user_mad.c

   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
  66#define MAX_UMAD_RECV_LIST_SIZE 200000
  67
  68enum {
  69	IB_UMAD_MAX_PORTS  = RDMA_MAX_PORTS,
  70	IB_UMAD_MAX_AGENTS = 32,
  71
  72	IB_UMAD_MAJOR      = 231,
  73	IB_UMAD_MINOR_BASE = 0,
  74	IB_UMAD_NUM_FIXED_MINOR = 64,
  75	IB_UMAD_NUM_DYNAMIC_MINOR = IB_UMAD_MAX_PORTS - IB_UMAD_NUM_FIXED_MINOR,
  76	IB_ISSM_MINOR_BASE        = IB_UMAD_NUM_FIXED_MINOR,
  77};
  78
  79/*
  80 * Our lifetime rules for these structs are the following:
  81 * device special file is opened, we take a reference on the
  82 * ib_umad_port's struct ib_umad_device. We drop these
  83 * references in the corresponding close().
  84 *
  85 * In addition to references coming from open character devices, there
  86 * is one more reference to each ib_umad_device representing the
  87 * module's reference taken when allocating the ib_umad_device in
  88 * ib_umad_add_one().
  89 *
  90 * When destroying an ib_umad_device, we drop the module's reference.
  91 */
  92
  93struct ib_umad_port {
  94	struct cdev           cdev;
  95	struct device	      dev;
  96	struct cdev           sm_cdev;
  97	struct device	      sm_dev;
  98	struct semaphore       sm_sem;
  99
 100	struct mutex	       file_mutex;
 101	struct list_head       file_list;
 102
 103	struct ib_device      *ib_dev;
 104	struct ib_umad_device *umad_dev;
 105	int                    dev_num;
 106	u32                     port_num;
 107};
 108
 109struct ib_umad_device {
 110	struct kref kref;
 111	struct ib_umad_port ports[];
 112};
 113
 114struct ib_umad_file {
 115	struct mutex		mutex;
 116	struct ib_umad_port    *port;
 117	struct list_head	recv_list;
 118	atomic_t		recv_list_size;
 119	struct list_head	send_list;
 120	struct list_head	port_list;
 121	spinlock_t		send_lock;
 122	wait_queue_head_t	recv_wait;
 123	struct ib_mad_agent    *agent[IB_UMAD_MAX_AGENTS];
 124	int			agents_dead;
 125	u8			use_pkey_index;
 126	u8			already_used;
 127};
 128
 129struct ib_umad_packet {
 130	struct ib_mad_send_buf *msg;
 131	struct ib_mad_recv_wc  *recv_wc;
 132	struct list_head   list;
 133	int		   length;
 134	struct ib_user_mad mad;
 135};
 136
 137struct ib_rmpp_mad_hdr {
 138	struct ib_mad_hdr	mad_hdr;
 139	struct ib_rmpp_hdr      rmpp_hdr;
 140} __packed;
 141
 142#define CREATE_TRACE_POINTS
 143#include <trace/events/ib_umad.h>
 144
 145static const dev_t base_umad_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
 146static const dev_t base_issm_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE) +
 147				   IB_UMAD_NUM_FIXED_MINOR;
 148static dev_t dynamic_umad_dev;
 149static dev_t dynamic_issm_dev;
 150
 151static DEFINE_IDA(umad_ida);
 152
 153static int ib_umad_add_one(struct ib_device *device);
 154static void ib_umad_remove_one(struct ib_device *device, void *client_data);
 155
 156static void ib_umad_dev_free(struct kref *kref)
 157{
 158	struct ib_umad_device *dev =
 159		container_of(kref, struct ib_umad_device, kref);
 160
 161	kfree(dev);
 162}
 163
 164static void ib_umad_dev_get(struct ib_umad_device *dev)
 165{
 166	kref_get(&dev->kref);
 167}
 168
 169static void ib_umad_dev_put(struct ib_umad_device *dev)
 170{
 171	kref_put(&dev->kref, ib_umad_dev_free);
 172}
 173
 174static int hdr_size(struct ib_umad_file *file)
 175{
 176	return file->use_pkey_index ? sizeof(struct ib_user_mad_hdr) :
 177				      sizeof(struct ib_user_mad_hdr_old);
 178}
 179
 180/* caller must hold file->mutex */
 181static struct ib_mad_agent *__get_agent(struct ib_umad_file *file, int id)
 182{
 183	return file->agents_dead ? NULL : file->agent[id];
 184}
 185
 186static int queue_packet(struct ib_umad_file *file, struct ib_mad_agent *agent,
 187			struct ib_umad_packet *packet, bool is_recv_mad)
 188{
 189	int ret = 1;
 190
 191	mutex_lock(&file->mutex);
 192
 193	if (is_recv_mad &&
 194	    atomic_read(&file->recv_list_size) > MAX_UMAD_RECV_LIST_SIZE)
 195		goto unlock;
 196
 197	for (packet->mad.hdr.id = 0;
 198	     packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
 199	     packet->mad.hdr.id++)
 200		if (agent == __get_agent(file, packet->mad.hdr.id)) {
 201			list_add_tail(&packet->list, &file->recv_list);
 202			atomic_inc(&file->recv_list_size);
 203			wake_up_interruptible(&file->recv_wait);
 204			ret = 0;
 205			break;
 206		}
 207unlock:
 208	mutex_unlock(&file->mutex);
 209
 210	return ret;
 211}
 212
 213static void dequeue_send(struct ib_umad_file *file,
 214			 struct ib_umad_packet *packet)
 215{
 216	spin_lock_irq(&file->send_lock);
 217	list_del(&packet->list);
 218	spin_unlock_irq(&file->send_lock);
 219}
 220
 221static void send_handler(struct ib_mad_agent *agent,
 222			 struct ib_mad_send_wc *send_wc)
 223{
 224	struct ib_umad_file *file = agent->context;
 225	struct ib_umad_packet *packet = send_wc->send_buf->context[0];
 226
 227	dequeue_send(file, packet);
 228	rdma_destroy_ah(packet->msg->ah, RDMA_DESTROY_AH_SLEEPABLE);
 229	ib_free_send_mad(packet->msg);
 230
 231	if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
 232		packet->length = IB_MGMT_MAD_HDR;
 233		packet->mad.hdr.status = ETIMEDOUT;
 234		if (!queue_packet(file, agent, packet, false))
 235			return;
 236	}
 237	kfree(packet);
 238}
 239
 240static void recv_handler(struct ib_mad_agent *agent,
 241			 struct ib_mad_send_buf *send_buf,
 242			 struct ib_mad_recv_wc *mad_recv_wc)
 243{
 244	struct ib_umad_file *file = agent->context;
 245	struct ib_umad_packet *packet;
 246
 247	if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
 248		goto err1;
 249
 250	packet = kzalloc(sizeof *packet, GFP_KERNEL);
 251	if (!packet)
 252		goto err1;
 253
 254	packet->length = mad_recv_wc->mad_len;
 255	packet->recv_wc = mad_recv_wc;
 256
 257	packet->mad.hdr.status	   = 0;
 258	packet->mad.hdr.length	   = hdr_size(file) + mad_recv_wc->mad_len;
 259	packet->mad.hdr.qpn	   = cpu_to_be32(mad_recv_wc->wc->src_qp);
 260	/*
 261	 * On OPA devices it is okay to lose the upper 16 bits of LID as this
 262	 * information is obtained elsewhere. Mask off the upper 16 bits.
