xref: /linux/drivers/infiniband/core/mad.c (revision 7b12b9137930eb821b68e1bfa11e9de692208620)
1 /*
2  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
3  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies Ltd.  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: mad.c 5596 2006-03-03 01:00:07Z sean.hefty $
35  */
36 #include <linux/dma-mapping.h>
37 
38 #include "mad_priv.h"
39 #include "mad_rmpp.h"
40 #include "smi.h"
41 #include "agent.h"
42 
43 MODULE_LICENSE("Dual BSD/GPL");
44 MODULE_DESCRIPTION("kernel IB MAD API");
45 MODULE_AUTHOR("Hal Rosenstock");
46 MODULE_AUTHOR("Sean Hefty");
47 
48 
49 kmem_cache_t *ib_mad_cache;
50 
51 static struct list_head ib_mad_port_list;
52 static u32 ib_mad_client_id = 0;
53 
54 /* Port list lock */
55 static spinlock_t ib_mad_port_list_lock;
56 
57 
58 /* Forward declarations */
59 static int method_in_use(struct ib_mad_mgmt_method_table **method,
60 			 struct ib_mad_reg_req *mad_reg_req);
61 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
62 static struct ib_mad_agent_private *find_mad_agent(
63 					struct ib_mad_port_private *port_priv,
64 					struct ib_mad *mad);
65 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
66 				    struct ib_mad_private *mad);
67 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
68 static void timeout_sends(void *data);
69 static void local_completions(void *data);
70 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
71 			      struct ib_mad_agent_private *agent_priv,
72 			      u8 mgmt_class);
73 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
74 			   struct ib_mad_agent_private *agent_priv);
75 
76 /*
77  * Returns a ib_mad_port_private structure or NULL for a device/port
78  * Assumes ib_mad_port_list_lock is being held
79  */
80 static inline struct ib_mad_port_private *
81 __ib_get_mad_port(struct ib_device *device, int port_num)
82 {
83 	struct ib_mad_port_private *entry;
84 
85 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
86 		if (entry->device == device && entry->port_num == port_num)
87 			return entry;
88 	}
89 	return NULL;
90 }
91 
92 /*
93  * Wrapper function to return a ib_mad_port_private structure or NULL
94  * for a device/port
95  */
96 static inline struct ib_mad_port_private *
97 ib_get_mad_port(struct ib_device *device, int port_num)
98 {
99 	struct ib_mad_port_private *entry;
100 	unsigned long flags;
101 
102 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
103 	entry = __ib_get_mad_port(device, port_num);
104 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
105 
106 	return entry;
107 }
108 
109 static inline u8 convert_mgmt_class(u8 mgmt_class)
110 {
111 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
112 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
113 		0 : mgmt_class;
114 }
115 
116 static int get_spl_qp_index(enum ib_qp_type qp_type)
117 {
118 	switch (qp_type)
119 	{
120 	case IB_QPT_SMI:
121 		return 0;
122 	case IB_QPT_GSI:
123 		return 1;
124 	default:
125 		return -1;
126 	}
127 }
128 
129 static int vendor_class_index(u8 mgmt_class)
130 {
131 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
132 }
133 
134 static int is_vendor_class(u8 mgmt_class)
135 {
136 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
137 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
138 		return 0;
139 	return 1;
140 }
141 
142 static int is_vendor_oui(char *oui)
143 {
144 	if (oui[0] || oui[1] || oui[2])
145 		return 1;
146 	return 0;
147 }
148 
149 static int is_vendor_method_in_use(
150 		struct ib_mad_mgmt_vendor_class *vendor_class,
151 		struct ib_mad_reg_req *mad_reg_req)
152 {
153 	struct ib_mad_mgmt_method_table *method;
154 	int i;
155 
156 	for (i = 0; i < MAX_MGMT_OUI; i++) {
157 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
158 			method = vendor_class->method_table[i];
159 			if (method) {
160 				if (method_in_use(&method, mad_reg_req))
161 					return 1;
162 				else
163 					break;
164 			}
165 		}
166 	}
167 	return 0;
168 }
169 
170 /*
171  * ib_register_mad_agent - Register to send/receive MADs
172  */
173 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
174 					   u8 port_num,
175 					   enum ib_qp_type qp_type,
176 					   struct ib_mad_reg_req *mad_reg_req,
177 					   u8 rmpp_version,
178 					   ib_mad_send_handler send_handler,
179 					   ib_mad_recv_handler recv_handler,
180 					   void *context)
181 {
182 	struct ib_mad_port_private *port_priv;
183 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
184 	struct ib_mad_agent_private *mad_agent_priv;
185 	struct ib_mad_reg_req *reg_req = NULL;
186 	struct ib_mad_mgmt_class_table *class;
187 	struct ib_mad_mgmt_vendor_class_table *vendor;
188 	struct ib_mad_mgmt_vendor_class *vendor_class;
189 	struct ib_mad_mgmt_method_table *method;
190 	int ret2, qpn;
191 	unsigned long flags;
192 	u8 mgmt_class, vclass;
193 
194 	/* Validate parameters */
195 	qpn = get_spl_qp_index(qp_type);
196 	if (qpn == -1)
197 		goto error1;
198 
199 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
200 		goto error1;
201 
202 	/* Validate MAD registration request if supplied */
203 	if (mad_reg_req) {
204 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
205 			goto error1;
206 		if (!recv_handler)
207 			goto error1;
208 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
209 			/*
210 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
211 			 * one in this range currently allowed
212 			 */
213 			if (mad_reg_req->mgmt_class !=
214 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
215 				goto error1;
216 		} else if (mad_reg_req->mgmt_class == 0) {
217 			/*
218 			 * Class 0 is reserved in IBA and is used for
219 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
220 			 */
221 			goto error1;
222 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
223 			/*
224 			 * If class is in "new" vendor range,
225 			 * ensure supplied OUI is not zero
226 			 */
227 			if (!is_vendor_oui(mad_reg_req->oui))
228 				goto error1;
229 		}
230 		/* Make sure class supplied is consistent with RMPP */
231 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
232 			if (rmpp_version)
233 				goto error1;
234 		}
235 		/* Make sure class supplied is consistent with QP type */
236 		if (qp_type == IB_QPT_SMI) {
237 			if ((mad_reg_req->mgmt_class !=
238 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
239 			    (mad_reg_req->mgmt_class !=
240 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
241 				goto error1;
242 		} else {
243 			if ((mad_reg_req->mgmt_class ==
244 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
245 			    (mad_reg_req->mgmt_class ==
246 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
247 				goto error1;
248 		}
249 	} else {
250 		/* No registration request supplied */
251 		if (!send_handler)
252 			goto error1;
253 	}
254 
255 	/* Validate device and port */
256 	port_priv = ib_get_mad_port(device, port_num);
257 	if (!port_priv) {
258 		ret = ERR_PTR(-ENODEV);
259 		goto error1;
260 	}
261 
262 	/* Allocate structures */
263 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
264 	if (!mad_agent_priv) {
265 		ret = ERR_PTR(-ENOMEM);
266 		goto error1;
267 	}
268 
269 	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
270 						 IB_ACCESS_LOCAL_WRITE);
271 	if (IS_ERR(mad_agent_priv->agent.mr)) {
272 		ret = ERR_PTR(-ENOMEM);
273 		goto error2;
274 	}
275 
276 	if (mad_reg_req) {
277 		reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
278 		if (!reg_req) {
279 			ret = ERR_PTR(-ENOMEM);
280 			goto error3;
281 		}
282 		/* Make a copy of the MAD registration request */
283 		memcpy(reg_req, mad_reg_req, sizeof *reg_req);
284 	}
285 
286 	/* Now, fill in the various structures */
287 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
288 	mad_agent_priv->reg_req = reg_req;
289 	mad_agent_priv->agent.rmpp_version = rmpp_version;
290 	mad_agent_priv->agent.device = device;
291 	mad_agent_priv->agent.recv_handler = recv_handler;
292 	mad_agent_priv->agent.send_handler = send_handler;
293 	mad_agent_priv->agent.context = context;
294 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
295 	mad_agent_priv->agent.port_num = port_num;
296 
297 	spin_lock_irqsave(&port_priv->reg_lock, flags);
298 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
299 
300 	/*
301 	 * Make sure MAD registration (if supplied)
302 	 * is non overlapping with any existing ones
303 	 */
304 	if (mad_reg_req) {
305 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
306 		if (!is_vendor_class(mgmt_class)) {
307 			class = port_priv->version[mad_reg_req->
308 						   mgmt_class_version].class;
309 			if (class) {
310 				method = class->method_table[mgmt_class];
311 				if (method) {
312 					if (method_in_use(&method,
313 							   mad_reg_req))
314 						goto error4;
315 				}
316 			}
317 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
318 						  mgmt_class);
319 		} else {
320 			/* "New" vendor class range */
321 			vendor = port_priv->version[mad_reg_req->
322 						    mgmt_class_version].vendor;
323 			if (vendor) {
324 				vclass = vendor_class_index(mgmt_class);
325 				vendor_class = vendor->vendor_class[vclass];
326 				if (vendor_class) {
327 					if (is_vendor_method_in_use(
328 							vendor_class,
329 							mad_reg_req))
330 						goto error4;
331 				}
332 			}
333 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
334 		}
335 		if (ret2) {
336 			ret = ERR_PTR(ret2);
337 			goto error4;
338 		}
339 	}
340 
341 	/* Add mad agent into port's agent list */
342 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
343 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
344 
345 	spin_lock_init(&mad_agent_priv->lock);
346 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
347 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
348 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
349 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
350 	INIT_WORK(&mad_agent_priv->timed_work, timeout_sends, mad_agent_priv);
351 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
352 	INIT_WORK(&mad_agent_priv->local_work, local_completions,
353 		   mad_agent_priv);
354 	atomic_set(&mad_agent_priv->refcount, 1);
355 	init_waitqueue_head(&mad_agent_priv->wait);
356 
