xref: /linux/drivers/infiniband/core/user_mad.c (revision c537b994505099b7197e7d3125b942ecbcc51eb6)
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  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  *
34  * $Id: user_mad.c 5596 2006-03-03 01:00:07Z sean.hefty $
35  */
36 
37 #include <linux/module.h>
38 #include <linux/init.h>
39 #include <linux/device.h>
40 #include <linux/err.h>
41 #include <linux/fs.h>
42 #include <linux/cdev.h>
43 #include <linux/pci.h>
44 #include <linux/dma-mapping.h>
45 #include <linux/poll.h>
46 #include <linux/rwsem.h>
47 #include <linux/kref.h>
48 
49 #include <asm/uaccess.h>
50 #include <asm/semaphore.h>
51 
52 #include <rdma/ib_mad.h>
53 #include <rdma/ib_user_mad.h>
54 
55 MODULE_AUTHOR("Roland Dreier");
56 MODULE_DESCRIPTION("InfiniBand userspace MAD packet access");
57 MODULE_LICENSE("Dual BSD/GPL");
58 
59 enum {
60 	IB_UMAD_MAX_PORTS  = 64,
61 	IB_UMAD_MAX_AGENTS = 32,
62 
63 	IB_UMAD_MAJOR      = 231,
64 	IB_UMAD_MINOR_BASE = 0
65 };
66 
67 /*
68  * Our lifetime rules for these structs are the following: each time a
69  * device special file is opened, we look up the corresponding struct
70  * ib_umad_port by minor in the umad_port[] table while holding the
71  * port_lock.  If this lookup succeeds, we take a reference on the
72  * ib_umad_port's struct ib_umad_device while still holding the
73  * port_lock; if the lookup fails, we fail the open().  We drop these
74  * references in the corresponding close().
75  *
76  * In addition to references coming from open character devices, there
77  * is one more reference to each ib_umad_device representing the
78  * module's reference taken when allocating the ib_umad_device in
79  * ib_umad_add_one().
80  *
81  * When destroying an ib_umad_device, we clear all of its
82  * ib_umad_ports from umad_port[] while holding port_lock before
83  * dropping the module's reference to the ib_umad_device.  This is
84  * always safe because any open() calls will either succeed and obtain
85  * a reference before we clear the umad_port[] entries, or fail after
86  * we clear the umad_port[] entries.
87  */
88 
89 struct ib_umad_port {
90 	struct cdev           *dev;
91 	struct class_device   *class_dev;
92 
93 	struct cdev           *sm_dev;
94 	struct class_device   *sm_class_dev;
95 	struct semaphore       sm_sem;
96 
97 	struct rw_semaphore    mutex;
98 	struct list_head       file_list;
99 
100 	struct ib_device      *ib_dev;
101 	struct ib_umad_device *umad_dev;
102 	int                    dev_num;
103 	u8                     port_num;
104 };
105 
106 struct ib_umad_device {
107 	int                  start_port, end_port;
108 	struct kref          ref;
109 	struct ib_umad_port  port[0];
110 };
111 
112 struct ib_umad_file {
113 	struct ib_umad_port    *port;
114 	struct list_head	recv_list;
115 	struct list_head	send_list;
116 	struct list_head	port_list;
117 	spinlock_t		recv_lock;
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 };
123 
124 struct ib_umad_packet {
125 	struct ib_mad_send_buf *msg;
126 	struct ib_mad_recv_wc  *recv_wc;
127 	struct list_head   list;
128 	int		   length;
129 	struct ib_user_mad mad;
130 };
131 
132 static struct class *umad_class;
133 
134 static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
135 
136 static DEFINE_SPINLOCK(port_lock);
137 static struct ib_umad_port *umad_port[IB_UMAD_MAX_PORTS];
138 static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS * 2);
139 
140 static void ib_umad_add_one(struct ib_device *device);
141 static void ib_umad_remove_one(struct ib_device *device);
142 
143 static void ib_umad_release_dev(struct kref *ref)
144 {
145 	struct ib_umad_device *dev =
146 		container_of(ref, struct ib_umad_device, ref);
147 
148 	kfree(dev);
149 }
150 
151 /* caller must hold port->mutex at least for reading */
