xref: /linux/drivers/infiniband/core/user_mad.c (revision de2fe5e07d58424bc286fff3fd3c1b0bf933cd58)
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	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 	int			agents_dead;
120 };
121 
122 struct ib_umad_packet {
123 	struct ib_mad_send_buf *msg;
124 	struct ib_mad_recv_wc  *recv_wc;
125 	struct list_head   list;
126 	int		   length;
127 	struct ib_user_mad mad;
128 };
129 
130 static struct class *umad_class;
131 
132 static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
133 
134 static DEFINE_SPINLOCK(port_lock);
135 static struct ib_umad_port *umad_port[IB_UMAD_MAX_PORTS];
136 static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS * 2);
137 
138 static void ib_umad_add_one(struct ib_device *device);
139 static void ib_umad_remove_one(struct ib_device *device);
140 
141 static void ib_umad_release_dev(struct kref *ref)
142 {
143 	struct ib_umad_device *dev =
144 		container_of(ref, struct ib_umad_device, ref);
145 
146 	kfree(dev);
147 }
148 
149 /* caller must hold port->mutex at least for reading */
150 static struct ib_mad_agent *__get_agent(struct ib_umad_file *file, int id)
151 {
152 	return file->agents_dead ? NULL : file->agent[id];
153 }
154 
155 static int queue_packet(struct ib_umad_file *file,
156 			struct ib_mad_agent *agent,
157 			struct ib_umad_packet *packet)
158 {
159 	int ret = 1;
160 
161 	down_read(&file->port->mutex);
162 
163 	for (packet->mad.hdr.id = 0;
164 	     packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
165 	     packet->mad.hdr.id++)
166 		if (agent == __get_agent(file, packet->mad.hdr.id)) {
167 			spin_lock_irq(&file->recv_lock);
168 			list_add_tail(&packet->list, &file->recv_list);
169 			spin_unlock_irq(&file->recv_lock);
170 			wake_up_interruptible(&file->recv_wait);
171 			ret = 0;
172 			break;
173 		}
174 
175 	up_read(&file->port->mutex);
176 
177 	return ret;
178 }
179 
180 static void send_handler(struct ib_mad_agent *agent,
181 			 struct ib_mad_send_wc *send_wc)
182 {
183 	struct ib_umad_file *file = agent->context;
184 	struct ib_umad_packet *packet = send_wc->send_buf->context[0];
185 
186 	ib_destroy_ah(packet->msg->ah);
187 	ib_free_send_mad(packet->msg);
188 
189 	if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
190 		packet->length = IB_MGMT_MAD_HDR;
191 		packet->mad.hdr.status = ETIMEDOUT;
192 		if (!queue_packet(file, agent, packet))
193 			return;
194 	}
195 	kfree(packet);
196 }
197 
198 static void recv_handler(struct ib_mad_agent *agent,
199 			 struct ib_mad_recv_wc *mad_recv_wc)
200 {
201 	struct ib_umad_file *file = agent->context;
202 	struct ib_umad_packet *packet;
203 
204 	if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
205 		goto err1;
206 
207 	packet = kzalloc(sizeof *packet, GFP_KERNEL);
208 	if (!packet)
209 		goto err1;
210 
211 	packet->length = mad_recv_wc->mad_len;
212 	packet->recv_wc = mad_recv_wc;
213 
214 	packet->mad.hdr.status    = 0;
215 	packet->mad.hdr.length    = sizeof (struct ib_user_mad) +
216 				    mad_recv_wc->mad_len;
217 	packet->mad.hdr.qpn 	  = cpu_to_be32(mad_recv_wc->wc->src_qp);
218 	packet->mad.hdr.lid 	  = cpu_to_be16(mad_recv_wc->wc->slid);
219 	packet->mad.hdr.sl  	  = mad_recv_wc->wc->sl;
220 	packet->mad.hdr.path_bits = mad_recv_wc->wc->dlid_path_bits;
221 	packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
222 	if (packet->mad.hdr.grh_present) {
