PTR(-EINVAL);
 288
 289	read_lock(&cm_id_priv->av.port->cm_dev->mad_agent_lock);
 290	mad_agent = rep_agent ? cm_id_priv->av.port->rep_agent :
 291				cm_id_priv->av.port->mad_agent;
 292	if (!mad_agent) {
 293		m = ERR_PTR(-EINVAL);
 294		goto out;
 295	}
 296
 297	ah = rdma_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr, 0);
 298	if (IS_ERR(ah)) {
 299		m = ERR_CAST(ah);
 300		goto out;
 301	}
 302
 303	m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 304			       cm_id_priv->av.pkey_index,
 305			       0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 306			       GFP_ATOMIC,
 307			       IB_MGMT_BASE_VERSION);
 308	if (IS_ERR(m)) {
 309		rdma_destroy_ah(ah, 0);
 310		goto out;
 311	}
 312
 313	m->ah = ah;
 314
 315out:
 316	read_unlock(&cm_id_priv->av.port->cm_dev->mad_agent_lock);
 317	return m;
 318}
 319
 320static struct ib_mad_send_buf *cm_alloc_msg(struct cm_id_private *cm_id_priv)
 321{
 322	return cm_alloc_msg_agent(cm_id_priv, false);
 323}
 324
 325static void cm_free_msg(struct ib_mad_send_buf *msg)
 326{
 327	if (msg->ah)
 328		rdma_destroy_ah(msg->ah, 0);
 329	ib_free_send_mad(msg);
 330}
 331
 332static struct ib_mad_send_buf *
 333cm_alloc_priv_msg_rep(struct cm_id_private *cm_id_priv, enum ib_cm_state state,
 334		      bool rep_agent)
 335{
 336	struct ib_mad_send_buf *msg;
 337
 338	lockdep_assert_held(&cm_id_priv->lock);
 339
 340	msg = cm_alloc_msg_agent(cm_id_priv, rep_agent);
 341	if (IS_ERR(msg))
 342		return msg;
 343
 344	cm_id_priv->msg = msg;
 345	refcount_inc(&cm_id_priv->refcount);
 346	msg->context[0] = cm_id_priv;
 347	msg->context[1] = (void *) (unsigned long) state;
 348
 349	msg->retries = cm_id_priv->max_cm_retries;
 350	msg->timeout_ms = cm_id_priv->timeout_ms;
 351
 352	return msg;
 353}
 354
 355static struct ib_mad_send_buf *
 356cm_alloc_priv_msg(struct cm_id_private *cm_id_priv, enum ib_cm_state state)
 357{
 358	return cm_alloc_priv_msg_rep(cm_id_priv, state, false);
 359}
 360
 361static void cm_free_priv_msg(struct ib_mad_send_buf *msg)
 362{
 363	struct cm_id_private *cm_id_priv = msg->context[0];
 364
 365	lockdep_assert_held(&cm_id_priv->lock);
 366
 367	if (!WARN_ON(cm_id_priv->msg != msg))
 368		cm_id_priv->msg = NULL;
 369
 370	if (msg->ah)
 371		rdma_destroy_ah(msg->ah, 0);
 372	cm_deref_id(cm_id_priv);
 373	ib_free_send_mad(msg);
 374}
 375
 376static struct ib_mad_send_buf *
 377cm_alloc_response_msg_no_ah(struct cm_port *port,
 378			    struct ib_mad_recv_wc *mad_recv_wc,
 379			    bool direct_retry)
 380{
 381	struct ib_mad_send_buf *m;
 382
 383	m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 384			       0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 385			       GFP_ATOMIC, IB_MGMT_BASE_VERSION);
 386	if (!IS_ERR(m))
 387		m->context[0] = direct_retry ? CM_DIRECT_RETRY_CTX : NULL;
 388
 389	return m;
 390}
 391
 392static int cm_create_response_msg_ah(struct cm_port *port,
 393				     struct ib_mad_recv_wc *mad_recv_wc,
 394				     struct ib_mad_send_buf *msg)
 395{
 396	struct ib_ah *ah;
 397
 398	ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 399				  mad_recv_wc->recv_buf.grh, port->port_num);
 400	if (IS_ERR(ah))
 401		return PTR_ERR(ah);
 402
 403	msg->ah = ah;
 404	return 0;
 405}
 406
 407static int cm_alloc_response_msg(struct cm_port *port,
 408				 struct ib_mad_recv_wc *mad_recv_wc,
 409				 bool direct_retry,
 410				 struct ib_mad_send_buf **msg)
 411{
 412	struct ib_mad_send_buf *m;
 413	int ret;
 414
 415	m = cm_alloc_response_msg_no_ah(port, mad_recv_wc, direct_retry);
 416	if (IS_ERR(m))
 417		return PTR_ERR(m);
 418
 419	ret = cm_create_response_msg_ah(port, mad_recv_wc, m);
 420	if (ret) {
 421		ib_free_send_mad(m);
 422		return ret;
 423	}
 424
 425	*msg = m;
 426	return 0;
 427}
 428
 429static void *cm_copy_private_data(const void *private_data, u8 private_data_len)
 430{
 431	void *data;
 432
 433	if (!private_data || !private_data_len)
 434		return NULL;
 435
 436	data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 437	if (!data)
 438		return ERR_PTR(-ENOMEM);
 439
 440	return data;
 441}
 442
 443static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 444				 void *private_data, u8 private_data_len)
 445{
 446	if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 447		kfree(cm_id_priv->private_data);
 448
 449	cm_id_priv->private_data = private_data;
 450	cm_id_priv->private_data_len = private_data_len;
 451}
 452
 453static void cm_set_av_port(struct cm_av *av, struct cm_port *port)
 454{
 455	struct cm_port *old_port = av->port;
 456
 457	if (old_port == port)
 458		return;
 459
 460	av->port = port;
 461	if (old_port)
 462		cm_device_put(old_port->cm_dev);
 463	if (port)
 464		kref_get(&port->cm_dev->kref);
 465}
 466
 467static void cm_init_av_for_lap(struct cm_port *port, struct ib_wc *wc,
 468			       struct rdma_ah_attr *ah_attr, struct cm_av *av)
 469{
 470	cm_set_av_port(av, port);
 471	av->pkey_index = wc->pkey_index;
 472	rdma_move_ah_attr(&av->ah_attr, ah_attr);
 473}
 474
 475static int cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 476				   struct ib_grh *grh, struct cm_av *av)
 477{
 478	cm_set_av_port(av, port);
 479	av->pkey_index = wc->pkey_index;
 480	return ib_init_ah_attr_from_wc(port->cm_dev->ib_device,
 481				       port->port_num, wc,
 482				       grh, &av->ah_attr);
 483}
 484
 485static struct cm_port *
 486get_cm_port_from_path(struct sa_path_rec *path, const struct ib_gid_attr *attr)
 487{
 488	struct cm_device *cm_dev;
 489	struct cm_port *port = NULL;
 490	unsigned long flags;
 491
 492	if (attr) {
 493		read_lock_irqsave(&cm.device_lock, flags);
 494		list_for_each_entry(cm_dev, &cm.device_list, list) {
 495			if (cm_dev->ib_device == attr->device) {
 496				port = cm_dev->port[attr->port_num - 1];
 497				break;
 498			}
 499		}
 500		read_unlock_irqrestore(&cm.device_lock, flags);
 501	} else {
 502		/* SGID attribute can be NULL in following
 503		 * conditions.
 504		 * (a) Alternative path
 505		 * (b) IB link layer without GRH
 506		 * (c) LAP send messages
 507		 */
 508		read_lock_irqsave(&cm.device_lock, flags);
 509		list_for_each_entry(cm_dev, &cm.device_list, list) {
 510			attr = rdma_find_gid(cm_dev->ib_device,
 511					     &path->sgid,
 512					     sa_conv_pathrec_to_gid_type(path),
 513					     NULL);
 514			if (!IS_ERR(attr)) {
 515				port = cm_dev->port[attr->port_num - 1];
 516				break;
 517			}
 518		}
 519		read_unlock_irqrestore(&cm.device_lock, flags);
 520		if (port)
 521			rdma_put_gid_attr(attr);
 522	}
 523	return port;
 524}
 525
 526static int cm_init_av_by_path(struct sa_path_rec *path,
 527			      const struct ib_gid_attr *sgid_attr,
 528			      struct cm_av *av)
 529{
 530	struct rdma_ah_attr new_ah_attr;
 531	struct cm_device *cm_dev;
 532	struct cm_port *port;
 533	int ret;
 534
 535	port = get_cm_port_from_path(path, sgid_attr);
 536	if (!port)
 537		return -EINVAL;
 538	cm_dev = port->cm_dev;
 539
 540	ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 541				  be16_to_cpu(path->pkey), &av->pkey_index);
 542	if (ret)
 543		return ret;
 544
 545	cm_set_av_port(av, port);
 546
 547	/*
 548	 * av->ah_attr might be initialized based on wc or during
 549	 * request processing time which might have reference to sgid_attr.
 550	 * So initialize a new ah_attr on stack.
 551	 * If initialization fails, old ah_attr is used for sending any
 552	 * responses. If initialization is successful, than new ah_attr
 553	 * is used by overwriting the old one. So that right ah_attr
 554	 * can be used to return an error response.
 555	 */
 556	ret = ib_init_ah_attr_from_path(cm_dev->ib_device, port->port_num, path,
 557					&new_ah_attr, sgid_attr);
 558	if (ret)
 559		return ret;
 560
 561	av->timeout = path->packet_life_time + 1;
 562	rdma_move_ah_attr(&av->ah_attr, &new_ah_attr);
 563	return 0;
 564}
 565
 566/* Move av created by cm_init_av_by_path(), so av.dgid is not moved */
 567static void cm_move_av_from_path(struct cm_av *dest, struct cm_av *src)
 568{
 569	cm_set_av_port(dest, src->port);
 570	cm_set_av_port(src, NULL);
 571	dest->pkey_index = src->pkey_index;
 572	rdma_move_ah_attr(&dest->ah_attr, &src->ah_attr);
 573	dest->timeout = src->timeout;
 574}
 575
 576static void cm_destroy_av(struct cm_av *av)
 577{
 578	rdma_destroy_ah_attr(&av->ah_attr);
 579	cm_set_av_port(av, NULL);
 580}
 581
 582static u32 cm_local_id(__be32 local_id)
 583{
 584	return (__force u32) (local_id ^ cm.random_id_operand);
 585}
 586
 587static struct cm_id_private *cm_acquire_id(__be32 local_id, __be32 remote_id)
 588{
 589	struct cm_id_private *cm_id_priv;
 590
 591	rcu_read_lock();
 592	cm_id_priv = xa_load(&cm.local_id_table, cm_local_id(local_id));
 593	if (!cm_id_priv || cm_id_priv->id.remote_id != remote_id ||
 594	    !refcount_inc_not_zero(&cm_id_priv->refcount))
 595		cm_id_priv = NULL;
 596	rcu_read_unlock();
 597
 598	return cm_id_priv;
 599}
 600
 601/*
 602 * Trivial helpers to strip endian annotation and compare; the
 603 * endianness doesn't actually matter since we just need a stable
 604 * order for the RB tree.
 605 */
 606static int be32_lt(__be32 a, __be32 b)
 607{
 608	return (__force u32) a < (__force u32) b;
 609}
 610
 611static int be32_gt(__be32 a, __be32 b)
 612{
 613	return (__force u32) a > (__force u32) b;
 614}
 615
 616static int be64_lt(__be64 a, __be64 b)
 617{
 618	return (__force u64) a < (__force u64) b;
 619}
 620
 621static int be64_gt(__be64 a, __be64 b)
 622{
 623	return (__force u64) a > (__force u64) b;
 624}
 625
 626/*
 627 * Inserts a new cm_id_priv into the listen_service_table. Returns cm_id_priv
 628 * if the new ID was inserted, NULL if it could not be inserted due to a
 629 * collision, or the existing cm_id_priv ready for shared usage.
 630 */
 631static struct cm_id_private *cm_insert_listen(struct cm_id_private *cm_id_priv,
 632					      ib_cm_handler shared_handler)
 633{
 634	struct rb_node **link = &cm.listen_service_table.rb_node;
 635	struct rb_node *parent = NULL;
 636	struct cm_id_private *cur_cm_id_priv;
 637	__be64 service_id = cm_id_priv->id.service_id;
 638	unsigned long flags;
 639
 640	spin_lock_irqsave(&cm.lock, flags);
 641	while (*link) {
 642		parent = *link;
 643		cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 644					  service_node);
 645
 646		if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 647			link = &(*link)->rb_left;
 648		else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 649			link = &(*link)->rb_right;
 650		else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 651			link = &(*link)->rb_left;
 652		else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 653			link = &(*link)->rb_right;
 654		else {
 655			/*
 656			 * Sharing an ib_cm_id with different handlers is not
 657			 * supported
 658			 */
 659			if (cur_cm_id_priv->id.cm_handler != shared_handler ||
 660			    cur_cm_id_priv->id.context ||
 661			    WARN_ON(!cur_cm_id_priv->id.cm_handler)) {
 662				spin_unlock_irqrestore(&cm.lock, flags);
 663				return NULL;
 664			}
 665			refcount_inc(&cur_cm_id_priv->refcount);
 666			cur_cm_id_priv->listen_sharecount++;
 667			spin_unlock_irqrestore(&cm.lock, flags);
 668			return cur_cm_id_priv;
 669		}
 670	}
 671	cm_id_priv->listen_sharecount++;
 672	rb_link_node(&cm_id_priv->service_node, parent, link);
 673	rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 674	spin_unlock_irqrestore(&cm.lock, flags);
 675	return cm_id_priv;
 676}
 677
 678static struct cm_id_private *cm_find_listen(struct ib_device *device,
 679					    __be64 service_id)
 680{
 681	struct rb_node *node = cm.listen_service_table.rb_node;
 682	struct cm_id_private *cm_id_priv;
 683
 684	while (node) {
 685		cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 686
 687		if (device < cm_id_priv->id.device)
 688			node = node->rb_left;
 689		else if (device > cm_id_priv->id.device)
 690			node = node->rb_right;
 691		else if (be64_lt(service_id, cm_id_priv->id.service_id))
 692			node = node->rb_left;
 693		else if (be64_gt(service_id, cm_id_priv->id.service_id))
 694			node = node->rb_right;
 695		else {
 696			refcount_inc(&cm_id_priv->refcount);
 697			return cm_id_priv;
 698		}
 699	}
 700	return NULL;
 701}
 702
 703static struct cm_timewait_info *
 704cm_insert_remote_id(struct cm_timewait_info *timewait_info)
 705{
 706	struct rb_node **link = &cm.remote_id_table.rb_node;
 707	struct rb_node *parent = NULL;
 708	struct cm_timewait_info *cur_timewait_info;
 709	__be64 remote_ca_guid = timewait_info->remote_ca_guid;
 710	__be32 remote_id = timewait_info->work.remote_id;
 711
 712	while (*link) {
 713		parent = *link;
 714		cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 715					     remote_id_node);
 716		if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 717			link = &(*link)->rb_left;
 718		else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 719			link = &(*link)->rb_right;
 720		else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 721			link = &(*link)->rb_left;
 722		else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 723			link = &(*link)->rb_right;
 724		else
 725			return cur_timewait_info;
 726	}
 727	timewait_info->inserted_remote_id = 1;
 728	rb_link_node(&timewait_info->remote_id_node, parent, link);
 729	rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 730	return NULL;
 731}
 732
 733static struct cm_id_private *cm_find_remote_id(__be64 remote_ca_guid,
 734					       __be32 remote_id)
 735{
 736	struct rb_node *node = cm.remote_id_table.rb_node;
 737	struct cm_timewait_info *timewait_info;
 738	struct cm_id_private *res = NULL;
 739
 740	spin_lock_irq(&cm.lock);
 741	while (node) {
 742		timewait_info = rb_entry(node, struct cm_timewait_info,
 743					 remote_id_node);
 744		if (be32_lt(remote_id, timewait_info->work.remote_id))
 745			node = node->rb_left;
 746		else if (be32_gt(remote_id, timewait_info->work.remote_id))
 747			node = node->rb_right;
 748		else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 749			node = node->rb_left;
 750		else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 751			node = node->rb_right;
 752		else {
 753			res = cm_acquire_id(timewait_info->work.local_id,
 754					     timewait_info->work.remote_id);
 755			break;
 756		}
 757	}
 758	spin_unlock_irq(&cm.lock);
 759	return res;
 760}
 761
 762static struct cm_timewait_info *
 763cm_insert_remote_qpn(struct cm_timewait_info *timewait_info)
 764{
 765	struct rb_node **link = &cm.remote_qp_table.rb_node;
 766	struct rb_node *parent = NULL;
 767	struct cm_timewait_info *cur_timewait_info;
 768	__be64 remote_ca_guid = timewait_info->remote_ca_guid;
 769	__be32 remote_qpn = timewait_info->remote_qpn;
 770
 771	while (*link) {
 772		parent = *link;
 773		cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 774					     remote_qp_node);
 775		if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 776			link = &(*link)->rb_left;
 777		else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 778			link = &(*link)->rb_right;
 779		else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 780			link = &(*link)->rb_left;
 781		else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 782			link = &(*link)->rb_right;
 783		else
 784			return cur_timewait_info;
 785	}
 786	timewait_info->inserted_remote_qp = 1;
 787	rb_link_node(&timewait_info->remote_qp_node, parent, link);
 788	rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 789	return NULL;
 790}
 791
 792static struct cm_id_private *
 793cm_insert_remote_sidr(struct cm_id_private *cm_id_priv)
 794{
 795	struct rb_node **link = &cm.remote_sidr_table.rb_node;
 796	struct rb_node *parent = NULL;
 797	struct cm_id_private *cur_cm_id_priv;
 798	__be32 remote_id = cm_id_priv->id.remote_id;
 799
 800	while (*link) {
 801		parent = *link;
 802		cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 803					  sidr_id_node);
 804		if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 805			link = &(*link)->rb_left;
 806		else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 807			link = &(*link)->rb_right;
 808		else {
 809			if (cur_cm_id_priv->sidr_slid < cm_id_priv->sidr_slid)
 810				link = &(*link)->rb_left;
 811			else if (cur_cm_id_priv->sidr_slid > cm_id_priv->sidr_slid)
 812				link = &(*link)->rb_right;
 813			else
 814				return cur_cm_id_priv;
 815		}
 816	}
 817	rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 818	rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 819	return NULL;
 820}
 821
 822static struct cm_id_private *cm_alloc_id_priv(struct ib_device *device,
 823					      ib_cm_handler cm_handler,
 824					      void *context)
 825{
 826	struct cm_id_private *cm_id_priv;
 827	u32 id;
 828	int ret;
 829
 830	cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 831	if (!cm_id_priv)
 832		return ERR_PTR(-ENOMEM);
 833
 834	cm_id_priv->id.state = IB_CM_IDLE;
 835	cm_id_priv->id.device = device;
 836	cm_id_priv->id.cm_handler = cm_handler;
 837	cm_id_priv->id.context = context;
 838	cm_id_priv->id.remote_cm_qpn = 1;
 839
 840	RB_CLEAR_NODE(&cm_id_priv->service_node);
 841	RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
 842	spin_lock_init(&cm_id_priv->lock);
 843	init_completion(&cm_id_priv->comp);
 844	INIT_LIST_HEAD(&cm_id_priv->work_list);
 845	atomic_set(&cm_id_priv->work_count, -1);
 846	refcount_set(&cm_id_priv->refcount, 1);
 847
 848	ret = xa_alloc_cyclic(&cm.local_id_table, &id, NULL, xa_limit_32b,
 849			      &cm.local_id_next, GFP_KERNEL);
 850	if (ret < 0)
 851		goto error;
 852	cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 853
 854	return cm_id_priv;
 855
 856error:
 857	kfree(cm_id_priv);
 858	return ERR_PTR(ret);
 859}
 860
 861/*
 862 * Make the ID visible to the MAD handlers and other threads that use the
 863 * xarray.