 263	 */
 264	if (rdma_cap_opa_mad(agent->device, agent->port_num))
 265		packet->mad.hdr.lid = ib_lid_be16(0xFFFF &
 266						  mad_recv_wc->wc->slid);
 267	else
 268		packet->mad.hdr.lid = ib_lid_be16(mad_recv_wc->wc->slid);
 269	packet->mad.hdr.sl	   = mad_recv_wc->wc->sl;
 270	packet->mad.hdr.path_bits  = mad_recv_wc->wc->dlid_path_bits;
 271	packet->mad.hdr.pkey_index = mad_recv_wc->wc->pkey_index;
 272	packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
 273	if (packet->mad.hdr.grh_present) {
 274		struct rdma_ah_attr ah_attr;
 275		const struct ib_global_route *grh;
 276		int ret;
 277
 278		ret = ib_init_ah_attr_from_wc(agent->device, agent->port_num,
 279					      mad_recv_wc->wc,
 280					      mad_recv_wc->recv_buf.grh,
 281					      &ah_attr);
 282		if (ret)
 283			goto err2;
 284
 285		grh = rdma_ah_read_grh(&ah_attr);
 286		packet->mad.hdr.gid_index = grh->sgid_index;
 287		packet->mad.hdr.hop_limit = grh->hop_limit;
 288		packet->mad.hdr.traffic_class = grh->traffic_class;
 289		memcpy(packet->mad.hdr.gid, &grh->dgid, 16);
 290		packet->mad.hdr.flow_label = cpu_to_be32(grh->flow_label);
 291		rdma_destroy_ah_attr(&ah_attr);
 292	}
 293
 294	if (queue_packet(file, agent, packet, true))
 295		goto err2;
 296	return;
 297
 298err2:
 299	kfree(packet);
 300err1:
 301	ib_free_recv_mad(mad_recv_wc);
 302}
 303
 304static ssize_t copy_recv_mad(struct ib_umad_file *file, char __user *buf,
 305			     struct ib_umad_packet *packet, size_t count)
 306{
 307	struct ib_mad_recv_buf *recv_buf;
 308	int left, seg_payload, offset, max_seg_payload;
 309	size_t seg_size;
 310
 311	recv_buf = &packet->recv_wc->recv_buf;
 312	seg_size = packet->recv_wc->mad_seg_size;
 313
 314	/* We need enough room to copy the first (or only) MAD segment. */
 315	if ((packet->length <= seg_size &&
 316	     count < hdr_size(file) + packet->length) ||
 317	    (packet->length > seg_size &&
 318	     count < hdr_size(file) + seg_size))
 319		return -EINVAL;
 320
 321	if (copy_to_user(buf, &packet->mad, hdr_size(file)))
 322		return -EFAULT;
 323
 324	buf += hdr_size(file);
 325	seg_payload = min_t(int, packet->length, seg_size);
 326	if (copy_to_user(buf, recv_buf->mad, seg_payload))
 327		return -EFAULT;
 328
 329	if (seg_payload < packet->length) {
 330		/*
 331		 * Multipacket RMPP MAD message. Copy remainder of message.
 332		 * Note that last segment may have a shorter payload.
 333		 */
 334		if (count < hdr_size(file) + packet->length) {
 335			/*
 336			 * The buffer is too small, return the first RMPP segment,
 337			 * which includes the RMPP message length.
 338			 */
 339			return -ENOSPC;
 340		}
 341		offset = ib_get_mad_data_offset(recv_buf->mad->mad_hdr.mgmt_class);
 342		max_seg_payload = seg_size - offset;
 343
 344		for (left = packet->length - seg_payload, buf += seg_payload;
 345		     left; left -= seg_payload, buf += seg_payload) {
 346			recv_buf = container_of(recv_buf->list.next,
 347						struct ib_mad_recv_buf, list);
 348			seg_payload = min(left, max_seg_payload);
 349			if (copy_to_user(buf, ((void *) recv_buf->mad) + offset,
 350					 seg_payload))
 351				return -EFAULT;
 352		}
 353	}
 354
 355	trace_ib_umad_read_recv(file, &packet->mad.hdr, &recv_buf->mad->mad_hdr);
 356
 357	return hdr_size(file) + packet->length;
 358}
 359
 360static ssize_t copy_send_mad(struct ib_umad_file *file, char __user *buf,
 361			     struct ib_umad_packet *packet, size_t count)
 362{
 363	ssize_t size = hdr_size(file) + packet->length;
 364
 365	if (count < size)
 366		return -EINVAL;
 367
 368	if (copy_to_user(buf, &packet->mad, hdr_size(file)))
 369		return -EFAULT;
 370
 371	buf += hdr_size(file);
 372
 373	if (copy_to_user(buf, packet->mad.data, packet->length))
 374		return -EFAULT;
 375
 376	trace_ib_umad_read_send(file, &packet->mad.hdr,
 377				(struct ib_mad_hdr *)&packet->mad.data);
 378
 379	return size;
 380}
 381
 382static ssize_t ib_umad_read(struct file *filp, char __user *buf,
 383			    size_t count, loff_t *pos)
 384{
 385	struct ib_umad_file *file = filp->private_data;
 386	struct ib_umad_packet *packet;
 387	ssize_t ret;
 388
 389	if (count < hdr_size(file))
 390		return -EINVAL;
 391
 392	mutex_lock(&file->mutex);
 393
 394	if (file->agents_dead) {
 395		mutex_unlock(&file->mutex);
 396		return -EIO;
 397	}
 398
 399	while (list_empty(&file->recv_list)) {
 400		mutex_unlock(&file->mutex);
 401
 402		if (filp->f_flags & O_NONBLOCK)
 403			return -EAGAIN;
 404
 405		if (wait_event_interruptible(file->recv_wait,
 406					     !list_empty(&file->recv_list)))
 407			return -ERESTARTSYS;
 408
 409		mutex_lock(&file->mutex);
 410	}
 411
 412	if (file->agents_dead) {
 413		mutex_unlock(&file->mutex);
 414		return -EIO;
 415	}
 416
 417	packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
 418	list_del(&packet->list);
 419	atomic_dec(&file->recv_list_size);
 420
 421	mutex_unlock(&file->mutex);
 422
 423	if (packet->recv_wc)
 424		ret = copy_recv_mad(file, buf, packet, count);
 425	else
 426		ret = copy_send_mad(file, buf, packet, count);
 427
 428	if (ret < 0) {
 429		/* Requeue packet */
 430		mutex_lock(&file->mutex);
 431		list_add(&packet->list, &file->recv_list);
 432		atomic_inc(&file->recv_list_size);
