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 * Copyright (c) 2009 HNR Consulting. All rights reserved.
6 *
7 * This software is available to you under a choice of one of two
8 * licenses.  You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
12 *
13 *     Redistribution and use in source and binary forms, with or
14 *     without modification, are permitted provided that the following
15 *     conditions are met:
16 *
17 *      - Redistributions of source code must retain the above
18 *        copyright notice, this list of conditions and the following
19 *        disclaimer.
20 *
21 *      - Redistributions in binary form must reproduce the above
22 *        copyright notice, this list of conditions and the following
23 *        disclaimer in the documentation and/or other materials
24 *        provided with the distribution.
25 *
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 * SOFTWARE.
34 *
35 */
36#include <linux/dma-mapping.h>
37#include <rdma/ib_cache.h>
38
39#include "mad_priv.h"
40#include "mad_rmpp.h"
41#include "smi.h"
42#include "agent.h"
43
44MODULE_LICENSE("Dual BSD/GPL");
45MODULE_DESCRIPTION("kernel IB MAD API");
46MODULE_AUTHOR("Hal Rosenstock");
47MODULE_AUTHOR("Sean Hefty");
48
49int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
50int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
51
52module_param_named(send_queue_size, mad_sendq_size, int, 0444);
53MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
54module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
55MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
56
57static struct kmem_cache *ib_mad_cache;
58
59static struct list_head ib_mad_port_list;
60static u32 ib_mad_client_id = 0;
61
62/* Port list lock */
63static spinlock_t ib_mad_port_list_lock;
64
65
66/* Forward declarations */
67static int method_in_use(struct ib_mad_mgmt_method_table **method,
68			 struct ib_mad_reg_req *mad_reg_req);
69static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
70static struct ib_mad_agent_private *find_mad_agent(
71					struct ib_mad_port_private *port_priv,
72					struct ib_mad *mad);
73static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
74				    struct ib_mad_private *mad);
75static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
76static void timeout_sends(struct work_struct *work);
77static void local_completions(struct work_struct *work);
78static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
79			      struct ib_mad_agent_private *agent_priv,
80			      u8 mgmt_class);
81static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
82			   struct ib_mad_agent_private *agent_priv);
83
84/*
85 * Returns a ib_mad_port_private structure or NULL for a device/port
86 * Assumes ib_mad_port_list_lock is being held
87 */
88static inline struct ib_mad_port_private *
89__ib_get_mad_port(struct ib_device *device, int port_num)
90{
91	struct ib_mad_port_private *entry;
92
93	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
94		if (entry->device == device && entry->port_num == port_num)
95			return entry;
96	}
97	return NULL;
98}
99
100/*
101 * Wrapper function to return a ib_mad_port_private structure or NULL
102 * for a device/port
103 */
104static inline struct ib_mad_port_private *
105ib_get_mad_port(struct ib_device *device, int port_num)
106{
107	struct ib_mad_port_private *entry;
108	unsigned long flags;
109
110	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
111	entry = __ib_get_mad_port(device, port_num);
112	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
113
114	return entry;
115}
116
117static inline u8 convert_mgmt_class(u8 mgmt_class)
118{
119	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
120	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
121		0 : mgmt_class;
122}
123
124static int get_spl_qp_index(enum ib_qp_type qp_type)
125{
126	switch (qp_type)
127	{
128	case IB_QPT_SMI:
129		return 0;
130	case IB_QPT_GSI:
131		return 1;
132	default:
133		return -1;
134	}
135}
136
137static int vendor_class_index(u8 mgmt_class)
138{
139	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
140}
141
142static int is_vendor_class(u8 mgmt_class)
143{
144	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
145	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
146		return 0;
147	return 1;
148}
149
150static int is_vendor_oui(char *oui)
151{
152	if (oui[0] || oui[1] || oui[2])
153		return 1;
154	return 0;
155}
156
157static int is_vendor_method_in_use(
158		struct ib_mad_mgmt_vendor_class *vendor_class,
159		struct ib_mad_reg_req *mad_reg_req)
160{
161	struct ib_mad_mgmt_method_table *method;
162	int i;
163
164	for (i = 0; i < MAX_MGMT_OUI; i++) {
165		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
166			method = vendor_class->method_table[i];
167			if (method) {
168				if (method_in_use(&method, mad_reg_req))
169					return 1;
170				else
171					break;
172			}
173		}
174	}
175	return 0;
176}
177
178int ib_response_mad(struct ib_mad *mad)
179{
180	return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
181		(mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
182		((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
183		 (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
184}
185EXPORT_SYMBOL(ib_response_mad);
186
187static void timeout_callback(unsigned long data)
188{
189	struct ib_mad_agent_private *mad_agent_priv =
190		(struct ib_mad_agent_private *) data;
191
192	queue_work(mad_agent_priv->qp_info->port_priv->wq,
193		   &mad_agent_priv->timeout_work);
194}
195
196/*
197 * ib_register_mad_agent - Register to send/receive MADs
198 */
199struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
200					   u8 port_num,
201					   enum ib_qp_type qp_type,
202					   struct ib_mad_reg_req *mad_reg_req,
203					   u8 rmpp_version,
204					   ib_mad_send_handler send_handler,
205					   ib_mad_recv_handler recv_handler,
206					   void *context)
207{
208	struct ib_mad_port_private *port_priv;
209	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
210	struct ib_mad_agent_private *mad_agent_priv;
211	struct ib_mad_reg_req *reg_req = NULL;
212	struct ib_mad_mgmt_class_table *class;
213	struct ib_mad_mgmt_vendor_class_table *vendor;
214	struct ib_mad_mgmt_vendor_class *vendor_class;
215	struct ib_mad_mgmt_method_table *method;
216	int ret2, qpn;
217	unsigned long flags;
218	u8 mgmt_class, vclass;
219
220	/* Validate parameters */
221	qpn = get_spl_qp_index(qp_type);
222	if (qpn == -1)
223		goto error1;
224
225	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
226		goto error1;
227
228	/* Validate MAD registration request if supplied */
229	if (mad_reg_req) {
230		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
231			goto error1;
232		if (!recv_handler)
233			goto error1;
234		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
235			/*
236			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
237			 * one in this range currently allowed
238			 */
239			if (mad_reg_req->mgmt_class !=
240			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
241				goto error1;
242		} else if (mad_reg_req->mgmt_class == 0) {
243			/*
244			 * Class 0 is reserved in IBA and is used for
245			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
246			 */
247			goto error1;
248		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
249			/*
250			 * If class is in "new" vendor range,
251			 * ensure supplied OUI is not zero
252			 */
253			if (!is_vendor_oui(mad_reg_req->oui))
254				goto error1;
255		}
256		/* Make sure class supplied is consistent with RMPP */
257		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
258			if (rmpp_version)
259				goto error1;
260		}
261		/* Make sure class supplied is consistent with QP type */
262		if (qp_type == IB_QPT_SMI) {
263			if ((mad_reg_req->mgmt_class !=
264					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
265			    (mad_reg_req->mgmt_class !=
266					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
267				goto error1;
268		} else {
269			if ((mad_reg_req->mgmt_class ==
270					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
271			    (mad_reg_req->mgmt_class ==
272					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
273				goto error1;
274		}
275	} else {
276		/* No registration request supplied */
277		if (!send_handler)
278			goto error1;
279	}
280
281	/* Validate device and port */
282	port_priv = ib_get_mad_port(device, port_num);
283	if (!port_priv) {
284		ret = ERR_PTR(-ENODEV);
285		goto error1;
286	}
287
288	/* Allocate structures */
289	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
290	if (!mad_agent_priv) {
291		ret = ERR_PTR(-ENOMEM);
292		goto error1;
293	}
294
295	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
296						 IB_ACCESS_LOCAL_WRITE);
297	if (IS_ERR(mad_agent_priv->agent.mr)) {
298		ret = ERR_PTR(-ENOMEM);
299		goto error2;
300	}
301
302	if (mad_reg_req) {
303		reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
304		if (!reg_req) {
305			ret = ERR_PTR(-ENOMEM);
306			goto error3;
307		}
308		/* Make a copy of the MAD registration request */
309		memcpy(reg_req, mad_reg_req, sizeof *reg_req);
310	}
311
312	/* Now, fill in the various structures */
313	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
314	mad_agent_priv->reg_req = reg_req;
315	mad_agent_priv->agent.rmpp_version = rmpp_version;
316	mad_agent_priv->agent.device = device;
317	mad_agent_priv->agent.recv_handler = recv_handler;
318	mad_agent_priv->agent.send_handler = send_handler;
319	mad_agent_priv->agent.context = context;
320	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
321	mad_agent_priv->agent.port_num = port_num;
322	spin_lock_init(&mad_agent_priv->lock);
323	INIT_LIST_HEAD(&mad_agent_priv->send_list);
324	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
325	INIT_LIST_HEAD(&mad_agent_priv->done_list);
326	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
327	INIT_WORK(&mad_agent_priv->timeout_work, timeout_sends);
328	setup_timer(&mad_agent_priv->timeout_timer, timeout_callback,
329		    (unsigned long) mad_agent_priv);
330	INIT_LIST_HEAD(&mad_agent_priv->local_list);
331	INIT_WORK(&mad_agent_priv->local_work, local_completions);
332	atomic_set(&mad_agent_priv->refcount, 1);
333	init_completion(&mad_agent_priv->comp);
334
335	spin_lock_irqsave(&port_priv->reg_lock, flags);
336	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
337
338	/*
339	 * Make sure MAD registration (if supplied)
340	 * is non overlapping with any existing ones
341	 */
342	if (mad_reg_req) {
343		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
344		if (!is_vendor_class(mgmt_class)) {
345			class = port_priv->version[mad_reg_req->
346						   mgmt_class_version].class;
347			if (class) {
348				method = class->method_table[mgmt_class];
349				if (method) {
350					if (method_in_use(&method,
351							   mad_reg_req))
352						goto error4;
353				}
354			}
355			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
356						  mgmt_class);
357		} else {
358			/* "New" vendor class range */
359			vendor = port_priv->version[mad_reg_req->
360						    mgmt_class_version].vendor;
361			if (vendor) {
362				vclass = vendor_class_index(mgmt_class);
363				vendor_class = vendor->vendor_class[vclass];
364				if (vendor_class) {
365					if (is_vendor_method_in_use(
366							vendor_class,
367							mad_reg_req))
368						goto error4;
369				}
370			}
371			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
372		}
373		if (ret2) {
374			ret = ERR_PTR(ret2);
375			goto error4;
376		}
377	}
378
379	/* Add mad agent into port's agent list */
380	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
381	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
382
383	return &mad_agent_priv->agent;
384
385error4:
386	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
387	kfree(reg_req);
388error3:
389	ib_dereg_mr(mad_agent_priv->agent.mr);
390error2:
391	kfree(mad_agent_priv);
392error1:
393	return ret;
394}
395EXPORT_SYMBOL(ib_register_mad_agent);
396
397static inline int is_snooping_sends(int mad_snoop_flags)
