xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision 020c5260)
1 /*
2  * Copyright (c) 2004-2007 Voltaire, Inc. All rights reserved.
3  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
5  * Copyright (c) 2009 HNR Consulting. All rights reserved.
6  * Copyright (c) 2014 Intel Corporation.  All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  */
37 
38 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
39 
40 #include <linux/dma-mapping.h>
41 #include <linux/slab.h>
42 #include <linux/module.h>
43 #include <rdma/ib_cache.h>
44 
45 #include "mad_priv.h"
46 #include "mad_rmpp.h"
47 #include "smi.h"
48 #include "opa_smi.h"
49 #include "agent.h"
50 #include "core_priv.h"
51 
52 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
53 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
54 
55 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
56 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
57 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
58 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
59 
60 static struct list_head ib_mad_port_list;
61 static u32 ib_mad_client_id = 0;
62 
63 /* Port list lock */
64 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
65 
66 /* Forward declarations */
67 static int method_in_use(struct ib_mad_mgmt_method_table **method,
68 			 struct ib_mad_reg_req *mad_reg_req);
69 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
70 static struct ib_mad_agent_private *find_mad_agent(
71 					struct ib_mad_port_private *port_priv,
72 					const struct ib_mad_hdr *mad);
73 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
74 				    struct ib_mad_private *mad);
75 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
76 static void timeout_sends(struct work_struct *work);
77 static void local_completions(struct work_struct *work);
78 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
79 			      struct ib_mad_agent_private *agent_priv,
80 			      u8 mgmt_class);
81 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
82 			   struct ib_mad_agent_private *agent_priv);
83 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
84 			      struct ib_wc *wc);
85 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
86 
87 /*
88  * Returns a ib_mad_port_private structure or NULL for a device/port
89  * Assumes ib_mad_port_list_lock is being held
90  */
91 static inline struct ib_mad_port_private *
92 __ib_get_mad_port(struct ib_device *device, int port_num)
93 {
94 	struct ib_mad_port_private *entry;
95 
96 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
97 		if (entry->device == device && entry->port_num == port_num)
98 			return entry;
99 	}
100 	return NULL;
101 }
102 
103 /*
104  * Wrapper function to return a ib_mad_port_private structure or NULL
105  * for a device/port
106  */
107 static inline struct ib_mad_port_private *
108 ib_get_mad_port(struct ib_device *device, int port_num)
109 {
110 	struct ib_mad_port_private *entry;
111 	unsigned long flags;
112 
113 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
114 	entry = __ib_get_mad_port(device, port_num);
115 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
116 
117 	return entry;
118 }
119 
120 static inline u8 convert_mgmt_class(u8 mgmt_class)
121 {
122 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
123 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
124 		0 : mgmt_class;
125 }
126 
127 static int get_spl_qp_index(enum ib_qp_type qp_type)
128 {
129 	switch (qp_type)
130 	{
131 	case IB_QPT_SMI:
132 		return 0;
133 	case IB_QPT_GSI:
134 		return 1;
135 	default:
136 		return -1;
137 	}
138 }
139 
140 static int vendor_class_index(u8 mgmt_class)
141 {
142 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
143 }
144 
145 static int is_vendor_class(u8 mgmt_class)
146 {
147 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
148 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
149 		return 0;
150 	return 1;
151 }
152 
153 static int is_vendor_oui(char *oui)
154 {
155 	if (oui[0] || oui[1] || oui[2])
156 		return 1;
157 	return 0;
158 }
159 
160 static int is_vendor_method_in_use(
161 		struct ib_mad_mgmt_vendor_class *vendor_class,
162 		struct ib_mad_reg_req *mad_reg_req)
163 {
164 	struct ib_mad_mgmt_method_table *method;
165 	int i;
166 
167 	for (i = 0; i < MAX_MGMT_OUI; i++) {
168 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
169 			method = vendor_class->method_table[i];
170 			if (method) {
171 				if (method_in_use(&method, mad_reg_req))
172 					return 1;
173 				else
174 					break;
175 			}
176 		}
177 	}
178 	return 0;
179 }
180 
181 int ib_response_mad(const struct ib_mad_hdr *hdr)
182 {
183 	return ((hdr->method & IB_MGMT_METHOD_RESP) ||
184 		(hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
185 		((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
186 		 (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
187 }
188 EXPORT_SYMBOL(ib_response_mad);
189 
190 /*
191  * ib_register_mad_agent - Register to send/receive MADs
192  */
193 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
194 					   u8 port_num,
195 					   enum ib_qp_type qp_type,
196 					   struct ib_mad_reg_req *mad_reg_req,
197 					   u8 rmpp_version,
198 					   ib_mad_send_handler send_handler,
199 					   ib_mad_recv_handler recv_handler,
200 					   void *context,
201 					   u32 registration_flags)
202 {
203 	struct ib_mad_port_private *port_priv;
204 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
205 	struct ib_mad_agent_private *mad_agent_priv;
206 	struct ib_mad_reg_req *reg_req = NULL;
207 	struct ib_mad_mgmt_class_table *class;
208 	struct ib_mad_mgmt_vendor_class_table *vendor;
209 	struct ib_mad_mgmt_vendor_class *vendor_class;
210 	struct ib_mad_mgmt_method_table *method;
211 	int ret2, qpn;
212 	unsigned long flags;
213 	u8 mgmt_class, vclass;
214 
215 	/* Validate parameters */
216 	qpn = get_spl_qp_index(qp_type);
217 	if (qpn == -1) {
218 		dev_notice(&device->dev,
219 			   "ib_register_mad_agent: invalid QP Type %d\n",
220 			   qp_type);
221 		goto error1;
222 	}
223 
224 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
225 		dev_notice(&device->dev,
226 			   "ib_register_mad_agent: invalid RMPP Version %u\n",
227 			   rmpp_version);
228 		goto error1;
229 	}
230 
231 	/* Validate MAD registration request if supplied */
232 	if (mad_reg_req) {
233 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
234 			dev_notice(&device->dev,
235 				   "ib_register_mad_agent: invalid Class Version %u\n",
236 				   mad_reg_req->mgmt_class_version);
237 			goto error1;
238 		}
239 		if (!recv_handler) {
240 			dev_notice(&device->dev,
241 				   "ib_register_mad_agent: no recv_handler\n");
242 			goto error1;
243 		}
244 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
245 			/*
246 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
247 			 * one in this range currently allowed
248 			 */
249 			if (mad_reg_req->mgmt_class !=
250 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
251 				dev_notice(&device->dev,
252 					   "ib_register_mad_agent: Invalid Mgmt Class 0x%x\n",
253 					   mad_reg_req->mgmt_class);
254 				goto error1;
255 			}
256 		} else if (mad_reg_req->mgmt_class == 0) {
257 			/*
258 			 * Class 0 is reserved in IBA and is used for
259 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
260 			 */
261 			dev_notice(&device->dev,
262 				   "ib_register_mad_agent: Invalid Mgmt Class 0\n");
263 			goto error1;
264 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
265 			/*
266 			 * If class is in "new" vendor range,
267 			 * ensure supplied OUI is not zero
268 			 */
269 			if (!is_vendor_oui(mad_reg_req->oui)) {
270 				dev_notice(&device->dev,
271 					   "ib_register_mad_agent: No OUI specified for class 0x%x\n",
272 					   mad_reg_req->mgmt_class);
273 				goto error1;
274 			}
275 		}
276 		/* Make sure class supplied is consistent with RMPP */
277 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
278 			if (rmpp_version) {
279 				dev_notice(&device->dev,
280 					   "ib_register_mad_agent: RMPP version for non-RMPP class 0x%x\n",
281 					   mad_reg_req->mgmt_class);
282 				goto error1;
283 			}
284 		}
285 
286 		/* Make sure class supplied is consistent with QP type */
287 		if (qp_type == IB_QPT_SMI) {
288 			if ((mad_reg_req->mgmt_class !=
289 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
290 			    (mad_reg_req->mgmt_class !=
291 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
292 				dev_notice(&device->dev,
293 					   "ib_register_mad_agent: Invalid SM QP type: class 0x%x\n",
294 					   mad_reg_req->mgmt_class);
295 				goto error1;
296 			}
297 		} else {
298 			if ((mad_reg_req->mgmt_class ==
299 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
300 			    (mad_reg_req->mgmt_class ==
301 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
302 				dev_notice(&device->dev,
303 					   "ib_register_mad_agent: Invalid GS QP type: class 0x%x\n",
304 					   mad_reg_req->mgmt_class);
305 				goto error1;
306 			}
307 		}
308 	} else {
309 		/* No registration request supplied */
310 		if (!send_handler)
311 			goto error1;
312 		if (registration_flags & IB_MAD_USER_RMPP)
313 			goto error1;
314 	}
315 
316 	/* Validate device and port */
317 	port_priv = ib_get_mad_port(device, port_num);
318 	if (!port_priv) {
319 		dev_notice(&device->dev,
320 			   "ib_register_mad_agent: Invalid port %d\n",
321 			   port_num);
322 		ret = ERR_PTR(-ENODEV);
323 		goto error1;
324 	}
325 
326 	/* Verify the QP requested is supported.  For example, Ethernet devices
327 	 * will not have QP0 */
328 	if (!port_priv->qp_info[qpn].qp) {
329 		dev_notice(&device->dev,
330 			   "ib_register_mad_agent: QP %d not supported\n", qpn);
331 		ret = ERR_PTR(-EPROTONOSUPPORT);
332 		goto error1;
333 	}
334 
335 	/* Allocate structures */
336 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
337 	if (!mad_agent_priv) {
338 		ret = ERR_PTR(-ENOMEM);
339 		goto error1;
340 	}
341 
342 	if (mad_reg_req) {
343 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
344 		if (!reg_req) {
345 			ret = ERR_PTR(-ENOMEM);
346 			goto error3;
347 		}
348 	}
349 
350 	/* Now, fill in the various structures */
351 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
352 	mad_agent_priv->reg_req = reg_req;
353 	mad_agent_priv->agent.rmpp_version = rmpp_version;
354 	mad_agent_priv->agent.device = device;
355 	mad_agent_priv->agent.recv_handler = recv_handler;
356 	mad_agent_priv->agent.send_handler = send_handler;
357 	mad_agent_priv->agent.context = context;
358 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
359 	mad_agent_priv->agent.port_num = port_num;
360 	mad_agent_priv->agent.flags = registration_flags;
361 	spin_lock_init(&mad_agent_priv->lock);
362 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
363 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
364 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
365 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
366 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
367 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
368 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
369 	atomic_set(&mad_agent_priv->refcount, 1);
370 	init_completion(&mad_agent_priv->comp);
371 
372 	spin_lock_irqsave(&port_priv->reg_lock, flags);
373 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
374 
375 	/*
376 	 * Make sure MAD registration (if supplied)
377 	 * is non overlapping with any existing ones
378 	 */
379 	if (mad_reg_req) {
380 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
381 		if (!is_vendor_class(mgmt_class)) {
382 			class = port_priv->version[mad_reg_req->
383 						   mgmt_class_version].class;
384 			if (class) {
385 				method = class->method_table[mgmt_class];
386 				if (method) {
387 					if (method_in_use(&method,
388 							   mad_reg_req))