 433		mutex_unlock(&file->mutex);
 434	} else {
 435		if (packet->recv_wc)
 436			ib_free_recv_mad(packet->recv_wc);
 437		kfree(packet);
 438	}
 439	return ret;
 440}
 441
 442static int copy_rmpp_mad(struct ib_mad_send_buf *msg, const char __user *buf)
 443{
 444	int left, seg;
 445
 446	/* Copy class specific header */
 447	if ((msg->hdr_len > IB_MGMT_RMPP_HDR) &&
 448	    copy_from_user(msg->mad + IB_MGMT_RMPP_HDR, buf + IB_MGMT_RMPP_HDR,
 449			   msg->hdr_len - IB_MGMT_RMPP_HDR))
 450		return -EFAULT;
 451
 452	/* All headers are in place.  Copy data segments. */
 453	for (seg = 1, left = msg->data_len, buf += msg->hdr_len; left > 0;
 454	     seg++, left -= msg->seg_size, buf += msg->seg_size) {
 455		if (copy_from_user(ib_get_rmpp_segment(msg, seg), buf,
 456				   min(left, msg->seg_size)))
 457			return -EFAULT;
 458	}
 459	return 0;
 460}
 461
 462static int same_destination(struct ib_user_mad_hdr *hdr1,
 463			    struct ib_user_mad_hdr *hdr2)
 464{
 465	if (!hdr1->grh_present && !hdr2->grh_present)
 466	   return (hdr1->lid == hdr2->lid);
 467
 468	if (hdr1->grh_present && hdr2->grh_present)
 469	   return !memcmp(hdr1->gid, hdr2->gid, 16);
 470
 471	return 0;
 472}
 473
 474static int is_duplicate(struct ib_umad_file *file,
 475			struct ib_umad_packet *packet)
 476{
 477	struct ib_umad_packet *sent_packet;
 478	struct ib_mad_hdr *sent_hdr, *hdr;
 479
 480	hdr = (struct ib_mad_hdr *) packet->mad.data;
 481	list_for_each_entry(sent_packet, &file->send_list, list) {
 482		sent_hdr = (struct ib_mad_hdr *) sent_packet->mad.data;
 483
 484		if ((hdr->tid != sent_hdr->tid) ||
 485		    (hdr->mgmt_class != sent_hdr->mgmt_class))
 486			continue;
 487
 488		/*
 489		 * No need to be overly clever here.  If two new operations have
 490		 * the same TID, reject the second as a duplicate.  This is more
 491		 * restrictive than required by the spec.
 492		 */
 493		if (!ib_response_mad(hdr)) {
 494			if (!ib_response_mad(sent_hdr))
 495				return 1;
 496			continue;
 497		} else if (!ib_response_mad(sent_hdr))
 498			continue;
 499
 500		if (same_destination(&packet->mad.hdr, &sent_packet->mad.hdr))
 501			return 1;
 502	}
 503
 504	return 0;
 505}
 506
 507static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
 508			     size_t count, loff_t *pos)
 509{
 510	struct ib_umad_file *file = filp->private_data;
 511	struct ib_rmpp_mad_hdr *rmpp_mad_hdr;
 512	struct ib_umad_packet *packet;
 513	struct ib_mad_agent *agent;
 514	struct rdma_ah_attr ah_attr;
 515	struct ib_ah *ah;
 516	__be64 *tid;
 517	int ret, hdr_len, copy_offset, rmpp_active;
 518	size_t data_len;
 519	u8 base_version;
 520
 521	if (count < hdr_size(file) + IB_MGMT_RMPP_HDR)
 522		return -EINVAL;
 523
 524	packet = kzalloc(sizeof(*packet) + IB_MGMT_RMPP_HDR, GFP_KERNEL);
 525	if (!packet)
 526		return -ENOMEM;
 527
 528	if (copy_from_user(&packet->mad, buf, hdr_size(file))) {
 529		ret = -EFAULT;
 530		goto err;
 531	}
 532
 533	if (packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
 534		ret = -EINVAL;
 535		goto err;
 536	}
 537
 538	buf += hdr_size(file);
 539
 540	if (copy_from_user(packet->mad.data, buf, IB_MGMT_RMPP_HDR)) {
 541		ret = -EFAULT;
 542		goto err;
 543	}
 544
 545	mutex_lock(&file->mutex);
 546
 547	trace_ib_umad_write(file, &packet->mad.hdr,
 548			    (struct ib_mad_hdr *)&packet->mad.data);
 549
 550	agent = __get_agent(file, packet->mad.hdr.id);
 551	if (!agent) {
 552		ret = -EIO;
 553		goto err_up;
 554	}
 555
 556	memset(&ah_attr, 0, sizeof ah_attr);
 557	ah_attr.type = rdma_ah_find_type(agent->device,
 558					 file->port->port_num);
 559	rdma_ah_set_dlid(&ah_attr, be16_to_cpu(packet->mad.hdr.lid));
 560	rdma_ah_set_sl(&ah_attr, packet->mad.hdr.sl);
 561	rdma_ah_set_path_bits(&ah_attr, packet->mad.hdr.path_bits);
 562	rdma_ah_set_port_num(&ah_attr, file->port->port_num);
 563	if (packet->mad.hdr.grh_present) {
 564		rdma_ah_set_grh(&ah_attr, NULL,
 565				be32_to_cpu(packet->mad.hdr.flow_label),
 566				packet->mad.hdr.gid_index,
 567				packet->mad.hdr.hop_limit,
 568				packet->mad.hdr.traffic_class);
 569		rdma_ah_set_dgid_raw(&ah_attr, packet->mad.hdr.gid);
 570	}
 571
 572	ah = rdma_create_user_ah(agent->qp->pd, &ah_attr, NULL);
 573	if (IS_ERR(ah)) {
 574		ret = PTR_ERR(ah);
 575		goto err_up;
 576	}
 577
 578	rmpp_mad_hdr = (struct ib_rmpp_mad_hdr *)packet->mad.data;
 579	hdr_len = ib_get_mad_data_offset(rmpp_mad_hdr->mad_hdr.mgmt_class);
 580
 581	if (ib_is_mad_class_rmpp(rmpp_mad_hdr->mad_hdr.mgmt_class)
 582	    && ib_mad_kernel_rmpp_agent(agent)) {
 583		copy_offset = IB_MGMT_RMPP_HDR;
 584		rmpp_active = ib_get_rmpp_flags(&rmpp_mad_hdr->rmpp_hdr) &
 585						IB_MGMT_RMPP_FLAG_ACTIVE;
 586	} else {
 587		copy_offset = IB_MGMT_MAD_HDR;
 588		rmpp_active = 0;
 589	}
 590
 591	base_version = ((struct ib_mad_hdr *)&packet->mad.data)->base_version;
 592	if (check_sub_overflow(count, hdr_size(file) + hdr_len, &data_len)) {
 593		ret = -EINVAL;
 594		goto err_ah;
 595	}
 596	packet->msg = ib_create_send_mad(agent,
 597					 be32_to_cpu(packet->mad.hdr.qpn),
 598					 packet->mad.hdr.pkey_index, rmpp_active,
 599					 hdr_len, data_len, GFP_KERNEL,
 600					 base_version);
 601	if (IS_ERR(packet->msg)) {
