xref: /linux/drivers/infiniband/core/user_mad.c (revision 858259cf7d1c443c836a2022b78cb281f0a9b95e)
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 /*
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     port_list;
116 	spinlock_t           recv_lock;
117 	wait_queue_head_t    recv_wait;
118 	struct ib_mad_agent *agent[IB_UMAD_MAX_AGENTS];
119 	struct ib_mr        *mr[IB_UMAD_MAX_AGENTS];
120 };
121 
122 struct ib_umad_packet {
123 	struct ib_mad_send_buf *msg;
124 	struct list_head   list;
125 	int		   length;
126 	struct ib_user_mad mad;
127 };
128 
129 static struct class *umad_class;
130 
131 static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
132 
133 static DEFINE_SPINLOCK(port_lock);
134 static struct ib_umad_port *umad_port[IB_UMAD_MAX_PORTS];
135 static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS * 2);
136 
137 static void ib_umad_add_one(struct ib_device *device);
138 static void ib_umad_remove_one(struct ib_device *device);
139 
140 static void ib_umad_release_dev(struct kref *ref)
141 {
142 	struct ib_umad_device *dev =
143 		container_of(ref, struct ib_umad_device, ref);
144 
145 	kfree(dev);
146 }
147 
148 static int queue_packet(struct ib_umad_file *file,
149 			struct ib_mad_agent *agent,
150 			struct ib_umad_packet *packet)
151 {
152 	int ret = 1;
153 
154 	down_read(&file->port->mutex);
155 	for (packet->mad.hdr.id = 0;
156 	     packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
157 	     packet->mad.hdr.id++)
158 		if (agent == file->agent[packet->mad.hdr.id]) {
159 			spin_lock_irq(&file->recv_lock);
160 			list_add_tail(&packet->list, &file->recv_list);
161 			spin_unlock_irq(&file->recv_lock);
162 			wake_up_interruptible(&file->recv_wait);
163 			ret = 0;
164 			break;
165 		}
166 
167 	up_read(&file->port->mutex);
168 
169 	return ret;
170 }
171 
172 static void send_handler(struct ib_mad_agent *agent,
173 			 struct ib_mad_send_wc *send_wc)
174 {
175 	struct ib_umad_file *file = agent->context;
176 	struct ib_umad_packet *timeout;
177 	struct ib_umad_packet *packet = send_wc->send_buf->context[0];
178 
179 	ib_destroy_ah(packet->msg->ah);
180 	ib_free_send_mad(packet->msg);
181 
182 	if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
183 		timeout = kzalloc(sizeof *timeout + IB_MGMT_MAD_HDR, GFP_KERNEL);
184 		if (!timeout)
185 			goto out;
186 
187 		timeout->length 	= IB_MGMT_MAD_HDR;
188 		timeout->mad.hdr.id 	= packet->mad.hdr.id;
189 		timeout->mad.hdr.status = ETIMEDOUT;
190 		memcpy(timeout->mad.data, packet->mad.data,
191 		       sizeof (struct ib_mad_hdr));
192 
193 		if (!queue_packet(file, agent, timeout))
194 				return;
195 	}
196 out:
197 	kfree(packet);
198 }
199 
200 static void recv_handler(struct ib_mad_agent *agent,
201 			 struct ib_mad_recv_wc *mad_recv_wc)
202 {
203 	struct ib_umad_file *file = agent->context;
204 	struct ib_umad_packet *packet;
205 	int length;
206 
207 	if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
208 		goto out;
209 
210 	length = mad_recv_wc->mad_len;
211 	packet = kzalloc(sizeof *packet + length, GFP_KERNEL);
212 	if (!packet)
213 		goto out;
214 
215 	packet->length = length;