389 						goto error4;
390 				}
391 			}
392 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
393 						  mgmt_class);
394 		} else {
395 			/* "New" vendor class range */
396 			vendor = port_priv->version[mad_reg_req->
397 						    mgmt_class_version].vendor;
398 			if (vendor) {
399 				vclass = vendor_class_index(mgmt_class);
400 				vendor_class = vendor->vendor_class[vclass];
401 				if (vendor_class) {
402 					if (is_vendor_method_in_use(
403 							vendor_class,
404 							mad_reg_req))
405 						goto error4;
406 				}
407 			}
408 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
409 		}
410 		if (ret2) {
411 			ret = ERR_PTR(ret2);
412 			goto error4;
413 		}
414 	}
415 
416 	/* Add mad agent into port's agent list */
417 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
418 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
419 
420 	return &mad_agent_priv->agent;
421 
422 error4:
423 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
424 	kfree(reg_req);
425 error3:
426 	kfree(mad_agent_priv);
427 error1:
428 	return ret;
429 }
430 EXPORT_SYMBOL(ib_register_mad_agent);
431 
432 static inline int is_snooping_sends(int mad_snoop_flags)
433 {
434 	return (mad_snoop_flags &
435 		(/*IB_MAD_SNOOP_POSTED_SENDS |
436 		 IB_MAD_SNOOP_RMPP_SENDS |*/
437 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
438 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
439 }
440 
441 static inline int is_snooping_recvs(int mad_snoop_flags)
442 {
443 	return (mad_snoop_flags &
444 		(IB_MAD_SNOOP_RECVS /*|
445 		 IB_MAD_SNOOP_RMPP_RECVS*/));
446 }
447 
448 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
449 				struct ib_mad_snoop_private *mad_snoop_priv)
450 {
451 	struct ib_mad_snoop_private **new_snoop_table;
452 	unsigned long flags;
453 	int i;
454 
455 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
456 	/* Check for empty slot in array. */
457 	for (i = 0; i < qp_info->snoop_table_size; i++)
458 		if (!qp_info->snoop_table[i])
459 			break;
460 
461 	if (i == qp_info->snoop_table_size) {
462 		/* Grow table. */
463 		new_snoop_table = krealloc(qp_info->snoop_table,
464 					   sizeof mad_snoop_priv *
465 					   (qp_info->snoop_table_size + 1),
466 					   GFP_ATOMIC);
467 		if (!new_snoop_table) {
468 			i = -ENOMEM;
469 			goto out;
470 		}
471 
472 		qp_info->snoop_table = new_snoop_table;
473 		qp_info->snoop_table_size++;
474 	}
475 	qp_info->snoop_table[i] = mad_snoop_priv;
476 	atomic_inc(&qp_info->snoop_count);
477 out:
478 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
479 	return i;
480 }
481 
482 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
483 					   u8 port_num,
484 					   enum ib_qp_type qp_type,
485 					   int mad_snoop_flags,
486 					   ib_mad_snoop_handler snoop_handler,
487 					   ib_mad_recv_handler recv_handler,
488 					   void *context)
489 {
490 	struct ib_mad_port_private *port_priv;
491 	struct ib_mad_agent *ret;
492 	struct ib_mad_snoop_private *mad_snoop_priv;
493 	int qpn;
494 
495 	/* Validate parameters */
496 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
497 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
498 		ret = ERR_PTR(-EINVAL);
499 		goto error1;
500 	}
501 	qpn = get_spl_qp_index(qp_type);
502 	if (qpn == -1) {
503 		ret = ERR_PTR(-EINVAL);
504 		goto error1;
505 	}
506 	port_priv = ib_get_mad_port(device, port_num);
507 	if (!port_priv) {
508 		ret = ERR_PTR(-ENODEV);
509 		goto error1;
510 	}
511 	/* Allocate structures */
512 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
513 	if (!mad_snoop_priv) {
514 		ret = ERR_PTR(-ENOMEM);
515 		goto error1;
516 	}
517 
518 	/* Now, fill in the various structures */
519 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
520 	mad_snoop_priv->agent.device = device;
521 	mad_snoop_priv->agent.recv_handler = recv_handler;
522 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
523 	mad_snoop_priv->agent.context = context;
524 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
525 	mad_snoop_priv->agent.port_num = port_num;
526 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
527 	init_completion(&mad_snoop_priv->comp);
528 	mad_snoop_priv->snoop_index = register_snoop_agent(
529 						&port_priv->qp_info[qpn],
530 						mad_snoop_priv);
531 	if (mad_snoop_priv->snoop_index < 0) {
532 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
533 		goto error2;
534 	}
535 
536 	atomic_set(&mad_snoop_priv->refcount, 1);
537 	return &mad_snoop_priv->agent;
538 
539 error2:
540 	kfree(mad_snoop_priv);
541 error1:
542 	return ret;
543 }
544 EXPORT_SYMBOL(ib_register_mad_snoop);
545 
546 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
547 {
548 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
549 		complete(&mad_agent_priv->comp);
550 }
551 
552 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
553 {
554 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
555 		complete(&mad_snoop_priv->comp);
556 }
557 
558 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
559 {
560 	struct ib_mad_port_private *port_priv;
561 	unsigned long flags;
562 
563 	/* Note that we could still be handling received MADs */
564 
565 	/*
566 	 * Canceling all sends results in dropping received response
567 	 * MADs, preventing us from queuing additional work
568 	 */
569 	cancel_mads(mad_agent_priv);
570 	port_priv = mad_agent_priv->qp_info->port_priv;
571 	cancel_delayed_work(&mad_agent_priv->timed_work);
572 
573 	spin_lock_irqsave(&port_priv->reg_lock, flags);
574 	remove_mad_reg_req(mad_agent_priv);
575 	list_del(&mad_agent_priv->agent_list);
576 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
577 
578 	flush_workqueue(port_priv->wq);
579 	ib_cancel_rmpp_recvs(mad_agent_priv);
580 
581 	deref_mad_agent(mad_agent_priv);
582 	wait_for_completion(&mad_agent_priv->comp);
583 
584 	kfree(mad_agent_priv->reg_req);
585 	kfree(mad_agent_priv);
586 }
587 
588 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
589 {
590 	struct ib_mad_qp_info *qp_info;
591 	unsigned long flags;
592 
593 	qp_info = mad_snoop_priv->qp_info;
594 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
595 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
596 	atomic_dec(&qp_info->snoop_count);
597 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
598 
599 	deref_snoop_agent(mad_snoop_priv);
600 	wait_for_completion(&mad_snoop_priv->comp);
601 
602 	kfree(mad_snoop_priv);
603 }
604 
605 /*
606  * ib_unregister_mad_agent - Unregisters a client from using MAD services
607  */
608 void ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
609 {
610 	struct ib_mad_agent_private *mad_agent_priv;
611 	struct ib_mad_snoop_private *mad_snoop_priv;
612 
613 	/* If the TID is zero, the agent can only snoop. */
614 	if (mad_agent->hi_tid) {
615 		mad_agent_priv = container_of(mad_agent,
616 					      struct ib_mad_agent_private,
617 					      agent);
618 		unregister_mad_agent(mad_agent_priv);
619 	} else {
620 		mad_snoop_priv = container_of(mad_agent,
621 					      struct ib_mad_snoop_private,
622 					      agent);
623 		unregister_mad_snoop(mad_snoop_priv);
624 	}
625 }
626 EXPORT_SYMBOL(ib_unregister_mad_agent);
627 
628 static void dequeue_mad(struct ib_mad_list_head *mad_list)
629 {
630 	struct ib_mad_queue *mad_queue;
631 	unsigned long flags;
632 
633 	BUG_ON(!mad_list->mad_queue);
634 	mad_queue = mad_list->mad_queue;
635 	spin_lock_irqsave(&mad_queue->lock, flags);
636 	list_del(&mad_list->list);
637 	mad_queue->count--;
638 	spin_unlock_irqrestore(&mad_queue->lock, flags);
639 }
640 
641 static void snoop_send(struct ib_mad_qp_info *qp_info,
642 		       struct ib_mad_send_buf *send_buf,
643 		       struct ib_mad_send_wc *mad_send_wc,
644 		       int mad_snoop_flags)
645 {
646 	struct ib_mad_snoop_private *mad_snoop_priv;
647 	unsigned long flags;
648 	int i;
649 
650 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
651 	for (i = 0; i < qp_info->snoop_table_size; i++) {
652 		mad_snoop_priv = qp_info->snoop_table[i];
653 		if (!mad_snoop_priv ||
654 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
655 			continue;
656 
657 		atomic_inc(&mad_snoop_priv->refcount);
658 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
659 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
660 						    send_buf, mad_send_wc);
661 		deref_snoop_agent(mad_snoop_priv);
662 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
663 	}
664 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
665 }
666 
667 static void snoop_recv(struct ib_mad_qp_info *qp_info,
668 		       struct ib_mad_recv_wc *mad_recv_wc,
669 		       int mad_snoop_flags)
670 {
671 	struct ib_mad_snoop_private *mad_snoop_priv;
672 	unsigned long flags;
673 	int i;
674 
675 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
676 	for (i = 0; i < qp_info->snoop_table_size; i++) {
677 		mad_snoop_priv = qp_info->snoop_table[i];
678 		if (!mad_snoop_priv ||
679 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
680 			continue;
681 
682 		atomic_inc(&mad_snoop_priv->refcount);
683 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
684 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
685 						   mad_recv_wc);
686 		deref_snoop_agent(mad_snoop_priv);
687 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
688 	}
689 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
690 }
691 
692 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
693 		u16 pkey_index, u8 port_num, struct ib_wc *wc)
694 {
695 	memset(wc, 0, sizeof *wc);
696 	wc->wr_cqe = cqe;
697 	wc->status = IB_WC_SUCCESS;
698 	wc->opcode = IB_WC_RECV;
699 	wc->pkey_index = pkey_index;
700 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
701 	wc->src_qp = IB_QP0;
702 	wc->qp = qp;
703 	wc->slid = slid;
704 	wc->sl = 0;
705 	wc->dlid_path_bits = 0;
706 	wc->port_num = port_num;
707 }
708 
709 static size_t mad_priv_size(const struct ib_mad_private *mp)
710 {
711 	return sizeof(struct ib_mad_private) + mp->mad_size;
712 }
713 
714 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
715 {
716 	size_t size = sizeof(struct ib_mad_private) + mad_size;
717 	struct ib_mad_private *ret = kzalloc(size, flags);
718 
719 	if (ret)
720 		ret->mad_size = mad_size;
721 
722 	return ret;
723 }
724 
725 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
726 {
727 	return rdma_max_mad_size(port_priv->device, port_priv->port_num);
728 }
729 
730 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
731 {
732 	return sizeof(struct ib_grh) + mp->mad_size;
733 }
734 
735 /*
736  * Return 0 if SMP is to be sent
737  * Return 1 if SMP was consumed locally (whether or not solicited)
738  * Return < 0 if error
739  */
740 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
741 				  struct ib_mad_send_wr_private *mad_send_wr)
742 {
743 	int ret = 0;
744 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
745 	struct opa_smp *opa_smp = (struct opa_smp *)smp;
746 	unsigned long flags;
747 	struct ib_mad_local_private *local;
748 	struct ib_mad_private *mad_priv;
749 	struct ib_mad_port_private *port_priv;
750 	struct ib_mad_agent_private *recv_mad_agent = NULL;
751 	struct ib_device *device = mad_agent_priv->agent.device;
752 	u8 port_num;
753 	struct ib_wc mad_wc;
754 	struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
755 	size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
756 	u16 out_mad_pkey_index = 0;
757 	u16 drslid;
758 	bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
759 				    mad_agent_priv->qp_info->port_priv->port_num);
760 
761 	if (rdma_cap_ib_switch(device) &&
762 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
763 		port_num = send_wr->port_num;
764 	else
765 		port_num = mad_agent_priv->agent.port_num;
766 
767 	/*
768 	 * Directed route handling starts if the initial LID routed part of
769 	 * a request or the ending LID routed part of a response is empty.
