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 	u32                     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 int 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 	if (file->agents_dead) {
383 		mutex_unlock(&file->mutex);
384 		return -EIO;
385 	}
386 
387 	while (list_empty(&file->recv_list)) {
388 		mutex_unlock(&file->mutex);
389 
390 		if (filp->f_flags & O_NONBLOCK)
391 			return -EAGAIN;
392 
393 		if (wait_event_interruptible(file->recv_wait,
394 					     !list_empty(&file->recv_list)))
395 			return -ERESTARTSYS;
396 
397 		mutex_lock(&file->mutex);
398 	}
399 
400 	if (file->agents_dead) {
401 		mutex_unlock(&file->mutex);
402 		return -EIO;
403 	}
404 
405 	packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
406 	list_del(&packet->list);
407 
408 	mutex_unlock(&file->mutex);
409 
410 	if (packet->recv_wc)
411 		ret = copy_recv_mad(file, buf, packet, count);
412 	else
413 		ret = copy_send_mad(file, buf, packet, count);
414 
415 	if (ret < 0) {
416 		/* Requeue packet */
417 		mutex_lock(&file->mutex);
418 		list_add(&packet->list, &file->recv_list);
419 		mutex_unlock(&file->mutex);
420 	} else {
421 		if (packet->recv_wc)
422 			ib_free_recv_mad(packet->recv_wc);
423 		kfree(packet);
424 	}
425 	return ret;
426 }
427 
428 static int copy_rmpp_mad(struct ib_mad_send_buf *msg, const char __user *buf)
429 {
430 	int left, seg;
431 
432 	/* Copy class specific header */
433 	if ((msg->hdr_len > IB_MGMT_RMPP_HDR) &&
434 	    copy_from_user(msg->mad + IB_MGMT_RMPP_HDR, buf + IB_MGMT_RMPP_HDR,
435 			   msg->hdr_len - IB_MGMT_RMPP_HDR))
436 		return -EFAULT;
437 
438 	/* All headers are in place.  Copy data segments. */
439 	for (seg = 1, left = msg->data_len, buf += msg->hdr_len; left > 0;
440 	     seg++, left -= msg->seg_size, buf += msg->seg_size) {
441 		if (copy_from_user(ib_get_rmpp_segment(msg, seg), buf,
442 				   min(left, msg->seg_size)))
443 			return -EFAULT;
444 	}
445 	return 0;
446 }
447 
448 static int same_destination(struct ib_user_mad_hdr *hdr1,
449 			    struct ib_user_mad_hdr *hdr2)
450 {
451 	if (!hdr1->grh_present && !hdr2->grh_present)
452 	   return (hdr1->lid == hdr2->lid);
453 
454 	if (hdr1->grh_present && hdr2->grh_present)
455 	   return !memcmp(hdr1->gid, hdr2->gid, 16);
456 
457 	return 0;
458 }
459 
460 static int is_duplicate(struct ib_umad_file *file,
461 			struct ib_umad_packet *packet)
462 {
463 	struct ib_umad_packet *sent_packet;
464 	struct ib_mad_hdr *sent_hdr, *hdr;
465 
466 	hdr = (struct ib_mad_hdr *) packet->mad.data;
467 	list_for_each_entry(sent_packet, &file->send_list, list) {
468 		sent_hdr = (struct ib_mad_hdr *) sent_packet->mad.data;
469 
470 		if ((hdr->tid != sent_hdr->tid) ||
471 		    (hdr->mgmt_class != sent_hdr->mgmt_class))
472 			continue;
473 
474 		/*
475 		 * No need to be overly clever here.  If two new operations have
476 		 * the same TID, reject the second as a duplicate.  This is more
477 		 * restrictive than required by the spec.
