xref: /linux/drivers/infiniband/core/user_mad.c (revision 13abf8130139c2ccd4962a7e5a8902be5e6cb5a7)
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 2814 2005-07-06 19:14:09Z halr $
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 struct ib_umad_port {
68 	int                    devnum;
69 	struct cdev            dev;
70 	struct class_device    class_dev;
71 
72 	int                    sm_devnum;
73 	struct cdev            sm_dev;
74 	struct class_device    sm_class_dev;
75 	struct semaphore       sm_sem;
76 
77 	struct ib_device      *ib_dev;
78 	struct ib_umad_device *umad_dev;
79 	u8                     port_num;
80 };
81 
82 struct ib_umad_device {
83 	int                  start_port, end_port;
84 	struct kref          ref;
85 	struct ib_umad_port  port[0];
86 };
87 
88 struct ib_umad_file {
89 	struct ib_umad_port *port;
90 	spinlock_t           recv_lock;
91 	struct list_head     recv_list;
92 	wait_queue_head_t    recv_wait;
93 	struct rw_semaphore  agent_mutex;
94 	struct ib_mad_agent *agent[IB_UMAD_MAX_AGENTS];
95 	struct ib_mr        *mr[IB_UMAD_MAX_AGENTS];
96 };
97 
98 struct ib_umad_packet {
99 	struct ib_ah      *ah;
100 	struct ib_mad_send_buf *msg;
101 	struct list_head   list;
102 	int		   length;
103 	DECLARE_PCI_UNMAP_ADDR(mapping)
104 	struct ib_user_mad mad;
105 };
106 
107 static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
108 static spinlock_t map_lock;
109 static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS * 2);
110 
111 static void ib_umad_add_one(struct ib_device *device);
112 static void ib_umad_remove_one(struct ib_device *device);
113 
114 static int queue_packet(struct ib_umad_file *file,
115 			struct ib_mad_agent *agent,
116 			struct ib_umad_packet *packet)
117 {
118 	int ret = 1;
119 
120 	down_read(&file->agent_mutex);
121 	for (packet->mad.hdr.id = 0;
122 	     packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
123 	     packet->mad.hdr.id++)
124 		if (agent == file->agent[packet->mad.hdr.id]) {
125 			spin_lock_irq(&file->recv_lock);
126 			list_add_tail(&packet->list, &file->recv_list);
127 			spin_unlock_irq(&file->recv_lock);
128 			wake_up_interruptible(&file->recv_wait);
129 			ret = 0;
130 			break;
131 		}
132 
133 	up_read(&file->agent_mutex);
134 
135 	return ret;
136 }
137 
138 static void send_handler(struct ib_mad_agent *agent,
139 			 struct ib_mad_send_wc *send_wc)
140 {
141 	struct ib_umad_file *file = agent->context;
142 	struct ib_umad_packet *timeout, *packet =
143 		(void *) (unsigned long) send_wc->wr_id;
144 
145 	ib_destroy_ah(packet->msg->send_wr.wr.ud.ah);
146 	ib_free_send_mad(packet->msg);
147 
148 	if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
149 		timeout = kmalloc(sizeof *timeout + sizeof (struct ib_mad_hdr),
150 				  GFP_KERNEL);
151 		if (!timeout)
152 			goto out;
153 
154 		memset(timeout, 0, sizeof *timeout + sizeof (struct ib_mad_hdr));
155 
156 		timeout->length = sizeof (struct ib_mad_hdr);
157 		timeout->mad.hdr.id = packet->mad.hdr.id;
158 		timeout->mad.hdr.status = ETIMEDOUT;
159 		memcpy(timeout->mad.data, packet->mad.data,
160 		       sizeof (struct ib_mad_hdr));
161 