770 	 * If we are at the start of the LID routed part, don't update the
771 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
772 	 */
773 	if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
774 		u32 opa_drslid;
775 
776 		if ((opa_get_smp_direction(opa_smp)
777 		     ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
778 		     OPA_LID_PERMISSIVE &&
779 		     opa_smi_handle_dr_smp_send(opa_smp,
780 						rdma_cap_ib_switch(device),
781 						port_num) == IB_SMI_DISCARD) {
782 			ret = -EINVAL;
783 			dev_err(&device->dev, "OPA Invalid directed route\n");
784 			goto out;
785 		}
786 		opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
787 		if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
788 		    opa_drslid & 0xffff0000) {
789 			ret = -EINVAL;
790 			dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
791 			       opa_drslid);
792 			goto out;
793 		}
794 		drslid = (u16)(opa_drslid & 0x0000ffff);
795 
796 		/* Check to post send on QP or process locally */
797 		if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
798 		    opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
799 			goto out;
800 	} else {
801 		if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
802 		     IB_LID_PERMISSIVE &&
803 		     smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
804 		     IB_SMI_DISCARD) {
805 			ret = -EINVAL;
806 			dev_err(&device->dev, "Invalid directed route\n");
807 			goto out;
808 		}
809 		drslid = be16_to_cpu(smp->dr_slid);
810 
811 		/* Check to post send on QP or process locally */
812 		if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
813 		    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
814 			goto out;
815 	}
816 
817 	local = kmalloc(sizeof *local, GFP_ATOMIC);
818 	if (!local) {
819 		ret = -ENOMEM;
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 		kfree(local);
828 		goto out;
829 	}
830 
831 	build_smp_wc(mad_agent_priv->agent.qp,
832 		     send_wr->wr.wr_cqe, drslid,
833 		     send_wr->pkey_index,
834 		     send_wr->port_num, &mad_wc);
835 
836 	if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
837 		mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
838 					+ mad_send_wr->send_buf.data_len
839 					+ sizeof(struct ib_grh);
840 	}
841 
842 	/* No GRH for DR SMP */
843 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
844 				  (const struct ib_mad_hdr *)smp, mad_size,
845 				  (struct ib_mad_hdr *)mad_priv->mad,
846 				  &mad_size, &out_mad_pkey_index);
847 	switch (ret)
848 	{
849 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
850 		if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
851 		    mad_agent_priv->agent.recv_handler) {
852 			local->mad_priv = mad_priv;
853 			local->recv_mad_agent = mad_agent_priv;
854 			/*
855 			 * Reference MAD agent until receive
856 			 * side of local completion handled
857 			 */
858 			atomic_inc(&mad_agent_priv->refcount);
859 		} else
860 			kfree(mad_priv);
861 		break;
862 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
863 		kfree(mad_priv);
864 		break;
865 	case IB_MAD_RESULT_SUCCESS:
866 		/* Treat like an incoming receive MAD */
867 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
868 					    mad_agent_priv->agent.port_num);
869 		if (port_priv) {
870 			memcpy(mad_priv->mad, smp, mad_priv->mad_size);
871 			recv_mad_agent = find_mad_agent(port_priv,
872 						        (const struct ib_mad_hdr *)mad_priv->mad);
873 		}
874 		if (!port_priv || !recv_mad_agent) {
875 			/*
876 			 * No receiving agent so drop packet and
877 			 * generate send completion.
878 			 */
879 			kfree(mad_priv);
880 			break;
881 		}
882 		local->mad_priv = mad_priv;
883 		local->recv_mad_agent = recv_mad_agent;
884 		break;
885 	default:
886 		kfree(mad_priv);
887 		kfree(local);
888 		ret = -EINVAL;
889 		goto out;
890 	}
891 
892 	local->mad_send_wr = mad_send_wr;
893 	if (opa) {
894 		local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
895 		local->return_wc_byte_len = mad_size;
896 	}
897 	/* Reference MAD agent until send side of local completion handled */
898 	atomic_inc(&mad_agent_priv->refcount);
899 	/* Queue local completion to local list */
900 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
901 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
902 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
903 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
904 		   &mad_agent_priv->local_work);
905 	ret = 1;
906 out:
907 	return ret;
908 }
909 
910 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
911 {
912 	int seg_size, pad;
913 
914 	seg_size = mad_size - hdr_len;
915 	if (data_len && seg_size) {
916 		pad = seg_size - data_len % seg_size;
917 		return pad == seg_size ? 0 : pad;
918 	} else
919 		return seg_size;
920 }
921 
922 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
923 {
924 	struct ib_rmpp_segment *s, *t;
925 
926 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
927 		list_del(&s->list);
928 		kfree(s);
929 	}
930 }
931 
932 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
933 				size_t mad_size, gfp_t gfp_mask)
934 {
935 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
936 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
937 	struct ib_rmpp_segment *seg = NULL;
938 	int left, seg_size, pad;
939 
940 	send_buf->seg_size = mad_size - send_buf->hdr_len;
941 	send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
942 	seg_size = send_buf->seg_size;
943 	pad = send_wr->pad;
944 
945 	/* Allocate data segments. */
946 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
947 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
948 		if (!seg) {
949 			free_send_rmpp_list(send_wr);
950 			return -ENOMEM;
951 		}
952 		seg->num = ++send_buf->seg_count;
953 		list_add_tail(&seg->list, &send_wr->rmpp_list);
954 	}
955 
956 	/* Zero any padding */
957 	if (pad)
958 		memset(seg->data + seg_size - pad, 0, pad);
959 
960 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
961 					  agent.rmpp_version;
962 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
963 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
964 
965 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
966 					struct ib_rmpp_segment, list);
967 	send_wr->last_ack_seg = send_wr->cur_seg;
968 	return 0;
969 }
970 
971 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
972 {
973 	return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
974 }
975 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
976 
977 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
978 					    u32 remote_qpn, u16 pkey_index,
979 					    int rmpp_active,
980 					    int hdr_len, int data_len,
981 					    gfp_t gfp_mask,
982 					    u8 base_version)
983 {
984 	struct ib_mad_agent_private *mad_agent_priv;
985 	struct ib_mad_send_wr_private *mad_send_wr;
986 	int pad, message_size, ret, size;
987 	void *buf;
988 	size_t mad_size;
989 	bool opa;
990 
991 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
992 				      agent);
993 
994 	opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
995 
996 	if (opa && base_version == OPA_MGMT_BASE_VERSION)
997 		mad_size = sizeof(struct opa_mad);
998 	else
999 		mad_size = sizeof(struct ib_mad);
1000 
1001 	pad = get_pad_size(hdr_len, data_len, mad_size);
1002 	message_size = hdr_len + data_len + pad;
1003 
1004 	if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1005 		if (!rmpp_active && message_size > mad_size)
1006 			return ERR_PTR(-EINVAL);
1007 	} else
1008 		if (rmpp_active || message_size > mad_size)
1009 			return ERR_PTR(-EINVAL);
1010 
1011 	size = rmpp_active ? hdr_len : mad_size;
1012 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1013 	if (!buf)
1014 		return ERR_PTR(-ENOMEM);
1015 
1016 	mad_send_wr = buf + size;
1017 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1018 	mad_send_wr->send_buf.mad = buf;
1019 	mad_send_wr->send_buf.hdr_len = hdr_len;
1020 	mad_send_wr->send_buf.data_len = data_len;
1021 	mad_send_wr->pad = pad;
1022 
1023 	mad_send_wr->mad_agent_priv = mad_agent_priv;
1024 	mad_send_wr->sg_list[0].length = hdr_len;
1025 	mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1026 
1027 	/* OPA MADs don't have to be the full 2048 bytes */
1028 	if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1029 	    data_len < mad_size - hdr_len)
1030 		mad_send_wr->sg_list[1].length = data_len;
1031 	else
1032 		mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1033 
1034 	mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1035 
1036 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1037 
1038 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1039 	mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1040 	mad_send_wr->send_wr.wr.num_sge = 2;
1041 	mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1042 	mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1043 	mad_send_wr->send_wr.remote_qpn = remote_qpn;
1044 	mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1045 	mad_send_wr->send_wr.pkey_index = pkey_index;
1046 
1047 	if (rmpp_active) {
1048 		ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1049 		if (ret) {
1050 			kfree(buf);
1051 			return ERR_PTR(ret);
1052 		}
1053 	}
1054 
1055 	mad_send_wr->send_buf.mad_agent = mad_agent;
1056 	atomic_inc(&mad_agent_priv->refcount);
1057 	return &mad_send_wr->send_buf;
1058 }
1059 EXPORT_SYMBOL(ib_create_send_mad);
1060 
1061 int ib_get_mad_data_offset(u8 mgmt_class)
1062 {
1063 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1064 		return IB_MGMT_SA_HDR;
1065 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1066 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1067 		 (mgmt_class == IB_MGMT_CLASS_BIS))
1068 		return IB_MGMT_DEVICE_HDR;
1069 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1070 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1071 		return IB_MGMT_VENDOR_HDR;
1072 	else
