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 #include <linux/module.h>
37 #include <linux/init.h>
38 #include <linux/device.h>
39 #include <linux/err.h>
40 #include <linux/fs.h>
41 #include <linux/cdev.h>
42 #include <linux/dma-mapping.h>
43 #include <linux/poll.h>
44 #include <linux/mutex.h>
45 #include <linux/kref.h>
46 #include <linux/compat.h>
47 #include <linux/sched.h>
48 #include <linux/semaphore.h>
49 #include <linux/slab.h>
50 
51 #include <asm/uaccess.h>
52 
53 #include <rdma/ib_mad.h>
54 #include <rdma/ib_user_mad.h>
55 
56 MODULE_AUTHOR("Roland Dreier");
57 MODULE_DESCRIPTION("InfiniBand userspace MAD packet access");
58 MODULE_LICENSE("Dual BSD/GPL");
59 
60 enum {
61 	IB_UMAD_MAX_PORTS  = 64,
62 	IB_UMAD_MAX_AGENTS = 32,
63 
64 	IB_UMAD_MAJOR      = 231,
65 	IB_UMAD_MINOR_BASE = 0
66 };
67 
68 /*
69  * Our lifetime rules for these structs are the following:
70  * device special file is opened, we take a reference on the
71  * ib_umad_port's struct ib_umad_device. We drop these
72  * references in the corresponding close().
73  *
74  * In addition to references coming from open character devices, there
75  * is one more reference to each ib_umad_device representing the
76  * module's reference taken when allocating the ib_umad_device in
77  * ib_umad_add_one().
78  *
79  * When destroying an ib_umad_device, we drop the module's reference.
80  */
81 
82 struct ib_umad_port {
83 	struct cdev           cdev;
84 	struct device	      *dev;
85 
86 	struct cdev           sm_cdev;
87 	struct device	      *sm_dev;
88 	struct semaphore       sm_sem;
89 
90 	struct mutex	       file_mutex;
91 	struct list_head       file_list;
92 
93 	struct ib_device      *ib_dev;
94 	struct ib_umad_device *umad_dev;
95 	int                    dev_num;
96 	u8                     port_num;
97 };
98 
99 struct ib_umad_device {
100 	int                  start_port, end_port;
101 	struct kref          ref;
102 	struct ib_umad_port  port[0];
103 };
104 
105 struct ib_umad_file {
106 	struct mutex		mutex;
107 	struct ib_umad_port    *port;
108 	struct list_head	recv_list;
109 	struct list_head	send_list;
110 	struct list_head	port_list;
111 	spinlock_t		send_lock;
112 	wait_queue_head_t	recv_wait;
113 	struct ib_mad_agent    *agent[IB_UMAD_MAX_AGENTS];
114 	int			agents_dead;
115 	u8			use_pkey_index;
116 	u8			already_used;
117 };
118 
119 struct ib_umad_packet {
120 	struct ib_mad_send_buf *msg;
121 	struct ib_mad_recv_wc  *recv_wc;
122 	struct list_head   list;
123 	int		   length;
124 	struct ib_user_mad mad;
125 };
126 
127 static struct class *umad_class;
128 
129 static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
130 
131 static DEFINE_SPINLOCK(port_lock);
132 static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS);
133 
134 static void ib_umad_add_one(struct ib_device *device);
135 static void ib_umad_remove_one(struct ib_device *device);
136 
137 static void ib_umad_release_dev(struct kref *ref)
138 {
139 	struct ib_umad_device *dev =
140 		container_of(ref, struct ib_umad_device, ref);
141 
142 	kfree(dev);
143 }
144 
145 static int hdr_size(struct ib_umad_file *file)
146 {
147 	return file->use_pkey_index ? sizeof (struct ib_user_mad_hdr) :
148 		sizeof (struct ib_user_mad_hdr_old);
149 }
150 
151 /* caller must hold file->mutex */
152 static struct ib_mad_agent *__get_agent(struct ib_umad_file *file, int id)
153 {
154 	return file->agents_dead ? NULL : file->agent[id];
155 }
156 
157 static int queue_packet(struct ib_umad_file *file,
158 			struct ib_mad_agent *agent,
159 			struct ib_umad_packet *packet)
160 {
161 	int ret = 1;
162 
163 	mutex_lock(&file->mutex);
164 
165 	for (packet->mad.hdr.id = 0;
166 	     packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
167 	     packet->mad.hdr.id++)
168 		if (agent == __get_agent(file, packet->mad.hdr.id)) {