 864 */
 865static void cm_finalize_id(struct cm_id_private *cm_id_priv)
 866{
 867	xa_store(&cm.local_id_table, cm_local_id(cm_id_priv->id.local_id),
 868		 cm_id_priv, GFP_ATOMIC);
 869}
 870
 871struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 872				 ib_cm_handler cm_handler,
 873				 void *context)
 874{
 875	struct cm_id_private *cm_id_priv;
 876
 877	cm_id_priv = cm_alloc_id_priv(device, cm_handler, context);
 878	if (IS_ERR(cm_id_priv))
 879		return ERR_CAST(cm_id_priv);
 880
 881	cm_finalize_id(cm_id_priv);
 882	return &cm_id_priv->id;
 883}
 884EXPORT_SYMBOL(ib_create_cm_id);
 885
 886static struct cm_work *cm_dequeue_work(struct cm_id_private *cm_id_priv)
 887{
 888	struct cm_work *work;
 889
 890	if (list_empty(&cm_id_priv->work_list))
 891		return NULL;
 892
 893	work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 894	list_del(&work->list);
 895	return work;
 896}
 897
 898static void cm_free_work(struct cm_work *work)
 899{
 900	if (work->mad_recv_wc)
 901		ib_free_recv_mad(work->mad_recv_wc);
 902	kfree(work);
 903}
 904
 905static void cm_queue_work_unlock(struct cm_id_private *cm_id_priv,
 906				 struct cm_work *work)
 907	__releases(&cm_id_priv->lock)
 908{
 909	bool immediate;
 910
 911	/*
 912	 * To deliver the event to the user callback we have the drop the
 913	 * spinlock, however, we need to ensure that the user callback is single
 914	 * threaded and receives events in the temporal order. If there are
 915	 * already events being processed then thread new events onto a list,
 916	 * the thread currently processing will pick them up.
 917	 */
 918	immediate = atomic_inc_and_test(&cm_id_priv->work_count);
 919	if (!immediate) {
 920		list_add_tail(&work->list, &cm_id_priv->work_list);
 921		/*
 922		 * This routine always consumes incoming reference. Once queued
 923		 * to the work_list then a reference is held by the thread
 924		 * currently running cm_process_work() and this reference is not
 925		 * needed.
 926		 */
 927		cm_deref_id(cm_id_priv);
 928	}
 929	spin_unlock_irq(&cm_id_priv->lock);
 930
 931	if (immediate)
 932		cm_process_work(cm_id_priv, work);
 933}
 934
 935static inline int cm_convert_to_ms(int iba_time)
 936{
 937	/* approximate conversion to ms from 4.096us x 2^iba_time */
 938	return 1 << max(iba_time - 8, 0);
 939}
 940
 941/*
 942 * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 943 * Because of how ack_timeout is stored, adding one doubles the timeout.
 944 * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 945 * increment it (round up) only if the other is within 50%.
 946 */
 947static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 948{
 949	int ack_timeout = packet_life_time + 1;
 950
 951	if (ack_timeout >= ca_ack_delay)
 952		ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 953	else
 954		ack_timeout = ca_ack_delay +
 955			      (ack_timeout >= (ca_ack_delay - 1));
 956
 957	return min(31, ack_timeout);
 958}
 959
 960static void cm_remove_remote(struct cm_id_private *cm_id_priv)
 961{
 962	struct cm_timewait_info *timewait_info = cm_id_priv->timewait_info;
 963
 964	if (timewait_info->inserted_remote_id) {
 965		rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 966		timewait_info->inserted_remote_id = 0;
 967	}
 968
 969	if (timewait_info->inserted_remote_qp) {
 970		rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 971		timewait_info->inserted_remote_qp = 0;
 972	}
 973}
 974
 975static struct cm_timewait_info *cm_create_timewait_info(__be32 local_id)
 976{
 977	struct cm_timewait_info *timewait_info;
 978
 979	timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 980	if (!timewait_info)
 981		return ERR_PTR(-ENOMEM);
 982
 983	timewait_info->work.local_id = local_id;
 984	INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 985	timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 986	return timewait_info;
 987}
 988
 989static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 990{
 991	int wait_time;
 992	unsigned long flags;
 993	struct cm_device *cm_dev;
 994
 995	lockdep_assert_held(&cm_id_priv->lock);
 996
 997	cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client);
 998	if (!cm_dev)
 999		return;
1000
1001	spin_lock_irqsave(&cm.lock, flags);
1002	cm_remove_remote(cm_id_priv);
1003	list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
1004	spin_unlock_irqrestore(&cm.lock, flags);
1005
1006	/*
1007	 * The cm_id could be destroyed by the user before we exit timewait.
1008	 * To protect against this, we search for the cm_id after exiting
1009	 * timewait before notifying the user that we've exited timewait.
1010	 */
1011	cm_id_priv->id.state = IB_CM_TIMEWAIT;
1012	wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
1013
1014	/* Check if the device started its remove_one */
1015	spin_lock_irqsave(&cm.lock, flags);
1016	if (!cm_dev->going_down)
1017		queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
1018				   msecs_to_jiffies(wait_time));
1019	spin_unlock_irqrestore(&cm.lock, flags);
1020
1021	/*
1022	 * The timewait_info is converted into a work and gets freed during
1023	 * cm_free_work() in cm_timewait_handler().
1024	 */
1025	BUILD_BUG_ON(offsetof(struct cm_timewait_info, work) != 0);
1026	cm_id_priv->timewait_info = NULL;
1027}
1028
1029static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
1030{
1031	unsigned long flags;
1032
1033	lockdep_assert_held(&cm_id_priv->lock);
1034
1035	cm_id_priv->id.state = IB_CM_IDLE;
1036	if (cm_id_priv->timewait_info) {
1037		spin_lock_irqsave(&cm.lock, flags);
1038		cm_remove_remote(cm_id_priv);
1039		spin_unlock_irqrestore(&cm.lock, flags);
1040		kfree(cm_id_priv->timewait_info);
1041		cm_id_priv->timewait_info = NULL;
1042	}
1043}
1044
1045static noinline void cm_destroy_id_wait_timeout(struct ib_cm_id *cm_id,
1046						enum ib_cm_state old_state)
1047{
1048	struct cm_id_private *cm_id_priv;
1049
1050	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1051	pr_err_ratelimited("%s: cm_id=%p timed out. state %d -> %d, refcnt=%d\n", __func__,
1052			   cm_id, old_state, cm_id->state, refcount_read(&cm_id_priv->refcount));
1053}
1054
1055static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
1056{
1057	struct cm_id_private *cm_id_priv;
1058	enum ib_cm_state old_state;
1059	unsigned long timeout;
1060	struct cm_work *work;
1061	int ret;
1062
1063	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1064	spin_lock_irq(&cm_id_priv->lock);
1065	old_state = cm_id->state;
1066retest:
1067	switch (cm_id->state) {
1068	case IB_CM_LISTEN:
1069		spin_lock(&cm.lock);
1070		if (--cm_id_priv->listen_sharecount > 0) {
1071			/* The id is still shared. */
1072			WARN_ON(refcount_read(&cm_id_priv->refcount) == 1);
1073			spin_unlock(&cm.lock);
1074			spin_unlock_irq(&cm_id_priv->lock);
1075			cm_deref_id(cm_id_priv);
1076			return;
1077		}
1078		cm_id->state = IB_CM_IDLE;
1079		rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
1080		RB_CLEAR_NODE(&cm_id_priv->service_node);
1081		spin_unlock(&cm.lock);
1082		break;
1083	case IB_CM_SIDR_REQ_SENT:
1084		cm_id->state = IB_CM_IDLE;
1085		ib_cancel_mad(cm_id_priv->msg);
1086		break;
1087	case IB_CM_SIDR_REQ_RCVD:
1088		cm_send_sidr_rep_locked(cm_id_priv,
1089					&(struct ib_cm_sidr_rep_param){
1090						.status = IB_SIDR_REJECT });
1091		/* cm_send_sidr_rep_locked will not move to IDLE if it fails */
1092		cm_id->state = IB_CM_IDLE;
1093		break;
1094	case IB_CM_REQ_SENT:
1095	case IB_CM_MRA_REQ_RCVD:
1096		ib_cancel_mad(cm_id_priv->msg);
1097		cm_send_rej_locked(cm_id_priv, IB_CM_REJ_TIMEOUT,
1098				   &cm_id_priv->id.device->node_guid,
1099				   sizeof(cm_id_priv->id.device->node_guid),
1100				   NULL, 0);
1101		break;
1102	case IB_CM_REQ_RCVD:
1103		if (err == -ENOMEM) {
1104			/* Do not reject to allow future retries. */
1105			cm_reset_to_idle(cm_id_priv);
1106		} else {
1107			cm_send_rej_locked(cm_id_priv,
1108					   IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
1109					   NULL, 0);
1110		}
1111		break;
1112	case IB_CM_REP_SENT:
1113	case IB_CM_MRA_REP_RCVD:
1114		ib_cancel_mad(cm_id_priv->msg);
1115		cm_send_rej_locked(cm_id_priv, IB_CM_REJ_CONSUMER_DEFINED, NULL,
1116				   0, NULL, 0);
1117		goto retest;
1118	case IB_CM_MRA_REQ_SENT:
1119	case IB_CM_REP_RCVD:
1120	case IB_CM_MRA_REP_SENT:
1121		cm_send_rej_locked(cm_id_priv, IB_CM_REJ_CONSUMER_DEFINED, NULL,
1122				   0, NULL, 0);
1123		break;
1124	case IB_CM_ESTABLISHED:
1125		if (cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
1126			cm_id->state = IB_CM_IDLE;
1127			break;
1128		}
1129		cm_issue_dreq(cm_id_priv);
1130		cm_enter_timewait(cm_id_priv);
1131		goto retest;
1132	case IB_CM_DREQ_SENT:
1133		ib_cancel_mad(cm_id_priv->msg);
1134		cm_enter_timewait(cm_id_priv);
1135		goto retest;
1136	case IB_CM_DREQ_RCVD:
1137		cm_send_drep_locked(cm_id_priv, NULL, 0);
1138		WARN_ON(cm_id->state != IB_CM_TIMEWAIT);
1139		goto retest;
1140	case IB_CM_TIMEWAIT:
1141		/*
1142		 * The cm_acquire_id in cm_timewait_handler will stop working
1143		 * once we do xa_erase below, so just move to idle here for
1144		 * consistency.
1145		 */
1146		cm_id->state = IB_CM_IDLE;
1147		break;
1148	case IB_CM_IDLE:
1149		break;
1150	}
1151	WARN_ON(cm_id->state != IB_CM_IDLE);
1152
1153	spin_lock(&cm.lock);
1154	/* Required for cleanup paths related cm_req_handler() */
1155	if (cm_id_priv->timewait_info) {
1156		cm_remove_remote(cm_id_priv);
1157		kfree(cm_id_priv->timewait_info);
1158		cm_id_priv->timewait_info = NULL;
1159	}
1160
1161	WARN_ON(cm_id_priv->listen_sharecount);
1162	WARN_ON(!RB_EMPTY_NODE(&cm_id_priv->service_node));
1163	if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node))
1164		rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
1165	spin_unlock(&cm.lock);
1166	spin_unlock_irq(&cm_id_priv->lock);
1167
1168	xa_erase(&cm.local_id_table, cm_local_id(cm_id->local_id));
1169	cm_deref_id(cm_id_priv);
1170	timeout = msecs_to_jiffies((cm_id_priv->max_cm_retries * cm_id_priv->timeout_ms * 5) / 4);
1171	do {
1172		ret = wait_for_completion_timeout(&cm_id_priv->comp, timeout);
1173		if (!ret) /* timeout happened */
1174			cm_destroy_id_wait_timeout(cm_id, old_state);
1175	} while (!ret);
1176
1177	while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
1178		cm_free_work(work);
1179
1180	cm_destroy_av(&cm_id_priv->av);
1181	cm_destroy_av(&cm_id_priv->alt_av);
1182	kfree(cm_id_priv->private_data);
1183	kfree_rcu(cm_id_priv, rcu);
1184}
1185
1186void ib_destroy_cm_id(struct ib_cm_id *cm_id)
1187{
1188	cm_destroy_id(cm_id, 0);
1189}
1190EXPORT_SYMBOL(ib_destroy_cm_id);
1191
1192static int cm_init_listen(struct cm_id_private *cm_id_priv, __be64 service_id)
1193{
1194	if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
1195	    (service_id != IB_CM_ASSIGN_SERVICE_ID))
1196		return -EINVAL;
1197
1198	if (service_id == IB_CM_ASSIGN_SERVICE_ID)
1199		cm_id_priv->id.service_id = cpu_to_be64(cm.listen_service_id++);
1200	else
1201		cm_id_priv->id.service_id = service_id;
1202
1203	return 0;
1204}
1205
1206/**
1207 * ib_cm_listen - Initiates listening on the specified service ID for
1208 *   connection and service ID resolution requests.
1209 * @cm_id: Connection identifier associated with the listen request.
1210 * @service_id: Service identifier matched against incoming connection
1211 *   and service ID resolution requests.  The service ID should be specified
1212 *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1213 *   assign a service ID to the caller.
1214 */
1215int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id)
1216{
1217	struct cm_id_private *cm_id_priv =
1218		container_of(cm_id, struct cm_id_private, id);
1219	unsigned long flags;
1220	int ret;
1221
1222	spin_lock_irqsave(&cm_id_priv->lock, flags);
1223	if (cm_id_priv->id.state != IB_CM_IDLE) {
1224		ret = -EINVAL;
1225		goto out;
1226	}
1227
1228	ret = cm_init_listen(cm_id_priv, service_id);
1229	if (ret)
1230		goto out;
1231
1232	if (!cm_insert_listen(cm_id_priv, NULL)) {
1233		ret = -EBUSY;
1234		goto out;
1235	}
1236
1237	cm_id_priv->id.state = IB_CM_LISTEN;
1238	ret = 0;
1239
1240out:
1241	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1242	return ret;
1243}
1244EXPORT_SYMBOL(ib_cm_listen);
1245
1246/**
1247 * ib_cm_insert_listen - Create a new listening ib_cm_id and listen on
1248 *			 the given service ID.
1249 *
1250 * If there's an existing ID listening on that same device and service ID,
1251 * return it.
1252 *
1253 * @device: Device associated with the cm_id.  All related communication will
1254 * be associated with the specified device.
1255 * @cm_handler: Callback invoked to notify the user of CM events.
1256 * @service_id: Service identifier matched against incoming connection
1257 *   and service ID resolution requests.  The service ID should be specified
1258 *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1259 *   assign a service ID to the caller.
1260 *
1261 * Callers should call ib_destroy_cm_id when done with the listener ID.
1262 */
1263struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
1264				     ib_cm_handler cm_handler,
1265				     __be64 service_id)
1266{
1267	struct cm_id_private *listen_id_priv;
1268	struct cm_id_private *cm_id_priv;
1269	int err = 0;
1270
1271	/* Create an ID in advance, since the creation may sleep */
1272	cm_id_priv = cm_alloc_id_priv(device, cm_handler, NULL);
1273	if (IS_ERR(cm_id_priv))
1274		return ERR_CAST(cm_id_priv);
1275
1276	err = cm_init_listen(cm_id_priv, service_id);
1277	if (err) {
1278		ib_destroy_cm_id(&cm_id_priv->id);
1279		return ERR_PTR(err);
1280	}
1281
1282	spin_lock_irq(&cm_id_priv->lock);
1283	listen_id_priv = cm_insert_listen(cm_id_priv, cm_handler);
1284	if (listen_id_priv != cm_id_priv) {
1285		spin_unlock_irq(&cm_id_priv->lock);
1286		ib_destroy_cm_id(&cm_id_priv->id);
1287		if (!listen_id_priv)
1288			return ERR_PTR(-EINVAL);
1289		return &listen_id_priv->id;
1290	}
1291	cm_id_priv->id.state = IB_CM_LISTEN;
1292	spin_unlock_irq(&cm_id_priv->lock);
1293
1294	/*
1295	 * A listen ID does not need to be in the xarray since it does not
1296	 * receive mads, is not placed in the remote_id or remote_qpn rbtree,
1297	 * and does not enter timewait.
