xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision 7bcae826)
1 /*
2  * Copyright (c) 2004-2007 Voltaire, Inc. All rights reserved.
3  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
5  * Copyright (c) 2009 HNR Consulting. All rights reserved.
6  * Copyright (c) 2014 Intel Corporation.  All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  */
37 
38 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
39 
40 #include <linux/dma-mapping.h>
41 #include <linux/slab.h>
42 #include <linux/module.h>
43 #include <rdma/ib_cache.h>
44 
45 #include "mad_priv.h"
46 #include "mad_rmpp.h"
47 #include "smi.h"
48 #include "opa_smi.h"
49 #include "agent.h"
50 #include "core_priv.h"
51 
52 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
53 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
54 
55 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
56 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
57 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
58 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
59 
60 static struct list_head ib_mad_port_list;
61 static u32 ib_mad_client_id = 0;
62 
63 /* Port list lock */
64 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
65 
66 /* Forward declarations */
67 static int method_in_use(struct ib_mad_mgmt_method_table **method,
68 			 struct ib_mad_reg_req *mad_reg_req);
69 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
70 static struct ib_mad_agent_private *find_mad_agent(
71 					struct ib_mad_port_private *port_priv,
72 					const struct ib_mad_hdr *mad);
73 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
74 				    struct ib_mad_private *mad);
75 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
76 static void timeout_sends(struct work_struct *work);
77 static void local_completions(struct work_struct *work);
78 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
79 			      struct ib_mad_agent_private *agent_priv,
80 			      u8 mgmt_class);
81 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
82 			   struct ib_mad_agent_private *agent_priv);
83 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
84 			      struct ib_wc *wc);
85 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
86 
87 /*
88  * Returns a ib_mad_port_private structure or NULL for a device/port
89  * Assumes ib_mad_port_list_lock is being held
90  */
91 static inline struct ib_mad_port_private *
92 __ib_get_mad_port(struct ib_device *device, int port_num)
93 {
94 	struct ib_mad_port_private *entry;
95 
96 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
97 		if (entry->device == device && entry->port_num == port_num)
98 			return entry;
99 	}
100 	return NULL;
101 }
102 
103 /*
104  * Wrapper function to return a ib_mad_port_private structure or NULL
105  * for a device/port
106  */
107 static inline struct ib_mad_port_private *
108 ib_get_mad_port(struct ib_device *device, int port_num)
109 {
110 	struct ib_mad_port_private *entry;
111 	unsigned long flags;
112 
113 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
114 	entry = __ib_get_mad_port(device, port_num);
115 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
116 
117 	return entry;
118 }
119 
120 static inline u8 convert_mgmt_class(u8 mgmt_class)
121 {
122 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
123 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
124 		0 : mgmt_class;
125 }
126 
127 static int get_spl_qp_index(enum ib_qp_type qp_type)
128 {
129 	switch (qp_type)
130 	{
131 	case IB_QPT_SMI:
132 		return 0;
133 	case IB_QPT_GSI:
134 		return 1;
135 	default:
136 		return -1;
137 	}
138 }
139 
140 static int vendor_class_index(u8 mgmt_class)
141 {
142 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
143 }
144 
145 static int is_vendor_class(u8 mgmt_class)
146 {
147 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
148 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
149 		return 0;
150 	return 1;
151 }
152 
153 static int is_vendor_oui(char *oui)
154 {
155 	if (oui[0] || oui[1] || oui[2])
156 		return 1;
157 	return 0;
158 }
159 
160 static int is_vendor_method_in_use(
161 		struct ib_mad_mgmt_vendor_class *vendor_class,
162 		struct ib_mad_reg_req *mad_reg_req)
163 {
164 	struct ib_mad_mgmt_method_table *method;
165 	int i;
166 
167 	for (i = 0; i < MAX_MGMT_OUI; i++) {
168 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
169 			method = vendor_class->method_table[i];
170 			if (method) {
171 				if (method_in_use(&method, mad_reg_req))
172 					return 1;
173 				else
174 					break;
175 			}
176 		}
177 	}
178 	return 0;
179 }
180 
181 int ib_response_mad(const struct ib_mad_hdr *hdr)
182 {
183 	return ((hdr->method & IB_MGMT_METHOD_RESP) ||
184 		(hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
185 		((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
186 		 (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
187 }
188 EXPORT_SYMBOL(ib_response_mad);
189 
190 /*
191  * ib_register_mad_agent - Register to send/receive MADs
192  */
193 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
194 					   u8 port_num,
195 					   enum ib_qp_type qp_type,
196 					   struct ib_mad_reg_req *mad_reg_req,
197 					   u8 rmpp_version,
198 					   ib_mad_send_handler send_handler,
199 					   ib_mad_recv_handler recv_handler,
200 					   void *context,
201 					   u32 registration_flags)
202 {
203 	struct ib_mad_port_private *port_priv;
204 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
205 	struct ib_mad_agent_private *mad_agent_priv;
206 	struct ib_mad_reg_req *reg_req = NULL;
207 	struct ib_mad_mgmt_class_table *class;
208 	struct ib_mad_mgmt_vendor_class_table *vendor;
209 	struct ib_mad_mgmt_vendor_class *vendor_class;
210 	struct ib_mad_mgmt_method_table *method;
211 	int ret2, qpn;
212 	unsigned long flags;
213 	u8 mgmt_class, vclass;
214 
215 	/* Validate parameters */
216 	qpn = get_spl_qp_index(qp_type);
217 	if (qpn == -1) {
218 		dev_notice(&device->dev,
219 			   "ib_register_mad_agent: invalid QP Type %d\n",
220 			   qp_type);
221 		goto error1;
222 	}
223 
224 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
225 		dev_notice(&device->dev,
226 			   "ib_register_mad_agent: invalid RMPP Version %u\n",
227 			   rmpp_version);
228 		goto error1;
229 	}
230 
231 	/* Validate MAD registration request if supplied */
232 	if (mad_reg_req) {
233 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
234 			dev_notice(&device->dev,
235 				   "ib_register_mad_agent: invalid Class Version %u\n",
236 				   mad_reg_req->mgmt_class_version);
237 			goto error1;
238 		}
239 		if (!recv_handler) {
240 			dev_notice(&device->dev,
241 				   "ib_register_mad_agent: no recv_handler\n");
242 			goto error1;
243 		}
244 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
245 			/*
246 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
247 			 * one in this range currently allowed
248 			 */
249 			if (mad_reg_req->mgmt_class !=
250 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
251 				dev_notice(&device->dev,
252 					   "ib_register_mad_agent: Invalid Mgmt Class 0x%x\n",
253 					   mad_reg_req->mgmt_class);
254 				goto error1;
255 			}
256 		} else if (mad_reg_req->mgmt_class == 0) {
257 			/*
258 			 * Class 0 is reserved in IBA and is used for
259 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
260 			 */
261 			dev_notice(&device->dev,
262 				   "ib_register_mad_agent: Invalid Mgmt Class 0\n");
263 			goto error1;
264 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
265 			/*
266 			 * If class is in "new" vendor range,
267 			 * ensure supplied OUI is not zero
268 			 */
269 			if (!is_vendor_oui(mad_reg_req->oui)) {
270 				dev_notice(&device->dev,
271 					   "ib_register_mad_agent: No OUI specified for class 0x%x\n",
272 					   mad_reg_req->mgmt_class);
273 				goto error1;
274 			}
275 		}
276 		/* Make sure class supplied is consistent with RMPP */
277 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
278 			if (rmpp_version) {
279 				dev_notice(&device->dev,
280 					   "ib_register_mad_agent: RMPP version for non-RMPP class 0x%x\n",
281 					   mad_reg_req->mgmt_class);
282 				goto error1;
283 			}
284 		}
285 
286 		/* Make sure class supplied is consistent with QP type */
287 		if (qp_type == IB_QPT_SMI) {
288 			if ((mad_reg_req->mgmt_class !=
289 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
290 			    (mad_reg_req->mgmt_class !=
291 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
292 				dev_notice(&device->dev,
293 					   "ib_register_mad_agent: Invalid SM QP type: class 0x%x\n",
294 					   mad_reg_req->mgmt_class);
295 				goto error1;
296 			}
297 		} else {
298 			if ((mad_reg_req->mgmt_class ==
299 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
300 			    (mad_reg_req->mgmt_class ==
301 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
302 				dev_notice(&device->dev,
303 					   "ib_register_mad_agent: Invalid GS QP type: class 0x%x\n",
304 					   mad_reg_req->mgmt_class);
305 				goto error1;
306 			}
307 		}
308 	} else {
309 		/* No registration request supplied */
310 		if (!send_handler)
311 			goto error1;
312 		if (registration_flags & IB_MAD_USER_RMPP)
313 			goto error1;
314 	}
315 
316 	/* Validate device and port */
317 	port_priv = ib_get_mad_port(device, port_num);
318 	if (!port_priv) {
319 		dev_notice(&device->dev,
320 			   "ib_register_mad_agent: Invalid port %d\n",
321 			   port_num);
322 		ret = ERR_PTR(-ENODEV);
323 		goto error1;
324 	}
325 
326 	/* Verify the QP requested is supported.  For example, Ethernet devices
327 	 * will not have QP0 */
328 	if (!port_priv->qp_info[qpn].qp) {
329 		dev_notice(&device->dev,
330 			   "ib_register_mad_agent: QP %d not supported\n", qpn);
331 		ret = ERR_PTR(-EPROTONOSUPPORT);
332 		goto error1;
333 	}
334 
335 	/* Allocate structures */
336 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
337 	if (!mad_agent_priv) {
338 		ret = ERR_PTR(-ENOMEM);
339 		goto error1;
340 	}
341 
342 	if (mad_reg_req) {
343 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
344 		if (!reg_req) {
345 			ret = ERR_PTR(-ENOMEM);
346 			goto error3;
347 		}
348 	}
349 
350 	/* Now, fill in the various structures */
351 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
352 	mad_agent_priv->reg_req = reg_req;
353 	mad_agent_priv->agent.rmpp_version = rmpp_version;
354 	mad_agent_priv->agent.device = device;
355 	mad_agent_priv->agent.recv_handler = recv_handler;
356 	mad_agent_priv->agent.send_handler = send_handler;
357 	mad_agent_priv->agent.context = context;
358 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
359 	mad_agent_priv->agent.port_num = port_num;
360 	mad_agent_priv->agent.flags = registration_flags;
361 	spin_lock_init(&mad_agent_priv->lock);
362 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
363 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
364 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
365 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
366 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
367 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
368 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
369 	atomic_set(&mad_agent_priv->refcount, 1);
370 	init_completion(&mad_agent_priv->comp);
371 
372 	spin_lock_irqsave(&port_priv->reg_lock, flags);
373 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
374 
375 	/*
376 	 * Make sure MAD registration (if supplied)
377 	 * is non overlapping with any existing ones
378 	 */
379 	if (mad_reg_req) {
380 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
381 		if (!is_vendor_class(mgmt_class)) {
382 			class = port_priv->version[mad_reg_req->
383 						   mgmt_class_version].class;
384 			if (class) {
385 				method = class->method_table[mgmt_class];
386 				if (method) {
387 					if (method_in_use(&method,
388 							   mad_reg_req))
389 						goto error4;
390 				}
391 			}
392 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
393 						  mgmt_class);
394 		} else {
395 			/* "New" vendor class range */
396 			vendor = port_priv->version[mad_reg_req->
397 						    mgmt_class_version].vendor;
398 			if (vendor) {
399 				vclass = vendor_class_index(mgmt_class);
400 				vendor_class = vendor->vendor_class[vclass];
401 				if (vendor_class) {
402 					if (is_vendor_method_in_use(
403 							vendor_class,
404 							mad_reg_req))
405 						goto error4;
406 				}
407 			}
408 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
409 		}
410 		if (ret2) {
411 			ret = ERR_PTR(ret2);
412 			goto error4;
413 		}
414 	}
415 
416 	/* Add mad agent into port's agent list */
417 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
418 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
419 
420 	return &mad_agent_priv->agent;
421 
422 error4:
423 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
424 	kfree(reg_req);
425 error3:
426 	kfree(mad_agent_priv);
427 error1:
428 	return ret;
429 }
430 EXPORT_SYMBOL(ib_register_mad_agent);
431 
432 static inline int is_snooping_sends(int mad_snoop_flags)
433 {
434 	return (mad_snoop_flags &
435 		(/*IB_MAD_SNOOP_POSTED_SENDS |
436 		 IB_MAD_SNOOP_RMPP_SENDS |*/
437 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
438 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
439 }
440 
441 static inline int is_snooping_recvs(int mad_snoop_flags)
442 {
443 	return (mad_snoop_flags &
444 		(IB_MAD_SNOOP_RECVS /*|
445 		 IB_MAD_SNOOP_RMPP_RECVS*/));
446 }
447 
448 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
449 				struct ib_mad_snoop_private *mad_snoop_priv)
450 {
451 	struct ib_mad_snoop_private **new_snoop_table;
452 	unsigned long flags;
453 	int i;
454 
455 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