1073 		return IB_MGMT_MAD_HDR;
1074 }
1075 EXPORT_SYMBOL(ib_get_mad_data_offset);
1076 
1077 int ib_is_mad_class_rmpp(u8 mgmt_class)
1078 {
1079 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1080 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1081 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1082 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
1083 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1084 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1085 		return 1;
1086 	return 0;
1087 }
1088 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1089 
1090 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1091 {
1092 	struct ib_mad_send_wr_private *mad_send_wr;
1093 	struct list_head *list;
1094 
1095 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1096 				   send_buf);
1097 	list = &mad_send_wr->cur_seg->list;
1098 
1099 	if (mad_send_wr->cur_seg->num < seg_num) {
1100 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
1101 			if (mad_send_wr->cur_seg->num == seg_num)
1102 				break;
1103 	} else if (mad_send_wr->cur_seg->num > seg_num) {
1104 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1105 			if (mad_send_wr->cur_seg->num == seg_num)
1106 				break;
1107 	}
1108 	return mad_send_wr->cur_seg->data;
1109 }
1110 EXPORT_SYMBOL(ib_get_rmpp_segment);
1111 
1112 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1113 {
1114 	if (mad_send_wr->send_buf.seg_count)
1115 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1116 					   mad_send_wr->seg_num);
1117 	else
1118 		return mad_send_wr->send_buf.mad +
1119 		       mad_send_wr->send_buf.hdr_len;
1120 }
1121 
1122 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1123 {
1124 	struct ib_mad_agent_private *mad_agent_priv;
1125 	struct ib_mad_send_wr_private *mad_send_wr;
1126 
1127 	mad_agent_priv = container_of(send_buf->mad_agent,
1128 				      struct ib_mad_agent_private, agent);
1129 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1130 				   send_buf);
1131 
1132 	free_send_rmpp_list(mad_send_wr);
1133 	kfree(send_buf->mad);
1134 	deref_mad_agent(mad_agent_priv);
1135 }
1136 EXPORT_SYMBOL(ib_free_send_mad);
1137 
1138 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1139 {
1140 	struct ib_mad_qp_info *qp_info;
1141 	struct list_head *list;
1142 	struct ib_send_wr *bad_send_wr;
1143 	struct ib_mad_agent *mad_agent;
1144 	struct ib_sge *sge;
1145 	unsigned long flags;
1146 	int ret;
1147 
1148 	/* Set WR ID to find mad_send_wr upon completion */
1149 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1150 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1151 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1152 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1153 
1154 	mad_agent = mad_send_wr->send_buf.mad_agent;
1155 	sge = mad_send_wr->sg_list;
1156 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1157 					mad_send_wr->send_buf.mad,
1158 					sge[0].length,
1159 					DMA_TO_DEVICE);
1160 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1161 		return -ENOMEM;
1162 
1163 	mad_send_wr->header_mapping = sge[0].addr;
1164 
1165 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1166 					ib_get_payload(mad_send_wr),
1167 					sge[1].length,
1168 					DMA_TO_DEVICE);
1169 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1170 		ib_dma_unmap_single(mad_agent->device,
1171 				    mad_send_wr->header_mapping,
1172 				    sge[0].length, DMA_TO_DEVICE);
1173 		return -ENOMEM;
1174 	}
1175 	mad_send_wr->payload_mapping = sge[1].addr;
1176 
1177 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1178 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1179 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1180 				   &bad_send_wr);
1181 		list = &qp_info->send_queue.list;
1182 	} else {
1183 		ret = 0;
1184 		list = &qp_info->overflow_list;
1185 	}
1186 
1187 	if (!ret) {
1188 		qp_info->send_queue.count++;
1189 		list_add_tail(&mad_send_wr->mad_list.list, list);
1190 	}
1191 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1192 	if (ret) {
1193 		ib_dma_unmap_single(mad_agent->device,
1194 				    mad_send_wr->header_mapping,
1195 				    sge[0].length, DMA_TO_DEVICE);
1196 		ib_dma_unmap_single(mad_agent->device,
1197 				    mad_send_wr->payload_mapping,
1198 				    sge[1].length, DMA_TO_DEVICE);
1199 	}
1200 	return ret;
1201 }
1202 
1203 /*
1204  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1205  *  with the registered client
1206  */
1207 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1208 		     struct ib_mad_send_buf **bad_send_buf)
1209 {
1210 	struct ib_mad_agent_private *mad_agent_priv;
1211 	struct ib_mad_send_buf *next_send_buf;
1212 	struct ib_mad_send_wr_private *mad_send_wr;
1213 	unsigned long flags;
1214 	int ret = -EINVAL;
1215 
1216 	/* Walk list of send WRs and post each on send list */
1217 	for (; send_buf; send_buf = next_send_buf) {
1218 
1219 		mad_send_wr = container_of(send_buf,
1220 					   struct ib_mad_send_wr_private,
1221 					   send_buf);
1222 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1223 
1224 		if (!send_buf->mad_agent->send_handler ||
1225 		    (send_buf->timeout_ms &&
1226 		     !send_buf->mad_agent->recv_handler)) {
1227 			ret = -EINVAL;
1228 			goto error;
1229 		}
1230 
1231 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1232 			if (mad_agent_priv->agent.rmpp_version) {
1233 				ret = -EINVAL;
1234 				goto error;
1235 			}
1236 		}
1237 
1238 		/*
1239 		 * Save pointer to next work request to post in case the
1240 		 * current one completes, and the user modifies the work
1241 		 * request associated with the completion
1242 		 */
1243 		next_send_buf = send_buf->next;
1244 		mad_send_wr->send_wr.ah = send_buf->ah;
1245 
1246 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1247 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1248 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1249 						     mad_send_wr);
1250 			if (ret < 0)		/* error */
1251 				goto error;
1252 			else if (ret == 1)	/* locally consumed */
1253 				continue;
1254 		}
1255 
1256 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1257 		/* Timeout will be updated after send completes */
1258 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1259 		mad_send_wr->max_retries = send_buf->retries;
1260 		mad_send_wr->retries_left = send_buf->retries;
1261 		send_buf->retries = 0;
1262 		/* Reference for work request to QP + response */
1263 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1264 		mad_send_wr->status = IB_WC_SUCCESS;
1265 
1266 		/* Reference MAD agent until send completes */
1267 		atomic_inc(&mad_agent_priv->refcount);
1268 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1269 		list_add_tail(&mad_send_wr->agent_list,
1270 			      &mad_agent_priv->send_list);
1271 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1272 
1273 		if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1274 			ret = ib_send_rmpp_mad(mad_send_wr);
1275 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1276 				ret = ib_send_mad(mad_send_wr);
1277 		} else
1278 			ret = ib_send_mad(mad_send_wr);
1279 		if (ret < 0) {
1280 			/* Fail send request */
1281 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1282 			list_del(&mad_send_wr->agent_list);
1283 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1284 			atomic_dec(&mad_agent_priv->refcount);
1285 			goto error;
1286 		}
1287 	}
1288 	return 0;
1289 error:
1290 	if (bad_send_buf)
1291 		*bad_send_buf = send_buf;
1292 	return ret;
1293 }
1294 EXPORT_SYMBOL(ib_post_send_mad);
1295 
1296 /*
1297  * ib_free_recv_mad - Returns data buffers used to receive
1298  *  a MAD to the access layer
1299  */
1300 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1301 {
1302 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1303 	struct ib_mad_private_header *mad_priv_hdr;
1304 	struct ib_mad_private *priv;
1305 	struct list_head free_list;
1306 
1307 	INIT_LIST_HEAD(&free_list);
1308 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1309 
1310 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1311 					&free_list, list) {
1312 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1313 					   recv_buf);
1314 		mad_priv_hdr = container_of(mad_recv_wc,
1315 					    struct ib_mad_private_header,
1316 					    recv_wc);
1317 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1318 				    header);
1319 		kfree(priv);
1320 	}
1321 }
1322 EXPORT_SYMBOL(ib_free_recv_mad);
1323 
1324 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1325 					u8 rmpp_version,
1326 					ib_mad_send_handler send_handler,
1327 					ib_mad_recv_handler recv_handler,
1328 					void *context)
1329 {
1330 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1331 }
1332 EXPORT_SYMBOL(ib_redirect_mad_qp);
1333 
1334 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1335 		      struct ib_wc *wc)
1336 {
1337 	dev_err(&mad_agent->device->dev,
1338 		"ib_process_mad_wc() not implemented yet\n");
1339 	return 0;
1340 }
1341 EXPORT_SYMBOL(ib_process_mad_wc);
1342 
1343 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1344 			 struct ib_mad_reg_req *mad_reg_req)
1345 {
1346 	int i;
1347 
1348 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1349 		if ((*method)->agent[i]) {
1350 			pr_err("Method %d already in use\n", i);
1351 			return -EINVAL;
1352 		}
1353 	}
1354 	return 0;
1355 }
1356 
1357 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1358 {
1359 	/* Allocate management method table */
1360 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1361 	return (*method) ? 0 : (-ENOMEM);
1362 }
1363 
1364 /*
1365  * Check to see if there are any methods still in use
1366  */
1367 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1368 {
1369 	int i;
1370 
1371 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1372 		if (method->agent[i])
1373 			return 1;
1374 	return 0;
1375 }
1376 
1377 /*