162 		if (!queue_packet(file, agent, timeout))
163 				return;
164 	}
165 out:
166 	kfree(packet);
167 }
168 
169 static void recv_handler(struct ib_mad_agent *agent,
170 			 struct ib_mad_recv_wc *mad_recv_wc)
171 {
172 	struct ib_umad_file *file = agent->context;
173 	struct ib_umad_packet *packet;
174 	int length;
175 
176 	if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
177 		goto out;
178 
179 	length = mad_recv_wc->mad_len;
180 	packet = kmalloc(sizeof *packet + length, GFP_KERNEL);
181 	if (!packet)
182 		goto out;
183 
184 	memset(packet, 0, sizeof *packet + length);
185 	packet->length = length;
186 
187 	ib_coalesce_recv_mad(mad_recv_wc, packet->mad.data);
188 
189 	packet->mad.hdr.status    = 0;
190 	packet->mad.hdr.length    = length + sizeof (struct ib_user_mad);
191 	packet->mad.hdr.qpn 	  = cpu_to_be32(mad_recv_wc->wc->src_qp);
192 	packet->mad.hdr.lid 	  = cpu_to_be16(mad_recv_wc->wc->slid);
193 	packet->mad.hdr.sl  	  = mad_recv_wc->wc->sl;
194 	packet->mad.hdr.path_bits = mad_recv_wc->wc->dlid_path_bits;
195 	packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
196 	if (packet->mad.hdr.grh_present) {
197 		/* XXX parse GRH */
198 		packet->mad.hdr.gid_index 	= 0;
199 		packet->mad.hdr.hop_limit 	= 0;
200 		packet->mad.hdr.traffic_class	= 0;
201 		memset(packet->mad.hdr.gid, 0, 16);
202 		packet->mad.hdr.flow_label	= 0;
203 	}
204 
205 	if (queue_packet(file, agent, packet))
206 		kfree(packet);
207 
208 out:
209 	ib_free_recv_mad(mad_recv_wc);
210 }
211 
212 static ssize_t ib_umad_read(struct file *filp, char __user *buf,
213 			    size_t count, loff_t *pos)
214 {
215 	struct ib_umad_file *file = filp->private_data;
216 	struct ib_umad_packet *packet;
217 	ssize_t ret;
218 
219 	if (count < sizeof (struct ib_user_mad) + sizeof (struct ib_mad))
220 		return -EINVAL;
221 
222 	spin_lock_irq(&file->recv_lock);
223 
224 	while (list_empty(&file->recv_list)) {
225 		spin_unlock_irq(&file->recv_lock);
226 
227 		if (filp->f_flags & O_NONBLOCK)
228 			return -EAGAIN;
229 
230 		if (wait_event_interruptible(file->recv_wait,
231 					     !list_empty(&file->recv_list)))
232 			return -ERESTARTSYS;
233 
234 		spin_lock_irq(&file->recv_lock);
235 	}
236 
237 	packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
238 	list_del(&packet->list);
239 
240 	spin_unlock_irq(&file->recv_lock);
241 
242 	if (count < packet->length + sizeof (struct ib_user_mad)) {
243 		/* Return length needed (and first RMPP segment) if too small */
244 		if (copy_to_user(buf, &packet->mad,
245 				 sizeof (struct ib_user_mad) + sizeof (struct ib_mad)))
246 			ret = -EFAULT;
247 		else
248 			ret = -ENOSPC;
249 	} else if (copy_to_user(buf, &packet->mad,
250 			      packet->length + sizeof (struct ib_user_mad)))
251 		ret = -EFAULT;
252 	else
253 		ret = packet->length + sizeof (struct ib_user_mad);
254 	if (ret < 0) {
255 		/* Requeue packet */
256 		spin_lock_irq(&file->recv_lock);
257 		list_add(&packet->list, &file->recv_list);
258 		spin_unlock_irq(&file->recv_lock);
259 	} else
260 		kfree(packet);
261 	return ret;
262 }
263 
264 static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