398{
399	return (mad_snoop_flags &
400		(/*IB_MAD_SNOOP_POSTED_SENDS |
401		 IB_MAD_SNOOP_RMPP_SENDS |*/
402		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
403		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
404}
405
406static inline int is_snooping_recvs(int mad_snoop_flags)
407{
408	return (mad_snoop_flags &
409		(IB_MAD_SNOOP_RECVS /*|
410		 IB_MAD_SNOOP_RMPP_RECVS*/));
411}
412
413static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
414				struct ib_mad_snoop_private *mad_snoop_priv)
415{
416	struct ib_mad_snoop_private **new_snoop_table;
417	unsigned long flags;
418	int i;
419
420	spin_lock_irqsave(&qp_info->snoop_lock, flags);
421	/* Check for empty slot in array. */
422	for (i = 0; i < qp_info->snoop_table_size; i++)
423		if (!qp_info->snoop_table[i])
424			break;
425
426	if (i == qp_info->snoop_table_size) {
427		/* Grow table. */
428		new_snoop_table = krealloc(qp_info->snoop_table,
429					   sizeof mad_snoop_priv *
430					   (qp_info->snoop_table_size + 1),
431					   GFP_ATOMIC);
432		if (!new_snoop_table) {
433			i = -ENOMEM;
434			goto out;
435		}
436
437		qp_info->snoop_table = new_snoop_table;
438		qp_info->snoop_table_size++;
439	}
440	qp_info->snoop_table[i] = mad_snoop_priv;
441	atomic_inc(&qp_info->snoop_count);
442out:
443	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
444	return i;
445}
446
447struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
448					   u8 port_num,
449					   enum ib_qp_type qp_type,
450					   int mad_snoop_flags,
451					   ib_mad_snoop_handler snoop_handler,
452					   ib_mad_recv_handler recv_handler,
453					   void *context)
454{
455	struct ib_mad_port_private *port_priv;
456	struct ib_mad_agent *ret;
457	struct ib_mad_snoop_private *mad_snoop_priv;
458	int qpn;
459
460	/* Validate parameters */
461	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
462	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
463		ret = ERR_PTR(-EINVAL);
464		goto error1;
465	}
466	qpn = get_spl_qp_index(qp_type);
467	if (qpn == -1) {
468		ret = ERR_PTR(-EINVAL);
469		goto error1;
470	}
471	port_priv = ib_get_mad_port(device, port_num);
472	if (!port_priv) {
473		ret = ERR_PTR(-ENODEV);
474		goto error1;
475	}
476	/* Allocate structures */
477	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
478	if (!mad_snoop_priv) {
479		ret = ERR_PTR(-ENOMEM);
480		goto error1;
481	}
482
483	/* Now, fill in the various structures */
484	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
485	mad_snoop_priv->agent.device = device;
486	mad_snoop_priv->agent.recv_handler = recv_handler;
487	mad_snoop_priv->agent.snoop_handler = snoop_handler;
488	mad_snoop_priv->agent.context = context;
489	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
490	mad_snoop_priv->agent.port_num = port_num;
491	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
492	init_completion(&mad_snoop_priv->comp);
493	mad_snoop_priv->snoop_index = register_snoop_agent(
494						&port_priv->qp_info[qpn],
495						mad_snoop_priv);
496	if (mad_snoop_priv->snoop_index < 0) {
497		ret = ERR_PTR(mad_snoop_priv->snoop_index);
498		goto error2;
499	}
500
501	atomic_set(&mad_snoop_priv->refcount, 1);
502	return &mad_snoop_priv->agent;
503
504error2:
505	kfree(mad_snoop_priv);
506error1:
507	return ret;
508}
509EXPORT_SYMBOL(ib_register_mad_snoop);
510
511static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
512{
513	if (atomic_dec_and_test(&mad_agent_priv->refcount))
514		complete(&mad_agent_priv->comp);
515}
516
517static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
518{
519	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
520		complete(&mad_snoop_priv->comp);
521}
522
523static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
524{
525	struct ib_mad_port_private *port_priv;
526	unsigned long flags;
527
528	/* Note that we could still be handling received MADs */
529
530	/*
531	 * Canceling all sends results in dropping received response
532	 * MADs, preventing us from queuing additional work
533	 */
534	cancel_mads(mad_agent_priv);
535	port_priv = mad_agent_priv->qp_info->port_priv;
536	del_timer_sync(&mad_agent_priv->timeout_timer);
537	cancel_work_sync(&mad_agent_priv->timeout_work);
538
539	spin_lock_irqsave(&port_priv->reg_lock, flags);
540	remove_mad_reg_req(mad_agent_priv);
541	list_del(&mad_agent_priv->agent_list);
542	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
543
544	flush_workqueue(port_priv->wq);
545	ib_cancel_rmpp_recvs(mad_agent_priv);
546
547	deref_mad_agent(mad_agent_priv);
548	wait_for_completion(&mad_agent_priv->comp);
549
550	kfree(mad_agent_priv->reg_req);
551	ib_dereg_mr(mad_agent_priv->agent.mr);
552	kfree(mad_agent_priv);
553}
554
555static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
556{
557	struct ib_mad_qp_info *qp_info;
558	unsigned long flags;
559
560	qp_info = mad_snoop_priv->qp_info;
561	spin_lock_irqsave(&qp_info->snoop_lock, flags);
562	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
563	atomic_dec(&qp_info->snoop_count);
564	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
565
566	deref_snoop_agent(mad_snoop_priv);
567	wait_for_completion(&mad_snoop_priv->comp);
568
569	kfree(mad_snoop_priv);
570}
571
572/*
573 * ib_unregister_mad_agent - Unregisters a client from using MAD services
574 */
575int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
576{
577	struct ib_mad_agent_private *mad_agent_priv;
578	struct ib_mad_snoop_private *mad_snoop_priv;
579
580	/* If the TID is zero, the agent can only snoop. */
581	if (mad_agent->hi_tid) {
582		mad_agent_priv = container_of(mad_agent,
583					      struct ib_mad_agent_private,
584					      agent);
585		unregister_mad_agent(mad_agent_priv);
586	} else {
587		mad_snoop_priv = container_of(mad_agent,
588					      struct ib_mad_snoop_private,
589					      agent);
590		unregister_mad_snoop(mad_snoop_priv);
591	}
592	return 0;
593}
594EXPORT_SYMBOL(ib_unregister_mad_agent);
595
596static void dequeue_mad(struct ib_mad_list_head *mad_list)
597{
598	struct ib_mad_queue *mad_queue;
599	unsigned long flags;
600
601	BUG_ON(!mad_list->mad_queue);
602	mad_queue = mad_list->mad_queue;
603	spin_lock_irqsave(&mad_queue->lock, flags);
604	list_del(&mad_list->list);
605	mad_queue->count--;
606	spin_unlock_irqrestore(&mad_queue->lock, flags);
607}
608
609static void snoop_send(struct ib_mad_qp_info *qp_info,
610		       struct ib_mad_send_buf *send_buf,
611		       struct ib_mad_send_wc *mad_send_wc,
612		       int mad_snoop_flags)
613{
614	struct ib_mad_snoop_private *mad_snoop_priv;
615	unsigned long flags;
616	int i;
617
618	spin_lock_irqsave(&qp_info->snoop_lock, flags);
619	for (i = 0; i < qp_info->snoop_table_size; i++) {
620		mad_snoop_priv = qp_info->snoop_table[i];
621		if (!mad_snoop_priv ||
622		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
623			continue;
624
625		atomic_inc(&mad_snoop_priv->refcount);
626		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
627		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
628						    send_buf, mad_send_wc);
629		deref_snoop_agent(mad_snoop_priv);
630		spin_lock_irqsave(&qp_info->snoop_lock, flags);
631	}
632	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
633}
634
635static void snoop_recv(struct ib_mad_qp_info *qp_info,
636		       struct ib_mad_recv_wc *mad_recv_wc,
637		       int mad_snoop_flags)
638{
639	struct ib_mad_snoop_private *mad_snoop_priv;
640	unsigned long flags;
641	int i;
642
643	spin_lock_irqsave(&qp_info->snoop_lock, flags);
644	for (i = 0; i < qp_info->snoop_table_size; i++) {
645		mad_snoop_priv = qp_info->snoop_table[i];
646		if (!mad_snoop_priv ||
647		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
648			continue;
649
650		atomic_inc(&mad_snoop_priv->refcount);
651		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
652		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
653						   mad_recv_wc);
654		deref_snoop_agent(mad_snoop_priv);
655		spin_lock_irqsave(&qp_info->snoop_lock, flags);
656	}
657	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
658}
659
660static void build_smp_wc(struct ib_qp *qp,
661			 u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
662			 struct ib_wc *wc)
663{
664	memset(wc, 0, sizeof *wc);
665	wc->wr_id = wr_id;
666	wc->status = IB_WC_SUCCESS;
667	wc->opcode = IB_WC_RECV;
668	wc->pkey_index = pkey_index;
669	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
670	wc->src_qp = IB_QP0;
671	wc->qp = qp;
672	wc->slid = slid;
673	wc->sl = 0;
674	wc->dlid_path_bits = 0;
675	wc->port_num = port_num;
676}
677
678/*
679 * Return 0 if SMP is to be sent
680 * Return 1 if SMP was consumed locally (whether or not solicited)
681 * Return < 0 if error
682 */
683static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
684				  struct ib_mad_send_wr_private *mad_send_wr)
685{
686	int ret = 0;
687	struct ib_smp *smp = mad_send_wr->send_buf.mad;
688	unsigned long flags;
689	struct ib_mad_local_private *local;
690	struct ib_mad_private *mad_priv;
691	struct ib_mad_port_private *port_priv;
692	struct ib_mad_agent_private *recv_mad_agent = NULL;
693	struct ib_device *device = mad_agent_priv->agent.device;
694	u8 port_num;
695	struct ib_wc mad_wc;
696	struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
697
698	if (device->node_type == RDMA_NODE_IB_SWITCH)
699		port_num = send_wr->wr.ud.port_num;
700	else
701		port_num = mad_agent_priv->agent.port_num;
702
703	/*
704	 * Directed route handling starts if the initial LID routed part of
705	 * a request or the ending LID routed part of a response is empty.
706	 * If we are at the start of the LID routed part, don't update the
707	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
708	 */
709	if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) !=
710	     IB_LID_PERMISSIVE)
711		goto out;
712	if (smi_handle_dr_smp_send(smp, device->node_type, port_num) ==
713	     IB_SMI_DISCARD) {
714		ret = -EINVAL;
715		printk(KERN_ERR PFX "Invalid directed route\n");
716		goto out;
717	}
718
719	/* Check to post send on QP or process locally */
720	if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
721	    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
722		goto out;
723
724	local = kmalloc(sizeof *local, GFP_ATOMIC);
725	if (!local) {
726		ret = -ENOMEM;
727		printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
728		goto out;
729	}
730	local->mad_priv = NULL;
731	local->recv_mad_agent = NULL;
732	mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
733	if (!mad_priv) {
734		ret = -ENOMEM;
735		printk(KERN_ERR PFX "No memory for local response MAD\n");
736		kfree(local);
737		goto out;
738	}
739
740	build_smp_wc(mad_agent_priv->agent.qp,
741		     send_wr->wr_id, be16_to_cpu(smp->dr_slid),
742		     send_wr->wr.ud.pkey_index,
743		     send_wr->wr.ud.port_num, &mad_wc);
744
745	/* No GRH for DR SMP */
746	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
747				  (struct ib_mad *)smp,
748				  (struct ib_mad *)&mad_priv->mad);
749	switch (ret)
750	{
751	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
752		if (ib_response_mad(&mad_priv->mad.mad) &&
753		    mad_agent_priv->agent.recv_handler) {
754			local->mad_priv = mad_priv;
755			local->recv_mad_agent = mad_agent_priv;
756			/*
757			 * Reference MAD agent until receive
758			 * side of local completion handled
759			 */
760			atomic_inc(&mad_agent_priv->refcount);
761		} else
762			kmem_cache_free(ib_mad_cache, mad_priv);
763		break;
764	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
765		kmem_cache_free(ib_mad_cache, mad_priv);
766		break;
767	case IB_MAD_RESULT_SUCCESS:
768		/* Treat like an incoming receive MAD */
769		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
770					    mad_agent_priv->agent.port_num);
771		if (port_priv) {
772			memcpy(&mad_priv->mad.mad, smp, sizeof(struct ib_mad));
773			recv_mad_agent = find_mad_agent(port_priv,
774						        &mad_priv->mad.mad);
775		}
776		if (!port_priv || !recv_mad_agent) {
777			/*
778			 * No receiving agent so drop packet and
779			 * generate send completion.