1298	 */
1299
1300	return &cm_id_priv->id;
1301}
1302EXPORT_SYMBOL(ib_cm_insert_listen);
1303
1304static __be64 cm_form_tid(struct cm_id_private *cm_id_priv)
1305{
1306	u64 hi_tid = 0, low_tid;
1307
1308	lockdep_assert_held(&cm_id_priv->lock);
1309
1310	low_tid = (u64)cm_id_priv->id.local_id;
1311	if (!cm_id_priv->av.port)
1312		return cpu_to_be64(low_tid);
1313
1314	read_lock(&cm_id_priv->av.port->cm_dev->mad_agent_lock);
1315	if (cm_id_priv->av.port->mad_agent)
1316		hi_tid = ((u64)cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1317	read_unlock(&cm_id_priv->av.port->cm_dev->mad_agent_lock);
1318	return cpu_to_be64(hi_tid | low_tid);
1319}
1320
1321static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1322			      __be16 attr_id, __be64 tid)
1323{
1324	hdr->base_version  = IB_MGMT_BASE_VERSION;
1325	hdr->mgmt_class	   = IB_MGMT_CLASS_CM;
1326	hdr->class_version = IB_CM_CLASS_VERSION;
1327	hdr->method	   = IB_MGMT_METHOD_SEND;
1328	hdr->attr_id	   = attr_id;
1329	hdr->tid	   = tid;
1330}
1331
1332static void cm_format_mad_ece_hdr(struct ib_mad_hdr *hdr, __be16 attr_id,
1333				  __be64 tid, u32 attr_mod)
1334{
1335	cm_format_mad_hdr(hdr, attr_id, tid);
1336	hdr->attr_mod = cpu_to_be32(attr_mod);
1337}
1338
1339static void cm_format_req(struct cm_req_msg *req_msg,
1340			  struct cm_id_private *cm_id_priv,
1341			  struct ib_cm_req_param *param)
1342{
1343	struct sa_path_rec *pri_path = param->primary_path;
1344	struct sa_path_rec *alt_path = param->alternate_path;
1345	bool pri_ext = false;
1346	__be16 lid;
1347
1348	if (pri_path->rec_type == SA_PATH_REC_TYPE_OPA)
1349		pri_ext = opa_is_extended_lid(pri_path->opa.dlid,
1350					      pri_path->opa.slid);
1351
1352	cm_format_mad_ece_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1353			      cm_form_tid(cm_id_priv), param->ece.attr_mod);
1354
1355	IBA_SET(CM_REQ_LOCAL_COMM_ID, req_msg,
1356		be32_to_cpu(cm_id_priv->id.local_id));
1357	IBA_SET(CM_REQ_SERVICE_ID, req_msg, be64_to_cpu(param->service_id));
1358	IBA_SET(CM_REQ_LOCAL_CA_GUID, req_msg,
1359		be64_to_cpu(cm_id_priv->id.device->node_guid));
1360	IBA_SET(CM_REQ_LOCAL_QPN, req_msg, param->qp_num);
1361	IBA_SET(CM_REQ_INITIATOR_DEPTH, req_msg, param->initiator_depth);
1362	IBA_SET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg,
1363		param->remote_cm_response_timeout);
1364	cm_req_set_qp_type(req_msg, param->qp_type);
1365	IBA_SET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg, param->flow_control);
1366	IBA_SET(CM_REQ_STARTING_PSN, req_msg, param->starting_psn);
1367	IBA_SET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg,
1368		param->local_cm_response_timeout);
1369	IBA_SET(CM_REQ_PARTITION_KEY, req_msg,
1370		be16_to_cpu(param->primary_path->pkey));
1371	IBA_SET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg,
1372		param->primary_path->mtu);
1373	IBA_SET(CM_REQ_MAX_CM_RETRIES, req_msg, param->max_cm_retries);
1374
1375	if (param->qp_type != IB_QPT_XRC_INI) {
1376		IBA_SET(CM_REQ_RESPONDER_RESOURCES, req_msg,
1377			param->responder_resources);
1378		IBA_SET(CM_REQ_RETRY_COUNT, req_msg, param->retry_count);
1379		IBA_SET(CM_REQ_RNR_RETRY_COUNT, req_msg,
1380			param->rnr_retry_count);
1381		IBA_SET(CM_REQ_SRQ, req_msg, param->srq);
1382	}
1383
1384	*IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg) =
1385		pri_path->sgid;
1386	*IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg) =
1387		pri_path->dgid;
1388	if (pri_ext) {
1389		IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg)
1390			->global.interface_id =
1391			OPA_MAKE_ID(be32_to_cpu(pri_path->opa.slid));
1392		IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg)
1393			->global.interface_id =
1394			OPA_MAKE_ID(be32_to_cpu(pri_path->opa.dlid));
1395	}
1396	if (pri_path->hop_limit <= 1) {
1397		IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
1398			be16_to_cpu(pri_ext ? 0 :
1399					      htons(ntohl(sa_path_get_slid(
1400						      pri_path)))));
1401		IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
1402			be16_to_cpu(pri_ext ? 0 :
1403					      htons(ntohl(sa_path_get_dlid(
1404						      pri_path)))));
1405	} else {
1406
1407		if (param->primary_path_inbound) {
1408			lid = param->primary_path_inbound->ib.dlid;
1409			IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
1410				be16_to_cpu(lid));
1411		} else
1412			IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
1413				be16_to_cpu(IB_LID_PERMISSIVE));
1414
1415		/* Work-around until there's a way to obtain remote LID info */
1416		IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
1417			be16_to_cpu(IB_LID_PERMISSIVE));
1418	}
1419	IBA_SET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg,
1420		be32_to_cpu(pri_path->flow_label));
1421	IBA_SET(CM_REQ_PRIMARY_PACKET_RATE, req_msg, pri_path->rate);
1422	IBA_SET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg, pri_path->traffic_class);
1423	IBA_SET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg, pri_path->hop_limit);
1424	IBA_SET(CM_REQ_PRIMARY_SL, req_msg, pri_path->sl);
1425	IBA_SET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg,
1426		(pri_path->hop_limit <= 1));
1427	IBA_SET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg,
1428		cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1429			       pri_path->packet_life_time));
1430
1431	if (alt_path) {
1432		bool alt_ext = false;
1433
1434		if (alt_path->rec_type == SA_PATH_REC_TYPE_OPA)
1435			alt_ext = opa_is_extended_lid(alt_path->opa.dlid,
1436						      alt_path->opa.slid);
1437
1438		*IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg) =
1439			alt_path->sgid;
1440		*IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg) =
1441			alt_path->dgid;
1442		if (alt_ext) {
1443			IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID,
1444					req_msg)
1445				->global.interface_id =
1446				OPA_MAKE_ID(be32_to_cpu(alt_path->opa.slid));
1447			IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID,
1448					req_msg)
1449				->global.interface_id =
1450				OPA_MAKE_ID(be32_to_cpu(alt_path->opa.dlid));
1451		}
1452		if (alt_path->hop_limit <= 1) {
1453			IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
1454				be16_to_cpu(
1455					alt_ext ? 0 :
1456						  htons(ntohl(sa_path_get_slid(
1457							  alt_path)))));
1458			IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
1459				be16_to_cpu(
1460					alt_ext ? 0 :
1461						  htons(ntohl(sa_path_get_dlid(
1462							  alt_path)))));
1463		} else {
1464			IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
1465				be16_to_cpu(IB_LID_PERMISSIVE));
1466			IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
1467				be16_to_cpu(IB_LID_PERMISSIVE));
1468		}
1469		IBA_SET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg,
1470			be32_to_cpu(alt_path->flow_label));
1471		IBA_SET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg, alt_path->rate);
1472		IBA_SET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg,
1473			alt_path->traffic_class);
1474		IBA_SET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg,
1475			alt_path->hop_limit);
1476		IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, alt_path->sl);
1477		IBA_SET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg,
1478			(alt_path->hop_limit <= 1));
1479		IBA_SET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg,
1480			cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1481				       alt_path->packet_life_time));
1482	}
1483	IBA_SET(CM_REQ_VENDOR_ID, req_msg, param->ece.vendor_id);
1484
1485	if (param->private_data && param->private_data_len)
1486		IBA_SET_MEM(CM_REQ_PRIVATE_DATA, req_msg, param->private_data,
1487			    param->private_data_len);
1488}
1489
1490static int cm_validate_req_param(struct ib_cm_req_param *param)
1491{
1492	if (!param->primary_path)
1493		return -EINVAL;
1494
1495	if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1496	    param->qp_type != IB_QPT_XRC_INI)
1497		return -EINVAL;
1498
1499	if (param->private_data &&
1500	    param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1501		return -EINVAL;
1502
1503	if (param->alternate_path &&
1504	    (param->alternate_path->pkey != param->primary_path->pkey ||
1505	     param->alternate_path->mtu != param->primary_path->mtu))
1506		return -EINVAL;
1507
1508	return 0;
1509}
1510
1511int ib_send_cm_req(struct ib_cm_id *cm_id,
1512		   struct ib_cm_req_param *param)
1513{
1514	struct cm_av av = {}, alt_av = {};
1515	struct cm_id_private *cm_id_priv;
1516	struct ib_mad_send_buf *msg;
1517	struct cm_req_msg *req_msg;
1518	unsigned long flags;
1519	int ret;
1520
1521	ret = cm_validate_req_param(param);
1522	if (ret)
1523		return ret;
1524
1525	/* Verify that we're not in timewait. */
1526	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1527	spin_lock_irqsave(&cm_id_priv->lock, flags);
1528	if (cm_id->state != IB_CM_IDLE || WARN_ON(cm_id_priv->timewait_info)) {
1529		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1530		return -EINVAL;
1531	}
1532	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1533
1534	cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1535							    id.local_id);
1536	if (IS_ERR(cm_id_priv->timewait_info)) {
1537		ret = PTR_ERR(cm_id_priv->timewait_info);
1538		cm_id_priv->timewait_info = NULL;
1539		return ret;
1540	}
1541
1542	ret = cm_init_av_by_path(param->primary_path,
1543				 param->ppath_sgid_attr, &av);
1544	if (ret)
1545		return ret;
1546	if (param->alternate_path) {
1547		ret = cm_init_av_by_path(param->alternate_path, NULL,
1548					 &alt_av);
1549		if (ret) {
1550			cm_destroy_av(&av);
1551			return ret;
1552		}
1553	}
1554	cm_id->service_id = param->service_id;
1555	cm_id_priv->timeout_ms = cm_convert_to_ms(
1556				    param->primary_path->packet_life_time) * 2 +
1557				 cm_convert_to_ms(
1558				    param->remote_cm_response_timeout);
1559	cm_id_priv->max_cm_retries = param->max_cm_retries;
1560	cm_id_priv->initiator_depth = param->initiator_depth;
1561	cm_id_priv->responder_resources = param->responder_resources;
1562	cm_id_priv->retry_count = param->retry_count;
1563	cm_id_priv->path_mtu = param->primary_path->mtu;
1564	cm_id_priv->pkey = param->primary_path->pkey;
1565	cm_id_priv->qp_type = param->qp_type;
1566
1567	spin_lock_irqsave(&cm_id_priv->lock, flags);
1568
1569	cm_move_av_from_path(&cm_id_priv->av, &av);
1570	if (param->primary_path_outbound)
1571		cm_id_priv->av.dlid_datapath =
1572			be16_to_cpu(param->primary_path_outbound->ib.dlid);
1573
1574	if (param->alternate_path)
1575		cm_move_av_from_path(&cm_id_priv->alt_av, &alt_av);
1576
1577	msg = cm_alloc_priv_msg(cm_id_priv, IB_CM_REQ_SENT);
1578	if (IS_ERR(msg)) {
1579		ret = PTR_ERR(msg);
1580		goto out_unlock;
1581	}
1582
1583	req_msg = (struct cm_req_msg *)msg->mad;
1584	cm_format_req(req_msg, cm_id_priv, param);
1585	cm_id_priv->tid = req_msg->hdr.tid;
1586
1587	cm_id_priv->local_qpn = cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg));
1588	cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg));
1589
1590	trace_icm_send_req(&cm_id_priv->id);
1591	ret = ib_post_send_mad(msg, NULL);
1592	if (ret)
1593		goto out_free;
1594	BUG_ON(cm_id->state != IB_CM_IDLE);
1595	cm_id->state = IB_CM_REQ_SENT;
1596	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1597	return 0;
1598out_free:
1599	cm_free_priv_msg(msg);
1600out_unlock:
1601	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1602	return ret;
1603}
1604EXPORT_SYMBOL(ib_send_cm_req);
1605
1606static int cm_issue_rej(struct cm_port *port,
1607			struct ib_mad_recv_wc *mad_recv_wc,
1608			enum ib_cm_rej_reason reason,
1609			enum cm_msg_response msg_rejected,
1610			void *ari, u8 ari_length)
1611{
1612	struct ib_mad_send_buf *msg = NULL;
1613	struct cm_rej_msg *rej_msg, *rcv_msg;
1614	int ret;
1615
1616	ret = cm_alloc_response_msg(port, mad_recv_wc, false, &msg);
1617	if (ret)
1618		return ret;
1619
1620	/* We just need common CM header information.  Cast to any message. */
1621	rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1622	rej_msg = (struct cm_rej_msg *) msg->mad;
1623
1624	cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1625	IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg,
1626		IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg));
1627	IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1628		IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg));
1629	IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, msg_rejected);
1630	IBA_SET(CM_REJ_REASON, rej_msg, reason);
1631
1632	if (ari && ari_length) {
1633		IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length);
1634		IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length);
1635	}
1636
1637	trace_icm_issue_rej(
1638		IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg),
1639		IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg));
1640	ret = ib_post_send_mad(msg, NULL);
1641	if (ret)
1642		cm_free_msg(msg);
1643
1644	return ret;
1645}
1646
1647static bool cm_req_has_alt_path(struct cm_req_msg *req_msg)
1648{
1649	return ((cpu_to_be16(
1650			IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg))) ||
1651		(ib_is_opa_gid(IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID,
1652					       req_msg))));
1653}
1654
1655static void cm_path_set_rec_type(struct ib_device *ib_device, u32 port_num,
1656				 struct sa_path_rec *path, union ib_gid *gid)
1657{
1658	if (ib_is_opa_gid(gid) && rdma_cap_opa_ah(ib_device, port_num))
1659		path->rec_type = SA_PATH_REC_TYPE_OPA;
1660	else
1661		path->rec_type = SA_PATH_REC_TYPE_IB;
1662}
1663
1664static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg,
1665					struct sa_path_rec *primary_path,
1666					struct sa_path_rec *alt_path,
1667					struct ib_wc *wc)
1668{
1669	u32 lid;
1670
1671	if (primary_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1672		sa_path_set_dlid(primary_path, wc->slid);
1673		sa_path_set_slid(primary_path,
1674				 IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
1675					 req_msg));
1676	} else {
1677		lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1678			CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg));
1679		sa_path_set_dlid(primary_path, lid);
1680
1681		lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1682			CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg));
1683		sa_path_set_slid(primary_path, lid);
1684	}
1685
1686	if (!cm_req_has_alt_path(req_msg))
1687		return;
1688
1689	if (alt_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1690		sa_path_set_dlid(alt_path,
1691				 IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID,
1692					 req_msg));
1693		sa_path_set_slid(alt_path,
1694				 IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID,
1695					 req_msg));
1696	} else {
1697		lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1698			CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg));
1699		sa_path_set_dlid(alt_path, lid);
1700
1701		lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1702			CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg));
1703		sa_path_set_slid(alt_path, lid);
1704	}
1705}
1706
1707static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1708				     struct sa_path_rec *primary_path,
1709				     struct sa_path_rec *alt_path,
1710				     struct ib_wc *wc)
1711{
1712	primary_path->dgid =
1713		*IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg);
1714	primary_path->sgid =
1715		*IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg);
1716	primary_path->flow_label =
1717		cpu_to_be32(IBA_GET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg));
1718	primary_path->hop_limit = IBA_GET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg);
1719	primary_path->traffic_class =
1720		IBA_GET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg);
1721	primary_path->reversible = 1;
1722	primary_path->pkey =
1723		cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
1724	primary_path->sl = IBA_GET(CM_REQ_PRIMARY_SL, req_msg);
1725	primary_path->mtu_selector = IB_SA_EQ;
1726	primary_path->mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
1727	primary_path->rate_selector = IB_SA_EQ;
1728	primary_path->rate = IBA_GET(CM_REQ_PRIMARY_PACKET_RATE, req_msg);
1729	primary_path->packet_life_time_selector = IB_SA_EQ;
1730	primary_path->packet_life_time =
1731		IBA_GET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg);
1732	primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1733	primary_path->service_id =
1734		cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
1735	if (sa_path_is_roce(primary_path))
1736		primary_path->roce.route_resolved = false;
1737
1738	if (cm_req_has_alt_path(req_msg)) {
1739		alt_path->dgid = *IBA_GET_MEM_PTR(
1740			CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg);
1741		alt_path->sgid = *IBA_GET_MEM_PTR(
1742			CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg);
1743		alt_path->flow_label = cpu_to_be32(
1744			IBA_GET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg));
1745		alt_path->hop_limit =
1746			IBA_GET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg);
1747		alt_path->traffic_class =
1748			IBA_GET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg);
1749		alt_path->reversible = 1;
1750		alt_path->pkey =
1751			cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
1752		alt_path->sl = IBA_GET(CM_REQ_ALTERNATE_SL, req_msg);
1753		alt_path->mtu_selector = IB_SA_EQ;
1754		alt_path->mtu =
1755			IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
1756		alt_path->rate_selector = IB_SA_EQ;
1757		alt_path->rate = IBA_GET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg);
1758		alt_path->packet_life_time_selector = IB_SA_EQ;
1759		alt_path->packet_life_time =
1760			IBA_GET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg);
1761		alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1762		alt_path->service_id =
1763			cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
1764
1765		if (sa_path_is_roce(alt_path))
1766			alt_path->roce.route_resolved = false;
1767	}
1768	cm_format_path_lid_from_req(req_msg, primary_path, alt_path, wc);
1769}
1770
1771static u16 cm_get_bth_pkey(struct cm_work *work)
1772{
1773	struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1774	u32 port_num = work->port->port_num;
1775	u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1776	u16 pkey;
1777	int ret;
1778
1779	ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1780	if (ret) {
1781		dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %u, pkey index %u). %d\n",
1782				     port_num, pkey_index, ret);
1783		return 0;
1784	}
1785
1786	return pkey;
1787}
1788
1789/**
1790 * cm_opa_to_ib_sgid - Convert OPA SGID to IB SGID
1791 * ULPs (such as IPoIB) do not understand OPA GIDs and will
1792 * reject them as the local_gid will not match the sgid. Therefore,
1793 * change the pathrec's SGID to an IB SGID.
