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 
36 #include <linux/module.h>
37 #include <linux/init.h>
38 #include <linux/device.h>
39 #include <linux/err.h>
40 #include <linux/fs.h>
41 #include <linux/cdev.h>
42 #include <linux/dma-mapping.h>
43 #include <linux/poll.h>
44 #include <linux/mutex.h>
45 #include <linux/kref.h>
46 #include <linux/compat.h>
47 #include <linux/semaphore.h>
48 
49 #include <asm/uaccess.h>
50 
51 #include <rdma/ib_mad.h>
52 #include <rdma/ib_user_mad.h>
53 
54 MODULE_AUTHOR("Roland Dreier");
55 MODULE_DESCRIPTION("InfiniBand userspace MAD packet access");
56 MODULE_LICENSE("Dual BSD/GPL");
57 
58 enum {
59 	IB_UMAD_MAX_PORTS  = 64,
60 	IB_UMAD_MAX_AGENTS = 32,
61 
62 	IB_UMAD_MAJOR      = 231,
63 	IB_UMAD_MINOR_BASE = 0
64 };
65 
66 /*
67  * Our lifetime rules for these structs are the following: each time a
68  * device special file is opened, we look up the corresponding struct
69  * ib_umad_port by minor in the umad_port[] table while holding the
70  * port_lock.  If this lookup succeeds, we take a reference on the
71  * ib_umad_port's struct ib_umad_device while still holding the
72  * port_lock; if the lookup fails, we fail the open().  We drop these
73  * references in the corresponding close().
74  *
75  * In addition to references coming from open character devices, there
76  * is one more reference to each ib_umad_device representing the
77  * module's reference taken when allocating the ib_umad_device in
78  * ib_umad_add_one().
79  *
80  * When destroying an ib_umad_device, we clear all of its
81  * ib_umad_ports from umad_port[] while holding port_lock before
82  * dropping the module's reference to the ib_umad_device.  This is
83  * always safe because any open() calls will either succeed and obtain
84  * a reference before we clear the umad_port[] entries, or fail after
85  * we clear the umad_port[] entries.
86  */
87 
88 struct ib_umad_port {
89 	struct cdev           *cdev;
90 	struct device	      *dev;
91 
92 	struct cdev           *sm_cdev;
93 	struct device	      *sm_dev;
94 	struct semaphore       sm_sem;
95 
96 	struct mutex	       file_mutex;
97 	struct list_head       file_list;
98 
99 	struct ib_device      *ib_dev;
100 	struct ib_umad_device *umad_dev;
101 	int                    dev_num;
102 	u8                     port_num;
103 };
104 
105 struct ib_umad_device {
106 	int                  start_port, end_port;
107 	struct kref          ref;
108 	struct ib_umad_port  port[0];
109 };
110 
111 struct ib_umad_file {
112 	struct mutex		mutex;
113 	struct ib_umad_port    *port;
114 	struct list_head	recv_list;
115 	struct list_head	send_list;
116 	struct list_head	port_list;
117 	spinlock_t		send_lock;
118 	wait_queue_head_t	recv_wait;
119 	struct ib_mad_agent    *agent[IB_UMAD_MAX_AGENTS];
120 	int			agents_dead;
121 	u8			use_pkey_index;
122 	u8			already_used;
123 };
124 
125 struct ib_umad_packet {
126 	struct ib_mad_send_buf *msg;
127 	struct ib_mad_recv_wc  *recv_wc;
128 	struct list_head   list;
129 	int		   length;
130 	struct ib_user_mad mad;
131 };
132 
133 static struct class *umad_class;
134 
135 static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
136 
137 static DEFINE_SPINLOCK(port_lock);
138 static struct ib_umad_port *umad_port[IB_UMAD_MAX_PORTS];
139 static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS);
140 
141 static void ib_umad_add_one(struct ib_device *device);
142 static void ib_umad_remove_one(struct ib_device *device);
143 
144 static void ib_umad_release_dev(struct kref *ref)
145 {
146 	struct ib_umad_device *dev =
147 		container_of(ref, struct ib_umad_device, ref);
148 
149 	kfree(dev);
150 }
151 
152 static int hdr_size(struct ib_umad_file *file)
153 {
154 	return file->use_pkey_index ? sizeof (struct ib_user_mad_hdr) :
155 		sizeof (struct ib_user_mad_hdr_old);
156 }
157 
158 /* caller must hold file->mutex */
159 static struct ib_mad_agent *__get_agent(struct ib_umad_file *file, int id)
160 {
161 	return file->agents_dead ? NULL : file->agent[id];
162 }
163 
164 static int queue_packet(struct ib_umad_file *file,
165 			struct ib_mad_agent *agent,
166 			struct ib_umad_packet *packet)