169 			list_add_tail(&packet->list, &file->recv_list);
170 			wake_up_interruptible(&file->recv_wait);
171 			ret = 0;
172 			break;
173 		}
174 
175 	mutex_unlock(&file->mutex);
176 
177 	return ret;
178 }
179 
180 static void dequeue_send(struct ib_umad_file *file,
181 			 struct ib_umad_packet *packet)
182 {
183 	spin_lock_irq(&file->send_lock);
184 	list_del(&packet->list);
185 	spin_unlock_irq(&file->send_lock);
186 }
187 
188 static void send_handler(struct ib_mad_agent *agent,
189 			 struct ib_mad_send_wc *send_wc)
190 {
191 	struct ib_umad_file *file = agent->context;
192 	struct ib_umad_packet *packet = send_wc->send_buf->context[0];
193 
194 	dequeue_send(file, packet);
195 	ib_destroy_ah(packet->msg->ah);
196 	ib_free_send_mad(packet->msg);
197 
198 	if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
199 		packet->length = IB_MGMT_MAD_HDR;
200 		packet->mad.hdr.status = ETIMEDOUT;
201 		if (!queue_packet(file, agent, packet))
202 			return;
203 	}
204 	kfree(packet);
205 }
206 
207 static void recv_handler(struct ib_mad_agent *agent,
208 			 struct ib_mad_recv_wc *mad_recv_wc)
209 {
210 	struct ib_umad_file *file = agent->context;
211 	struct ib_umad_packet *packet;
212 
213 	if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
214 		goto err1;
215 
216 	packet = kzalloc(sizeof *packet, GFP_KERNEL);
217 	if (!packet)
218 		goto err1;
219 
220 	packet->length = mad_recv_wc->mad_len;
221 	packet->recv_wc = mad_recv_wc;
222 
223 	packet->mad.hdr.status	   = 0;
224 	packet->mad.hdr.length	   = hdr_size(file) + mad_recv_wc->mad_len;
225 	packet->mad.hdr.qpn	   = cpu_to_be32(mad_recv_wc->wc->src_qp);
226 	packet->mad.hdr.lid	   = cpu_to_be16(mad_recv_wc->wc->slid);
227 	packet->mad.hdr.sl	   = mad_recv_wc->wc->sl;
228 	packet->mad.hdr.path_bits  = mad_recv_wc->wc->dlid_path_bits;
229 	packet->mad.hdr.pkey_index = mad_recv_wc->wc->pkey_index;
230 	packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
231 	if (packet->mad.hdr.grh_present) {
232 		struct ib_ah_attr ah_attr;
233 
234 		ib_init_ah_from_wc(agent->device, agent->port_num,
235 				   mad_recv_wc->wc, mad_recv_wc->recv_buf.grh,
236 				   &ah_attr);
237 
238 		packet->mad.hdr.gid_index = ah_attr.grh.sgid_index;
239 		packet->mad.hdr.hop_limit = ah_attr.grh.hop_limit;
240 		packet->mad.hdr.traffic_class = ah_attr.grh.traffic_class;
241 		memcpy(packet->mad.hdr.gid, &ah_attr.grh.dgid, 16);
242 		packet->mad.hdr.flow_label = cpu_to_be32(ah_attr.grh.flow_label);
243 	}
244 
245 	if (queue_packet(file, agent, packet))
246 		goto err2;
247 	return;
248 
249 err2:
250 	kfree(packet);
251 err1:
252 	ib_free_recv_mad(mad_recv_wc);
253 }
254 
255 static ssize_t copy_recv_mad(struct ib_umad_file *file, char __user *buf,
256 			     struct ib_umad_packet *packet, size_t count)
257 {
258 	struct ib_mad_recv_buf *recv_buf;
259 	int left, seg_payload, offset, max_seg_payload;
260 
261 	/* We need enough room to copy the first (or only) MAD segment. */
262 	recv_buf = &packet->recv_wc->recv_buf;
263 	if ((packet->length <= sizeof (*recv_buf->mad) &&
264 	     count < hdr_size(file) + packet->length) ||
265 	    (packet->length > sizeof (*recv_buf->mad) &&
266 	     count < hdr_size(file) + sizeof (*recv_buf->mad)))
267 		return -EINVAL;
268 
269 	if (copy_to_user(buf, &packet->mad, hdr_size(file)))
270 		return -EFAULT;
271 
272 	buf += hdr_size(file);
273 	seg_payload = min_t(int, packet->length, sizeof (*recv_buf->mad));
274 	if (copy_to_user(buf, recv_buf->mad, seg_payload))
275 		return -EFAULT;
276 
277 	if (seg_payload < packet->length) {
278 		/*
279 		 * Multipacket RMPP MAD message. Copy remainder of message.
280 		 * Note that last segment may have a shorter payload.
281 		 */
282 		if (count < hdr_size(file) + packet->length) {
283 			/*
284 			 * The buffer is too small, return the first RMPP segment,
285 			 * which includes the RMPP message length.