357 	return &mad_agent_priv->agent;
358 
359 error4:
360 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
361 	kfree(reg_req);
362 error3:
363 	ib_dereg_mr(mad_agent_priv->agent.mr);
364 error2:
365 	kfree(mad_agent_priv);
366 error1:
367 	return ret;
368 }
369 EXPORT_SYMBOL(ib_register_mad_agent);
370 
371 static inline int is_snooping_sends(int mad_snoop_flags)
372 {
373 	return (mad_snoop_flags &
374 		(/*IB_MAD_SNOOP_POSTED_SENDS |
375 		 IB_MAD_SNOOP_RMPP_SENDS |*/
376 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
377 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
378 }
379 
380 static inline int is_snooping_recvs(int mad_snoop_flags)
381 {
382 	return (mad_snoop_flags &
383 		(IB_MAD_SNOOP_RECVS /*|
384 		 IB_MAD_SNOOP_RMPP_RECVS*/));
385 }
386 
387 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
388 				struct ib_mad_snoop_private *mad_snoop_priv)
389 {
390 	struct ib_mad_snoop_private **new_snoop_table;
391 	unsigned long flags;
392 	int i;
393 
394 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
395 	/* Check for empty slot in array. */
396 	for (i = 0; i < qp_info->snoop_table_size; i++)
397 		if (!qp_info->snoop_table[i])
398 			break;
399 
400 	if (i == qp_info->snoop_table_size) {
401 		/* Grow table. */
402 		new_snoop_table = kmalloc(sizeof mad_snoop_priv *
403 					  qp_info->snoop_table_size + 1,
404 					  GFP_ATOMIC);
405 		if (!new_snoop_table) {
406 			i = -ENOMEM;
407 			goto out;
408 		}
409 		if (qp_info->snoop_table) {
410 			memcpy(new_snoop_table, qp_info->snoop_table,
411 			       sizeof mad_snoop_priv *
412 			       qp_info->snoop_table_size);
413 			kfree(qp_info->snoop_table);
414 		}
415 		qp_info->snoop_table = new_snoop_table;
416 		qp_info->snoop_table_size++;
417 	}
418 	qp_info->snoop_table[i] = mad_snoop_priv;
419 	atomic_inc(&qp_info->snoop_count);
420 out:
421 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
422 	return i;
423 }
424 
425 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
426 					   u8 port_num,
427 					   enum ib_qp_type qp_type,
428 					   int mad_snoop_flags,
429 					   ib_mad_snoop_handler snoop_handler,
430 					   ib_mad_recv_handler recv_handler,
431 					   void *context)
432 {
433 	struct ib_mad_port_private *port_priv;
434 	struct ib_mad_agent *ret;
435 	struct ib_mad_snoop_private *mad_snoop_priv;
436 	int qpn;
437 
438 	/* Validate parameters */
439 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
440 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
441 		ret = ERR_PTR(-EINVAL);
442 		goto error1;
443 	}
444 	qpn = get_spl_qp_index(qp_type);
445 	if (qpn == -1) {
446 		ret = ERR_PTR(-EINVAL);
447 		goto error1;
448 	}
449 	port_priv = ib_get_mad_port(device, port_num);
450 	if (!port_priv) {
451 		ret = ERR_PTR(-ENODEV);
452 		goto error1;
453 	}
454 	/* Allocate structures */
455 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
456 	if (!mad_snoop_priv) {
457 		ret = ERR_PTR(-ENOMEM);
458 		goto error1;
459 	}
460 
461 	/* Now, fill in the various structures */
462 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
463 	mad_snoop_priv->agent.device = device;
464 	mad_snoop_priv->agent.recv_handler = recv_handler;
465 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
466 	mad_snoop_priv->agent.context = context;
467 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
468 	mad_snoop_priv->agent.port_num = port_num;
469 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
470 	init_waitqueue_head(&mad_snoop_priv->wait);
471 	mad_snoop_priv->snoop_index = register_snoop_agent(
472 						&port_priv->qp_info[qpn],
473 						mad_snoop_priv);
474 	if (mad_snoop_priv->snoop_index < 0) {
475 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
476 		goto error2;
477 	}
478 
479 	atomic_set(&mad_snoop_priv->refcount, 1);
480 	return &mad_snoop_priv->agent;
481 
482 error2:
483 	kfree(mad_snoop_priv);
484 error1:
485 	return ret;
486 }
487 EXPORT_SYMBOL(ib_register_mad_snoop);
488 
489 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
490 {
491 	struct ib_mad_port_private *port_priv;
492 	unsigned long flags;
493 
494 	/* Note that we could still be handling received MADs */
495 
496 	/*
497 	 * Canceling all sends results in dropping received response
498 	 * MADs, preventing us from queuing additional work
499 	 */
500 	cancel_mads(mad_agent_priv);
501 	port_priv = mad_agent_priv->qp_info->port_priv;
502 	cancel_delayed_work(&mad_agent_priv->timed_work);
503 
504 	spin_lock_irqsave(&port_priv->reg_lock, flags);
505 	remove_mad_reg_req(mad_agent_priv);
506 	list_del(&mad_agent_priv->agent_list);
507 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
508 
509 	flush_workqueue(port_priv->wq);
510 	ib_cancel_rmpp_recvs(mad_agent_priv);
511 
512 	atomic_dec(&mad_agent_priv->refcount);
513 	wait_event(mad_agent_priv->wait,
514 		   !atomic_read(&mad_agent_priv->refcount));
515 
516 	kfree(mad_agent_priv->reg_req);
517 	ib_dereg_mr(mad_agent_priv->agent.mr);
518 	kfree(mad_agent_priv);
519 }
520 
521 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
522 {
523 	struct ib_mad_qp_info *qp_info;
524 	unsigned long flags;
525 
526 	qp_info = mad_snoop_priv->qp_info;
527 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
528 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
529 	atomic_dec(&qp_info->snoop_count);
530 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
531 
532 	atomic_dec(&mad_snoop_priv->refcount);
533 	wait_event(mad_snoop_priv->wait,
534 		   !atomic_read(&mad_snoop_priv->refcount));
535 
536 	kfree(mad_snoop_priv);
537 }
538 
539 /*
540  * ib_unregister_mad_agent - Unregisters a client from using MAD services
541  */
542 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
543 {
544 	struct ib_mad_agent_private *mad_agent_priv;
545 	struct ib_mad_snoop_private *mad_snoop_priv;
546 
547 	/* If the TID is zero, the agent can only snoop. */
548 	if (mad_agent->hi_tid) {
549 		mad_agent_priv = container_of(mad_agent,
550 					      struct ib_mad_agent_private,
551 					      agent);
552 		unregister_mad_agent(mad_agent_priv);
553 	} else {
554 		mad_snoop_priv = container_of(mad_agent,
555 					      struct ib_mad_snoop_private,
556 					      agent);
557 		unregister_mad_snoop(mad_snoop_priv);
558 	}
559 	return 0;
560 }
561 EXPORT_SYMBOL(ib_unregister_mad_agent);
562 
563 static inline int response_mad(struct ib_mad *mad)
564 {
565 	/* Trap represses are responses although response bit is reset */
566 	return ((mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
567 		(mad->mad_hdr.method & IB_MGMT_METHOD_RESP));
568 }
569 
570 static void dequeue_mad(struct ib_mad_list_head *mad_list)
571 {
572 	struct ib_mad_queue *mad_queue;
573 	unsigned long flags;
574 
575 	BUG_ON(!mad_list->mad_queue);
576 	mad_queue = mad_list->mad_queue;
577 	spin_lock_irqsave(&mad_queue->lock, flags);
578 	list_del(&mad_list->list);
579 	mad_queue->count--;
580 	spin_unlock_irqrestore(&mad_queue->lock, flags);
581 }
582 
583 static void snoop_send(struct ib_mad_qp_info *qp_info,
584 		       struct ib_mad_send_buf *send_buf,
585 		       struct ib_mad_send_wc *mad_send_wc,
586 		       int mad_snoop_flags)
587 {
588 	struct ib_mad_snoop_private *mad_snoop_priv;
589 	unsigned long flags;
590 	int i;
591 
592 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
593 	for (i = 0; i < qp_info->snoop_table_size; i++) {
594 		mad_snoop_priv = qp_info->snoop_table[i];
595 		if (!mad_snoop_priv ||
596 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
597 			continue;
598 
599 		atomic_inc(&mad_snoop_priv->refcount);
600 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
601 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
602 						    send_buf, mad_send_wc);
603 		if (atomic_dec_and_test(&mad_snoop_priv->refcount))
604 			wake_up(&mad_snoop_priv->wait);
605 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
606 	}
607 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
608 }
609 
610 static void snoop_recv(struct ib_mad_qp_info *qp_info,
611 		       struct ib_mad_recv_wc *mad_recv_wc,
612 		       int mad_snoop_flags)
613 {
614 	struct ib_mad_snoop_private *mad_snoop_priv;
615 	unsigned long flags;
616 	int i;
617 
618 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
619 	for (i = 0; i < qp_info->snoop_table_size; i++) {
620 		mad_snoop_priv = qp_info->snoop_table[i];
621 		if (!mad_snoop_priv ||
622 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
623 			continue;
624 
625 		atomic_inc(&mad_snoop_priv->refcount);
626 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
627 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
628 						   mad_recv_wc);
629 		if (atomic_dec_and_test(&mad_snoop_priv->refcount))
630 			wake_up(&mad_snoop_priv->wait);
631 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
632 	}
633 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
634 }
635 
636 static void build_smp_wc(u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
637 			 struct ib_wc *wc)
638 {
639 	memset(wc, 0, sizeof *wc);
640 	wc->wr_id = wr_id;
641 	wc->status = IB_WC_SUCCESS;
642 	wc->opcode = IB_WC_RECV;
643 	wc->pkey_index = pkey_index;
644 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
645 	wc->src_qp = IB_QP0;
646 	wc->qp_num = IB_QP0;
647 	wc->slid = slid;
648 	wc->sl = 0;
649 	wc->dlid_path_bits = 0;
650 	wc->port_num = port_num;
651 }
652 
653 /*
654  * Return 0 if SMP is to be sent
655  * Return 1 if SMP was consumed locally (whether or not solicited)
656  * Return < 0 if error
657  */
658 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
659 				  struct ib_mad_send_wr_private *mad_send_wr)
660 {
661 	int ret;
662 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
663 	unsigned long flags;
664 	struct ib_mad_local_private *local;
665 	struct ib_mad_private *mad_priv;
666 	struct ib_mad_port_private *port_priv;
667 	struct ib_mad_agent_private *recv_mad_agent = NULL;
668 	struct ib_device *device = mad_agent_priv->agent.device;
669 	u8 port_num = mad_agent_priv->agent.port_num;
670 	struct ib_wc mad_wc;
671 	struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
672 
673 	/*
674 	 * Directed route handling starts if the initial LID routed part of
675 	 * a request or the ending LID routed part of a response is empty.