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 	down_read(&file->port->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 			spin_lock_irq(&file->recv_lock);
170 			list_add_tail(&packet->list, &file->recv_list);
171 			spin_unlock_irq(&file->recv_lock);
172 			wake_up_interruptible(&file->recv_wait);
173 			ret = 0;
174 			break;
175 		}
176 
177 	up_read(&file->port->mutex);
178 
179 	return ret;
180 }
181 
182 static void dequeue_send(struct ib_umad_file *file,
183 			 struct ib_umad_packet *packet)
184  {
185 	spin_lock_irq(&file->send_lock);
186 	list_del(&packet->list);
187 	spin_unlock_irq(&file->send_lock);
188  }
189 
190 static void send_handler(struct ib_mad_agent *agent,
191 			 struct ib_mad_send_wc *send_wc)
192 {
193 	struct ib_umad_file *file = agent->context;
194 	struct ib_umad_packet *packet = send_wc->send_buf->context[0];
195 
196 	dequeue_send(file, packet);
197 	ib_destroy_ah(packet->msg->ah);
198 	ib_free_send_mad(packet->msg);
199 
200 	if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
201 		packet->length = IB_MGMT_MAD_HDR;
202 		packet->mad.hdr.status = ETIMEDOUT;
203 		if (!queue_packet(file, agent, packet))
204 			return;
205 	}
206 	kfree(packet);
207 }
208 
209 static void recv_handler(struct ib_mad_agent *agent,
210 			 struct ib_mad_recv_wc *mad_recv_wc)
211 {
212 	struct ib_umad_file *file = agent->context;
213 	struct ib_umad_packet *packet;
214 
215 	if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
216 		goto err1;
217 
218 	packet = kzalloc(sizeof *packet, GFP_KERNEL);
219 	if (!packet)
220 		goto err1;
221 
222 	packet->length = mad_recv_wc->mad_len;
223 	packet->recv_wc = mad_recv_wc;
224 
225 	packet->mad.hdr.status    = 0;
226 	packet->mad.hdr.length    = sizeof (struct ib_user_mad) +
227 				    mad_recv_wc->mad_len;
228 	packet->mad.hdr.qpn 	  = cpu_to_be32(mad_recv_wc->wc->src_qp);
229 	packet->mad.hdr.lid 	  = cpu_to_be16(mad_recv_wc->wc->slid);
230 	packet->mad.hdr.sl  	  = mad_recv_wc->wc->sl;
231 	packet->mad.hdr.path_bits = mad_recv_wc->wc->dlid_path_bits;
232 	packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
233 	if (packet->mad.hdr.grh_present) {
234 		/* XXX parse GRH */
235 		packet->mad.hdr.gid_index 	= 0;
236 		packet->mad.hdr.hop_limit 	= 0;
237 		packet->mad.hdr.traffic_class	= 0;
238 		memset(packet->mad.hdr.gid, 0, 16);
239 		packet->mad.hdr.flow_label	= 0;
240 	}
241 
242 	if (queue_packet(file, agent, packet))
243 		goto err2;
244 	return;
245 
246 err2:
247 	kfree(packet);
248 err1:
249 	ib_free_recv_mad(mad_recv_wc);
250 }
251 
252 static ssize_t copy_recv_mad(char __user *buf, struct ib_umad_packet *packet,
253 			     size_t count)
254 {
255 	struct ib_mad_recv_buf *recv_buf;
256 	int left, seg_payload, offset, max_seg_payload;
257 
258 	/* We need enough room to copy the first (or only) MAD segment. */
259 	recv_buf = &packet->recv_wc->recv_buf;
260 	if ((packet->length <= sizeof (*recv_buf->mad) &&
261 	     count < sizeof (packet->mad) + packet->length) ||
262 	    (packet->length > sizeof (*recv_buf->mad) &&
263 	     count < sizeof (packet->mad) + sizeof (*recv_buf->mad)))
264 		return -EINVAL;
265 
266 	if (copy_to_user(buf, &packet->mad, sizeof (packet->mad)))
267 		return -EFAULT;
268 
269 	buf += sizeof (packet->mad);
270 	seg_payload = min_t(int, packet->length, sizeof (*recv_buf->mad));
271 	if (copy_to_user(buf, recv_buf->mad, seg_payload))
272 		return -EFAULT;
273 
274 	if (seg_payload < packet->length) {
275 		/*
276 		 * Multipacket RMPP MAD message. Copy remainder of message.
277 		 * Note that last segment may have a shorter payload.
278 		 */
279 		if (count < sizeof (packet->mad) + packet->length) {
280 			/*
281 			 * The buffer is too small, return the first RMPP segment,
282 			 * which includes the RMPP message length.