 602		ret = PTR_ERR(packet->msg);
 603		goto err_ah;
 604	}
 605
 606	packet->msg->ah		= ah;
 607	packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
 608	packet->msg->retries	= packet->mad.hdr.retries;
 609	packet->msg->context[0] = packet;
 610
 611	/* Copy MAD header.  Any RMPP header is already in place. */
 612	memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
 613
 614	if (!rmpp_active) {
 615		if (copy_from_user(packet->msg->mad + copy_offset,
 616				   buf + copy_offset,
 617				   hdr_len + data_len - copy_offset)) {
 618			ret = -EFAULT;
 619			goto err_msg;
 620		}
 621	} else {
 622		ret = copy_rmpp_mad(packet->msg, buf);
 623		if (ret)
 624			goto err_msg;
 625	}
 626
 627	/*
 628	 * Set the high-order part of the transaction ID to make MADs from
 629	 * different agents unique, and allow routing responses back to the
 630	 * original requestor.
 631	 */
 632	if (!ib_response_mad(packet->msg->mad)) {
 633		tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
 634		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
 635				   (be64_to_cpup(tid) & 0xffffffff));
 636		rmpp_mad_hdr->mad_hdr.tid = *tid;
 637	}
 638
 639	if (!ib_mad_kernel_rmpp_agent(agent)
 640	    && ib_is_mad_class_rmpp(rmpp_mad_hdr->mad_hdr.mgmt_class)
 641	    && (ib_get_rmpp_flags(&rmpp_mad_hdr->rmpp_hdr) & IB_MGMT_RMPP_FLAG_ACTIVE)) {
 642		spin_lock_irq(&file->send_lock);
 643		list_add_tail(&packet->list, &file->send_list);
 644		spin_unlock_irq(&file->send_lock);
 645	} else {
 646		spin_lock_irq(&file->send_lock);
 647		ret = is_duplicate(file, packet);
 648		if (!ret)
 649			list_add_tail(&packet->list, &file->send_list);
 650		spin_unlock_irq(&file->send_lock);
 651		if (ret) {
 652			ret = -EINVAL;
 653			goto err_msg;
 654		}
 655	}
 656
 657	ret = ib_post_send_mad(packet->msg, NULL);
 658	if (ret)
 659		goto err_send;
 660
 661	mutex_unlock(&file->mutex);
 662	return count;
 663
 664err_send:
 665	dequeue_send(file, packet);
 666err_msg:
 667	ib_free_send_mad(packet->msg);
 668err_ah:
 669	rdma_destroy_ah(ah, RDMA_DESTROY_AH_SLEEPABLE);
 670err_up:
 671	mutex_unlock(&file->mutex);
 672err:
 673	kfree(packet);
 674	return ret;
 675}
 676
 677static __poll_t ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
 678{
 679	struct ib_umad_file *file = filp->private_data;
 680
 681	/* we will always be able to post a MAD send */
 682	__poll_t mask = EPOLLOUT | EPOLLWRNORM;
 683
 684	mutex_lock(&file->mutex);
 685	poll_wait(filp, &file->recv_wait, wait);
 686
 687	if (!list_empty(&file->recv_list))
 688		mask |= EPOLLIN | EPOLLRDNORM;
 689	if (file->agents_dead)
 690		mask = EPOLLERR;
 691	mutex_unlock(&file->mutex);
 692
 693	return mask;
 694}
 695
 696static int ib_umad_reg_agent(struct ib_umad_file *file, void __user *arg,
 697			     int compat_method_mask)
 698{
 699	struct ib_user_mad_reg_req ureq;
 700	struct ib_mad_reg_req req;
 701	struct ib_mad_agent *agent = NULL;
 702	int agent_id;
 703	int ret;
 704
 705	mutex_lock(&file->port->file_mutex);
 706	mutex_lock(&file->mutex);
 707
 708	if (!file->port->ib_dev) {
 709		dev_notice(&file->port->dev, "%s: invalid device\n", __func__);
 710		ret = -EPIPE;
 711		goto out;
 712	}
 713
 714	if (copy_from_user(&ureq, arg, sizeof ureq)) {
 715		ret = -EFAULT;
 716		goto out;
 717	}
 718
 719	if (ureq.qpn != 0 && ureq.qpn != 1) {
 720		dev_notice(&file->port->dev,
 721			   "%s: invalid QPN %u specified\n", __func__,
 722			   ureq.qpn);
 723		ret = -EINVAL;
 724		goto out;
 725	}
 726
 727	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
 728		if (!__get_agent(file, agent_id))
 729			goto found;
 730
 731	dev_notice(&file->port->dev, "%s: Max Agents (%u) reached\n", __func__,
 732		   IB_UMAD_MAX_AGENTS);
 733
 734	ret = -ENOMEM;
 735	goto out;
 736
 737found:
 738	if (ureq.mgmt_class) {
 739		memset(&req, 0, sizeof(req));
 740		req.mgmt_class         = ureq.mgmt_class;
 741		req.mgmt_class_version = ureq.mgmt_class_version;
 742		memcpy(req.oui, ureq.oui, sizeof req.oui);
 743
 744		if (compat_method_mask) {
 745			u32 *umm = (u32 *) ureq.method_mask;
 746			int i;
 747
 748			for (i = 0; i < BITS_TO_LONGS(IB_MGMT_MAX_METHODS); ++i)
 749				req.method_mask[i] =
 750					umm[i * 2] | ((u64) umm[i * 2 + 1] << 32);
 751		} else
 752			memcpy(req.method_mask, ureq.method_mask,
 753			       sizeof req.method_mask);
 754	}
 755
 756	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
 757				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
 758				      ureq.mgmt_class ? &req : NULL,
 759				      ureq.rmpp_version,
 760				      send_handler, recv_handler, file, 0);
 761	if (IS_ERR(agent)) {
 762		ret = PTR_ERR(agent);
 763		agent = NULL;
 764		goto out;
 765	}
 766
 767	if (put_user(agent_id,
 768		     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
 769		ret = -EFAULT;
 770		goto out;
 771	}
 772
 773	if (!file->already_used) {
 774		file->already_used = 1;
 775		if (!file->use_pkey_index) {
 776			dev_warn(&file->port->dev,
 777				"process %s did not enable P_Key index support.\n",
 778				current->comm);
 779			dev_warn(&file->port->dev,
 780				"   Documentation/infiniband/user_mad.rst has info on the new ABI.\n");
 781		}
 782	}
 783