780			 */
781			kmem_cache_free(ib_mad_cache, mad_priv);
782			break;
783		}
784		local->mad_priv = mad_priv;
785		local->recv_mad_agent = recv_mad_agent;
786		break;
787	default:
788		kmem_cache_free(ib_mad_cache, mad_priv);
789		kfree(local);
790		ret = -EINVAL;
791		goto out;
792	}
793
794	local->mad_send_wr = mad_send_wr;
795	/* Reference MAD agent until send side of local completion handled */
796	atomic_inc(&mad_agent_priv->refcount);
797	/* Queue local completion to local list */
798	spin_lock_irqsave(&mad_agent_priv->lock, flags);
799	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
800	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
801	queue_work(mad_agent_priv->qp_info->port_priv->wq,
802		   &mad_agent_priv->local_work);
803	ret = 1;
804out:
805	return ret;
806}
807
808static int get_pad_size(int hdr_len, int data_len)
809{
810	int seg_size, pad;
811
812	seg_size = sizeof(struct ib_mad) - hdr_len;
813	if (data_len && seg_size) {
814		pad = seg_size - data_len % seg_size;
815		return pad == seg_size ? 0 : pad;
816	} else
817		return seg_size;
818}
819
820static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
821{
822	struct ib_rmpp_segment *s, *t;
823
824	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
825		list_del(&s->list);
826		kfree(s);
827	}
828}
829
830static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
831				gfp_t gfp_mask)
832{
833	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
834	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
835	struct ib_rmpp_segment *seg = NULL;
836	int left, seg_size, pad;
837
838	send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
839	seg_size = send_buf->seg_size;
840	pad = send_wr->pad;
841
842	/* Allocate data segments. */
843	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
844		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
845		if (!seg) {
846			printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
847			       "alloc failed for len %zd, gfp %#x\n",
848			       sizeof (*seg) + seg_size, gfp_mask);
849			free_send_rmpp_list(send_wr);
850			return -ENOMEM;
851		}
852		seg->num = ++send_buf->seg_count;
853		list_add_tail(&seg->list, &send_wr->rmpp_list);
854	}
855
856	/* Zero any padding */
857	if (pad)
858		memset(seg->data + seg_size - pad, 0, pad);
859
860	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
861					  agent.rmpp_version;
862	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
863	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
864
865	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
866					struct ib_rmpp_segment, list);
867	send_wr->last_ack_seg = send_wr->cur_seg;
868	return 0;
869}
870
871struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
872					    u32 remote_qpn, u16 pkey_index,
873					    int rmpp_active,
874					    int hdr_len, int data_len,
875					    gfp_t gfp_mask)
876{
877	struct ib_mad_agent_private *mad_agent_priv;
878	struct ib_mad_send_wr_private *mad_send_wr;
879	int pad, message_size, ret, size;
880	void *buf;
881
882	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
883				      agent);
884	pad = get_pad_size(hdr_len, data_len);
885	message_size = hdr_len + data_len + pad;
886
887	if ((!mad_agent->rmpp_version &&
888	     (rmpp_active || message_size > sizeof(struct ib_mad))) ||
889	    (!rmpp_active && message_size > sizeof(struct ib_mad)))
890		return ERR_PTR(-EINVAL);
891
892	size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
893	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
894	if (!buf)
895		return ERR_PTR(-ENOMEM);
896
897	mad_send_wr = buf + size;
898	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
899	mad_send_wr->send_buf.mad = buf;
900	mad_send_wr->send_buf.hdr_len = hdr_len;
901	mad_send_wr->send_buf.data_len = data_len;
902	mad_send_wr->pad = pad;
903
904	mad_send_wr->mad_agent_priv = mad_agent_priv;
905	mad_send_wr->sg_list[0].length = hdr_len;
906	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
907	mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
908	mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
909
910	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
911	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
912	mad_send_wr->send_wr.num_sge = 2;
913	mad_send_wr->send_wr.opcode = IB_WR_SEND;
914	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
915	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
916	mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
917	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
918
919	if (rmpp_active) {
920		ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
921		if (ret) {
922			kfree(buf);
923			return ERR_PTR(ret);
924		}
925	}
926
927	mad_send_wr->send_buf.mad_agent = mad_agent;
928	atomic_inc(&mad_agent_priv->refcount);
929	return &mad_send_wr->send_buf;
930}
931EXPORT_SYMBOL(ib_create_send_mad);
932
933int ib_get_mad_data_offset(u8 mgmt_class)
934{
935	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
936		return IB_MGMT_SA_HDR;
937	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
938		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
939		 (mgmt_class == IB_MGMT_CLASS_BIS))
940		return IB_MGMT_DEVICE_HDR;
941	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
942		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
943		return IB_MGMT_VENDOR_HDR;
944	else
945		return IB_MGMT_MAD_HDR;
946}
947EXPORT_SYMBOL(ib_get_mad_data_offset);
948
949int ib_is_mad_class_rmpp(u8 mgmt_class)
950{
951	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
952	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
953	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
954	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
955	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
956	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
957		return 1;
958	return 0;
959}
960EXPORT_SYMBOL(ib_is_mad_class_rmpp);
961
962void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
963{
964	struct ib_mad_send_wr_private *mad_send_wr;
965	struct list_head *list;
966
967	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
968				   send_buf);
969	list = &mad_send_wr->cur_seg->list;
970
971	if (mad_send_wr->cur_seg->num < seg_num) {
972		list_for_each_entry(mad_send_wr->cur_seg, list, list)
973			if (mad_send_wr->cur_seg->num == seg_num)
974				break;
975	} else if (mad_send_wr->cur_seg->num > seg_num) {
976		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
977			if (mad_send_wr->cur_seg->num == seg_num)
978				break;
979	}
980	return mad_send_wr->cur_seg->data;
981}
982EXPORT_SYMBOL(ib_get_rmpp_segment);
983
984static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
985{
986	if (mad_send_wr->send_buf.seg_count)
987		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
988					   mad_send_wr->seg_num);
989	else
990		return mad_send_wr->send_buf.mad +
991		       mad_send_wr->send_buf.hdr_len;
992}
993
994void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
995{
996	struct ib_mad_agent_private *mad_agent_priv;
997	struct ib_mad_send_wr_private *mad_send_wr;
998
999	mad_agent_priv = container_of(send_buf->mad_agent,
1000				      struct ib_mad_agent_private, agent);
1001	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1002				   send_buf);
1003
1004	free_send_rmpp_list(mad_send_wr);
1005	kfree(send_buf->mad);
1006	deref_mad_agent(mad_agent_priv);
1007}
1008EXPORT_SYMBOL(ib_free_send_mad);
1009
1010int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1011{
1012	struct ib_mad_qp_info *qp_info;
1013	struct list_head *list;
1014	struct ib_send_wr *bad_send_wr;
1015	struct ib_mad_agent *mad_agent;
1016	struct ib_sge *sge;
1017	unsigned long flags;
1018	int ret;
1019
1020	/* Set WR ID to find mad_send_wr upon completion */
1021	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1022	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
1023	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1024
1025	mad_agent = mad_send_wr->send_buf.mad_agent;
1026	sge = mad_send_wr->sg_list;
1027	sge[0].addr = ib_dma_map_single(mad_agent->device,
1028					mad_send_wr->send_buf.mad,
1029					sge[0].length,
1030					DMA_TO_DEVICE);
1031	mad_send_wr->header_mapping = sge[0].addr;
1032
1033	sge[1].addr = ib_dma_map_single(mad_agent->device,
1034					ib_get_payload(mad_send_wr),
1035					sge[1].length,
1036					DMA_TO_DEVICE);
1037	mad_send_wr->payload_mapping = sge[1].addr;
1038
1039	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1040	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1041		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1042				   &bad_send_wr);
1043		list = &qp_info->send_queue.list;
1044	} else {
1045		ret = 0;
1046		list = &qp_info->overflow_list;
1047	}
1048
1049	if (!ret) {
1050		qp_info->send_queue.count++;
1051		list_add_tail(&mad_send_wr->mad_list.list, list);
1052	}
1053	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1054	if (ret) {
1055		ib_dma_unmap_single(mad_agent->device,
1056				    mad_send_wr->header_mapping,
1057				    sge[0].length, DMA_TO_DEVICE);
1058		ib_dma_unmap_single(mad_agent->device,
1059				    mad_send_wr->payload_mapping,
1060				    sge[1].length, DMA_TO_DEVICE);
1061	}
1062	return ret;
1063}
1064
1065/*
1066 * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1067 *  with the registered client
1068 */
1069int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1070		     struct ib_mad_send_buf **bad_send_buf)
1071{
1072	struct ib_mad_agent_private *mad_agent_priv;
1073	struct ib_mad_send_buf *next_send_buf;
1074	struct ib_mad_send_wr_private *mad_send_wr;
1075	unsigned long flags;
1076	int ret = -EINVAL;
1077
1078	/* Walk list of send WRs and post each on send list */
1079	for (; send_buf; send_buf = next_send_buf) {
1080
1081		mad_send_wr = container_of(send_buf,
1082					   struct ib_mad_send_wr_private,
1083					   send_buf);
1084		mad_agent_priv = mad_send_wr->mad_agent_priv;
1085
1086		if (!send_buf->mad_agent->send_handler ||
1087		    (send_buf->timeout_ms &&
1088		     !send_buf->mad_agent->recv_handler)) {
1089			ret = -EINVAL;
1090			goto error;
1091		}
1092
1093		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1094			if (mad_agent_priv->agent.rmpp_version) {
1095				ret = -EINVAL;
1096				goto error;
1097			}
1098		}
1099
1100		/*
1101		 * Save pointer to next work request to post in case the
1102		 * current one completes, and the user modifies the work
1103		 * request associated with the completion
1104		 */
1105		next_send_buf = send_buf->next;
1106		mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1107
1108		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1109		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1110			ret = handle_outgoing_dr_smp(mad_agent_priv,
1111						     mad_send_wr);
1112			if (ret < 0)		/* error */
1113				goto error;
1114			else if (ret == 1)	/* locally consumed */
1115				continue;
1116		}
1117
1118		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1119		/* Timeout will be updated after send completes */
1120		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1121		mad_send_wr->max_retries = send_buf->retries;
1122		mad_send_wr->retries_left = send_buf->retries;
1123		send_buf->retries = 0;
1124		/* Reference for work request to QP + response */
1125		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1126		mad_send_wr->status = IB_WC_SUCCESS;
1127
1128		/* Reference MAD agent until send completes */
1129		atomic_inc(&mad_agent_priv->refcount);
1130		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1131		list_add_tail(&mad_send_wr->agent_list,
1132			      &mad_agent_priv->send_list);
1133		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1134
1135		if (mad_agent_priv->agent.rmpp_version) {
1136			ret = ib_send_rmpp_mad(mad_send_wr);
1137			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1138				ret = ib_send_mad(mad_send_wr);
1139		} else
1140			ret = ib_send_mad(mad_send_wr);
1141		if (ret < 0) {
1142			/* Fail send request */
1143			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1144			list_del(&mad_send_wr->agent_list);
1145			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1146			atomic_dec(&mad_agent_priv->refcount);
1147			goto error;
1148		}
1149	}
1150	return 0;
1151error:
1152	if (bad_send_buf)