283 			 */
284 			return -ENOSPC;
285 		}
286 		offset = ib_get_mad_data_offset(recv_buf->mad->mad_hdr.mgmt_class);
287 		max_seg_payload = sizeof (struct ib_mad) - offset;
288 
289 		for (left = packet->length - seg_payload, buf += seg_payload;
290 		     left; left -= seg_payload, buf += seg_payload) {
291 			recv_buf = container_of(recv_buf->list.next,
292 						struct ib_mad_recv_buf, list);
293 			seg_payload = min(left, max_seg_payload);
294 			if (copy_to_user(buf, ((void *) recv_buf->mad) + offset,
295 					 seg_payload))
296 				return -EFAULT;
297 		}
298 	}
299 	return sizeof (packet->mad) + packet->length;
300 }
301 
302 static ssize_t copy_send_mad(char __user *buf, struct ib_umad_packet *packet,
303 			     size_t count)
304 {
305 	ssize_t size = sizeof (packet->mad) + packet->length;
306 
307 	if (count < size)
308 		return -EINVAL;
309 
310 	if (copy_to_user(buf, &packet->mad, size))
311 		return -EFAULT;
312 
313 	return size;
314 }
315 
316 static ssize_t ib_umad_read(struct file *filp, char __user *buf,
317 			    size_t count, loff_t *pos)
318 {
319 	struct ib_umad_file *file = filp->private_data;
320 	struct ib_umad_packet *packet;
321 	ssize_t ret;
322 
323 	if (count < sizeof (struct ib_user_mad))
324 		return -EINVAL;
325 
326 	spin_lock_irq(&file->recv_lock);
327 
328 	while (list_empty(&file->recv_list)) {
329 		spin_unlock_irq(&file->recv_lock);
330 
331 		if (filp->f_flags & O_NONBLOCK)
332 			return -EAGAIN;
333 
334 		if (wait_event_interruptible(file->recv_wait,
335 					     !list_empty(&file->recv_list)))
336 			return -ERESTARTSYS;
337 
338 		spin_lock_irq(&file->recv_lock);
339 	}
340 
341 	packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
342 	list_del(&packet->list);
343 
344 	spin_unlock_irq(&file->recv_lock);
345 
346 	if (packet->recv_wc)
347 		ret = copy_recv_mad(buf, packet, count);
348 	else
349 		ret = copy_send_mad(buf, packet, count);
350 
351 	if (ret < 0) {
352 		/* Requeue packet */
353 		spin_lock_irq(&file->recv_lock);
354 		list_add(&packet->list, &file->recv_list);
355 		spin_unlock_irq(&file->recv_lock);
356 	} else {
357 		if (packet->recv_wc)
358 			ib_free_recv_mad(packet->recv_wc);
359 		kfree(packet);
360 	}
361 	return ret;
362 }
363 
364 static int copy_rmpp_mad(struct ib_mad_send_buf *msg, const char __user *buf)
365 {
366 	int left, seg;
367 
368 	/* Copy class specific header */
369 	if ((msg->hdr_len > IB_MGMT_RMPP_HDR) &&
370 	    copy_from_user(msg->mad + IB_MGMT_RMPP_HDR, buf + IB_MGMT_RMPP_HDR,
371 			   msg->hdr_len - IB_MGMT_RMPP_HDR))
372 		return -EFAULT;
373 
374 	/* All headers are in place.  Copy data segments. */
375 	for (seg = 1, left = msg->data_len, buf += msg->hdr_len; left > 0;
376 	     seg++, left -= msg->seg_size, buf += msg->seg_size) {
377 		if (copy_from_user(ib_get_rmpp_segment(msg, seg), buf,
378 				   min(left, msg->seg_size)))
379 			return -EFAULT;
380 	}
381 	return 0;
382 }
383 
384 static int same_destination(struct ib_user_mad_hdr *hdr1,
385 			    struct ib_user_mad_hdr *hdr2)
386 {
387 	if (!hdr1->grh_present && !hdr2->grh_present)
388 	   return (hdr1->lid == hdr2->lid);
389 
390 	if (hdr1->grh_present && hdr2->grh_present)
391 	   return !memcmp(hdr1->gid, hdr2->gid, 16);
392 
393 	return 0;
394 }
395 
396 static int is_duplicate(struct ib_umad_file *file,
397 			struct ib_umad_packet *packet)
398 {
399 	struct ib_umad_packet *sent_packet;
400 	struct ib_mad_hdr *sent_hdr, *hdr;
401 
402 	hdr = (struct ib_mad_hdr *) packet->mad.data;
403 	list_for_each_entry(sent_packet, &file->send_list, list) {
404 		sent_hdr = (struct ib_mad_hdr *) sent_packet->mad.data;
405 
406 		if ((hdr->tid != sent_hdr->tid) ||
407 		    (hdr->mgmt_class != sent_hdr->mgmt_class))
408 			continue;
409 
410 		/*
411 		 * No need to be overly clever here.  If two new operations have
412 		 * the same TID, reject the second as a duplicate.  This is more
413 		 * restrictive than required by the spec.