286 			 */
287 			return -ENOSPC;
288 		}
289 		offset = ib_get_mad_data_offset(recv_buf->mad->mad_hdr.mgmt_class);
290 		max_seg_payload = sizeof (struct ib_mad) - offset;
291 
292 		for (left = packet->length - seg_payload, buf += seg_payload;
293 		     left; left -= seg_payload, buf += seg_payload) {
294 			recv_buf = container_of(recv_buf->list.next,
295 						struct ib_mad_recv_buf, list);
296 			seg_payload = min(left, max_seg_payload);
297 			if (copy_to_user(buf, ((void *) recv_buf->mad) + offset,
298 					 seg_payload))
299 				return -EFAULT;
300 		}
301 	}
302 	return hdr_size(file) + packet->length;
303 }
304 
305 static ssize_t copy_send_mad(struct ib_umad_file *file, char __user *buf,
306 			     struct ib_umad_packet *packet, size_t count)
307 {
308 	ssize_t size = hdr_size(file) + packet->length;
309 
310 	if (count < size)
311 		return -EINVAL;
312 
313 	if (copy_to_user(buf, &packet->mad, hdr_size(file)))
314 		return -EFAULT;
315 
316 	buf += hdr_size(file);
317 
318 	if (copy_to_user(buf, packet->mad.data, packet->length))
319 		return -EFAULT;
320 
321 	return size;
322 }
323 
324 static ssize_t ib_umad_read(struct file *filp, char __user *buf,
325 			    size_t count, loff_t *pos)
326 {
327 	struct ib_umad_file *file = filp->private_data;
328 	struct ib_umad_packet *packet;
329 	ssize_t ret;
330 
331 	if (count < hdr_size(file))
332 		return -EINVAL;
333 
334 	mutex_lock(&file->mutex);
335 
336 	while (list_empty(&file->recv_list)) {
337 		mutex_unlock(&file->mutex);
338 
339 		if (filp->f_flags & O_NONBLOCK)
340 			return -EAGAIN;
341 
342 		if (wait_event_interruptible(file->recv_wait,
343 					     !list_empty(&file->recv_list)))
344 			return -ERESTARTSYS;
345 
346 		mutex_lock(&file->mutex);
347 	}
348 
349 	packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
350 	list_del(&packet->list);
351 
352 	mutex_unlock(&file->mutex);
353 
354 	if (packet->recv_wc)
355 		ret = copy_recv_mad(file, buf, packet, count);
356 	else
357 		ret = copy_send_mad(file, buf, packet, count);
358 
359 	if (ret < 0) {
360 		/* Requeue packet */
361 		mutex_lock(&file->mutex);
362 		list_add(&packet->list, &file->recv_list);
363 		mutex_unlock(&file->mutex);
364 	} else {
365 		if (packet->recv_wc)
366 			ib_free_recv_mad(packet->recv_wc);
367 		kfree(packet);
368 	}
369 	return ret;
370 }
371 
372 static int copy_rmpp_mad(struct ib_mad_send_buf *msg, const char __user *buf)
373 {
374 	int left, seg;
375 
376 	/* Copy class specific header */
377 	if ((msg->hdr_len > IB_MGMT_RMPP_HDR) &&
378 	    copy_from_user(msg->mad + IB_MGMT_RMPP_HDR, buf + IB_MGMT_RMPP_HDR,
379 			   msg->hdr_len - IB_MGMT_RMPP_HDR))
380 		return -EFAULT;
381 
382 	/* All headers are in place.  Copy data segments. */
383 	for (seg = 1, left = msg->data_len, buf += msg->hdr_len; left > 0;
384 	     seg++, left -= msg->seg_size, buf += msg->seg_size) {
385 		if (copy_from_user(ib_get_rmpp_segment(msg, seg), buf,
386 				   min(left, msg->seg_size)))
387 			return -EFAULT;
388 	}
389 	return 0;
390 }
391 
392 static int same_destination(struct ib_user_mad_hdr *hdr1,
393 			    struct ib_user_mad_hdr *hdr2)
394 {
395 	if (!hdr1->grh_present && !hdr2->grh_present)
396 	   return (hdr1->lid == hdr2->lid);
397 
398 	if (hdr1->grh_present && hdr2->grh_present)
399 	   return !memcmp(hdr1->gid, hdr2->gid, 16);
400 
401 	return 0;
402 }
403 
404 static int is_duplicate(struct ib_umad_file *file,
405 			struct ib_umad_packet *packet)
406 {
407 	struct ib_umad_packet *sent_packet;
408 	struct ib_mad_hdr *sent_hdr, *hdr;
409 
410 	hdr = (struct ib_mad_hdr *) packet->mad.data;
411 	list_for_each_entry(sent_packet, &file->send_list, list) {
412 		sent_hdr = (struct ib_mad_hdr *) sent_packet->mad.data;
413 
414 		if ((hdr->tid != sent_hdr->tid) ||
415 		    (hdr->mgmt_class != sent_hdr->mgmt_class))
416 			continue;
417 
418 		/*
419 		 * No need to be overly clever here.  If two new operations have
420 		 * the same TID, reject the second as a duplicate.  This is more
421 		 * restrictive than required by the spec.