265 			     size_t count, loff_t *pos)
266 {
267 	struct ib_umad_file *file = filp->private_data;
268 	struct ib_umad_packet *packet;
269 	struct ib_mad_agent *agent;
270 	struct ib_ah_attr ah_attr;
271 	struct ib_send_wr *bad_wr;
272 	struct ib_rmpp_mad *rmpp_mad;
273 	u8 method;
274 	__be64 *tid;
275 	int ret, length, hdr_len, data_len, rmpp_hdr_size;
276 	int rmpp_active = 0;
277 
278 	if (count < sizeof (struct ib_user_mad))
279 		return -EINVAL;
280 
281 	length = count - sizeof (struct ib_user_mad);
282 	packet = kmalloc(sizeof *packet + sizeof(struct ib_mad_hdr) +
283 			 sizeof(struct ib_rmpp_hdr), GFP_KERNEL);
284 	if (!packet)
285 		return -ENOMEM;
286 
287 	if (copy_from_user(&packet->mad, buf,
288 			    sizeof (struct ib_user_mad) +
289 			    sizeof(struct ib_mad_hdr) +
290 			    sizeof(struct ib_rmpp_hdr))) {
291 		ret = -EFAULT;
292 		goto err;
293 	}
294 
295 	if (packet->mad.hdr.id < 0 ||
296 	    packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
297 		ret = -EINVAL;
298 		goto err;
299 	}
300 
301 	packet->length = length;
302 
303 	down_read(&file->agent_mutex);
304 
305 	agent = file->agent[packet->mad.hdr.id];
306 	if (!agent) {
307 		ret = -EINVAL;
308 		goto err_up;
309 	}
310 
311 	memset(&ah_attr, 0, sizeof ah_attr);
312 	ah_attr.dlid          = be16_to_cpu(packet->mad.hdr.lid);
313 	ah_attr.sl            = packet->mad.hdr.sl;
314 	ah_attr.src_path_bits = packet->mad.hdr.path_bits;
315 	ah_attr.port_num      = file->port->port_num;
316 	if (packet->mad.hdr.grh_present) {
317 		ah_attr.ah_flags = IB_AH_GRH;
318 		memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
319 		ah_attr.grh.flow_label 	   = be32_to_cpu(packet->mad.hdr.flow_label);
320 		ah_attr.grh.hop_limit  	   = packet->mad.hdr.hop_limit;
321 		ah_attr.grh.traffic_class  = packet->mad.hdr.traffic_class;
322 	}
323 
324 	packet->ah = ib_create_ah(agent->qp->pd, &ah_attr);
325 	if (IS_ERR(packet->ah)) {
326 		ret = PTR_ERR(packet->ah);
327 		goto err_up;
328 	}
329 
330 	rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
331 	if (ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) & IB_MGMT_RMPP_FLAG_ACTIVE) {
332 		/* RMPP active */
333 		if (!agent->rmpp_version) {
334 			ret = -EINVAL;
335 			goto err_ah;
336 		}
337 		/* Validate that management class can support RMPP */
338 		if (rmpp_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_ADM) {
339 			hdr_len = offsetof(struct ib_sa_mad, data);
340 			data_len = length;
341 		} else if ((rmpp_mad->mad_hdr.mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
342 			    (rmpp_mad->mad_hdr.mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)) {
343 				hdr_len = offsetof(struct ib_vendor_mad, data);
344 				data_len = length - hdr_len;
345 		} else {
346 			ret = -EINVAL;
347 			goto err_ah;
348 		}
349 		rmpp_active = 1;
350 	} else {
351 		if (length > sizeof(struct ib_mad)) {
352 			ret = -EINVAL;
353 			goto err_ah;
354 		}
355 		hdr_len = offsetof(struct ib_mad, data);
356 		data_len = length - hdr_len;
357 	}
358 
359 	packet->msg = ib_create_send_mad(agent,
360 					 be32_to_cpu(packet->mad.hdr.qpn),
361 					 0, packet->ah, rmpp_active,