1378  * Check to see if there are any method tables for this class still in use
1379  */
1380 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1381 {
1382 	int i;
1383 
1384 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1385 		if (class->method_table[i])
1386 			return 1;
1387 	return 0;
1388 }
1389 
1390 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1391 {
1392 	int i;
1393 
1394 	for (i = 0; i < MAX_MGMT_OUI; i++)
1395 		if (vendor_class->method_table[i])
1396 			return 1;
1397 	return 0;
1398 }
1399 
1400 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1401 			   const char *oui)
1402 {
1403 	int i;
1404 
1405 	for (i = 0; i < MAX_MGMT_OUI; i++)
1406 		/* Is there matching OUI for this vendor class ? */
1407 		if (!memcmp(vendor_class->oui[i], oui, 3))
1408 			return i;
1409 
1410 	return -1;
1411 }
1412 
1413 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1414 {
1415 	int i;
1416 
1417 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1418 		if (vendor->vendor_class[i])
1419 			return 1;
1420 
1421 	return 0;
1422 }
1423 
1424 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1425 				     struct ib_mad_agent_private *agent)
1426 {
1427 	int i;
1428 
1429 	/* Remove any methods for this mad agent */
1430 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1431 		if (method->agent[i] == agent) {
1432 			method->agent[i] = NULL;
1433 		}
1434 	}
1435 }
1436 
1437 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1438 			      struct ib_mad_agent_private *agent_priv,
1439 			      u8 mgmt_class)
1440 {
1441 	struct ib_mad_port_private *port_priv;
1442 	struct ib_mad_mgmt_class_table **class;
1443 	struct ib_mad_mgmt_method_table **method;
1444 	int i, ret;
1445 
1446 	port_priv = agent_priv->qp_info->port_priv;
1447 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1448 	if (!*class) {
1449 		/* Allocate management class table for "new" class version */
1450 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1451 		if (!*class) {
1452 			ret = -ENOMEM;
1453 			goto error1;
1454 		}
1455 
1456 		/* Allocate method table for this management class */
1457 		method = &(*class)->method_table[mgmt_class];
1458 		if ((ret = allocate_method_table(method)))
1459 			goto error2;
1460 	} else {
1461 		method = &(*class)->method_table[mgmt_class];
1462 		if (!*method) {
1463 			/* Allocate method table for this management class */
1464 			if ((ret = allocate_method_table(method)))
1465 				goto error1;
1466 		}
1467 	}
1468 
1469 	/* Now, make sure methods are not already in use */
1470 	if (method_in_use(method, mad_reg_req))
1471 		goto error3;
1472 
1473 	/* Finally, add in methods being registered */
1474 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1475 		(*method)->agent[i] = agent_priv;
1476 
1477 	return 0;
1478 
1479 error3:
1480 	/* Remove any methods for this mad agent */
1481 	remove_methods_mad_agent(*method, agent_priv);
1482 	/* Now, check to see if there are any methods in use */
1483 	if (!check_method_table(*method)) {
1484 		/* If not, release management method table */
1485 		kfree(*method);
1486 		*method = NULL;
1487 	}
1488 	ret = -EINVAL;
1489 	goto error1;
1490 error2:
1491 	kfree(*class);
1492 	*class = NULL;
1493 error1:
1494 	return ret;
1495 }
1496 
1497 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1498 			   struct ib_mad_agent_private *agent_priv)
1499 {
1500 	struct ib_mad_port_private *port_priv;
1501 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1502 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1503 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1504 	struct ib_mad_mgmt_method_table **method;
1505 	int i, ret = -ENOMEM;
1506 	u8 vclass;
1507 
1508 	/* "New" vendor (with OUI) class */
1509 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1510 	port_priv = agent_priv->qp_info->port_priv;
1511 	vendor_table = &port_priv->version[
1512 				mad_reg_req->mgmt_class_version].vendor;
1513 	if (!*vendor_table) {
1514 		/* Allocate mgmt vendor class table for "new" class version */
1515 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1516 		if (!vendor)
1517 			goto error1;
1518 
1519 		*vendor_table = vendor;
1520 	}
1521 	if (!(*vendor_table)->vendor_class[vclass]) {
1522 		/* Allocate table for this management vendor class */
1523 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1524 		if (!vendor_class)
1525 			goto error2;
1526 
1527 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1528 	}
1529 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1530 		/* Is there matching OUI for this vendor class ? */
1531 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1532 			    mad_reg_req->oui, 3)) {
1533 			method = &(*vendor_table)->vendor_class[
1534 						vclass]->method_table[i];
1535 			BUG_ON(!*method);
1536 			goto check_in_use;
1537 		}
1538 	}
1539 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1540 		/* OUI slot available ? */
1541 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1542 				vclass]->oui[i])) {
1543 			method = &(*vendor_table)->vendor_class[
1544 				vclass]->method_table[i];
1545 			BUG_ON(*method);
1546 			/* Allocate method table for this OUI */
1547 			if ((ret = allocate_method_table(method)))
1548 				goto error3;
1549 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1550 			       mad_reg_req->oui, 3);
1551 			goto check_in_use;
1552 		}
1553 	}
1554 	dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1555 	goto error3;
1556 
1557 check_in_use:
1558 	/* Now, make sure methods are not already in use */
1559 	if (method_in_use(method, mad_reg_req))
1560 		goto error4;
1561 
1562 	/* Finally, add in methods being registered */
1563 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1564 		(*method)->agent[i] = agent_priv;
1565 
1566 	return 0;
1567 
1568 error4:
1569 	/* Remove any methods for this mad agent */
1570 	remove_methods_mad_agent(*method, agent_priv);
1571 	/* Now, check to see if there are any methods in use */
1572 	if (!check_method_table(*method)) {
1573 		/* If not, release management method table */
1574 		kfree(*method);
1575 		*method = NULL;
1576 	}
1577 	ret = -EINVAL;
1578 error3:
1579 	if (vendor_class) {
1580 		(*vendor_table)->vendor_class[vclass] = NULL;
1581 		kfree(vendor_class);
1582 	}
1583 error2:
1584 	if (vendor) {
1585 		*vendor_table = NULL;
1586 		kfree(vendor);
1587 	}
1588 error1:
1589 	return ret;
1590 }
1591 
1592 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1593 {
1594 	struct ib_mad_port_private *port_priv;
1595 	struct ib_mad_mgmt_class_table *class;
1596 	struct ib_mad_mgmt_method_table *method;
1597 	struct ib_mad_mgmt_vendor_class_table *vendor;
1598 	struct ib_mad_mgmt_vendor_class *vendor_class;
1599 	int index;
1600 	u8 mgmt_class;
1601 
1602 	/*
1603 	 * Was MAD registration request supplied
1604 	 * with original registration ?
1605 	 */
1606 	if (!agent_priv->reg_req) {
1607 		goto out;
1608 	}
1609 
1610 	port_priv = agent_priv->qp_info->port_priv;
1611 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1612 	class = port_priv->version[
1613 			agent_priv->reg_req->mgmt_class_version].class;
1614 	if (!class)
1615 		goto vendor_check;
1616 
1617 	method = class->method_table[mgmt_class];
1618 	if (method) {
1619 		/* Remove any methods for this mad agent */
1620 		remove_methods_mad_agent(method, agent_priv);
1621 		/* Now, check to see if there are any methods still in use */
1622 		if (!check_method_table(method)) {
1623 			/* If not, release management method table */
1624 			kfree(method);
1625 			class->method_table[mgmt_class] = NULL;
1626 			/* Any management classes left ? */
1627 			if (!check_class_table(class)) {
1628 				/* If not, release management class table */
1629 				kfree(class);
1630 				port_priv->version[
1631 					agent_priv->reg_req->
1632 					mgmt_class_version].class = NULL;
1633 			}
1634 		}
1635 	}
1636 
1637 vendor_check:
1638 	if (!is_vendor_class(mgmt_class))
1639 		goto out;
1640 
1641 	/* normalize mgmt_class to vendor range 2 */
1642 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1643 	vendor = port_priv->version[
1644 			agent_priv->reg_req->mgmt_class_version].vendor;
1645 
1646 	if (!vendor)
1647 		goto out;
1648 
1649 	vendor_class = vendor->vendor_class[mgmt_class];
1650 	if (vendor_class) {
1651 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1652 		if (index < 0)
1653 			goto out;
1654 		method = vendor_class->method_table[index];
1655 		if (method) {
1656 			/* Remove any methods for this mad agent */
1657 			remove_methods_mad_agent(method, agent_priv);
1658 			/*
1659 			 * Now, check to see if there are
1660 			 * any methods still in use
1661 			 */
1662 			if (!check_method_table(method)) {
1663 				/* If not, release management method table */
1664 				kfree(method);
1665 				vendor_class->method_table[index] = NULL;
1666 				memset(vendor_class->oui[index], 0, 3);
1667 				/* Any OUIs left ? */
1668 				if (!check_vendor_class(vendor_class)) {
1669 					/* If not, release vendor class table */
1670 					kfree(vendor_class);
1671 					vendor->vendor_class[mgmt_class] = NULL;
1672 					/* Any other vendor classes left ? */
1673 					if (!check_vendor_table(vendor)) {
1674 						kfree(vendor);
1675 						port_priv->version[
1676 							agent_priv->reg_req->
1677 							mgmt_class_version].
1678 							vendor = NULL;
1679 					}
1680 				}
1681 			}
1682 		}
1683 	}
1684 
1685 out:
1686 	return;
1687 }
1688 
1689 static struct ib_mad_agent_private *
1690 find_mad_agent(struct ib_mad_port_private *port_priv,
1691 	       const struct ib_mad_hdr *mad_hdr)
1692 {
1693 	struct ib_mad_agent_private *mad_agent = NULL;
1694 	unsigned long flags;
1695 
1696 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1697 	if (ib_response_mad(mad_hdr)) {
1698 		u32 hi_tid;
1699 		struct ib_mad_agent_private *entry;
1700 
1701 		/*
1702 		 * Routing is based on high 32 bits of transaction ID
1703 		 * of MAD.