414 		 */
415 		if (!ib_response_mad((struct ib_mad *) hdr)) {
416 			if (!ib_response_mad((struct ib_mad *) sent_hdr))
417 				return 1;
418 			continue;
419 		} else if (!ib_response_mad((struct ib_mad *) sent_hdr))
420 			continue;
421 
422 		if (same_destination(&packet->mad.hdr, &sent_packet->mad.hdr))
423 			return 1;
424 	}
425 
426 	return 0;
427 }
428 
429 static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
430 			     size_t count, loff_t *pos)
431 {
432 	struct ib_umad_file *file = filp->private_data;
433 	struct ib_umad_packet *packet;
434 	struct ib_mad_agent *agent;
435 	struct ib_ah_attr ah_attr;
436 	struct ib_ah *ah;
437 	struct ib_rmpp_mad *rmpp_mad;
438 	__be64 *tid;
439 	int ret, data_len, hdr_len, copy_offset, rmpp_active;
440 
441 	if (count < sizeof (struct ib_user_mad) + IB_MGMT_RMPP_HDR)
442 		return -EINVAL;
443 
444 	packet = kzalloc(sizeof *packet + IB_MGMT_RMPP_HDR, GFP_KERNEL);
445 	if (!packet)
446 		return -ENOMEM;
447 
448 	if (copy_from_user(&packet->mad, buf,
449 			    sizeof (struct ib_user_mad) + IB_MGMT_RMPP_HDR)) {
450 		ret = -EFAULT;
451 		goto err;
452 	}
453 
454 	if (packet->mad.hdr.id < 0 ||
455 	    packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
456 		ret = -EINVAL;
457 		goto err;
458 	}
459 
460 	down_read(&file->port->mutex);
461 
462 	agent = __get_agent(file, packet->mad.hdr.id);
463 	if (!agent) {
464 		ret = -EINVAL;
465 		goto err_up;
466 	}
467 
468 	memset(&ah_attr, 0, sizeof ah_attr);
469 	ah_attr.dlid          = be16_to_cpu(packet->mad.hdr.lid);
470 	ah_attr.sl            = packet->mad.hdr.sl;
471 	ah_attr.src_path_bits = packet->mad.hdr.path_bits;
472 	ah_attr.port_num      = file->port->port_num;
473 	if (packet->mad.hdr.grh_present) {
474 		ah_attr.ah_flags = IB_AH_GRH;
475 		memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
476 		ah_attr.grh.flow_label 	   = be32_to_cpu(packet->mad.hdr.flow_label);
477 		ah_attr.grh.hop_limit  	   = packet->mad.hdr.hop_limit;
478 		ah_attr.grh.traffic_class  = packet->mad.hdr.traffic_class;
479 	}
480 
481 	ah = ib_create_ah(agent->qp->pd, &ah_attr);
482 	if (IS_ERR(ah)) {
483 		ret = PTR_ERR(ah);
484 		goto err_up;
485 	}
486 
487 	rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
488 	hdr_len = ib_get_mad_data_offset(rmpp_mad->mad_hdr.mgmt_class);
489 	if (!ib_is_mad_class_rmpp(rmpp_mad->mad_hdr.mgmt_class)) {
490 		copy_offset = IB_MGMT_MAD_HDR;
491 		rmpp_active = 0;
492 	} else {
493 		copy_offset = IB_MGMT_RMPP_HDR;
494 		rmpp_active = ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
495 			      IB_MGMT_RMPP_FLAG_ACTIVE;
496 	}
497 
498 	data_len = count - sizeof (struct ib_user_mad) - hdr_len;
499 	packet->msg = ib_create_send_mad(agent,
500 					 be32_to_cpu(packet->mad.hdr.qpn),
501 					 0, rmpp_active, hdr_len,
502 					 data_len, GFP_KERNEL);
503 	if (IS_ERR(packet->msg)) {
504 		ret = PTR_ERR(packet->msg);
505 		goto err_ah;
506 	}
507 
508 	packet->msg->ah 	= ah;
509 	packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
510 	packet->msg->retries 	= packet->mad.hdr.retries;
511 	packet->msg->context[0] = packet;
512 
513 	/* Copy MAD header.  Any RMPP header is already in place. */
514 	memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
515 	buf += sizeof (struct ib_user_mad);
516 
517 	if (!rmpp_active) {
518 		if (copy_from_user(packet->msg->mad + copy_offset,
519 				   buf + copy_offset,
520 				   hdr_len + data_len - copy_offset)) {
521 			ret = -EFAULT;
522 			goto err_msg;
523 		}
524 	} else {
525 		ret = copy_rmpp_mad(packet->msg, buf);
526 		if (ret)
527 			goto err_msg;
528 	}
529 
530 	/*
531 	 * Set the high-order part of the transaction ID to make MADs from
532 	 * different agents unique, and allow routing responses back to the
533 	 * original requestor.
