xref: /linux/drivers/infiniband/core/mad.c (revision b0148a98ec5151fec82064d95f11eb9efbc628ea)
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 #include <rdma/ib_cache.h>
38 
39 #include "mad_priv.h"
40 #include "mad_rmpp.h"
41 #include "smi.h"
42 #include "agent.h"
43 
44 MODULE_LICENSE("Dual BSD/GPL");
45 MODULE_DESCRIPTION("kernel IB MAD API");
46 MODULE_AUTHOR("Hal Rosenstock");
47 MODULE_AUTHOR("Sean Hefty");
48 
49 static struct kmem_cache *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(struct work_struct *work);
69 static void local_completions(struct work_struct *work);
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 int ib_response_mad(struct ib_mad *mad)
171 {
172 	return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
173 		(mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
174 		((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
175 		 (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
176 }
177 EXPORT_SYMBOL(ib_response_mad);
178 
179 /*
180  * ib_register_mad_agent - Register to send/receive MADs
181  */
182 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
183 					   u8 port_num,
184 					   enum ib_qp_type qp_type,
185 					   struct ib_mad_reg_req *mad_reg_req,
186 					   u8 rmpp_version,
187 					   ib_mad_send_handler send_handler,
188 					   ib_mad_recv_handler recv_handler,
189 					   void *context)
190 {
191 	struct ib_mad_port_private *port_priv;
192 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
193 	struct ib_mad_agent_private *mad_agent_priv;
194 	struct ib_mad_reg_req *reg_req = NULL;
195 	struct ib_mad_mgmt_class_table *class;
196 	struct ib_mad_mgmt_vendor_class_table *vendor;
197 	struct ib_mad_mgmt_vendor_class *vendor_class;
198 	struct ib_mad_mgmt_method_table *method;
199 	int ret2, qpn;
200 	unsigned long flags;
201 	u8 mgmt_class, vclass;
202 
203 	/* Validate parameters */
204 	qpn = get_spl_qp_index(qp_type);
205 	if (qpn == -1)
206 		goto error1;
207 
208 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
209 		goto error1;
210 
211 	/* Validate MAD registration request if supplied */
212 	if (mad_reg_req) {
213 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
214 			goto error1;
215 		if (!recv_handler)
216 			goto error1;
217 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
218 			/*
219 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
220 			 * one in this range currently allowed
221 			 */
222 			if (mad_reg_req->mgmt_class !=
223 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
224 				goto error1;
225 		} else if (mad_reg_req->mgmt_class == 0) {
226 			/*
227 			 * Class 0 is reserved in IBA and is used for
228 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
229 			 */
230 			goto error1;
231 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
232 			/*
233 			 * If class is in "new" vendor range,
234 			 * ensure supplied OUI is not zero
235 			 */
236 			if (!is_vendor_oui(mad_reg_req->oui))
237 				goto error1;
238 		}
239 		/* Make sure class supplied is consistent with RMPP */
240 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
241 			if (rmpp_version)
242 				goto error1;
243 		}
244 		/* Make sure class supplied is consistent with QP type */
245 		if (qp_type == IB_QPT_SMI) {
246 			if ((mad_reg_req->mgmt_class !=
247 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
248 			    (mad_reg_req->mgmt_class !=
249 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
250 				goto error1;
251 		} else {
252 			if ((mad_reg_req->mgmt_class ==
253 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
254 			    (mad_reg_req->mgmt_class ==
255 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
256 				goto error1;
257 		}
258 	} else {
259 		/* No registration request supplied */
260 		if (!send_handler)
261 			goto error1;
262 	}
263 
264 	/* Validate device and port */
265 	port_priv = ib_get_mad_port(device, port_num);
266 	if (!port_priv) {
267 		ret = ERR_PTR(-ENODEV);
268 		goto error1;
269 	}
270 
271 	/* Allocate structures */
272 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
273 	if (!mad_agent_priv) {
274 		ret = ERR_PTR(-ENOMEM);
275 		goto error1;
276 	}
277 
278 	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
279 						 IB_ACCESS_LOCAL_WRITE);
280 	if (IS_ERR(mad_agent_priv->agent.mr)) {
281 		ret = ERR_PTR(-ENOMEM);
282 		goto error2;
283 	}
284 
285 	if (mad_reg_req) {
286 		reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
287 		if (!reg_req) {
288 			ret = ERR_PTR(-ENOMEM);
289 			goto error3;
290 		}
291 		/* Make a copy of the MAD registration request */
292 		memcpy(reg_req, mad_reg_req, sizeof *reg_req);
293 	}
294 
295 	/* Now, fill in the various structures */
296 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
297 	mad_agent_priv->reg_req = reg_req;
298 	mad_agent_priv->agent.rmpp_version = rmpp_version;
299 	mad_agent_priv->agent.device = device;
300 	mad_agent_priv->agent.recv_handler = recv_handler;
301 	mad_agent_priv->agent.send_handler = send_handler;
302 	mad_agent_priv->agent.context = context;
303 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
304 	mad_agent_priv->agent.port_num = port_num;
305 
306 	spin_lock_irqsave(&port_priv->reg_lock, flags);
307 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
308 
309 	/*
310 	 * Make sure MAD registration (if supplied)
311 	 * is non overlapping with any existing ones
312 	 */
313 	if (mad_reg_req) {
314 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
315 		if (!is_vendor_class(mgmt_class)) {
316 			class = port_priv->version[mad_reg_req->
317 						   mgmt_class_version].class;
318 			if (class) {
319 				method = class->method_table[mgmt_class];
320 				if (method) {
321 					if (method_in_use(&method,
322 							   mad_reg_req))
323 						goto error4;
324 				}
325 			}
326 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
327 						  mgmt_class);
328 		} else {
329 			/* "New" vendor class range */
330 			vendor = port_priv->version[mad_reg_req->
331 						    mgmt_class_version].vendor;
332 			if (vendor) {
333 				vclass = vendor_class_index(mgmt_class);
334 				vendor_class = vendor->vendor_class[vclass];
335 				if (vendor_class) {
336 					if (is_vendor_method_in_use(
337 							vendor_class,
338 							mad_reg_req))
339 						goto error4;
340 				}
341 			}
342 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
343 		}
344 		if (ret2) {
345 			ret = ERR_PTR(ret2);
346 			goto error4;
347 		}
348 	}
349 
350 	/* Add mad agent into port's agent list */
351 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
352 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
353 
354 	spin_lock_init(&mad_agent_priv->lock);
355 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
356 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
357 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
358 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
359 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
360 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
361 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
362 	atomic_set(&mad_agent_priv->refcount, 1);
363 	init_completion(&mad_agent_priv->comp);
364 
365 	return &mad_agent_priv->agent;
366 
367 error4:
368 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
369 	kfree(reg_req);
370 error3:
371 	ib_dereg_mr(mad_agent_priv->agent.mr);
372 error2:
373 	kfree(mad_agent_priv);
374 error1:
375 	return ret;
376 }
377 EXPORT_SYMBOL(ib_register_mad_agent);
378 
379 static inline int is_snooping_sends(int mad_snoop_flags)
380 {
381 	return (mad_snoop_flags &
382 		(/*IB_MAD_SNOOP_POSTED_SENDS |
383 		 IB_MAD_SNOOP_RMPP_SENDS |*/
384 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
385 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
386 }
387 
388 static inline int is_snooping_recvs(int mad_snoop_flags)
389 {
390 	return (mad_snoop_flags &
391 		(IB_MAD_SNOOP_RECVS /*|
392 		 IB_MAD_SNOOP_RMPP_RECVS*/));
393 }
394 
395 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
396 				struct ib_mad_snoop_private *mad_snoop_priv)
397 {
398 	struct ib_mad_snoop_private **new_snoop_table;
399 	unsigned long flags;
400 	int i;
401 
402 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
403 	/* Check for empty slot in array. */
404 	for (i = 0; i < qp_info->snoop_table_size; i++)
405 		if (!qp_info->snoop_table[i])
406 			break;
407 
408 	if (i == qp_info->snoop_table_size) {
409 		/* Grow table. */
410 		new_snoop_table = kmalloc(sizeof mad_snoop_priv *
411 					  qp_info->snoop_table_size + 1,
412 					  GFP_ATOMIC);
413 		if (!new_snoop_table) {
414 			i = -ENOMEM;
415 			goto out;
416 		}
417 		if (qp_info->snoop_table) {
418 			memcpy(new_snoop_table, qp_info->snoop_table,
419 			       sizeof mad_snoop_priv *
420 			       qp_info->snoop_table_size);
421 			kfree(qp_info->snoop_table);
422 		}
423 		qp_info->snoop_table = new_snoop_table;
424 		qp_info->snoop_table_size++;
425 	}
426 	qp_info->snoop_table[i] = mad_snoop_priv;
427 	atomic_inc(&qp_info->snoop_count);
428 out:
429 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
430 	return i;
431 }
432 
433 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
434 					   u8 port_num,
435 					   enum ib_qp_type qp_type,
436 					   int mad_snoop_flags,
437 					   ib_mad_snoop_handler snoop_handler,
438 					   ib_mad_recv_handler recv_handler,
439 					   void *context)
440 {
441 	struct ib_mad_port_private *port_priv;
442 	struct ib_mad_agent *ret;
443 	struct ib_mad_snoop_private *mad_snoop_priv;
444 	int qpn;
445 
446 	/* Validate parameters */
447 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
448 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
449 		ret = ERR_PTR(-EINVAL);
450 		goto error1;
451 	}
452 	qpn = get_spl_qp_index(qp_type);
453 	if (qpn == -1) {
454 		ret = ERR_PTR(-EINVAL);
455 		goto error1;
456 	}
457 	port_priv = ib_get_mad_port(device, port_num);
458 	if (!port_priv) {
459 		ret = ERR_PTR(-ENODEV);
460 		goto error1;
461 	}
462 	/* Allocate structures */
463 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
464 	if (!mad_snoop_priv) {
465 		ret = ERR_PTR(-ENOMEM);
466 		goto error1;
467 	}
468 
469 	/* Now, fill in the various structures */
470 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
471 	mad_snoop_priv->agent.device = device;
472 	mad_snoop_priv->agent.recv_handler = recv_handler;
473 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
474 	mad_snoop_priv->agent.context = context;
475 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
476 	mad_snoop_priv->agent.port_num = port_num;
477 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
478 	init_completion(&mad_snoop_priv->comp);
479 	mad_snoop_priv->snoop_index = register_snoop_agent(
480 						&port_priv->qp_info[qpn],
481 						mad_snoop_priv);
482 	if (mad_snoop_priv->snoop_index < 0) {
483 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
484 		goto error2;
485 	}
486 
487 	atomic_set(&mad_snoop_priv->refcount, 1);
488 	return &mad_snoop_priv->agent;
489 
490 error2:
491 	kfree(mad_snoop_priv);
492 error1:
493 	return ret;
494 }
495 EXPORT_SYMBOL(ib_register_mad_snoop);
496 
497 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
498 {
499 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
500 		complete(&mad_agent_priv->comp);
501 }
502 