167 {
168 	int ret = 1;
169 
170 	mutex_lock(&file->mutex);
171 
172 	for (packet->mad.hdr.id = 0;
173 	     packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
174 	     packet->mad.hdr.id++)
175 		if (agent == __get_agent(file, packet->mad.hdr.id)) {
176 			list_add_tail(&packet->list, &file->recv_list);
177 			wake_up_interruptible(&file->recv_wait);
178 			ret = 0;
179 			break;
180 		}
181 
182 	mutex_unlock(&file->mutex);
183 
184 	return ret;
185 }
186 
187 static void dequeue_send(struct ib_umad_file *file,
188 			 struct ib_umad_packet *packet)
189 {
190 	spin_lock_irq(&file->send_lock);
191 	list_del(&packet->list);
192 	spin_unlock_irq(&file->send_lock);
193 }
194 
195 static void send_handler(struct ib_mad_agent *agent,
196 			 struct ib_mad_send_wc *send_wc)
197 {
198 	struct ib_umad_file *file = agent->context;
199 	struct ib_umad_packet *packet = send_wc->send_buf->context[0];
200 
201 	dequeue_send(file, packet);
202 	ib_destroy_ah(packet->msg->ah);
203 	ib_free_send_mad(packet->msg);
204 
205 	if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
206 		packet->length = IB_MGMT_MAD_HDR;
207 		packet->mad.hdr.status = ETIMEDOUT;
208 		if (!queue_packet(file, agent, packet))
209 			return;
210 	}
211 	kfree(packet);
212 }
213 
214 static void recv_handler(struct ib_mad_agent *agent,
215 			 struct ib_mad_recv_wc *mad_recv_wc)
216 {
217 	struct ib_umad_file *file = agent->context;
218 	struct ib_umad_packet *packet;
219 
220 	if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
221 		goto err1;
222 
223 	packet = kzalloc(sizeof *packet, GFP_KERNEL);
224 	if (!packet)
225 		goto err1;
226 
227 	packet->length = mad_recv_wc->mad_len;
228 	packet->recv_wc = mad_recv_wc;
229 
230 	packet->mad.hdr.status	   = 0;
231 	packet->mad.hdr.length	   = hdr_size(file) + mad_recv_wc->mad_len;
232 	packet->mad.hdr.qpn	   = cpu_to_be32(mad_recv_wc->wc->src_qp);
233 	packet->mad.hdr.lid	   = cpu_to_be16(mad_recv_wc->wc->slid);
234 	packet->mad.hdr.sl	   = mad_recv_wc->wc->sl;
235 	packet->mad.hdr.path_bits  = mad_recv_wc->wc->dlid_path_bits;
236 	packet->mad.hdr.pkey_index = mad_recv_wc->wc->pkey_index;
237 	packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
238 	if (packet->mad.hdr.grh_present) {
239 		struct ib_ah_attr ah_attr;
240 
241 		ib_init_ah_from_wc(agent->device, agent->port_num,
242 				   mad_recv_wc->wc, mad_recv_wc->recv_buf.grh,
243 				   &ah_attr);
244 
245 		packet->mad.hdr.gid_index = ah_attr.grh.sgid_index;
246 		packet->mad.hdr.hop_limit = ah_attr.grh.hop_limit;
247 		packet->mad.hdr.traffic_class = ah_attr.grh.traffic_class;
248 		memcpy(packet->mad.hdr.gid, &ah_attr.grh.dgid, 16);
249 		packet->mad.hdr.flow_label = cpu_to_be32(ah_attr.grh.flow_label);
250 	}
251 
252 	if (queue_packet(file, agent, packet))
253 		goto err2;
254 	return;
255 
256 err2:
257 	kfree(packet);
258 err1:
259 	ib_free_recv_mad(mad_recv_wc);
260 }
261 
262 static ssize_t copy_recv_mad(struct ib_umad_file *file, char __user *buf,
263 			     struct ib_umad_packet *packet, size_t count)
264 {
265 	struct ib_mad_recv_buf *recv_buf;
266 	int left, seg_payload, offset, max_seg_payload;
267 
268 	/* We need enough room to copy the first (or only) MAD segment. */
269 	recv_buf = &packet->recv_wc->recv_buf;
270 	if ((packet->length <= sizeof (*recv_buf->mad) &&
271 	     count < hdr_size(file) + packet->length) ||
272 	    (packet->length > sizeof (*recv_buf->mad) &&
273 	     count < hdr_size(file) + sizeof (*recv_buf->mad)))
274 		return -EINVAL;
275 
276 	if (copy_to_user(buf, &packet->mad, hdr_size(file)))
277 		return -EFAULT;
278 
279 	buf += hdr_size(file);
280 	seg_payload = min_t(int, packet->length, sizeof (*recv_buf->mad));
281 	if (copy_to_user(buf, recv_buf->mad, seg_payload))
282 		return -EFAULT;
283 
284 	if (seg_payload < packet->length) {
285 		/*
286 		 * Multipacket RMPP MAD message. Copy remainder of message.
287 		 * Note that last segment may have a shorter payload.
288 		 */
289 		if (count < hdr_size(file) + packet->length) {
290 			/*
291 			 * The buffer is too small, return the first RMPP segment,
292 			 * which includes the RMPP message length.