534 	 */
535 	if (!ib_response_mad(packet->msg->mad)) {
536 		tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
537 		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
538 				   (be64_to_cpup(tid) & 0xffffffff));
539 		rmpp_mad->mad_hdr.tid = *tid;
540 	}
541 
542 	spin_lock_irq(&file->send_lock);
543 	ret = is_duplicate(file, packet);
544 	if (!ret)
545 		list_add_tail(&packet->list, &file->send_list);
546 	spin_unlock_irq(&file->send_lock);
547 	if (ret) {
548 		ret = -EINVAL;
549 		goto err_msg;
550 	}
551 
552 	ret = ib_post_send_mad(packet->msg, NULL);
553 	if (ret)
554 		goto err_send;
555 
556 	up_read(&file->port->mutex);
557 	return count;
558 
559 err_send:
560 	dequeue_send(file, packet);
561 err_msg:
562 	ib_free_send_mad(packet->msg);
563 err_ah:
564 	ib_destroy_ah(ah);
565 err_up:
566 	up_read(&file->port->mutex);
567 err:
568 	kfree(packet);
569 	return ret;
570 }
571 
572 static unsigned int ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
573 {
574 	struct ib_umad_file *file = filp->private_data;
575 
576 	/* we will always be able to post a MAD send */
577 	unsigned int mask = POLLOUT | POLLWRNORM;
578 
579 	poll_wait(filp, &file->recv_wait, wait);
580 
581 	if (!list_empty(&file->recv_list))
582 		mask |= POLLIN | POLLRDNORM;
583 
584 	return mask;
585 }
586 
587 static int ib_umad_reg_agent(struct ib_umad_file *file, unsigned long arg)
588 {
589 	struct ib_user_mad_reg_req ureq;
590 	struct ib_mad_reg_req req;
591 	struct ib_mad_agent *agent;
592 	int agent_id;
593 	int ret;
594 
595 	down_write(&file->port->mutex);
596 
597 	if (!file->port->ib_dev) {
598 		ret = -EPIPE;
599 		goto out;
600 	}
601 
602 	if (copy_from_user(&ureq, (void __user *) arg, sizeof ureq)) {
603 		ret = -EFAULT;
604 		goto out;
605 	}
606 
607 	if (ureq.qpn != 0 && ureq.qpn != 1) {
608 		ret = -EINVAL;
609 		goto out;
610 	}
611 
612 	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
613 		if (!__get_agent(file, agent_id))
614 			goto found;
615 
616 	ret = -ENOMEM;
617 	goto out;
618 
619 found:
620 	if (ureq.mgmt_class) {
621 		req.mgmt_class         = ureq.mgmt_class;
622 		req.mgmt_class_version = ureq.mgmt_class_version;
623 		memcpy(req.method_mask, ureq.method_mask, sizeof req.method_mask);
624 		memcpy(req.oui,         ureq.oui,         sizeof req.oui);
625 	}
626 
627 	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
628 				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
629 				      ureq.mgmt_class ? &req : NULL,
630 				      ureq.rmpp_version,
631 				      send_handler, recv_handler, file);
632 	if (IS_ERR(agent)) {
633 		ret = PTR_ERR(agent);
634 		goto out;
635 	}
636 
637 	if (put_user(agent_id,
638 		     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
639 		ret = -EFAULT;
640 		ib_unregister_mad_agent(agent);
641 		goto out;
642 	}
643 
644 	file->agent[agent_id] = agent;
645 	ret = 0;
646 
647 out:
648 	up_write(&file->port->mutex);
649 	return ret;
650 }
651 
652 static int ib_umad_unreg_agent(struct ib_umad_file *file, unsigned long arg)
653 {
654 	struct ib_mad_agent *agent = NULL;
655 	u32 id;
656 	int ret = 0;
657 
658 	if (get_user(id, (u32 __user *) arg))
659 		return -EFAULT;
660 
661 	down_write(&file->port->mutex);
662 
663 	if (id < 0 || id >= IB_UMAD_MAX_AGENTS || !__get_agent(file, id)) {
664 		ret = -EINVAL;
665 		goto out;
666 	}
667 
668 	agent = file->agent[id];
669 	file->agent[id] = NULL;
670 
671 out:
672 	up_write(&file->port->mutex);
673 
674 	if (agent)
675 		ib_unregister_mad_agent(agent);
676 
677 	return ret;
678 }
679 
680 static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
681 			  unsigned long arg)
682 {
683 	switch (cmd) {
684 	case IB_USER_MAD_REGISTER_AGENT:
685 		return ib_umad_reg_agent(filp->private_data, arg);
686 	case IB_USER_MAD_UNREGISTER_AGENT:
687 		return ib_umad_unreg_agent(filp->private_data, arg);
688 	default:
689 		return -ENOIOCTLCMD;
690 	}
691 }
692 
693 static int ib_umad_open(struct inode *inode, struct file *filp)
694 {
695 	struct ib_umad_port *port;
696 	struct ib_umad_file *file;
697 	int ret = 0;
698 
699 	spin_lock(&port_lock);
700 	port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE];
