xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision 87c2ce3b)
1 /*
2  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
3  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  *
34  * $Id: mad.c 2817 2005-07-07 11:29:26Z halr $
35  */
36 #include <linux/dma-mapping.h>
37 
38 #include "mad_priv.h"
39 #include "mad_rmpp.h"
40 #include "smi.h"
41 #include "agent.h"
42 
43 MODULE_LICENSE("Dual BSD/GPL");
44 MODULE_DESCRIPTION("kernel IB MAD API");
45 MODULE_AUTHOR("Hal Rosenstock");
46 MODULE_AUTHOR("Sean Hefty");
47 
48 
49 kmem_cache_t *ib_mad_cache;
50 
51 static struct list_head ib_mad_port_list;
52 static u32 ib_mad_client_id = 0;
53 
54 /* Port list lock */
55 static spinlock_t ib_mad_port_list_lock;
56 
57 
58 /* Forward declarations */
59 static int method_in_use(struct ib_mad_mgmt_method_table **method,
60 			 struct ib_mad_reg_req *mad_reg_req);
61 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
62 static struct ib_mad_agent_private *find_mad_agent(
63 					struct ib_mad_port_private *port_priv,
64 					struct ib_mad *mad);
65 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
66 				    struct ib_mad_private *mad);
67 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
68 static void timeout_sends(void *data);
69 static void local_completions(void *data);
70 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
71 			      struct ib_mad_agent_private *agent_priv,
72 			      u8 mgmt_class);
73 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
74 			   struct ib_mad_agent_private *agent_priv);
75 
76 /*
77  * Returns a ib_mad_port_private structure or NULL for a device/port
78  * Assumes ib_mad_port_list_lock is being held
79  */
80 static inline struct ib_mad_port_private *
81 __ib_get_mad_port(struct ib_device *device, int port_num)
82 {
83 	struct ib_mad_port_private *entry;
84 
85 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
86 		if (entry->device == device && entry->port_num == port_num)
87 			return entry;
88 	}
89 	return NULL;
90 }
91 
92 /*
93  * Wrapper function to return a ib_mad_port_private structure or NULL
94  * for a device/port
95  */
96 static inline struct ib_mad_port_private *
97 ib_get_mad_port(struct ib_device *device, int port_num)
98 {
99 	struct ib_mad_port_private *entry;
100 	unsigned long flags;
101 
102 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
103 	entry = __ib_get_mad_port(device, port_num);
104 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
105 
106 	return entry;
107 }
108 
109 static inline u8 convert_mgmt_class(u8 mgmt_class)
110 {
111 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
112 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
113 		0 : mgmt_class;
114 }
115 
116 static int get_spl_qp_index(enum ib_qp_type qp_type)
117 {
118 	switch (qp_type)
119 	{
120 	case IB_QPT_SMI:
121 		return 0;
122 	case IB_QPT_GSI:
123 		return 1;
124 	default:
125 		return -1;
126 	}
127 }
128 
129 static int vendor_class_index(u8 mgmt_class)
130 {
131 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
132 }
133 
134 static int is_vendor_class(u8 mgmt_class)
135 {
136 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
137 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
138 		return 0;
139 	return 1;
140 }
141 
142 static int is_vendor_oui(char *oui)
143 {
144 	if (oui[0] || oui[1] || oui[2])
145 		return 1;
146 	return 0;
147 }
148 
149 static int is_vendor_method_in_use(
150 		struct ib_mad_mgmt_vendor_class *vendor_class,
151 		struct ib_mad_reg_req *mad_reg_req)
152 {
153 	struct ib_mad_mgmt_method_table *method;
154 	int i;
155 
156 	for (i = 0; i < MAX_MGMT_OUI; i++) {
157 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
158 			method = vendor_class->method_table[i];
159 			if (method) {
160 				if (method_in_use(&method, mad_reg_req))
161 					return 1;
162 				else
163 					break;
164 			}
165 		}
166 	}
167 	return 0;
168 }
169 
170 /*
171  * ib_register_mad_agent - Register to send/receive MADs
172  */
173 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
174 					   u8 port_num,
175 					   enum ib_qp_type qp_type,
176 					   struct ib_mad_reg_req *mad_reg_req,
177 					   u8 rmpp_version,
178 					   ib_mad_send_handler send_handler,
179 					   ib_mad_recv_handler recv_handler,
180 					   void *context)
181 {
182 	struct ib_mad_port_private *port_priv;
183 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
184 	struct ib_mad_agent_private *mad_agent_priv;
185 	struct ib_mad_reg_req *reg_req = NULL;
186 	struct ib_mad_mgmt_class_table *class;
187 	struct ib_mad_mgmt_vendor_class_table *vendor;
188 	struct ib_mad_mgmt_vendor_class *vendor_class;
189 	struct ib_mad_mgmt_method_table *method;
190 	int ret2, qpn;
191 	unsigned long flags;
192 	u8 mgmt_class, vclass;
193 
194 	/* Validate parameters */
195 	qpn = get_spl_qp_index(qp_type);
196 	if (qpn == -1)
197 		goto error1;
198 
199 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
200 		goto error1;
201 
202 	/* Validate MAD registration request if supplied */
203 	if (mad_reg_req) {
204 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
205 			goto error1;
206 		if (!recv_handler)
207 			goto error1;
208 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
209 			/*
210 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
211 			 * one in this range currently allowed
212 			 */
213 			if (mad_reg_req->mgmt_class !=
214 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
215 				goto error1;
216 		} else if (mad_reg_req->mgmt_class == 0) {
217 			/*
218 			 * Class 0 is reserved in IBA and is used for
219 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
220 			 */
221 			goto error1;
222 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
223 			/*
224 			 * If class is in "new" vendor range,
225 			 * ensure supplied OUI is not zero
226 			 */
227 			if (!is_vendor_oui(mad_reg_req->oui))
228 				goto error1;
229 		}
230 		/* Make sure class supplied is consistent with QP type */
231 		if (qp_type == IB_QPT_SMI) {
232 			if ((mad_reg_req->mgmt_class !=
233 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
234 			    (mad_reg_req->mgmt_class !=
235 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
236 				goto error1;
237 		} else {
238 			if ((mad_reg_req->mgmt_class ==
239 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
240 			    (mad_reg_req->mgmt_class ==
241 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
242 				goto error1;
243 		}
244 	} else {
245 		/* No registration request supplied */
246 		if (!send_handler)
247 			goto error1;
248 	}
249 
250 	/* Validate device and port */
251 	port_priv = ib_get_mad_port(device, port_num);
252 	if (!port_priv) {
253 		ret = ERR_PTR(-ENODEV);
254 		goto error1;
255 	}
256 
257 	/* Allocate structures */
258 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
259 	if (!mad_agent_priv) {
260 		ret = ERR_PTR(-ENOMEM);
261 		goto error1;
262 	}
263 
264 	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
265 						 IB_ACCESS_LOCAL_WRITE);
266 	if (IS_ERR(mad_agent_priv->agent.mr)) {
267 		ret = ERR_PTR(-ENOMEM);
268 		goto error2;
269 	}
270 
271 	if (mad_reg_req) {
272 		reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
273 		if (!reg_req) {
274 			ret = ERR_PTR(-ENOMEM);
275 			goto error3;
276 		}
277 		/* Make a copy of the MAD registration request */
278 		memcpy(reg_req, mad_reg_req, sizeof *reg_req);
279 	}
280 
281 	/* Now, fill in the various structures */
282 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
283 	mad_agent_priv->reg_req = reg_req;
284 	mad_agent_priv->agent.rmpp_version = rmpp_version;
285 	mad_agent_priv->agent.device = device;
286 	mad_agent_priv->agent.recv_handler = recv_handler;
287 	mad_agent_priv->agent.send_handler = send_handler;
288 	mad_agent_priv->agent.context = context;
289 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
290 	mad_agent_priv->agent.port_num = port_num;
291 
292 	spin_lock_irqsave(&port_priv->reg_lock, flags);
293 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
294 
295 	/*
296 	 * Make sure MAD registration (if supplied)
297 	 * is non overlapping with any existing ones
298 	 */
299 	if (mad_reg_req) {
300 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
301 		if (!is_vendor_class(mgmt_class)) {
302 			class = port_priv->version[mad_reg_req->
303 						   mgmt_class_version].class;
304 			if (class) {
305 				method = class->method_table[mgmt_class];
306 				if (method) {
307 					if (method_in_use(&method,
308 							   mad_reg_req))
309 						goto error4;
310 				}
311 			}
312 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
313 						  mgmt_class);
314 		} else {
315 			/* "New" vendor class range */
316 			vendor = port_priv->version[mad_reg_req->
317 						    mgmt_class_version].vendor;
318 			if (vendor) {
319 				vclass = vendor_class_index(mgmt_class);
320 				vendor_class = vendor->vendor_class[vclass];
321 				if (vendor_class) {
322 					if (is_vendor_method_in_use(
323 							vendor_class,
324 							mad_reg_req))
325 						goto error4;
326 				}
327 			}
328 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
329 		}
330 		if (ret2) {
331 			ret = ERR_PTR(ret2);
332 			goto error4;
333 		}
334 	}
335 
336 	/* Add mad agent into port's agent list */
337 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
338 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
339 
340 	spin_lock_init(&mad_agent_priv->lock);
341 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
342 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
343 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
344 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
345 	INIT_WORK(&mad_agent_priv->timed_work, timeout_sends, mad_agent_priv);
346 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
347 	INIT_WORK(&mad_agent_priv->local_work, local_completions,
348 		   mad_agent_priv);
349 	atomic_set(&mad_agent_priv->refcount, 1);
350 	init_waitqueue_head(&mad_agent_priv->wait);
351 
352 	return &mad_agent_priv->agent;
353 
354 error4:
355 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
356 	kfree(reg_req);
357 error3:
358 	ib_dereg_mr(mad_agent_priv->agent.mr);
359 error2:
360 	kfree(mad_agent_priv);
361 error1:
362 	return ret;
363 }
364 EXPORT_SYMBOL(ib_register_mad_agent);
365 
366 static inline int is_snooping_sends(int mad_snoop_flags)
367 {
368 	return (mad_snoop_flags &
369 		(/*IB_MAD_SNOOP_POSTED_SENDS |
370 		 IB_MAD_SNOOP_RMPP_SENDS |*/
371 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
372 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
373 }
374 
375 static inline int is_snooping_recvs(int mad_snoop_flags)
376 {
377 	return (mad_snoop_flags &
378 		(IB_MAD_SNOOP_RECVS /*|
379 		 IB_MAD_SNOOP_RMPP_RECVS*/));
380 }
381 
382 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