 784	file->agent[agent_id] = agent;
 785	ret = 0;
 786
 787out:
 788	mutex_unlock(&file->mutex);
 789
 790	if (ret && agent)
 791		ib_unregister_mad_agent(agent);
 792
 793	mutex_unlock(&file->port->file_mutex);
 794
 795	return ret;
 796}
 797
 798static int ib_umad_reg_agent2(struct ib_umad_file *file, void __user *arg)
 799{
 800	struct ib_user_mad_reg_req2 ureq;
 801	struct ib_mad_reg_req req;
 802	struct ib_mad_agent *agent = NULL;
 803	int agent_id;
 804	int ret;
 805
 806	mutex_lock(&file->port->file_mutex);
 807	mutex_lock(&file->mutex);
 808
 809	if (!file->port->ib_dev) {
 810		dev_notice(&file->port->dev, "%s: invalid device\n", __func__);
 811		ret = -EPIPE;
 812		goto out;
 813	}
 814
 815	if (copy_from_user(&ureq, arg, sizeof(ureq))) {
 816		ret = -EFAULT;
 817		goto out;
 818	}
 819
 820	if (ureq.qpn != 0 && ureq.qpn != 1) {
 821		dev_notice(&file->port->dev, "%s: invalid QPN %u specified\n",
 822			   __func__, ureq.qpn);
 823		ret = -EINVAL;
 824		goto out;
 825	}
 826
 827	if (ureq.flags & ~IB_USER_MAD_REG_FLAGS_CAP) {
 828		dev_notice(&file->port->dev,
 829			   "%s failed: invalid registration flags specified 0x%x; supported 0x%x\n",
 830			   __func__, ureq.flags, IB_USER_MAD_REG_FLAGS_CAP);
 831		ret = -EINVAL;
 832
 833		if (put_user((u32)IB_USER_MAD_REG_FLAGS_CAP,
 834				(u32 __user *) (arg + offsetof(struct
 835				ib_user_mad_reg_req2, flags))))
 836			ret = -EFAULT;
 837
 838		goto out;
 839	}
 840
 841	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
 842		if (!__get_agent(file, agent_id))
 843			goto found;
 844
 845	dev_notice(&file->port->dev, "%s: Max Agents (%u) reached\n", __func__,
 846		   IB_UMAD_MAX_AGENTS);
 847	ret = -ENOMEM;
 848	goto out;
 849
 850found:
 851	if (ureq.mgmt_class) {
 852		memset(&req, 0, sizeof(req));
 853		req.mgmt_class         = ureq.mgmt_class;
 854		req.mgmt_class_version = ureq.mgmt_class_version;
 855		if (ureq.oui & 0xff000000) {
 856			dev_notice(&file->port->dev,
 857				   "%s failed: oui invalid 0x%08x\n", __func__,
 858				   ureq.oui);
 859			ret = -EINVAL;
 860			goto out;
 861		}
 862		req.oui[2] =  ureq.oui & 0x0000ff;
 863		req.oui[1] = (ureq.oui & 0x00ff00) >> 8;
 864		req.oui[0] = (ureq.oui & 0xff0000) >> 16;
 865		memcpy(req.method_mask, ureq.method_mask,
 866			sizeof(req.method_mask));
 867	}
 868
 869	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
 870				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
 871				      ureq.mgmt_class ? &req : NULL,
 872				      ureq.rmpp_version,
 873				      send_handler, recv_handler, file,
 874				      ureq.flags);
 875	if (IS_ERR(agent)) {
 876		ret = PTR_ERR(agent);
 877		agent = NULL;
 878		goto out;
 879	}
 880
 881	if (put_user(agent_id,
 882		     (u32 __user *)(arg +
 883				offsetof(struct ib_user_mad_reg_req2, id)))) {
 884		ret = -EFAULT;
 885		goto out;
 886	}
 887
 888	if (!file->already_used) {
 889		file->already_used = 1;
 890		file->use_pkey_index = 1;
 891	}
 892
 893	file->agent[agent_id] = agent;
 894	ret = 0;
 895
 896out:
 897	mutex_unlock(&file->mutex);
 898
 899	if (ret && agent)
 900		ib_unregister_mad_agent(agent);
 901
 902	mutex_unlock(&file->port->file_mutex);
 903
 904	return ret;
 905}
 906
 907
 908static int ib_umad_unreg_agent(struct ib_umad_file *file, u32 __user *arg)
 909{
 910	struct ib_mad_agent *agent = NULL;
 911	u32 id;
 912	int ret = 0;
 913
 914	if (get_user(id, arg))
 915		return -EFAULT;
 916	if (id >= IB_UMAD_MAX_AGENTS)
 917		return -EINVAL;
 918
 919	mutex_lock(&file->port->file_mutex);
 920	mutex_lock(&file->mutex);
 921
 922	id = array_index_nospec(id, IB_UMAD_MAX_AGENTS);
 923	if (!__get_agent(file, id)) {
 924		ret = -EINVAL;
 925		goto out;
 926	}
 927
 928	agent = file->agent[id];
 929	file->agent[id] = NULL;
 930
 931out:
 932	mutex_unlock(&file->mutex);
 933
 934	if (agent)
 935		ib_unregister_mad_agent(agent);
 936
 937	mutex_unlock(&file->port->file_mutex);
 938
 939	return ret;
 940}
 941
 942static long ib_umad_enable_pkey(struct ib_umad_file *file)
 943{
 944	int ret = 0;
 945
 946	mutex_lock(&file->mutex);
 947	if (file->already_used)
 948		ret = -EINVAL;
 949	else
 950		file->use_pkey_index = 1;
 951	mutex_unlock(&file->mutex);
 952
 953	return ret;
 954}
 955
 956static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
 957			  unsigned long arg)
 958{
 959	switch (cmd) {
 960	case IB_USER_MAD_REGISTER_AGENT:
 961		return ib_umad_reg_agent(filp->private_data, (void __user *) arg, 0);
 962	case IB_USER_MAD_UNREGISTER_AGENT:
 963		return ib_umad_unreg_agent(filp->private_data, (__u32 __user *) arg);
 964	case IB_USER_MAD_ENABLE_PKEY:
 965		return ib_umad_enable_pkey(filp->private_data);
 966	case IB_USER_MAD_REGISTER_AGENT2:
 967		return ib_umad_reg_agent2(filp->private_data, (void __user *) arg);
 968	default:
 969		return -ENOIOCTLCMD;
 970	}
 971}
 972
 973#ifdef CONFIG_COMPAT
 974static long ib_umad_compat_ioctl(struct file *filp, unsigned int cmd,
 975				 unsigned long arg)
 976{
 977	switch (cmd) {
 978	case IB_USER_MAD_REGISTER_AGENT:
 979		return ib_umad_reg_agent(filp->private_data, compat_ptr(arg), 1);
 980	case IB_USER_MAD_UNREGISTER_AGENT:
 981		return ib_umad_unreg_agent(filp->private_data, compat_ptr(arg));