1704 		 */
1705 		hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1706 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1707 			if (entry->agent.hi_tid == hi_tid) {
1708 				mad_agent = entry;
1709 				break;
1710 			}
1711 		}
1712 	} else {
1713 		struct ib_mad_mgmt_class_table *class;
1714 		struct ib_mad_mgmt_method_table *method;
1715 		struct ib_mad_mgmt_vendor_class_table *vendor;
1716 		struct ib_mad_mgmt_vendor_class *vendor_class;
1717 		const struct ib_vendor_mad *vendor_mad;
1718 		int index;
1719 
1720 		/*
1721 		 * Routing is based on version, class, and method
1722 		 * For "newer" vendor MADs, also based on OUI
1723 		 */
1724 		if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1725 			goto out;
1726 		if (!is_vendor_class(mad_hdr->mgmt_class)) {
1727 			class = port_priv->version[
1728 					mad_hdr->class_version].class;
1729 			if (!class)
1730 				goto out;
1731 			if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1732 			    ARRAY_SIZE(class->method_table))
1733 				goto out;
1734 			method = class->method_table[convert_mgmt_class(
1735 							mad_hdr->mgmt_class)];
1736 			if (method)
1737 				mad_agent = method->agent[mad_hdr->method &
1738 							  ~IB_MGMT_METHOD_RESP];
1739 		} else {
1740 			vendor = port_priv->version[
1741 					mad_hdr->class_version].vendor;
1742 			if (!vendor)
1743 				goto out;
1744 			vendor_class = vendor->vendor_class[vendor_class_index(
1745 						mad_hdr->mgmt_class)];
1746 			if (!vendor_class)
1747 				goto out;
1748 			/* Find matching OUI */
1749 			vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1750 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1751 			if (index == -1)
1752 				goto out;
1753 			method = vendor_class->method_table[index];
1754 			if (method) {
1755 				mad_agent = method->agent[mad_hdr->method &
1756 							  ~IB_MGMT_METHOD_RESP];
1757 			}
1758 		}
1759 	}
1760 
1761 	if (mad_agent) {
1762 		if (mad_agent->agent.recv_handler)
1763 			atomic_inc(&mad_agent->refcount);
1764 		else {
1765 			dev_notice(&port_priv->device->dev,
1766 				   "No receive handler for client %p on port %d\n",
1767 				   &mad_agent->agent, port_priv->port_num);
1768 			mad_agent = NULL;
1769 		}
1770 	}
1771 out:
1772 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1773 
1774 	return mad_agent;
1775 }
1776 
1777 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1778 			const struct ib_mad_qp_info *qp_info,
1779 			bool opa)
1780 {
1781 	int valid = 0;
1782 	u32 qp_num = qp_info->qp->qp_num;
1783 
1784 	/* Make sure MAD base version is understood */
1785 	if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1786 	    (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1787 		pr_err("MAD received with unsupported base version %d %s\n",
1788 		       mad_hdr->base_version, opa ? "(opa)" : "");
1789 		goto out;
1790 	}
1791 
1792 	/* Filter SMI packets sent to other than QP0 */
1793 	if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1794 	    (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1795 		if (qp_num == 0)
1796 			valid = 1;
1797 	} else {
1798 		/* CM attributes other than ClassPortInfo only use Send method */
1799 		if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1800 		    (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1801 		    (mad_hdr->method != IB_MGMT_METHOD_SEND))
1802 			goto out;
1803 		/* Filter GSI packets sent to QP0 */
1804 		if (qp_num != 0)
1805 			valid = 1;
1806 	}
1807 
1808 out:
1809 	return valid;
1810 }
1811 
1812 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1813 			    const struct ib_mad_hdr *mad_hdr)
1814 {
1815 	struct ib_rmpp_mad *rmpp_mad;
1816 
1817 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1818 	return !mad_agent_priv->agent.rmpp_version ||
1819 		!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1820 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1821 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1822 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1823 }
1824 
1825 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1826 				     const struct ib_mad_recv_wc *rwc)
1827 {
1828 	return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1829 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1830 }
1831 
1832 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1833 				   const struct ib_mad_send_wr_private *wr,
1834 				   const struct ib_mad_recv_wc *rwc )
1835 {
1836 	struct rdma_ah_attr attr;
1837 	u8 send_resp, rcv_resp;
1838 	union ib_gid sgid;
1839 	struct ib_device *device = mad_agent_priv->agent.device;
1840 	u8 port_num = mad_agent_priv->agent.port_num;
1841 	u8 lmc;
1842 	bool has_grh;
1843 
1844 	send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1845 	rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1846 
1847 	if (send_resp == rcv_resp)
1848 		/* both requests, or both responses. GIDs different */
1849 		return 0;
1850 
1851 	if (rdma_query_ah(wr->send_buf.ah, &attr))
1852 		/* Assume not equal, to avoid false positives. */
1853 		return 0;
1854 
1855 	has_grh = !!(rdma_ah_get_ah_flags(&attr) & IB_AH_GRH);
1856 	if (has_grh != !!(rwc->wc->wc_flags & IB_WC_GRH))
1857 		/* one has GID, other does not.  Assume different */
1858 		return 0;
1859 
1860 	if (!send_resp && rcv_resp) {
1861 		/* is request/response. */
1862 		if (!has_grh) {
1863 			if (ib_get_cached_lmc(device, port_num, &lmc))
1864 				return 0;
1865 			return (!lmc || !((rdma_ah_get_path_bits(&attr) ^
1866 					   rwc->wc->dlid_path_bits) &
1867 					  ((1 << lmc) - 1)));
1868 		} else {
1869 			const struct ib_global_route *grh =
1870 					rdma_ah_read_grh(&attr);
1871 
1872 			if (ib_get_cached_gid(device, port_num,
1873 					      grh->sgid_index, &sgid, NULL))
1874 				return 0;
1875 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1876 				       16);
1877 		}
1878 	}
1879 
1880 	if (!has_grh)
1881 		return rdma_ah_get_dlid(&attr) == rwc->wc->slid;
1882 	else
1883 		return !memcmp(rdma_ah_read_grh(&attr)->dgid.raw,
1884 			       rwc->recv_buf.grh->sgid.raw,
1885 			       16);
1886 }
1887 
1888 static inline int is_direct(u8 class)
1889 {
1890 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1891 }
1892 
1893 struct ib_mad_send_wr_private*
1894 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1895 		 const struct ib_mad_recv_wc *wc)
1896 {
1897 	struct ib_mad_send_wr_private *wr;
1898 	const struct ib_mad_hdr *mad_hdr;
1899 
1900 	mad_hdr = &wc->recv_buf.mad->mad_hdr;
1901 
1902 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1903 		if ((wr->tid == mad_hdr->tid) &&
1904 		    rcv_has_same_class(wr, wc) &&
1905 		    /*
1906 		     * Don't check GID for direct routed MADs.
1907 		     * These might have permissive LIDs.
1908 		     */
1909 		    (is_direct(mad_hdr->mgmt_class) ||
1910 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1911 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1912 	}
1913 
1914 	/*
1915 	 * It's possible to receive the response before we've
1916 	 * been notified that the send has completed
1917 	 */
1918 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1919 		if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1920 		    wr->tid == mad_hdr->tid &&
1921 		    wr->timeout &&
1922 		    rcv_has_same_class(wr, wc) &&
1923 		    /*
1924 		     * Don't check GID for direct routed MADs.
1925 		     * These might have permissive LIDs.
1926 		     */
1927 		    (is_direct(mad_hdr->mgmt_class) ||
1928 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1929 			/* Verify request has not been canceled */
1930 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1931 	}
1932 	return NULL;
1933 }
1934 
1935 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1936 {
1937 	mad_send_wr->timeout = 0;
1938 	if (mad_send_wr->refcount == 1)
1939 		list_move_tail(&mad_send_wr->agent_list,
1940 			      &mad_send_wr->mad_agent_priv->done_list);
1941 }
1942 
1943 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1944 				 struct ib_mad_recv_wc *mad_recv_wc)
1945 {
1946 	struct ib_mad_send_wr_private *mad_send_wr;
1947 	struct ib_mad_send_wc mad_send_wc;
1948 	unsigned long flags;
1949 
1950 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1951 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1952 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1953 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1954 						      mad_recv_wc);
1955 		if (!mad_recv_wc) {
1956 			deref_mad_agent(mad_agent_priv);
1957 			return;
1958 		}
1959 	}
1960 
1961 	/* Complete corresponding request */
1962 	if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1963 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1964 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1965 		if (!mad_send_wr) {
1966 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1967 			if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
1968 			   && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
1969 			   && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
1970 					& IB_MGMT_RMPP_FLAG_ACTIVE)) {
1971 				/* user rmpp is in effect
1972 				 * and this is an active RMPP MAD
1973 				 */
1974 				mad_agent_priv->agent.recv_handler(
1975 						&mad_agent_priv->agent, NULL,
1976 						mad_recv_wc);
1977 				atomic_dec(&mad_agent_priv->refcount);
1978 			} else {
1979 				/* not user rmpp, revert to normal behavior and
1980 				 * drop the mad */
1981 				ib_free_recv_mad(mad_recv_wc);
1982 				deref_mad_agent(mad_agent_priv);
1983 				return;
1984 			}
1985 		} else {
1986 			ib_mark_mad_done(mad_send_wr);
1987 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1988 
1989 			/* Defined behavior is to complete response before request */
1990 			mad_agent_priv->agent.recv_handler(
1991 					&mad_agent_priv->agent,
1992 					&mad_send_wr->send_buf,
1993 					mad_recv_wc);
1994 			atomic_dec(&mad_agent_priv->refcount);
1995 
1996 			mad_send_wc.status = IB_WC_SUCCESS;
1997 			mad_send_wc.vendor_err = 0;
1998 			mad_send_wc.send_buf = &mad_send_wr->send_buf;
1999 			ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2000 		}
2001 	} else {
2002 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2003 						   mad_recv_wc);
2004 		deref_mad_agent(mad_agent_priv);
2005 	}
2006 }
2007 
2008 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2009 				     const struct ib_mad_qp_info *qp_info,
2010 				     const struct ib_wc *wc,
2011 				     int port_num,
2012 				     struct ib_mad_private *recv,
2013 				     struct ib_mad_private *response)
2014 {
2015 	enum smi_forward_action retsmi;
2016 	struct ib_smp *smp = (struct ib_smp *)recv->mad;
2017 
2018 	if (smi_handle_dr_smp_recv(smp,
2019 				   rdma_cap_ib_switch(port_priv->device),
2020 				   port_num,
2021 				   port_priv->device->phys_port_cnt) ==
2022 				   IB_SMI_DISCARD)
2023 		return IB_SMI_DISCARD;
2024 
2025 	retsmi = smi_check_forward_dr_smp(smp);
2026 	if (retsmi == IB_SMI_LOCAL)
2027 		return IB_SMI_HANDLE;
2028 
2029 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2030 		if (smi_handle_dr_smp_send(smp,
2031 					   rdma_cap_ib_switch(port_priv->device),
2032 					   port_num) == IB_SMI_DISCARD)