362 					 hdr_len, data_len,
363 					 GFP_KERNEL);
364 	if (IS_ERR(packet->msg)) {
365 		ret = PTR_ERR(packet->msg);
366 		goto err_ah;
367 	}
368 
369 	packet->msg->send_wr.wr.ud.timeout_ms  = packet->mad.hdr.timeout_ms;
370 	packet->msg->send_wr.wr.ud.retries = packet->mad.hdr.retries;
371 
372 	/* Override send WR WRID initialized in ib_create_send_mad */
373 	packet->msg->send_wr.wr_id = (unsigned long) packet;
374 
375 	if (!rmpp_active) {
376 		/* Copy message from user into send buffer */
377 		if (copy_from_user(packet->msg->mad,
378 				   buf + sizeof(struct ib_user_mad), length)) {
379 			ret = -EFAULT;
380 			goto err_msg;
381 		}
382 	} else {
383 		rmpp_hdr_size = sizeof(struct ib_mad_hdr) +
384 				sizeof(struct ib_rmpp_hdr);
385 
386 		/* Only copy MAD headers (RMPP header in place) */
387 		memcpy(packet->msg->mad, packet->mad.data,
388 		       sizeof(struct ib_mad_hdr));
389 
390 		/* Now, copy rest of message from user into send buffer */
391 		if (copy_from_user(((struct ib_rmpp_mad *) packet->msg->mad)->data,
392 				   buf + sizeof (struct ib_user_mad) + rmpp_hdr_size,
393 				   length - rmpp_hdr_size)) {
394 			ret = -EFAULT;
395 			goto err_msg;
396 		}
397 	}
398 
399 	/*
400 	 * If userspace is generating a request that will generate a
401 	 * response, we need to make sure the high-order part of the
402 	 * transaction ID matches the agent being used to send the
403 	 * MAD.
404 	 */
405 	method = packet->msg->mad->mad_hdr.method;
406 
407 	if (!(method & IB_MGMT_METHOD_RESP)       &&
408 	    method != IB_MGMT_METHOD_TRAP_REPRESS &&
409 	    method != IB_MGMT_METHOD_SEND) {
410 		tid = &packet->msg->mad->mad_hdr.tid;
411 		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
412 				   (be64_to_cpup(tid) & 0xffffffff));
413 	}
414 
415 	ret = ib_post_send_mad(agent, &packet->msg->send_wr, &bad_wr);
416 	if (ret)
417 		goto err_msg;
418 
419 	up_read(&file->agent_mutex);
420 
421 	return sizeof (struct ib_user_mad_hdr) + packet->length;
422 
423 err_msg:
424 	ib_free_send_mad(packet->msg);
425 
426 err_ah:
427 	ib_destroy_ah(packet->ah);
428 
429 err_up:
430 	up_read(&file->agent_mutex);
431 
432 err:
433 	kfree(packet);
434 	return ret;
435 }
436 
437 static unsigned int ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
438 {
439 	struct ib_umad_file *file = filp->private_data;
440 
441 	/* we will always be able to post a MAD send */
442 	unsigned int mask = POLLOUT | POLLWRNORM;
443 
444 	poll_wait(filp, &file->recv_wait, wait);
445 
446 	if (!list_empty(&file->recv_list))
447 		mask |= POLLIN | POLLRDNORM;
448 
449 	return mask;
450 }
451 
452 static int ib_umad_reg_agent(struct ib_umad_file *file, unsigned long arg)
453 {
454 	struct ib_user_mad_reg_req ureq;
455 	struct ib_mad_reg_req req;
456 	struct ib_mad_agent *agent;
457 	int agent_id;
458 	int ret;
459 
460 	down_write(&file->agent_mutex);
461 
462 	if (copy_from_user(&ureq, (void __user *) arg, sizeof ureq)) {
463 		ret = -EFAULT;
464 		goto out;
465 	}
466 
467 	if (ureq.qpn != 0 && ureq.qpn != 1) {
468 		ret = -EINVAL;
469 		goto out;
470 	}
471 
472 	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