676 	 * If we are at the start of the LID routed part, don't update the
677 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
678 	 */
679 	if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
680 	     IB_LID_PERMISSIVE &&
681 	    !smi_handle_dr_smp_send(smp, device->node_type, port_num)) {
682 		ret = -EINVAL;
683 		printk(KERN_ERR PFX "Invalid directed route\n");
684 		goto out;
685 	}
686 	/* Check to post send on QP or process locally */
687 	ret = smi_check_local_smp(smp, device);
688 	if (!ret)
689 		goto out;
690 
691 	local = kmalloc(sizeof *local, GFP_ATOMIC);
692 	if (!local) {
693 		ret = -ENOMEM;
694 		printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
695 		goto out;
696 	}
697 	local->mad_priv = NULL;
698 	local->recv_mad_agent = NULL;
699 	mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
700 	if (!mad_priv) {
701 		ret = -ENOMEM;
702 		printk(KERN_ERR PFX "No memory for local response MAD\n");
703 		kfree(local);
704 		goto out;
705 	}
706 
707 	build_smp_wc(send_wr->wr_id, be16_to_cpu(smp->dr_slid),
708 		     send_wr->wr.ud.pkey_index,
709 		     send_wr->wr.ud.port_num, &mad_wc);
710 
711 	/* No GRH for DR SMP */
712 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
713 				  (struct ib_mad *)smp,
714 				  (struct ib_mad *)&mad_priv->mad);
715 	switch (ret)
716 	{
717 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
718 		if (response_mad(&mad_priv->mad.mad) &&
719 		    mad_agent_priv->agent.recv_handler) {
720 			local->mad_priv = mad_priv;
721 			local->recv_mad_agent = mad_agent_priv;
722 			/*
723 			 * Reference MAD agent until receive
724 			 * side of local completion handled
725 			 */
726 			atomic_inc(&mad_agent_priv->refcount);
727 		} else
728 			kmem_cache_free(ib_mad_cache, mad_priv);
729 		break;
730 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
731 		kmem_cache_free(ib_mad_cache, mad_priv);
732 		break;
733 	case IB_MAD_RESULT_SUCCESS:
734 		/* Treat like an incoming receive MAD */
735 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
736 					    mad_agent_priv->agent.port_num);
737 		if (port_priv) {
738 			mad_priv->mad.mad.mad_hdr.tid =
739 				((struct ib_mad *)smp)->mad_hdr.tid;
740 			recv_mad_agent = find_mad_agent(port_priv,
741 						        &mad_priv->mad.mad);
742 		}
743 		if (!port_priv || !recv_mad_agent) {
744 			kmem_cache_free(ib_mad_cache, mad_priv);
745 			kfree(local);
746 			ret = 0;
747 			goto out;
748 		}
749 		local->mad_priv = mad_priv;
750 		local->recv_mad_agent = recv_mad_agent;
751 		break;
752 	default:
753 		kmem_cache_free(ib_mad_cache, mad_priv);
754 		kfree(local);
755 		ret = -EINVAL;
756 		goto out;
757 	}
758 
759 	local->mad_send_wr = mad_send_wr;
760 	/* Reference MAD agent until send side of local completion handled */
761 	atomic_inc(&mad_agent_priv->refcount);
762 	/* Queue local completion to local list */
763 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
764 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
765 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
766 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
767 		   &mad_agent_priv->local_work);
768 	ret = 1;
769 out:
770 	return ret;
771 }
772 
773 static int get_pad_size(int hdr_len, int data_len)
774 {
775 	int seg_size, pad;
776 
777 	seg_size = sizeof(struct ib_mad) - hdr_len;
778 	if (data_len && seg_size) {
779 		pad = seg_size - data_len % seg_size;
780 		return pad == seg_size ? 0 : pad;
781 	} else
782 		return seg_size;
783 }
784 
785 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
786 {
787 	struct ib_rmpp_segment *s, *t;
788 
789 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
790 		list_del(&s->list);
791 		kfree(s);
792 	}
793 }
794 
795 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
796 				gfp_t gfp_mask)
797 {
798 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
799 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
800 	struct ib_rmpp_segment *seg = NULL;
801 	int left, seg_size, pad;
802 
803 	send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
804 	seg_size = send_buf->seg_size;
805 	pad = send_wr->pad;
806 
807 	/* Allocate data segments. */
808 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
809 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
810 		if (!seg) {
811 			printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
812 			       "alloc failed for len %zd, gfp %#x\n",
813 			       sizeof (*seg) + seg_size, gfp_mask);
814 			free_send_rmpp_list(send_wr);
815 			return -ENOMEM;
816 		}
817 		seg->num = ++send_buf->seg_count;
818 		list_add_tail(&seg->list, &send_wr->rmpp_list);
819 	}
820 
821 	/* Zero any padding */
822 	if (pad)
823 		memset(seg->data + seg_size - pad, 0, pad);
824 
825 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
826 					  agent.rmpp_version;
827 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
828 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
829 
830 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
831 					struct ib_rmpp_segment, list);
832 	send_wr->last_ack_seg = send_wr->cur_seg;
833 	return 0;
834 }
835 
836 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
837 					    u32 remote_qpn, u16 pkey_index,
838 					    int rmpp_active,
839 					    int hdr_len, int data_len,
840 					    gfp_t gfp_mask)
841 {
842 	struct ib_mad_agent_private *mad_agent_priv;
843 	struct ib_mad_send_wr_private *mad_send_wr;
844 	int pad, message_size, ret, size;
845 	void *buf;
846 
847 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
848 				      agent);
849 	pad = get_pad_size(hdr_len, data_len);
850 	message_size = hdr_len + data_len + pad;
851 
852 	if ((!mad_agent->rmpp_version &&
853 	     (rmpp_active || message_size > sizeof(struct ib_mad))) ||
854 	    (!rmpp_active && message_size > sizeof(struct ib_mad)))
855 		return ERR_PTR(-EINVAL);
856 
857 	size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
858 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
859 	if (!buf)
860 		return ERR_PTR(-ENOMEM);
861 
862 	mad_send_wr = buf + size;
863 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
864 	mad_send_wr->send_buf.mad = buf;
865 	mad_send_wr->send_buf.hdr_len = hdr_len;
866 	mad_send_wr->send_buf.data_len = data_len;
867 	mad_send_wr->pad = pad;
868 
869 	mad_send_wr->mad_agent_priv = mad_agent_priv;
870 	mad_send_wr->sg_list[0].length = hdr_len;
871 	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
872 	mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
873 	mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
874 
875 	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
876 	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
877 	mad_send_wr->send_wr.num_sge = 2;
878 	mad_send_wr->send_wr.opcode = IB_WR_SEND;
879 	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
880 	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
881 	mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
882 	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
883 
884 	if (rmpp_active) {
885 		ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
886 		if (ret) {
887 			kfree(buf);
888 			return ERR_PTR(ret);
889 		}
890 	}
891 
892 	mad_send_wr->send_buf.mad_agent = mad_agent;
893 	atomic_inc(&mad_agent_priv->refcount);
894 	return &mad_send_wr->send_buf;
895 }
896 EXPORT_SYMBOL(ib_create_send_mad);
897 
898 int ib_get_mad_data_offset(u8 mgmt_class)
899 {
900 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
901 		return IB_MGMT_SA_HDR;
902 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
903 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
904 		 (mgmt_class == IB_MGMT_CLASS_BIS))
905 		return IB_MGMT_DEVICE_HDR;
906 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
907 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
908 		return IB_MGMT_VENDOR_HDR;
909 	else
910 		return IB_MGMT_MAD_HDR;
911 }
912 EXPORT_SYMBOL(ib_get_mad_data_offset);
913 
914 int ib_is_mad_class_rmpp(u8 mgmt_class)
915 {
916 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
917 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
918 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
919 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
920 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
921 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
922 		return 1;
923 	return 0;
924 }
925 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
926 
927 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
928 {
929 	struct ib_mad_send_wr_private *mad_send_wr;
930 	struct list_head *list;
931 
932 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
933 				   send_buf);
934 	list = &mad_send_wr->cur_seg->list;
935 
936 	if (mad_send_wr->cur_seg->num < seg_num) {
937 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
938 			if (mad_send_wr->cur_seg->num == seg_num)
939 				break;
940 	} else if (mad_send_wr->cur_seg->num > seg_num) {
941 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
942 			if (mad_send_wr->cur_seg->num == seg_num)
943 				break;
944 	}
945 	return mad_send_wr->cur_seg->data;
946 }
947 EXPORT_SYMBOL(ib_get_rmpp_segment);
948 
949 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
950 {
951 	if (mad_send_wr->send_buf.seg_count)
952 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
953 					   mad_send_wr->seg_num);
954 	else
955 		return mad_send_wr->send_buf.mad +
956 		       mad_send_wr->send_buf.hdr_len;
957 }
958 
959 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
960 {
961 	struct ib_mad_agent_private *mad_agent_priv;
962 	struct ib_mad_send_wr_private *mad_send_wr;
963 
964 	mad_agent_priv = container_of(send_buf->mad_agent,
965 				      struct ib_mad_agent_private, agent);
966 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
967 				   send_buf);
968 
969 	free_send_rmpp_list(mad_send_wr);
970 	kfree(send_buf->mad);
971 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
972 		wake_up(&mad_agent_priv->wait);
973 }
974 EXPORT_SYMBOL(ib_free_send_mad);
975 
976 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
977 {
978 	struct ib_mad_qp_info *qp_info;
979 	struct list_head *list;
980 	struct ib_send_wr *bad_send_wr;
981 	struct ib_mad_agent *mad_agent;
982 	struct ib_sge *sge;
983 	unsigned long flags;
984 	int ret;
985 
986 	/* Set WR ID to find mad_send_wr upon completion */
987 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
988 	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
989 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
990 
991 	mad_agent = mad_send_wr->send_buf.mad_agent;
992 	sge = mad_send_wr->sg_list;
993 	sge[0].addr = dma_map_single(mad_agent->device->dma_device,
994 				     mad_send_wr->send_buf.mad,
995 				     sge[0].length,
996 				     DMA_TO_DEVICE);
997 	pci_unmap_addr_set(mad_send_wr, header_mapping, sge[0].addr);
998 
999 	sge[1].addr = dma_map_single(mad_agent->device->dma_device,
1000 				     ib_get_payload(mad_send_wr),
1001 				     sge[1].length,
1002 				     DMA_TO_DEVICE);
1003 	pci_unmap_addr_set(mad_send_wr, payload_mapping, sge[1].addr);
1004 
1005 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1006 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1007 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1008 				   &bad_send_wr);
1009 		list = &qp_info->send_queue.list;
1010 	} else {
1011 		ret = 0;
1012 		list = &qp_info->overflow_list;
1013 	}
1014 
1015 	if (!ret) {
1016 		qp_info->send_queue.count++;
1017 		list_add_tail(&mad_send_wr->mad_list.list, list);
1018 	}
1019 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1020 	if (ret) {
1021 		dma_unmap_single(mad_agent->device->dma_device,
1022 				 pci_unmap_addr(mad_send_wr, header_mapping),
1023 				 sge[0].length, DMA_TO_DEVICE);
1024 		dma_unmap_single(mad_agent->device->dma_device,
1025 				 pci_unmap_addr(mad_send_wr, payload_mapping),
1026 				 sge[1].length, DMA_TO_DEVICE);
1027 	}
1028 	return ret;
1029 }
1030 
1031 /*
1032  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1033  *  with the registered client
1034  */
1035 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1036 		     struct ib_mad_send_buf **bad_send_buf)
1037 {
1038 	struct ib_mad_agent_private *mad_agent_priv;
1039 	struct ib_mad_send_buf *next_send_buf;
1040 	struct ib_mad_send_wr_private *mad_send_wr;
1041 	unsigned long flags;
1042 	int ret = -EINVAL;
1043 
1044 	/* Walk list of send WRs and post each on send list */
1045 	for (; send_buf; send_buf = next_send_buf) {
1046 
1047 		mad_send_wr = container_of(send_buf,
1048 					   struct ib_mad_send_wr_private,
1049 					   send_buf);
1050 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1051 
1052 		if (!send_buf->mad_agent->send_handler ||
1053 		    (send_buf->timeout_ms &&
1054 		     !send_buf->mad_agent->recv_handler)) {
1055 			ret = -EINVAL;
1056 			goto error;
1057 		}
1058 
1059 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1060 			if (mad_agent_priv->agent.rmpp_version) {
1061 				ret = -EINVAL;
1062 				goto error;
1063 			}
1064 		}
1065 
1066 		/*
1067 		 * Save pointer to next work request to post in case the
1068 		 * current one completes, and the user modifies the work
1069 		 * request associated with the completion
1070 		 */
1071 		next_send_buf = send_buf->next;
1072 		mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1073 
1074 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1075 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1076 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1077 						     mad_send_wr);
1078 			if (ret < 0)		/* error */
1079 				goto error;
1080 			else if (ret == 1)	/* locally consumed */
1081 				continue;
1082 		}
1083 
1084 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1085 		/* Timeout will be updated after send completes */