1153		*bad_send_buf = send_buf;
1154	return ret;
1155}
1156EXPORT_SYMBOL(ib_post_send_mad);
1157
1158/*
1159 * ib_free_recv_mad - Returns data buffers used to receive
1160 *  a MAD to the access layer
1161 */
1162void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1163{
1164	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1165	struct ib_mad_private_header *mad_priv_hdr;
1166	struct ib_mad_private *priv;
1167	struct list_head free_list;
1168
1169	INIT_LIST_HEAD(&free_list);
1170	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1171
1172	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1173					&free_list, list) {
1174		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1175					   recv_buf);
1176		mad_priv_hdr = container_of(mad_recv_wc,
1177					    struct ib_mad_private_header,
1178					    recv_wc);
1179		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1180				    header);
1181		kmem_cache_free(ib_mad_cache, priv);
1182	}
1183}
1184EXPORT_SYMBOL(ib_free_recv_mad);
1185
1186struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1187					u8 rmpp_version,
1188					ib_mad_send_handler send_handler,
1189					ib_mad_recv_handler recv_handler,
1190					void *context)
1191{
1192	return ERR_PTR(-EINVAL);	/* XXX: for now */
1193}
1194EXPORT_SYMBOL(ib_redirect_mad_qp);
1195
1196int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1197		      struct ib_wc *wc)
1198{
1199	printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1200	return 0;
1201}
1202EXPORT_SYMBOL(ib_process_mad_wc);
1203
1204static int method_in_use(struct ib_mad_mgmt_method_table **method,
1205			 struct ib_mad_reg_req *mad_reg_req)
1206{
1207	int i;
1208
1209	for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1210	     i < IB_MGMT_MAX_METHODS;
1211	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1212			       1+i)) {
1213		if ((*method)->agent[i]) {
1214			printk(KERN_ERR PFX "Method %d already in use\n", i);
1215			return -EINVAL;
1216		}
1217	}
1218	return 0;
1219}
1220
1221static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1222{
1223	/* Allocate management method table */
1224	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1225	if (!*method) {
1226		printk(KERN_ERR PFX "No memory for "
1227		       "ib_mad_mgmt_method_table\n");
1228		return -ENOMEM;
1229	}
1230
1231	return 0;
1232}
1233
1234/*
1235 * Check to see if there are any methods still in use
1236 */
1237static int check_method_table(struct ib_mad_mgmt_method_table *method)
1238{
1239	int i;
1240
1241	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1242		if (method->agent[i])
1243			return 1;
1244	return 0;
1245}
1246
1247/*
1248 * Check to see if there are any method tables for this class still in use
1249 */
1250static int check_class_table(struct ib_mad_mgmt_class_table *class)
1251{
1252	int i;
1253
1254	for (i = 0; i < MAX_MGMT_CLASS; i++)
1255		if (class->method_table[i])
1256			return 1;
1257	return 0;
1258}
1259
1260static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1261{
1262	int i;
1263
1264	for (i = 0; i < MAX_MGMT_OUI; i++)
1265		if (vendor_class->method_table[i])
1266			return 1;
1267	return 0;
1268}
1269
1270static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1271			   char *oui)
1272{
1273	int i;
1274
1275	for (i = 0; i < MAX_MGMT_OUI; i++)
1276		/* Is there matching OUI for this vendor class ? */
1277		if (!memcmp(vendor_class->oui[i], oui, 3))
1278			return i;
1279
1280	return -1;
1281}
1282
1283static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1284{
1285	int i;
1286
1287	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1288		if (vendor->vendor_class[i])
1289			return 1;
1290
1291	return 0;
1292}
1293
1294static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1295				     struct ib_mad_agent_private *agent)
1296{
1297	int i;
1298
1299	/* Remove any methods for this mad agent */
1300	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1301		if (method->agent[i] == agent) {
1302			method->agent[i] = NULL;
1303		}
1304	}
1305}
1306
1307static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1308			      struct ib_mad_agent_private *agent_priv,
1309			      u8 mgmt_class)
1310{
1311	struct ib_mad_port_private *port_priv;
1312	struct ib_mad_mgmt_class_table **class;
1313	struct ib_mad_mgmt_method_table **method;
1314	int i, ret;
1315
1316	port_priv = agent_priv->qp_info->port_priv;
1317	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1318	if (!*class) {
1319		/* Allocate management class table for "new" class version */
1320		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1321		if (!*class) {
1322			printk(KERN_ERR PFX "No memory for "
1323			       "ib_mad_mgmt_class_table\n");
1324			ret = -ENOMEM;
1325			goto error1;
1326		}
1327
1328		/* Allocate method table for this management class */
1329		method = &(*class)->method_table[mgmt_class];
1330		if ((ret = allocate_method_table(method)))
1331			goto error2;
1332	} else {
1333		method = &(*class)->method_table[mgmt_class];
1334		if (!*method) {
1335			/* Allocate method table for this management class */
1336			if ((ret = allocate_method_table(method)))
1337				goto error1;
1338		}
1339	}
1340
1341	/* Now, make sure methods are not already in use */
1342	if (method_in_use(method, mad_reg_req))
1343		goto error3;
1344
1345	/* Finally, add in methods being registered */
1346	for (i = find_first_bit(mad_reg_req->method_mask,
1347				IB_MGMT_MAX_METHODS);
1348	     i < IB_MGMT_MAX_METHODS;
1349	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1350			       1+i)) {
1351		(*method)->agent[i] = agent_priv;
1352	}
1353	return 0;
1354
1355error3:
1356	/* Remove any methods for this mad agent */
1357	remove_methods_mad_agent(*method, agent_priv);
1358	/* Now, check to see if there are any methods in use */
1359	if (!check_method_table(*method)) {
1360		/* If not, release management method table */
1361		kfree(*method);
1362		*method = NULL;
1363	}
1364	ret = -EINVAL;
1365	goto error1;
1366error2:
1367	kfree(*class);
1368	*class = NULL;
1369error1:
1370	return ret;
1371}
1372
1373static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1374			   struct ib_mad_agent_private *agent_priv)
1375{
1376	struct ib_mad_port_private *port_priv;
1377	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1378	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1379	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1380	struct ib_mad_mgmt_method_table **method;
1381	int i, ret = -ENOMEM;
1382	u8 vclass;
1383
1384	/* "New" vendor (with OUI) class */
1385	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1386	port_priv = agent_priv->qp_info->port_priv;
1387	vendor_table = &port_priv->version[
1388				mad_reg_req->mgmt_class_version].vendor;
1389	if (!*vendor_table) {
1390		/* Allocate mgmt vendor class table for "new" class version */
1391		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1392		if (!vendor) {
1393			printk(KERN_ERR PFX "No memory for "
1394			       "ib_mad_mgmt_vendor_class_table\n");
1395			goto error1;
1396		}
1397
1398		*vendor_table = vendor;
1399	}
1400	if (!(*vendor_table)->vendor_class[vclass]) {
1401		/* Allocate table for this management vendor class */
1402		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1403		if (!vendor_class) {
1404			printk(KERN_ERR PFX "No memory for "
1405			       "ib_mad_mgmt_vendor_class\n");
1406			goto error2;
1407		}
1408
1409		(*vendor_table)->vendor_class[vclass] = vendor_class;
1410	}
1411	for (i = 0; i < MAX_MGMT_OUI; i++) {
1412		/* Is there matching OUI for this vendor class ? */
1413		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1414			    mad_reg_req->oui, 3)) {
1415			method = &(*vendor_table)->vendor_class[
1416						vclass]->method_table[i];
1417			BUG_ON(!*method);
1418			goto check_in_use;
1419		}
1420	}
1421	for (i = 0; i < MAX_MGMT_OUI; i++) {
1422		/* OUI slot available ? */
1423		if (!is_vendor_oui((*vendor_table)->vendor_class[
1424				vclass]->oui[i])) {
1425			method = &(*vendor_table)->vendor_class[
1426				vclass]->method_table[i];
1427			BUG_ON(*method);
1428			/* Allocate method table for this OUI */
1429			if ((ret = allocate_method_table(method)))
1430				goto error3;
1431			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1432			       mad_reg_req->oui, 3);
1433			goto check_in_use;
1434		}
1435	}
1436	printk(KERN_ERR PFX "All OUI slots in use\n");
1437	goto error3;
1438
1439check_in_use:
1440	/* Now, make sure methods are not already in use */
1441	if (method_in_use(method, mad_reg_req))
1442		goto error4;
1443
1444	/* Finally, add in methods being registered */
1445	for (i = find_first_bit(mad_reg_req->method_mask,
1446				IB_MGMT_MAX_METHODS);
1447	     i < IB_MGMT_MAX_METHODS;
1448	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1449			       1+i)) {
1450		(*method)->agent[i] = agent_priv;
1451	}
1452	return 0;
1453
1454error4:
1455	/* Remove any methods for this mad agent */
1456	remove_methods_mad_agent(*method, agent_priv);
1457	/* Now, check to see if there are any methods in use */
1458	if (!check_method_table(*method)) {
1459		/* If not, release management method table */
1460		kfree(*method);
1461		*method = NULL;
1462	}
1463	ret = -EINVAL;
1464error3:
1465	if (vendor_class) {
1466		(*vendor_table)->vendor_class[vclass] = NULL;
1467		kfree(vendor_class);
1468	}
1469error2:
1470	if (vendor) {
1471		*vendor_table = NULL;
1472		kfree(vendor);
1473	}
1474error1:
1475	return ret;
1476}
1477
1478static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1479{
1480	struct ib_mad_port_private *port_priv;
1481	struct ib_mad_mgmt_class_table *class;
1482	struct ib_mad_mgmt_method_table *method;
1483	struct ib_mad_mgmt_vendor_class_table *vendor;
1484	struct ib_mad_mgmt_vendor_class *vendor_class;
1485	int index;
1486	u8 mgmt_class;
1487
1488	/*
1489	 * Was MAD registration request supplied
1490	 * with original registration ?
1491	 */
1492	if (!agent_priv->reg_req) {
1493		goto out;
1494	}
1495
1496	port_priv = agent_priv->qp_info->port_priv;
1497	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1498	class = port_priv->version[
1499			agent_priv->reg_req->mgmt_class_version].class;
1500	if (!class)
1501		goto vendor_check;
1502
1503	method = class->method_table[mgmt_class];
1504	if (method) {
1505		/* Remove any methods for this mad agent */
1506		remove_methods_mad_agent(method, agent_priv);
1507		/* Now, check to see if there are any methods still in use */
1508		if (!check_method_table(method)) {
1509			/* If not, release management method table */
1510			 kfree(method);
1511			 class->method_table[mgmt_class] = NULL;
1512			 /* Any management classes left ? */
1513			if (!check_class_table(class)) {
1514				/* If not, release management class table */
1515				kfree(class);
1516				port_priv->version[
1517					agent_priv->reg_req->
1518					mgmt_class_version].class = NULL;
1519			}
1520		}
1521	}
1522
1523vendor_check:
1524	if (!is_vendor_class(mgmt_class))
1525		goto out;
1526
1527	/* normalize mgmt_class to vendor range 2 */
1528	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1529	vendor = port_priv->version[
1530			agent_priv->reg_req->mgmt_class_version].vendor;
1531
1532	if (!vendor)
1533		goto out;
1534
1535	vendor_class = vendor->vendor_class[mgmt_class];
1536	if (vendor_class) {
1537		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1538		if (index < 0)
1539			goto out;
1540		method = vendor_class->method_table[index];
1541		if (method) {
1542			/* Remove any methods for this mad agent */
1543			remove_methods_mad_agent(method, agent_priv);
1544			/*
1545			 * Now, check to see if there are
1546			 * any methods still in use
1547			 */
1548			if (!check_method_table(method)) {
1549				/* If not, release management method table */
1550				kfree(method);
1551				vendor_class->method_table[index] = NULL;
1552				memset(vendor_class->oui[index], 0, 3);
1553				/* Any OUIs left ? */
1554				if (!check_vendor_class(vendor_class)) {
1555					/* If not, release vendor class table */
1556					kfree(vendor_class);
1557					vendor->vendor_class[mgmt_class] = NULL;
1558					/* Any other vendor classes left ? */
1559					if (!check_vendor_table(vendor)) {
1560						kfree(vendor);
1561						port_priv->version[
1562							agent_priv->reg_req->
1563							mgmt_class_version].
