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  * $Id: user_mad.c 5596 2006-03-03 01:00:07Z sean.hefty $
36  */
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 
50 #include <asm/uaccess.h>
51 #include <asm/semaphore.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: each time a
70  * device special file is opened, we look up the corresponding struct
71  * ib_umad_port by minor in the umad_port[] table while holding the
72  * port_lock.  If this lookup succeeds, we take a reference on the
73  * ib_umad_port's struct ib_umad_device while still holding the
74  * port_lock; if the lookup fails, we fail the open().  We drop these
75  * references in the corresponding close().
76  *
77  * In addition to references coming from open character devices, there
78  * is one more reference to each ib_umad_device representing the
79  * module's reference taken when allocating the ib_umad_device in
80  * ib_umad_add_one().
81  *
82  * When destroying an ib_umad_device, we clear all of its
83  * ib_umad_ports from umad_port[] while holding port_lock before
84  * dropping the module's reference to the ib_umad_device.  This is
85  * always safe because any open() calls will either succeed and obtain
86  * a reference before we clear the umad_port[] entries, or fail after
87  * we clear the umad_port[] entries.
88  */
89 
90 struct ib_umad_port {
91 	struct cdev           *dev;
92 	struct class_device   *class_dev;
93 
94 	struct cdev           *sm_dev;
95 	struct class_device   *sm_class_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 	u8                     port_num;
105 };
106 
107 struct ib_umad_device {
108 	int                  start_port, end_port;
109 	struct kref          ref;
110 	struct ib_umad_port  port[0];
111 };
112 
113 struct ib_umad_file {
114 	struct mutex		mutex;
115 	struct ib_umad_port    *port;
116 	struct list_head	recv_list;
117 	struct list_head	send_list;
118 	struct list_head	port_list;
119 	spinlock_t		send_lock;
120 	wait_queue_head_t	recv_wait;
121 	struct ib_mad_agent    *agent[IB_UMAD_MAX_AGENTS];
122 	int			agents_dead;
123 	u8			use_pkey_index;
124 	u8			already_used;
125 };
126 
127 struct ib_umad_packet {
128 	struct ib_mad_send_buf *msg;
129 	struct ib_mad_recv_wc  *recv_wc;
130 	struct list_head   list;
131 	int		   length;
132 	struct ib_user_mad mad;
133 };
134 
135 static struct class *umad_class;
136 
137 static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
138 
139 static DEFINE_SPINLOCK(port_lock);
140 static struct ib_umad_port *umad_port[IB_UMAD_MAX_PORTS];
141 static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS);
142 
143 static void ib_umad_add_one(struct ib_device *device);
144 static void ib_umad_remove_one(struct ib_device *device);
145 
146 static void ib_umad_release_dev(struct kref *ref)
147 {
148 	struct ib_umad_device *dev =
149 		container_of(ref, struct ib_umad_device, ref);
150 
151 	kfree(dev);
152 }
153 
154 static int hdr_size(struct ib_umad_file *file)
155 {
156 	return file->use_pkey_index ? sizeof (struct ib_user_mad_hdr) :
157 		sizeof (struct ib_user_mad_hdr_old);
158 }
159 
160 /* caller must hold file->mutex */
161 static struct ib_mad_agent *__get_agent(struct ib_umad_file *file, int id)
162 {
163 	return file->agents_dead ? NULL : file->agent[id];
164 }
165 
166 static int queue_packet(struct ib_umad_file *file,
167 			struct ib_mad_agent *agent,
168 			struct ib_umad_packet *packet)
169 {
170 	int ret = 1;
171 
172 	mutex_lock(&file->mutex);
173 
174 	for (packet->mad.hdr.id = 0;
175 	     packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
176 	     packet->mad.hdr.id++)
177 		if (agent == __get_agent(file, packet->mad.hdr.id)) {
178 			list_add_tail(&packet->list, &file->recv_list);
179 			wake_up_interruptible(&file->recv_wait);