503 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
504 {
505 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
506 		complete(&mad_snoop_priv->comp);
507 }
508 
509 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
510 {
511 	struct ib_mad_port_private *port_priv;
512 	unsigned long flags;
513 
514 	/* Note that we could still be handling received MADs */
515 
516 	/*
517 	 * Canceling all sends results in dropping received response
518 	 * MADs, preventing us from queuing additional work
519 	 */
520 	cancel_mads(mad_agent_priv);
521 	port_priv = mad_agent_priv->qp_info->port_priv;
522 	cancel_delayed_work(&mad_agent_priv->timed_work);
523 
524 	spin_lock_irqsave(&port_priv->reg_lock, flags);
525 	remove_mad_reg_req(mad_agent_priv);
526 	list_del(&mad_agent_priv->agent_list);
527 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
528 
529 	flush_workqueue(port_priv->wq);
530 	ib_cancel_rmpp_recvs(mad_agent_priv);
531 
532 	deref_mad_agent(mad_agent_priv);
533 	wait_for_completion(&mad_agent_priv->comp);
534 
535 	kfree(mad_agent_priv->reg_req);
536 	ib_dereg_mr(mad_agent_priv->agent.mr);
537 	kfree(mad_agent_priv);
538 }
539 
540 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
541 {
542 	struct ib_mad_qp_info *qp_info;
543 	unsigned long flags;
544 
545 	qp_info = mad_snoop_priv->qp_info;
546 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
547 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
548 	atomic_dec(&qp_info->snoop_count);
549 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
550 
551 	deref_snoop_agent(mad_snoop_priv);
552 	wait_for_completion(&mad_snoop_priv->comp);
553 
554 	kfree(mad_snoop_priv);
555 }
556 
557 /*
558  * ib_unregister_mad_agent - Unregisters a client from using MAD services
559  */
560 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
561 {
562 	struct ib_mad_agent_private *mad_agent_priv;
563 	struct ib_mad_snoop_private *mad_snoop_priv;
564 
565 	/* If the TID is zero, the agent can only snoop. */
566 	if (mad_agent->hi_tid) {
567 		mad_agent_priv = container_of(mad_agent,
568 					      struct ib_mad_agent_private,
569 					      agent);
570 		unregister_mad_agent(mad_agent_priv);
571 	} else {
572 		mad_snoop_priv = container_of(mad_agent,
573 					      struct ib_mad_snoop_private,
574 					      agent);
575 		unregister_mad_snoop(mad_snoop_priv);
576 	}
577 	return 0;
578 }
579 EXPORT_SYMBOL(ib_unregister_mad_agent);
580 
581 static void dequeue_mad(struct ib_mad_list_head *mad_list)
582 {
583 	struct ib_mad_queue *mad_queue;
584 	unsigned long flags;
585 
586 	BUG_ON(!mad_list->mad_queue);
587 	mad_queue = mad_list->mad_queue;
588 	spin_lock_irqsave(&mad_queue->lock, flags);
589 	list_del(&mad_list->list);
590 	mad_queue->count--;
591 	spin_unlock_irqrestore(&mad_queue->lock, flags);
592 }
593 
594 static void snoop_send(struct ib_mad_qp_info *qp_info,
595 		       struct ib_mad_send_buf *send_buf,
596 		       struct ib_mad_send_wc *mad_send_wc,
597 		       int mad_snoop_flags)
598 {
599 	struct ib_mad_snoop_private *mad_snoop_priv;
600 	unsigned long flags;
601 	int i;
602 
603 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
604 	for (i = 0; i < qp_info->snoop_table_size; i++) {
605 		mad_snoop_priv = qp_info->snoop_table[i];
606 		if (!mad_snoop_priv ||
607 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
608 			continue;
609 
610 		atomic_inc(&mad_snoop_priv->refcount);
611 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
612 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
613 						    send_buf, mad_send_wc);
614 		deref_snoop_agent(mad_snoop_priv);
615 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
616 	}
617 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
618 }
619 
620 static void snoop_recv(struct ib_mad_qp_info *qp_info,
621 		       struct ib_mad_recv_wc *mad_recv_wc,
622 		       int mad_snoop_flags)
623 {
624 	struct ib_mad_snoop_private *mad_snoop_priv;
625 	unsigned long flags;
626 	int i;
627 
628 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
629 	for (i = 0; i < qp_info->snoop_table_size; i++) {
630 		mad_snoop_priv = qp_info->snoop_table[i];
631 		if (!mad_snoop_priv ||
632 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
633 			continue;
634 
635 		atomic_inc(&mad_snoop_priv->refcount);
636 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
637 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
638 						   mad_recv_wc);
639 		deref_snoop_agent(mad_snoop_priv);
640 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
641 	}
642 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
643 }
644 
645 static void build_smp_wc(struct ib_qp *qp,
646 			 u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
647 			 struct ib_wc *wc)
648 {
649 	memset(wc, 0, sizeof *wc);
650 	wc->wr_id = wr_id;
651 	wc->status = IB_WC_SUCCESS;
652 	wc->opcode = IB_WC_RECV;
653 	wc->pkey_index = pkey_index;
654 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
655 	wc->src_qp = IB_QP0;
656 	wc->qp = qp;
657 	wc->slid = slid;
658 	wc->sl = 0;
659 	wc->dlid_path_bits = 0;
660 	wc->port_num = port_num;
661 }
662 
663 /*
664  * Return 0 if SMP is to be sent
665  * Return 1 if SMP was consumed locally (whether or not solicited)
666  * Return < 0 if error
667  */
668 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
669 				  struct ib_mad_send_wr_private *mad_send_wr)
670 {
671 	int ret;
672 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
673 	unsigned long flags;
674 	struct ib_mad_local_private *local;
675 	struct ib_mad_private *mad_priv;
676 	struct ib_mad_port_private *port_priv;
677 	struct ib_mad_agent_private *recv_mad_agent = NULL;
678 	struct ib_device *device = mad_agent_priv->agent.device;
679 	u8 port_num = mad_agent_priv->agent.port_num;
680 	struct ib_wc mad_wc;
681 	struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
682 
683 	/*
684 	 * Directed route handling starts if the initial LID routed part of
685 	 * a request or the ending LID routed part of a response is empty.
686 	 * If we are at the start of the LID routed part, don't update the
687 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
688 	 */
689 	if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
690 	     IB_LID_PERMISSIVE &&
691 	    !smi_handle_dr_smp_send(smp, device->node_type, port_num)) {
692 		ret = -EINVAL;
693 		printk(KERN_ERR PFX "Invalid directed route\n");
694 		goto out;
695 	}
696 	/* Check to post send on QP or process locally */
697 	ret = smi_check_local_smp(smp, device);
698 	if (!ret)
699 		goto out;
700 
701 	local = kmalloc(sizeof *local, GFP_ATOMIC);
702 	if (!local) {
703 		ret = -ENOMEM;
704 		printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
705 		goto out;
706 	}
707 	local->mad_priv = NULL;
708 	local->recv_mad_agent = NULL;
709 	mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
710 	if (!mad_priv) {
711 		ret = -ENOMEM;
712 		printk(KERN_ERR PFX "No memory for local response MAD\n");
713 		kfree(local);
714 		goto out;
715 	}
716 
717 	build_smp_wc(mad_agent_priv->agent.qp,
718 		     send_wr->wr_id, be16_to_cpu(smp->dr_slid),
719 		     send_wr->wr.ud.pkey_index,
720 		     send_wr->wr.ud.port_num, &mad_wc);
721 
722 	/* No GRH for DR SMP */
723 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
724 				  (struct ib_mad *)smp,
725 				  (struct ib_mad *)&mad_priv->mad);
726 	switch (ret)
727 	{
728 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
729 		if (ib_response_mad(&mad_priv->mad.mad) &&
730 		    mad_agent_priv->agent.recv_handler) {
731 			local->mad_priv = mad_priv;
732 			local->recv_mad_agent = mad_agent_priv;
733 			/*
734 			 * Reference MAD agent until receive
735 			 * side of local completion handled
736 			 */
737 			atomic_inc(&mad_agent_priv->refcount);
738 		} else
739 			kmem_cache_free(ib_mad_cache, mad_priv);
740 		break;
741 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
742 		kmem_cache_free(ib_mad_cache, mad_priv);
743 		break;
744 	case IB_MAD_RESULT_SUCCESS:
745 		/* Treat like an incoming receive MAD */
746 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
747 					    mad_agent_priv->agent.port_num);
748 		if (port_priv) {
749 			mad_priv->mad.mad.mad_hdr.tid =
750 				((struct ib_mad *)smp)->mad_hdr.tid;
751 			recv_mad_agent = find_mad_agent(port_priv,
752 						        &mad_priv->mad.mad);
753 		}
754 		if (!port_priv || !recv_mad_agent) {
755 			kmem_cache_free(ib_mad_cache, mad_priv);
756 			kfree(local);
757 			ret = 0;
758 			goto out;
759 		}
760 		local->mad_priv = mad_priv;
761 		local->recv_mad_agent = recv_mad_agent;
762 		break;
763 	default:
764 		kmem_cache_free(ib_mad_cache, mad_priv);
765 		kfree(local);
766 		ret = -EINVAL;
767 		goto out;
768 	}
769 
770 	local->mad_send_wr = mad_send_wr;
771 	/* Reference MAD agent until send side of local completion handled */
772 	atomic_inc(&mad_agent_priv->refcount);
773 	/* Queue local completion to local list */
774 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
775 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
776 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
777 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
778 		   &mad_agent_priv->local_work);
779 	ret = 1;
780 out:
781 	return ret;
782 }
783 
784 static int get_pad_size(int hdr_len, int data_len)
785 {
786 	int seg_size, pad;
787 
788 	seg_size = sizeof(struct ib_mad) - hdr_len;
789 	if (data_len && seg_size) {
790 		pad = seg_size - data_len % seg_size;
791 		return pad == seg_size ? 0 : pad;
792 	} else
793 		return seg_size;
794 }
795 
796 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
797 {
798 	struct ib_rmpp_segment *s, *t;
799 
800 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
801 		list_del(&s->list);
802 		kfree(s);
803 	}
804 }
805 
806 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
807 				gfp_t gfp_mask)
808 {
809 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
810 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
811 	struct ib_rmpp_segment *seg = NULL;
812 	int left, seg_size, pad;
813 
814 	send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
815 	seg_size = send_buf->seg_size;
816 	pad = send_wr->pad;
817 
818 	/* Allocate data segments. */
819 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
820 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
821 		if (!seg) {
822 			printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
823 			       "alloc failed for len %zd, gfp %#x\n",
824 			       sizeof (*seg) + seg_size, gfp_mask);
825 			free_send_rmpp_list(send_wr);
826 			return -ENOMEM;
827 		}
828 		seg->num = ++send_buf->seg_count;
829 		list_add_tail(&seg->list, &send_wr->rmpp_list);
830 	}
831 
832 	/* Zero any padding */
833 	if (pad)
834 		memset(seg->data + seg_size - pad, 0, pad);
835 
836 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
837 					  agent.rmpp_version;
838 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
839 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
840 
841 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
842 					struct ib_rmpp_segment, list);
843 	send_wr->last_ack_seg = send_wr->cur_seg;
844 	return 0;
845 }
846 
847 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
848 					    u32 remote_qpn, u16 pkey_index,
849 					    int rmpp_active,
850 					    int hdr_len, int data_len,
851 					    gfp_t gfp_mask)
