xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision 7587eb18)
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, "ib_register_mad_agent: Invalid port\n");
320 		ret = ERR_PTR(-ENODEV);
321 		goto error1;
322 	}
323 
324 	/* Verify the QP requested is supported.  For example, Ethernet devices
325 	 * will not have QP0 */
326 	if (!port_priv->qp_info[qpn].qp) {
327 		dev_notice(&device->dev,
328 			   "ib_register_mad_agent: QP %d not supported\n", qpn);
329 		ret = ERR_PTR(-EPROTONOSUPPORT);
330 		goto error1;
331 	}
332 
333 	/* Allocate structures */
334 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
335 	if (!mad_agent_priv) {
336 		ret = ERR_PTR(-ENOMEM);
337 		goto error1;
338 	}
339 
340 	if (mad_reg_req) {
341 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
342 		if (!reg_req) {
343 			ret = ERR_PTR(-ENOMEM);
344 			goto error3;
345 		}
346 	}
347 
348 	/* Now, fill in the various structures */
349 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
350 	mad_agent_priv->reg_req = reg_req;
351 	mad_agent_priv->agent.rmpp_version = rmpp_version;
352 	mad_agent_priv->agent.device = device;
353 	mad_agent_priv->agent.recv_handler = recv_handler;
354 	mad_agent_priv->agent.send_handler = send_handler;
355 	mad_agent_priv->agent.context = context;
356 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
357 	mad_agent_priv->agent.port_num = port_num;
358 	mad_agent_priv->agent.flags = registration_flags;
359 	spin_lock_init(&mad_agent_priv->lock);
360 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
361 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
362 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
363 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
364 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
365 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
366 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
367 	atomic_set(&mad_agent_priv->refcount, 1);
368 	init_completion(&mad_agent_priv->comp);
369 
370 	spin_lock_irqsave(&port_priv->reg_lock, flags);
371 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
372 
373 	/*
374 	 * Make sure MAD registration (if supplied)
375 	 * is non overlapping with any existing ones
376 	 */
377 	if (mad_reg_req) {
378 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
379 		if (!is_vendor_class(mgmt_class)) {
380 			class = port_priv->version[mad_reg_req->
381 						   mgmt_class_version].class;
382 			if (class) {
383 				method = class->method_table[mgmt_class];
384 				if (method) {
385 					if (method_in_use(&method,
386 							   mad_reg_req))
387 						goto error4;
388 				}
389 			}
390 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
391 						  mgmt_class);
392 		} else {
393 			/* "New" vendor class range */
394 			vendor = port_priv->version[mad_reg_req->
395 						    mgmt_class_version].vendor;
396 			if (vendor) {
397 				vclass = vendor_class_index(mgmt_class);
398 				vendor_class = vendor->vendor_class[vclass];
399 				if (vendor_class) {
400 					if (is_vendor_method_in_use(
401 							vendor_class,
402 							mad_reg_req))
403 						goto error4;
404 				}
405 			}
406 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
407 		}
408 		if (ret2) {
409 			ret = ERR_PTR(ret2);
410 			goto error4;
411 		}
412 	}
413 
414 	/* Add mad agent into port's agent list */
415 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
416 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
417 
418 	return &mad_agent_priv->agent;
419 
420 error4:
421 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
422 	kfree(reg_req);
423 error3:
424 	kfree(mad_agent_priv);
425 error1:
426 	return ret;
427 }
428 EXPORT_SYMBOL(ib_register_mad_agent);
429 
430 static inline int is_snooping_sends(int mad_snoop_flags)
431 {
432 	return (mad_snoop_flags &
433 		(/*IB_MAD_SNOOP_POSTED_SENDS |
434 		 IB_MAD_SNOOP_RMPP_SENDS |*/
435 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
436 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
437 }
438 
439 static inline int is_snooping_recvs(int mad_snoop_flags)
440 {
441 	return (mad_snoop_flags &
442 		(IB_MAD_SNOOP_RECVS /*|
443 		 IB_MAD_SNOOP_RMPP_RECVS*/));
444 }
445 
446 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
447 				struct ib_mad_snoop_private *mad_snoop_priv)
448 {
449 	struct ib_mad_snoop_private **new_snoop_table;
450 	unsigned long flags;
451 	int i;
452 
453 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
454 	/* Check for empty slot in array. */
455 	for (i = 0; i < qp_info->snoop_table_size; i++)
456 		if (!qp_info->snoop_table[i])
457 			break;
458 
459 	if (i == qp_info->snoop_table_size) {
460 		/* Grow table. */
461 		new_snoop_table = krealloc(qp_info->snoop_table,
462 					   sizeof mad_snoop_priv *
463 					   (qp_info->snoop_table_size + 1),
464 					   GFP_ATOMIC);
465 		if (!new_snoop_table) {
466 			i = -ENOMEM;
467 			goto out;
468 		}
469 
470 		qp_info->snoop_table = new_snoop_table;
471 		qp_info->snoop_table_size++;
472 	}
473 	qp_info->snoop_table[i] = mad_snoop_priv;
474 	atomic_inc(&qp_info->snoop_count);
475 out:
476 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
477 	return i;
478 }
479 
480 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
481 					   u8 port_num,
482 					   enum ib_qp_type qp_type,
483 					   int mad_snoop_flags,
484 					   ib_mad_snoop_handler snoop_handler,
485 					   ib_mad_recv_handler recv_handler,
486 					   void *context)
487 {
488 	struct ib_mad_port_private *port_priv;
489 	struct ib_mad_agent *ret;
490 	struct ib_mad_snoop_private *mad_snoop_priv;
491 	int qpn;
492 
493 	/* Validate parameters */
494 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
495 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
496 		ret = ERR_PTR(-EINVAL);
497 		goto error1;
498 	}
499 	qpn = get_spl_qp_index(qp_type);
500 	if (qpn == -1) {
501 		ret = ERR_PTR(-EINVAL);
502 		goto error1;
503 	}
504 	port_priv = ib_get_mad_port(device, port_num);
505 	if (!port_priv) {
506 		ret = ERR_PTR(-ENODEV);
507 		goto error1;
508 	}
509 	/* Allocate structures */
510 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
511 	if (!mad_snoop_priv) {
512 		ret = ERR_PTR(-ENOMEM);
513 		goto error1;
514 	}
515 
516 	/* Now, fill in the various structures */
517 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
518 	mad_snoop_priv->agent.device = device;
519 	mad_snoop_priv->agent.recv_handler = recv_handler;
520 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
521 	mad_snoop_priv->agent.context = context;
522 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
523 	mad_snoop_priv->agent.port_num = port_num;
524 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
525 	init_completion(&mad_snoop_priv->comp);
526 	mad_snoop_priv->snoop_index = register_snoop_agent(
527 						&port_priv->qp_info[qpn],
528 						mad_snoop_priv);
529 	if (mad_snoop_priv->snoop_index < 0) {
530 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
531 		goto error2;
532 	}
533 
534 	atomic_set(&mad_snoop_priv->refcount, 1);
535 	return &mad_snoop_priv->agent;
536 
537 error2:
538 	kfree(mad_snoop_priv);
539 error1:
540 	return ret;
541 }
542 EXPORT_SYMBOL(ib_register_mad_snoop);
543 
544 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
545 {
546 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
547 		complete(&mad_agent_priv->comp);
548 }
549 
550 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
551 {
552 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
553 		complete(&mad_snoop_priv->comp);
554 }
555 
556 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
557 {
558 	struct ib_mad_port_private *port_priv;
559 	unsigned long flags;
560 
561 	/* Note that we could still be handling received MADs */
562 
563 	/*
564 	 * Canceling all sends results in dropping received response
565 	 * MADs, preventing us from queuing additional work
566 	 */
567 	cancel_mads(mad_agent_priv);
568 	port_priv = mad_agent_priv->qp_info->port_priv;
569 	cancel_delayed_work(&mad_agent_priv->timed_work);
570 
571 	spin_lock_irqsave(&port_priv->reg_lock, flags);
572 	remove_mad_reg_req(mad_agent_priv);
573 	list_del(&mad_agent_priv->agent_list);
574 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
575 
576 	flush_workqueue(port_priv->wq);
577 	ib_cancel_rmpp_recvs(mad_agent_priv);
578 
579 	deref_mad_agent(mad_agent_priv);
580 	wait_for_completion(&mad_agent_priv->comp);
581 
582 	kfree(mad_agent_priv->reg_req);
583 	kfree(mad_agent_priv);
584 }
585 
586 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
587 {
588 	struct ib_mad_qp_info *qp_info;
589 	unsigned long flags;
590 
591 	qp_info = mad_snoop_priv->qp_info;
592 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
593 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
594 	atomic_dec(&qp_info->snoop_count);
595 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
596 
597 	deref_snoop_agent(mad_snoop_priv);
598 	wait_for_completion(&mad_snoop_priv->comp);
599 
600 	kfree(mad_snoop_priv);
601 }
602 
603 /*
604  * ib_unregister_mad_agent - Unregisters a client from using MAD services
605  */
606 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
607 {
608 	struct ib_mad_agent_private *mad_agent_priv;
609 	struct ib_mad_snoop_private *mad_snoop_priv;
610 
611 	/* If the TID is zero, the agent can only snoop. */
612 	if (mad_agent->hi_tid) {
613 		mad_agent_priv = container_of(mad_agent,
614 					      struct ib_mad_agent_private,
615 					      agent);
616 		unregister_mad_agent(mad_agent_priv);
617 	} else {
618 		mad_snoop_priv = container_of(mad_agent,
619 					      struct ib_mad_snoop_private,
620 					      agent);
621 		unregister_mad_snoop(mad_snoop_priv);
622 	}
623 	return 0;
624 }
625 EXPORT_SYMBOL(ib_unregister_mad_agent);
626 
627 static void dequeue_mad(struct ib_mad_list_head *mad_list)
628 {
629 	struct ib_mad_queue *mad_queue;
630 	unsigned long flags;
631 
632 	BUG_ON(!mad_list->mad_queue);
633 	mad_queue = mad_list->mad_queue;
634 	spin_lock_irqsave(&mad_queue->lock, flags);
635 	list_del(&mad_list->list);
636 	mad_queue->count--;
637 	spin_unlock_irqrestore(&mad_queue->lock, flags);
638 }
639 
640 static void snoop_send(struct ib_mad_qp_info *qp_info,
641 		       struct ib_mad_send_buf *send_buf,
642 		       struct ib_mad_send_wc *mad_send_wc,
643 		       int mad_snoop_flags)
644 {
645 	struct ib_mad_snoop_private *mad_snoop_priv;
646 	unsigned long flags;
647 	int i;
648 
649 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
650 	for (i = 0; i < qp_info->snoop_table_size; i++) {
651 		mad_snoop_priv = qp_info->snoop_table[i];
652 		if (!mad_snoop_priv ||
653 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
654 			continue;
655 
656 		atomic_inc(&mad_snoop_priv->refcount);
657 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
658 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
659 						    send_buf, mad_send_wc);
660 		deref_snoop_agent(mad_snoop_priv);
661 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
662 	}
663 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
664 }
665 
666 static void snoop_recv(struct ib_mad_qp_info *qp_info,
667 		       struct ib_mad_recv_wc *mad_recv_wc,
668 		       int mad_snoop_flags)
669 {
670 	struct ib_mad_snoop_private *mad_snoop_priv;
671 	unsigned long flags;
672 	int i;
673 
674 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
675 	for (i = 0; i < qp_info->snoop_table_size; i++) {
676 		mad_snoop_priv = qp_info->snoop_table[i];
677 		if (!mad_snoop_priv ||
678 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
679 			continue;
680 
681 		atomic_inc(&mad_snoop_priv->refcount);
682 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
683 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
684 						   mad_recv_wc);
685 		deref_snoop_agent(mad_snoop_priv);
686 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
687 	}
688 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
689 }
690 
691 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
692 		u16 pkey_index, u8 port_num, struct ib_wc *wc)
693 {
694 	memset(wc, 0, sizeof *wc);
695 	wc->wr_cqe = cqe;
696 	wc->status = IB_WC_SUCCESS;
697 	wc->opcode = IB_WC_RECV;
698 	wc->pkey_index = pkey_index;
699 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
700 	wc->src_qp = IB_QP0;
701 	wc->qp = qp;
702 	wc->slid = slid;
703 	wc->sl = 0;
704 	wc->dlid_path_bits = 0;
705 	wc->port_num = port_num;
706 }
707 
708 static size_t mad_priv_size(const struct ib_mad_private *mp)
709 {
710 	return sizeof(struct ib_mad_private) + mp->mad_size;
711 }
712 
713 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
714 {
715 	size_t size = sizeof(struct ib_mad_private) + mad_size;
716 	struct ib_mad_private *ret = kzalloc(size, flags);
717 
718 	if (ret)
719 		ret->mad_size = mad_size;
720 
721 	return ret;
722 }
723 
724 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
725 {
726 	return rdma_max_mad_size(port_priv->device, port_priv->port_num);
727 }
728 
729 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
730 {
731 	return sizeof(struct ib_grh) + mp->mad_size;
732 }
733 
734 /*
735  * Return 0 if SMP is to be sent
736  * Return 1 if SMP was consumed locally (whether or not solicited)
737  * Return < 0 if error
738  */
739 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
740 				  struct ib_mad_send_wr_private *mad_send_wr)
741 {
742 	int ret = 0;
743 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
744 	struct opa_smp *opa_smp = (struct opa_smp *)smp;
745 	unsigned long flags;
746 	struct ib_mad_local_private *local;
747 	struct ib_mad_private *mad_priv;
748 	struct ib_mad_port_private *port_priv;
749 	struct ib_mad_agent_private *recv_mad_agent = NULL;
750 	struct ib_device *device = mad_agent_priv->agent.device;
751 	u8 port_num;
752 	struct ib_wc mad_wc;
753 	struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
754 	size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
755 	u16 out_mad_pkey_index = 0;
756 	u16 drslid;
757 	bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
758 				    mad_agent_priv->qp_info->port_priv->port_num);
759 
760 	if (rdma_cap_ib_switch(device) &&
761 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
762 		port_num = send_wr->port_num;
763 	else
764 		port_num = mad_agent_priv->agent.port_num;
765 
766 	/*
767 	 * Directed route handling starts if the initial LID routed part of
768 	 * a request or the ending LID routed part of a response is empty.