293 			 */
294 			return -ENOSPC;
295 		}
296 		offset = ib_get_mad_data_offset(recv_buf->mad->mad_hdr.mgmt_class);
297 		max_seg_payload = sizeof (struct ib_mad) - offset;
298 
299 		for (left = packet->length - seg_payload, buf += seg_payload;
300 		     left; left -= seg_payload, buf += seg_payload) {
301 			recv_buf = container_of(recv_buf->list.next,
302 						struct ib_mad_recv_buf, list);
303 			seg_payload = min(left, max_seg_payload);
304 			if (copy_to_user(buf, ((void *) recv_buf->mad) + offset,
305 					 seg_payload))
306 				return -EFAULT;
307 		}
308 	}
309 	return hdr_size(file) + packet->length;
310 }
311 
312 static ssize_t copy_send_mad(struct ib_umad_file *file, char __user *buf,
313 			     struct ib_umad_packet *packet, size_t count)
314 {
315 	ssize_t size = hdr_size(file) + packet->length;
316 
317 	if (count < size)
318 		return -EINVAL;
319 
320 	if (copy_to_user(buf, &packet->mad, hdr_size(file)))
321 		return -EFAULT;
322 
323 	buf += hdr_size(file);
324 
325 	if (copy_to_user(buf, packet->mad.data, packet->length))
326 		return -EFAULT;
327 
328 	return size;
329 }
330 
331 static ssize_t ib_umad_read(struct file *filp, char __user *buf,
332 			    size_t count, loff_t *pos)
333 {
334 	struct ib_umad_file *file = filp->private_data;
335 	struct ib_umad_packet *packet;
336 	ssize_t ret;
337 
338 	if (count < hdr_size(file))
339 		return -EINVAL;
340 
341 	mutex_lock(&file->mutex);
342 
343 	while (list_empty(&file->recv_list)) {
344 		mutex_unlock(&file->mutex);
345 
346 		if (filp->f_flags & O_NONBLOCK)
347 			return -EAGAIN;
348 
349 		if (wait_event_interruptible(file->recv_wait,
350 					     !list_empty(&file->recv_list)))
351 			return -ERESTARTSYS;
352 
353 		mutex_lock(&file->mutex);
354 	}
355 
356 	packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
357 	list_del(&packet->list);
358 
359 	mutex_unlock(&file->mutex);
360 
361 	if (packet->recv_wc)
362 		ret = copy_recv_mad(file, buf, packet, count);
363 	else
364 		ret = copy_send_mad(file, buf, packet, count);
365 
366 	if (ret < 0) {
367 		/* Requeue packet */
368 		mutex_lock(&file->mutex);
369 		list_add(&packet->list, &file->recv_list);
370 		mutex_unlock(&file->mutex);
371 	} else {
372 		if (packet->recv_wc)
373 			ib_free_recv_mad(packet->recv_wc);
374 		kfree(packet);
375 	}
376 	return ret;
377 }
378 
379 static int copy_rmpp_mad(struct ib_mad_send_buf *msg, const char __user *buf)
380 {
381 	int left, seg;
382 
383 	/* Copy class specific header */
384 	if ((msg->hdr_len > IB_MGMT_RMPP_HDR) &&
385 	    copy_from_user(msg->mad + IB_MGMT_RMPP_HDR, buf + IB_MGMT_RMPP_HDR,
386 			   msg->hdr_len - IB_MGMT_RMPP_HDR))
387 		return -EFAULT;
388 
389 	/* All headers are in place.  Copy data segments. */
390 	for (seg = 1, left = msg->data_len, buf += msg->hdr_len; left > 0;
391 	     seg++, left -= msg->seg_size, buf += msg->seg_size) {
392 		if (copy_from_user(ib_get_rmpp_segment(msg, seg), buf,
393 				   min(left, msg->seg_size)))
394 			return -EFAULT;
395 	}
396 	return 0;
397 }
398 
399 static int same_destination(struct ib_user_mad_hdr *hdr1,
400 			    struct ib_user_mad_hdr *hdr2)
401 {
402 	if (!hdr1->grh_present && !hdr2->grh_present)
403 	   return (hdr1->lid == hdr2->lid);
404 
405 	if (hdr1->grh_present && hdr2->grh_present)
406 	   return !memcmp(hdr1->gid, hdr2->gid, 16);
407 
408 	return 0;
409 }
410 
411 static int is_duplicate(struct ib_umad_file *file,
412 			struct ib_umad_packet *packet)
413 {
414 	struct ib_umad_packet *sent_packet;
415 	struct ib_mad_hdr *sent_hdr, *hdr;
416 
417 	hdr = (struct ib_mad_hdr *) packet->mad.data;
418 	list_for_each_entry(sent_packet, &file->send_list, list) {
419 		sent_hdr = (struct ib_mad_hdr *) sent_packet->mad.data;
420 
421 		if ((hdr->tid != sent_hdr->tid) ||
422 		    (hdr->mgmt_class != sent_hdr->mgmt_class))
423 			continue;
424 
425 		/*
426 		 * No need to be overly clever here.  If two new operations have
427 		 * the same TID, reject the second as a duplicate.  This is more
428 		 * restrictive than required by the spec.