852 {
853 	struct ib_mad_agent_private *mad_agent_priv;
854 	struct ib_mad_send_wr_private *mad_send_wr;
855 	int pad, message_size, ret, size;
856 	void *buf;
857 
858 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
859 				      agent);
860 	pad = get_pad_size(hdr_len, data_len);
861 	message_size = hdr_len + data_len + pad;
862 
863 	if ((!mad_agent->rmpp_version &&
864 	     (rmpp_active || message_size > sizeof(struct ib_mad))) ||
865 	    (!rmpp_active && message_size > sizeof(struct ib_mad)))
866 		return ERR_PTR(-EINVAL);
867 
868 	size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
869 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
870 	if (!buf)
871 		return ERR_PTR(-ENOMEM);
872 
873 	mad_send_wr = buf + size;
874 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
875 	mad_send_wr->send_buf.mad = buf;
876 	mad_send_wr->send_buf.hdr_len = hdr_len;
877 	mad_send_wr->send_buf.data_len = data_len;
878 	mad_send_wr->pad = pad;
879 
880 	mad_send_wr->mad_agent_priv = mad_agent_priv;
881 	mad_send_wr->sg_list[0].length = hdr_len;
882 	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
883 	mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
884 	mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
885 
886 	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
887 	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
888 	mad_send_wr->send_wr.num_sge = 2;
889 	mad_send_wr->send_wr.opcode = IB_WR_SEND;
890 	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
891 	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
892 	mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
893 	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
894 
895 	if (rmpp_active) {
896 		ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
897 		if (ret) {
898 			kfree(buf);
899 			return ERR_PTR(ret);
900 		}
901 	}
902 
903 	mad_send_wr->send_buf.mad_agent = mad_agent;
904 	atomic_inc(&mad_agent_priv->refcount);
905 	return &mad_send_wr->send_buf;
906 }
907 EXPORT_SYMBOL(ib_create_send_mad);
908 
909 int ib_get_mad_data_offset(u8 mgmt_class)
910 {
911 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
912 		return IB_MGMT_SA_HDR;
913 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
914 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
915 		 (mgmt_class == IB_MGMT_CLASS_BIS))
916 		return IB_MGMT_DEVICE_HDR;
917 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
918 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
919 		return IB_MGMT_VENDOR_HDR;
920 	else
921 		return IB_MGMT_MAD_HDR;
922 }
923 EXPORT_SYMBOL(ib_get_mad_data_offset);
924 
925 int ib_is_mad_class_rmpp(u8 mgmt_class)
926 {
927 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
928 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
929 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
930 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
931 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
932 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
933 		return 1;
934 	return 0;
935 }
936 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
937 
938 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
939 {
940 	struct ib_mad_send_wr_private *mad_send_wr;
941 	struct list_head *list;
942 
943 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
944 				   send_buf);
945 	list = &mad_send_wr->cur_seg->list;
946 
947 	if (mad_send_wr->cur_seg->num < seg_num) {
948 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
949 			if (mad_send_wr->cur_seg->num == seg_num)
950 				break;
951 	} else if (mad_send_wr->cur_seg->num > seg_num) {
952 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
953 			if (mad_send_wr->cur_seg->num == seg_num)
954 				break;
955 	}
956 	return mad_send_wr->cur_seg->data;
957 }
958 EXPORT_SYMBOL(ib_get_rmpp_segment);
959 
960 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
961 {
962 	if (mad_send_wr->send_buf.seg_count)
963 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
964 					   mad_send_wr->seg_num);
965 	else
966 		return mad_send_wr->send_buf.mad +
967 		       mad_send_wr->send_buf.hdr_len;
968 }
969 
970 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
971 {
972 	struct ib_mad_agent_private *mad_agent_priv;
973 	struct ib_mad_send_wr_private *mad_send_wr;
974 
975 	mad_agent_priv = container_of(send_buf->mad_agent,
976 				      struct ib_mad_agent_private, agent);
977 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
978 				   send_buf);
979 
980 	free_send_rmpp_list(mad_send_wr);
981 	kfree(send_buf->mad);
982 	deref_mad_agent(mad_agent_priv);
983 }
984 EXPORT_SYMBOL(ib_free_send_mad);
985 
986 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
987 {
988 	struct ib_mad_qp_info *qp_info;
989 	struct list_head *list;
990 	struct ib_send_wr *bad_send_wr;
991 	struct ib_mad_agent *mad_agent;
992 	struct ib_sge *sge;
993 	unsigned long flags;
994 	int ret;
995 
996 	/* Set WR ID to find mad_send_wr upon completion */
997 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
998 	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
999 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1000 
1001 	mad_agent = mad_send_wr->send_buf.mad_agent;
1002 	sge = mad_send_wr->sg_list;
1003 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1004 					mad_send_wr->send_buf.mad,
1005 					sge[0].length,
1006 					DMA_TO_DEVICE);
1007 	mad_send_wr->header_mapping = sge[0].addr;
1008 
1009 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1010 					ib_get_payload(mad_send_wr),
1011 					sge[1].length,
1012 					DMA_TO_DEVICE);
1013 	mad_send_wr->payload_mapping = sge[1].addr;
1014 
1015 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1016 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1017 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1018 				   &bad_send_wr);
1019 		list = &qp_info->send_queue.list;
1020 	} else {
1021 		ret = 0;
1022 		list = &qp_info->overflow_list;
1023 	}
1024 
1025 	if (!ret) {
1026 		qp_info->send_queue.count++;
1027 		list_add_tail(&mad_send_wr->mad_list.list, list);
1028 	}
1029 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1030 	if (ret) {
1031 		ib_dma_unmap_single(mad_agent->device,
1032 				    mad_send_wr->header_mapping,
1033 				    sge[0].length, DMA_TO_DEVICE);
1034 		ib_dma_unmap_single(mad_agent->device,
1035 				    mad_send_wr->payload_mapping,
1036 				    sge[1].length, DMA_TO_DEVICE);
1037 	}
1038 	return ret;
1039 }
1040 
1041 /*
1042  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1043  *  with the registered client
1044  */
1045 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1046 		     struct ib_mad_send_buf **bad_send_buf)
1047 {
1048 	struct ib_mad_agent_private *mad_agent_priv;
1049 	struct ib_mad_send_buf *next_send_buf;
1050 	struct ib_mad_send_wr_private *mad_send_wr;
1051 	unsigned long flags;
1052 	int ret = -EINVAL;
1053 
1054 	/* Walk list of send WRs and post each on send list */
1055 	for (; send_buf; send_buf = next_send_buf) {
1056 
1057 		mad_send_wr = container_of(send_buf,
1058 					   struct ib_mad_send_wr_private,
1059 					   send_buf);
1060 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1061 
1062 		if (!send_buf->mad_agent->send_handler ||
1063 		    (send_buf->timeout_ms &&
1064 		     !send_buf->mad_agent->recv_handler)) {
1065 			ret = -EINVAL;
1066 			goto error;
1067 		}
1068 
1069 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1070 			if (mad_agent_priv->agent.rmpp_version) {
1071 				ret = -EINVAL;
1072 				goto error;
1073 			}
1074 		}
1075 
1076 		/*
1077 		 * Save pointer to next work request to post in case the
1078 		 * current one completes, and the user modifies the work
1079 		 * request associated with the completion
1080 		 */
1081 		next_send_buf = send_buf->next;
1082 		mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1083 
1084 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1085 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1086 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1087 						     mad_send_wr);
1088 			if (ret < 0)		/* error */
1089 				goto error;
1090 			else if (ret == 1)	/* locally consumed */
1091 				continue;
1092 		}
1093 
1094 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1095 		/* Timeout will be updated after send completes */
1096 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1097 		mad_send_wr->retries = send_buf->retries;
1098 		/* Reference for work request to QP + response */
1099 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1100 		mad_send_wr->status = IB_WC_SUCCESS;
1101 
1102 		/* Reference MAD agent until send completes */
1103 		atomic_inc(&mad_agent_priv->refcount);
1104 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1105 		list_add_tail(&mad_send_wr->agent_list,
1106 			      &mad_agent_priv->send_list);
1107 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1108 
1109 		if (mad_agent_priv->agent.rmpp_version) {
1110 			ret = ib_send_rmpp_mad(mad_send_wr);
1111 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1112 				ret = ib_send_mad(mad_send_wr);
1113 		} else
1114 			ret = ib_send_mad(mad_send_wr);
1115 		if (ret < 0) {
1116 			/* Fail send request */
1117 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1118 			list_del(&mad_send_wr->agent_list);
1119 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1120 			atomic_dec(&mad_agent_priv->refcount);
1121 			goto error;
1122 		}
1123 	}
1124 	return 0;
1125 error:
1126 	if (bad_send_buf)
1127 		*bad_send_buf = send_buf;
1128 	return ret;
1129 }
1130 EXPORT_SYMBOL(ib_post_send_mad);
1131 
1132 /*
1133  * ib_free_recv_mad - Returns data buffers used to receive
1134  *  a MAD to the access layer
1135  */
1136 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1137 {
1138 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1139 	struct ib_mad_private_header *mad_priv_hdr;
1140 	struct ib_mad_private *priv;
1141 	struct list_head free_list;
1142 
1143 	INIT_LIST_HEAD(&free_list);
1144 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1145 
1146 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1147 					&free_list, list) {
1148 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1149 					   recv_buf);
1150 		mad_priv_hdr = container_of(mad_recv_wc,
1151 					    struct ib_mad_private_header,
1152 					    recv_wc);
1153 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1154 				    header);
1155 		kmem_cache_free(ib_mad_cache, priv);
1156 	}
1157 }
1158 EXPORT_SYMBOL(ib_free_recv_mad);
1159 
1160 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1161 					u8 rmpp_version,
1162 					ib_mad_send_handler send_handler,
1163 					ib_mad_recv_handler recv_handler,
1164 					void *context)
1165 {
1166 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1167 }
1168 EXPORT_SYMBOL(ib_redirect_mad_qp);
1169 
1170 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1171 		      struct ib_wc *wc)
1172 {
1173 	printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1174 	return 0;
1175 }
1176 EXPORT_SYMBOL(ib_process_mad_wc);
1177 
1178 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1179 			 struct ib_mad_reg_req *mad_reg_req)
1180 {
1181 	int i;
1182 
1183 	for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1184 	     i < IB_MGMT_MAX_METHODS;
1185 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1186 			       1+i)) {