478 		 */
479 		if (!ib_response_mad(hdr)) {
480 			if (!ib_response_mad(sent_hdr))
481 				return 1;
482 			continue;
483 		} else if (!ib_response_mad(sent_hdr))
484 			continue;
485 
486 		if (same_destination(&packet->mad.hdr, &sent_packet->mad.hdr))
487 			return 1;
488 	}
489 
490 	return 0;
491 }
492 
493 static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
494 			     size_t count, loff_t *pos)
495 {
496 	struct ib_umad_file *file = filp->private_data;
497 	struct ib_umad_packet *packet;
498 	struct ib_mad_agent *agent;
499 	struct rdma_ah_attr ah_attr;
500 	struct ib_ah *ah;
501 	struct ib_rmpp_mad *rmpp_mad;
502 	__be64 *tid;
503 	int ret, data_len, hdr_len, copy_offset, rmpp_active;
504 	u8 base_version;
505 
506 	if (count < hdr_size(file) + IB_MGMT_RMPP_HDR)
507 		return -EINVAL;
508 
509 	packet = kzalloc(sizeof *packet + IB_MGMT_RMPP_HDR, GFP_KERNEL);
510 	if (!packet)
511 		return -ENOMEM;
512 
513 	if (copy_from_user(&packet->mad, buf, hdr_size(file))) {
514 		ret = -EFAULT;
515 		goto err;
516 	}
517 
518 	if (packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
519 		ret = -EINVAL;
520 		goto err;
521 	}
522 
523 	buf += hdr_size(file);
524 
525 	if (copy_from_user(packet->mad.data, buf, IB_MGMT_RMPP_HDR)) {
526 		ret = -EFAULT;
527 		goto err;
528 	}
529 
530 	mutex_lock(&file->mutex);
531 
532 	trace_ib_umad_write(file, &packet->mad.hdr,
533 			    (struct ib_mad_hdr *)&packet->mad.data);
534 
535 	agent = __get_agent(file, packet->mad.hdr.id);
536 	if (!agent) {
537 		ret = -EIO;
538 		goto err_up;
539 	}
540 
541 	memset(&ah_attr, 0, sizeof ah_attr);
542 	ah_attr.type = rdma_ah_find_type(agent->device,
543 					 file->port->port_num);
544 	rdma_ah_set_dlid(&ah_attr, be16_to_cpu(packet->mad.hdr.lid));
545 	rdma_ah_set_sl(&ah_attr, packet->mad.hdr.sl);
546 	rdma_ah_set_path_bits(&ah_attr, packet->mad.hdr.path_bits);
547 	rdma_ah_set_port_num(&ah_attr, file->port->port_num);
548 	if (packet->mad.hdr.grh_present) {
549 		rdma_ah_set_grh(&ah_attr, NULL,
550 				be32_to_cpu(packet->mad.hdr.flow_label),
551 				packet->mad.hdr.gid_index,
552 				packet->mad.hdr.hop_limit,
553 				packet->mad.hdr.traffic_class);
554 		rdma_ah_set_dgid_raw(&ah_attr, packet->mad.hdr.gid);
555 	}
556 
557 	ah = rdma_create_user_ah(agent->qp->pd, &ah_attr, NULL);
558 	if (IS_ERR(ah)) {
559 		ret = PTR_ERR(ah);
560 		goto err_up;
561 	}
562 
563 	rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
564 	hdr_len = ib_get_mad_data_offset(rmpp_mad->mad_hdr.mgmt_class);
565 
566 	if (ib_is_mad_class_rmpp(rmpp_mad->mad_hdr.mgmt_class)
567 	    && ib_mad_kernel_rmpp_agent(agent)) {
568 		copy_offset = IB_MGMT_RMPP_HDR;
569 		rmpp_active = ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
570 						IB_MGMT_RMPP_FLAG_ACTIVE;
571 	} else {
572 		copy_offset = IB_MGMT_MAD_HDR;
573 		rmpp_active = 0;
574 	}
575 
576 	base_version = ((struct ib_mad_hdr *)&packet->mad.data)->base_version;
577 	data_len = count - hdr_size(file) - hdr_len;
578 	packet->msg = ib_create_send_mad(agent,
579 					 be32_to_cpu(packet->mad.hdr.qpn),
580 					 packet->mad.hdr.pkey_index, rmpp_active,
581 					 hdr_len, data_len, GFP_KERNEL,
582 					 base_version);
583 	if (IS_ERR(packet->msg)) {
584 		ret = PTR_ERR(packet->msg);
585 		goto err_ah;
586 	}
587 
588 	packet->msg->ah		= ah;
589 	packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
590 	packet->msg->retries	= packet->mad.hdr.retries;
591 	packet->msg->context[0] = packet;
592 
593 	/* Copy MAD header.  Any RMPP header is already in place. */
594 	memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
595 
596 	if (!rmpp_active) {
597 		if (copy_from_user(packet->msg->mad + copy_offset,
598 				   buf + copy_offset,
599 				   hdr_len + data_len - copy_offset)) {
600 			ret = -EFAULT;
601 			goto err_msg;
602 		}
603 	} else {
604 		ret = copy_rmpp_mad(packet->msg, buf);
605 		if (ret)
606 			goto err_msg;
607 	}
608 
609 	/*
610 	 * Set the high-order part of the transaction ID to make MADs from
611 	 * different agents unique, and allow routing responses back to the
612 	 * original requestor.