180 			ret = 0;
181 			break;
182 		}
183 
184 	mutex_unlock(&file->mutex);
185 
186 	return ret;
187 }
188 
189 static void dequeue_send(struct ib_umad_file *file,
190 			 struct ib_umad_packet *packet)
191 {
192 	spin_lock_irq(&file->send_lock);
193 	list_del(&packet->list);
194 	spin_unlock_irq(&file->send_lock);
195 }
196 
197 static void send_handler(struct ib_mad_agent *agent,
198 			 struct ib_mad_send_wc *send_wc)
199 {
200 	struct ib_umad_file *file = agent->context;
201 	struct ib_umad_packet *packet = send_wc->send_buf->context[0];
202 
203 	dequeue_send(file, packet);
204 	ib_destroy_ah(packet->msg->ah);
205 	ib_free_send_mad(packet->msg);
206 
207 	if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
208 		packet->length = IB_MGMT_MAD_HDR;
209 		packet->mad.hdr.status = ETIMEDOUT;
210 		if (!queue_packet(file, agent, packet))
211 			return;
212 	}
213 	kfree(packet);
214 }
215 
216 static void recv_handler(struct ib_mad_agent *agent,
217 			 struct ib_mad_recv_wc *mad_recv_wc)
218 {
219 	struct ib_umad_file *file = agent->context;
220 	struct ib_umad_packet *packet;
221 
222 	if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
223 		goto err1;
224 
225 	packet = kzalloc(sizeof *packet, GFP_KERNEL);
226 	if (!packet)
227 		goto err1;
228 
229 	packet->length = mad_recv_wc->mad_len;
230 	packet->recv_wc = mad_recv_wc;
231 
232 	packet->mad.hdr.status	   = 0;
233 	packet->mad.hdr.length	   = hdr_size(file) + mad_recv_wc->mad_len;
234 	packet->mad.hdr.qpn	   = cpu_to_be32(mad_recv_wc->wc->src_qp);
235 	packet->mad.hdr.lid	   = cpu_to_be16(mad_recv_wc->wc->slid);
236 	packet->mad.hdr.sl	   = mad_recv_wc->wc->sl;
237 	packet->mad.hdr.path_bits  = mad_recv_wc->wc->dlid_path_bits;
238 	packet->mad.hdr.pkey_index = mad_recv_wc->wc->pkey_index;
239 	packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
240 	if (packet->mad.hdr.grh_present) {
241 		struct ib_ah_attr ah_attr;
242 
243 		ib_init_ah_from_wc(agent->device, agent->port_num,
244 				   mad_recv_wc->wc, mad_recv_wc->recv_buf.grh,
245 				   &ah_attr);
246 
247 		packet->mad.hdr.gid_index = ah_attr.grh.sgid_index;
248 		packet->mad.hdr.hop_limit = ah_attr.grh.hop_limit;
249 		packet->mad.hdr.traffic_class = ah_attr.grh.traffic_class;
250 		memcpy(packet->mad.hdr.gid, &ah_attr.grh.dgid, 16);
251 		packet->mad.hdr.flow_label = cpu_to_be32(ah_attr.grh.flow_label);
252 	}
253 
254 	if (queue_packet(file, agent, packet))
255 		goto err2;
256 	return;
257 
258 err2:
259 	kfree(packet);
260 err1:
261 	ib_free_recv_mad(mad_recv_wc);
262 }
263 
264 static ssize_t copy_recv_mad(struct ib_umad_file *file, char __user *buf,
265 			     struct ib_umad_packet *packet, size_t count)
266 {
267 	struct ib_mad_recv_buf *recv_buf;
268 	int left, seg_payload, offset, max_seg_payload;
269 
270 	/* We need enough room to copy the first (or only) MAD segment. */
271 	recv_buf = &packet->recv_wc->recv_buf;
272 	if ((packet->length <= sizeof (*recv_buf->mad) &&
273 	     count < hdr_size(file) + packet->length) ||
274 	    (packet->length > sizeof (*recv_buf->mad) &&
275 	     count < hdr_size(file) + sizeof (*recv_buf->mad)))
276 		return -EINVAL;
277 
278 	if (copy_to_user(buf, &packet->mad, hdr_size(file)))
279 		return -EFAULT;
280 
281 	buf += hdr_size(file);
282 	seg_payload = min_t(int, packet->length, sizeof (*recv_buf->mad));
283 	if (copy_to_user(buf, recv_buf->mad, seg_payload))
284 		return -EFAULT;
285 
286 	if (seg_payload < packet->length) {
287 		/*
288 		 * Multipacket RMPP MAD message. Copy remainder of message.
289 		 * Note that last segment may have a shorter payload.
290 		 */
291 		if (count < hdr_size(file) + packet->length) {
292 			/*
293 			 * The buffer is too small, return the first RMPP segment,
294 			 * which includes the RMPP message length.