223 		/* XXX parse GRH */
224 		packet->mad.hdr.gid_index 	= 0;
225 		packet->mad.hdr.hop_limit 	= 0;
226 		packet->mad.hdr.traffic_class	= 0;
227 		memset(packet->mad.hdr.gid, 0, 16);
228 		packet->mad.hdr.flow_label	= 0;
229 	}
230 
231 	if (queue_packet(file, agent, packet))
232 		goto err2;
233 	return;
234 
235 err2:
236 	kfree(packet);
237 err1:
238 	ib_free_recv_mad(mad_recv_wc);
239 }
240 
241 static ssize_t copy_recv_mad(char __user *buf, struct ib_umad_packet *packet,
242 			     size_t count)
243 {
244 	struct ib_mad_recv_buf *recv_buf;
245 	int left, seg_payload, offset, max_seg_payload;
246 
247 	/* We need enough room to copy the first (or only) MAD segment. */
248 	recv_buf = &packet->recv_wc->recv_buf;
249 	if ((packet->length <= sizeof (*recv_buf->mad) &&
250 	     count < sizeof (packet->mad) + packet->length) ||
251 	    (packet->length > sizeof (*recv_buf->mad) &&
252 	     count < sizeof (packet->mad) + sizeof (*recv_buf->mad)))
253 		return -EINVAL;
254 
255 	if (copy_to_user(buf, &packet->mad, sizeof (packet->mad)))
256 		return -EFAULT;
257 
258 	buf += sizeof (packet->mad);
259 	seg_payload = min_t(int, packet->length, sizeof (*recv_buf->mad));
260 	if (copy_to_user(buf, recv_buf->mad, seg_payload))
261 		return -EFAULT;
262 
263 	if (seg_payload < packet->length) {
264 		/*
265 		 * Multipacket RMPP MAD message. Copy remainder of message.
266 		 * Note that last segment may have a shorter payload.
267 		 */
268 		if (count < sizeof (packet->mad) + packet->length) {
269 			/*
270 			 * The buffer is too small, return the first RMPP segment,
271 			 * which includes the RMPP message length.
272 			 */
273 			return -ENOSPC;
274 		}
275 		offset = ib_get_mad_data_offset(recv_buf->mad->mad_hdr.mgmt_class);
276 		max_seg_payload = sizeof (struct ib_mad) - offset;
277 
278 		for (left = packet->length - seg_payload, buf += seg_payload;
279 		     left; left -= seg_payload, buf += seg_payload) {
280 			recv_buf = container_of(recv_buf->list.next,
281 						struct ib_mad_recv_buf, list);
282 			seg_payload = min(left, max_seg_payload);
283 			if (copy_to_user(buf, ((void *) recv_buf->mad) + offset,
284 					 seg_payload))
285 				return -EFAULT;
286 		}
287 	}
288 	return sizeof (packet->mad) + packet->length;
289 }
290 
291 static ssize_t copy_send_mad(char __user *buf, struct ib_umad_packet *packet,
292 			     size_t count)
293 {
294 	ssize_t size = sizeof (packet->mad) + packet->length;
295 
296 	if (count < size)
297 		return -EINVAL;
298 
299 	if (copy_to_user(buf, &packet->mad, size))
300 		return -EFAULT;
301 
302 	return size;
303 }
304 
305 static ssize_t ib_umad_read(struct file *filp, char __user *buf,
306 			    size_t count, loff_t *pos)
307 {
308 	struct ib_umad_file *file = filp->private_data;
309 	struct ib_umad_packet *packet;
310 	ssize_t ret;
311 
312 	if (count < sizeof (struct ib_user_mad))
313 		return -EINVAL;
314 
315 	spin_lock_irq(&file->recv_lock);
316 
317 	while (list_empty(&file->recv_list)) {
318 		spin_unlock_irq(&file->recv_lock);
319 
320 		if (filp->f_flags & O_NONBLOCK)
321 			return -EAGAIN;
322 
323 		if (wait_event_interruptible(file->recv_wait,
324 					     !list_empty(&file->recv_list)))
325 			return -ERESTARTSYS;
326 
327 		spin_lock_irq(&file->recv_lock);
328 	}
329 
330 	packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
331 	list_del(&packet->list);
332 
333 	spin_unlock_irq(&file->recv_lock);