1794 *
1795 * @work: Work completion
1796 * @path: Path record
1797 */
1798static void cm_opa_to_ib_sgid(struct cm_work *work,
1799			      struct sa_path_rec *path)
1800{
1801	struct ib_device *dev = work->port->cm_dev->ib_device;
1802	u32 port_num = work->port->port_num;
1803
1804	if (rdma_cap_opa_ah(dev, port_num) &&
1805	    (ib_is_opa_gid(&path->sgid))) {
1806		union ib_gid sgid;
1807
1808		if (rdma_query_gid(dev, port_num, 0, &sgid)) {
1809			dev_warn(&dev->dev,
1810				 "Error updating sgid in CM request\n");
1811			return;
1812		}
1813
1814		path->sgid = sgid;
1815	}
1816}
1817
1818static void cm_format_req_event(struct cm_work *work,
1819				struct cm_id_private *cm_id_priv,
1820				struct ib_cm_id *listen_id)
1821{
1822	struct cm_req_msg *req_msg;
1823	struct ib_cm_req_event_param *param;
1824
1825	req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1826	param = &work->cm_event.param.req_rcvd;
1827	param->listen_id = listen_id;
1828	param->bth_pkey = cm_get_bth_pkey(work);
1829	param->port = cm_id_priv->av.port->port_num;
1830	param->primary_path = &work->path[0];
1831	cm_opa_to_ib_sgid(work, param->primary_path);
1832	if (cm_req_has_alt_path(req_msg)) {
1833		param->alternate_path = &work->path[1];
1834		cm_opa_to_ib_sgid(work, param->alternate_path);
1835	} else {
1836		param->alternate_path = NULL;
1837	}
1838	param->remote_ca_guid =
1839		cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg));
1840	param->remote_qkey = IBA_GET(CM_REQ_LOCAL_Q_KEY, req_msg);
1841	param->remote_qpn = IBA_GET(CM_REQ_LOCAL_QPN, req_msg);
1842	param->qp_type = cm_req_get_qp_type(req_msg);
1843	param->starting_psn = IBA_GET(CM_REQ_STARTING_PSN, req_msg);
1844	param->responder_resources = IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg);
1845	param->initiator_depth = IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg);
1846	param->local_cm_response_timeout =
1847		IBA_GET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg);
1848	param->flow_control = IBA_GET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg);
1849	param->remote_cm_response_timeout =
1850		IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg);
1851	param->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg);
1852	param->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg);
1853	param->srq = IBA_GET(CM_REQ_SRQ, req_msg);
1854	param->ppath_sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr;
1855	param->ece.vendor_id = IBA_GET(CM_REQ_VENDOR_ID, req_msg);
1856	param->ece.attr_mod = be32_to_cpu(req_msg->hdr.attr_mod);
1857
1858	work->cm_event.private_data =
1859		IBA_GET_MEM_PTR(CM_REQ_PRIVATE_DATA, req_msg);
1860}
1861
1862static void cm_process_work(struct cm_id_private *cm_id_priv,
1863			    struct cm_work *work)
1864{
1865	int ret;
1866
1867	/* We will typically only have the current event to report. */
1868	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1869	cm_free_work(work);
1870
1871	while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1872		spin_lock_irq(&cm_id_priv->lock);
1873		work = cm_dequeue_work(cm_id_priv);
1874		spin_unlock_irq(&cm_id_priv->lock);
1875		if (!work)
1876			return;
1877
1878		ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1879						&work->cm_event);
1880		cm_free_work(work);
1881	}
1882	cm_deref_id(cm_id_priv);
1883	if (ret)
1884		cm_destroy_id(&cm_id_priv->id, ret);
1885}
1886
1887static void cm_format_mra(struct cm_mra_msg *mra_msg,
1888			  struct cm_id_private *cm_id_priv,
1889			  enum cm_msg_response msg_mraed,
1890			  const void *private_data, u8 private_data_len)
1891{
1892	cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1893	IBA_SET(CM_MRA_MESSAGE_MRAED, mra_msg, msg_mraed);
1894	IBA_SET(CM_MRA_LOCAL_COMM_ID, mra_msg,
1895		be32_to_cpu(cm_id_priv->id.local_id));
1896	IBA_SET(CM_MRA_REMOTE_COMM_ID, mra_msg,
1897		be32_to_cpu(cm_id_priv->id.remote_id));
1898	IBA_SET(CM_MRA_SERVICE_TIMEOUT, mra_msg, CM_MRA_SETTING);
1899
1900	if (private_data && private_data_len)
1901		IBA_SET_MEM(CM_MRA_PRIVATE_DATA, mra_msg, private_data,
1902			    private_data_len);
1903}
1904
1905static void cm_format_rej(struct cm_rej_msg *rej_msg,
1906			  struct cm_id_private *cm_id_priv,
1907			  enum ib_cm_rej_reason reason, void *ari,
1908			  u8 ari_length, const void *private_data,
1909			  u8 private_data_len, enum ib_cm_state state)
1910{
1911	lockdep_assert_held(&cm_id_priv->lock);
1912
1913	cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1914	IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg,
1915		be32_to_cpu(cm_id_priv->id.remote_id));
1916
1917	switch (state) {
1918	case IB_CM_REQ_RCVD:
1919		IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, be32_to_cpu(0));
1920		IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ);
1921		break;
1922	case IB_CM_MRA_REQ_SENT:
1923		IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1924			be32_to_cpu(cm_id_priv->id.local_id));
1925		IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ);
1926		break;
1927	case IB_CM_REP_RCVD:
1928	case IB_CM_MRA_REP_SENT:
1929		IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1930			be32_to_cpu(cm_id_priv->id.local_id));
1931		IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REP);
1932		break;
1933	default:
1934		IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1935			be32_to_cpu(cm_id_priv->id.local_id));
1936		IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg,
1937			CM_MSG_RESPONSE_OTHER);
1938		break;
1939	}
1940
1941	IBA_SET(CM_REJ_REASON, rej_msg, reason);
1942	if (ari && ari_length) {
1943		IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length);
1944		IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length);
1945	}
1946
1947	if (private_data && private_data_len)
1948		IBA_SET_MEM(CM_REJ_PRIVATE_DATA, rej_msg, private_data,
1949			    private_data_len);
1950}
1951
1952static void cm_dup_req_handler(struct cm_work *work,
1953			       struct cm_id_private *cm_id_priv)
1954{
1955	struct ib_mad_send_buf *msg = NULL;
1956	int ret;
1957
1958	atomic_long_inc(
1959		&work->port->counters[CM_RECV_DUPLICATES][CM_REQ_COUNTER]);
1960
1961	/* Quick state check to discard duplicate REQs. */
1962	spin_lock_irq(&cm_id_priv->lock);
1963	if (cm_id_priv->id.state == IB_CM_REQ_RCVD) {
1964		spin_unlock_irq(&cm_id_priv->lock);
1965		return;
1966	}
1967	spin_unlock_irq(&cm_id_priv->lock);
1968
1969	ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, true, &msg);
1970	if (ret)
1971		return;
1972
1973	spin_lock_irq(&cm_id_priv->lock);
1974	switch (cm_id_priv->id.state) {
1975	case IB_CM_MRA_REQ_SENT:
1976		cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1977			      CM_MSG_RESPONSE_REQ,
1978			      cm_id_priv->private_data,
1979			      cm_id_priv->private_data_len);
1980		break;
1981	case IB_CM_TIMEWAIT:
1982		cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv,
1983			      IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0,
1984			      IB_CM_TIMEWAIT);
1985		break;
1986	default:
1987		goto unlock;
1988	}
1989	spin_unlock_irq(&cm_id_priv->lock);
1990
1991	trace_icm_send_dup_req(&cm_id_priv->id);
1992	ret = ib_post_send_mad(msg, NULL);
1993	if (ret)
1994		goto free;
1995	return;
1996
1997unlock:	spin_unlock_irq(&cm_id_priv->lock);
1998free:	cm_free_msg(msg);
1999}
2000
2001static struct cm_id_private *cm_match_req(struct cm_work *work,
2002					  struct cm_id_private *cm_id_priv)
2003{
2004	struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
2005	struct cm_timewait_info *timewait_info;
2006	struct cm_req_msg *req_msg;
2007
2008	req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
2009
2010	/* Check for possible duplicate REQ. */
2011	spin_lock_irq(&cm.lock);
2012	timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
2013	if (timewait_info) {
2014		cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2015					   timewait_info->work.remote_id);
2016		spin_unlock_irq(&cm.lock);
2017		if (cur_cm_id_priv) {
2018			cm_dup_req_handler(work, cur_cm_id_priv);
2019			cm_deref_id(cur_cm_id_priv);
2020		}
2021		return NULL;
2022	}
2023
2024	/* Check for stale connections. */
2025	timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2026	if (timewait_info) {
2027		cm_remove_remote(cm_id_priv);
2028		cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2029					   timewait_info->work.remote_id);
2030
2031		spin_unlock_irq(&cm.lock);
2032		cm_issue_rej(work->port, work->mad_recv_wc,
2033			     IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
2034			     NULL, 0);
2035		if (cur_cm_id_priv) {
2036			ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0);
2037			cm_deref_id(cur_cm_id_priv);
2038		}
2039		return NULL;
2040	}
2041
2042	/* Find matching listen request. */
2043	listen_cm_id_priv = cm_find_listen(
2044		cm_id_priv->id.device,
2045		cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)));
2046	if (!listen_cm_id_priv) {
2047		cm_remove_remote(cm_id_priv);
2048		spin_unlock_irq(&cm.lock);
2049		cm_issue_rej(work->port, work->mad_recv_wc,
2050			     IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
2051			     NULL, 0);
2052		return NULL;
2053	}
2054	spin_unlock_irq(&cm.lock);
2055	return listen_cm_id_priv;
2056}
2057
2058/*
2059 * Work-around for inter-subnet connections.  If the LIDs are permissive,
2060 * we need to override the LID/SL data in the REQ with the LID information
2061 * in the work completion.
2062 */
2063static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
2064{
2065	if (!IBA_GET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg)) {
2066		if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID,
2067					req_msg)) == IB_LID_PERMISSIVE) {
2068			IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
2069				be16_to_cpu(ib_lid_be16(wc->slid)));
2070			IBA_SET(CM_REQ_PRIMARY_SL, req_msg, wc->sl);
2071		}
2072
2073		if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
2074					req_msg)) == IB_LID_PERMISSIVE)
2075			IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
2076				wc->dlid_path_bits);
2077	}
2078
2079	if (!IBA_GET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg)) {
2080		if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID,
2081					req_msg)) == IB_LID_PERMISSIVE) {
2082			IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
2083				be16_to_cpu(ib_lid_be16(wc->slid)));
2084			IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, wc->sl);
2085		}
2086
2087		if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID,
2088					req_msg)) == IB_LID_PERMISSIVE)
2089			IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
2090				wc->dlid_path_bits);
2091	}
2092}
2093
2094static int cm_req_handler(struct cm_work *work)
2095{
2096	struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
2097	struct cm_req_msg *req_msg;
2098	const struct ib_global_route *grh;
2099	const struct ib_gid_attr *gid_attr;
2100	int ret;
2101
2102	req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
2103
2104	cm_id_priv =
2105		cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL);
2106	if (IS_ERR(cm_id_priv))
2107		return PTR_ERR(cm_id_priv);
2108
2109	cm_id_priv->id.remote_id =
2110		cpu_to_be32(IBA_GET(CM_REQ_LOCAL_COMM_ID, req_msg));
2111	cm_id_priv->id.service_id =
2112		cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
2113	cm_id_priv->tid = req_msg->hdr.tid;
2114	cm_id_priv->timeout_ms = cm_convert_to_ms(
2115		IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg));
2116	cm_id_priv->max_cm_retries = IBA_GET(CM_REQ_MAX_CM_RETRIES, req_msg);
2117	cm_id_priv->remote_qpn =
2118		cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg));
2119	cm_id_priv->initiator_depth =
2120		IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg);
2121	cm_id_priv->responder_resources =
2122		IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg);
2123	cm_id_priv->path_mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
2124	cm_id_priv->pkey = cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
2125	cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg));
2126	cm_id_priv->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg);
2127	cm_id_priv->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg);
2128	cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
2129
2130	ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2131				      work->mad_recv_wc->recv_buf.grh,
2132				      &cm_id_priv->av);
2133	if (ret)
2134		goto destroy;
2135	cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
2136							    id.local_id);
2137	if (IS_ERR(cm_id_priv->timewait_info)) {
2138		ret = PTR_ERR(cm_id_priv->timewait_info);
2139		cm_id_priv->timewait_info = NULL;
2140		goto destroy;
2141	}
2142	cm_id_priv->timewait_info->work.remote_id = cm_id_priv->id.remote_id;
2143	cm_id_priv->timewait_info->remote_ca_guid =
2144		cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg));
2145	cm_id_priv->timewait_info->remote_qpn = cm_id_priv->remote_qpn;
2146
2147	/*
2148	 * Note that the ID pointer is not in the xarray at this point,
2149	 * so this set is only visible to the local thread.
2150	 */
2151	cm_id_priv->id.state = IB_CM_REQ_RCVD;
2152
2153	listen_cm_id_priv = cm_match_req(work, cm_id_priv);
2154	if (!listen_cm_id_priv) {
2155		trace_icm_no_listener_err(&cm_id_priv->id);
2156		cm_id_priv->id.state = IB_CM_IDLE;
2157		ret = -EINVAL;
2158		goto destroy;
2159	}
2160
2161	memset(&work->path[0], 0, sizeof(work->path[0]));
2162	if (cm_req_has_alt_path(req_msg))
2163		memset(&work->path[1], 0, sizeof(work->path[1]));
2164	grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr);
2165	gid_attr = grh->sgid_attr;
2166
2167	if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE) {
2168		work->path[0].rec_type =
2169			sa_conv_gid_to_pathrec_type(gid_attr->gid_type);
2170	} else {
2171		cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
2172		cm_path_set_rec_type(
2173			work->port->cm_dev->ib_device, work->port->port_num,
2174			&work->path[0],
2175			IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID,
2176					req_msg));
2177	}
2178	if (cm_req_has_alt_path(req_msg))
2179		work->path[1].rec_type = work->path[0].rec_type;
2180	cm_format_paths_from_req(req_msg, &work->path[0],
2181				 &work->path[1], work->mad_recv_wc->wc);
2182	if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
2183		sa_path_set_dmac(&work->path[0],
2184				 cm_id_priv->av.ah_attr.roce.dmac);
2185	work->path[0].hop_limit = grh->hop_limit;
2186
2187	/* This destroy call is needed to pair with cm_init_av_for_response */
2188	cm_destroy_av(&cm_id_priv->av);
2189	ret = cm_init_av_by_path(&work->path[0], gid_attr, &cm_id_priv->av);
2190	if (ret) {
2191		int err;
2192
2193		err = rdma_query_gid(work->port->cm_dev->ib_device,
2194				     work->port->port_num, 0,
2195				     &work->path[0].sgid);
2196		if (err)
2197			ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID,
2198				       NULL, 0, NULL, 0);
2199		else
2200			ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID,
2201				       &work->path[0].sgid,
2202				       sizeof(work->path[0].sgid),
2203				       NULL, 0);
2204		goto rejected;
2205	}
2206	if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_IB)
2207		cm_id_priv->av.dlid_datapath =
2208			IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg);
2209
2210	if (cm_req_has_alt_path(req_msg)) {
2211		ret = cm_init_av_by_path(&work->path[1], NULL,
2212					 &cm_id_priv->alt_av);
2213		if (ret) {
2214			ib_send_cm_rej(&cm_id_priv->id,
2215				       IB_CM_REJ_INVALID_ALT_GID,
2216				       &work->path[0].sgid,
2217				       sizeof(work->path[0].sgid), NULL, 0);
2218			goto rejected;
2219		}
2220	}
2221
2222	cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
2223	cm_id_priv->id.context = listen_cm_id_priv->id.context;
2224	cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
2225
2226	/* Now MAD handlers can see the new ID */
2227	spin_lock_irq(&cm_id_priv->lock);
2228	cm_finalize_id(cm_id_priv);
2229
2230	/* Refcount belongs to the event, pairs with cm_process_work() */
2231	refcount_inc(&cm_id_priv->refcount);
2232	cm_queue_work_unlock(cm_id_priv, work);
2233	/*
2234	 * Since this ID was just created and was not made visible to other MAD
2235	 * handlers until the cm_finalize_id() above we know that the
2236	 * cm_process_work() will deliver the event and the listen_cm_id
2237	 * embedded in the event can be derefed here.
2238	 */
2239	cm_deref_id(listen_cm_id_priv);
2240	return 0;
2241
2242rejected:
2243	cm_deref_id(listen_cm_id_priv);
2244destroy:
2245	ib_destroy_cm_id(&cm_id_priv->id);
2246	return ret;
2247}
2248
2249static void cm_format_rep(struct cm_rep_msg *rep_msg,
2250			  struct cm_id_private *cm_id_priv,
2251			  struct ib_cm_rep_param *param)
2252{
2253	cm_format_mad_ece_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid,
2254			      param->ece.attr_mod);
2255	IBA_SET(CM_REP_LOCAL_COMM_ID, rep_msg,
2256		be32_to_cpu(cm_id_priv->id.local_id));
2257	IBA_SET(CM_REP_REMOTE_COMM_ID, rep_msg,
2258		be32_to_cpu(cm_id_priv->id.remote_id));
2259	IBA_SET(CM_REP_STARTING_PSN, rep_msg, param->starting_psn);
2260	IBA_SET(CM_REP_RESPONDER_RESOURCES, rep_msg,
2261		param->responder_resources);
2262	IBA_SET(CM_REP_TARGET_ACK_DELAY, rep_msg,
2263		cm_id_priv->av.port->cm_dev->ack_delay);
2264	IBA_SET(CM_REP_FAILOVER_ACCEPTED, rep_msg, param->failover_accepted);
2265	IBA_SET(CM_REP_RNR_RETRY_COUNT, rep_msg, param->rnr_retry_count);
2266	IBA_SET(CM_REP_LOCAL_CA_GUID, rep_msg,
2267		be64_to_cpu(cm_id_priv->id.device->node_guid));
2268
2269	if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
2270		IBA_SET(CM_REP_INITIATOR_DEPTH, rep_msg,
2271			param->initiator_depth);
2272		IBA_SET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg,
2273			param->flow_control);
2274		IBA_SET(CM_REP_SRQ, rep_msg, param->srq);
2275		IBA_SET(CM_REP_LOCAL_QPN, rep_msg, param->qp_num);
2276	} else {
2277		IBA_SET(CM_REP_SRQ, rep_msg, 1);
2278		IBA_SET(CM_REP_LOCAL_EE_CONTEXT_NUMBER, rep_msg, param->qp_num);
2279	}
2280
2281	IBA_SET(CM_REP_VENDOR_ID_L, rep_msg, param->ece.vendor_id);
2282	IBA_SET(CM_REP_VENDOR_ID_M, rep_msg, param->ece.vendor_id >> 8);
2283	IBA_SET(CM_REP_VENDOR_ID_H, rep_msg, param->ece.vendor_id >> 16);
2284
2285	if (param->private_data && param->private_data_len)
2286		IBA_SET_MEM(CM_REP_PRIVATE_DATA, rep_msg, param->private_data,
2287			    param->private_data_len);
2288}
2289
2290int ib_send_cm_rep(struct ib_cm_id *cm_id,
2291		   struct ib_cm_rep_param *param)
2292{
2293	struct cm_id_private *cm_id_priv;
2294	struct ib_mad_send_buf *msg;
2295	struct cm_rep_msg *rep_msg;
2296	unsigned long flags;
2297	int ret;
2298
2299	if (param->private_data &&
2300	    param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
2301		return -EINVAL;
2302
2303	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2304	spin_lock_irqsave(&cm_id_priv->lock, flags);
2305	if (cm_id->state != IB_CM_REQ_RCVD &&
2306	    cm_id->state != IB_CM_MRA_REQ_SENT) {
2307		trace_icm_send_rep_err(cm_id_priv->id.local_id, cm_id->state);
2308		ret = -EINVAL;
2309		goto out;
2310	}
2311
2312	msg = cm_alloc_priv_msg_rep(cm_id_priv, IB_CM_REP_SENT, true);
2313	if (IS_ERR(msg)) {
2314		ret = PTR_ERR(msg);
2315		goto out;
2316	}
2317
2318	rep_msg = (struct cm_rep_msg *) msg->mad;
2319	cm_format_rep(rep_msg, cm_id_priv, param);
2320
2321	trace_icm_send_rep(cm_id);
2322	ret = ib_post_send_mad(msg, NULL);
2323	if (ret)
2324		goto out_free;
2325
2326	cm_id->state = IB_CM_REP_SENT;
2327	cm_id_priv->initiator_depth = param->initiator_depth;
2328	cm_id_priv->responder_resources = param->responder_resources;
2329	cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg));
2330	WARN_ONCE(param->qp_num & 0xFF000000,
2331		  "IBTA declares QPN to be 24 bits, but it is 0x%X\n",
2332		  param->qp_num);
2333	cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
2334	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2335	return 0;
2336
2337out_free:
2338	cm_free_priv_msg(msg);
2339out:
2340	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2341	return ret;
2342}
2343EXPORT_SYMBOL(ib_send_cm_rep);
2344
2345static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
2346			  struct cm_id_private *cm_id_priv,
2347			  const void *private_data,
2348			  u8 private_data_len)
2349{
2350	cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
2351	IBA_SET(CM_RTU_LOCAL_COMM_ID, rtu_msg,
2352		be32_to_cpu(cm_id_priv->id.local_id));
2353	IBA_SET(CM_RTU_REMOTE_COMM_ID, rtu_msg,
2354		be32_to_cpu(cm_id_priv->id.remote_id));
2355
2356	if (private_data && private_data_len)
2357		IBA_SET_MEM(CM_RTU_PRIVATE_DATA, rtu_msg, private_data,