216 
217 	ib_coalesce_recv_mad(mad_recv_wc, packet->mad.data);
218 
219 	packet->mad.hdr.status    = 0;
220 	packet->mad.hdr.length    = length + sizeof (struct ib_user_mad);
221 	packet->mad.hdr.qpn 	  = cpu_to_be32(mad_recv_wc->wc->src_qp);
222 	packet->mad.hdr.lid 	  = cpu_to_be16(mad_recv_wc->wc->slid);
223 	packet->mad.hdr.sl  	  = mad_recv_wc->wc->sl;
224 	packet->mad.hdr.path_bits = mad_recv_wc->wc->dlid_path_bits;
225 	packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
226 	if (packet->mad.hdr.grh_present) {
227 		/* XXX parse GRH */
228 		packet->mad.hdr.gid_index 	= 0;
229 		packet->mad.hdr.hop_limit 	= 0;
230 		packet->mad.hdr.traffic_class	= 0;
231 		memset(packet->mad.hdr.gid, 0, 16);
232 		packet->mad.hdr.flow_label	= 0;
233 	}
234 
235 	if (queue_packet(file, agent, packet))
236 		kfree(packet);
237 
238 out:
239 	ib_free_recv_mad(mad_recv_wc);
240 }
241 
242 static ssize_t ib_umad_read(struct file *filp, char __user *buf,
243 			    size_t count, loff_t *pos)
244 {
245 	struct ib_umad_file *file = filp->private_data;
246 	struct ib_umad_packet *packet;
247 	ssize_t ret;
248 
249 	if (count < sizeof (struct ib_user_mad) + sizeof (struct ib_mad))
250 		return -EINVAL;
251 
252 	spin_lock_irq(&file->recv_lock);
253 
254 	while (list_empty(&file->recv_list)) {
255 		spin_unlock_irq(&file->recv_lock);
256 
257 		if (filp->f_flags & O_NONBLOCK)
258 			return -EAGAIN;
259 
260 		if (wait_event_interruptible(file->recv_wait,
261 					     !list_empty(&file->recv_list)))
262 			return -ERESTARTSYS;
263 
264 		spin_lock_irq(&file->recv_lock);
265 	}
266 
267 	packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
268 	list_del(&packet->list);
269 
270 	spin_unlock_irq(&file->recv_lock);
271 
272 	if (count < packet->length + sizeof (struct ib_user_mad)) {
273 		/* Return length needed (and first RMPP segment) if too small */
274 		if (copy_to_user(buf, &packet->mad,
275 				 sizeof (struct ib_user_mad) + sizeof (struct ib_mad)))
276 			ret = -EFAULT;
277 		else
278 			ret = -ENOSPC;
279 	} else if (copy_to_user(buf, &packet->mad,
280 				packet->length + sizeof (struct ib_user_mad)))
281 		ret = -EFAULT;
282 	else
283 		ret = packet->length + sizeof (struct ib_user_mad);
284 	if (ret < 0) {
285 		/* Requeue packet */
286 		spin_lock_irq(&file->recv_lock);
287 		list_add(&packet->list, &file->recv_list);
288 		spin_unlock_irq(&file->recv_lock);
289 	} else
290 		kfree(packet);
291 	return ret;
292 }
293 
294 static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
295 			     size_t count, loff_t *pos)
296 {
297 	struct ib_umad_file *file = filp->private_data;
298 	struct ib_umad_packet *packet;
299 	struct ib_mad_agent *agent;
300 	struct ib_ah_attr ah_attr;
301 	struct ib_ah *ah;
302 	struct ib_rmpp_mad *rmpp_mad;
303 	u8 method;
304 	__be64 *tid;
305 	int ret, length, hdr_len, copy_offset;
306 	int rmpp_active = 0;
307 
308 	if (count < sizeof (struct ib_user_mad))
309 		return -EINVAL;
310 
311 	length = count - sizeof (struct ib_user_mad);
312 	packet = kmalloc(sizeof *packet + IB_MGMT_RMPP_HDR, GFP_KERNEL);
313 	if (!packet)
314 		return -ENOMEM;
315 
