Merge git://git.infradead.org/mtd-2.6
[linux-2.6] / drivers / infiniband / core / mad.c
1 /*
2  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
3  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  *
34  * $Id: mad.c 5596 2006-03-03 01:00:07Z sean.hefty $
35  */
36 #include <linux/dma-mapping.h>
37
38 #include "mad_priv.h"
39 #include "mad_rmpp.h"
40 #include "smi.h"
41 #include "agent.h"
42
43 MODULE_LICENSE("Dual BSD/GPL");
44 MODULE_DESCRIPTION("kernel IB MAD API");
45 MODULE_AUTHOR("Hal Rosenstock");
46 MODULE_AUTHOR("Sean Hefty");
47
48
49 kmem_cache_t *ib_mad_cache;
50
51 static struct list_head ib_mad_port_list;
52 static u32 ib_mad_client_id = 0;
53
54 /* Port list lock */
55 static spinlock_t ib_mad_port_list_lock;
56
57
58 /* Forward declarations */
59 static int method_in_use(struct ib_mad_mgmt_method_table **method,
60                          struct ib_mad_reg_req *mad_reg_req);
61 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
62 static struct ib_mad_agent_private *find_mad_agent(
63                                         struct ib_mad_port_private *port_priv,
64                                         struct ib_mad *mad);
65 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
66                                     struct ib_mad_private *mad);
67 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
68 static void timeout_sends(void *data);
69 static void local_completions(void *data);
70 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
71                               struct ib_mad_agent_private *agent_priv,
72                               u8 mgmt_class);
73 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
74                            struct ib_mad_agent_private *agent_priv);
75
76 /*
77  * Returns a ib_mad_port_private structure or NULL for a device/port
78  * Assumes ib_mad_port_list_lock is being held
79  */
80 static inline struct ib_mad_port_private *
81 __ib_get_mad_port(struct ib_device *device, int port_num)
82 {
83         struct ib_mad_port_private *entry;
84
85         list_for_each_entry(entry, &ib_mad_port_list, port_list) {
86                 if (entry->device == device && entry->port_num == port_num)
87                         return entry;
88         }
89         return NULL;
90 }
91
92 /*
93  * Wrapper function to return a ib_mad_port_private structure or NULL
94  * for a device/port
95  */
96 static inline struct ib_mad_port_private *
97 ib_get_mad_port(struct ib_device *device, int port_num)
98 {
99         struct ib_mad_port_private *entry;
100         unsigned long flags;
101
102         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
103         entry = __ib_get_mad_port(device, port_num);
104         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
105
106         return entry;
107 }
108
109 static inline u8 convert_mgmt_class(u8 mgmt_class)
110 {
111         /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
112         return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
113                 0 : mgmt_class;
114 }
115
116 static int get_spl_qp_index(enum ib_qp_type qp_type)
117 {
118         switch (qp_type)
119         {
120         case IB_QPT_SMI:
121                 return 0;
122         case IB_QPT_GSI:
123                 return 1;
124         default:
125                 return -1;
126         }
127 }
128
129 static int vendor_class_index(u8 mgmt_class)
130 {
131         return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
132 }
133
134 static int is_vendor_class(u8 mgmt_class)
135 {
136         if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
137             (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
138                 return 0;
139         return 1;
140 }
141
142 static int is_vendor_oui(char *oui)
143 {
144         if (oui[0] || oui[1] || oui[2])
145                 return 1;
146         return 0;
147 }
148
149 static int is_vendor_method_in_use(
150                 struct ib_mad_mgmt_vendor_class *vendor_class,
151                 struct ib_mad_reg_req *mad_reg_req)
152 {
153         struct ib_mad_mgmt_method_table *method;
154         int i;
155
156         for (i = 0; i < MAX_MGMT_OUI; i++) {
157                 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
158                         method = vendor_class->method_table[i];
159                         if (method) {
160                                 if (method_in_use(&method, mad_reg_req))
161                                         return 1;
162                                 else
163                                         break;
164                         }
165                 }
166         }
167         return 0;
168 }
169
170 /*
171  * ib_register_mad_agent - Register to send/receive MADs
172  */
173 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
174                                            u8 port_num,
175                                            enum ib_qp_type qp_type,
176                                            struct ib_mad_reg_req *mad_reg_req,
177                                            u8 rmpp_version,
178                                            ib_mad_send_handler send_handler,
179                                            ib_mad_recv_handler recv_handler,
180                                            void *context)
181 {
182         struct ib_mad_port_private *port_priv;
183         struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
184         struct ib_mad_agent_private *mad_agent_priv;
185         struct ib_mad_reg_req *reg_req = NULL;
186         struct ib_mad_mgmt_class_table *class;
187         struct ib_mad_mgmt_vendor_class_table *vendor;
188         struct ib_mad_mgmt_vendor_class *vendor_class;
189         struct ib_mad_mgmt_method_table *method;
190         int ret2, qpn;
191         unsigned long flags;
192         u8 mgmt_class, vclass;
193
194         /* Validate parameters */
195         qpn = get_spl_qp_index(qp_type);
196         if (qpn == -1)
197                 goto error1;
198
199         if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
200                 goto error1;
201
202         /* Validate MAD registration request if supplied */
203         if (mad_reg_req) {
204                 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
205                         goto error1;
206                 if (!recv_handler)
207                         goto error1;
208                 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
209                         /*
210                          * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
211                          * one in this range currently allowed
212                          */
213                         if (mad_reg_req->mgmt_class !=
214                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
215                                 goto error1;
216                 } else if (mad_reg_req->mgmt_class == 0) {
217                         /*
218                          * Class 0 is reserved in IBA and is used for
219                          * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
220                          */
221                         goto error1;
222                 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
223                         /*
224                          * If class is in "new" vendor range,
225                          * ensure supplied OUI is not zero
226                          */
227                         if (!is_vendor_oui(mad_reg_req->oui))
228                                 goto error1;
229                 }
230                 /* Make sure class supplied is consistent with RMPP */
231                 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
232                         if (rmpp_version)
233                                 goto error1;
234                 }
235                 /* Make sure class supplied is consistent with QP type */
236                 if (qp_type == IB_QPT_SMI) {
237                         if ((mad_reg_req->mgmt_class !=
238                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
239                             (mad_reg_req->mgmt_class !=
240                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
241                                 goto error1;
242                 } else {
243                         if ((mad_reg_req->mgmt_class ==
244                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
245                             (mad_reg_req->mgmt_class ==
246                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
247                                 goto error1;
248                 }
249         } else {
250                 /* No registration request supplied */
251                 if (!send_handler)
252                         goto error1;
253         }
254
255         /* Validate device and port */
256         port_priv = ib_get_mad_port(device, port_num);
257         if (!port_priv) {
258                 ret = ERR_PTR(-ENODEV);
259                 goto error1;
260         }
261
262         /* Allocate structures */
263         mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
264         if (!mad_agent_priv) {
265                 ret = ERR_PTR(-ENOMEM);
266                 goto error1;
267         }
268
269         mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
270                                                  IB_ACCESS_LOCAL_WRITE);
271         if (IS_ERR(mad_agent_priv->agent.mr)) {
272                 ret = ERR_PTR(-ENOMEM);
273                 goto error2;
274         }
275
276         if (mad_reg_req) {
277                 reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
278                 if (!reg_req) {
279                         ret = ERR_PTR(-ENOMEM);
280                         goto error3;
281                 }
282                 /* Make a copy of the MAD registration request */
283                 memcpy(reg_req, mad_reg_req, sizeof *reg_req);
284         }
285
286         /* Now, fill in the various structures */
287         mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
288         mad_agent_priv->reg_req = reg_req;
289         mad_agent_priv->agent.rmpp_version = rmpp_version;
290         mad_agent_priv->agent.device = device;
291         mad_agent_priv->agent.recv_handler = recv_handler;
292         mad_agent_priv->agent.send_handler = send_handler;
293         mad_agent_priv->agent.context = context;
294         mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
295         mad_agent_priv->agent.port_num = port_num;
296
297         spin_lock_irqsave(&port_priv->reg_lock, flags);
298         mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
299
300         /*
301          * Make sure MAD registration (if supplied)
302          * is non overlapping with any existing ones
303          */
304         if (mad_reg_req) {
305                 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
306                 if (!is_vendor_class(mgmt_class)) {
307                         class = port_priv->version[mad_reg_req->
308                                                    mgmt_class_version].class;
309                         if (class) {
310                                 method = class->method_table[mgmt_class];
311                                 if (method) {
312                                         if (method_in_use(&method,
313                                                            mad_reg_req))
314                                                 goto error4;
315                                 }
316                         }
317                         ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
318                                                   mgmt_class);
319                 } else {
320                         /* "New" vendor class range */
321                         vendor = port_priv->version[mad_reg_req->
322                                                     mgmt_class_version].vendor;
323                         if (vendor) {
324                                 vclass = vendor_class_index(mgmt_class);
325                                 vendor_class = vendor->vendor_class[vclass];
326                                 if (vendor_class) {
327                                         if (is_vendor_method_in_use(
328                                                         vendor_class,
329                                                         mad_reg_req))
330                                                 goto error4;
331                                 }
332                         }
333                         ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
334                 }
335                 if (ret2) {
336                         ret = ERR_PTR(ret2);
337                         goto error4;
338                 }
339         }
340
341         /* Add mad agent into port's agent list */
342         list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
343         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
344
345         spin_lock_init(&mad_agent_priv->lock);
346         INIT_LIST_HEAD(&mad_agent_priv->send_list);
347         INIT_LIST_HEAD(&mad_agent_priv->wait_list);
348         INIT_LIST_HEAD(&mad_agent_priv->done_list);