456 	/* Check for empty slot in array. */
457 	for (i = 0; i < qp_info->snoop_table_size; i++)
458 		if (!qp_info->snoop_table[i])
459 			break;
460 
461 	if (i == qp_info->snoop_table_size) {
462 		/* Grow table. */
463 		new_snoop_table = krealloc(qp_info->snoop_table,
464 					   sizeof mad_snoop_priv *
465 					   (qp_info->snoop_table_size + 1),
466 					   GFP_ATOMIC);
467 		if (!new_snoop_table) {
468 			i = -ENOMEM;
469 			goto out;
470 		}
471 
472 		qp_info->snoop_table = new_snoop_table;
473 		qp_info->snoop_table_size++;
474 	}
475 	qp_info->snoop_table[i] = mad_snoop_priv;
476 	atomic_inc(&qp_info->snoop_count);
477 out:
478 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
479 	return i;
480 }
481 
482 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
483 					   u8 port_num,
484 					   enum ib_qp_type qp_type,
485 					   int mad_snoop_flags,
486 					   ib_mad_snoop_handler snoop_handler,
487 					   ib_mad_recv_handler recv_handler,
488 					   void *context)
489 {
490 	struct ib_mad_port_private *port_priv;
491 	struct ib_mad_agent *ret;
492 	struct ib_mad_snoop_private *mad_snoop_priv;
493 	int qpn;
494 
495 	/* Validate parameters */
496 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
497 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
498 		ret = ERR_PTR(-EINVAL);
499 		goto error1;
500 	}
501 	qpn = get_spl_qp_index(qp_type);
502 	if (qpn == -1) {
503 		ret = ERR_PTR(-EINVAL);
504 		goto error1;
505 	}
506 	port_priv = ib_get_mad_port(device, port_num);
507 	if (!port_priv) {
508 		ret = ERR_PTR(-ENODEV);
509 		goto error1;
510 	}
511 	/* Allocate structures */
512 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
513 	if (!mad_snoop_priv) {
514 		ret = ERR_PTR(-ENOMEM);
515 		goto error1;
516 	}
517 
518 	/* Now, fill in the various structures */
519 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
520 	mad_snoop_priv->agent.device = device;
521 	mad_snoop_priv->agent.recv_handler = recv_handler;
522 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
523 	mad_snoop_priv->agent.context = context;
524 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
525 	mad_snoop_priv->agent.port_num = port_num;
526 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
527 	init_completion(&mad_snoop_priv->comp);
528 	mad_snoop_priv->snoop_index = register_snoop_agent(
529 						&port_priv->qp_info[qpn],
530 						mad_snoop_priv);
531 	if (mad_snoop_priv->snoop_index < 0) {
532 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
533 		goto error2;
534 	}
535 
536 	atomic_set(&mad_snoop_priv->refcount, 1);
537 	return &mad_snoop_priv->agent;
538 
539 error2:
540 	kfree(mad_snoop_priv);
541 error1:
542 	return ret;
543 }
544 EXPORT_SYMBOL(ib_register_mad_snoop);
545 
546 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
547 {
548 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
549 		complete(&mad_agent_priv->comp);
550 }
551 
552 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
553 {
554 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
555 		complete(&mad_snoop_priv->comp);
556 }
557 
558 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
559 {
560 	struct ib_mad_port_private *port_priv;
561 	unsigned long flags;
562 
563 	/* Note that we could still be handling received MADs */
564 
565 	/*
566 	 * Canceling all sends results in dropping received response
567 	 * MADs, preventing us from queuing additional work
568 	 */
569 	cancel_mads(mad_agent_priv);
570 	port_priv = mad_agent_priv->qp_info->port_priv;
571 	cancel_delayed_work(&mad_agent_priv->timed_work);
572 
573 	spin_lock_irqsave(&port_priv->reg_lock, flags);
574 	remove_mad_reg_req(mad_agent_priv);
575 	list_del(&mad_agent_priv->agent_list);
576 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
577 
578 	flush_workqueue(port_priv->wq);
579 	ib_cancel_rmpp_recvs(mad_agent_priv);
580 
581 	deref_mad_agent(mad_agent_priv);
582 	wait_for_completion(&mad_agent_priv->comp);
583 
584 	kfree(mad_agent_priv->reg_req);
585 	kfree(mad_agent_priv);
586 }
587 
588 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
589 {
590 	struct ib_mad_qp_info *qp_info;
591 	unsigned long flags;
592 
593 	qp_info = mad_snoop_priv->qp_info;
594 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
595 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
596 	atomic_dec(&qp_info->snoop_count);
597 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
598 
599 	deref_snoop_agent(mad_snoop_priv);
600 	wait_for_completion(&mad_snoop_priv->comp);
601 
602 	kfree(mad_snoop_priv);
603 }
604 
605 /*
606  * ib_unregister_mad_agent - Unregisters a client from using MAD services
607  */
608 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
609 {
610 	struct ib_mad_agent_private *mad_agent_priv;
611 	struct ib_mad_snoop_private *mad_snoop_priv;
612 
613 	/* If the TID is zero, the agent can only snoop. */
614 	if (mad_agent->hi_tid) {
615 		mad_agent_priv = container_of(mad_agent,
616 					      struct ib_mad_agent_private,
617 					      agent);
618 		unregister_mad_agent(mad_agent_priv);
619 	} else {
620 		mad_snoop_priv = container_of(mad_agent,
621 					      struct ib_mad_snoop_private,
622 					      agent);
623 		unregister_mad_snoop(mad_snoop_priv);
624 	}
625 	return 0;
626 }
627 EXPORT_SYMBOL(ib_unregister_mad_agent);
628 
629 static void dequeue_mad(struct ib_mad_list_head *mad_list)
630 {
631 	struct ib_mad_queue *mad_queue;
632 	unsigned long flags;
633 
634 	BUG_ON(!mad_list->mad_queue);
635 	mad_queue = mad_list->mad_queue;
636 	spin_lock_irqsave(&mad_queue->lock, flags);
637 	list_del(&mad_list->list);
638 	mad_queue->count--;
639 	spin_unlock_irqrestore(&mad_queue->lock, flags);
640 }
641 
642 static void snoop_send(struct ib_mad_qp_info *qp_info,
643 		       struct ib_mad_send_buf *send_buf,
644 		       struct ib_mad_send_wc *mad_send_wc,
645 		       int mad_snoop_flags)
646 {
647 	struct ib_mad_snoop_private *mad_snoop_priv;
648 	unsigned long flags;
649 	int i;
650 
651 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
652 	for (i = 0; i < qp_info->snoop_table_size; i++) {
653 		mad_snoop_priv = qp_info->snoop_table[i];
654 		if (!mad_snoop_priv ||
655 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
656 			continue;
657 
658 		atomic_inc(&mad_snoop_priv->refcount);
659 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
660 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
661 						    send_buf, mad_send_wc);
662 		deref_snoop_agent(mad_snoop_priv);
663 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
664 	}
665 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
666 }
667 
668 static void snoop_recv(struct ib_mad_qp_info *qp_info,
669 		       struct ib_mad_recv_wc *mad_recv_wc,
670 		       int mad_snoop_flags)
671 {
672 	struct ib_mad_snoop_private *mad_snoop_priv;
673 	unsigned long flags;
674 	int i;
675 
676 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
677 	for (i = 0; i < qp_info->snoop_table_size; i++) {
678 		mad_snoop_priv = qp_info->snoop_table[i];
679 		if (!mad_snoop_priv ||
680 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
681 			continue;
682 
683 		atomic_inc(&mad_snoop_priv->refcount);
684 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
685 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
686 						   mad_recv_wc);
687 		deref_snoop_agent(mad_snoop_priv);
688 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
689 	}
690 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
691 }
692 
693 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
694 		u16 pkey_index, u8 port_num, struct ib_wc *wc)
695 {
696 	memset(wc, 0, sizeof *wc);
697 	wc->wr_cqe = cqe;
698 	wc->status = IB_WC_SUCCESS;
699 	wc->opcode = IB_WC_RECV;
700 	wc->pkey_index = pkey_index;
701 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
702 	wc->src_qp = IB_QP0;
703 	wc->qp = qp;
704 	wc->slid = slid;
705 	wc->sl = 0;
706 	wc->dlid_path_bits = 0;
707 	wc->port_num = port_num;
708 }
709 
710 static size_t mad_priv_size(const struct ib_mad_private *mp)
711 {
712 	return sizeof(struct ib_mad_private) + mp->mad_size;
713 }
714 
715 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
716 {
717 	size_t size = sizeof(struct ib_mad_private) + mad_size;
718 	struct ib_mad_private *ret = kzalloc(size, flags);
719 
720 	if (ret)
721 		ret->mad_size = mad_size;
722 
723 	return ret;
724 }
725 
726 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
727 {
728 	return rdma_max_mad_size(port_priv->device, port_priv->port_num);
729 }
730 
731 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
732 {
733 	return sizeof(struct ib_grh) + mp->mad_size;
734 }
735 
736 /*
737  * Return 0 if SMP is to be sent
738  * Return 1 if SMP was consumed locally (whether or not solicited)
739  * Return < 0 if error
740  */
741 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
742 				  struct ib_mad_send_wr_private *mad_send_wr)
743 {
744 	int ret = 0;
745 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
746 	struct opa_smp *opa_smp = (struct opa_smp *)smp;
747 	unsigned long flags;
748 	struct ib_mad_local_private *local;
749 	struct ib_mad_private *mad_priv;
750 	struct ib_mad_port_private *port_priv;
751 	struct ib_mad_agent_private *recv_mad_agent = NULL;
752 	struct ib_device *device = mad_agent_priv->agent.device;
753 	u8 port_num;
754 	struct ib_wc mad_wc;
755 	struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
756 	size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
757 	u16 out_mad_pkey_index = 0;
758 	u16 drslid;
759 	bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
760 				    mad_agent_priv->qp_info->port_priv->port_num);
761 
762 	if (rdma_cap_ib_switch(device) &&
763 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
764 		port_num = send_wr->port_num;
765 	else
766 		port_num = mad_agent_priv->agent.port_num;
767 
768 	/*
769 	 * Directed route handling starts if the initial LID routed part of
770 	 * a request or the ending LID routed part of a response is empty.
771 	 * If we are at the start of the LID routed part, don't update the
772 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
773 	 */
774 	if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
775 		u32 opa_drslid;
776 
777 		if ((opa_get_smp_direction(opa_smp)
778 		     ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
779 		     OPA_LID_PERMISSIVE &&
780 		     opa_smi_handle_dr_smp_send(opa_smp,
781 						rdma_cap_ib_switch(device),
782 						port_num) == IB_SMI_DISCARD) {
783 			ret = -EINVAL;
784 			dev_err(&device->dev, "OPA Invalid directed route\n");
785 			goto out;
786 		}
787 		opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
788 		if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
789 		    opa_drslid & 0xffff0000) {
790 			ret = -EINVAL;
791 			dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
792 			       opa_drslid);
793 			goto out;
794 		}
795 		drslid = (u16)(opa_drslid & 0x0000ffff);
796 
797 		/* Check to post send on QP or process locally */
798 		if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
799 		    opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
800 			goto out;
801 	} else {
802 		if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
803 		     IB_LID_PERMISSIVE &&
804 		     smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
805 		     IB_SMI_DISCARD) {
806 			ret = -EINVAL;
807 			dev_err(&device->dev, "Invalid directed route\n");
808 			goto out;
809 		}
810 		drslid = be16_to_cpu(smp->dr_slid);
811 
812 		/* Check to post send on QP or process locally */
813 		if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
814 		    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
815 			goto out;
816 	}
817 
818 	local = kmalloc(sizeof *local, GFP_ATOMIC);
819 	if (!local) {
820 		ret = -ENOMEM;
821 		goto out;
822 	}
823 	local->mad_priv = NULL;
824 	local->recv_mad_agent = NULL;
825 	mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
826 	if (!mad_priv) {
827 		ret = -ENOMEM;
828 		kfree(local);
829 		goto out;
830 	}
831 
832 	build_smp_wc(mad_agent_priv->agent.qp,
833 		     send_wr->wr.wr_cqe, drslid,
834 		     send_wr->pkey_index,
835 		     send_wr->port_num, &mad_wc);
836 
837 	if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
838 		mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
839 					+ mad_send_wr->send_buf.data_len
840 					+ sizeof(struct ib_grh);
841 	}
842 
843 	/* No GRH for DR SMP */
844 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
845 				  (const struct ib_mad_hdr *)smp, mad_size,
846 				  (struct ib_mad_hdr *)mad_priv->mad,
847 				  &mad_size, &out_mad_pkey_index);
848 	switch (ret)
849 	{
850 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
851 		if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
852 		    mad_agent_priv->agent.recv_handler) {
853 			local->mad_priv = mad_priv;
854 			local->recv_mad_agent = mad_agent_priv;
855 			/*
856 			 * Reference MAD agent until receive
857 			 * side of local completion handled
858 			 */
859 			atomic_inc(&mad_agent_priv->refcount);
860 		} else
861 			kfree(mad_priv);
862 		break;
863 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
864 		kfree(mad_priv);
865 		break;
866 	case IB_MAD_RESULT_SUCCESS:
867 		/* Treat like an incoming receive MAD */
868 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
869 					    mad_agent_priv->agent.port_num);
870 		if (port_priv) {
871 			memcpy(mad_priv->mad, smp, mad_priv->mad_size);
872 			recv_mad_agent = find_mad_agent(port_priv,
873 						        (const struct ib_mad_hdr *)mad_priv->mad);
874 		}
875 		if (!port_priv || !recv_mad_agent) {
876 			/*
877 			 * No receiving agent so drop packet and
878 			 * generate send completion.