295 			 */
296 			return -ENOSPC;
297 		}
298 		offset = ib_get_mad_data_offset(recv_buf->mad->mad_hdr.mgmt_class);
299 		max_seg_payload = sizeof (struct ib_mad) - offset;
300 
301 		for (left = packet->length - seg_payload, buf += seg_payload;
302 		     left; left -= seg_payload, buf += seg_payload) {
303 			recv_buf = container_of(recv_buf->list.next,
304 						struct ib_mad_recv_buf, list);
305 			seg_payload = min(left, max_seg_payload);
306 			if (copy_to_user(buf, ((void *) recv_buf->mad) + offset,
307 					 seg_payload))
308 				return -EFAULT;
309 		}
310 	}
311 	return hdr_size(file) + packet->length;
312 }
313 
314 static ssize_t copy_send_mad(struct ib_umad_file *file, char __user *buf,
315 			     struct ib_umad_packet *packet, size_t count)
316 {
317 	ssize_t size = hdr_size(file) + packet->length;
318 
319 	if (count < size)
320 		return -EINVAL;
321 
322 	if (copy_to_user(buf, &packet->mad, hdr_size(file)))
323 		return -EFAULT;
324 
325 	buf += hdr_size(file);
326 
327 	if (copy_to_user(buf, packet->mad.data, packet->length))
328 		return -EFAULT;
329 
330 	return size;
331 }
332 
333 static ssize_t ib_umad_read(struct file *filp, char __user *buf,
334 			    size_t count, loff_t *pos)
335 {
336 	struct ib_umad_file *file = filp->private_data;
337 	struct ib_umad_packet *packet;
338 	ssize_t ret;
339 
340 	if (count < hdr_size(file))
341 		return -EINVAL;
342 
343 	mutex_lock(&file->mutex);
344 
345 	while (list_empty(&file->recv_list)) {
346 		mutex_unlock(&file->mutex);
347 
348 		if (filp->f_flags & O_NONBLOCK)
349 			return -EAGAIN;
350 
351 		if (wait_event_interruptible(file->recv_wait,
352 					     !list_empty(&file->recv_list)))
353 			return -ERESTARTSYS;
354 
355 		mutex_lock(&file->mutex);
356 	}
357 
358 	packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
359 	list_del(&packet->list);
360 
361 	mutex_unlock(&file->mutex);
362 
363 	if (packet->recv_wc)
364 		ret = copy_recv_mad(file, buf, packet, count);
365 	else
366 		ret = copy_send_mad(file, buf, packet, count);
367 
368 	if (ret < 0) {
369 		/* Requeue packet */
370 		mutex_lock(&file->mutex);
371 		list_add(&packet->list, &file->recv_list);
372 		mutex_unlock(&file->mutex);
373 	} else {
374 		if (packet->recv_wc)
375 			ib_free_recv_mad(packet->recv_wc);
376 		kfree(packet);
377 	}
378 	return ret;
379 }
380 
381 static int copy_rmpp_mad(struct ib_mad_send_buf *msg, const char __user *buf)
382 {
383 	int left, seg;
384 
385 	/* Copy class specific header */
386 	if ((msg->hdr_len > IB_MGMT_RMPP_HDR) &&
387 	    copy_from_user(msg->mad + IB_MGMT_RMPP_HDR, buf + IB_MGMT_RMPP_HDR,
388 			   msg->hdr_len - IB_MGMT_RMPP_HDR))
389 		return -EFAULT;
390 
391 	/* All headers are in place.  Copy data segments. */
392 	for (seg = 1, left = msg->data_len, buf += msg->hdr_len; left > 0;
393 	     seg++, left -= msg->seg_size, buf += msg->seg_size) {
394 		if (copy_from_user(ib_get_rmpp_segment(msg, seg), buf,
395 				   min(left, msg->seg_size)))
396 			return -EFAULT;
397 	}
398 	return 0;
399 }
400 
401 static int same_destination(struct ib_user_mad_hdr *hdr1,
402 			    struct ib_user_mad_hdr *hdr2)
403 {
404 	if (!hdr1->grh_present && !hdr2->grh_present)
405 	   return (hdr1->lid == hdr2->lid);
406 
407 	if (hdr1->grh_present && hdr2->grh_present)
408 	   return !memcmp(hdr1->gid, hdr2->gid, 16);
409 
410 	return 0;
411 }
412 
413 static int is_duplicate(struct ib_umad_file *file,
414 			struct ib_umad_packet *packet)
415 {
416 	struct ib_umad_packet *sent_packet;
417 	struct ib_mad_hdr *sent_hdr, *hdr;
418 
419 	hdr = (struct ib_mad_hdr *) packet->mad.data;
420 	list_for_each_entry(sent_packet, &file->send_list, list) {
421 		sent_hdr = (struct ib_mad_hdr *) sent_packet->mad.data;
422 
423 		if ((hdr->tid != sent_hdr->tid) ||
424 		    (hdr->mgmt_class != sent_hdr->mgmt_class))
425 			continue;
426 
427 		/*
428 		 * No need to be overly clever here.  If two new operations have
429 		 * the same TID, reject the second as a duplicate.  This is more
430 		 * restrictive than required by the spec.