769 	 * If we are at the start of the LID routed part, don't update the
770 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
771 	 */
772 	if (opa && smp->class_version == OPA_SMP_CLASS_VERSION) {
773 		u32 opa_drslid;
774 
775 		if ((opa_get_smp_direction(opa_smp)
776 		     ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
777 		     OPA_LID_PERMISSIVE &&
778 		     opa_smi_handle_dr_smp_send(opa_smp,
779 						rdma_cap_ib_switch(device),
780 						port_num) == IB_SMI_DISCARD) {
781 			ret = -EINVAL;
782 			dev_err(&device->dev, "OPA Invalid directed route\n");
783 			goto out;
784 		}
785 		opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
786 		if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
787 		    opa_drslid & 0xffff0000) {
788 			ret = -EINVAL;
789 			dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
790 			       opa_drslid);
791 			goto out;
792 		}
793 		drslid = (u16)(opa_drslid & 0x0000ffff);
794 
795 		/* Check to post send on QP or process locally */
796 		if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
797 		    opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
798 			goto out;
799 	} else {
800 		if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
801 		     IB_LID_PERMISSIVE &&
802 		     smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
803 		     IB_SMI_DISCARD) {
804 			ret = -EINVAL;
805 			dev_err(&device->dev, "Invalid directed route\n");
806 			goto out;
807 		}
808 		drslid = be16_to_cpu(smp->dr_slid);
809 
810 		/* Check to post send on QP or process locally */
811 		if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
812 		    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
813 			goto out;
814 	}
815 
816 	local = kmalloc(sizeof *local, GFP_ATOMIC);
817 	if (!local) {
818 		ret = -ENOMEM;
819 		dev_err(&device->dev, "No memory for ib_mad_local_private\n");
820 		goto out;
821 	}
822 	local->mad_priv = NULL;
823 	local->recv_mad_agent = NULL;
824 	mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
825 	if (!mad_priv) {
826 		ret = -ENOMEM;
827 		dev_err(&device->dev, "No memory for local response MAD\n");
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 			dev_err(&send_buf->mad_agent->device->dev,
951 				"alloc_send_rmpp_segs: RMPP mem alloc failed for len %zd, gfp %#x\n",
952 				sizeof (*seg) + seg_size, gfp_mask);
953 			free_send_rmpp_list(send_wr);
954 			return -ENOMEM;
955 		}
956 		seg->num = ++send_buf->seg_count;
957 		list_add_tail(&seg->list, &send_wr->rmpp_list);
958 	}
959 
960 	/* Zero any padding */
961 	if (pad)
962 		memset(seg->data + seg_size - pad, 0, pad);
963 
964 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
965 					  agent.rmpp_version;
966 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
967 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
968 
969 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
970 					struct ib_rmpp_segment, list);
971 	send_wr->last_ack_seg = send_wr->cur_seg;
972 	return 0;
973 }
974 
975 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
976 {
977 	return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
978 }
979 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
980 
981 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
982 					    u32 remote_qpn, u16 pkey_index,
983 					    int rmpp_active,
984 					    int hdr_len, int data_len,
985 					    gfp_t gfp_mask,
986 					    u8 base_version)
987 {
988 	struct ib_mad_agent_private *mad_agent_priv;
989 	struct ib_mad_send_wr_private *mad_send_wr;
990 	int pad, message_size, ret, size;
991 	void *buf;
992 	size_t mad_size;
993 	bool opa;
994 
995 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
996 				      agent);
997 
998 	opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
999 
1000 	if (opa && base_version == OPA_MGMT_BASE_VERSION)
1001 		mad_size = sizeof(struct opa_mad);
1002 	else
1003 		mad_size = sizeof(struct ib_mad);
1004 
1005 	pad = get_pad_size(hdr_len, data_len, mad_size);
1006 	message_size = hdr_len + data_len + pad;
1007 
1008 	if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1009 		if (!rmpp_active && message_size > mad_size)
1010 			return ERR_PTR(-EINVAL);
1011 	} else
1012 		if (rmpp_active || message_size > mad_size)
1013 			return ERR_PTR(-EINVAL);
1014 
1015 	size = rmpp_active ? hdr_len : mad_size;
1016 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1017 	if (!buf)
1018 		return ERR_PTR(-ENOMEM);
1019 
1020 	mad_send_wr = buf + size;
1021 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1022 	mad_send_wr->send_buf.mad = buf;
1023 	mad_send_wr->send_buf.hdr_len = hdr_len;
1024 	mad_send_wr->send_buf.data_len = data_len;
1025 	mad_send_wr->pad = pad;
1026 
1027 	mad_send_wr->mad_agent_priv = mad_agent_priv;
1028 	mad_send_wr->sg_list[0].length = hdr_len;
1029 	mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1030 
1031 	/* OPA MADs don't have to be the full 2048 bytes */
1032 	if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1033 	    data_len < mad_size - hdr_len)
1034 		mad_send_wr->sg_list[1].length = data_len;
1035 	else
1036 		mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1037 
1038 	mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1039 
1040 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1041 
1042 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1043 	mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1044 	mad_send_wr->send_wr.wr.num_sge = 2;
1045 	mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1046 	mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1047 	mad_send_wr->send_wr.remote_qpn = remote_qpn;
1048 	mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1049 	mad_send_wr->send_wr.pkey_index = pkey_index;
1050 
1051 	if (rmpp_active) {
1052 		ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1053 		if (ret) {
1054 			kfree(buf);
1055 			return ERR_PTR(ret);
1056 		}
1057 	}
1058 
1059 	mad_send_wr->send_buf.mad_agent = mad_agent;
1060 	atomic_inc(&mad_agent_priv->refcount);
1061 	return &mad_send_wr->send_buf;
1062 }
1063 EXPORT_SYMBOL(ib_create_send_mad);
1064 
1065 int ib_get_mad_data_offset(u8 mgmt_class)
1066 {
1067 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1068 		return IB_MGMT_SA_HDR;
1069 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1070 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1071 		 (mgmt_class == IB_MGMT_CLASS_BIS))
1072 		return IB_MGMT_DEVICE_HDR;
1073 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1074 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1075 		return IB_MGMT_VENDOR_HDR;
1076 	else
1077 		return IB_MGMT_MAD_HDR;
1078 }
1079 EXPORT_SYMBOL(ib_get_mad_data_offset);
1080 
1081 int ib_is_mad_class_rmpp(u8 mgmt_class)
1082 {
1083 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1084 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1085 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1086 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
1087 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1088 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1089 		return 1;
1090 	return 0;
1091 }
1092 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1093 
1094 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1095 {
1096 	struct ib_mad_send_wr_private *mad_send_wr;
1097 	struct list_head *list;
1098 
1099 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1100 				   send_buf);
1101 	list = &mad_send_wr->cur_seg->list;
1102 
1103 	if (mad_send_wr->cur_seg->num < seg_num) {
1104 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
1105 			if (mad_send_wr->cur_seg->num == seg_num)
1106 				break;
1107 	} else if (mad_send_wr->cur_seg->num > seg_num) {
1108 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1109 			if (mad_send_wr->cur_seg->num == seg_num)
1110 				break;
1111 	}
1112 	return mad_send_wr->cur_seg->data;
1113 }
1114 EXPORT_SYMBOL(ib_get_rmpp_segment);
1115 
1116 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1117 {
1118 	if (mad_send_wr->send_buf.seg_count)
1119 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1120 					   mad_send_wr->seg_num);
1121 	else
1122 		return mad_send_wr->send_buf.mad +
1123 		       mad_send_wr->send_buf.hdr_len;
1124 }
1125 
1126 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1127 {
1128 	struct ib_mad_agent_private *mad_agent_priv;
1129 	struct ib_mad_send_wr_private *mad_send_wr;
1130 
1131 	mad_agent_priv = container_of(send_buf->mad_agent,
1132 				      struct ib_mad_agent_private, agent);
1133 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1134 				   send_buf);
1135 
1136 	free_send_rmpp_list(mad_send_wr);
1137 	kfree(send_buf->mad);
1138 	deref_mad_agent(mad_agent_priv);
1139 }
1140 EXPORT_SYMBOL(ib_free_send_mad);
1141 
1142 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1143 {
1144 	struct ib_mad_qp_info *qp_info;
1145 	struct list_head *list;
1146 	struct ib_send_wr *bad_send_wr;
1147 	struct ib_mad_agent *mad_agent;
1148 	struct ib_sge *sge;
1149 	unsigned long flags;
1150 	int ret;
1151 
1152 	/* Set WR ID to find mad_send_wr upon completion */
1153 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1154 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1155 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1156 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1157 
1158 	mad_agent = mad_send_wr->send_buf.mad_agent;
1159 	sge = mad_send_wr->sg_list;
1160 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1161 					mad_send_wr->send_buf.mad,
1162 					sge[0].length,
1163 					DMA_TO_DEVICE);
1164 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1165 		return -ENOMEM;
1166 
1167 	mad_send_wr->header_mapping = sge[0].addr;
1168 
1169 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1170 					ib_get_payload(mad_send_wr),
1171 					sge[1].length,
1172 					DMA_TO_DEVICE);
1173 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1174 		ib_dma_unmap_single(mad_agent->device,
1175 				    mad_send_wr->header_mapping,
1176 				    sge[0].length, DMA_TO_DEVICE);
1177 		return -ENOMEM;
1178 	}
1179 	mad_send_wr->payload_mapping = sge[1].addr;
1180 
1181 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1182 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1183 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1184 				   &bad_send_wr);
1185 		list = &qp_info->send_queue.list;
1186 	} else {
1187 		ret = 0;
1188 		list = &qp_info->overflow_list;
1189 	}
1190 
1191 	if (!ret) {
1192 		qp_info->send_queue.count++;
1193 		list_add_tail(&mad_send_wr->mad_list.list, list);
1194 	}
1195 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1196 	if (ret) {
1197 		ib_dma_unmap_single(mad_agent->device,
1198 				    mad_send_wr->header_mapping,
1199 				    sge[0].length, DMA_TO_DEVICE);
1200 		ib_dma_unmap_single(mad_agent->device,
1201 				    mad_send_wr->payload_mapping,
1202 				    sge[1].length, DMA_TO_DEVICE);
1203 	}
1204 	return ret;
1205 }
1206 
1207 /*
1208  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1209  *  with the registered client
1210  */
1211 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1212 		     struct ib_mad_send_buf **bad_send_buf)
1213 {
1214 	struct ib_mad_agent_private *mad_agent_priv;
1215 	struct ib_mad_send_buf *next_send_buf;
1216 	struct ib_mad_send_wr_private *mad_send_wr;
1217 	unsigned long flags;
1218 	int ret = -EINVAL;
1219 
1220 	/* Walk list of send WRs and post each on send list */
1221 	for (; send_buf; send_buf = next_send_buf) {
1222 
1223 		mad_send_wr = container_of(send_buf,
1224 					   struct ib_mad_send_wr_private,
1225 					   send_buf);
1226 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1227 
1228 		if (!send_buf->mad_agent->send_handler ||
1229 		    (send_buf->timeout_ms &&
1230 		     !send_buf->mad_agent->recv_handler)) {
1231 			ret = -EINVAL;
1232 			goto error;
1233 		}
1234 