349         INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
350         INIT_WORK(&mad_agent_priv->timed_work, timeout_sends, mad_agent_priv);
351         INIT_LIST_HEAD(&mad_agent_priv->local_list);
352         INIT_WORK(&mad_agent_priv->local_work, local_completions,
353                    mad_agent_priv);
354         atomic_set(&mad_agent_priv->refcount, 1);
355         init_completion(&mad_agent_priv->comp);
356
357         return &mad_agent_priv->agent;
358
359 error4:
360         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
361         kfree(reg_req);
362 error3:
363         ib_dereg_mr(mad_agent_priv->agent.mr);
364 error2:
365         kfree(mad_agent_priv);
366 error1:
367         return ret;
368 }
369 EXPORT_SYMBOL(ib_register_mad_agent);
370
371 static inline int is_snooping_sends(int mad_snoop_flags)
372 {
373         return (mad_snoop_flags &
374                 (/*IB_MAD_SNOOP_POSTED_SENDS |
375                  IB_MAD_SNOOP_RMPP_SENDS |*/
376                  IB_MAD_SNOOP_SEND_COMPLETIONS /*|
377                  IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
378 }
379
380 static inline int is_snooping_recvs(int mad_snoop_flags)
381 {
382         return (mad_snoop_flags &
383                 (IB_MAD_SNOOP_RECVS /*|
384                  IB_MAD_SNOOP_RMPP_RECVS*/));
385 }
386
387 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
388                                 struct ib_mad_snoop_private *mad_snoop_priv)
389 {
390         struct ib_mad_snoop_private **new_snoop_table;
391         unsigned long flags;
392         int i;
393
394         spin_lock_irqsave(&qp_info->snoop_lock, flags);
395         /* Check for empty slot in array. */
396         for (i = 0; i < qp_info->snoop_table_size; i++)
397                 if (!qp_info->snoop_table[i])
398                         break;
399
400         if (i == qp_info->snoop_table_size) {
401                 /* Grow table. */
402                 new_snoop_table = kmalloc(sizeof mad_snoop_priv *
403                                           qp_info->snoop_table_size + 1,
404                                           GFP_ATOMIC);
405                 if (!new_snoop_table) {
406                         i = -ENOMEM;
407                         goto out;
408                 }
409                 if (qp_info->snoop_table) {
410                         memcpy(new_snoop_table, qp_info->snoop_table,
411                                sizeof mad_snoop_priv *
412                                qp_info->snoop_table_size);
413                         kfree(qp_info->snoop_table);
414                 }
415                 qp_info->snoop_table = new_snoop_table;
416                 qp_info->snoop_table_size++;
417         }
418         qp_info->snoop_table[i] = mad_snoop_priv;
419         atomic_inc(&qp_info->snoop_count);
420 out:
421         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
422         return i;
423 }
424
425 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
426                                            u8 port_num,
427                                            enum ib_qp_type qp_type,
428                                            int mad_snoop_flags,
429                                            ib_mad_snoop_handler snoop_handler,
430                                            ib_mad_recv_handler recv_handler,
431                                            void *context)
432 {
433         struct ib_mad_port_private *port_priv;
434         struct ib_mad_agent *ret;
435         struct ib_mad_snoop_private *mad_snoop_priv;
436         int qpn;
437
438         /* Validate parameters */
439         if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
440             (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
441                 ret = ERR_PTR(-EINVAL);
442                 goto error1;
443         }
444         qpn = get_spl_qp_index(qp_type);
445         if (qpn == -1) {
446                 ret = ERR_PTR(-EINVAL);
447                 goto error1;
448         }
449         port_priv = ib_get_mad_port(device, port_num);
450         if (!port_priv) {
451                 ret = ERR_PTR(-ENODEV);
452                 goto error1;
453         }
454         /* Allocate structures */
455         mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
456         if (!mad_snoop_priv) {
457                 ret = ERR_PTR(-ENOMEM);
458                 goto error1;
459         }
460
461         /* Now, fill in the various structures */
462         mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
463         mad_snoop_priv->agent.device = device;
464         mad_snoop_priv->agent.recv_handler = recv_handler;
465         mad_snoop_priv->agent.snoop_handler = snoop_handler;
466         mad_snoop_priv->agent.context = context;
467         mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
468         mad_snoop_priv->agent.port_num = port_num;
469         mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
470         init_completion(&mad_snoop_priv->comp);
471         mad_snoop_priv->snoop_index = register_snoop_agent(
472                                                 &port_priv->qp_info[qpn],
473                                                 mad_snoop_priv);
474         if (mad_snoop_priv->snoop_index < 0) {
475                 ret = ERR_PTR(mad_snoop_priv->snoop_index);
476                 goto error2;
477         }
478
479         atomic_set(&mad_snoop_priv->refcount, 1);
480         return &mad_snoop_priv->agent;
481
482 error2:
483         kfree(mad_snoop_priv);
484 error1:
485         return ret;
486 }
487 EXPORT_SYMBOL(ib_register_mad_snoop);
488
489 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
490 {
491         if (atomic_dec_and_test(&mad_agent_priv->refcount))
492                 complete(&mad_agent_priv->comp);
493 }
494
495 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
496 {
497         if (atomic_dec_and_test(&mad_snoop_priv->refcount))
498                 complete(&mad_snoop_priv->comp);
499 }
500
501 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
502 {
503         struct ib_mad_port_private *port_priv;
504         unsigned long flags;
505
506         /* Note that we could still be handling received MADs */
507
508         /*
509          * Canceling all sends results in dropping received response
510          * MADs, preventing us from queuing additional work
511          */
512         cancel_mads(mad_agent_priv);
513         port_priv = mad_agent_priv->qp_info->port_priv;
514         cancel_delayed_work(&mad_agent_priv->timed_work);
515
516         spin_lock_irqsave(&port_priv->reg_lock, flags);
517         remove_mad_reg_req(mad_agent_priv);
518         list_del(&mad_agent_priv->agent_list);
519         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
520
521         flush_workqueue(port_priv->wq);
522         ib_cancel_rmpp_recvs(mad_agent_priv);
523
524         deref_mad_agent(mad_agent_priv);
525         wait_for_completion(&mad_agent_priv->comp);
526
527         kfree(mad_agent_priv->reg_req);
528         ib_dereg_mr(mad_agent_priv->agent.mr);
529         kfree(mad_agent_priv);
530 }
531
532 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
533 {
534         struct ib_mad_qp_info *qp_info;
535         unsigned long flags;
536
537         qp_info = mad_snoop_priv->qp_info;
538         spin_lock_irqsave(&qp_info->snoop_lock, flags);
539         qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
540         atomic_dec(&qp_info->snoop_count);
541         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
542
543         deref_snoop_agent(mad_snoop_priv);
544         wait_for_completion(&mad_snoop_priv->comp);
545
546         kfree(mad_snoop_priv);
547 }
548
549 /*
550  * ib_unregister_mad_agent - Unregisters a client from using MAD services
551  */
552 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
553 {
554         struct ib_mad_agent_private *mad_agent_priv;
555         struct ib_mad_snoop_private *mad_snoop_priv;
556
557         /* If the TID is zero, the agent can only snoop. */
558         if (mad_agent->hi_tid) {
559                 mad_agent_priv = container_of(mad_agent,
560                                               struct ib_mad_agent_private,
561                                               agent);
562                 unregister_mad_agent(mad_agent_priv);
563         } else {
564                 mad_snoop_priv = container_of(mad_agent,
565                                               struct ib_mad_snoop_private,
566                                               agent);
567                 unregister_mad_snoop(mad_snoop_priv);
568         }
569         return 0;
570 }
571 EXPORT_SYMBOL(ib_unregister_mad_agent);
572
573 static inline int response_mad(struct ib_mad *mad)
574 {
575         /* Trap represses are responses although response bit is reset */
576         return ((mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
577                 (mad->mad_hdr.method & IB_MGMT_METHOD_RESP));
578 }
579
580 static void dequeue_mad(struct ib_mad_list_head *mad_list)
581 {
582         struct ib_mad_queue *mad_queue;
583         unsigned long flags;
584
585         BUG_ON(!mad_list->mad_queue);
586         mad_queue = mad_list->mad_queue;
587         spin_lock_irqsave(&mad_queue->lock, flags);
588         list_del(&mad_list->list);
589         mad_queue->count--;
590         spin_unlock_irqrestore(&mad_queue->lock, flags);
591 }
592
593 static void snoop_send(struct ib_mad_qp_info *qp_info,
594                        struct ib_mad_send_buf *send_buf,
595                        struct ib_mad_send_wc *mad_send_wc,
596                        int mad_snoop_flags)
597 {
598         struct ib_mad_snoop_private *mad_snoop_priv;
599         unsigned long flags;
600         int i;
601
602         spin_lock_irqsave(&qp_info->snoop_lock, flags);
603         for (i = 0; i < qp_info->snoop_table_size; i++) {
604                 mad_snoop_priv = qp_info->snoop_table[i];
605                 if (!mad_snoop_priv ||
606                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
607                         continue;
608
609                 atomic_inc(&mad_snoop_priv->refcount);
610                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
611                 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
612                                                     send_buf, mad_send_wc);
613                 deref_snoop_agent(mad_snoop_priv);
614                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
615         }
616         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
617 }
618
619 static void snoop_recv(struct ib_mad_qp_info *qp_info,
620                        struct ib_mad_recv_wc *mad_recv_wc,
621                        int mad_snoop_flags)
622 {
623         struct ib_mad_snoop_private *mad_snoop_priv;
624         unsigned long flags;
625         int i;
626
627         spin_lock_irqsave(&qp_info->snoop_lock, flags);
628         for (i = 0; i < qp_info->snoop_table_size; i++) {
629                 mad_snoop_priv = qp_info->snoop_table[i];
630                 if (!mad_snoop_priv ||
631                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
632                         continue;
633
634                 atomic_inc(&mad_snoop_priv->refcount);
635                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
636                 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
637                                                    mad_recv_wc);
638                 deref_snoop_agent(mad_snoop_priv);
639                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
640         }
641         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
642 }
643
644 static void build_smp_wc(u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
645                          struct ib_wc *wc)
646 {
647         memset(wc, 0, sizeof *wc);
648         wc->wr_id = wr_id;
649         wc->status = IB_WC_SUCCESS;
650         wc->opcode = IB_WC_RECV;
651         wc->pkey_index = pkey_index;
652         wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
653         wc->src_qp = IB_QP0;
654         wc->qp_num = IB_QP0;
655         wc->slid = slid;
656         wc->sl = 0;
657         wc->dlid_path_bits = 0;
658         wc->port_num = port_num;
659 }
660
661 /*
662  * Return 0 if SMP is to be sent
663  * Return 1 if SMP was consumed locally (whether or not solicited)
664  * Return < 0 if error
665  */
666 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
667                                   struct ib_mad_send_wr_private *mad_send_wr)
668 {
669         int ret;
670         struct ib_smp *smp = mad_send_wr->send_buf.mad;
671         unsigned long flags;
672         struct ib_mad_local_private *local;
673         struct ib_mad_private *mad_priv;
674         struct ib_mad_port_private *port_priv;
675         struct ib_mad_agent_private *recv_mad_agent = NULL;
676         struct ib_device *device = mad_agent_priv->agent.device;
677         u8 port_num = mad_agent_priv->agent.port_num;
678         struct ib_wc mad_wc;
679         struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
680
681         /*
682          * Directed route handling starts if the initial LID routed part of
683          * a request or the ending LID routed part of a response is empty.