429 		 */
430 		if (!ib_response_mad((struct ib_mad *) hdr)) {
431 			if (!ib_response_mad((struct ib_mad *) sent_hdr))
432 				return 1;
433 			continue;
434 		} else if (!ib_response_mad((struct ib_mad *) sent_hdr))
435 			continue;
436 
437 		if (same_destination(&packet->mad.hdr, &sent_packet->mad.hdr))
438 			return 1;
439 	}
440 
441 	return 0;
442 }
443 
444 static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
445 			     size_t count, loff_t *pos)
446 {
447 	struct ib_umad_file *file = filp->private_data;
448 	struct ib_umad_packet *packet;
449 	struct ib_mad_agent *agent;
450 	struct ib_ah_attr ah_attr;
451 	struct ib_ah *ah;
452 	struct ib_rmpp_mad *rmpp_mad;
453 	__be64 *tid;
454 	int ret, data_len, hdr_len, copy_offset, rmpp_active;
455 
456 	if (count < hdr_size(file) + IB_MGMT_RMPP_HDR)
457 		return -EINVAL;
458 
459 	packet = kzalloc(sizeof *packet + IB_MGMT_RMPP_HDR, GFP_KERNEL);
460 	if (!packet)
461 		return -ENOMEM;
462 
463 	if (copy_from_user(&packet->mad, buf, hdr_size(file))) {
464 		ret = -EFAULT;
465 		goto err;
466 	}
467 
468 	if (packet->mad.hdr.id < 0 ||
469 	    packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
470 		ret = -EINVAL;
471 		goto err;
472 	}
473 
474 	buf += hdr_size(file);
475 
476 	if (copy_from_user(packet->mad.data, buf, IB_MGMT_RMPP_HDR)) {
477 		ret = -EFAULT;
478 		goto err;
479 	}
480 
481 	mutex_lock(&file->mutex);
482 
483 	agent = __get_agent(file, packet->mad.hdr.id);
484 	if (!agent) {
485 		ret = -EINVAL;
486 		goto err_up;
487 	}
488 
489 	memset(&ah_attr, 0, sizeof ah_attr);
490 	ah_attr.dlid          = be16_to_cpu(packet->mad.hdr.lid);
491 	ah_attr.sl            = packet->mad.hdr.sl;
492 	ah_attr.src_path_bits = packet->mad.hdr.path_bits;
493 	ah_attr.port_num      = file->port->port_num;
494 	if (packet->mad.hdr.grh_present) {
495 		ah_attr.ah_flags = IB_AH_GRH;
496 		memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
497 		ah_attr.grh.sgid_index	   = packet->mad.hdr.gid_index;
498 		ah_attr.grh.flow_label 	   = be32_to_cpu(packet->mad.hdr.flow_label);
499 		ah_attr.grh.hop_limit  	   = packet->mad.hdr.hop_limit;
500 		ah_attr.grh.traffic_class  = packet->mad.hdr.traffic_class;
501 	}
502 
503 	ah = ib_create_ah(agent->qp->pd, &ah_attr);
504 	if (IS_ERR(ah)) {
505 		ret = PTR_ERR(ah);
506 		goto err_up;
507 	}
508 
509 	rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
510 	hdr_len = ib_get_mad_data_offset(rmpp_mad->mad_hdr.mgmt_class);
511 	if (!ib_is_mad_class_rmpp(rmpp_mad->mad_hdr.mgmt_class)) {
512 		copy_offset = IB_MGMT_MAD_HDR;
513 		rmpp_active = 0;
514 	} else {
515 		copy_offset = IB_MGMT_RMPP_HDR;
516 		rmpp_active = ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
517 			      IB_MGMT_RMPP_FLAG_ACTIVE;
518 	}
519 
520 	data_len = count - hdr_size(file) - hdr_len;
521 	packet->msg = ib_create_send_mad(agent,
522 					 be32_to_cpu(packet->mad.hdr.qpn),
523 					 packet->mad.hdr.pkey_index, rmpp_active,
524 					 hdr_len, data_len, GFP_KERNEL);
525 	if (IS_ERR(packet->msg)) {
526 		ret = PTR_ERR(packet->msg);
527 		goto err_ah;
528 	}
529 
530 	packet->msg->ah 	= ah;
531 	packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
532 	packet->msg->retries 	= packet->mad.hdr.retries;
533 	packet->msg->context[0] = packet;
534 
535 	/* Copy MAD header.  Any RMPP header is already in place. */
536 	memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
537 
538 	if (!rmpp_active) {
539 		if (copy_from_user(packet->msg->mad + copy_offset,
540 				   buf + copy_offset,
541 				   hdr_len + data_len - copy_offset)) {
542 			ret = -EFAULT;
543 			goto err_msg;
544 		}
545 	} else {
546 		ret = copy_rmpp_mad(packet->msg, buf);
547 		if (ret)
548 			goto err_msg;
549 	}
550 
551 	/*
552 	 * Set the high-order part of the transaction ID to make MADs from
553 	 * different agents unique, and allow routing responses back to the
554 	 * original requestor.