383 				struct ib_mad_snoop_private *mad_snoop_priv)
384 {
385 	struct ib_mad_snoop_private **new_snoop_table;
386 	unsigned long flags;
387 	int i;
388 
389 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
390 	/* Check for empty slot in array. */
391 	for (i = 0; i < qp_info->snoop_table_size; i++)
392 		if (!qp_info->snoop_table[i])
393 			break;
394 
395 	if (i == qp_info->snoop_table_size) {
396 		/* Grow table. */
397 		new_snoop_table = kmalloc(sizeof mad_snoop_priv *
398 					  qp_info->snoop_table_size + 1,
399 					  GFP_ATOMIC);
400 		if (!new_snoop_table) {
401 			i = -ENOMEM;
402 			goto out;
403 		}
404 		if (qp_info->snoop_table) {
405 			memcpy(new_snoop_table, qp_info->snoop_table,
406 			       sizeof mad_snoop_priv *
407 			       qp_info->snoop_table_size);
408 			kfree(qp_info->snoop_table);
409 		}
410 		qp_info->snoop_table = new_snoop_table;
411 		qp_info->snoop_table_size++;
412 	}
413 	qp_info->snoop_table[i] = mad_snoop_priv;
414 	atomic_inc(&qp_info->snoop_count);
415 out:
416 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
417 	return i;
418 }
419 
420 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
421 					   u8 port_num,
422 					   enum ib_qp_type qp_type,
423 					   int mad_snoop_flags,
424 					   ib_mad_snoop_handler snoop_handler,
425 					   ib_mad_recv_handler recv_handler,
426 					   void *context)
427 {
428 	struct ib_mad_port_private *port_priv;
429 	struct ib_mad_agent *ret;
430 	struct ib_mad_snoop_private *mad_snoop_priv;
431 	int qpn;
432 
433 	/* Validate parameters */
434 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
435 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
436 		ret = ERR_PTR(-EINVAL);
437 		goto error1;
438 	}
439 	qpn = get_spl_qp_index(qp_type);
440 	if (qpn == -1) {
441 		ret = ERR_PTR(-EINVAL);
442 		goto error1;
443 	}
444 	port_priv = ib_get_mad_port(device, port_num);
445 	if (!port_priv) {
446 		ret = ERR_PTR(-ENODEV);
447 		goto error1;
448 	}
449 	/* Allocate structures */
450 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
451 	if (!mad_snoop_priv) {
452 		ret = ERR_PTR(-ENOMEM);
453 		goto error1;
454 	}
455 
456 	/* Now, fill in the various structures */
457 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
458 	mad_snoop_priv->agent.device = device;
459 	mad_snoop_priv->agent.recv_handler = recv_handler;
460 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
461 	mad_snoop_priv->agent.context = context;
462 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
463 	mad_snoop_priv->agent.port_num = port_num;
464 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
465 	init_waitqueue_head(&mad_snoop_priv->wait);
466 	mad_snoop_priv->snoop_index = register_snoop_agent(
467 						&port_priv->qp_info[qpn],
468 						mad_snoop_priv);
469 	if (mad_snoop_priv->snoop_index < 0) {
470 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
471 		goto error2;
472 	}
473 
474 	atomic_set(&mad_snoop_priv->refcount, 1);
475 	return &mad_snoop_priv->agent;
476 
477 error2:
478 	kfree(mad_snoop_priv);
479 error1:
480 	return ret;
481 }
482 EXPORT_SYMBOL(ib_register_mad_snoop);
483 
484 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
485 {
486 	struct ib_mad_port_private *port_priv;
487 	unsigned long flags;
488 
489 	/* Note that we could still be handling received MADs */
490 
491 	/*
492 	 * Canceling all sends results in dropping received response
493 	 * MADs, preventing us from queuing additional work
494 	 */
495 	cancel_mads(mad_agent_priv);
496 	port_priv = mad_agent_priv->qp_info->port_priv;
497 	cancel_delayed_work(&mad_agent_priv->timed_work);
498 
499 	spin_lock_irqsave(&port_priv->reg_lock, flags);
500 	remove_mad_reg_req(mad_agent_priv);
501 	list_del(&mad_agent_priv->agent_list);
502 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
503 
504 	flush_workqueue(port_priv->wq);
505 	ib_cancel_rmpp_recvs(mad_agent_priv);
506 
507 	atomic_dec(&mad_agent_priv->refcount);
508 	wait_event(mad_agent_priv->wait,
509 		   !atomic_read(&mad_agent_priv->refcount));
510 
511 	kfree(mad_agent_priv->reg_req);
512 	ib_dereg_mr(mad_agent_priv->agent.mr);
513 	kfree(mad_agent_priv);
514 }
515 
516 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
517 {
518 	struct ib_mad_qp_info *qp_info;
519 	unsigned long flags;
520 
521 	qp_info = mad_snoop_priv->qp_info;
522 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
523 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
524 	atomic_dec(&qp_info->snoop_count);
525 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
526 
527 	atomic_dec(&mad_snoop_priv->refcount);
528 	wait_event(mad_snoop_priv->wait,
529 		   !atomic_read(&mad_snoop_priv->refcount));
530 
531 	kfree(mad_snoop_priv);
532 }
533 
534 /*
535  * ib_unregister_mad_agent - Unregisters a client from using MAD services
536  */
537 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
538 {
539 	struct ib_mad_agent_private *mad_agent_priv;
540 	struct ib_mad_snoop_private *mad_snoop_priv;
541 
542 	/* If the TID is zero, the agent can only snoop. */
543 	if (mad_agent->hi_tid) {
544 		mad_agent_priv = container_of(mad_agent,
545 					      struct ib_mad_agent_private,
546 					      agent);
547 		unregister_mad_agent(mad_agent_priv);
548 	} else {
549 		mad_snoop_priv = container_of(mad_agent,
550 					      struct ib_mad_snoop_private,
551 					      agent);
552 		unregister_mad_snoop(mad_snoop_priv);
553 	}
554 	return 0;
555 }
556 EXPORT_SYMBOL(ib_unregister_mad_agent);
557 
558 static inline int response_mad(struct ib_mad *mad)
559 {
560 	/* Trap represses are responses although response bit is reset */
561 	return ((mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
562 		(mad->mad_hdr.method & IB_MGMT_METHOD_RESP));
563 }
564 
565 static void dequeue_mad(struct ib_mad_list_head *mad_list)
566 {
567 	struct ib_mad_queue *mad_queue;
568 	unsigned long flags;
569 
570 	BUG_ON(!mad_list->mad_queue);
571 	mad_queue = mad_list->mad_queue;
572 	spin_lock_irqsave(&mad_queue->lock, flags);
573 	list_del(&mad_list->list);
574 	mad_queue->count--;
575 	spin_unlock_irqrestore(&mad_queue->lock, flags);
576 }
577 
578 static void snoop_send(struct ib_mad_qp_info *qp_info,
579 		       struct ib_mad_send_buf *send_buf,
580 		       struct ib_mad_send_wc *mad_send_wc,
581 		       int mad_snoop_flags)
582 {
583 	struct ib_mad_snoop_private *mad_snoop_priv;
584 	unsigned long flags;
585 	int i;
586 
587 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
588 	for (i = 0; i < qp_info->snoop_table_size; i++) {
589 		mad_snoop_priv = qp_info->snoop_table[i];
590 		if (!mad_snoop_priv ||
591 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
592 			continue;
593 
594 		atomic_inc(&mad_snoop_priv->refcount);
595 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
596 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
597 						    send_buf, mad_send_wc);
598 		if (atomic_dec_and_test(&mad_snoop_priv->refcount))
599 			wake_up(&mad_snoop_priv->wait);
600 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
601 	}
602 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
603 }
604 
605 static void snoop_recv(struct ib_mad_qp_info *qp_info,
606 		       struct ib_mad_recv_wc *mad_recv_wc,
607 		       int mad_snoop_flags)
608 {
609 	struct ib_mad_snoop_private *mad_snoop_priv;
610 	unsigned long flags;
611 	int i;
612 
613 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
614 	for (i = 0; i < qp_info->snoop_table_size; i++) {
615 		mad_snoop_priv = qp_info->snoop_table[i];
616 		if (!mad_snoop_priv ||
617 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
618 			continue;
619 
620 		atomic_inc(&mad_snoop_priv->refcount);
621 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
622 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
623 						   mad_recv_wc);
624 		if (atomic_dec_and_test(&mad_snoop_priv->refcount))
625 			wake_up(&mad_snoop_priv->wait);
626 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
627 	}
628 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
629 }
630 
631 static void build_smp_wc(u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
632 			 struct ib_wc *wc)
633 {
634 	memset(wc, 0, sizeof *wc);
635 	wc->wr_id = wr_id;
636 	wc->status = IB_WC_SUCCESS;
637 	wc->opcode = IB_WC_RECV;
638 	wc->pkey_index = pkey_index;
639 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
640 	wc->src_qp = IB_QP0;
641 	wc->qp_num = IB_QP0;
642 	wc->slid = slid;
643 	wc->sl = 0;
644 	wc->dlid_path_bits = 0;
645 	wc->port_num = port_num;
646 }
647 
648 /*
649  * Return 0 if SMP is to be sent
650  * Return 1 if SMP was consumed locally (whether or not solicited)
651  * Return < 0 if error
652  */
653 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
654 				  struct ib_mad_send_wr_private *mad_send_wr)
655 {
656 	int ret;
657 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
658 	unsigned long flags;
659 	struct ib_mad_local_private *local;
660 	struct ib_mad_private *mad_priv;
661 	struct ib_mad_port_private *port_priv;
662 	struct ib_mad_agent_private *recv_mad_agent = NULL;
663 	struct ib_device *device = mad_agent_priv->agent.device;
664 	u8 port_num = mad_agent_priv->agent.port_num;
665 	struct ib_wc mad_wc;
666 	struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
667 
668 	if (!smi_handle_dr_smp_send(smp, device->node_type, port_num)) {
669 		ret = -EINVAL;
670 		printk(KERN_ERR PFX "Invalid directed route\n");
671 		goto out;
672 	}
673 	/* Check to post send on QP or process locally */
674 	ret = smi_check_local_dr_smp(smp, device, port_num);
675 	if (!ret || !device->process_mad)
676 		goto out;
677 
678 	local = kmalloc(sizeof *local, GFP_ATOMIC);
679 	if (!local) {
680 		ret = -ENOMEM;
681 		printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
682 		goto out;
683 	}
684 	local->mad_priv = NULL;
685 	local->recv_mad_agent = NULL;
686 	mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
687 	if (!mad_priv) {
688 		ret = -ENOMEM;
689 		printk(KERN_ERR PFX "No memory for local response MAD\n");
690 		kfree(local);
691 		goto out;
692 	}
693 
694 	build_smp_wc(send_wr->wr_id, be16_to_cpu(smp->dr_slid),
695 		     send_wr->wr.ud.pkey_index,
696 		     send_wr->wr.ud.port_num, &mad_wc);
697 
698 	/* No GRH for DR SMP */
699 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
700 				  (struct ib_mad *)smp,
701 				  (struct ib_mad *)&mad_priv->mad);
702 	switch (ret)
703 	{
704 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
705 		if (response_mad(&mad_priv->mad.mad) &&
706 		    mad_agent_priv->agent.recv_handler) {
707 			local->mad_priv = mad_priv;
708 			local->recv_mad_agent = mad_agent_priv;
709 			/*
710 			 * Reference MAD agent until receive
711 			 * side of local completion handled
712 			 */
713 			atomic_inc(&mad_agent_priv->refcount);
714 		} else
715 			kmem_cache_free(ib_mad_cache, mad_priv);