684          * If we are at the start of the LID routed part, don't update the
685          * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
686          */
687         if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
688              IB_LID_PERMISSIVE &&
689             !smi_handle_dr_smp_send(smp, device->node_type, port_num)) {
690                 ret = -EINVAL;
691                 printk(KERN_ERR PFX "Invalid directed route\n");
692                 goto out;
693         }
694         /* Check to post send on QP or process locally */
695         ret = smi_check_local_smp(smp, device);
696         if (!ret)
697                 goto out;
698
699         local = kmalloc(sizeof *local, GFP_ATOMIC);
700         if (!local) {
701                 ret = -ENOMEM;
702                 printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
703                 goto out;
704         }
705         local->mad_priv = NULL;
706         local->recv_mad_agent = NULL;
707         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
708         if (!mad_priv) {
709                 ret = -ENOMEM;
710                 printk(KERN_ERR PFX "No memory for local response MAD\n");
711                 kfree(local);
712                 goto out;
713         }
714
715         build_smp_wc(send_wr->wr_id, be16_to_cpu(smp->dr_slid),
716                      send_wr->wr.ud.pkey_index,
717                      send_wr->wr.ud.port_num, &mad_wc);
718
719         /* No GRH for DR SMP */
720         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
721                                   (struct ib_mad *)smp,
722                                   (struct ib_mad *)&mad_priv->mad);
723         switch (ret)
724         {
725         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
726                 if (response_mad(&mad_priv->mad.mad) &&
727                     mad_agent_priv->agent.recv_handler) {
728                         local->mad_priv = mad_priv;
729                         local->recv_mad_agent = mad_agent_priv;
730                         /*
731                          * Reference MAD agent until receive
732                          * side of local completion handled
733                          */
734                         atomic_inc(&mad_agent_priv->refcount);
735                 } else
736                         kmem_cache_free(ib_mad_cache, mad_priv);
737                 break;
738         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
739                 kmem_cache_free(ib_mad_cache, mad_priv);
740                 break;
741         case IB_MAD_RESULT_SUCCESS:
742                 /* Treat like an incoming receive MAD */
743                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
744                                             mad_agent_priv->agent.port_num);
745                 if (port_priv) {
746                         mad_priv->mad.mad.mad_hdr.tid =
747                                 ((struct ib_mad *)smp)->mad_hdr.tid;
748                         recv_mad_agent = find_mad_agent(port_priv,
749                                                         &mad_priv->mad.mad);
750                 }
751                 if (!port_priv || !recv_mad_agent) {
752                         kmem_cache_free(ib_mad_cache, mad_priv);
753                         kfree(local);
754                         ret = 0;
755                         goto out;
756                 }
757                 local->mad_priv = mad_priv;
758                 local->recv_mad_agent = recv_mad_agent;
759                 break;
760         default:
761                 kmem_cache_free(ib_mad_cache, mad_priv);
762                 kfree(local);
763                 ret = -EINVAL;
764                 goto out;
765         }
766
767         local->mad_send_wr = mad_send_wr;
768         /* Reference MAD agent until send side of local completion handled */
769         atomic_inc(&mad_agent_priv->refcount);
770         /* Queue local completion to local list */
771         spin_lock_irqsave(&mad_agent_priv->lock, flags);
772         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
773         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
774         queue_work(mad_agent_priv->qp_info->port_priv->wq,
775                    &mad_agent_priv->local_work);
776         ret = 1;
777 out:
778         return ret;
779 }
780
781 static int get_pad_size(int hdr_len, int data_len)
782 {
783         int seg_size, pad;
784
785         seg_size = sizeof(struct ib_mad) - hdr_len;
786         if (data_len && seg_size) {
787                 pad = seg_size - data_len % seg_size;
788                 return pad == seg_size ? 0 : pad;
789         } else
790                 return seg_size;
791 }
792
793 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
794 {
795         struct ib_rmpp_segment *s, *t;
796
797         list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
798                 list_del(&s->list);
799                 kfree(s);
800         }
801 }
802
803 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
804                                 gfp_t gfp_mask)
805 {
806         struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
807         struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
808         struct ib_rmpp_segment *seg = NULL;
809         int left, seg_size, pad;
810
811         send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
812         seg_size = send_buf->seg_size;
813         pad = send_wr->pad;
814
815         /* Allocate data segments. */
816         for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
817                 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
818                 if (!seg) {
819                         printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
820                                "alloc failed for len %zd, gfp %#x\n",
821                                sizeof (*seg) + seg_size, gfp_mask);
822                         free_send_rmpp_list(send_wr);
823                         return -ENOMEM;
824                 }
825                 seg->num = ++send_buf->seg_count;
826                 list_add_tail(&seg->list, &send_wr->rmpp_list);
827         }
828
829         /* Zero any padding */
830         if (pad)
831                 memset(seg->data + seg_size - pad, 0, pad);
832
833         rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
834                                           agent.rmpp_version;
835         rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
836         ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
837
838         send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
839                                         struct ib_rmpp_segment, list);
840         send_wr->last_ack_seg = send_wr->cur_seg;
841         return 0;
842 }
843
844 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
845                                             u32 remote_qpn, u16 pkey_index,
846                                             int rmpp_active,
847                                             int hdr_len, int data_len,
848                                             gfp_t gfp_mask)
849 {
850         struct ib_mad_agent_private *mad_agent_priv;
851         struct ib_mad_send_wr_private *mad_send_wr;
852         int pad, message_size, ret, size;
853         void *buf;
854
855         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
856                                       agent);
857         pad = get_pad_size(hdr_len, data_len);
858         message_size = hdr_len + data_len + pad;
859
860         if ((!mad_agent->rmpp_version &&
861              (rmpp_active || message_size > sizeof(struct ib_mad))) ||
862             (!rmpp_active && message_size > sizeof(struct ib_mad)))
863                 return ERR_PTR(-EINVAL);
864
865         size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
866         buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
867         if (!buf)
868                 return ERR_PTR(-ENOMEM);
869
870         mad_send_wr = buf + size;
871         INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
872         mad_send_wr->send_buf.mad = buf;
873         mad_send_wr->send_buf.hdr_len = hdr_len;
874         mad_send_wr->send_buf.data_len = data_len;
875         mad_send_wr->pad = pad;
876
877         mad_send_wr->mad_agent_priv = mad_agent_priv;
878         mad_send_wr->sg_list[0].length = hdr_len;
879         mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
880         mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
881         mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
882
883         mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
884         mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
885         mad_send_wr->send_wr.num_sge = 2;
886         mad_send_wr->send_wr.opcode = IB_WR_SEND;
887         mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
888         mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
889         mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
890         mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
891
892         if (rmpp_active) {
893                 ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
894                 if (ret) {
895                         kfree(buf);
896                         return ERR_PTR(ret);
897                 }
898         }
899
900         mad_send_wr->send_buf.mad_agent = mad_agent;
901         atomic_inc(&mad_agent_priv->refcount);
902         return &mad_send_wr->send_buf;
903 }
904 EXPORT_SYMBOL(ib_create_send_mad);
905
906 int ib_get_mad_data_offset(u8 mgmt_class)
907 {
908         if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
909                 return IB_MGMT_SA_HDR;
910         else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
911                  (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
912                  (mgmt_class == IB_MGMT_CLASS_BIS))
913                 return IB_MGMT_DEVICE_HDR;
914         else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
915                  (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
916                 return IB_MGMT_VENDOR_HDR;
917         else
918                 return IB_MGMT_MAD_HDR;
919 }
920 EXPORT_SYMBOL(ib_get_mad_data_offset);
921
922 int ib_is_mad_class_rmpp(u8 mgmt_class)
923 {
924         if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
925             (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
926             (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
927             (mgmt_class == IB_MGMT_CLASS_BIS) ||
928             ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
929              (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
930                 return 1;
931         return 0;
932 }
933 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
934
935 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
936 {
937         struct ib_mad_send_wr_private *mad_send_wr;
938         struct list_head *list;
939
940         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
941                                    send_buf);
942         list = &mad_send_wr->cur_seg->list;
943
944         if (mad_send_wr->cur_seg->num < seg_num) {
945                 list_for_each_entry(mad_send_wr->cur_seg, list, list)
946                         if (mad_send_wr->cur_seg->num == seg_num)
947                                 break;
948         } else if (mad_send_wr->cur_seg->num > seg_num) {
949                 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
950                         if (mad_send_wr->cur_seg->num == seg_num)
951                                 break;
952         }
953         return mad_send_wr->cur_seg->data;
954 }
955 EXPORT_SYMBOL(ib_get_rmpp_segment);
956
957 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
958 {
959         if (mad_send_wr->send_buf.seg_count)
960                 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
961                                            mad_send_wr->seg_num);
962         else
963                 return mad_send_wr->send_buf.mad +
964                        mad_send_wr->send_buf.hdr_len;
965 }
966
967 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
968 {
969         struct ib_mad_agent_private *mad_agent_priv;
970         struct ib_mad_send_wr_private *mad_send_wr;
971
972         mad_agent_priv = container_of(send_buf->mad_agent,
973                                       struct ib_mad_agent_private, agent);
974         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
975                                    send_buf);
976
977         free_send_rmpp_list(mad_send_wr);
978         kfree(send_buf->mad);
979         deref_mad_agent(mad_agent_priv);
980 }
981 EXPORT_SYMBOL(ib_free_send_mad);
982
983 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
984 {
985         struct ib_mad_qp_info *qp_info;
986         struct list_head *list;
987         struct ib_send_wr *bad_send_wr;
988         struct ib_mad_agent *mad_agent;
989         struct ib_sge *sge;
990         unsigned long flags;
991         int ret;
992
993         /* Set WR ID to find mad_send_wr upon completion */
994         qp_info = mad_send_wr->mad_agent_priv->qp_info;
995         mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
996         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
997
998         mad_agent = mad_send_wr->send_buf.mad_agent;
999         sge = mad_send_wr->sg_list;
1000         sge[0].addr = dma_map_single(mad_agent->device->dma_device,
1001                                      mad_send_wr->send_buf.mad,
1002                                      sge[0].length,
1003                                      DMA_TO_DEVICE);
1004         pci_unmap_addr_set(mad_send_wr, header_mapping, sge[0].addr);
1005
1006         sge[1].addr = dma_map_single(mad_agent->device->dma_device,
1007                                      ib_get_payload(mad_send_wr),
1008                                      sge[1].length,
1009                                      DMA_TO_DEVICE);
1010         pci_unmap_addr_set(mad_send_wr, payload_mapping, sge[1].addr);
1011
1012         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1013         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1014                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1015                                    &bad_send_wr);
1016                 list = &qp_info->send_queue.list;
1017         } else {
1018                 ret = 0;
1019                 list = &qp_info->overflow_list;
1020         }
1021
1022         if (!ret) {
1023                 qp_info->send_queue.count++;
1024                 list_add_tail(&mad_send_wr->mad_list.list, list);
1025         }
1026         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1027         if (ret) {
1028                 dma_unmap_single(mad_agent->device->dma_device,
1029                                  pci_unmap_addr(mad_send_wr, header_mapping),
1030                                  sge[0].length, DMA_TO_DEVICE);
1031                 dma_unmap_single(mad_agent->device->dma_device,
1032                                  pci_unmap_addr(mad_send_wr, payload_mapping),
1033                                  sge[1].length, DMA_TO_DEVICE);
1034         }
1035         return ret;
1036 }
1037
1038 /*
1039  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1040  *  with the registered client
1041  */
1042 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1043                      struct ib_mad_send_buf **bad_send_buf)
1044 {
1045         struct ib_mad_agent_private *mad_agent_priv;
1046         struct ib_mad_send_buf *next_send_buf;
1047         struct ib_mad_send_wr_private *mad_send_wr;
1048         unsigned long flags;
1049         int ret = -EINVAL;
1050
1051         /* Walk list of send WRs and post each on send list */
1052         for (; send_buf; send_buf = next_send_buf) {
1053
1054                 mad_send_wr = container_of(send_buf,
1055                                            struct ib_mad_send_wr_private,
1056                                            send_buf);
1057                 mad_agent_priv = mad_send_wr->mad_agent_priv;
1058
1059                 if (!send_buf->mad_agent->send_handler ||
1060                     (send_buf->timeout_ms &&
1061                      !send_buf->mad_agent->recv_handler)) {
1062                         ret = -EINVAL;
1063                         goto error;
1064                 }
1065
1066                 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1067                         if (mad_agent_priv->agent.rmpp_version) {
1068                                 ret = -EINVAL;
1069                                 goto error;
1070                         }
1071                 }
1072
1073                 /*
1074                  * Save pointer to next work request to post in case the
1075                  * current one completes, and the user modifies the work
1076                  * request associated with the completion
1077                  */
1078                 next_send_buf = send_buf->next;
1079                 mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1080
1081                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1082                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1083                         ret = handle_outgoing_dr_smp(mad_agent_priv,
1084                                                      mad_send_wr);
1085                         if (ret < 0)            /* error */
1086                                 goto error;
1087                         else if (ret == 1)      /* locally consumed */
1088                                 continue;
1089                 }
1090
1091                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1092                 /* Timeout will be updated after send completes */
1093                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1094                 mad_send_wr->retries = send_buf->retries;
1095                 /* Reference for work request to QP + response */
1096                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1097                 mad_send_wr->status = IB_WC_SUCCESS;