879 			 */
880 			kfree(mad_priv);
881 			break;
882 		}
883 		local->mad_priv = mad_priv;
884 		local->recv_mad_agent = recv_mad_agent;
885 		break;
886 	default:
887 		kfree(mad_priv);
888 		kfree(local);
889 		ret = -EINVAL;
890 		goto out;
891 	}
892 
893 	local->mad_send_wr = mad_send_wr;
894 	if (opa) {
895 		local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
896 		local->return_wc_byte_len = mad_size;
897 	}
898 	/* Reference MAD agent until send side of local completion handled */
899 	atomic_inc(&mad_agent_priv->refcount);
900 	/* Queue local completion to local list */
901 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
902 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
903 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
904 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
905 		   &mad_agent_priv->local_work);
906 	ret = 1;
907 out:
908 	return ret;
909 }
910 
911 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
912 {
913 	int seg_size, pad;
914 
915 	seg_size = mad_size - hdr_len;
916 	if (data_len && seg_size) {
917 		pad = seg_size - data_len % seg_size;
918 		return pad == seg_size ? 0 : pad;
919 	} else
920 		return seg_size;
921 }
922 
923 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
924 {
925 	struct ib_rmpp_segment *s, *t;
926 
927 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
928 		list_del(&s->list);
929 		kfree(s);
930 	}
931 }
932 
933 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
934 				size_t mad_size, gfp_t gfp_mask)
935 {
936 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
937 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
938 	struct ib_rmpp_segment *seg = NULL;
939 	int left, seg_size, pad;
940 
941 	send_buf->seg_size = mad_size - send_buf->hdr_len;
942 	send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
943 	seg_size = send_buf->seg_size;
944 	pad = send_wr->pad;
945 
946 	/* Allocate data segments. */
947 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
948 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
949 		if (!seg) {
950 			free_send_rmpp_list(send_wr);
951 			return -ENOMEM;
952 		}
953 		seg->num = ++send_buf->seg_count;
954 		list_add_tail(&seg->list, &send_wr->rmpp_list);
955 	}
956 
957 	/* Zero any padding */
958 	if (pad)
959 		memset(seg->data + seg_size - pad, 0, pad);
960 
961 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
962 					  agent.rmpp_version;
963 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
964 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
965 
966 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
967 					struct ib_rmpp_segment, list);
968 	send_wr->last_ack_seg = send_wr->cur_seg;
969 	return 0;
970 }
971 
972 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
973 {
974 	return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
975 }
976 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
977 
978 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
979 					    u32 remote_qpn, u16 pkey_index,
980 					    int rmpp_active,
981 					    int hdr_len, int data_len,
982 					    gfp_t gfp_mask,
983 					    u8 base_version)
984 {
985 	struct ib_mad_agent_private *mad_agent_priv;
986 	struct ib_mad_send_wr_private *mad_send_wr;
987 	int pad, message_size, ret, size;
988 	void *buf;
989 	size_t mad_size;
990 	bool opa;
991 
992 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
993 				      agent);
994 
995 	opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
996 
997 	if (opa && base_version == OPA_MGMT_BASE_VERSION)
998 		mad_size = sizeof(struct opa_mad);
999 	else
1000 		mad_size = sizeof(struct ib_mad);
1001 
1002 	pad = get_pad_size(hdr_len, data_len, mad_size);
1003 	message_size = hdr_len + data_len + pad;
1004 
1005 	if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1006 		if (!rmpp_active && message_size > mad_size)
1007 			return ERR_PTR(-EINVAL);
1008 	} else
1009 		if (rmpp_active || message_size > mad_size)
1010 			return ERR_PTR(-EINVAL);
1011 
1012 	size = rmpp_active ? hdr_len : mad_size;
1013 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1014 	if (!buf)
1015 		return ERR_PTR(-ENOMEM);
1016 
1017 	mad_send_wr = buf + size;
1018 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1019 	mad_send_wr->send_buf.mad = buf;
1020 	mad_send_wr->send_buf.hdr_len = hdr_len;
1021 	mad_send_wr->send_buf.data_len = data_len;
1022 	mad_send_wr->pad = pad;
1023 
1024 	mad_send_wr->mad_agent_priv = mad_agent_priv;
1025 	mad_send_wr->sg_list[0].length = hdr_len;
1026 	mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1027 
1028 	/* OPA MADs don't have to be the full 2048 bytes */
1029 	if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1030 	    data_len < mad_size - hdr_len)
1031 		mad_send_wr->sg_list[1].length = data_len;
1032 	else
1033 		mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1034 
1035 	mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1036 
1037 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1038 
1039 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1040 	mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1041 	mad_send_wr->send_wr.wr.num_sge = 2;
1042 	mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1043 	mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1044 	mad_send_wr->send_wr.remote_qpn = remote_qpn;
1045 	mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1046 	mad_send_wr->send_wr.pkey_index = pkey_index;
1047 
1048 	if (rmpp_active) {
1049 		ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1050 		if (ret) {
1051 			kfree(buf);
1052 			return ERR_PTR(ret);
1053 		}
1054 	}
1055 
1056 	mad_send_wr->send_buf.mad_agent = mad_agent;
1057 	atomic_inc(&mad_agent_priv->refcount);
1058 	return &mad_send_wr->send_buf;
1059 }
1060 EXPORT_SYMBOL(ib_create_send_mad);
1061 
1062 int ib_get_mad_data_offset(u8 mgmt_class)
1063 {
1064 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1065 		return IB_MGMT_SA_HDR;
1066 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1067 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1068 		 (mgmt_class == IB_MGMT_CLASS_BIS))
1069 		return IB_MGMT_DEVICE_HDR;
1070 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1071 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1072 		return IB_MGMT_VENDOR_HDR;
1073 	else
1074 		return IB_MGMT_MAD_HDR;
1075 }
1076 EXPORT_SYMBOL(ib_get_mad_data_offset);
1077 
1078 int ib_is_mad_class_rmpp(u8 mgmt_class)
1079 {
1080 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1081 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1082 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1083 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
1084 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1085 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1086 		return 1;
1087 	return 0;
1088 }
1089 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1090 
1091 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1092 {
1093 	struct ib_mad_send_wr_private *mad_send_wr;
1094 	struct list_head *list;
1095 
1096 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1097 				   send_buf);
1098 	list = &mad_send_wr->cur_seg->list;
1099 
1100 	if (mad_send_wr->cur_seg->num < seg_num) {
1101 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
1102 			if (mad_send_wr->cur_seg->num == seg_num)
1103 				break;
1104 	} else if (mad_send_wr->cur_seg->num > seg_num) {
1105 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1106 			if (mad_send_wr->cur_seg->num == seg_num)
1107 				break;
1108 	}
1109 	return mad_send_wr->cur_seg->data;
1110 }
1111 EXPORT_SYMBOL(ib_get_rmpp_segment);
1112 
1113 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1114 {
1115 	if (mad_send_wr->send_buf.seg_count)
1116 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1117 					   mad_send_wr->seg_num);
1118 	else
1119 		return mad_send_wr->send_buf.mad +
1120 		       mad_send_wr->send_buf.hdr_len;
1121 }
1122 
1123 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1124 {
1125 	struct ib_mad_agent_private *mad_agent_priv;
1126 	struct ib_mad_send_wr_private *mad_send_wr;
1127 
1128 	mad_agent_priv = container_of(send_buf->mad_agent,
1129 				      struct ib_mad_agent_private, agent);
1130 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1131 				   send_buf);
1132 
1133 	free_send_rmpp_list(mad_send_wr);
1134 	kfree(send_buf->mad);
1135 	deref_mad_agent(mad_agent_priv);
1136 }
1137 EXPORT_SYMBOL(ib_free_send_mad);
1138 
1139 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1140 {
1141 	struct ib_mad_qp_info *qp_info;
1142 	struct list_head *list;
1143 	struct ib_send_wr *bad_send_wr;
1144 	struct ib_mad_agent *mad_agent;
1145 	struct ib_sge *sge;
1146 	unsigned long flags;
1147 	int ret;
1148 
1149 	/* Set WR ID to find mad_send_wr upon completion */
1150 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1151 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1152 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1153 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1154 
1155 	mad_agent = mad_send_wr->send_buf.mad_agent;
1156 	sge = mad_send_wr->sg_list;
1157 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1158 					mad_send_wr->send_buf.mad,
1159 					sge[0].length,
1160 					DMA_TO_DEVICE);
1161 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1162 		return -ENOMEM;
1163 
1164 	mad_send_wr->header_mapping = sge[0].addr;
1165 
1166 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1167 					ib_get_payload(mad_send_wr),
1168 					sge[1].length,
1169 					DMA_TO_DEVICE);
1170 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1171 		ib_dma_unmap_single(mad_agent->device,
1172 				    mad_send_wr->header_mapping,
1173 				    sge[0].length, DMA_TO_DEVICE);
1174 		return -ENOMEM;
1175 	}
1176 	mad_send_wr->payload_mapping = sge[1].addr;
1177 
1178 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1179 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1180 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1181 				   &bad_send_wr);
1182 		list = &qp_info->send_queue.list;
1183 	} else {
1184 		ret = 0;
1185 		list = &qp_info->overflow_list;
1186 	}
1187 
1188 	if (!ret) {
1189 		qp_info->send_queue.count++;
1190 		list_add_tail(&mad_send_wr->mad_list.list, list);
1191 	}
1192 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1193 	if (ret) {
1194 		ib_dma_unmap_single(mad_agent->device,
1195 				    mad_send_wr->header_mapping,
1196 				    sge[0].length, DMA_TO_DEVICE);
1197 		ib_dma_unmap_single(mad_agent->device,
1198 				    mad_send_wr->payload_mapping,
1199 				    sge[1].length, DMA_TO_DEVICE);
1200 	}
1201 	return ret;
1202 }
1203 
1204 /*
1205  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1206  *  with the registered client
1207  */
1208 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1209 		     struct ib_mad_send_buf **bad_send_buf)
1210 {
1211 	struct ib_mad_agent_private *mad_agent_priv;
1212 	struct ib_mad_send_buf *next_send_buf;
1213 	struct ib_mad_send_wr_private *mad_send_wr;
1214 	unsigned long flags;
1215 	int ret = -EINVAL;
1216 
1217 	/* Walk list of send WRs and post each on send list */
1218 	for (; send_buf; send_buf = next_send_buf) {
1219 
1220 		mad_send_wr = container_of(send_buf,
1221 					   struct ib_mad_send_wr_private,
1222 					   send_buf);
1223 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1224 
1225 		if (!send_buf->mad_agent->send_handler ||
1226 		    (send_buf->timeout_ms &&
1227 		     !send_buf->mad_agent->recv_handler)) {
1228 			ret = -EINVAL;
1229 			goto error;
1230 		}
1231 
1232 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1233 			if (mad_agent_priv->agent.rmpp_version) {
1234 				ret = -EINVAL;
1235 				goto error;
1236 			}
1237 		}
1238 
1239 		/*
1240 		 * Save pointer to next work request to post in case the
1241 		 * current one completes, and the user modifies the work
1242 		 * request associated with the completion
1243 		 */
1244 		next_send_buf = send_buf->next;
1245 		mad_send_wr->send_wr.ah = send_buf->ah;
1246 
1247 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1248 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1249 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1250 						     mad_send_wr);
1251 			if (ret < 0)		/* error */
1252 				goto error;
1253 			else if (ret == 1)	/* locally consumed */
1254 				continue;
1255 		}
1256 
1257 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1258 		/* Timeout will be updated after send completes */
1259 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1260 		mad_send_wr->max_retries = send_buf->retries;
1261 		mad_send_wr->retries_left = send_buf->retries;
1262 		send_buf->retries = 0;
1263 		/* Reference for work request to QP + response */
1264 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1265 		mad_send_wr->status = IB_WC_SUCCESS;
1266 
1267 		/* Reference MAD agent until send completes */
1268 		atomic_inc(&mad_agent_priv->refcount);
1269 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1270 		list_add_tail(&mad_send_wr->agent_list,