716 		break;
717 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
718 		kmem_cache_free(ib_mad_cache, mad_priv);
719 		break;
720 	case IB_MAD_RESULT_SUCCESS:
721 		/* Treat like an incoming receive MAD */
722 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
723 					    mad_agent_priv->agent.port_num);
724 		if (port_priv) {
725 			mad_priv->mad.mad.mad_hdr.tid =
726 				((struct ib_mad *)smp)->mad_hdr.tid;
727 			recv_mad_agent = find_mad_agent(port_priv,
728 						        &mad_priv->mad.mad);
729 		}
730 		if (!port_priv || !recv_mad_agent) {
731 			kmem_cache_free(ib_mad_cache, mad_priv);
732 			kfree(local);
733 			ret = 0;
734 			goto out;
735 		}
736 		local->mad_priv = mad_priv;
737 		local->recv_mad_agent = recv_mad_agent;
738 		break;
739 	default:
740 		kmem_cache_free(ib_mad_cache, mad_priv);
741 		kfree(local);
742 		ret = -EINVAL;
743 		goto out;
744 	}
745 
746 	local->mad_send_wr = mad_send_wr;
747 	/* Reference MAD agent until send side of local completion handled */
748 	atomic_inc(&mad_agent_priv->refcount);
749 	/* Queue local completion to local list */
750 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
751 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
752 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
753 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
754 		   &mad_agent_priv->local_work);
755 	ret = 1;
756 out:
757 	return ret;
758 }
759 
760 static int get_buf_length(int hdr_len, int data_len)
761 {
762 	int seg_size, pad;
763 
764 	seg_size = sizeof(struct ib_mad) - hdr_len;
765 	if (data_len && seg_size) {
766 		pad = seg_size - data_len % seg_size;
767 		if (pad == seg_size)
768 			pad = 0;
769 	} else
770 		pad = seg_size;
771 	return hdr_len + data_len + pad;
772 }
773 
774 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
775 					    u32 remote_qpn, u16 pkey_index,
776 					    int rmpp_active,
777 					    int hdr_len, int data_len,
778 					    gfp_t gfp_mask)
779 {
780 	struct ib_mad_agent_private *mad_agent_priv;
781 	struct ib_mad_send_wr_private *mad_send_wr;
782 	int buf_size;
783 	void *buf;
784 
785 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
786 				      agent);
787 	buf_size = get_buf_length(hdr_len, data_len);
788 
789 	if ((!mad_agent->rmpp_version &&
790 	     (rmpp_active || buf_size > sizeof(struct ib_mad))) ||
791 	    (!rmpp_active && buf_size > sizeof(struct ib_mad)))
792 		return ERR_PTR(-EINVAL);
793 
794 	buf = kzalloc(sizeof *mad_send_wr + buf_size, gfp_mask);
795 	if (!buf)
796 		return ERR_PTR(-ENOMEM);
797 
798 	mad_send_wr = buf + buf_size;
799 	mad_send_wr->send_buf.mad = buf;
800 
801 	mad_send_wr->mad_agent_priv = mad_agent_priv;
802 	mad_send_wr->sg_list[0].length = buf_size;
803 	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
804 
805 	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
806 	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
807 	mad_send_wr->send_wr.num_sge = 1;
808 	mad_send_wr->send_wr.opcode = IB_WR_SEND;
809 	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
810 	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
811 	mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
812 	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
813 
814 	if (rmpp_active) {
815 		struct ib_rmpp_mad *rmpp_mad = mad_send_wr->send_buf.mad;
816 		rmpp_mad->rmpp_hdr.paylen_newwin = cpu_to_be32(hdr_len -
817 						   IB_MGMT_RMPP_HDR + data_len);
818 		rmpp_mad->rmpp_hdr.rmpp_version = mad_agent->rmpp_version;
819 		rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
820 		ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr,
821 				  IB_MGMT_RMPP_FLAG_ACTIVE);
822 	}
823 
824 	mad_send_wr->send_buf.mad_agent = mad_agent;
825 	atomic_inc(&mad_agent_priv->refcount);
826 	return &mad_send_wr->send_buf;
827 }
828 EXPORT_SYMBOL(ib_create_send_mad);
829 
830 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
831 {
832 	struct ib_mad_agent_private *mad_agent_priv;
833 
834 	mad_agent_priv = container_of(send_buf->mad_agent,
835 				      struct ib_mad_agent_private, agent);
836 	kfree(send_buf->mad);
837 
838 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
839 		wake_up(&mad_agent_priv->wait);
840 }
841 EXPORT_SYMBOL(ib_free_send_mad);
842 
843 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
844 {
845 	struct ib_mad_qp_info *qp_info;
846 	struct list_head *list;
847 	struct ib_send_wr *bad_send_wr;
848 	struct ib_mad_agent *mad_agent;
849 	struct ib_sge *sge;
850 	unsigned long flags;
851 	int ret;
852 
853 	/* Set WR ID to find mad_send_wr upon completion */
854 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
855 	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
856 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
857 
858 	mad_agent = mad_send_wr->send_buf.mad_agent;
859 	sge = mad_send_wr->sg_list;
860 	sge->addr = dma_map_single(mad_agent->device->dma_device,
861 				   mad_send_wr->send_buf.mad, sge->length,
862 				   DMA_TO_DEVICE);
863 	pci_unmap_addr_set(mad_send_wr, mapping, sge->addr);
864 
865 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
866 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
867 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
868 				   &bad_send_wr);
869 		list = &qp_info->send_queue.list;
870 	} else {
871 		ret = 0;
872 		list = &qp_info->overflow_list;
873 	}
874 
875 	if (!ret) {
876 		qp_info->send_queue.count++;
877 		list_add_tail(&mad_send_wr->mad_list.list, list);
878 	}
879 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
880 	if (ret)
881 		dma_unmap_single(mad_agent->device->dma_device,
882 				 pci_unmap_addr(mad_send_wr, mapping),
883 				 sge->length, DMA_TO_DEVICE);
884 
885 	return ret;
886 }
887 
888 /*
889  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
890  *  with the registered client
891  */
892 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
893 		     struct ib_mad_send_buf **bad_send_buf)
894 {
895 	struct ib_mad_agent_private *mad_agent_priv;
896 	struct ib_mad_send_buf *next_send_buf;
897 	struct ib_mad_send_wr_private *mad_send_wr;
898 	unsigned long flags;
899 	int ret = -EINVAL;
900 
901 	/* Walk list of send WRs and post each on send list */
902 	for (; send_buf; send_buf = next_send_buf) {
903 
904 		mad_send_wr = container_of(send_buf,
905 					   struct ib_mad_send_wr_private,
906 					   send_buf);
907 		mad_agent_priv = mad_send_wr->mad_agent_priv;
908 
909 		if (!send_buf->mad_agent->send_handler ||
910 		    (send_buf->timeout_ms &&
911 		     !send_buf->mad_agent->recv_handler)) {
912 			ret = -EINVAL;
913 			goto error;
914 		}
915 
916 		/*
917 		 * Save pointer to next work request to post in case the
918 		 * current one completes, and the user modifies the work
919 		 * request associated with the completion
920 		 */
921 		next_send_buf = send_buf->next;
922 		mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
923 
924 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
925 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
926 			ret = handle_outgoing_dr_smp(mad_agent_priv,
927 						     mad_send_wr);
928 			if (ret < 0)		/* error */
929 				goto error;
930 			else if (ret == 1)	/* locally consumed */
931 				continue;
932 		}
933 
934 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
935 		/* Timeout will be updated after send completes */
936 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
937 		mad_send_wr->retries = send_buf->retries;
938 		/* Reference for work request to QP + response */
939 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
940 		mad_send_wr->status = IB_WC_SUCCESS;
941 
942 		/* Reference MAD agent until send completes */
943 		atomic_inc(&mad_agent_priv->refcount);
944 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
945 		list_add_tail(&mad_send_wr->agent_list,
946 			      &mad_agent_priv->send_list);
947 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
948 
949 		if (mad_agent_priv->agent.rmpp_version) {
950 			ret = ib_send_rmpp_mad(mad_send_wr);
951 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
952 				ret = ib_send_mad(mad_send_wr);
953 		} else
954 			ret = ib_send_mad(mad_send_wr);
955 		if (ret < 0) {
956 			/* Fail send request */
957 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
958 			list_del(&mad_send_wr->agent_list);
959 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
960 			atomic_dec(&mad_agent_priv->refcount);
961 			goto error;
962 		}
963 	}
964 	return 0;
965 error:
966 	if (bad_send_buf)
967 		*bad_send_buf = send_buf;
968 	return ret;
969 }
970 EXPORT_SYMBOL(ib_post_send_mad);
971 
972 /*
973  * ib_free_recv_mad - Returns data buffers used to receive
974  *  a MAD to the access layer
975  */
976 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
977 {
978 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
979 	struct ib_mad_private_header *mad_priv_hdr;
980 	struct ib_mad_private *priv;
981 	struct list_head free_list;
982 
983 	INIT_LIST_HEAD(&free_list);
984 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
985 
986 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
987 					&free_list, list) {
988 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
989 					   recv_buf);
990 		mad_priv_hdr = container_of(mad_recv_wc,
991 					    struct ib_mad_private_header,
992 					    recv_wc);
993 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
994 				    header);
995 		kmem_cache_free(ib_mad_cache, priv);
996 	}
997 }
998 EXPORT_SYMBOL(ib_free_recv_mad);
999 
1000 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1001 					u8 rmpp_version,
1002 					ib_mad_send_handler send_handler,
1003 					ib_mad_recv_handler recv_handler,
1004 					void *context)
1005 {
1006 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1007 }
1008 EXPORT_SYMBOL(ib_redirect_mad_qp);
1009 
1010 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1011 		      struct ib_wc *wc)
1012 {
1013 	printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1014 	return 0;
1015 }
1016 EXPORT_SYMBOL(ib_process_mad_wc);
1017 
1018 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1019 			 struct ib_mad_reg_req *mad_reg_req)
1020 {
1021 	int i;
1022 
1023 	for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1024 	     i < IB_MGMT_MAX_METHODS;
1025 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1026 			       1+i)) {
1027 		if ((*method)->agent[i]) {
1028 			printk(KERN_ERR PFX "Method %d already in use\n", i);
1029 			return -EINVAL;
1030 		}
1031 	}
1032 	return 0;
1033 }
1034 
1035 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1036 {
1037 	/* Allocate management method table */
1038 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1039 	if (!*method) {
1040 		printk(KERN_ERR PFX "No memory for "
1041 		       "ib_mad_mgmt_method_table\n");
1042 		return -ENOMEM;
1043 	}
1044 
1045 	return 0;