1564							vendor = NULL;
1565					}
1566				}
1567			}
1568		}
1569	}
1570
1571out:
1572	return;
1573}
1574
1575static struct ib_mad_agent_private *
1576find_mad_agent(struct ib_mad_port_private *port_priv,
1577	       struct ib_mad *mad)
1578{
1579	struct ib_mad_agent_private *mad_agent = NULL;
1580	unsigned long flags;
1581
1582	spin_lock_irqsave(&port_priv->reg_lock, flags);
1583	if (ib_response_mad(mad)) {
1584		u32 hi_tid;
1585		struct ib_mad_agent_private *entry;
1586
1587		/*
1588		 * Routing is based on high 32 bits of transaction ID
1589		 * of MAD.
1590		 */
1591		hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1592		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1593			if (entry->agent.hi_tid == hi_tid) {
1594				mad_agent = entry;
1595				break;
1596			}
1597		}
1598	} else {
1599		struct ib_mad_mgmt_class_table *class;
1600		struct ib_mad_mgmt_method_table *method;
1601		struct ib_mad_mgmt_vendor_class_table *vendor;
1602		struct ib_mad_mgmt_vendor_class *vendor_class;
1603		struct ib_vendor_mad *vendor_mad;
1604		int index;
1605
1606		/*
1607		 * Routing is based on version, class, and method
1608		 * For "newer" vendor MADs, also based on OUI
1609		 */
1610		if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1611			goto out;
1612		if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1613			class = port_priv->version[
1614					mad->mad_hdr.class_version].class;
1615			if (!class)
1616				goto out;
1617			method = class->method_table[convert_mgmt_class(
1618							mad->mad_hdr.mgmt_class)];
1619			if (method)
1620				mad_agent = method->agent[mad->mad_hdr.method &
1621							  ~IB_MGMT_METHOD_RESP];
1622		} else {
1623			vendor = port_priv->version[
1624					mad->mad_hdr.class_version].vendor;
1625			if (!vendor)
1626				goto out;
1627			vendor_class = vendor->vendor_class[vendor_class_index(
1628						mad->mad_hdr.mgmt_class)];
1629			if (!vendor_class)
1630				goto out;
1631			/* Find matching OUI */
1632			vendor_mad = (struct ib_vendor_mad *)mad;
1633			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1634			if (index == -1)
1635				goto out;
1636			method = vendor_class->method_table[index];
1637			if (method) {
1638				mad_agent = method->agent[mad->mad_hdr.method &
1639							  ~IB_MGMT_METHOD_RESP];
1640			}
1641		}
1642	}
1643
1644	if (mad_agent) {
1645		if (mad_agent->agent.recv_handler)
1646			atomic_inc(&mad_agent->refcount);
1647		else {
1648			printk(KERN_NOTICE PFX "No receive handler for client "
1649			       "%p on port %d\n",
1650			       &mad_agent->agent, port_priv->port_num);
1651			mad_agent = NULL;
1652		}
1653	}
1654out:
1655	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1656
1657	return mad_agent;
1658}
1659
1660static int validate_mad(struct ib_mad *mad, u32 qp_num)
1661{
1662	int valid = 0;
1663
1664	/* Make sure MAD base version is understood */
1665	if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1666		printk(KERN_ERR PFX "MAD received with unsupported base "
1667		       "version %d\n", mad->mad_hdr.base_version);
1668		goto out;
1669	}
1670
1671	/* Filter SMI packets sent to other than QP0 */
1672	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1673	    (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1674		if (qp_num == 0)
1675			valid = 1;
1676	} else {
1677		/* Filter GSI packets sent to QP0 */
1678		if (qp_num != 0)
1679			valid = 1;
1680	}
1681
1682out:
1683	return valid;
1684}
1685
1686static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1687		       struct ib_mad_hdr *mad_hdr)
1688{
1689	struct ib_rmpp_mad *rmpp_mad;
1690
1691	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1692	return !mad_agent_priv->agent.rmpp_version ||
1693		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1694				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1695		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1696}
1697
1698static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1699				     struct ib_mad_recv_wc *rwc)
1700{
1701	return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1702		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1703}
1704
1705static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1706				   struct ib_mad_send_wr_private *wr,
1707				   struct ib_mad_recv_wc *rwc )
1708{
1709	struct ib_ah_attr attr;
1710	u8 send_resp, rcv_resp;
1711	union ib_gid sgid;
1712	struct ib_device *device = mad_agent_priv->agent.device;
1713	u8 port_num = mad_agent_priv->agent.port_num;
1714	u8 lmc;
1715
1716	send_resp = ib_response_mad((struct ib_mad *)wr->send_buf.mad);
1717	rcv_resp = ib_response_mad(rwc->recv_buf.mad);
1718
1719	if (send_resp == rcv_resp)
1720		/* both requests, or both responses. GIDs different */
1721		return 0;
1722
1723	if (ib_query_ah(wr->send_buf.ah, &attr))
1724		/* Assume not equal, to avoid false positives. */
1725		return 0;
1726
1727	if (!!(attr.ah_flags & IB_AH_GRH) !=
1728	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1729		/* one has GID, other does not.  Assume different */
1730		return 0;
1731
1732	if (!send_resp && rcv_resp) {
1733		/* is request/response. */
1734		if (!(attr.ah_flags & IB_AH_GRH)) {
1735			if (ib_get_cached_lmc(device, port_num, &lmc))
1736				return 0;
1737			return (!lmc || !((attr.src_path_bits ^
1738					   rwc->wc->dlid_path_bits) &
1739					  ((1 << lmc) - 1)));
1740		} else {
1741			if (ib_get_cached_gid(device, port_num,
1742					      attr.grh.sgid_index, &sgid))
1743				return 0;
1744			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1745				       16);
1746		}
1747	}
1748
1749	if (!(attr.ah_flags & IB_AH_GRH))
1750		return attr.dlid == rwc->wc->slid;
1751	else
1752		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1753			       16);
1754}
1755
1756static inline int is_direct(u8 class)
1757{
1758	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1759}
1760
1761struct ib_mad_send_wr_private*
1762ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1763		 struct ib_mad_recv_wc *wc)
1764{
1765	struct ib_mad_send_wr_private *wr;
1766	struct ib_mad *mad;
1767
1768	mad = (struct ib_mad *)wc->recv_buf.mad;
1769
1770	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1771		if ((wr->tid == mad->mad_hdr.tid) &&
1772		    rcv_has_same_class(wr, wc) &&
1773		    /*
1774		     * Don't check GID for direct routed MADs.
1775		     * These might have permissive LIDs.
1776		     */
1777		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1778		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1779			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1780	}
1781
1782	/*
1783	 * It's possible to receive the response before we've
1784	 * been notified that the send has completed
1785	 */
1786	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1787		if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1788		    wr->tid == mad->mad_hdr.tid &&
1789		    wr->timeout &&
1790		    rcv_has_same_class(wr, wc) &&
1791		    /*
1792		     * Don't check GID for direct routed MADs.
1793		     * These might have permissive LIDs.