613 	 */
614 	if (!ib_response_mad(packet->msg->mad)) {
615 		tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
616 		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
617 				   (be64_to_cpup(tid) & 0xffffffff));
618 		rmpp_mad->mad_hdr.tid = *tid;
619 	}
620 
621 	if (!ib_mad_kernel_rmpp_agent(agent)
622 	   && ib_is_mad_class_rmpp(rmpp_mad->mad_hdr.mgmt_class)
623 	   && (ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) & IB_MGMT_RMPP_FLAG_ACTIVE)) {
624 		spin_lock_irq(&file->send_lock);
625 		list_add_tail(&packet->list, &file->send_list);
626 		spin_unlock_irq(&file->send_lock);
627 	} else {
628 		spin_lock_irq(&file->send_lock);
629 		ret = is_duplicate(file, packet);
630 		if (!ret)
631 			list_add_tail(&packet->list, &file->send_list);
632 		spin_unlock_irq(&file->send_lock);
633 		if (ret) {
634 			ret = -EINVAL;
635 			goto err_msg;
636 		}
637 	}
638 
639 	ret = ib_post_send_mad(packet->msg, NULL);
640 	if (ret)
641 		goto err_send;
642 
643 	mutex_unlock(&file->mutex);
644 	return count;
645 
646 err_send:
647 	dequeue_send(file, packet);
648 err_msg:
649 	ib_free_send_mad(packet->msg);
650 err_ah:
651 	rdma_destroy_ah(ah, RDMA_DESTROY_AH_SLEEPABLE);
652 err_up:
653 	mutex_unlock(&file->mutex);
654 err:
655 	kfree(packet);
656 	return ret;
657 }
658 
659 static __poll_t ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
660 {
661 	struct ib_umad_file *file = filp->private_data;
662 
663 	/* we will always be able to post a MAD send */
664 	__poll_t mask = EPOLLOUT | EPOLLWRNORM;
665 
666 	mutex_lock(&file->mutex);
667 	poll_wait(filp, &file->recv_wait, wait);
668 
669 	if (!list_empty(&file->recv_list))
670 		mask |= EPOLLIN | EPOLLRDNORM;
671 	if (file->agents_dead)
672 		mask = EPOLLERR;
673 	mutex_unlock(&file->mutex);
674 
675 	return mask;
676 }
677 
678 static int ib_umad_reg_agent(struct ib_umad_file *file, void __user *arg,
679 			     int compat_method_mask)
680 {
681 	struct ib_user_mad_reg_req ureq;
682 	struct ib_mad_reg_req req;
683 	struct ib_mad_agent *agent = NULL;
684 	int agent_id;
685 	int ret;
686 
687 	mutex_lock(&file->port->file_mutex);
688 	mutex_lock(&file->mutex);
689 
690 	if (!file->port->ib_dev) {
691 		dev_notice(&file->port->dev, "%s: invalid device\n", __func__);
692 		ret = -EPIPE;
693 		goto out;
694 	}
695 
696 	if (copy_from_user(&ureq, arg, sizeof ureq)) {
697 		ret = -EFAULT;
698 		goto out;
699 	}
700 
701 	if (ureq.qpn != 0 && ureq.qpn != 1) {
702 		dev_notice(&file->port->dev,
703 			   "%s: invalid QPN %d specified\n", __func__,
704 			   ureq.qpn);
705 		ret = -EINVAL;
706 		goto out;
707 	}
708 
709 	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
710 		if (!__get_agent(file, agent_id))
711 			goto found;
712 
713 	dev_notice(&file->port->dev, "%s: Max Agents (%u) reached\n", __func__,
714 		   IB_UMAD_MAX_AGENTS);
715 
716 	ret = -ENOMEM;
717 	goto out;
718 
719 found:
720 	if (ureq.mgmt_class) {
721 		memset(&req, 0, sizeof(req));
722 		req.mgmt_class         = ureq.mgmt_class;
723 		req.mgmt_class_version = ureq.mgmt_class_version;
724 		memcpy(req.oui, ureq.oui, sizeof req.oui);
725 
726 		if (compat_method_mask) {
727 			u32 *umm = (u32 *) ureq.method_mask;
728 			int i;
729 
730 			for (i = 0; i < BITS_TO_LONGS(IB_MGMT_MAX_METHODS); ++i)
731 				req.method_mask[i] =
732 					umm[i * 2] | ((u64) umm[i * 2 + 1] << 32);
733 		} else
734 			memcpy(req.method_mask, ureq.method_mask,
735 			       sizeof req.method_mask);
736 	}
737 
738 	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
739 				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
740 				      ureq.mgmt_class ? &req : NULL,
741 				      ureq.rmpp_version,
742 				      send_handler, recv_handler, file, 0);
743 	if (IS_ERR(agent)) {
744 		ret = PTR_ERR(agent);
745 		agent = NULL;
746 		goto out;
747 	}
748 
749 	if (put_user(agent_id,
750 		     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
751 		ret = -EFAULT;
752 		goto out;
753 	}
754 
755 	if (!file->already_used) {
756 		file->already_used = 1;