1046 }
1047 
1048 /*
1049  * Check to see if there are any methods still in use
1050  */
1051 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1052 {
1053 	int i;
1054 
1055 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1056 		if (method->agent[i])
1057 			return 1;
1058 	return 0;
1059 }
1060 
1061 /*
1062  * Check to see if there are any method tables for this class still in use
1063  */
1064 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1065 {
1066 	int i;
1067 
1068 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1069 		if (class->method_table[i])
1070 			return 1;
1071 	return 0;
1072 }
1073 
1074 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1075 {
1076 	int i;
1077 
1078 	for (i = 0; i < MAX_MGMT_OUI; i++)
1079 		if (vendor_class->method_table[i])
1080 			return 1;
1081 	return 0;
1082 }
1083 
1084 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1085 			   char *oui)
1086 {
1087 	int i;
1088 
1089 	for (i = 0; i < MAX_MGMT_OUI; i++)
1090                 /* Is there matching OUI for this vendor class ? */
1091                 if (!memcmp(vendor_class->oui[i], oui, 3))
1092 			return i;
1093 
1094 	return -1;
1095 }
1096 
1097 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1098 {
1099 	int i;
1100 
1101 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1102 		if (vendor->vendor_class[i])
1103 			return 1;
1104 
1105 	return 0;
1106 }
1107 
1108 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1109 				     struct ib_mad_agent_private *agent)
1110 {
1111 	int i;
1112 
1113 	/* Remove any methods for this mad agent */
1114 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1115 		if (method->agent[i] == agent) {
1116 			method->agent[i] = NULL;
1117 		}
1118 	}
1119 }
1120 
1121 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1122 			      struct ib_mad_agent_private *agent_priv,
1123 			      u8 mgmt_class)
1124 {
1125 	struct ib_mad_port_private *port_priv;
1126 	struct ib_mad_mgmt_class_table **class;
1127 	struct ib_mad_mgmt_method_table **method;
1128 	int i, ret;
1129 
1130 	port_priv = agent_priv->qp_info->port_priv;
1131 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1132 	if (!*class) {
1133 		/* Allocate management class table for "new" class version */
1134 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1135 		if (!*class) {
1136 			printk(KERN_ERR PFX "No memory for "
1137 			       "ib_mad_mgmt_class_table\n");
1138 			ret = -ENOMEM;
1139 			goto error1;
1140 		}
1141 
1142 		/* Allocate method table for this management class */
1143 		method = &(*class)->method_table[mgmt_class];
1144 		if ((ret = allocate_method_table(method)))
1145 			goto error2;
1146 	} else {
1147 		method = &(*class)->method_table[mgmt_class];
1148 		if (!*method) {
1149 			/* Allocate method table for this management class */
1150 			if ((ret = allocate_method_table(method)))
1151 				goto error1;
1152 		}
1153 	}
1154 
1155 	/* Now, make sure methods are not already in use */
1156 	if (method_in_use(method, mad_reg_req))
1157 		goto error3;
1158 
1159 	/* Finally, add in methods being registered */
1160 	for (i = find_first_bit(mad_reg_req->method_mask,
1161 				IB_MGMT_MAX_METHODS);
1162 	     i < IB_MGMT_MAX_METHODS;
1163 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1164 			       1+i)) {
1165 		(*method)->agent[i] = agent_priv;
1166 	}
1167 	return 0;
1168 
1169 error3:
1170 	/* Remove any methods for this mad agent */
1171 	remove_methods_mad_agent(*method, agent_priv);
1172 	/* Now, check to see if there are any methods in use */
1173 	if (!check_method_table(*method)) {
1174 		/* If not, release management method table */
1175 		kfree(*method);
1176 		*method = NULL;
1177 	}
1178 	ret = -EINVAL;
1179 	goto error1;
1180 error2:
1181 	kfree(*class);
1182 	*class = NULL;
1183 error1:
1184 	return ret;
1185 }
1186 
1187 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1188 			   struct ib_mad_agent_private *agent_priv)
1189 {
1190 	struct ib_mad_port_private *port_priv;
1191 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1192 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1193 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1194 	struct ib_mad_mgmt_method_table **method;
1195 	int i, ret = -ENOMEM;
1196 	u8 vclass;
1197 
1198 	/* "New" vendor (with OUI) class */
1199 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1200 	port_priv = agent_priv->qp_info->port_priv;
1201 	vendor_table = &port_priv->version[
1202 				mad_reg_req->mgmt_class_version].vendor;
1203 	if (!*vendor_table) {
1204 		/* Allocate mgmt vendor class table for "new" class version */
1205 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1206 		if (!vendor) {
1207 			printk(KERN_ERR PFX "No memory for "
1208 			       "ib_mad_mgmt_vendor_class_table\n");
1209 			goto error1;
1210 		}
1211 
1212 		*vendor_table = vendor;
1213 	}
1214 	if (!(*vendor_table)->vendor_class[vclass]) {
1215 		/* Allocate table for this management vendor class */
1216 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1217 		if (!vendor_class) {
1218 			printk(KERN_ERR PFX "No memory for "
1219 			       "ib_mad_mgmt_vendor_class\n");
1220 			goto error2;
1221 		}
1222 
1223 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1224 	}
1225 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1226 		/* Is there matching OUI for this vendor class ? */
1227 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1228 			    mad_reg_req->oui, 3)) {
1229 			method = &(*vendor_table)->vendor_class[
1230 						vclass]->method_table[i];
1231 			BUG_ON(!*method);
1232 			goto check_in_use;
1233 		}
1234 	}
1235 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1236 		/* OUI slot available ? */
1237 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1238 				vclass]->oui[i])) {
1239 			method = &(*vendor_table)->vendor_class[
1240 				vclass]->method_table[i];
1241 			BUG_ON(*method);
1242 			/* Allocate method table for this OUI */
1243 			if ((ret = allocate_method_table(method)))
1244 				goto error3;
1245 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1246 			       mad_reg_req->oui, 3);
1247 			goto check_in_use;
1248 		}
1249 	}
1250 	printk(KERN_ERR PFX "All OUI slots in use\n");
1251 	goto error3;
1252 
1253 check_in_use:
1254 	/* Now, make sure methods are not already in use */
1255 	if (method_in_use(method, mad_reg_req))
1256 		goto error4;
1257 
1258 	/* Finally, add in methods being registered */
1259 	for (i = find_first_bit(mad_reg_req->method_mask,
1260 				IB_MGMT_MAX_METHODS);
1261 	     i < IB_MGMT_MAX_METHODS;
1262 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1263 			       1+i)) {
1264 		(*method)->agent[i] = agent_priv;
1265 	}
1266 	return 0;
1267 
1268 error4:
1269 	/* Remove any methods for this mad agent */
1270 	remove_methods_mad_agent(*method, agent_priv);
1271 	/* Now, check to see if there are any methods in use */
1272 	if (!check_method_table(*method)) {
1273 		/* If not, release management method table */
1274 		kfree(*method);
1275 		*method = NULL;
1276 	}
1277 	ret = -EINVAL;
1278 error3:
1279 	if (vendor_class) {
1280 		(*vendor_table)->vendor_class[vclass] = NULL;
1281 		kfree(vendor_class);
1282 	}
1283 error2:
1284 	if (vendor) {
1285 		*vendor_table = NULL;
1286 		kfree(vendor);
1287 	}
1288 error1:
1289 	return ret;
1290 }
1291 
1292 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1293 {
1294 	struct ib_mad_port_private *port_priv;
1295 	struct ib_mad_mgmt_class_table *class;
1296 	struct ib_mad_mgmt_method_table *method;
1297 	struct ib_mad_mgmt_vendor_class_table *vendor;
1298 	struct ib_mad_mgmt_vendor_class *vendor_class;
1299 	int index;
1300 	u8 mgmt_class;
1301 
1302 	/*
1303 	 * Was MAD registration request supplied
1304 	 * with original registration ?
1305 	 */
1306 	if (!agent_priv->reg_req) {
1307 		goto out;
1308 	}
1309 
1310 	port_priv = agent_priv->qp_info->port_priv;
1311 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1312 	class = port_priv->version[
1313 			agent_priv->reg_req->mgmt_class_version].class;
1314 	if (!class)
1315 		goto vendor_check;
1316 
1317 	method = class->method_table[mgmt_class];
1318 	if (method) {
1319 		/* Remove any methods for this mad agent */
1320 		remove_methods_mad_agent(method, agent_priv);
1321 		/* Now, check to see if there are any methods still in use */
1322 		if (!check_method_table(method)) {
1323 			/* If not, release management method table */
1324 			 kfree(method);
1325 			 class->method_table[mgmt_class] = NULL;
1326 			 /* Any management classes left ? */
1327 			if (!check_class_table(class)) {
1328 				/* If not, release management class table */
1329 				kfree(class);
1330 				port_priv->version[
1331 					agent_priv->reg_req->
1332 					mgmt_class_version].class = NULL;
1333 			}
1334 		}
1335 	}
1336 
1337 vendor_check:
1338 	if (!is_vendor_class(mgmt_class))
1339 		goto out;
1340 
1341 	/* normalize mgmt_class to vendor range 2 */
1342 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1343 	vendor = port_priv->version[
1344 			agent_priv->reg_req->mgmt_class_version].vendor;
1345 
1346 	if (!vendor)
1347 		goto out;
1348 
1349 	vendor_class = vendor->vendor_class[mgmt_class];
1350 	if (vendor_class) {
1351 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1352 		if (index < 0)
1353 			goto out;
1354 		method = vendor_class->method_table[index];
1355 		if (method) {
1356 			/* Remove any methods for this mad agent */
1357 			remove_methods_mad_agent(method, agent_priv);
1358 			/*
1359 			 * Now, check to see if there are
1360 			 * any methods still in use
1361 			 */
1362 			if (!check_method_table(method)) {
1363 				/* If not, release management method table */
1364 				kfree(method);
1365 				vendor_class->method_table[index] = NULL;
1366 				memset(vendor_class->oui[index], 0, 3);
1367 				/* Any OUIs left ? */
1368 				if (!check_vendor_class(vendor_class)) {
1369 					/* If not, release vendor class table */
1370 					kfree(vendor_class);
1371 					vendor->vendor_class[mgmt_class] = NULL;
1372 					/* Any other vendor classes left ? */
1373 					if (!check_vendor_table(vendor)) {
1374 						kfree(vendor);
1375 						port_priv->version[
1376 							agent_priv->reg_req->
1377 							mgmt_class_version].
1378 							vendor = NULL;
1379 					}
1380 				}
1381 			}
1382 		}
1383 	}
1384 
1385 out:
1386 	return;
1387 }
1388 
1389 static struct ib_mad_agent_private *
1390 find_mad_agent(struct ib_mad_port_private *port_priv,
1391 	       struct ib_mad *mad)
1392 {
1393 	struct ib_mad_agent_private *mad_agent = NULL;
1394 	unsigned long flags;
1395 
1396 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1397 	if (response_mad(mad)) {
1398 		u32 hi_tid;
1399 		struct ib_mad_agent_private *entry;
1400 
1401 		/*
1402 		 * Routing is based on high 32 bits of transaction ID
1403 		 * of MAD.