334 
335 	if (packet->recv_wc)
336 		ret = copy_recv_mad(buf, packet, count);
337 	else
338 		ret = copy_send_mad(buf, packet, count);
339 
340 	if (ret < 0) {
341 		/* Requeue packet */
342 		spin_lock_irq(&file->recv_lock);
343 		list_add(&packet->list, &file->recv_list);
344 		spin_unlock_irq(&file->recv_lock);
345 	} else {
346 		if (packet->recv_wc)
347 			ib_free_recv_mad(packet->recv_wc);
348 		kfree(packet);
349 	}
350 	return ret;
351 }
352 
353 static int copy_rmpp_mad(struct ib_mad_send_buf *msg, const char __user *buf)
354 {
355 	int left, seg;
356 
357 	/* Copy class specific header */
358 	if ((msg->hdr_len > IB_MGMT_RMPP_HDR) &&
359 	    copy_from_user(msg->mad + IB_MGMT_RMPP_HDR, buf + IB_MGMT_RMPP_HDR,
360 			   msg->hdr_len - IB_MGMT_RMPP_HDR))
361 		return -EFAULT;
362 
363 	/* All headers are in place.  Copy data segments. */
364 	for (seg = 1, left = msg->data_len, buf += msg->hdr_len; left > 0;
365 	     seg++, left -= msg->seg_size, buf += msg->seg_size) {
366 		if (copy_from_user(ib_get_rmpp_segment(msg, seg), buf,
367 				   min(left, msg->seg_size)))
368 			return -EFAULT;
369 	}
370 	return 0;
371 }
372 
373 static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
374 			     size_t count, loff_t *pos)
375 {
376 	struct ib_umad_file *file = filp->private_data;
377 	struct ib_umad_packet *packet;
378 	struct ib_mad_agent *agent;
379 	struct ib_ah_attr ah_attr;
380 	struct ib_ah *ah;
381 	struct ib_rmpp_mad *rmpp_mad;
382 	u8 method;
383 	__be64 *tid;
384 	int ret, data_len, hdr_len, copy_offset, rmpp_active;
385 
386 	if (count < sizeof (struct ib_user_mad) + IB_MGMT_RMPP_HDR)
387 		return -EINVAL;
388 
389 	packet = kzalloc(sizeof *packet + IB_MGMT_RMPP_HDR, GFP_KERNEL);
390 	if (!packet)
391 		return -ENOMEM;
392 
393 	if (copy_from_user(&packet->mad, buf,
394 			    sizeof (struct ib_user_mad) + IB_MGMT_RMPP_HDR)) {
395 		ret = -EFAULT;
396 		goto err;
397 	}
398 
399 	if (packet->mad.hdr.id < 0 ||
400 	    packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
401 		ret = -EINVAL;
402 		goto err;
403 	}
404 
405 	down_read(&file->port->mutex);
406 
407 	agent = __get_agent(file, packet->mad.hdr.id);
408 	if (!agent) {
409 		ret = -EINVAL;
410 		goto err_up;
411 	}
412 
413 	memset(&ah_attr, 0, sizeof ah_attr);
414 	ah_attr.dlid          = be16_to_cpu(packet->mad.hdr.lid);
415 	ah_attr.sl            = packet->mad.hdr.sl;
416 	ah_attr.src_path_bits = packet->mad.hdr.path_bits;
417 	ah_attr.port_num      = file->port->port_num;
418 	if (packet->mad.hdr.grh_present) {
419 		ah_attr.ah_flags = IB_AH_GRH;
420 		memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
421 		ah_attr.grh.flow_label 	   = be32_to_cpu(packet->mad.hdr.flow_label);
422 		ah_attr.grh.hop_limit  	   = packet->mad.hdr.hop_limit;
423 		ah_attr.grh.traffic_class  = packet->mad.hdr.traffic_class;
424 	}
425 
426 	ah = ib_create_ah(agent->qp->pd, &ah_attr);
427 	if (IS_ERR(ah)) {
428 		ret = PTR_ERR(ah);
429 		goto err_up;
430 	}
431 
432 	rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
433 	hdr_len = ib_get_mad_data_offset(rmpp_mad->mad_hdr.mgmt_class);
434 	if (!ib_is_mad_class_rmpp(rmpp_mad->mad_hdr.mgmt_class)) {
435 		copy_offset = IB_MGMT_MAD_HDR;
436 		rmpp_active = 0;
437 	} else {
438 		copy_offset = IB_MGMT_RMPP_HDR;
439 		rmpp_active = ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
440 			      IB_MGMT_RMPP_FLAG_ACTIVE;
441 	}