1098
1099                 /* Reference MAD agent until send completes */
1100                 atomic_inc(&mad_agent_priv->refcount);
1101                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1102                 list_add_tail(&mad_send_wr->agent_list,
1103                               &mad_agent_priv->send_list);
1104                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1105
1106                 if (mad_agent_priv->agent.rmpp_version) {
1107                         ret = ib_send_rmpp_mad(mad_send_wr);
1108                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1109                                 ret = ib_send_mad(mad_send_wr);
1110                 } else
1111                         ret = ib_send_mad(mad_send_wr);
1112                 if (ret < 0) {
1113                         /* Fail send request */
1114                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1115                         list_del(&mad_send_wr->agent_list);
1116                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1117                         atomic_dec(&mad_agent_priv->refcount);
1118                         goto error;
1119                 }
1120         }
1121         return 0;
1122 error:
1123         if (bad_send_buf)
1124                 *bad_send_buf = send_buf;
1125         return ret;
1126 }
1127 EXPORT_SYMBOL(ib_post_send_mad);
1128
1129 /*
1130  * ib_free_recv_mad - Returns data buffers used to receive
1131  *  a MAD to the access layer
1132  */
1133 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1134 {
1135         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1136         struct ib_mad_private_header *mad_priv_hdr;
1137         struct ib_mad_private *priv;
1138         struct list_head free_list;
1139
1140         INIT_LIST_HEAD(&free_list);
1141         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1142
1143         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1144                                         &free_list, list) {
1145                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1146                                            recv_buf);
1147                 mad_priv_hdr = container_of(mad_recv_wc,
1148                                             struct ib_mad_private_header,
1149                                             recv_wc);
1150                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1151                                     header);
1152                 kmem_cache_free(ib_mad_cache, priv);
1153         }
1154 }
1155 EXPORT_SYMBOL(ib_free_recv_mad);
1156
1157 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1158                                         u8 rmpp_version,
1159                                         ib_mad_send_handler send_handler,
1160                                         ib_mad_recv_handler recv_handler,
1161                                         void *context)
1162 {
1163         return ERR_PTR(-EINVAL);        /* XXX: for now */
1164 }
1165 EXPORT_SYMBOL(ib_redirect_mad_qp);
1166
1167 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1168                       struct ib_wc *wc)
1169 {
1170         printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1171         return 0;
1172 }
1173 EXPORT_SYMBOL(ib_process_mad_wc);
1174
1175 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1176                          struct ib_mad_reg_req *mad_reg_req)
1177 {
1178         int i;
1179
1180         for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1181              i < IB_MGMT_MAX_METHODS;
1182              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1183                                1+i)) {
1184                 if ((*method)->agent[i]) {
1185                         printk(KERN_ERR PFX "Method %d already in use\n", i);
1186                         return -EINVAL;
1187                 }
1188         }
1189         return 0;
1190 }
1191
1192 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1193 {
1194         /* Allocate management method table */
1195         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1196         if (!*method) {
1197                 printk(KERN_ERR PFX "No memory for "
1198                        "ib_mad_mgmt_method_table\n");
1199                 return -ENOMEM;
1200         }
1201
1202         return 0;
1203 }
1204
1205 /*
1206  * Check to see if there are any methods still in use
1207  */
1208 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1209 {
1210         int i;
1211
1212         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1213                 if (method->agent[i])
1214                         return 1;
1215         return 0;
1216 }
1217
1218 /*
1219  * Check to see if there are any method tables for this class still in use
1220  */
1221 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1222 {
1223         int i;
1224
1225         for (i = 0; i < MAX_MGMT_CLASS; i++)
1226                 if (class->method_table[i])
1227                         return 1;
1228         return 0;
1229 }
1230
1231 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1232 {
1233         int i;
1234
1235         for (i = 0; i < MAX_MGMT_OUI; i++)
1236                 if (vendor_class->method_table[i])
1237                         return 1;
1238         return 0;
1239 }
1240
1241 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1242                            char *oui)
1243 {
1244         int i;
1245
1246         for (i = 0; i < MAX_MGMT_OUI; i++)
1247                 /* Is there matching OUI for this vendor class ? */
1248                 if (!memcmp(vendor_class->oui[i], oui, 3))
1249                         return i;
1250
1251         return -1;
1252 }
1253
1254 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1255 {
1256         int i;
1257
1258         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1259                 if (vendor->vendor_class[i])
1260                         return 1;
1261
1262         return 0;
1263 }
1264
1265 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1266                                      struct ib_mad_agent_private *agent)
1267 {
1268         int i;
1269
1270         /* Remove any methods for this mad agent */
1271         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1272                 if (method->agent[i] == agent) {
1273                         method->agent[i] = NULL;
1274                 }
1275         }
1276 }
1277
1278 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1279                               struct ib_mad_agent_private *agent_priv,
1280                               u8 mgmt_class)
1281 {
1282         struct ib_mad_port_private *port_priv;
1283         struct ib_mad_mgmt_class_table **class;
1284         struct ib_mad_mgmt_method_table **method;
1285         int i, ret;
1286
1287         port_priv = agent_priv->qp_info->port_priv;
1288         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1289         if (!*class) {
1290                 /* Allocate management class table for "new" class version */
1291                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1292                 if (!*class) {
1293                         printk(KERN_ERR PFX "No memory for "
1294                                "ib_mad_mgmt_class_table\n");
1295                         ret = -ENOMEM;
1296                         goto error1;
1297                 }
1298
1299                 /* Allocate method table for this management class */
1300                 method = &(*class)->method_table[mgmt_class];
1301                 if ((ret = allocate_method_table(method)))
1302                         goto error2;
1303         } else {
1304                 method = &(*class)->method_table[mgmt_class];
1305                 if (!*method) {
1306                         /* Allocate method table for this management class */
1307                         if ((ret = allocate_method_table(method)))
1308                                 goto error1;
1309                 }
1310         }
1311
1312         /* Now, make sure methods are not already in use */
1313         if (method_in_use(method, mad_reg_req))
1314                 goto error3;
1315
1316         /* Finally, add in methods being registered */
1317         for (i = find_first_bit(mad_reg_req->method_mask,
1318                                 IB_MGMT_MAX_METHODS);
1319              i < IB_MGMT_MAX_METHODS;
1320              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1321                                1+i)) {
1322                 (*method)->agent[i] = agent_priv;
1323         }
1324         return 0;
1325
1326 error3:
1327         /* Remove any methods for this mad agent */
1328         remove_methods_mad_agent(*method, agent_priv);
1329         /* Now, check to see if there are any methods in use */
1330         if (!check_method_table(*method)) {
1331                 /* If not, release management method table */
1332                 kfree(*method);
1333                 *method = NULL;
1334         }
1335         ret = -EINVAL;
1336         goto error1;
1337 error2:
1338         kfree(*class);
1339         *class = NULL;
1340 error1:
1341         return ret;
1342 }
1343
1344 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1345                            struct ib_mad_agent_private *agent_priv)
1346 {
1347         struct ib_mad_port_private *port_priv;
1348         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1349         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1350         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1351         struct ib_mad_mgmt_method_table **method;
1352         int i, ret = -ENOMEM;
1353         u8 vclass;
1354
1355         /* "New" vendor (with OUI) class */
1356         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1357         port_priv = agent_priv->qp_info->port_priv;
1358         vendor_table = &port_priv->version[
1359                                 mad_reg_req->mgmt_class_version].vendor;
1360         if (!*vendor_table) {
1361                 /* Allocate mgmt vendor class table for "new" class version */
1362                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1363                 if (!vendor) {
1364                         printk(KERN_ERR PFX "No memory for "
1365                                "ib_mad_mgmt_vendor_class_table\n");
1366                         goto error1;
1367                 }
1368
1369                 *vendor_table = vendor;
1370         }
1371         if (!(*vendor_table)->vendor_class[vclass]) {
1372                 /* Allocate table for this management vendor class */
1373                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1374                 if (!vendor_class) {
1375                         printk(KERN_ERR PFX "No memory for "
1376                                "ib_mad_mgmt_vendor_class\n");
1377                         goto error2;
1378                 }
1379
1380                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1381         }
1382         for (i = 0; i < MAX_MGMT_OUI; i++) {
1383                 /* Is there matching OUI for this vendor class ? */
1384                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1385                             mad_reg_req->oui, 3)) {
1386                         method = &(*vendor_table)->vendor_class[
1387                                                 vclass]->method_table[i];
1388                         BUG_ON(!*method);
1389                         goto check_in_use;
1390                 }
1391         }
1392         for (i = 0; i < MAX_MGMT_OUI; i++) {
1393                 /* OUI slot available ? */
1394                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1395                                 vclass]->oui[i])) {
1396                         method = &(*vendor_table)->vendor_class[
1397                                 vclass]->method_table[i];
1398                         BUG_ON(*method);
1399                         /* Allocate method table for this OUI */
1400                         if ((ret = allocate_method_table(method)))
1401                                 goto error3;
1402                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1403                                mad_reg_req->oui, 3);
1404                         goto check_in_use;
1405                 }
1406         }
1407         printk(KERN_ERR PFX "All OUI slots in use\n");
1408         goto error3;
1409
1410 check_in_use:
1411         /* Now, make sure methods are not already in use */
1412         if (method_in_use(method, mad_reg_req))
1413                 goto error4;
1414
1415         /* Finally, add in methods being registered */
1416         for (i = find_first_bit(mad_reg_req->method_mask,
1417                                 IB_MGMT_MAX_METHODS);
1418              i < IB_MGMT_MAX_METHODS;
1419              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1420                                1+i)) {
1421                 (*method)->agent[i] = agent_priv;
1422         }
1423         return 0;
1424
1425 error4:
1426         /* Remove any methods for this mad agent */
1427         remove_methods_mad_agent(*method, agent_priv);
1428         /* Now, check to see if there are any methods in use */
1429         if (!check_method_table(*method)) {
1430                 /* If not, release management method table */
1431                 kfree(*method);
1432                 *method = NULL;
1433         }
1434         ret = -EINVAL;
1435 error3:
1436         if (vendor_class) {
1437                 (*vendor_table)->vendor_class[vclass] = NULL;
1438                 kfree(vendor_class);
1439         }
1440 error2:
1441         if (vendor) {
1442                 *vendor_table = NULL;
1443                 kfree(vendor);
1444         }
1445 error1:
1446         return ret;
1447 }
1448
1449 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1450 {
1451         struct ib_mad_port_private *port_priv;
1452         struct ib_mad_mgmt_class_table *class;
1453         struct ib_mad_mgmt_method_table *method;
1454         struct ib_mad_mgmt_vendor_class_table *vendor;
1455         struct ib_mad_mgmt_vendor_class *vendor_class;
1456         int index;
1457         u8 mgmt_class;
1458
1459         /*
1460          * Was MAD registration request supplied
1461          * with original registration ?
1462          */
1463         if (!agent_priv->reg_req) {
1464                 goto out;
1465         }
1466
1467         port_priv = agent_priv->qp_info->port_priv;
1468         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1469         class = port_priv->version[
1470                         agent_priv->reg_req->mgmt_class_version].class;
1471         if (!class)
1472                 goto vendor_check;
1473
1474         method = class->method_table[mgmt_class];
1475         if (method) {
1476                 /* Remove any methods for this mad agent */
1477                 remove_methods_mad_agent(method, agent_priv);
1478                 /* Now, check to see if there are any methods still in use */
1479                 if (!check_method_table(method)) {
1480                         /* If not, release management method table */
1481                          kfree(method);
1482                          class->method_table[mgmt_class] = NULL;
1483                          /* Any management classes left ? */
1484                         if (!check_class_table(class)) {
1485                                 /* If not, release management class table */
1486                                 kfree(class);
1487                                 port_priv->version[
1488                                         agent_priv->reg_req->
1489                                         mgmt_class_version].class = NULL;
1490                         }
1491                 }
1492         }
1493
1494 vendor_check:
1495         if (!is_vendor_class(mgmt_class))
1496                 goto out;
1497
1498         /* normalize mgmt_class to vendor range 2 */
1499         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1500         vendor = port_priv->version[
1501                         agent_priv->reg_req->mgmt_class_version].vendor;
1502
1503         if (!vendor)
1504                 goto out;
1505
1506         vendor_class = vendor->vendor_class[mgmt_class];
1507         if (vendor_class) {
1508                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1509                 if (index < 0)
1510                         goto out;
1511                 method = vendor_class->method_table[index];
1512                 if (method) {
1513                         /* Remove any methods for this mad agent */
1514                         remove_methods_mad_agent(method, agent_priv);
1515                         /*
1516                          * Now, check to see if there are
1517                          * any methods still in use
1518                          */
1519                         if (!check_method_table(method)) {
1520                                 /* If not, release management method table */
1521                                 kfree(method);
1522                                 vendor_class->method_table[index] = NULL;
1523                                 memset(vendor_class->oui[index], 0, 3);
1524                                 /* Any OUIs left ? */
1525                                 if (!check_vendor_class(vendor_class)) {
1526                                         /* If not, release vendor class table */
1527                                         kfree(vendor_class);
1528                                         vendor->vendor_class[mgmt_class] = NULL;
1529                                         /* Any other vendor classes left ? */
1530                                         if (!check_vendor_table(vendor)) {
1531                                                 kfree(vendor);
1532                                                 port_priv->version[
1533                                                         agent_priv->reg_req->
1534                                                         mgmt_class_version].