1404 		 */
1405 		hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1406 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1407 			if (entry->agent.hi_tid == hi_tid) {
1408 				mad_agent = entry;
1409 				break;
1410 			}
1411 		}
1412 	} else {
1413 		struct ib_mad_mgmt_class_table *class;
1414 		struct ib_mad_mgmt_method_table *method;
1415 		struct ib_mad_mgmt_vendor_class_table *vendor;
1416 		struct ib_mad_mgmt_vendor_class *vendor_class;
1417 		struct ib_vendor_mad *vendor_mad;
1418 		int index;
1419 
1420 		/*
1421 		 * Routing is based on version, class, and method
1422 		 * For "newer" vendor MADs, also based on OUI
1423 		 */
1424 		if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1425 			goto out;
1426 		if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1427 			class = port_priv->version[
1428 					mad->mad_hdr.class_version].class;
1429 			if (!class)
1430 				goto out;
1431 			method = class->method_table[convert_mgmt_class(
1432 							mad->mad_hdr.mgmt_class)];
1433 			if (method)
1434 				mad_agent = method->agent[mad->mad_hdr.method &
1435 							  ~IB_MGMT_METHOD_RESP];
1436 		} else {
1437 			vendor = port_priv->version[
1438 					mad->mad_hdr.class_version].vendor;
1439 			if (!vendor)
1440 				goto out;
1441 			vendor_class = vendor->vendor_class[vendor_class_index(
1442 						mad->mad_hdr.mgmt_class)];
1443 			if (!vendor_class)
1444 				goto out;
1445 			/* Find matching OUI */
1446 			vendor_mad = (struct ib_vendor_mad *)mad;
1447 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1448 			if (index == -1)
1449 				goto out;
1450 			method = vendor_class->method_table[index];
1451 			if (method) {
1452 				mad_agent = method->agent[mad->mad_hdr.method &
1453 							  ~IB_MGMT_METHOD_RESP];
1454 			}
1455 		}
1456 	}
1457 
1458 	if (mad_agent) {
1459 		if (mad_agent->agent.recv_handler)
1460 			atomic_inc(&mad_agent->refcount);
1461 		else {
1462 			printk(KERN_NOTICE PFX "No receive handler for client "
1463 			       "%p on port %d\n",
1464 			       &mad_agent->agent, port_priv->port_num);
1465 			mad_agent = NULL;
1466 		}
1467 	}
1468 out:
1469 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1470 
1471 	return mad_agent;
1472 }
1473 
1474 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1475 {
1476 	int valid = 0;
1477 
1478 	/* Make sure MAD base version is understood */
1479 	if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1480 		printk(KERN_ERR PFX "MAD received with unsupported base "
1481 		       "version %d\n", mad->mad_hdr.base_version);
1482 		goto out;
1483 	}
1484 
1485 	/* Filter SMI packets sent to other than QP0 */
1486 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1487 	    (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1488 		if (qp_num == 0)
1489 			valid = 1;
1490 	} else {
1491 		/* Filter GSI packets sent to QP0 */
1492 		if (qp_num != 0)
1493 			valid = 1;
1494 	}
1495 
1496 out:
1497 	return valid;
1498 }
1499 
1500 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1501 		       struct ib_mad_hdr *mad_hdr)
1502 {
1503 	struct ib_rmpp_mad *rmpp_mad;
1504 
1505 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1506 	return !mad_agent_priv->agent.rmpp_version ||
1507 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1508 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1509 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1510 }
1511 
1512 struct ib_mad_send_wr_private*
1513 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv, __be64 tid)
1514 {
1515 	struct ib_mad_send_wr_private *mad_send_wr;
1516 
1517 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
1518 			    agent_list) {
1519 		if (mad_send_wr->tid == tid)
1520 			return mad_send_wr;
1521 	}
1522 
1523 	/*
1524 	 * It's possible to receive the response before we've
1525 	 * been notified that the send has completed
1526 	 */
1527 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
1528 			    agent_list) {
1529 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
1530 		    mad_send_wr->tid == tid && mad_send_wr->timeout) {
1531 			/* Verify request has not been canceled */
1532 			return (mad_send_wr->status == IB_WC_SUCCESS) ?
1533 				mad_send_wr : NULL;
1534 		}
1535 	}
1536 	return NULL;
1537 }
1538 
1539 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1540 {
1541 	mad_send_wr->timeout = 0;
1542 	if (mad_send_wr->refcount == 1) {
1543 		list_del(&mad_send_wr->agent_list);
1544 		list_add_tail(&mad_send_wr->agent_list,
1545 			      &mad_send_wr->mad_agent_priv->done_list);
1546 	}
1547 }
1548 
1549 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1550 				 struct ib_mad_recv_wc *mad_recv_wc)
1551 {
1552 	struct ib_mad_send_wr_private *mad_send_wr;
1553 	struct ib_mad_send_wc mad_send_wc;
1554 	unsigned long flags;
1555 	__be64 tid;
1556 
1557 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1558 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1559 	if (mad_agent_priv->agent.rmpp_version) {
1560 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1561 						      mad_recv_wc);
1562 		if (!mad_recv_wc) {
1563 			if (atomic_dec_and_test(&mad_agent_priv->refcount))
1564 				wake_up(&mad_agent_priv->wait);
1565 			return;
1566 		}
1567 	}
1568 
1569 	/* Complete corresponding request */
1570 	if (response_mad(mad_recv_wc->recv_buf.mad)) {
1571 		tid = mad_recv_wc->recv_buf.mad->mad_hdr.tid;
1572 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1573 		mad_send_wr = ib_find_send_mad(mad_agent_priv, tid);
1574 		if (!mad_send_wr) {
1575 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1576 			ib_free_recv_mad(mad_recv_wc);
1577 			if (atomic_dec_and_test(&mad_agent_priv->refcount))
1578 				wake_up(&mad_agent_priv->wait);
1579 			return;
1580 		}
1581 		ib_mark_mad_done(mad_send_wr);
1582 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1583 
1584 		/* Defined behavior is to complete response before request */
1585 		mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1586 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1587 						   mad_recv_wc);
1588 		atomic_dec(&mad_agent_priv->refcount);
1589 
1590 		mad_send_wc.status = IB_WC_SUCCESS;
1591 		mad_send_wc.vendor_err = 0;
1592 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
1593 		ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1594 	} else {
1595 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1596 						   mad_recv_wc);
1597 		if (atomic_dec_and_test(&mad_agent_priv->refcount))
1598 			wake_up(&mad_agent_priv->wait);
1599 	}
1600 }
1601 
1602 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1603 				     struct ib_wc *wc)
1604 {
1605 	struct ib_mad_qp_info *qp_info;
1606 	struct ib_mad_private_header *mad_priv_hdr;
1607 	struct ib_mad_private *recv, *response;
1608 	struct ib_mad_list_head *mad_list;
1609 	struct ib_mad_agent_private *mad_agent;
1610 
1611 	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1612 	if (!response)
1613 		printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1614 		       "for response buffer\n");
1615 
1616 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1617 	qp_info = mad_list->mad_queue->qp_info;
1618 	dequeue_mad(mad_list);
1619 
1620 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1621 				    mad_list);
1622 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1623 	dma_unmap_single(port_priv->device->dma_device,
1624 			 pci_unmap_addr(&recv->header, mapping),
1625 			 sizeof(struct ib_mad_private) -
1626 			 sizeof(struct ib_mad_private_header),
1627 			 DMA_FROM_DEVICE);
1628 
1629 	/* Setup MAD receive work completion from "normal" work completion */
1630 	recv->header.wc = *wc;
1631 	recv->header.recv_wc.wc = &recv->header.wc;
1632 	recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1633 	recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1634 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
1635 
1636 	if (atomic_read(&qp_info->snoop_count))
1637 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1638 
1639 	/* Validate MAD */
1640 	if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1641 		goto out;
1642 
1643 	if (recv->mad.mad.mad_hdr.mgmt_class ==
1644 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1645 		if (!smi_handle_dr_smp_recv(&recv->mad.smp,
1646 					    port_priv->device->node_type,
1647 					    port_priv->port_num,
1648 					    port_priv->device->phys_port_cnt))
1649 			goto out;
1650 		if (!smi_check_forward_dr_smp(&recv->mad.smp))
1651 			goto local;
1652 		if (!smi_handle_dr_smp_send(&recv->mad.smp,
1653 					    port_priv->device->node_type,
1654 					    port_priv->port_num))
1655 			goto out;
1656 		if (!smi_check_local_dr_smp(&recv->mad.smp,
1657 					    port_priv->device,
1658 					    port_priv->port_num))
1659 			goto out;
1660 	}
1661 
1662 local:
1663 	/* Give driver "right of first refusal" on incoming MAD */
1664 	if (port_priv->device->process_mad) {
1665 		int ret;
1666 
1667 		if (!response) {
1668 			printk(KERN_ERR PFX "No memory for response MAD\n");
1669 			/*
1670 			 * Is it better to assume that
1671 			 * it wouldn't be processed ?