1086 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1087 		mad_send_wr->retries = send_buf->retries;
1088 		/* Reference for work request to QP + response */
1089 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1090 		mad_send_wr->status = IB_WC_SUCCESS;
1091 
1092 		/* Reference MAD agent until send completes */
1093 		atomic_inc(&mad_agent_priv->refcount);
1094 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1095 		list_add_tail(&mad_send_wr->agent_list,
1096 			      &mad_agent_priv->send_list);
1097 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1098 
1099 		if (mad_agent_priv->agent.rmpp_version) {
1100 			ret = ib_send_rmpp_mad(mad_send_wr);
1101 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1102 				ret = ib_send_mad(mad_send_wr);
1103 		} else
1104 			ret = ib_send_mad(mad_send_wr);
1105 		if (ret < 0) {
1106 			/* Fail send request */
1107 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1108 			list_del(&mad_send_wr->agent_list);
1109 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1110 			atomic_dec(&mad_agent_priv->refcount);
1111 			goto error;
1112 		}
1113 	}
1114 	return 0;
1115 error:
1116 	if (bad_send_buf)
1117 		*bad_send_buf = send_buf;
1118 	return ret;
1119 }
1120 EXPORT_SYMBOL(ib_post_send_mad);
1121 
1122 /*
1123  * ib_free_recv_mad - Returns data buffers used to receive
1124  *  a MAD to the access layer
1125  */
1126 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1127 {
1128 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1129 	struct ib_mad_private_header *mad_priv_hdr;
1130 	struct ib_mad_private *priv;
1131 	struct list_head free_list;
1132 
1133 	INIT_LIST_HEAD(&free_list);
1134 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1135 
1136 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1137 					&free_list, list) {
1138 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1139 					   recv_buf);
1140 		mad_priv_hdr = container_of(mad_recv_wc,
1141 					    struct ib_mad_private_header,
1142 					    recv_wc);
1143 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1144 				    header);
1145 		kmem_cache_free(ib_mad_cache, priv);
1146 	}
1147 }
1148 EXPORT_SYMBOL(ib_free_recv_mad);
1149 
1150 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1151 					u8 rmpp_version,
1152 					ib_mad_send_handler send_handler,
1153 					ib_mad_recv_handler recv_handler,
1154 					void *context)
1155 {
1156 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1157 }
1158 EXPORT_SYMBOL(ib_redirect_mad_qp);
1159 
1160 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1161 		      struct ib_wc *wc)
1162 {
1163 	printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1164 	return 0;
1165 }
1166 EXPORT_SYMBOL(ib_process_mad_wc);
1167 
1168 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1169 			 struct ib_mad_reg_req *mad_reg_req)
1170 {
1171 	int i;
1172 
1173 	for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1174 	     i < IB_MGMT_MAX_METHODS;
1175 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1176 			       1+i)) {
1177 		if ((*method)->agent[i]) {
1178 			printk(KERN_ERR PFX "Method %d already in use\n", i);
1179 			return -EINVAL;
1180 		}
1181 	}
1182 	return 0;
1183 }
1184 
1185 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1186 {
1187 	/* Allocate management method table */
1188 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1189 	if (!*method) {
1190 		printk(KERN_ERR PFX "No memory for "
1191 		       "ib_mad_mgmt_method_table\n");
1192 		return -ENOMEM;
1193 	}
1194 
1195 	return 0;
1196 }
1197 
1198 /*
1199  * Check to see if there are any methods still in use
1200  */
1201 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1202 {
1203 	int i;
1204 
1205 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1206 		if (method->agent[i])
1207 			return 1;
1208 	return 0;
1209 }
1210 
1211 /*
1212  * Check to see if there are any method tables for this class still in use
1213  */
1214 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1215 {
1216 	int i;
1217 
1218 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1219 		if (class->method_table[i])
1220 			return 1;
1221 	return 0;
1222 }
1223 
1224 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1225 {
1226 	int i;
1227 
1228 	for (i = 0; i < MAX_MGMT_OUI; i++)
1229 		if (vendor_class->method_table[i])
1230 			return 1;
1231 	return 0;
1232 }
1233 
1234 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1235 			   char *oui)
1236 {
1237 	int i;
1238 
1239 	for (i = 0; i < MAX_MGMT_OUI; i++)
1240                 /* Is there matching OUI for this vendor class ? */
1241                 if (!memcmp(vendor_class->oui[i], oui, 3))
1242 			return i;
1243 
1244 	return -1;
1245 }
1246 
1247 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1248 {
1249 	int i;
1250 
1251 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1252 		if (vendor->vendor_class[i])
1253 			return 1;
1254 
1255 	return 0;
1256 }
1257 
1258 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1259 				     struct ib_mad_agent_private *agent)
1260 {
1261 	int i;
1262 
1263 	/* Remove any methods for this mad agent */
1264 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1265 		if (method->agent[i] == agent) {
1266 			method->agent[i] = NULL;
1267 		}
1268 	}
1269 }
1270 
1271 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1272 			      struct ib_mad_agent_private *agent_priv,
1273 			      u8 mgmt_class)
1274 {
1275 	struct ib_mad_port_private *port_priv;
1276 	struct ib_mad_mgmt_class_table **class;
1277 	struct ib_mad_mgmt_method_table **method;
1278 	int i, ret;
1279 
1280 	port_priv = agent_priv->qp_info->port_priv;
1281 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1282 	if (!*class) {
1283 		/* Allocate management class table for "new" class version */
1284 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1285 		if (!*class) {
1286 			printk(KERN_ERR PFX "No memory for "
1287 			       "ib_mad_mgmt_class_table\n");
1288 			ret = -ENOMEM;
1289 			goto error1;
1290 		}
1291 
1292 		/* Allocate method table for this management class */
1293 		method = &(*class)->method_table[mgmt_class];
1294 		if ((ret = allocate_method_table(method)))
1295 			goto error2;
1296 	} else {
1297 		method = &(*class)->method_table[mgmt_class];
1298 		if (!*method) {
1299 			/* Allocate method table for this management class */
1300 			if ((ret = allocate_method_table(method)))
1301 				goto error1;
1302 		}
1303 	}
1304 
1305 	/* Now, make sure methods are not already in use */
1306 	if (method_in_use(method, mad_reg_req))
1307 		goto error3;
1308 
1309 	/* Finally, add in methods being registered */
1310 	for (i = find_first_bit(mad_reg_req->method_mask,
1311 				IB_MGMT_MAX_METHODS);
1312 	     i < IB_MGMT_MAX_METHODS;
1313 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1314 			       1+i)) {
1315 		(*method)->agent[i] = agent_priv;
1316 	}
1317 	return 0;
1318 
1319 error3:
1320 	/* Remove any methods for this mad agent */
1321 	remove_methods_mad_agent(*method, agent_priv);
1322 	/* Now, check to see if there are any methods in use */
1323 	if (!check_method_table(*method)) {
1324 		/* If not, release management method table */
1325 		kfree(*method);
1326 		*method = NULL;
1327 	}
1328 	ret = -EINVAL;
1329 	goto error1;
1330 error2:
1331 	kfree(*class);
1332 	*class = NULL;
1333 error1:
1334 	return ret;
1335 }
1336 
1337 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1338 			   struct ib_mad_agent_private *agent_priv)
1339 {
1340 	struct ib_mad_port_private *port_priv;
1341 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1342 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1343 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1344 	struct ib_mad_mgmt_method_table **method;
1345 	int i, ret = -ENOMEM;
1346 	u8 vclass;
1347 
1348 	/* "New" vendor (with OUI) class */
1349 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1350 	port_priv = agent_priv->qp_info->port_priv;
1351 	vendor_table = &port_priv->version[
1352 				mad_reg_req->mgmt_class_version].vendor;
1353 	if (!*vendor_table) {
1354 		/* Allocate mgmt vendor class table for "new" class version */
1355 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1356 		if (!vendor) {
1357 			printk(KERN_ERR PFX "No memory for "
1358 			       "ib_mad_mgmt_vendor_class_table\n");
1359 			goto error1;
1360 		}
1361 
1362 		*vendor_table = vendor;
1363 	}
1364 	if (!(*vendor_table)->vendor_class[vclass]) {
1365 		/* Allocate table for this management vendor class */
1366 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1367 		if (!vendor_class) {
1368 			printk(KERN_ERR PFX "No memory for "
1369 			       "ib_mad_mgmt_vendor_class\n");
1370 			goto error2;
1371 		}
1372 
1373 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1374 	}
1375 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1376 		/* Is there matching OUI for this vendor class ? */
1377 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1378 			    mad_reg_req->oui, 3)) {
1379 			method = &(*vendor_table)->vendor_class[
1380 						vclass]->method_table[i];
1381 			BUG_ON(!*method);
1382 			goto check_in_use;
1383 		}
1384 	}
1385 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1386 		/* OUI slot available ? */
1387 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1388 				vclass]->oui[i])) {
1389 			method = &(*vendor_table)->vendor_class[
1390 				vclass]->method_table[i];
1391 			BUG_ON(*method);
1392 			/* Allocate method table for this OUI */
1393 			if ((ret = allocate_method_table(method)))
1394 				goto error3;
1395 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1396 			       mad_reg_req->oui, 3);
1397 			goto check_in_use;
1398 		}
1399 	}
1400 	printk(KERN_ERR PFX "All OUI slots in use\n");
1401 	goto error3;
1402 
1403 check_in_use:
1404 	/* Now, make sure methods are not already in use */
1405 	if (method_in_use(method, mad_reg_req))
1406 		goto error4;
1407 
1408 	/* Finally, add in methods being registered */
1409 	for (i = find_first_bit(mad_reg_req->method_mask,
1410 				IB_MGMT_MAX_METHODS);
1411 	     i < IB_MGMT_MAX_METHODS;
1412 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1413 			       1+i)) {
1414 		(*method)->agent[i] = agent_priv;
1415 	}
1416 	return 0;
1417 
1418 error4:
1419 	/* Remove any methods for this mad agent */
1420 	remove_methods_mad_agent(*method, agent_priv);
1421 	/* Now, check to see if there are any methods in use */
1422 	if (!check_method_table(*method)) {
1423 		/* If not, release management method table */
1424 		kfree(*method);
1425 		*method = NULL;
1426 	}
1427 	ret = -EINVAL;
1428 error3:
1429 	if (vendor_class) {
1430 		(*vendor_table)->vendor_class[vclass] = NULL;
1431 		kfree(vendor_class);
1432 	}
1433 error2:
1434 	if (vendor) {
1435 		*vendor_table = NULL;
1436 		kfree(vendor);
1437 	}
1438 error1:
1439 	return ret;
1440 }
1441 
1442 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1443 {
1444 	struct ib_mad_port_private *port_priv;
1445 	struct ib_mad_mgmt_class_table *class;
1446 	struct ib_mad_mgmt_method_table *method;
1447 	struct ib_mad_mgmt_vendor_class_table *vendor;
1448 	struct ib_mad_mgmt_vendor_class *vendor_class;
1449 	int index;
1450 	u8 mgmt_class;
1451 
1452 	/*
1453 	 * Was MAD registration request supplied
1454 	 * with original registration ?
1455 	 */
1456 	if (!agent_priv->reg_req) {
1457 		goto out;
1458 	}
1459 
1460 	port_priv = agent_priv->qp_info->port_priv;
1461 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1462 	class = port_priv->version[
1463 			agent_priv->reg_req->mgmt_class_version].class;
1464 	if (!class)
1465 		goto vendor_check;
1466 
1467 	method = class->method_table[mgmt_class];
1468 	if (method) {
1469 		/* Remove any methods for this mad agent */
1470 		remove_methods_mad_agent(method, agent_priv);
1471 		/* Now, check to see if there are any methods still in use */
1472 		if (!check_method_table(method)) {
1473 			/* If not, release management method table */
1474 			 kfree(method);
1475 			 class->method_table[mgmt_class] = NULL;
1476 			 /* Any management classes left ? */
1477 			if (!check_class_table(class)) {
1478 				/* If not, release management class table */
1479 				kfree(class);
1480 				port_priv->version[
1481 					agent_priv->reg_req->
1482 					mgmt_class_version].class = NULL;
1483 			}
1484 		}
1485 	}
1486 
1487 vendor_check:
1488 	if (!is_vendor_class(mgmt_class))
1489 		goto out;
1490 
1491 	/* normalize mgmt_class to vendor range 2 */
1492 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1493 	vendor = port_priv->version[
1494 			agent_priv->reg_req->mgmt_class_version].vendor;
1495 
1496 	if (!vendor)
1497 		goto out;
1498 
1499 	vendor_class = vendor->vendor_class[mgmt_class];
1500 	if (vendor_class) {
1501 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1502 		if (index < 0)
1503 			goto out;
1504 		method = vendor_class->method_table[index];
1505 		if (method) {
1506 			/* Remove any methods for this mad agent */
1507 			remove_methods_mad_agent(method, agent_priv);
1508 			/*
1509 			 * Now, check to see if there are
1510 			 * any methods still in use
1511 			 */
1512 			if (!check_method_table(method)) {
1513 				/* If not, release management method table */
1514 				kfree(method);
1515 				vendor_class->method_table[index] = NULL;
1516 				memset(vendor_class->oui[index], 0, 3);
1517 				/* Any OUIs left ? */
1518 				if (!check_vendor_class(vendor_class)) {
1519 					/* If not, release vendor class table */
1520 					kfree(vendor_class);
1521 					vendor->vendor_class[mgmt_class] = NULL;
1522 					/* Any other vendor classes left ? */
1523 					if (!check_vendor_table(vendor)) {
1524 						kfree(vendor);
1525 						port_priv->version[
1526 							agent_priv->reg_req->
1527 							mgmt_class_version].