2033 			return IB_SMI_DISCARD;
2034 
2035 		if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2036 			return IB_SMI_DISCARD;
2037 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2038 		/* forward case for switches */
2039 		memcpy(response, recv, mad_priv_size(response));
2040 		response->header.recv_wc.wc = &response->header.wc;
2041 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2042 		response->header.recv_wc.recv_buf.grh = &response->grh;
2043 
2044 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2045 				    &response->grh, wc,
2046 				    port_priv->device,
2047 				    smi_get_fwd_port(smp),
2048 				    qp_info->qp->qp_num,
2049 				    response->mad_size,
2050 				    false);
2051 
2052 		return IB_SMI_DISCARD;
2053 	}
2054 	return IB_SMI_HANDLE;
2055 }
2056 
2057 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2058 				    struct ib_mad_private *response,
2059 				    size_t *resp_len, bool opa)
2060 {
2061 	const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2062 	struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2063 
2064 	if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2065 	    recv_hdr->method == IB_MGMT_METHOD_SET) {
2066 		memcpy(response, recv, mad_priv_size(response));
2067 		response->header.recv_wc.wc = &response->header.wc;
2068 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2069 		response->header.recv_wc.recv_buf.grh = &response->grh;
2070 		resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2071 		resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2072 		if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2073 			resp_hdr->status |= IB_SMP_DIRECTION;
2074 
2075 		if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2076 			if (recv_hdr->mgmt_class ==
2077 			    IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2078 			    recv_hdr->mgmt_class ==
2079 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2080 				*resp_len = opa_get_smp_header_size(
2081 							(struct opa_smp *)recv->mad);
2082 			else
2083 				*resp_len = sizeof(struct ib_mad_hdr);
2084 		}
2085 
2086 		return true;
2087 	} else {
2088 		return false;
2089 	}
2090 }
2091 
2092 static enum smi_action
2093 handle_opa_smi(struct ib_mad_port_private *port_priv,
2094 	       struct ib_mad_qp_info *qp_info,
2095 	       struct ib_wc *wc,
2096 	       int port_num,
2097 	       struct ib_mad_private *recv,
2098 	       struct ib_mad_private *response)
2099 {
2100 	enum smi_forward_action retsmi;
2101 	struct opa_smp *smp = (struct opa_smp *)recv->mad;
2102 
2103 	if (opa_smi_handle_dr_smp_recv(smp,
2104 				   rdma_cap_ib_switch(port_priv->device),
2105 				   port_num,
2106 				   port_priv->device->phys_port_cnt) ==
2107 				   IB_SMI_DISCARD)
2108 		return IB_SMI_DISCARD;
2109 
2110 	retsmi = opa_smi_check_forward_dr_smp(smp);
2111 	if (retsmi == IB_SMI_LOCAL)
2112 		return IB_SMI_HANDLE;
2113 
2114 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2115 		if (opa_smi_handle_dr_smp_send(smp,
2116 					   rdma_cap_ib_switch(port_priv->device),
2117 					   port_num) == IB_SMI_DISCARD)
2118 			return IB_SMI_DISCARD;
2119 
2120 		if (opa_smi_check_local_smp(smp, port_priv->device) ==
2121 		    IB_SMI_DISCARD)
2122 			return IB_SMI_DISCARD;
2123 
2124 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2125 		/* forward case for switches */
2126 		memcpy(response, recv, mad_priv_size(response));
2127 		response->header.recv_wc.wc = &response->header.wc;
2128 		response->header.recv_wc.recv_buf.opa_mad =
2129 				(struct opa_mad *)response->mad;
2130 		response->header.recv_wc.recv_buf.grh = &response->grh;
2131 
2132 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2133 				    &response->grh, wc,
2134 				    port_priv->device,
2135 				    opa_smi_get_fwd_port(smp),
2136 				    qp_info->qp->qp_num,
2137 				    recv->header.wc.byte_len,
2138 				    true);
2139 
2140 		return IB_SMI_DISCARD;
2141 	}
2142 
2143 	return IB_SMI_HANDLE;
2144 }
2145 
2146 static enum smi_action
2147 handle_smi(struct ib_mad_port_private *port_priv,
2148 	   struct ib_mad_qp_info *qp_info,
2149 	   struct ib_wc *wc,
2150 	   int port_num,
2151 	   struct ib_mad_private *recv,
2152 	   struct ib_mad_private *response,
2153 	   bool opa)
2154 {
2155 	struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2156 
2157 	if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2158 	    mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2159 		return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2160 				      response);
2161 
2162 	return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2163 }
2164 
2165 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2166 {
2167 	struct ib_mad_port_private *port_priv = cq->cq_context;
2168 	struct ib_mad_list_head *mad_list =
2169 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2170 	struct ib_mad_qp_info *qp_info;
2171 	struct ib_mad_private_header *mad_priv_hdr;
2172 	struct ib_mad_private *recv, *response = NULL;
2173 	struct ib_mad_agent_private *mad_agent;
2174 	int port_num;
2175 	int ret = IB_MAD_RESULT_SUCCESS;
2176 	size_t mad_size;
2177 	u16 resp_mad_pkey_index = 0;
2178 	bool opa;
2179 
2180 	if (list_empty_careful(&port_priv->port_list))
2181 		return;
2182 
2183 	if (wc->status != IB_WC_SUCCESS) {
2184 		/*
2185 		 * Receive errors indicate that the QP has entered the error
2186 		 * state - error handling/shutdown code will cleanup
2187 		 */
2188 		return;
2189 	}
2190 
2191 	qp_info = mad_list->mad_queue->qp_info;
2192 	dequeue_mad(mad_list);
2193 
2194 	opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2195 			       qp_info->port_priv->port_num);
2196 
2197 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2198 				    mad_list);
2199 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2200 	ib_dma_unmap_single(port_priv->device,
2201 			    recv->header.mapping,
2202 			    mad_priv_dma_size(recv),
2203 			    DMA_FROM_DEVICE);
2204 
2205 	/* Setup MAD receive work completion from "normal" work completion */
2206 	recv->header.wc = *wc;
2207 	recv->header.recv_wc.wc = &recv->header.wc;
2208 
2209 	if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2210 		recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2211 		recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2212 	} else {
2213 		recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2214 		recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2215 	}
2216 
2217 	recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2218 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
2219 
2220 	if (atomic_read(&qp_info->snoop_count))
2221 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2222 
2223 	/* Validate MAD */
2224 	if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2225 		goto out;
2226 
2227 	mad_size = recv->mad_size;
2228 	response = alloc_mad_private(mad_size, GFP_KERNEL);
2229 	if (!response)
2230 		goto out;
2231 
2232 	if (rdma_cap_ib_switch(port_priv->device))
2233 		port_num = wc->port_num;
2234 	else
2235 		port_num = port_priv->port_num;
2236 
2237 	if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2238 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2239 		if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2240 			       response, opa)
2241 		    == IB_SMI_DISCARD)
2242 			goto out;
2243 	}
2244 
2245 	/* Give driver "right of first refusal" on incoming MAD */
2246 	if (port_priv->device->process_mad) {
2247 		ret = port_priv->device->process_mad(port_priv->device, 0,
2248 						     port_priv->port_num,
2249 						     wc, &recv->grh,
2250 						     (const struct ib_mad_hdr *)recv->mad,
2251 						     recv->mad_size,
2252 						     (struct ib_mad_hdr *)response->mad,
2253 						     &mad_size, &resp_mad_pkey_index);
2254 
2255 		if (opa)
2256 			wc->pkey_index = resp_mad_pkey_index;
2257 
2258 		if (ret & IB_MAD_RESULT_SUCCESS) {
2259 			if (ret & IB_MAD_RESULT_CONSUMED)
2260 				goto out;
2261 			if (ret & IB_MAD_RESULT_REPLY) {
2262 				agent_send_response((const struct ib_mad_hdr *)response->mad,
2263 						    &recv->grh, wc,
2264 						    port_priv->device,
2265 						    port_num,
2266 						    qp_info->qp->qp_num,
2267 						    mad_size, opa);
2268 				goto out;
2269 			}
2270 		}
2271 	}
2272 
2273 	mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2274 	if (mad_agent) {
2275 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2276 		/*
2277 		 * recv is freed up in error cases in ib_mad_complete_recv
2278 		 * or via recv_handler in ib_mad_complete_recv()
2279 		 */
2280 		recv = NULL;
2281 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2282 		   generate_unmatched_resp(recv, response, &mad_size, opa)) {
2283 		agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2284 				    port_priv->device, port_num,
2285 				    qp_info->qp->qp_num, mad_size, opa);
2286 	}
2287 
2288 out:
2289 	/* Post another receive request for this QP */
2290 	if (response) {
2291 		ib_mad_post_receive_mads(qp_info, response);
2292 		kfree(recv);
2293 	} else
2294 		ib_mad_post_receive_mads(qp_info, recv);
2295 }
2296 
2297 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2298 {
2299 	struct ib_mad_send_wr_private *mad_send_wr;
2300 	unsigned long delay;
2301 
2302 	if (list_empty(&mad_agent_priv->wait_list)) {
2303 		cancel_delayed_work(&mad_agent_priv->timed_work);
2304 	} else {
2305 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2306 					 struct ib_mad_send_wr_private,
2307 					 agent_list);
2308 
2309 		if (time_after(mad_agent_priv->timeout,
2310 			       mad_send_wr->timeout)) {
2311 			mad_agent_priv->timeout = mad_send_wr->timeout;
2312 			delay = mad_send_wr->timeout - jiffies;
2313 			if ((long)delay <= 0)
2314 				delay = 1;
2315 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2316 					 &mad_agent_priv->timed_work, delay);
2317 		}
2318 	}
2319 }
2320 
2321 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2322 {
2323 	struct ib_mad_agent_private *mad_agent_priv;
2324 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2325 	struct list_head *list_item;
2326 	unsigned long delay;
2327 
2328 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2329 	list_del(&mad_send_wr->agent_list);
2330 
2331 	delay = mad_send_wr->timeout;
2332 	mad_send_wr->timeout += jiffies;
2333 
2334 	if (delay) {
2335 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2336 			temp_mad_send_wr = list_entry(list_item,
2337 						struct ib_mad_send_wr_private,
2338 						agent_list);
2339 			if (time_after(mad_send_wr->timeout,
2340 				       temp_mad_send_wr->timeout))
2341 				break;
2342 		}
2343 	}
2344 	else
2345 		list_item = &mad_agent_priv->wait_list;
2346 	list_add(&mad_send_wr->agent_list, list_item);
2347 
2348 	/* Reschedule a work item if we have a shorter timeout */
2349 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2350 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2351 				 &mad_agent_priv->timed_work, delay);
2352 }
2353 
2354 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2355 			  int timeout_ms)
2356 {
2357 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2358 	wait_for_response(mad_send_wr);