701 	if (port)
702 		kref_get(&port->umad_dev->ref);
703 	spin_unlock(&port_lock);
704 
705 	if (!port)
706 		return -ENXIO;
707 
708 	down_write(&port->mutex);
709 
710 	if (!port->ib_dev) {
711 		ret = -ENXIO;
712 		goto out;
713 	}
714 
715 	file = kzalloc(sizeof *file, GFP_KERNEL);
716 	if (!file) {
717 		kref_put(&port->umad_dev->ref, ib_umad_release_dev);
718 		ret = -ENOMEM;
719 		goto out;
720 	}
721 
722 	spin_lock_init(&file->recv_lock);
723 	spin_lock_init(&file->send_lock);
724 	INIT_LIST_HEAD(&file->recv_list);
725 	INIT_LIST_HEAD(&file->send_list);
726 	init_waitqueue_head(&file->recv_wait);
727 
728 	file->port = port;
729 	filp->private_data = file;
730 
731 	list_add_tail(&file->port_list, &port->file_list);
732 
733 out:
734 	up_write(&port->mutex);
735 	return ret;
736 }
737 
738 static int ib_umad_close(struct inode *inode, struct file *filp)
739 {
740 	struct ib_umad_file *file = filp->private_data;
741 	struct ib_umad_device *dev = file->port->umad_dev;
742 	struct ib_umad_packet *packet, *tmp;
743 	int already_dead;
744 	int i;
745 
746 	down_write(&file->port->mutex);
747 
748 	already_dead = file->agents_dead;
749 	file->agents_dead = 1;
750 
751 	list_for_each_entry_safe(packet, tmp, &file->recv_list, list) {
752 		if (packet->recv_wc)
753 			ib_free_recv_mad(packet->recv_wc);
754 		kfree(packet);
755 	}
756 
757 	list_del(&file->port_list);
758 
759 	downgrade_write(&file->port->mutex);
760 
761 	if (!already_dead)
762 		for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
763 			if (file->agent[i])
764 				ib_unregister_mad_agent(file->agent[i]);
765 
766 	up_read(&file->port->mutex);
767 
768 	kfree(file);
769 	kref_put(&dev->ref, ib_umad_release_dev);
770 
771 	return 0;
772 }
773 
774 static const struct file_operations umad_fops = {
775 	.owner 	 	= THIS_MODULE,
776 	.read 	 	= ib_umad_read,
777 	.write 	 	= ib_umad_write,
778 	.poll 	 	= ib_umad_poll,
779 	.unlocked_ioctl = ib_umad_ioctl,
780 	.compat_ioctl 	= ib_umad_ioctl,
781 	.open 	 	= ib_umad_open,
782 	.release 	= ib_umad_close
783 };
784 
785 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
786 {
787 	struct ib_umad_port *port;
788 	struct ib_port_modify props = {
789 		.set_port_cap_mask = IB_PORT_SM
790 	};
791 	int ret;
792 
793 	spin_lock(&port_lock);
794 	port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE - IB_UMAD_MAX_PORTS];
795 	if (port)
796 		kref_get(&port->umad_dev->ref);
797 	spin_unlock(&port_lock);
798 
799 	if (!port)
800 		return -ENXIO;
801 
802 	if (filp->f_flags & O_NONBLOCK) {
803 		if (down_trylock(&port->sm_sem)) {
804 			ret = -EAGAIN;
805 			goto fail;
806 		}
807 	} else {
808 		if (down_interruptible(&port->sm_sem)) {
809 			ret = -ERESTARTSYS;
810 			goto fail;
811 		}
812 	}
813 
814 	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
815 	if (ret) {
816 		up(&port->sm_sem);
817 		goto fail;
818 	}
819 
820 	filp->private_data = port;
821 
822 	return 0;
823 
824 fail:
825 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
826 	return ret;
827 }
828 
829 static int ib_umad_sm_close(struct inode *inode, struct file *filp)
830 {
831 	struct ib_umad_port *port = filp->private_data;
832 	struct ib_port_modify props = {
833 		.clr_port_cap_mask = IB_PORT_SM
834 	};
835 	int ret = 0;
836 
837 	down_write(&port->mutex);
838 	if (port->ib_dev)
839 		ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
840 	up_write(&port->mutex);
841 
842 	up(&port->sm_sem);
843 
844 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
845 
846 	return ret;
847 }
848 
849 static const struct file_operations umad_sm_fops = {
850 	.owner 	 = THIS_MODULE,
851 	.open 	 = ib_umad_sm_open,
852 	.release = ib_umad_sm_close
853 };
854 
855 static struct ib_client umad_client = {
856 	.name   = "umad",
857 	.add    = ib_umad_add_one,
858 	.remove = ib_umad_remove_one
859 };
860 
861 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
862 {
863 	struct ib_umad_port *port = class_get_devdata(class_dev);
864 
865 	if (!port)
866 		return -ENODEV;
867 
868 	return sprintf(buf, "%s\n", port->ib_dev->name);
869 }
870 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
871 