1528 							vendor = NULL;
1529 					}
1530 				}
1531 			}
1532 		}
1533 	}
1534 
1535 out:
1536 	return;
1537 }
1538 
1539 static struct ib_mad_agent_private *
1540 find_mad_agent(struct ib_mad_port_private *port_priv,
1541 	       struct ib_mad *mad)
1542 {
1543 	struct ib_mad_agent_private *mad_agent = NULL;
1544 	unsigned long flags;
1545 
1546 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1547 	if (response_mad(mad)) {
1548 		u32 hi_tid;
1549 		struct ib_mad_agent_private *entry;
1550 
1551 		/*
1552 		 * Routing is based on high 32 bits of transaction ID
1553 		 * of MAD.
1554 		 */
1555 		hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1556 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1557 			if (entry->agent.hi_tid == hi_tid) {
1558 				mad_agent = entry;
1559 				break;
1560 			}
1561 		}
1562 	} else {
1563 		struct ib_mad_mgmt_class_table *class;
1564 		struct ib_mad_mgmt_method_table *method;
1565 		struct ib_mad_mgmt_vendor_class_table *vendor;
1566 		struct ib_mad_mgmt_vendor_class *vendor_class;
1567 		struct ib_vendor_mad *vendor_mad;
1568 		int index;
1569 
1570 		/*
1571 		 * Routing is based on version, class, and method
1572 		 * For "newer" vendor MADs, also based on OUI
1573 		 */
1574 		if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1575 			goto out;
1576 		if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1577 			class = port_priv->version[
1578 					mad->mad_hdr.class_version].class;
1579 			if (!class)
1580 				goto out;
1581 			method = class->method_table[convert_mgmt_class(
1582 							mad->mad_hdr.mgmt_class)];
1583 			if (method)
1584 				mad_agent = method->agent[mad->mad_hdr.method &
1585 							  ~IB_MGMT_METHOD_RESP];
1586 		} else {
1587 			vendor = port_priv->version[
1588 					mad->mad_hdr.class_version].vendor;
1589 			if (!vendor)
1590 				goto out;
1591 			vendor_class = vendor->vendor_class[vendor_class_index(
1592 						mad->mad_hdr.mgmt_class)];
1593 			if (!vendor_class)
1594 				goto out;
1595 			/* Find matching OUI */
1596 			vendor_mad = (struct ib_vendor_mad *)mad;
1597 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1598 			if (index == -1)
1599 				goto out;
1600 			method = vendor_class->method_table[index];
1601 			if (method) {
1602 				mad_agent = method->agent[mad->mad_hdr.method &
1603 							  ~IB_MGMT_METHOD_RESP];
1604 			}
1605 		}
1606 	}
1607 
1608 	if (mad_agent) {
1609 		if (mad_agent->agent.recv_handler)
1610 			atomic_inc(&mad_agent->refcount);
1611 		else {
1612 			printk(KERN_NOTICE PFX "No receive handler for client "
1613 			       "%p on port %d\n",
1614 			       &mad_agent->agent, port_priv->port_num);
1615 			mad_agent = NULL;
1616 		}
1617 	}
1618 out:
1619 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1620 
1621 	return mad_agent;
1622 }
1623 
1624 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1625 {
1626 	int valid = 0;
1627 
1628 	/* Make sure MAD base version is understood */
1629 	if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1630 		printk(KERN_ERR PFX "MAD received with unsupported base "
1631 		       "version %d\n", mad->mad_hdr.base_version);
1632 		goto out;
1633 	}
1634 
1635 	/* Filter SMI packets sent to other than QP0 */
1636 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1637 	    (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1638 		if (qp_num == 0)
1639 			valid = 1;
1640 	} else {
1641 		/* Filter GSI packets sent to QP0 */
1642 		if (qp_num != 0)
1643 			valid = 1;
1644 	}
1645 
1646 out:
1647 	return valid;
1648 }
1649 
1650 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1651 		       struct ib_mad_hdr *mad_hdr)
1652 {
1653 	struct ib_rmpp_mad *rmpp_mad;
1654 
1655 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1656 	return !mad_agent_priv->agent.rmpp_version ||
1657 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1658 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1659 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1660 }
1661 
1662 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1663 				     struct ib_mad_recv_wc *rwc)
1664 {
1665 	return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1666 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1667 }
1668 
1669 static inline int rcv_has_same_gid(struct ib_mad_send_wr_private *wr,
1670 				   struct ib_mad_recv_wc *rwc )
1671 {
1672 	struct ib_ah_attr attr;
1673 	u8 send_resp, rcv_resp;
1674 
1675 	send_resp = ((struct ib_mad *)(wr->send_buf.mad))->
1676 		     mad_hdr.method & IB_MGMT_METHOD_RESP;
1677 	rcv_resp = rwc->recv_buf.mad->mad_hdr.method & IB_MGMT_METHOD_RESP;
1678 
1679 	if (!send_resp && rcv_resp)
1680 		/* is request/response. GID/LIDs are both local (same). */
1681 		return 1;
1682 
1683 	if (send_resp == rcv_resp)
1684 		/* both requests, or both responses. GIDs different */
1685 		return 0;
1686 
1687 	if (ib_query_ah(wr->send_buf.ah, &attr))
1688 		/* Assume not equal, to avoid false positives. */
1689 		return 0;
1690 
1691 	if (!(attr.ah_flags & IB_AH_GRH) && !(rwc->wc->wc_flags & IB_WC_GRH))
1692 		return attr.dlid == rwc->wc->slid;
1693 	else if ((attr.ah_flags & IB_AH_GRH) &&
1694 		 (rwc->wc->wc_flags & IB_WC_GRH))
1695 		return memcmp(attr.grh.dgid.raw,
1696 			      rwc->recv_buf.grh->sgid.raw, 16) == 0;
1697 	else
1698 		/* one has GID, other does not.  Assume different */
1699 		return 0;
1700 }
1701 struct ib_mad_send_wr_private*
1702 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1703 		 struct ib_mad_recv_wc *mad_recv_wc)
1704 {
1705 	struct ib_mad_send_wr_private *mad_send_wr;
1706 	struct ib_mad *mad;
1707 
1708 	mad = (struct ib_mad *)mad_recv_wc->recv_buf.mad;
1709 
1710 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
1711 			    agent_list) {
1712 		if ((mad_send_wr->tid == mad->mad_hdr.tid) &&
1713 		    rcv_has_same_class(mad_send_wr, mad_recv_wc) &&
1714 		    rcv_has_same_gid(mad_send_wr, mad_recv_wc))
1715 			return mad_send_wr;
1716 	}
1717 
1718 	/*
1719 	 * It's possible to receive the response before we've
1720 	 * been notified that the send has completed
1721 	 */
1722 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
1723 			    agent_list) {
1724 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
1725 		    mad_send_wr->tid == mad->mad_hdr.tid &&
1726 		    mad_send_wr->timeout &&
1727 		    rcv_has_same_class(mad_send_wr, mad_recv_wc) &&
1728 		    rcv_has_same_gid(mad_send_wr, mad_recv_wc)) {
1729 			/* Verify request has not been canceled */
1730 			return (mad_send_wr->status == IB_WC_SUCCESS) ?
1731 				mad_send_wr : NULL;
1732 		}
1733 	}
1734 	return NULL;
1735 }
1736 
1737 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1738 {
1739 	mad_send_wr->timeout = 0;
1740 	if (mad_send_wr->refcount == 1) {
1741 		list_del(&mad_send_wr->agent_list);
1742 		list_add_tail(&mad_send_wr->agent_list,
1743 			      &mad_send_wr->mad_agent_priv->done_list);
1744 	}
1745 }
1746 
1747 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1748 				 struct ib_mad_recv_wc *mad_recv_wc)
1749 {
1750 	struct ib_mad_send_wr_private *mad_send_wr;
1751 	struct ib_mad_send_wc mad_send_wc;
1752 	unsigned long flags;
1753 
1754 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1755 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1756 	if (mad_agent_priv->agent.rmpp_version) {
1757 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1758 						      mad_recv_wc);
1759 		if (!mad_recv_wc) {
1760 			if (atomic_dec_and_test(&mad_agent_priv->refcount))
1761 				wake_up(&mad_agent_priv->wait);
1762 			return;
1763 		}
1764 	}
1765 
1766 	/* Complete corresponding request */
1767 	if (response_mad(mad_recv_wc->recv_buf.mad)) {
1768 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1769 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1770 		if (!mad_send_wr) {
1771 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1772 			ib_free_recv_mad(mad_recv_wc);
1773 			if (atomic_dec_and_test(&mad_agent_priv->refcount))
1774 				wake_up(&mad_agent_priv->wait);
1775 			return;
1776 		}
1777 		ib_mark_mad_done(mad_send_wr);
1778 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1779 
1780 		/* Defined behavior is to complete response before request */
1781 		mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1782 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1783 						   mad_recv_wc);
1784 		atomic_dec(&mad_agent_priv->refcount);
1785 
1786 		mad_send_wc.status = IB_WC_SUCCESS;
1787 		mad_send_wc.vendor_err = 0;
1788 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
1789 		ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1790 	} else {
1791 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1792 						   mad_recv_wc);
1793 		if (atomic_dec_and_test(&mad_agent_priv->refcount))
1794 			wake_up(&mad_agent_priv->wait);
1795 	}
1796 }
1797 
1798 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1799 				     struct ib_wc *wc)
1800 {
1801 	struct ib_mad_qp_info *qp_info;
1802 	struct ib_mad_private_header *mad_priv_hdr;
1803 	struct ib_mad_private *recv, *response;
1804 	struct ib_mad_list_head *mad_list;
1805 	struct ib_mad_agent_private *mad_agent;
1806 
1807 	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1808 	if (!response)
1809 		printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1810 		       "for response buffer\n");
1811 
1812 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1813 	qp_info = mad_list->mad_queue->qp_info;
1814 	dequeue_mad(mad_list);
1815 
1816 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1817 				    mad_list);
1818 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1819 	dma_unmap_single(port_priv->device->dma_device,
1820 			 pci_unmap_addr(&recv->header, mapping),
1821 			 sizeof(struct ib_mad_private) -
1822 			 sizeof(struct ib_mad_private_header),
1823 			 DMA_FROM_DEVICE);
1824 
1825 	/* Setup MAD receive work completion from "normal" work completion */
1826 	recv->header.wc = *wc;
1827 	recv->header.recv_wc.wc = &recv->header.wc;
1828 	recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1829 	recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1830 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
1831 
1832 	if (atomic_read(&qp_info->snoop_count))
1833 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1834 
1835 	/* Validate MAD */
1836 	if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1837 		goto out;
1838 
1839 	if (recv->mad.mad.mad_hdr.mgmt_class ==
1840 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1841 		if (!smi_handle_dr_smp_recv(&recv->mad.smp,
1842 					    port_priv->device->node_type,
1843 					    port_priv->port_num,
1844 					    port_priv->device->phys_port_cnt))
1845 			goto out;
1846 		if (!smi_check_forward_dr_smp(&recv->mad.smp))
1847 			goto local;
1848 		if (!smi_handle_dr_smp_send(&recv->mad.smp,
1849 					    port_priv->device->node_type,
1850 					    port_priv->port_num))
1851 			goto out;
1852 		if (!smi_check_local_smp(&recv->mad.smp, port_priv->device))
1853 			goto out;
1854 	}
1855 
1856 local:
1857 	/* Give driver "right of first refusal" on incoming MAD */
1858 	if (port_priv->device->process_mad) {
1859 		int ret;
1860 
1861 		if (!response) {
1862 			printk(KERN_ERR PFX "No memory for response MAD\n");
1863 			/*
1864 			 * Is it better to assume that
1865 			 * it wouldn't be processed ?