1271 			      &mad_agent_priv->send_list);
1272 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1273 
1274 		if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1275 			ret = ib_send_rmpp_mad(mad_send_wr);
1276 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1277 				ret = ib_send_mad(mad_send_wr);
1278 		} else
1279 			ret = ib_send_mad(mad_send_wr);
1280 		if (ret < 0) {
1281 			/* Fail send request */
1282 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1283 			list_del(&mad_send_wr->agent_list);
1284 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1285 			atomic_dec(&mad_agent_priv->refcount);
1286 			goto error;
1287 		}
1288 	}
1289 	return 0;
1290 error:
1291 	if (bad_send_buf)
1292 		*bad_send_buf = send_buf;
1293 	return ret;
1294 }
1295 EXPORT_SYMBOL(ib_post_send_mad);
1296 
1297 /*
1298  * ib_free_recv_mad - Returns data buffers used to receive
1299  *  a MAD to the access layer
1300  */
1301 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1302 {
1303 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1304 	struct ib_mad_private_header *mad_priv_hdr;
1305 	struct ib_mad_private *priv;
1306 	struct list_head free_list;
1307 
1308 	INIT_LIST_HEAD(&free_list);
1309 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1310 
1311 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1312 					&free_list, list) {
1313 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1314 					   recv_buf);
1315 		mad_priv_hdr = container_of(mad_recv_wc,
1316 					    struct ib_mad_private_header,
1317 					    recv_wc);
1318 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1319 				    header);
1320 		kfree(priv);
1321 	}
1322 }
1323 EXPORT_SYMBOL(ib_free_recv_mad);
1324 
1325 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1326 					u8 rmpp_version,
1327 					ib_mad_send_handler send_handler,
1328 					ib_mad_recv_handler recv_handler,
1329 					void *context)
1330 {
1331 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1332 }
1333 EXPORT_SYMBOL(ib_redirect_mad_qp);
1334 
1335 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1336 		      struct ib_wc *wc)
1337 {
1338 	dev_err(&mad_agent->device->dev,
1339 		"ib_process_mad_wc() not implemented yet\n");
1340 	return 0;
1341 }
1342 EXPORT_SYMBOL(ib_process_mad_wc);
1343 
1344 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1345 			 struct ib_mad_reg_req *mad_reg_req)
1346 {
1347 	int i;
1348 
1349 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1350 		if ((*method)->agent[i]) {
1351 			pr_err("Method %d already in use\n", i);
1352 			return -EINVAL;
1353 		}
1354 	}
1355 	return 0;
1356 }
1357 
1358 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1359 {
1360 	/* Allocate management method table */
1361 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1362 	return (*method) ? 0 : (-ENOMEM);
1363 }
1364 
1365 /*
1366  * Check to see if there are any methods still in use
1367  */
1368 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1369 {
1370 	int i;
1371 
1372 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1373 		if (method->agent[i])
1374 			return 1;
1375 	return 0;
1376 }
1377 
1378 /*
1379  * Check to see if there are any method tables for this class still in use
1380  */
1381 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1382 {
1383 	int i;
1384 
1385 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1386 		if (class->method_table[i])
1387 			return 1;
1388 	return 0;
1389 }
1390 
1391 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1392 {
1393 	int i;
1394 
1395 	for (i = 0; i < MAX_MGMT_OUI; i++)
1396 		if (vendor_class->method_table[i])
1397 			return 1;
1398 	return 0;
1399 }
1400 
1401 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1402 			   const char *oui)
1403 {
1404 	int i;
1405 
1406 	for (i = 0; i < MAX_MGMT_OUI; i++)
1407 		/* Is there matching OUI for this vendor class ? */
1408 		if (!memcmp(vendor_class->oui[i], oui, 3))
1409 			return i;
1410 
1411 	return -1;
1412 }
1413 
1414 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1415 {
1416 	int i;
1417 
1418 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1419 		if (vendor->vendor_class[i])
1420 			return 1;
1421 
1422 	return 0;
1423 }
1424 
1425 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1426 				     struct ib_mad_agent_private *agent)
1427 {
1428 	int i;
1429 
1430 	/* Remove any methods for this mad agent */
1431 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1432 		if (method->agent[i] == agent) {
1433 			method->agent[i] = NULL;
1434 		}
1435 	}
1436 }
1437 
1438 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1439 			      struct ib_mad_agent_private *agent_priv,
1440 			      u8 mgmt_class)
1441 {
1442 	struct ib_mad_port_private *port_priv;
1443 	struct ib_mad_mgmt_class_table **class;
1444 	struct ib_mad_mgmt_method_table **method;
1445 	int i, ret;
1446 
1447 	port_priv = agent_priv->qp_info->port_priv;
1448 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1449 	if (!*class) {
1450 		/* Allocate management class table for "new" class version */
1451 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1452 		if (!*class) {
1453 			ret = -ENOMEM;
1454 			goto error1;
1455 		}
1456 
1457 		/* Allocate method table for this management class */
1458 		method = &(*class)->method_table[mgmt_class];
1459 		if ((ret = allocate_method_table(method)))
1460 			goto error2;
1461 	} else {
1462 		method = &(*class)->method_table[mgmt_class];
1463 		if (!*method) {
1464 			/* Allocate method table for this management class */
1465 			if ((ret = allocate_method_table(method)))
1466 				goto error1;
1467 		}
1468 	}
1469 
1470 	/* Now, make sure methods are not already in use */
1471 	if (method_in_use(method, mad_reg_req))
1472 		goto error3;
1473 
1474 	/* Finally, add in methods being registered */
1475 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1476 		(*method)->agent[i] = agent_priv;
1477 
1478 	return 0;
1479 
1480 error3:
1481 	/* Remove any methods for this mad agent */
1482 	remove_methods_mad_agent(*method, agent_priv);
1483 	/* Now, check to see if there are any methods in use */
1484 	if (!check_method_table(*method)) {
1485 		/* If not, release management method table */
1486 		kfree(*method);
1487 		*method = NULL;
1488 	}
1489 	ret = -EINVAL;
1490 	goto error1;
1491 error2:
1492 	kfree(*class);
1493 	*class = NULL;
1494 error1:
1495 	return ret;
1496 }
1497 
1498 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1499 			   struct ib_mad_agent_private *agent_priv)
1500 {
1501 	struct ib_mad_port_private *port_priv;
1502 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1503 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1504 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1505 	struct ib_mad_mgmt_method_table **method;
1506 	int i, ret = -ENOMEM;
1507 	u8 vclass;
1508 
1509 	/* "New" vendor (with OUI) class */
1510 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1511 	port_priv = agent_priv->qp_info->port_priv;
1512 	vendor_table = &port_priv->version[
1513 				mad_reg_req->mgmt_class_version].vendor;
1514 	if (!*vendor_table) {
1515 		/* Allocate mgmt vendor class table for "new" class version */
1516 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1517 		if (!vendor)
1518 			goto error1;
1519 
1520 		*vendor_table = vendor;
1521 	}
1522 	if (!(*vendor_table)->vendor_class[vclass]) {
1523 		/* Allocate table for this management vendor class */
1524 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1525 		if (!vendor_class)
1526 			goto error2;
1527 
1528 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1529 	}
1530 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1531 		/* Is there matching OUI for this vendor class ? */
1532 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1533 			    mad_reg_req->oui, 3)) {
1534 			method = &(*vendor_table)->vendor_class[
1535 						vclass]->method_table[i];
1536 			BUG_ON(!*method);
1537 			goto check_in_use;
1538 		}
1539 	}
1540 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1541 		/* OUI slot available ? */
1542 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1543 				vclass]->oui[i])) {
1544 			method = &(*vendor_table)->vendor_class[
1545 				vclass]->method_table[i];
1546 			BUG_ON(*method);
1547 			/* Allocate method table for this OUI */
1548 			if ((ret = allocate_method_table(method)))
1549 				goto error3;
1550 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1551 			       mad_reg_req->oui, 3);
1552 			goto check_in_use;
1553 		}
1554 	}
1555 	dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1556 	goto error3;
1557 
1558 check_in_use:
1559 	/* Now, make sure methods are not already in use */
1560 	if (method_in_use(method, mad_reg_req))
1561 		goto error4;
1562 
1563 	/* Finally, add in methods being registered */
1564 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1565 		(*method)->agent[i] = agent_priv;
1566 
1567 	return 0;
1568 
1569 error4:
1570 	/* Remove any methods for this mad agent */
1571 	remove_methods_mad_agent(*method, agent_priv);
1572 	/* Now, check to see if there are any methods in use */
1573 	if (!check_method_table(*method)) {
1574 		/* If not, release management method table */
1575 		kfree(*method);
1576 		*method = NULL;
1577 	}
1578 	ret = -EINVAL;
1579 error3:
1580 	if (vendor_class) {
1581 		(*vendor_table)->vendor_class[vclass] = NULL;
1582 		kfree(vendor_class);
1583 	}
1584 error2:
1585 	if (vendor) {
1586 		*vendor_table = NULL;
1587 		kfree(vendor);
1588 	}
1589 error1:
1590 	return ret;
1591 }
1592 
1593 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1594 {
1595 	struct ib_mad_port_private *port_priv;
1596 	struct ib_mad_mgmt_class_table *class;
1597 	struct ib_mad_mgmt_method_table *method;
1598 	struct ib_mad_mgmt_vendor_class_table *vendor;
1599 	struct ib_mad_mgmt_vendor_class *vendor_class;
1600 	int index;
1601 	u8 mgmt_class;
1602 
1603 	/*
1604 	 * Was MAD registration request supplied
1605 	 * with original registration ?
1606 	 */
1607 	if (!agent_priv->reg_req) {
1608 		goto out;
1609 	}
1610 
1611 	port_priv = agent_priv->qp_info->port_priv;
1612 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1613 	class = port_priv->version[
1614 			agent_priv->reg_req->mgmt_class_version].class;
1615 	if (!class)
1616 		goto vendor_check;
1617 
1618 	method = class->method_table[mgmt_class];
1619 	if (method) {
1620 		/* Remove any methods for this mad agent */
1621 		remove_methods_mad_agent(method, agent_priv);
1622 		/* Now, check to see if there are any methods still in use */
1623 		if (!check_method_table(method)) {
1624 			/* If not, release management method table */
1625 			kfree(method);
1626 			class->method_table[mgmt_class] = NULL;
1627 			/* Any management classes left ? */
1628 			if (!check_class_table(class)) {
1629 				/* If not, release management class table */
1630 				kfree(class);
1631 				port_priv->version[
1632 					agent_priv->reg_req->
1633 					mgmt_class_version].class = NULL;
1634 			}
1635 		}
1636 	}
1637 
1638 vendor_check:
1639 	if (!is_vendor_class(mgmt_class))
1640 		goto out;
1641 
1642 	/* normalize mgmt_class to vendor range 2 */
1643 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1644 	vendor = port_priv->version[
1645 			agent_priv->reg_req->mgmt_class_version].vendor;
1646 
1647 	if (!vendor)
1648 		goto out;
1649 
1650 	vendor_class = vendor->vendor_class[mgmt_class];
1651 	if (vendor_class) {
1652 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1653 		if (index < 0)
1654 			goto out;
1655 		method = vendor_class->method_table[index];
1656 		if (method) {
1657 			/* Remove any methods for this mad agent */
1658 			remove_methods_mad_agent(method, agent_priv);
1659 			/*
1660 			 * Now, check to see if there are
1661 			 * any methods still in use
1662 			 */
1663 			if (!check_method_table(method)) {
1664 				/* If not, release management method table */
1665 				kfree(method);
1666 				vendor_class->method_table[index] = NULL;
1667 				memset(vendor_class->oui[index], 0, 3);
1668 				/* Any OUIs left ? */
1669 				if (!check_vendor_class(vendor_class)) {
1670 					/* If not, release vendor class table */
1671 					kfree(vendor_class);
1672 					vendor->vendor_class[mgmt_class] = NULL;
1673 					/* Any other vendor classes left ? */
1674 					if (!check_vendor_table(vendor)) {
1675 						kfree(vendor);
1676 						port_priv->version[
1677 							agent_priv->reg_req->
1678 							mgmt_class_version].
1679 							vendor = NULL;
1680 					}
1681 				}
1682 			}
1683 		}
1684 	}
1685 
1686 out:
1687 	return;
1688 }
1689 
1690 static struct ib_mad_agent_private *
1691 find_mad_agent(struct ib_mad_port_private *port_priv,
1692 	       const struct ib_mad_hdr *mad_hdr)
1693 {
1694 	struct ib_mad_agent_private *mad_agent = NULL;
1695 	unsigned long flags;
1696 
1697 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1698 	if (ib_response_mad(mad_hdr)) {
1699 		u32 hi_tid;
1700 		struct ib_mad_agent_private *entry;
1701 
1702 		/*
1703 		 * Routing is based on high 32 bits of transaction ID
1704 		 * of MAD.