473 		if (!file->agent[agent_id])
474 			goto found;
475 
476 	ret = -ENOMEM;
477 	goto out;
478 
479 found:
480 	if (ureq.mgmt_class) {
481 		req.mgmt_class         = ureq.mgmt_class;
482 		req.mgmt_class_version = ureq.mgmt_class_version;
483 		memcpy(req.method_mask, ureq.method_mask, sizeof req.method_mask);
484 		memcpy(req.oui,         ureq.oui,         sizeof req.oui);
485 	}
486 
487 	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
488 				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
489 				      ureq.mgmt_class ? &req : NULL,
490 				      ureq.rmpp_version,
491 				      send_handler, recv_handler, file);
492 	if (IS_ERR(agent)) {
493 		ret = PTR_ERR(agent);
494 		goto out;
495 	}
496 
497 	file->agent[agent_id] = agent;
498 
499 	file->mr[agent_id] = ib_get_dma_mr(agent->qp->pd, IB_ACCESS_LOCAL_WRITE);
500 	if (IS_ERR(file->mr[agent_id])) {
501 		ret = -ENOMEM;
502 		goto err;
503 	}
504 
505 	if (put_user(agent_id,
506 		     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
507 		ret = -EFAULT;
508 		goto err_mr;
509 	}
510 
511 	ret = 0;
512 	goto out;
513 
514 err_mr:
515 	ib_dereg_mr(file->mr[agent_id]);
516 
517 err:
518 	file->agent[agent_id] = NULL;
519 	ib_unregister_mad_agent(agent);
520 
521 out:
522 	up_write(&file->agent_mutex);
523 	return ret;
524 }
525 
526 static int ib_umad_unreg_agent(struct ib_umad_file *file, unsigned long arg)
527 {
528 	u32 id;
529 	int ret = 0;
530 
531 	down_write(&file->agent_mutex);
532 
533 	if (get_user(id, (u32 __user *) arg)) {
534 		ret = -EFAULT;
535 		goto out;
536 	}
537 
538 	if (id < 0 || id >= IB_UMAD_MAX_AGENTS || !file->agent[id]) {
539 		ret = -EINVAL;
540 		goto out;
541 	}
542 
543 	ib_dereg_mr(file->mr[id]);
544 	ib_unregister_mad_agent(file->agent[id]);
545 	file->agent[id] = NULL;
546 
547 out:
548 	up_write(&file->agent_mutex);
549 	return ret;
550 }
551 
552 static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
553 			  unsigned long arg)
554 {
555 	switch (cmd) {
556 	case IB_USER_MAD_REGISTER_AGENT:
557 		return ib_umad_reg_agent(filp->private_data, arg);
558 	case IB_USER_MAD_UNREGISTER_AGENT:
559 		return ib_umad_unreg_agent(filp->private_data, arg);
560 	default:
561 		return -ENOIOCTLCMD;
562 	}
563 }
564 
565 static int ib_umad_open(struct inode *inode, struct file *filp)
566 {
567 	struct ib_umad_port *port =
568 		container_of(inode->i_cdev, struct ib_umad_port, dev);
569 	struct ib_umad_file *file;
570 
571 	file = kmalloc(sizeof *file, GFP_KERNEL);
572 	if (!file)
573 		return -ENOMEM;
574 
575 	memset(file, 0, sizeof *file);
576 
577 	spin_lock_init(&file->recv_lock);
578 	init_rwsem(&file->agent_mutex);
579 	INIT_LIST_HEAD(&file->recv_list);
580 	init_waitqueue_head(&file->recv_wait);
581 
582 	file->port = port;
583 	filp->private_data = file;
584 
585 	return 0;
586 }
587 
588 static int ib_umad_close(struct inode *inode, struct file *filp)
589 {
590 	struct ib_umad_file *file = filp->private_data;
591 	struct ib_umad_packet *packet, *tmp;
592 	int i;
593 
594 	for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
595 		if (file->agent[i]) {
596 			ib_dereg_mr(file->mr[i]);