1672 			 */
1673 			goto out;
1674 		}
1675 
1676 		ret = port_priv->device->process_mad(port_priv->device, 0,
1677 						     port_priv->port_num,
1678 						     wc, &recv->grh,
1679 						     &recv->mad.mad,
1680 						     &response->mad.mad);
1681 		if (ret & IB_MAD_RESULT_SUCCESS) {
1682 			if (ret & IB_MAD_RESULT_CONSUMED)
1683 				goto out;
1684 			if (ret & IB_MAD_RESULT_REPLY) {
1685 				agent_send_response(&response->mad.mad,
1686 						    &recv->grh, wc,
1687 						    port_priv->device,
1688 						    port_priv->port_num,
1689 						    qp_info->qp->qp_num);
1690 				goto out;
1691 			}
1692 		}
1693 	}
1694 
1695 	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1696 	if (mad_agent) {
1697 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1698 		/*
1699 		 * recv is freed up in error cases in ib_mad_complete_recv
1700 		 * or via recv_handler in ib_mad_complete_recv()
1701 		 */
1702 		recv = NULL;
1703 	}
1704 
1705 out:
1706 	/* Post another receive request for this QP */
1707 	if (response) {
1708 		ib_mad_post_receive_mads(qp_info, response);
1709 		if (recv)
1710 			kmem_cache_free(ib_mad_cache, recv);
1711 	} else
1712 		ib_mad_post_receive_mads(qp_info, recv);
1713 }
1714 
1715 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1716 {
1717 	struct ib_mad_send_wr_private *mad_send_wr;
1718 	unsigned long delay;
1719 
1720 	if (list_empty(&mad_agent_priv->wait_list)) {
1721 		cancel_delayed_work(&mad_agent_priv->timed_work);
1722 	} else {
1723 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1724 					 struct ib_mad_send_wr_private,
1725 					 agent_list);
1726 
1727 		if (time_after(mad_agent_priv->timeout,
1728 			       mad_send_wr->timeout)) {
1729 			mad_agent_priv->timeout = mad_send_wr->timeout;
1730 			cancel_delayed_work(&mad_agent_priv->timed_work);
1731 			delay = mad_send_wr->timeout - jiffies;
1732 			if ((long)delay <= 0)
1733 				delay = 1;
1734 			queue_delayed_work(mad_agent_priv->qp_info->
1735 					   port_priv->wq,
1736 					   &mad_agent_priv->timed_work, delay);
1737 		}
1738 	}
1739 }
1740 
1741 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1742 {
1743 	struct ib_mad_agent_private *mad_agent_priv;
1744 	struct ib_mad_send_wr_private *temp_mad_send_wr;
1745 	struct list_head *list_item;
1746 	unsigned long delay;
1747 
1748 	mad_agent_priv = mad_send_wr->mad_agent_priv;
1749 	list_del(&mad_send_wr->agent_list);
1750 
1751 	delay = mad_send_wr->timeout;
1752 	mad_send_wr->timeout += jiffies;
1753 
1754 	if (delay) {
1755 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1756 			temp_mad_send_wr = list_entry(list_item,
1757 						struct ib_mad_send_wr_private,
1758 						agent_list);
1759 			if (time_after(mad_send_wr->timeout,
1760 				       temp_mad_send_wr->timeout))
1761 				break;
1762 		}
1763 	}
1764 	else
1765 		list_item = &mad_agent_priv->wait_list;
1766 	list_add(&mad_send_wr->agent_list, list_item);
1767 
1768 	/* Reschedule a work item if we have a shorter timeout */
1769 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
1770 		cancel_delayed_work(&mad_agent_priv->timed_work);
1771 		queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
1772 				   &mad_agent_priv->timed_work, delay);
1773 	}
1774 }
1775 
1776 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
1777 			  int timeout_ms)
1778 {
1779 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
1780 	wait_for_response(mad_send_wr);
1781 }
1782 
1783 /*
1784  * Process a send work completion
1785  */
1786 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
1787 			     struct ib_mad_send_wc *mad_send_wc)
1788 {
1789 	struct ib_mad_agent_private	*mad_agent_priv;
1790 	unsigned long			flags;
1791 	int				ret;
1792 
1793 	mad_agent_priv = mad_send_wr->mad_agent_priv;
1794 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
1795 	if (mad_agent_priv->agent.rmpp_version) {
1796 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
1797 		if (ret == IB_RMPP_RESULT_CONSUMED)
1798 			goto done;
1799 	} else
1800 		ret = IB_RMPP_RESULT_UNHANDLED;
1801 
1802 	if (mad_send_wc->status != IB_WC_SUCCESS &&
1803 	    mad_send_wr->status == IB_WC_SUCCESS) {
1804 		mad_send_wr->status = mad_send_wc->status;
1805 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
1806 	}
1807 
1808 	if (--mad_send_wr->refcount > 0) {
1809 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
1810 		    mad_send_wr->status == IB_WC_SUCCESS) {
1811 			wait_for_response(mad_send_wr);
1812 		}
1813 		goto done;
1814 	}
1815 
1816 	/* Remove send from MAD agent and notify client of completion */
1817 	list_del(&mad_send_wr->agent_list);
1818 	adjust_timeout(mad_agent_priv);
1819 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1820 
1821 	if (mad_send_wr->status != IB_WC_SUCCESS )
1822 		mad_send_wc->status = mad_send_wr->status;
1823 	if (ret == IB_RMPP_RESULT_INTERNAL)
1824 		ib_rmpp_send_handler(mad_send_wc);
1825 	else
1826 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
1827 						   mad_send_wc);
1828 
1829 	/* Release reference on agent taken when sending */
1830 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
1831 		wake_up(&mad_agent_priv->wait);
1832 	return;
1833 done:
1834 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1835 }
1836 
1837 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
1838 				     struct ib_wc *wc)
1839 {
1840 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
1841 	struct ib_mad_list_head		*mad_list;
1842 	struct ib_mad_qp_info		*qp_info;
1843 	struct ib_mad_queue		*send_queue;
1844 	struct ib_send_wr		*bad_send_wr;
1845 	struct ib_mad_send_wc		mad_send_wc;
1846 	unsigned long flags;
1847 	int ret;
1848 
1849 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1850 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1851 				   mad_list);
1852 	send_queue = mad_list->mad_queue;
1853 	qp_info = send_queue->qp_info;
1854 
1855 retry:
1856 	dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
1857 			 pci_unmap_addr(mad_send_wr, mapping),
1858 			 mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
1859 	queued_send_wr = NULL;
1860 	spin_lock_irqsave(&send_queue->lock, flags);
1861 	list_del(&mad_list->list);
1862 
1863 	/* Move queued send to the send queue */
1864 	if (send_queue->count-- > send_queue->max_active) {
1865 		mad_list = container_of(qp_info->overflow_list.next,
1866 					struct ib_mad_list_head, list);
1867 		queued_send_wr = container_of(mad_list,
1868 					struct ib_mad_send_wr_private,
1869 					mad_list);
1870 		list_del(&mad_list->list);
1871 		list_add_tail(&mad_list->list, &send_queue->list);
1872 	}
1873 	spin_unlock_irqrestore(&send_queue->lock, flags);
1874 
1875 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
1876 	mad_send_wc.status = wc->status;
1877 	mad_send_wc.vendor_err = wc->vendor_err;
1878 	if (atomic_read(&qp_info->snoop_count))
1879 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
1880 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
1881 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1882 
1883 	if (queued_send_wr) {
1884 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
1885 				   &bad_send_wr);
1886 		if (ret) {
1887 			printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
1888 			mad_send_wr = queued_send_wr;
1889 			wc->status = IB_WC_LOC_QP_OP_ERR;
1890 			goto retry;
1891 		}
1892 	}
1893 }
1894 
1895 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
1896 {
1897 	struct ib_mad_send_wr_private *mad_send_wr;
1898 	struct ib_mad_list_head *mad_list;
1899 	unsigned long flags;
1900 
1901 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1902 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
1903 		mad_send_wr = container_of(mad_list,
1904 					   struct ib_mad_send_wr_private,
1905 					   mad_list);
1906 		mad_send_wr->retry = 1;
1907 	}
1908 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1909 }
1910 
1911 static void mad_error_handler(struct ib_mad_port_private *port_priv,
1912 			      struct ib_wc *wc)
1913 {
1914 	struct ib_mad_list_head *mad_list;
1915 	struct ib_mad_qp_info *qp_info;
1916 	struct ib_mad_send_wr_private *mad_send_wr;
1917 	int ret;
1918 
1919 	/* Determine if failure was a send or receive */
1920 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1921 	qp_info = mad_list->mad_queue->qp_info;
1922 	if (mad_list->mad_queue == &qp_info->recv_queue)
1923 		/*
1924 		 * Receive errors indicate that the QP has entered the error
1925 		 * state - error handling/shutdown code will cleanup
1926 		 */
1927 		return;
1928 
1929 	/*
1930 	 * Send errors will transition the QP to SQE - move
1931 	 * QP to RTS and repost flushed work requests
1932 	 */
1933 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1934 				   mad_list);
1935 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
1936 		if (mad_send_wr->retry) {
1937 			/* Repost send */
1938 			struct ib_send_wr *bad_send_wr;
1939 
1940 			mad_send_wr->retry = 0;
1941 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
1942 					&bad_send_wr);
1943 			if (ret)
1944 				ib_mad_send_done_handler(port_priv, wc);
1945 		} else
1946 			ib_mad_send_done_handler(port_priv, wc);
1947 	} else {
1948 		struct ib_qp_attr *attr;
1949 
1950 		/* Transition QP to RTS and fail offending send */
1951 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
1952 		if (attr) {
1953 			attr->qp_state = IB_QPS_RTS;
1954 			attr->cur_qp_state = IB_QPS_SQE;
1955 			ret = ib_modify_qp(qp_info->qp, attr,
1956 					   IB_QP_STATE | IB_QP_CUR_STATE);
1957 			kfree(attr);
1958 			if (ret)
1959 				printk(KERN_ERR PFX "mad_error_handler - "
1960 				       "ib_modify_qp to RTS : %d\n", ret);
1961 			else
1962 				mark_sends_for_retry(qp_info);
1963 		}
1964 		ib_mad_send_done_handler(port_priv, wc);
1965 	}
1966 }
1967 
1968 /*
1969  * IB MAD completion callback
1970  */
1971 static void ib_mad_completion_handler(void *data)
1972 {
1973 	struct ib_mad_port_private *port_priv;
1974 	struct ib_wc wc;
1975 
1976 	port_priv = (struct ib_mad_port_private *)data;
1977 	ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
1978 
1979 	while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
1980 		if (wc.status == IB_WC_SUCCESS) {
1981 			switch (wc.opcode) {
1982 			case IB_WC_SEND:
1983 				ib_mad_send_done_handler(port_priv, &wc);
1984 				break;
1985 			case IB_WC_RECV:
1986 				ib_mad_recv_done_handler(port_priv, &wc);
1987 				break;
1988 			default:
1989 				BUG_ON(1);
1990 				break;
1991 			}
1992 		} else
1993 			mad_error_handler(port_priv, &wc);
1994 	}
1995 }
1996 
1997 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
1998 {
1999 	unsigned long flags;
2000 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2001 	struct ib_mad_send_wc mad_send_wc;
2002 	struct list_head cancel_list;
2003 
2004 	INIT_LIST_HEAD(&cancel_list);
2005 
2006 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2007 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2008 				 &mad_agent_priv->send_list, agent_list) {
2009 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2010  			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2011 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2012 		}
2013 	}
2014 
2015 	/* Empty wait list to prevent receives from finding a request */
2016 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2017 	/* Empty local completion list as well */
2018 	list_splice_init(&mad_agent_priv->local_list, &cancel_list);
2019 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2020 
2021 	/* Report all cancelled requests */
2022 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2023 	mad_send_wc.vendor_err = 0;
2024 