1794		     */
1795		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1796		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1797			/* Verify request has not been canceled */
1798			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1799	}
1800	return NULL;
1801}
1802
1803void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1804{
1805	mad_send_wr->timeout = 0;
1806	if (mad_send_wr->refcount == 1)
1807		list_move_tail(&mad_send_wr->agent_list,
1808			      &mad_send_wr->mad_agent_priv->done_list);
1809}
1810
1811static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1812				 struct ib_mad_recv_wc *mad_recv_wc)
1813{
1814	struct ib_mad_send_wr_private *mad_send_wr;
1815	struct ib_mad_send_wc mad_send_wc;
1816	unsigned long flags;
1817
1818	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1819	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1820	if (mad_agent_priv->agent.rmpp_version) {
1821		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1822						      mad_recv_wc);
1823		if (!mad_recv_wc) {
1824			deref_mad_agent(mad_agent_priv);
1825			return;
1826		}
1827	}
1828
1829	/* Complete corresponding request */
1830	if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1831		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1832		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1833		if (!mad_send_wr) {
1834			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1835			ib_free_recv_mad(mad_recv_wc);
1836			deref_mad_agent(mad_agent_priv);
1837			return;
1838		}
1839		ib_mark_mad_done(mad_send_wr);
1840		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1841
1842		/* Defined behavior is to complete response before request */
1843		mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1844		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1845						   mad_recv_wc);
1846		atomic_dec(&mad_agent_priv->refcount);
1847
1848		mad_send_wc.status = IB_WC_SUCCESS;
1849		mad_send_wc.vendor_err = 0;
1850		mad_send_wc.send_buf = &mad_send_wr->send_buf;
1851		ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1852	} else {
1853		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1854						   mad_recv_wc);
1855		deref_mad_agent(mad_agent_priv);
1856	}
1857}
1858
1859static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1860				     struct ib_wc *wc)
1861{
1862	struct ib_mad_qp_info *qp_info;
1863	struct ib_mad_private_header *mad_priv_hdr;
1864	struct ib_mad_private *recv, *response = NULL;
1865	struct ib_mad_list_head *mad_list;
1866	struct ib_mad_agent_private *mad_agent;
1867	int port_num;
1868
1869	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1870	qp_info = mad_list->mad_queue->qp_info;
1871	dequeue_mad(mad_list);
1872
1873	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1874				    mad_list);
1875	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1876	ib_dma_unmap_single(port_priv->device,
1877			    recv->header.mapping,
1878			    sizeof(struct ib_mad_private) -
1879			      sizeof(struct ib_mad_private_header),
1880			    DMA_FROM_DEVICE);
1881
1882	/* Setup MAD receive work completion from "normal" work completion */
1883	recv->header.wc = *wc;
1884	recv->header.recv_wc.wc = &recv->header.wc;
1885	recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1886	recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1887	recv->header.recv_wc.recv_buf.grh = &recv->grh;
1888
1889	if (atomic_read(&qp_info->snoop_count))
1890		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1891
1892	/* Validate MAD */
1893	if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1894		goto out;
1895
1896	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1897	if (!response) {
1898		printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1899		       "for response buffer\n");
1900		goto out;
1901	}
1902
1903	if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH)
1904		port_num = wc->port_num;
1905	else
1906		port_num = port_priv->port_num;
1907
1908	if (recv->mad.mad.mad_hdr.mgmt_class ==
1909	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1910		enum smi_forward_action retsmi;
1911
1912		if (smi_handle_dr_smp_recv(&recv->mad.smp,
1913					   port_priv->device->node_type,
1914					   port_num,
1915					   port_priv->device->phys_port_cnt) ==
1916					   IB_SMI_DISCARD)
1917			goto out;
1918
1919		retsmi = smi_check_forward_dr_smp(&recv->mad.smp);
1920		if (retsmi == IB_SMI_LOCAL)
1921			goto local;
1922
1923		if (retsmi == IB_SMI_SEND) { /* don't forward */
1924			if (smi_handle_dr_smp_send(&recv->mad.smp,
1925						   port_priv->device->node_type,
1926						   port_num) == IB_SMI_DISCARD)
1927				goto out;
1928
1929			if (smi_check_local_smp(&recv->mad.smp, port_priv->device) == IB_SMI_DISCARD)
1930				goto out;
1931		} else if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH) {
1932			/* forward case for switches */
1933			memcpy(response, recv, sizeof(*response));
1934			response->header.recv_wc.wc = &response->header.wc;
1935			response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1936			response->header.recv_wc.recv_buf.grh = &response->grh;
1937
1938			agent_send_response(&response->mad.mad,
1939					    &response->grh, wc,
1940					    port_priv->device,
1941					    smi_get_fwd_port(&recv->mad.smp),
1942					    qp_info->qp->qp_num);
1943
1944			goto out;
1945		}
1946	}
1947
1948local:
1949	/* Give driver "right of first refusal" on incoming MAD */
1950	if (port_priv->device->process_mad) {
1951		int ret;
1952
1953		ret = port_priv->device->process_mad(port_priv->device, 0,
1954						     port_priv->port_num,
1955						     wc, &recv->grh,
1956						     &recv->mad.mad,
1957						     &response->mad.mad);
1958		if (ret & IB_MAD_RESULT_SUCCESS) {
1959			if (ret & IB_MAD_RESULT_CONSUMED)
1960				goto out;
1961			if (ret & IB_MAD_RESULT_REPLY) {
1962				agent_send_response(&response->mad.mad,
1963						    &recv->grh, wc,
1964						    port_priv->device,
1965						    port_num,
1966						    qp_info->qp->qp_num);
1967				goto out;
1968			}
1969		}
1970	}
1971
1972	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1973	if (mad_agent) {
1974		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1975		/*
1976		 * recv is freed up in error cases in ib_mad_complete_recv
1977		 * or via recv_handler in ib_mad_complete_recv()
1978		 */
1979		recv = NULL;
1980	}
1981
1982out:
1983	/* Post another receive request for this QP */
1984	if (response) {
1985		ib_mad_post_receive_mads(qp_info, response);
1986		if (recv)
1987			kmem_cache_free(ib_mad_cache, recv);
1988	} else
1989		ib_mad_post_receive_mads(qp_info, recv);
1990}
1991
1992static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1993{
1994	struct ib_mad_send_wr_private *mad_send_wr;
1995
1996	if (list_empty(&mad_agent_priv->wait_list)) {
1997		del_timer(&mad_agent_priv->timeout_timer);
1998	} else {
1999		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2000					 struct ib_mad_send_wr_private,
2001					 agent_list);
2002
2003		if (time_after(mad_agent_priv->timeout,
2004			       mad_send_wr->timeout)) {
2005			mad_agent_priv->timeout = mad_send_wr->timeout;
2006			mod_timer(&mad_agent_priv->timeout_timer,
2007				  mad_send_wr->timeout);
2008		}
2009	}
2010}
2011
2012static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2013{
2014	struct ib_mad_agent_private *mad_agent_priv;
2015	struct ib_mad_send_wr_private *temp_mad_send_wr;
2016	struct list_head *list_item;
2017	unsigned long delay;
2018
2019	mad_agent_priv = mad_send_wr->mad_agent_priv;
2020	list_del(&mad_send_wr->agent_list);
2021
2022	delay = mad_send_wr->timeout;
2023	mad_send_wr->timeout += jiffies;
2024
2025	if (delay) {
2026		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2027			temp_mad_send_wr = list_entry(list_item,
2028						struct ib_mad_send_wr_private,
2029						agent_list);
2030			if (time_after(mad_send_wr->timeout,
2031				       temp_mad_send_wr->timeout))
2032				break;
2033		}
2034	} else
2035		list_item = &mad_agent_priv->wait_list;
2036	list_add(&mad_send_wr->agent_list, list_item);
2037
2038	/* Reschedule a work item if we have a shorter timeout */
2039	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2040		mod_timer(&mad_agent_priv->timeout_timer,
2041			  mad_send_wr->timeout);
2042}
2043
2044void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2045			  int timeout_ms)
2046{
2047	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2048	wait_for_response(mad_send_wr);
2049}
2050
2051/*
2052 * Process a send work completion
2053 */
2054void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2055			     struct ib_mad_send_wc *mad_send_wc)
2056{
2057	struct ib_mad_agent_private	*mad_agent_priv;
2058	unsigned long			flags;
2059	int				ret;
2060
2061	mad_agent_priv = mad_send_wr->mad_agent_priv;
2062	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2063	if (mad_agent_priv->agent.rmpp_version) {
2064		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2065		if (ret == IB_RMPP_RESULT_CONSUMED)
2066			goto done;
2067	} else
2068		ret = IB_RMPP_RESULT_UNHANDLED;
2069
2070	if (mad_send_wc->status != IB_WC_SUCCESS &&
2071	    mad_send_wr->status == IB_WC_SUCCESS) {
2072		mad_send_wr->status = mad_send_wc->status;
2073		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2074	}
2075
2076	if (--mad_send_wr->refcount > 0) {
2077		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2078		    mad_send_wr->status == IB_WC_SUCCESS) {
2079			wait_for_response(mad_send_wr);
2080		}
2081		goto done;
2082	}
2083
2084	/* Remove send from MAD agent and notify client of completion */
2085	list_del(&mad_send_wr->agent_list);
2086	adjust_timeout(mad_agent_priv);
2087	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2088
2089	if (mad_send_wr->status != IB_WC_SUCCESS )
2090		mad_send_wc->status = mad_send_wr->status;
2091	if (ret == IB_RMPP_RESULT_INTERNAL)
2092		ib_rmpp_send_handler(mad_send_wc);
2093	else
2094		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2095						   mad_send_wc);
2096
2097	/* Release reference on agent taken when sending */
2098	deref_mad_agent(mad_agent_priv);
2099	return;
2100done:
2101	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2102}
2103
2104static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2105				     struct ib_wc *wc)
2106{
2107	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2108	struct ib_mad_list_head		*mad_list;
2109	struct ib_mad_qp_info		*qp_info;
2110	struct ib_mad_queue		*send_queue;
2111	struct ib_send_wr		*bad_send_wr;
2112	struct ib_mad_send_wc		mad_send_wc;
2113	unsigned long flags;
2114	int ret;
2115
2116	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2117	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2118				   mad_list);
2119	send_queue = mad_list->mad_queue;
2120	qp_info = send_queue->qp_info;
2121
2122retry:
2123	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2124			    mad_send_wr->header_mapping,
2125			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2126	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2127			    mad_send_wr->payload_mapping,
2128			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2129	queued_send_wr = NULL;
2130	spin_lock_irqsave(&send_queue->lock, flags);
2131	list_del(&mad_list->list);
2132
2133	/* Move queued send to the send queue */
2134	if (send_queue->count-- > send_queue->max_active) {
2135		mad_list = container_of(qp_info->overflow_list.next,
2136					struct ib_mad_list_head, list);
2137		queued_send_wr = container_of(mad_list,
2138					struct ib_mad_send_wr_private,
2139					mad_list);
2140		list_move_tail(&mad_list->list, &send_queue->list);
2141	}
2142	spin_unlock_irqrestore(&send_queue->lock, flags);
2143
2144	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2145	mad_send_wc.status = wc->status;
2146	mad_send_wc.vendor_err = wc->vendor_err;
2147	if (atomic_read(&qp_info->snoop_count))
2148		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2149			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2150	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2151
2152	if (queued_send_wr) {
2153		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2154				   &bad_send_wr);
2155		if (ret) {
2156			printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2157			mad_send_wr = queued_send_wr;
2158			wc->status = IB_WC_LOC_QP_OP_ERR;
2159			goto retry;
2160		}
2161	}
2162}
2163
2164static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2165{
2166	struct ib_mad_send_wr_private *mad_send_wr;
2167	struct ib_mad_list_head *mad_list;
2168	unsigned long flags;
2169
2170	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2171	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2172		mad_send_wr = container_of(mad_list,
2173					   struct ib_mad_send_wr_private,
2174					   mad_list);
2175		mad_send_wr->retry = 1;
2176	}
2177	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2178}
2179
2180static void mad_error_handler(struct ib_mad_port_private *port_priv,
2181			      struct ib_wc *wc)
2182{
2183	struct ib_mad_list_head *mad_list;
2184	struct ib_mad_qp_info *qp_info;
2185	struct ib_mad_send_wr_private *mad_send_wr;
2186	int ret;
2187
2188	/* Determine if failure was a send or receive */
2189	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2190	qp_info = mad_list->mad_queue->qp_info;
2191	if (mad_list->mad_queue == &qp_info->recv_queue)
2192		/*