316 	if (copy_from_user(&packet->mad, buf,
317 			    sizeof (struct ib_user_mad) + IB_MGMT_RMPP_HDR)) {
318 		ret = -EFAULT;
319 		goto err;
320 	}
321 
322 	if (packet->mad.hdr.id < 0 ||
323 	    packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
324 		ret = -EINVAL;
325 		goto err;
326 	}
327 
328 	down_read(&file->port->mutex);
329 
330 	agent = file->agent[packet->mad.hdr.id];
331 	if (!agent) {
332 		ret = -EINVAL;
333 		goto err_up;
334 	}
335 
336 	memset(&ah_attr, 0, sizeof ah_attr);
337 	ah_attr.dlid          = be16_to_cpu(packet->mad.hdr.lid);
338 	ah_attr.sl            = packet->mad.hdr.sl;
339 	ah_attr.src_path_bits = packet->mad.hdr.path_bits;
340 	ah_attr.port_num      = file->port->port_num;
341 	if (packet->mad.hdr.grh_present) {
342 		ah_attr.ah_flags = IB_AH_GRH;
343 		memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
344 		ah_attr.grh.flow_label 	   = be32_to_cpu(packet->mad.hdr.flow_label);
345 		ah_attr.grh.hop_limit  	   = packet->mad.hdr.hop_limit;
346 		ah_attr.grh.traffic_class  = packet->mad.hdr.traffic_class;
347 	}
348 
349 	ah = ib_create_ah(agent->qp->pd, &ah_attr);
350 	if (IS_ERR(ah)) {
351 		ret = PTR_ERR(ah);
352 		goto err_up;
353 	}
354 
355 	rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
356 	if (ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) & IB_MGMT_RMPP_FLAG_ACTIVE) {
357 		/* RMPP active */
358 		if (!agent->rmpp_version) {
359 			ret = -EINVAL;
360 			goto err_ah;
361 		}
362 
363 		/* Validate that the management class can support RMPP */
364 		if (rmpp_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_ADM) {
365 			hdr_len = IB_MGMT_SA_HDR;
366 		} else if ((rmpp_mad->mad_hdr.mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
367 			    (rmpp_mad->mad_hdr.mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)) {
368 				hdr_len = IB_MGMT_VENDOR_HDR;
369 		} else {
370 			ret = -EINVAL;
371 			goto err_ah;
372 		}
373 		rmpp_active = 1;
374 		copy_offset = IB_MGMT_RMPP_HDR;
375 	} else {
376 		hdr_len = IB_MGMT_MAD_HDR;
377 		copy_offset = IB_MGMT_MAD_HDR;
378 	}
379 
380 	packet->msg = ib_create_send_mad(agent,
381 					 be32_to_cpu(packet->mad.hdr.qpn),
382 					 0, rmpp_active,
383 					 hdr_len, length - hdr_len,
384 					 GFP_KERNEL);
385 	if (IS_ERR(packet->msg)) {
386 		ret = PTR_ERR(packet->msg);
387 		goto err_ah;
388 	}
389 
390 	packet->msg->ah 	= ah;
391 	packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
392 	packet->msg->retries 	= packet->mad.hdr.retries;
393 	packet->msg->context[0] = packet;
394 
395 	/* Copy MAD headers (RMPP header in place) */
396 	memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
397 	/* Now, copy rest of message from user into send buffer */
398 	if (copy_from_user(packet->msg->mad + copy_offset,
399 			   buf + sizeof (struct ib_user_mad) + copy_offset,
400 			   length - copy_offset)) {
401 		ret = -EFAULT;
402 		goto err_msg;
403 	}
404 
405 	/*
406 	 * If userspace is generating a request that will generate a
407 	 * response, we need to make sure the high-order part of the
408 	 * transaction ID matches the agent being used to send the
409 	 * MAD.