1187 		if ((*method)->agent[i]) {
1188 			printk(KERN_ERR PFX "Method %d already in use\n", i);
1189 			return -EINVAL;
1190 		}
1191 	}
1192 	return 0;
1193 }
1194 
1195 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1196 {
1197 	/* Allocate management method table */
1198 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1199 	if (!*method) {
1200 		printk(KERN_ERR PFX "No memory for "
1201 		       "ib_mad_mgmt_method_table\n");
1202 		return -ENOMEM;
1203 	}
1204 
1205 	return 0;
1206 }
1207 
1208 /*
1209  * Check to see if there are any methods still in use
1210  */
1211 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1212 {
1213 	int i;
1214 
1215 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1216 		if (method->agent[i])
1217 			return 1;
1218 	return 0;
1219 }
1220 
1221 /*
1222  * Check to see if there are any method tables for this class still in use
1223  */
1224 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1225 {
1226 	int i;
1227 
1228 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1229 		if (class->method_table[i])
1230 			return 1;
1231 	return 0;
1232 }
1233 
1234 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1235 {
1236 	int i;
1237 
1238 	for (i = 0; i < MAX_MGMT_OUI; i++)
1239 		if (vendor_class->method_table[i])
1240 			return 1;
1241 	return 0;
1242 }
1243 
1244 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1245 			   char *oui)
1246 {
1247 	int i;
1248 
1249 	for (i = 0; i < MAX_MGMT_OUI; i++)
1250 		/* Is there matching OUI for this vendor class ? */
1251 		if (!memcmp(vendor_class->oui[i], oui, 3))
1252 			return i;
1253 
1254 	return -1;
1255 }
1256 
1257 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1258 {
1259 	int i;
1260 
1261 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1262 		if (vendor->vendor_class[i])
1263 			return 1;
1264 
1265 	return 0;
1266 }
1267 
1268 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1269 				     struct ib_mad_agent_private *agent)
1270 {
1271 	int i;
1272 
1273 	/* Remove any methods for this mad agent */
1274 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1275 		if (method->agent[i] == agent) {
1276 			method->agent[i] = NULL;
1277 		}
1278 	}
1279 }
1280 
1281 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1282 			      struct ib_mad_agent_private *agent_priv,
1283 			      u8 mgmt_class)
1284 {
1285 	struct ib_mad_port_private *port_priv;
1286 	struct ib_mad_mgmt_class_table **class;
1287 	struct ib_mad_mgmt_method_table **method;
1288 	int i, ret;
1289 
1290 	port_priv = agent_priv->qp_info->port_priv;
1291 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1292 	if (!*class) {
1293 		/* Allocate management class table for "new" class version */
1294 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1295 		if (!*class) {
1296 			printk(KERN_ERR PFX "No memory for "
1297 			       "ib_mad_mgmt_class_table\n");
1298 			ret = -ENOMEM;
1299 			goto error1;
1300 		}
1301 
1302 		/* Allocate method table for this management class */
1303 		method = &(*class)->method_table[mgmt_class];
1304 		if ((ret = allocate_method_table(method)))
1305 			goto error2;
1306 	} else {
1307 		method = &(*class)->method_table[mgmt_class];
1308 		if (!*method) {
1309 			/* Allocate method table for this management class */
1310 			if ((ret = allocate_method_table(method)))
1311 				goto error1;
1312 		}
1313 	}
1314 
1315 	/* Now, make sure methods are not already in use */
1316 	if (method_in_use(method, mad_reg_req))
1317 		goto error3;
1318 
1319 	/* Finally, add in methods being registered */
1320 	for (i = find_first_bit(mad_reg_req->method_mask,
1321 				IB_MGMT_MAX_METHODS);
1322 	     i < IB_MGMT_MAX_METHODS;
1323 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1324 			       1+i)) {
1325 		(*method)->agent[i] = agent_priv;
1326 	}
1327 	return 0;
1328 
1329 error3:
1330 	/* Remove any methods for this mad agent */
1331 	remove_methods_mad_agent(*method, agent_priv);
1332 	/* Now, check to see if there are any methods in use */
1333 	if (!check_method_table(*method)) {
1334 		/* If not, release management method table */
1335 		kfree(*method);
1336 		*method = NULL;
1337 	}
1338 	ret = -EINVAL;
1339 	goto error1;
1340 error2:
1341 	kfree(*class);
1342 	*class = NULL;
1343 error1:
1344 	return ret;
1345 }
1346 
1347 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1348 			   struct ib_mad_agent_private *agent_priv)
1349 {
1350 	struct ib_mad_port_private *port_priv;
1351 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1352 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1353 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1354 	struct ib_mad_mgmt_method_table **method;
1355 	int i, ret = -ENOMEM;
1356 	u8 vclass;
1357 
1358 	/* "New" vendor (with OUI) class */
1359 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1360 	port_priv = agent_priv->qp_info->port_priv;
1361 	vendor_table = &port_priv->version[
1362 				mad_reg_req->mgmt_class_version].vendor;
1363 	if (!*vendor_table) {
1364 		/* Allocate mgmt vendor class table for "new" class version */
1365 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1366 		if (!vendor) {
1367 			printk(KERN_ERR PFX "No memory for "
1368 			       "ib_mad_mgmt_vendor_class_table\n");
1369 			goto error1;
1370 		}
1371 
1372 		*vendor_table = vendor;
1373 	}
1374 	if (!(*vendor_table)->vendor_class[vclass]) {
1375 		/* Allocate table for this management vendor class */
1376 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1377 		if (!vendor_class) {
1378 			printk(KERN_ERR PFX "No memory for "
1379 			       "ib_mad_mgmt_vendor_class\n");
1380 			goto error2;
1381 		}
1382 
1383 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1384 	}
1385 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1386 		/* Is there matching OUI for this vendor class ? */
1387 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1388 			    mad_reg_req->oui, 3)) {
1389 			method = &(*vendor_table)->vendor_class[
1390 						vclass]->method_table[i];
1391 			BUG_ON(!*method);
1392 			goto check_in_use;
1393 		}
1394 	}
1395 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1396 		/* OUI slot available ? */
1397 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1398 				vclass]->oui[i])) {
1399 			method = &(*vendor_table)->vendor_class[
1400 				vclass]->method_table[i];
1401 			BUG_ON(*method);
1402 			/* Allocate method table for this OUI */
1403 			if ((ret = allocate_method_table(method)))
1404 				goto error3;
1405 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1406 			       mad_reg_req->oui, 3);
1407 			goto check_in_use;
1408 		}
1409 	}
1410 	printk(KERN_ERR PFX "All OUI slots in use\n");
1411 	goto error3;
1412 
1413 check_in_use:
1414 	/* Now, make sure methods are not already in use */
1415 	if (method_in_use(method, mad_reg_req))
1416 		goto error4;
1417 
1418 	/* Finally, add in methods being registered */
1419 	for (i = find_first_bit(mad_reg_req->method_mask,
1420 				IB_MGMT_MAX_METHODS);
1421 	     i < IB_MGMT_MAX_METHODS;
1422 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1423 			       1+i)) {
1424 		(*method)->agent[i] = agent_priv;
1425 	}
1426 	return 0;
1427 
1428 error4:
1429 	/* Remove any methods for this mad agent */
1430 	remove_methods_mad_agent(*method, agent_priv);
1431 	/* Now, check to see if there are any methods in use */
1432 	if (!check_method_table(*method)) {
1433 		/* If not, release management method table */
1434 		kfree(*method);
1435 		*method = NULL;
1436 	}
1437 	ret = -EINVAL;
1438 error3:
1439 	if (vendor_class) {
1440 		(*vendor_table)->vendor_class[vclass] = NULL;
1441 		kfree(vendor_class);
1442 	}
1443 error2:
1444 	if (vendor) {
1445 		*vendor_table = NULL;
1446 		kfree(vendor);
1447 	}
1448 error1:
1449 	return ret;
1450 }
1451 
1452 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1453 {
1454 	struct ib_mad_port_private *port_priv;
1455 	struct ib_mad_mgmt_class_table *class;
1456 	struct ib_mad_mgmt_method_table *method;
1457 	struct ib_mad_mgmt_vendor_class_table *vendor;
1458 	struct ib_mad_mgmt_vendor_class *vendor_class;
1459 	int index;
1460 	u8 mgmt_class;
1461 
1462 	/*
1463 	 * Was MAD registration request supplied
1464 	 * with original registration ?
1465 	 */
1466 	if (!agent_priv->reg_req) {
1467 		goto out;
1468 	}
1469 
1470 	port_priv = agent_priv->qp_info->port_priv;
1471 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1472 	class = port_priv->version[
1473 			agent_priv->reg_req->mgmt_class_version].class;
1474 	if (!class)
1475 		goto vendor_check;
1476 
1477 	method = class->method_table[mgmt_class];
1478 	if (method) {
1479 		/* Remove any methods for this mad agent */
1480 		remove_methods_mad_agent(method, agent_priv);
1481 		/* Now, check to see if there are any methods still in use */
1482 		if (!check_method_table(method)) {
1483 			/* If not, release management method table */
1484 			 kfree(method);
1485 			 class->method_table[mgmt_class] = NULL;
1486 			 /* Any management classes left ? */
1487 			if (!check_class_table(class)) {
1488 				/* If not, release management class table */
1489 				kfree(class);
1490 				port_priv->version[
1491 					agent_priv->reg_req->
1492 					mgmt_class_version].class = NULL;
1493 			}
1494 		}
1495 	}
1496 
1497 vendor_check:
1498 	if (!is_vendor_class(mgmt_class))
1499 		goto out;
1500 
1501 	/* normalize mgmt_class to vendor range 2 */
1502 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1503 	vendor = port_priv->version[
1504 			agent_priv->reg_req->mgmt_class_version].vendor;
1505 
1506 	if (!vendor)
1507 		goto out;
1508 
1509 	vendor_class = vendor->vendor_class[mgmt_class];
1510 	if (vendor_class) {
1511 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1512 		if (index < 0)
1513 			goto out;
1514 		method = vendor_class->method_table[index];
1515 		if (method) {
1516 			/* Remove any methods for this mad agent */
1517 			remove_methods_mad_agent(method, agent_priv);
1518 			/*
1519 			 * Now, check to see if there are
1520 			 * any methods still in use
1521 			 */
1522 			if (!check_method_table(method)) {
1523 				/* If not, release management method table */
1524 				kfree(method);
1525 				vendor_class->method_table[index] = NULL;
1526 				memset(vendor_class->oui[index], 0, 3);
1527 				/* Any OUIs left ? */
1528 				if (!check_vendor_class(vendor_class)) {
1529 					/* If not, release vendor class table */
1530 					kfree(vendor_class);
1531 					vendor->vendor_class[mgmt_class] = NULL;
1532 					/* Any other vendor classes left ? */
1533 					if (!check_vendor_table(vendor)) {
1534 						kfree(vendor);
1535 						port_priv->version[
1536 							agent_priv->reg_req->
1537 							mgmt_class_version].
1538 							vendor = NULL;
1539 					}
1540 				}
1541 			}
1542 		}
1543 	}
1544 
1545 out:
1546 	return;
1547 }
1548 
1549 static struct ib_mad_agent_private *
1550 find_mad_agent(struct ib_mad_port_private *port_priv,
1551 	       struct ib_mad *mad)
1552 {
1553 	struct ib_mad_agent_private *mad_agent = NULL;
1554 	unsigned long flags;
1555 
1556 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1557 	if (ib_response_mad(mad)) {
1558 		u32 hi_tid;
1559 		struct ib_mad_agent_private *entry;
1560 
1561 		/*
1562 		 * Routing is based on high 32 bits of transaction ID
1563 		 * of MAD.