872 static ssize_t show_port(struct class_device *class_dev, char *buf)
873 {
874 	struct ib_umad_port *port = class_get_devdata(class_dev);
875 
876 	if (!port)
877 		return -ENODEV;
878 
879 	return sprintf(buf, "%d\n", port->port_num);
880 }
881 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
882 
883 static ssize_t show_abi_version(struct class *class, char *buf)
884 {
885 	return sprintf(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
886 }
887 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
888 
889 static int ib_umad_init_port(struct ib_device *device, int port_num,
890 			     struct ib_umad_port *port)
891 {
892 	spin_lock(&port_lock);
893 	port->dev_num = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
894 	if (port->dev_num >= IB_UMAD_MAX_PORTS) {
895 		spin_unlock(&port_lock);
896 		return -1;
897 	}
898 	set_bit(port->dev_num, dev_map);
899 	spin_unlock(&port_lock);
900 
901 	port->ib_dev   = device;
902 	port->port_num = port_num;
903 	init_MUTEX(&port->sm_sem);
904 	init_rwsem(&port->mutex);
905 	INIT_LIST_HEAD(&port->file_list);
906 
907 	port->dev = cdev_alloc();
908 	if (!port->dev)
909 		return -1;
910 	port->dev->owner = THIS_MODULE;
911 	port->dev->ops   = &umad_fops;
912 	kobject_set_name(&port->dev->kobj, "umad%d", port->dev_num);
913 	if (cdev_add(port->dev, base_dev + port->dev_num, 1))
914 		goto err_cdev;
915 
916 	port->class_dev = class_device_create(umad_class, NULL, port->dev->dev,
917 					      device->dma_device,
918 					      "umad%d", port->dev_num);
919 	if (IS_ERR(port->class_dev))
920 		goto err_cdev;
921 
922 	if (class_device_create_file(port->class_dev, &class_device_attr_ibdev))
923 		goto err_class;
924 	if (class_device_create_file(port->class_dev, &class_device_attr_port))
925 		goto err_class;
926 
927 	port->sm_dev = cdev_alloc();
928 	if (!port->sm_dev)
929 		goto err_class;
930 	port->sm_dev->owner = THIS_MODULE;
931 	port->sm_dev->ops   = &umad_sm_fops;
932 	kobject_set_name(&port->sm_dev->kobj, "issm%d", port->dev_num);
933 	if (cdev_add(port->sm_dev, base_dev + port->dev_num + IB_UMAD_MAX_PORTS, 1))
934 		goto err_sm_cdev;
935 
936 	port->sm_class_dev = class_device_create(umad_class, NULL, port->sm_dev->dev,
937 						 device->dma_device,
938 						 "issm%d", port->dev_num);
939 	if (IS_ERR(port->sm_class_dev))
940 		goto err_sm_cdev;
941 
942 	class_set_devdata(port->class_dev,    port);
943 	class_set_devdata(port->sm_class_dev, port);
944 
945 	if (class_device_create_file(port->sm_class_dev, &class_device_attr_ibdev))
946 		goto err_sm_class;
947 	if (class_device_create_file(port->sm_class_dev, &class_device_attr_port))
948 		goto err_sm_class;
949 
950 	spin_lock(&port_lock);
951 	umad_port[port->dev_num] = port;
952 	spin_unlock(&port_lock);
953 
954 	return 0;
955 
956 err_sm_class:
957 	class_device_destroy(umad_class, port->sm_dev->dev);
958 
959 err_sm_cdev:
960 	cdev_del(port->sm_dev);
961 
962 err_class:
963 	class_device_destroy(umad_class, port->dev->dev);
964 
965 err_cdev:
966 	cdev_del(port->dev);
967 	clear_bit(port->dev_num, dev_map);
968 
969 	return -1;
970 }
971 
972 static void ib_umad_kill_port(struct ib_umad_port *port)
973 {
974 	struct ib_umad_file *file;
975 	int id;
976 
977 	class_set_devdata(port->class_dev,    NULL);
978 	class_set_devdata(port->sm_class_dev, NULL);
979 
980 	class_device_destroy(umad_class, port->dev->dev);
981 	class_device_destroy(umad_class, port->sm_dev->dev);
982 
983 	cdev_del(port->dev);
984 	cdev_del(port->sm_dev);
985 
986 	spin_lock(&port_lock);
987 	umad_port[port->dev_num] = NULL;
988 	spin_unlock(&port_lock);
989 
990 	down_write(&port->mutex);
991 
992 	port->ib_dev = NULL;
993 
994 	/*
995 	 * Now go through the list of files attached to this port and
996 	 * unregister all of their MAD agents.  We need to hold
997 	 * port->mutex while doing this to avoid racing with
998 	 * ib_umad_close(), but we can't hold the mutex for writing
999 	 * while calling ib_unregister_mad_agent(), since that might
1000 	 * deadlock by calling back into queue_packet().  So we
1001 	 * downgrade our lock to a read lock, and then drop and
1002 	 * reacquire the write lock for the next iteration.