442 
443 	data_len = count - sizeof (struct ib_user_mad) - hdr_len;
444 	packet->msg = ib_create_send_mad(agent,
445 					 be32_to_cpu(packet->mad.hdr.qpn),
446 					 0, rmpp_active, hdr_len,
447 					 data_len, GFP_KERNEL);
448 	if (IS_ERR(packet->msg)) {
449 		ret = PTR_ERR(packet->msg);
450 		goto err_ah;
451 	}
452 
453 	packet->msg->ah 	= ah;
454 	packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
455 	packet->msg->retries 	= packet->mad.hdr.retries;
456 	packet->msg->context[0] = packet;
457 
458 	/* Copy MAD header.  Any RMPP header is already in place. */
459 	memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
460 	buf += sizeof (struct ib_user_mad);
461 
462 	if (!rmpp_active) {
463 		if (copy_from_user(packet->msg->mad + copy_offset,
464 				   buf + copy_offset,
465 				   hdr_len + data_len - copy_offset)) {
466 			ret = -EFAULT;
467 			goto err_msg;
468 		}
469 	} else {
470 		ret = copy_rmpp_mad(packet->msg, buf);
471 		if (ret)
472 			goto err_msg;
473 	}
474 
475 	/*
476 	 * If userspace is generating a request that will generate a
477 	 * response, we need to make sure the high-order part of the
478 	 * transaction ID matches the agent being used to send the
479 	 * MAD.
480 	 */
481 	method = ((struct ib_mad_hdr *) packet->msg->mad)->method;
482 
483 	if (!(method & IB_MGMT_METHOD_RESP)       &&
484 	    method != IB_MGMT_METHOD_TRAP_REPRESS &&
485 	    method != IB_MGMT_METHOD_SEND) {
486 		tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
487 		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
488 				   (be64_to_cpup(tid) & 0xffffffff));
489 	}
490 
491 	ret = ib_post_send_mad(packet->msg, NULL);
492 	if (ret)
493 		goto err_msg;
494 
495 	up_read(&file->port->mutex);
496 	return count;
497 
498 err_msg:
499 	ib_free_send_mad(packet->msg);
500 err_ah:
501 	ib_destroy_ah(ah);
502 err_up:
503 	up_read(&file->port->mutex);
504 err:
505 	kfree(packet);
506 	return ret;
507 }
508 
509 static unsigned int ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
510 {
511 	struct ib_umad_file *file = filp->private_data;
512 
513 	/* we will always be able to post a MAD send */
514 	unsigned int mask = POLLOUT | POLLWRNORM;
515 
516 	poll_wait(filp, &file->recv_wait, wait);
517 
518 	if (!list_empty(&file->recv_list))
519 		mask |= POLLIN | POLLRDNORM;
520 
521 	return mask;
522 }
523 
524 static int ib_umad_reg_agent(struct ib_umad_file *file, unsigned long arg)
525 {
526 	struct ib_user_mad_reg_req ureq;
527 	struct ib_mad_reg_req req;
528 	struct ib_mad_agent *agent;
529 	int agent_id;
530 	int ret;
531 
532 	down_write(&file->port->mutex);
533 
534 	if (!file->port->ib_dev) {
535 		ret = -EPIPE;
536 		goto out;
537 	}
538 
539 	if (copy_from_user(&ureq, (void __user *) arg, sizeof ureq)) {
540 		ret = -EFAULT;
541 		goto out;
542 	}
543 
544 	if (ureq.qpn != 0 && ureq.qpn != 1) {
545 		ret = -EINVAL;
546 		goto out;
547 	}
548 
549 	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
550 		if (!__get_agent(file, agent_id))
551 			goto found;
552 
553 	ret = -ENOMEM;
554 	goto out;
555 
556 found:
557 	if (ureq.mgmt_class) {
558 		req.mgmt_class         = ureq.mgmt_class;
559 		req.mgmt_class_version = ureq.mgmt_class_version;
560 		memcpy(req.method_mask, ureq.method_mask, sizeof req.method_mask);
561 		memcpy(req.oui,         ureq.oui,         sizeof req.oui);
562 	}
563 
564 	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