1564 		 */
1565 		hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1566 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1567 			if (entry->agent.hi_tid == hi_tid) {
1568 				mad_agent = entry;
1569 				break;
1570 			}
1571 		}
1572 	} else {
1573 		struct ib_mad_mgmt_class_table *class;
1574 		struct ib_mad_mgmt_method_table *method;
1575 		struct ib_mad_mgmt_vendor_class_table *vendor;
1576 		struct ib_mad_mgmt_vendor_class *vendor_class;
1577 		struct ib_vendor_mad *vendor_mad;
1578 		int index;
1579 
1580 		/*
1581 		 * Routing is based on version, class, and method
1582 		 * For "newer" vendor MADs, also based on OUI
1583 		 */
1584 		if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1585 			goto out;
1586 		if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1587 			class = port_priv->version[
1588 					mad->mad_hdr.class_version].class;
1589 			if (!class)
1590 				goto out;
1591 			method = class->method_table[convert_mgmt_class(
1592 							mad->mad_hdr.mgmt_class)];
1593 			if (method)
1594 				mad_agent = method->agent[mad->mad_hdr.method &
1595 							  ~IB_MGMT_METHOD_RESP];
1596 		} else {
1597 			vendor = port_priv->version[
1598 					mad->mad_hdr.class_version].vendor;
1599 			if (!vendor)
1600 				goto out;
1601 			vendor_class = vendor->vendor_class[vendor_class_index(
1602 						mad->mad_hdr.mgmt_class)];
1603 			if (!vendor_class)
1604 				goto out;
1605 			/* Find matching OUI */
1606 			vendor_mad = (struct ib_vendor_mad *)mad;
1607 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1608 			if (index == -1)
1609 				goto out;
1610 			method = vendor_class->method_table[index];
1611 			if (method) {
1612 				mad_agent = method->agent[mad->mad_hdr.method &
1613 							  ~IB_MGMT_METHOD_RESP];
1614 			}
1615 		}
1616 	}
1617 
1618 	if (mad_agent) {
1619 		if (mad_agent->agent.recv_handler)
1620 			atomic_inc(&mad_agent->refcount);
1621 		else {
1622 			printk(KERN_NOTICE PFX "No receive handler for client "
1623 			       "%p on port %d\n",
1624 			       &mad_agent->agent, port_priv->port_num);
1625 			mad_agent = NULL;
1626 		}
1627 	}
1628 out:
1629 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1630 
1631 	return mad_agent;
1632 }
1633 
1634 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1635 {
1636 	int valid = 0;
1637 
1638 	/* Make sure MAD base version is understood */
1639 	if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1640 		printk(KERN_ERR PFX "MAD received with unsupported base "
1641 		       "version %d\n", mad->mad_hdr.base_version);
1642 		goto out;
1643 	}
1644 
1645 	/* Filter SMI packets sent to other than QP0 */
1646 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1647 	    (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1648 		if (qp_num == 0)
1649 			valid = 1;
1650 	} else {
1651 		/* Filter GSI packets sent to QP0 */
1652 		if (qp_num != 0)
1653 			valid = 1;
1654 	}
1655 
1656 out:
1657 	return valid;
1658 }
1659 
1660 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1661 		       struct ib_mad_hdr *mad_hdr)
1662 {
1663 	struct ib_rmpp_mad *rmpp_mad;
1664 
1665 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1666 	return !mad_agent_priv->agent.rmpp_version ||
1667 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1668 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1669 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1670 }
1671 
1672 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1673 				     struct ib_mad_recv_wc *rwc)
1674 {
1675 	return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1676 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1677 }
1678 
1679 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1680 				   struct ib_mad_send_wr_private *wr,
1681 				   struct ib_mad_recv_wc *rwc )
1682 {
1683 	struct ib_ah_attr attr;
1684 	u8 send_resp, rcv_resp;
1685 	union ib_gid sgid;
1686 	struct ib_device *device = mad_agent_priv->agent.device;
1687 	u8 port_num = mad_agent_priv->agent.port_num;
1688 	u8 lmc;
1689 
1690 	send_resp = ((struct ib_mad *)(wr->send_buf.mad))->
1691 		     mad_hdr.method & IB_MGMT_METHOD_RESP;
1692 	rcv_resp = rwc->recv_buf.mad->mad_hdr.method & IB_MGMT_METHOD_RESP;
1693 
1694 	if (send_resp == rcv_resp)
1695 		/* both requests, or both responses. GIDs different */
1696 		return 0;
1697 
1698 	if (ib_query_ah(wr->send_buf.ah, &attr))
1699 		/* Assume not equal, to avoid false positives. */
1700 		return 0;
1701 
1702 	if (!!(attr.ah_flags & IB_AH_GRH) !=
1703 	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1704 		/* one has GID, other does not.  Assume different */
1705 		return 0;
1706 
1707 	if (!send_resp && rcv_resp) {
1708 		/* is request/response. */
1709 		if (!(attr.ah_flags & IB_AH_GRH)) {
1710 			if (ib_get_cached_lmc(device, port_num, &lmc))
1711 				return 0;
1712 			return (!lmc || !((attr.src_path_bits ^
1713 					   rwc->wc->dlid_path_bits) &
1714 					  ((1 << lmc) - 1)));
1715 		} else {
1716 			if (ib_get_cached_gid(device, port_num,
1717 					      attr.grh.sgid_index, &sgid))
1718 				return 0;
1719 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1720 				       16);
1721 		}
1722 	}
1723 
1724 	if (!(attr.ah_flags & IB_AH_GRH))
1725 		return attr.dlid == rwc->wc->slid;
1726 	else
1727 		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1728 			       16);
1729 }
1730 
1731 static inline int is_direct(u8 class)
1732 {
1733 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1734 }
1735 
1736 struct ib_mad_send_wr_private*
1737 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1738 		 struct ib_mad_recv_wc *wc)
1739 {
1740 	struct ib_mad_send_wr_private *wr;
1741 	struct ib_mad *mad;
1742 
1743 	mad = (struct ib_mad *)wc->recv_buf.mad;
1744 
1745 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1746 		if ((wr->tid == mad->mad_hdr.tid) &&
1747 		    rcv_has_same_class(wr, wc) &&
1748 		    /*
1749 		     * Don't check GID for direct routed MADs.
1750 		     * These might have permissive LIDs.
1751 		     */
1752 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1753 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1754 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1755 	}
1756 
1757 	/*
1758 	 * It's possible to receive the response before we've
1759 	 * been notified that the send has completed
1760 	 */
1761 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1762 		if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1763 		    wr->tid == mad->mad_hdr.tid &&
1764 		    wr->timeout &&
1765 		    rcv_has_same_class(wr, wc) &&
1766 		    /*
1767 		     * Don't check GID for direct routed MADs.
1768 		     * These might have permissive LIDs.
1769 		     */
1770 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1771 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1772 			/* Verify request has not been canceled */
1773 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1774 	}
1775 	return NULL;
1776 }
1777 
1778 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1779 {
1780 	mad_send_wr->timeout = 0;
1781 	if (mad_send_wr->refcount == 1)
1782 		list_move_tail(&mad_send_wr->agent_list,
1783 			      &mad_send_wr->mad_agent_priv->done_list);
1784 }
1785 
1786 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1787 				 struct ib_mad_recv_wc *mad_recv_wc)
1788 {
1789 	struct ib_mad_send_wr_private *mad_send_wr;
1790 	struct ib_mad_send_wc mad_send_wc;
1791 	unsigned long flags;
1792 
1793 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1794 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1795 	if (mad_agent_priv->agent.rmpp_version) {
1796 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1797 						      mad_recv_wc);
1798 		if (!mad_recv_wc) {
1799 			deref_mad_agent(mad_agent_priv);
1800 			return;
1801 		}
1802 	}
1803 
1804 	/* Complete corresponding request */
1805 	if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1806 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1807 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1808 		if (!mad_send_wr) {
1809 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1810 			ib_free_recv_mad(mad_recv_wc);
1811 			deref_mad_agent(mad_agent_priv);
1812 			return;
1813 		}
1814 		ib_mark_mad_done(mad_send_wr);
1815 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1816 
1817 		/* Defined behavior is to complete response before request */
1818 		mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1819 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1820 						   mad_recv_wc);
1821 		atomic_dec(&mad_agent_priv->refcount);
1822 
1823 		mad_send_wc.status = IB_WC_SUCCESS;
1824 		mad_send_wc.vendor_err = 0;
1825 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
1826 		ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1827 	} else {
1828 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1829 						   mad_recv_wc);
1830 		deref_mad_agent(mad_agent_priv);
1831 	}
1832 }
1833 
1834 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1835 				     struct ib_wc *wc)
1836 {
1837 	struct ib_mad_qp_info *qp_info;
1838 	struct ib_mad_private_header *mad_priv_hdr;
1839 	struct ib_mad_private *recv, *response;
1840 	struct ib_mad_list_head *mad_list;
1841 	struct ib_mad_agent_private *mad_agent;
1842 
1843 	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1844 	if (!response)
1845 		printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1846 		       "for response buffer\n");
1847 
1848 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1849 	qp_info = mad_list->mad_queue->qp_info;
1850 	dequeue_mad(mad_list);
1851 
1852 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1853 				    mad_list);
1854 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1855 	ib_dma_unmap_single(port_priv->device,
1856 			    recv->header.mapping,
1857 			    sizeof(struct ib_mad_private) -
1858 			      sizeof(struct ib_mad_private_header),
1859 			    DMA_FROM_DEVICE);
1860 
1861 	/* Setup MAD receive work completion from "normal" work completion */
1862 	recv->header.wc = *wc;
1863 	recv->header.recv_wc.wc = &recv->header.wc;
1864 	recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1865 	recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1866 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
1867 
1868 	if (atomic_read(&qp_info->snoop_count))
1869 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1870 
1871 	/* Validate MAD */
1872 	if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1873 		goto out;
1874 
1875 	if (recv->mad.mad.mad_hdr.mgmt_class ==
1876 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1877 		if (!smi_handle_dr_smp_recv(&recv->mad.smp,
1878 					    port_priv->device->node_type,
1879 					    port_priv->port_num,
1880 					    port_priv->device->phys_port_cnt))
1881 			goto out;
1882 		if (!smi_check_forward_dr_smp(&recv->mad.smp))
1883 			goto local;
1884 		if (!smi_handle_dr_smp_send(&recv->mad.smp,
1885 					    port_priv->device->node_type,
1886 					    port_priv->port_num))
1887 			goto out;
1888 		if (!smi_check_local_smp(&recv->mad.smp, port_priv->device))
1889 			goto out;
1890 	}
1891 
1892 local:
1893 	/* Give driver "right of first refusal" on incoming MAD */
1894 	if (port_priv->device->process_mad) {
1895 		int ret;
1896 
1897 		if (!response) {
1898 			printk(KERN_ERR PFX "No memory for response MAD\n");
1899 			/*
1900 			 * Is it better to assume that
1901 			 * it wouldn't be processed ?
