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