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