2358			    private_data_len);
2359}
2360
2361int ib_send_cm_rtu(struct ib_cm_id *cm_id,
2362		   const void *private_data,
2363		   u8 private_data_len)
2364{
2365	struct cm_id_private *cm_id_priv;
2366	struct ib_mad_send_buf *msg;
2367	unsigned long flags;
2368	void *data;
2369	int ret;
2370
2371	if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
2372		return -EINVAL;
2373
2374	data = cm_copy_private_data(private_data, private_data_len);
2375	if (IS_ERR(data))
2376		return PTR_ERR(data);
2377
2378	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2379	spin_lock_irqsave(&cm_id_priv->lock, flags);
2380	if (cm_id->state != IB_CM_REP_RCVD &&
2381	    cm_id->state != IB_CM_MRA_REP_SENT) {
2382		trace_icm_send_cm_rtu_err(cm_id);
2383		ret = -EINVAL;
2384		goto error;
2385	}
2386
2387	msg = cm_alloc_msg(cm_id_priv);
2388	if (IS_ERR(msg)) {
2389		ret = PTR_ERR(msg);
2390		goto error;
2391	}
2392
2393	cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2394		      private_data, private_data_len);
2395
2396	trace_icm_send_rtu(cm_id);
2397	ret = ib_post_send_mad(msg, NULL);
2398	if (ret) {
2399		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2400		cm_free_msg(msg);
2401		kfree(data);
2402		return ret;
2403	}
2404
2405	cm_id->state = IB_CM_ESTABLISHED;
2406	cm_set_private_data(cm_id_priv, data, private_data_len);
2407	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2408	return 0;
2409
2410error:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2411	kfree(data);
2412	return ret;
2413}
2414EXPORT_SYMBOL(ib_send_cm_rtu);
2415
2416static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
2417{
2418	struct cm_rep_msg *rep_msg;
2419	struct ib_cm_rep_event_param *param;
2420
2421	rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2422	param = &work->cm_event.param.rep_rcvd;
2423	param->remote_ca_guid =
2424		cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg));
2425	param->remote_qkey = IBA_GET(CM_REP_LOCAL_Q_KEY, rep_msg);
2426	param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
2427	param->starting_psn = IBA_GET(CM_REP_STARTING_PSN, rep_msg);
2428	param->responder_resources = IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg);
2429	param->initiator_depth = IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg);
2430	param->target_ack_delay = IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg);
2431	param->failover_accepted = IBA_GET(CM_REP_FAILOVER_ACCEPTED, rep_msg);
2432	param->flow_control = IBA_GET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg);
2433	param->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg);
2434	param->srq = IBA_GET(CM_REP_SRQ, rep_msg);
2435	param->ece.vendor_id = IBA_GET(CM_REP_VENDOR_ID_H, rep_msg) << 16;
2436	param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_M, rep_msg) << 8;
2437	param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_L, rep_msg);
2438	param->ece.attr_mod = be32_to_cpu(rep_msg->hdr.attr_mod);
2439
2440	work->cm_event.private_data =
2441		IBA_GET_MEM_PTR(CM_REP_PRIVATE_DATA, rep_msg);
2442}
2443
2444static void cm_dup_rep_handler(struct cm_work *work)
2445{
2446	struct cm_id_private *cm_id_priv;
2447	struct cm_rep_msg *rep_msg;
2448	struct ib_mad_send_buf *msg = NULL;
2449	int ret;
2450
2451	rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
2452	cm_id_priv = cm_acquire_id(
2453		cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)),
2454		cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg)));
2455	if (!cm_id_priv)
2456		return;
2457
2458	atomic_long_inc(
2459		&work->port->counters[CM_RECV_DUPLICATES][CM_REP_COUNTER]);
2460	ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, true, &msg);
2461	if (ret)
2462		goto deref;
2463
2464	spin_lock_irq(&cm_id_priv->lock);
2465	if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
2466		cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2467			      cm_id_priv->private_data,
2468			      cm_id_priv->private_data_len);
2469	else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
2470		cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2471			      CM_MSG_RESPONSE_REP,
2472			      cm_id_priv->private_data,
2473			      cm_id_priv->private_data_len);
2474	else
2475		goto unlock;
2476	spin_unlock_irq(&cm_id_priv->lock);
2477
2478	trace_icm_send_dup_rep(&cm_id_priv->id);
2479	ret = ib_post_send_mad(msg, NULL);
2480	if (ret)
2481		goto free;
2482	goto deref;
2483
2484unlock:	spin_unlock_irq(&cm_id_priv->lock);
2485free:	cm_free_msg(msg);
2486deref:	cm_deref_id(cm_id_priv);
2487}
2488
2489static int cm_rep_handler(struct cm_work *work)
2490{
2491	struct cm_id_private *cm_id_priv;
2492	struct cm_rep_msg *rep_msg;
2493	int ret;
2494	struct cm_id_private *cur_cm_id_priv;
2495	struct cm_timewait_info *timewait_info;
2496
2497	rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2498	cm_id_priv = cm_acquire_id(
2499		cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)), 0);
2500	if (!cm_id_priv) {
2501		cm_dup_rep_handler(work);
2502		trace_icm_remote_no_priv_err(
2503			 IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2504		return -EINVAL;
2505	}
2506
2507	cm_format_rep_event(work, cm_id_priv->qp_type);
2508
2509	spin_lock_irq(&cm_id_priv->lock);
2510	switch (cm_id_priv->id.state) {
2511	case IB_CM_REQ_SENT:
2512	case IB_CM_MRA_REQ_RCVD:
2513		break;
2514	default:
2515		ret = -EINVAL;
2516		trace_icm_rep_unknown_err(
2517			IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg),
2518			IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg),
2519			cm_id_priv->id.state);
2520		spin_unlock_irq(&cm_id_priv->lock);
2521		goto error;
2522	}
2523
2524	cm_id_priv->timewait_info->work.remote_id =
2525		cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg));
2526	cm_id_priv->timewait_info->remote_ca_guid =
2527		cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg));
2528	cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2529
2530	spin_lock(&cm.lock);
2531	/* Check for duplicate REP. */
2532	if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2533		spin_unlock(&cm.lock);
2534		spin_unlock_irq(&cm_id_priv->lock);
2535		ret = -EINVAL;
2536		trace_icm_insert_failed_err(
2537			 IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2538		goto error;
2539	}
2540	/* Check for a stale connection. */
2541	timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2542	if (timewait_info) {
2543		cm_remove_remote(cm_id_priv);
2544		cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2545					   timewait_info->work.remote_id);
2546
2547		spin_unlock(&cm.lock);
2548		spin_unlock_irq(&cm_id_priv->lock);
2549		cm_issue_rej(work->port, work->mad_recv_wc,
2550			     IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2551			     NULL, 0);
2552		ret = -EINVAL;
2553		trace_icm_staleconn_err(
2554			IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg),
2555			IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2556
2557		if (cur_cm_id_priv) {
2558			ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0);
2559			cm_deref_id(cur_cm_id_priv);
2560		}
2561
2562		goto error;
2563	}
2564	spin_unlock(&cm.lock);
2565
2566	cm_id_priv->id.state = IB_CM_REP_RCVD;
2567	cm_id_priv->id.remote_id =
2568		cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg));
2569	cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2570	cm_id_priv->initiator_depth =
2571		IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg);
2572	cm_id_priv->responder_resources =
2573		IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg);
2574	cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg));
2575	cm_id_priv->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg);
2576	cm_id_priv->target_ack_delay =
2577		IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg);
2578	cm_id_priv->av.timeout =
2579			cm_ack_timeout(cm_id_priv->target_ack_delay,
2580				       cm_id_priv->av.timeout - 1);
2581	cm_id_priv->alt_av.timeout =
2582			cm_ack_timeout(cm_id_priv->target_ack_delay,
2583				       cm_id_priv->alt_av.timeout - 1);
2584
2585	ib_cancel_mad(cm_id_priv->msg);
2586	cm_queue_work_unlock(cm_id_priv, work);
2587	return 0;
2588
2589error:
2590	cm_deref_id(cm_id_priv);
2591	return ret;
2592}
2593
2594static int cm_establish_handler(struct cm_work *work)
2595{
2596	struct cm_id_private *cm_id_priv;
2597
2598	/* See comment in cm_establish about lookup. */
2599	cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2600	if (!cm_id_priv)
2601		return -EINVAL;
2602
2603	spin_lock_irq(&cm_id_priv->lock);
2604	if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2605		spin_unlock_irq(&cm_id_priv->lock);
2606		goto out;
2607	}
2608
2609	ib_cancel_mad(cm_id_priv->msg);
2610	cm_queue_work_unlock(cm_id_priv, work);
2611	return 0;
2612out:
2613	cm_deref_id(cm_id_priv);
2614	return -EINVAL;
2615}
2616
2617static int cm_rtu_handler(struct cm_work *work)
2618{
2619	struct cm_id_private *cm_id_priv;
2620	struct cm_rtu_msg *rtu_msg;
2621
2622	rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2623	cm_id_priv = cm_acquire_id(
2624		cpu_to_be32(IBA_GET(CM_RTU_REMOTE_COMM_ID, rtu_msg)),
2625		cpu_to_be32(IBA_GET(CM_RTU_LOCAL_COMM_ID, rtu_msg)));
2626	if (!cm_id_priv)
2627		return -EINVAL;
2628
2629	work->cm_event.private_data =
2630		IBA_GET_MEM_PTR(CM_RTU_PRIVATE_DATA, rtu_msg);
2631
2632	spin_lock_irq(&cm_id_priv->lock);
2633	if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2634	    cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2635		spin_unlock_irq(&cm_id_priv->lock);
2636		atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
2637						     [CM_RTU_COUNTER]);
2638		goto out;
2639	}
2640	cm_id_priv->id.state = IB_CM_ESTABLISHED;
2641
2642	ib_cancel_mad(cm_id_priv->msg);
2643	cm_queue_work_unlock(cm_id_priv, work);
2644	return 0;
2645out:
2646	cm_deref_id(cm_id_priv);
2647	return -EINVAL;
2648}
2649
2650static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2651			  struct cm_id_private *cm_id_priv,
2652			  const void *private_data,
2653			  u8 private_data_len)
2654{
2655	cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2656			  cm_form_tid(cm_id_priv));
2657	IBA_SET(CM_DREQ_LOCAL_COMM_ID, dreq_msg,
2658		be32_to_cpu(cm_id_priv->id.local_id));
2659	IBA_SET(CM_DREQ_REMOTE_COMM_ID, dreq_msg,
2660		be32_to_cpu(cm_id_priv->id.remote_id));
2661	IBA_SET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg,
2662		be32_to_cpu(cm_id_priv->remote_qpn));
2663
2664	if (private_data && private_data_len)
2665		IBA_SET_MEM(CM_DREQ_PRIVATE_DATA, dreq_msg, private_data,
2666			    private_data_len);
2667}
2668
2669static void cm_issue_dreq(struct cm_id_private *cm_id_priv)
2670{
2671	struct ib_mad_send_buf *msg;
2672	int ret;
2673
2674	lockdep_assert_held(&cm_id_priv->lock);
2675
2676	msg = cm_alloc_msg(cm_id_priv);
2677	if (IS_ERR(msg))
2678		return;
2679
2680	cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv, NULL, 0);
2681
2682	trace_icm_send_dreq(&cm_id_priv->id);
2683	ret = ib_post_send_mad(msg, NULL);
2684	if (ret)
2685		cm_free_msg(msg);
2686}
2687
2688int ib_send_cm_dreq(struct ib_cm_id *cm_id, const void *private_data,
2689		    u8 private_data_len)
2690{
2691	struct cm_id_private *cm_id_priv =
2692		container_of(cm_id, struct cm_id_private, id);
2693	struct ib_mad_send_buf *msg;
2694	unsigned long flags;
2695	int ret;
2696
2697	if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2698		return -EINVAL;
2699
2700	spin_lock_irqsave(&cm_id_priv->lock, flags);
2701	if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2702		trace_icm_dreq_skipped(&cm_id_priv->id);
2703		ret = -EINVAL;
2704		goto unlock;
2705	}
2706
2707	if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2708	    cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2709		ib_cancel_mad(cm_id_priv->msg);
2710
2711	msg = cm_alloc_priv_msg(cm_id_priv, IB_CM_DREQ_SENT);
2712	if (IS_ERR(msg)) {
2713		cm_enter_timewait(cm_id_priv);
2714		ret = PTR_ERR(msg);
2715		goto unlock;
2716	}
2717
2718	cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2719		       private_data, private_data_len);
2720
2721	trace_icm_send_dreq(&cm_id_priv->id);
2722	ret = ib_post_send_mad(msg, NULL);
2723	if (ret) {
2724		cm_enter_timewait(cm_id_priv);
2725		cm_free_priv_msg(msg);
2726		goto unlock;
2727	}
2728
2729	cm_id_priv->id.state = IB_CM_DREQ_SENT;
2730unlock:
2731	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2732	return ret;
2733}
2734EXPORT_SYMBOL(ib_send_cm_dreq);
2735
2736static void cm_format_drep(struct cm_drep_msg *drep_msg,
2737			  struct cm_id_private *cm_id_priv,
2738			  const void *private_data,
2739			  u8 private_data_len)
2740{
2741	cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2742	IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg,
2743		be32_to_cpu(cm_id_priv->id.local_id));
2744	IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg,
2745		be32_to_cpu(cm_id_priv->id.remote_id));
2746
2747	if (private_data && private_data_len)
2748		IBA_SET_MEM(CM_DREP_PRIVATE_DATA, drep_msg, private_data,
2749			    private_data_len);
2750}
2751
2752static int cm_send_drep_locked(struct cm_id_private *cm_id_priv,
2753			       void *private_data, u8 private_data_len)
2754{
2755	struct ib_mad_send_buf *msg;
2756	int ret;
2757
2758	lockdep_assert_held(&cm_id_priv->lock);
2759
2760	if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2761		return -EINVAL;
2762
2763	if (cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2764		trace_icm_send_drep_err(&cm_id_priv->id);
2765		kfree(private_data);
2766		return -EINVAL;
2767	}
2768
2769	cm_set_private_data(cm_id_priv, private_data, private_data_len);
2770	cm_enter_timewait(cm_id_priv);
2771
2772	msg = cm_alloc_msg(cm_id_priv);
2773	if (IS_ERR(msg))
2774		return PTR_ERR(msg);
2775
2776	cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2777		       private_data, private_data_len);
2778
2779	trace_icm_send_drep(&cm_id_priv->id);
2780	ret = ib_post_send_mad(msg, NULL);
2781	if (ret) {
2782		cm_free_msg(msg);
2783		return ret;
2784	}
2785	return 0;
2786}
2787
2788int ib_send_cm_drep(struct ib_cm_id *cm_id, const void *private_data,
2789		    u8 private_data_len)
2790{
2791	struct cm_id_private *cm_id_priv =
2792		container_of(cm_id, struct cm_id_private, id);
2793	unsigned long flags;
2794	void *data;
2795	int ret;
2796
2797	data = cm_copy_private_data(private_data, private_data_len);
2798	if (IS_ERR(data))
2799		return PTR_ERR(data);
2800
2801	spin_lock_irqsave(&cm_id_priv->lock, flags);
2802	ret = cm_send_drep_locked(cm_id_priv, data, private_data_len);
2803	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2804	return ret;
2805}
2806EXPORT_SYMBOL(ib_send_cm_drep);
2807
2808static int cm_issue_drep(struct cm_port *port,
2809			 struct ib_mad_recv_wc *mad_recv_wc)
2810{
2811	struct ib_mad_send_buf *msg = NULL;
2812	struct cm_dreq_msg *dreq_msg;
2813	struct cm_drep_msg *drep_msg;
2814	int ret;
2815
2816	ret = cm_alloc_response_msg(port, mad_recv_wc, true, &msg);
2817	if (ret)
2818		return ret;
2819
2820	dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2821	drep_msg = (struct cm_drep_msg *) msg->mad;
2822
2823	cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2824	IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg,
2825		IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg));
2826	IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg,
2827		IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2828
2829	trace_icm_issue_drep(
2830		IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg),
2831		IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2832	ret = ib_post_send_mad(msg, NULL);
2833	if (ret)
2834		cm_free_msg(msg);
2835
2836	return ret;
2837}
2838
2839static int cm_dreq_handler(struct cm_work *work)
2840{
2841	struct cm_id_private *cm_id_priv;
2842	struct cm_dreq_msg *dreq_msg;
2843	struct ib_mad_send_buf *msg = NULL;
2844
2845	dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2846	cm_id_priv = cm_acquire_id(
2847		cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)),
2848		cpu_to_be32(IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg)));
2849	if (!cm_id_priv) {
2850		atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
2851						     [CM_DREQ_COUNTER]);
2852		cm_issue_drep(work->port, work->mad_recv_wc);
2853		trace_icm_no_priv_err(
2854			IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg),
2855			IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2856		return -EINVAL;
2857	}
2858
2859	work->cm_event.private_data =
2860		IBA_GET_MEM_PTR(CM_DREQ_PRIVATE_DATA, dreq_msg);
2861
2862	spin_lock_irq(&cm_id_priv->lock);
2863	if (cm_id_priv->local_qpn !=
2864	    cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg)))
2865		goto unlock;
2866
2867	switch (cm_id_priv->id.state) {
2868	case IB_CM_REP_SENT:
2869	case IB_CM_DREQ_SENT:
2870	case IB_CM_MRA_REP_RCVD:
2871		ib_cancel_mad(cm_id_priv->msg);
2872		break;
2873	case IB_CM_ESTABLISHED:
2874		if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2875		    cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2876			ib_cancel_mad(cm_id_priv->msg);
2877		break;
2878	case IB_CM_TIMEWAIT:
2879		atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
2880						     [CM_DREQ_COUNTER]);
2881		msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc,
2882						  true);
2883		if (IS_ERR(msg))
2884			goto unlock;
2885
2886		cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2887			       cm_id_priv->private_data,
2888			       cm_id_priv->private_data_len);
2889		spin_unlock_irq(&cm_id_priv->lock);
2890
2891		if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
2892		    ib_post_send_mad(msg, NULL))
2893			cm_free_msg(msg);
2894		goto deref;
2895	case IB_CM_DREQ_RCVD:
2896		atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
2897						     [CM_DREQ_COUNTER]);
2898		goto unlock;
2899	default:
2900		trace_icm_dreq_unknown_err(&cm_id_priv->id);
2901		goto unlock;
2902	}
2903	cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2904	cm_id_priv->tid = dreq_msg->hdr.tid;
2905	cm_queue_work_unlock(cm_id_priv, work);
2906	return 0;
2907
2908unlock:	spin_unlock_irq(&cm_id_priv->lock);
2909deref:	cm_deref_id(cm_id_priv);
2910	return -EINVAL;
2911}
2912
2913static int cm_drep_handler(struct cm_work *work)
2914{
2915	struct cm_id_private *cm_id_priv;
2916	struct cm_drep_msg *drep_msg;
2917
2918	drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2919	cm_id_priv = cm_acquire_id(
2920		cpu_to_be32(IBA_GET(CM_DREP_REMOTE_COMM_ID, drep_msg)),
2921		cpu_to_be32(IBA_GET(CM_DREP_LOCAL_COMM_ID, drep_msg)));
2922	if (!cm_id_priv)
2923		return -EINVAL;
2924
2925	work->cm_event.private_data =
2926		IBA_GET_MEM_PTR(CM_DREP_PRIVATE_DATA, drep_msg);
2927
2928	spin_lock_irq(&cm_id_priv->lock);
2929	if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2930	    cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2931		spin_unlock_irq(&cm_id_priv->lock);
2932		goto out;
2933	}
2934	cm_enter_timewait(cm_id_priv);
2935
2936	ib_cancel_mad(cm_id_priv->msg);
2937	cm_queue_work_unlock(cm_id_priv, work);
2938	return 0;
2939out:
2940	cm_deref_id(cm_id_priv);
2941	return -EINVAL;
2942}
2943
2944static int cm_send_rej_locked(struct cm_id_private *cm_id_priv,
2945			      enum ib_cm_rej_reason reason, void *ari,
2946			      u8 ari_length, const void *private_data,
2947			      u8 private_data_len)
2948{
2949	enum ib_cm_state state = cm_id_priv->id.state;
2950	struct ib_mad_send_buf *msg;
2951	int ret;
2952
2953	lockdep_assert_held(&cm_id_priv->lock);
2954
2955	if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2956	    (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2957		return -EINVAL;
2958
2959	trace_icm_send_rej(&cm_id_priv->id, reason);
2960
2961	switch (state) {
2962	case IB_CM_REQ_SENT:
2963	case IB_CM_MRA_REQ_RCVD:
2964	case IB_CM_REQ_RCVD:
2965	case IB_CM_MRA_REQ_SENT:
2966	case IB_CM_REP_RCVD:
2967	case IB_CM_MRA_REP_SENT:
2968		cm_reset_to_idle(cm_id_priv);
2969		msg = cm_alloc_msg(cm_id_priv);
2970		if (IS_ERR(msg))
2971			return PTR_ERR(msg);
2972		cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason,
2973			      ari, ari_length, private_data, private_data_len,
2974			      state);
2975		break;
2976	case IB_CM_REP_SENT:
2977	case IB_CM_MRA_REP_RCVD:
2978		cm_enter_timewait(cm_id_priv);
2979		msg = cm_alloc_msg(cm_id_priv);
2980		if (IS_ERR(msg))
2981			return PTR_ERR(msg);
2982		cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason,
2983			      ari, ari_length, private_data, private_data_len,
2984			      state);
2985		break;
2986	default:
2987		trace_icm_send_unknown_rej_err(&cm_id_priv->id);
