xref: /linux/drivers/infiniband/core/mad.c (revision 367b8112fe2ea5c39a7bb4d263dcdd9b612fae18)
1 /*
2  * Copyright (c) 2004-2007 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  */
35 #include <linux/dma-mapping.h>
36 #include <rdma/ib_cache.h>
37 
38 #include "mad_priv.h"
39 #include "mad_rmpp.h"
40 #include "smi.h"
41 #include "agent.h"
42 
43 MODULE_LICENSE("Dual BSD/GPL");
44 MODULE_DESCRIPTION("kernel IB MAD API");
45 MODULE_AUTHOR("Hal Rosenstock");
46 MODULE_AUTHOR("Sean Hefty");
47 
48 static struct kmem_cache *ib_mad_cache;
49 
50 static struct list_head ib_mad_port_list;
51 static u32 ib_mad_client_id = 0;
52 
53 /* Port list lock */
54 static spinlock_t ib_mad_port_list_lock;
55 
56 
57 /* Forward declarations */
58 static int method_in_use(struct ib_mad_mgmt_method_table **method,
59 			 struct ib_mad_reg_req *mad_reg_req);
60 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
61 static struct ib_mad_agent_private *find_mad_agent(
62 					struct ib_mad_port_private *port_priv,
63 					struct ib_mad *mad);
64 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
65 				    struct ib_mad_private *mad);
66 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
67 static void timeout_sends(struct work_struct *work);
68 static void local_completions(struct work_struct *work);
69 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
70 			      struct ib_mad_agent_private *agent_priv,
71 			      u8 mgmt_class);
72 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
73 			   struct ib_mad_agent_private *agent_priv);
74 
75 /*
76  * Returns a ib_mad_port_private structure or NULL for a device/port
77  * Assumes ib_mad_port_list_lock is being held
78  */
79 static inline struct ib_mad_port_private *
80 __ib_get_mad_port(struct ib_device *device, int port_num)
81 {
82 	struct ib_mad_port_private *entry;
83 
84 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
85 		if (entry->device == device && entry->port_num == port_num)
86 			return entry;
87 	}
88 	return NULL;
89 }
90 
91 /*
92  * Wrapper function to return a ib_mad_port_private structure or NULL
93  * for a device/port
94  */
95 static inline struct ib_mad_port_private *
96 ib_get_mad_port(struct ib_device *device, int port_num)
97 {
98 	struct ib_mad_port_private *entry;
99 	unsigned long flags;
100 
101 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
102 	entry = __ib_get_mad_port(device, port_num);
103 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
104 
105 	return entry;
106 }
107 
108 static inline u8 convert_mgmt_class(u8 mgmt_class)
109 {
110 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
111 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
112 		0 : mgmt_class;
113 }
114 
115 static int get_spl_qp_index(enum ib_qp_type qp_type)
116 {
117 	switch (qp_type)
118 	{
119 	case IB_QPT_SMI:
120 		return 0;
121 	case IB_QPT_GSI:
122 		return 1;
123 	default:
124 		return -1;
125 	}
126 }
127 
128 static int vendor_class_index(u8 mgmt_class)
129 {
130 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
131 }
132 
133 static int is_vendor_class(u8 mgmt_class)
134 {
135 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
136 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
137 		return 0;
138 	return 1;
139 }
140 
141 static int is_vendor_oui(char *oui)
142 {
143 	if (oui[0] || oui[1] || oui[2])
144 		return 1;
145 	return 0;
146 }
147 
148 static int is_vendor_method_in_use(
149 		struct ib_mad_mgmt_vendor_class *vendor_class,
150 		struct ib_mad_reg_req *mad_reg_req)
151 {
152 	struct ib_mad_mgmt_method_table *method;
153 	int i;
154 
155 	for (i = 0; i < MAX_MGMT_OUI; i++) {
156 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
157 			method = vendor_class->method_table[i];
158 			if (method) {
159 				if (method_in_use(&method, mad_reg_req))
160 					return 1;
161 				else
162 					break;
163 			}
164 		}
165 	}
166 	return 0;
167 }
168 
169 int ib_response_mad(struct ib_mad *mad)
170 {
171 	return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
172 		(mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
173 		((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
174 		 (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
175 }
176 EXPORT_SYMBOL(ib_response_mad);
177 
178 /*
179  * ib_register_mad_agent - Register to send/receive MADs
180  */
181 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
182 					   u8 port_num,
183 					   enum ib_qp_type qp_type,
184 					   struct ib_mad_reg_req *mad_reg_req,
185 					   u8 rmpp_version,
186 					   ib_mad_send_handler send_handler,
187 					   ib_mad_recv_handler recv_handler,
188 					   void *context)
189 {
190 	struct ib_mad_port_private *port_priv;
191 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
192 	struct ib_mad_agent_private *mad_agent_priv;
193 	struct ib_mad_reg_req *reg_req = NULL;
194 	struct ib_mad_mgmt_class_table *class;
195 	struct ib_mad_mgmt_vendor_class_table *vendor;
196 	struct ib_mad_mgmt_vendor_class *vendor_class;
197 	struct ib_mad_mgmt_method_table *method;
198 	int ret2, qpn;
199 	unsigned long flags;
200 	u8 mgmt_class, vclass;
201 
202 	/* Validate parameters */
203 	qpn = get_spl_qp_index(qp_type);
204 	if (qpn == -1)
205 		goto error1;
206 
207 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
208 		goto error1;
209 
210 	/* Validate MAD registration request if supplied */
211 	if (mad_reg_req) {
212 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
213 			goto error1;
214 		if (!recv_handler)
215 			goto error1;
216 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
217 			/*
218 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
219 			 * one in this range currently allowed
220 			 */
221 			if (mad_reg_req->mgmt_class !=
222 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
223 				goto error1;
224 		} else if (mad_reg_req->mgmt_class == 0) {
225 			/*
226 			 * Class 0 is reserved in IBA and is used for
227 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
228 			 */
229 			goto error1;
230 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
231 			/*
232 			 * If class is in "new" vendor range,
233 			 * ensure supplied OUI is not zero
234 			 */
235 			if (!is_vendor_oui(mad_reg_req->oui))
236 				goto error1;
237 		}
238 		/* Make sure class supplied is consistent with RMPP */
239 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
240 			if (rmpp_version)
241 				goto error1;
242 		}
243 		/* Make sure class supplied is consistent with QP type */
244 		if (qp_type == IB_QPT_SMI) {
245 			if ((mad_reg_req->mgmt_class !=
246 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
247 			    (mad_reg_req->mgmt_class !=
248 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
249 				goto error1;
250 		} else {
251 			if ((mad_reg_req->mgmt_class ==
252 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
253 			    (mad_reg_req->mgmt_class ==
254 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
255 				goto error1;
256 		}
257 	} else {
258 		/* No registration request supplied */
259 		if (!send_handler)
260 			goto error1;
261 	}
262 
263 	/* Validate device and port */
264 	port_priv = ib_get_mad_port(device, port_num);
265 	if (!port_priv) {
266 		ret = ERR_PTR(-ENODEV);
267 		goto error1;
268 	}
269 
270 	/* Allocate structures */
271 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
272 	if (!mad_agent_priv) {
273 		ret = ERR_PTR(-ENOMEM);
274 		goto error1;
275 	}
276 
277 	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
278 						 IB_ACCESS_LOCAL_WRITE);
279 	if (IS_ERR(mad_agent_priv->agent.mr)) {
280 		ret = ERR_PTR(-ENOMEM);
281 		goto error2;
282 	}
283 
284 	if (mad_reg_req) {
285 		reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
286 		if (!reg_req) {
287 			ret = ERR_PTR(-ENOMEM);
288 			goto error3;
289 		}
290 		/* Make a copy of the MAD registration request */
291 		memcpy(reg_req, mad_reg_req, sizeof *reg_req);
292 	}
293 
294 	/* Now, fill in the various structures */
295 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
296 	mad_agent_priv->reg_req = reg_req;
297 	mad_agent_priv->agent.rmpp_version = rmpp_version;
298 	mad_agent_priv->agent.device = device;
299 	mad_agent_priv->agent.recv_handler = recv_handler;
300 	mad_agent_priv->agent.send_handler = send_handler;
301 	mad_agent_priv->agent.context = context;
302 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
303 	mad_agent_priv->agent.port_num = port_num;
304 
305 	spin_lock_irqsave(&port_priv->reg_lock, flags);
306 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
307 
308 	/*
309 	 * Make sure MAD registration (if supplied)
310 	 * is non overlapping with any existing ones
311 	 */
312 	if (mad_reg_req) {
313 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
314 		if (!is_vendor_class(mgmt_class)) {
315 			class = port_priv->version[mad_reg_req->
316 						   mgmt_class_version].class;
317 			if (class) {
318 				method = class->method_table[mgmt_class];
319 				if (method) {
320 					if (method_in_use(&method,
321 							   mad_reg_req))
322 						goto error4;
323 				}
324 			}
325 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
326 						  mgmt_class);
327 		} else {
328 			/* "New" vendor class range */
329 			vendor = port_priv->version[mad_reg_req->
330 						    mgmt_class_version].vendor;
331 			if (vendor) {
332 				vclass = vendor_class_index(mgmt_class);
333 				vendor_class = vendor->vendor_class[vclass];
334 				if (vendor_class) {
335 					if (is_vendor_method_in_use(
336 							vendor_class,
337 							mad_reg_req))
338 						goto error4;
339 				}
340 			}
341 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
342 		}
343 		if (ret2) {
344 			ret = ERR_PTR(ret2);
345 			goto error4;
346 		}
347 	}
348 
349 	/* Add mad agent into port's agent list */
350 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
351 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
352 
353 	spin_lock_init(&mad_agent_priv->lock);
354 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
355 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
356 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
357 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
358 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
