1/*
   2 * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3 * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
   4 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   5 * Copyright (c) 2008 Cisco. All rights reserved.
   6 *
   7 * This software is available to you under a choice of one of two
   8 * licenses.  You may choose to be licensed under the terms of the GNU
   9 * General Public License (GPL) Version 2, available from the file
  10 * COPYING in the main directory of this source tree, or the
  11 * OpenIB.org BSD license below:
  12 *
  13 *     Redistribution and use in source and binary forms, with or
  14 *     without modification, are permitted provided that the following
  15 *     conditions are met:
  16 *
  17 *      - Redistributions of source code must retain the above
  18 *        copyright notice, this list of conditions and the following
  19 *        disclaimer.
  20 *
  21 *      - Redistributions in binary form must reproduce the above
  22 *        copyright notice, this list of conditions and the following
  23 *        disclaimer in the documentation and/or other materials
  24 *        provided with the distribution.
  25 *
  26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33 * SOFTWARE.
  34 */
  35
  36#define pr_fmt(fmt) "user_mad: " fmt
  37
  38#include <linux/module.h>
  39#include <linux/init.h>
  40#include <linux/device.h>
  41#include <linux/err.h>
  42#include <linux/fs.h>
  43#include <linux/cdev.h>
  44#include <linux/dma-mapping.h>
  45#include <linux/poll.h>
  46#include <linux/mutex.h>
  47#include <linux/kref.h>
  48#include <linux/compat.h>
  49#include <linux/sched.h>
  50#include <linux/semaphore.h>
  51#include <linux/slab.h>
  52#include <linux/nospec.h>
  53
  54#include <linux/uaccess.h>
  55
  56#include <rdma/ib_mad.h>
  57#include <rdma/ib_user_mad.h>
  58#include <rdma/rdma_netlink.h>
  59
  60#include "core_priv.h"
  61
  62MODULE_AUTHOR("Roland Dreier");
  63MODULE_DESCRIPTION("InfiniBand userspace MAD packet access");
  64MODULE_LICENSE("Dual BSD/GPL");
  65
  66enum {
  67	IB_UMAD_MAX_PORTS  = RDMA_MAX_PORTS,
  68	IB_UMAD_MAX_AGENTS = 32,
  69
  70	IB_UMAD_MAJOR      = 231,
  71	IB_UMAD_MINOR_BASE = 0,
  72	IB_UMAD_NUM_FIXED_MINOR = 64,
  73	IB_UMAD_NUM_DYNAMIC_MINOR = IB_UMAD_MAX_PORTS - IB_UMAD_NUM_FIXED_MINOR,
  74	IB_ISSM_MINOR_BASE        = IB_UMAD_NUM_FIXED_MINOR,
  75};
  76
  77/*
  78 * Our lifetime rules for these structs are the following:
  79 * device special file is opened, we take a reference on the
  80 * ib_umad_port's struct ib_umad_device. We drop these
  81 * references in the corresponding close().
  82 *
  83 * In addition to references coming from open character devices, there
  84 * is one more reference to each ib_umad_device representing the
  85 * module's reference taken when allocating the ib_umad_device in
  86 * ib_umad_add_one().
  87 *
  88 * When destroying an ib_umad_device, we drop the module's reference.
  89 */
  90
  91struct ib_umad_port {
  92	struct cdev           cdev;
  93	struct device	      dev;
  94	struct cdev           sm_cdev;
  95	struct device	      sm_dev;
  96	struct semaphore       sm_sem;
  97
  98	struct mutex	       file_mutex;
  99	struct list_head       file_list;
 100
 101	struct ib_device      *ib_dev;
 102	struct ib_umad_device *umad_dev;
 103	int                    dev_num;
 104	u32                     port_num;
 105};
 106
 107struct ib_umad_device {
 108	struct kref kref;
 109	struct ib_umad_port ports[];
 110};
 111
 112struct ib_umad_file {
 113	struct mutex		mutex;
 114	struct ib_umad_port    *port;
 115	struct list_head	recv_list;
 116	struct list_head	send_list;
 117	struct list_head	port_list;
 118	spinlock_t		send_lock;
 119	wait_queue_head_t	recv_wait;
 120	struct ib_mad_agent    *agent[IB_UMAD_MAX_AGENTS];
 121	int			agents_dead;
 122	u8			use_pkey_index;
 123	u8			already_used;
 124};
 125
 126struct ib_umad_packet {
 127	struct ib_mad_send_buf *msg;
 128	struct ib_mad_recv_wc  *recv_wc;
 129	struct list_head   list;
 130	int		   length;
 131	struct ib_user_mad mad;
 132};
 133
 134struct ib_rmpp_mad_hdr {
 135	struct ib_mad_hdr	mad_hdr;
 136	struct ib_rmpp_hdr      rmpp_hdr;
 137} __packed;
 138
 139#define CREATE_TRACE_POINTS
 140#include <trace/events/ib_umad.h>
 141
 142static const dev_t base_umad_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
 143static const dev_t base_issm_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE) +
 144				   IB_UMAD_NUM_FIXED_MINOR;
 145static dev_t dynamic_umad_dev;
 146static dev_t dynamic_issm_dev;
 147
 148static DEFINE_IDA(umad_ida);
 149
 150static int ib_umad_add_one(struct ib_device *device);
 151static void ib_umad_remove_one(struct ib_device *device, void *client_data);
 152
 153static void ib_umad_dev_free(struct kref *kref)
 154{
 155	struct ib_umad_device *dev =
 156		container_of(kref, struct ib_umad_device, kref);
 157
 158	kfree(dev);
 159}
 160
 161static void ib_umad_dev_get(struct ib_umad_device *dev)
 162{
 163	kref_get(&dev->kref);
 164}
 165
 166static void ib_umad_dev_put(struct ib_umad_device *dev)
 167{
 168	kref_put(&dev->kref, ib_umad_dev_free);
 169}
 170