555 	 */
556 	if (!ib_response_mad(packet->msg->mad)) {
557 		tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
558 		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
559 				   (be64_to_cpup(tid) & 0xffffffff));
560 		rmpp_mad->mad_hdr.tid = *tid;
561 	}
562 
563 	spin_lock_irq(&file->send_lock);
564 	ret = is_duplicate(file, packet);
565 	if (!ret)
566 		list_add_tail(&packet->list, &file->send_list);
567 	spin_unlock_irq(&file->send_lock);
568 	if (ret) {
569 		ret = -EINVAL;
570 		goto err_msg;
571 	}
572 
573 	ret = ib_post_send_mad(packet->msg, NULL);
574 	if (ret)
575 		goto err_send;
576 
577 	mutex_unlock(&file->mutex);
578 	return count;
579 
580 err_send:
581 	dequeue_send(file, packet);
582 err_msg:
583 	ib_free_send_mad(packet->msg);
584 err_ah:
585 	ib_destroy_ah(ah);
586 err_up:
587 	mutex_unlock(&file->mutex);
588 err:
589 	kfree(packet);
590 	return ret;
591 }
592 
593 static unsigned int ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
594 {
595 	struct ib_umad_file *file = filp->private_data;
596 
597 	/* we will always be able to post a MAD send */
598 	unsigned int mask = POLLOUT | POLLWRNORM;
599 
600 	poll_wait(filp, &file->recv_wait, wait);
601 
602 	if (!list_empty(&file->recv_list))
603 		mask |= POLLIN | POLLRDNORM;
604 
605 	return mask;
606 }
607 
608 static int ib_umad_reg_agent(struct ib_umad_file *file, void __user *arg,
609 			     int compat_method_mask)
610 {
611 	struct ib_user_mad_reg_req ureq;
612 	struct ib_mad_reg_req req;
613 	struct ib_mad_agent *agent = NULL;
614 	int agent_id;
615 	int ret;
616 
617 	mutex_lock(&file->port->file_mutex);
618 	mutex_lock(&file->mutex);
619 
620 	if (!file->port->ib_dev) {
621 		ret = -EPIPE;
622 		goto out;
623 	}
624 
625 	if (copy_from_user(&ureq, arg, sizeof ureq)) {
626 		ret = -EFAULT;
627 		goto out;
628 	}
629 
630 	if (ureq.qpn != 0 && ureq.qpn != 1) {
631 		ret = -EINVAL;
632 		goto out;
633 	}
634 
635 	for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
636 		if (!__get_agent(file, agent_id))
637 			goto found;
638 
639 	ret = -ENOMEM;
640 	goto out;
641 
642 found:
643 	if (ureq.mgmt_class) {
644 		req.mgmt_class         = ureq.mgmt_class;
645 		req.mgmt_class_version = ureq.mgmt_class_version;
646 		memcpy(req.oui, ureq.oui, sizeof req.oui);
647 
648 		if (compat_method_mask) {
649 			u32 *umm = (u32 *) ureq.method_mask;
650 			int i;
651 
652 			for (i = 0; i < BITS_TO_LONGS(IB_MGMT_MAX_METHODS); ++i)
653 				req.method_mask[i] =
654 					umm[i * 2] | ((u64) umm[i * 2 + 1] << 32);
655 		} else
656 			memcpy(req.method_mask, ureq.method_mask,
657 			       sizeof req.method_mask);
658 	}
659 
660 	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
661 				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
662 				      ureq.mgmt_class ? &req : NULL,
663 				      ureq.rmpp_version,
664 				      send_handler, recv_handler, file);
665 	if (IS_ERR(agent)) {
666 		ret = PTR_ERR(agent);
667 		agent = NULL;
668 		goto out;
669 	}
670 
671 	if (put_user(agent_id,
672 		     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
673 		ret = -EFAULT;
674 		goto out;
675 	}
676 
677 	if (!file->already_used) {
678 		file->already_used = 1;
679 		if (!file->use_pkey_index) {
680 			printk(KERN_WARNING "user_mad: process %s did not enable "
681 			       "P_Key index support.\n", current->comm);
682 			printk(KERN_WARNING "user_mad:   Documentation/infiniband/user_mad.txt "
683 			       "has info on the new ABI.\n");
684 		}
685 	}
686 
687 	file->agent[agent_id] = agent;
688 	ret = 0;
689 
690 out:
691 	mutex_unlock(&file->mutex);
692 
693 	if (ret && agent)
694 		ib_unregister_mad_agent(agent);
695 
696 	mutex_unlock(&file->port->file_mutex);
697 
698 	return ret;
699 }
700 
701 static int ib_umad_unreg_agent(struct ib_umad_file *file, u32 __user *arg)
702 {
703 	struct ib_mad_agent *agent = NULL;
704 	u32 id;
705 	int ret = 0;
706 
707 	if (get_user(id, arg))
708 		return -EFAULT;
709 
710 	mutex_lock(&file->port->file_mutex);
711 	mutex_lock(&file->mutex);
712 
713 	if (id < 0 || id >= IB_UMAD_MAX_AGENTS || !__get_agent(file, id)) {
714 		ret = -EINVAL;
715 		goto out;
716 	}
717 
718 	agent = file->agent[id];
719 	file->agent[id] = NULL;
720 
721 out:
722 	mutex_unlock(&file->mutex);
723 
724 	if (agent)
725 		ib_unregister_mad_agent(agent);
726 
727 	mutex_unlock(&file->port->file_mutex);
728 
729 	return ret;
730 }
731 
732 static long ib_umad_enable_pkey(struct ib_umad_file *file)
733 {
734 	int ret = 0;
735 
736 	mutex_lock(&file->mutex);
737 	if (file->already_used)
738 		ret = -EINVAL;
739 	else
740 		file->use_pkey_index = 1;
741 	mutex_unlock(&file->mutex);
742 
743 	return ret;
744 }
745 
746 static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
747 			  unsigned long arg)
748 {
749 	switch (cmd) {
750 	case IB_USER_MAD_REGISTER_AGENT:
751 		return ib_umad_reg_agent(filp->private_data, (void __user *) arg, 0);
752 	case IB_USER_MAD_UNREGISTER_AGENT:
753 		return ib_umad_unreg_agent(filp->private_data, (__u32 __user *) arg);
754 	case IB_USER_MAD_ENABLE_PKEY:
755 		return ib_umad_enable_pkey(filp->private_data);
756 	default:
757 		return -ENOIOCTLCMD;
758 	}
759 }
760 
761 #ifdef CONFIG_COMPAT
762 static long ib_umad_compat_ioctl(struct file *filp, unsigned int cmd,
763 				 unsigned long arg)
764 {
765 	switch (cmd) {
766 	case IB_USER_MAD_REGISTER_AGENT:
767 		return ib_umad_reg_agent(filp->private_data, compat_ptr(arg), 1);
768 	case IB_USER_MAD_UNREGISTER_AGENT:
769 		return ib_umad_unreg_agent(filp->private_data, compat_ptr(arg));
770 	case IB_USER_MAD_ENABLE_PKEY:
771 		return ib_umad_enable_pkey(filp->private_data);
772 	default:
773 		return -ENOIOCTLCMD;
774 	}
775 }
776 #endif
777 
778 /*
779  * ib_umad_open() does not need the BKL:
780  *
781  *  - umad_port[] accesses are protected by port_lock, the
782  *    ib_umad_port structures are properly reference counted, and
783  *    everything else is purely local to the file being created, so
784  *    races against other open calls are not a problem;
785  *  - the ioctl method does not affect any global state outside of the
786  *    file structure being operated on;
787  *  - the port is added to umad_port[] as the last part of module
788  *    initialization so the open method will either immediately run
789  *    -ENXIO, or all required initialization will be done.