359 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
360 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
361 	atomic_set(&mad_agent_priv->refcount, 1);
362 	init_completion(&mad_agent_priv->comp);
363 
364 	return &mad_agent_priv->agent;
365 
366 error4:
367 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
368 	kfree(reg_req);
369 error3:
370 	ib_dereg_mr(mad_agent_priv->agent.mr);
371 error2:
372 	kfree(mad_agent_priv);
373 error1:
374 	return ret;
375 }
376 EXPORT_SYMBOL(ib_register_mad_agent);
377 
378 static inline int is_snooping_sends(int mad_snoop_flags)
379 {
380 	return (mad_snoop_flags &
381 		(/*IB_MAD_SNOOP_POSTED_SENDS |
382 		 IB_MAD_SNOOP_RMPP_SENDS |*/
383 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
384 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
385 }
386 
387 static inline int is_snooping_recvs(int mad_snoop_flags)
388 {
389 	return (mad_snoop_flags &
390 		(IB_MAD_SNOOP_RECVS /*|
391 		 IB_MAD_SNOOP_RMPP_RECVS*/));
392 }
393 
394 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
395 				struct ib_mad_snoop_private *mad_snoop_priv)
396 {
397 	struct ib_mad_snoop_private **new_snoop_table;
398 	unsigned long flags;
399 	int i;
400 
401 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
402 	/* Check for empty slot in array. */
403 	for (i = 0; i < qp_info->snoop_table_size; i++)
404 		if (!qp_info->snoop_table[i])
405 			break;
406 
407 	if (i == qp_info->snoop_table_size) {
408 		/* Grow table. */
409 		new_snoop_table = krealloc(qp_info->snoop_table,
410 					   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 
418 		qp_info->snoop_table = new_snoop_table;
419 		qp_info->snoop_table_size++;
420 	}
421 	qp_info->snoop_table[i] = mad_snoop_priv;
422 	atomic_inc(&qp_info->snoop_count);
423 out:
424 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
425 	return i;
426 }
427 
428 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
429 					   u8 port_num,
430 					   enum ib_qp_type qp_type,
431 					   int mad_snoop_flags,
432 					   ib_mad_snoop_handler snoop_handler,
433 					   ib_mad_recv_handler recv_handler,
434 					   void *context)
435 {
436 	struct ib_mad_port_private *port_priv;
437 	struct ib_mad_agent *ret;
438 	struct ib_mad_snoop_private *mad_snoop_priv;
439 	int qpn;
440 
441 	/* Validate parameters */
442 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
443 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
444 		ret = ERR_PTR(-EINVAL);
445 		goto error1;
446 	}
447 	qpn = get_spl_qp_index(qp_type);
448 	if (qpn == -1) {
449 		ret = ERR_PTR(-EINVAL);
450 		goto error1;
451 	}
452 	port_priv = ib_get_mad_port(device, port_num);
453 	if (!port_priv) {
454 		ret = ERR_PTR(-ENODEV);
455 		goto error1;
456 	}
457 	/* Allocate structures */
458 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
459 	if (!mad_snoop_priv) {
460 		ret = ERR_PTR(-ENOMEM);
461 		goto error1;
462 	}
463 
464 	/* Now, fill in the various structures */
465 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
466 	mad_snoop_priv->agent.device = device;
467 	mad_snoop_priv->agent.recv_handler = recv_handler;
468 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
469 	mad_snoop_priv->agent.context = context;
470 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
471 	mad_snoop_priv->agent.port_num = port_num;
472 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
473 	init_completion(&mad_snoop_priv->comp);
474 	mad_snoop_priv->snoop_index = register_snoop_agent(
475 						&port_priv->qp_info[qpn],
476 						mad_snoop_priv);
477 	if (mad_snoop_priv->snoop_index < 0) {
478 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
479 		goto error2;
480 	}
481 
482 	atomic_set(&mad_snoop_priv->refcount, 1);
483 	return &mad_snoop_priv->agent;
484 
485 error2:
486 	kfree(mad_snoop_priv);
487 error1:
488 	return ret;
489 }
490 EXPORT_SYMBOL(ib_register_mad_snoop);
491 
492 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
493 {
494 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
495 		complete(&mad_agent_priv->comp);
496 }
497 
498 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
499 {
500 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
501 		complete(&mad_snoop_priv->comp);
502 }
503 
504 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
505 {
506 	struct ib_mad_port_private *port_priv;
507 	unsigned long flags;
508 
509 	/* Note that we could still be handling received MADs */
510 
511 	/*
512 	 * Canceling all sends results in dropping received response
513 	 * MADs, preventing us from queuing additional work
514 	 */
515 	cancel_mads(mad_agent_priv);
516 	port_priv = mad_agent_priv->qp_info->port_priv;
517 	cancel_delayed_work(&mad_agent_priv->timed_work);
518 
519 	spin_lock_irqsave(&port_priv->reg_lock, flags);
520 	remove_mad_reg_req(mad_agent_priv);
521 	list_del(&mad_agent_priv->agent_list);
522 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
523 
524 	flush_workqueue(port_priv->wq);
525 	ib_cancel_rmpp_recvs(mad_agent_priv);
526 
527 	deref_mad_agent(mad_agent_priv);
528 	wait_for_completion(&mad_agent_priv->comp);
529 
530 	kfree(mad_agent_priv->reg_req);
531 	ib_dereg_mr(mad_agent_priv->agent.mr);
532 	kfree(mad_agent_priv);
533 }
534 
535 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
536 {
537 	struct ib_mad_qp_info *qp_info;
538 	unsigned long flags;
539 
540 	qp_info = mad_snoop_priv->qp_info;
541 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
542 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
543 	atomic_dec(&qp_info->snoop_count);
544 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
545 
546 	deref_snoop_agent(mad_snoop_priv);
547 	wait_for_completion(&mad_snoop_priv->comp);
548 
549 	kfree(mad_snoop_priv);
550 }
551 
552 /*
553  * ib_unregister_mad_agent - Unregisters a client from using MAD services
554  */
555 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
556 {
557 	struct ib_mad_agent_private *mad_agent_priv;
558 	struct ib_mad_snoop_private *mad_snoop_priv;
559 
560 	/* If the TID is zero, the agent can only snoop. */
561 	if (mad_agent->hi_tid) {
562 		mad_agent_priv = container_of(mad_agent,
563 					      struct ib_mad_agent_private,
564 					      agent);
565 		unregister_mad_agent(mad_agent_priv);
566 	} else {
567 		mad_snoop_priv = container_of(mad_agent,
568 					      struct ib_mad_snoop_private,
569 					      agent);
570 		unregister_mad_snoop(mad_snoop_priv);
571 	}
572 	return 0;
573 }
574 EXPORT_SYMBOL(ib_unregister_mad_agent);
575 
576 static void dequeue_mad(struct ib_mad_list_head *mad_list)
577 {
578 	struct ib_mad_queue *mad_queue;
579 	unsigned long flags;
580 
581 	BUG_ON(!mad_list->mad_queue);
582 	mad_queue = mad_list->mad_queue;
583 	spin_lock_irqsave(&mad_queue->lock, flags);
584 	list_del(&mad_list->list);
585 	mad_queue->count--;
586 	spin_unlock_irqrestore(&mad_queue->lock, flags);
587 }
588 
589 static void snoop_send(struct ib_mad_qp_info *qp_info,
590 		       struct ib_mad_send_buf *send_buf,
591 		       struct ib_mad_send_wc *mad_send_wc,
592 		       int mad_snoop_flags)
593 {
594 	struct ib_mad_snoop_private *mad_snoop_priv;
595 	unsigned long flags;
596 	int i;
597 
598 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
599 	for (i = 0; i < qp_info->snoop_table_size; i++) {
600 		mad_snoop_priv = qp_info->snoop_table[i];
601 		if (!mad_snoop_priv ||
602 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
603 			continue;
604 
605 		atomic_inc(&mad_snoop_priv->refcount);
606 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
607 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
608 						    send_buf, mad_send_wc);
609 		deref_snoop_agent(mad_snoop_priv);
610 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
611 	}
612 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
613 }
614 
615 static void snoop_recv(struct ib_mad_qp_info *qp_info,
616 		       struct ib_mad_recv_wc *mad_recv_wc,
617 		       int mad_snoop_flags)
618 {
619 	struct ib_mad_snoop_private *mad_snoop_priv;
620 	unsigned long flags;
621 	int i;
622 
623 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
624 	for (i = 0; i < qp_info->snoop_table_size; i++) {
625 		mad_snoop_priv = qp_info->snoop_table[i];
626 		if (!mad_snoop_priv ||
627 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
628 			continue;
629 
630 		atomic_inc(&mad_snoop_priv->refcount);
631 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
632 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
633 						   mad_recv_wc);
634 		deref_snoop_agent(mad_snoop_priv);
635 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
636 	}
637 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
638 }
639 
640 static void build_smp_wc(struct ib_qp *qp,
641 			 u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
642 			 struct ib_wc *wc)
643 {
644 	memset(wc, 0, sizeof *wc);
645 	wc->wr_id = wr_id;
646 	wc->status = IB_WC_SUCCESS;
647 	wc->opcode = IB_WC_RECV;
648 	wc->pkey_index = pkey_index;
649 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
650 	wc->src_qp = IB_QP0;
651 	wc->qp = qp;
652 	wc->slid = slid;
653 	wc->sl = 0;
654 	wc->dlid_path_bits = 0;
655 	wc->port_num = port_num;
656 }
657 
658 /*
659  * Return 0 if SMP is to be sent
660  * Return 1 if SMP was consumed locally (whether or not solicited)
661  * Return < 0 if error
662  */
663 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
664 				  struct ib_mad_send_wr_private *mad_send_wr)
665 {
666 	int ret = 0;
667 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
668 	unsigned long flags;
669 	struct ib_mad_local_private *local;
670 	struct ib_mad_private *mad_priv;
671 	struct ib_mad_port_private *port_priv;
672 	struct ib_mad_agent_private *recv_mad_agent = NULL;
673 	struct ib_device *device = mad_agent_priv->agent.device;
674 	u8 port_num;
675 	struct ib_wc mad_wc;
676 	struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
677 
678 	if (device->node_type == RDMA_NODE_IB_SWITCH &&
679 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
680 		port_num = send_wr->wr.ud.port_num;
681 	else
682 		port_num = mad_agent_priv->agent.port_num;
683 
684 	/*
685 	 * Directed route handling starts if the initial LID routed part of
686 	 * a request or the ending LID routed part of a response is empty.