757 		if (!file->use_pkey_index) {
758 			dev_warn(&file->port->dev,
759 				"process %s did not enable P_Key index support.\n",
760 				current->comm);
761 			dev_warn(&file->port->dev,
762 				"   Documentation/infiniband/user_mad.rst has info on the new ABI.\n");
763 		}
764 	}
765 
766 	file->agent[agent_id] = agent;
767 	ret = 0;
768 
769 out:
770 	mutex_unlock(&file->mutex);
771 
772 	if (ret && agent)
773 		ib_unregister_mad_agent(agent);
774 
775 	mutex_unlock(&file->port->file_mutex);
776 
777 	return ret;
778 }
779 
780 static int ib_umad_reg_agent2(struct ib_umad_file *file, void __user *arg)
781 {
782 	struct ib_user_mad_reg_req2 ureq;
783 	struct ib_mad_reg_req req;
784 	struct ib_mad_agent *agent = NULL;
785 	int agent_id;
786 	int ret;
787 
788 	mutex_lock(&file->port->file_mutex);
789 	mutex_lock(&file->mutex);
790 
791 	if (!file->port->ib_dev) {
792 		dev_notice(&file->port->dev, "%s: invalid device\n", __func__);
793 		ret = -EPIPE;
794 		goto out;
795 	}
796 
797 	if (copy_from_user(&ureq, arg, sizeof(ureq))) {
798 		ret = -EFAULT;
799 		goto out;
800 	}
801 
802 	if (ureq.qpn != 0 && ureq.qpn != 1) {
803 		dev_notice(&file->port->dev, "%s: invalid QPN %d specified\n",
804 			   __func__, ureq.qpn);
805 		ret = -EINVAL;
806 		goto out;
807 	}
808 
809 	if (ureq.flags & ~IB_USER_MAD_REG_FLAGS_CAP) {
810 		dev_notice(&file->port->dev,
811 			   "%s failed: invalid registration flags specified 0x%x; supported 0x%x\n",
812 			   __func__, ureq.flags, IB_USER_MAD_REG_FLAGS_CAP);
813 		ret = -EINVAL;
814 
815 		if (put_user((u32)IB_USER_MAD_REG_FLAGS_CAP,
816 				(u32 __user *) (arg + offsetof(struct
817 				ib_user_mad_reg_req2, flags))))
818 			ret = -EFAULT;
819 
820 		goto out;
821 	}
822 
823 	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
824 		if (!__get_agent(file, agent_id))
825 			goto found;
826 
827 	dev_notice(&file->port->dev, "%s: Max Agents (%u) reached\n", __func__,
828 		   IB_UMAD_MAX_AGENTS);
829 	ret = -ENOMEM;
830 	goto out;
831 
832 found:
833 	if (ureq.mgmt_class) {
834 		memset(&req, 0, sizeof(req));
835 		req.mgmt_class         = ureq.mgmt_class;
836 		req.mgmt_class_version = ureq.mgmt_class_version;
837 		if (ureq.oui & 0xff000000) {
838 			dev_notice(&file->port->dev,
839 				   "%s failed: oui invalid 0x%08x\n", __func__,
840 				   ureq.oui);
841 			ret = -EINVAL;
842 			goto out;
843 		}
844 		req.oui[2] =  ureq.oui & 0x0000ff;
845 		req.oui[1] = (ureq.oui & 0x00ff00) >> 8;
846 		req.oui[0] = (ureq.oui & 0xff0000) >> 16;
847 		memcpy(req.method_mask, ureq.method_mask,
848 			sizeof(req.method_mask));
849 	}
850 
851 	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
852 				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
853 				      ureq.mgmt_class ? &req : NULL,
854 				      ureq.rmpp_version,
855 				      send_handler, recv_handler, file,
856 				      ureq.flags);
857 	if (IS_ERR(agent)) {
858 		ret = PTR_ERR(agent);
859 		agent = NULL;
860 		goto out;
861 	}
862 
863 	if (put_user(agent_id,
864 		     (u32 __user *)(arg +
865 				offsetof(struct ib_user_mad_reg_req2, id)))) {
866 		ret = -EFAULT;
867 		goto out;
868 	}
869 
870 	if (!file->already_used) {
871 		file->already_used = 1;
872 		file->use_pkey_index = 1;
873 	}
874 
875 	file->agent[agent_id] = agent;
876 	ret = 0;
877 
878 out:
879 	mutex_unlock(&file->mutex);
880 
881 	if (ret && agent)
882 		ib_unregister_mad_agent(agent);
883 
884 	mutex_unlock(&file->port->file_mutex);
885 
886 	return ret;
887 }
888 
889 
890 static int ib_umad_unreg_agent(struct ib_umad_file *file, u32 __user *arg)
891 {
892 	struct ib_mad_agent *agent = NULL;
893 	u32 id;
894 	int ret = 0;
895 
896 	if (get_user(id, arg))
897 		return -EFAULT;
898 	if (id >= IB_UMAD_MAX_AGENTS)
899 		return -EINVAL;
900 
901 	mutex_lock(&file->port->file_mutex);
902 	mutex_lock(&file->mutex);
903 
904 	id = array_index_nospec(id, IB_UMAD_MAX_AGENTS);
905 	if (!__get_agent(file, id)) {
906 		ret = -EINVAL;
907 		goto out;
908 	}
909 
910 	agent = file->agent[id];