431 		 */
432 		if (!ib_response_mad((struct ib_mad *) hdr)) {
433 			if (!ib_response_mad((struct ib_mad *) sent_hdr))
434 				return 1;
435 			continue;
436 		} else if (!ib_response_mad((struct ib_mad *) sent_hdr))
437 			continue;
438 
439 		if (same_destination(&packet->mad.hdr, &sent_packet->mad.hdr))
440 			return 1;
441 	}
442 
443 	return 0;
444 }
445 
446 static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
447 			     size_t count, loff_t *pos)
448 {
449 	struct ib_umad_file *file = filp->private_data;
450 	struct ib_umad_packet *packet;
451 	struct ib_mad_agent *agent;
452 	struct ib_ah_attr ah_attr;
453 	struct ib_ah *ah;
454 	struct ib_rmpp_mad *rmpp_mad;
455 	__be64 *tid;
456 	int ret, data_len, hdr_len, copy_offset, rmpp_active;
457 
458 	if (count < hdr_size(file) + IB_MGMT_RMPP_HDR)
459 		return -EINVAL;
460 
461 	packet = kzalloc(sizeof *packet + IB_MGMT_RMPP_HDR, GFP_KERNEL);
462 	if (!packet)
463 		return -ENOMEM;
464 
465 	if (copy_from_user(&packet->mad, buf, hdr_size(file))) {
466 		ret = -EFAULT;
467 		goto err;
468 	}
469 
470 	if (packet->mad.hdr.id < 0 ||
471 	    packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
472 		ret = -EINVAL;
473 		goto err;
474 	}
475 
476 	buf += hdr_size(file);
477 
478 	if (copy_from_user(packet->mad.data, buf, IB_MGMT_RMPP_HDR)) {
479 		ret = -EFAULT;
480 		goto err;
481 	}
482 
483 	mutex_lock(&file->mutex);
484 
485 	agent = __get_agent(file, packet->mad.hdr.id);
486 	if (!agent) {
487 		ret = -EINVAL;
488 		goto err_up;
489 	}
490 
491 	memset(&ah_attr, 0, sizeof ah_attr);
492 	ah_attr.dlid          = be16_to_cpu(packet->mad.hdr.lid);
493 	ah_attr.sl            = packet->mad.hdr.sl;
494 	ah_attr.src_path_bits = packet->mad.hdr.path_bits;
495 	ah_attr.port_num      = file->port->port_num;
496 	if (packet->mad.hdr.grh_present) {
497 		ah_attr.ah_flags = IB_AH_GRH;
498 		memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
499 		ah_attr.grh.sgid_index	   = packet->mad.hdr.gid_index;
500 		ah_attr.grh.flow_label 	   = be32_to_cpu(packet->mad.hdr.flow_label);
501 		ah_attr.grh.hop_limit  	   = packet->mad.hdr.hop_limit;
502 		ah_attr.grh.traffic_class  = packet->mad.hdr.traffic_class;
503 	}
504 
505 	ah = ib_create_ah(agent->qp->pd, &ah_attr);
506 	if (IS_ERR(ah)) {
507 		ret = PTR_ERR(ah);
508 		goto err_up;
509 	}
510 
511 	rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
512 	hdr_len = ib_get_mad_data_offset(rmpp_mad->mad_hdr.mgmt_class);
513 	if (!ib_is_mad_class_rmpp(rmpp_mad->mad_hdr.mgmt_class)) {
514 		copy_offset = IB_MGMT_MAD_HDR;
515 		rmpp_active = 0;
516 	} else {
517 		copy_offset = IB_MGMT_RMPP_HDR;
518 		rmpp_active = ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
519 			      IB_MGMT_RMPP_FLAG_ACTIVE;
520 	}
521 
522 	data_len = count - hdr_size(file) - hdr_len;
523 	packet->msg = ib_create_send_mad(agent,
524 					 be32_to_cpu(packet->mad.hdr.qpn),
525 					 packet->mad.hdr.pkey_index, rmpp_active,
526 					 hdr_len, data_len, GFP_KERNEL);
527 	if (IS_ERR(packet->msg)) {
528 		ret = PTR_ERR(packet->msg);
529 		goto err_ah;
530 	}
531 
532 	packet->msg->ah 	= ah;
533 	packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
534 	packet->msg->retries 	= packet->mad.hdr.retries;
535 	packet->msg->context[0] = packet;
536 
537 	/* Copy MAD header.  Any RMPP header is already in place. */
538 	memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
539 
540 	if (!rmpp_active) {
541 		if (copy_from_user(packet->msg->mad + copy_offset,
542 				   buf + copy_offset,
543 				   hdr_len + data_len - copy_offset)) {
544 			ret = -EFAULT;
545 			goto err_msg;
546 		}
547 	} else {
548 		ret = copy_rmpp_mad(packet->msg, buf);
549 		if (ret)
550 			goto err_msg;
551 	}
552 
553 	/*
554 	 * Set the high-order part of the transaction ID to make MADs from
555 	 * different agents unique, and allow routing responses back to the
556 	 * original requestor.