422 		 */
423 		if (!ib_response_mad((struct ib_mad *) hdr)) {
424 			if (!ib_response_mad((struct ib_mad *) sent_hdr))
425 				return 1;
426 			continue;
427 		} else if (!ib_response_mad((struct ib_mad *) sent_hdr))
428 			continue;
429 
430 		if (same_destination(&packet->mad.hdr, &sent_packet->mad.hdr))
431 			return 1;
432 	}
433 
434 	return 0;
435 }
436 
437 static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
438 			     size_t count, loff_t *pos)
439 {
440 	struct ib_umad_file *file = filp->private_data;
441 	struct ib_umad_packet *packet;
442 	struct ib_mad_agent *agent;
443 	struct ib_ah_attr ah_attr;
444 	struct ib_ah *ah;
445 	struct ib_rmpp_mad *rmpp_mad;
446 	__be64 *tid;
447 	int ret, data_len, hdr_len, copy_offset, rmpp_active;
448 
449 	if (count < hdr_size(file) + IB_MGMT_RMPP_HDR)
450 		return -EINVAL;
451 
452 	packet = kzalloc(sizeof *packet + IB_MGMT_RMPP_HDR, GFP_KERNEL);
453 	if (!packet)
454 		return -ENOMEM;
455 
456 	if (copy_from_user(&packet->mad, buf, hdr_size(file))) {
457 		ret = -EFAULT;
458 		goto err;
459 	}
460 
461 	if (packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
462 		ret = -EINVAL;
463 		goto err;
464 	}
465 
466 	buf += hdr_size(file);
467 
468 	if (copy_from_user(packet->mad.data, buf, IB_MGMT_RMPP_HDR)) {
469 		ret = -EFAULT;
470 		goto err;
471 	}
472 
473 	mutex_lock(&file->mutex);
474 
475 	agent = __get_agent(file, packet->mad.hdr.id);
476 	if (!agent) {
477 		ret = -EINVAL;
478 		goto err_up;
479 	}
480 
481 	memset(&ah_attr, 0, sizeof ah_attr);
482 	ah_attr.dlid          = be16_to_cpu(packet->mad.hdr.lid);
483 	ah_attr.sl            = packet->mad.hdr.sl;
484 	ah_attr.src_path_bits = packet->mad.hdr.path_bits;
485 	ah_attr.port_num      = file->port->port_num;
486 	if (packet->mad.hdr.grh_present) {
487 		ah_attr.ah_flags = IB_AH_GRH;
488 		memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
489 		ah_attr.grh.sgid_index	   = packet->mad.hdr.gid_index;
490 		ah_attr.grh.flow_label	   = be32_to_cpu(packet->mad.hdr.flow_label);
491 		ah_attr.grh.hop_limit	   = packet->mad.hdr.hop_limit;
492 		ah_attr.grh.traffic_class  = packet->mad.hdr.traffic_class;
493 	}
494 
495 	ah = ib_create_ah(agent->qp->pd, &ah_attr);
496 	if (IS_ERR(ah)) {
497 		ret = PTR_ERR(ah);
498 		goto err_up;
499 	}
500 
501 	rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
502 	hdr_len = ib_get_mad_data_offset(rmpp_mad->mad_hdr.mgmt_class);
503 	if (!ib_is_mad_class_rmpp(rmpp_mad->mad_hdr.mgmt_class)) {
504 		copy_offset = IB_MGMT_MAD_HDR;
505 		rmpp_active = 0;
506 	} else {
507 		copy_offset = IB_MGMT_RMPP_HDR;
508 		rmpp_active = ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
509 			      IB_MGMT_RMPP_FLAG_ACTIVE;
510 	}
511 
512 	data_len = count - hdr_size(file) - hdr_len;
513 	packet->msg = ib_create_send_mad(agent,
514 					 be32_to_cpu(packet->mad.hdr.qpn),
515 					 packet->mad.hdr.pkey_index, rmpp_active,
516 					 hdr_len, data_len, GFP_KERNEL);
517 	if (IS_ERR(packet->msg)) {
518 		ret = PTR_ERR(packet->msg);
519 		goto err_ah;
520 	}
521 
522 	packet->msg->ah		= ah;
523 	packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
524 	packet->msg->retries	= packet->mad.hdr.retries;
525 	packet->msg->context[0] = packet;
526 
527 	/* Copy MAD header.  Any RMPP header is already in place. */
528 	memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
529 
530 	if (!rmpp_active) {
531 		if (copy_from_user(packet->msg->mad + copy_offset,
532 				   buf + copy_offset,
533 				   hdr_len + data_len - copy_offset)) {
534 			ret = -EFAULT;
535 			goto err_msg;
536 		}
537 	} else {
538 		ret = copy_rmpp_mad(packet->msg, buf);
539 		if (ret)
540 			goto err_msg;
541 	}
542 
543 	/*
544 	 * Set the high-order part of the transaction ID to make MADs from
545 	 * different agents unique, and allow routing responses back to the
546 	 * original requestor.
547 	 */
548 	if (!ib_response_mad(packet->msg->mad)) {
549 		tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
550 		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
551 				   (be64_to_cpup(tid) & 0xffffffff));
552 		rmpp_mad->mad_hdr.tid = *tid;
553 	}
554 
555 	spin_lock_irq(&file->send_lock);
556 	ret = is_duplicate(file, packet);
557 	if (!ret)
558 		list_add_tail(&packet->list, &file->send_list);
559 	spin_unlock_irq(&file->send_lock);
560 	if (ret) {
561 		ret = -EINVAL;
562 		goto err_msg;
563 	}
564 
565 	ret = ib_post_send_mad(packet->msg, NULL);