911 	file->agent[id] = NULL;
912 
913 out:
914 	mutex_unlock(&file->mutex);
915 
916 	if (agent)
917 		ib_unregister_mad_agent(agent);
918 
919 	mutex_unlock(&file->port->file_mutex);
920 
921 	return ret;
922 }
923 
924 static long ib_umad_enable_pkey(struct ib_umad_file *file)
925 {
926 	int ret = 0;
927 
928 	mutex_lock(&file->mutex);
929 	if (file->already_used)
930 		ret = -EINVAL;
931 	else
932 		file->use_pkey_index = 1;
933 	mutex_unlock(&file->mutex);
934 
935 	return ret;
936 }
937 
938 static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
939 			  unsigned long arg)
940 {
941 	switch (cmd) {
942 	case IB_USER_MAD_REGISTER_AGENT:
943 		return ib_umad_reg_agent(filp->private_data, (void __user *) arg, 0);
944 	case IB_USER_MAD_UNREGISTER_AGENT:
945 		return ib_umad_unreg_agent(filp->private_data, (__u32 __user *) arg);
946 	case IB_USER_MAD_ENABLE_PKEY:
947 		return ib_umad_enable_pkey(filp->private_data);
948 	case IB_USER_MAD_REGISTER_AGENT2:
949 		return ib_umad_reg_agent2(filp->private_data, (void __user *) arg);
950 	default:
951 		return -ENOIOCTLCMD;
952 	}
953 }
954 
955 #ifdef CONFIG_COMPAT
956 static long ib_umad_compat_ioctl(struct file *filp, unsigned int cmd,
957 				 unsigned long arg)
958 {
959 	switch (cmd) {
960 	case IB_USER_MAD_REGISTER_AGENT:
961 		return ib_umad_reg_agent(filp->private_data, compat_ptr(arg), 1);
962 	case IB_USER_MAD_UNREGISTER_AGENT:
963 		return ib_umad_unreg_agent(filp->private_data, compat_ptr(arg));
964 	case IB_USER_MAD_ENABLE_PKEY:
965 		return ib_umad_enable_pkey(filp->private_data);
966 	case IB_USER_MAD_REGISTER_AGENT2:
967 		return ib_umad_reg_agent2(filp->private_data, compat_ptr(arg));
968 	default:
969 		return -ENOIOCTLCMD;
970 	}
971 }
972 #endif
973 
974 /*
975  * ib_umad_open() does not need the BKL:
976  *
977  *  - the ib_umad_port structures are properly reference counted, and
978  *    everything else is purely local to the file being created, so
979  *    races against other open calls are not a problem;
980  *  - the ioctl method does not affect any global state outside of the
981  *    file structure being operated on;
982  */
983 static int ib_umad_open(struct inode *inode, struct file *filp)
984 {
985 	struct ib_umad_port *port;
986 	struct ib_umad_file *file;
987 	int ret = 0;
988 
989 	port = container_of(inode->i_cdev, struct ib_umad_port, cdev);
990 
991 	mutex_lock(&port->file_mutex);
992 
993 	if (!port->ib_dev) {
994 		ret = -ENXIO;
995 		goto out;
996 	}
997 
998 	if (!rdma_dev_access_netns(port->ib_dev, current->nsproxy->net_ns)) {
999 		ret = -EPERM;
1000 		goto out;
1001 	}
1002 
1003 	file = kzalloc(sizeof(*file), GFP_KERNEL);
1004 	if (!file) {
1005 		ret = -ENOMEM;
1006 		goto out;
1007 	}
1008 
1009 	mutex_init(&file->mutex);
1010 	spin_lock_init(&file->send_lock);
1011 	INIT_LIST_HEAD(&file->recv_list);
1012 	INIT_LIST_HEAD(&file->send_list);
1013 	init_waitqueue_head(&file->recv_wait);
1014 
1015 	file->port = port;
1016 	filp->private_data = file;
1017 
1018 	list_add_tail(&file->port_list, &port->file_list);
1019 
1020 	stream_open(inode, filp);
1021 out:
1022 	mutex_unlock(&port->file_mutex);
1023 	return ret;
1024 }
1025 
1026 static int ib_umad_close(struct inode *inode, struct file *filp)
1027 {
1028 	struct ib_umad_file *file = filp->private_data;
1029 	struct ib_umad_packet *packet, *tmp;
1030 	int already_dead;
1031 	int i;
1032 
1033 	mutex_lock(&file->port->file_mutex);
1034 	mutex_lock(&file->mutex);
1035 
1036 	already_dead = file->agents_dead;
1037 	file->agents_dead = 1;
1038 
1039 	list_for_each_entry_safe(packet, tmp, &file->recv_list, list) {
1040 		if (packet->recv_wc)
1041 			ib_free_recv_mad(packet->recv_wc);
1042 		kfree(packet);
1043 	}
1044 
1045 	list_del(&file->port_list);
1046 
1047 	mutex_unlock(&file->mutex);
1048 
1049 	if (!already_dead)
1050 		for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
1051 			if (file->agent[i])
1052 				ib_unregister_mad_agent(file->agent[i]);
1053 
1054 	mutex_unlock(&file->port->file_mutex);
1055 	mutex_destroy(&file->mutex);