2988		return -EINVAL;
2989	}
2990
2991	ret = ib_post_send_mad(msg, NULL);
2992	if (ret) {
2993		cm_free_msg(msg);
2994		return ret;
2995	}
2996
2997	return 0;
2998}
2999
3000int ib_send_cm_rej(struct ib_cm_id *cm_id, enum ib_cm_rej_reason reason,
3001		   void *ari, u8 ari_length, const void *private_data,
3002		   u8 private_data_len)
3003{
3004	struct cm_id_private *cm_id_priv =
3005		container_of(cm_id, struct cm_id_private, id);
3006	unsigned long flags;
3007	int ret;
3008
3009	spin_lock_irqsave(&cm_id_priv->lock, flags);
3010	ret = cm_send_rej_locked(cm_id_priv, reason, ari, ari_length,
3011				 private_data, private_data_len);
3012	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3013	return ret;
3014}
3015EXPORT_SYMBOL(ib_send_cm_rej);
3016
3017static void cm_format_rej_event(struct cm_work *work)
3018{
3019	struct cm_rej_msg *rej_msg;
3020	struct ib_cm_rej_event_param *param;
3021
3022	rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
3023	param = &work->cm_event.param.rej_rcvd;
3024	param->ari = IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg);
3025	param->ari_length = IBA_GET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg);
3026	param->reason = IBA_GET(CM_REJ_REASON, rej_msg);
3027	work->cm_event.private_data =
3028		IBA_GET_MEM_PTR(CM_REJ_PRIVATE_DATA, rej_msg);
3029}
3030
3031static struct cm_id_private *cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
3032{
3033	struct cm_id_private *cm_id_priv;
3034	__be32 remote_id;
3035
3036	remote_id = cpu_to_be32(IBA_GET(CM_REJ_LOCAL_COMM_ID, rej_msg));
3037
3038	if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_TIMEOUT) {
3039		cm_id_priv = cm_find_remote_id(
3040			*((__be64 *)IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg)),
3041			remote_id);
3042	} else if (IBA_GET(CM_REJ_MESSAGE_REJECTED, rej_msg) ==
3043		   CM_MSG_RESPONSE_REQ)
3044		cm_id_priv = cm_acquire_id(
3045			cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)),
3046			0);
3047	else
3048		cm_id_priv = cm_acquire_id(
3049			cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)),
3050			remote_id);
3051
3052	return cm_id_priv;
3053}
3054
3055static int cm_rej_handler(struct cm_work *work)
3056{
3057	struct cm_id_private *cm_id_priv;
3058	struct cm_rej_msg *rej_msg;
3059
3060	rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
3061	cm_id_priv = cm_acquire_rejected_id(rej_msg);
3062	if (!cm_id_priv)
3063		return -EINVAL;
3064
3065	cm_format_rej_event(work);
3066
3067	spin_lock_irq(&cm_id_priv->lock);
3068	switch (cm_id_priv->id.state) {
3069	case IB_CM_REQ_SENT:
3070	case IB_CM_MRA_REQ_RCVD:
3071	case IB_CM_REP_SENT:
3072	case IB_CM_MRA_REP_RCVD:
3073		ib_cancel_mad(cm_id_priv->msg);
3074		fallthrough;
3075	case IB_CM_REQ_RCVD:
3076	case IB_CM_MRA_REQ_SENT:
3077		if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_STALE_CONN)
3078			cm_enter_timewait(cm_id_priv);
3079		else
3080			cm_reset_to_idle(cm_id_priv);
3081		break;
3082	case IB_CM_DREQ_SENT:
3083		ib_cancel_mad(cm_id_priv->msg);
3084		fallthrough;
3085	case IB_CM_REP_RCVD:
3086	case IB_CM_MRA_REP_SENT:
3087		cm_enter_timewait(cm_id_priv);
3088		break;
3089	case IB_CM_ESTABLISHED:
3090		if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
3091		    cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
3092			if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
3093				ib_cancel_mad(cm_id_priv->msg);
3094			cm_enter_timewait(cm_id_priv);
3095			break;
3096		}
3097		fallthrough;
3098	default:
3099		trace_icm_rej_unknown_err(&cm_id_priv->id);
3100		spin_unlock_irq(&cm_id_priv->lock);
3101		goto out;
3102	}
3103
3104	cm_queue_work_unlock(cm_id_priv, work);
3105	return 0;
3106out:
3107	cm_deref_id(cm_id_priv);
3108	return -EINVAL;
3109}
3110
3111int ib_prepare_cm_mra(struct ib_cm_id *cm_id)
3112{
3113	struct cm_id_private *cm_id_priv;
3114	enum ib_cm_state cm_state;
3115	enum ib_cm_lap_state lap_state;
3116	unsigned long flags;
3117	int ret = 0;
3118
3119	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3120
3121	spin_lock_irqsave(&cm_id_priv->lock, flags);
3122	switch (cm_id_priv->id.state) {
3123	case IB_CM_REQ_RCVD:
3124		cm_state = IB_CM_MRA_REQ_SENT;
3125		lap_state = cm_id->lap_state;
3126		break;
3127	case IB_CM_REP_RCVD:
3128		cm_state = IB_CM_MRA_REP_SENT;
3129		lap_state = cm_id->lap_state;
3130		break;
3131	case IB_CM_ESTABLISHED:
3132		if (cm_id->lap_state == IB_CM_LAP_RCVD) {
3133			cm_state = cm_id->state;
3134			lap_state = IB_CM_MRA_LAP_SENT;
3135			break;
3136		}
3137		fallthrough;
3138	default:
3139		trace_icm_prepare_mra_unknown_err(&cm_id_priv->id);
3140		ret = -EINVAL;
3141		goto error_unlock;
3142	}
3143
3144	cm_id->state = cm_state;
3145	cm_id->lap_state = lap_state;
3146	cm_set_private_data(cm_id_priv, NULL, 0);
3147
3148error_unlock:
3149	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3150	return ret;
3151}
3152EXPORT_SYMBOL(ib_prepare_cm_mra);
3153
3154static struct cm_id_private *cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
3155{
3156	switch (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg)) {
3157	case CM_MSG_RESPONSE_REQ:
3158		return cm_acquire_id(
3159			cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)),
3160			0);
3161	case CM_MSG_RESPONSE_REP:
3162	case CM_MSG_RESPONSE_OTHER:
3163		return cm_acquire_id(
3164			cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)),
3165			cpu_to_be32(IBA_GET(CM_MRA_LOCAL_COMM_ID, mra_msg)));
3166	default:
3167		return NULL;
3168	}
3169}
3170
3171static int cm_mra_handler(struct cm_work *work)
3172{
3173	struct cm_id_private *cm_id_priv;
3174	struct cm_mra_msg *mra_msg;
3175	int timeout;
3176
3177	mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
3178	cm_id_priv = cm_acquire_mraed_id(mra_msg);
3179	if (!cm_id_priv)
3180		return -EINVAL;
3181
3182	work->cm_event.private_data =
3183		IBA_GET_MEM_PTR(CM_MRA_PRIVATE_DATA, mra_msg);
3184	work->cm_event.param.mra_rcvd.service_timeout =
3185		IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg);
3186	timeout = cm_convert_to_ms(IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg)) +
3187		  cm_convert_to_ms(cm_id_priv->av.timeout);
3188
3189	spin_lock_irq(&cm_id_priv->lock);
3190	switch (cm_id_priv->id.state) {
3191	case IB_CM_REQ_SENT:
3192		if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3193			    CM_MSG_RESPONSE_REQ ||
3194		    ib_modify_mad(cm_id_priv->msg, timeout))
3195			goto out;
3196		cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
3197		break;
3198	case IB_CM_REP_SENT:
3199		if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3200			    CM_MSG_RESPONSE_REP ||
3201		    ib_modify_mad(cm_id_priv->msg, timeout))
3202			goto out;
3203		cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
3204		break;
3205	case IB_CM_ESTABLISHED:
3206		if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3207			    CM_MSG_RESPONSE_OTHER ||
3208		    cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
3209		    ib_modify_mad(cm_id_priv->msg, timeout)) {
3210			if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
3211				atomic_long_inc(
3212					&work->port->counters[CM_RECV_DUPLICATES]
3213							     [CM_MRA_COUNTER]);
3214			goto out;
3215		}
3216		cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
3217		break;
3218	case IB_CM_MRA_REQ_RCVD:
3219	case IB_CM_MRA_REP_RCVD:
3220		atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
3221						     [CM_MRA_COUNTER]);
3222		fallthrough;
3223	default:
3224		trace_icm_mra_unknown_err(&cm_id_priv->id);
3225		goto out;
3226	}
3227
3228	cm_id_priv->msg->context[1] = (void *) (unsigned long)
3229				      cm_id_priv->id.state;
3230	cm_queue_work_unlock(cm_id_priv, work);
3231	return 0;
3232out:
3233	spin_unlock_irq(&cm_id_priv->lock);
3234	cm_deref_id(cm_id_priv);
3235	return -EINVAL;
3236}
3237
3238static void cm_format_path_lid_from_lap(struct cm_lap_msg *lap_msg,
3239					struct sa_path_rec *path)
3240{
3241	u32 lid;
3242
3243	if (path->rec_type != SA_PATH_REC_TYPE_OPA) {
3244		sa_path_set_dlid(path, IBA_GET(CM_LAP_ALTERNATE_LOCAL_PORT_LID,
3245					       lap_msg));
3246		sa_path_set_slid(path, IBA_GET(CM_LAP_ALTERNATE_REMOTE_PORT_LID,
3247					       lap_msg));
3248	} else {
3249		lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
3250			CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg));
3251		sa_path_set_dlid(path, lid);
3252
3253		lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
3254			CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg));
3255		sa_path_set_slid(path, lid);
3256	}
3257}
3258
3259static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
3260				    struct sa_path_rec *path,
3261				    struct cm_lap_msg *lap_msg)
3262{
3263	path->dgid = *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg);
3264	path->sgid =
3265		*IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg);
3266	path->flow_label =
3267		cpu_to_be32(IBA_GET(CM_LAP_ALTERNATE_FLOW_LABEL, lap_msg));
3268	path->hop_limit = IBA_GET(CM_LAP_ALTERNATE_HOP_LIMIT, lap_msg);
3269	path->traffic_class = IBA_GET(CM_LAP_ALTERNATE_TRAFFIC_CLASS, lap_msg);
3270	path->reversible = 1;
3271	path->pkey = cm_id_priv->pkey;
3272	path->sl = IBA_GET(CM_LAP_ALTERNATE_SL, lap_msg);
3273	path->mtu_selector = IB_SA_EQ;
3274	path->mtu = cm_id_priv->path_mtu;
3275	path->rate_selector = IB_SA_EQ;
3276	path->rate = IBA_GET(CM_LAP_ALTERNATE_PACKET_RATE, lap_msg);
3277	path->packet_life_time_selector = IB_SA_EQ;
3278	path->packet_life_time =
3279		IBA_GET(CM_LAP_ALTERNATE_LOCAL_ACK_TIMEOUT, lap_msg);
3280	path->packet_life_time -= (path->packet_life_time > 0);
3281	cm_format_path_lid_from_lap(lap_msg, path);
3282}
3283
3284static int cm_lap_handler(struct cm_work *work)
3285{
3286	struct cm_id_private *cm_id_priv;
3287	struct cm_lap_msg *lap_msg;
3288	struct ib_cm_lap_event_param *param;
3289	struct ib_mad_send_buf *msg = NULL;
3290	struct rdma_ah_attr ah_attr;
3291	struct cm_av alt_av = {};
3292	int ret;
3293
3294	/* Currently Alternate path messages are not supported for
3295	 * RoCE link layer.
3296	 */
3297	if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3298			       work->port->port_num))
3299		return -EINVAL;
3300
3301	/* todo: verify LAP request and send reject APR if invalid. */
3302	lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
3303	cm_id_priv = cm_acquire_id(
3304		cpu_to_be32(IBA_GET(CM_LAP_REMOTE_COMM_ID, lap_msg)),
3305		cpu_to_be32(IBA_GET(CM_LAP_LOCAL_COMM_ID, lap_msg)));
3306	if (!cm_id_priv)
3307		return -EINVAL;
3308
3309	param = &work->cm_event.param.lap_rcvd;
3310	memset(&work->path[0], 0, sizeof(work->path[1]));
3311	cm_path_set_rec_type(work->port->cm_dev->ib_device,
3312			     work->port->port_num, &work->path[0],
3313			     IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID,
3314					     lap_msg));
3315	param->alternate_path = &work->path[0];
3316	cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
3317	work->cm_event.private_data =
3318		IBA_GET_MEM_PTR(CM_LAP_PRIVATE_DATA, lap_msg);
3319
3320	ret = ib_init_ah_attr_from_wc(work->port->cm_dev->ib_device,
3321				      work->port->port_num,
3322				      work->mad_recv_wc->wc,
3323				      work->mad_recv_wc->recv_buf.grh,
3324				      &ah_attr);
3325	if (ret)
3326		goto deref;
3327
3328	ret = cm_init_av_by_path(param->alternate_path, NULL, &alt_av);
3329	if (ret) {
3330		rdma_destroy_ah_attr(&ah_attr);
3331		goto deref;
3332	}
3333
3334	spin_lock_irq(&cm_id_priv->lock);
3335	cm_init_av_for_lap(work->port, work->mad_recv_wc->wc,
3336			   &ah_attr, &cm_id_priv->av);
3337	cm_move_av_from_path(&cm_id_priv->alt_av, &alt_av);
3338
3339	if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
3340		goto unlock;
3341
3342	switch (cm_id_priv->id.lap_state) {
3343	case IB_CM_LAP_UNINIT:
3344	case IB_CM_LAP_IDLE:
3345		break;
3346	case IB_CM_MRA_LAP_SENT:
3347		atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
3348						     [CM_LAP_COUNTER]);
3349		msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc,
3350						  true);
3351		if (IS_ERR(msg))
3352			goto unlock;
3353
3354		cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3355			      CM_MSG_RESPONSE_OTHER,
3356			      cm_id_priv->private_data,
3357			      cm_id_priv->private_data_len);
3358		spin_unlock_irq(&cm_id_priv->lock);
3359
3360		if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
3361		    ib_post_send_mad(msg, NULL))
3362			cm_free_msg(msg);
3363		goto deref;
3364	case IB_CM_LAP_RCVD:
3365		atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
3366						     [CM_LAP_COUNTER]);
3367		goto unlock;
3368	default:
3369		goto unlock;
3370	}
3371
3372	cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
3373	cm_id_priv->tid = lap_msg->hdr.tid;
3374	cm_queue_work_unlock(cm_id_priv, work);
3375	return 0;
3376
3377unlock:	spin_unlock_irq(&cm_id_priv->lock);
3378deref:	cm_deref_id(cm_id_priv);
3379	return -EINVAL;
3380}
3381
3382static int cm_apr_handler(struct cm_work *work)
3383{
3384	struct cm_id_private *cm_id_priv;
3385	struct cm_apr_msg *apr_msg;
3386
3387	/* Currently Alternate path messages are not supported for
3388	 * RoCE link layer.
3389	 */
3390	if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3391			       work->port->port_num))
3392		return -EINVAL;
3393
3394	apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3395	cm_id_priv = cm_acquire_id(
3396		cpu_to_be32(IBA_GET(CM_APR_REMOTE_COMM_ID, apr_msg)),
3397		cpu_to_be32(IBA_GET(CM_APR_LOCAL_COMM_ID, apr_msg)));
3398	if (!cm_id_priv)
3399		return -EINVAL; /* Unmatched reply. */
3400
3401	work->cm_event.param.apr_rcvd.ap_status =
3402		IBA_GET(CM_APR_AR_STATUS, apr_msg);
3403	work->cm_event.param.apr_rcvd.apr_info =
3404		IBA_GET_MEM_PTR(CM_APR_ADDITIONAL_INFORMATION, apr_msg);
3405	work->cm_event.param.apr_rcvd.info_len =
3406		IBA_GET(CM_APR_ADDITIONAL_INFORMATION_LENGTH, apr_msg);
3407	work->cm_event.private_data =
3408		IBA_GET_MEM_PTR(CM_APR_PRIVATE_DATA, apr_msg);
3409
3410	spin_lock_irq(&cm_id_priv->lock);
3411	if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3412	    (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3413	     cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3414		spin_unlock_irq(&cm_id_priv->lock);
3415		goto out;
3416	}
3417	cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3418	ib_cancel_mad(cm_id_priv->msg);
3419	cm_queue_work_unlock(cm_id_priv, work);
3420	return 0;
3421out:
3422	cm_deref_id(cm_id_priv);
3423	return -EINVAL;
3424}
3425
3426static int cm_timewait_handler(struct cm_work *work)
3427{
3428	struct cm_timewait_info *timewait_info;
3429	struct cm_id_private *cm_id_priv;
3430
3431	timewait_info = container_of(work, struct cm_timewait_info, work);
3432	spin_lock_irq(&cm.lock);
3433	list_del(&timewait_info->list);
3434	spin_unlock_irq(&cm.lock);
3435
3436	cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3437				   timewait_info->work.remote_id);
3438	if (!cm_id_priv)
3439		return -EINVAL;
3440
3441	spin_lock_irq(&cm_id_priv->lock);
3442	if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3443	    cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3444		spin_unlock_irq(&cm_id_priv->lock);
3445		goto out;
3446	}
3447	cm_id_priv->id.state = IB_CM_IDLE;
3448	cm_queue_work_unlock(cm_id_priv, work);
3449	return 0;
3450out:
3451	cm_deref_id(cm_id_priv);
3452	return -EINVAL;
3453}
3454
3455static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3456			       struct cm_id_private *cm_id_priv,
3457			       struct ib_cm_sidr_req_param *param)
3458{
3459	cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3460			  cm_form_tid(cm_id_priv));
3461	IBA_SET(CM_SIDR_REQ_REQUESTID, sidr_req_msg,
3462		be32_to_cpu(cm_id_priv->id.local_id));
3463	IBA_SET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg,
3464		be16_to_cpu(param->path->pkey));
3465	IBA_SET(CM_SIDR_REQ_SERVICEID, sidr_req_msg,
3466		be64_to_cpu(param->service_id));
3467
3468	if (param->private_data && param->private_data_len)
3469		IBA_SET_MEM(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg,
3470			    param->private_data, param->private_data_len);
3471}
3472
3473int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3474			struct ib_cm_sidr_req_param *param)
3475{
3476	struct cm_id_private *cm_id_priv;
3477	struct ib_mad_send_buf *msg;
3478	struct cm_av av = {};
3479	unsigned long flags;
3480	int ret;
3481
3482	if (!param->path || (param->private_data &&
3483	     param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3484		return -EINVAL;
3485
3486	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3487	ret = cm_init_av_by_path(param->path, param->sgid_attr, &av);
3488	if (ret)
3489		return ret;
3490
3491	spin_lock_irqsave(&cm_id_priv->lock, flags);
3492	cm_move_av_from_path(&cm_id_priv->av, &av);
3493	cm_id->service_id = param->service_id;
3494	cm_id_priv->timeout_ms = param->timeout_ms;
3495	cm_id_priv->max_cm_retries = param->max_cm_retries;
3496	if (cm_id->state != IB_CM_IDLE) {
3497		ret = -EINVAL;
3498		goto out_unlock;
3499	}
3500
3501	msg = cm_alloc_priv_msg(cm_id_priv, IB_CM_SIDR_REQ_SENT);
3502	if (IS_ERR(msg)) {
3503		ret = PTR_ERR(msg);
3504		goto out_unlock;
3505	}
3506
3507	cm_format_sidr_req((struct cm_sidr_req_msg *)msg->mad, cm_id_priv,
3508			   param);
3509
3510	trace_icm_send_sidr_req(&cm_id_priv->id);
3511	ret = ib_post_send_mad(msg, NULL);
3512	if (ret)
3513		goto out_free;
3514	cm_id->state = IB_CM_SIDR_REQ_SENT;
3515	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3516	return 0;
3517out_free:
3518	cm_free_priv_msg(msg);
3519out_unlock:
3520	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3521	return ret;
3522}
3523EXPORT_SYMBOL(ib_send_cm_sidr_req);
3524
3525static void cm_format_sidr_req_event(struct cm_work *work,
3526				     const struct cm_id_private *rx_cm_id,
3527				     struct ib_cm_id *listen_id)
3528{
3529	struct cm_sidr_req_msg *sidr_req_msg;
3530	struct ib_cm_sidr_req_event_param *param;
3531
3532	sidr_req_msg = (struct cm_sidr_req_msg *)
3533				work->mad_recv_wc->recv_buf.mad;
3534	param = &work->cm_event.param.sidr_req_rcvd;
3535	param->pkey = IBA_GET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg);
3536	param->listen_id = listen_id;
3537	param->service_id =
3538		cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg));
3539	param->bth_pkey = cm_get_bth_pkey(work);
3540	param->port = work->port->port_num;
3541	param->sgid_attr = rx_cm_id->av.ah_attr.grh.sgid_attr;
3542	work->cm_event.private_data =
3543		IBA_GET_MEM_PTR(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg);
3544}
3545
3546static int cm_sidr_req_handler(struct cm_work *work)
3547{
3548	struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
3549	struct cm_sidr_req_msg *sidr_req_msg;
3550	struct ib_wc *wc;
3551	int ret;
3552
3553	cm_id_priv =
3554		cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL);
3555	if (IS_ERR(cm_id_priv))
3556		return PTR_ERR(cm_id_priv);
3557
3558	/* Record SGID/SLID and request ID for lookup. */
3559	sidr_req_msg = (struct cm_sidr_req_msg *)
3560				work->mad_recv_wc->recv_buf.mad;
3561
3562	cm_id_priv->id.remote_id =
3563		cpu_to_be32(IBA_GET(CM_SIDR_REQ_REQUESTID, sidr_req_msg));
3564	cm_id_priv->id.service_id =
3565		cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg));
3566	cm_id_priv->tid = sidr_req_msg->hdr.tid;
3567
3568	wc = work->mad_recv_wc->wc;
3569	cm_id_priv->sidr_slid = wc->slid;
3570	ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3571				      work->mad_recv_wc->recv_buf.grh,
3572				      &cm_id_priv->av);
3573	if (ret)
3574		goto out;
3575
3576	spin_lock_irq(&cm.lock);
3577	listen_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3578	if (listen_cm_id_priv) {
3579		spin_unlock_irq(&cm.lock);
3580		atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
3581						     [CM_SIDR_REQ_COUNTER]);
3582		goto out; /* Duplicate message. */
3583	}
3584	cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3585	listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
3586					   cm_id_priv->id.service_id);
3587	if (!listen_cm_id_priv) {
3588		spin_unlock_irq(&cm.lock);
3589		ib_send_cm_sidr_rep(&cm_id_priv->id,
3590				    &(struct ib_cm_sidr_rep_param){
3591					    .status = IB_SIDR_UNSUPPORTED });
3592		goto out; /* No match. */
3593	}
3594	spin_unlock_irq(&cm.lock);
3595
3596	cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
3597	cm_id_priv->id.context = listen_cm_id_priv->id.context;
3598
3599	/*
3600	 * A SIDR ID does not need to be in the xarray since it does not receive
3601	 * mads, is not placed in the remote_id or remote_qpn rbtree, and does
3602	 * not enter timewait.