1235 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1236 			if (mad_agent_priv->agent.rmpp_version) {
1237 				ret = -EINVAL;
1238 				goto error;
1239 			}
1240 		}
1241 
1242 		/*
1243 		 * Save pointer to next work request to post in case the
1244 		 * current one completes, and the user modifies the work
1245 		 * request associated with the completion
1246 		 */
1247 		next_send_buf = send_buf->next;
1248 		mad_send_wr->send_wr.ah = send_buf->ah;
1249 
1250 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1251 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1252 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1253 						     mad_send_wr);
1254 			if (ret < 0)		/* error */
1255 				goto error;
1256 			else if (ret == 1)	/* locally consumed */
1257 				continue;
1258 		}
1259 
1260 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1261 		/* Timeout will be updated after send completes */
1262 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1263 		mad_send_wr->max_retries = send_buf->retries;
1264 		mad_send_wr->retries_left = send_buf->retries;
1265 		send_buf->retries = 0;
1266 		/* Reference for work request to QP + response */
1267 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1268 		mad_send_wr->status = IB_WC_SUCCESS;
1269 
1270 		/* Reference MAD agent until send completes */
1271 		atomic_inc(&mad_agent_priv->refcount);
1272 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1273 		list_add_tail(&mad_send_wr->agent_list,
1274 			      &mad_agent_priv->send_list);
1275 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1276 
1277 		if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1278 			ret = ib_send_rmpp_mad(mad_send_wr);
1279 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1280 				ret = ib_send_mad(mad_send_wr);
1281 		} else
1282 			ret = ib_send_mad(mad_send_wr);
1283 		if (ret < 0) {
1284 			/* Fail send request */
1285 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1286 			list_del(&mad_send_wr->agent_list);
1287 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1288 			atomic_dec(&mad_agent_priv->refcount);
1289 			goto error;
1290 		}
1291 	}
1292 	return 0;
1293 error:
1294 	if (bad_send_buf)
1295 		*bad_send_buf = send_buf;
1296 	return ret;
1297 }
1298 EXPORT_SYMBOL(ib_post_send_mad);
1299 
1300 /*
1301  * ib_free_recv_mad - Returns data buffers used to receive
1302  *  a MAD to the access layer
1303  */
1304 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1305 {
1306 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1307 	struct ib_mad_private_header *mad_priv_hdr;
1308 	struct ib_mad_private *priv;
1309 	struct list_head free_list;
1310 
1311 	INIT_LIST_HEAD(&free_list);
1312 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1313 
1314 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1315 					&free_list, list) {
1316 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1317 					   recv_buf);
1318 		mad_priv_hdr = container_of(mad_recv_wc,
1319 					    struct ib_mad_private_header,
1320 					    recv_wc);
1321 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1322 				    header);
1323 		kfree(priv);
1324 	}
1325 }
1326 EXPORT_SYMBOL(ib_free_recv_mad);
1327 
1328 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1329 					u8 rmpp_version,
1330 					ib_mad_send_handler send_handler,
1331 					ib_mad_recv_handler recv_handler,
1332 					void *context)
1333 {
1334 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1335 }
1336 EXPORT_SYMBOL(ib_redirect_mad_qp);
1337 
1338 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1339 		      struct ib_wc *wc)
1340 {
1341 	dev_err(&mad_agent->device->dev,
1342 		"ib_process_mad_wc() not implemented yet\n");
1343 	return 0;
1344 }
1345 EXPORT_SYMBOL(ib_process_mad_wc);
1346 
1347 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1348 			 struct ib_mad_reg_req *mad_reg_req)
1349 {
1350 	int i;
1351 
1352 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1353 		if ((*method)->agent[i]) {
1354 			pr_err("Method %d already in use\n", i);
1355 			return -EINVAL;
1356 		}
1357 	}
1358 	return 0;
1359 }
1360 
1361 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1362 {
1363 	/* Allocate management method table */
1364 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1365 	if (!*method) {
1366 		pr_err("No memory for ib_mad_mgmt_method_table\n");
1367 		return -ENOMEM;
1368 	}
1369 
1370 	return 0;
1371 }
1372 
1373 /*
1374  * Check to see if there are any methods still in use
1375  */
1376 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1377 {
1378 	int i;
1379 
1380 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1381 		if (method->agent[i])
1382 			return 1;
1383 	return 0;
1384 }
1385 
1386 /*
1387  * Check to see if there are any method tables for this class still in use
1388  */
1389 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1390 {
1391 	int i;
1392 
1393 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1394 		if (class->method_table[i])
1395 			return 1;
1396 	return 0;
1397 }
1398 
1399 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1400 {
1401 	int i;
1402 
1403 	for (i = 0; i < MAX_MGMT_OUI; i++)
1404 		if (vendor_class->method_table[i])
1405 			return 1;
1406 	return 0;
1407 }
1408 
1409 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1410 			   const char *oui)
1411 {
1412 	int i;
1413 
1414 	for (i = 0; i < MAX_MGMT_OUI; i++)
1415 		/* Is there matching OUI for this vendor class ? */
1416 		if (!memcmp(vendor_class->oui[i], oui, 3))
1417 			return i;
1418 
1419 	return -1;
1420 }
1421 
1422 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1423 {
1424 	int i;
1425 
1426 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1427 		if (vendor->vendor_class[i])
1428 			return 1;
1429 
1430 	return 0;
1431 }
1432 
1433 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1434 				     struct ib_mad_agent_private *agent)
1435 {
1436 	int i;
1437 
1438 	/* Remove any methods for this mad agent */
1439 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1440 		if (method->agent[i] == agent) {
1441 			method->agent[i] = NULL;
1442 		}
1443 	}
1444 }
1445 
1446 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1447 			      struct ib_mad_agent_private *agent_priv,
1448 			      u8 mgmt_class)
1449 {
1450 	struct ib_mad_port_private *port_priv;
1451 	struct ib_mad_mgmt_class_table **class;
1452 	struct ib_mad_mgmt_method_table **method;
1453 	int i, ret;
1454 
1455 	port_priv = agent_priv->qp_info->port_priv;
1456 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1457 	if (!*class) {
1458 		/* Allocate management class table for "new" class version */
1459 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1460 		if (!*class) {
1461 			dev_err(&agent_priv->agent.device->dev,
1462 				"No memory for ib_mad_mgmt_class_table\n");
1463 			ret = -ENOMEM;
1464 			goto error1;
1465 		}
1466 
1467 		/* Allocate method table for this management class */
1468 		method = &(*class)->method_table[mgmt_class];
1469 		if ((ret = allocate_method_table(method)))
1470 			goto error2;
1471 	} else {
1472 		method = &(*class)->method_table[mgmt_class];
1473 		if (!*method) {
1474 			/* Allocate method table for this management class */
1475 			if ((ret = allocate_method_table(method)))
1476 				goto error1;
1477 		}
1478 	}
1479 
1480 	/* Now, make sure methods are not already in use */
1481 	if (method_in_use(method, mad_reg_req))
1482 		goto error3;
1483 
1484 	/* Finally, add in methods being registered */
1485 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1486 		(*method)->agent[i] = agent_priv;
1487 
1488 	return 0;
1489 
1490 error3:
1491 	/* Remove any methods for this mad agent */
1492 	remove_methods_mad_agent(*method, agent_priv);
1493 	/* Now, check to see if there are any methods in use */
1494 	if (!check_method_table(*method)) {
1495 		/* If not, release management method table */
1496 		kfree(*method);
1497 		*method = NULL;
1498 	}
1499 	ret = -EINVAL;
1500 	goto error1;
1501 error2:
1502 	kfree(*class);
1503 	*class = NULL;
1504 error1:
1505 	return ret;
1506 }
1507 
1508 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1509 			   struct ib_mad_agent_private *agent_priv)
1510 {
1511 	struct ib_mad_port_private *port_priv;
1512 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1513 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1514 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1515 	struct ib_mad_mgmt_method_table **method;
1516 	int i, ret = -ENOMEM;
1517 	u8 vclass;
1518 
1519 	/* "New" vendor (with OUI) class */
1520 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1521 	port_priv = agent_priv->qp_info->port_priv;
1522 	vendor_table = &port_priv->version[
1523 				mad_reg_req->mgmt_class_version].vendor;
1524 	if (!*vendor_table) {
1525 		/* Allocate mgmt vendor class table for "new" class version */
1526 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1527 		if (!vendor) {
1528 			dev_err(&agent_priv->agent.device->dev,
1529 				"No memory for ib_mad_mgmt_vendor_class_table\n");
1530 			goto error1;
1531 		}
1532 
1533 		*vendor_table = vendor;
1534 	}
1535 	if (!(*vendor_table)->vendor_class[vclass]) {
1536 		/* Allocate table for this management vendor class */
1537 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1538 		if (!vendor_class) {
1539 			dev_err(&agent_priv->agent.device->dev,
1540 				"No memory for ib_mad_mgmt_vendor_class\n");
1541 			goto error2;
1542 		}
1543 
1544 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1545 	}
1546 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1547 		/* Is there matching OUI for this vendor class ? */
1548 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1549 			    mad_reg_req->oui, 3)) {
1550 			method = &(*vendor_table)->vendor_class[
1551 						vclass]->method_table[i];
1552 			BUG_ON(!*method);
1553 			goto check_in_use;
1554 		}
1555 	}
1556 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1557 		/* OUI slot available ? */
1558 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1559 				vclass]->oui[i])) {
1560 			method = &(*vendor_table)->vendor_class[
1561 				vclass]->method_table[i];
1562 			BUG_ON(*method);
1563 			/* Allocate method table for this OUI */
1564 			if ((ret = allocate_method_table(method)))
1565 				goto error3;
1566 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1567 			       mad_reg_req->oui, 3);
1568 			goto check_in_use;
1569 		}
1570 	}
1571 	dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1572 	goto error3;
1573 
1574 check_in_use:
1575 	/* Now, make sure methods are not already in use */
1576 	if (method_in_use(method, mad_reg_req))
1577 		goto error4;
1578 
1579 	/* Finally, add in methods being registered */
1580 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1581 		(*method)->agent[i] = agent_priv;
1582 
1583 	return 0;
1584 
1585 error4:
1586 	/* Remove any methods for this mad agent */
1587 	remove_methods_mad_agent(*method, agent_priv);
1588 	/* Now, check to see if there are any methods in use */
1589 	if (!check_method_table(*method)) {
1590 		/* If not, release management method table */
1591 		kfree(*method);
1592 		*method = NULL;
1593 	}
1594 	ret = -EINVAL;
1595 error3:
1596 	if (vendor_class) {
1597 		(*vendor_table)->vendor_class[vclass] = NULL;
1598 		kfree(vendor_class);
1599 	}
1600 error2:
1601 	if (vendor) {
1602 		*vendor_table = NULL;
1603 		kfree(vendor);
1604 	}
1605 error1:
1606 	return ret;
1607 }
1608 
1609 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1610 {
1611 	struct ib_mad_port_private *port_priv;
1612 	struct ib_mad_mgmt_class_table *class;
1613 	struct ib_mad_mgmt_method_table *method;
1614 	struct ib_mad_mgmt_vendor_class_table *vendor;
1615 	struct ib_mad_mgmt_vendor_class *vendor_class;
1616 	int index;
1617 	u8 mgmt_class;
1618 
1619 	/*
1620 	 * Was MAD registration request supplied
1621 	 * with original registration ?