410 	 */
411 	method = ((struct ib_mad_hdr *) packet->msg->mad)->method;
412 
413 	if (!(method & IB_MGMT_METHOD_RESP)       &&
414 	    method != IB_MGMT_METHOD_TRAP_REPRESS &&
415 	    method != IB_MGMT_METHOD_SEND) {
416 		tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
417 		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
418 				   (be64_to_cpup(tid) & 0xffffffff));
419 	}
420 
421 	ret = ib_post_send_mad(packet->msg, NULL);
422 	if (ret)
423 		goto err_msg;
424 
425 	up_read(&file->port->mutex);
426 
427 	return count;
428 
429 err_msg:
430 	ib_free_send_mad(packet->msg);
431 
432 err_ah:
433 	ib_destroy_ah(ah);
434 
435 err_up:
436 	up_read(&file->port->mutex);
437 
438 err:
439 	kfree(packet);
440 	return ret;
441 }
442 
443 static unsigned int ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
444 {
445 	struct ib_umad_file *file = filp->private_data;
446 
447 	/* we will always be able to post a MAD send */
448 	unsigned int mask = POLLOUT | POLLWRNORM;
449 
450 	poll_wait(filp, &file->recv_wait, wait);
451 
452 	if (!list_empty(&file->recv_list))
453 		mask |= POLLIN | POLLRDNORM;
454 
455 	return mask;
456 }
457 
458 static int ib_umad_reg_agent(struct ib_umad_file *file, unsigned long arg)
459 {
460 	struct ib_user_mad_reg_req ureq;
461 	struct ib_mad_reg_req req;
462 	struct ib_mad_agent *agent;
463 	int agent_id;
464 	int ret;
465 
466 	down_write(&file->port->mutex);
467 
468 	if (!file->port->ib_dev) {
469 		ret = -EPIPE;
470 		goto out;
471 	}
472 
473 	if (copy_from_user(&ureq, (void __user *) arg, sizeof ureq)) {
474 		ret = -EFAULT;
475 		goto out;
476 	}
477 
478 	if (ureq.qpn != 0 && ureq.qpn != 1) {
479 		ret = -EINVAL;
480 		goto out;
481 	}
482 
483 	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
484 		if (!file->agent[agent_id])
485 			goto found;
486 
487 	ret = -ENOMEM;
488 	goto out;
489 
490 found:
491 	if (ureq.mgmt_class) {
492 		req.mgmt_class         = ureq.mgmt_class;
493 		req.mgmt_class_version = ureq.mgmt_class_version;
494 		memcpy(req.method_mask, ureq.method_mask, sizeof req.method_mask);
495 		memcpy(req.oui,         ureq.oui,         sizeof req.oui);
496 	}
497 
498 	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
499 				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
500 				      ureq.mgmt_class ? &req : NULL,
501 				      ureq.rmpp_version,
502 				      send_handler, recv_handler, file);
503 	if (IS_ERR(agent)) {
504 		ret = PTR_ERR(agent);
505 		goto out;
506 	}
507 
508 	file->agent[agent_id] = agent;
509 
510 	file->mr[agent_id] = ib_get_dma_mr(agent->qp->pd, IB_ACCESS_LOCAL_WRITE);
511 	if (IS_ERR(file->mr[agent_id])) {
512 		ret = -ENOMEM;
513 		goto err;
514 	}
515 
516 	if (put_user(agent_id,
517 		     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
518 		ret = -EFAULT;
519 		goto err_mr;
520 	}
521 
522 	ret = 0;
523 	goto out;
524 
525 err_mr:
526 	ib_dereg_mr(file->mr[agent_id]);
527 
528 err:
529 	file->agent[agent_id] = NULL;
530 	ib_unregister_mad_agent(agent);
531 
532 out:
533 	up_write(&file->port->mutex);
534 	return ret;
535 }
536 
537 static int ib_umad_unreg_agent(struct ib_umad_file *file, unsigned long arg)
538 {
539 	u32 id;
540 	int ret = 0;
541 
542 	down_write(&file->port->mutex);
543 
544 	if (get_user(id, (u32 __user *) arg)) {
545 		ret = -EFAULT;
546 		goto out;
547 	}
548 
549 	if (id < 0 || id >= IB_UMAD_MAX_AGENTS || !file->agent[id]) {
550 		ret = -EINVAL;
551 		goto out;
552 	}
553 
554 	ib_dereg_mr(file->mr[id]);
555 	ib_unregister_mad_agent(file->agent[id]);
556 	file->agent[id] = NULL;
557 
558 out:
559 	up_write(&file->port->mutex);
560 	return ret;
561 }
562 
563 static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