790  */
791 static int ib_umad_open(struct inode *inode, struct file *filp)
792 {
793 	struct ib_umad_port *port;
794 	struct ib_umad_file *file;
795 	int ret = 0;
796 
797 	spin_lock(&port_lock);
798 	port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE];
799 	if (port)
800 		kref_get(&port->umad_dev->ref);
801 	spin_unlock(&port_lock);
802 
803 	if (!port)
804 		return -ENXIO;
805 
806 	mutex_lock(&port->file_mutex);
807 
808 	if (!port->ib_dev) {
809 		ret = -ENXIO;
810 		goto out;
811 	}
812 
813 	file = kzalloc(sizeof *file, GFP_KERNEL);
814 	if (!file) {
815 		kref_put(&port->umad_dev->ref, ib_umad_release_dev);
816 		ret = -ENOMEM;
817 		goto out;
818 	}
819 
820 	mutex_init(&file->mutex);
821 	spin_lock_init(&file->send_lock);
822 	INIT_LIST_HEAD(&file->recv_list);
823 	INIT_LIST_HEAD(&file->send_list);
824 	init_waitqueue_head(&file->recv_wait);
825 
826 	file->port = port;
827 	filp->private_data = file;
828 
829 	list_add_tail(&file->port_list, &port->file_list);
830 
831 out:
832 	mutex_unlock(&port->file_mutex);
833 	return ret;
834 }
835 
836 static int ib_umad_close(struct inode *inode, struct file *filp)
837 {
838 	struct ib_umad_file *file = filp->private_data;
839 	struct ib_umad_device *dev = file->port->umad_dev;
840 	struct ib_umad_packet *packet, *tmp;
841 	int already_dead;
842 	int i;
843 
844 	mutex_lock(&file->port->file_mutex);
845 	mutex_lock(&file->mutex);
846 
847 	already_dead = file->agents_dead;
848 	file->agents_dead = 1;
849 
850 	list_for_each_entry_safe(packet, tmp, &file->recv_list, list) {
851 		if (packet->recv_wc)
852 			ib_free_recv_mad(packet->recv_wc);
853 		kfree(packet);
854 	}
855 
856 	list_del(&file->port_list);
857 
858 	mutex_unlock(&file->mutex);
859 
860 	if (!already_dead)
861 		for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
862 			if (file->agent[i])
863 				ib_unregister_mad_agent(file->agent[i]);
864 
865 	mutex_unlock(&file->port->file_mutex);
866 
867 	kfree(file);
868 	kref_put(&dev->ref, ib_umad_release_dev);
869 
870 	return 0;
871 }
872 
873 static const struct file_operations umad_fops = {
874 	.owner 	 	= THIS_MODULE,
875 	.read 	 	= ib_umad_read,
876 	.write 	 	= ib_umad_write,
877 	.poll 	 	= ib_umad_poll,
878 	.unlocked_ioctl = ib_umad_ioctl,
879 #ifdef CONFIG_COMPAT
880 	.compat_ioctl 	= ib_umad_compat_ioctl,
881 #endif
882 	.open 	 	= ib_umad_open,
883 	.release 	= ib_umad_close
884 };
885 
886 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
887 {
888 	struct ib_umad_port *port;
889 	struct ib_port_modify props = {
890 		.set_port_cap_mask = IB_PORT_SM
891 	};
892 	int ret;
893 
894 	spin_lock(&port_lock);
895 	port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE - IB_UMAD_MAX_PORTS];
896 	if (port)
897 		kref_get(&port->umad_dev->ref);
898 	spin_unlock(&port_lock);
899 
900 	if (!port)
901 		return -ENXIO;
902 
903 	if (filp->f_flags & O_NONBLOCK) {
904 		if (down_trylock(&port->sm_sem)) {
905 			ret = -EAGAIN;
906 			goto fail;
907 		}
908 	} else {
909 		if (down_interruptible(&port->sm_sem)) {
910 			ret = -ERESTARTSYS;
911 			goto fail;
912 		}
913 	}
914 
915 	ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
916 	if (ret) {
917 		up(&port->sm_sem);
918 		goto fail;
919 	}
920 
921 	filp->private_data = port;
922 
923 	return 0;
924 
925 fail:
926 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
927 	return ret;
928 }
929 
930 static int ib_umad_sm_close(struct inode *inode, struct file *filp)
931 {
932 	struct ib_umad_port *port = filp->private_data;
933 	struct ib_port_modify props = {
934 		.clr_port_cap_mask = IB_PORT_SM
935 	};
936 	int ret = 0;
937 
938 	mutex_lock(&port->file_mutex);
939 	if (port->ib_dev)
940 		ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
941 	mutex_unlock(&port->file_mutex);
942 