1535                                                         vendor = NULL;
1536                                         }
1537                                 }
1538                         }
1539                 }
1540         }
1541
1542 out:
1543         return;
1544 }
1545
1546 static struct ib_mad_agent_private *
1547 find_mad_agent(struct ib_mad_port_private *port_priv,
1548                struct ib_mad *mad)
1549 {
1550         struct ib_mad_agent_private *mad_agent = NULL;
1551         unsigned long flags;
1552
1553         spin_lock_irqsave(&port_priv->reg_lock, flags);
1554         if (response_mad(mad)) {
1555                 u32 hi_tid;
1556                 struct ib_mad_agent_private *entry;
1557
1558                 /*
1559                  * Routing is based on high 32 bits of transaction ID
1560                  * of MAD.
1561                  */
1562                 hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1563                 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1564                         if (entry->agent.hi_tid == hi_tid) {
1565                                 mad_agent = entry;
1566                                 break;
1567                         }
1568                 }
1569         } else {
1570                 struct ib_mad_mgmt_class_table *class;
1571                 struct ib_mad_mgmt_method_table *method;
1572                 struct ib_mad_mgmt_vendor_class_table *vendor;
1573                 struct ib_mad_mgmt_vendor_class *vendor_class;
1574                 struct ib_vendor_mad *vendor_mad;
1575                 int index;
1576
1577                 /*
1578                  * Routing is based on version, class, and method
1579                  * For "newer" vendor MADs, also based on OUI
1580                  */
1581                 if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1582                         goto out;
1583                 if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1584                         class = port_priv->version[
1585                                         mad->mad_hdr.class_version].class;
1586                         if (!class)
1587                                 goto out;
1588                         method = class->method_table[convert_mgmt_class(
1589                                                         mad->mad_hdr.mgmt_class)];
1590                         if (method)
1591                                 mad_agent = method->agent[mad->mad_hdr.method &
1592                                                           ~IB_MGMT_METHOD_RESP];
1593                 } else {
1594                         vendor = port_priv->version[
1595                                         mad->mad_hdr.class_version].vendor;
1596                         if (!vendor)
1597                                 goto out;
1598                         vendor_class = vendor->vendor_class[vendor_class_index(
1599                                                 mad->mad_hdr.mgmt_class)];
1600                         if (!vendor_class)
1601                                 goto out;
1602                         /* Find matching OUI */
1603                         vendor_mad = (struct ib_vendor_mad *)mad;
1604                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1605                         if (index == -1)
1606                                 goto out;
1607                         method = vendor_class->method_table[index];
1608                         if (method) {
1609                                 mad_agent = method->agent[mad->mad_hdr.method &
1610                                                           ~IB_MGMT_METHOD_RESP];
1611                         }
1612                 }
1613         }
1614
1615         if (mad_agent) {
1616                 if (mad_agent->agent.recv_handler)
1617                         atomic_inc(&mad_agent->refcount);
1618                 else {
1619                         printk(KERN_NOTICE PFX "No receive handler for client "
1620                                "%p on port %d\n",
1621                                &mad_agent->agent, port_priv->port_num);
1622                         mad_agent = NULL;
1623                 }
1624         }
1625 out:
1626         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1627
1628         return mad_agent;
1629 }
1630
1631 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1632 {
1633         int valid = 0;
1634
1635         /* Make sure MAD base version is understood */
1636         if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1637                 printk(KERN_ERR PFX "MAD received with unsupported base "
1638                        "version %d\n", mad->mad_hdr.base_version);
1639                 goto out;
1640         }
1641
1642         /* Filter SMI packets sent to other than QP0 */
1643         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1644             (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1645                 if (qp_num == 0)
1646                         valid = 1;
1647         } else {
1648                 /* Filter GSI packets sent to QP0 */
1649                 if (qp_num != 0)
1650                         valid = 1;
1651         }
1652
1653 out:
1654         return valid;
1655 }
1656
1657 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1658                        struct ib_mad_hdr *mad_hdr)
1659 {
1660         struct ib_rmpp_mad *rmpp_mad;
1661
1662         rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1663         return !mad_agent_priv->agent.rmpp_version ||
1664                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1665                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1666                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1667 }
1668
1669 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1670                                      struct ib_mad_recv_wc *rwc)
1671 {
1672         return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1673                 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1674 }
1675
1676 static inline int rcv_has_same_gid(struct ib_mad_send_wr_private *wr,
1677                                    struct ib_mad_recv_wc *rwc )
1678 {
1679         struct ib_ah_attr attr;
1680         u8 send_resp, rcv_resp;
1681
1682         send_resp = ((struct ib_mad *)(wr->send_buf.mad))->
1683                      mad_hdr.method & IB_MGMT_METHOD_RESP;
1684         rcv_resp = rwc->recv_buf.mad->mad_hdr.method & IB_MGMT_METHOD_RESP;
1685
1686         if (!send_resp && rcv_resp)
1687                 /* is request/response. GID/LIDs are both local (same). */
1688                 return 1;
1689
1690         if (send_resp == rcv_resp)
1691                 /* both requests, or both responses. GIDs different */
1692                 return 0;
1693
1694         if (ib_query_ah(wr->send_buf.ah, &attr))
1695                 /* Assume not equal, to avoid false positives. */
1696                 return 0;
1697
1698         if (!(attr.ah_flags & IB_AH_GRH) && !(rwc->wc->wc_flags & IB_WC_GRH))
1699                 return attr.dlid == rwc->wc->slid;
1700         else if ((attr.ah_flags & IB_AH_GRH) &&
1701                  (rwc->wc->wc_flags & IB_WC_GRH))
1702                 return memcmp(attr.grh.dgid.raw,
1703                               rwc->recv_buf.grh->sgid.raw, 16) == 0;
1704         else
1705                 /* one has GID, other does not.  Assume different */
1706                 return 0;
1707 }
1708 struct ib_mad_send_wr_private*
1709 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1710                  struct ib_mad_recv_wc *mad_recv_wc)
1711 {
1712         struct ib_mad_send_wr_private *mad_send_wr;
1713         struct ib_mad *mad;
1714
1715         mad = (struct ib_mad *)mad_recv_wc->recv_buf.mad;
1716
1717         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
1718                             agent_list) {
1719                 if ((mad_send_wr->tid == mad->mad_hdr.tid) &&
1720                     rcv_has_same_class(mad_send_wr, mad_recv_wc) &&
1721                     rcv_has_same_gid(mad_send_wr, mad_recv_wc))
1722                         return mad_send_wr;
1723         }
1724
1725         /*
1726          * It's possible to receive the response before we've
1727          * been notified that the send has completed
1728          */
1729         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
1730                             agent_list) {
1731                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
1732                     mad_send_wr->tid == mad->mad_hdr.tid &&
1733                     mad_send_wr->timeout &&
1734                     rcv_has_same_class(mad_send_wr, mad_recv_wc) &&
1735                     rcv_has_same_gid(mad_send_wr, mad_recv_wc)) {
1736                         /* Verify request has not been canceled */
1737                         return (mad_send_wr->status == IB_WC_SUCCESS) ?
1738                                 mad_send_wr : NULL;
1739                 }
1740         }
1741         return NULL;
1742 }
1743
1744 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1745 {
1746         mad_send_wr->timeout = 0;
1747         if (mad_send_wr->refcount == 1) {
1748                 list_del(&mad_send_wr->agent_list);
1749                 list_add_tail(&mad_send_wr->agent_list,
1750                               &mad_send_wr->mad_agent_priv->done_list);
1751         }
1752 }
1753
1754 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1755                                  struct ib_mad_recv_wc *mad_recv_wc)
1756 {
1757         struct ib_mad_send_wr_private *mad_send_wr;
1758         struct ib_mad_send_wc mad_send_wc;
1759         unsigned long flags;
1760
1761         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1762         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1763         if (mad_agent_priv->agent.rmpp_version) {
1764                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1765                                                       mad_recv_wc);
1766                 if (!mad_recv_wc) {
1767                         deref_mad_agent(mad_agent_priv);
1768                         return;
1769                 }
1770         }
1771
1772         /* Complete corresponding request */
1773         if (response_mad(mad_recv_wc->recv_buf.mad)) {
1774                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1775                 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1776                 if (!mad_send_wr) {
1777                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1778                         ib_free_recv_mad(mad_recv_wc);
1779                         deref_mad_agent(mad_agent_priv);
1780                         return;
1781                 }
1782                 ib_mark_mad_done(mad_send_wr);
1783                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1784
1785                 /* Defined behavior is to complete response before request */
1786                 mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1787                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1788                                                    mad_recv_wc);
1789                 atomic_dec(&mad_agent_priv->refcount);
1790
1791                 mad_send_wc.status = IB_WC_SUCCESS;
1792                 mad_send_wc.vendor_err = 0;
1793                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
1794                 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1795         } else {
1796                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1797                                                    mad_recv_wc);
1798                 deref_mad_agent(mad_agent_priv);
1799         }
1800 }
1801
1802 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1803                                      struct ib_wc *wc)
1804 {
1805         struct ib_mad_qp_info *qp_info;
1806         struct ib_mad_private_header *mad_priv_hdr;
1807         struct ib_mad_private *recv, *response;
1808         struct ib_mad_list_head *mad_list;
1809         struct ib_mad_agent_private *mad_agent;
1810
1811         response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1812         if (!response)
1813                 printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1814                        "for response buffer\n");
1815
1816         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1817         qp_info = mad_list->mad_queue->qp_info;
1818         dequeue_mad(mad_list);
1819
1820         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1821                                     mad_list);
1822         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1823         dma_unmap_single(port_priv->device->dma_device,
1824                          pci_unmap_addr(&recv->header, mapping),
1825                          sizeof(struct ib_mad_private) -
1826                          sizeof(struct ib_mad_private_header),
1827                          DMA_FROM_DEVICE);
1828
1829         /* Setup MAD receive work completion from "normal" work completion */
1830         recv->header.wc = *wc;
1831         recv->header.recv_wc.wc = &recv->header.wc;
1832         recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1833         recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1834         recv->header.recv_wc.recv_buf.grh = &recv->grh;
1835
1836         if (atomic_read(&qp_info->snoop_count))
1837                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1838
1839         /* Validate MAD */
1840         if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1841                 goto out;
1842
1843         if (recv->mad.mad.mad_hdr.mgmt_class ==
1844             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1845                 if (!smi_handle_dr_smp_recv(&recv->mad.smp,
1846                                             port_priv->device->node_type,
1847                                             port_priv->port_num,
1848                                             port_priv->device->phys_port_cnt))
1849                         goto out;
1850                 if (!smi_check_forward_dr_smp(&recv->mad.smp))
1851                         goto local;
1852                 if (!smi_handle_dr_smp_send(&recv->mad.smp,
1853                                             port_priv->device->node_type,
1854                                             port_priv->port_num))
1855                         goto out;
1856                 if (!smi_check_local_smp(&recv->mad.smp, port_priv->device))
1857                         goto out;
1858         }
1859
1860 local:
1861         /* Give driver "right of first refusal" on incoming MAD */
1862         if (port_priv->device->process_mad) {
1863                 int ret;
1864
1865                 if (!response) {
1866                         printk(KERN_ERR PFX "No memory for response MAD\n");
1867                         /*
1868                          * Is it better to assume that
1869                          * it wouldn't be processed ?