1705 		 */
1706 		hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1707 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1708 			if (entry->agent.hi_tid == hi_tid) {
1709 				mad_agent = entry;
1710 				break;
1711 			}
1712 		}
1713 	} else {
1714 		struct ib_mad_mgmt_class_table *class;
1715 		struct ib_mad_mgmt_method_table *method;
1716 		struct ib_mad_mgmt_vendor_class_table *vendor;
1717 		struct ib_mad_mgmt_vendor_class *vendor_class;
1718 		const struct ib_vendor_mad *vendor_mad;
1719 		int index;
1720 
1721 		/*
1722 		 * Routing is based on version, class, and method
1723 		 * For "newer" vendor MADs, also based on OUI
1724 		 */
1725 		if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1726 			goto out;
1727 		if (!is_vendor_class(mad_hdr->mgmt_class)) {
1728 			class = port_priv->version[
1729 					mad_hdr->class_version].class;
1730 			if (!class)
1731 				goto out;
1732 			if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1733 			    ARRAY_SIZE(class->method_table))
1734 				goto out;
1735 			method = class->method_table[convert_mgmt_class(
1736 							mad_hdr->mgmt_class)];
1737 			if (method)
1738 				mad_agent = method->agent[mad_hdr->method &
1739 							  ~IB_MGMT_METHOD_RESP];
1740 		} else {
1741 			vendor = port_priv->version[
1742 					mad_hdr->class_version].vendor;
1743 			if (!vendor)
1744 				goto out;
1745 			vendor_class = vendor->vendor_class[vendor_class_index(
1746 						mad_hdr->mgmt_class)];
1747 			if (!vendor_class)
1748 				goto out;
1749 			/* Find matching OUI */
1750 			vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1751 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1752 			if (index == -1)
1753 				goto out;
1754 			method = vendor_class->method_table[index];
1755 			if (method) {
1756 				mad_agent = method->agent[mad_hdr->method &
1757 							  ~IB_MGMT_METHOD_RESP];
1758 			}
1759 		}
1760 	}
1761 
1762 	if (mad_agent) {
1763 		if (mad_agent->agent.recv_handler)
1764 			atomic_inc(&mad_agent->refcount);
1765 		else {
1766 			dev_notice(&port_priv->device->dev,
1767 				   "No receive handler for client %p on port %d\n",
1768 				   &mad_agent->agent, port_priv->port_num);
1769 			mad_agent = NULL;
1770 		}
1771 	}
1772 out:
1773 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1774 
1775 	return mad_agent;
1776 }
1777 
1778 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1779 			const struct ib_mad_qp_info *qp_info,
1780 			bool opa)
1781 {
1782 	int valid = 0;
1783 	u32 qp_num = qp_info->qp->qp_num;
1784 
1785 	/* Make sure MAD base version is understood */
1786 	if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1787 	    (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1788 		pr_err("MAD received with unsupported base version %d %s\n",
1789 		       mad_hdr->base_version, opa ? "(opa)" : "");
1790 		goto out;
1791 	}
1792 
1793 	/* Filter SMI packets sent to other than QP0 */
1794 	if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1795 	    (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1796 		if (qp_num == 0)
1797 			valid = 1;
1798 	} else {
1799 		/* CM attributes other than ClassPortInfo only use Send method */
1800 		if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1801 		    (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1802 		    (mad_hdr->method != IB_MGMT_METHOD_SEND))
1803 			goto out;
1804 		/* Filter GSI packets sent to QP0 */
1805 		if (qp_num != 0)
1806 			valid = 1;
1807 	}
1808 
1809 out:
1810 	return valid;
1811 }
1812 
1813 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1814 			    const struct ib_mad_hdr *mad_hdr)
1815 {
1816 	struct ib_rmpp_mad *rmpp_mad;
1817 
1818 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1819 	return !mad_agent_priv->agent.rmpp_version ||
1820 		!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1821 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1822 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1823 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1824 }
1825 
1826 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1827 				     const struct ib_mad_recv_wc *rwc)
1828 {
1829 	return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1830 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1831 }
1832 
1833 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1834 				   const struct ib_mad_send_wr_private *wr,
1835 				   const struct ib_mad_recv_wc *rwc )
1836 {
1837 	struct ib_ah_attr attr;
1838 	u8 send_resp, rcv_resp;
1839 	union ib_gid sgid;
1840 	struct ib_device *device = mad_agent_priv->agent.device;
1841 	u8 port_num = mad_agent_priv->agent.port_num;
1842 	u8 lmc;
1843 
1844 	send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1845 	rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1846 
1847 	if (send_resp == rcv_resp)
1848 		/* both requests, or both responses. GIDs different */
1849 		return 0;
1850 
1851 	if (ib_query_ah(wr->send_buf.ah, &attr))
1852 		/* Assume not equal, to avoid false positives. */
1853 		return 0;
1854 
1855 	if (!!(attr.ah_flags & IB_AH_GRH) !=
1856 	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1857 		/* one has GID, other does not.  Assume different */
1858 		return 0;
1859 
1860 	if (!send_resp && rcv_resp) {
1861 		/* is request/response. */
1862 		if (!(attr.ah_flags & IB_AH_GRH)) {
1863 			if (ib_get_cached_lmc(device, port_num, &lmc))
1864 				return 0;
1865 			return (!lmc || !((attr.src_path_bits ^
1866 					   rwc->wc->dlid_path_bits) &
1867 					  ((1 << lmc) - 1)));
1868 		} else {
1869 			if (ib_get_cached_gid(device, port_num,
1870 					      attr.grh.sgid_index, &sgid, NULL))
1871 				return 0;
1872 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1873 				       16);
1874 		}
1875 	}
1876 
1877 	if (!(attr.ah_flags & IB_AH_GRH))
1878 		return attr.dlid == rwc->wc->slid;
1879 	else
1880 		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1881 			       16);
1882 }
1883 
1884 static inline int is_direct(u8 class)
1885 {
1886 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1887 }
1888 
1889 struct ib_mad_send_wr_private*
1890 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1891 		 const struct ib_mad_recv_wc *wc)
1892 {
1893 	struct ib_mad_send_wr_private *wr;
1894 	const struct ib_mad_hdr *mad_hdr;
1895 
1896 	mad_hdr = &wc->recv_buf.mad->mad_hdr;
1897 
1898 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1899 		if ((wr->tid == mad_hdr->tid) &&
1900 		    rcv_has_same_class(wr, wc) &&
1901 		    /*
1902 		     * Don't check GID for direct routed MADs.
1903 		     * These might have permissive LIDs.
1904 		     */
1905 		    (is_direct(mad_hdr->mgmt_class) ||
1906 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1907 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1908 	}
1909 
1910 	/*
1911 	 * It's possible to receive the response before we've
1912 	 * been notified that the send has completed
1913 	 */
1914 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1915 		if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1916 		    wr->tid == mad_hdr->tid &&
1917 		    wr->timeout &&
1918 		    rcv_has_same_class(wr, wc) &&
1919 		    /*
1920 		     * Don't check GID for direct routed MADs.
1921 		     * These might have permissive LIDs.
1922 		     */
1923 		    (is_direct(mad_hdr->mgmt_class) ||
1924 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1925 			/* Verify request has not been canceled */
1926 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1927 	}
1928 	return NULL;
1929 }
1930 
1931 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1932 {
1933 	mad_send_wr->timeout = 0;
1934 	if (mad_send_wr->refcount == 1)
1935 		list_move_tail(&mad_send_wr->agent_list,
1936 			      &mad_send_wr->mad_agent_priv->done_list);
1937 }
1938 
1939 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1940 				 struct ib_mad_recv_wc *mad_recv_wc)
1941 {
1942 	struct ib_mad_send_wr_private *mad_send_wr;
1943 	struct ib_mad_send_wc mad_send_wc;
1944 	unsigned long flags;
1945 
1946 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1947 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1948 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1949 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1950 						      mad_recv_wc);
1951 		if (!mad_recv_wc) {
1952 			deref_mad_agent(mad_agent_priv);
1953 			return;
1954 		}
1955 	}
1956 
1957 	/* Complete corresponding request */
1958 	if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1959 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1960 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1961 		if (!mad_send_wr) {
1962 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1963 			if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
1964 			   && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
1965 			   && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
1966 					& IB_MGMT_RMPP_FLAG_ACTIVE)) {
1967 				/* user rmpp is in effect
1968 				 * and this is an active RMPP MAD
1969 				 */
1970 				mad_agent_priv->agent.recv_handler(
1971 						&mad_agent_priv->agent, NULL,
1972 						mad_recv_wc);
1973 				atomic_dec(&mad_agent_priv->refcount);
1974 			} else {
1975 				/* not user rmpp, revert to normal behavior and
1976 				 * drop the mad */
1977 				ib_free_recv_mad(mad_recv_wc);
1978 				deref_mad_agent(mad_agent_priv);
1979 				return;
1980 			}
1981 		} else {
1982 			ib_mark_mad_done(mad_send_wr);
1983 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1984 
1985 			/* Defined behavior is to complete response before request */
1986 			mad_agent_priv->agent.recv_handler(
1987 					&mad_agent_priv->agent,
1988 					&mad_send_wr->send_buf,
1989 					mad_recv_wc);
1990 			atomic_dec(&mad_agent_priv->refcount);
1991 
1992 			mad_send_wc.status = IB_WC_SUCCESS;
1993 			mad_send_wc.vendor_err = 0;
1994 			mad_send_wc.send_buf = &mad_send_wr->send_buf;
1995 			ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1996 		}
1997 	} else {
1998 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
1999 						   mad_recv_wc);
2000 		deref_mad_agent(mad_agent_priv);
2001 	}
2002 }
2003 
2004 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2005 				     const struct ib_mad_qp_info *qp_info,
2006 				     const struct ib_wc *wc,
2007 				     int port_num,
2008 				     struct ib_mad_private *recv,
2009 				     struct ib_mad_private *response)
2010 {
2011 	enum smi_forward_action retsmi;
2012 	struct ib_smp *smp = (struct ib_smp *)recv->mad;
2013 
2014 	if (smi_handle_dr_smp_recv(smp,
2015 				   rdma_cap_ib_switch(port_priv->device),
2016 				   port_num,
2017 				   port_priv->device->phys_port_cnt) ==
2018 				   IB_SMI_DISCARD)
2019 		return IB_SMI_DISCARD;
2020 
2021 	retsmi = smi_check_forward_dr_smp(smp);
2022 	if (retsmi == IB_SMI_LOCAL)
2023 		return IB_SMI_HANDLE;
2024 
2025 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2026 		if (smi_handle_dr_smp_send(smp,
2027 					   rdma_cap_ib_switch(port_priv->device),
2028 					   port_num) == IB_SMI_DISCARD)
2029 			return IB_SMI_DISCARD;
2030 
2031 		if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2032 			return IB_SMI_DISCARD;
2033 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2034 		/* forward case for switches */
2035 		memcpy(response, recv, mad_priv_size(response));
2036 		response->header.recv_wc.wc = &response->header.wc;
2037 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2038 		response->header.recv_wc.recv_buf.grh = &response->grh;
2039 
2040 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2041 				    &response->grh, wc,
2042 				    port_priv->device,
2043 				    smi_get_fwd_port(smp),
2044 				    qp_info->qp->qp_num,
2045 				    response->mad_size,
2046 				    false);
2047 
2048 		return IB_SMI_DISCARD;
2049 	}
2050 	return IB_SMI_HANDLE;
2051 }
2052 
2053 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2054 				    struct ib_mad_private *response,
2055 				    size_t *resp_len, bool opa)
2056 {
2057 	const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2058 	struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2059 
2060 	if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2061 	    recv_hdr->method == IB_MGMT_METHOD_SET) {
2062 		memcpy(response, recv, mad_priv_size(response));
2063 		response->header.recv_wc.wc = &response->header.wc;
2064 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2065 		response->header.recv_wc.recv_buf.grh = &response->grh;
2066 		resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2067 		resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2068 		if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2069 			resp_hdr->status |= IB_SMP_DIRECTION;
2070 
2071 		if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2072 			if (recv_hdr->mgmt_class ==
2073 			    IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2074 			    recv_hdr->mgmt_class ==
2075 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2076 				*resp_len = opa_get_smp_header_size(
2077 							(struct opa_smp *)recv->mad);
2078 			else
2079 				*resp_len = sizeof(struct ib_mad_hdr);
2080 		}
2081 
2082 		return true;
2083 	} else {
2084 		return false;
2085 	}
2086 }
2087 