2193		 * Receive errors indicate that the QP has entered the error
2194		 * state - error handling/shutdown code will cleanup
2195		 */
2196		return;
2197
2198	/*
2199	 * Send errors will transition the QP to SQE - move
2200	 * QP to RTS and repost flushed work requests
2201	 */
2202	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2203				   mad_list);
2204	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2205		if (mad_send_wr->retry) {
2206			/* Repost send */
2207			struct ib_send_wr *bad_send_wr;
2208
2209			mad_send_wr->retry = 0;
2210			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2211					&bad_send_wr);
2212			if (ret)
2213				ib_mad_send_done_handler(port_priv, wc);
2214		} else
2215			ib_mad_send_done_handler(port_priv, wc);
2216	} else {
2217		struct ib_qp_attr *attr;
2218
2219		/* Transition QP to RTS and fail offending send */
2220		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2221		if (attr) {
2222			attr->qp_state = IB_QPS_RTS;
2223			attr->cur_qp_state = IB_QPS_SQE;
2224			ret = ib_modify_qp(qp_info->qp, attr,
2225					   IB_QP_STATE | IB_QP_CUR_STATE);
2226			kfree(attr);
2227			if (ret)
2228				printk(KERN_ERR PFX "mad_error_handler - "
2229				       "ib_modify_qp to RTS : %d\n", ret);
2230			else
2231				mark_sends_for_retry(qp_info);
2232		}
2233		ib_mad_send_done_handler(port_priv, wc);
2234	}
2235}
2236
2237/*
2238 * IB MAD completion callback
2239 */
2240static void ib_mad_completion_handler(struct work_struct *work)
2241{
2242	struct ib_mad_port_private *port_priv;
2243	struct ib_wc wc;
2244
2245	port_priv = container_of(work, struct ib_mad_port_private, work);
2246	ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2247
2248	while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2249		if (wc.status == IB_WC_SUCCESS) {
2250			switch (wc.opcode) {
2251			case IB_WC_SEND:
2252				ib_mad_send_done_handler(port_priv, &wc);
2253				break;
2254			case IB_WC_RECV:
2255				ib_mad_recv_done_handler(port_priv, &wc);
2256				break;
2257			default:
2258				BUG_ON(1);
2259				break;
2260			}
2261		} else
2262			mad_error_handler(port_priv, &wc);
2263	}
2264}
2265
2266static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2267{
2268	unsigned long flags;
2269	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2270	struct ib_mad_send_wc mad_send_wc;
2271	struct list_head cancel_list;
2272
2273	INIT_LIST_HEAD(&cancel_list);
2274
2275	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2276	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2277				 &mad_agent_priv->send_list, agent_list) {
2278		if (mad_send_wr->status == IB_WC_SUCCESS) {
2279			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2280			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2281		}
2282	}
2283
2284	/* Empty wait list to prevent receives from finding a request */
2285	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2286	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2287
2288	/* Report all cancelled requests */
2289	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2290	mad_send_wc.vendor_err = 0;
2291
2292	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2293				 &cancel_list, agent_list) {
2294		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2295		list_del(&mad_send_wr->agent_list);
2296		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2297						   &mad_send_wc);
2298		atomic_dec(&mad_agent_priv->refcount);
2299	}
2300}
2301
2302static struct ib_mad_send_wr_private*
2303find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2304	     struct ib_mad_send_buf *send_buf)
2305{
2306	struct ib_mad_send_wr_private *mad_send_wr;
2307
2308	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2309			    agent_list) {
2310		if (&mad_send_wr->send_buf == send_buf)
2311			return mad_send_wr;
2312	}
2313
2314	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2315			    agent_list) {
2316		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2317		    &mad_send_wr->send_buf == send_buf)
2318			return mad_send_wr;
2319	}
2320	return NULL;
2321}
2322
2323int ib_modify_mad(struct ib_mad_agent *mad_agent,
2324		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2325{
2326	struct ib_mad_agent_private *mad_agent_priv;
2327	struct ib_mad_send_wr_private *mad_send_wr;
2328	unsigned long flags;
2329	int active;
2330
2331	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2332				      agent);
2333	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2334	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2335	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2336		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2337		return -EINVAL;
2338	}
2339
2340	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2341	if (!timeout_ms) {
2342		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2343		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2344	}
2345
2346	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2347	if (active)
2348		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2349	else
2350		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2351
2352	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2353	return 0;
2354}
2355EXPORT_SYMBOL(ib_modify_mad);
2356
2357void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2358		   struct ib_mad_send_buf *send_buf)
2359{
2360	ib_modify_mad(mad_agent, send_buf, 0);
2361}
2362EXPORT_SYMBOL(ib_cancel_mad);
2363
2364static void local_completions(struct work_struct *work)
2365{
2366	struct ib_mad_agent_private *mad_agent_priv;
2367	struct ib_mad_local_private *local;
2368	struct ib_mad_agent_private *recv_mad_agent;
2369	unsigned long flags;
2370	int free_mad;
2371	struct ib_wc wc;
2372	struct ib_mad_send_wc mad_send_wc;
2373
2374	mad_agent_priv =
2375		container_of(work, struct ib_mad_agent_private, local_work);
2376
2377	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2378	while (!list_empty(&mad_agent_priv->local_list)) {
2379		local = list_entry(mad_agent_priv->local_list.next,
2380				   struct ib_mad_local_private,
2381				   completion_list);
2382		list_del(&local->completion_list);
2383		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2384		free_mad = 0;
2385		if (local->mad_priv) {
2386			recv_mad_agent = local->recv_mad_agent;
2387			if (!recv_mad_agent) {
2388				printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2389				free_mad = 1;
2390				goto local_send_completion;
2391			}
2392
2393			/*
2394			 * Defined behavior is to complete response
2395			 * before request
2396			 */
2397			build_smp_wc(recv_mad_agent->agent.qp,
2398				     (unsigned long) local->mad_send_wr,
2399				     be16_to_cpu(IB_LID_PERMISSIVE),
2400				     0, recv_mad_agent->agent.port_num, &wc);
2401
2402			local->mad_priv->header.recv_wc.wc = &wc;
2403			local->mad_priv->header.recv_wc.mad_len =
2404						sizeof(struct ib_mad);
2405			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2406			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2407				 &local->mad_priv->header.recv_wc.rmpp_list);
2408			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2409			local->mad_priv->header.recv_wc.recv_buf.mad =
2410						&local->mad_priv->mad.mad;
2411			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2412				snoop_recv(recv_mad_agent->qp_info,
2413					  &local->mad_priv->header.recv_wc,
2414					   IB_MAD_SNOOP_RECVS);
2415			recv_mad_agent->agent.recv_handler(
2416						&recv_mad_agent->agent,
2417						&local->mad_priv->header.recv_wc);
2418			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2419			atomic_dec(&recv_mad_agent->refcount);
2420			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2421		}
2422
2423local_send_completion:
2424		/* Complete send */
2425		mad_send_wc.status = IB_WC_SUCCESS;
2426		mad_send_wc.vendor_err = 0;
2427		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2428		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2429			snoop_send(mad_agent_priv->qp_info,
2430				   &local->mad_send_wr->send_buf,
2431				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2432		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2433						   &mad_send_wc);
2434
2435		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2436		atomic_dec(&mad_agent_priv->refcount);
2437		if (free_mad)
2438			kmem_cache_free(ib_mad_cache, local->mad_priv);
2439		kfree(local);
2440	}
2441	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2442}
2443
2444static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2445{
2446	int ret;
2447
2448	if (!mad_send_wr->retries_left)
2449		return -ETIMEDOUT;
2450
2451	mad_send_wr->retries_left--;
2452	mad_send_wr->send_buf.retries++;
2453
2454	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2455
2456	if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2457		ret = ib_retry_rmpp(mad_send_wr);
2458		switch (ret) {
2459		case IB_RMPP_RESULT_UNHANDLED:
2460			ret = ib_send_mad(mad_send_wr);
2461			break;
2462		case IB_RMPP_RESULT_CONSUMED:
2463			ret = 0;
2464			break;
2465		default:
2466			ret = -ECOMM;
2467			break;
2468		}
2469	} else
2470		ret = ib_send_mad(mad_send_wr);
2471
2472	if (!ret) {
2473		mad_send_wr->refcount++;
2474		list_add_tail(&mad_send_wr->agent_list,
2475			      &mad_send_wr->mad_agent_priv->send_list);
2476	}
2477	return ret;
2478}
2479
2480static void timeout_sends(struct work_struct *work)
2481{
2482	struct ib_mad_agent_private *mad_agent_priv;
2483	struct ib_mad_send_wr_private *mad_send_wr;
2484	struct ib_mad_send_wc mad_send_wc;
2485	unsigned long flags;
2486
2487	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2488				      timeout_work);
2489	mad_send_wc.vendor_err = 0;
2490
2491	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2492	while (!list_empty(&mad_agent_priv->wait_list)) {
2493		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2494					 struct ib_mad_send_wr_private,
2495					 agent_list);
2496
2497		if (time_after(mad_send_wr->timeout, jiffies)) {
2498			mod_timer(&mad_agent_priv->timeout_timer,
2499				  mad_send_wr->timeout);
2500			break;
2501		}
2502
2503		list_del(&mad_send_wr->agent_list);
2504		if (mad_send_wr->status == IB_WC_SUCCESS &&
2505		    !retry_send(mad_send_wr))
2506			continue;
2507
2508		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2509
2510		if (mad_send_wr->status == IB_WC_SUCCESS)
2511			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2512		else
2513			mad_send_wc.status = mad_send_wr->status;
2514		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2515		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2516						   &mad_send_wc);
2517
2518		atomic_dec(&mad_agent_priv->refcount);
2519		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2520	}
2521	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2522}
2523
2524static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2525{
2526	struct ib_mad_port_private *port_priv = cq->cq_context;
2527	unsigned long flags;
2528
2529	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2530	if (!list_empty(&port_priv->port_list))
2531		queue_work(port_priv->wq, &port_priv->work);
2532	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2533}
2534
2535/*
2536 * Allocate receive MADs and post receive WRs for them
2537 */
2538static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2539				    struct ib_mad_private *mad)
2540{
2541	unsigned long flags;
2542	int post, ret;
2543	struct ib_mad_private *mad_priv;
2544	struct ib_sge sg_list;
2545	struct ib_recv_wr recv_wr, *bad_recv_wr;
2546	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2547
2548	/* Initialize common scatter list fields */
2549	sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2550	sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2551
2552	/* Initialize common receive WR fields */
2553	recv_wr.next = NULL;
2554	recv_wr.sg_list = &sg_list;
2555	recv_wr.num_sge = 1;
2556
2557	do {
2558		/* Allocate and map receive buffer */
2559		if (mad) {
2560			mad_priv = mad;
2561			mad = NULL;
2562		} else {
2563			mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2564			if (!mad_priv) {
2565				printk(KERN_ERR PFX "No memory for receive buffer\n");
2566				ret = -ENOMEM;
2567				break;
2568			}
2569		}
2570		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2571						 &mad_priv->grh,
2572						 sizeof *mad_priv -
2573						   sizeof mad_priv->header,
2574						 DMA_FROM_DEVICE);
2575		mad_priv->header.mapping = sg_list.addr;
2576		recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2577		mad_priv->header.mad_list.mad_queue = recv_queue;
2578
2579		/* Post receive WR */
2580		spin_lock_irqsave(&recv_queue->lock, flags);
2581		post = (++recv_queue->count < recv_queue->max_active);
2582		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2583		spin_unlock_irqrestore(&recv_queue->lock, flags);
2584		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2585		if (ret) {
2586			spin_lock_irqsave(&recv_queue->lock, flags);
2587			list_del(&mad_priv->header.mad_list.list);
2588			recv_queue->count--;
2589			spin_unlock_irqrestore(&recv_queue->lock, flags);
2590			ib_dma_unmap_single(qp_info->port_priv->device,
2591					    mad_priv->header.mapping,
2592					    sizeof *mad_priv -
2593					      sizeof mad_priv->header,
2594					    DMA_FROM_DEVICE);
2595			kmem_cache_free(ib_mad_cache, mad_priv);
2596			printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2597			break;
