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