597 			ib_unregister_mad_agent(file->agent[i]);
598 		}
599 
600 	list_for_each_entry_safe(packet, tmp, &file->recv_list, list)
601 		kfree(packet);
602 
603 	kfree(file);
604 
605 	return 0;
606 }
607 
608 static struct file_operations umad_fops = {
609 	.owner 	 	= THIS_MODULE,
610 	.read 	 	= ib_umad_read,
611 	.write 	 	= ib_umad_write,
612 	.poll 	 	= ib_umad_poll,
613 	.unlocked_ioctl = ib_umad_ioctl,
614 	.compat_ioctl 	= ib_umad_ioctl,
615 	.open 	 	= ib_umad_open,
616 	.release 	= ib_umad_close
617 };
618 
619 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
620 {
621 	struct ib_umad_port *port =
622 		container_of(inode->i_cdev, struct ib_umad_port, sm_dev);
623 	struct ib_port_modify props = {
624 		.set_port_cap_mask = IB_PORT_SM
625 	};
626 	int ret;
627 
628 	if (filp->f_flags & O_NONBLOCK) {
629 		if (down_trylock(&port->sm_sem))
630 			return -EAGAIN;
631 	} else {
632 		if (down_interruptible(&port->sm_sem))
633 			return -ERESTARTSYS;
634 	}
635 
636 	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
637 	if (ret) {
638 		up(&port->sm_sem);
639 		return ret;
640 	}
641 
642 	filp->private_data = port;
643 
644 	return 0;
645 }
646 
647 static int ib_umad_sm_close(struct inode *inode, struct file *filp)
648 {
649 	struct ib_umad_port *port = filp->private_data;
650 	struct ib_port_modify props = {
651 		.clr_port_cap_mask = IB_PORT_SM
652 	};
653 	int ret;
654 
655 	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
656 	up(&port->sm_sem);
657 
658 	return ret;
659 }
660 
661 static struct file_operations umad_sm_fops = {
662 	.owner 	 = THIS_MODULE,
663 	.open 	 = ib_umad_sm_open,
664 	.release = ib_umad_sm_close
665 };
666 
667 static struct ib_client umad_client = {
668 	.name   = "umad",
669 	.add    = ib_umad_add_one,
670 	.remove = ib_umad_remove_one
671 };
672 
673 static ssize_t show_dev(struct class_device *class_dev, char *buf)
674 {
675 	struct ib_umad_port *port = class_get_devdata(class_dev);
676 
677 	if (class_dev == &port->class_dev)
678 		return print_dev_t(buf, port->dev.dev);
679 	else
680 		return print_dev_t(buf, port->sm_dev.dev);
681 }
682 static CLASS_DEVICE_ATTR(dev, S_IRUGO, show_dev, NULL);
683 
684 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
685 {
686 	struct ib_umad_port *port = class_get_devdata(class_dev);
687 
688 	return sprintf(buf, "%s\n", port->ib_dev->name);
689 }
690 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
691 
692 static ssize_t show_port(struct class_device *class_dev, char *buf)
693 {
694 	struct ib_umad_port *port = class_get_devdata(class_dev);
695 
696 	return sprintf(buf, "%d\n", port->port_num);
697 }
698 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
699 
700 static void ib_umad_release_dev(struct kref *ref)
701 {
702 	struct ib_umad_device *dev =
703 		container_of(ref, struct ib_umad_device, ref);
704 
705 	kfree(dev);
706 }
707 
708 static void ib_umad_release_port(struct class_device *class_dev)
709 {
710 	struct ib_umad_port *port = class_get_devdata(class_dev);
711 
712 	if (class_dev == &port->class_dev) {
713 		cdev_del(&port->dev);
714 		clear_bit(port->devnum, dev_map);