 982	case IB_USER_MAD_ENABLE_PKEY:
 983		return ib_umad_enable_pkey(filp->private_data);
 984	case IB_USER_MAD_REGISTER_AGENT2:
 985		return ib_umad_reg_agent2(filp->private_data, compat_ptr(arg));
 986	default:
 987		return -ENOIOCTLCMD;
 988	}
 989}
 990#endif
 991
 992/*
 993 * ib_umad_open() does not need the BKL:
 994 *
 995 *  - the ib_umad_port structures are properly reference counted, and
 996 *    everything else is purely local to the file being created, so
 997 *    races against other open calls are not a problem;
 998 *  - the ioctl method does not affect any global state outside of the
 999 *    file structure being operated on;
1000 */
1001static int ib_umad_open(struct inode *inode, struct file *filp)
1002{
1003	struct ib_umad_port *port;
1004	struct ib_umad_file *file;
1005	int ret = 0;
1006
1007	port = container_of(inode->i_cdev, struct ib_umad_port, cdev);
1008
1009	mutex_lock(&port->file_mutex);
1010
1011	if (!port->ib_dev) {
1012		ret = -ENXIO;
1013		goto out;
1014	}
1015
1016	if (!rdma_dev_access_netns(port->ib_dev, current->nsproxy->net_ns)) {
1017		ret = -EPERM;
1018		goto out;
1019	}
1020
1021	file = kzalloc(sizeof(*file), GFP_KERNEL);
1022	if (!file) {
1023		ret = -ENOMEM;
1024		goto out;
1025	}
1026
1027	mutex_init(&file->mutex);
1028	spin_lock_init(&file->send_lock);
1029	INIT_LIST_HEAD(&file->recv_list);
1030	INIT_LIST_HEAD(&file->send_list);
1031	init_waitqueue_head(&file->recv_wait);
1032
1033	file->port = port;
1034	filp->private_data = file;
1035
1036	list_add_tail(&file->port_list, &port->file_list);
1037
1038	stream_open(inode, filp);
1039out:
1040	mutex_unlock(&port->file_mutex);
1041	return ret;
1042}
1043
1044static int ib_umad_close(struct inode *inode, struct file *filp)
1045{
1046	struct ib_umad_file *file = filp->private_data;
1047	struct ib_umad_packet *packet, *tmp;
1048	int already_dead;
1049	int i;
1050
1051	mutex_lock(&file->port->file_mutex);
1052	mutex_lock(&file->mutex);
1053
1054	already_dead = file->agents_dead;
1055	file->agents_dead = 1;
1056
1057	list_for_each_entry_safe(packet, tmp, &file->recv_list, list) {
1058		if (packet->recv_wc)
1059			ib_free_recv_mad(packet->recv_wc);
1060		kfree(packet);
1061	}
1062
1063	list_del(&file->port_list);
1064
1065	mutex_unlock(&file->mutex);
1066
1067	if (!already_dead)
1068		for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
1069			if (file->agent[i])
1070				ib_unregister_mad_agent(file->agent[i]);
1071
1072	mutex_unlock(&file->port->file_mutex);
1073	mutex_destroy(&file->mutex);
1074	kfree(file);
1075	return 0;
1076}
1077
1078static const struct file_operations umad_fops = {
1079	.owner		= THIS_MODULE,
1080	.read		= ib_umad_read,
1081	.write		= ib_umad_write,
1082	.poll		= ib_umad_poll,
1083	.unlocked_ioctl = ib_umad_ioctl,
1084#ifdef CONFIG_COMPAT
1085	.compat_ioctl	= ib_umad_compat_ioctl,
1086#endif
1087	.open		= ib_umad_open,
1088	.release	= ib_umad_close,
1089};
1090
1091static int ib_umad_sm_open(struct inode *inode, struct file *filp)
1092{
1093	struct ib_umad_port *port;
1094	struct ib_port_modify props = {
1095		.set_port_cap_mask = IB_PORT_SM
1096	};
1097	int ret;
1098
1099	port = container_of(inode->i_cdev, struct ib_umad_port, sm_cdev);
1100
1101	if (filp->f_flags & O_NONBLOCK) {
1102		if (down_trylock(&port->sm_sem)) {
1103			ret = -EAGAIN;
1104			goto fail;
1105		}
1106	} else {
1107		if (down_interruptible(&port->sm_sem)) {
1108			ret = -ERESTARTSYS;
1109			goto fail;
1110		}
1111	}
1112
1113	if (!rdma_dev_access_netns(port->ib_dev, current->nsproxy->net_ns)) {
1114		ret = -EPERM;
1115		goto err_up_sem;
1116	}
1117
1118	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
1119	if (ret)
1120		goto err_up_sem;
1121
1122	filp->private_data = port;
1123
1124	nonseekable_open(inode, filp);
1125	return 0;
1126
1127err_up_sem:
1128	up(&port->sm_sem);
1129
1130fail:
1131	return ret;
1132}
1133
1134static int ib_umad_sm_close(struct inode *inode, struct file *filp)
1135{
1136	struct ib_umad_port *port = filp->private_data;
1137	struct ib_port_modify props = {
1138		.clr_port_cap_mask = IB_PORT_SM
1139	};
1140	int ret = 0;
1141
1142	mutex_lock(&port->file_mutex);
1143	if (port->ib_dev)
1144		ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
1145	mutex_unlock(&port->file_mutex);
1146
1147	up(&port->sm_sem);
1148
1149	return ret;
1150}
1151
1152static const struct file_operations umad_sm_fops = {
1153	.owner	 = THIS_MODULE,
1154	.open	 = ib_umad_sm_open,
1155	.release = ib_umad_sm_close,
1156};
1157
1158static struct ib_umad_port *get_port(struct ib_device *ibdev,
1159				     struct ib_umad_device *umad_dev,
1160				     u32 port)
1161{
1162	if (!umad_dev)
1163		return ERR_PTR(-EOPNOTSUPP);
1164	if (!rdma_is_port_valid(ibdev, port))
1165		return ERR_PTR(-EINVAL);
1166	if (!rdma_cap_ib_mad(ibdev, port))
1167		return ERR_PTR(-EOPNOTSUPP);
1168
1169	return &umad_dev->ports[port - rdma_start_port(ibdev)];
1170}
1171
1172static int ib_umad_get_nl_info(struct ib_device *ibdev, void *client_data,
1173			       struct ib_client_nl_info *res)
1174{
1175	struct ib_umad_port *port = get_port(ibdev, client_data, res->port);
1176
1177	if (IS_ERR(port))
1178		return PTR_ERR(port);
1179
1180	res->abi = IB_USER_MAD_ABI_VERSION;