557 	 */
558 	if (!ib_response_mad(packet->msg->mad)) {
559 		tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
560 		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
561 				   (be64_to_cpup(tid) & 0xffffffff));
562 		rmpp_mad->mad_hdr.tid = *tid;
563 	}
564 
565 	spin_lock_irq(&file->send_lock);
566 	ret = is_duplicate(file, packet);
567 	if (!ret)
568 		list_add_tail(&packet->list, &file->send_list);
569 	spin_unlock_irq(&file->send_lock);
570 	if (ret) {
571 		ret = -EINVAL;
572 		goto err_msg;
573 	}
574 
575 	ret = ib_post_send_mad(packet->msg, NULL);
576 	if (ret)
577 		goto err_send;
578 
579 	mutex_unlock(&file->mutex);
580 	return count;
581 
582 err_send:
583 	dequeue_send(file, packet);
584 err_msg:
585 	ib_free_send_mad(packet->msg);
586 err_ah:
587 	ib_destroy_ah(ah);
588 err_up:
589 	mutex_unlock(&file->mutex);
590 err:
591 	kfree(packet);
592 	return ret;
593 }
594 
595 static unsigned int ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
596 {
597 	struct ib_umad_file *file = filp->private_data;
598 
599 	/* we will always be able to post a MAD send */
600 	unsigned int mask = POLLOUT | POLLWRNORM;
601 
602 	poll_wait(filp, &file->recv_wait, wait);
603 
604 	if (!list_empty(&file->recv_list))
605 		mask |= POLLIN | POLLRDNORM;
606 
607 	return mask;
608 }
609 
610 static int ib_umad_reg_agent(struct ib_umad_file *file, void __user *arg,
611 			     int compat_method_mask)
612 {
613 	struct ib_user_mad_reg_req ureq;
614 	struct ib_mad_reg_req req;
615 	struct ib_mad_agent *agent = NULL;
616 	int agent_id;
617 	int ret;
618 
619 	mutex_lock(&file->port->file_mutex);
620 	mutex_lock(&file->mutex);
621 
622 	if (!file->port->ib_dev) {
623 		ret = -EPIPE;
624 		goto out;
625 	}
626 
627 	if (copy_from_user(&ureq, arg, sizeof ureq)) {
628 		ret = -EFAULT;
629 		goto out;
630 	}
631 
632 	if (ureq.qpn != 0 && ureq.qpn != 1) {
633 		ret = -EINVAL;
634 		goto out;
635 	}
636 
637 	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
638 		if (!__get_agent(file, agent_id))
639 			goto found;
640 
641 	ret = -ENOMEM;
642 	goto out;
643 
644 found:
645 	if (ureq.mgmt_class) {
646 		req.mgmt_class         = ureq.mgmt_class;
647 		req.mgmt_class_version = ureq.mgmt_class_version;
648 		memcpy(req.oui, ureq.oui, sizeof req.oui);
649 
650 		if (compat_method_mask) {
651 			u32 *umm = (u32 *) ureq.method_mask;
652 			int i;
653 
654 			for (i = 0; i < BITS_TO_LONGS(IB_MGMT_MAX_METHODS); ++i)
655 				req.method_mask[i] =
656 					umm[i * 2] | ((u64) umm[i * 2 + 1] << 32);
657 		} else
658 			memcpy(req.method_mask, ureq.method_mask,
659 			       sizeof req.method_mask);
660 	}
661 
662 	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
663 				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
664 				      ureq.mgmt_class ? &req : NULL,
665 				      ureq.rmpp_version,
666 				      send_handler, recv_handler, file);
667 	if (IS_ERR(agent)) {
668 		ret = PTR_ERR(agent);
669 		agent = NULL;
670 		goto out;
671 	}
672 
673 	if (put_user(agent_id,
674 		     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
675 		ret = -EFAULT;
676 		goto out;
677 	}
678 
679 	if (!file->already_used) {
680 		file->already_used = 1;
681 		if (!file->use_pkey_index) {
682 			printk(KERN_WARNING "user_mad: process %s did not enable "
683 			       "P_Key index support.\n", current->comm);
684 			printk(KERN_WARNING "user_mad:   Documentation/infiniband/user_mad.txt "
685 			       "has info on the new ABI.\n");
686 		}
687 	}
688 
689 	file->agent[agent_id] = agent;
690 	ret = 0;
691 
692 out:
693 	mutex_unlock(&file->mutex);
694 
695 	if (ret && agent)
696 		ib_unregister_mad_agent(agent);
697 
698 	mutex_unlock(&file->port->file_mutex);
699 
700 	return ret;
701 }
702 
703 static int ib_umad_unreg_agent(struct ib_umad_file *file, u32 __user *arg)
704 {
705 	struct ib_mad_agent *agent = NULL;
706 	u32 id;
707 	int ret = 0;
708 
709 	if (get_user(id, arg))
710 		return -EFAULT;
711 
712 	mutex_lock(&file->port->file_mutex);
713 	mutex_lock(&file->mutex);
714 
715 	if (id < 0 || id >= IB_UMAD_MAX_AGENTS || !__get_agent(file, id)) {
716 		ret = -EINVAL;
717 		goto out;
718 	}
719 
720 	agent = file->agent[id];
721 	file->agent[id] = NULL;
722 
723 out:
724 	mutex_unlock(&file->mutex);
725 
726 	if (agent)
727 		ib_unregister_mad_agent(agent);
728 