565 				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
566 				      ureq.mgmt_class ? &req : NULL,
567 				      ureq.rmpp_version,
568 				      send_handler, recv_handler, file);
569 	if (IS_ERR(agent)) {
570 		ret = PTR_ERR(agent);
571 		goto out;
572 	}
573 
574 	if (put_user(agent_id,
575 		     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
576 		ret = -EFAULT;
577 		ib_unregister_mad_agent(agent);
578 		goto out;
579 	}
580 
581 	file->agent[agent_id] = agent;
582 	ret = 0;
583 
584 out:
585 	up_write(&file->port->mutex);
586 	return ret;
587 }
588 
589 static int ib_umad_unreg_agent(struct ib_umad_file *file, unsigned long arg)
590 {
591 	struct ib_mad_agent *agent = NULL;
592 	u32 id;
593 	int ret = 0;
594 
595 	if (get_user(id, (u32 __user *) arg))
596 		return -EFAULT;
597 
598 	down_write(&file->port->mutex);
599 
600 	if (id < 0 || id >= IB_UMAD_MAX_AGENTS || !__get_agent(file, id)) {
601 		ret = -EINVAL;
602 		goto out;
603 	}
604 
605 	agent = file->agent[id];
606 	file->agent[id] = NULL;
607 
608 out:
609 	up_write(&file->port->mutex);
610 
611 	if (agent)
612 		ib_unregister_mad_agent(agent);
613 
614 	return ret;
615 }
616 
617 static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
618 			  unsigned long arg)
619 {
620 	switch (cmd) {
621 	case IB_USER_MAD_REGISTER_AGENT:
622 		return ib_umad_reg_agent(filp->private_data, arg);
623 	case IB_USER_MAD_UNREGISTER_AGENT:
624 		return ib_umad_unreg_agent(filp->private_data, arg);
625 	default:
626 		return -ENOIOCTLCMD;
627 	}
628 }
629 
630 static int ib_umad_open(struct inode *inode, struct file *filp)
631 {
632 	struct ib_umad_port *port;
633 	struct ib_umad_file *file;
634 	int ret = 0;
635 
636 	spin_lock(&port_lock);
637 	port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE];
638 	if (port)
639 		kref_get(&port->umad_dev->ref);
640 	spin_unlock(&port_lock);
641 
642 	if (!port)
643 		return -ENXIO;
644 
645 	down_write(&port->mutex);
646 
647 	if (!port->ib_dev) {
648 		ret = -ENXIO;
649 		goto out;
650 	}
651 
652 	file = kzalloc(sizeof *file, GFP_KERNEL);
653 	if (!file) {
654 		kref_put(&port->umad_dev->ref, ib_umad_release_dev);
655 		ret = -ENOMEM;
656 		goto out;
657 	}
658 
659 	spin_lock_init(&file->recv_lock);
660 	INIT_LIST_HEAD(&file->recv_list);
661 	init_waitqueue_head(&file->recv_wait);
662 
663 	file->port = port;
664 	filp->private_data = file;
665 
666 	list_add_tail(&file->port_list, &port->file_list);
667 
668 out:
669 	up_write(&port->mutex);
670 	return ret;
671 }
672 
673 static int ib_umad_close(struct inode *inode, struct file *filp)
674 {
675 	struct ib_umad_file *file = filp->private_data;
676 	struct ib_umad_device *dev = file->port->umad_dev;
677 	struct ib_umad_packet *packet, *tmp;
678 	int already_dead;
679 	int i;
680 
681 	down_write(&file->port->mutex);
682 
683 	already_dead = file->agents_dead;
684 	file->agents_dead = 1;
685 
686 	list_for_each_entry_safe(packet, tmp, &file->recv_list, list) {
687 		if (packet->recv_wc)
688 			ib_free_recv_mad(packet->recv_wc);
689 		kfree(packet);
690 	}
691 
692 	list_del(&file->port_list);
693 
694 	downgrade_write(&file->port->mutex);
695 
696 	if (!already_dead)
697 		for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
698 			if (file->agent[i])
699 				ib_unregister_mad_agent(file->agent[i]);
700 
701 	up_read(&file->port->mutex);
702 
703 	kfree(file);