1181	res->cdev = &port->dev;
1182	return 0;
1183}
1184
1185static struct ib_client umad_client = {
1186	.name   = "umad",
1187	.add    = ib_umad_add_one,
1188	.remove = ib_umad_remove_one,
1189	.get_nl_info = ib_umad_get_nl_info,
1190};
1191MODULE_ALIAS_RDMA_CLIENT("umad");
1192
1193static int ib_issm_get_nl_info(struct ib_device *ibdev, void *client_data,
1194			       struct ib_client_nl_info *res)
1195{
1196	struct ib_umad_port *port = get_port(ibdev, client_data, res->port);
1197
1198	if (IS_ERR(port))
1199		return PTR_ERR(port);
1200
1201	res->abi = IB_USER_MAD_ABI_VERSION;
1202	res->cdev = &port->sm_dev;
1203	return 0;
1204}
1205
1206static struct ib_client issm_client = {
1207	.name = "issm",
1208	.get_nl_info = ib_issm_get_nl_info,
1209};
1210MODULE_ALIAS_RDMA_CLIENT("issm");
1211
1212static ssize_t ibdev_show(struct device *dev, struct device_attribute *attr,
1213			  char *buf)
1214{
1215	struct ib_umad_port *port = dev_get_drvdata(dev);
1216
1217	if (!port)
1218		return -ENODEV;
1219
1220	return sysfs_emit(buf, "%s\n", dev_name(&port->ib_dev->dev));
1221}
1222static DEVICE_ATTR_RO(ibdev);
1223
1224static ssize_t port_show(struct device *dev, struct device_attribute *attr,
1225			 char *buf)
1226{
1227	struct ib_umad_port *port = dev_get_drvdata(dev);
1228
1229	if (!port)
1230		return -ENODEV;
1231
1232	return sysfs_emit(buf, "%d\n", port->port_num);
1233}
1234static DEVICE_ATTR_RO(port);
1235
1236static struct attribute *umad_class_dev_attrs[] = {
1237	&dev_attr_ibdev.attr,
1238	&dev_attr_port.attr,
1239	NULL,
1240};
1241ATTRIBUTE_GROUPS(umad_class_dev);
1242
1243static char *umad_devnode(const struct device *dev, umode_t *mode)
1244{
1245	return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1246}
1247
1248static ssize_t abi_version_show(const struct class *class,
1249				const struct class_attribute *attr, char *buf)
1250{
1251	return sysfs_emit(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
1252}
1253static CLASS_ATTR_RO(abi_version);
1254
1255static struct attribute *umad_class_attrs[] = {
1256	&class_attr_abi_version.attr,
1257	NULL,
1258};
1259ATTRIBUTE_GROUPS(umad_class);
1260
1261static struct class umad_class = {
1262	.name		= "infiniband_mad",
1263	.devnode	= umad_devnode,
1264	.class_groups	= umad_class_groups,
1265	.dev_groups	= umad_class_dev_groups,
1266};
1267
1268static void ib_umad_release_port(struct device *device)
1269{
1270	struct ib_umad_port *port = dev_get_drvdata(device);
1271	struct ib_umad_device *umad_dev = port->umad_dev;
1272
1273	ib_umad_dev_put(umad_dev);
1274}
1275
1276static void ib_umad_init_port_dev(struct device *dev,
1277				  struct ib_umad_port *port,
1278				  const struct ib_device *device)
1279{
1280	device_initialize(dev);
1281	ib_umad_dev_get(port->umad_dev);
1282	dev->class = &umad_class;
1283	dev->parent = device->dev.parent;
1284	dev_set_drvdata(dev, port);
1285	dev->release = ib_umad_release_port;
1286}
1287
1288static int ib_umad_init_port(struct ib_device *device, int port_num,
1289			     struct ib_umad_device *umad_dev,
1290			     struct ib_umad_port *port)
1291{
1292	int devnum;
1293	dev_t base_umad;
1294	dev_t base_issm;
1295	int ret;
1296
1297	devnum = ida_alloc_max(&umad_ida, IB_UMAD_MAX_PORTS - 1, GFP_KERNEL);
1298	if (devnum < 0)
1299		return -1;
1300	port->dev_num = devnum;
1301	if (devnum >= IB_UMAD_NUM_FIXED_MINOR) {
1302		base_umad = dynamic_umad_dev + devnum - IB_UMAD_NUM_FIXED_MINOR;
1303		base_issm = dynamic_issm_dev + devnum - IB_UMAD_NUM_FIXED_MINOR;
1304	} else {
1305		base_umad = devnum + base_umad_dev;
1306		base_issm = devnum + base_issm_dev;
1307	}
1308
1309	port->ib_dev   = device;
1310	port->umad_dev = umad_dev;
1311	port->port_num = port_num;
1312	sema_init(&port->sm_sem, 1);
1313	mutex_init(&port->file_mutex);
1314	INIT_LIST_HEAD(&port->file_list);
1315
1316	ib_umad_init_port_dev(&port->dev, port, device);
1317	port->dev.devt = base_umad;
1318	dev_set_name(&port->dev, "umad%d", port->dev_num);
1319	cdev_init(&port->cdev, &umad_fops);
1320	port->cdev.owner = THIS_MODULE;
1321
1322	ret = cdev_device_add(&port->cdev, &port->dev);
1323	if (ret)
1324		goto err_cdev;
1325
1326	if (rdma_cap_ib_smi(device, port_num)) {
1327		ib_umad_init_port_dev(&port->sm_dev, port, device);
1328		port->sm_dev.devt = base_issm;
1329		dev_set_name(&port->sm_dev, "issm%d", port->dev_num);
1330		cdev_init(&port->sm_cdev, &umad_sm_fops);
1331		port->sm_cdev.owner = THIS_MODULE;
1332
1333		ret = cdev_device_add(&port->sm_cdev, &port->sm_dev);
1334		if (ret)
1335			goto err_dev;
1336	}
1337
1338	return 0;
1339
1340err_dev:
1341	put_device(&port->sm_dev);
1342	cdev_device_del(&port->cdev, &port->dev);
1343err_cdev:
1344	put_device(&port->dev);
1345	ida_free(&umad_ida, devnum);
1346	return ret;
1347}
1348
1349static void ib_umad_kill_port(struct ib_umad_port *port)
1350{
1351	struct ib_umad_file *file;
1352	bool has_smi = false;
1353	int id;
1354
1355	if (rdma_cap_ib_smi(port->ib_dev, port->port_num)) {
1356		cdev_device_del(&port->sm_cdev, &port->sm_dev);
1357		has_smi = true;
1358	}
1359	cdev_device_del(&port->cdev, &port->dev);
1360
1361	mutex_lock(&port->file_mutex);
1362
1363	/* Mark ib_dev NULL and block ioctl or other file ops to progress
1364	 * further.