1902 			 */
1903 			goto out;
1904 		}
1905 
1906 		ret = port_priv->device->process_mad(port_priv->device, 0,
1907 						     port_priv->port_num,
1908 						     wc, &recv->grh,
1909 						     &recv->mad.mad,
1910 						     &response->mad.mad);
1911 		if (ret & IB_MAD_RESULT_SUCCESS) {
1912 			if (ret & IB_MAD_RESULT_CONSUMED)
1913 				goto out;
1914 			if (ret & IB_MAD_RESULT_REPLY) {
1915 				agent_send_response(&response->mad.mad,
1916 						    &recv->grh, wc,
1917 						    port_priv->device,
1918 						    port_priv->port_num,
1919 						    qp_info->qp->qp_num);
1920 				goto out;
1921 			}
1922 		}
1923 	}
1924 
1925 	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1926 	if (mad_agent) {
1927 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1928 		/*
1929 		 * recv is freed up in error cases in ib_mad_complete_recv
1930 		 * or via recv_handler in ib_mad_complete_recv()
1931 		 */
1932 		recv = NULL;
1933 	}
1934 
1935 out:
1936 	/* Post another receive request for this QP */
1937 	if (response) {
1938 		ib_mad_post_receive_mads(qp_info, response);
1939 		if (recv)
1940 			kmem_cache_free(ib_mad_cache, recv);
1941 	} else
1942 		ib_mad_post_receive_mads(qp_info, recv);
1943 }
1944 
1945 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1946 {
1947 	struct ib_mad_send_wr_private *mad_send_wr;
1948 	unsigned long delay;
1949 
1950 	if (list_empty(&mad_agent_priv->wait_list)) {
1951 		cancel_delayed_work(&mad_agent_priv->timed_work);
1952 	} else {
1953 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1954 					 struct ib_mad_send_wr_private,
1955 					 agent_list);
1956 
1957 		if (time_after(mad_agent_priv->timeout,
1958 			       mad_send_wr->timeout)) {
1959 			mad_agent_priv->timeout = mad_send_wr->timeout;
1960 			cancel_delayed_work(&mad_agent_priv->timed_work);
1961 			delay = mad_send_wr->timeout - jiffies;
1962 			if ((long)delay <= 0)
1963 				delay = 1;
1964 			queue_delayed_work(mad_agent_priv->qp_info->
1965 					   port_priv->wq,
1966 					   &mad_agent_priv->timed_work, delay);
1967 		}
1968 	}
1969 }
1970 
1971 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1972 {
1973 	struct ib_mad_agent_private *mad_agent_priv;
1974 	struct ib_mad_send_wr_private *temp_mad_send_wr;
1975 	struct list_head *list_item;
1976 	unsigned long delay;
1977 
1978 	mad_agent_priv = mad_send_wr->mad_agent_priv;
1979 	list_del(&mad_send_wr->agent_list);
1980 
1981 	delay = mad_send_wr->timeout;
1982 	mad_send_wr->timeout += jiffies;
1983 
1984 	if (delay) {
1985 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1986 			temp_mad_send_wr = list_entry(list_item,
1987 						struct ib_mad_send_wr_private,
1988 						agent_list);
1989 			if (time_after(mad_send_wr->timeout,
1990 				       temp_mad_send_wr->timeout))
1991 				break;
1992 		}
1993 	}
1994 	else
1995 		list_item = &mad_agent_priv->wait_list;
1996 	list_add(&mad_send_wr->agent_list, list_item);
1997 
1998 	/* Reschedule a work item if we have a shorter timeout */
1999 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
2000 		cancel_delayed_work(&mad_agent_priv->timed_work);
2001 		queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2002 				   &mad_agent_priv->timed_work, delay);
2003 	}
2004 }
2005 
2006 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2007 			  int timeout_ms)
2008 {
2009 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2010 	wait_for_response(mad_send_wr);
2011 }
2012 
2013 /*
2014  * Process a send work completion
2015  */
2016 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2017 			     struct ib_mad_send_wc *mad_send_wc)
2018 {
2019 	struct ib_mad_agent_private	*mad_agent_priv;
2020 	unsigned long			flags;
2021 	int				ret;
2022 
2023 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2024 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2025 	if (mad_agent_priv->agent.rmpp_version) {
2026 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2027 		if (ret == IB_RMPP_RESULT_CONSUMED)
2028 			goto done;
2029 	} else
2030 		ret = IB_RMPP_RESULT_UNHANDLED;
2031 
2032 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2033 	    mad_send_wr->status == IB_WC_SUCCESS) {
2034 		mad_send_wr->status = mad_send_wc->status;
2035 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2036 	}
2037 
2038 	if (--mad_send_wr->refcount > 0) {
2039 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2040 		    mad_send_wr->status == IB_WC_SUCCESS) {
2041 			wait_for_response(mad_send_wr);
2042 		}
2043 		goto done;
2044 	}
2045 
2046 	/* Remove send from MAD agent and notify client of completion */
2047 	list_del(&mad_send_wr->agent_list);
2048 	adjust_timeout(mad_agent_priv);
2049 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2050 
2051 	if (mad_send_wr->status != IB_WC_SUCCESS )
2052 		mad_send_wc->status = mad_send_wr->status;
2053 	if (ret == IB_RMPP_RESULT_INTERNAL)
2054 		ib_rmpp_send_handler(mad_send_wc);
2055 	else
2056 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2057 						   mad_send_wc);
2058 
2059 	/* Release reference on agent taken when sending */
2060 	deref_mad_agent(mad_agent_priv);
2061 	return;
2062 done:
2063 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2064 }
2065 
2066 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2067 				     struct ib_wc *wc)
2068 {
2069 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2070 	struct ib_mad_list_head		*mad_list;
2071 	struct ib_mad_qp_info		*qp_info;
2072 	struct ib_mad_queue		*send_queue;
2073 	struct ib_send_wr		*bad_send_wr;
2074 	struct ib_mad_send_wc		mad_send_wc;
2075 	unsigned long flags;
2076 	int ret;
2077 
2078 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2079 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2080 				   mad_list);
2081 	send_queue = mad_list->mad_queue;
2082 	qp_info = send_queue->qp_info;
2083 
2084 retry:
2085 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2086 			    mad_send_wr->header_mapping,
2087 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2088 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2089 			    mad_send_wr->payload_mapping,
2090 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2091 	queued_send_wr = NULL;
2092 	spin_lock_irqsave(&send_queue->lock, flags);
2093 	list_del(&mad_list->list);
2094 
2095 	/* Move queued send to the send queue */
2096 	if (send_queue->count-- > send_queue->max_active) {
2097 		mad_list = container_of(qp_info->overflow_list.next,
2098 					struct ib_mad_list_head, list);
2099 		queued_send_wr = container_of(mad_list,
2100 					struct ib_mad_send_wr_private,
2101 					mad_list);
2102 		list_move_tail(&mad_list->list, &send_queue->list);
2103 	}
2104 	spin_unlock_irqrestore(&send_queue->lock, flags);
2105 
2106 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2107 	mad_send_wc.status = wc->status;
2108 	mad_send_wc.vendor_err = wc->vendor_err;
2109 	if (atomic_read(&qp_info->snoop_count))
2110 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2111 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2112 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2113 
2114 	if (queued_send_wr) {
2115 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2116 				   &bad_send_wr);
2117 		if (ret) {
2118 			printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2119 			mad_send_wr = queued_send_wr;
2120 			wc->status = IB_WC_LOC_QP_OP_ERR;
2121 			goto retry;
2122 		}
2123 	}
2124 }
2125 
2126 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2127 {
2128 	struct ib_mad_send_wr_private *mad_send_wr;
2129 	struct ib_mad_list_head *mad_list;
2130 	unsigned long flags;
2131 
2132 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2133 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2134 		mad_send_wr = container_of(mad_list,
2135 					   struct ib_mad_send_wr_private,
2136 					   mad_list);
2137 		mad_send_wr->retry = 1;
2138 	}
2139 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2140 }
2141 
2142 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2143 			      struct ib_wc *wc)
2144 {
2145 	struct ib_mad_list_head *mad_list;
2146 	struct ib_mad_qp_info *qp_info;
2147 	struct ib_mad_send_wr_private *mad_send_wr;
2148 	int ret;
2149 
2150 	/* Determine if failure was a send or receive */
2151 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2152 	qp_info = mad_list->mad_queue->qp_info;
2153 	if (mad_list->mad_queue == &qp_info->recv_queue)
2154 		/*
2155 		 * Receive errors indicate that the QP has entered the error
2156 		 * state - error handling/shutdown code will cleanup
2157 		 */
2158 		return;
2159 
2160 	/*
2161 	 * Send errors will transition the QP to SQE - move
2162 	 * QP to RTS and repost flushed work requests
2163 	 */
2164 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2165 				   mad_list);
2166 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2167 		if (mad_send_wr->retry) {
2168 			/* Repost send */
2169 			struct ib_send_wr *bad_send_wr;
2170 
2171 			mad_send_wr->retry = 0;
2172 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2173 					&bad_send_wr);
2174 			if (ret)
2175 				ib_mad_send_done_handler(port_priv, wc);
2176 		} else
2177 			ib_mad_send_done_handler(port_priv, wc);
2178 	} else {
2179 		struct ib_qp_attr *attr;
2180 
2181 		/* Transition QP to RTS and fail offending send */
2182 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2183 		if (attr) {
2184 			attr->qp_state = IB_QPS_RTS;
2185 			attr->cur_qp_state = IB_QPS_SQE;
2186 			ret = ib_modify_qp(qp_info->qp, attr,
2187 					   IB_QP_STATE | IB_QP_CUR_STATE);
2188 			kfree(attr);
2189 			if (ret)
2190 				printk(KERN_ERR PFX "mad_error_handler - "
2191 				       "ib_modify_qp to RTS : %d\n", ret);
2192 			else
2193 				mark_sends_for_retry(qp_info);
2194 		}
2195 		ib_mad_send_done_handler(port_priv, wc);
2196 	}
2197 }
2198 
2199 /*
2200  * IB MAD completion callback
2201  */
2202 static void ib_mad_completion_handler(struct work_struct *work)
2203 {
2204 	struct ib_mad_port_private *port_priv;
2205 	struct ib_wc wc;
2206 
2207 	port_priv = container_of(work, struct ib_mad_port_private, work);
2208 	ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2209 
2210 	while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2211 		if (wc.status == IB_WC_SUCCESS) {
2212 			switch (wc.opcode) {
2213 			case IB_WC_SEND:
2214 				ib_mad_send_done_handler(port_priv, &wc);
2215 				break;
2216 			case IB_WC_RECV:
2217 				ib_mad_recv_done_handler(port_priv, &wc);
2218 				break;
2219 			default:
2220 				BUG_ON(1);
2221 				break;
2222 			}
2223 		} else
2224 			mad_error_handler(port_priv, &wc);
2225 	}
2226 }
2227 
2228 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2229 {
2230 	unsigned long flags;
2231 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2232 	struct ib_mad_send_wc mad_send_wc;
2233 	struct list_head cancel_list;
2234 
2235 	INIT_LIST_HEAD(&cancel_list);
2236 
2237 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2238 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2239 				 &mad_agent_priv->send_list, agent_list) {
2240 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2241 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2242 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2243 		}
2244 	}
2245 
2246 	/* Empty wait list to prevent receives from finding a request */
2247 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2248 	/* Empty local completion list as well */
2249 	list_splice_init(&mad_agent_priv->local_list, &cancel_list);
2250 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2251 
2252 	/* Report all cancelled requests */
2253 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2254 	mad_send_wc.vendor_err = 0;
2255 
2256 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2257 				 &cancel_list, agent_list) {