1622 	 */
1623 	if (!agent_priv->reg_req) {
1624 		goto out;
1625 	}
1626 
1627 	port_priv = agent_priv->qp_info->port_priv;
1628 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1629 	class = port_priv->version[
1630 			agent_priv->reg_req->mgmt_class_version].class;
1631 	if (!class)
1632 		goto vendor_check;
1633 
1634 	method = class->method_table[mgmt_class];
1635 	if (method) {
1636 		/* Remove any methods for this mad agent */
1637 		remove_methods_mad_agent(method, agent_priv);
1638 		/* Now, check to see if there are any methods still in use */
1639 		if (!check_method_table(method)) {
1640 			/* If not, release management method table */
1641 			kfree(method);
1642 			class->method_table[mgmt_class] = NULL;
1643 			/* Any management classes left ? */
1644 			if (!check_class_table(class)) {
1645 				/* If not, release management class table */
1646 				kfree(class);
1647 				port_priv->version[
1648 					agent_priv->reg_req->
1649 					mgmt_class_version].class = NULL;
1650 			}
1651 		}
1652 	}
1653 
1654 vendor_check:
1655 	if (!is_vendor_class(mgmt_class))
1656 		goto out;
1657 
1658 	/* normalize mgmt_class to vendor range 2 */
1659 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1660 	vendor = port_priv->version[
1661 			agent_priv->reg_req->mgmt_class_version].vendor;
1662 
1663 	if (!vendor)
1664 		goto out;
1665 
1666 	vendor_class = vendor->vendor_class[mgmt_class];
1667 	if (vendor_class) {
1668 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1669 		if (index < 0)
1670 			goto out;
1671 		method = vendor_class->method_table[index];
1672 		if (method) {
1673 			/* Remove any methods for this mad agent */
1674 			remove_methods_mad_agent(method, agent_priv);
1675 			/*
1676 			 * Now, check to see if there are
1677 			 * any methods still in use
1678 			 */
1679 			if (!check_method_table(method)) {
1680 				/* If not, release management method table */
1681 				kfree(method);
1682 				vendor_class->method_table[index] = NULL;
1683 				memset(vendor_class->oui[index], 0, 3);
1684 				/* Any OUIs left ? */
1685 				if (!check_vendor_class(vendor_class)) {
1686 					/* If not, release vendor class table */
1687 					kfree(vendor_class);
1688 					vendor->vendor_class[mgmt_class] = NULL;
1689 					/* Any other vendor classes left ? */
1690 					if (!check_vendor_table(vendor)) {
1691 						kfree(vendor);
1692 						port_priv->version[
1693 							agent_priv->reg_req->
1694 							mgmt_class_version].
1695 							vendor = NULL;
1696 					}
1697 				}
1698 			}
1699 		}
1700 	}
1701 
1702 out:
1703 	return;
1704 }
1705 
1706 static struct ib_mad_agent_private *
1707 find_mad_agent(struct ib_mad_port_private *port_priv,
1708 	       const struct ib_mad_hdr *mad_hdr)
1709 {
1710 	struct ib_mad_agent_private *mad_agent = NULL;
1711 	unsigned long flags;
1712 
1713 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1714 	if (ib_response_mad(mad_hdr)) {
1715 		u32 hi_tid;
1716 		struct ib_mad_agent_private *entry;
1717 
1718 		/*
1719 		 * Routing is based on high 32 bits of transaction ID
1720 		 * of MAD.
1721 		 */
1722 		hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1723 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1724 			if (entry->agent.hi_tid == hi_tid) {
1725 				mad_agent = entry;
1726 				break;
1727 			}
1728 		}
1729 	} else {
1730 		struct ib_mad_mgmt_class_table *class;
1731 		struct ib_mad_mgmt_method_table *method;
1732 		struct ib_mad_mgmt_vendor_class_table *vendor;
1733 		struct ib_mad_mgmt_vendor_class *vendor_class;
1734 		const struct ib_vendor_mad *vendor_mad;
1735 		int index;
1736 
1737 		/*
1738 		 * Routing is based on version, class, and method
1739 		 * For "newer" vendor MADs, also based on OUI
1740 		 */
1741 		if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1742 			goto out;
1743 		if (!is_vendor_class(mad_hdr->mgmt_class)) {
1744 			class = port_priv->version[
1745 					mad_hdr->class_version].class;
1746 			if (!class)
1747 				goto out;
1748 			if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1749 			    IB_MGMT_MAX_METHODS)
1750 				goto out;
1751 			method = class->method_table[convert_mgmt_class(
1752 							mad_hdr->mgmt_class)];
1753 			if (method)
1754 				mad_agent = method->agent[mad_hdr->method &
1755 							  ~IB_MGMT_METHOD_RESP];
1756 		} else {
1757 			vendor = port_priv->version[
1758 					mad_hdr->class_version].vendor;
1759 			if (!vendor)
1760 				goto out;
1761 			vendor_class = vendor->vendor_class[vendor_class_index(
1762 						mad_hdr->mgmt_class)];
1763 			if (!vendor_class)
1764 				goto out;
1765 			/* Find matching OUI */
1766 			vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1767 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1768 			if (index == -1)
1769 				goto out;
1770 			method = vendor_class->method_table[index];
1771 			if (method) {
1772 				mad_agent = method->agent[mad_hdr->method &
1773 							  ~IB_MGMT_METHOD_RESP];
1774 			}
1775 		}
1776 	}
1777 
1778 	if (mad_agent) {
1779 		if (mad_agent->agent.recv_handler)
1780 			atomic_inc(&mad_agent->refcount);
1781 		else {
1782 			dev_notice(&port_priv->device->dev,
1783 				   "No receive handler for client %p on port %d\n",
1784 				   &mad_agent->agent, port_priv->port_num);
1785 			mad_agent = NULL;
1786 		}
1787 	}
1788 out:
1789 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1790 
1791 	return mad_agent;
1792 }
1793 
1794 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1795 			const struct ib_mad_qp_info *qp_info,
1796 			bool opa)
1797 {
1798 	int valid = 0;
1799 	u32 qp_num = qp_info->qp->qp_num;
1800 
1801 	/* Make sure MAD base version is understood */
1802 	if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1803 	    (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1804 		pr_err("MAD received with unsupported base version %d %s\n",
1805 		       mad_hdr->base_version, opa ? "(opa)" : "");
1806 		goto out;
1807 	}
1808 
1809 	/* Filter SMI packets sent to other than QP0 */
1810 	if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1811 	    (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1812 		if (qp_num == 0)
1813 			valid = 1;
1814 	} else {
1815 		/* CM attributes other than ClassPortInfo only use Send method */
1816 		if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1817 		    (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1818 		    (mad_hdr->method != IB_MGMT_METHOD_SEND))
1819 			goto out;
1820 		/* Filter GSI packets sent to QP0 */
1821 		if (qp_num != 0)
1822 			valid = 1;
1823 	}
1824 
1825 out:
1826 	return valid;
1827 }
1828 
1829 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1830 			    const struct ib_mad_hdr *mad_hdr)
1831 {
1832 	struct ib_rmpp_mad *rmpp_mad;
1833 
1834 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1835 	return !mad_agent_priv->agent.rmpp_version ||
1836 		!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1837 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1838 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1839 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1840 }
1841 
1842 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1843 				     const struct ib_mad_recv_wc *rwc)
1844 {
1845 	return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1846 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1847 }
1848 
1849 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1850 				   const struct ib_mad_send_wr_private *wr,
1851 				   const struct ib_mad_recv_wc *rwc )
1852 {
1853 	struct ib_ah_attr attr;
1854 	u8 send_resp, rcv_resp;
1855 	union ib_gid sgid;
1856 	struct ib_device *device = mad_agent_priv->agent.device;
1857 	u8 port_num = mad_agent_priv->agent.port_num;
1858 	u8 lmc;
1859 
1860 	send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1861 	rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1862 
1863 	if (send_resp == rcv_resp)
1864 		/* both requests, or both responses. GIDs different */
1865 		return 0;
1866 
1867 	if (ib_query_ah(wr->send_buf.ah, &attr))
1868 		/* Assume not equal, to avoid false positives. */
1869 		return 0;
1870 
1871 	if (!!(attr.ah_flags & IB_AH_GRH) !=
1872 	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1873 		/* one has GID, other does not.  Assume different */
1874 		return 0;
1875 
1876 	if (!send_resp && rcv_resp) {
1877 		/* is request/response. */
1878 		if (!(attr.ah_flags & IB_AH_GRH)) {
1879 			if (ib_get_cached_lmc(device, port_num, &lmc))
1880 				return 0;
1881 			return (!lmc || !((attr.src_path_bits ^
1882 					   rwc->wc->dlid_path_bits) &
1883 					  ((1 << lmc) - 1)));
1884 		} else {
1885 			if (ib_get_cached_gid(device, port_num,
1886 					      attr.grh.sgid_index, &sgid, NULL))
1887 				return 0;
1888 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1889 				       16);
1890 		}
1891 	}
1892 
1893 	if (!(attr.ah_flags & IB_AH_GRH))
1894 		return attr.dlid == rwc->wc->slid;
1895 	else
1896 		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1897 			       16);
1898 }
1899 
1900 static inline int is_direct(u8 class)
1901 {
1902 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1903 }
1904 
1905 struct ib_mad_send_wr_private*
1906 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1907 		 const struct ib_mad_recv_wc *wc)
1908 {
1909 	struct ib_mad_send_wr_private *wr;
1910 	const struct ib_mad_hdr *mad_hdr;
1911 
1912 	mad_hdr = &wc->recv_buf.mad->mad_hdr;
1913 
1914 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1915 		if ((wr->tid == mad_hdr->tid) &&
1916 		    rcv_has_same_class(wr, wc) &&
1917 		    /*
1918 		     * Don't check GID for direct routed MADs.