564 			  unsigned long arg)
565 {
566 	switch (cmd) {
567 	case IB_USER_MAD_REGISTER_AGENT:
568 		return ib_umad_reg_agent(filp->private_data, arg);
569 	case IB_USER_MAD_UNREGISTER_AGENT:
570 		return ib_umad_unreg_agent(filp->private_data, arg);
571 	default:
572 		return -ENOIOCTLCMD;
573 	}
574 }
575 
576 static int ib_umad_open(struct inode *inode, struct file *filp)
577 {
578 	struct ib_umad_port *port;
579 	struct ib_umad_file *file;
580 	int ret = 0;
581 
582 	spin_lock(&port_lock);
583 	port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE];
584 	if (port)
585 		kref_get(&port->umad_dev->ref);
586 	spin_unlock(&port_lock);
587 
588 	if (!port)
589 		return -ENXIO;
590 
591 	down_write(&port->mutex);
592 
593 	if (!port->ib_dev) {
594 		ret = -ENXIO;
595 		goto out;
596 	}
597 
598 	file = kzalloc(sizeof *file, GFP_KERNEL);
599 	if (!file) {
600 		kref_put(&port->umad_dev->ref, ib_umad_release_dev);
601 		ret = -ENOMEM;
602 		goto out;
603 	}
604 
605 	spin_lock_init(&file->recv_lock);
606 	INIT_LIST_HEAD(&file->recv_list);
607 	init_waitqueue_head(&file->recv_wait);
608 
609 	file->port = port;
610 	filp->private_data = file;
611 
612 	list_add_tail(&file->port_list, &port->file_list);
613 
614 out:
615 	up_write(&port->mutex);
616 	return ret;
617 }
618 
619 static int ib_umad_close(struct inode *inode, struct file *filp)
620 {
621 	struct ib_umad_file *file = filp->private_data;
622 	struct ib_umad_device *dev = file->port->umad_dev;
623 	struct ib_umad_packet *packet, *tmp;
624 	int i;
625 
626 	down_write(&file->port->mutex);
627 	for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
628 		if (file->agent[i]) {
629 			ib_dereg_mr(file->mr[i]);
630 			ib_unregister_mad_agent(file->agent[i]);
631 		}
632 
633 	list_for_each_entry_safe(packet, tmp, &file->recv_list, list)
634 		kfree(packet);
635 
636 	list_del(&file->port_list);
637 	up_write(&file->port->mutex);
638 
639 	kfree(file);
640 
641 	kref_put(&dev->ref, ib_umad_release_dev);
642 
643 	return 0;
644 }
645 
646 static struct file_operations umad_fops = {
647 	.owner 	 	= THIS_MODULE,
648 	.read 	 	= ib_umad_read,
649 	.write 	 	= ib_umad_write,
650 	.poll 	 	= ib_umad_poll,
651 	.unlocked_ioctl = ib_umad_ioctl,
652 	.compat_ioctl 	= ib_umad_ioctl,
653 	.open 	 	= ib_umad_open,
654 	.release 	= ib_umad_close
655 };
656 
657 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
658 {
659 	struct ib_umad_port *port;
660 	struct ib_port_modify props = {
661 		.set_port_cap_mask = IB_PORT_SM
662 	};
663 	int ret;
664 
665 	spin_lock(&port_lock);
666 	port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE - IB_UMAD_MAX_PORTS];
667 	if (port)
668 		kref_get(&port->umad_dev->ref);
669 	spin_unlock(&port_lock);
670 
671 	if (!port)
672 		return -ENXIO;
673 
674 	if (filp->f_flags & O_NONBLOCK) {
675 		if (down_trylock(&port->sm_sem)) {
676 			ret = -EAGAIN;
677 			goto fail;
678 		}
679 	} else {
680 		if (down_interruptible(&port->sm_sem)) {
681 			ret = -ERESTARTSYS;
682 			goto fail;
683 		}
684 	}
685 
686 	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
687 	if (ret) {
688 		up(&port->sm_sem);
689 		goto fail;
690 	}
691 
692 	filp->private_data = port;
693 
694 	return 0;
695 
696 fail:
697 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
698 	return ret;
699 }
700 
701 static int ib_umad_sm_close(struct inode *inode, struct file *filp)
702 {
703 	struct ib_umad_port *port = filp->private_data;
704 	struct ib_port_modify props = {
705 		.clr_port_cap_mask = IB_PORT_SM
706 	};
707 	int ret = 0;
708 
709 	down_write(&port->mutex);
710 	if (port->ib_dev)
711 		ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
712 	up_write(&port->mutex);
713 