704 	kref_put(&dev->ref, ib_umad_release_dev);
705 
706 	return 0;
707 }
708 
709 static struct file_operations umad_fops = {
710 	.owner 	 	= THIS_MODULE,
711 	.read 	 	= ib_umad_read,
712 	.write 	 	= ib_umad_write,
713 	.poll 	 	= ib_umad_poll,
714 	.unlocked_ioctl = ib_umad_ioctl,
715 	.compat_ioctl 	= ib_umad_ioctl,
716 	.open 	 	= ib_umad_open,
717 	.release 	= ib_umad_close
718 };
719 
720 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
721 {
722 	struct ib_umad_port *port;
723 	struct ib_port_modify props = {
724 		.set_port_cap_mask = IB_PORT_SM
725 	};
726 	int ret;
727 
728 	spin_lock(&port_lock);
729 	port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE - IB_UMAD_MAX_PORTS];
730 	if (port)
731 		kref_get(&port->umad_dev->ref);
732 	spin_unlock(&port_lock);
733 
734 	if (!port)
735 		return -ENXIO;
736 
737 	if (filp->f_flags & O_NONBLOCK) {
738 		if (down_trylock(&port->sm_sem)) {
739 			ret = -EAGAIN;
740 			goto fail;
741 		}
742 	} else {
743 		if (down_interruptible(&port->sm_sem)) {
744 			ret = -ERESTARTSYS;
745 			goto fail;
746 		}
747 	}
748 
749 	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
750 	if (ret) {
751 		up(&port->sm_sem);
752 		goto fail;
753 	}
754 
755 	filp->private_data = port;
756 
757 	return 0;
758 
759 fail:
760 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
761 	return ret;
762 }
763 
764 static int ib_umad_sm_close(struct inode *inode, struct file *filp)
765 {
766 	struct ib_umad_port *port = filp->private_data;
767 	struct ib_port_modify props = {
768 		.clr_port_cap_mask = IB_PORT_SM
769 	};
770 	int ret = 0;
771 
772 	down_write(&port->mutex);
773 	if (port->ib_dev)
774 		ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
775 	up_write(&port->mutex);
776 
777 	up(&port->sm_sem);
778 
779 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
780 
781 	return ret;
782 }
783 
784 static struct file_operations umad_sm_fops = {
785 	.owner 	 = THIS_MODULE,
786 	.open 	 = ib_umad_sm_open,
787 	.release = ib_umad_sm_close
788 };
789 
790 static struct ib_client umad_client = {
791 	.name   = "umad",
792 	.add    = ib_umad_add_one,
793 	.remove = ib_umad_remove_one
794 };
795 
796 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
797 {
798 	struct ib_umad_port *port = class_get_devdata(class_dev);
799 
800 	if (!port)
801 		return -ENODEV;
802 
803 	return sprintf(buf, "%s\n", port->ib_dev->name);
804 }
805 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
806 
807 static ssize_t show_port(struct class_device *class_dev, char *buf)
808 {
809 	struct ib_umad_port *port = class_get_devdata(class_dev);
810 
811 	if (!port)
812 		return -ENODEV;
813 
814 	return sprintf(buf, "%d\n", port->port_num);
815 }
816 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
817 
818 static ssize_t show_abi_version(struct class *class, char *buf)
819 {
820 	return sprintf(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
821 }
822 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
823 
824 static int ib_umad_init_port(struct ib_device *device, int port_num,
825 			     struct ib_umad_port *port)
826 {
827 	spin_lock(&port_lock);
828 	port->dev_num = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
829 	if (port->dev_num >= IB_UMAD_MAX_PORTS) {
830 		spin_unlock(&port_lock);
831 		return -1;
832 	}
833 	set_bit(port->dev_num, dev_map);
834 	spin_unlock(&port_lock);
835 
836 	port->ib_dev   = device;