2088 static enum smi_action
2089 handle_opa_smi(struct ib_mad_port_private *port_priv,
2090 	       struct ib_mad_qp_info *qp_info,
2091 	       struct ib_wc *wc,
2092 	       int port_num,
2093 	       struct ib_mad_private *recv,
2094 	       struct ib_mad_private *response)
2095 {
2096 	enum smi_forward_action retsmi;
2097 	struct opa_smp *smp = (struct opa_smp *)recv->mad;
2098 
2099 	if (opa_smi_handle_dr_smp_recv(smp,
2100 				   rdma_cap_ib_switch(port_priv->device),
2101 				   port_num,
2102 				   port_priv->device->phys_port_cnt) ==
2103 				   IB_SMI_DISCARD)
2104 		return IB_SMI_DISCARD;
2105 
2106 	retsmi = opa_smi_check_forward_dr_smp(smp);
2107 	if (retsmi == IB_SMI_LOCAL)
2108 		return IB_SMI_HANDLE;
2109 
2110 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2111 		if (opa_smi_handle_dr_smp_send(smp,
2112 					   rdma_cap_ib_switch(port_priv->device),
2113 					   port_num) == IB_SMI_DISCARD)
2114 			return IB_SMI_DISCARD;
2115 
2116 		if (opa_smi_check_local_smp(smp, port_priv->device) ==
2117 		    IB_SMI_DISCARD)
2118 			return IB_SMI_DISCARD;
2119 
2120 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2121 		/* forward case for switches */
2122 		memcpy(response, recv, mad_priv_size(response));
2123 		response->header.recv_wc.wc = &response->header.wc;
2124 		response->header.recv_wc.recv_buf.opa_mad =
2125 				(struct opa_mad *)response->mad;
2126 		response->header.recv_wc.recv_buf.grh = &response->grh;
2127 
2128 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2129 				    &response->grh, wc,
2130 				    port_priv->device,
2131 				    opa_smi_get_fwd_port(smp),
2132 				    qp_info->qp->qp_num,
2133 				    recv->header.wc.byte_len,
2134 				    true);
2135 
2136 		return IB_SMI_DISCARD;
2137 	}
2138 
2139 	return IB_SMI_HANDLE;
2140 }
2141 
2142 static enum smi_action
2143 handle_smi(struct ib_mad_port_private *port_priv,
2144 	   struct ib_mad_qp_info *qp_info,
2145 	   struct ib_wc *wc,
2146 	   int port_num,
2147 	   struct ib_mad_private *recv,
2148 	   struct ib_mad_private *response,
2149 	   bool opa)
2150 {
2151 	struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2152 
2153 	if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2154 	    mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2155 		return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2156 				      response);
2157 
2158 	return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2159 }
2160 
2161 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2162 {
2163 	struct ib_mad_port_private *port_priv = cq->cq_context;
2164 	struct ib_mad_list_head *mad_list =
2165 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2166 	struct ib_mad_qp_info *qp_info;
2167 	struct ib_mad_private_header *mad_priv_hdr;
2168 	struct ib_mad_private *recv, *response = NULL;
2169 	struct ib_mad_agent_private *mad_agent;
2170 	int port_num;
2171 	int ret = IB_MAD_RESULT_SUCCESS;
2172 	size_t mad_size;
2173 	u16 resp_mad_pkey_index = 0;
2174 	bool opa;
2175 
2176 	if (list_empty_careful(&port_priv->port_list))
2177 		return;
2178 
2179 	if (wc->status != IB_WC_SUCCESS) {
2180 		/*
2181 		 * Receive errors indicate that the QP has entered the error
2182 		 * state - error handling/shutdown code will cleanup
2183 		 */
2184 		return;
2185 	}
2186 
2187 	qp_info = mad_list->mad_queue->qp_info;
2188 	dequeue_mad(mad_list);
2189 
2190 	opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2191 			       qp_info->port_priv->port_num);
2192 
2193 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2194 				    mad_list);
2195 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2196 	ib_dma_unmap_single(port_priv->device,
2197 			    recv->header.mapping,
2198 			    mad_priv_dma_size(recv),
2199 			    DMA_FROM_DEVICE);
2200 
2201 	/* Setup MAD receive work completion from "normal" work completion */
2202 	recv->header.wc = *wc;
2203 	recv->header.recv_wc.wc = &recv->header.wc;
2204 
2205 	if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2206 		recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2207 		recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2208 	} else {
2209 		recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2210 		recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2211 	}
2212 
2213 	recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2214 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
2215 
2216 	if (atomic_read(&qp_info->snoop_count))
2217 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2218 
2219 	/* Validate MAD */
2220 	if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2221 		goto out;
2222 
2223 	mad_size = recv->mad_size;
2224 	response = alloc_mad_private(mad_size, GFP_KERNEL);
2225 	if (!response)
2226 		goto out;
2227 
2228 	if (rdma_cap_ib_switch(port_priv->device))
2229 		port_num = wc->port_num;
2230 	else
2231 		port_num = port_priv->port_num;
2232 
2233 	if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2234 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2235 		if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2236 			       response, opa)
2237 		    == IB_SMI_DISCARD)
2238 			goto out;
2239 	}
2240 
2241 	/* Give driver "right of first refusal" on incoming MAD */
2242 	if (port_priv->device->process_mad) {
2243 		ret = port_priv->device->process_mad(port_priv->device, 0,
2244 						     port_priv->port_num,
2245 						     wc, &recv->grh,
2246 						     (const struct ib_mad_hdr *)recv->mad,
2247 						     recv->mad_size,
2248 						     (struct ib_mad_hdr *)response->mad,
2249 						     &mad_size, &resp_mad_pkey_index);
2250 
2251 		if (opa)
2252 			wc->pkey_index = resp_mad_pkey_index;
2253 
2254 		if (ret & IB_MAD_RESULT_SUCCESS) {
2255 			if (ret & IB_MAD_RESULT_CONSUMED)
2256 				goto out;
2257 			if (ret & IB_MAD_RESULT_REPLY) {
2258 				agent_send_response((const struct ib_mad_hdr *)response->mad,
2259 						    &recv->grh, wc,
2260 						    port_priv->device,
2261 						    port_num,
2262 						    qp_info->qp->qp_num,
2263 						    mad_size, opa);
2264 				goto out;
2265 			}
2266 		}
2267 	}
2268 
2269 	mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2270 	if (mad_agent) {
2271 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2272 		/*
2273 		 * recv is freed up in error cases in ib_mad_complete_recv
2274 		 * or via recv_handler in ib_mad_complete_recv()
2275 		 */
2276 		recv = NULL;
2277 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2278 		   generate_unmatched_resp(recv, response, &mad_size, opa)) {
2279 		agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2280 				    port_priv->device, port_num,
2281 				    qp_info->qp->qp_num, mad_size, opa);
2282 	}
2283 
2284 out:
2285 	/* Post another receive request for this QP */
2286 	if (response) {
2287 		ib_mad_post_receive_mads(qp_info, response);
2288 		kfree(recv);
2289 	} else
2290 		ib_mad_post_receive_mads(qp_info, recv);
2291 }
2292 
2293 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2294 {
2295 	struct ib_mad_send_wr_private *mad_send_wr;
2296 	unsigned long delay;
2297 
2298 	if (list_empty(&mad_agent_priv->wait_list)) {
2299 		cancel_delayed_work(&mad_agent_priv->timed_work);
2300 	} else {
2301 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2302 					 struct ib_mad_send_wr_private,
2303 					 agent_list);
2304 
2305 		if (time_after(mad_agent_priv->timeout,
2306 			       mad_send_wr->timeout)) {
2307 			mad_agent_priv->timeout = mad_send_wr->timeout;
2308 			delay = mad_send_wr->timeout - jiffies;
2309 			if ((long)delay <= 0)
2310 				delay = 1;
2311 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2312 					 &mad_agent_priv->timed_work, delay);
2313 		}
2314 	}
2315 }
2316 
2317 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2318 {
2319 	struct ib_mad_agent_private *mad_agent_priv;
2320 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2321 	struct list_head *list_item;
2322 	unsigned long delay;
2323 
2324 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2325 	list_del(&mad_send_wr->agent_list);
2326 
2327 	delay = mad_send_wr->timeout;
2328 	mad_send_wr->timeout += jiffies;
2329 
2330 	if (delay) {
2331 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2332 			temp_mad_send_wr = list_entry(list_item,
2333 						struct ib_mad_send_wr_private,
2334 						agent_list);
2335 			if (time_after(mad_send_wr->timeout,
2336 				       temp_mad_send_wr->timeout))
2337 				break;
2338 		}
2339 	}
2340 	else
2341 		list_item = &mad_agent_priv->wait_list;
2342 	list_add(&mad_send_wr->agent_list, list_item);
2343 
2344 	/* Reschedule a work item if we have a shorter timeout */
2345 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2346 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2347 				 &mad_agent_priv->timed_work, delay);
2348 }
2349 
2350 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2351 			  int timeout_ms)
2352 {
2353 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2354 	wait_for_response(mad_send_wr);
2355 }
2356 
2357 /*
2358  * Process a send work completion
2359  */
2360 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2361 			     struct ib_mad_send_wc *mad_send_wc)
2362 {
2363 	struct ib_mad_agent_private	*mad_agent_priv;
2364 	unsigned long			flags;
2365 	int				ret;
2366 
2367 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2368 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2369 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2370 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2371 		if (ret == IB_RMPP_RESULT_CONSUMED)
2372 			goto done;
2373 	} else
2374 		ret = IB_RMPP_RESULT_UNHANDLED;
2375 
2376 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2377 	    mad_send_wr->status == IB_WC_SUCCESS) {
2378 		mad_send_wr->status = mad_send_wc->status;
2379 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2380 	}
2381 
2382 	if (--mad_send_wr->refcount > 0) {
2383 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2384 		    mad_send_wr->status == IB_WC_SUCCESS) {
2385 			wait_for_response(mad_send_wr);
2386 		}
2387 		goto done;
2388 	}
2389 
2390 	/* Remove send from MAD agent and notify client of completion */
2391 	list_del(&mad_send_wr->agent_list);
2392 	adjust_timeout(mad_agent_priv);
2393 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2394 
2395 	if (mad_send_wr->status != IB_WC_SUCCESS )
2396 		mad_send_wc->status = mad_send_wr->status;
2397 	if (ret == IB_RMPP_RESULT_INTERNAL)
2398 		ib_rmpp_send_handler(mad_send_wc);
2399 	else
2400 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2401 						   mad_send_wc);
2402 
2403 	/* Release reference on agent taken when sending */
2404 	deref_mad_agent(mad_agent_priv);
2405 	return;
2406 done:
2407 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2408 }
2409 
2410 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2411 {
2412 	struct ib_mad_port_private *port_priv = cq->cq_context;
2413 	struct ib_mad_list_head *mad_list =
2414 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2415 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2416 	struct ib_mad_qp_info		*qp_info;
2417 	struct ib_mad_queue		*send_queue;
2418 	struct ib_send_wr		*bad_send_wr;
2419 	struct ib_mad_send_wc		mad_send_wc;
2420 	unsigned long flags;
2421 	int ret;
2422 
2423 	if (list_empty_careful(&port_priv->port_list))
2424 		return;
2425 
2426 	if (wc->status != IB_WC_SUCCESS) {
2427 		if (!ib_mad_send_error(port_priv, wc))
2428 			return;
2429 	}
2430 
2431 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2432 				   mad_list);
2433 	send_queue = mad_list->mad_queue;
2434 	qp_info = send_queue->qp_info;
2435 
2436 retry:
2437 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2438 			    mad_send_wr->header_mapping,
2439 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2440 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2441 			    mad_send_wr->payload_mapping,
2442 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2443 	queued_send_wr = NULL;
2444 	spin_lock_irqsave(&send_queue->lock, flags);
2445 	list_del(&mad_list->list);
2446 
2447 	/* Move queued send to the send queue */
2448 	if (send_queue->count-- > send_queue->max_active) {
2449 		mad_list = container_of(qp_info->overflow_list.next,
2450 					struct ib_mad_list_head, list);
2451 		queued_send_wr = container_of(mad_list,
2452 					struct ib_mad_send_wr_private,
2453 					mad_list);
2454 		list_move_tail(&mad_list->list, &send_queue->list);
2455 	}
2456 	spin_unlock_irqrestore(&send_queue->lock, flags);
2457 
2458 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2459 	mad_send_wc.status = wc->status;
2460 	mad_send_wc.vendor_err = wc->vendor_err;
2461 	if (atomic_read(&qp_info->snoop_count))
2462 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2463 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2464 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2465 
2466 	if (queued_send_wr) {
2467 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2468 				   &bad_send_wr);
2469 		if (ret) {
2470 			dev_err(&port_priv->device->dev,
2471 				"ib_post_send failed: %d\n", ret);
2472 			mad_send_wr = queued_send_wr;
2473 			wc->status = IB_WC_LOC_QP_OP_ERR;
2474 			goto retry;