1056 	kfree(file);
1057 	return 0;
1058 }
1059 
1060 static const struct file_operations umad_fops = {
1061 	.owner		= THIS_MODULE,
1062 	.read		= ib_umad_read,
1063 	.write		= ib_umad_write,
1064 	.poll		= ib_umad_poll,
1065 	.unlocked_ioctl = ib_umad_ioctl,
1066 #ifdef CONFIG_COMPAT
1067 	.compat_ioctl	= ib_umad_compat_ioctl,
1068 #endif
1069 	.open		= ib_umad_open,
1070 	.release	= ib_umad_close,
1071 	.llseek		= no_llseek,
1072 };
1073 
1074 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
1075 {
1076 	struct ib_umad_port *port;
1077 	struct ib_port_modify props = {
1078 		.set_port_cap_mask = IB_PORT_SM
1079 	};
1080 	int ret;
1081 
1082 	port = container_of(inode->i_cdev, struct ib_umad_port, sm_cdev);
1083 
1084 	if (filp->f_flags & O_NONBLOCK) {
1085 		if (down_trylock(&port->sm_sem)) {
1086 			ret = -EAGAIN;
1087 			goto fail;
1088 		}
1089 	} else {
1090 		if (down_interruptible(&port->sm_sem)) {
1091 			ret = -ERESTARTSYS;
1092 			goto fail;
1093 		}
1094 	}
1095 
1096 	if (!rdma_dev_access_netns(port->ib_dev, current->nsproxy->net_ns)) {
1097 		ret = -EPERM;
1098 		goto err_up_sem;
1099 	}
1100 
1101 	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
1102 	if (ret)
1103 		goto err_up_sem;
1104 
1105 	filp->private_data = port;
1106 
1107 	nonseekable_open(inode, filp);
1108 	return 0;
1109 
1110 err_up_sem:
1111 	up(&port->sm_sem);
1112 
1113 fail:
1114 	return ret;
1115 }
1116 
1117 static int ib_umad_sm_close(struct inode *inode, struct file *filp)
1118 {
1119 	struct ib_umad_port *port = filp->private_data;
1120 	struct ib_port_modify props = {
1121 		.clr_port_cap_mask = IB_PORT_SM
1122 	};
1123 	int ret = 0;
1124 
1125 	mutex_lock(&port->file_mutex);
1126 	if (port->ib_dev)
1127 		ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
1128 	mutex_unlock(&port->file_mutex);
1129 
1130 	up(&port->sm_sem);
1131 
1132 	return ret;
1133 }
1134 
1135 static const struct file_operations umad_sm_fops = {
1136 	.owner	 = THIS_MODULE,
1137 	.open	 = ib_umad_sm_open,
1138 	.release = ib_umad_sm_close,
1139 	.llseek	 = no_llseek,
1140 };
1141 
1142 static struct ib_umad_port *get_port(struct ib_device *ibdev,
1143 				     struct ib_umad_device *umad_dev,
1144 				     u32 port)
1145 {
1146 	if (!umad_dev)
1147 		return ERR_PTR(-EOPNOTSUPP);
1148 	if (!rdma_is_port_valid(ibdev, port))
1149 		return ERR_PTR(-EINVAL);
1150 	if (!rdma_cap_ib_mad(ibdev, port))
1151 		return ERR_PTR(-EOPNOTSUPP);
1152 
1153 	return &umad_dev->ports[port - rdma_start_port(ibdev)];
1154 }
1155 
1156 static int ib_umad_get_nl_info(struct ib_device *ibdev, void *client_data,
1157 			       struct ib_client_nl_info *res)
1158 {
1159 	struct ib_umad_port *port = get_port(ibdev, client_data, res->port);
1160 
1161 	if (IS_ERR(port))
1162 		return PTR_ERR(port);
1163 
1164 	res->abi = IB_USER_MAD_ABI_VERSION;
1165 	res->cdev = &port->dev;
1166 	return 0;
1167 }
1168 
1169 static struct ib_client umad_client = {
1170 	.name   = "umad",
1171 	.add    = ib_umad_add_one,
1172 	.remove = ib_umad_remove_one,
1173 	.get_nl_info = ib_umad_get_nl_info,
1174 };
1175 MODULE_ALIAS_RDMA_CLIENT("umad");
1176 
1177 static int ib_issm_get_nl_info(struct ib_device *ibdev, void *client_data,
1178 			       struct ib_client_nl_info *res)
1179 {
1180 	struct ib_umad_port *port = get_port(ibdev, client_data, res->port);
1181 
1182 	if (IS_ERR(port))
1183 		return PTR_ERR(port);
1184 
1185 	res->abi = IB_USER_MAD_ABI_VERSION;
1186 	res->cdev = &port->sm_dev;
1187 	return 0;
1188 }
1189 
1190 static struct ib_client issm_client = {
1191 	.name = "issm",
1192 	.get_nl_info = ib_issm_get_nl_info,
1193 };
1194 MODULE_ALIAS_RDMA_CLIENT("issm");
1195 
1196 static ssize_t ibdev_show(struct device *dev, struct device_attribute *attr,
1197 			  char *buf)
1198 {
1199 	struct ib_umad_port *port = dev_get_drvdata(dev);
1200 
1201 	if (!port)
1202 		return -ENODEV;
1203 
1204 	return sysfs_emit(buf, "%s\n", dev_name(&port->ib_dev->dev));
1205 }
1206 static DEVICE_ATTR_RO(ibdev);
1207 