2025 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2026 				 &cancel_list, agent_list) {
2027 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2028 		list_del(&mad_send_wr->agent_list);
2029 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2030 						   &mad_send_wc);
2031 		atomic_dec(&mad_agent_priv->refcount);
2032 	}
2033 }
2034 
2035 static struct ib_mad_send_wr_private*
2036 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2037 	     struct ib_mad_send_buf *send_buf)
2038 {
2039 	struct ib_mad_send_wr_private *mad_send_wr;
2040 
2041 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2042 			    agent_list) {
2043 		if (&mad_send_wr->send_buf == send_buf)
2044 			return mad_send_wr;
2045 	}
2046 
2047 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2048 			    agent_list) {
2049 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2050 		    &mad_send_wr->send_buf == send_buf)
2051 			return mad_send_wr;
2052 	}
2053 	return NULL;
2054 }
2055 
2056 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2057 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2058 {
2059 	struct ib_mad_agent_private *mad_agent_priv;
2060 	struct ib_mad_send_wr_private *mad_send_wr;
2061 	unsigned long flags;
2062 	int active;
2063 
2064 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2065 				      agent);
2066 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2067 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2068 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2069 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2070 		return -EINVAL;
2071 	}
2072 
2073 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2074 	if (!timeout_ms) {
2075 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2076 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2077 	}
2078 
2079 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2080 	if (active)
2081 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2082 	else
2083 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2084 
2085 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2086 	return 0;
2087 }
2088 EXPORT_SYMBOL(ib_modify_mad);
2089 
2090 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2091 		   struct ib_mad_send_buf *send_buf)
2092 {
2093 	ib_modify_mad(mad_agent, send_buf, 0);
2094 }
2095 EXPORT_SYMBOL(ib_cancel_mad);
2096 
2097 static void local_completions(void *data)
2098 {
2099 	struct ib_mad_agent_private *mad_agent_priv;
2100 	struct ib_mad_local_private *local;
2101 	struct ib_mad_agent_private *recv_mad_agent;
2102 	unsigned long flags;
2103 	int recv = 0;
2104 	struct ib_wc wc;
2105 	struct ib_mad_send_wc mad_send_wc;
2106 
2107 	mad_agent_priv = (struct ib_mad_agent_private *)data;
2108 
2109 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2110 	while (!list_empty(&mad_agent_priv->local_list)) {
2111 		local = list_entry(mad_agent_priv->local_list.next,
2112 				   struct ib_mad_local_private,
2113 				   completion_list);
2114 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2115 		if (local->mad_priv) {
2116 			recv_mad_agent = local->recv_mad_agent;
2117 			if (!recv_mad_agent) {
2118 				printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2119 				goto local_send_completion;
2120 			}
2121 
2122 			recv = 1;
2123 			/*
2124 			 * Defined behavior is to complete response
2125 			 * before request
2126 			 */
2127 			build_smp_wc((unsigned long) local->mad_send_wr,
2128 				     be16_to_cpu(IB_LID_PERMISSIVE),
2129 				     0, recv_mad_agent->agent.port_num, &wc);
2130 
2131 			local->mad_priv->header.recv_wc.wc = &wc;
2132 			local->mad_priv->header.recv_wc.mad_len =
2133 						sizeof(struct ib_mad);
2134 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2135 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2136 				 &local->mad_priv->header.recv_wc.rmpp_list);
2137 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2138 			local->mad_priv->header.recv_wc.recv_buf.mad =
2139 						&local->mad_priv->mad.mad;
2140 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2141 				snoop_recv(recv_mad_agent->qp_info,
2142 					  &local->mad_priv->header.recv_wc,
2143 					   IB_MAD_SNOOP_RECVS);
2144 			recv_mad_agent->agent.recv_handler(
2145 						&recv_mad_agent->agent,
2146 						&local->mad_priv->header.recv_wc);
2147 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2148 			atomic_dec(&recv_mad_agent->refcount);
2149 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2150 		}
2151 
2152 local_send_completion:
2153 		/* Complete send */
2154 		mad_send_wc.status = IB_WC_SUCCESS;
2155 		mad_send_wc.vendor_err = 0;
2156 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2157 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2158 			snoop_send(mad_agent_priv->qp_info,
2159 				   &local->mad_send_wr->send_buf,
2160 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2161 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2162 						   &mad_send_wc);
2163 
2164 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2165 		list_del(&local->completion_list);
2166 		atomic_dec(&mad_agent_priv->refcount);
2167 		if (!recv)
2168 			kmem_cache_free(ib_mad_cache, local->mad_priv);
2169 		kfree(local);
2170 	}
2171 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2172 }
2173 
2174 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2175 {
2176 	int ret;
2177 
2178 	if (!mad_send_wr->retries--)
2179 		return -ETIMEDOUT;
2180 
2181 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2182 
2183 	if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2184 		ret = ib_retry_rmpp(mad_send_wr);
2185 		switch (ret) {
2186 		case IB_RMPP_RESULT_UNHANDLED:
2187 			ret = ib_send_mad(mad_send_wr);
2188 			break;
2189 		case IB_RMPP_RESULT_CONSUMED:
2190 			ret = 0;
2191 			break;
2192 		default:
2193 			ret = -ECOMM;
2194 			break;
2195 		}
2196 	} else
2197 		ret = ib_send_mad(mad_send_wr);
2198 
2199 	if (!ret) {
2200 		mad_send_wr->refcount++;
2201 		list_add_tail(&mad_send_wr->agent_list,
2202 			      &mad_send_wr->mad_agent_priv->send_list);
2203 	}
2204 	return ret;
2205 }
2206 
2207 static void timeout_sends(void *data)
2208 {
2209 	struct ib_mad_agent_private *mad_agent_priv;
2210 	struct ib_mad_send_wr_private *mad_send_wr;
2211 	struct ib_mad_send_wc mad_send_wc;
2212 	unsigned long flags, delay;
2213 
2214 	mad_agent_priv = (struct ib_mad_agent_private *)data;
2215 	mad_send_wc.vendor_err = 0;
2216 
2217 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2218 	while (!list_empty(&mad_agent_priv->wait_list)) {
2219 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2220 					 struct ib_mad_send_wr_private,
2221 					 agent_list);
2222 
2223 		if (time_after(mad_send_wr->timeout, jiffies)) {
2224 			delay = mad_send_wr->timeout - jiffies;
2225 			if ((long)delay <= 0)
2226 				delay = 1;
2227 			queue_delayed_work(mad_agent_priv->qp_info->
2228 					   port_priv->wq,
2229 					   &mad_agent_priv->timed_work, delay);
2230 			break;
2231 		}
2232 
2233 		list_del(&mad_send_wr->agent_list);
2234 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2235 		    !retry_send(mad_send_wr))
2236 			continue;
2237 
2238 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2239 
2240 		if (mad_send_wr->status == IB_WC_SUCCESS)
2241 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2242 		else
2243 			mad_send_wc.status = mad_send_wr->status;
2244 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2245 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2246 						   &mad_send_wc);
2247 
2248 		atomic_dec(&mad_agent_priv->refcount);
2249 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2250 	}
2251 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2252 }
2253 
2254 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2255 {
2256 	struct ib_mad_port_private *port_priv = cq->cq_context;
2257 
2258 	queue_work(port_priv->wq, &port_priv->work);
2259 }
2260 
2261 /*
2262  * Allocate receive MADs and post receive WRs for them
2263  */
2264 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2265 				    struct ib_mad_private *mad)
2266 {
2267 	unsigned long flags;
2268 	int post, ret;
2269 	struct ib_mad_private *mad_priv;
2270 	struct ib_sge sg_list;
2271 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2272 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2273 
2274 	/* Initialize common scatter list fields */
2275 	sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2276 	sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2277 
2278 	/* Initialize common receive WR fields */
2279 	recv_wr.next = NULL;
2280 	recv_wr.sg_list = &sg_list;
2281 	recv_wr.num_sge = 1;
2282 
2283 	do {
2284 		/* Allocate and map receive buffer */
2285 		if (mad) {
2286 			mad_priv = mad;
2287 			mad = NULL;
2288 		} else {
2289 			mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2290 			if (!mad_priv) {
2291 				printk(KERN_ERR PFX "No memory for receive buffer\n");
2292 				ret = -ENOMEM;
2293 				break;
2294 			}
2295 		}
2296 		sg_list.addr = dma_map_single(qp_info->port_priv->
2297 						device->dma_device,
2298 					&mad_priv->grh,
2299 					sizeof *mad_priv -
2300 						sizeof mad_priv->header,
2301 					DMA_FROM_DEVICE);
2302 		pci_unmap_addr_set(&mad_priv->header, mapping, sg_list.addr);
2303 		recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2304 		mad_priv->header.mad_list.mad_queue = recv_queue;
2305 
2306 		/* Post receive WR */
2307 		spin_lock_irqsave(&recv_queue->lock, flags);
2308 		post = (++recv_queue->count < recv_queue->max_active);
2309 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2310 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2311 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2312 		if (ret) {
2313 			spin_lock_irqsave(&recv_queue->lock, flags);
2314 			list_del(&mad_priv->header.mad_list.list);
2315 			recv_queue->count--;
2316 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2317 			dma_unmap_single(qp_info->port_priv->device->dma_device,
2318 					 pci_unmap_addr(&mad_priv->header,
2319 							mapping),
2320 					 sizeof *mad_priv -
2321 					   sizeof mad_priv->header,
2322 					 DMA_FROM_DEVICE);
2323 			kmem_cache_free(ib_mad_cache, mad_priv);
2324 			printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2325 			break;
2326 		}
2327 	} while (post);
2328 
2329 	return ret;
2330 }
2331 
2332 /*
2333  * Return all the posted receive MADs
2334  */
2335 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2336 {
2337 	struct ib_mad_private_header *mad_priv_hdr;
2338 	struct ib_mad_private *recv;
2339 	struct ib_mad_list_head *mad_list;
2340 
2341 	while (!list_empty(&qp_info->recv_queue.list)) {
2342 
2343 		mad_list = list_entry(qp_info->recv_queue.list.next,
2344 				      struct ib_mad_list_head, list);
2345 		mad_priv_hdr = container_of(mad_list,
2346 					    struct ib_mad_private_header,
2347 					    mad_list);
2348 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2349 				    header);
2350 
2351 		/* Remove from posted receive MAD list */
2352 		list_del(&mad_list->list);
2353 
2354 		dma_unmap_single(qp_info->port_priv->device->dma_device,
2355 				 pci_unmap_addr(&recv->header, mapping),
2356 				 sizeof(struct ib_mad_private) -
2357 				 sizeof(struct ib_mad_private_header),
2358 				 DMA_FROM_DEVICE);
2359 		kmem_cache_free(ib_mad_cache, recv);
2360 	}
2361 
2362 	qp_info->recv_queue.count = 0;
2363 }
2364 
2365 /*
2366  * Start the port
2367  */
2368 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2369 {
2370 	int ret, i;
2371 	struct ib_qp_attr *attr;
2372 	struct ib_qp *qp;
2373 
2374 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2375  	if (!attr) {