2475 		}
2476 	}
2477 }
2478 
2479 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2480 {
2481 	struct ib_mad_send_wr_private *mad_send_wr;
2482 	struct ib_mad_list_head *mad_list;
2483 	unsigned long flags;
2484 
2485 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2486 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2487 		mad_send_wr = container_of(mad_list,
2488 					   struct ib_mad_send_wr_private,
2489 					   mad_list);
2490 		mad_send_wr->retry = 1;
2491 	}
2492 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2493 }
2494 
2495 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2496 		struct ib_wc *wc)
2497 {
2498 	struct ib_mad_list_head *mad_list =
2499 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2500 	struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2501 	struct ib_mad_send_wr_private *mad_send_wr;
2502 	int ret;
2503 
2504 	/*
2505 	 * Send errors will transition the QP to SQE - move
2506 	 * QP to RTS and repost flushed work requests
2507 	 */
2508 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2509 				   mad_list);
2510 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2511 		if (mad_send_wr->retry) {
2512 			/* Repost send */
2513 			struct ib_send_wr *bad_send_wr;
2514 
2515 			mad_send_wr->retry = 0;
2516 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2517 					&bad_send_wr);
2518 			if (!ret)
2519 				return false;
2520 		}
2521 	} else {
2522 		struct ib_qp_attr *attr;
2523 
2524 		/* Transition QP to RTS and fail offending send */
2525 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2526 		if (attr) {
2527 			attr->qp_state = IB_QPS_RTS;
2528 			attr->cur_qp_state = IB_QPS_SQE;
2529 			ret = ib_modify_qp(qp_info->qp, attr,
2530 					   IB_QP_STATE | IB_QP_CUR_STATE);
2531 			kfree(attr);
2532 			if (ret)
2533 				dev_err(&port_priv->device->dev,
2534 					"%s - ib_modify_qp to RTS: %d\n",
2535 					__func__, ret);
2536 			else
2537 				mark_sends_for_retry(qp_info);
2538 		}
2539 	}
2540 
2541 	return true;
2542 }
2543 
2544 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2545 {
2546 	unsigned long flags;
2547 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2548 	struct ib_mad_send_wc mad_send_wc;
2549 	struct list_head cancel_list;
2550 
2551 	INIT_LIST_HEAD(&cancel_list);
2552 
2553 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2554 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2555 				 &mad_agent_priv->send_list, agent_list) {
2556 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2557 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2558 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2559 		}
2560 	}
2561 
2562 	/* Empty wait list to prevent receives from finding a request */
2563 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2564 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2565 
2566 	/* Report all cancelled requests */
2567 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2568 	mad_send_wc.vendor_err = 0;
2569 
2570 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2571 				 &cancel_list, agent_list) {
2572 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2573 		list_del(&mad_send_wr->agent_list);
2574 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2575 						   &mad_send_wc);
2576 		atomic_dec(&mad_agent_priv->refcount);
2577 	}
2578 }
2579 
2580 static struct ib_mad_send_wr_private*
2581 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2582 	     struct ib_mad_send_buf *send_buf)
2583 {
2584 	struct ib_mad_send_wr_private *mad_send_wr;
2585 
2586 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2587 			    agent_list) {
2588 		if (&mad_send_wr->send_buf == send_buf)
2589 			return mad_send_wr;
2590 	}
2591 
2592 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2593 			    agent_list) {
2594 		if (is_rmpp_data_mad(mad_agent_priv,
2595 				     mad_send_wr->send_buf.mad) &&
2596 		    &mad_send_wr->send_buf == send_buf)
2597 			return mad_send_wr;
2598 	}
2599 	return NULL;
2600 }
2601 
2602 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2603 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2604 {
2605 	struct ib_mad_agent_private *mad_agent_priv;
2606 	struct ib_mad_send_wr_private *mad_send_wr;
2607 	unsigned long flags;
2608 	int active;
2609 
2610 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2611 				      agent);
2612 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2613 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2614 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2615 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2616 		return -EINVAL;
2617 	}
2618 
2619 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2620 	if (!timeout_ms) {
2621 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2622 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2623 	}
2624 
2625 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2626 	if (active)
2627 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2628 	else
2629 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2630 
2631 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2632 	return 0;
2633 }
2634 EXPORT_SYMBOL(ib_modify_mad);
2635 
2636 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2637 		   struct ib_mad_send_buf *send_buf)
2638 {
2639 	ib_modify_mad(mad_agent, send_buf, 0);
2640 }
2641 EXPORT_SYMBOL(ib_cancel_mad);
2642 
2643 static void local_completions(struct work_struct *work)
2644 {
2645 	struct ib_mad_agent_private *mad_agent_priv;
2646 	struct ib_mad_local_private *local;
2647 	struct ib_mad_agent_private *recv_mad_agent;
2648 	unsigned long flags;
2649 	int free_mad;
2650 	struct ib_wc wc;
2651 	struct ib_mad_send_wc mad_send_wc;
2652 	bool opa;
2653 
2654 	mad_agent_priv =
2655 		container_of(work, struct ib_mad_agent_private, local_work);
2656 
2657 	opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2658 			       mad_agent_priv->qp_info->port_priv->port_num);
2659 
2660 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2661 	while (!list_empty(&mad_agent_priv->local_list)) {
2662 		local = list_entry(mad_agent_priv->local_list.next,
2663 				   struct ib_mad_local_private,
2664 				   completion_list);
2665 		list_del(&local->completion_list);
2666 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2667 		free_mad = 0;
2668 		if (local->mad_priv) {
2669 			u8 base_version;
2670 			recv_mad_agent = local->recv_mad_agent;
2671 			if (!recv_mad_agent) {
2672 				dev_err(&mad_agent_priv->agent.device->dev,
2673 					"No receive MAD agent for local completion\n");
2674 				free_mad = 1;
2675 				goto local_send_completion;
2676 			}
2677 
2678 			/*
2679 			 * Defined behavior is to complete response
2680 			 * before request
2681 			 */
2682 			build_smp_wc(recv_mad_agent->agent.qp,
2683 				     local->mad_send_wr->send_wr.wr.wr_cqe,
2684 				     be16_to_cpu(IB_LID_PERMISSIVE),
2685 				     local->mad_send_wr->send_wr.pkey_index,
2686 				     recv_mad_agent->agent.port_num, &wc);
2687 
2688 			local->mad_priv->header.recv_wc.wc = &wc;
2689 
2690 			base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2691 			if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2692 				local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2693 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2694 			} else {
2695 				local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2696 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2697 			}
2698 
2699 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2700 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2701 				 &local->mad_priv->header.recv_wc.rmpp_list);
2702 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2703 			local->mad_priv->header.recv_wc.recv_buf.mad =
2704 						(struct ib_mad *)local->mad_priv->mad;
2705 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2706 				snoop_recv(recv_mad_agent->qp_info,
2707 					  &local->mad_priv->header.recv_wc,
2708 					   IB_MAD_SNOOP_RECVS);
2709 			recv_mad_agent->agent.recv_handler(
2710 						&recv_mad_agent->agent,
2711 						&local->mad_send_wr->send_buf,
2712 						&local->mad_priv->header.recv_wc);
2713 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2714 			atomic_dec(&recv_mad_agent->refcount);
2715 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2716 		}
2717 
2718 local_send_completion:
2719 		/* Complete send */
2720 		mad_send_wc.status = IB_WC_SUCCESS;
2721 		mad_send_wc.vendor_err = 0;
2722 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2723 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2724 			snoop_send(mad_agent_priv->qp_info,
2725 				   &local->mad_send_wr->send_buf,
2726 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2727 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2728 						   &mad_send_wc);
2729 
2730 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2731 		atomic_dec(&mad_agent_priv->refcount);
2732 		if (free_mad)
2733 			kfree(local->mad_priv);
2734 		kfree(local);
2735 	}
2736 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2737 }
2738 
2739 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2740 {
2741 	int ret;
2742 
2743 	if (!mad_send_wr->retries_left)
2744 		return -ETIMEDOUT;
2745 
2746 	mad_send_wr->retries_left--;
2747 	mad_send_wr->send_buf.retries++;
2748 
2749 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2750 
2751 	if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2752 		ret = ib_retry_rmpp(mad_send_wr);
2753 		switch (ret) {
2754 		case IB_RMPP_RESULT_UNHANDLED:
2755 			ret = ib_send_mad(mad_send_wr);
2756 			break;
2757 		case IB_RMPP_RESULT_CONSUMED:
2758 			ret = 0;
2759 			break;
2760 		default:
2761 			ret = -ECOMM;
2762 			break;
2763 		}
2764 	} else
2765 		ret = ib_send_mad(mad_send_wr);
2766 
2767 	if (!ret) {
2768 		mad_send_wr->refcount++;
2769 		list_add_tail(&mad_send_wr->agent_list,
2770 			      &mad_send_wr->mad_agent_priv->send_list);
2771 	}
2772 	return ret;
2773 }
2774 
2775 static void timeout_sends(struct work_struct *work)
2776 {
2777 	struct ib_mad_agent_private *mad_agent_priv;
2778 	struct ib_mad_send_wr_private *mad_send_wr;
2779 	struct ib_mad_send_wc mad_send_wc;
2780 	unsigned long flags, delay;
2781 
2782 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2783 				      timed_work.work);
2784 	mad_send_wc.vendor_err = 0;
2785 
2786 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2787 	while (!list_empty(&mad_agent_priv->wait_list)) {
2788 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2789 					 struct ib_mad_send_wr_private,
2790 					 agent_list);
2791 
2792 		if (time_after(mad_send_wr->timeout, jiffies)) {
2793 			delay = mad_send_wr->timeout - jiffies;
2794 			if ((long)delay <= 0)
2795 				delay = 1;
2796 			queue_delayed_work(mad_agent_priv->qp_info->
2797 					   port_priv->wq,
2798 					   &mad_agent_priv->timed_work, delay);
2799 			break;
2800 		}
2801 
2802 		list_del(&mad_send_wr->agent_list);
2803 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2804 		    !retry_send(mad_send_wr))
2805 			continue;
2806 
2807 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2808 
2809 		if (mad_send_wr->status == IB_WC_SUCCESS)
2810 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2811 		else
2812 			mad_send_wc.status = mad_send_wr->status;
2813 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2814 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2815 						   &mad_send_wc);
2816 
2817 		atomic_dec(&mad_agent_priv->refcount);
2818 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2819 	}
2820 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2821 }
2822 
2823 /*
2824  * Allocate receive MADs and post receive WRs for them
2825  */
2826 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2827 				    struct ib_mad_private *mad)
2828 {
2829 	unsigned long flags;
2830 	int post, ret;
2831 	struct ib_mad_private *mad_priv;
2832 	struct ib_sge sg_list;
2833 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2834 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2835 
2836 	/* Initialize common scatter list fields */
2837 	sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2838 
2839 	/* Initialize common receive WR fields */
2840 	recv_wr.next = NULL;
2841 	recv_wr.sg_list = &sg_list;
2842 	recv_wr.num_sge = 1;
2843 
2844 	do {
2845 		/* Allocate and map receive buffer */
2846 		if (mad) {
2847 			mad_priv = mad;
2848 			mad = NULL;
2849 		} else {
2850 			mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2851 						     GFP_ATOMIC);
2852 			if (!mad_priv) {
2853 				ret = -ENOMEM;
2854 				break;
2855 			}
2856 		}
2857 		sg_list.length = mad_priv_dma_size(mad_priv);
2858 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2859 						 &mad_priv->grh,
2860 						 mad_priv_dma_size(mad_priv),
2861 						 DMA_FROM_DEVICE);
2862 		if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2863 						  sg_list.addr))) {
2864 			ret = -ENOMEM;
2865 			break;
2866 		}
2867 		mad_priv->header.mapping = sg_list.addr;
2868 		mad_priv->header.mad_list.mad_queue = recv_queue;
2869 		mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2870 		recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2871 