1866 			 */
1867 			goto out;
1868 		}
1869 
1870 		ret = port_priv->device->process_mad(port_priv->device, 0,
1871 						     port_priv->port_num,
1872 						     wc, &recv->grh,
1873 						     &recv->mad.mad,
1874 						     &response->mad.mad);
1875 		if (ret & IB_MAD_RESULT_SUCCESS) {
1876 			if (ret & IB_MAD_RESULT_CONSUMED)
1877 				goto out;
1878 			if (ret & IB_MAD_RESULT_REPLY) {
1879 				agent_send_response(&response->mad.mad,
1880 						    &recv->grh, wc,
1881 						    port_priv->device,
1882 						    port_priv->port_num,
1883 						    qp_info->qp->qp_num);
1884 				goto out;
1885 			}
1886 		}
1887 	}
1888 
1889 	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1890 	if (mad_agent) {
1891 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1892 		/*
1893 		 * recv is freed up in error cases in ib_mad_complete_recv
1894 		 * or via recv_handler in ib_mad_complete_recv()
1895 		 */
1896 		recv = NULL;
1897 	}
1898 
1899 out:
1900 	/* Post another receive request for this QP */
1901 	if (response) {
1902 		ib_mad_post_receive_mads(qp_info, response);
1903 		if (recv)
1904 			kmem_cache_free(ib_mad_cache, recv);
1905 	} else
1906 		ib_mad_post_receive_mads(qp_info, recv);
1907 }
1908 
1909 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1910 {
1911 	struct ib_mad_send_wr_private *mad_send_wr;
1912 	unsigned long delay;
1913 
1914 	if (list_empty(&mad_agent_priv->wait_list)) {
1915 		cancel_delayed_work(&mad_agent_priv->timed_work);
1916 	} else {
1917 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1918 					 struct ib_mad_send_wr_private,
1919 					 agent_list);
1920 
1921 		if (time_after(mad_agent_priv->timeout,
1922 			       mad_send_wr->timeout)) {
1923 			mad_agent_priv->timeout = mad_send_wr->timeout;
1924 			cancel_delayed_work(&mad_agent_priv->timed_work);
1925 			delay = mad_send_wr->timeout - jiffies;
1926 			if ((long)delay <= 0)
1927 				delay = 1;
1928 			queue_delayed_work(mad_agent_priv->qp_info->
1929 					   port_priv->wq,
1930 					   &mad_agent_priv->timed_work, delay);
1931 		}
1932 	}
1933 }
1934 
1935 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1936 {
1937 	struct ib_mad_agent_private *mad_agent_priv;
1938 	struct ib_mad_send_wr_private *temp_mad_send_wr;
1939 	struct list_head *list_item;
1940 	unsigned long delay;
1941 
1942 	mad_agent_priv = mad_send_wr->mad_agent_priv;
1943 	list_del(&mad_send_wr->agent_list);
1944 
1945 	delay = mad_send_wr->timeout;
1946 	mad_send_wr->timeout += jiffies;
1947 
1948 	if (delay) {
1949 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1950 			temp_mad_send_wr = list_entry(list_item,
1951 						struct ib_mad_send_wr_private,
1952 						agent_list);
1953 			if (time_after(mad_send_wr->timeout,
1954 				       temp_mad_send_wr->timeout))
1955 				break;
1956 		}
1957 	}
1958 	else
1959 		list_item = &mad_agent_priv->wait_list;
1960 	list_add(&mad_send_wr->agent_list, list_item);
1961 
1962 	/* Reschedule a work item if we have a shorter timeout */
1963 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
1964 		cancel_delayed_work(&mad_agent_priv->timed_work);
1965 		queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
1966 				   &mad_agent_priv->timed_work, delay);
1967 	}
1968 }
1969 
1970 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
1971 			  int timeout_ms)
1972 {
1973 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
1974 	wait_for_response(mad_send_wr);
1975 }
1976 
1977 /*
1978  * Process a send work completion
1979  */
1980 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
1981 			     struct ib_mad_send_wc *mad_send_wc)
1982 {
1983 	struct ib_mad_agent_private	*mad_agent_priv;
1984 	unsigned long			flags;
1985 	int				ret;
1986 
1987 	mad_agent_priv = mad_send_wr->mad_agent_priv;
1988 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
1989 	if (mad_agent_priv->agent.rmpp_version) {
1990 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
1991 		if (ret == IB_RMPP_RESULT_CONSUMED)
1992 			goto done;
1993 	} else
1994 		ret = IB_RMPP_RESULT_UNHANDLED;
1995 
1996 	if (mad_send_wc->status != IB_WC_SUCCESS &&
1997 	    mad_send_wr->status == IB_WC_SUCCESS) {
1998 		mad_send_wr->status = mad_send_wc->status;
1999 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2000 	}
2001 
2002 	if (--mad_send_wr->refcount > 0) {
2003 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2004 		    mad_send_wr->status == IB_WC_SUCCESS) {
2005 			wait_for_response(mad_send_wr);
2006 		}
2007 		goto done;
2008 	}
2009 
2010 	/* Remove send from MAD agent and notify client of completion */
2011 	list_del(&mad_send_wr->agent_list);
2012 	adjust_timeout(mad_agent_priv);
2013 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2014 
2015 	if (mad_send_wr->status != IB_WC_SUCCESS )
2016 		mad_send_wc->status = mad_send_wr->status;
2017 	if (ret == IB_RMPP_RESULT_INTERNAL)
2018 		ib_rmpp_send_handler(mad_send_wc);
2019 	else
2020 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2021 						   mad_send_wc);
2022 
2023 	/* Release reference on agent taken when sending */
2024 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
2025 		wake_up(&mad_agent_priv->wait);
2026 	return;
2027 done:
2028 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2029 }
2030 
2031 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2032 				     struct ib_wc *wc)
2033 {
2034 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2035 	struct ib_mad_list_head		*mad_list;
2036 	struct ib_mad_qp_info		*qp_info;
2037 	struct ib_mad_queue		*send_queue;
2038 	struct ib_send_wr		*bad_send_wr;
2039 	struct ib_mad_send_wc		mad_send_wc;
2040 	unsigned long flags;
2041 	int ret;
2042 
2043 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2044 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2045 				   mad_list);
2046 	send_queue = mad_list->mad_queue;
2047 	qp_info = send_queue->qp_info;
2048 
2049 retry:
2050 	dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
2051 			 pci_unmap_addr(mad_send_wr, header_mapping),
2052 			 mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2053 	dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
2054 			 pci_unmap_addr(mad_send_wr, payload_mapping),
2055 			 mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2056 	queued_send_wr = NULL;
2057 	spin_lock_irqsave(&send_queue->lock, flags);
2058 	list_del(&mad_list->list);
2059 
2060 	/* Move queued send to the send queue */
2061 	if (send_queue->count-- > send_queue->max_active) {
2062 		mad_list = container_of(qp_info->overflow_list.next,
2063 					struct ib_mad_list_head, list);
2064 		queued_send_wr = container_of(mad_list,
2065 					struct ib_mad_send_wr_private,
2066 					mad_list);
2067 		list_del(&mad_list->list);
2068 		list_add_tail(&mad_list->list, &send_queue->list);
2069 	}
2070 	spin_unlock_irqrestore(&send_queue->lock, flags);
2071 
2072 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2073 	mad_send_wc.status = wc->status;
2074 	mad_send_wc.vendor_err = wc->vendor_err;
2075 	if (atomic_read(&qp_info->snoop_count))
2076 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2077 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2078 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2079 
2080 	if (queued_send_wr) {
2081 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2082 				   &bad_send_wr);
2083 		if (ret) {
2084 			printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2085 			mad_send_wr = queued_send_wr;
2086 			wc->status = IB_WC_LOC_QP_OP_ERR;
2087 			goto retry;
2088 		}
2089 	}
2090 }
2091 
2092 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2093 {
2094 	struct ib_mad_send_wr_private *mad_send_wr;
2095 	struct ib_mad_list_head *mad_list;
2096 	unsigned long flags;
2097 
2098 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2099 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2100 		mad_send_wr = container_of(mad_list,
2101 					   struct ib_mad_send_wr_private,
2102 					   mad_list);
2103 		mad_send_wr->retry = 1;
2104 	}
2105 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2106 }
2107 
2108 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2109 			      struct ib_wc *wc)
2110 {
2111 	struct ib_mad_list_head *mad_list;
2112 	struct ib_mad_qp_info *qp_info;
2113 	struct ib_mad_send_wr_private *mad_send_wr;
2114 	int ret;
2115 
2116 	/* Determine if failure was a send or receive */
2117 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2118 	qp_info = mad_list->mad_queue->qp_info;
2119 	if (mad_list->mad_queue == &qp_info->recv_queue)
2120 		/*
2121 		 * Receive errors indicate that the QP has entered the error
2122 		 * state - error handling/shutdown code will cleanup
2123 		 */
2124 		return;
2125 
2126 	/*
2127 	 * Send errors will transition the QP to SQE - move
2128 	 * QP to RTS and repost flushed work requests
2129 	 */
2130 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2131 				   mad_list);
2132 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2133 		if (mad_send_wr->retry) {
2134 			/* Repost send */
2135 			struct ib_send_wr *bad_send_wr;
2136 
2137 			mad_send_wr->retry = 0;
2138 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2139 					&bad_send_wr);
2140 			if (ret)
2141 				ib_mad_send_done_handler(port_priv, wc);
2142 		} else
2143 			ib_mad_send_done_handler(port_priv, wc);
2144 	} else {
2145 		struct ib_qp_attr *attr;
2146 
2147 		/* Transition QP to RTS and fail offending send */
2148 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2149 		if (attr) {
2150 			attr->qp_state = IB_QPS_RTS;
2151 			attr->cur_qp_state = IB_QPS_SQE;
2152 			ret = ib_modify_qp(qp_info->qp, attr,
2153 					   IB_QP_STATE | IB_QP_CUR_STATE);
2154 			kfree(attr);
2155 			if (ret)
2156 				printk(KERN_ERR PFX "mad_error_handler - "
2157 				       "ib_modify_qp to RTS : %d\n", ret);
2158 			else
2159 				mark_sends_for_retry(qp_info);
2160 		}
2161 		ib_mad_send_done_handler(port_priv, wc);
2162 	}
2163 }
2164 
2165 /*
2166  * IB MAD completion callback
2167  */
2168 static void ib_mad_completion_handler(void *data)
2169 {
2170 	struct ib_mad_port_private *port_priv;
2171 	struct ib_wc wc;
2172 
2173 	port_priv = (struct ib_mad_port_private *)data;
2174 	ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2175 
2176 	while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2177 		if (wc.status == IB_WC_SUCCESS) {
2178 			switch (wc.opcode) {
2179 			case IB_WC_SEND:
2180 				ib_mad_send_done_handler(port_priv, &wc);
2181 				break;
2182 			case IB_WC_RECV:
2183 				ib_mad_recv_done_handler(port_priv, &wc);
2184 				break;
2185 			default:
2186 				BUG_ON(1);
2187 				break;
2188 			}
2189 		} else
2190 			mad_error_handler(port_priv, &wc);
2191 	}
2192 }
2193 
2194 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2195 {
2196 	unsigned long flags;
2197 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2198 	struct ib_mad_send_wc mad_send_wc;
2199 	struct list_head cancel_list;
2200 
2201 	INIT_LIST_HEAD(&cancel_list);
2202 
2203 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2204 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2205 				 &mad_agent_priv->send_list, agent_list) {
2206 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2207  			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2208 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2209 		}
2210 	}
2211 
2212 	/* Empty wait list to prevent receives from finding a request */
2213 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2214 	/* Empty local completion list as well */
2215 	list_splice_init(&mad_agent_priv->local_list, &cancel_list);
2216 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2217 
2218 	/* Report all cancelled requests */
2219 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2220 	mad_send_wc.vendor_err = 0;
2221 