 171static int hdr_size(struct ib_umad_file *file)
 172{
 173	return file->use_pkey_index ? sizeof(struct ib_user_mad_hdr) :
 174				      sizeof(struct ib_user_mad_hdr_old);
 175}
 176
 177/* caller must hold file->mutex */
 178static struct ib_mad_agent *__get_agent(struct ib_umad_file *file, int id)
 179{
 180	return file->agents_dead ? NULL : file->agent[id];
 181}
 182
 183static int queue_packet(struct ib_umad_file *file,
 184			struct ib_mad_agent *agent,
 185			struct ib_umad_packet *packet)
 186{
 187	int ret = 1;
 188
 189	mutex_lock(&file->mutex);
 190
 191	for (packet->mad.hdr.id = 0;
 192	     packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
 193	     packet->mad.hdr.id++)
 194		if (agent == __get_agent(file, packet->mad.hdr.id)) {
 195			list_add_tail(&packet->list, &file->recv_list);
 196			wake_up_interruptible(&file->recv_wait);
 197			ret = 0;
 198			break;
 199		}
 200
 201	mutex_unlock(&file->mutex);
 202
 203	return ret;
 204}
 205
 206static void dequeue_send(struct ib_umad_file *file,
 207			 struct ib_umad_packet *packet)
 208{
 209	spin_lock_irq(&file->send_lock);
 210	list_del(&packet->list);
 211	spin_unlock_irq(&file->send_lock);
 212}
 213
 214static void send_handler(struct ib_mad_agent *agent,
 215			 struct ib_mad_send_wc *send_wc)
 216{
 217	struct ib_umad_file *file = agent->context;
 218	struct ib_umad_packet *packet = send_wc->send_buf->context[0];
 219
 220	dequeue_send(file, packet);
 221	rdma_destroy_ah(packet->msg->ah, RDMA_DESTROY_AH_SLEEPABLE);
 222	ib_free_send_mad(packet->msg);
 223
 224	if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
 225		packet->length = IB_MGMT_MAD_HDR;
 226		packet->mad.hdr.status = ETIMEDOUT;
 227		if (!queue_packet(file, agent, packet))
 228			return;
 229	}
 230	kfree(packet);
 231}
 232
 233static void recv_handler(struct ib_mad_agent *agent,
 234			 struct ib_mad_send_buf *send_buf,
 235			 struct ib_mad_recv_wc *mad_recv_wc)
 236{
 237	struct ib_umad_file *file = agent->context;
 238	struct ib_umad_packet *packet;
 239
 240	if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
 241		goto err1;
 242
 243	packet = kzalloc(sizeof *packet, GFP_KERNEL);
 244	if (!packet)
 245		goto err1;
 246
 247	packet->length = mad_recv_wc->mad_len;
 248	packet->recv_wc = mad_recv_wc;
 249
 250	packet->mad.hdr.status	   = 0;
 251	packet->mad.hdr.length	   = hdr_size(file) + mad_recv_wc->mad_len;
 252	packet->mad.hdr.qpn	   = cpu_to_be32(mad_recv_wc->wc->src_qp);
 253	/*
 254	 * On OPA devices it is okay to lose the upper 16 bits of LID as this
 255	 * information is obtained elsewhere. Mask off the upper 16 bits.
 256	 */
 257	if (rdma_cap_opa_mad(agent->device, agent->port_num))
 258		packet->mad.hdr.lid = ib_lid_be16(0xFFFF &
 259						  mad_recv_wc->wc->slid);
 260	else
 261		packet->mad.hdr.lid = ib_lid_be16(mad_recv_wc->wc->slid);
 262	packet->mad.hdr.sl	   = mad_recv_wc->wc->sl;
 263	packet->mad.hdr.path_bits  = mad_recv_wc->wc->dlid_path_bits;
 264	packet->mad.hdr.pkey_index = mad_recv_wc->wc->pkey_index;
 265	packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
 266	if (packet->mad.hdr.grh_present) {
 267		struct rdma_ah_attr ah_attr;
 268		const struct ib_global_route *grh;
 269		int ret;
 270
 271		ret = ib_init_ah_attr_from_wc(agent->device, agent->port_num,
 272					      mad_recv_wc->wc,
 273					      mad_recv_wc->recv_buf.grh,
 274					      &ah_attr);
 275		if (ret)
 276			goto err2;
 277
 278		grh = rdma_ah_read_grh(&ah_attr);
 279		packet->mad.hdr.gid_index = grh->sgid_index;
 280		packet->mad.hdr.hop_limit = grh->hop_limit;
 281		packet->mad.hdr.traffic_class = grh->traffic_class;
 282		memcpy(packet->mad.hdr.gid, &grh->dgid, 16);
 283		packet->mad.hdr.flow_label = cpu_to_be32(grh->flow_label);
 284		rdma_destroy_ah_attr(&ah_attr);
 285	}
 286
 287	if (queue_packet(file, agent, packet))
 288		goto err2;
 289	return;
 290
 291err2:
 292	kfree(packet);
 293err1:
 294	ib_free_recv_mad(mad_recv_wc);
 295}
 296
 297static ssize_t copy_recv_mad(struct ib_umad_file *file, char __user *buf,
 298			     struct ib_umad_packet *packet, size_t count)
 299{
 300	struct ib_mad_recv_buf *recv_buf;
 301	int left, seg_payload, offset, max_seg_payload;
 302	size_t seg_size;
 303
 304	recv_buf = &packet->recv_wc->recv_buf;
 305	seg_size = packet->recv_wc->mad_seg_size;
 306
 307	/* We need enough room to copy the first (or only) MAD segment. */
 308	if ((packet->length <= seg_size &&
 309	     count < hdr_size(file) + packet->length) ||
 310	    (packet->length > seg_size &&
 311	     count < hdr_size(file) + seg_size))
 312		return -EINVAL;
 313
 314	if (copy_to_user(buf, &packet->mad, hdr_size(file)))
 315		return -EFAULT;
 316
 317	buf += hdr_size(file);
 318	seg_payload = min_t(int, packet->length, seg_size);
 319	if (copy_to_user(buf, recv_buf->mad, seg_payload))
 320		return -EFAULT;
 321
 322	if (seg_payload < packet->length) {
 323		/*
 324		 * Multipacket RMPP MAD message. Copy remainder of message.
 325		 * Note that last segment may have a shorter payload.
 326		 */
 327		if (count < hdr_size(file) + packet->length) {
 328			/*
 329			 * The buffer is too small, return the first RMPP segment,
 330			 * which includes the RMPP message length.