943 	up(&port->sm_sem);
944 
945 	kref_put(&port->umad_dev->ref, ib_umad_release_dev);
946 
947 	return ret;
948 }
949 
950 static const struct file_operations umad_sm_fops = {
951 	.owner 	 = THIS_MODULE,
952 	.open 	 = ib_umad_sm_open,
953 	.release = ib_umad_sm_close
954 };
955 
956 static struct ib_client umad_client = {
957 	.name   = "umad",
958 	.add    = ib_umad_add_one,
959 	.remove = ib_umad_remove_one
960 };
961 
962 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
963 			  char *buf)
964 {
965 	struct ib_umad_port *port = dev_get_drvdata(dev);
966 
967 	if (!port)
968 		return -ENODEV;
969 
970 	return sprintf(buf, "%s\n", port->ib_dev->name);
971 }
972 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
973 
974 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
975 			 char *buf)
976 {
977 	struct ib_umad_port *port = dev_get_drvdata(dev);
978 
979 	if (!port)
980 		return -ENODEV;
981 
982 	return sprintf(buf, "%d\n", port->port_num);
983 }
984 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
985 
986 static ssize_t show_abi_version(struct class *class, char *buf)
987 {
988 	return sprintf(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
989 }
990 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
991 
992 static int ib_umad_init_port(struct ib_device *device, int port_num,
993 			     struct ib_umad_port *port)
994 {
995 	spin_lock(&port_lock);
996 	port->dev_num = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
997 	if (port->dev_num >= IB_UMAD_MAX_PORTS) {
998 		spin_unlock(&port_lock);
999 		return -1;
1000 	}
1001 	set_bit(port->dev_num, dev_map);
1002 	spin_unlock(&port_lock);
1003 
1004 	port->ib_dev   = device;
1005 	port->port_num = port_num;
1006 	init_MUTEX(&port->sm_sem);
1007 	mutex_init(&port->file_mutex);
1008 	INIT_LIST_HEAD(&port->file_list);
1009 
1010 	port->cdev = cdev_alloc();
1011 	if (!port->cdev)
1012 		return -1;
1013 	port->cdev->owner = THIS_MODULE;
1014 	port->cdev->ops   = &umad_fops;
1015 	kobject_set_name(&port->cdev->kobj, "umad%d", port->dev_num);
1016 	if (cdev_add(port->cdev, base_dev + port->dev_num, 1))
1017 		goto err_cdev;
1018 
1019 	port->dev = device_create(umad_class, device->dma_device,
1020 				  port->cdev->dev, port,
1021 				  "umad%d", port->dev_num);
1022 	if (IS_ERR(port->dev))
1023 		goto err_cdev;
1024 
1025 	if (device_create_file(port->dev, &dev_attr_ibdev))
1026 		goto err_dev;
1027 	if (device_create_file(port->dev, &dev_attr_port))
1028 		goto err_dev;
1029 
1030 	port->sm_cdev = cdev_alloc();
1031 	if (!port->sm_cdev)
1032 		goto err_dev;
1033 	port->sm_cdev->owner = THIS_MODULE;
1034 	port->sm_cdev->ops   = &umad_sm_fops;
1035 	kobject_set_name(&port->sm_cdev->kobj, "issm%d", port->dev_num);
1036 	if (cdev_add(port->sm_cdev, base_dev + port->dev_num + IB_UMAD_MAX_PORTS, 1))
1037 		goto err_sm_cdev;
1038 
1039 	port->sm_dev = device_create(umad_class, device->dma_device,
1040 				     port->sm_cdev->dev, port,
1041 				     "issm%d", port->dev_num);
1042 	if (IS_ERR(port->sm_dev))
1043 		goto err_sm_cdev;
1044 
1045 	if (device_create_file(port->sm_dev, &dev_attr_ibdev))
1046 		goto err_sm_dev;
1047 	if (device_create_file(port->sm_dev, &dev_attr_port))
1048 		goto err_sm_dev;
1049 
1050 	spin_lock(&port_lock);
1051 	umad_port[port->dev_num] = port;
1052 	spin_unlock(&port_lock);
1053 
1054 	return 0;
1055 
1056 err_sm_dev:
1057 	device_destroy(umad_class, port->sm_cdev->dev);
1058 
1059 err_sm_cdev:
1060 	cdev_del(port->sm_cdev);
1061 
1062 err_dev:
1063 	device_destroy(umad_class, port->cdev->dev);
1064 
1065 err_cdev:
1066 	cdev_del(port->cdev);
1067 	clear_bit(port->dev_num, dev_map);
1068 
1069 	return -1;
1070 }
1071 
1072 static void ib_umad_kill_port(struct ib_umad_port *port)
1073 {
1074 	struct ib_umad_file *file;