566 	if (ret)
567 		goto err_send;
568 
569 	mutex_unlock(&file->mutex);
570 	return count;
571 
572 err_send:
573 	dequeue_send(file, packet);
574 err_msg:
575 	ib_free_send_mad(packet->msg);
576 err_ah:
577 	ib_destroy_ah(ah);
578 err_up:
579 	mutex_unlock(&file->mutex);
580 err:
581 	kfree(packet);
582 	return ret;
583 }
584 
585 static unsigned int ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
586 {
587 	struct ib_umad_file *file = filp->private_data;
588 
589 	/* we will always be able to post a MAD send */
590 	unsigned int mask = POLLOUT | POLLWRNORM;
591 
592 	poll_wait(filp, &file->recv_wait, wait);
593 
594 	if (!list_empty(&file->recv_list))
595 		mask |= POLLIN | POLLRDNORM;
596 
597 	return mask;
598 }
599 
600 static int ib_umad_reg_agent(struct ib_umad_file *file, void __user *arg,
601 			     int compat_method_mask)
602 {
603 	struct ib_user_mad_reg_req ureq;
604 	struct ib_mad_reg_req req;
605 	struct ib_mad_agent *agent = NULL;
606 	int agent_id;
607 	int ret;
608 
609 	mutex_lock(&file->port->file_mutex);
610 	mutex_lock(&file->mutex);
611 
612 	if (!file->port->ib_dev) {
613 		ret = -EPIPE;
614 		goto out;
615 	}
616 
617 	if (copy_from_user(&ureq, arg, sizeof ureq)) {
618 		ret = -EFAULT;
619 		goto out;
620 	}
621 
622 	if (ureq.qpn != 0 && ureq.qpn != 1) {
623 		ret = -EINVAL;
624 		goto out;
625 	}
626 
627 	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
628 		if (!__get_agent(file, agent_id))
629 			goto found;
630 
631 	ret = -ENOMEM;
632 	goto out;
633 
634 found:
635 	if (ureq.mgmt_class) {
636 		req.mgmt_class         = ureq.mgmt_class;
637 		req.mgmt_class_version = ureq.mgmt_class_version;
638 		memcpy(req.oui, ureq.oui, sizeof req.oui);
639 
640 		if (compat_method_mask) {
641 			u32 *umm = (u32 *) ureq.method_mask;
642 			int i;
643 
644 			for (i = 0; i < BITS_TO_LONGS(IB_MGMT_MAX_METHODS); ++i)
645 				req.method_mask[i] =
646 					umm[i * 2] | ((u64) umm[i * 2 + 1] << 32);
647 		} else
648 			memcpy(req.method_mask, ureq.method_mask,
649 			       sizeof req.method_mask);
650 	}
651 
652 	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
653 				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
654 				      ureq.mgmt_class ? &req : NULL,
655 				      ureq.rmpp_version,
656 				      send_handler, recv_handler, file);
657 	if (IS_ERR(agent)) {
658 		ret = PTR_ERR(agent);
659 		agent = NULL;
660 		goto out;
661 	}
662 
663 	if (put_user(agent_id,
664 		     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
665 		ret = -EFAULT;
666 		goto out;
667 	}
668 
669 	if (!file->already_used) {
670 		file->already_used = 1;
671 		if (!file->use_pkey_index) {
672 			printk(KERN_WARNING "user_mad: process %s did not enable "
673 			       "P_Key index support.\n", current->comm);
674 			printk(KERN_WARNING "user_mad:   Documentation/infiniband/user_mad.txt "
675 			       "has info on the new ABI.\n");
676 		}
677 	}
678 
679 	file->agent[agent_id] = agent;
680 	ret = 0;
681 
682 out:
683 	mutex_unlock(&file->mutex);
684 
685 	if (ret && agent)
686 		ib_unregister_mad_agent(agent);
687 
688 	mutex_unlock(&file->port->file_mutex);
689 
690 	return ret;
691 }
692 
693 static int ib_umad_unreg_agent(struct ib_umad_file *file, u32 __user *arg)
694 {
695 	struct ib_mad_agent *agent = NULL;
696 	u32 id;
697 	int ret = 0;
698 
699 	if (get_user(id, arg))
700 		return -EFAULT;
701 
702 	mutex_lock(&file->port->file_mutex);
703 	mutex_lock(&file->mutex);
704 
705 	if (id >= IB_UMAD_MAX_AGENTS || !__get_agent(file, id)) {
706 		ret = -EINVAL;
707 		goto out;
708 	}
709 
710 	agent = file->agent[id];
711 	file->agent[id] = NULL;
712 
713 out:
714 	mutex_unlock(&file->mutex);
715 
716 	if (agent)
717 		ib_unregister_mad_agent(agent);
718 
719 	mutex_unlock(&file->port->file_mutex);
720 
721 	return ret;
722 }
723 
724 static long ib_umad_enable_pkey(struct ib_umad_file *file)
725 {
726 	int ret = 0;
727 
728 	mutex_lock(&file->mutex);
729 	if (file->already_used)
730 		ret = -EINVAL;
731 	else
732 		file->use_pkey_index = 1;
733 	mutex_unlock(&file->mutex);
734 
735 	return ret;
736 }
737 
738 static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
739 			  unsigned long arg)
740 {
741 	switch (cmd) {
742 	case IB_USER_MAD_REGISTER_AGENT:
743 		return ib_umad_reg_agent(filp->private_data, (void __user *) arg, 0);
744 	case IB_USER_MAD_UNREGISTER_AGENT:
745 		return ib_umad_unreg_agent(filp->private_data, (__u32 __user *) arg);