 331			 */
 332			return -ENOSPC;
 333		}
 334		offset = ib_get_mad_data_offset(recv_buf->mad->mad_hdr.mgmt_class);
 335		max_seg_payload = seg_size - offset;
 336
 337		for (left = packet->length - seg_payload, buf += seg_payload;
 338		     left; left -= seg_payload, buf += seg_payload) {
 339			recv_buf = container_of(recv_buf->list.next,
 340						struct ib_mad_recv_buf, list);
 341			seg_payload = min(left, max_seg_payload);
 342			if (copy_to_user(buf, ((void *) recv_buf->mad) + offset,
 343					 seg_payload))
 344				return -EFAULT;
 345		}
 346	}
 347
 348	trace_ib_umad_read_recv(file, &packet->mad.hdr, &recv_buf->mad->mad_hdr);
 349
 350	return hdr_size(file) + packet->length;
 351}
 352
 353static ssize_t copy_send_mad(struct ib_umad_file *file, char __user *buf,
 354			     struct ib_umad_packet *packet, size_t count)
 355{
 356	ssize_t size = hdr_size(file) + packet->length;
 357
 358	if (count < size)
 359		return -EINVAL;
 360
 361	if (copy_to_user(buf, &packet->mad, hdr_size(file)))
 362		return -EFAULT;
 363
 364	buf += hdr_size(file);
 365
 366	if (copy_to_user(buf, packet->mad.data, packet->length))
 367		return -EFAULT;
 368
 369	trace_ib_umad_read_send(file, &packet->mad.hdr,
 370				(struct ib_mad_hdr *)&packet->mad.data);
 371
 372	return size;
 373}
 374
 375static ssize_t ib_umad_read(struct file *filp, char __user *buf,
 376			    size_t count, loff_t *pos)
 377{
 378	struct ib_umad_file *file = filp->private_data;
 379	struct ib_umad_packet *packet;
 380	ssize_t ret;
 381
 382	if (count < hdr_size(file))
 383		return -EINVAL;
 384
 385	mutex_lock(&file->mutex);
 386
 387	if (file->agents_dead) {
 388		mutex_unlock(&file->mutex);
 389		return -EIO;
 390	}
 391
 392	while (list_empty(&file->recv_list)) {
 393		mutex_unlock(&file->mutex);
 394
 395		if (filp->f_flags & O_NONBLOCK)
 396			return -EAGAIN;
 397
 398		if (wait_event_interruptible(file->recv_wait,
 399					     !list_empty(&file->recv_list)))
 400			return -ERESTARTSYS;
 401
 402		mutex_lock(&file->mutex);
 403	}
 404
 405	if (file->agents_dead) {
 406		mutex_unlock(&file->mutex);
 407		return -EIO;
 408	}
 409
 410	packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
 411	list_del(&packet->list);
 412
 413	mutex_unlock(&file->mutex);
 414
 415	if (packet->recv_wc)
 416		ret = copy_recv_mad(file, buf, packet, count);
 417	else
 418		ret = copy_send_mad(file, buf, packet, count);
 419
 420	if (ret < 0) {
 421		/* Requeue packet */
 422		mutex_lock(&file->mutex);
 423		list_add(&packet->list, &file->recv_list);
 424		mutex_unlock(&file->mutex);
 425	} else {
 426		if (packet->recv_wc)
 427			ib_free_recv_mad(packet->recv_wc);
 428		kfree(packet);
 429	}
 430	return ret;
 431}
 432
 433static int copy_rmpp_mad(struct ib_mad_send_buf *msg, const char __user *buf)
 434{
 435	int left, seg;
 436
 437	/* Copy class specific header */
 438	if ((msg->hdr_len > IB_MGMT_RMPP_HDR) &&
 439	    copy_from_user(msg->mad + IB_MGMT_RMPP_HDR, buf + IB_MGMT_RMPP_HDR,
 440			   msg->hdr_len - IB_MGMT_RMPP_HDR))
 441		return -EFAULT;
 442
 443	/* All headers are in place.  Copy data segments. */
 444	for (seg = 1, left = msg->data_len, buf += msg->hdr_len; left > 0;
 445	     seg++, left -= msg->seg_size, buf += msg->seg_size) {
 446		if (copy_from_user(ib_get_rmpp_segment(msg, seg), buf,
 447				   min(left, msg->seg_size)))
 448			return -EFAULT;
 449	}
 450	return 0;
 451}
 452
 453static int same_destination(struct ib_user_mad_hdr *hdr1,
 454			    struct ib_user_mad_hdr *hdr2)
 455{
 456	if (!hdr1->grh_present && !hdr2->grh_present)
 457	   return (hdr1->lid == hdr2->lid);
 458
 459	if (hdr1->grh_present && hdr2->grh_present)
 460	   return !memcmp(hdr1->gid, hdr2->gid, 16);
 461
 462	return 0;
 463}
 464
 465static int is_duplicate(struct ib_umad_file *file,
 466			struct ib_umad_packet *packet)
 467{
 468	struct ib_umad_packet *sent_packet;
 469	struct ib_mad_hdr *sent_hdr, *hdr;
 470
 471	hdr = (struct ib_mad_hdr *) packet->mad.data;
 472	list_for_each_entry(sent_packet, &file->send_list, list) {
 473		sent_hdr = (struct ib_mad_hdr *) sent_packet->mad.data;
 474
 475		if ((hdr->tid != sent_hdr->tid) ||
 476		    (hdr->mgmt_class != sent_hdr->mgmt_class))
 477			continue;
 478
 479		/*
 480		 * No need to be overly clever here.  If two new operations have
 481		 * the same TID, reject the second as a duplicate.  This is more
 482		 * restrictive than required by the spec.
 483		 */
 484		if (!ib_response_mad(hdr)) {
 485			if (!ib_response_mad(sent_hdr))
 486				return 1;
 487			continue;
 488		} else if (!ib_response_mad(sent_hdr))
 489			continue;
 490
 491		if (same_destination(&packet->mad.hdr, &sent_packet->mad.hdr))
 492			return 1;
 493	}
 494
 495	return 0;
 496}
 497
 498static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
 499			     size_t count, loff_t *pos)
 500{
 501	struct ib_umad_file *file = filp->private_data;
 502	struct ib_rmpp_mad_hdr *rmpp_mad_hdr;
 503	struct ib_umad_packet *packet;
 504	struct ib_mad_agent *agent;
 505	struct rdma_ah_attr ah_attr;
 506	struct ib_ah *ah;
 507	__be64 *tid;
 508	int ret, data_len, hdr_len, copy_offset, rmpp_active;
 509	u8 base_version;
 510
 511	if (count < hdr_size(file) + IB_MGMT_RMPP_HDR)
 512		return -EINVAL;
 513
 514	packet = kzalloc(sizeof(*packet) + IB_MGMT_RMPP_HDR, GFP_KERNEL);
 515	if (!packet)
 516		return -ENOMEM;
 517
 518	if (copy_from_user(&packet->mad, buf, hdr_size(file))) {
 519		ret = -EFAULT;
 520		goto err;
 521	}
 522
 523	if (packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
 524		ret = -EINVAL;
 525		goto err;
 526	}
 527
 528	buf += hdr_size(file);
 529
 530	if (copy_from_user(packet->mad.data, buf, IB_MGMT_RMPP_HDR)) {
