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