3603	 */
3604
3605	cm_format_sidr_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
3606	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
3607	cm_free_work(work);
3608	/*
3609	 * A pointer to the listen_cm_id is held in the event, so this deref
3610	 * must be after the event is delivered above.
3611	 */
3612	cm_deref_id(listen_cm_id_priv);
3613	if (ret)
3614		cm_destroy_id(&cm_id_priv->id, ret);
3615	return 0;
3616out:
3617	ib_destroy_cm_id(&cm_id_priv->id);
3618	return -EINVAL;
3619}
3620
3621static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3622			       struct cm_id_private *cm_id_priv,
3623			       struct ib_cm_sidr_rep_param *param)
3624{
3625	cm_format_mad_ece_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3626			      cm_id_priv->tid, param->ece.attr_mod);
3627	IBA_SET(CM_SIDR_REP_REQUESTID, sidr_rep_msg,
3628		be32_to_cpu(cm_id_priv->id.remote_id));
3629	IBA_SET(CM_SIDR_REP_STATUS, sidr_rep_msg, param->status);
3630	IBA_SET(CM_SIDR_REP_QPN, sidr_rep_msg, param->qp_num);
3631	IBA_SET(CM_SIDR_REP_SERVICEID, sidr_rep_msg,
3632		be64_to_cpu(cm_id_priv->id.service_id));
3633	IBA_SET(CM_SIDR_REP_Q_KEY, sidr_rep_msg, param->qkey);
3634	IBA_SET(CM_SIDR_REP_VENDOR_ID_L, sidr_rep_msg,
3635		param->ece.vendor_id & 0xFF);
3636	IBA_SET(CM_SIDR_REP_VENDOR_ID_H, sidr_rep_msg,
3637		(param->ece.vendor_id >> 8) & 0xFF);
3638
3639	if (param->info && param->info_length)
3640		IBA_SET_MEM(CM_SIDR_REP_ADDITIONAL_INFORMATION, sidr_rep_msg,
3641			    param->info, param->info_length);
3642
3643	if (param->private_data && param->private_data_len)
3644		IBA_SET_MEM(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg,
3645			    param->private_data, param->private_data_len);
3646}
3647
3648static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv,
3649				   struct ib_cm_sidr_rep_param *param)
3650{
3651	struct ib_mad_send_buf *msg;
3652	unsigned long flags;
3653	int ret;
3654
3655	lockdep_assert_held(&cm_id_priv->lock);
3656
3657	if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3658	    (param->private_data &&
3659	     param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3660		return -EINVAL;
3661
3662	if (cm_id_priv->id.state != IB_CM_SIDR_REQ_RCVD)
3663		return -EINVAL;
3664
3665	msg = cm_alloc_msg(cm_id_priv);
3666	if (IS_ERR(msg))
3667		return PTR_ERR(msg);
3668
3669	cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3670			   param);
3671	trace_icm_send_sidr_rep(&cm_id_priv->id);
3672	ret = ib_post_send_mad(msg, NULL);
3673	if (ret) {
3674		cm_free_msg(msg);
3675		return ret;
3676	}
3677	cm_id_priv->id.state = IB_CM_IDLE;
3678	spin_lock_irqsave(&cm.lock, flags);
3679	if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3680		rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3681		RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3682	}
3683	spin_unlock_irqrestore(&cm.lock, flags);
3684	return 0;
3685}
3686
3687int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3688			struct ib_cm_sidr_rep_param *param)
3689{
3690	struct cm_id_private *cm_id_priv =
3691		container_of(cm_id, struct cm_id_private, id);
3692	unsigned long flags;
3693	int ret;
3694
3695	spin_lock_irqsave(&cm_id_priv->lock, flags);
3696	ret = cm_send_sidr_rep_locked(cm_id_priv, param);
3697	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3698	return ret;
3699}
3700EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3701
3702static void cm_format_sidr_rep_event(struct cm_work *work,
3703				     const struct cm_id_private *cm_id_priv)
3704{
3705	struct cm_sidr_rep_msg *sidr_rep_msg;
3706	struct ib_cm_sidr_rep_event_param *param;
3707
3708	sidr_rep_msg = (struct cm_sidr_rep_msg *)
3709				work->mad_recv_wc->recv_buf.mad;
3710	param = &work->cm_event.param.sidr_rep_rcvd;
3711	param->status = IBA_GET(CM_SIDR_REP_STATUS, sidr_rep_msg);
3712	param->qkey = IBA_GET(CM_SIDR_REP_Q_KEY, sidr_rep_msg);
3713	param->qpn = IBA_GET(CM_SIDR_REP_QPN, sidr_rep_msg);
3714	param->info = IBA_GET_MEM_PTR(CM_SIDR_REP_ADDITIONAL_INFORMATION,
3715				      sidr_rep_msg);
3716	param->info_len = IBA_GET(CM_SIDR_REP_ADDITIONAL_INFORMATION_LENGTH,
3717				  sidr_rep_msg);
3718	param->sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr;
3719	work->cm_event.private_data =
3720		IBA_GET_MEM_PTR(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg);
3721}
3722
3723static int cm_sidr_rep_handler(struct cm_work *work)
3724{
3725	struct cm_sidr_rep_msg *sidr_rep_msg;
3726	struct cm_id_private *cm_id_priv;
3727
3728	sidr_rep_msg = (struct cm_sidr_rep_msg *)
3729				work->mad_recv_wc->recv_buf.mad;
3730	cm_id_priv = cm_acquire_id(
3731		cpu_to_be32(IBA_GET(CM_SIDR_REP_REQUESTID, sidr_rep_msg)), 0);
3732	if (!cm_id_priv)
3733		return -EINVAL; /* Unmatched reply. */
3734
3735	spin_lock_irq(&cm_id_priv->lock);
3736	if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3737		spin_unlock_irq(&cm_id_priv->lock);
3738		goto out;
3739	}
3740	cm_id_priv->id.state = IB_CM_IDLE;
3741	ib_cancel_mad(cm_id_priv->msg);
3742	spin_unlock_irq(&cm_id_priv->lock);
3743
3744	cm_format_sidr_rep_event(work, cm_id_priv);
3745	cm_process_work(cm_id_priv, work);
3746	return 0;
3747out:
3748	cm_deref_id(cm_id_priv);
3749	return -EINVAL;
3750}
3751
3752static void cm_process_send_error(struct cm_id_private *cm_id_priv,
3753				  struct ib_mad_send_buf *msg,
3754				  enum ib_wc_status wc_status)
3755{
3756	enum ib_cm_state state = (unsigned long) msg->context[1];
3757	struct ib_cm_event cm_event = {};
3758	int ret;
3759
3760	/* Discard old sends. */
3761	spin_lock_irq(&cm_id_priv->lock);
3762	if (msg != cm_id_priv->msg) {
3763		spin_unlock_irq(&cm_id_priv->lock);
3764		cm_free_msg(msg);
3765		cm_deref_id(cm_id_priv);
3766		return;
3767	}
3768	cm_free_priv_msg(msg);
3769
3770	if (state != cm_id_priv->id.state || wc_status == IB_WC_SUCCESS ||
3771	    wc_status == IB_WC_WR_FLUSH_ERR)
3772		goto out_unlock;
3773
3774	trace_icm_mad_send_err(state, wc_status);
3775	switch (state) {
3776	case IB_CM_REQ_SENT:
3777	case IB_CM_MRA_REQ_RCVD:
3778		cm_reset_to_idle(cm_id_priv);
3779		cm_event.event = IB_CM_REQ_ERROR;
3780		break;
3781	case IB_CM_REP_SENT:
3782	case IB_CM_MRA_REP_RCVD:
3783		cm_reset_to_idle(cm_id_priv);
3784		cm_event.event = IB_CM_REP_ERROR;
3785		break;
3786	case IB_CM_DREQ_SENT:
3787		cm_enter_timewait(cm_id_priv);
3788		cm_event.event = IB_CM_DREQ_ERROR;
3789		break;
3790	case IB_CM_SIDR_REQ_SENT:
3791		cm_id_priv->id.state = IB_CM_IDLE;
3792		cm_event.event = IB_CM_SIDR_REQ_ERROR;
3793		break;
3794	default:
3795		goto out_unlock;
3796	}
3797	spin_unlock_irq(&cm_id_priv->lock);
3798	cm_event.param.send_status = wc_status;
3799
3800	/* No other events can occur on the cm_id at this point. */
3801	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3802	if (ret)
3803		ib_destroy_cm_id(&cm_id_priv->id);
3804	return;
3805out_unlock:
3806	spin_unlock_irq(&cm_id_priv->lock);
3807}
3808
3809static void cm_send_handler(struct ib_mad_agent *mad_agent,
3810			    struct ib_mad_send_wc *mad_send_wc)
3811{
3812	struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3813	struct cm_id_private *cm_id_priv;
3814	struct cm_port *port;
3815	u16 attr_index;
3816
3817	port = mad_agent->context;
3818	attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3819				  msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3820
3821	if (msg->context[0] == CM_DIRECT_RETRY_CTX) {
3822		msg->retries = 1;
3823		cm_id_priv = NULL;
3824	} else {
3825		cm_id_priv = msg->context[0];
3826	}
3827
3828	atomic_long_add(1 + msg->retries, &port->counters[CM_XMIT][attr_index]);
3829	if (msg->retries)
3830		atomic_long_add(msg->retries,
3831				&port->counters[CM_XMIT_RETRIES][attr_index]);
3832
3833	if (cm_id_priv)
3834		cm_process_send_error(cm_id_priv, msg, mad_send_wc->status);
3835	else
3836		cm_free_msg(msg);
3837}
3838
3839static void cm_work_handler(struct work_struct *_work)
3840{
3841	struct cm_work *work = container_of(_work, struct cm_work, work.work);
3842	int ret;
3843
3844	switch (work->cm_event.event) {
3845	case IB_CM_REQ_RECEIVED:
3846		ret = cm_req_handler(work);
3847		break;
3848	case IB_CM_MRA_RECEIVED:
3849		ret = cm_mra_handler(work);
3850		break;
3851	case IB_CM_REJ_RECEIVED:
3852		ret = cm_rej_handler(work);
3853		break;
3854	case IB_CM_REP_RECEIVED:
3855		ret = cm_rep_handler(work);
3856		break;
3857	case IB_CM_RTU_RECEIVED:
3858		ret = cm_rtu_handler(work);
3859		break;
3860	case IB_CM_USER_ESTABLISHED:
3861		ret = cm_establish_handler(work);
3862		break;
3863	case IB_CM_DREQ_RECEIVED:
3864		ret = cm_dreq_handler(work);
3865		break;
3866	case IB_CM_DREP_RECEIVED:
3867		ret = cm_drep_handler(work);
3868		break;
3869	case IB_CM_SIDR_REQ_RECEIVED:
3870		ret = cm_sidr_req_handler(work);
3871		break;
3872	case IB_CM_SIDR_REP_RECEIVED:
3873		ret = cm_sidr_rep_handler(work);
3874		break;
3875	case IB_CM_LAP_RECEIVED:
3876		ret = cm_lap_handler(work);
3877		break;
3878	case IB_CM_APR_RECEIVED:
3879		ret = cm_apr_handler(work);
3880		break;
3881	case IB_CM_TIMEWAIT_EXIT:
3882		ret = cm_timewait_handler(work);
3883		break;
3884	default:
3885		trace_icm_handler_err(work->cm_event.event);
3886		ret = -EINVAL;
3887		break;
3888	}
3889	if (ret)
3890		cm_free_work(work);
3891}
3892
3893static int cm_establish(struct ib_cm_id *cm_id)
3894{
3895	struct cm_id_private *cm_id_priv;
3896	struct cm_work *work;
3897	unsigned long flags;
3898	int ret = 0;
3899	struct cm_device *cm_dev;
3900
3901	cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3902	if (!cm_dev)
3903		return -ENODEV;
3904
3905	work = kmalloc(sizeof *work, GFP_ATOMIC);
3906	if (!work)
3907		return -ENOMEM;
3908
3909	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3910	spin_lock_irqsave(&cm_id_priv->lock, flags);
3911	switch (cm_id->state) {
3912	case IB_CM_REP_SENT:
3913	case IB_CM_MRA_REP_RCVD:
3914		cm_id->state = IB_CM_ESTABLISHED;
3915		break;
3916	case IB_CM_ESTABLISHED:
3917		ret = -EISCONN;
3918		break;
3919	default:
3920		trace_icm_establish_err(cm_id);
3921		ret = -EINVAL;
3922		break;
3923	}
3924	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3925
3926	if (ret) {
3927		kfree(work);
3928		goto out;
3929	}
3930
3931	/*
3932	 * The CM worker thread may try to destroy the cm_id before it
3933	 * can execute this work item.  To prevent potential deadlock,
3934	 * we need to find the cm_id once we're in the context of the
3935	 * worker thread, rather than holding a reference on it.
3936	 */
3937	INIT_DELAYED_WORK(&work->work, cm_work_handler);
3938	work->local_id = cm_id->local_id;
3939	work->remote_id = cm_id->remote_id;
3940	work->mad_recv_wc = NULL;
3941	work->cm_event.event = IB_CM_USER_ESTABLISHED;
3942
3943	/* Check if the device started its remove_one */
3944	spin_lock_irqsave(&cm.lock, flags);
3945	if (!cm_dev->going_down) {
3946		queue_delayed_work(cm.wq, &work->work, 0);
3947	} else {
3948		kfree(work);
3949		ret = -ENODEV;
3950	}
3951	spin_unlock_irqrestore(&cm.lock, flags);
3952
3953out:
3954	return ret;
3955}
3956
3957static int cm_migrate(struct ib_cm_id *cm_id)
3958{
3959	struct cm_id_private *cm_id_priv;
3960	unsigned long flags;
3961	int ret = 0;
3962
3963	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3964	spin_lock_irqsave(&cm_id_priv->lock, flags);
3965	if (cm_id->state == IB_CM_ESTABLISHED &&
3966	    (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3967	     cm_id->lap_state == IB_CM_LAP_IDLE)) {
3968		cm_id->lap_state = IB_CM_LAP_IDLE;
3969		cm_id_priv->av = cm_id_priv->alt_av;
3970	} else
3971		ret = -EINVAL;
3972	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3973
3974	return ret;
3975}
3976
3977int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3978{
3979	int ret;
3980
3981	switch (event) {
3982	case IB_EVENT_COMM_EST:
3983		ret = cm_establish(cm_id);
3984		break;
3985	case IB_EVENT_PATH_MIG:
3986		ret = cm_migrate(cm_id);
3987		break;
3988	default:
3989		ret = -EINVAL;
3990	}
3991	return ret;
3992}
3993EXPORT_SYMBOL(ib_cm_notify);
3994
3995static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3996			    struct ib_mad_send_buf *send_buf,
3997			    struct ib_mad_recv_wc *mad_recv_wc)
3998{
3999	struct cm_port *port = mad_agent->context;
4000	struct cm_work *work;
4001	enum ib_cm_event_type event;
4002	bool alt_path = false;
4003	u16 attr_id;
4004	int paths = 0;
4005	int going_down = 0;
4006
4007	switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
4008	case CM_REQ_ATTR_ID:
4009		alt_path = cm_req_has_alt_path((struct cm_req_msg *)
4010						mad_recv_wc->recv_buf.mad);
4011		paths = 1 + (alt_path != 0);
4012		event = IB_CM_REQ_RECEIVED;
4013		break;
4014	case CM_MRA_ATTR_ID:
4015		event = IB_CM_MRA_RECEIVED;
4016		break;
4017	case CM_REJ_ATTR_ID:
4018		event = IB_CM_REJ_RECEIVED;
4019		break;
4020	case CM_REP_ATTR_ID:
4021		event = IB_CM_REP_RECEIVED;
4022		break;
4023	case CM_RTU_ATTR_ID:
4024		event = IB_CM_RTU_RECEIVED;
4025		break;
4026	case CM_DREQ_ATTR_ID:
4027		event = IB_CM_DREQ_RECEIVED;
4028		break;
4029	case CM_DREP_ATTR_ID:
4030		event = IB_CM_DREP_RECEIVED;
4031		break;
4032	case CM_SIDR_REQ_ATTR_ID:
4033		event = IB_CM_SIDR_REQ_RECEIVED;
4034		break;
4035	case CM_SIDR_REP_ATTR_ID:
4036		event = IB_CM_SIDR_REP_RECEIVED;
4037		break;
4038	case CM_LAP_ATTR_ID:
4039		paths = 1;
4040		event = IB_CM_LAP_RECEIVED;
4041		break;
4042	case CM_APR_ATTR_ID:
4043		event = IB_CM_APR_RECEIVED;
4044		break;
4045	default:
4046		ib_free_recv_mad(mad_recv_wc);
4047		return;
4048	}
4049
4050	attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
4051	atomic_long_inc(&port->counters[CM_RECV][attr_id - CM_ATTR_ID_OFFSET]);
4052
4053	work = kmalloc(struct_size(work, path, paths), GFP_KERNEL);
4054	if (!work) {
4055		ib_free_recv_mad(mad_recv_wc);
4056		return;
4057	}
4058
4059	INIT_DELAYED_WORK(&work->work, cm_work_handler);
4060	work->cm_event.event = event;
4061	work->mad_recv_wc = mad_recv_wc;
4062	work->port = port;
4063
4064	/* Check if the device started its remove_one */
4065	spin_lock_irq(&cm.lock);
4066	if (!port->cm_dev->going_down)
4067		queue_delayed_work(cm.wq, &work->work, 0);
4068	else
4069		going_down = 1;
4070	spin_unlock_irq(&cm.lock);
4071
4072	if (going_down) {
4073		kfree(work);
4074		ib_free_recv_mad(mad_recv_wc);
4075	}
4076}
4077
4078static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
4079				struct ib_qp_attr *qp_attr,
4080				int *qp_attr_mask)
4081{
4082	unsigned long flags;
4083	int ret;
4084
4085	spin_lock_irqsave(&cm_id_priv->lock, flags);
4086	switch (cm_id_priv->id.state) {
4087	case IB_CM_REQ_SENT:
4088	case IB_CM_MRA_REQ_RCVD:
4089	case IB_CM_REQ_RCVD:
4090	case IB_CM_MRA_REQ_SENT:
4091	case IB_CM_REP_RCVD:
4092	case IB_CM_MRA_REP_SENT:
4093	case IB_CM_REP_SENT:
4094	case IB_CM_MRA_REP_RCVD:
4095	case IB_CM_ESTABLISHED:
4096		*qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
4097				IB_QP_PKEY_INDEX | IB_QP_PORT;
4098		qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
4099		if (cm_id_priv->responder_resources) {
4100			struct ib_device *ib_dev = cm_id_priv->id.device;
4101			u64 support_flush = ib_dev->attrs.device_cap_flags &
4102			  (IB_DEVICE_FLUSH_GLOBAL | IB_DEVICE_FLUSH_PERSISTENT);
4103			u32 flushable = support_flush ?