 531		ret = -EFAULT;
 532		goto err;
 533	}
 534
 535	mutex_lock(&file->mutex);
 536
 537	trace_ib_umad_write(file, &packet->mad.hdr,
 538			    (struct ib_mad_hdr *)&packet->mad.data);
 539
 540	agent = __get_agent(file, packet->mad.hdr.id);
 541	if (!agent) {
 542		ret = -EIO;
 543		goto err_up;
 544	}
 545
 546	memset(&ah_attr, 0, sizeof ah_attr);
 547	ah_attr.type = rdma_ah_find_type(agent->device,
 548					 file->port->port_num);
 549	rdma_ah_set_dlid(&ah_attr, be16_to_cpu(packet->mad.hdr.lid));
 550	rdma_ah_set_sl(&ah_attr, packet->mad.hdr.sl);
 551	rdma_ah_set_path_bits(&ah_attr, packet->mad.hdr.path_bits);
 552	rdma_ah_set_port_num(&ah_attr, file->port->port_num);
 553	if (packet->mad.hdr.grh_present) {
 554		rdma_ah_set_grh(&ah_attr, NULL,
 555				be32_to_cpu(packet->mad.hdr.flow_label),
 556				packet->mad.hdr.gid_index,
 557				packet->mad.hdr.hop_limit,
 558				packet->mad.hdr.traffic_class);
 559		rdma_ah_set_dgid_raw(&ah_attr, packet->mad.hdr.gid);
 560	}
 561
 562	ah = rdma_create_user_ah(agent->qp->pd, &ah_attr, NULL);
 563	if (IS_ERR(ah)) {
 564		ret = PTR_ERR(ah);
 565		goto err_up;
 566	}
 567
 568	rmpp_mad_hdr = (struct ib_rmpp_mad_hdr *)packet->mad.data;
 569	hdr_len = ib_get_mad_data_offset(rmpp_mad_hdr->mad_hdr.mgmt_class);
 570
 571	if (ib_is_mad_class_rmpp(rmpp_mad_hdr->mad_hdr.mgmt_class)
 572	    && ib_mad_kernel_rmpp_agent(agent)) {
 573		copy_offset = IB_MGMT_RMPP_HDR;
 574		rmpp_active = ib_get_rmpp_flags(&rmpp_mad_hdr->rmpp_hdr) &
 575						IB_MGMT_RMPP_FLAG_ACTIVE;
 576	} else {
 577		copy_offset = IB_MGMT_MAD_HDR;
 578		rmpp_active = 0;
 579	}
 580
 581	base_version = ((struct ib_mad_hdr *)&packet->mad.data)->base_version;
 582	data_len = count - hdr_size(file) - hdr_len;
 583	packet->msg = ib_create_send_mad(agent,
 584					 be32_to_cpu(packet->mad.hdr.qpn),
 585					 packet->mad.hdr.pkey_index, rmpp_active,
 586					 hdr_len, data_len, GFP_KERNEL,
 587					 base_version);
 588	if (IS_ERR(packet->msg)) {
 589		ret = PTR_ERR(packet->msg);
 590		goto err_ah;
 591	}
 592
 593	packet->msg->ah		= ah;
 594	packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
 595	packet->msg->retries	= packet->mad.hdr.retries;
 596	packet->msg->context[0] = packet;
 597
 598	/* Copy MAD header.  Any RMPP header is already in place. */
 599	memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
 600
 601	if (!rmpp_active) {
 602		if (copy_from_user(packet->msg->mad + copy_offset,
 603				   buf + copy_offset,
 604				   hdr_len + data_len - copy_offset)) {
 605			ret = -EFAULT;
 606			goto err_msg;
 607		}
 608	} else {
 609		ret = copy_rmpp_mad(packet->msg, buf);
 610		if (ret)
 611			goto err_msg;
 612	}
 613
 614	/*
 615	 * Set the high-order part of the transaction ID to make MADs from
 616	 * different agents unique, and allow routing responses back to the
 617	 * original requestor.
 618	 */
 619	if (!ib_response_mad(packet->msg->mad)) {
 620		tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
 621		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
 622				   (be64_to_cpup(tid) & 0xffffffff));
 623		rmpp_mad_hdr->mad_hdr.tid = *tid;
 624	}
 625
 626	if (!ib_mad_kernel_rmpp_agent(agent)
 627	    && ib_is_mad_class_rmpp(rmpp_mad_hdr->mad_hdr.mgmt_class)
 628	    && (ib_get_rmpp_flags(&rmpp_mad_hdr->rmpp_hdr) & IB_MGMT_RMPP_FLAG_ACTIVE)) {
 629		spin_lock_irq(&file->send_lock);
 630		list_add_tail(&packet->list, &file->send_list);
 631		spin_unlock_irq(&file->send_lock);
 632	} else {
 633		spin_lock_irq(&file->send_lock);
 634		ret = is_duplicate(file, packet);
 635		if (!ret)
 636			list_add_tail(&packet->list, &file->send_list);
 637		spin_unlock_irq(&file->send_lock);
 638		if (ret) {
 639			ret = -EINVAL;
 640			goto err_msg;
 641		}
 642	}
 643
 644	ret = ib_post_send_mad(packet->msg, NULL);
 645	if (ret)
 646		goto err_send;
 647
 648	mutex_unlock(&file->mutex);
 649	return count;
 650
 651err_send:
 652	dequeue_send(file, packet);
 653err_msg:
 654	ib_free_send_mad(packet->msg);
 655err_ah:
 656	rdma_destroy_ah(ah, RDMA_DESTROY_AH_SLEEPABLE);
 657err_up:
 658	mutex_unlock(&file->mutex);
 659err:
 660	kfree(packet);
 661	return ret;
 662}
 663
 664static __poll_t ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
 665{
 666	struct ib_umad_file *file = filp->private_data;
 667
 668	/* we will always be able to post a MAD send */
 669	__poll_t mask = EPOLLOUT | EPOLLWRNORM;
 670
 671	mutex_lock(&file->mutex);
 672	poll_wait(filp, &file->recv_wait, wait);
 673
 674	if (!list_empty(&file->recv_list))
 675		mask |= EPOLLIN | EPOLLRDNORM;
 676	if (file->agents_dead)
 677		mask = EPOLLERR;
 678	mutex_unlock(&file->mutex);
 679
 680	return mask;
 681}
 682
 683static int ib_umad_reg_agent(struct ib_umad_file *file, void __user *arg,
 684			     int compat_method_mask)
 685{
 686	struct ib_user_mad_reg_req ureq;
 687	struct ib_mad_reg_req req;
 688	struct ib_mad_agent *agent = NULL;
 689	int agent_id;
 690	int ret;
 691
 692	mutex_lock(&file->port->file_mutex);
 693	mutex_lock(&file->mutex);
 694
 695	if (!file->port->ib_dev) {
 696		dev_notice(&file->port->dev, "%s: invalid device\n", __func__);
 697		ret = -EPIPE;
 698		goto out;
 699	}
 700
 701	if (copy_from_user(&ureq, arg, sizeof ureq)) {
 702		ret = -EFAULT;
 703		goto out;
 704	}
 705
 706	if (ureq.qpn != 0 && ureq.qpn != 1) {
 707		dev_notice(&file->port->dev,
 708			   "%s: invalid QPN %u specified\n", __func__,
 709			   ureq.qpn);
 710		ret = -EINVAL;
 711		goto out;
 712	}
 713
 714	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
 715		if (!__get_agent(file, agent_id))
 716			goto found;
 717
 718	dev_notice(&file->port->dev, "%s: Max Agents (%u) reached\n", __func__,