837 	port->port_num = port_num;
838 	init_MUTEX(&port->sm_sem);
839 	init_rwsem(&port->mutex);
840 	INIT_LIST_HEAD(&port->file_list);
841 
842 	port->dev = cdev_alloc();
843 	if (!port->dev)
844 		return -1;
845 	port->dev->owner = THIS_MODULE;
846 	port->dev->ops   = &umad_fops;
847 	kobject_set_name(&port->dev->kobj, "umad%d", port->dev_num);
848 	if (cdev_add(port->dev, base_dev + port->dev_num, 1))
849 		goto err_cdev;
850 
851 	port->class_dev = class_device_create(umad_class, NULL, port->dev->dev,
852 					      device->dma_device,
853 					      "umad%d", port->dev_num);
854 	if (IS_ERR(port->class_dev))
855 		goto err_cdev;
856 
857 	if (class_device_create_file(port->class_dev, &class_device_attr_ibdev))
858 		goto err_class;
859 	if (class_device_create_file(port->class_dev, &class_device_attr_port))
860 		goto err_class;
861 
862 	port->sm_dev = cdev_alloc();
863 	if (!port->sm_dev)
864 		goto err_class;
865 	port->sm_dev->owner = THIS_MODULE;
866 	port->sm_dev->ops   = &umad_sm_fops;
867 	kobject_set_name(&port->sm_dev->kobj, "issm%d", port->dev_num);
868 	if (cdev_add(port->sm_dev, base_dev + port->dev_num + IB_UMAD_MAX_PORTS, 1))
869 		goto err_sm_cdev;
870 
871 	port->sm_class_dev = class_device_create(umad_class, NULL, port->sm_dev->dev,
872 						 device->dma_device,
873 						 "issm%d", port->dev_num);
874 	if (IS_ERR(port->sm_class_dev))
875 		goto err_sm_cdev;
876 
877 	class_set_devdata(port->class_dev,    port);
878 	class_set_devdata(port->sm_class_dev, port);
879 
880 	if (class_device_create_file(port->sm_class_dev, &class_device_attr_ibdev))
881 		goto err_sm_class;
882 	if (class_device_create_file(port->sm_class_dev, &class_device_attr_port))
883 		goto err_sm_class;
884 
885 	spin_lock(&port_lock);
886 	umad_port[port->dev_num] = port;
887 	spin_unlock(&port_lock);
888 
889 	return 0;
890 
891 err_sm_class:
892 	class_device_destroy(umad_class, port->sm_dev->dev);
893 
894 err_sm_cdev:
895 	cdev_del(port->sm_dev);
896 
897 err_class:
898 	class_device_destroy(umad_class, port->dev->dev);
899 
900 err_cdev:
901 	cdev_del(port->dev);
902 	clear_bit(port->dev_num, dev_map);
903 
904 	return -1;
905 }
906 
907 static void ib_umad_kill_port(struct ib_umad_port *port)
908 {
909 	struct ib_umad_file *file;
910 	int id;
911 
912 	class_set_devdata(port->class_dev,    NULL);
913 	class_set_devdata(port->sm_class_dev, NULL);
914 
915 	class_device_destroy(umad_class, port->dev->dev);
916 	class_device_destroy(umad_class, port->sm_dev->dev);
917 
918 	cdev_del(port->dev);
919 	cdev_del(port->sm_dev);
920 
921 	spin_lock(&port_lock);
922 	umad_port[port->dev_num] = NULL;
923 	spin_unlock(&port_lock);
924 
925 	down_write(&port->mutex);
926 
927 	port->ib_dev = NULL;
928 
929 	/*
930 	 * Now go through the list of files attached to this port and
931 	 * unregister all of their MAD agents.  We need to hold
932 	 * port->mutex while doing this to avoid racing with
933 	 * ib_umad_close(), but we can't hold the mutex for writing
934 	 * while calling ib_unregister_mad_agent(), since that might
935 	 * deadlock by calling back into queue_packet().  So we
936 	 * downgrade our lock to a read lock, and then drop and
937 	 * reacquire the write lock for the next iteration.
938 	 *
939 	 * We do list_del_init() on the file's list_head so that the
940 	 * list_del in ib_umad_close() is still OK, even after the
941 	 * file is removed from the list.