687 	 * If we are at the start of the LID routed part, don't update the
688 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
689 	 */
690 	if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
691 	     IB_LID_PERMISSIVE &&
692 	     smi_handle_dr_smp_send(smp, device->node_type, port_num) ==
693 	     IB_SMI_DISCARD) {
694 		ret = -EINVAL;
695 		printk(KERN_ERR PFX "Invalid directed route\n");
696 		goto out;
697 	}
698 
699 	/* Check to post send on QP or process locally */
700 	if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
701 	    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
702 		goto out;
703 
704 	local = kmalloc(sizeof *local, GFP_ATOMIC);
705 	if (!local) {
706 		ret = -ENOMEM;
707 		printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
708 		goto out;
709 	}
710 	local->mad_priv = NULL;
711 	local->recv_mad_agent = NULL;
712 	mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
713 	if (!mad_priv) {
714 		ret = -ENOMEM;
715 		printk(KERN_ERR PFX "No memory for local response MAD\n");
716 		kfree(local);
717 		goto out;
718 	}
719 
720 	build_smp_wc(mad_agent_priv->agent.qp,
721 		     send_wr->wr_id, be16_to_cpu(smp->dr_slid),
722 		     send_wr->wr.ud.pkey_index,
723 		     send_wr->wr.ud.port_num, &mad_wc);
724 
725 	/* No GRH for DR SMP */
726 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
727 				  (struct ib_mad *)smp,
728 				  (struct ib_mad *)&mad_priv->mad);
729 	switch (ret)
730 	{
731 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
732 		if (ib_response_mad(&mad_priv->mad.mad) &&
733 		    mad_agent_priv->agent.recv_handler) {
734 			local->mad_priv = mad_priv;
735 			local->recv_mad_agent = mad_agent_priv;
736 			/*
737 			 * Reference MAD agent until receive
738 			 * side of local completion handled
739 			 */
740 			atomic_inc(&mad_agent_priv->refcount);
741 		} else
742 			kmem_cache_free(ib_mad_cache, mad_priv);
743 		break;
744 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
745 		kmem_cache_free(ib_mad_cache, mad_priv);
746 		kfree(local);
747 		ret = 1;
748 		goto out;
749 	case IB_MAD_RESULT_SUCCESS:
750 		/* Treat like an incoming receive MAD */
751 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
752 					    mad_agent_priv->agent.port_num);
753 		if (port_priv) {
754 			memcpy(&mad_priv->mad.mad, smp, sizeof(struct ib_mad));
755 			recv_mad_agent = find_mad_agent(port_priv,
756 						        &mad_priv->mad.mad);
757 		}
758 		if (!port_priv || !recv_mad_agent) {
759 			kmem_cache_free(ib_mad_cache, mad_priv);
760 			kfree(local);
761 			ret = 0;
762 			goto out;
763 		}
764 		local->mad_priv = mad_priv;
765 		local->recv_mad_agent = recv_mad_agent;
766 		break;
767 	default:
768 		kmem_cache_free(ib_mad_cache, mad_priv);
769 		kfree(local);
770 		ret = -EINVAL;
771 		goto out;
772 	}
773 
774 	local->mad_send_wr = mad_send_wr;
775 	/* Reference MAD agent until send side of local completion handled */
776 	atomic_inc(&mad_agent_priv->refcount);
777 	/* Queue local completion to local list */
778 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
779 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
780 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
781 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
782 		   &mad_agent_priv->local_work);
783 	ret = 1;
784 out:
785 	return ret;
786 }
787 
788 static int get_pad_size(int hdr_len, int data_len)
789 {
790 	int seg_size, pad;
791 
792 	seg_size = sizeof(struct ib_mad) - hdr_len;
793 	if (data_len && seg_size) {
794 		pad = seg_size - data_len % seg_size;
795 		return pad == seg_size ? 0 : pad;
796 	} else
797 		return seg_size;
798 }
799 
800 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
801 {
802 	struct ib_rmpp_segment *s, *t;
803 
804 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
805 		list_del(&s->list);
806 		kfree(s);
807 	}
808 }
809 
810 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
811 				gfp_t gfp_mask)
812 {
813 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
814 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
815 	struct ib_rmpp_segment *seg = NULL;
816 	int left, seg_size, pad;
817 
818 	send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
819 	seg_size = send_buf->seg_size;
820 	pad = send_wr->pad;
821 
822 	/* Allocate data segments. */
823 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
824 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
825 		if (!seg) {
826 			printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
827 			       "alloc failed for len %zd, gfp %#x\n",
828 			       sizeof (*seg) + seg_size, gfp_mask);
829 			free_send_rmpp_list(send_wr);
830 			return -ENOMEM;
831 		}
832 		seg->num = ++send_buf->seg_count;
833 		list_add_tail(&seg->list, &send_wr->rmpp_list);
834 	}
835 
836 	/* Zero any padding */
837 	if (pad)
838 		memset(seg->data + seg_size - pad, 0, pad);
839 
840 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
841 					  agent.rmpp_version;
842 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
843 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
844 
845 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
846 					struct ib_rmpp_segment, list);
847 	send_wr->last_ack_seg = send_wr->cur_seg;
848 	return 0;
849 }
850 
851 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
852 					    u32 remote_qpn, u16 pkey_index,
853 					    int rmpp_active,
854 					    int hdr_len, int data_len,
855 					    gfp_t gfp_mask)
856 {
857 	struct ib_mad_agent_private *mad_agent_priv;
858 	struct ib_mad_send_wr_private *mad_send_wr;
859 	int pad, message_size, ret, size;
860 	void *buf;
861 
862 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
863 				      agent);
864 	pad = get_pad_size(hdr_len, data_len);
865 	message_size = hdr_len + data_len + pad;
866 
867 	if ((!mad_agent->rmpp_version &&
868 	     (rmpp_active || message_size > sizeof(struct ib_mad))) ||
869 	    (!rmpp_active && message_size > sizeof(struct ib_mad)))
870 		return ERR_PTR(-EINVAL);
871 
872 	size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
873 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
874 	if (!buf)
875 		return ERR_PTR(-ENOMEM);
876 
877 	mad_send_wr = buf + size;
878 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
879 	mad_send_wr->send_buf.mad = buf;
880 	mad_send_wr->send_buf.hdr_len = hdr_len;
881 	mad_send_wr->send_buf.data_len = data_len;
882 	mad_send_wr->pad = pad;
883 
884 	mad_send_wr->mad_agent_priv = mad_agent_priv;
885 	mad_send_wr->sg_list[0].length = hdr_len;
886 	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
887 	mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
888 	mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
889 
890 	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
891 	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
892 	mad_send_wr->send_wr.num_sge = 2;
893 	mad_send_wr->send_wr.opcode = IB_WR_SEND;
894 	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
895 	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
896 	mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
897 	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
898 
899 	if (rmpp_active) {
900 		ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
901 		if (ret) {
902 			kfree(buf);
903 			return ERR_PTR(ret);
904 		}
905 	}
906 
907 	mad_send_wr->send_buf.mad_agent = mad_agent;
908 	atomic_inc(&mad_agent_priv->refcount);
909 	return &mad_send_wr->send_buf;
910 }
911 EXPORT_SYMBOL(ib_create_send_mad);
912 
913 int ib_get_mad_data_offset(u8 mgmt_class)
914 {
915 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
916 		return IB_MGMT_SA_HDR;
917 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
918 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
919 		 (mgmt_class == IB_MGMT_CLASS_BIS))
920 		return IB_MGMT_DEVICE_HDR;
921 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
922 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
923 		return IB_MGMT_VENDOR_HDR;
924 	else
925 		return IB_MGMT_MAD_HDR;
926 }
927 EXPORT_SYMBOL(ib_get_mad_data_offset);
928 
929 int ib_is_mad_class_rmpp(u8 mgmt_class)
930 {
931 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
932 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
933 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
934 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
935 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
936 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
937 		return 1;
938 	return 0;
939 }
940 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
941 
942 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
943 {
944 	struct ib_mad_send_wr_private *mad_send_wr;
945 	struct list_head *list;
946 
947 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
948 				   send_buf);
949 	list = &mad_send_wr->cur_seg->list;
950 
951 	if (mad_send_wr->cur_seg->num < seg_num) {
952 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
953 			if (mad_send_wr->cur_seg->num == seg_num)
954 				break;
955 	} else if (mad_send_wr->cur_seg->num > seg_num) {
956 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
957 			if (mad_send_wr->cur_seg->num == seg_num)
958 				break;
959 	}
960 	return mad_send_wr->cur_seg->data;
961 }
962 EXPORT_SYMBOL(ib_get_rmpp_segment);
963 
964 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
965 {
966 	if (mad_send_wr->send_buf.seg_count)
967 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
968 					   mad_send_wr->seg_num);
969 	else
970 		return mad_send_wr->send_buf.mad +
971 		       mad_send_wr->send_buf.hdr_len;
972 }
973 
974 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
975 {
976 	struct ib_mad_agent_private *mad_agent_priv;
977 	struct ib_mad_send_wr_private *mad_send_wr;
978 
979 	mad_agent_priv = container_of(send_buf->mad_agent,
980 				      struct ib_mad_agent_private, agent);
981 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
982 				   send_buf);
983 
984 	free_send_rmpp_list(mad_send_wr);
985 	kfree(send_buf->mad);
986 	deref_mad_agent(mad_agent_priv);
987 }
988 EXPORT_SYMBOL(ib_free_send_mad);
989 
990 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
991 {
992 	struct ib_mad_qp_info *qp_info;
993 	struct list_head *list;
994 	struct ib_send_wr *bad_send_wr;
995 	struct ib_mad_agent *mad_agent;
996 	struct ib_sge *sge;
997 	unsigned long flags;
998 	int ret;
999 
1000 	/* Set WR ID to find mad_send_wr upon completion */
1001 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1002 	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
1003 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1004 
1005 	mad_agent = mad_send_wr->send_buf.mad_agent;
1006 	sge = mad_send_wr->sg_list;
1007 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1008 					mad_send_wr->send_buf.mad,
1009 					sge[0].length,
1010 					DMA_TO_DEVICE);
1011 	mad_send_wr->header_mapping = sge[0].addr;
1012 
1013 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1014 					ib_get_payload(mad_send_wr),
1015 					sge[1].length,
1016 					DMA_TO_DEVICE);
1017 	mad_send_wr->payload_mapping = sge[1].addr;
1018 
1019 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1020 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1021 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1022 				   &bad_send_wr);
1023 		list = &qp_info->send_queue.list;
1024 	} else {
1025 		ret = 0;
1026 		list = &qp_info->overflow_list;
1027 	}
1028 
1029 	if (!ret) {
1030 		qp_info->send_queue.count++;
1031 		list_add_tail(&mad_send_wr->mad_list.list, list);
1032 	}
1033 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1034 	if (ret) {
1035 		ib_dma_unmap_single(mad_agent->device,
1036 				    mad_send_wr->header_mapping,
1037 				    sge[0].length, DMA_TO_DEVICE);
1038 		ib_dma_unmap_single(mad_agent->device,
1039 				    mad_send_wr->payload_mapping,
1040 				    sge[1].length, DMA_TO_DEVICE);
1041 	}
1042 	return ret;
1043 }
1044 
1045 /*
1046  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1047  *  with the registered client
1048  */
1049 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1050 		     struct ib_mad_send_buf **bad_send_buf)
1051 {
1052 	struct ib_mad_agent_private *mad_agent_priv;
1053 	struct ib_mad_send_buf *next_send_buf;
1054 	struct ib_mad_send_wr_private *mad_send_wr;
1055 	unsigned long flags;
1056 	int ret = -EINVAL;
1057 
1058 	/* Walk list of send WRs and post each on send list */
1059 	for (; send_buf; send_buf = next_send_buf) {
1060 
1061 		mad_send_wr = container_of(send_buf,
1062 					   struct ib_mad_send_wr_private,
1063 					   send_buf);
1064 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1065 
1066 		if (!send_buf->mad_agent->send_handler ||
1067 		    (send_buf->timeout_ms &&
1068 		     !send_buf->mad_agent->recv_handler)) {
1069 			ret = -EINVAL;
1070 			goto error;
1071 		}
1072 
1073 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1074 			if (mad_agent_priv->agent.rmpp_version) {
1075 				ret = -EINVAL;
1076 				goto error;
1077 			}
1078 		}
1079 
1080 		/*
1081 		 * Save pointer to next work request to post in case the
1082 		 * current one completes, and the user modifies the work
1083 		 * request associated with the completion
1084 		 */
1085 		next_send_buf = send_buf->next;
1086 		mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1087 
1088 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1089 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1090 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1091 						     mad_send_wr);
1092 			if (ret < 0)		/* error */
1093 				goto error;
1094 			else if (ret == 1)	/* locally consumed */
1095 				continue;
1096 		}
1097 
1098 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1099 		/* Timeout will be updated after send completes */