715 	} else {
716 		cdev_del(&port->sm_dev);
717 		clear_bit(port->sm_devnum, dev_map);
718 	}
719 
720 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
721 }
722 
723 static struct class umad_class = {
724 	.name    = "infiniband_mad",
725 	.release = ib_umad_release_port
726 };
727 
728 static ssize_t show_abi_version(struct class *class, char *buf)
729 {
730 	return sprintf(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
731 }
732 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
733 
734 static int ib_umad_init_port(struct ib_device *device, int port_num,
735 			     struct ib_umad_port *port)
736 {
737 	spin_lock(&map_lock);
738 	port->devnum = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
739 	if (port->devnum >= IB_UMAD_MAX_PORTS) {
740 		spin_unlock(&map_lock);
741 		return -1;
742 	}
743 	port->sm_devnum = find_next_zero_bit(dev_map, IB_UMAD_MAX_PORTS * 2, IB_UMAD_MAX_PORTS);
744 	if (port->sm_devnum >= IB_UMAD_MAX_PORTS * 2) {
745 		spin_unlock(&map_lock);
746 		return -1;
747 	}
748 	set_bit(port->devnum, dev_map);
749 	set_bit(port->sm_devnum, dev_map);
750 	spin_unlock(&map_lock);
751 
752 	port->ib_dev   = device;
753 	port->port_num = port_num;
754 	init_MUTEX(&port->sm_sem);
755 
756 	cdev_init(&port->dev, &umad_fops);
757 	port->dev.owner = THIS_MODULE;
758 	kobject_set_name(&port->dev.kobj, "umad%d", port->devnum);
759 	if (cdev_add(&port->dev, base_dev + port->devnum, 1))
760 		return -1;
761 
762 	port->class_dev.class = &umad_class;
763 	port->class_dev.dev   = device->dma_device;
764 
765 	snprintf(port->class_dev.class_id, BUS_ID_SIZE, "umad%d", port->devnum);
766 
767 	if (class_device_register(&port->class_dev))
768 		goto err_cdev;
769 
770 	class_set_devdata(&port->class_dev, port);
771 	kref_get(&port->umad_dev->ref);
772 
773 	if (class_device_create_file(&port->class_dev, &class_device_attr_dev))
774 		goto err_class;
775 	if (class_device_create_file(&port->class_dev, &class_device_attr_ibdev))
776 		goto err_class;
777 	if (class_device_create_file(&port->class_dev, &class_device_attr_port))
778 		goto err_class;
779 
780 	cdev_init(&port->sm_dev, &umad_sm_fops);
781 	port->sm_dev.owner = THIS_MODULE;
782 	kobject_set_name(&port->dev.kobj, "issm%d", port->sm_devnum - IB_UMAD_MAX_PORTS);
783 	if (cdev_add(&port->sm_dev, base_dev + port->sm_devnum, 1))
784 		return -1;
785 
786 	port->sm_class_dev.class = &umad_class;
787 	port->sm_class_dev.dev   = device->dma_device;
788 
789 	snprintf(port->sm_class_dev.class_id, BUS_ID_SIZE, "issm%d", port->sm_devnum - IB_UMAD_MAX_PORTS);
790 
791 	if (class_device_register(&port->sm_class_dev))
792 		goto err_sm_cdev;
793 
794 	class_set_devdata(&port->sm_class_dev, port);
795 	kref_get(&port->umad_dev->ref);
796 
797 	if (class_device_create_file(&port->sm_class_dev, &class_device_attr_dev))
798 		goto err_sm_class;
799 	if (class_device_create_file(&port->sm_class_dev, &class_device_attr_ibdev))
800 		goto err_sm_class;
801 	if (class_device_create_file(&port->sm_class_dev, &class_device_attr_port))
802 		goto err_sm_class;
803 
804 	return 0;
805 
806 err_sm_class:
807 	class_device_unregister(&port->sm_class_dev);