1365	 */
1366	port->ib_dev = NULL;
1367
1368	list_for_each_entry(file, &port->file_list, port_list) {
1369		mutex_lock(&file->mutex);
1370		file->agents_dead = 1;
1371		wake_up_interruptible(&file->recv_wait);
1372		mutex_unlock(&file->mutex);
1373
1374		for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id)
1375			if (file->agent[id])
1376				ib_unregister_mad_agent(file->agent[id]);
1377	}
1378
1379	mutex_unlock(&port->file_mutex);
1380
1381	ida_free(&umad_ida, port->dev_num);
1382
1383	/* balances device_initialize() */
1384	if (has_smi)
1385		put_device(&port->sm_dev);
1386	put_device(&port->dev);
1387}
1388
1389static int ib_umad_add_one(struct ib_device *device)
1390{
1391	struct ib_umad_device *umad_dev;
1392	int s, e, i;
1393	int count = 0;
1394	int ret;
1395
1396	s = rdma_start_port(device);
1397	e = rdma_end_port(device);
1398
1399	umad_dev = kzalloc(struct_size(umad_dev, ports,
1400				       size_add(size_sub(e, s), 1)),
1401			   GFP_KERNEL);
1402	if (!umad_dev)
1403		return -ENOMEM;
1404
1405	kref_init(&umad_dev->kref);
1406	for (i = s; i <= e; ++i) {
1407		if (!rdma_cap_ib_mad(device, i))
1408			continue;
1409
1410		ret = ib_umad_init_port(device, i, umad_dev,
1411					&umad_dev->ports[i - s]);
1412		if (ret)
1413			goto err;
1414
1415		count++;
1416	}
1417
1418	if (!count) {
1419		ret = -EOPNOTSUPP;
1420		goto free;
1421	}
1422
1423	ib_set_client_data(device, &umad_client, umad_dev);
1424
1425	return 0;
1426
1427err:
1428	while (--i >= s) {
1429		if (!rdma_cap_ib_mad(device, i))
1430			continue;
1431
1432		ib_umad_kill_port(&umad_dev->ports[i - s]);
1433	}
1434free:
1435	/* balances kref_init */
1436	ib_umad_dev_put(umad_dev);
1437	return ret;
1438}
1439
1440static void ib_umad_remove_one(struct ib_device *device, void *client_data)
1441{
1442	struct ib_umad_device *umad_dev = client_data;
1443	unsigned int i;
1444
1445	rdma_for_each_port (device, i) {
1446		if (rdma_cap_ib_mad(device, i))
1447			ib_umad_kill_port(
1448				&umad_dev->ports[i - rdma_start_port(device)]);
1449	}
1450	/* balances kref_init() */
1451	ib_umad_dev_put(umad_dev);
1452}
1453
1454static int __init ib_umad_init(void)
1455{
1456	int ret;
1457
1458	ret = register_chrdev_region(base_umad_dev,
1459				     IB_UMAD_NUM_FIXED_MINOR * 2,
1460				     umad_class.name);
1461	if (ret) {
1462		pr_err("couldn't register device number\n");
1463		goto out;
1464	}
1465
1466	ret = alloc_chrdev_region(&dynamic_umad_dev, 0,
1467				  IB_UMAD_NUM_DYNAMIC_MINOR * 2,
1468				  umad_class.name);
1469	if (ret) {
1470		pr_err("couldn't register dynamic device number\n");
1471		goto out_alloc;
1472	}
1473	dynamic_issm_dev = dynamic_umad_dev + IB_UMAD_NUM_DYNAMIC_MINOR;
1474
1475	ret = class_register(&umad_class);
1476	if (ret) {
1477		pr_err("couldn't create class infiniband_mad\n");
1478		goto out_chrdev;
1479	}
1480
1481	ret = ib_register_client(&umad_client);
1482	if (ret)
1483		goto out_class;
1484
1485	ret = ib_register_client(&issm_client);
1486	if (ret)
1487		goto out_client;
1488
1489	return 0;
1490
1491out_client:
1492	ib_unregister_client(&umad_client);
1493out_class:
1494	class_unregister(&umad_class);
1495
1496out_chrdev:
1497	unregister_chrdev_region(dynamic_umad_dev,
1498				 IB_UMAD_NUM_DYNAMIC_MINOR * 2);
1499
1500out_alloc:
1501	unregister_chrdev_region(base_umad_dev,
1502				 IB_UMAD_NUM_FIXED_MINOR * 2);
1503
1504out:
1505	return ret;
1506}
1507
1508static void __exit ib_umad_cleanup(void)
1509{
1510	ib_unregister_client(&issm_client);
1511	ib_unregister_client(&umad_client);
1512	class_unregister(&umad_class);
1513	unregister_chrdev_region(base_umad_dev,
1514				 IB_UMAD_NUM_FIXED_MINOR * 2);
1515	unregister_chrdev_region(dynamic_umad_dev,
1516				 IB_UMAD_NUM_DYNAMIC_MINOR * 2);
1517}
1518
1519module_init(ib_umad_init);
1520module_exit(ib_umad_cleanup);