1100 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1101 		mad_send_wr->max_retries = send_buf->retries;
1102 		mad_send_wr->retries_left = send_buf->retries;
1103 		send_buf->retries = 0;
1104 		/* Reference for work request to QP + response */
1105 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1106 		mad_send_wr->status = IB_WC_SUCCESS;
1107 
1108 		/* Reference MAD agent until send completes */
1109 		atomic_inc(&mad_agent_priv->refcount);
1110 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1111 		list_add_tail(&mad_send_wr->agent_list,
1112 			      &mad_agent_priv->send_list);
1113 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1114 
1115 		if (mad_agent_priv->agent.rmpp_version) {
1116 			ret = ib_send_rmpp_mad(mad_send_wr);
1117 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1118 				ret = ib_send_mad(mad_send_wr);
1119 		} else
1120 			ret = ib_send_mad(mad_send_wr);
1121 		if (ret < 0) {
1122 			/* Fail send request */
1123 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1124 			list_del(&mad_send_wr->agent_list);
1125 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1126 			atomic_dec(&mad_agent_priv->refcount);
1127 			goto error;
1128 		}
1129 	}
1130 	return 0;
1131 error:
1132 	if (bad_send_buf)
1133 		*bad_send_buf = send_buf;
1134 	return ret;
1135 }
1136 EXPORT_SYMBOL(ib_post_send_mad);
1137 
1138 /*
1139  * ib_free_recv_mad - Returns data buffers used to receive
1140  *  a MAD to the access layer
1141  */
1142 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1143 {
1144 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1145 	struct ib_mad_private_header *mad_priv_hdr;
1146 	struct ib_mad_private *priv;
1147 	struct list_head free_list;
1148 
1149 	INIT_LIST_HEAD(&free_list);
1150 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1151 
1152 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1153 					&free_list, list) {
1154 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1155 					   recv_buf);
1156 		mad_priv_hdr = container_of(mad_recv_wc,
1157 					    struct ib_mad_private_header,
1158 					    recv_wc);
1159 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1160 				    header);
1161 		kmem_cache_free(ib_mad_cache, priv);
1162 	}
1163 }
1164 EXPORT_SYMBOL(ib_free_recv_mad);
1165 
1166 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1167 					u8 rmpp_version,
1168 					ib_mad_send_handler send_handler,
1169 					ib_mad_recv_handler recv_handler,
1170 					void *context)
1171 {
1172 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1173 }
1174 EXPORT_SYMBOL(ib_redirect_mad_qp);
1175 
1176 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1177 		      struct ib_wc *wc)
1178 {
1179 	printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1180 	return 0;
1181 }
1182 EXPORT_SYMBOL(ib_process_mad_wc);
1183 
1184 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1185 			 struct ib_mad_reg_req *mad_reg_req)
1186 {
1187 	int i;
1188 
1189 	for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1190 	     i < IB_MGMT_MAX_METHODS;
1191 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1192 			       1+i)) {
1193 		if ((*method)->agent[i]) {
1194 			printk(KERN_ERR PFX "Method %d already in use\n", i);
1195 			return -EINVAL;
1196 		}
1197 	}
1198 	return 0;
1199 }
1200 
1201 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1202 {
1203 	/* Allocate management method table */
1204 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1205 	if (!*method) {
1206 		printk(KERN_ERR PFX "No memory for "
1207 		       "ib_mad_mgmt_method_table\n");
1208 		return -ENOMEM;
1209 	}
1210 
1211 	return 0;
1212 }
1213 
1214 /*
1215  * Check to see if there are any methods still in use
1216  */
1217 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1218 {
1219 	int i;
1220 
1221 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1222 		if (method->agent[i])
1223 			return 1;
1224 	return 0;
1225 }
1226 
1227 /*
1228  * Check to see if there are any method tables for this class still in use
1229  */
1230 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1231 {
1232 	int i;
1233 
1234 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1235 		if (class->method_table[i])
1236 			return 1;
1237 	return 0;
1238 }
1239 
1240 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1241 {
1242 	int i;
1243 
1244 	for (i = 0; i < MAX_MGMT_OUI; i++)
1245 		if (vendor_class->method_table[i])
1246 			return 1;
1247 	return 0;
1248 }
1249 
1250 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1251 			   char *oui)
1252 {
1253 	int i;
1254 
1255 	for (i = 0; i < MAX_MGMT_OUI; i++)
1256 		/* Is there matching OUI for this vendor class ? */
1257 		if (!memcmp(vendor_class->oui[i], oui, 3))
1258 			return i;
1259 
1260 	return -1;
1261 }
1262 
1263 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1264 {
1265 	int i;
1266 
1267 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1268 		if (vendor->vendor_class[i])
1269 			return 1;
1270 
1271 	return 0;
1272 }
1273 
1274 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1275 				     struct ib_mad_agent_private *agent)
1276 {
1277 	int i;
1278 
1279 	/* Remove any methods for this mad agent */
1280 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1281 		if (method->agent[i] == agent) {
1282 			method->agent[i] = NULL;
1283 		}
1284 	}
1285 }
1286 
1287 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1288 			      struct ib_mad_agent_private *agent_priv,
1289 			      u8 mgmt_class)
1290 {
1291 	struct ib_mad_port_private *port_priv;
1292 	struct ib_mad_mgmt_class_table **class;
1293 	struct ib_mad_mgmt_method_table **method;
1294 	int i, ret;
1295 
1296 	port_priv = agent_priv->qp_info->port_priv;
1297 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1298 	if (!*class) {
1299 		/* Allocate management class table for "new" class version */
1300 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1301 		if (!*class) {
1302 			printk(KERN_ERR PFX "No memory for "
1303 			       "ib_mad_mgmt_class_table\n");
1304 			ret = -ENOMEM;
1305 			goto error1;
1306 		}
1307 
1308 		/* Allocate method table for this management class */
1309 		method = &(*class)->method_table[mgmt_class];
1310 		if ((ret = allocate_method_table(method)))
1311 			goto error2;
1312 	} else {
1313 		method = &(*class)->method_table[mgmt_class];
1314 		if (!*method) {
1315 			/* Allocate method table for this management class */
1316 			if ((ret = allocate_method_table(method)))
1317 				goto error1;
1318 		}
1319 	}
1320 
1321 	/* Now, make sure methods are not already in use */
1322 	if (method_in_use(method, mad_reg_req))
1323 		goto error3;
1324 
1325 	/* Finally, add in methods being registered */
1326 	for (i = find_first_bit(mad_reg_req->method_mask,
1327 				IB_MGMT_MAX_METHODS);
1328 	     i < IB_MGMT_MAX_METHODS;
1329 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1330 			       1+i)) {
1331 		(*method)->agent[i] = agent_priv;
1332 	}
1333 	return 0;
1334 
1335 error3:
1336 	/* Remove any methods for this mad agent */
1337 	remove_methods_mad_agent(*method, agent_priv);
1338 	/* Now, check to see if there are any methods in use */
1339 	if (!check_method_table(*method)) {
1340 		/* If not, release management method table */
1341 		kfree(*method);
1342 		*method = NULL;
1343 	}
1344 	ret = -EINVAL;
1345 	goto error1;
1346 error2:
1347 	kfree(*class);
1348 	*class = NULL;
1349 error1:
1350 	return ret;
1351 }
1352 
1353 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1354 			   struct ib_mad_agent_private *agent_priv)
1355 {
1356 	struct ib_mad_port_private *port_priv;
1357 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1358 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1359 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1360 	struct ib_mad_mgmt_method_table **method;
1361 	int i, ret = -ENOMEM;
1362 	u8 vclass;
1363 
1364 	/* "New" vendor (with OUI) class */
1365 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1366 	port_priv = agent_priv->qp_info->port_priv;
1367 	vendor_table = &port_priv->version[
1368 				mad_reg_req->mgmt_class_version].vendor;
1369 	if (!*vendor_table) {
1370 		/* Allocate mgmt vendor class table for "new" class version */
1371 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1372 		if (!vendor) {
1373 			printk(KERN_ERR PFX "No memory for "
1374 			       "ib_mad_mgmt_vendor_class_table\n");
1375 			goto error1;
1376 		}
1377 
1378 		*vendor_table = vendor;
1379 	}
1380 	if (!(*vendor_table)->vendor_class[vclass]) {
1381 		/* Allocate table for this management vendor class */
1382 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1383 		if (!vendor_class) {
1384 			printk(KERN_ERR PFX "No memory for "
1385 			       "ib_mad_mgmt_vendor_class\n");
1386 			goto error2;
1387 		}
1388 
1389 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1390 	}
1391 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1392 		/* Is there matching OUI for this vendor class ? */
1393 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1394 			    mad_reg_req->oui, 3)) {
1395 			method = &(*vendor_table)->vendor_class[
1396 						vclass]->method_table[i];
1397 			BUG_ON(!*method);
1398 			goto check_in_use;
1399 		}
1400 	}
1401 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1402 		/* OUI slot available ? */
1403 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1404 				vclass]->oui[i])) {
1405 			method = &(*vendor_table)->vendor_class[
1406 				vclass]->method_table[i];
1407 			BUG_ON(*method);
1408 			/* Allocate method table for this OUI */
1409 			if ((ret = allocate_method_table(method)))
1410 				goto error3;
1411 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1412 			       mad_reg_req->oui, 3);
1413 			goto check_in_use;
1414 		}
1415 	}
1416 	printk(KERN_ERR PFX "All OUI slots in use\n");
1417 	goto error3;
1418 
1419 check_in_use:
1420 	/* Now, make sure methods are not already in use */
1421 	if (method_in_use(method, mad_reg_req))
1422 		goto error4;
1423 
1424 	/* Finally, add in methods being registered */
1425 	for (i = find_first_bit(mad_reg_req->method_mask,
1426 				IB_MGMT_MAX_METHODS);
1427 	     i < IB_MGMT_MAX_METHODS;
1428 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1429 			       1+i)) {
1430 		(*method)->agent[i] = agent_priv;
1431 	}
1432 	return 0;
1433 
1434 error4:
1435 	/* Remove any methods for this mad agent */
1436 	remove_methods_mad_agent(*method, agent_priv);
1437 	/* Now, check to see if there are any methods in use */
1438 	if (!check_method_table(*method)) {
1439 		/* If not, release management method table */
1440 		kfree(*method);
1441 		*method = NULL;
1442 	}
1443 	ret = -EINVAL;
1444 error3:
1445 	if (vendor_class) {
1446 		(*vendor_table)->vendor_class[vclass] = NULL;
1447 		kfree(vendor_class);
1448 	}
1449 error2:
1450 	if (vendor) {
1451 		*vendor_table = NULL;
1452 		kfree(vendor);
1453 	}
1454 error1:
1455 	return ret;
1456 }
1457 
1458 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1459 {
1460 	struct ib_mad_port_private *port_priv;
1461 	struct ib_mad_mgmt_class_table *class;
1462 	struct ib_mad_mgmt_method_table *method;
1463 	struct ib_mad_mgmt_vendor_class_table *vendor;
1464 	struct ib_mad_mgmt_vendor_class *vendor_class;
1465 	int index;
1466 	u8 mgmt_class;
1467 
1468 	/*
1469 	 * Was MAD registration request supplied
1470 	 * with original registration ?
1471 	 */
1472 	if (!agent_priv->reg_req) {
1473 		goto out;
1474 	}
1475 
1476 	port_priv = agent_priv->qp_info->port_priv;
1477 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1478 	class = port_priv->version[
1479 			agent_priv->reg_req->mgmt_class_version].class;
1480 	if (!class)
1481 		goto vendor_check;
1482 
1483 	method = class->method_table[mgmt_class];
1484 	if (method) {
1485 		/* Remove any methods for this mad agent */
1486 		remove_methods_mad_agent(method, agent_priv);
1487 		/* Now, check to see if there are any methods still in use */
1488 		if (!check_method_table(method)) {
1489 			/* If not, release management method table */
1490 			 kfree(method);
1491 			 class->method_table[mgmt_class] = NULL;
1492 			 /* Any management classes left ? */
1493 			if (!check_class_table(class)) {
1494 				/* If not, release management class table */
1495 				kfree(class);
1496 				port_priv->version[
1497 					agent_priv->reg_req->
1498 					mgmt_class_version].class = NULL;
1499 			}
1500 		}
1501 	}
1502 
1503 vendor_check:
1504 	if (!is_vendor_class(mgmt_class))
1505 		goto out;
1506 
1507 	/* normalize mgmt_class to vendor range 2 */
1508 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1509 	vendor = port_priv->version[
1510 			agent_priv->reg_req->mgmt_class_version].vendor;
1511 
1512 	if (!vendor)
1513 		goto out;
1514 
1515 	vendor_class = vendor->vendor_class[mgmt_class];
1516 	if (vendor_class) {
1517 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1518 		if (index < 0)
1519 			goto out;
1520 		method = vendor_class->method_table[index];
1521 		if (method) {
1522 			/* Remove any methods for this mad agent */
1523 			remove_methods_mad_agent(method, agent_priv);
1524 			/*
1525 			 * Now, check to see if there are
1526 			 * any methods still in use
1527 			 */
1528 			if (!check_method_table(method)) {
1529 				/* If not, release management method table */
1530 				kfree(method);
1531 				vendor_class->method_table[index] = NULL;
1532 				memset(vendor_class->oui[index], 0, 3);
1533 				/* Any OUIs left ? */
1534 				if (!check_vendor_class(vendor_class)) {
1535 					/* If not, release vendor class table */
1536 					kfree(vendor_class);
1537 					vendor->vendor_class[mgmt_class] = NULL;
1538 					/* Any other vendor classes left ? */
1539 					if (!check_vendor_table(vendor)) {
1540 						kfree(vendor);
1541 						port_priv->version[
1542 							agent_priv->reg_req->
1543 							mgmt_class_version].