1919 		     * These might have permissive LIDs.
1920 		     */
1921 		    (is_direct(mad_hdr->mgmt_class) ||
1922 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1923 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1924 	}
1925 
1926 	/*
1927 	 * It's possible to receive the response before we've
1928 	 * been notified that the send has completed
1929 	 */
1930 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1931 		if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1932 		    wr->tid == mad_hdr->tid &&
1933 		    wr->timeout &&
1934 		    rcv_has_same_class(wr, wc) &&
1935 		    /*
1936 		     * Don't check GID for direct routed MADs.
1937 		     * These might have permissive LIDs.
1938 		     */
1939 		    (is_direct(mad_hdr->mgmt_class) ||
1940 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1941 			/* Verify request has not been canceled */
1942 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1943 	}
1944 	return NULL;
1945 }
1946 
1947 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1948 {
1949 	mad_send_wr->timeout = 0;
1950 	if (mad_send_wr->refcount == 1)
1951 		list_move_tail(&mad_send_wr->agent_list,
1952 			      &mad_send_wr->mad_agent_priv->done_list);
1953 }
1954 
1955 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1956 				 struct ib_mad_recv_wc *mad_recv_wc)
1957 {
1958 	struct ib_mad_send_wr_private *mad_send_wr;
1959 	struct ib_mad_send_wc mad_send_wc;
1960 	unsigned long flags;
1961 
1962 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1963 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1964 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1965 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1966 						      mad_recv_wc);
1967 		if (!mad_recv_wc) {
1968 			deref_mad_agent(mad_agent_priv);
1969 			return;
1970 		}
1971 	}
1972 
1973 	/* Complete corresponding request */
1974 	if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1975 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1976 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1977 		if (!mad_send_wr) {
1978 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1979 			if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
1980 			   && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
1981 			   && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
1982 					& IB_MGMT_RMPP_FLAG_ACTIVE)) {
1983 				/* user rmpp is in effect
1984 				 * and this is an active RMPP MAD
1985 				 */
1986 				mad_agent_priv->agent.recv_handler(
1987 						&mad_agent_priv->agent, NULL,
1988 						mad_recv_wc);
1989 				atomic_dec(&mad_agent_priv->refcount);
1990 			} else {
1991 				/* not user rmpp, revert to normal behavior and
1992 				 * drop the mad */
1993 				ib_free_recv_mad(mad_recv_wc);
1994 				deref_mad_agent(mad_agent_priv);
1995 				return;
1996 			}
1997 		} else {
1998 			ib_mark_mad_done(mad_send_wr);
1999 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2000 
2001 			/* Defined behavior is to complete response before request */
2002 			mad_agent_priv->agent.recv_handler(
2003 					&mad_agent_priv->agent,
2004 					&mad_send_wr->send_buf,
2005 					mad_recv_wc);
2006 			atomic_dec(&mad_agent_priv->refcount);
2007 
2008 			mad_send_wc.status = IB_WC_SUCCESS;
2009 			mad_send_wc.vendor_err = 0;
2010 			mad_send_wc.send_buf = &mad_send_wr->send_buf;
2011 			ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2012 		}
2013 	} else {
2014 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2015 						   mad_recv_wc);
2016 		deref_mad_agent(mad_agent_priv);
2017 	}
2018 }
2019 
2020 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2021 				     const struct ib_mad_qp_info *qp_info,
2022 				     const struct ib_wc *wc,
2023 				     int port_num,
2024 				     struct ib_mad_private *recv,
2025 				     struct ib_mad_private *response)
2026 {
2027 	enum smi_forward_action retsmi;
2028 	struct ib_smp *smp = (struct ib_smp *)recv->mad;
2029 
2030 	if (smi_handle_dr_smp_recv(smp,
2031 				   rdma_cap_ib_switch(port_priv->device),
2032 				   port_num,
2033 				   port_priv->device->phys_port_cnt) ==
2034 				   IB_SMI_DISCARD)
2035 		return IB_SMI_DISCARD;
2036 
2037 	retsmi = smi_check_forward_dr_smp(smp);
2038 	if (retsmi == IB_SMI_LOCAL)
2039 		return IB_SMI_HANDLE;
2040 
2041 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2042 		if (smi_handle_dr_smp_send(smp,
2043 					   rdma_cap_ib_switch(port_priv->device),
2044 					   port_num) == IB_SMI_DISCARD)
2045 			return IB_SMI_DISCARD;
2046 
2047 		if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2048 			return IB_SMI_DISCARD;
2049 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2050 		/* forward case for switches */
2051 		memcpy(response, recv, mad_priv_size(response));
2052 		response->header.recv_wc.wc = &response->header.wc;
2053 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2054 		response->header.recv_wc.recv_buf.grh = &response->grh;
2055 
2056 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2057 				    &response->grh, wc,
2058 				    port_priv->device,
2059 				    smi_get_fwd_port(smp),
2060 				    qp_info->qp->qp_num,
2061 				    response->mad_size,
2062 				    false);
2063 
2064 		return IB_SMI_DISCARD;
2065 	}
2066 	return IB_SMI_HANDLE;
2067 }
2068 
2069 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2070 				    struct ib_mad_private *response,
2071 				    size_t *resp_len, bool opa)
2072 {
2073 	const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2074 	struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2075 
2076 	if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2077 	    recv_hdr->method == IB_MGMT_METHOD_SET) {
2078 		memcpy(response, recv, mad_priv_size(response));
2079 		response->header.recv_wc.wc = &response->header.wc;
2080 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2081 		response->header.recv_wc.recv_buf.grh = &response->grh;
2082 		resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2083 		resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2084 		if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2085 			resp_hdr->status |= IB_SMP_DIRECTION;
2086 
2087 		if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2088 			if (recv_hdr->mgmt_class ==
2089 			    IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2090 			    recv_hdr->mgmt_class ==
2091 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2092 				*resp_len = opa_get_smp_header_size(
2093 							(struct opa_smp *)recv->mad);
2094 			else
2095 				*resp_len = sizeof(struct ib_mad_hdr);
2096 		}
2097 
2098 		return true;
2099 	} else {
2100 		return false;
2101 	}
2102 }
2103 
2104 static enum smi_action
2105 handle_opa_smi(struct ib_mad_port_private *port_priv,
2106 	       struct ib_mad_qp_info *qp_info,
2107 	       struct ib_wc *wc,
2108 	       int port_num,
2109 	       struct ib_mad_private *recv,
2110 	       struct ib_mad_private *response)
2111 {
2112 	enum smi_forward_action retsmi;
2113 	struct opa_smp *smp = (struct opa_smp *)recv->mad;
2114 
2115 	if (opa_smi_handle_dr_smp_recv(smp,
2116 				   rdma_cap_ib_switch(port_priv->device),
2117 				   port_num,
2118 				   port_priv->device->phys_port_cnt) ==
2119 				   IB_SMI_DISCARD)
2120 		return IB_SMI_DISCARD;
2121 
2122 	retsmi = opa_smi_check_forward_dr_smp(smp);
2123 	if (retsmi == IB_SMI_LOCAL)
2124 		return IB_SMI_HANDLE;
2125 
2126 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2127 		if (opa_smi_handle_dr_smp_send(smp,
2128 					   rdma_cap_ib_switch(port_priv->device),
2129 					   port_num) == IB_SMI_DISCARD)
2130 			return IB_SMI_DISCARD;
2131 
2132 		if (opa_smi_check_local_smp(smp, port_priv->device) ==
2133 		    IB_SMI_DISCARD)
2134 			return IB_SMI_DISCARD;
2135 
2136 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2137 		/* forward case for switches */
2138 		memcpy(response, recv, mad_priv_size(response));
2139 		response->header.recv_wc.wc = &response->header.wc;
2140 		response->header.recv_wc.recv_buf.opa_mad =
2141 				(struct opa_mad *)response->mad;
2142 		response->header.recv_wc.recv_buf.grh = &response->grh;
2143 
2144 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2145 				    &response->grh, wc,
2146 				    port_priv->device,
2147 				    opa_smi_get_fwd_port(smp),
2148 				    qp_info->qp->qp_num,
2149 				    recv->header.wc.byte_len,
2150 				    true);
2151 
2152 		return IB_SMI_DISCARD;
2153 	}
2154 
2155 	return IB_SMI_HANDLE;
2156 }
2157 
2158 static enum smi_action
2159 handle_smi(struct ib_mad_port_private *port_priv,
2160 	   struct ib_mad_qp_info *qp_info,
2161 	   struct ib_wc *wc,
2162 	   int port_num,
2163 	   struct ib_mad_private *recv,
2164 	   struct ib_mad_private *response,
2165 	   bool opa)
2166 {
2167 	struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2168 
2169 	if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2170 	    mad_hdr->class_version == OPA_SMI_CLASS_VERSION)
2171 		return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2172 				      response);
2173 
2174 	return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2175 }
2176 
2177 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2178 {
2179 	struct ib_mad_port_private *port_priv = cq->cq_context;
2180 	struct ib_mad_list_head *mad_list =
2181 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2182 	struct ib_mad_qp_info *qp_info;
2183 	struct ib_mad_private_header *mad_priv_hdr;
2184 	struct ib_mad_private *recv, *response = NULL;
2185 	struct ib_mad_agent_private *mad_agent;
2186 	int port_num;
2187 	int ret = IB_MAD_RESULT_SUCCESS;
2188 	size_t mad_size;
2189 	u16 resp_mad_pkey_index = 0;
2190 	bool opa;
2191 
2192 	if (list_empty_careful(&port_priv->port_list))
2193 		return;
2194 
2195 	if (wc->status != IB_WC_SUCCESS) {
2196 		/*
2197 		 * Receive errors indicate that the QP has entered the error
2198 		 * state - error handling/shutdown code will cleanup
2199 		 */
2200 		return;
2201 	}
2202 
2203 	qp_info = mad_list->mad_queue->qp_info;
2204 	dequeue_mad(mad_list);
2205 
2206 	opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2207 			       qp_info->port_priv->port_num);
2208 
2209 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2210 				    mad_list);
2211 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2212 	ib_dma_unmap_single(port_priv->device,
2213 			    recv->header.mapping,
2214 			    mad_priv_dma_size(recv),
2215 			    DMA_FROM_DEVICE);
2216 
2217 	/* Setup MAD receive work completion from "normal" work completion */
2218 	recv->header.wc = *wc;
2219 	recv->header.recv_wc.wc = &recv->header.wc;
2220 
2221 	if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2222 		recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2223 		recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2224 	} else {
2225 		recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2226 		recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2227 	}
2228 
2229 	recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2230 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
2231 
2232 	if (atomic_read(&qp_info->snoop_count))
2233 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2234 
2235 	/* Validate MAD */
2236 	if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2237 		goto out;
2238 
2239 	mad_size = recv->mad_size;
2240 	response = alloc_mad_private(mad_size, GFP_KERNEL);
2241 	if (!response) {
2242 		dev_err(&port_priv->device->dev,
2243 			"%s: no memory for response buffer\n", __func__);
2244 		goto out;
2245 	}
2246 
2247 	if (rdma_cap_ib_switch(port_priv->device))
2248 		port_num = wc->port_num;
2249 	else
2250 		port_num = port_priv->port_num;
2251 