729 	mutex_unlock(&file->port->file_mutex);
730 
731 	return ret;
732 }
733 
734 static long ib_umad_enable_pkey(struct ib_umad_file *file)
735 {
736 	int ret = 0;
737 
738 	mutex_lock(&file->mutex);
739 	if (file->already_used)
740 		ret = -EINVAL;
741 	else
742 		file->use_pkey_index = 1;
743 	mutex_unlock(&file->mutex);
744 
745 	return ret;
746 }
747 
748 static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
749 			  unsigned long arg)
750 {
751 	switch (cmd) {
752 	case IB_USER_MAD_REGISTER_AGENT:
753 		return ib_umad_reg_agent(filp->private_data, (void __user *) arg, 0);
754 	case IB_USER_MAD_UNREGISTER_AGENT:
755 		return ib_umad_unreg_agent(filp->private_data, (__u32 __user *) arg);
756 	case IB_USER_MAD_ENABLE_PKEY:
757 		return ib_umad_enable_pkey(filp->private_data);
758 	default:
759 		return -ENOIOCTLCMD;
760 	}
761 }
762 
763 #ifdef CONFIG_COMPAT
764 static long ib_umad_compat_ioctl(struct file *filp, unsigned int cmd,
765 				 unsigned long arg)
766 {
767 	switch (cmd) {
768 	case IB_USER_MAD_REGISTER_AGENT:
769 		return ib_umad_reg_agent(filp->private_data, compat_ptr(arg), 1);
770 	case IB_USER_MAD_UNREGISTER_AGENT:
771 		return ib_umad_unreg_agent(filp->private_data, compat_ptr(arg));
772 	case IB_USER_MAD_ENABLE_PKEY:
773 		return ib_umad_enable_pkey(filp->private_data);
774 	default:
775 		return -ENOIOCTLCMD;
776 	}
777 }
778 #endif
779 
780 static int ib_umad_open(struct inode *inode, struct file *filp)
781 {
782 	struct ib_umad_port *port;
783 	struct ib_umad_file *file;
784 	int ret = 0;
785 
786 	spin_lock(&port_lock);
787 	port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE];
788 	if (port)
789 		kref_get(&port->umad_dev->ref);
790 	spin_unlock(&port_lock);
791 
792 	if (!port)
793 		return -ENXIO;
794 
795 	mutex_lock(&port->file_mutex);
796 
797 	if (!port->ib_dev) {
798 		ret = -ENXIO;
799 		goto out;
800 	}
801 
802 	file = kzalloc(sizeof *file, GFP_KERNEL);
803 	if (!file) {
804 		kref_put(&port->umad_dev->ref, ib_umad_release_dev);
805 		ret = -ENOMEM;
806 		goto out;
807 	}
808 
809 	mutex_init(&file->mutex);
810 	spin_lock_init(&file->send_lock);
811 	INIT_LIST_HEAD(&file->recv_list);
812 	INIT_LIST_HEAD(&file->send_list);
813 	init_waitqueue_head(&file->recv_wait);
814 
815 	file->port = port;
816 	filp->private_data = file;
817 
818 	list_add_tail(&file->port_list, &port->file_list);
819 
820 out:
821 	mutex_unlock(&port->file_mutex);
822 	return ret;
823 }
824 
825 static int ib_umad_close(struct inode *inode, struct file *filp)
826 {
827 	struct ib_umad_file *file = filp->private_data;
828 	struct ib_umad_device *dev = file->port->umad_dev;
829 	struct ib_umad_packet *packet, *tmp;
830 	int already_dead;
831 	int i;
832 
833 	mutex_lock(&file->port->file_mutex);
834 	mutex_lock(&file->mutex);
835 
836 	already_dead = file->agents_dead;
837 	file->agents_dead = 1;
838 
839 	list_for_each_entry_safe(packet, tmp, &file->recv_list, list) {
840 		if (packet->recv_wc)
841 			ib_free_recv_mad(packet->recv_wc);
842 		kfree(packet);
843 	}
844 
845 	list_del(&file->port_list);
846 
847 	mutex_unlock(&file->mutex);
848 
849 	if (!already_dead)
850 		for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
851 			if (file->agent[i])
852 				ib_unregister_mad_agent(file->agent[i]);
853 
854 	mutex_unlock(&file->port->file_mutex);
855 
856 	kfree(file);
857 	kref_put(&dev->ref, ib_umad_release_dev);
858 
859 	return 0;
860 }
861 
862 static const struct file_operations umad_fops = {
863 	.owner 	 	= THIS_MODULE,
864 	.read 	 	= ib_umad_read,
865 	.write 	 	= ib_umad_write,
866 	.poll 	 	= ib_umad_poll,
867 	.unlocked_ioctl = ib_umad_ioctl,
868 #ifdef CONFIG_COMPAT
869 	.compat_ioctl 	= ib_umad_compat_ioctl,
870 #endif
871 	.open 	 	= ib_umad_open,
872 	.release 	= ib_umad_close
873 };
874 
875 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
876 {
877 	struct ib_umad_port *port;
878 	struct ib_port_modify props = {
879 		.set_port_cap_mask = IB_PORT_SM
880 	};
881 	int ret;
882 
883 	spin_lock(&port_lock);
884 	port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE - IB_UMAD_MAX_PORTS];
885 	if (port)
886 		kref_get(&port->umad_dev->ref);
887 	spin_unlock(&port_lock);
888 
889 	if (!port)
890 		return -ENXIO;
891 
892 	if (filp->f_flags & O_NONBLOCK) {
893 		if (down_trylock(&port->sm_sem)) {
894 			ret = -EAGAIN;