1208 static ssize_t port_show(struct device *dev, struct device_attribute *attr,
1209 			 char *buf)
1210 {
1211 	struct ib_umad_port *port = dev_get_drvdata(dev);
1212 
1213 	if (!port)
1214 		return -ENODEV;
1215 
1216 	return sysfs_emit(buf, "%d\n", port->port_num);
1217 }
1218 static DEVICE_ATTR_RO(port);
1219 
1220 static struct attribute *umad_class_dev_attrs[] = {
1221 	&dev_attr_ibdev.attr,
1222 	&dev_attr_port.attr,
1223 	NULL,
1224 };
1225 ATTRIBUTE_GROUPS(umad_class_dev);
1226 
1227 static char *umad_devnode(struct device *dev, umode_t *mode)
1228 {
1229 	return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1230 }
1231 
1232 static ssize_t abi_version_show(struct class *class,
1233 				struct class_attribute *attr, char *buf)
1234 {
1235 	return sysfs_emit(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
1236 }
1237 static CLASS_ATTR_RO(abi_version);
1238 
1239 static struct attribute *umad_class_attrs[] = {
1240 	&class_attr_abi_version.attr,
1241 	NULL,
1242 };
1243 ATTRIBUTE_GROUPS(umad_class);
1244 
1245 static struct class umad_class = {
1246 	.name		= "infiniband_mad",
1247 	.devnode	= umad_devnode,
1248 	.class_groups	= umad_class_groups,
1249 	.dev_groups	= umad_class_dev_groups,
1250 };
1251 
1252 static void ib_umad_release_port(struct device *device)
1253 {
1254 	struct ib_umad_port *port = dev_get_drvdata(device);
1255 	struct ib_umad_device *umad_dev = port->umad_dev;
1256 
1257 	ib_umad_dev_put(umad_dev);
1258 }
1259 
1260 static void ib_umad_init_port_dev(struct device *dev,
1261 				  struct ib_umad_port *port,
1262 				  const struct ib_device *device)
1263 {
1264 	device_initialize(dev);
1265 	ib_umad_dev_get(port->umad_dev);
1266 	dev->class = &umad_class;
1267 	dev->parent = device->dev.parent;
1268 	dev_set_drvdata(dev, port);
1269 	dev->release = ib_umad_release_port;
1270 }
1271 
1272 static int ib_umad_init_port(struct ib_device *device, int port_num,
1273 			     struct ib_umad_device *umad_dev,
1274 			     struct ib_umad_port *port)
1275 {
1276 	int devnum;
1277 	dev_t base_umad;
1278 	dev_t base_issm;
1279 	int ret;
1280 
1281 	devnum = ida_alloc_max(&umad_ida, IB_UMAD_MAX_PORTS - 1, GFP_KERNEL);
1282 	if (devnum < 0)
1283 		return -1;
1284 	port->dev_num = devnum;
1285 	if (devnum >= IB_UMAD_NUM_FIXED_MINOR) {
1286 		base_umad = dynamic_umad_dev + devnum - IB_UMAD_NUM_FIXED_MINOR;
1287 		base_issm = dynamic_issm_dev + devnum - IB_UMAD_NUM_FIXED_MINOR;
1288 	} else {
1289 		base_umad = devnum + base_umad_dev;
1290 		base_issm = devnum + base_issm_dev;
1291 	}
1292 
1293 	port->ib_dev   = device;
1294 	port->umad_dev = umad_dev;
1295 	port->port_num = port_num;
1296 	sema_init(&port->sm_sem, 1);
1297 	mutex_init(&port->file_mutex);
1298 	INIT_LIST_HEAD(&port->file_list);
1299 
1300 	ib_umad_init_port_dev(&port->dev, port, device);
1301 	port->dev.devt = base_umad;
1302 	dev_set_name(&port->dev, "umad%d", port->dev_num);
1303 	cdev_init(&port->cdev, &umad_fops);
1304 	port->cdev.owner = THIS_MODULE;
1305 
1306 	ret = cdev_device_add(&port->cdev, &port->dev);
1307 	if (ret)
1308 		goto err_cdev;
1309 
1310 	ib_umad_init_port_dev(&port->sm_dev, port, device);
1311 	port->sm_dev.devt = base_issm;
1312 	dev_set_name(&port->sm_dev, "issm%d", port->dev_num);
1313 	cdev_init(&port->sm_cdev, &umad_sm_fops);
1314 	port->sm_cdev.owner = THIS_MODULE;
1315 
1316 	ret = cdev_device_add(&port->sm_cdev, &port->sm_dev);
1317 	if (ret)
1318 		goto err_dev;
1319 
1320 	return 0;
1321 
1322 err_dev:
1323 	put_device(&port->sm_dev);
1324 	cdev_device_del(&port->cdev, &port->dev);
1325 err_cdev:
1326 	put_device(&port->dev);
1327 	ida_free(&umad_ida, devnum);
1328 	return ret;
1329 }
1330 
1331 static void ib_umad_kill_port(struct ib_umad_port *port)
1332 {
1333 	struct ib_umad_file *file;
1334 	int id;
1335 
1336 	cdev_device_del(&port->sm_cdev, &port->sm_dev);
1337 	cdev_device_del(&port->cdev, &port->dev);
1338 
1339 	mutex_lock(&port->file_mutex);
1340 
1341 	/* Mark ib_dev NULL and block ioctl or other file ops to progress
1342 	 * further.