942 	 */
943 	while (!list_empty(&port->file_list)) {
944 		file = list_entry(port->file_list.next, struct ib_umad_file,
945 				  port_list);
946 
947 		file->agents_dead = 1;
948 		list_del_init(&file->port_list);
949 
950 		downgrade_write(&port->mutex);
951 
952 		for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id)
953 			if (file->agent[id])
954 				ib_unregister_mad_agent(file->agent[id]);
955 
956 		up_read(&port->mutex);
957 		down_write(&port->mutex);
958 	}
959 
960 	up_write(&port->mutex);
961 
962 	clear_bit(port->dev_num, dev_map);
963 }
964 
965 static void ib_umad_add_one(struct ib_device *device)
966 {
967 	struct ib_umad_device *umad_dev;
968 	int s, e, i;
969 
970 	if (device->node_type == IB_NODE_SWITCH)
971 		s = e = 0;
972 	else {
973 		s = 1;
974 		e = device->phys_port_cnt;
975 	}
976 
977 	umad_dev = kzalloc(sizeof *umad_dev +
978 			   (e - s + 1) * sizeof (struct ib_umad_port),
979 			   GFP_KERNEL);
980 	if (!umad_dev)
981 		return;
982 
983 	kref_init(&umad_dev->ref);
984 
985 	umad_dev->start_port = s;
986 	umad_dev->end_port   = e;
987 
988 	for (i = s; i <= e; ++i) {
989 		umad_dev->port[i - s].umad_dev = umad_dev;
990 
991 		if (ib_umad_init_port(device, i, &umad_dev->port[i - s]))
992 			goto err;
993 	}
994 
995 	ib_set_client_data(device, &umad_client, umad_dev);
996 
997 	return;
998 
999 err:
1000 	while (--i >= s)
1001 		ib_umad_kill_port(&umad_dev->port[i - s]);
1002 
1003 	kref_put(&umad_dev->ref, ib_umad_release_dev);
1004 }
1005 
1006 static void ib_umad_remove_one(struct ib_device *device)
1007 {
1008 	struct ib_umad_device *umad_dev = ib_get_client_data(device, &umad_client);
1009 	int i;
1010 
1011 	if (!umad_dev)
1012 		return;
1013 
1014 	for (i = 0; i <= umad_dev->end_port - umad_dev->start_port; ++i)
1015 		ib_umad_kill_port(&umad_dev->port[i]);
1016 
1017 	kref_put(&umad_dev->ref, ib_umad_release_dev);
1018 }
1019 
1020 static int __init ib_umad_init(void)
1021 {
1022 	int ret;
1023 
1024 	ret = register_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2,
1025 				     "infiniband_mad");
1026 	if (ret) {
1027 		printk(KERN_ERR "user_mad: couldn't register device number\n");
1028 		goto out;
1029 	}
1030 
1031 	umad_class = class_create(THIS_MODULE, "infiniband_mad");
1032 	if (IS_ERR(umad_class)) {
1033 		ret = PTR_ERR(umad_class);
1034 		printk(KERN_ERR "user_mad: couldn't create class infiniband_mad\n");
1035 		goto out_chrdev;
1036 	}
1037 
1038 	ret = class_create_file(umad_class, &class_attr_abi_version);
1039 	if (ret) {
1040 		printk(KERN_ERR "user_mad: couldn't create abi_version attribute\n");
1041 		goto out_class;
1042 	}
1043 
1044 	ret = ib_register_client(&umad_client);
1045 	if (ret) {
1046 		printk(KERN_ERR "user_mad: couldn't register ib_umad client\n");
1047 		goto out_class;
1048 	}
1049 
1050 	return 0;
1051 
1052 out_class:
1053 	class_destroy(umad_class);
1054 
1055 out_chrdev:
1056 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
1057 
1058 out:
1059 	return ret;
1060 }
1061 
1062 static void __exit ib_umad_cleanup(void)
1063 {
1064 	ib_unregister_client(&umad_client);
1065 	class_destroy(umad_class);
1066 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
1067 }
1068 
1069 module_init(ib_umad_init);
1070 module_exit(ib_umad_cleanup);
1071