1870                          */
1871                         goto out;
1872                 }
1873
1874                 ret = port_priv->device->process_mad(port_priv->device, 0,
1875                                                      port_priv->port_num,
1876                                                      wc, &recv->grh,
1877                                                      &recv->mad.mad,
1878                                                      &response->mad.mad);
1879                 if (ret & IB_MAD_RESULT_SUCCESS) {
1880                         if (ret & IB_MAD_RESULT_CONSUMED)
1881                                 goto out;
1882                         if (ret & IB_MAD_RESULT_REPLY) {
1883                                 agent_send_response(&response->mad.mad,
1884                                                     &recv->grh, wc,
1885                                                     port_priv->device,
1886                                                     port_priv->port_num,
1887                                                     qp_info->qp->qp_num);
1888                                 goto out;
1889                         }
1890                 }
1891         }
1892
1893         mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1894         if (mad_agent) {
1895                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1896                 /*
1897                  * recv is freed up in error cases in ib_mad_complete_recv
1898                  * or via recv_handler in ib_mad_complete_recv()
1899                  */
1900                 recv = NULL;
1901         }
1902
1903 out:
1904         /* Post another receive request for this QP */
1905         if (response) {
1906                 ib_mad_post_receive_mads(qp_info, response);
1907                 if (recv)
1908                         kmem_cache_free(ib_mad_cache, recv);
1909         } else
1910                 ib_mad_post_receive_mads(qp_info, recv);
1911 }
1912
1913 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1914 {
1915         struct ib_mad_send_wr_private *mad_send_wr;
1916         unsigned long delay;
1917
1918         if (list_empty(&mad_agent_priv->wait_list)) {
1919                 cancel_delayed_work(&mad_agent_priv->timed_work);
1920         } else {
1921                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1922                                          struct ib_mad_send_wr_private,
1923                                          agent_list);
1924
1925                 if (time_after(mad_agent_priv->timeout,
1926                                mad_send_wr->timeout)) {
1927                         mad_agent_priv->timeout = mad_send_wr->timeout;
1928                         cancel_delayed_work(&mad_agent_priv->timed_work);
1929                         delay = mad_send_wr->timeout - jiffies;
1930                         if ((long)delay <= 0)
1931                                 delay = 1;
1932                         queue_delayed_work(mad_agent_priv->qp_info->
1933                                            port_priv->wq,
1934                                            &mad_agent_priv->timed_work, delay);
1935                 }
1936         }
1937 }
1938
1939 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1940 {
1941         struct ib_mad_agent_private *mad_agent_priv;
1942         struct ib_mad_send_wr_private *temp_mad_send_wr;
1943         struct list_head *list_item;
1944         unsigned long delay;
1945
1946         mad_agent_priv = mad_send_wr->mad_agent_priv;
1947         list_del(&mad_send_wr->agent_list);
1948
1949         delay = mad_send_wr->timeout;
1950         mad_send_wr->timeout += jiffies;
1951
1952         if (delay) {
1953                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1954                         temp_mad_send_wr = list_entry(list_item,
1955                                                 struct ib_mad_send_wr_private,
1956                                                 agent_list);
1957                         if (time_after(mad_send_wr->timeout,
1958                                        temp_mad_send_wr->timeout))
1959                                 break;
1960                 }
1961         }
1962         else
1963                 list_item = &mad_agent_priv->wait_list;
1964         list_add(&mad_send_wr->agent_list, list_item);
1965
1966         /* Reschedule a work item if we have a shorter timeout */
1967         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
1968                 cancel_delayed_work(&mad_agent_priv->timed_work);
1969                 queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
1970                                    &mad_agent_priv->timed_work, delay);
1971         }
1972 }
1973
1974 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
1975                           int timeout_ms)
1976 {
1977         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
1978         wait_for_response(mad_send_wr);
1979 }
1980
1981 /*
1982  * Process a send work completion
1983  */
1984 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
1985                              struct ib_mad_send_wc *mad_send_wc)
1986 {
1987         struct ib_mad_agent_private     *mad_agent_priv;
1988         unsigned long                   flags;
1989         int                             ret;
1990
1991         mad_agent_priv = mad_send_wr->mad_agent_priv;
1992         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1993         if (mad_agent_priv->agent.rmpp_version) {
1994                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
1995                 if (ret == IB_RMPP_RESULT_CONSUMED)
1996                         goto done;
1997         } else
1998                 ret = IB_RMPP_RESULT_UNHANDLED;
1999
2000         if (mad_send_wc->status != IB_WC_SUCCESS &&
2001             mad_send_wr->status == IB_WC_SUCCESS) {
2002                 mad_send_wr->status = mad_send_wc->status;
2003                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2004         }
2005
2006         if (--mad_send_wr->refcount > 0) {
2007                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2008                     mad_send_wr->status == IB_WC_SUCCESS) {
2009                         wait_for_response(mad_send_wr);
2010                 }
2011                 goto done;
2012         }
2013
2014         /* Remove send from MAD agent and notify client of completion */
2015         list_del(&mad_send_wr->agent_list);
2016         adjust_timeout(mad_agent_priv);
2017         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2018
2019         if (mad_send_wr->status != IB_WC_SUCCESS )
2020                 mad_send_wc->status = mad_send_wr->status;
2021         if (ret == IB_RMPP_RESULT_INTERNAL)
2022                 ib_rmpp_send_handler(mad_send_wc);
2023         else
2024                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2025                                                    mad_send_wc);
2026
2027         /* Release reference on agent taken when sending */
2028         deref_mad_agent(mad_agent_priv);
2029         return;
2030 done:
2031         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2032 }
2033
2034 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2035                                      struct ib_wc *wc)
2036 {
2037         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
2038         struct ib_mad_list_head         *mad_list;
2039         struct ib_mad_qp_info           *qp_info;
2040         struct ib_mad_queue             *send_queue;
2041         struct ib_send_wr               *bad_send_wr;
2042         struct ib_mad_send_wc           mad_send_wc;
2043         unsigned long flags;
2044         int ret;
2045
2046         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2047         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2048                                    mad_list);
2049         send_queue = mad_list->mad_queue;
2050         qp_info = send_queue->qp_info;
2051
2052 retry:
2053         dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
2054                          pci_unmap_addr(mad_send_wr, header_mapping),
2055                          mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2056         dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
2057                          pci_unmap_addr(mad_send_wr, payload_mapping),
2058                          mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2059         queued_send_wr = NULL;
2060         spin_lock_irqsave(&send_queue->lock, flags);
2061         list_del(&mad_list->list);
2062
2063         /* Move queued send to the send queue */
2064         if (send_queue->count-- > send_queue->max_active) {
2065                 mad_list = container_of(qp_info->overflow_list.next,
2066                                         struct ib_mad_list_head, list);
2067                 queued_send_wr = container_of(mad_list,
2068                                         struct ib_mad_send_wr_private,
2069                                         mad_list);
2070                 list_del(&mad_list->list);
2071                 list_add_tail(&mad_list->list, &send_queue->list);
2072         }
2073         spin_unlock_irqrestore(&send_queue->lock, flags);
2074
2075         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2076         mad_send_wc.status = wc->status;
2077         mad_send_wc.vendor_err = wc->vendor_err;
2078         if (atomic_read(&qp_info->snoop_count))
2079                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2080                            IB_MAD_SNOOP_SEND_COMPLETIONS);
2081         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2082
2083         if (queued_send_wr) {
2084                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2085                                    &bad_send_wr);
2086                 if (ret) {
2087                         printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2088                         mad_send_wr = queued_send_wr;
2089                         wc->status = IB_WC_LOC_QP_OP_ERR;
2090                         goto retry;
2091                 }
2092         }
2093 }
2094
2095 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2096 {
2097         struct ib_mad_send_wr_private *mad_send_wr;
2098         struct ib_mad_list_head *mad_list;
2099         unsigned long flags;
2100
2101         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2102         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2103                 mad_send_wr = container_of(mad_list,
2104                                            struct ib_mad_send_wr_private,
2105                                            mad_list);
2106                 mad_send_wr->retry = 1;
2107         }
2108         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2109 }
2110
2111 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2112                               struct ib_wc *wc)
2113 {
2114         struct ib_mad_list_head *mad_list;
2115         struct ib_mad_qp_info *qp_info;
2116         struct ib_mad_send_wr_private *mad_send_wr;
2117         int ret;
2118
2119         /* Determine if failure was a send or receive */
2120         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2121         qp_info = mad_list->mad_queue->qp_info;
2122         if (mad_list->mad_queue == &qp_info->recv_queue)
2123                 /*
2124                  * Receive errors indicate that the QP has entered the error
2125                  * state - error handling/shutdown code will cleanup
2126                  */
2127                 return;
2128
2129         /*
2130          * Send errors will transition the QP to SQE - move
2131          * QP to RTS and repost flushed work requests
2132          */
2133         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2134                                    mad_list);
2135         if (wc->status == IB_WC_WR_FLUSH_ERR) {
2136                 if (mad_send_wr->retry) {
2137                         /* Repost send */
2138                         struct ib_send_wr *bad_send_wr;
2139
2140                         mad_send_wr->retry = 0;
2141                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2142                                         &bad_send_wr);
2143                         if (ret)
2144                                 ib_mad_send_done_handler(port_priv, wc);
2145                 } else
2146                         ib_mad_send_done_handler(port_priv, wc);
2147         } else {
2148                 struct ib_qp_attr *attr;
2149
2150                 /* Transition QP to RTS and fail offending send */
2151                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2152                 if (attr) {
2153                         attr->qp_state = IB_QPS_RTS;
2154                         attr->cur_qp_state = IB_QPS_SQE;
2155                         ret = ib_modify_qp(qp_info->qp, attr,
2156                                            IB_QP_STATE | IB_QP_CUR_STATE);
2157                         kfree(attr);
2158                         if (ret)
2159                                 printk(KERN_ERR PFX "mad_error_handler - "
2160                                        "ib_modify_qp to RTS : %d\n", ret);
2161                         else
2162                                 mark_sends_for_retry(qp_info);
2163                 }
2164                 ib_mad_send_done_handler(port_priv, wc);
2165         }
2166 }
2167
2168 /*
2169  * IB MAD completion callback
2170  */
2171 static void ib_mad_completion_handler(void *data)
2172 {
2173         struct ib_mad_port_private *port_priv;
2174         struct ib_wc wc;
2175
2176         port_priv = (struct ib_mad_port_private *)data;
2177         ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2178
2179         while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2180                 if (wc.status == IB_WC_SUCCESS) {
2181                         switch (wc.opcode) {
2182                         case IB_WC_SEND:
2183                                 ib_mad_send_done_handler(port_priv, &wc);
2184                                 break;
2185                         case IB_WC_RECV:
2186                                 ib_mad_recv_done_handler(port_priv, &wc);
2187                                 break;
2188                         default:
2189                                 BUG_ON(1);
2190                                 break;
2191                         }
2192                 } else
2193                         mad_error_handler(port_priv, &wc);
2194         }
2195 }
2196
2197 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2198 {
2199         unsigned long flags;
2200         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2201         struct ib_mad_send_wc mad_send_wc;
2202         struct list_head cancel_list;
2203
2204         INIT_LIST_HEAD(&cancel_list);
2205
2206         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2207         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2208                                  &mad_agent_priv->send_list, agent_list) {
2209                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2210                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2211                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2212                 }
2213         }
2214
2215         /* Empty wait list to prevent receives from finding a request */
2216         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2217         /* Empty local completion list as well */