1544 							vendor = NULL;
1545 					}
1546 				}
1547 			}
1548 		}
1549 	}
1550 
1551 out:
1552 	return;
1553 }
1554 
1555 static struct ib_mad_agent_private *
1556 find_mad_agent(struct ib_mad_port_private *port_priv,
1557 	       struct ib_mad *mad)
1558 {
1559 	struct ib_mad_agent_private *mad_agent = NULL;
1560 	unsigned long flags;
1561 
1562 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1563 	if (ib_response_mad(mad)) {
1564 		u32 hi_tid;
1565 		struct ib_mad_agent_private *entry;
1566 
1567 		/*
1568 		 * Routing is based on high 32 bits of transaction ID
1569 		 * of MAD.
1570 		 */
1571 		hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1572 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1573 			if (entry->agent.hi_tid == hi_tid) {
1574 				mad_agent = entry;
1575 				break;
1576 			}
1577 		}
1578 	} else {
1579 		struct ib_mad_mgmt_class_table *class;
1580 		struct ib_mad_mgmt_method_table *method;
1581 		struct ib_mad_mgmt_vendor_class_table *vendor;
1582 		struct ib_mad_mgmt_vendor_class *vendor_class;
1583 		struct ib_vendor_mad *vendor_mad;
1584 		int index;
1585 
1586 		/*
1587 		 * Routing is based on version, class, and method
1588 		 * For "newer" vendor MADs, also based on OUI
1589 		 */
1590 		if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1591 			goto out;
1592 		if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1593 			class = port_priv->version[
1594 					mad->mad_hdr.class_version].class;
1595 			if (!class)
1596 				goto out;
1597 			method = class->method_table[convert_mgmt_class(
1598 							mad->mad_hdr.mgmt_class)];
1599 			if (method)
1600 				mad_agent = method->agent[mad->mad_hdr.method &
1601 							  ~IB_MGMT_METHOD_RESP];
1602 		} else {
1603 			vendor = port_priv->version[
1604 					mad->mad_hdr.class_version].vendor;
1605 			if (!vendor)
1606 				goto out;
1607 			vendor_class = vendor->vendor_class[vendor_class_index(
1608 						mad->mad_hdr.mgmt_class)];
1609 			if (!vendor_class)
1610 				goto out;
1611 			/* Find matching OUI */
1612 			vendor_mad = (struct ib_vendor_mad *)mad;
1613 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1614 			if (index == -1)
1615 				goto out;
1616 			method = vendor_class->method_table[index];
1617 			if (method) {
1618 				mad_agent = method->agent[mad->mad_hdr.method &
1619 							  ~IB_MGMT_METHOD_RESP];
1620 			}
1621 		}
1622 	}
1623 
1624 	if (mad_agent) {
1625 		if (mad_agent->agent.recv_handler)
1626 			atomic_inc(&mad_agent->refcount);
1627 		else {
1628 			printk(KERN_NOTICE PFX "No receive handler for client "
1629 			       "%p on port %d\n",
1630 			       &mad_agent->agent, port_priv->port_num);
1631 			mad_agent = NULL;
1632 		}
1633 	}
1634 out:
1635 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1636 
1637 	return mad_agent;
1638 }
1639 
1640 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1641 {
1642 	int valid = 0;
1643 
1644 	/* Make sure MAD base version is understood */
1645 	if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1646 		printk(KERN_ERR PFX "MAD received with unsupported base "
1647 		       "version %d\n", mad->mad_hdr.base_version);
1648 		goto out;
1649 	}
1650 
1651 	/* Filter SMI packets sent to other than QP0 */
1652 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1653 	    (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1654 		if (qp_num == 0)
1655 			valid = 1;
1656 	} else {
1657 		/* Filter GSI packets sent to QP0 */
1658 		if (qp_num != 0)
1659 			valid = 1;
1660 	}
1661 
1662 out:
1663 	return valid;
1664 }
1665 
1666 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1667 		       struct ib_mad_hdr *mad_hdr)
1668 {
1669 	struct ib_rmpp_mad *rmpp_mad;
1670 
1671 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1672 	return !mad_agent_priv->agent.rmpp_version ||
1673 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1674 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1675 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1676 }
1677 
1678 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1679 				     struct ib_mad_recv_wc *rwc)
1680 {
1681 	return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1682 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1683 }
1684 
1685 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1686 				   struct ib_mad_send_wr_private *wr,
1687 				   struct ib_mad_recv_wc *rwc )
1688 {
1689 	struct ib_ah_attr attr;
1690 	u8 send_resp, rcv_resp;
1691 	union ib_gid sgid;
1692 	struct ib_device *device = mad_agent_priv->agent.device;
1693 	u8 port_num = mad_agent_priv->agent.port_num;
1694 	u8 lmc;
1695 
1696 	send_resp = ib_response_mad((struct ib_mad *)wr->send_buf.mad);
1697 	rcv_resp = ib_response_mad(rwc->recv_buf.mad);
1698 
1699 	if (send_resp == rcv_resp)
1700 		/* both requests, or both responses. GIDs different */
1701 		return 0;
1702 
1703 	if (ib_query_ah(wr->send_buf.ah, &attr))
1704 		/* Assume not equal, to avoid false positives. */
1705 		return 0;
1706 
1707 	if (!!(attr.ah_flags & IB_AH_GRH) !=
1708 	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1709 		/* one has GID, other does not.  Assume different */
1710 		return 0;
1711 
1712 	if (!send_resp && rcv_resp) {
1713 		/* is request/response. */
1714 		if (!(attr.ah_flags & IB_AH_GRH)) {
1715 			if (ib_get_cached_lmc(device, port_num, &lmc))
1716 				return 0;
1717 			return (!lmc || !((attr.src_path_bits ^
1718 					   rwc->wc->dlid_path_bits) &
1719 					  ((1 << lmc) - 1)));
1720 		} else {
1721 			if (ib_get_cached_gid(device, port_num,
1722 					      attr.grh.sgid_index, &sgid))
1723 				return 0;
1724 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1725 				       16);
1726 		}
1727 	}
1728 
1729 	if (!(attr.ah_flags & IB_AH_GRH))
1730 		return attr.dlid == rwc->wc->slid;
1731 	else
1732 		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1733 			       16);
1734 }
1735 
1736 static inline int is_direct(u8 class)
1737 {
1738 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1739 }
1740 
1741 struct ib_mad_send_wr_private*
1742 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1743 		 struct ib_mad_recv_wc *wc)
1744 {
1745 	struct ib_mad_send_wr_private *wr;
1746 	struct ib_mad *mad;
1747 
1748 	mad = (struct ib_mad *)wc->recv_buf.mad;
1749 
1750 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1751 		if ((wr->tid == mad->mad_hdr.tid) &&
1752 		    rcv_has_same_class(wr, wc) &&
1753 		    /*
1754 		     * Don't check GID for direct routed MADs.
1755 		     * These might have permissive LIDs.
1756 		     */
1757 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1758 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1759 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1760 	}
1761 
1762 	/*
1763 	 * It's possible to receive the response before we've
1764 	 * been notified that the send has completed
1765 	 */
1766 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1767 		if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1768 		    wr->tid == mad->mad_hdr.tid &&
1769 		    wr->timeout &&
1770 		    rcv_has_same_class(wr, wc) &&
1771 		    /*
1772 		     * Don't check GID for direct routed MADs.
1773 		     * These might have permissive LIDs.
1774 		     */
1775 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1776 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1777 			/* Verify request has not been canceled */
1778 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1779 	}
1780 	return NULL;
1781 }
1782 
1783 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1784 {
1785 	mad_send_wr->timeout = 0;
1786 	if (mad_send_wr->refcount == 1)
1787 		list_move_tail(&mad_send_wr->agent_list,
1788 			      &mad_send_wr->mad_agent_priv->done_list);
1789 }
1790 
1791 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1792 				 struct ib_mad_recv_wc *mad_recv_wc)
1793 {
1794 	struct ib_mad_send_wr_private *mad_send_wr;
1795 	struct ib_mad_send_wc mad_send_wc;
1796 	unsigned long flags;
1797 
1798 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1799 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1800 	if (mad_agent_priv->agent.rmpp_version) {
1801 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1802 						      mad_recv_wc);
1803 		if (!mad_recv_wc) {
1804 			deref_mad_agent(mad_agent_priv);
1805 			return;
1806 		}
1807 	}
1808 
1809 	/* Complete corresponding request */
1810 	if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1811 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1812 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1813 		if (!mad_send_wr) {
1814 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1815 			ib_free_recv_mad(mad_recv_wc);
1816 			deref_mad_agent(mad_agent_priv);
1817 			return;
1818 		}
1819 		ib_mark_mad_done(mad_send_wr);
1820 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1821 
1822 		/* Defined behavior is to complete response before request */
1823 		mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1824 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1825 						   mad_recv_wc);
1826 		atomic_dec(&mad_agent_priv->refcount);
1827 
1828 		mad_send_wc.status = IB_WC_SUCCESS;
1829 		mad_send_wc.vendor_err = 0;
1830 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
1831 		ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1832 	} else {
1833 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1834 						   mad_recv_wc);
1835 		deref_mad_agent(mad_agent_priv);
1836 	}
1837 }
1838 
1839 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1840 				     struct ib_wc *wc)
1841 {
1842 	struct ib_mad_qp_info *qp_info;
1843 	struct ib_mad_private_header *mad_priv_hdr;
1844 	struct ib_mad_private *recv, *response = NULL;
1845 	struct ib_mad_list_head *mad_list;
1846 	struct ib_mad_agent_private *mad_agent;
1847 	int port_num;
1848 
1849 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1850 	qp_info = mad_list->mad_queue->qp_info;
1851 	dequeue_mad(mad_list);
1852 
1853 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1854 				    mad_list);
1855 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1856 	ib_dma_unmap_single(port_priv->device,
1857 			    recv->header.mapping,
1858 			    sizeof(struct ib_mad_private) -
1859 			      sizeof(struct ib_mad_private_header),
1860 			    DMA_FROM_DEVICE);
1861 
1862 	/* Setup MAD receive work completion from "normal" work completion */
1863 	recv->header.wc = *wc;
1864 	recv->header.recv_wc.wc = &recv->header.wc;
1865 	recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1866 	recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1867 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
1868 
1869 	if (atomic_read(&qp_info->snoop_count))
1870 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1871 
1872 	/* Validate MAD */
1873 	if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1874 		goto out;
1875 
1876 	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1877 	if (!response) {
1878 		printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1879 		       "for response buffer\n");
1880 		goto out;
1881 	}
1882 
1883 	if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH)
1884 		port_num = wc->port_num;
1885 	else
1886 		port_num = port_priv->port_num;
1887 
1888 	if (recv->mad.mad.mad_hdr.mgmt_class ==
1889 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1890 		enum smi_forward_action retsmi;
1891 
1892 		if (smi_handle_dr_smp_recv(&recv->mad.smp,
1893 					   port_priv->device->node_type,
1894 					   port_num,
1895 					   port_priv->device->phys_port_cnt) ==
1896 					   IB_SMI_DISCARD)
1897 			goto out;
1898 
1899 		retsmi = smi_check_forward_dr_smp(&recv->mad.smp);
1900 		if (retsmi == IB_SMI_LOCAL)
1901 			goto local;
1902 
1903 		if (retsmi == IB_SMI_SEND) { /* don't forward */
1904 			if (smi_handle_dr_smp_send(&recv->mad.smp,
1905 						   port_priv->device->node_type,
1906 						   port_num) == IB_SMI_DISCARD)
1907 				goto out;
1908 
1909 			if (smi_check_local_smp(&recv->mad.smp, port_priv->device) == IB_SMI_DISCARD)