808 
809 err_sm_cdev:
810 	cdev_del(&port->sm_dev);
811 
812 err_class:
813 	class_device_unregister(&port->class_dev);
814 
815 err_cdev:
816 	cdev_del(&port->dev);
817 	clear_bit(port->devnum, dev_map);
818 
819 	return -1;
820 }
821 
822 static void ib_umad_add_one(struct ib_device *device)
823 {
824 	struct ib_umad_device *umad_dev;
825 	int s, e, i;
826 
827 	if (device->node_type == IB_NODE_SWITCH)
828 		s = e = 0;
829 	else {
830 		s = 1;
831 		e = device->phys_port_cnt;
832 	}
833 
834 	umad_dev = kmalloc(sizeof *umad_dev +
835 			   (e - s + 1) * sizeof (struct ib_umad_port),
836 			   GFP_KERNEL);
837 	if (!umad_dev)
838 		return;
839 
840 	memset(umad_dev, 0, sizeof *umad_dev +
841 	       (e - s + 1) * sizeof (struct ib_umad_port));
842 
843 	kref_init(&umad_dev->ref);
844 
845 	umad_dev->start_port = s;
846 	umad_dev->end_port   = e;
847 
848 	for (i = s; i <= e; ++i) {
849 		umad_dev->port[i - s].umad_dev = umad_dev;
850 
851 		if (ib_umad_init_port(device, i, &umad_dev->port[i - s]))
852 			goto err;
853 	}
854 
855 	ib_set_client_data(device, &umad_client, umad_dev);
856 
857 	return;
858 
859 err:
860 	while (--i >= s) {
861 		class_device_unregister(&umad_dev->port[i - s].class_dev);
862 		class_device_unregister(&umad_dev->port[i - s].sm_class_dev);
863 	}
864 
865 	kref_put(&umad_dev->ref, ib_umad_release_dev);
866 }
867 
868 static void ib_umad_remove_one(struct ib_device *device)
869 {
870 	struct ib_umad_device *umad_dev = ib_get_client_data(device, &umad_client);
871 	int i;
872 
873 	if (!umad_dev)
874 		return;
875 
876 	for (i = 0; i <= umad_dev->end_port - umad_dev->start_port; ++i) {
877 		class_device_unregister(&umad_dev->port[i].class_dev);
878 		class_device_unregister(&umad_dev->port[i].sm_class_dev);
879 	}
880 
881 	kref_put(&umad_dev->ref, ib_umad_release_dev);
882 }
883 
884 static int __init ib_umad_init(void)
885 {
886 	int ret;
887 
888 	spin_lock_init(&map_lock);
889 
890 	ret = register_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2,
891 				     "infiniband_mad");
892 	if (ret) {
893 		printk(KERN_ERR "user_mad: couldn't register device number\n");
894 		goto out;
895 	}
896 
897 	ret = class_register(&umad_class);
898 	if (ret) {
899 		printk(KERN_ERR "user_mad: couldn't create class infiniband_mad\n");
900 		goto out_chrdev;
901 	}
902 
903 	ret = class_create_file(&umad_class, &class_attr_abi_version);
904 	if (ret) {
905 		printk(KERN_ERR "user_mad: couldn't create abi_version attribute\n");
906 		goto out_class;
907 	}
908 
909 	ret = ib_register_client(&umad_client);
910 	if (ret) {
911 		printk(KERN_ERR "user_mad: couldn't register ib_umad client\n");
912 		goto out_class;
913 	}
914 
915 	return 0;
916 
917 out_class:
918 	class_unregister(&umad_class);
919 
920 out_chrdev:
921 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
922 
923 out:
924 	return ret;
925 }
926 
927 static void __exit ib_umad_cleanup(void)
928 {
929 	ib_unregister_client(&umad_client);
930 	class_unregister(&umad_class);
931 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
932 }
933 
934 module_init(ib_umad_init);
935 module_exit(ib_umad_cleanup);
936