1075 	int already_dead;
1076 	int id;
1077 
1078 	dev_set_drvdata(port->dev,    NULL);
1079 	dev_set_drvdata(port->sm_dev, NULL);
1080 
1081 	device_destroy(umad_class, port->cdev->dev);
1082 	device_destroy(umad_class, port->sm_cdev->dev);
1083 
1084 	cdev_del(port->cdev);
1085 	cdev_del(port->sm_cdev);
1086 
1087 	spin_lock(&port_lock);
1088 	umad_port[port->dev_num] = NULL;
1089 	spin_unlock(&port_lock);
1090 
1091 	mutex_lock(&port->file_mutex);
1092 
1093 	port->ib_dev = NULL;
1094 
1095 	list_for_each_entry(file, &port->file_list, port_list) {
1096 		mutex_lock(&file->mutex);
1097 		already_dead = file->agents_dead;
1098 		file->agents_dead = 1;
1099 		mutex_unlock(&file->mutex);
1100 
1101 		for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id)
1102 			if (file->agent[id])
1103 				ib_unregister_mad_agent(file->agent[id]);
1104 	}
1105 
1106 	mutex_unlock(&port->file_mutex);
1107 
1108 	clear_bit(port->dev_num, dev_map);
1109 }
1110 
1111 static void ib_umad_add_one(struct ib_device *device)
1112 {
1113 	struct ib_umad_device *umad_dev;
1114 	int s, e, i;
1115 
1116 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1117 		return;
1118 
1119 	if (device->node_type == RDMA_NODE_IB_SWITCH)
1120 		s = e = 0;
1121 	else {
1122 		s = 1;
1123 		e = device->phys_port_cnt;
1124 	}
1125 
1126 	umad_dev = kzalloc(sizeof *umad_dev +
1127 			   (e - s + 1) * sizeof (struct ib_umad_port),
1128 			   GFP_KERNEL);
1129 	if (!umad_dev)
1130 		return;
1131 
1132 	kref_init(&umad_dev->ref);
1133 
1134 	umad_dev->start_port = s;
1135 	umad_dev->end_port   = e;
1136 
1137 	for (i = s; i <= e; ++i) {
1138 		umad_dev->port[i - s].umad_dev = umad_dev;
1139 
1140 		if (ib_umad_init_port(device, i, &umad_dev->port[i - s]))
1141 			goto err;
1142 	}
1143 
1144 	ib_set_client_data(device, &umad_client, umad_dev);
1145 
1146 	return;
1147 
1148 err:
1149 	while (--i >= s)
1150 		ib_umad_kill_port(&umad_dev->port[i - s]);
1151 
1152 	kref_put(&umad_dev->ref, ib_umad_release_dev);
1153 }
1154 
1155 static void ib_umad_remove_one(struct ib_device *device)
1156 {
1157 	struct ib_umad_device *umad_dev = ib_get_client_data(device, &umad_client);
1158 	int i;
1159 
1160 	if (!umad_dev)
1161 		return;
1162 
1163 	for (i = 0; i <= umad_dev->end_port - umad_dev->start_port; ++i)
1164 		ib_umad_kill_port(&umad_dev->port[i]);
1165 
1166 	kref_put(&umad_dev->ref, ib_umad_release_dev);
1167 }
1168 
1169 static int __init ib_umad_init(void)
1170 {
1171 	int ret;
1172 
1173 	ret = register_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2,
1174 				     "infiniband_mad");
1175 	if (ret) {
1176 		printk(KERN_ERR "user_mad: couldn't register device number\n");
1177 		goto out;
1178 	}
1179 
1180 	umad_class = class_create(THIS_MODULE, "infiniband_mad");
1181 	if (IS_ERR(umad_class)) {
1182 		ret = PTR_ERR(umad_class);
1183 		printk(KERN_ERR "user_mad: couldn't create class infiniband_mad\n");
1184 		goto out_chrdev;
1185 	}
1186 
1187 	ret = class_create_file(umad_class, &class_attr_abi_version);
1188 	if (ret) {
1189 		printk(KERN_ERR "user_mad: couldn't create abi_version attribute\n");
1190 		goto out_class;
1191 	}
1192 
1193 	ret = ib_register_client(&umad_client);
1194 	if (ret) {
1195 		printk(KERN_ERR "user_mad: couldn't register ib_umad client\n");
1196 		goto out_class;
1197 	}
1198 
1199 	return 0;
1200 
1201 out_class:
1202 	class_destroy(umad_class);
1203 
1204 out_chrdev:
1205 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
1206 
1207 out:
1208 	return ret;
1209 }
1210 
1211 static void __exit ib_umad_cleanup(void)
1212 {
1213 	ib_unregister_client(&umad_client);
1214 	class_destroy(umad_class);
1215 	unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
1216 }
1217 
1218 module_init(ib_umad_init);
1219 module_exit(ib_umad_cleanup);
1220