 719		   IB_UMAD_MAX_AGENTS);
 720
 721	ret = -ENOMEM;
 722	goto out;
 723
 724found:
 725	if (ureq.mgmt_class) {
 726		memset(&req, 0, sizeof(req));
 727		req.mgmt_class         = ureq.mgmt_class;
 728		req.mgmt_class_version = ureq.mgmt_class_version;
 729		memcpy(req.oui, ureq.oui, sizeof req.oui);
 730
 731		if (compat_method_mask) {
 732			u32 *umm = (u32 *) ureq.method_mask;
 733			int i;
 734
 735			for (i = 0; i < BITS_TO_LONGS(IB_MGMT_MAX_METHODS); ++i)
 736				req.method_mask[i] =
 737					umm[i * 2] | ((u64) umm[i * 2 + 1] << 32);
 738		} else
 739			memcpy(req.method_mask, ureq.method_mask,
 740			       sizeof req.method_mask);
 741	}
 742
 743	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
 744				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
 745				      ureq.mgmt_class ? &req : NULL,
 746				      ureq.rmpp_version,
 747				      send_handler, recv_handler, file, 0);
 748	if (IS_ERR(agent)) {
 749		ret = PTR_ERR(agent);
 750		agent = NULL;
 751		goto out;
 752	}
 753
 754	if (put_user(agent_id,
 755		     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
 756		ret = -EFAULT;
 757		goto out;
 758	}
 759
 760	if (!file->already_used) {
 761		file->already_used = 1;
 762		if (!file->use_pkey_index) {
 763			dev_warn(&file->port->dev,
 764				"process %s did not enable P_Key index support.\n",
 765				current->comm);
 766			dev_warn(&file->port->dev,
 767				"   Documentation/infiniband/user_mad.rst has info on the new ABI.\n");
 768		}
 769	}
 770
 771	file->agent[agent_id] = agent;
 772	ret = 0;
 773
 774out:
 775	mutex_unlock(&file->mutex);
 776
 777	if (ret && agent)
 778		ib_unregister_mad_agent(agent);
 779
 780	mutex_unlock(&file->port->file_mutex);
 781
 782	return ret;
 783}
 784
 785static int ib_umad_reg_agent2(struct ib_umad_file *file, void __user *arg)
 786{
 787	struct ib_user_mad_reg_req2 ureq;
 788	struct ib_mad_reg_req req;
 789	struct ib_mad_agent *agent = NULL;
 790	int agent_id;
 791	int ret;
 792
 793	mutex_lock(&file->port->file_mutex);
 794	mutex_lock(&file->mutex);
 795
 796	if (!file->port->ib_dev) {
 797		dev_notice(&file->port->dev, "%s: invalid device\n", __func__);
 798		ret = -EPIPE;
 799		goto out;
 800	}
 801
 802	if (copy_from_user(&ureq, arg, sizeof(ureq))) {
 803		ret = -EFAULT;
 804		goto out;
 805	}
 806
 807	if (ureq.qpn != 0 && ureq.qpn != 1) {
 808		dev_notice(&file->port->dev, "%s: invalid QPN %u specified\n",
 809			   __func__, ureq.qpn);
 810		ret = -EINVAL;
 811		goto out;
 812	}
 813
 814	if (ureq.flags & ~IB_USER_MAD_REG_FLAGS_CAP) {
 815		dev_notice(&file->port->dev,
 816			   "%s failed: invalid registration flags specified 0x%x; supported 0x%x\n",
 817			   __func__, ureq.flags, IB_USER_MAD_REG_FLAGS_CAP);
 818		ret = -EINVAL;
 819
 820		if (put_user((u32)IB_USER_MAD_REG_FLAGS_CAP,
 821				(u32 __user *) (arg + offsetof(struct
 822				ib_user_mad_reg_req2, flags))))
 823			ret = -EFAULT;
 824
 825		goto out;
 826	}
 827
 828	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
 829		if (!__get_agent(file, agent_id))
 830			goto found;
 831
 832	dev_notice(&file->port->dev, "%s: Max Agents (%u) reached\n", __func__,
 833		   IB_UMAD_MAX_AGENTS);
 834	ret = -ENOMEM;
 835	goto out;
 836
 837found:
 838	if (ureq.mgmt_class) {
 839		memset(&req, 0, sizeof(req));
 840		req.mgmt_class         = ureq.mgmt_class;
 841		req.mgmt_class_version = ureq.mgmt_class_version;
 842		if (ureq.oui & 0xff000000) {
 843			dev_notice(&file->port->dev,
 844				   "%s failed: oui invalid 0x%08x\n", __func__,
 845				   ureq.oui);
 846			ret = -EINVAL;
 847			goto out;
 848		}
 849		req.oui[2] =  ureq.oui & 0x0000ff;
 850		req.oui[1] = (ureq.oui & 0x00ff00) >> 8;
 851		req.oui[0] = (ureq.oui & 0xff0000) >> 16;
 852		memcpy(req.method_mask, ureq.method_mask,
 853			sizeof(req.method_mask));
 854	}
 855
 856	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
 857				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
 858				      ureq.mgmt_class ? &req : NULL,
 859				      ureq.rmpp_version,
 860				      send_handler, recv_handler, file,
 861				      ureq.flags);
 862	if (IS_ERR(agent)) {
 863		ret = PTR_ERR(agent);
 864		agent = NULL;
 865		goto out;
 866	}
 867
 868	if (put_user(agent_id,
 869		     (u32 __user *)(arg +
 870				offsetof(struct ib_user_mad_reg_req2, id)))) {
 871		ret = -EFAULT;
 872		goto out;
 873	}
 874
 875	if (!file->already_used) {
 876		file->already_used = 1;
 877		file->use_pkey_index = 1;
 878	}
 879
 880	file->agent[agent_id] = agent;
 881	ret = 0;
 882
 883out:
 884	mutex_unlock(&file->mutex);
 885
 886	if (ret && agent)
 887		ib_unregister_mad_agent(agent);
 888
 889	mutex_unlock(&file->port->file_mutex);
 890
 891	return ret;
 892}
 893
 894
 895static int ib_umad_unreg_agent(struct ib_umad_file *file, u32 __user *arg)
 896{
 897	struct ib_mad_agent *agent = NULL;
 898	u32 id;
 899	int ret = 0;
 900
 901	if (get_user(id, arg))
 902		return -EFAULT;
 903	if (id >= IB_UMAD_MAX_AGENTS)
 904		return -EINVAL;
 905
 906	mutex_lock(&file->port->file_mutex);
 907	mutex_lock(&file->mutex);
 908
 909	id = array_index_nospec(id, IB_UMAD_MAX_AGENTS);
 910	if (!__get_agent(file, id)) {
 911		ret = -EINVAL;
 912		goto out;
 913	}
 914
 915	agent = file->agent[id];
 916	file->agent[id] = NULL;
 917
 918out:
 919	mutex_unlock(&file->mutex);
 920
 921	if (agent)
 922		ib_unregister_mad_agent(agent);
 923
 924	mutex_unlock(&file->port->file_mutex);
 925
 926	return ret;
 927}
 928
 929static long ib_umad_enable_pkey(struct ib_umad_file *file)
 930{
 931	int ret = 0;
 932
 933	mutex_lock(&file->mutex);
 934	if (file->already_used)
 935		ret = -EINVAL;
 936	else
 937		file->use_pkey_index = 1;