714 	up(&port->sm_sem);
715 
716 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
717 
718 	return ret;
719 }
720 
721 static struct file_operations umad_sm_fops = {
722 	.owner 	 = THIS_MODULE,
723 	.open 	 = ib_umad_sm_open,
724 	.release = ib_umad_sm_close
725 };
726 
727 static struct ib_client umad_client = {
728 	.name   = "umad",
729 	.add    = ib_umad_add_one,
730 	.remove = ib_umad_remove_one
731 };
732 
733 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
734 {
735 	struct ib_umad_port *port = class_get_devdata(class_dev);
736 
737 	if (!port)
738 		return -ENODEV;
739 
740 	return sprintf(buf, "%s\n", port->ib_dev->name);
741 }
742 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
743 
744 static ssize_t show_port(struct class_device *class_dev, char *buf)
745 {
746 	struct ib_umad_port *port = class_get_devdata(class_dev);
747 
748 	if (!port)
749 		return -ENODEV;
750 
751 	return sprintf(buf, "%d\n", port->port_num);
752 }
753 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
754 
755 static ssize_t show_abi_version(struct class *class, char *buf)
756 {
757 	return sprintf(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
758 }
759 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
760 
761 static int ib_umad_init_port(struct ib_device *device, int port_num,
762 			     struct ib_umad_port *port)
763 {
764 	spin_lock(&port_lock);
765 	port->dev_num = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
766 	if (port->dev_num >= IB_UMAD_MAX_PORTS) {
767 		spin_unlock(&port_lock);
768 		return -1;
769 	}
770 	set_bit(port->dev_num, dev_map);
771 	spin_unlock(&port_lock);
772 
773 	port->ib_dev   = device;
774 	port->port_num = port_num;
775 	init_MUTEX(&port->sm_sem);
776 	init_rwsem(&port->mutex);
777 	INIT_LIST_HEAD(&port->file_list);
778 
779 	port->dev = cdev_alloc();
780 	if (!port->dev)
781 		return -1;
782 	port->dev->owner = THIS_MODULE;
783 	port->dev->ops   = &umad_fops;
784 	kobject_set_name(&port->dev->kobj, "umad%d", port->dev_num);
785 	if (cdev_add(port->dev, base_dev + port->dev_num, 1))
786 		goto err_cdev;
787 
788 	port->class_dev = class_device_create(umad_class, NULL, port->dev->dev,
789 					      device->dma_device,
790 					      "umad%d", port->dev_num);
791 	if (IS_ERR(port->class_dev))
792 		goto err_cdev;
793 
794 	if (class_device_create_file(port->class_dev, &class_device_attr_ibdev))
795 		goto err_class;
796 	if (class_device_create_file(port->class_dev, &class_device_attr_port))
797 		goto err_class;
798 
799 	port->sm_dev = cdev_alloc();
800 	if (!port->sm_dev)
801 		goto err_class;
802 	port->sm_dev->owner = THIS_MODULE;
803 	port->sm_dev->ops   = &umad_sm_fops;
804 	kobject_set_name(&port->dev->kobj, "issm%d", port->dev_num);
805 	if (cdev_add(port->sm_dev, base_dev + port->dev_num + IB_UMAD_MAX_PORTS, 1))
806 		goto err_sm_cdev;
807 
808 	port->sm_class_dev = class_device_create(umad_class, NULL, port->sm_dev->dev,
809 						 device->dma_device,
810 						 "issm%d", port->dev_num);
811 	if (IS_ERR(port->sm_class_dev))
812 		goto err_sm_cdev;
813 
814 	class_set_devdata(port->class_dev,    port);
815 	class_set_devdata(port->sm_class_dev, port);
816 
817 	if (class_device_create_file(port->sm_class_dev, &class_device_attr_ibdev))
818 		goto err_sm_class;
819 	if (class_device_create_file(port->sm_class_dev, &class_device_attr_port))
820 		goto err_sm_class;
821 
822 	spin_lock(&port_lock);
823 	umad_port[port->dev_num] = port;
824 	spin_unlock(&port_lock);
825 
826 	return 0;
827 
828 err_sm_class:
829 	class_device_destroy(umad_class, port->sm_dev->dev);
830 
831 err_sm_cdev:
832 	cdev_del(port->sm_dev);
833 
834 err_class:
835 	class_device_destroy(umad_class, port->dev->dev);