2258 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2259 		list_del(&mad_send_wr->agent_list);
2260 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2261 						   &mad_send_wc);
2262 		atomic_dec(&mad_agent_priv->refcount);
2263 	}
2264 }
2265 
2266 static struct ib_mad_send_wr_private*
2267 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2268 	     struct ib_mad_send_buf *send_buf)
2269 {
2270 	struct ib_mad_send_wr_private *mad_send_wr;
2271 
2272 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2273 			    agent_list) {
2274 		if (&mad_send_wr->send_buf == send_buf)
2275 			return mad_send_wr;
2276 	}
2277 
2278 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2279 			    agent_list) {
2280 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2281 		    &mad_send_wr->send_buf == send_buf)
2282 			return mad_send_wr;
2283 	}
2284 	return NULL;
2285 }
2286 
2287 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2288 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2289 {
2290 	struct ib_mad_agent_private *mad_agent_priv;
2291 	struct ib_mad_send_wr_private *mad_send_wr;
2292 	unsigned long flags;
2293 	int active;
2294 
2295 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2296 				      agent);
2297 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2298 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2299 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2300 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2301 		return -EINVAL;
2302 	}
2303 
2304 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2305 	if (!timeout_ms) {
2306 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2307 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2308 	}
2309 
2310 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2311 	if (active)
2312 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2313 	else
2314 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2315 
2316 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2317 	return 0;
2318 }
2319 EXPORT_SYMBOL(ib_modify_mad);
2320 
2321 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2322 		   struct ib_mad_send_buf *send_buf)
2323 {
2324 	ib_modify_mad(mad_agent, send_buf, 0);
2325 }
2326 EXPORT_SYMBOL(ib_cancel_mad);
2327 
2328 static void local_completions(struct work_struct *work)
2329 {
2330 	struct ib_mad_agent_private *mad_agent_priv;
2331 	struct ib_mad_local_private *local;
2332 	struct ib_mad_agent_private *recv_mad_agent;
2333 	unsigned long flags;
2334 	int recv = 0;
2335 	struct ib_wc wc;
2336 	struct ib_mad_send_wc mad_send_wc;
2337 
2338 	mad_agent_priv =
2339 		container_of(work, struct ib_mad_agent_private, local_work);
2340 
2341 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2342 	while (!list_empty(&mad_agent_priv->local_list)) {
2343 		local = list_entry(mad_agent_priv->local_list.next,
2344 				   struct ib_mad_local_private,
2345 				   completion_list);
2346 		list_del(&local->completion_list);
2347 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2348 		if (local->mad_priv) {
2349 			recv_mad_agent = local->recv_mad_agent;
2350 			if (!recv_mad_agent) {
2351 				printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2352 				goto local_send_completion;
2353 			}
2354 
2355 			recv = 1;
2356 			/*
2357 			 * Defined behavior is to complete response
2358 			 * before request
2359 			 */
2360 			build_smp_wc(recv_mad_agent->agent.qp,
2361 				     (unsigned long) local->mad_send_wr,
2362 				     be16_to_cpu(IB_LID_PERMISSIVE),
2363 				     0, recv_mad_agent->agent.port_num, &wc);
2364 
2365 			local->mad_priv->header.recv_wc.wc = &wc;
2366 			local->mad_priv->header.recv_wc.mad_len =
2367 						sizeof(struct ib_mad);
2368 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2369 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2370 				 &local->mad_priv->header.recv_wc.rmpp_list);
2371 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2372 			local->mad_priv->header.recv_wc.recv_buf.mad =
2373 						&local->mad_priv->mad.mad;
2374 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2375 				snoop_recv(recv_mad_agent->qp_info,
2376 					  &local->mad_priv->header.recv_wc,
2377 					   IB_MAD_SNOOP_RECVS);
2378 			recv_mad_agent->agent.recv_handler(
2379 						&recv_mad_agent->agent,
2380 						&local->mad_priv->header.recv_wc);
2381 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2382 			atomic_dec(&recv_mad_agent->refcount);
2383 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2384 		}
2385 
2386 local_send_completion:
2387 		/* Complete send */
2388 		mad_send_wc.status = IB_WC_SUCCESS;
2389 		mad_send_wc.vendor_err = 0;
2390 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2391 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2392 			snoop_send(mad_agent_priv->qp_info,
2393 				   &local->mad_send_wr->send_buf,
2394 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2395 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2396 						   &mad_send_wc);
2397 
2398 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2399 		atomic_dec(&mad_agent_priv->refcount);
2400 		if (!recv)
2401 			kmem_cache_free(ib_mad_cache, local->mad_priv);
2402 		kfree(local);
2403 	}
2404 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2405 }
2406 
2407 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2408 {
2409 	int ret;
2410 
2411 	if (!mad_send_wr->retries--)
2412 		return -ETIMEDOUT;
2413 
2414 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2415 
2416 	if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2417 		ret = ib_retry_rmpp(mad_send_wr);
2418 		switch (ret) {
2419 		case IB_RMPP_RESULT_UNHANDLED:
2420 			ret = ib_send_mad(mad_send_wr);
2421 			break;
2422 		case IB_RMPP_RESULT_CONSUMED:
2423 			ret = 0;
2424 			break;
2425 		default:
2426 			ret = -ECOMM;
2427 			break;
2428 		}
2429 	} else
2430 		ret = ib_send_mad(mad_send_wr);
2431 
2432 	if (!ret) {
2433 		mad_send_wr->refcount++;
2434 		list_add_tail(&mad_send_wr->agent_list,
2435 			      &mad_send_wr->mad_agent_priv->send_list);
2436 	}
2437 	return ret;
2438 }
2439 
2440 static void timeout_sends(struct work_struct *work)
2441 {
2442 	struct ib_mad_agent_private *mad_agent_priv;
2443 	struct ib_mad_send_wr_private *mad_send_wr;
2444 	struct ib_mad_send_wc mad_send_wc;
2445 	unsigned long flags, delay;
2446 
2447 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2448 				      timed_work.work);
2449 	mad_send_wc.vendor_err = 0;
2450 
2451 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2452 	while (!list_empty(&mad_agent_priv->wait_list)) {
2453 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2454 					 struct ib_mad_send_wr_private,
2455 					 agent_list);
2456 
2457 		if (time_after(mad_send_wr->timeout, jiffies)) {
2458 			delay = mad_send_wr->timeout - jiffies;
2459 			if ((long)delay <= 0)
2460 				delay = 1;
2461 			queue_delayed_work(mad_agent_priv->qp_info->
2462 					   port_priv->wq,
2463 					   &mad_agent_priv->timed_work, delay);
2464 			break;
2465 		}
2466 
2467 		list_del(&mad_send_wr->agent_list);
2468 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2469 		    !retry_send(mad_send_wr))
2470 			continue;
2471 
2472 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2473 
2474 		if (mad_send_wr->status == IB_WC_SUCCESS)
2475 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2476 		else
2477 			mad_send_wc.status = mad_send_wr->status;
2478 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2479 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2480 						   &mad_send_wc);
2481 
2482 		atomic_dec(&mad_agent_priv->refcount);
2483 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2484 	}
2485 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2486 }
2487 
2488 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2489 {
2490 	struct ib_mad_port_private *port_priv = cq->cq_context;
2491 	unsigned long flags;
2492 
2493 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2494 	if (!list_empty(&port_priv->port_list))
2495 		queue_work(port_priv->wq, &port_priv->work);
2496 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2497 }
2498 
2499 /*
2500  * Allocate receive MADs and post receive WRs for them
2501  */
2502 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2503 				    struct ib_mad_private *mad)
2504 {
2505 	unsigned long flags;
2506 	int post, ret;
2507 	struct ib_mad_private *mad_priv;
2508 	struct ib_sge sg_list;
2509 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2510 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2511 
2512 	/* Initialize common scatter list fields */
2513 	sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2514 	sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2515 
2516 	/* Initialize common receive WR fields */
2517 	recv_wr.next = NULL;
2518 	recv_wr.sg_list = &sg_list;
2519 	recv_wr.num_sge = 1;
2520 
2521 	do {
2522 		/* Allocate and map receive buffer */
2523 		if (mad) {
2524 			mad_priv = mad;
2525 			mad = NULL;
2526 		} else {
2527 			mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2528 			if (!mad_priv) {
2529 				printk(KERN_ERR PFX "No memory for receive buffer\n");
2530 				ret = -ENOMEM;
2531 				break;
2532 			}
2533 		}
2534 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2535 						 &mad_priv->grh,
2536 						 sizeof *mad_priv -
2537 						   sizeof mad_priv->header,
2538 						 DMA_FROM_DEVICE);
2539 		mad_priv->header.mapping = sg_list.addr;
2540 		recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2541 		mad_priv->header.mad_list.mad_queue = recv_queue;
2542 
2543 		/* Post receive WR */
2544 		spin_lock_irqsave(&recv_queue->lock, flags);
2545 		post = (++recv_queue->count < recv_queue->max_active);
2546 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2547 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2548 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2549 		if (ret) {
2550 			spin_lock_irqsave(&recv_queue->lock, flags);
2551 			list_del(&mad_priv->header.mad_list.list);
2552 			recv_queue->count--;
2553 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2554 			ib_dma_unmap_single(qp_info->port_priv->device,
2555 					    mad_priv->header.mapping,
2556 					    sizeof *mad_priv -
2557 					      sizeof mad_priv->header,
2558 					    DMA_FROM_DEVICE);
2559 			kmem_cache_free(ib_mad_cache, mad_priv);
2560 			printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2561 			break;
2562 		}
2563 	} while (post);
2564 
2565 	return ret;
2566 }
2567 
2568 /*
2569  * Return all the posted receive MADs
2570  */
2571 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2572 {
2573 	struct ib_mad_private_header *mad_priv_hdr;
2574 	struct ib_mad_private *recv;
2575 	struct ib_mad_list_head *mad_list;
2576 
2577 	while (!list_empty(&qp_info->recv_queue.list)) {
2578 
2579 		mad_list = list_entry(qp_info->recv_queue.list.next,
2580 				      struct ib_mad_list_head, list);
2581 		mad_priv_hdr = container_of(mad_list,
2582 					    struct ib_mad_private_header,
2583 					    mad_list);
2584 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2585 				    header);
2586 
2587 		/* Remove from posted receive MAD list */
2588 		list_del(&mad_list->list);
2589 
2590 		ib_dma_unmap_single(qp_info->port_priv->device,
2591 				    recv->header.mapping,
2592 				    sizeof(struct ib_mad_private) -
2593 				      sizeof(struct ib_mad_private_header),
2594 				    DMA_FROM_DEVICE);
2595 		kmem_cache_free(ib_mad_cache, recv);
2596 	}
2597 
2598 	qp_info->recv_queue.count = 0;
2599 }
2600 
2601 /*
2602  * Start the port
2603  */
2604 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2605 {
2606 	int ret, i;
2607 	struct ib_qp_attr *attr;
2608 	struct ib_qp *qp;
2609 
2610 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2611 	if (!attr) {