1343 	 */
1344 	port->ib_dev = NULL;
1345 
1346 	list_for_each_entry(file, &port->file_list, port_list) {
1347 		mutex_lock(&file->mutex);
1348 		file->agents_dead = 1;
1349 		wake_up_interruptible(&file->recv_wait);
1350 		mutex_unlock(&file->mutex);
1351 
1352 		for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id)
1353 			if (file->agent[id])
1354 				ib_unregister_mad_agent(file->agent[id]);
1355 	}
1356 
1357 	mutex_unlock(&port->file_mutex);
1358 
1359 	ida_free(&umad_ida, port->dev_num);
1360 
1361 	/* balances device_initialize() */
1362 	put_device(&port->sm_dev);
1363 	put_device(&port->dev);
1364 }
1365 
1366 static int ib_umad_add_one(struct ib_device *device)
1367 {
1368 	struct ib_umad_device *umad_dev;
1369 	int s, e, i;
1370 	int count = 0;
1371 	int ret;
1372 
1373 	s = rdma_start_port(device);
1374 	e = rdma_end_port(device);
1375 
1376 	umad_dev = kzalloc(struct_size(umad_dev, ports, e - s + 1), GFP_KERNEL);
1377 	if (!umad_dev)
1378 		return -ENOMEM;
1379 
1380 	kref_init(&umad_dev->kref);
1381 	for (i = s; i <= e; ++i) {
1382 		if (!rdma_cap_ib_mad(device, i))
1383 			continue;
1384 
1385 		ret = ib_umad_init_port(device, i, umad_dev,
1386 					&umad_dev->ports[i - s]);
1387 		if (ret)
1388 			goto err;
1389 
1390 		count++;
1391 	}
1392 
1393 	if (!count) {
1394 		ret = -EOPNOTSUPP;
1395 		goto free;
1396 	}
1397 
1398 	ib_set_client_data(device, &umad_client, umad_dev);
1399 
1400 	return 0;
1401 
1402 err:
1403 	while (--i >= s) {
1404 		if (!rdma_cap_ib_mad(device, i))
1405 			continue;
1406 
1407 		ib_umad_kill_port(&umad_dev->ports[i - s]);
1408 	}
1409 free:
1410 	/* balances kref_init */
1411 	ib_umad_dev_put(umad_dev);
1412 	return ret;
1413 }
1414 
1415 static void ib_umad_remove_one(struct ib_device *device, void *client_data)
1416 {
1417 	struct ib_umad_device *umad_dev = client_data;
1418 	unsigned int i;
1419 
1420 	rdma_for_each_port (device, i) {
1421 		if (rdma_cap_ib_mad(device, i))
1422 			ib_umad_kill_port(
1423 				&umad_dev->ports[i - rdma_start_port(device)]);
1424 	}
1425 	/* balances kref_init() */
1426 	ib_umad_dev_put(umad_dev);
1427 }
1428 
1429 static int __init ib_umad_init(void)
1430 {
1431 	int ret;
1432 
1433 	ret = register_chrdev_region(base_umad_dev,
1434 				     IB_UMAD_NUM_FIXED_MINOR * 2,
1435 				     umad_class.name);
1436 	if (ret) {
1437 		pr_err("couldn't register device number\n");
1438 		goto out;
1439 	}
1440 
1441 	ret = alloc_chrdev_region(&dynamic_umad_dev, 0,
1442 				  IB_UMAD_NUM_DYNAMIC_MINOR * 2,
1443 				  umad_class.name);
1444 	if (ret) {
1445 		pr_err("couldn't register dynamic device number\n");
1446 		goto out_alloc;
1447 	}
1448 	dynamic_issm_dev = dynamic_umad_dev + IB_UMAD_NUM_DYNAMIC_MINOR;
1449 
1450 	ret = class_register(&umad_class);
1451 	if (ret) {
1452 		pr_err("couldn't create class infiniband_mad\n");
1453 		goto out_chrdev;
1454 	}
1455 
1456 	ret = ib_register_client(&umad_client);
1457 	if (ret)
1458 		goto out_class;
1459 
1460 	ret = ib_register_client(&issm_client);
1461 	if (ret)
1462 		goto out_client;
1463 
1464 	return 0;
1465 
1466 out_client:
1467 	ib_unregister_client(&umad_client);
1468 out_class:
1469 	class_unregister(&umad_class);
1470 
1471 out_chrdev:
1472 	unregister_chrdev_region(dynamic_umad_dev,
1473 				 IB_UMAD_NUM_DYNAMIC_MINOR * 2);
1474 
1475 out_alloc:
1476 	unregister_chrdev_region(base_umad_dev,
1477 				 IB_UMAD_NUM_FIXED_MINOR * 2);
1478 
1479 out:
1480 	return ret;
1481 }
1482 
1483 static void __exit ib_umad_cleanup(void)
1484 {
1485 	ib_unregister_client(&issm_client);
1486 	ib_unregister_client(&umad_client);
1487 	class_unregister(&umad_class);
1488 	unregister_chrdev_region(base_umad_dev,
1489 				 IB_UMAD_NUM_FIXED_MINOR * 2);
1490 	unregister_chrdev_region(dynamic_umad_dev,
1491 				 IB_UMAD_NUM_DYNAMIC_MINOR * 2);
1492 }
1493 
1494 module_init(ib_umad_init);
1495 module_exit(ib_umad_cleanup);
1496