1003 	 *
1004 	 * We do list_del_init() on the file's list_head so that the
1005 	 * list_del in ib_umad_close() is still OK, even after the
1006 	 * file is removed from the list.
1007 	 */
1008 	while (!list_empty(&port->file_list)) {
1009 		file = list_entry(port->file_list.next, struct ib_umad_file,
1010 				  port_list);
1011 
1012 		file->agents_dead = 1;
1013 		list_del_init(&file->port_list);
1014 
1015 		downgrade_write(&port->mutex);
1016 
1017 		for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id)
1018 			if (file->agent[id])
1019 				ib_unregister_mad_agent(file->agent[id]);
1020 
1021 		up_read(&port->mutex);
1022 		down_write(&port->mutex);
1023 	}
1024 
1025 	up_write(&port->mutex);
1026 
1027 	clear_bit(port->dev_num, dev_map);
1028 }
1029 
1030 static void ib_umad_add_one(struct ib_device *device)
1031 {
1032 	struct ib_umad_device *umad_dev;
1033 	int s, e, i;
1034 
1035 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1036 		return;
1037 
1038 	if (device->node_type == RDMA_NODE_IB_SWITCH)
1039 		s = e = 0;
1040 	else {
1041 		s = 1;
1042 		e = device->phys_port_cnt;
1043 	}
1044 
1045 	umad_dev = kzalloc(sizeof *umad_dev +
1046 			   (e - s + 1) * sizeof (struct ib_umad_port),
1047 			   GFP_KERNEL);
1048 	if (!umad_dev)
1049 		return;
1050 
1051 	kref_init(&umad_dev->ref);
1052 
1053 	umad_dev->start_port = s;
1054 	umad_dev->end_port   = e;
1055 
1056 	for (i = s; i <= e; ++i) {
1057 		umad_dev->port[i - s].umad_dev = umad_dev;
1058 
1059 		if (ib_umad_init_port(device, i, &umad_dev->port[i - s]))
1060 			goto err;
1061 	}
1062 
1063 	ib_set_client_data(device, &umad_client, umad_dev);
1064 
1065 	return;
1066 
1067 err:
1068 	while (--i >= s)
1069 		ib_umad_kill_port(&umad_dev->port[i - s]);
1070 
1071 	kref_put(&umad_dev->ref, ib_umad_release_dev);
1072 }
1073 
1074 static void ib_umad_remove_one(struct ib_device *device)
1075 {
1076 	struct ib_umad_device *umad_dev = ib_get_client_data(device, &umad_client);
1077 	int i;
1078 
1079 	if (!umad_dev)
1080 		return;
1081 
1082 	for (i = 0; i <= umad_dev->end_port - umad_dev->start_port; ++i)
1083 		ib_umad_kill_port(&umad_dev->port[i]);
1084 
1085 	kref_put(&umad_dev->ref, ib_umad_release_dev);
1086 }
1087 
1088 static int __init ib_umad_init(void)
1089 {
1090 	int ret;
1091 
1092 	ret = register_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2,
1093 				     "infiniband_mad");
1094 	if (ret) {
1095 		printk(KERN_ERR "user_mad: couldn't register device number\n");
1096 		goto out;
1097 	}
1098 
1099 	umad_class = class_create(THIS_MODULE, "infiniband_mad");
1100 	if (IS_ERR(umad_class)) {
1101 		ret = PTR_ERR(umad_class);
1102 		printk(KERN_ERR "user_mad: couldn't create class infiniband_mad\n");
1103 		goto out_chrdev;
1104 	}
1105 
1106 	ret = class_create_file(umad_class, &class_attr_abi_version);
1107 	if (ret) {
1108 		printk(KERN_ERR "user_mad: couldn't create abi_version attribute\n");
1109 		goto out_class;
1110 	}
1111 
1112 	ret = ib_register_client(&umad_client);
1113 	if (ret) {
1114 		printk(KERN_ERR "user_mad: couldn't register ib_umad client\n");
1115 		goto out_class;
1116 	}
1117 
1118 	return 0;
1119 
1120 out_class:
1121 	class_destroy(umad_class);
1122 
1123 out_chrdev:
1124 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
1125 
1126 out:
1127 	return ret;
1128 }
1129 
1130 static void __exit ib_umad_cleanup(void)
1131 {
1132 	ib_unregister_client(&umad_client);
1133 	class_destroy(umad_class);
1134 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
1135 }
1136 
1137 module_init(ib_umad_init);
1138 module_exit(ib_umad_cleanup);
1139