895 			goto fail;
896 		}
897 	} else {
898 		if (down_interruptible(&port->sm_sem)) {
899 			ret = -ERESTARTSYS;
900 			goto fail;
901 		}
902 	}
903 
904 	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
905 	if (ret) {
906 		up(&port->sm_sem);
907 		goto fail;
908 	}
909 
910 	filp->private_data = port;
911 
912 	return 0;
913 
914 fail:
915 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
916 	return ret;
917 }
918 
919 static int ib_umad_sm_close(struct inode *inode, struct file *filp)
920 {
921 	struct ib_umad_port *port = filp->private_data;
922 	struct ib_port_modify props = {
923 		.clr_port_cap_mask = IB_PORT_SM
924 	};
925 	int ret = 0;
926 
927 	mutex_lock(&port->file_mutex);
928 	if (port->ib_dev)
929 		ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
930 	mutex_unlock(&port->file_mutex);
931 
932 	up(&port->sm_sem);
933 
934 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
935 
936 	return ret;
937 }
938 
939 static const struct file_operations umad_sm_fops = {
940 	.owner 	 = THIS_MODULE,
941 	.open 	 = ib_umad_sm_open,
942 	.release = ib_umad_sm_close
943 };
944 
945 static struct ib_client umad_client = {
946 	.name   = "umad",
947 	.add    = ib_umad_add_one,
948 	.remove = ib_umad_remove_one
949 };
950 
951 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
952 {
953 	struct ib_umad_port *port = class_get_devdata(class_dev);
954 
955 	if (!port)
956 		return -ENODEV;
957 
958 	return sprintf(buf, "%s\n", port->ib_dev->name);
959 }
960 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
961 
962 static ssize_t show_port(struct class_device *class_dev, char *buf)
963 {
964 	struct ib_umad_port *port = class_get_devdata(class_dev);
965 
966 	if (!port)
967 		return -ENODEV;
968 
969 	return sprintf(buf, "%d\n", port->port_num);
970 }
971 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
972 
973 static ssize_t show_abi_version(struct class *class, char *buf)
974 {
975 	return sprintf(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
976 }
977 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
978 
979 static int ib_umad_init_port(struct ib_device *device, int port_num,
980 			     struct ib_umad_port *port)
981 {
982 	spin_lock(&port_lock);
983 	port->dev_num = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
984 	if (port->dev_num >= IB_UMAD_MAX_PORTS) {
985 		spin_unlock(&port_lock);
986 		return -1;
987 	}
988 	set_bit(port->dev_num, dev_map);
989 	spin_unlock(&port_lock);
990 
991 	port->ib_dev   = device;
992 	port->port_num = port_num;
993 	init_MUTEX(&port->sm_sem);
994 	mutex_init(&port->file_mutex);
995 	INIT_LIST_HEAD(&port->file_list);
996 
997 	port->dev = cdev_alloc();
998 	if (!port->dev)
999 		return -1;
1000 	port->dev->owner = THIS_MODULE;
1001 	port->dev->ops   = &umad_fops;
1002 	kobject_set_name(&port->dev->kobj, "umad%d", port->dev_num);
1003 	if (cdev_add(port->dev, base_dev + port->dev_num, 1))
1004 		goto err_cdev;
1005 
1006 	port->class_dev = class_device_create(umad_class, NULL, port->dev->dev,
1007 					      device->dma_device,
1008 					      "umad%d", port->dev_num);
1009 	if (IS_ERR(port->class_dev))
1010 		goto err_cdev;
1011 
1012 	if (class_device_create_file(port->class_dev, &class_device_attr_ibdev))
1013 		goto err_class;
1014 	if (class_device_create_file(port->class_dev, &class_device_attr_port))
1015 		goto err_class;
1016 
1017 	port->sm_dev = cdev_alloc();
1018 	if (!port->sm_dev)
1019 		goto err_class;
1020 	port->sm_dev->owner = THIS_MODULE;
1021 	port->sm_dev->ops   = &umad_sm_fops;
1022 	kobject_set_name(&port->sm_dev->kobj, "issm%d", port->dev_num);
1023 	if (cdev_add(port->sm_dev, base_dev + port->dev_num + IB_UMAD_MAX_PORTS, 1))
1024 		goto err_sm_cdev;
1025 
1026 	port->sm_class_dev = class_device_create(umad_class, NULL, port->sm_dev->dev,
1027 						 device->dma_device,
1028 						 "issm%d", port->dev_num);
1029 	if (IS_ERR(port->sm_class_dev))
1030 		goto err_sm_cdev;
1031 
1032 	class_set_devdata(port->class_dev,    port);
1033 	class_set_devdata(port->sm_class_dev, port);
1034 
1035 	if (class_device_create_file(port->sm_class_dev, &class_device_attr_ibdev))
1036 		goto err_sm_class;
1037 	if (class_device_create_file(port->sm_class_dev, &class_device_attr_port))
1038 		goto err_sm_class;
1039 
1040 	spin_lock(&port_lock);
1041 	umad_port[port->dev_num] = port;