1910 				goto out;
1911 		} else if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH) {
1912 			/* forward case for switches */
1913 			memcpy(response, recv, sizeof(*response));
1914 			response->header.recv_wc.wc = &response->header.wc;
1915 			response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1916 			response->header.recv_wc.recv_buf.grh = &response->grh;
1917 
1918 			agent_send_response(&response->mad.mad,
1919 					    &response->grh, wc,
1920 					    port_priv->device,
1921 					    smi_get_fwd_port(&recv->mad.smp),
1922 					    qp_info->qp->qp_num);
1923 
1924 			goto out;
1925 		}
1926 	}
1927 
1928 local:
1929 	/* Give driver "right of first refusal" on incoming MAD */
1930 	if (port_priv->device->process_mad) {
1931 		int ret;
1932 
1933 		ret = port_priv->device->process_mad(port_priv->device, 0,
1934 						     port_priv->port_num,
1935 						     wc, &recv->grh,
1936 						     &recv->mad.mad,
1937 						     &response->mad.mad);
1938 		if (ret & IB_MAD_RESULT_SUCCESS) {
1939 			if (ret & IB_MAD_RESULT_CONSUMED)
1940 				goto out;
1941 			if (ret & IB_MAD_RESULT_REPLY) {
1942 				agent_send_response(&response->mad.mad,
1943 						    &recv->grh, wc,
1944 						    port_priv->device,
1945 						    port_num,
1946 						    qp_info->qp->qp_num);
1947 				goto out;
1948 			}
1949 		}
1950 	}
1951 
1952 	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1953 	if (mad_agent) {
1954 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1955 		/*
1956 		 * recv is freed up in error cases in ib_mad_complete_recv
1957 		 * or via recv_handler in ib_mad_complete_recv()
1958 		 */
1959 		recv = NULL;
1960 	}
1961 
1962 out:
1963 	/* Post another receive request for this QP */
1964 	if (response) {
1965 		ib_mad_post_receive_mads(qp_info, response);
1966 		if (recv)
1967 			kmem_cache_free(ib_mad_cache, recv);
1968 	} else
1969 		ib_mad_post_receive_mads(qp_info, recv);
1970 }
1971 
1972 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1973 {
1974 	struct ib_mad_send_wr_private *mad_send_wr;
1975 	unsigned long delay;
1976 
1977 	if (list_empty(&mad_agent_priv->wait_list)) {
1978 		cancel_delayed_work(&mad_agent_priv->timed_work);
1979 	} else {
1980 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1981 					 struct ib_mad_send_wr_private,
1982 					 agent_list);
1983 
1984 		if (time_after(mad_agent_priv->timeout,
1985 			       mad_send_wr->timeout)) {
1986 			mad_agent_priv->timeout = mad_send_wr->timeout;
1987 			cancel_delayed_work(&mad_agent_priv->timed_work);
1988 			delay = mad_send_wr->timeout - jiffies;
1989 			if ((long)delay <= 0)
1990 				delay = 1;
1991 			queue_delayed_work(mad_agent_priv->qp_info->
1992 					   port_priv->wq,
1993 					   &mad_agent_priv->timed_work, delay);
1994 		}
1995 	}
1996 }
1997 
1998 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1999 {
2000 	struct ib_mad_agent_private *mad_agent_priv;
2001 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2002 	struct list_head *list_item;
2003 	unsigned long delay;
2004 
2005 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2006 	list_del(&mad_send_wr->agent_list);
2007 
2008 	delay = mad_send_wr->timeout;
2009 	mad_send_wr->timeout += jiffies;
2010 
2011 	if (delay) {
2012 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2013 			temp_mad_send_wr = list_entry(list_item,
2014 						struct ib_mad_send_wr_private,
2015 						agent_list);
2016 			if (time_after(mad_send_wr->timeout,
2017 				       temp_mad_send_wr->timeout))
2018 				break;
2019 		}
2020 	}
2021 	else
2022 		list_item = &mad_agent_priv->wait_list;
2023 	list_add(&mad_send_wr->agent_list, list_item);
2024 
2025 	/* Reschedule a work item if we have a shorter timeout */
2026 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
2027 		cancel_delayed_work(&mad_agent_priv->timed_work);
2028 		queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2029 				   &mad_agent_priv->timed_work, delay);
2030 	}
2031 }
2032 
2033 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2034 			  int timeout_ms)
2035 {
2036 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2037 	wait_for_response(mad_send_wr);
2038 }
2039 
2040 /*
2041  * Process a send work completion
2042  */
2043 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2044 			     struct ib_mad_send_wc *mad_send_wc)
2045 {
2046 	struct ib_mad_agent_private	*mad_agent_priv;
2047 	unsigned long			flags;
2048 	int				ret;
2049 
2050 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2051 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2052 	if (mad_agent_priv->agent.rmpp_version) {
2053 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2054 		if (ret == IB_RMPP_RESULT_CONSUMED)
2055 			goto done;
2056 	} else
2057 		ret = IB_RMPP_RESULT_UNHANDLED;
2058 
2059 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2060 	    mad_send_wr->status == IB_WC_SUCCESS) {
2061 		mad_send_wr->status = mad_send_wc->status;
2062 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2063 	}
2064 
2065 	if (--mad_send_wr->refcount > 0) {
2066 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2067 		    mad_send_wr->status == IB_WC_SUCCESS) {
2068 			wait_for_response(mad_send_wr);
2069 		}
2070 		goto done;
2071 	}
2072 
2073 	/* Remove send from MAD agent and notify client of completion */
2074 	list_del(&mad_send_wr->agent_list);
2075 	adjust_timeout(mad_agent_priv);
2076 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2077 
2078 	if (mad_send_wr->status != IB_WC_SUCCESS )
2079 		mad_send_wc->status = mad_send_wr->status;
2080 	if (ret == IB_RMPP_RESULT_INTERNAL)
2081 		ib_rmpp_send_handler(mad_send_wc);
2082 	else
2083 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2084 						   mad_send_wc);
2085 
2086 	/* Release reference on agent taken when sending */
2087 	deref_mad_agent(mad_agent_priv);
2088 	return;
2089 done:
2090 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2091 }
2092 
2093 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2094 				     struct ib_wc *wc)
2095 {
2096 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2097 	struct ib_mad_list_head		*mad_list;
2098 	struct ib_mad_qp_info		*qp_info;
2099 	struct ib_mad_queue		*send_queue;
2100 	struct ib_send_wr		*bad_send_wr;
2101 	struct ib_mad_send_wc		mad_send_wc;
2102 	unsigned long flags;
2103 	int ret;
2104 
2105 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2106 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2107 				   mad_list);
2108 	send_queue = mad_list->mad_queue;
2109 	qp_info = send_queue->qp_info;
2110 
2111 retry:
2112 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2113 			    mad_send_wr->header_mapping,
2114 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2115 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2116 			    mad_send_wr->payload_mapping,
2117 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2118 	queued_send_wr = NULL;
2119 	spin_lock_irqsave(&send_queue->lock, flags);
2120 	list_del(&mad_list->list);
2121 
2122 	/* Move queued send to the send queue */
2123 	if (send_queue->count-- > send_queue->max_active) {
2124 		mad_list = container_of(qp_info->overflow_list.next,
2125 					struct ib_mad_list_head, list);
2126 		queued_send_wr = container_of(mad_list,
2127 					struct ib_mad_send_wr_private,
2128 					mad_list);
2129 		list_move_tail(&mad_list->list, &send_queue->list);
2130 	}
2131 	spin_unlock_irqrestore(&send_queue->lock, flags);
2132 
2133 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2134 	mad_send_wc.status = wc->status;
2135 	mad_send_wc.vendor_err = wc->vendor_err;
2136 	if (atomic_read(&qp_info->snoop_count))
2137 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2138 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2139 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2140 
2141 	if (queued_send_wr) {
2142 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2143 				   &bad_send_wr);
2144 		if (ret) {
2145 			printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2146 			mad_send_wr = queued_send_wr;
2147 			wc->status = IB_WC_LOC_QP_OP_ERR;
2148 			goto retry;
2149 		}
2150 	}
2151 }
2152 
2153 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2154 {
2155 	struct ib_mad_send_wr_private *mad_send_wr;
2156 	struct ib_mad_list_head *mad_list;
2157 	unsigned long flags;
2158 
2159 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2160 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2161 		mad_send_wr = container_of(mad_list,
2162 					   struct ib_mad_send_wr_private,
2163 					   mad_list);
2164 		mad_send_wr->retry = 1;
2165 	}
2166 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2167 }
2168 
2169 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2170 			      struct ib_wc *wc)
2171 {
2172 	struct ib_mad_list_head *mad_list;
2173 	struct ib_mad_qp_info *qp_info;
2174 	struct ib_mad_send_wr_private *mad_send_wr;
2175 	int ret;
2176 
2177 	/* Determine if failure was a send or receive */
2178 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2179 	qp_info = mad_list->mad_queue->qp_info;
2180 	if (mad_list->mad_queue == &qp_info->recv_queue)
2181 		/*
2182 		 * Receive errors indicate that the QP has entered the error
2183 		 * state - error handling/shutdown code will cleanup
2184 		 */
2185 		return;
2186 
2187 	/*
2188 	 * Send errors will transition the QP to SQE - move
2189 	 * QP to RTS and repost flushed work requests
2190 	 */
2191 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2192 				   mad_list);
2193 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2194 		if (mad_send_wr->retry) {
2195 			/* Repost send */
2196 			struct ib_send_wr *bad_send_wr;
2197 
2198 			mad_send_wr->retry = 0;
2199 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2200 					&bad_send_wr);
2201 			if (ret)
2202 				ib_mad_send_done_handler(port_priv, wc);
2203 		} else
2204 			ib_mad_send_done_handler(port_priv, wc);
2205 	} else {
2206 		struct ib_qp_attr *attr;
2207 
2208 		/* Transition QP to RTS and fail offending send */
2209 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2210 		if (attr) {
2211 			attr->qp_state = IB_QPS_RTS;
2212 			attr->cur_qp_state = IB_QPS_SQE;
2213 			ret = ib_modify_qp(qp_info->qp, attr,
2214 					   IB_QP_STATE | IB_QP_CUR_STATE);
2215 			kfree(attr);
2216 			if (ret)
2217 				printk(KERN_ERR PFX "mad_error_handler - "
2218 				       "ib_modify_qp to RTS : %d\n", ret);
2219 			else
2220 				mark_sends_for_retry(qp_info);
2221 		}
2222 		ib_mad_send_done_handler(port_priv, wc);
2223 	}
2224 }
2225 
2226 /*
2227  * IB MAD completion callback
2228  */
2229 static void ib_mad_completion_handler(struct work_struct *work)
2230 {
2231 	struct ib_mad_port_private *port_priv;
2232 	struct ib_wc wc;
2233 
2234 	port_priv = container_of(work, struct ib_mad_port_private, work);
2235 	ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2236 
2237 	while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2238 		if (wc.status == IB_WC_SUCCESS) {
2239 			switch (wc.opcode) {
2240 			case IB_WC_SEND:
2241 				ib_mad_send_done_handler(port_priv, &wc);
2242 				break;
2243 			case IB_WC_RECV:
2244 				ib_mad_recv_done_handler(port_priv, &wc);
2245 				break;
2246 			default:
2247 				BUG_ON(1);
2248 				break;
2249 			}
2250 		} else
2251 			mad_error_handler(port_priv, &wc);
2252 	}
2253 }
2254 
2255 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2256 {
2257 	unsigned long flags;
2258 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2259 	struct ib_mad_send_wc mad_send_wc;
2260 	struct list_head cancel_list;
2261 
2262 	INIT_LIST_HEAD(&cancel_list);
2263 
2264 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2265 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2266 				 &mad_agent_priv->send_list, agent_list) {
2267 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2268 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2269 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2270 		}
2271 	}
2272 