2598		}
2599	} while (post);
2600
2601	return ret;
2602}
2603
2604/*
2605 * Return all the posted receive MADs
2606 */
2607static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2608{
2609	struct ib_mad_private_header *mad_priv_hdr;
2610	struct ib_mad_private *recv;
2611	struct ib_mad_list_head *mad_list;
2612
2613	if (!qp_info->qp)
2614		return;
2615
2616	while (!list_empty(&qp_info->recv_queue.list)) {
2617
2618		mad_list = list_entry(qp_info->recv_queue.list.next,
2619				      struct ib_mad_list_head, list);
2620		mad_priv_hdr = container_of(mad_list,
2621					    struct ib_mad_private_header,
2622					    mad_list);
2623		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2624				    header);
2625
2626		/* Remove from posted receive MAD list */
2627		list_del(&mad_list->list);
2628
2629		ib_dma_unmap_single(qp_info->port_priv->device,
2630				    recv->header.mapping,
2631				    sizeof(struct ib_mad_private) -
2632				      sizeof(struct ib_mad_private_header),
2633				    DMA_FROM_DEVICE);
2634		kmem_cache_free(ib_mad_cache, recv);
2635	}
2636
2637	qp_info->recv_queue.count = 0;
2638}
2639
2640/*
2641 * Start the port
2642 */
2643static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2644{
2645	int ret, i;
2646	struct ib_qp_attr *attr;
2647	struct ib_qp *qp;
2648
2649	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2650	if (!attr) {
2651		printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2652		return -ENOMEM;
2653	}
2654
2655	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2656		qp = port_priv->qp_info[i].qp;
2657		if (!qp)
2658			continue;
2659
2660		/*
2661		 * PKey index for QP1 is irrelevant but
2662		 * one is needed for the Reset to Init transition
2663		 */
2664		attr->qp_state = IB_QPS_INIT;
2665		attr->pkey_index = 0;
2666		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2667		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2668					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2669		if (ret) {
2670			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2671			       "INIT: %d\n", i, ret);
2672			goto out;
2673		}
2674
2675		attr->qp_state = IB_QPS_RTR;
2676		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2677		if (ret) {
2678			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2679			       "RTR: %d\n", i, ret);
2680			goto out;
2681		}
2682
2683		attr->qp_state = IB_QPS_RTS;
2684		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2685		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2686		if (ret) {
2687			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2688			       "RTS: %d\n", i, ret);
2689			goto out;
2690		}
2691	}
2692
2693	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2694	if (ret) {
2695		printk(KERN_ERR PFX "Failed to request completion "
2696		       "notification: %d\n", ret);
2697		goto out;
2698	}
2699
2700	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2701		if (!port_priv->qp_info[i].qp)
2702			continue;
2703
2704		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2705		if (ret) {
2706			printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2707			goto out;
2708		}
2709	}
2710out:
2711	kfree(attr);
2712	return ret;
2713}
2714
2715static void qp_event_handler(struct ib_event *event, void *qp_context)
2716{
2717	struct ib_mad_qp_info	*qp_info = qp_context;
2718
2719	/* It's worse than that! He's dead, Jim! */
2720	printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2721		event->event, qp_info->qp->qp_num);
2722}
2723
2724static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2725			   struct ib_mad_queue *mad_queue)
2726{
2727	mad_queue->qp_info = qp_info;
2728	mad_queue->count = 0;
2729	spin_lock_init(&mad_queue->lock);
2730	INIT_LIST_HEAD(&mad_queue->list);
2731}
2732
2733static void init_mad_qp(struct ib_mad_port_private *port_priv,
2734			struct ib_mad_qp_info *qp_info)
2735{
2736	qp_info->port_priv = port_priv;
2737	init_mad_queue(qp_info, &qp_info->send_queue);
2738	init_mad_queue(qp_info, &qp_info->recv_queue);
2739	INIT_LIST_HEAD(&qp_info->overflow_list);
2740	spin_lock_init(&qp_info->snoop_lock);
2741	qp_info->snoop_table = NULL;
2742	qp_info->snoop_table_size = 0;
2743	atomic_set(&qp_info->snoop_count, 0);
2744}
2745
2746static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2747			 enum ib_qp_type qp_type)
2748{
2749	struct ib_qp_init_attr	qp_init_attr;
2750	int ret;
2751
2752	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2753	qp_init_attr.send_cq = qp_info->port_priv->cq;
2754	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2755	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2756	qp_init_attr.cap.max_send_wr = mad_sendq_size;
2757	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
2758	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2759	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2760	qp_init_attr.qp_type = qp_type;
2761	qp_init_attr.port_num = qp_info->port_priv->port_num;
2762	qp_init_attr.qp_context = qp_info;
2763	qp_init_attr.event_handler = qp_event_handler;
2764	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2765	if (IS_ERR(qp_info->qp)) {
2766		printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2767		       get_spl_qp_index(qp_type));
2768		ret = PTR_ERR(qp_info->qp);
2769		goto error;
2770	}
2771	/* Use minimum queue sizes unless the CQ is resized */
2772	qp_info->send_queue.max_active = mad_sendq_size;
2773	qp_info->recv_queue.max_active = mad_recvq_size;
2774	return 0;
2775
2776error:
2777	return ret;
2778}
2779
2780static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2781{
2782	if (!qp_info->qp)
2783		return;
2784
2785	ib_destroy_qp(qp_info->qp);
2786	kfree(qp_info->snoop_table);
2787}
2788
2789/*
2790 * Open the port
2791 * Create the QP, PD, MR, and CQ if needed
2792 */
2793static int ib_mad_port_open(struct ib_device *device,
2794			    int port_num)
2795{
2796	int ret, cq_size;
2797	struct ib_mad_port_private *port_priv;
2798	unsigned long flags;
2799	char name[sizeof "ib_mad123"];
2800	int has_smi;
2801
2802	/* Create new device info */
2803	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2804	if (!port_priv) {
2805		printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2806		return -ENOMEM;
2807	}
2808
2809	port_priv->device = device;
2810	port_priv->port_num = port_num;
2811	spin_lock_init(&port_priv->reg_lock);
2812	INIT_LIST_HEAD(&port_priv->agent_list);
2813	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2814	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2815
2816	cq_size = mad_sendq_size + mad_recvq_size;
2817	has_smi = rdma_port_get_link_layer(device, port_num) == IB_LINK_LAYER_INFINIBAND;
2818	if (has_smi)
2819		cq_size *= 2;
2820
2821	port_priv->cq = ib_create_cq(port_priv->device,
2822				     ib_mad_thread_completion_handler,
2823				     NULL, port_priv, cq_size, 0);
2824	if (IS_ERR(port_priv->cq)) {
2825		printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2826		ret = PTR_ERR(port_priv->cq);
2827		goto error3;
2828	}
2829
2830	port_priv->pd = ib_alloc_pd(device);
2831	if (IS_ERR(port_priv->pd)) {
2832		printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2833		ret = PTR_ERR(port_priv->pd);
2834		goto error4;
2835	}
2836
2837	port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2838	if (IS_ERR(port_priv->mr)) {
2839		printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2840		ret = PTR_ERR(port_priv->mr);
2841		goto error5;
2842	}
2843
2844	if (has_smi) {
2845		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2846		if (ret)
2847			goto error6;
2848	}
2849	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2850	if (ret)
2851		goto error7;
2852
2853	snprintf(name, sizeof name, "ib_mad%d", port_num);
2854	port_priv->wq = create_singlethread_workqueue(name);
2855	if (!port_priv->wq) {
2856		ret = -ENOMEM;
2857		goto error8;
2858	}
2859	INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2860
2861	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2862	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2863	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2864
2865	ret = ib_mad_port_start(port_priv);
2866	if (ret) {
2867		printk(KERN_ERR PFX "Couldn't start port\n");
2868		goto error9;
2869	}
2870
2871	return 0;
2872
2873error9:
2874	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2875	list_del_init(&port_priv->port_list);
2876	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2877
2878	destroy_workqueue(port_priv->wq);
2879error8:
2880	destroy_mad_qp(&port_priv->qp_info[1]);
2881error7:
2882	destroy_mad_qp(&port_priv->qp_info[0]);
2883error6:
2884	ib_dereg_mr(port_priv->mr);
2885error5:
2886	ib_dealloc_pd(port_priv->pd);
2887error4:
2888	ib_destroy_cq(port_priv->cq);
2889	cleanup_recv_queue(&port_priv->qp_info[1]);
2890	cleanup_recv_queue(&port_priv->qp_info[0]);
2891error3:
2892	kfree(port_priv);
2893
2894	return ret;
2895}
2896
2897/*
2898 * Close the port
2899 * If there are no classes using the port, free the port
2900 * resources (CQ, MR, PD, QP) and remove the port's info structure
2901 */
2902static int ib_mad_port_close(struct ib_device *device, int port_num)
2903{
2904	struct ib_mad_port_private *port_priv;
2905	unsigned long flags;
2906
2907	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2908	port_priv = __ib_get_mad_port(device, port_num);
2909	if (port_priv == NULL) {
2910		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2911		printk(KERN_ERR PFX "Port %d not found\n", port_num);
2912		return -ENODEV;
2913	}
2914	list_del_init(&port_priv->port_list);
2915	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2916
2917	destroy_workqueue(port_priv->wq);
2918	destroy_mad_qp(&port_priv->qp_info[1]);
2919	destroy_mad_qp(&port_priv->qp_info[0]);
2920	ib_dereg_mr(port_priv->mr);
2921	ib_dealloc_pd(port_priv->pd);
2922	ib_destroy_cq(port_priv->cq);
2923	cleanup_recv_queue(&port_priv->qp_info[1]);
2924	cleanup_recv_queue(&port_priv->qp_info[0]);
2925	/* XXX: Handle deallocation of MAD registration tables */
2926
2927	kfree(port_priv);
2928
2929	return 0;
2930}
2931
2932static void ib_mad_init_device(struct ib_device *device)
2933{
2934	int start, end, i;
2935
2936	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2937		return;
2938
2939	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2940		start = 0;
2941		end   = 0;
2942	} else {
2943		start = 1;
2944		end   = device->phys_port_cnt;
2945	}
2946
2947	for (i = start; i <= end; i++) {
2948		if (ib_mad_port_open(device, i)) {
2949			printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2950			       device->name, i);
2951			goto error;
2952		}
2953		if (ib_agent_port_open(device, i)) {
2954			printk(KERN_ERR PFX "Couldn't open %s port %d "
2955			       "for agents\n",
2956			       device->name, i);
2957			goto error_agent;
2958		}
2959	}
2960	return;
2961
2962error_agent:
2963	if (ib_mad_port_close(device, i))
2964		printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2965		       device->name, i);
2966
2967error:
2968	i--;
2969
2970	while (i >= start) {
2971		if (ib_agent_port_close(device, i))
2972			printk(KERN_ERR PFX "Couldn't close %s port %d "
2973			       "for agents\n",
2974			       device->name, i);
2975		if (ib_mad_port_close(device, i))
2976			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2977			       device->name, i);
2978		i--;
2979	}
2980}
2981
2982static void ib_mad_remove_device(struct ib_device *device)
2983{
2984	int i, num_ports, cur_port;
2985
2986	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2987		num_ports = 1;
2988		cur_port = 0;
2989	} else {
2990		num_ports = device->phys_port_cnt;
2991		cur_port = 1;
2992	}
2993	for (i = 0; i < num_ports; i++, cur_port++) {
2994		if (ib_agent_port_close(device, cur_port))
2995			printk(KERN_ERR PFX "Couldn't close %s port %d "
2996			       "for agents\n",
2997			       device->name, cur_port);
2998		if (ib_mad_port_close(device, cur_port))
2999			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
3000			       device->name, cur_port);
3001	}
3002}
3003
3004static struct ib_client mad_client = {
3005	.name   = "mad",
3006	.add = ib_mad_init_device,
3007	.remove = ib_mad_remove_device
3008};
3009
3010static int __init ib_mad_init_module(void)
3011{
3012	int ret;
3013
3014	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3015	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3016
3017	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3018	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3019
3020	spin_lock_init(&ib_mad_port_list_lock);
3021
3022	ib_mad_cache = kmem_cache_create("ib_mad",
3023					 sizeof(struct ib_mad_private),
3024					 0,
3025					 SLAB_HWCACHE_ALIGN,
3026					 NULL);
3027	if (!ib_mad_cache) {
3028		printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
3029		ret = -ENOMEM;
3030		goto error1;
3031	}
3032
3033	INIT_LIST_HEAD(&ib_mad_port_list);
3034
3035	if (ib_register_client(&mad_client)) {
3036		printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
3037		ret = -EINVAL;
3038		goto error2;
3039	}
3040
3041	return 0;
3042
3043error2:
3044	kmem_cache_destroy(ib_mad_cache);
3045error1:
3046	return ret;
3047}
3048
3049static void __exit ib_mad_cleanup_module(void)
3050{
3051	ib_unregister_client(&mad_client);
3052	kmem_cache_destroy(ib_mad_cache);
3053}
3054
3055module_init(ib_mad_init_module);
3056module_exit(ib_mad_cleanup_module);
3057
3058