2376 		printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2377 		return -ENOMEM;
2378 	}
2379 
2380 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2381 		qp = port_priv->qp_info[i].qp;
2382 		/*
2383 		 * PKey index for QP1 is irrelevant but
2384 		 * one is needed for the Reset to Init transition
2385 		 */
2386 		attr->qp_state = IB_QPS_INIT;
2387 		attr->pkey_index = 0;
2388 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2389 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2390 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2391 		if (ret) {
2392 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2393 			       "INIT: %d\n", i, ret);
2394 			goto out;
2395 		}
2396 
2397 		attr->qp_state = IB_QPS_RTR;
2398 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2399 		if (ret) {
2400 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2401 			       "RTR: %d\n", i, ret);
2402 			goto out;
2403 		}
2404 
2405 		attr->qp_state = IB_QPS_RTS;
2406 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2407 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2408 		if (ret) {
2409 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2410 			       "RTS: %d\n", i, ret);
2411 			goto out;
2412 		}
2413 	}
2414 
2415 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2416 	if (ret) {
2417 		printk(KERN_ERR PFX "Failed to request completion "
2418 		       "notification: %d\n", ret);
2419 		goto out;
2420 	}
2421 
2422 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2423 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2424 		if (ret) {
2425 			printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2426 			goto out;
2427 		}
2428 	}
2429 out:
2430 	kfree(attr);
2431 	return ret;
2432 }
2433 
2434 static void qp_event_handler(struct ib_event *event, void *qp_context)
2435 {
2436 	struct ib_mad_qp_info	*qp_info = qp_context;
2437 
2438 	/* It's worse than that! He's dead, Jim! */
2439 	printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2440 		event->event, qp_info->qp->qp_num);
2441 }
2442 
2443 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2444 			   struct ib_mad_queue *mad_queue)
2445 {
2446 	mad_queue->qp_info = qp_info;
2447 	mad_queue->count = 0;
2448 	spin_lock_init(&mad_queue->lock);
2449 	INIT_LIST_HEAD(&mad_queue->list);
2450 }
2451 
2452 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2453 			struct ib_mad_qp_info *qp_info)
2454 {
2455 	qp_info->port_priv = port_priv;
2456 	init_mad_queue(qp_info, &qp_info->send_queue);
2457 	init_mad_queue(qp_info, &qp_info->recv_queue);
2458 	INIT_LIST_HEAD(&qp_info->overflow_list);
2459 	spin_lock_init(&qp_info->snoop_lock);
2460 	qp_info->snoop_table = NULL;
2461 	qp_info->snoop_table_size = 0;
2462 	atomic_set(&qp_info->snoop_count, 0);
2463 }
2464 
2465 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2466 			 enum ib_qp_type qp_type)
2467 {
2468 	struct ib_qp_init_attr	qp_init_attr;
2469 	int ret;
2470 
2471 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2472 	qp_init_attr.send_cq = qp_info->port_priv->cq;
2473 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2474 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2475 	qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2476 	qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2477 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2478 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2479 	qp_init_attr.qp_type = qp_type;
2480 	qp_init_attr.port_num = qp_info->port_priv->port_num;
2481 	qp_init_attr.qp_context = qp_info;
2482 	qp_init_attr.event_handler = qp_event_handler;
2483 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2484 	if (IS_ERR(qp_info->qp)) {
2485 		printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2486 		       get_spl_qp_index(qp_type));
2487 		ret = PTR_ERR(qp_info->qp);
2488 		goto error;
2489 	}
2490 	/* Use minimum queue sizes unless the CQ is resized */
2491 	qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2492 	qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2493 	return 0;
2494 
2495 error:
2496 	return ret;
2497 }
2498 
2499 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2500 {
2501 	ib_destroy_qp(qp_info->qp);
2502 	kfree(qp_info->snoop_table);
2503 }
2504 
2505 /*
2506  * Open the port
2507  * Create the QP, PD, MR, and CQ if needed
2508  */
2509 static int ib_mad_port_open(struct ib_device *device,
2510 			    int port_num)
2511 {
2512 	int ret, cq_size;
2513 	struct ib_mad_port_private *port_priv;
2514 	unsigned long flags;
2515 	char name[sizeof "ib_mad123"];
2516 
2517 	/* Create new device info */
2518 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2519 	if (!port_priv) {
2520 		printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2521 		return -ENOMEM;
2522 	}
2523 
2524 	port_priv->device = device;
2525 	port_priv->port_num = port_num;
2526 	spin_lock_init(&port_priv->reg_lock);
2527 	INIT_LIST_HEAD(&port_priv->agent_list);
2528 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2529 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2530 
2531 	cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2532 	port_priv->cq = ib_create_cq(port_priv->device,
2533 				     ib_mad_thread_completion_handler,
2534 				     NULL, port_priv, cq_size);
2535 	if (IS_ERR(port_priv->cq)) {
2536 		printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2537 		ret = PTR_ERR(port_priv->cq);
2538 		goto error3;
2539 	}
2540 
2541 	port_priv->pd = ib_alloc_pd(device);
2542 	if (IS_ERR(port_priv->pd)) {
2543 		printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2544 		ret = PTR_ERR(port_priv->pd);
2545 		goto error4;
2546 	}
2547 
2548 	port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2549 	if (IS_ERR(port_priv->mr)) {
2550 		printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2551 		ret = PTR_ERR(port_priv->mr);
2552 		goto error5;
2553 	}
2554 
2555 	ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2556 	if (ret)
2557 		goto error6;
2558 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2559 	if (ret)
2560 		goto error7;
2561 
2562 	snprintf(name, sizeof name, "ib_mad%d", port_num);
2563 	port_priv->wq = create_singlethread_workqueue(name);
2564 	if (!port_priv->wq) {
2565 		ret = -ENOMEM;
2566 		goto error8;
2567 	}
2568 	INIT_WORK(&port_priv->work, ib_mad_completion_handler, port_priv);
2569 
2570 	ret = ib_mad_port_start(port_priv);
2571 	if (ret) {
2572 		printk(KERN_ERR PFX "Couldn't start port\n");
2573 		goto error9;
2574 	}
2575 
2576 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2577 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2578 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2579 	return 0;
2580 
2581 error9:
2582 	destroy_workqueue(port_priv->wq);
2583 error8:
2584 	destroy_mad_qp(&port_priv->qp_info[1]);
2585 error7:
2586 	destroy_mad_qp(&port_priv->qp_info[0]);
2587 error6:
2588 	ib_dereg_mr(port_priv->mr);
2589 error5:
2590 	ib_dealloc_pd(port_priv->pd);
2591 error4:
2592 	ib_destroy_cq(port_priv->cq);
2593 	cleanup_recv_queue(&port_priv->qp_info[1]);
2594 	cleanup_recv_queue(&port_priv->qp_info[0]);
2595 error3:
2596 	kfree(port_priv);
2597 
2598 	return ret;
2599 }
2600 
2601 /*
2602  * Close the port
2603  * If there are no classes using the port, free the port
2604  * resources (CQ, MR, PD, QP) and remove the port's info structure
2605  */
2606 static int ib_mad_port_close(struct ib_device *device, int port_num)
2607 {
2608 	struct ib_mad_port_private *port_priv;
2609 	unsigned long flags;
2610 
2611 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2612 	port_priv = __ib_get_mad_port(device, port_num);
2613 	if (port_priv == NULL) {
2614 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2615 		printk(KERN_ERR PFX "Port %d not found\n", port_num);
2616 		return -ENODEV;
2617 	}
2618 	list_del(&port_priv->port_list);
2619 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2620 
2621 	/* Stop processing completions. */
2622 	flush_workqueue(port_priv->wq);
2623 	destroy_workqueue(port_priv->wq);
2624 	destroy_mad_qp(&port_priv->qp_info[1]);
2625 	destroy_mad_qp(&port_priv->qp_info[0]);
2626 	ib_dereg_mr(port_priv->mr);
2627 	ib_dealloc_pd(port_priv->pd);
2628 	ib_destroy_cq(port_priv->cq);
2629 	cleanup_recv_queue(&port_priv->qp_info[1]);
2630 	cleanup_recv_queue(&port_priv->qp_info[0]);
2631 	/* XXX: Handle deallocation of MAD registration tables */
2632 
2633 	kfree(port_priv);
2634 
2635 	return 0;
2636 }
2637 
2638 static void ib_mad_init_device(struct ib_device *device)
2639 {
2640 	int start, end, i;
2641 
2642 	if (device->node_type == IB_NODE_SWITCH) {
2643 		start = 0;
2644 		end   = 0;
2645 	} else {
2646 		start = 1;
2647 		end   = device->phys_port_cnt;
2648 	}
2649 
2650 	for (i = start; i <= end; i++) {
2651 		if (ib_mad_port_open(device, i)) {
2652 			printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2653 			       device->name, i);
2654 			goto error;
2655 		}
2656 		if (ib_agent_port_open(device, i)) {
2657 			printk(KERN_ERR PFX "Couldn't open %s port %d "
2658 			       "for agents\n",
2659 			       device->name, i);
2660 			goto error_agent;
2661 		}
2662 	}
2663 	return;
2664 
2665 error_agent:
2666 	if (ib_mad_port_close(device, i))
2667 		printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2668 		       device->name, i);
2669 
2670 error:
2671 	i--;
2672 
2673 	while (i >= start) {
2674 		if (ib_agent_port_close(device, i))
2675 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2676 			       "for agents\n",
2677 			       device->name, i);
2678 		if (ib_mad_port_close(device, i))
2679 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2680 			       device->name, i);
2681 		i--;
2682 	}
2683 }
2684 
2685 static void ib_mad_remove_device(struct ib_device *device)
2686 {
2687 	int i, num_ports, cur_port;
2688 
2689 	if (device->node_type == IB_NODE_SWITCH) {
2690 		num_ports = 1;
2691 		cur_port = 0;
2692 	} else {
2693 		num_ports = device->phys_port_cnt;
2694 		cur_port = 1;
2695 	}
2696 	for (i = 0; i < num_ports; i++, cur_port++) {
2697 		if (ib_agent_port_close(device, cur_port))
2698 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2699 			       "for agents\n",
2700 			       device->name, cur_port);
2701 		if (ib_mad_port_close(device, cur_port))
2702 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2703 			       device->name, cur_port);
2704 	}
2705 }
2706 
2707 static struct ib_client mad_client = {
2708 	.name   = "mad",
2709 	.add = ib_mad_init_device,
2710 	.remove = ib_mad_remove_device
2711 };
2712 
2713 static int __init ib_mad_init_module(void)
2714 {
2715 	int ret;
2716 
2717 	spin_lock_init(&ib_mad_port_list_lock);
2718 
2719 	ib_mad_cache = kmem_cache_create("ib_mad",
2720 					 sizeof(struct ib_mad_private),
2721 					 0,
2722 					 SLAB_HWCACHE_ALIGN,
2723 					 NULL,
2724 					 NULL);
2725 	if (!ib_mad_cache) {
2726 		printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2727 		ret = -ENOMEM;
2728 		goto error1;
2729 	}
2730 
2731 	INIT_LIST_HEAD(&ib_mad_port_list);
2732 
2733 	if (ib_register_client(&mad_client)) {
2734 		printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2735 		ret = -EINVAL;
2736 		goto error2;
2737 	}
2738 
2739 	return 0;
2740 
2741 error2:
2742 	kmem_cache_destroy(ib_mad_cache);
2743 error1:
2744 	return ret;
2745 }
2746 
2747 static void __exit ib_mad_cleanup_module(void)
2748 {
2749 	ib_unregister_client(&mad_client);
2750 
2751 	if (kmem_cache_destroy(ib_mad_cache)) {
2752 		printk(KERN_DEBUG PFX "Failed to destroy ib_mad cache\n");
2753 	}
2754 }
2755 
2756 module_init(ib_mad_init_module);
2757 module_exit(ib_mad_cleanup_module);
2758 
2759