2218         list_splice_init(&mad_agent_priv->local_list, &cancel_list);
2219         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2220
2221         /* Report all cancelled requests */
2222         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2223         mad_send_wc.vendor_err = 0;
2224
2225         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2226                                  &cancel_list, agent_list) {
2227                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2228                 list_del(&mad_send_wr->agent_list);
2229                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2230                                                    &mad_send_wc);
2231                 atomic_dec(&mad_agent_priv->refcount);
2232         }
2233 }
2234
2235 static struct ib_mad_send_wr_private*
2236 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2237              struct ib_mad_send_buf *send_buf)
2238 {
2239         struct ib_mad_send_wr_private *mad_send_wr;
2240
2241         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2242                             agent_list) {
2243                 if (&mad_send_wr->send_buf == send_buf)
2244                         return mad_send_wr;
2245         }
2246
2247         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2248                             agent_list) {
2249                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2250                     &mad_send_wr->send_buf == send_buf)
2251                         return mad_send_wr;
2252         }
2253         return NULL;
2254 }
2255
2256 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2257                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2258 {
2259         struct ib_mad_agent_private *mad_agent_priv;
2260         struct ib_mad_send_wr_private *mad_send_wr;
2261         unsigned long flags;
2262         int active;
2263
2264         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2265                                       agent);
2266         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2267         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2268         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2269                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2270                 return -EINVAL;
2271         }
2272
2273         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2274         if (!timeout_ms) {
2275                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2276                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2277         }
2278
2279         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2280         if (active)
2281                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2282         else
2283                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2284
2285         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2286         return 0;
2287 }
2288 EXPORT_SYMBOL(ib_modify_mad);
2289
2290 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2291                    struct ib_mad_send_buf *send_buf)
2292 {
2293         ib_modify_mad(mad_agent, send_buf, 0);
2294 }
2295 EXPORT_SYMBOL(ib_cancel_mad);
2296
2297 static void local_completions(void *data)
2298 {
2299         struct ib_mad_agent_private *mad_agent_priv;
2300         struct ib_mad_local_private *local;
2301         struct ib_mad_agent_private *recv_mad_agent;
2302         unsigned long flags;
2303         int recv = 0;
2304         struct ib_wc wc;
2305         struct ib_mad_send_wc mad_send_wc;
2306
2307         mad_agent_priv = (struct ib_mad_agent_private *)data;
2308
2309         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2310         while (!list_empty(&mad_agent_priv->local_list)) {
2311                 local = list_entry(mad_agent_priv->local_list.next,
2312                                    struct ib_mad_local_private,
2313                                    completion_list);
2314                 list_del(&local->completion_list);
2315                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2316                 if (local->mad_priv) {
2317                         recv_mad_agent = local->recv_mad_agent;
2318                         if (!recv_mad_agent) {
2319                                 printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2320                                 goto local_send_completion;
2321                         }
2322
2323                         recv = 1;
2324                         /*
2325                          * Defined behavior is to complete response
2326                          * before request
2327                          */
2328                         build_smp_wc((unsigned long) local->mad_send_wr,
2329                                      be16_to_cpu(IB_LID_PERMISSIVE),
2330                                      0, recv_mad_agent->agent.port_num, &wc);
2331
2332                         local->mad_priv->header.recv_wc.wc = &wc;
2333                         local->mad_priv->header.recv_wc.mad_len =
2334                                                 sizeof(struct ib_mad);
2335                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2336                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2337                                  &local->mad_priv->header.recv_wc.rmpp_list);
2338                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2339                         local->mad_priv->header.recv_wc.recv_buf.mad =
2340                                                 &local->mad_priv->mad.mad;
2341                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2342                                 snoop_recv(recv_mad_agent->qp_info,
2343                                           &local->mad_priv->header.recv_wc,
2344                                            IB_MAD_SNOOP_RECVS);
2345                         recv_mad_agent->agent.recv_handler(
2346                                                 &recv_mad_agent->agent,
2347                                                 &local->mad_priv->header.recv_wc);
2348                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2349                         atomic_dec(&recv_mad_agent->refcount);
2350                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2351                 }
2352
2353 local_send_completion:
2354                 /* Complete send */
2355                 mad_send_wc.status = IB_WC_SUCCESS;
2356                 mad_send_wc.vendor_err = 0;
2357                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2358                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2359                         snoop_send(mad_agent_priv->qp_info,
2360                                    &local->mad_send_wr->send_buf,
2361                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2362                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2363                                                    &mad_send_wc);
2364
2365                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2366                 atomic_dec(&mad_agent_priv->refcount);
2367                 if (!recv)
2368                         kmem_cache_free(ib_mad_cache, local->mad_priv);
2369                 kfree(local);
2370         }
2371         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2372 }
2373
2374 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2375 {
2376         int ret;
2377
2378         if (!mad_send_wr->retries--)
2379                 return -ETIMEDOUT;
2380
2381         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2382
2383         if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2384                 ret = ib_retry_rmpp(mad_send_wr);
2385                 switch (ret) {
2386                 case IB_RMPP_RESULT_UNHANDLED:
2387                         ret = ib_send_mad(mad_send_wr);
2388                         break;
2389                 case IB_RMPP_RESULT_CONSUMED:
2390                         ret = 0;
2391                         break;
2392                 default:
2393                         ret = -ECOMM;
2394                         break;
2395                 }
2396         } else
2397                 ret = ib_send_mad(mad_send_wr);
2398
2399         if (!ret) {
2400                 mad_send_wr->refcount++;
2401                 list_add_tail(&mad_send_wr->agent_list,
2402                               &mad_send_wr->mad_agent_priv->send_list);
2403         }
2404         return ret;
2405 }
2406
2407 static void timeout_sends(void *data)
2408 {
2409         struct ib_mad_agent_private *mad_agent_priv;
2410         struct ib_mad_send_wr_private *mad_send_wr;
2411         struct ib_mad_send_wc mad_send_wc;
2412         unsigned long flags, delay;
2413
2414         mad_agent_priv = (struct ib_mad_agent_private *)data;
2415         mad_send_wc.vendor_err = 0;
2416
2417         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2418         while (!list_empty(&mad_agent_priv->wait_list)) {
2419                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2420                                          struct ib_mad_send_wr_private,
2421                                          agent_list);
2422
2423                 if (time_after(mad_send_wr->timeout, jiffies)) {
2424                         delay = mad_send_wr->timeout - jiffies;
2425                         if ((long)delay <= 0)
2426                                 delay = 1;
2427                         queue_delayed_work(mad_agent_priv->qp_info->
2428                                            port_priv->wq,
2429                                            &mad_agent_priv->timed_work, delay);
2430                         break;
2431                 }
2432
2433                 list_del(&mad_send_wr->agent_list);
2434                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2435                     !retry_send(mad_send_wr))
2436                         continue;
2437
2438                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2439
2440                 if (mad_send_wr->status == IB_WC_SUCCESS)
2441                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2442                 else
2443                         mad_send_wc.status = mad_send_wr->status;
2444                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2445                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2446                                                    &mad_send_wc);
2447
2448                 atomic_dec(&mad_agent_priv->refcount);
2449                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2450         }
2451         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2452 }
2453
2454 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2455 {
2456         struct ib_mad_port_private *port_priv = cq->cq_context;
2457         unsigned long flags;
2458
2459         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2460         if (!list_empty(&port_priv->port_list))
2461                 queue_work(port_priv->wq, &port_priv->work);
2462         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2463 }
2464
2465 /*
2466  * Allocate receive MADs and post receive WRs for them
2467  */
2468 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2469                                     struct ib_mad_private *mad)
2470 {
2471         unsigned long flags;
2472         int post, ret;
2473         struct ib_mad_private *mad_priv;
2474         struct ib_sge sg_list;
2475         struct ib_recv_wr recv_wr, *bad_recv_wr;
2476         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2477
2478         /* Initialize common scatter list fields */
2479         sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2480         sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2481
2482         /* Initialize common receive WR fields */
2483         recv_wr.next = NULL;
2484         recv_wr.sg_list = &sg_list;
2485         recv_wr.num_sge = 1;
2486
2487         do {
2488                 /* Allocate and map receive buffer */
2489                 if (mad) {
2490                         mad_priv = mad;
2491                         mad = NULL;
2492                 } else {
2493                         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2494                         if (!mad_priv) {
2495                                 printk(KERN_ERR PFX "No memory for receive buffer\n");
2496                                 ret = -ENOMEM;
2497                                 break;
2498                         }
2499                 }
2500                 sg_list.addr = dma_map_single(qp_info->port_priv->
2501                                                 device->dma_device,
2502                                               &mad_priv->grh,
2503                                               sizeof *mad_priv -
2504                                                 sizeof mad_priv->header,
2505                                               DMA_FROM_DEVICE);
2506                 pci_unmap_addr_set(&mad_priv->header, mapping, sg_list.addr);
2507                 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2508                 mad_priv->header.mad_list.mad_queue = recv_queue;
2509
2510                 /* Post receive WR */
2511                 spin_lock_irqsave(&recv_queue->lock, flags);
2512                 post = (++recv_queue->count < recv_queue->max_active);
2513                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2514                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2515                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2516                 if (ret) {
2517                         spin_lock_irqsave(&recv_queue->lock, flags);
2518                         list_del(&mad_priv->header.mad_list.list);
2519                         recv_queue->count--;
2520                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2521                         dma_unmap_single(qp_info->port_priv->device->dma_device,
2522                                          pci_unmap_addr(&mad_priv->header,
2523                                                         mapping),
2524                                          sizeof *mad_priv -
2525                                            sizeof mad_priv->header,
2526                                          DMA_FROM_DEVICE);
2527                         kmem_cache_free(ib_mad_cache, mad_priv);
2528                         printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2529                         break;
2530                 }
2531         } while (post);
2532
2533         return ret;
2534 }
2535
2536 /*
2537  * Return all the posted receive MADs
2538  */
2539 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2540 {
2541         struct ib_mad_private_header *mad_priv_hdr;
2542         struct ib_mad_private *recv;
2543         struct ib_mad_list_head *mad_list;
2544
2545         while (!list_empty(&qp_info->recv_queue.list)) {
2546
2547                 mad_list = list_entry(qp_info->recv_queue.list.next,
2548                                       struct ib_mad_list_head, list);
2549                 mad_priv_hdr = container_of(mad_list,
2550                                             struct ib_mad_private_header,
2551                                             mad_list);
2552                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2553                                     header);
2554