4104					(IB_ACCESS_FLUSH_GLOBAL |
4105					 IB_ACCESS_FLUSH_PERSISTENT) : 0;
4106
4107			qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
4108						    IB_ACCESS_REMOTE_ATOMIC |
4109						    flushable;
4110		}
4111		qp_attr->pkey_index = cm_id_priv->av.pkey_index;
4112		if (cm_id_priv->av.port)
4113			qp_attr->port_num = cm_id_priv->av.port->port_num;
4114		ret = 0;
4115		break;
4116	default:
4117		trace_icm_qp_init_err(&cm_id_priv->id);
4118		ret = -EINVAL;
4119		break;
4120	}
4121	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4122	return ret;
4123}
4124
4125static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
4126			       struct ib_qp_attr *qp_attr,
4127			       int *qp_attr_mask)
4128{
4129	unsigned long flags;
4130	int ret;
4131
4132	spin_lock_irqsave(&cm_id_priv->lock, flags);
4133	switch (cm_id_priv->id.state) {
4134	case IB_CM_REQ_RCVD:
4135	case IB_CM_MRA_REQ_SENT:
4136	case IB_CM_REP_RCVD:
4137	case IB_CM_MRA_REP_SENT:
4138	case IB_CM_REP_SENT:
4139	case IB_CM_MRA_REP_RCVD:
4140	case IB_CM_ESTABLISHED:
4141		*qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
4142				IB_QP_DEST_QPN | IB_QP_RQ_PSN;
4143		qp_attr->ah_attr = cm_id_priv->av.ah_attr;
4144		if ((qp_attr->ah_attr.type == RDMA_AH_ATTR_TYPE_IB) &&
4145		    cm_id_priv->av.dlid_datapath &&
4146		    (cm_id_priv->av.dlid_datapath != 0xffff))
4147			qp_attr->ah_attr.ib.dlid = cm_id_priv->av.dlid_datapath;
4148		qp_attr->path_mtu = cm_id_priv->path_mtu;
4149		qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
4150		qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
4151		if (cm_id_priv->qp_type == IB_QPT_RC ||
4152		    cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
4153			*qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
4154					 IB_QP_MIN_RNR_TIMER;
4155			qp_attr->max_dest_rd_atomic =
4156					cm_id_priv->responder_resources;
4157			qp_attr->min_rnr_timer = 0;
4158		}
4159		if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr) &&
4160		    cm_id_priv->alt_av.port) {
4161			*qp_attr_mask |= IB_QP_ALT_PATH;
4162			qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4163			qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4164			qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4165			qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4166		}
4167		ret = 0;
4168		break;
4169	default:
4170		trace_icm_qp_rtr_err(&cm_id_priv->id);
4171		ret = -EINVAL;
4172		break;
4173	}
4174	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4175	return ret;
4176}
4177
4178static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
4179			       struct ib_qp_attr *qp_attr,
4180			       int *qp_attr_mask)
4181{
4182	unsigned long flags;
4183	int ret;
4184
4185	spin_lock_irqsave(&cm_id_priv->lock, flags);
4186	switch (cm_id_priv->id.state) {
4187	/* Allow transition to RTS before sending REP */
4188	case IB_CM_REQ_RCVD:
4189	case IB_CM_MRA_REQ_SENT:
4190
4191	case IB_CM_REP_RCVD:
4192	case IB_CM_MRA_REP_SENT:
4193	case IB_CM_REP_SENT:
4194	case IB_CM_MRA_REP_RCVD:
4195	case IB_CM_ESTABLISHED:
4196		if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
4197			*qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
4198			qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
4199			switch (cm_id_priv->qp_type) {
4200			case IB_QPT_RC:
4201			case IB_QPT_XRC_INI:
4202				*qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4203						 IB_QP_MAX_QP_RD_ATOMIC;
4204				qp_attr->retry_cnt = cm_id_priv->retry_count;
4205				qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
4206				qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
4207				fallthrough;
4208			case IB_QPT_XRC_TGT:
4209				*qp_attr_mask |= IB_QP_TIMEOUT;
4210				qp_attr->timeout = cm_id_priv->av.timeout;
4211				break;
4212			default:
4213				break;
4214			}
4215			if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4216				*qp_attr_mask |= IB_QP_PATH_MIG_STATE;
4217				qp_attr->path_mig_state = IB_MIG_REARM;
4218			}
4219		} else {
4220			*qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
4221			if (cm_id_priv->alt_av.port)
4222				qp_attr->alt_port_num =
4223					cm_id_priv->alt_av.port->port_num;
4224			qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4225			qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4226			qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4227			qp_attr->path_mig_state = IB_MIG_REARM;
4228		}
4229		ret = 0;
4230		break;
4231	default:
4232		trace_icm_qp_rts_err(&cm_id_priv->id);
4233		ret = -EINVAL;
4234		break;
4235	}
4236	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4237	return ret;
4238}
4239
4240int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
4241		       struct ib_qp_attr *qp_attr,
4242		       int *qp_attr_mask)
4243{
4244	struct cm_id_private *cm_id_priv;
4245	int ret;
4246
4247	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
4248	switch (qp_attr->qp_state) {
4249	case IB_QPS_INIT:
4250		ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
4251		break;
4252	case IB_QPS_RTR:
4253		ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
4254		break;
4255	case IB_QPS_RTS:
4256		ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
4257		break;
4258	default:
4259		ret = -EINVAL;
4260		break;
4261	}
4262	return ret;
4263}
4264EXPORT_SYMBOL(ib_cm_init_qp_attr);
4265
4266static ssize_t cm_show_counter(struct ib_device *ibdev, u32 port_num,
4267			       struct ib_port_attribute *attr, char *buf)
4268{
4269	struct cm_counter_attribute *cm_attr =
4270		container_of(attr, struct cm_counter_attribute, attr);
4271	struct cm_device *cm_dev = ib_get_client_data(ibdev, &cm_client);
4272
4273	if (WARN_ON(!cm_dev))
4274		return -EINVAL;
4275
4276	return sysfs_emit(
4277		buf, "%ld\n",
4278		atomic_long_read(
4279			&cm_dev->port[port_num - 1]
4280				 ->counters[cm_attr->group][cm_attr->index]));
4281}
4282
4283#define CM_COUNTER_ATTR(_name, _group, _index)                                 \
4284	{                                                                      \
4285		.attr = __ATTR(_name, 0444, cm_show_counter, NULL),            \
4286		.group = _group, .index = _index                               \
4287	}
4288
4289#define CM_COUNTER_GROUP(_group, _name)                                        \
4290	static struct cm_counter_attribute cm_counter_attr_##_group[] = {      \
4291		CM_COUNTER_ATTR(req, _group, CM_REQ_COUNTER),                  \
4292		CM_COUNTER_ATTR(mra, _group, CM_MRA_COUNTER),                  \
4293		CM_COUNTER_ATTR(rej, _group, CM_REJ_COUNTER),                  \
4294		CM_COUNTER_ATTR(rep, _group, CM_REP_COUNTER),                  \
4295		CM_COUNTER_ATTR(rtu, _group, CM_RTU_COUNTER),                  \
4296		CM_COUNTER_ATTR(dreq, _group, CM_DREQ_COUNTER),                \
4297		CM_COUNTER_ATTR(drep, _group, CM_DREP_COUNTER),                \
4298		CM_COUNTER_ATTR(sidr_req, _group, CM_SIDR_REQ_COUNTER),        \
4299		CM_COUNTER_ATTR(sidr_rep, _group, CM_SIDR_REP_COUNTER),        \
4300		CM_COUNTER_ATTR(lap, _group, CM_LAP_COUNTER),                  \
4301		CM_COUNTER_ATTR(apr, _group, CM_APR_COUNTER),                  \
4302	};                                                                     \
4303	static struct attribute *cm_counter_attrs_##_group[] = {               \
4304		&cm_counter_attr_##_group[0].attr.attr,                        \
4305		&cm_counter_attr_##_group[1].attr.attr,                        \
4306		&cm_counter_attr_##_group[2].attr.attr,                        \
4307		&cm_counter_attr_##_group[3].attr.attr,                        \
4308		&cm_counter_attr_##_group[4].attr.attr,                        \
4309		&cm_counter_attr_##_group[5].attr.attr,                        \
4310		&cm_counter_attr_##_group[6].attr.attr,                        \
4311		&cm_counter_attr_##_group[7].attr.attr,                        \
4312		&cm_counter_attr_##_group[8].attr.attr,                        \
4313		&cm_counter_attr_##_group[9].attr.attr,                        \
4314		&cm_counter_attr_##_group[10].attr.attr,                       \
4315		NULL,                                                          \
4316	};                                                                     \
4317	static const struct attribute_group cm_counter_group_##_group = {      \
4318		.name = _name,                                                 \
4319		.attrs = cm_counter_attrs_##_group,                            \
4320	};
4321
4322CM_COUNTER_GROUP(CM_XMIT, "cm_tx_msgs")
4323CM_COUNTER_GROUP(CM_XMIT_RETRIES, "cm_tx_retries")
4324CM_COUNTER_GROUP(CM_RECV, "cm_rx_msgs")
4325CM_COUNTER_GROUP(CM_RECV_DUPLICATES, "cm_rx_duplicates")
4326
4327static const struct attribute_group *cm_counter_groups[] = {
4328	&cm_counter_group_CM_XMIT,
4329	&cm_counter_group_CM_XMIT_RETRIES,
4330	&cm_counter_group_CM_RECV,
4331	&cm_counter_group_CM_RECV_DUPLICATES,
4332	NULL,
4333};
4334
4335static int cm_add_one(struct ib_device *ib_device)
4336{
4337	struct cm_device *cm_dev;
4338	struct cm_port *port;
4339	struct ib_mad_reg_req reg_req = {
4340		.mgmt_class = IB_MGMT_CLASS_CM,
4341		.mgmt_class_version = IB_CM_CLASS_VERSION,
4342	};
4343	struct ib_port_modify port_modify = {
4344		.set_port_cap_mask = IB_PORT_CM_SUP
4345	};
4346	unsigned long flags;
4347	int ret;
4348	int count = 0;
4349	u32 i;
4350
4351	cm_dev = kzalloc(struct_size(cm_dev, port, ib_device->phys_port_cnt),
4352			 GFP_KERNEL);
4353	if (!cm_dev)
4354		return -ENOMEM;
4355
4356	kref_init(&cm_dev->kref);
4357	rwlock_init(&cm_dev->mad_agent_lock);
4358	cm_dev->ib_device = ib_device;
4359	cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
4360	cm_dev->going_down = 0;
4361
4362	ib_set_client_data(ib_device, &cm_client, cm_dev);
4363
4364	set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
4365	rdma_for_each_port (ib_device, i) {
4366		if (!rdma_cap_ib_cm(ib_device, i))
4367			continue;
4368
4369		port = kzalloc(sizeof *port, GFP_KERNEL);
4370		if (!port) {
4371			ret = -ENOMEM;
4372			goto error1;
4373		}
4374
4375		cm_dev->port[i-1] = port;
4376		port->cm_dev = cm_dev;
4377		port->port_num = i;
4378
4379		ret = ib_port_register_client_groups(ib_device, i,
4380						     cm_counter_groups);
4381		if (ret)
4382			goto error1;
4383
4384		port->mad_agent = ib_register_mad_agent(ib_device, i,
4385							IB_QPT_GSI,
4386							&reg_req,
4387							0,
4388							cm_send_handler,
4389							cm_recv_handler,
4390							port,
4391							0);
4392		if (IS_ERR(port->mad_agent)) {
4393			ret = PTR_ERR(port->mad_agent);
4394			goto error2;
4395		}
4396
4397		port->rep_agent = ib_register_mad_agent(ib_device, i,
4398							IB_QPT_GSI,
4399							NULL,
4400							0,
4401							cm_send_handler,
4402							NULL,
4403							port,
4404							0);
4405		if (IS_ERR(port->rep_agent)) {
4406			ret = PTR_ERR(port->rep_agent);
4407			goto error3;
4408		}
4409
4410		ret = ib_modify_port(ib_device, i, 0, &port_modify);
4411		if (ret)
4412			goto error4;
4413
4414		count++;
4415	}
4416
4417	if (!count) {
4418		ret = -EOPNOTSUPP;
4419		goto free;
4420	}
4421
4422	write_lock_irqsave(&cm.device_lock, flags);
4423	list_add_tail(&cm_dev->list, &cm.device_list);
4424	write_unlock_irqrestore(&cm.device_lock, flags);
4425	return 0;
4426
4427error4:
4428	ib_unregister_mad_agent(port->rep_agent);
4429error3:
4430	ib_unregister_mad_agent(port->mad_agent);
4431error2:
4432	ib_port_unregister_client_groups(ib_device, i, cm_counter_groups);
4433error1:
4434	port_modify.set_port_cap_mask = 0;
4435	port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4436	while (--i) {
4437		if (!rdma_cap_ib_cm(ib_device, i))
4438			continue;
4439
4440		port = cm_dev->port[i-1];
4441		ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4442		ib_unregister_mad_agent(port->rep_agent);
4443		ib_unregister_mad_agent(port->mad_agent);
4444		ib_port_unregister_client_groups(ib_device, i,
4445						 cm_counter_groups);
4446	}
4447free:
4448	cm_device_put(cm_dev);
4449	return ret;
4450}
4451
4452static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4453{
4454	struct cm_device *cm_dev = client_data;
4455	struct cm_port *port;
4456	struct ib_port_modify port_modify = {
4457		.clr_port_cap_mask = IB_PORT_CM_SUP
4458	};
4459	unsigned long flags;
4460	u32 i;
4461
4462	write_lock_irqsave(&cm.device_lock, flags);
4463	list_del(&cm_dev->list);
4464	write_unlock_irqrestore(&cm.device_lock, flags);
4465
4466	spin_lock_irq(&cm.lock);
4467	cm_dev->going_down = 1;
4468	spin_unlock_irq(&cm.lock);
4469
4470	rdma_for_each_port (ib_device, i) {
4471		struct ib_mad_agent *mad_agent;
4472		struct ib_mad_agent *rep_agent;
4473
4474		if (!rdma_cap_ib_cm(ib_device, i))
4475			continue;
4476
4477		port = cm_dev->port[i-1];
4478		mad_agent = port->mad_agent;
4479		rep_agent = port->rep_agent;
4480		ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4481		/*
4482		 * We flush the queue here after the going_down set, this
4483		 * verify that no new works will be queued in the recv handler,
4484		 * after that we can call the unregister_mad_agent
4485		 */
4486		flush_workqueue(cm.wq);
4487		/*
4488		 * The above ensures no call paths from the work are running,
4489		 * the remaining paths all take the mad_agent_lock.
4490		 */
4491		write_lock(&cm_dev->mad_agent_lock);
4492		port->mad_agent = NULL;
4493		port->rep_agent = NULL;
4494		write_unlock(&cm_dev->mad_agent_lock);
4495		ib_unregister_mad_agent(mad_agent);
4496		ib_unregister_mad_agent(rep_agent);
4497		ib_port_unregister_client_groups(ib_device, i,
4498						 cm_counter_groups);
4499	}
4500
4501	cm_device_put(cm_dev);
4502}
4503
4504static int __init ib_cm_init(void)
4505{
4506	int ret;
4507
4508	INIT_LIST_HEAD(&cm.device_list);
4509	rwlock_init(&cm.device_lock);
4510	spin_lock_init(&cm.lock);
4511	cm.listen_service_table = RB_ROOT;
4512	cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4513	cm.remote_id_table = RB_ROOT;
4514	cm.remote_qp_table = RB_ROOT;
4515	cm.remote_sidr_table = RB_ROOT;
4516	xa_init_flags(&cm.local_id_table, XA_FLAGS_ALLOC);
4517	get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4518	INIT_LIST_HEAD(&cm.timewait_list);
4519
4520	cm.wq = alloc_workqueue("ib_cm", WQ_PERCPU, 1);
4521	if (!cm.wq) {
4522		ret = -ENOMEM;
4523		goto error2;
4524	}
4525
4526	ret = ib_register_client(&cm_client);
4527	if (ret)
4528		goto error3;
4529
4530	return 0;
4531error3:
4532	destroy_workqueue(cm.wq);
4533error2:
4534	return ret;
4535}
4536
4537static void __exit ib_cm_cleanup(void)
4538{
4539	struct cm_timewait_info *timewait_info, *tmp;
4540
4541	spin_lock_irq(&cm.lock);
4542	list_for_each_entry(timewait_info, &cm.timewait_list, list)
4543		cancel_delayed_work(&timewait_info->work.work);
4544	spin_unlock_irq(&cm.lock);
4545
4546	ib_unregister_client(&cm_client);
4547	destroy_workqueue(cm.wq);
4548
4549	list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4550		list_del(&timewait_info->list);
4551		kfree(timewait_info);
4552	}
4553
4554	WARN_ON(!xa_empty(&cm.local_id_table));
4555}
4556
4557module_init(ib_cm_init);
4558module_exit(ib_cm_cleanup);