2612 		printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2613 		return -ENOMEM;
2614 	}
2615 
2616 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2617 		qp = port_priv->qp_info[i].qp;
2618 		/*
2619 		 * PKey index for QP1 is irrelevant but
2620 		 * one is needed for the Reset to Init transition
2621 		 */
2622 		attr->qp_state = IB_QPS_INIT;
2623 		attr->pkey_index = 0;
2624 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2625 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2626 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2627 		if (ret) {
2628 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2629 			       "INIT: %d\n", i, ret);
2630 			goto out;
2631 		}
2632 
2633 		attr->qp_state = IB_QPS_RTR;
2634 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2635 		if (ret) {
2636 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2637 			       "RTR: %d\n", i, ret);
2638 			goto out;
2639 		}
2640 
2641 		attr->qp_state = IB_QPS_RTS;
2642 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2643 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2644 		if (ret) {
2645 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2646 			       "RTS: %d\n", i, ret);
2647 			goto out;
2648 		}
2649 	}
2650 
2651 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2652 	if (ret) {
2653 		printk(KERN_ERR PFX "Failed to request completion "
2654 		       "notification: %d\n", ret);
2655 		goto out;
2656 	}
2657 
2658 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2659 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2660 		if (ret) {
2661 			printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2662 			goto out;
2663 		}
2664 	}
2665 out:
2666 	kfree(attr);
2667 	return ret;
2668 }
2669 
2670 static void qp_event_handler(struct ib_event *event, void *qp_context)
2671 {
2672 	struct ib_mad_qp_info	*qp_info = qp_context;
2673 
2674 	/* It's worse than that! He's dead, Jim! */
2675 	printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2676 		event->event, qp_info->qp->qp_num);
2677 }
2678 
2679 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2680 			   struct ib_mad_queue *mad_queue)
2681 {
2682 	mad_queue->qp_info = qp_info;
2683 	mad_queue->count = 0;
2684 	spin_lock_init(&mad_queue->lock);
2685 	INIT_LIST_HEAD(&mad_queue->list);
2686 }
2687 
2688 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2689 			struct ib_mad_qp_info *qp_info)
2690 {
2691 	qp_info->port_priv = port_priv;
2692 	init_mad_queue(qp_info, &qp_info->send_queue);
2693 	init_mad_queue(qp_info, &qp_info->recv_queue);
2694 	INIT_LIST_HEAD(&qp_info->overflow_list);
2695 	spin_lock_init(&qp_info->snoop_lock);
2696 	qp_info->snoop_table = NULL;
2697 	qp_info->snoop_table_size = 0;
2698 	atomic_set(&qp_info->snoop_count, 0);
2699 }
2700 
2701 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2702 			 enum ib_qp_type qp_type)
2703 {
2704 	struct ib_qp_init_attr	qp_init_attr;
2705 	int ret;
2706 
2707 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2708 	qp_init_attr.send_cq = qp_info->port_priv->cq;
2709 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2710 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2711 	qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2712 	qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2713 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2714 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2715 	qp_init_attr.qp_type = qp_type;
2716 	qp_init_attr.port_num = qp_info->port_priv->port_num;
2717 	qp_init_attr.qp_context = qp_info;
2718 	qp_init_attr.event_handler = qp_event_handler;
2719 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2720 	if (IS_ERR(qp_info->qp)) {
2721 		printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2722 		       get_spl_qp_index(qp_type));
2723 		ret = PTR_ERR(qp_info->qp);
2724 		goto error;
2725 	}
2726 	/* Use minimum queue sizes unless the CQ is resized */
2727 	qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2728 	qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2729 	return 0;
2730 
2731 error:
2732 	return ret;
2733 }
2734 
2735 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2736 {
2737 	ib_destroy_qp(qp_info->qp);
2738 	kfree(qp_info->snoop_table);
2739 }
2740 
2741 /*
2742  * Open the port
2743  * Create the QP, PD, MR, and CQ if needed
2744  */
2745 static int ib_mad_port_open(struct ib_device *device,
2746 			    int port_num)
2747 {
2748 	int ret, cq_size;
2749 	struct ib_mad_port_private *port_priv;
2750 	unsigned long flags;
2751 	char name[sizeof "ib_mad123"];
2752 
2753 	/* Create new device info */
2754 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2755 	if (!port_priv) {
2756 		printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2757 		return -ENOMEM;
2758 	}
2759 
2760 	port_priv->device = device;
2761 	port_priv->port_num = port_num;
2762 	spin_lock_init(&port_priv->reg_lock);
2763 	INIT_LIST_HEAD(&port_priv->agent_list);
2764 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2765 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2766 
2767 	cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2768 	port_priv->cq = ib_create_cq(port_priv->device,
2769 				     ib_mad_thread_completion_handler,
2770 				     NULL, port_priv, cq_size);
2771 	if (IS_ERR(port_priv->cq)) {
2772 		printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2773 		ret = PTR_ERR(port_priv->cq);
2774 		goto error3;
2775 	}
2776 
2777 	port_priv->pd = ib_alloc_pd(device);
2778 	if (IS_ERR(port_priv->pd)) {
2779 		printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2780 		ret = PTR_ERR(port_priv->pd);
2781 		goto error4;
2782 	}
2783 
2784 	port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2785 	if (IS_ERR(port_priv->mr)) {
2786 		printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2787 		ret = PTR_ERR(port_priv->mr);
2788 		goto error5;
2789 	}
2790 
2791 	ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2792 	if (ret)
2793 		goto error6;
2794 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2795 	if (ret)
2796 		goto error7;
2797 
2798 	snprintf(name, sizeof name, "ib_mad%d", port_num);
2799 	port_priv->wq = create_singlethread_workqueue(name);
2800 	if (!port_priv->wq) {
2801 		ret = -ENOMEM;
2802 		goto error8;
2803 	}
2804 	INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2805 
2806 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2807 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2808 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2809 
2810 	ret = ib_mad_port_start(port_priv);
2811 	if (ret) {
2812 		printk(KERN_ERR PFX "Couldn't start port\n");
2813 		goto error9;
2814 	}
2815 
2816 	return 0;
2817 
2818 error9:
2819 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2820 	list_del_init(&port_priv->port_list);
2821 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2822 
2823 	destroy_workqueue(port_priv->wq);
2824 error8:
2825 	destroy_mad_qp(&port_priv->qp_info[1]);
2826 error7:
2827 	destroy_mad_qp(&port_priv->qp_info[0]);
2828 error6:
2829 	ib_dereg_mr(port_priv->mr);
2830 error5:
2831 	ib_dealloc_pd(port_priv->pd);
2832 error4:
2833 	ib_destroy_cq(port_priv->cq);
2834 	cleanup_recv_queue(&port_priv->qp_info[1]);
2835 	cleanup_recv_queue(&port_priv->qp_info[0]);
2836 error3:
2837 	kfree(port_priv);
2838 
2839 	return ret;
2840 }
2841 
2842 /*
2843  * Close the port
2844  * If there are no classes using the port, free the port
2845  * resources (CQ, MR, PD, QP) and remove the port's info structure
2846  */
2847 static int ib_mad_port_close(struct ib_device *device, int port_num)
2848 {
2849 	struct ib_mad_port_private *port_priv;
2850 	unsigned long flags;
2851 
2852 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2853 	port_priv = __ib_get_mad_port(device, port_num);
2854 	if (port_priv == NULL) {
2855 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2856 		printk(KERN_ERR PFX "Port %d not found\n", port_num);
2857 		return -ENODEV;
2858 	}
2859 	list_del_init(&port_priv->port_list);
2860 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2861 
2862 	destroy_workqueue(port_priv->wq);
2863 	destroy_mad_qp(&port_priv->qp_info[1]);
2864 	destroy_mad_qp(&port_priv->qp_info[0]);
2865 	ib_dereg_mr(port_priv->mr);
2866 	ib_dealloc_pd(port_priv->pd);
2867 	ib_destroy_cq(port_priv->cq);
2868 	cleanup_recv_queue(&port_priv->qp_info[1]);
2869 	cleanup_recv_queue(&port_priv->qp_info[0]);
2870 	/* XXX: Handle deallocation of MAD registration tables */
2871 
2872 	kfree(port_priv);
2873 
2874 	return 0;
2875 }
2876 
2877 static void ib_mad_init_device(struct ib_device *device)
2878 {
2879 	int start, end, i;
2880 
2881 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2882 		return;
2883 
2884 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2885 		start = 0;
2886 		end   = 0;
2887 	} else {
2888 		start = 1;
2889 		end   = device->phys_port_cnt;
2890 	}
2891 
2892 	for (i = start; i <= end; i++) {
2893 		if (ib_mad_port_open(device, i)) {
2894 			printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2895 			       device->name, i);
2896 			goto error;
2897 		}
2898 		if (ib_agent_port_open(device, i)) {
2899 			printk(KERN_ERR PFX "Couldn't open %s port %d "
2900 			       "for agents\n",
2901 			       device->name, i);
2902 			goto error_agent;
2903 		}
2904 	}
2905 	return;
2906 
2907 error_agent:
2908 	if (ib_mad_port_close(device, i))
2909 		printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2910 		       device->name, i);
2911 
2912 error:
2913 	i--;
2914 
2915 	while (i >= start) {
2916 		if (ib_agent_port_close(device, i))
2917 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2918 			       "for agents\n",
2919 			       device->name, i);
2920 		if (ib_mad_port_close(device, i))
2921 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2922 			       device->name, i);
2923 		i--;
2924 	}
2925 }
2926 
2927 static void ib_mad_remove_device(struct ib_device *device)
2928 {
2929 	int i, num_ports, cur_port;
2930 
2931 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2932 		num_ports = 1;
2933 		cur_port = 0;
2934 	} else {
2935 		num_ports = device->phys_port_cnt;
2936 		cur_port = 1;
2937 	}
2938 	for (i = 0; i < num_ports; i++, cur_port++) {
2939 		if (ib_agent_port_close(device, cur_port))
2940 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2941 			       "for agents\n",
2942 			       device->name, cur_port);
2943 		if (ib_mad_port_close(device, cur_port))
2944 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2945 			       device->name, cur_port);
2946 	}
2947 }
2948 
2949 static struct ib_client mad_client = {
2950 	.name   = "mad",
2951 	.add = ib_mad_init_device,
2952 	.remove = ib_mad_remove_device
2953 };
2954 
2955 static int __init ib_mad_init_module(void)
2956 {
2957 	int ret;
2958 
2959 	spin_lock_init(&ib_mad_port_list_lock);
2960 
2961 	ib_mad_cache = kmem_cache_create("ib_mad",
2962 					 sizeof(struct ib_mad_private),
2963 					 0,
2964 					 SLAB_HWCACHE_ALIGN,
2965 					 NULL,
2966 					 NULL);
2967 	if (!ib_mad_cache) {
2968 		printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2969 		ret = -ENOMEM;
2970 		goto error1;
2971 	}
2972 
2973 	INIT_LIST_HEAD(&ib_mad_port_list);
2974 
2975 	if (ib_register_client(&mad_client)) {
2976 		printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2977 		ret = -EINVAL;
2978 		goto error2;
2979 	}
2980 
2981 	return 0;
2982 
2983 error2:
2984 	kmem_cache_destroy(ib_mad_cache);
2985 error1:
2986 	return ret;
2987 }
2988 
2989 static void __exit ib_mad_cleanup_module(void)
2990 {
2991 	ib_unregister_client(&mad_client);
2992 	kmem_cache_destroy(ib_mad_cache);
2993 }
2994 
2995 module_init(ib_mad_init_module);
2996 module_exit(ib_mad_cleanup_module);
2997 
2998