746 	case IB_USER_MAD_ENABLE_PKEY:
747 		return ib_umad_enable_pkey(filp->private_data);
748 	default:
749 		return -ENOIOCTLCMD;
750 	}
751 }
752 
753 #ifdef CONFIG_COMPAT
754 static long ib_umad_compat_ioctl(struct file *filp, unsigned int cmd,
755 				 unsigned long arg)
756 {
757 	switch (cmd) {
758 	case IB_USER_MAD_REGISTER_AGENT:
759 		return ib_umad_reg_agent(filp->private_data, compat_ptr(arg), 1);
760 	case IB_USER_MAD_UNREGISTER_AGENT:
761 		return ib_umad_unreg_agent(filp->private_data, compat_ptr(arg));
762 	case IB_USER_MAD_ENABLE_PKEY:
763 		return ib_umad_enable_pkey(filp->private_data);
764 	default:
765 		return -ENOIOCTLCMD;
766 	}
767 }
768 #endif
769 
770 /*
771  * ib_umad_open() does not need the BKL:
772  *
773  *  - the ib_umad_port structures are properly reference counted, and
774  *    everything else is purely local to the file being created, so
775  *    races against other open calls are not a problem;
776  *  - the ioctl method does not affect any global state outside of the
777  *    file structure being operated on;
778  */
779 static int ib_umad_open(struct inode *inode, struct file *filp)
780 {
781 	struct ib_umad_port *port;
782 	struct ib_umad_file *file;
783 	int ret;
784 
785 	port = container_of(inode->i_cdev, struct ib_umad_port, cdev);
786 	if (port)
787 		kref_get(&port->umad_dev->ref);
788 	else
789 		return -ENXIO;
790 
791 	mutex_lock(&port->file_mutex);
792 
793 	if (!port->ib_dev) {
794 		ret = -ENXIO;
795 		goto out;
796 	}
797 
798 	file = kzalloc(sizeof *file, GFP_KERNEL);
799 	if (!file) {
800 		kref_put(&port->umad_dev->ref, ib_umad_release_dev);
801 		ret = -ENOMEM;
802 		goto out;
803 	}
804 
805 	mutex_init(&file->mutex);
806 	spin_lock_init(&file->send_lock);
807 	INIT_LIST_HEAD(&file->recv_list);
808 	INIT_LIST_HEAD(&file->send_list);
809 	init_waitqueue_head(&file->recv_wait);
810 
811 	file->port = port;
812 	filp->private_data = file;
813 
814 	list_add_tail(&file->port_list, &port->file_list);
815 
816 	ret = nonseekable_open(inode, filp);
817 
818 out:
819 	mutex_unlock(&port->file_mutex);
820 	return ret;
821 }
822 
823 static int ib_umad_close(struct inode *inode, struct file *filp)
824 {
825 	struct ib_umad_file *file = filp->private_data;
826 	struct ib_umad_device *dev = file->port->umad_dev;
827 	struct ib_umad_packet *packet, *tmp;
828 	int already_dead;
829 	int i;
830 
831 	mutex_lock(&file->port->file_mutex);
832 	mutex_lock(&file->mutex);
833 
834 	already_dead = file->agents_dead;
835 	file->agents_dead = 1;
836 
837 	list_for_each_entry_safe(packet, tmp, &file->recv_list, list) {
838 		if (packet->recv_wc)
839 			ib_free_recv_mad(packet->recv_wc);
840 		kfree(packet);
841 	}
842 
843 	list_del(&file->port_list);
844 
845 	mutex_unlock(&file->mutex);
846 
847 	if (!already_dead)
848 		for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
849 			if (file->agent[i])
850 				ib_unregister_mad_agent(file->agent[i]);
851 
852 	mutex_unlock(&file->port->file_mutex);
853 
854 	kfree(file);
855 	kref_put(&dev->ref, ib_umad_release_dev);
856 
857 	return 0;
858 }
859 
860 static const struct file_operations umad_fops = {
861 	.owner		= THIS_MODULE,
862 	.read		= ib_umad_read,
863 	.write		= ib_umad_write,
864 	.poll		= ib_umad_poll,
865 	.unlocked_ioctl = ib_umad_ioctl,
866 #ifdef CONFIG_COMPAT
867 	.compat_ioctl	= ib_umad_compat_ioctl,
868 #endif
869 	.open		= ib_umad_open,
870 	.release	= ib_umad_close,
871 	.llseek		= no_llseek,
872 };
873 
874 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
875 {
876 	struct ib_umad_port *port;
877 	struct ib_port_modify props = {
878 		.set_port_cap_mask = IB_PORT_SM
879 	};
880 	int ret;
881 
882 	port = container_of(inode->i_cdev, struct ib_umad_port, sm_cdev);
883 	if (port)
884 		kref_get(&port->umad_dev->ref);
885 	else
886 		return -ENXIO;
887 
888 	if (filp->f_flags & O_NONBLOCK) {
889 		if (down_trylock(&port->sm_sem)) {
890 			ret = -EAGAIN;
891 			goto fail;
892 		}
893 	} else {
894 		if (down_interruptible(&port->sm_sem)) {
895 			ret = -ERESTARTSYS;
896 			goto fail;
897 		}
898 	}
899 
900 	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
901 	if (ret) {
902 		up(&port->sm_sem);
903 		goto fail;
904 	}
905 
906 	filp->private_data = port;
907 
908 	return nonseekable_open(inode, filp);
909 
910 fail:
911 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
912 	return ret;
913 }
914 