836 
837 err_cdev:
838 	cdev_del(port->dev);
839 	clear_bit(port->dev_num, dev_map);
840 
841 	return -1;
842 }
843 
844 static void ib_umad_kill_port(struct ib_umad_port *port)
845 {
846 	struct ib_umad_file *file;
847 	int id;
848 
849 	class_set_devdata(port->class_dev,    NULL);
850 	class_set_devdata(port->sm_class_dev, NULL);
851 
852 	class_device_destroy(umad_class, port->dev->dev);
853 	class_device_destroy(umad_class, port->sm_dev->dev);
854 
855 	cdev_del(port->dev);
856 	cdev_del(port->sm_dev);
857 
858 	spin_lock(&port_lock);
859 	umad_port[port->dev_num] = NULL;
860 	spin_unlock(&port_lock);
861 
862 	down_write(&port->mutex);
863 
864 	port->ib_dev = NULL;
865 
866 	list_for_each_entry(file, &port->file_list, port_list)
867 		for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id) {
868 			if (!file->agent[id])
869 				continue;
870 			ib_dereg_mr(file->mr[id]);
871 			ib_unregister_mad_agent(file->agent[id]);
872 			file->agent[id] = NULL;
873 		}
874 
875 	up_write(&port->mutex);
876 
877 	clear_bit(port->dev_num, dev_map);
878 }
879 
880 static void ib_umad_add_one(struct ib_device *device)
881 {
882 	struct ib_umad_device *umad_dev;
883 	int s, e, i;
884 
885 	if (device->node_type == IB_NODE_SWITCH)
886 		s = e = 0;
887 	else {
888 		s = 1;
889 		e = device->phys_port_cnt;
890 	}
891 
892 	umad_dev = kzalloc(sizeof *umad_dev +
893 			   (e - s + 1) * sizeof (struct ib_umad_port),
894 			   GFP_KERNEL);
895 	if (!umad_dev)
896 		return;
897 
898 	kref_init(&umad_dev->ref);
899 
900 	umad_dev->start_port = s;
901 	umad_dev->end_port   = e;
902 
903 	for (i = s; i <= e; ++i) {
904 		umad_dev->port[i - s].umad_dev = umad_dev;
905 
906 		if (ib_umad_init_port(device, i, &umad_dev->port[i - s]))
907 			goto err;
908 	}
909 
910 	ib_set_client_data(device, &umad_client, umad_dev);
911 
912 	return;
913 
914 err:
915 	while (--i >= s)
916 		ib_umad_kill_port(&umad_dev->port[i]);
917 
918 	kref_put(&umad_dev->ref, ib_umad_release_dev);
919 }
920 
921 static void ib_umad_remove_one(struct ib_device *device)
922 {
923 	struct ib_umad_device *umad_dev = ib_get_client_data(device, &umad_client);
924 	int i;
925 
926 	if (!umad_dev)
927 		return;
928 
929 	for (i = 0; i <= umad_dev->end_port - umad_dev->start_port; ++i)
930 		ib_umad_kill_port(&umad_dev->port[i]);
931 
932 	kref_put(&umad_dev->ref, ib_umad_release_dev);
933 }
934 
935 static int __init ib_umad_init(void)
936 {
937 	int ret;
938 
939 	ret = register_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2,
940 				     "infiniband_mad");
941 	if (ret) {
942 		printk(KERN_ERR "user_mad: couldn't register device number\n");
943 		goto out;
944 	}
945 
946 	umad_class = class_create(THIS_MODULE, "infiniband_mad");
947 	if (IS_ERR(umad_class)) {
948 		ret = PTR_ERR(umad_class);
949 		printk(KERN_ERR "user_mad: couldn't create class infiniband_mad\n");
950 		goto out_chrdev;
951 	}
952 
953 	ret = class_create_file(umad_class, &class_attr_abi_version);
954 	if (ret) {
955 		printk(KERN_ERR "user_mad: couldn't create abi_version attribute\n");
956 		goto out_class;
957 	}
958 
959 	ret = ib_register_client(&umad_client);
960 	if (ret) {
961 		printk(KERN_ERR "user_mad: couldn't register ib_umad client\n");
962 		goto out_class;
963 	}
964 
965 	return 0;
966 
967 out_class:
968 	class_destroy(umad_class);
969 
970 out_chrdev:
971 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
972 
973 out:
974 	return ret;
975 }
976 
977 static void __exit ib_umad_cleanup(void)
978 {
979 	ib_unregister_client(&umad_client);
980 	class_destroy(umad_class);
981 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
982 }
983 
984 module_init(ib_umad_init);
985 module_exit(ib_umad_cleanup);
986