2222 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2223 				 &cancel_list, agent_list) {
2224 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2225 		list_del(&mad_send_wr->agent_list);
2226 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2227 						   &mad_send_wc);
2228 		atomic_dec(&mad_agent_priv->refcount);
2229 	}
2230 }
2231 
2232 static struct ib_mad_send_wr_private*
2233 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2234 	     struct ib_mad_send_buf *send_buf)
2235 {
2236 	struct ib_mad_send_wr_private *mad_send_wr;
2237 
2238 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2239 			    agent_list) {
2240 		if (&mad_send_wr->send_buf == send_buf)
2241 			return mad_send_wr;
2242 	}
2243 
2244 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2245 			    agent_list) {
2246 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2247 		    &mad_send_wr->send_buf == send_buf)
2248 			return mad_send_wr;
2249 	}
2250 	return NULL;
2251 }
2252 
2253 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2254 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2255 {
2256 	struct ib_mad_agent_private *mad_agent_priv;
2257 	struct ib_mad_send_wr_private *mad_send_wr;
2258 	unsigned long flags;
2259 	int active;
2260 
2261 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2262 				      agent);
2263 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2264 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2265 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2266 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2267 		return -EINVAL;
2268 	}
2269 
2270 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2271 	if (!timeout_ms) {
2272 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2273 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2274 	}
2275 
2276 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2277 	if (active)
2278 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2279 	else
2280 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2281 
2282 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2283 	return 0;
2284 }
2285 EXPORT_SYMBOL(ib_modify_mad);
2286 
2287 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2288 		   struct ib_mad_send_buf *send_buf)
2289 {
2290 	ib_modify_mad(mad_agent, send_buf, 0);
2291 }
2292 EXPORT_SYMBOL(ib_cancel_mad);
2293 
2294 static void local_completions(void *data)
2295 {
2296 	struct ib_mad_agent_private *mad_agent_priv;
2297 	struct ib_mad_local_private *local;
2298 	struct ib_mad_agent_private *recv_mad_agent;
2299 	unsigned long flags;
2300 	int recv = 0;
2301 	struct ib_wc wc;
2302 	struct ib_mad_send_wc mad_send_wc;
2303 
2304 	mad_agent_priv = (struct ib_mad_agent_private *)data;
2305 
2306 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2307 	while (!list_empty(&mad_agent_priv->local_list)) {
2308 		local = list_entry(mad_agent_priv->local_list.next,
2309 				   struct ib_mad_local_private,
2310 				   completion_list);
2311 		list_del(&local->completion_list);
2312 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2313 		if (local->mad_priv) {
2314 			recv_mad_agent = local->recv_mad_agent;
2315 			if (!recv_mad_agent) {
2316 				printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2317 				goto local_send_completion;
2318 			}
2319 
2320 			recv = 1;
2321 			/*
2322 			 * Defined behavior is to complete response
2323 			 * before request
2324 			 */
2325 			build_smp_wc((unsigned long) local->mad_send_wr,
2326 				     be16_to_cpu(IB_LID_PERMISSIVE),
2327 				     0, recv_mad_agent->agent.port_num, &wc);
2328 
2329 			local->mad_priv->header.recv_wc.wc = &wc;
2330 			local->mad_priv->header.recv_wc.mad_len =
2331 						sizeof(struct ib_mad);
2332 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2333 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2334 				 &local->mad_priv->header.recv_wc.rmpp_list);
2335 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2336 			local->mad_priv->header.recv_wc.recv_buf.mad =
2337 						&local->mad_priv->mad.mad;
2338 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2339 				snoop_recv(recv_mad_agent->qp_info,
2340 					  &local->mad_priv->header.recv_wc,
2341 					   IB_MAD_SNOOP_RECVS);
2342 			recv_mad_agent->agent.recv_handler(
2343 						&recv_mad_agent->agent,
2344 						&local->mad_priv->header.recv_wc);
2345 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2346 			atomic_dec(&recv_mad_agent->refcount);
2347 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2348 		}
2349 
2350 local_send_completion:
2351 		/* Complete send */
2352 		mad_send_wc.status = IB_WC_SUCCESS;
2353 		mad_send_wc.vendor_err = 0;
2354 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2355 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2356 			snoop_send(mad_agent_priv->qp_info,
2357 				   &local->mad_send_wr->send_buf,
2358 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2359 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2360 						   &mad_send_wc);
2361 
2362 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2363 		atomic_dec(&mad_agent_priv->refcount);
2364 		if (!recv)
2365 			kmem_cache_free(ib_mad_cache, local->mad_priv);
2366 		kfree(local);
2367 	}
2368 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2369 }
2370 
2371 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2372 {
2373 	int ret;
2374 
2375 	if (!mad_send_wr->retries--)
2376 		return -ETIMEDOUT;
2377 
2378 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2379 
2380 	if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2381 		ret = ib_retry_rmpp(mad_send_wr);
2382 		switch (ret) {
2383 		case IB_RMPP_RESULT_UNHANDLED:
2384 			ret = ib_send_mad(mad_send_wr);
2385 			break;
2386 		case IB_RMPP_RESULT_CONSUMED:
2387 			ret = 0;
2388 			break;
2389 		default:
2390 			ret = -ECOMM;
2391 			break;
2392 		}
2393 	} else
2394 		ret = ib_send_mad(mad_send_wr);
2395 
2396 	if (!ret) {
2397 		mad_send_wr->refcount++;
2398 		list_add_tail(&mad_send_wr->agent_list,
2399 			      &mad_send_wr->mad_agent_priv->send_list);
2400 	}
2401 	return ret;
2402 }
2403 
2404 static void timeout_sends(void *data)
2405 {
2406 	struct ib_mad_agent_private *mad_agent_priv;
2407 	struct ib_mad_send_wr_private *mad_send_wr;
2408 	struct ib_mad_send_wc mad_send_wc;
2409 	unsigned long flags, delay;
2410 
2411 	mad_agent_priv = (struct ib_mad_agent_private *)data;
2412 	mad_send_wc.vendor_err = 0;
2413 
2414 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2415 	while (!list_empty(&mad_agent_priv->wait_list)) {
2416 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2417 					 struct ib_mad_send_wr_private,
2418 					 agent_list);
2419 
2420 		if (time_after(mad_send_wr->timeout, jiffies)) {
2421 			delay = mad_send_wr->timeout - jiffies;
2422 			if ((long)delay <= 0)
2423 				delay = 1;
2424 			queue_delayed_work(mad_agent_priv->qp_info->
2425 					   port_priv->wq,
2426 					   &mad_agent_priv->timed_work, delay);
2427 			break;
2428 		}
2429 
2430 		list_del(&mad_send_wr->agent_list);
2431 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2432 		    !retry_send(mad_send_wr))
2433 			continue;
2434 
2435 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2436 
2437 		if (mad_send_wr->status == IB_WC_SUCCESS)
2438 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2439 		else
2440 			mad_send_wc.status = mad_send_wr->status;
2441 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2442 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2443 						   &mad_send_wc);
2444 
2445 		atomic_dec(&mad_agent_priv->refcount);
2446 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2447 	}
2448 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2449 }
2450 
2451 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2452 {
2453 	struct ib_mad_port_private *port_priv = cq->cq_context;
2454 	unsigned long flags;
2455 
2456 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2457 	if (!list_empty(&port_priv->port_list))
2458 		queue_work(port_priv->wq, &port_priv->work);
2459 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2460 }
2461 
2462 /*
2463  * Allocate receive MADs and post receive WRs for them
2464  */
2465 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2466 				    struct ib_mad_private *mad)
2467 {
2468 	unsigned long flags;
2469 	int post, ret;
2470 	struct ib_mad_private *mad_priv;
2471 	struct ib_sge sg_list;
2472 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2473 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2474 
2475 	/* Initialize common scatter list fields */
2476 	sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2477 	sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2478 
2479 	/* Initialize common receive WR fields */
2480 	recv_wr.next = NULL;
2481 	recv_wr.sg_list = &sg_list;
2482 	recv_wr.num_sge = 1;
2483 
2484 	do {
2485 		/* Allocate and map receive buffer */
2486 		if (mad) {
2487 			mad_priv = mad;
2488 			mad = NULL;
2489 		} else {
2490 			mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2491 			if (!mad_priv) {
2492 				printk(KERN_ERR PFX "No memory for receive buffer\n");
2493 				ret = -ENOMEM;
2494 				break;
2495 			}
2496 		}
2497 		sg_list.addr = dma_map_single(qp_info->port_priv->
2498 					      	device->dma_device,
2499 					      &mad_priv->grh,
2500 					      sizeof *mad_priv -
2501 					      	sizeof mad_priv->header,
2502 					      DMA_FROM_DEVICE);
2503 		pci_unmap_addr_set(&mad_priv->header, mapping, sg_list.addr);
2504 		recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2505 		mad_priv->header.mad_list.mad_queue = recv_queue;
2506 
2507 		/* Post receive WR */
2508 		spin_lock_irqsave(&recv_queue->lock, flags);
2509 		post = (++recv_queue->count < recv_queue->max_active);
2510 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2511 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2512 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2513 		if (ret) {
2514 			spin_lock_irqsave(&recv_queue->lock, flags);
2515 			list_del(&mad_priv->header.mad_list.list);
2516 			recv_queue->count--;
2517 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2518 			dma_unmap_single(qp_info->port_priv->device->dma_device,
2519 					 pci_unmap_addr(&mad_priv->header,
2520 							mapping),
2521 					 sizeof *mad_priv -
2522 					   sizeof mad_priv->header,
2523 					 DMA_FROM_DEVICE);
2524 			kmem_cache_free(ib_mad_cache, mad_priv);
2525 			printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2526 			break;
2527 		}
2528 	} while (post);
2529 
2530 	return ret;
2531 }
2532 
2533 /*
2534  * Return all the posted receive MADs
2535  */
2536 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2537 {
2538 	struct ib_mad_private_header *mad_priv_hdr;
2539 	struct ib_mad_private *recv;
2540 	struct ib_mad_list_head *mad_list;
2541 
2542 	while (!list_empty(&qp_info->recv_queue.list)) {
2543 
2544 		mad_list = list_entry(qp_info->recv_queue.list.next,
2545 				      struct ib_mad_list_head, list);
2546 		mad_priv_hdr = container_of(mad_list,
2547 					    struct ib_mad_private_header,
2548 					    mad_list);
2549 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2550 				    header);
2551 
2552 		/* Remove from posted receive MAD list */
2553 		list_del(&mad_list->list);
2554 
2555 		dma_unmap_single(qp_info->port_priv->device->dma_device,
2556 				 pci_unmap_addr(&recv->header, mapping),
2557 				 sizeof(struct ib_mad_private) -
2558 				 sizeof(struct ib_mad_private_header),
2559 				 DMA_FROM_DEVICE);
2560 		kmem_cache_free(ib_mad_cache, recv);
2561 	}
2562 
2563 	qp_info->recv_queue.count = 0;
2564 }
2565 
2566 /*
2567  * Start the port
2568  */
2569 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2570 {
2571 	int ret, i;
2572 	struct ib_qp_attr *attr;
2573 	struct ib_qp *qp;
2574 