2273 	/* Empty wait list to prevent receives from finding a request */
2274 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2275 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2276 
2277 	/* Report all cancelled requests */
2278 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2279 	mad_send_wc.vendor_err = 0;
2280 
2281 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2282 				 &cancel_list, agent_list) {
2283 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2284 		list_del(&mad_send_wr->agent_list);
2285 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2286 						   &mad_send_wc);
2287 		atomic_dec(&mad_agent_priv->refcount);
2288 	}
2289 }
2290 
2291 static struct ib_mad_send_wr_private*
2292 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2293 	     struct ib_mad_send_buf *send_buf)
2294 {
2295 	struct ib_mad_send_wr_private *mad_send_wr;
2296 
2297 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2298 			    agent_list) {
2299 		if (&mad_send_wr->send_buf == send_buf)
2300 			return mad_send_wr;
2301 	}
2302 
2303 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2304 			    agent_list) {
2305 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2306 		    &mad_send_wr->send_buf == send_buf)
2307 			return mad_send_wr;
2308 	}
2309 	return NULL;
2310 }
2311 
2312 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2313 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2314 {
2315 	struct ib_mad_agent_private *mad_agent_priv;
2316 	struct ib_mad_send_wr_private *mad_send_wr;
2317 	unsigned long flags;
2318 	int active;
2319 
2320 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2321 				      agent);
2322 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2323 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2324 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2325 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2326 		return -EINVAL;
2327 	}
2328 
2329 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2330 	if (!timeout_ms) {
2331 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2332 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2333 	}
2334 
2335 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2336 	if (active)
2337 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2338 	else
2339 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2340 
2341 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2342 	return 0;
2343 }
2344 EXPORT_SYMBOL(ib_modify_mad);
2345 
2346 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2347 		   struct ib_mad_send_buf *send_buf)
2348 {
2349 	ib_modify_mad(mad_agent, send_buf, 0);
2350 }
2351 EXPORT_SYMBOL(ib_cancel_mad);
2352 
2353 static void local_completions(struct work_struct *work)
2354 {
2355 	struct ib_mad_agent_private *mad_agent_priv;
2356 	struct ib_mad_local_private *local;
2357 	struct ib_mad_agent_private *recv_mad_agent;
2358 	unsigned long flags;
2359 	int recv = 0;
2360 	struct ib_wc wc;
2361 	struct ib_mad_send_wc mad_send_wc;
2362 
2363 	mad_agent_priv =
2364 		container_of(work, struct ib_mad_agent_private, local_work);
2365 
2366 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2367 	while (!list_empty(&mad_agent_priv->local_list)) {
2368 		local = list_entry(mad_agent_priv->local_list.next,
2369 				   struct ib_mad_local_private,
2370 				   completion_list);
2371 		list_del(&local->completion_list);
2372 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2373 		if (local->mad_priv) {
2374 			recv_mad_agent = local->recv_mad_agent;
2375 			if (!recv_mad_agent) {
2376 				printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2377 				goto local_send_completion;
2378 			}
2379 
2380 			recv = 1;
2381 			/*
2382 			 * Defined behavior is to complete response
2383 			 * before request
2384 			 */
2385 			build_smp_wc(recv_mad_agent->agent.qp,
2386 				     (unsigned long) local->mad_send_wr,
2387 				     be16_to_cpu(IB_LID_PERMISSIVE),
2388 				     0, recv_mad_agent->agent.port_num, &wc);
2389 
2390 			local->mad_priv->header.recv_wc.wc = &wc;
2391 			local->mad_priv->header.recv_wc.mad_len =
2392 						sizeof(struct ib_mad);
2393 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2394 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2395 				 &local->mad_priv->header.recv_wc.rmpp_list);
2396 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2397 			local->mad_priv->header.recv_wc.recv_buf.mad =
2398 						&local->mad_priv->mad.mad;
2399 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2400 				snoop_recv(recv_mad_agent->qp_info,
2401 					  &local->mad_priv->header.recv_wc,
2402 					   IB_MAD_SNOOP_RECVS);
2403 			recv_mad_agent->agent.recv_handler(
2404 						&recv_mad_agent->agent,
2405 						&local->mad_priv->header.recv_wc);
2406 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2407 			atomic_dec(&recv_mad_agent->refcount);
2408 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2409 		}
2410 
2411 local_send_completion:
2412 		/* Complete send */
2413 		mad_send_wc.status = IB_WC_SUCCESS;
2414 		mad_send_wc.vendor_err = 0;
2415 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2416 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2417 			snoop_send(mad_agent_priv->qp_info,
2418 				   &local->mad_send_wr->send_buf,
2419 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2420 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2421 						   &mad_send_wc);
2422 
2423 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2424 		atomic_dec(&mad_agent_priv->refcount);
2425 		if (!recv)
2426 			kmem_cache_free(ib_mad_cache, local->mad_priv);
2427 		kfree(local);
2428 	}
2429 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2430 }
2431 
2432 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2433 {
2434 	int ret;
2435 
2436 	if (!mad_send_wr->retries_left)
2437 		return -ETIMEDOUT;
2438 
2439 	mad_send_wr->retries_left--;
2440 	mad_send_wr->send_buf.retries++;
2441 
2442 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2443 
2444 	if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2445 		ret = ib_retry_rmpp(mad_send_wr);
2446 		switch (ret) {
2447 		case IB_RMPP_RESULT_UNHANDLED:
2448 			ret = ib_send_mad(mad_send_wr);
2449 			break;
2450 		case IB_RMPP_RESULT_CONSUMED:
2451 			ret = 0;
2452 			break;
2453 		default:
2454 			ret = -ECOMM;
2455 			break;
2456 		}
2457 	} else
2458 		ret = ib_send_mad(mad_send_wr);
2459 
2460 	if (!ret) {
2461 		mad_send_wr->refcount++;
2462 		list_add_tail(&mad_send_wr->agent_list,
2463 			      &mad_send_wr->mad_agent_priv->send_list);
2464 	}
2465 	return ret;
2466 }
2467 
2468 static void timeout_sends(struct work_struct *work)
2469 {
2470 	struct ib_mad_agent_private *mad_agent_priv;
2471 	struct ib_mad_send_wr_private *mad_send_wr;
2472 	struct ib_mad_send_wc mad_send_wc;
2473 	unsigned long flags, delay;
2474 
2475 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2476 				      timed_work.work);
2477 	mad_send_wc.vendor_err = 0;
2478 
2479 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2480 	while (!list_empty(&mad_agent_priv->wait_list)) {
2481 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2482 					 struct ib_mad_send_wr_private,
2483 					 agent_list);
2484 
2485 		if (time_after(mad_send_wr->timeout, jiffies)) {
2486 			delay = mad_send_wr->timeout - jiffies;
2487 			if ((long)delay <= 0)
2488 				delay = 1;
2489 			queue_delayed_work(mad_agent_priv->qp_info->
2490 					   port_priv->wq,
2491 					   &mad_agent_priv->timed_work, delay);
2492 			break;
2493 		}
2494 
2495 		list_del(&mad_send_wr->agent_list);
2496 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2497 		    !retry_send(mad_send_wr))
2498 			continue;
2499 
2500 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2501 
2502 		if (mad_send_wr->status == IB_WC_SUCCESS)
2503 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2504 		else
2505 			mad_send_wc.status = mad_send_wr->status;
2506 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2507 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2508 						   &mad_send_wc);
2509 
2510 		atomic_dec(&mad_agent_priv->refcount);
2511 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2512 	}
2513 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2514 }
2515 
2516 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2517 {
2518 	struct ib_mad_port_private *port_priv = cq->cq_context;
2519 	unsigned long flags;
2520 
2521 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2522 	if (!list_empty(&port_priv->port_list))
2523 		queue_work(port_priv->wq, &port_priv->work);
2524 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2525 }
2526 
2527 /*
2528  * Allocate receive MADs and post receive WRs for them
2529  */
2530 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2531 				    struct ib_mad_private *mad)
2532 {
2533 	unsigned long flags;
2534 	int post, ret;
2535 	struct ib_mad_private *mad_priv;
2536 	struct ib_sge sg_list;
2537 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2538 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2539 
2540 	/* Initialize common scatter list fields */
2541 	sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2542 	sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2543 
2544 	/* Initialize common receive WR fields */
2545 	recv_wr.next = NULL;
2546 	recv_wr.sg_list = &sg_list;
2547 	recv_wr.num_sge = 1;
2548 
2549 	do {
2550 		/* Allocate and map receive buffer */
2551 		if (mad) {
2552 			mad_priv = mad;
2553 			mad = NULL;
2554 		} else {
2555 			mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2556 			if (!mad_priv) {
2557 				printk(KERN_ERR PFX "No memory for receive buffer\n");
2558 				ret = -ENOMEM;
2559 				break;
2560 			}
2561 		}
2562 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2563 						 &mad_priv->grh,
2564 						 sizeof *mad_priv -
2565 						   sizeof mad_priv->header,
2566 						 DMA_FROM_DEVICE);
2567 		mad_priv->header.mapping = sg_list.addr;
2568 		recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2569 		mad_priv->header.mad_list.mad_queue = recv_queue;
2570 
2571 		/* Post receive WR */
2572 		spin_lock_irqsave(&recv_queue->lock, flags);
2573 		post = (++recv_queue->count < recv_queue->max_active);
2574 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2575 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2576 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2577 		if (ret) {
2578 			spin_lock_irqsave(&recv_queue->lock, flags);
2579 			list_del(&mad_priv->header.mad_list.list);
2580 			recv_queue->count--;
2581 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2582 			ib_dma_unmap_single(qp_info->port_priv->device,
2583 					    mad_priv->header.mapping,
2584 					    sizeof *mad_priv -
2585 					      sizeof mad_priv->header,
2586 					    DMA_FROM_DEVICE);
2587 			kmem_cache_free(ib_mad_cache, mad_priv);
2588 			printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2589 			break;
2590 		}
2591 	} while (post);
2592 
2593 	return ret;
2594 }
2595 
2596 /*
2597  * Return all the posted receive MADs
2598  */
2599 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2600 {
2601 	struct ib_mad_private_header *mad_priv_hdr;
2602 	struct ib_mad_private *recv;
2603 	struct ib_mad_list_head *mad_list;
2604 
2605 	while (!list_empty(&qp_info->recv_queue.list)) {
2606 
2607 		mad_list = list_entry(qp_info->recv_queue.list.next,
2608 				      struct ib_mad_list_head, list);
2609 		mad_priv_hdr = container_of(mad_list,
2610 					    struct ib_mad_private_header,
2611 					    mad_list);
2612 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2613 				    header);
2614 
2615 		/* Remove from posted receive MAD list */
2616 		list_del(&mad_list->list);
2617 
2618 		ib_dma_unmap_single(qp_info->port_priv->device,
2619 				    recv->header.mapping,
2620 				    sizeof(struct ib_mad_private) -
2621 				      sizeof(struct ib_mad_private_header),
2622 				    DMA_FROM_DEVICE);
2623 		kmem_cache_free(ib_mad_cache, recv);
2624 	}
2625 
2626 	qp_info->recv_queue.count = 0;