2872 		/* Post receive WR */
2873 		spin_lock_irqsave(&recv_queue->lock, flags);
2874 		post = (++recv_queue->count < recv_queue->max_active);
2875 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2876 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2877 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2878 		if (ret) {
2879 			spin_lock_irqsave(&recv_queue->lock, flags);
2880 			list_del(&mad_priv->header.mad_list.list);
2881 			recv_queue->count--;
2882 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2883 			ib_dma_unmap_single(qp_info->port_priv->device,
2884 					    mad_priv->header.mapping,
2885 					    mad_priv_dma_size(mad_priv),
2886 					    DMA_FROM_DEVICE);
2887 			kfree(mad_priv);
2888 			dev_err(&qp_info->port_priv->device->dev,
2889 				"ib_post_recv failed: %d\n", ret);
2890 			break;
2891 		}
2892 	} while (post);
2893 
2894 	return ret;
2895 }
2896 
2897 /*
2898  * Return all the posted receive MADs
2899  */
2900 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2901 {
2902 	struct ib_mad_private_header *mad_priv_hdr;
2903 	struct ib_mad_private *recv;
2904 	struct ib_mad_list_head *mad_list;
2905 
2906 	if (!qp_info->qp)
2907 		return;
2908 
2909 	while (!list_empty(&qp_info->recv_queue.list)) {
2910 
2911 		mad_list = list_entry(qp_info->recv_queue.list.next,
2912 				      struct ib_mad_list_head, list);
2913 		mad_priv_hdr = container_of(mad_list,
2914 					    struct ib_mad_private_header,
2915 					    mad_list);
2916 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2917 				    header);
2918 
2919 		/* Remove from posted receive MAD list */
2920 		list_del(&mad_list->list);
2921 
2922 		ib_dma_unmap_single(qp_info->port_priv->device,
2923 				    recv->header.mapping,
2924 				    mad_priv_dma_size(recv),
2925 				    DMA_FROM_DEVICE);
2926 		kfree(recv);
2927 	}
2928 
2929 	qp_info->recv_queue.count = 0;
2930 }
2931 
2932 /*
2933  * Start the port
2934  */
2935 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2936 {
2937 	int ret, i;
2938 	struct ib_qp_attr *attr;
2939 	struct ib_qp *qp;
2940 	u16 pkey_index;
2941 
2942 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2943 	if (!attr)
2944 		return -ENOMEM;
2945 
2946 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2947 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
2948 	if (ret)
2949 		pkey_index = 0;
2950 
2951 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2952 		qp = port_priv->qp_info[i].qp;
2953 		if (!qp)
2954 			continue;
2955 
2956 		/*
2957 		 * PKey index for QP1 is irrelevant but
2958 		 * one is needed for the Reset to Init transition
2959 		 */
2960 		attr->qp_state = IB_QPS_INIT;
2961 		attr->pkey_index = pkey_index;
2962 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2963 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2964 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2965 		if (ret) {
2966 			dev_err(&port_priv->device->dev,
2967 				"Couldn't change QP%d state to INIT: %d\n",
2968 				i, ret);
2969 			goto out;
2970 		}
2971 
2972 		attr->qp_state = IB_QPS_RTR;
2973 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2974 		if (ret) {
2975 			dev_err(&port_priv->device->dev,
2976 				"Couldn't change QP%d state to RTR: %d\n",
2977 				i, ret);
2978 			goto out;
2979 		}
2980 
2981 		attr->qp_state = IB_QPS_RTS;
2982 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2983 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2984 		if (ret) {
2985 			dev_err(&port_priv->device->dev,
2986 				"Couldn't change QP%d state to RTS: %d\n",
2987 				i, ret);
2988 			goto out;
2989 		}
2990 	}
2991 
2992 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2993 	if (ret) {
2994 		dev_err(&port_priv->device->dev,
2995 			"Failed to request completion notification: %d\n",
2996 			ret);
2997 		goto out;
2998 	}
2999 
3000 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3001 		if (!port_priv->qp_info[i].qp)
3002 			continue;
3003 
3004 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3005 		if (ret) {
3006 			dev_err(&port_priv->device->dev,
3007 				"Couldn't post receive WRs\n");
3008 			goto out;
3009 		}
3010 	}
3011 out:
3012 	kfree(attr);
3013 	return ret;
3014 }
3015 
3016 static void qp_event_handler(struct ib_event *event, void *qp_context)
3017 {
3018 	struct ib_mad_qp_info	*qp_info = qp_context;
3019 
3020 	/* It's worse than that! He's dead, Jim! */
3021 	dev_err(&qp_info->port_priv->device->dev,
3022 		"Fatal error (%d) on MAD QP (%d)\n",
3023 		event->event, qp_info->qp->qp_num);
3024 }
3025 
3026 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3027 			   struct ib_mad_queue *mad_queue)
3028 {
3029 	mad_queue->qp_info = qp_info;
3030 	mad_queue->count = 0;
3031 	spin_lock_init(&mad_queue->lock);
3032 	INIT_LIST_HEAD(&mad_queue->list);
3033 }
3034 
3035 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3036 			struct ib_mad_qp_info *qp_info)
3037 {
3038 	qp_info->port_priv = port_priv;
3039 	init_mad_queue(qp_info, &qp_info->send_queue);
3040 	init_mad_queue(qp_info, &qp_info->recv_queue);
3041 	INIT_LIST_HEAD(&qp_info->overflow_list);
3042 	spin_lock_init(&qp_info->snoop_lock);
3043 	qp_info->snoop_table = NULL;
3044 	qp_info->snoop_table_size = 0;
3045 	atomic_set(&qp_info->snoop_count, 0);
3046 }
3047 
3048 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3049 			 enum ib_qp_type qp_type)
3050 {
3051 	struct ib_qp_init_attr	qp_init_attr;
3052 	int ret;
3053 
3054 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
3055 	qp_init_attr.send_cq = qp_info->port_priv->cq;
3056 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
3057 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3058 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
3059 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3060 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3061 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3062 	qp_init_attr.qp_type = qp_type;
3063 	qp_init_attr.port_num = qp_info->port_priv->port_num;
3064 	qp_init_attr.qp_context = qp_info;
3065 	qp_init_attr.event_handler = qp_event_handler;
3066 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3067 	if (IS_ERR(qp_info->qp)) {
3068 		dev_err(&qp_info->port_priv->device->dev,
3069 			"Couldn't create ib_mad QP%d\n",
3070 			get_spl_qp_index(qp_type));
3071 		ret = PTR_ERR(qp_info->qp);
3072 		goto error;
3073 	}
3074 	/* Use minimum queue sizes unless the CQ is resized */
3075 	qp_info->send_queue.max_active = mad_sendq_size;
3076 	qp_info->recv_queue.max_active = mad_recvq_size;
3077 	return 0;
3078 
3079 error:
3080 	return ret;
3081 }
3082 
3083 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3084 {
3085 	if (!qp_info->qp)
3086 		return;
3087 
3088 	ib_destroy_qp(qp_info->qp);
3089 	kfree(qp_info->snoop_table);
3090 }
3091 
3092 /*
3093  * Open the port
3094  * Create the QP, PD, MR, and CQ if needed
3095  */
3096 static int ib_mad_port_open(struct ib_device *device,
3097 			    int port_num)
3098 {
3099 	int ret, cq_size;
3100 	struct ib_mad_port_private *port_priv;
3101 	unsigned long flags;
3102 	char name[sizeof "ib_mad123"];
3103 	int has_smi;
3104 
3105 	if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3106 		return -EFAULT;
3107 
3108 	if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3109 		    rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3110 		return -EFAULT;
3111 
3112 	/* Create new device info */
3113 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3114 	if (!port_priv)
3115 		return -ENOMEM;
3116 
3117 	port_priv->device = device;
3118 	port_priv->port_num = port_num;
3119 	spin_lock_init(&port_priv->reg_lock);
3120 	INIT_LIST_HEAD(&port_priv->agent_list);
3121 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
3122 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
3123 
3124 	cq_size = mad_sendq_size + mad_recvq_size;
3125 	has_smi = rdma_cap_ib_smi(device, port_num);
3126 	if (has_smi)
3127 		cq_size *= 2;
3128 
3129 	port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3130 			IB_POLL_WORKQUEUE);
3131 	if (IS_ERR(port_priv->cq)) {
3132 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3133 		ret = PTR_ERR(port_priv->cq);
3134 		goto error3;
3135 	}
3136 
3137 	port_priv->pd = ib_alloc_pd(device, 0);
3138 	if (IS_ERR(port_priv->pd)) {
3139 		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3140 		ret = PTR_ERR(port_priv->pd);
3141 		goto error4;
3142 	}
3143 
3144 	if (has_smi) {
3145 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3146 		if (ret)
3147 			goto error6;
3148 	}
3149 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3150 	if (ret)
3151 		goto error7;
3152 
3153 	snprintf(name, sizeof name, "ib_mad%d", port_num);
3154 	port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3155 	if (!port_priv->wq) {
3156 		ret = -ENOMEM;
3157 		goto error8;
3158 	}
3159 
3160 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3161 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3162 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3163 
3164 	ret = ib_mad_port_start(port_priv);
3165 	if (ret) {
3166 		dev_err(&device->dev, "Couldn't start port\n");
3167 		goto error9;
3168 	}
3169 
3170 	return 0;
3171 
3172 error9:
3173 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3174 	list_del_init(&port_priv->port_list);
3175 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3176 
3177 	destroy_workqueue(port_priv->wq);
3178 error8:
3179 	destroy_mad_qp(&port_priv->qp_info[1]);
3180 error7:
3181 	destroy_mad_qp(&port_priv->qp_info[0]);
3182 error6:
3183 	ib_dealloc_pd(port_priv->pd);
3184 error4:
3185 	ib_free_cq(port_priv->cq);
3186 	cleanup_recv_queue(&port_priv->qp_info[1]);
3187 	cleanup_recv_queue(&port_priv->qp_info[0]);
3188 error3:
3189 	kfree(port_priv);
3190 
3191 	return ret;
3192 }
3193 
3194 /*
3195  * Close the port
3196  * If there are no classes using the port, free the port
3197  * resources (CQ, MR, PD, QP) and remove the port's info structure
3198  */
3199 static int ib_mad_port_close(struct ib_device *device, int port_num)
3200 {
3201 	struct ib_mad_port_private *port_priv;
3202 	unsigned long flags;
3203 
3204 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3205 	port_priv = __ib_get_mad_port(device, port_num);
3206 	if (port_priv == NULL) {
3207 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3208 		dev_err(&device->dev, "Port %d not found\n", port_num);
3209 		return -ENODEV;
3210 	}
3211 	list_del_init(&port_priv->port_list);
3212 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3213 
3214 	destroy_workqueue(port_priv->wq);
3215 	destroy_mad_qp(&port_priv->qp_info[1]);
3216 	destroy_mad_qp(&port_priv->qp_info[0]);
3217 	ib_dealloc_pd(port_priv->pd);
3218 	ib_free_cq(port_priv->cq);
3219 	cleanup_recv_queue(&port_priv->qp_info[1]);
3220 	cleanup_recv_queue(&port_priv->qp_info[0]);
3221 	/* XXX: Handle deallocation of MAD registration tables */
3222 
3223 	kfree(port_priv);
3224 
3225 	return 0;
3226 }
3227 
3228 static void ib_mad_init_device(struct ib_device *device)
3229 {
3230 	int start, i;
3231 
3232 	start = rdma_start_port(device);
3233 
3234 	for (i = start; i <= rdma_end_port(device); i++) {
3235 		if (!rdma_cap_ib_mad(device, i))
3236 			continue;
3237 
3238 		if (ib_mad_port_open(device, i)) {
3239 			dev_err(&device->dev, "Couldn't open port %d\n", i);
3240 			goto error;
3241 		}
3242 		if (ib_agent_port_open(device, i)) {
3243 			dev_err(&device->dev,
3244 				"Couldn't open port %d for agents\n", i);
3245 			goto error_agent;
3246 		}
3247 	}
3248 	return;
3249 
3250 error_agent:
3251 	if (ib_mad_port_close(device, i))
3252 		dev_err(&device->dev, "Couldn't close port %d\n", i);
3253 
3254 error:
3255 	while (--i >= start) {
3256 		if (!rdma_cap_ib_mad(device, i))
3257 			continue;
3258 
3259 		if (ib_agent_port_close(device, i))
3260 			dev_err(&device->dev,
3261 				"Couldn't close port %d for agents\n", i);
3262 		if (ib_mad_port_close(device, i))
3263 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3264 	}
3265 }
3266 
3267 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3268 {
3269 	int i;
3270 
3271 	for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3272 		if (!rdma_cap_ib_mad(device, i))
3273 			continue;
3274 
3275 		if (ib_agent_port_close(device, i))
3276 			dev_err(&device->dev,
3277 				"Couldn't close port %d for agents\n", i);
3278 		if (ib_mad_port_close(device, i))
3279 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3280 	}
3281 }
3282 
3283 static struct ib_client mad_client = {
3284 	.name   = "mad",
3285 	.add = ib_mad_init_device,
3286 	.remove = ib_mad_remove_device
3287 };
3288 
3289 int ib_mad_init(void)
3290 {
3291 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3292 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3293 
3294 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3295 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3296 
3297 	INIT_LIST_HEAD(&ib_mad_port_list);
3298 
3299 	if (ib_register_client(&mad_client)) {
3300 		pr_err("Couldn't register ib_mad client\n");
3301 		return -EINVAL;
3302 	}
3303 
3304 	return 0;
3305 }
3306 
3307 void ib_mad_cleanup(void)
3308 {
3309 	ib_unregister_client(&mad_client);
3310 }
3311