915 static int ib_umad_sm_close(struct inode *inode, struct file *filp)
916 {
917 	struct ib_umad_port *port = filp->private_data;
918 	struct ib_port_modify props = {
919 		.clr_port_cap_mask = IB_PORT_SM
920 	};
921 	int ret = 0;
922 
923 	mutex_lock(&port->file_mutex);
924 	if (port->ib_dev)
925 		ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
926 	mutex_unlock(&port->file_mutex);
927 
928 	up(&port->sm_sem);
929 
930 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
931 
932 	return ret;
933 }
934 
935 static const struct file_operations umad_sm_fops = {
936 	.owner	 = THIS_MODULE,
937 	.open	 = ib_umad_sm_open,
938 	.release = ib_umad_sm_close,
939 	.llseek	 = no_llseek,
940 };
941 
942 static struct ib_client umad_client = {
943 	.name   = "umad",
944 	.add    = ib_umad_add_one,
945 	.remove = ib_umad_remove_one
946 };
947 
948 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
949 			  char *buf)
950 {
951 	struct ib_umad_port *port = dev_get_drvdata(dev);
952 
953 	if (!port)
954 		return -ENODEV;
955 
956 	return sprintf(buf, "%s\n", port->ib_dev->name);
957 }
958 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
959 
960 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
961 			 char *buf)
962 {
963 	struct ib_umad_port *port = dev_get_drvdata(dev);
964 
965 	if (!port)
966 		return -ENODEV;
967 
968 	return sprintf(buf, "%d\n", port->port_num);
969 }
970 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
971 
972 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
973 			 __stringify(IB_USER_MAD_ABI_VERSION));
974 
975 static dev_t overflow_maj;
976 static DECLARE_BITMAP(overflow_map, IB_UMAD_MAX_PORTS);
977 static int find_overflow_devnum(void)
978 {
979 	int ret;
980 
981 	if (!overflow_maj) {
982 		ret = alloc_chrdev_region(&overflow_maj, 0, IB_UMAD_MAX_PORTS * 2,
983 					  "infiniband_mad");
984 		if (ret) {
985 			printk(KERN_ERR "user_mad: couldn't register dynamic device number\n");
986 			return ret;
987 		}
988 	}
989 
990 	ret = find_first_zero_bit(overflow_map, IB_UMAD_MAX_PORTS);
991 	if (ret >= IB_UMAD_MAX_PORTS)
992 		return -1;
993 
994 	return ret;
995 }
996 
997 static int ib_umad_init_port(struct ib_device *device, int port_num,
998 			     struct ib_umad_port *port)
999 {
1000 	int devnum;
1001 	dev_t base;
1002 
1003 	spin_lock(&port_lock);
1004 	devnum = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
1005 	if (devnum >= IB_UMAD_MAX_PORTS) {
1006 		spin_unlock(&port_lock);
1007 		devnum = find_overflow_devnum();
1008 		if (devnum < 0)
1009 			return -1;
1010 
1011 		spin_lock(&port_lock);
1012 		port->dev_num = devnum + IB_UMAD_MAX_PORTS;
1013 		base = devnum + overflow_maj;
1014 		set_bit(devnum, overflow_map);
1015 	} else {
1016 		port->dev_num = devnum;
1017 		base = devnum + base_dev;
1018 		set_bit(devnum, dev_map);
1019 	}
1020 	spin_unlock(&port_lock);
1021 
1022 	port->ib_dev   = device;
1023 	port->port_num = port_num;
1024 	sema_init(&port->sm_sem, 1);
1025 	mutex_init(&port->file_mutex);
1026 	INIT_LIST_HEAD(&port->file_list);
1027 
1028 	cdev_init(&port->cdev, &umad_fops);
1029 	port->cdev.owner = THIS_MODULE;
1030 	kobject_set_name(&port->cdev.kobj, "umad%d", port->dev_num);
1031 	if (cdev_add(&port->cdev, base, 1))
1032 		goto err_cdev;
1033 
1034 	port->dev = device_create(umad_class, device->dma_device,
1035 				  port->cdev.dev, port,
1036 				  "umad%d", port->dev_num);
1037 	if (IS_ERR(port->dev))
1038 		goto err_cdev;
1039 
1040 	if (device_create_file(port->dev, &dev_attr_ibdev))
1041 		goto err_dev;
1042 	if (device_create_file(port->dev, &dev_attr_port))
1043 		goto err_dev;
1044 
1045 	base += IB_UMAD_MAX_PORTS;
1046 	cdev_init(&port->sm_cdev, &umad_sm_fops);
1047 	port->sm_cdev.owner = THIS_MODULE;
1048 	kobject_set_name(&port->sm_cdev.kobj, "issm%d", port->dev_num);
1049 	if (cdev_add(&port->sm_cdev, base, 1))
1050 		goto err_sm_cdev;
1051 
1052 	port->sm_dev = device_create(umad_class, device->dma_device,
1053 				     port->sm_cdev.dev, port,
1054 				     "issm%d", port->dev_num);
1055 	if (IS_ERR(port->sm_dev))
1056 		goto err_sm_cdev;
1057 
1058 	if (device_create_file(port->sm_dev, &dev_attr_ibdev))
1059 		goto err_sm_dev;
1060 	if (device_create_file(port->sm_dev, &dev_attr_port))
1061 		goto err_sm_dev;
1062 
1063 	return 0;
1064 
1065 err_sm_dev:
1066 	device_destroy(umad_class, port->sm_cdev.dev);
1067 
1068 err_sm_cdev:
1069 	cdev_del(&port->sm_cdev);
1070 