2627 }
2628 
2629 /*
2630  * Start the port
2631  */
2632 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2633 {
2634 	int ret, i;
2635 	struct ib_qp_attr *attr;
2636 	struct ib_qp *qp;
2637 
2638 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2639 	if (!attr) {
2640 		printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2641 		return -ENOMEM;
2642 	}
2643 
2644 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2645 		qp = port_priv->qp_info[i].qp;
2646 		/*
2647 		 * PKey index for QP1 is irrelevant but
2648 		 * one is needed for the Reset to Init transition
2649 		 */
2650 		attr->qp_state = IB_QPS_INIT;
2651 		attr->pkey_index = 0;
2652 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2653 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2654 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2655 		if (ret) {
2656 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2657 			       "INIT: %d\n", i, ret);
2658 			goto out;
2659 		}
2660 
2661 		attr->qp_state = IB_QPS_RTR;
2662 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2663 		if (ret) {
2664 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2665 			       "RTR: %d\n", i, ret);
2666 			goto out;
2667 		}
2668 
2669 		attr->qp_state = IB_QPS_RTS;
2670 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2671 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2672 		if (ret) {
2673 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2674 			       "RTS: %d\n", i, ret);
2675 			goto out;
2676 		}
2677 	}
2678 
2679 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2680 	if (ret) {
2681 		printk(KERN_ERR PFX "Failed to request completion "
2682 		       "notification: %d\n", ret);
2683 		goto out;
2684 	}
2685 
2686 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2687 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2688 		if (ret) {
2689 			printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2690 			goto out;
2691 		}
2692 	}
2693 out:
2694 	kfree(attr);
2695 	return ret;
2696 }
2697 
2698 static void qp_event_handler(struct ib_event *event, void *qp_context)
2699 {
2700 	struct ib_mad_qp_info	*qp_info = qp_context;
2701 
2702 	/* It's worse than that! He's dead, Jim! */
2703 	printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2704 		event->event, qp_info->qp->qp_num);
2705 }
2706 
2707 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2708 			   struct ib_mad_queue *mad_queue)
2709 {
2710 	mad_queue->qp_info = qp_info;
2711 	mad_queue->count = 0;
2712 	spin_lock_init(&mad_queue->lock);
2713 	INIT_LIST_HEAD(&mad_queue->list);
2714 }
2715 
2716 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2717 			struct ib_mad_qp_info *qp_info)
2718 {
2719 	qp_info->port_priv = port_priv;
2720 	init_mad_queue(qp_info, &qp_info->send_queue);
2721 	init_mad_queue(qp_info, &qp_info->recv_queue);
2722 	INIT_LIST_HEAD(&qp_info->overflow_list);
2723 	spin_lock_init(&qp_info->snoop_lock);
2724 	qp_info->snoop_table = NULL;
2725 	qp_info->snoop_table_size = 0;
2726 	atomic_set(&qp_info->snoop_count, 0);
2727 }
2728 
2729 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2730 			 enum ib_qp_type qp_type)
2731 {
2732 	struct ib_qp_init_attr	qp_init_attr;
2733 	int ret;
2734 
2735 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2736 	qp_init_attr.send_cq = qp_info->port_priv->cq;
2737 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2738 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2739 	qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2740 	qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2741 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2742 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2743 	qp_init_attr.qp_type = qp_type;
2744 	qp_init_attr.port_num = qp_info->port_priv->port_num;
2745 	qp_init_attr.qp_context = qp_info;
2746 	qp_init_attr.event_handler = qp_event_handler;
2747 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2748 	if (IS_ERR(qp_info->qp)) {
2749 		printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2750 		       get_spl_qp_index(qp_type));
2751 		ret = PTR_ERR(qp_info->qp);
2752 		goto error;
2753 	}
2754 	/* Use minimum queue sizes unless the CQ is resized */
2755 	qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2756 	qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2757 	return 0;
2758 
2759 error:
2760 	return ret;
2761 }
2762 
2763 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2764 {
2765 	ib_destroy_qp(qp_info->qp);
2766 	kfree(qp_info->snoop_table);
2767 }
2768 
2769 /*
2770  * Open the port
2771  * Create the QP, PD, MR, and CQ if needed
2772  */
2773 static int ib_mad_port_open(struct ib_device *device,
2774 			    int port_num)
2775 {
2776 	int ret, cq_size;
2777 	struct ib_mad_port_private *port_priv;
2778 	unsigned long flags;
2779 	char name[sizeof "ib_mad123"];
2780 
2781 	/* Create new device info */
2782 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2783 	if (!port_priv) {
2784 		printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2785 		return -ENOMEM;
2786 	}
2787 
2788 	port_priv->device = device;
2789 	port_priv->port_num = port_num;
2790 	spin_lock_init(&port_priv->reg_lock);
2791 	INIT_LIST_HEAD(&port_priv->agent_list);
2792 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2793 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2794 
2795 	cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2796 	port_priv->cq = ib_create_cq(port_priv->device,
2797 				     ib_mad_thread_completion_handler,
2798 				     NULL, port_priv, cq_size, 0);
2799 	if (IS_ERR(port_priv->cq)) {
2800 		printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2801 		ret = PTR_ERR(port_priv->cq);
2802 		goto error3;
2803 	}
2804 
2805 	port_priv->pd = ib_alloc_pd(device);
2806 	if (IS_ERR(port_priv->pd)) {
2807 		printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2808 		ret = PTR_ERR(port_priv->pd);
2809 		goto error4;
2810 	}
2811 
2812 	port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2813 	if (IS_ERR(port_priv->mr)) {
2814 		printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2815 		ret = PTR_ERR(port_priv->mr);
2816 		goto error5;
2817 	}
2818 
2819 	ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2820 	if (ret)
2821 		goto error6;
2822 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2823 	if (ret)
2824 		goto error7;
2825 
2826 	snprintf(name, sizeof name, "ib_mad%d", port_num);
2827 	port_priv->wq = create_singlethread_workqueue(name);
2828 	if (!port_priv->wq) {
2829 		ret = -ENOMEM;
2830 		goto error8;
2831 	}
2832 	INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2833 
2834 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2835 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2836 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2837 
2838 	ret = ib_mad_port_start(port_priv);
2839 	if (ret) {
2840 		printk(KERN_ERR PFX "Couldn't start port\n");
2841 		goto error9;
2842 	}
2843 
2844 	return 0;
2845 
2846 error9:
2847 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2848 	list_del_init(&port_priv->port_list);
2849 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2850 
2851 	destroy_workqueue(port_priv->wq);
2852 error8:
2853 	destroy_mad_qp(&port_priv->qp_info[1]);
2854 error7:
2855 	destroy_mad_qp(&port_priv->qp_info[0]);
2856 error6:
2857 	ib_dereg_mr(port_priv->mr);
2858 error5:
2859 	ib_dealloc_pd(port_priv->pd);
2860 error4:
2861 	ib_destroy_cq(port_priv->cq);
2862 	cleanup_recv_queue(&port_priv->qp_info[1]);
2863 	cleanup_recv_queue(&port_priv->qp_info[0]);
2864 error3:
2865 	kfree(port_priv);
2866 
2867 	return ret;
2868 }
2869 
2870 /*
2871  * Close the port
2872  * If there are no classes using the port, free the port
2873  * resources (CQ, MR, PD, QP) and remove the port's info structure
2874  */
2875 static int ib_mad_port_close(struct ib_device *device, int port_num)
2876 {
2877 	struct ib_mad_port_private *port_priv;
2878 	unsigned long flags;
2879 
2880 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2881 	port_priv = __ib_get_mad_port(device, port_num);
2882 	if (port_priv == NULL) {
2883 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2884 		printk(KERN_ERR PFX "Port %d not found\n", port_num);
2885 		return -ENODEV;
2886 	}
2887 	list_del_init(&port_priv->port_list);
2888 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2889 
2890 	destroy_workqueue(port_priv->wq);
2891 	destroy_mad_qp(&port_priv->qp_info[1]);
2892 	destroy_mad_qp(&port_priv->qp_info[0]);
2893 	ib_dereg_mr(port_priv->mr);
2894 	ib_dealloc_pd(port_priv->pd);
2895 	ib_destroy_cq(port_priv->cq);
2896 	cleanup_recv_queue(&port_priv->qp_info[1]);
2897 	cleanup_recv_queue(&port_priv->qp_info[0]);
2898 	/* XXX: Handle deallocation of MAD registration tables */
2899 
2900 	kfree(port_priv);
2901 
2902 	return 0;
2903 }
2904 
2905 static void ib_mad_init_device(struct ib_device *device)
2906 {
2907 	int start, end, i;
2908 
2909 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2910 		return;
2911 
2912 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2913 		start = 0;
2914 		end   = 0;
2915 	} else {
2916 		start = 1;
2917 		end   = device->phys_port_cnt;
2918 	}
2919 
2920 	for (i = start; i <= end; i++) {
2921 		if (ib_mad_port_open(device, i)) {
2922 			printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2923 			       device->name, i);
2924 			goto error;
2925 		}
2926 		if (ib_agent_port_open(device, i)) {
2927 			printk(KERN_ERR PFX "Couldn't open %s port %d "
2928 			       "for agents\n",
2929 			       device->name, i);
2930 			goto error_agent;
2931 		}
2932 	}
2933 	return;
2934 
2935 error_agent:
2936 	if (ib_mad_port_close(device, i))
2937 		printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2938 		       device->name, i);
2939 
2940 error:
2941 	i--;
2942 
2943 	while (i >= start) {
2944 		if (ib_agent_port_close(device, i))
2945 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2946 			       "for agents\n",
2947 			       device->name, i);
2948 		if (ib_mad_port_close(device, i))
2949 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2950 			       device->name, i);
2951 		i--;
2952 	}
2953 }
2954 
2955 static void ib_mad_remove_device(struct ib_device *device)
2956 {
2957 	int i, num_ports, cur_port;
2958 
2959 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2960 		num_ports = 1;
2961 		cur_port = 0;
2962 	} else {
2963 		num_ports = device->phys_port_cnt;
2964 		cur_port = 1;
2965 	}
2966 	for (i = 0; i < num_ports; i++, cur_port++) {
2967 		if (ib_agent_port_close(device, cur_port))
2968 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2969 			       "for agents\n",
2970 			       device->name, cur_port);
2971 		if (ib_mad_port_close(device, cur_port))
2972 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2973 			       device->name, cur_port);
2974 	}
2975 }
2976 
2977 static struct ib_client mad_client = {
2978 	.name   = "mad",
2979 	.add = ib_mad_init_device,
2980 	.remove = ib_mad_remove_device
2981 };
2982 
2983 static int __init ib_mad_init_module(void)
2984 {
2985 	int ret;
2986 
2987 	spin_lock_init(&ib_mad_port_list_lock);
2988 
2989 	ib_mad_cache = kmem_cache_create("ib_mad",
2990 					 sizeof(struct ib_mad_private),
2991 					 0,
2992 					 SLAB_HWCACHE_ALIGN,
2993 					 NULL);
2994 	if (!ib_mad_cache) {
2995 		printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2996 		ret = -ENOMEM;
2997 		goto error1;
2998 	}
2999 
3000 	INIT_LIST_HEAD(&ib_mad_port_list);
3001 
3002 	if (ib_register_client(&mad_client)) {
3003 		printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
3004 		ret = -EINVAL;
3005 		goto error2;
3006 	}
3007 
3008 	return 0;
3009 
3010 error2:
3011 	kmem_cache_destroy(ib_mad_cache);
3012 error1:
3013 	return ret;
3014 }
3015 
3016 static void __exit ib_mad_cleanup_module(void)
3017 {
3018 	ib_unregister_client(&mad_client);
3019 	kmem_cache_destroy(ib_mad_cache);
3020 }
3021 
3022 module_init(ib_mad_init_module);
3023 module_exit(ib_mad_cleanup_module);
3024 
3025