2555                 /* Remove from posted receive MAD list */
2556                 list_del(&mad_list->list);
2557
2558                 dma_unmap_single(qp_info->port_priv->device->dma_device,
2559                                  pci_unmap_addr(&recv->header, mapping),
2560                                  sizeof(struct ib_mad_private) -
2561                                  sizeof(struct ib_mad_private_header),
2562                                  DMA_FROM_DEVICE);
2563                 kmem_cache_free(ib_mad_cache, recv);
2564         }
2565
2566         qp_info->recv_queue.count = 0;
2567 }
2568
2569 /*
2570  * Start the port
2571  */
2572 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2573 {
2574         int ret, i;
2575         struct ib_qp_attr *attr;
2576         struct ib_qp *qp;
2577
2578         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2579         if (!attr) {
2580                 printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2581                 return -ENOMEM;
2582         }
2583
2584         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2585                 qp = port_priv->qp_info[i].qp;
2586                 /*
2587                  * PKey index for QP1 is irrelevant but
2588                  * one is needed for the Reset to Init transition
2589                  */
2590                 attr->qp_state = IB_QPS_INIT;
2591                 attr->pkey_index = 0;
2592                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2593                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2594                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2595                 if (ret) {
2596                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2597                                "INIT: %d\n", i, ret);
2598                         goto out;
2599                 }
2600
2601                 attr->qp_state = IB_QPS_RTR;
2602                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2603                 if (ret) {
2604                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2605                                "RTR: %d\n", i, ret);
2606                         goto out;
2607                 }
2608
2609                 attr->qp_state = IB_QPS_RTS;
2610                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2611                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2612                 if (ret) {
2613                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2614                                "RTS: %d\n", i, ret);
2615                         goto out;
2616                 }
2617         }
2618
2619         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2620         if (ret) {
2621                 printk(KERN_ERR PFX "Failed to request completion "
2622                        "notification: %d\n", ret);
2623                 goto out;
2624         }
2625
2626         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2627                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2628                 if (ret) {
2629                         printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2630                         goto out;
2631                 }
2632         }
2633 out:
2634         kfree(attr);
2635         return ret;
2636 }
2637
2638 static void qp_event_handler(struct ib_event *event, void *qp_context)
2639 {
2640         struct ib_mad_qp_info   *qp_info = qp_context;
2641
2642         /* It's worse than that! He's dead, Jim! */
2643         printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2644                 event->event, qp_info->qp->qp_num);
2645 }
2646
2647 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2648                            struct ib_mad_queue *mad_queue)
2649 {
2650         mad_queue->qp_info = qp_info;
2651         mad_queue->count = 0;
2652         spin_lock_init(&mad_queue->lock);
2653         INIT_LIST_HEAD(&mad_queue->list);
2654 }
2655
2656 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2657                         struct ib_mad_qp_info *qp_info)
2658 {
2659         qp_info->port_priv = port_priv;
2660         init_mad_queue(qp_info, &qp_info->send_queue);
2661         init_mad_queue(qp_info, &qp_info->recv_queue);
2662         INIT_LIST_HEAD(&qp_info->overflow_list);
2663         spin_lock_init(&qp_info->snoop_lock);
2664         qp_info->snoop_table = NULL;
2665         qp_info->snoop_table_size = 0;
2666         atomic_set(&qp_info->snoop_count, 0);
2667 }
2668
2669 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2670                          enum ib_qp_type qp_type)
2671 {
2672         struct ib_qp_init_attr  qp_init_attr;
2673         int ret;
2674
2675         memset(&qp_init_attr, 0, sizeof qp_init_attr);
2676         qp_init_attr.send_cq = qp_info->port_priv->cq;
2677         qp_init_attr.recv_cq = qp_info->port_priv->cq;
2678         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2679         qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2680         qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2681         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2682         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2683         qp_init_attr.qp_type = qp_type;
2684         qp_init_attr.port_num = qp_info->port_priv->port_num;
2685         qp_init_attr.qp_context = qp_info;
2686         qp_init_attr.event_handler = qp_event_handler;
2687         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2688         if (IS_ERR(qp_info->qp)) {
2689                 printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2690                        get_spl_qp_index(qp_type));
2691                 ret = PTR_ERR(qp_info->qp);
2692                 goto error;
2693         }
2694         /* Use minimum queue sizes unless the CQ is resized */
2695         qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2696         qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2697         return 0;
2698
2699 error:
2700         return ret;
2701 }
2702
2703 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2704 {
2705         ib_destroy_qp(qp_info->qp);
2706         kfree(qp_info->snoop_table);
2707 }
2708
2709 /*
2710  * Open the port
2711  * Create the QP, PD, MR, and CQ if needed
2712  */
2713 static int ib_mad_port_open(struct ib_device *device,
2714                             int port_num)
2715 {
2716         int ret, cq_size;
2717         struct ib_mad_port_private *port_priv;
2718         unsigned long flags;
2719         char name[sizeof "ib_mad123"];
2720
2721         /* Create new device info */
2722         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2723         if (!port_priv) {
2724                 printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2725                 return -ENOMEM;
2726         }
2727
2728         port_priv->device = device;
2729         port_priv->port_num = port_num;
2730         spin_lock_init(&port_priv->reg_lock);
2731         INIT_LIST_HEAD(&port_priv->agent_list);
2732         init_mad_qp(port_priv, &port_priv->qp_info[0]);
2733         init_mad_qp(port_priv, &port_priv->qp_info[1]);
2734
2735         cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2736         port_priv->cq = ib_create_cq(port_priv->device,
2737                                      ib_mad_thread_completion_handler,
2738                                      NULL, port_priv, cq_size);
2739         if (IS_ERR(port_priv->cq)) {
2740                 printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2741                 ret = PTR_ERR(port_priv->cq);
2742                 goto error3;
2743         }
2744
2745         port_priv->pd = ib_alloc_pd(device);
2746         if (IS_ERR(port_priv->pd)) {
2747                 printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2748                 ret = PTR_ERR(port_priv->pd);
2749                 goto error4;
2750         }
2751
2752         port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2753         if (IS_ERR(port_priv->mr)) {
2754                 printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2755                 ret = PTR_ERR(port_priv->mr);
2756                 goto error5;
2757         }
2758
2759         ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2760         if (ret)
2761                 goto error6;
2762         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2763         if (ret)
2764                 goto error7;
2765
2766         snprintf(name, sizeof name, "ib_mad%d", port_num);
2767         port_priv->wq = create_singlethread_workqueue(name);
2768         if (!port_priv->wq) {
2769                 ret = -ENOMEM;
2770                 goto error8;
2771         }
2772         INIT_WORK(&port_priv->work, ib_mad_completion_handler, port_priv);
2773
2774         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2775         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2776         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2777
2778         ret = ib_mad_port_start(port_priv);
2779         if (ret) {
2780                 printk(KERN_ERR PFX "Couldn't start port\n");
2781                 goto error9;
2782         }
2783
2784         return 0;
2785
2786 error9:
2787         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2788         list_del_init(&port_priv->port_list);
2789         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2790
2791         destroy_workqueue(port_priv->wq);
2792 error8:
2793         destroy_mad_qp(&port_priv->qp_info[1]);
2794 error7:
2795         destroy_mad_qp(&port_priv->qp_info[0]);
2796 error6:
2797         ib_dereg_mr(port_priv->mr);
2798 error5:
2799         ib_dealloc_pd(port_priv->pd);
2800 error4:
2801         ib_destroy_cq(port_priv->cq);
2802         cleanup_recv_queue(&port_priv->qp_info[1]);
2803         cleanup_recv_queue(&port_priv->qp_info[0]);
2804 error3:
2805         kfree(port_priv);
2806
2807         return ret;
2808 }
2809
2810 /*
2811  * Close the port
2812  * If there are no classes using the port, free the port
2813  * resources (CQ, MR, PD, QP) and remove the port's info structure
2814  */
2815 static int ib_mad_port_close(struct ib_device *device, int port_num)
2816 {
2817         struct ib_mad_port_private *port_priv;
2818         unsigned long flags;
2819
2820         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2821         port_priv = __ib_get_mad_port(device, port_num);
2822         if (port_priv == NULL) {
2823                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2824                 printk(KERN_ERR PFX "Port %d not found\n", port_num);
2825                 return -ENODEV;
2826         }
2827         list_del_init(&port_priv->port_list);
2828         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2829
2830         destroy_workqueue(port_priv->wq);
2831         destroy_mad_qp(&port_priv->qp_info[1]);
2832         destroy_mad_qp(&port_priv->qp_info[0]);
2833         ib_dereg_mr(port_priv->mr);
2834         ib_dealloc_pd(port_priv->pd);
2835         ib_destroy_cq(port_priv->cq);
2836         cleanup_recv_queue(&port_priv->qp_info[1]);
2837         cleanup_recv_queue(&port_priv->qp_info[0]);
2838         /* XXX: Handle deallocation of MAD registration tables */
2839
2840         kfree(port_priv);
2841
2842         return 0;
2843 }
2844
2845 static void ib_mad_init_device(struct ib_device *device)
2846 {
2847         int start, end, i;
2848
2849         if (device->node_type == IB_NODE_SWITCH) {
2850                 start = 0;
2851                 end   = 0;
2852         } else {
2853                 start = 1;
2854                 end   = device->phys_port_cnt;
2855         }
2856
2857         for (i = start; i <= end; i++) {
2858                 if (ib_mad_port_open(device, i)) {
2859                         printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2860                                device->name, i);
2861                         goto error;
2862                 }
2863                 if (ib_agent_port_open(device, i)) {
2864                         printk(KERN_ERR PFX "Couldn't open %s port %d "
2865                                "for agents\n",
2866                                device->name, i);
2867                         goto error_agent;
2868                 }
2869         }
2870         return;
2871
2872 error_agent:
2873         if (ib_mad_port_close(device, i))
2874                 printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2875                        device->name, i);
2876
2877 error:
2878         i--;
2879
2880         while (i >= start) {
2881                 if (ib_agent_port_close(device, i))
2882                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2883                                "for agents\n",
2884                                device->name, i);
2885                 if (ib_mad_port_close(device, i))
2886                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2887                                device->name, i);
2888                 i--;
2889         }
2890 }
2891
2892 static void ib_mad_remove_device(struct ib_device *device)
2893 {
2894         int i, num_ports, cur_port;
2895
2896         if (device->node_type == IB_NODE_SWITCH) {
2897                 num_ports = 1;
2898                 cur_port = 0;
2899         } else {
2900                 num_ports = device->phys_port_cnt;
2901                 cur_port = 1;
2902         }
2903         for (i = 0; i < num_ports; i++, cur_port++) {
2904                 if (ib_agent_port_close(device, cur_port))
2905                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2906                                "for agents\n",
2907                                device->name, cur_port);
2908                 if (ib_mad_port_close(device, cur_port))
2909                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2910                                device->name, cur_port);
2911         }
2912 }
2913
2914 static struct ib_client mad_client = {
2915         .name   = "mad",
2916         .add = ib_mad_init_device,
2917         .remove = ib_mad_remove_device
2918 };
2919
2920 static int __init ib_mad_init_module(void)
2921 {
2922         int ret;
2923
2924         spin_lock_init(&ib_mad_port_list_lock);
2925
2926         ib_mad_cache = kmem_cache_create("ib_mad",
2927                                          sizeof(struct ib_mad_private),
2928                                          0,
2929                                          SLAB_HWCACHE_ALIGN,
2930                                          NULL,
2931                                          NULL);
2932         if (!ib_mad_cache) {
2933                 printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2934                 ret = -ENOMEM;
2935                 goto error1;
2936         }
2937
2938         INIT_LIST_HEAD(&ib_mad_port_list);
2939
2940         if (ib_register_client(&mad_client)) {
2941                 printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2942                 ret = -EINVAL;
2943                 goto error2;
2944         }
2945
2946         return 0;
2947
2948 error2:
2949         kmem_cache_destroy(ib_mad_cache);
2950 error1:
2951         return ret;
2952 }
2953
2954 static void __exit ib_mad_cleanup_module(void)
2955 {
2956         ib_unregister_client(&mad_client);
2957
2958         if (kmem_cache_destroy(ib_mad_cache)) {
2959                 printk(KERN_DEBUG PFX "Failed to destroy ib_mad cache\n");
2960         }
2961 }
2962
2963 module_init(ib_mad_init_module);
2964 module_exit(ib_mad_cleanup_module);
2965