2252 	if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2253 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2254 		if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2255 			       response, opa)
2256 		    == IB_SMI_DISCARD)
2257 			goto out;
2258 	}
2259 
2260 	/* Give driver "right of first refusal" on incoming MAD */
2261 	if (port_priv->device->process_mad) {
2262 		ret = port_priv->device->process_mad(port_priv->device, 0,
2263 						     port_priv->port_num,
2264 						     wc, &recv->grh,
2265 						     (const struct ib_mad_hdr *)recv->mad,
2266 						     recv->mad_size,
2267 						     (struct ib_mad_hdr *)response->mad,
2268 						     &mad_size, &resp_mad_pkey_index);
2269 
2270 		if (opa)
2271 			wc->pkey_index = resp_mad_pkey_index;
2272 
2273 		if (ret & IB_MAD_RESULT_SUCCESS) {
2274 			if (ret & IB_MAD_RESULT_CONSUMED)
2275 				goto out;
2276 			if (ret & IB_MAD_RESULT_REPLY) {
2277 				agent_send_response((const struct ib_mad_hdr *)response->mad,
2278 						    &recv->grh, wc,
2279 						    port_priv->device,
2280 						    port_num,
2281 						    qp_info->qp->qp_num,
2282 						    mad_size, opa);
2283 				goto out;
2284 			}
2285 		}
2286 	}
2287 
2288 	mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2289 	if (mad_agent) {
2290 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2291 		/*
2292 		 * recv is freed up in error cases in ib_mad_complete_recv
2293 		 * or via recv_handler in ib_mad_complete_recv()
2294 		 */
2295 		recv = NULL;
2296 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2297 		   generate_unmatched_resp(recv, response, &mad_size, opa)) {
2298 		agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2299 				    port_priv->device, port_num,
2300 				    qp_info->qp->qp_num, mad_size, opa);
2301 	}
2302 
2303 out:
2304 	/* Post another receive request for this QP */
2305 	if (response) {
2306 		ib_mad_post_receive_mads(qp_info, response);
2307 		kfree(recv);
2308 	} else
2309 		ib_mad_post_receive_mads(qp_info, recv);
2310 }
2311 
2312 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2313 {
2314 	struct ib_mad_send_wr_private *mad_send_wr;
2315 	unsigned long delay;
2316 
2317 	if (list_empty(&mad_agent_priv->wait_list)) {
2318 		cancel_delayed_work(&mad_agent_priv->timed_work);
2319 	} else {
2320 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2321 					 struct ib_mad_send_wr_private,
2322 					 agent_list);
2323 
2324 		if (time_after(mad_agent_priv->timeout,
2325 			       mad_send_wr->timeout)) {
2326 			mad_agent_priv->timeout = mad_send_wr->timeout;
2327 			delay = mad_send_wr->timeout - jiffies;
2328 			if ((long)delay <= 0)
2329 				delay = 1;
2330 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2331 					 &mad_agent_priv->timed_work, delay);
2332 		}
2333 	}
2334 }
2335 
2336 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2337 {
2338 	struct ib_mad_agent_private *mad_agent_priv;
2339 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2340 	struct list_head *list_item;
2341 	unsigned long delay;
2342 
2343 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2344 	list_del(&mad_send_wr->agent_list);
2345 
2346 	delay = mad_send_wr->timeout;
2347 	mad_send_wr->timeout += jiffies;
2348 
2349 	if (delay) {
2350 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2351 			temp_mad_send_wr = list_entry(list_item,
2352 						struct ib_mad_send_wr_private,
2353 						agent_list);
2354 			if (time_after(mad_send_wr->timeout,
2355 				       temp_mad_send_wr->timeout))
2356 				break;
2357 		}
2358 	}
2359 	else
2360 		list_item = &mad_agent_priv->wait_list;
2361 	list_add(&mad_send_wr->agent_list, list_item);
2362 
2363 	/* Reschedule a work item if we have a shorter timeout */
2364 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2365 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2366 				 &mad_agent_priv->timed_work, delay);
2367 }
2368 
2369 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2370 			  int timeout_ms)
2371 {
2372 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2373 	wait_for_response(mad_send_wr);
2374 }
2375 
2376 /*
2377  * Process a send work completion
2378  */
2379 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2380 			     struct ib_mad_send_wc *mad_send_wc)
2381 {
2382 	struct ib_mad_agent_private	*mad_agent_priv;
2383 	unsigned long			flags;
2384 	int				ret;
2385 
2386 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2387 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2388 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2389 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2390 		if (ret == IB_RMPP_RESULT_CONSUMED)
2391 			goto done;
2392 	} else
2393 		ret = IB_RMPP_RESULT_UNHANDLED;
2394 
2395 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2396 	    mad_send_wr->status == IB_WC_SUCCESS) {
2397 		mad_send_wr->status = mad_send_wc->status;
2398 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2399 	}
2400 
2401 	if (--mad_send_wr->refcount > 0) {
2402 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2403 		    mad_send_wr->status == IB_WC_SUCCESS) {
2404 			wait_for_response(mad_send_wr);
2405 		}
2406 		goto done;
2407 	}
2408 
2409 	/* Remove send from MAD agent and notify client of completion */
2410 	list_del(&mad_send_wr->agent_list);
2411 	adjust_timeout(mad_agent_priv);
2412 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2413 
2414 	if (mad_send_wr->status != IB_WC_SUCCESS )
2415 		mad_send_wc->status = mad_send_wr->status;
2416 	if (ret == IB_RMPP_RESULT_INTERNAL)
2417 		ib_rmpp_send_handler(mad_send_wc);
2418 	else
2419 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2420 						   mad_send_wc);
2421 
2422 	/* Release reference on agent taken when sending */
2423 	deref_mad_agent(mad_agent_priv);
2424 	return;
2425 done:
2426 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2427 }
2428 
2429 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2430 {
2431 	struct ib_mad_port_private *port_priv = cq->cq_context;
2432 	struct ib_mad_list_head *mad_list =
2433 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2434 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2435 	struct ib_mad_qp_info		*qp_info;
2436 	struct ib_mad_queue		*send_queue;
2437 	struct ib_send_wr		*bad_send_wr;
2438 	struct ib_mad_send_wc		mad_send_wc;
2439 	unsigned long flags;
2440 	int ret;
2441 
2442 	if (list_empty_careful(&port_priv->port_list))
2443 		return;
2444 
2445 	if (wc->status != IB_WC_SUCCESS) {
2446 		if (!ib_mad_send_error(port_priv, wc))
2447 			return;
2448 	}
2449 
2450 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2451 				   mad_list);
2452 	send_queue = mad_list->mad_queue;
2453 	qp_info = send_queue->qp_info;
2454 
2455 retry:
2456 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2457 			    mad_send_wr->header_mapping,
2458 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2459 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2460 			    mad_send_wr->payload_mapping,
2461 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2462 	queued_send_wr = NULL;
2463 	spin_lock_irqsave(&send_queue->lock, flags);
2464 	list_del(&mad_list->list);
2465 
2466 	/* Move queued send to the send queue */
2467 	if (send_queue->count-- > send_queue->max_active) {
2468 		mad_list = container_of(qp_info->overflow_list.next,
2469 					struct ib_mad_list_head, list);
2470 		queued_send_wr = container_of(mad_list,
2471 					struct ib_mad_send_wr_private,
2472 					mad_list);
2473 		list_move_tail(&mad_list->list, &send_queue->list);
2474 	}
2475 	spin_unlock_irqrestore(&send_queue->lock, flags);
2476 
2477 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2478 	mad_send_wc.status = wc->status;
2479 	mad_send_wc.vendor_err = wc->vendor_err;
2480 	if (atomic_read(&qp_info->snoop_count))
2481 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2482 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2483 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2484 
2485 	if (queued_send_wr) {
2486 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2487 				   &bad_send_wr);
2488 		if (ret) {
2489 			dev_err(&port_priv->device->dev,
2490 				"ib_post_send failed: %d\n", ret);
2491 			mad_send_wr = queued_send_wr;
2492 			wc->status = IB_WC_LOC_QP_OP_ERR;
2493 			goto retry;
2494 		}
2495 	}
2496 }
2497 
2498 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2499 {
2500 	struct ib_mad_send_wr_private *mad_send_wr;
2501 	struct ib_mad_list_head *mad_list;
2502 	unsigned long flags;
2503 
2504 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2505 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2506 		mad_send_wr = container_of(mad_list,
2507 					   struct ib_mad_send_wr_private,
2508 					   mad_list);
2509 		mad_send_wr->retry = 1;
2510 	}
2511 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2512 }
2513 
2514 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2515 		struct ib_wc *wc)
2516 {
2517 	struct ib_mad_list_head *mad_list =
2518 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2519 	struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2520 	struct ib_mad_send_wr_private *mad_send_wr;
2521 	int ret;
2522 
2523 	/*
2524 	 * Send errors will transition the QP to SQE - move
2525 	 * QP to RTS and repost flushed work requests
2526 	 */
2527 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2528 				   mad_list);
2529 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2530 		if (mad_send_wr->retry) {
2531 			/* Repost send */
2532 			struct ib_send_wr *bad_send_wr;
2533 
2534 			mad_send_wr->retry = 0;
2535 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2536 					&bad_send_wr);
2537 			if (!ret)
2538 				return false;
2539 		}
2540 	} else {
2541 		struct ib_qp_attr *attr;
2542 
2543 		/* Transition QP to RTS and fail offending send */
2544 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2545 		if (attr) {
2546 			attr->qp_state = IB_QPS_RTS;
2547 			attr->cur_qp_state = IB_QPS_SQE;
2548 			ret = ib_modify_qp(qp_info->qp, attr,
2549 					   IB_QP_STATE | IB_QP_CUR_STATE);
2550 			kfree(attr);
2551 			if (ret)
2552 				dev_err(&port_priv->device->dev,
2553 					"%s - ib_modify_qp to RTS: %d\n",
2554 					__func__, ret);
2555 			else
2556 				mark_sends_for_retry(qp_info);
2557 		}
2558 	}
2559 
2560 	return true;
2561 }
2562 
2563 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2564 {
2565 	unsigned long flags;
2566 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2567 	struct ib_mad_send_wc mad_send_wc;
2568 	struct list_head cancel_list;
2569 
2570 	INIT_LIST_HEAD(&cancel_list);
2571 
2572 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2573 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2574 				 &mad_agent_priv->send_list, agent_list) {
2575 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2576 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2577 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2578 		}
2579 	}
2580 
2581 	/* Empty wait list to prevent receives from finding a request */
2582 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2583 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2584 
2585 	/* Report all cancelled requests */
2586 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2587 	mad_send_wc.vendor_err = 0;
2588 
2589 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2590 				 &cancel_list, agent_list) {
2591 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2592 		list_del(&mad_send_wr->agent_list);
2593 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2594 						   &mad_send_wc);
2595 		atomic_dec(&mad_agent_priv->refcount);
2596 	}
2597 }
2598 