 938	mutex_unlock(&file->mutex);
 939
 940	return ret;
 941}
 942
 943static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
 944			  unsigned long arg)
 945{
 946	switch (cmd) {
 947	case IB_USER_MAD_REGISTER_AGENT:
 948		return ib_umad_reg_agent(filp->private_data, (void __user *) arg, 0);
 949	case IB_USER_MAD_UNREGISTER_AGENT:
 950		return ib_umad_unreg_agent(filp->private_data, (__u32 __user *) arg);
 951	case IB_USER_MAD_ENABLE_PKEY:
 952		return ib_umad_enable_pkey(filp->private_data);
 953	case IB_USER_MAD_REGISTER_AGENT2:
 954		return ib_umad_reg_agent2(filp->private_data, (void __user *) arg);
 955	default:
 956		return -ENOIOCTLCMD;
 957	}
 958}
 959
 960#ifdef CONFIG_COMPAT
 961static long ib_umad_compat_ioctl(struct file *filp, unsigned int cmd,
 962				 unsigned long arg)
 963{
 964	switch (cmd) {
 965	case IB_USER_MAD_REGISTER_AGENT:
 966		return ib_umad_reg_agent(filp->private_data, compat_ptr(arg), 1);
 967	case IB_USER_MAD_UNREGISTER_AGENT:
 968		return ib_umad_unreg_agent(filp->private_data, compat_ptr(arg));
 969	case IB_USER_MAD_ENABLE_PKEY:
 970		return ib_umad_enable_pkey(filp->private_data);
 971	case IB_USER_MAD_REGISTER_AGENT2:
 972		return ib_umad_reg_agent2(filp->private_data, compat_ptr(arg));
 973	default:
 974		return -ENOIOCTLCMD;
 975	}
 976}
 977#endif
 978
 979/*
 980 * ib_umad_open() does not need the BKL:
 981 *
 982 *  - the ib_umad_port structures are properly reference counted, and
 983 *    everything else is purely local to the file being created, so
 984 *    races against other open calls are not a problem;
 985 *  - the ioctl method does not affect any global state outside of the
 986 *    file structure being operated on;
 987 */
 988static int ib_umad_open(struct inode *inode, struct file *filp)
 989{
 990	struct ib_umad_port *port;
 991	struct ib_umad_file *file;
 992	int ret = 0;
 993
 994	port = container_of(inode->i_cdev, struct ib_umad_port, cdev);
 995
 996	mutex_lock(&port->file_mutex);
 997
 998	if (!port->ib_dev) {
 999		ret = -ENXIO;
1000		goto out;
1001	}
1002
1003	if (!rdma_dev_access_netns(port->ib_dev, current->nsproxy->net_ns)) {
1004		ret = -EPERM;
1005		goto out;
1006	}
1007
1008	file = kzalloc(sizeof(*file), GFP_KERNEL);
1009	if (!file) {
1010		ret = -ENOMEM;
1011		goto out;
1012	}
1013
1014	mutex_init(&file->mutex);
1015	spin_lock_init(&file->send_lock);
1016	INIT_LIST_HEAD(&file->recv_list);
1017	INIT_LIST_HEAD(&file->send_list);
1018	init_waitqueue_head(&file->recv_wait);
1019
1020	file->port = port;
1021	filp->private_data = file;
1022
1023	list_add_tail(&file->port_list, &port->file_list);
1024
1025	stream_open(inode, filp);
1026out:
1027	mutex_unlock(&port->file_mutex);
1028	return ret;
1029}
1030
1031static int ib_umad_close(struct inode *inode, struct file *filp)
1032{
1033	struct ib_umad_file *file = filp->private_data;
1034	struct ib_umad_packet *packet, *tmp;
1035	int already_dead;
1036	int i;
1037
1038	mutex_lock(&file->port->file_mutex);
1039	mutex_lock(&file->mutex);
1040
1041	already_dead = file->agents_dead;
1042	file->agents_dead = 1;
1043
1044	list_for_each_entry_safe(packet, tmp, &file->recv_list, list) {
1045		if (packet->recv_wc)
1046			ib_free_recv_mad(packet->recv_wc);
1047		kfree(packet);
1048	}
1049
1050	list_del(&file->port_list);
1051
1052	mutex_unlock(&file->mutex);
1053
1054	if (!already_dead)
1055		for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
1056			if (file->agent[i])
1057				ib_unregister_mad_agent(file->agent[i]);
1058
1059	mutex_unlock(&file->port->file_mutex);
1060	mutex_destroy(&file->mutex);
1061	kfree(file);
1062	return 0;
1063}
1064
1065static const struct file_operations umad_fops = {
1066	.owner		= THIS_MODULE,
1067	.read		= ib_umad_read,
1068	.write		= ib_umad_write,
1069	.poll		= ib_umad_poll,
1070	.unlocked_ioctl = ib_umad_ioctl,
1071#ifdef CONFIG_COMPAT
1072	.compat_ioctl	= ib_umad_compat_ioctl,
1073#endif
1074	.open		= ib_umad_open,
1075	.release	= ib_umad_close,
1076	.llseek		= no_llseek,
1077};
1078
1079static int ib_umad_sm_open(struct inode *inode, struct file *filp)
1080{
1081	struct ib_umad_port *port;
1082	struct ib_port_modify props = {
1083		.set_port_cap_mask = IB_PORT_SM
1084	};
1085	int ret;
1086
1087	port = container_of(inode->i_cdev, struct ib_umad_port, sm_cdev);
1088
1089	if (filp->f_flags & O_NONBLOCK) {
1090		if (down_trylock(&port->sm_sem)) {
1091			ret = -EAGAIN;
1092			goto fail;
1093		}
1094	} else {
1095		if (down_interruptible(&port->sm_sem)) {
1096			ret = -ERESTARTSYS;
1097			goto fail;
1098		}
1099	}
1100
1101	if (!rdma_dev_access_netns(port->ib_dev, current->nsproxy->net_ns)) {
1102		ret = -EPERM;
1103		goto err_up_sem;
1104	}
1105
1106	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
1107	if (ret)
1108		goto err_up_sem;
1109
1110	filp->private_data = port;
1111
1112	nonseekable_open(inode, filp);
1113	return 0;
1114
1115err_up_sem:
1116	up(&port->sm_sem);
1117
1118fail:
1119	return ret;
1120}
1121
1122static int ib_umad_sm_close(struct inode *inode, struct file *filp)
1123{
1124	struct ib_umad_port *port = filp->private_data;
1125	struct ib_port_modify props = {
1126		.clr_port_cap_mask = IB_PORT_SM
1127	};
1128	int ret = 0;
1129
1130	mutex_lock(&port->file_mutex);
1131	if (port->ib_dev)
1132		ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
1133	mutex_unlock(&port->file_mutex);
1134
1135	up(&port->sm_sem);
1136
1137	return ret;
1138}
1139
1140static const struct file_operations umad_sm_fops = {
1141	.owner	 = THIS_MODULE,
1142	.open	 = ib_umad_sm_open,
1143	.release = ib_umad_sm_close,
1144	.llseek	 = no_llseek,
1145};
1146
1147static struct ib_umad_port *get_port(struct ib_device *ibdev,
1148				     struct ib_umad_device *umad_dev,