1042 	spin_unlock(&port_lock);
1043 
1044 	return 0;
1045 
1046 err_sm_class:
1047 	class_device_destroy(umad_class, port->sm_dev->dev);
1048 
1049 err_sm_cdev:
1050 	cdev_del(port->sm_dev);
1051 
1052 err_class:
1053 	class_device_destroy(umad_class, port->dev->dev);
1054 
1055 err_cdev:
1056 	cdev_del(port->dev);
1057 	clear_bit(port->dev_num, dev_map);
1058 
1059 	return -1;
1060 }
1061 
1062 static void ib_umad_kill_port(struct ib_umad_port *port)
1063 {
1064 	struct ib_umad_file *file;
1065 	int already_dead;
1066 	int id;
1067 
1068 	class_set_devdata(port->class_dev,    NULL);
1069 	class_set_devdata(port->sm_class_dev, NULL);
1070 
1071 	class_device_destroy(umad_class, port->dev->dev);
1072 	class_device_destroy(umad_class, port->sm_dev->dev);
1073 
1074 	cdev_del(port->dev);
1075 	cdev_del(port->sm_dev);
1076 
1077 	spin_lock(&port_lock);
1078 	umad_port[port->dev_num] = NULL;
1079 	spin_unlock(&port_lock);
1080 
1081 	mutex_lock(&port->file_mutex);
1082 
1083 	port->ib_dev = NULL;
1084 
1085 	list_for_each_entry(file, &port->file_list, port_list) {
1086 		mutex_lock(&file->mutex);
1087 		already_dead = file->agents_dead;
1088 		file->agents_dead = 1;
1089 		mutex_unlock(&file->mutex);
1090 
1091 		for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id)
1092 			if (file->agent[id])
1093 				ib_unregister_mad_agent(file->agent[id]);
1094 	}
1095 
1096 	mutex_unlock(&port->file_mutex);
1097 
1098 	clear_bit(port->dev_num, dev_map);
1099 }
1100 
1101 static void ib_umad_add_one(struct ib_device *device)
1102 {
1103 	struct ib_umad_device *umad_dev;
1104 	int s, e, i;
1105 
1106 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1107 		return;
1108 
1109 	if (device->node_type == RDMA_NODE_IB_SWITCH)
1110 		s = e = 0;
1111 	else {
1112 		s = 1;
1113 		e = device->phys_port_cnt;
1114 	}
1115 
1116 	umad_dev = kzalloc(sizeof *umad_dev +
1117 			   (e - s + 1) * sizeof (struct ib_umad_port),
1118 			   GFP_KERNEL);
1119 	if (!umad_dev)
1120 		return;
1121 
1122 	kref_init(&umad_dev->ref);
1123 
1124 	umad_dev->start_port = s;
1125 	umad_dev->end_port   = e;
1126 
1127 	for (i = s; i <= e; ++i) {
1128 		umad_dev->port[i - s].umad_dev = umad_dev;
1129 
1130 		if (ib_umad_init_port(device, i, &umad_dev->port[i - s]))
1131 			goto err;
1132 	}
1133 
1134 	ib_set_client_data(device, &umad_client, umad_dev);
1135 
1136 	return;
1137 
1138 err:
1139 	while (--i >= s)
1140 		ib_umad_kill_port(&umad_dev->port[i - s]);
1141 
1142 	kref_put(&umad_dev->ref, ib_umad_release_dev);
1143 }
1144 
1145 static void ib_umad_remove_one(struct ib_device *device)
1146 {
1147 	struct ib_umad_device *umad_dev = ib_get_client_data(device, &umad_client);
1148 	int i;
1149 
1150 	if (!umad_dev)
1151 		return;
1152 
1153 	for (i = 0; i <= umad_dev->end_port - umad_dev->start_port; ++i)
1154 		ib_umad_kill_port(&umad_dev->port[i]);
1155 
1156 	kref_put(&umad_dev->ref, ib_umad_release_dev);
1157 }
1158 
1159 static int __init ib_umad_init(void)
1160 {
1161 	int ret;
1162 
1163 	ret = register_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2,
1164 				     "infiniband_mad");
1165 	if (ret) {
1166 		printk(KERN_ERR "user_mad: couldn't register device number\n");
1167 		goto out;
1168 	}
1169 
1170 	umad_class = class_create(THIS_MODULE, "infiniband_mad");
1171 	if (IS_ERR(umad_class)) {
1172 		ret = PTR_ERR(umad_class);
1173 		printk(KERN_ERR "user_mad: couldn't create class infiniband_mad\n");
1174 		goto out_chrdev;
1175 	}
1176 
1177 	ret = class_create_file(umad_class, &class_attr_abi_version);
1178 	if (ret) {
1179 		printk(KERN_ERR "user_mad: couldn't create abi_version attribute\n");
1180 		goto out_class;
1181 	}
1182 
1183 	ret = ib_register_client(&umad_client);
1184 	if (ret) {
1185 		printk(KERN_ERR "user_mad: couldn't register ib_umad client\n");
1186 		goto out_class;
1187 	}
1188 
1189 	return 0;
1190 
1191 out_class:
1192 	class_destroy(umad_class);
1193 
1194 out_chrdev:
1195 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
1196 
1197 out:
1198 	return ret;
1199 }
1200 
1201 static void __exit ib_umad_cleanup(void)
1202 {
1203 	ib_unregister_client(&umad_client);
1204 	class_destroy(umad_class);
1205 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
1206 }
1207 
1208 module_init(ib_umad_init);
1209 module_exit(ib_umad_cleanup);
1210