2599 static struct ib_mad_send_wr_private*
2600 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2601 	     struct ib_mad_send_buf *send_buf)
2602 {
2603 	struct ib_mad_send_wr_private *mad_send_wr;
2604 
2605 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2606 			    agent_list) {
2607 		if (&mad_send_wr->send_buf == send_buf)
2608 			return mad_send_wr;
2609 	}
2610 
2611 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2612 			    agent_list) {
2613 		if (is_rmpp_data_mad(mad_agent_priv,
2614 				     mad_send_wr->send_buf.mad) &&
2615 		    &mad_send_wr->send_buf == send_buf)
2616 			return mad_send_wr;
2617 	}
2618 	return NULL;
2619 }
2620 
2621 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2622 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2623 {
2624 	struct ib_mad_agent_private *mad_agent_priv;
2625 	struct ib_mad_send_wr_private *mad_send_wr;
2626 	unsigned long flags;
2627 	int active;
2628 
2629 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2630 				      agent);
2631 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2632 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2633 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2634 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2635 		return -EINVAL;
2636 	}
2637 
2638 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2639 	if (!timeout_ms) {
2640 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2641 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2642 	}
2643 
2644 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2645 	if (active)
2646 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2647 	else
2648 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2649 
2650 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2651 	return 0;
2652 }
2653 EXPORT_SYMBOL(ib_modify_mad);
2654 
2655 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2656 		   struct ib_mad_send_buf *send_buf)
2657 {
2658 	ib_modify_mad(mad_agent, send_buf, 0);
2659 }
2660 EXPORT_SYMBOL(ib_cancel_mad);
2661 
2662 static void local_completions(struct work_struct *work)
2663 {
2664 	struct ib_mad_agent_private *mad_agent_priv;
2665 	struct ib_mad_local_private *local;
2666 	struct ib_mad_agent_private *recv_mad_agent;
2667 	unsigned long flags;
2668 	int free_mad;
2669 	struct ib_wc wc;
2670 	struct ib_mad_send_wc mad_send_wc;
2671 	bool opa;
2672 
2673 	mad_agent_priv =
2674 		container_of(work, struct ib_mad_agent_private, local_work);
2675 
2676 	opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2677 			       mad_agent_priv->qp_info->port_priv->port_num);
2678 
2679 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2680 	while (!list_empty(&mad_agent_priv->local_list)) {
2681 		local = list_entry(mad_agent_priv->local_list.next,
2682 				   struct ib_mad_local_private,
2683 				   completion_list);
2684 		list_del(&local->completion_list);
2685 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2686 		free_mad = 0;
2687 		if (local->mad_priv) {
2688 			u8 base_version;
2689 			recv_mad_agent = local->recv_mad_agent;
2690 			if (!recv_mad_agent) {
2691 				dev_err(&mad_agent_priv->agent.device->dev,
2692 					"No receive MAD agent for local completion\n");
2693 				free_mad = 1;
2694 				goto local_send_completion;
2695 			}
2696 
2697 			/*
2698 			 * Defined behavior is to complete response
2699 			 * before request
2700 			 */
2701 			build_smp_wc(recv_mad_agent->agent.qp,
2702 				     local->mad_send_wr->send_wr.wr.wr_cqe,
2703 				     be16_to_cpu(IB_LID_PERMISSIVE),
2704 				     local->mad_send_wr->send_wr.pkey_index,
2705 				     recv_mad_agent->agent.port_num, &wc);
2706 
2707 			local->mad_priv->header.recv_wc.wc = &wc;
2708 
2709 			base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2710 			if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2711 				local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2712 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2713 			} else {
2714 				local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2715 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2716 			}
2717 
2718 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2719 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2720 				 &local->mad_priv->header.recv_wc.rmpp_list);
2721 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2722 			local->mad_priv->header.recv_wc.recv_buf.mad =
2723 						(struct ib_mad *)local->mad_priv->mad;
2724 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2725 				snoop_recv(recv_mad_agent->qp_info,
2726 					  &local->mad_priv->header.recv_wc,
2727 					   IB_MAD_SNOOP_RECVS);
2728 			recv_mad_agent->agent.recv_handler(
2729 						&recv_mad_agent->agent,
2730 						&local->mad_send_wr->send_buf,
2731 						&local->mad_priv->header.recv_wc);
2732 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2733 			atomic_dec(&recv_mad_agent->refcount);
2734 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2735 		}
2736 
2737 local_send_completion:
2738 		/* Complete send */
2739 		mad_send_wc.status = IB_WC_SUCCESS;
2740 		mad_send_wc.vendor_err = 0;
2741 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2742 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2743 			snoop_send(mad_agent_priv->qp_info,
2744 				   &local->mad_send_wr->send_buf,
2745 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2746 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2747 						   &mad_send_wc);
2748 
2749 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2750 		atomic_dec(&mad_agent_priv->refcount);
2751 		if (free_mad)
2752 			kfree(local->mad_priv);
2753 		kfree(local);
2754 	}
2755 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2756 }
2757 
2758 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2759 {
2760 	int ret;
2761 
2762 	if (!mad_send_wr->retries_left)
2763 		return -ETIMEDOUT;
2764 
2765 	mad_send_wr->retries_left--;
2766 	mad_send_wr->send_buf.retries++;
2767 
2768 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2769 
2770 	if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2771 		ret = ib_retry_rmpp(mad_send_wr);
2772 		switch (ret) {
2773 		case IB_RMPP_RESULT_UNHANDLED:
2774 			ret = ib_send_mad(mad_send_wr);
2775 			break;
2776 		case IB_RMPP_RESULT_CONSUMED:
2777 			ret = 0;
2778 			break;
2779 		default:
2780 			ret = -ECOMM;
2781 			break;
2782 		}
2783 	} else
2784 		ret = ib_send_mad(mad_send_wr);
2785 
2786 	if (!ret) {
2787 		mad_send_wr->refcount++;
2788 		list_add_tail(&mad_send_wr->agent_list,
2789 			      &mad_send_wr->mad_agent_priv->send_list);
2790 	}
2791 	return ret;
2792 }
2793 
2794 static void timeout_sends(struct work_struct *work)
2795 {
2796 	struct ib_mad_agent_private *mad_agent_priv;
2797 	struct ib_mad_send_wr_private *mad_send_wr;
2798 	struct ib_mad_send_wc mad_send_wc;
2799 	unsigned long flags, delay;
2800 
2801 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2802 				      timed_work.work);
2803 	mad_send_wc.vendor_err = 0;
2804 
2805 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2806 	while (!list_empty(&mad_agent_priv->wait_list)) {
2807 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2808 					 struct ib_mad_send_wr_private,
2809 					 agent_list);
2810 
2811 		if (time_after(mad_send_wr->timeout, jiffies)) {
2812 			delay = mad_send_wr->timeout - jiffies;
2813 			if ((long)delay <= 0)
2814 				delay = 1;
2815 			queue_delayed_work(mad_agent_priv->qp_info->
2816 					   port_priv->wq,
2817 					   &mad_agent_priv->timed_work, delay);
2818 			break;
2819 		}
2820 
2821 		list_del(&mad_send_wr->agent_list);
2822 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2823 		    !retry_send(mad_send_wr))
2824 			continue;
2825 
2826 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2827 
2828 		if (mad_send_wr->status == IB_WC_SUCCESS)
2829 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2830 		else
2831 			mad_send_wc.status = mad_send_wr->status;
2832 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2833 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2834 						   &mad_send_wc);
2835 
2836 		atomic_dec(&mad_agent_priv->refcount);
2837 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2838 	}
2839 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2840 }
2841 
2842 /*
2843  * Allocate receive MADs and post receive WRs for them
2844  */
2845 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2846 				    struct ib_mad_private *mad)
2847 {
2848 	unsigned long flags;
2849 	int post, ret;
2850 	struct ib_mad_private *mad_priv;
2851 	struct ib_sge sg_list;
2852 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2853 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2854 
2855 	/* Initialize common scatter list fields */
2856 	sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2857 
2858 	/* Initialize common receive WR fields */
2859 	recv_wr.next = NULL;
2860 	recv_wr.sg_list = &sg_list;
2861 	recv_wr.num_sge = 1;
2862 
2863 	do {
2864 		/* Allocate and map receive buffer */
2865 		if (mad) {
2866 			mad_priv = mad;
2867 			mad = NULL;
2868 		} else {
2869 			mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2870 						     GFP_ATOMIC);
2871 			if (!mad_priv) {
2872 				dev_err(&qp_info->port_priv->device->dev,
2873 					"No memory for receive buffer\n");
2874 				ret = -ENOMEM;
2875 				break;
2876 			}
2877 		}
2878 		sg_list.length = mad_priv_dma_size(mad_priv);
2879 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2880 						 &mad_priv->grh,
2881 						 mad_priv_dma_size(mad_priv),
2882 						 DMA_FROM_DEVICE);
2883 		if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2884 						  sg_list.addr))) {
2885 			ret = -ENOMEM;
2886 			break;
2887 		}
2888 		mad_priv->header.mapping = sg_list.addr;
2889 		mad_priv->header.mad_list.mad_queue = recv_queue;
2890 		mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2891 		recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2892 
2893 		/* Post receive WR */
2894 		spin_lock_irqsave(&recv_queue->lock, flags);
2895 		post = (++recv_queue->count < recv_queue->max_active);
2896 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2897 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2898 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2899 		if (ret) {
2900 			spin_lock_irqsave(&recv_queue->lock, flags);
2901 			list_del(&mad_priv->header.mad_list.list);
2902 			recv_queue->count--;
2903 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2904 			ib_dma_unmap_single(qp_info->port_priv->device,
2905 					    mad_priv->header.mapping,
2906 					    mad_priv_dma_size(mad_priv),
2907 					    DMA_FROM_DEVICE);
2908 			kfree(mad_priv);
2909 			dev_err(&qp_info->port_priv->device->dev,
2910 				"ib_post_recv failed: %d\n", ret);
2911 			break;
2912 		}
2913 	} while (post);
2914 
2915 	return ret;
2916 }
2917 
2918 /*
2919  * Return all the posted receive MADs
2920  */
2921 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2922 {
2923 	struct ib_mad_private_header *mad_priv_hdr;
2924 	struct ib_mad_private *recv;
2925 	struct ib_mad_list_head *mad_list;
2926 
2927 	if (!qp_info->qp)
2928 		return;
2929 
2930 	while (!list_empty(&qp_info->recv_queue.list)) {
2931 
2932 		mad_list = list_entry(qp_info->recv_queue.list.next,
2933 				      struct ib_mad_list_head, list);
2934 		mad_priv_hdr = container_of(mad_list,
2935 					    struct ib_mad_private_header,
2936 					    mad_list);
2937 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2938 				    header);
2939 
2940 		/* Remove from posted receive MAD list */
2941 		list_del(&mad_list->list);
2942 
2943 		ib_dma_unmap_single(qp_info->port_priv->device,