1149				     u32 port)
1150{
1151	if (!umad_dev)
1152		return ERR_PTR(-EOPNOTSUPP);
1153	if (!rdma_is_port_valid(ibdev, port))
1154		return ERR_PTR(-EINVAL);
1155	if (!rdma_cap_ib_mad(ibdev, port))
1156		return ERR_PTR(-EOPNOTSUPP);
1157
1158	return &umad_dev->ports[port - rdma_start_port(ibdev)];
1159}
1160
1161static int ib_umad_get_nl_info(struct ib_device *ibdev, void *client_data,
1162			       struct ib_client_nl_info *res)
1163{
1164	struct ib_umad_port *port = get_port(ibdev, client_data, res->port);
1165
1166	if (IS_ERR(port))
1167		return PTR_ERR(port);
1168
1169	res->abi = IB_USER_MAD_ABI_VERSION;
1170	res->cdev = &port->dev;
1171	return 0;
1172}
1173
1174static struct ib_client umad_client = {
1175	.name   = "umad",
1176	.add    = ib_umad_add_one,
1177	.remove = ib_umad_remove_one,
1178	.get_nl_info = ib_umad_get_nl_info,
1179};
1180MODULE_ALIAS_RDMA_CLIENT("umad");
1181
1182static int ib_issm_get_nl_info(struct ib_device *ibdev, void *client_data,
1183			       struct ib_client_nl_info *res)
1184{
1185	struct ib_umad_port *port = get_port(ibdev, client_data, res->port);
1186
1187	if (IS_ERR(port))
1188		return PTR_ERR(port);
1189
1190	res->abi = IB_USER_MAD_ABI_VERSION;
1191	res->cdev = &port->sm_dev;
1192	return 0;
1193}
1194
1195static struct ib_client issm_client = {
1196	.name = "issm",
1197	.get_nl_info = ib_issm_get_nl_info,
1198};
1199MODULE_ALIAS_RDMA_CLIENT("issm");
1200
1201static ssize_t ibdev_show(struct device *dev, struct device_attribute *attr,
1202			  char *buf)
1203{
1204	struct ib_umad_port *port = dev_get_drvdata(dev);
1205
1206	if (!port)
1207		return -ENODEV;
1208
1209	return sysfs_emit(buf, "%s\n", dev_name(&port->ib_dev->dev));
1210}
1211static DEVICE_ATTR_RO(ibdev);
1212
1213static ssize_t port_show(struct device *dev, struct device_attribute *attr,
1214			 char *buf)
1215{
1216	struct ib_umad_port *port = dev_get_drvdata(dev);
1217
1218	if (!port)
1219		return -ENODEV;
1220
1221	return sysfs_emit(buf, "%d\n", port->port_num);
1222}
1223static DEVICE_ATTR_RO(port);
1224
1225static struct attribute *umad_class_dev_attrs[] = {
1226	&dev_attr_ibdev.attr,
1227	&dev_attr_port.attr,
1228	NULL,
1229};
1230ATTRIBUTE_GROUPS(umad_class_dev);
1231
1232static char *umad_devnode(const struct device *dev, umode_t *mode)
1233{
1234	return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1235}
1236
1237static ssize_t abi_version_show(const struct class *class,
1238				const struct class_attribute *attr, char *buf)
1239{
1240	return sysfs_emit(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
1241}
1242static CLASS_ATTR_RO(abi_version);
1243
1244static struct attribute *umad_class_attrs[] = {
1245	&class_attr_abi_version.attr,
1246	NULL,
1247};
1248ATTRIBUTE_GROUPS(umad_class);
1249
1250static struct class umad_class = {
1251	.name		= "infiniband_mad",
1252	.devnode	= umad_devnode,
1253	.class_groups	= umad_class_groups,
1254	.dev_groups	= umad_class_dev_groups,
1255};
1256
1257static void ib_umad_release_port(struct device *device)
1258{
1259	struct ib_umad_port *port = dev_get_drvdata(device);
1260	struct ib_umad_device *umad_dev = port->umad_dev;
1261
1262	ib_umad_dev_put(umad_dev);
1263}
1264
1265static void ib_umad_init_port_dev(struct device *dev,
1266				  struct ib_umad_port *port,
1267				  const struct ib_device *device)
1268{
1269	device_initialize(dev);
1270	ib_umad_dev_get(port->umad_dev);
1271	dev->class = &umad_class;
1272	dev->parent = device->dev.parent;
1273	dev_set_drvdata(dev, port);
1274	dev->release = ib_umad_release_port;
1275}
1276
1277static int ib_umad_init_port(struct ib_device *device, int port_num,
1278			     struct ib_umad_device *umad_dev,
1279			     struct ib_umad_port *port)
1280{
1281	int devnum;
1282	dev_t base_umad;
1283	dev_t base_issm;
1284	int ret;
1285
1286	devnum = ida_alloc_max(&umad_ida, IB_UMAD_MAX_PORTS - 1, GFP_KERNEL);
1287	if (devnum < 0)
1288		return -1;
1289	port->dev_num = devnum;
1290	if (devnum >= IB_UMAD_NUM_FIXED_MINOR) {
1291		base_umad = dynamic_umad_dev + devnum - IB_UMAD_NUM_FIXED_MINOR;
1292		base_issm = dynamic_issm_dev + devnum - IB_UMAD_NUM_FIXED_MINOR;
1293	} else {
1294		base_umad = devnum + base_umad_dev;
1295		base_issm = devnum + base_issm_dev;
1296	}
1297
1298	port->ib_dev   = device;
1299	port->umad_dev = umad_dev;
1300	port->port_num = port_num;
1301	sema_init(&port->sm_sem, 1);
1302	mutex_init(&port->file_mutex);
1303	INIT_LIST_HEAD(&port->file_list);
1304
1305	ib_umad_init_port_dev(&port->dev, port, device);
1306	port->dev.devt = base_umad;
1307	dev_set_name(&port->dev, "umad%d", port->dev_num);
1308	cdev_init(&port->cdev, &umad_fops);
1309	port->cdev.owner = THIS_MODULE;
1310
1311	ret = cdev_device_add(&port->cdev, &port->dev);
1312	if (ret)
1313		goto err_cdev;
1314
1315	ib_umad_init_port_dev(&port->sm_dev, port, device);
1316	port->sm_dev.devt = base_issm;
1317	dev_set_name(&port->sm_dev, "issm%d", port->dev_num);
1318	cdev_init(&port->sm_cdev, &umad_sm_fops);
1319	port->sm_cdev.owner = THIS_MODULE;
1320
1321	ret = cdev_device_add(&port->sm_cdev, &port->sm_dev);
1322	if (ret)
1323		goto err_dev;
1324
1325	return 0;
1326
1327err_dev:
1328	put_device(&port->sm_dev);
1329	cdev_device_del(&port->cdev, &port->dev);
1330err_cdev:
1331	put_device(&port->dev);
1332	ida_free(&umad_ida, devnum);
1333	return ret;
1334}
1335
1336static void ib_umad_kill_port(struct ib_umad_port *port)
1337{
1338	struct ib_umad_file *file;
1339	int id;
1340
1341	cdev_device_del(&port->sm_cdev, &port->sm_dev);
1342	cdev_device_del(&port->cdev, &port->dev);
1343
1344	mutex_lock(&port->file_mutex);
1345
1346	/* Mark ib_dev NULL and block ioctl or other file ops to progress
1347	 * further.