1071 err_dev:
1072 	device_destroy(umad_class, port->cdev.dev);
1073 
1074 err_cdev:
1075 	cdev_del(&port->cdev);
1076 	if (port->dev_num < IB_UMAD_MAX_PORTS)
1077 		clear_bit(devnum, dev_map);
1078 	else
1079 		clear_bit(devnum, overflow_map);
1080 
1081 	return -1;
1082 }
1083 
1084 static void ib_umad_kill_port(struct ib_umad_port *port)
1085 {
1086 	struct ib_umad_file *file;
1087 	int id;
1088 
1089 	dev_set_drvdata(port->dev,    NULL);
1090 	dev_set_drvdata(port->sm_dev, NULL);
1091 
1092 	device_destroy(umad_class, port->cdev.dev);
1093 	device_destroy(umad_class, port->sm_cdev.dev);
1094 
1095 	cdev_del(&port->cdev);
1096 	cdev_del(&port->sm_cdev);
1097 
1098 	mutex_lock(&port->file_mutex);
1099 
1100 	port->ib_dev = NULL;
1101 
1102 	list_for_each_entry(file, &port->file_list, port_list) {
1103 		mutex_lock(&file->mutex);
1104 		file->agents_dead = 1;
1105 		mutex_unlock(&file->mutex);
1106 
1107 		for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id)
1108 			if (file->agent[id])
1109 				ib_unregister_mad_agent(file->agent[id]);
1110 	}
1111 
1112 	mutex_unlock(&port->file_mutex);
1113 
1114 	if (port->dev_num < IB_UMAD_MAX_PORTS)
1115 		clear_bit(port->dev_num, dev_map);
1116 	else
1117 		clear_bit(port->dev_num - IB_UMAD_MAX_PORTS, overflow_map);
1118 }
1119 
1120 static void ib_umad_add_one(struct ib_device *device)
1121 {
1122 	struct ib_umad_device *umad_dev;
1123 	int s, e, i;
1124 
1125 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1126 		return;
1127 
1128 	if (device->node_type == RDMA_NODE_IB_SWITCH)
1129 		s = e = 0;
1130 	else {
1131 		s = 1;
1132 		e = device->phys_port_cnt;
1133 	}
1134 
1135 	umad_dev = kzalloc(sizeof *umad_dev +
1136 			   (e - s + 1) * sizeof (struct ib_umad_port),
1137 			   GFP_KERNEL);
1138 	if (!umad_dev)
1139 		return;
1140 
1141 	kref_init(&umad_dev->ref);
1142 
1143 	umad_dev->start_port = s;
1144 	umad_dev->end_port   = e;
1145 
1146 	for (i = s; i <= e; ++i) {
1147 		umad_dev->port[i - s].umad_dev = umad_dev;
1148 
1149 		if (ib_umad_init_port(device, i, &umad_dev->port[i - s]))
1150 			goto err;
1151 	}
1152 
1153 	ib_set_client_data(device, &umad_client, umad_dev);
1154 
1155 	return;
1156 
1157 err:
1158 	while (--i >= s)
1159 		ib_umad_kill_port(&umad_dev->port[i - s]);
1160 
1161 	kref_put(&umad_dev->ref, ib_umad_release_dev);
1162 }
1163 
1164 static void ib_umad_remove_one(struct ib_device *device)
1165 {
1166 	struct ib_umad_device *umad_dev = ib_get_client_data(device, &umad_client);
1167 	int i;
1168 
1169 	if (!umad_dev)
1170 		return;
1171 
1172 	for (i = 0; i <= umad_dev->end_port - umad_dev->start_port; ++i)
1173 		ib_umad_kill_port(&umad_dev->port[i]);
1174 
1175 	kref_put(&umad_dev->ref, ib_umad_release_dev);
1176 }
1177 
1178 static char *umad_devnode(struct device *dev, umode_t *mode)
1179 {
1180 	return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1181 }
1182 
1183 static int __init ib_umad_init(void)
1184 {
1185 	int ret;
1186 
1187 	ret = register_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2,
1188 				     "infiniband_mad");
1189 	if (ret) {
1190 		printk(KERN_ERR "user_mad: couldn't register device number\n");
1191 		goto out;
1192 	}
1193 
1194 	umad_class = class_create(THIS_MODULE, "infiniband_mad");
1195 	if (IS_ERR(umad_class)) {
1196 		ret = PTR_ERR(umad_class);
1197 		printk(KERN_ERR "user_mad: couldn't create class infiniband_mad\n");
1198 		goto out_chrdev;
1199 	}
1200 
1201 	umad_class->devnode = umad_devnode;
1202 
1203 	ret = class_create_file(umad_class, &class_attr_abi_version.attr);
1204 	if (ret) {
1205 		printk(KERN_ERR "user_mad: couldn't create abi_version attribute\n");
1206 		goto out_class;
1207 	}
1208 
1209 	ret = ib_register_client(&umad_client);
1210 	if (ret) {
1211 		printk(KERN_ERR "user_mad: couldn't register ib_umad client\n");
1212 		goto out_class;
1213 	}
1214 
1215 	return 0;
1216 
1217 out_class:
1218 	class_destroy(umad_class);
1219 
1220 out_chrdev:
1221 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
1222 
1223 out:
1224 	return ret;
1225 }
1226 
1227 static void __exit ib_umad_cleanup(void)
1228 {
1229 	ib_unregister_client(&umad_client);
1230 	class_destroy(umad_class);
1231 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
1232 	if (overflow_maj)
1233 		unregister_chrdev_region(overflow_maj, IB_UMAD_MAX_PORTS * 2);
1234 }
1235 
1236 module_init(ib_umad_init);
1237 module_exit(ib_umad_cleanup);
1238