2944 				    recv->header.mapping,
2945 				    mad_priv_dma_size(recv),
2946 				    DMA_FROM_DEVICE);
2947 		kfree(recv);
2948 	}
2949 
2950 	qp_info->recv_queue.count = 0;
2951 }
2952 
2953 /*
2954  * Start the port
2955  */
2956 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2957 {
2958 	int ret, i;
2959 	struct ib_qp_attr *attr;
2960 	struct ib_qp *qp;
2961 	u16 pkey_index;
2962 
2963 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2964 	if (!attr) {
2965 		dev_err(&port_priv->device->dev,
2966 			"Couldn't kmalloc ib_qp_attr\n");
2967 		return -ENOMEM;
2968 	}
2969 
2970 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2971 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
2972 	if (ret)
2973 		pkey_index = 0;
2974 
2975 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2976 		qp = port_priv->qp_info[i].qp;
2977 		if (!qp)
2978 			continue;
2979 
2980 		/*
2981 		 * PKey index for QP1 is irrelevant but
2982 		 * one is needed for the Reset to Init transition
2983 		 */
2984 		attr->qp_state = IB_QPS_INIT;
2985 		attr->pkey_index = pkey_index;
2986 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2987 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2988 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2989 		if (ret) {
2990 			dev_err(&port_priv->device->dev,
2991 				"Couldn't change QP%d state to INIT: %d\n",
2992 				i, ret);
2993 			goto out;
2994 		}
2995 
2996 		attr->qp_state = IB_QPS_RTR;
2997 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2998 		if (ret) {
2999 			dev_err(&port_priv->device->dev,
3000 				"Couldn't change QP%d state to RTR: %d\n",
3001 				i, ret);
3002 			goto out;
3003 		}
3004 
3005 		attr->qp_state = IB_QPS_RTS;
3006 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
3007 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3008 		if (ret) {
3009 			dev_err(&port_priv->device->dev,
3010 				"Couldn't change QP%d state to RTS: %d\n",
3011 				i, ret);
3012 			goto out;
3013 		}
3014 	}
3015 
3016 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3017 	if (ret) {
3018 		dev_err(&port_priv->device->dev,
3019 			"Failed to request completion notification: %d\n",
3020 			ret);
3021 		goto out;
3022 	}
3023 
3024 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3025 		if (!port_priv->qp_info[i].qp)
3026 			continue;
3027 
3028 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3029 		if (ret) {
3030 			dev_err(&port_priv->device->dev,
3031 				"Couldn't post receive WRs\n");
3032 			goto out;
3033 		}
3034 	}
3035 out:
3036 	kfree(attr);
3037 	return ret;
3038 }
3039 
3040 static void qp_event_handler(struct ib_event *event, void *qp_context)
3041 {
3042 	struct ib_mad_qp_info	*qp_info = qp_context;
3043 
3044 	/* It's worse than that! He's dead, Jim! */
3045 	dev_err(&qp_info->port_priv->device->dev,
3046 		"Fatal error (%d) on MAD QP (%d)\n",
3047 		event->event, qp_info->qp->qp_num);
3048 }
3049 
3050 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3051 			   struct ib_mad_queue *mad_queue)
3052 {
3053 	mad_queue->qp_info = qp_info;
3054 	mad_queue->count = 0;
3055 	spin_lock_init(&mad_queue->lock);
3056 	INIT_LIST_HEAD(&mad_queue->list);
3057 }
3058 
3059 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3060 			struct ib_mad_qp_info *qp_info)
3061 {
3062 	qp_info->port_priv = port_priv;
3063 	init_mad_queue(qp_info, &qp_info->send_queue);
3064 	init_mad_queue(qp_info, &qp_info->recv_queue);
3065 	INIT_LIST_HEAD(&qp_info->overflow_list);
3066 	spin_lock_init(&qp_info->snoop_lock);
3067 	qp_info->snoop_table = NULL;
3068 	qp_info->snoop_table_size = 0;
3069 	atomic_set(&qp_info->snoop_count, 0);
3070 }
3071 
3072 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3073 			 enum ib_qp_type qp_type)
3074 {
3075 	struct ib_qp_init_attr	qp_init_attr;
3076 	int ret;
3077 
3078 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
3079 	qp_init_attr.send_cq = qp_info->port_priv->cq;
3080 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
3081 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3082 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
3083 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3084 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3085 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3086 	qp_init_attr.qp_type = qp_type;
3087 	qp_init_attr.port_num = qp_info->port_priv->port_num;
3088 	qp_init_attr.qp_context = qp_info;
3089 	qp_init_attr.event_handler = qp_event_handler;
3090 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3091 	if (IS_ERR(qp_info->qp)) {
3092 		dev_err(&qp_info->port_priv->device->dev,
3093 			"Couldn't create ib_mad QP%d\n",
3094 			get_spl_qp_index(qp_type));
3095 		ret = PTR_ERR(qp_info->qp);
3096 		goto error;
3097 	}
3098 	/* Use minimum queue sizes unless the CQ is resized */
3099 	qp_info->send_queue.max_active = mad_sendq_size;
3100 	qp_info->recv_queue.max_active = mad_recvq_size;
3101 	return 0;
3102 
3103 error:
3104 	return ret;
3105 }
3106 
3107 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3108 {
3109 	if (!qp_info->qp)
3110 		return;
3111 
3112 	ib_destroy_qp(qp_info->qp);
3113 	kfree(qp_info->snoop_table);
3114 }
3115 
3116 /*
3117  * Open the port
3118  * Create the QP, PD, MR, and CQ if needed
3119  */
3120 static int ib_mad_port_open(struct ib_device *device,
3121 			    int port_num)
3122 {
3123 	int ret, cq_size;
3124 	struct ib_mad_port_private *port_priv;
3125 	unsigned long flags;
3126 	char name[sizeof "ib_mad123"];
3127 	int has_smi;
3128 
3129 	if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3130 		return -EFAULT;
3131 
3132 	if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3133 		    rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3134 		return -EFAULT;
3135 
3136 	/* Create new device info */
3137 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3138 	if (!port_priv) {
3139 		dev_err(&device->dev, "No memory for ib_mad_port_private\n");
3140 		return -ENOMEM;
3141 	}
3142 
3143 	port_priv->device = device;
3144 	port_priv->port_num = port_num;
3145 	spin_lock_init(&port_priv->reg_lock);
3146 	INIT_LIST_HEAD(&port_priv->agent_list);
3147 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
3148 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
3149 
3150 	cq_size = mad_sendq_size + mad_recvq_size;
3151 	has_smi = rdma_cap_ib_smi(device, port_num);
3152 	if (has_smi)
3153 		cq_size *= 2;
3154 
3155 	port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3156 			IB_POLL_WORKQUEUE);
3157 	if (IS_ERR(port_priv->cq)) {
3158 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3159 		ret = PTR_ERR(port_priv->cq);
3160 		goto error3;
3161 	}
3162 
3163 	port_priv->pd = ib_alloc_pd(device);
3164 	if (IS_ERR(port_priv->pd)) {
3165 		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3166 		ret = PTR_ERR(port_priv->pd);
3167 		goto error4;
3168 	}
3169 
3170 	if (has_smi) {
3171 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3172 		if (ret)
3173 			goto error6;
3174 	}
3175 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3176 	if (ret)
3177 		goto error7;
3178 
3179 	snprintf(name, sizeof name, "ib_mad%d", port_num);
3180 	port_priv->wq = create_singlethread_workqueue(name);
3181 	if (!port_priv->wq) {
3182 		ret = -ENOMEM;
3183 		goto error8;
3184 	}
3185 
3186 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3187 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3188 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3189 
3190 	ret = ib_mad_port_start(port_priv);
3191 	if (ret) {
3192 		dev_err(&device->dev, "Couldn't start port\n");
3193 		goto error9;
3194 	}
3195 
3196 	return 0;
3197 
3198 error9:
3199 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3200 	list_del_init(&port_priv->port_list);
3201 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3202 
3203 	destroy_workqueue(port_priv->wq);
3204 error8:
3205 	destroy_mad_qp(&port_priv->qp_info[1]);
3206 error7:
3207 	destroy_mad_qp(&port_priv->qp_info[0]);
3208 error6:
3209 	ib_dealloc_pd(port_priv->pd);
3210 error4:
3211 	ib_free_cq(port_priv->cq);
3212 	cleanup_recv_queue(&port_priv->qp_info[1]);
3213 	cleanup_recv_queue(&port_priv->qp_info[0]);
3214 error3:
3215 	kfree(port_priv);
3216 
3217 	return ret;
3218 }
3219 
3220 /*
3221  * Close the port
3222  * If there are no classes using the port, free the port
3223  * resources (CQ, MR, PD, QP) and remove the port's info structure
3224  */
3225 static int ib_mad_port_close(struct ib_device *device, int port_num)
3226 {
3227 	struct ib_mad_port_private *port_priv;
3228 	unsigned long flags;
3229 
3230 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3231 	port_priv = __ib_get_mad_port(device, port_num);
3232 	if (port_priv == NULL) {
3233 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3234 		dev_err(&device->dev, "Port %d not found\n", port_num);
3235 		return -ENODEV;
3236 	}
3237 	list_del_init(&port_priv->port_list);
3238 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3239 
3240 	destroy_workqueue(port_priv->wq);
3241 	destroy_mad_qp(&port_priv->qp_info[1]);
3242 	destroy_mad_qp(&port_priv->qp_info[0]);
3243 	ib_dealloc_pd(port_priv->pd);
3244 	ib_free_cq(port_priv->cq);
3245 	cleanup_recv_queue(&port_priv->qp_info[1]);
3246 	cleanup_recv_queue(&port_priv->qp_info[0]);
3247 	/* XXX: Handle deallocation of MAD registration tables */
3248 
3249 	kfree(port_priv);
3250 
3251 	return 0;
3252 }
3253 
3254 static void ib_mad_init_device(struct ib_device *device)
3255 {
3256 	int start, i;
3257 
3258 	start = rdma_start_port(device);
3259 
3260 	for (i = start; i <= rdma_end_port(device); i++) {
3261 		if (!rdma_cap_ib_mad(device, i))
3262 			continue;
3263 
3264 		if (ib_mad_port_open(device, i)) {
3265 			dev_err(&device->dev, "Couldn't open port %d\n", i);
3266 			goto error;
3267 		}
3268 		if (ib_agent_port_open(device, i)) {
3269 			dev_err(&device->dev,
3270 				"Couldn't open port %d for agents\n", i);
3271 			goto error_agent;
3272 		}
3273 	}
3274 	return;
3275 
3276 error_agent:
3277 	if (ib_mad_port_close(device, i))
3278 		dev_err(&device->dev, "Couldn't close port %d\n", i);
3279 
3280 error:
3281 	while (--i >= start) {
3282 		if (!rdma_cap_ib_mad(device, i))
3283 			continue;
3284 
3285 		if (ib_agent_port_close(device, i))
3286 			dev_err(&device->dev,
3287 				"Couldn't close port %d for agents\n", i);
3288 		if (ib_mad_port_close(device, i))
3289 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3290 	}
3291 }
3292 
3293 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3294 {
3295 	int i;
3296 
3297 	for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3298 		if (!rdma_cap_ib_mad(device, i))
3299 			continue;
3300 
3301 		if (ib_agent_port_close(device, i))
3302 			dev_err(&device->dev,
3303 				"Couldn't close port %d for agents\n", i);
3304 		if (ib_mad_port_close(device, i))
3305 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3306 	}
3307 }
3308 
3309 static struct ib_client mad_client = {
3310 	.name   = "mad",
3311 	.add = ib_mad_init_device,
3312 	.remove = ib_mad_remove_device
3313 };
3314 
3315 int ib_mad_init(void)
3316 {
3317 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3318 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3319 
3320 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3321 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3322 
3323 	INIT_LIST_HEAD(&ib_mad_port_list);
3324 
3325 	if (ib_register_client(&mad_client)) {
3326 		pr_err("Couldn't register ib_mad client\n");
3327 		return -EINVAL;
3328 	}
3329 
3330 	return 0;
3331 }
3332 
3333 void ib_mad_cleanup(void)
3334 {
3335 	ib_unregister_client(&mad_client);
3336 }
3337