1348	 */
1349	port->ib_dev = NULL;
1350
1351	list_for_each_entry(file, &port->file_list, port_list) {
1352		mutex_lock(&file->mutex);
1353		file->agents_dead = 1;
1354		wake_up_interruptible(&file->recv_wait);
1355		mutex_unlock(&file->mutex);
1356
1357		for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id)
1358			if (file->agent[id])
1359				ib_unregister_mad_agent(file->agent[id]);
1360	}
1361
1362	mutex_unlock(&port->file_mutex);
1363
1364	ida_free(&umad_ida, port->dev_num);
1365
1366	/* balances device_initialize() */
1367	put_device(&port->sm_dev);
1368	put_device(&port->dev);
1369}
1370
1371static int ib_umad_add_one(struct ib_device *device)
1372{
1373	struct ib_umad_device *umad_dev;
1374	int s, e, i;
1375	int count = 0;
1376	int ret;
1377
1378	s = rdma_start_port(device);
1379	e = rdma_end_port(device);
1380
1381	umad_dev = kzalloc(struct_size(umad_dev, ports,
1382				       size_add(size_sub(e, s), 1)),
1383			   GFP_KERNEL);
1384	if (!umad_dev)
1385		return -ENOMEM;
1386
1387	kref_init(&umad_dev->kref);
1388	for (i = s; i <= e; ++i) {
1389		if (!rdma_cap_ib_mad(device, i))
1390			continue;
1391
1392		ret = ib_umad_init_port(device, i, umad_dev,
1393					&umad_dev->ports[i - s]);
1394		if (ret)
1395			goto err;
1396
1397		count++;
1398	}
1399
1400	if (!count) {
1401		ret = -EOPNOTSUPP;
1402		goto free;
1403	}
1404
1405	ib_set_client_data(device, &umad_client, umad_dev);
1406
1407	return 0;
1408
1409err:
1410	while (--i >= s) {
1411		if (!rdma_cap_ib_mad(device, i))
1412			continue;
1413
1414		ib_umad_kill_port(&umad_dev->ports[i - s]);
1415	}
1416free:
1417	/* balances kref_init */
1418	ib_umad_dev_put(umad_dev);
1419	return ret;
1420}
1421
1422static void ib_umad_remove_one(struct ib_device *device, void *client_data)
1423{
1424	struct ib_umad_device *umad_dev = client_data;
1425	unsigned int i;
1426
1427	rdma_for_each_port (device, i) {
1428		if (rdma_cap_ib_mad(device, i))
1429			ib_umad_kill_port(
1430				&umad_dev->ports[i - rdma_start_port(device)]);
1431	}
1432	/* balances kref_init() */
1433	ib_umad_dev_put(umad_dev);
1434}
1435
1436static int __init ib_umad_init(void)
1437{
1438	int ret;
1439
1440	ret = register_chrdev_region(base_umad_dev,
1441				     IB_UMAD_NUM_FIXED_MINOR * 2,
1442				     umad_class.name);
1443	if (ret) {
1444		pr_err("couldn't register device number\n");
1445		goto out;
1446	}
1447
1448	ret = alloc_chrdev_region(&dynamic_umad_dev, 0,
1449				  IB_UMAD_NUM_DYNAMIC_MINOR * 2,
1450				  umad_class.name);
1451	if (ret) {
1452		pr_err("couldn't register dynamic device number\n");
1453		goto out_alloc;
1454	}
1455	dynamic_issm_dev = dynamic_umad_dev + IB_UMAD_NUM_DYNAMIC_MINOR;
1456
1457	ret = class_register(&umad_class);
1458	if (ret) {
1459		pr_err("couldn't create class infiniband_mad\n");
1460		goto out_chrdev;
1461	}
1462
1463	ret = ib_register_client(&umad_client);
1464	if (ret)
1465		goto out_class;
1466
1467	ret = ib_register_client(&issm_client);
1468	if (ret)
1469		goto out_client;
1470
1471	return 0;
1472
1473out_client:
1474	ib_unregister_client(&umad_client);
1475out_class:
1476	class_unregister(&umad_class);
1477
1478out_chrdev:
1479	unregister_chrdev_region(dynamic_umad_dev,
1480				 IB_UMAD_NUM_DYNAMIC_MINOR * 2);
1481
1482out_alloc:
1483	unregister_chrdev_region(base_umad_dev,
1484				 IB_UMAD_NUM_FIXED_MINOR * 2);
1485
1486out:
1487	return ret;
1488}
1489
1490static void __exit ib_umad_cleanup(void)
1491{
1492	ib_unregister_client(&issm_client);
1493	ib_unregister_client(&umad_client);
1494	class_unregister(&umad_class);
1495	unregister_chrdev_region(base_umad_dev,
1496				 IB_UMAD_NUM_FIXED_MINOR * 2);
1497	unregister_chrdev_region(dynamic_umad_dev,
1498				 IB_UMAD_NUM_DYNAMIC_MINOR * 2);
1499}
1500
1501module_init(ib_umad_init);
1502module_exit(ib_umad_cleanup);