2575 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2576  	if (!attr) {
2577 		printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2578 		return -ENOMEM;
2579 	}
2580 
2581 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2582 		qp = port_priv->qp_info[i].qp;
2583 		/*
2584 		 * PKey index for QP1 is irrelevant but
2585 		 * one is needed for the Reset to Init transition
2586 		 */
2587 		attr->qp_state = IB_QPS_INIT;
2588 		attr->pkey_index = 0;
2589 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2590 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2591 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2592 		if (ret) {
2593 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2594 			       "INIT: %d\n", i, ret);
2595 			goto out;
2596 		}
2597 
2598 		attr->qp_state = IB_QPS_RTR;
2599 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2600 		if (ret) {
2601 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2602 			       "RTR: %d\n", i, ret);
2603 			goto out;
2604 		}
2605 
2606 		attr->qp_state = IB_QPS_RTS;
2607 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2608 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2609 		if (ret) {
2610 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2611 			       "RTS: %d\n", i, ret);
2612 			goto out;
2613 		}
2614 	}
2615 
2616 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2617 	if (ret) {
2618 		printk(KERN_ERR PFX "Failed to request completion "
2619 		       "notification: %d\n", ret);
2620 		goto out;
2621 	}
2622 
2623 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2624 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2625 		if (ret) {
2626 			printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2627 			goto out;
2628 		}
2629 	}
2630 out:
2631 	kfree(attr);
2632 	return ret;
2633 }
2634 
2635 static void qp_event_handler(struct ib_event *event, void *qp_context)
2636 {
2637 	struct ib_mad_qp_info	*qp_info = qp_context;
2638 
2639 	/* It's worse than that! He's dead, Jim! */
2640 	printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2641 		event->event, qp_info->qp->qp_num);
2642 }
2643 
2644 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2645 			   struct ib_mad_queue *mad_queue)
2646 {
2647 	mad_queue->qp_info = qp_info;
2648 	mad_queue->count = 0;
2649 	spin_lock_init(&mad_queue->lock);
2650 	INIT_LIST_HEAD(&mad_queue->list);
2651 }
2652 
2653 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2654 			struct ib_mad_qp_info *qp_info)
2655 {
2656 	qp_info->port_priv = port_priv;
2657 	init_mad_queue(qp_info, &qp_info->send_queue);
2658 	init_mad_queue(qp_info, &qp_info->recv_queue);
2659 	INIT_LIST_HEAD(&qp_info->overflow_list);
2660 	spin_lock_init(&qp_info->snoop_lock);
2661 	qp_info->snoop_table = NULL;
2662 	qp_info->snoop_table_size = 0;
2663 	atomic_set(&qp_info->snoop_count, 0);
2664 }
2665 
2666 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2667 			 enum ib_qp_type qp_type)
2668 {
2669 	struct ib_qp_init_attr	qp_init_attr;
2670 	int ret;
2671 
2672 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2673 	qp_init_attr.send_cq = qp_info->port_priv->cq;
2674 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2675 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2676 	qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2677 	qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2678 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2679 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2680 	qp_init_attr.qp_type = qp_type;
2681 	qp_init_attr.port_num = qp_info->port_priv->port_num;
2682 	qp_init_attr.qp_context = qp_info;
2683 	qp_init_attr.event_handler = qp_event_handler;
2684 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2685 	if (IS_ERR(qp_info->qp)) {
2686 		printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2687 		       get_spl_qp_index(qp_type));
2688 		ret = PTR_ERR(qp_info->qp);
2689 		goto error;
2690 	}
2691 	/* Use minimum queue sizes unless the CQ is resized */
2692 	qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2693 	qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2694 	return 0;
2695 
2696 error:
2697 	return ret;
2698 }
2699 
2700 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2701 {
2702 	ib_destroy_qp(qp_info->qp);
2703 	kfree(qp_info->snoop_table);
2704 }
2705 
2706 /*
2707  * Open the port
2708  * Create the QP, PD, MR, and CQ if needed
2709  */
2710 static int ib_mad_port_open(struct ib_device *device,
2711 			    int port_num)
2712 {
2713 	int ret, cq_size;
2714 	struct ib_mad_port_private *port_priv;
2715 	unsigned long flags;
2716 	char name[sizeof "ib_mad123"];
2717 
2718 	/* Create new device info */
2719 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2720 	if (!port_priv) {
2721 		printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2722 		return -ENOMEM;
2723 	}
2724 
2725 	port_priv->device = device;
2726 	port_priv->port_num = port_num;
2727 	spin_lock_init(&port_priv->reg_lock);
2728 	INIT_LIST_HEAD(&port_priv->agent_list);
2729 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2730 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2731 
2732 	cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2733 	port_priv->cq = ib_create_cq(port_priv->device,
2734 				     ib_mad_thread_completion_handler,
2735 				     NULL, port_priv, cq_size);
2736 	if (IS_ERR(port_priv->cq)) {
2737 		printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2738 		ret = PTR_ERR(port_priv->cq);
2739 		goto error3;
2740 	}
2741 
2742 	port_priv->pd = ib_alloc_pd(device);
2743 	if (IS_ERR(port_priv->pd)) {
2744 		printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2745 		ret = PTR_ERR(port_priv->pd);
2746 		goto error4;
2747 	}
2748 
2749 	port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2750 	if (IS_ERR(port_priv->mr)) {
2751 		printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2752 		ret = PTR_ERR(port_priv->mr);
2753 		goto error5;
2754 	}
2755 
2756 	ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2757 	if (ret)
2758 		goto error6;
2759 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2760 	if (ret)
2761 		goto error7;
2762 
2763 	snprintf(name, sizeof name, "ib_mad%d", port_num);
2764 	port_priv->wq = create_singlethread_workqueue(name);
2765 	if (!port_priv->wq) {
2766 		ret = -ENOMEM;
2767 		goto error8;
2768 	}
2769 	INIT_WORK(&port_priv->work, ib_mad_completion_handler, port_priv);
2770 
2771 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2772 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2773 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2774 
2775 	ret = ib_mad_port_start(port_priv);
2776 	if (ret) {
2777 		printk(KERN_ERR PFX "Couldn't start port\n");
2778 		goto error9;
2779 	}
2780 
2781 	return 0;
2782 
2783 error9:
2784 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2785 	list_del_init(&port_priv->port_list);
2786 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2787 
2788 	destroy_workqueue(port_priv->wq);
2789 error8:
2790 	destroy_mad_qp(&port_priv->qp_info[1]);
2791 error7:
2792 	destroy_mad_qp(&port_priv->qp_info[0]);
2793 error6:
2794 	ib_dereg_mr(port_priv->mr);
2795 error5:
2796 	ib_dealloc_pd(port_priv->pd);
2797 error4:
2798 	ib_destroy_cq(port_priv->cq);
2799 	cleanup_recv_queue(&port_priv->qp_info[1]);
2800 	cleanup_recv_queue(&port_priv->qp_info[0]);
2801 error3:
2802 	kfree(port_priv);
2803 
2804 	return ret;
2805 }
2806 
2807 /*
2808  * Close the port
2809  * If there are no classes using the port, free the port
2810  * resources (CQ, MR, PD, QP) and remove the port's info structure
2811  */
2812 static int ib_mad_port_close(struct ib_device *device, int port_num)
2813 {
2814 	struct ib_mad_port_private *port_priv;
2815 	unsigned long flags;
2816 
2817 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2818 	port_priv = __ib_get_mad_port(device, port_num);
2819 	if (port_priv == NULL) {
2820 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2821 		printk(KERN_ERR PFX "Port %d not found\n", port_num);
2822 		return -ENODEV;
2823 	}
2824 	list_del_init(&port_priv->port_list);
2825 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2826 
2827 	destroy_workqueue(port_priv->wq);
2828 	destroy_mad_qp(&port_priv->qp_info[1]);
2829 	destroy_mad_qp(&port_priv->qp_info[0]);
2830 	ib_dereg_mr(port_priv->mr);
2831 	ib_dealloc_pd(port_priv->pd);
2832 	ib_destroy_cq(port_priv->cq);
2833 	cleanup_recv_queue(&port_priv->qp_info[1]);
2834 	cleanup_recv_queue(&port_priv->qp_info[0]);
2835 	/* XXX: Handle deallocation of MAD registration tables */
2836 
2837 	kfree(port_priv);
2838 
2839 	return 0;
2840 }
2841 
2842 static void ib_mad_init_device(struct ib_device *device)
2843 {
2844 	int start, end, i;
2845 
2846 	if (device->node_type == IB_NODE_SWITCH) {
2847 		start = 0;
2848 		end   = 0;
2849 	} else {
2850 		start = 1;
2851 		end   = device->phys_port_cnt;
2852 	}
2853 
2854 	for (i = start; i <= end; i++) {
2855 		if (ib_mad_port_open(device, i)) {
2856 			printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2857 			       device->name, i);
2858 			goto error;
2859 		}
2860 		if (ib_agent_port_open(device, i)) {
2861 			printk(KERN_ERR PFX "Couldn't open %s port %d "
2862 			       "for agents\n",
2863 			       device->name, i);
2864 			goto error_agent;
2865 		}
2866 	}
2867 	return;
2868 
2869 error_agent:
2870 	if (ib_mad_port_close(device, i))
2871 		printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2872 		       device->name, i);
2873 
2874 error:
2875 	i--;
2876 
2877 	while (i >= start) {
2878 		if (ib_agent_port_close(device, i))
2879 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2880 			       "for agents\n",
2881 			       device->name, i);
2882 		if (ib_mad_port_close(device, i))
2883 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2884 			       device->name, i);
2885 		i--;
2886 	}
2887 }
2888 
2889 static void ib_mad_remove_device(struct ib_device *device)
2890 {
2891 	int i, num_ports, cur_port;
2892 
2893 	if (device->node_type == IB_NODE_SWITCH) {
2894 		num_ports = 1;
2895 		cur_port = 0;
2896 	} else {
2897 		num_ports = device->phys_port_cnt;
2898 		cur_port = 1;
2899 	}
2900 	for (i = 0; i < num_ports; i++, cur_port++) {
2901 		if (ib_agent_port_close(device, cur_port))
2902 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2903 			       "for agents\n",
2904 			       device->name, cur_port);
2905 		if (ib_mad_port_close(device, cur_port))
2906 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2907 			       device->name, cur_port);
2908 	}
2909 }
2910 
2911 static struct ib_client mad_client = {
2912 	.name   = "mad",
2913 	.add = ib_mad_init_device,
2914 	.remove = ib_mad_remove_device
2915 };
2916 
2917 static int __init ib_mad_init_module(void)
2918 {
2919 	int ret;
2920 
2921 	spin_lock_init(&ib_mad_port_list_lock);
2922 
2923 	ib_mad_cache = kmem_cache_create("ib_mad",
2924 					 sizeof(struct ib_mad_private),
2925 					 0,
2926 					 SLAB_HWCACHE_ALIGN,
2927 					 NULL,
2928 					 NULL);
2929 	if (!ib_mad_cache) {
2930 		printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2931 		ret = -ENOMEM;
2932 		goto error1;
2933 	}
2934 
2935 	INIT_LIST_HEAD(&ib_mad_port_list);
2936 
2937 	if (ib_register_client(&mad_client)) {
2938 		printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2939 		ret = -EINVAL;
2940 		goto error2;
2941 	}
2942 
2943 	return 0;
2944 
2945 error2:
2946 	kmem_cache_destroy(ib_mad_cache);
2947 error1:
2948 	return ret;
2949 }
2950 
2951 static void __exit ib_mad_cleanup_module(void)
2952 {
2953 	ib_unregister_client(&mad_client);
2954 
2955 	if (kmem_cache_destroy(ib_mad_cache)) {
2956 		printk(KERN_DEBUG PFX "Failed to destroy ib_mad cache\n");
2957 	}
2958 }
2959 
2960 module_init(ib_mad_init_module);
2961 module_exit(ib_mad_cleanup_module);
2962 
2963