2359 }
2360 
2361 /*
2362  * Process a send work completion
2363  */
2364 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2365 			     struct ib_mad_send_wc *mad_send_wc)
2366 {
2367 	struct ib_mad_agent_private	*mad_agent_priv;
2368 	unsigned long			flags;
2369 	int				ret;
2370 
2371 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2372 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2373 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2374 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2375 		if (ret == IB_RMPP_RESULT_CONSUMED)
2376 			goto done;
2377 	} else
2378 		ret = IB_RMPP_RESULT_UNHANDLED;
2379 
2380 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2381 	    mad_send_wr->status == IB_WC_SUCCESS) {
2382 		mad_send_wr->status = mad_send_wc->status;
2383 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2384 	}
2385 
2386 	if (--mad_send_wr->refcount > 0) {
2387 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2388 		    mad_send_wr->status == IB_WC_SUCCESS) {
2389 			wait_for_response(mad_send_wr);
2390 		}
2391 		goto done;
2392 	}
2393 
2394 	/* Remove send from MAD agent and notify client of completion */
2395 	list_del(&mad_send_wr->agent_list);
2396 	adjust_timeout(mad_agent_priv);
2397 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2398 
2399 	if (mad_send_wr->status != IB_WC_SUCCESS )
2400 		mad_send_wc->status = mad_send_wr->status;
2401 	if (ret == IB_RMPP_RESULT_INTERNAL)
2402 		ib_rmpp_send_handler(mad_send_wc);
2403 	else
2404 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2405 						   mad_send_wc);
2406 
2407 	/* Release reference on agent taken when sending */
2408 	deref_mad_agent(mad_agent_priv);
2409 	return;
2410 done:
2411 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2412 }
2413 
2414 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2415 {
2416 	struct ib_mad_port_private *port_priv = cq->cq_context;
2417 	struct ib_mad_list_head *mad_list =
2418 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2419 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2420 	struct ib_mad_qp_info		*qp_info;
2421 	struct ib_mad_queue		*send_queue;
2422 	struct ib_send_wr		*bad_send_wr;
2423 	struct ib_mad_send_wc		mad_send_wc;
2424 	unsigned long flags;
2425 	int ret;
2426 
2427 	if (list_empty_careful(&port_priv->port_list))
2428 		return;
2429 
2430 	if (wc->status != IB_WC_SUCCESS) {
2431 		if (!ib_mad_send_error(port_priv, wc))
2432 			return;
2433 	}
2434 
2435 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2436 				   mad_list);
2437 	send_queue = mad_list->mad_queue;
2438 	qp_info = send_queue->qp_info;
2439 
2440 retry:
2441 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2442 			    mad_send_wr->header_mapping,
2443 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2444 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2445 			    mad_send_wr->payload_mapping,
2446 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2447 	queued_send_wr = NULL;
2448 	spin_lock_irqsave(&send_queue->lock, flags);
2449 	list_del(&mad_list->list);
2450 
2451 	/* Move queued send to the send queue */
2452 	if (send_queue->count-- > send_queue->max_active) {
2453 		mad_list = container_of(qp_info->overflow_list.next,
2454 					struct ib_mad_list_head, list);
2455 		queued_send_wr = container_of(mad_list,
2456 					struct ib_mad_send_wr_private,
2457 					mad_list);
2458 		list_move_tail(&mad_list->list, &send_queue->list);
2459 	}
2460 	spin_unlock_irqrestore(&send_queue->lock, flags);
2461 
2462 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2463 	mad_send_wc.status = wc->status;
2464 	mad_send_wc.vendor_err = wc->vendor_err;
2465 	if (atomic_read(&qp_info->snoop_count))
2466 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2467 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2468 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2469 
2470 	if (queued_send_wr) {
2471 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2472 				   &bad_send_wr);
2473 		if (ret) {
2474 			dev_err(&port_priv->device->dev,
2475 				"ib_post_send failed: %d\n", ret);
2476 			mad_send_wr = queued_send_wr;
2477 			wc->status = IB_WC_LOC_QP_OP_ERR;
2478 			goto retry;
2479 		}
2480 	}
2481 }
2482 
2483 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2484 {
2485 	struct ib_mad_send_wr_private *mad_send_wr;
2486 	struct ib_mad_list_head *mad_list;
2487 	unsigned long flags;
2488 
2489 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2490 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2491 		mad_send_wr = container_of(mad_list,
2492 					   struct ib_mad_send_wr_private,
2493 					   mad_list);
2494 		mad_send_wr->retry = 1;
2495 	}
2496 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2497 }
2498 
2499 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2500 		struct ib_wc *wc)
2501 {
2502 	struct ib_mad_list_head *mad_list =
2503 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2504 	struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2505 	struct ib_mad_send_wr_private *mad_send_wr;
2506 	int ret;
2507 
2508 	/*
2509 	 * Send errors will transition the QP to SQE - move
2510 	 * QP to RTS and repost flushed work requests
2511 	 */
2512 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2513 				   mad_list);
2514 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2515 		if (mad_send_wr->retry) {
2516 			/* Repost send */
2517 			struct ib_send_wr *bad_send_wr;
2518 
2519 			mad_send_wr->retry = 0;
2520 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2521 					&bad_send_wr);
2522 			if (!ret)
2523 				return false;
2524 		}
2525 	} else {
2526 		struct ib_qp_attr *attr;
2527 
2528 		/* Transition QP to RTS and fail offending send */
2529 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2530 		if (attr) {
2531 			attr->qp_state = IB_QPS_RTS;
2532 			attr->cur_qp_state = IB_QPS_SQE;
2533 			ret = ib_modify_qp(qp_info->qp, attr,
2534 					   IB_QP_STATE | IB_QP_CUR_STATE);
2535 			kfree(attr);
2536 			if (ret)
2537 				dev_err(&port_priv->device->dev,
2538 					"%s - ib_modify_qp to RTS: %d\n",
2539 					__func__, ret);
2540 			else
2541 				mark_sends_for_retry(qp_info);
2542 		}
2543 	}
2544 
2545 	return true;
2546 }
2547 
2548 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2549 {
2550 	unsigned long flags;
2551 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2552 	struct ib_mad_send_wc mad_send_wc;
2553 	struct list_head cancel_list;
2554 
2555 	INIT_LIST_HEAD(&cancel_list);
2556 
2557 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2558 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2559 				 &mad_agent_priv->send_list, agent_list) {
2560 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2561 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2562 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2563 		}
2564 	}
2565 
2566 	/* Empty wait list to prevent receives from finding a request */
2567 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2568 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2569 
2570 	/* Report all cancelled requests */
2571 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2572 	mad_send_wc.vendor_err = 0;
2573 
2574 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2575 				 &cancel_list, agent_list) {
2576 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2577 		list_del(&mad_send_wr->agent_list);
2578 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2579 						   &mad_send_wc);
2580 		atomic_dec(&mad_agent_priv->refcount);
2581 	}
2582 }
2583 
2584 static struct ib_mad_send_wr_private*
2585 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2586 	     struct ib_mad_send_buf *send_buf)
2587 {
2588 	struct ib_mad_send_wr_private *mad_send_wr;
2589 
2590 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2591 			    agent_list) {
2592 		if (&mad_send_wr->send_buf == send_buf)
2593 			return mad_send_wr;
2594 	}
2595 
2596 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2597 			    agent_list) {
2598 		if (is_rmpp_data_mad(mad_agent_priv,
2599 				     mad_send_wr->send_buf.mad) &&
2600 		    &mad_send_wr->send_buf == send_buf)
2601 			return mad_send_wr;
2602 	}
2603 	return NULL;
2604 }
2605 
2606 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2607 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2608 {
2609 	struct ib_mad_agent_private *mad_agent_priv;
2610 	struct ib_mad_send_wr_private *mad_send_wr;
2611 	unsigned long flags;
2612 	int active;
2613 
2614 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2615 				      agent);
2616 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2617 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2618 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2619 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2620 		return -EINVAL;
2621 	}
2622 
2623 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2624 	if (!timeout_ms) {
2625 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2626 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2627 	}
2628 
2629 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2630 	if (active)
2631 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2632 	else
2633 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2634 
2635 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2636 	return 0;
2637 }
2638 EXPORT_SYMBOL(ib_modify_mad);
2639 
2640 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2641 		   struct ib_mad_send_buf *send_buf)
2642 {
2643 	ib_modify_mad(mad_agent, send_buf, 0);
2644 }
2645 EXPORT_SYMBOL(ib_cancel_mad);
2646 
2647 static void local_completions(struct work_struct *work)
2648 {
2649 	struct ib_mad_agent_private *mad_agent_priv;
2650 	struct ib_mad_local_private *local;
2651 	struct ib_mad_agent_private *recv_mad_agent;
2652 	unsigned long flags;
2653 	int free_mad;
2654 	struct ib_wc wc;
2655 	struct ib_mad_send_wc mad_send_wc;
2656 	bool opa;
2657 
2658 	mad_agent_priv =
2659 		container_of(work, struct ib_mad_agent_private, local_work);
2660 
2661 	opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2662 			       mad_agent_priv->qp_info->port_priv->port_num);
2663 
2664 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2665 	while (!list_empty(&mad_agent_priv->local_list)) {
2666 		local = list_entry(mad_agent_priv->local_list.next,
2667 				   struct ib_mad_local_private,
2668 				   completion_list);
2669 		list_del(&local->completion_list);
2670 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2671 		free_mad = 0;
2672 		if (local->mad_priv) {
2673 			u8 base_version;
2674 			recv_mad_agent = local->recv_mad_agent;
2675 			if (!recv_mad_agent) {
2676 				dev_err(&mad_agent_priv->agent.device->dev,
2677 					"No receive MAD agent for local completion\n");
2678 				free_mad = 1;
2679 				goto local_send_completion;
2680 			}
2681 
2682 			/*
2683 			 * Defined behavior is to complete response
2684 			 * before request
2685 			 */
2686 			build_smp_wc(recv_mad_agent->agent.qp,
2687 				     local->mad_send_wr->send_wr.wr.wr_cqe,
2688 				     be16_to_cpu(IB_LID_PERMISSIVE),
2689 				     local->mad_send_wr->send_wr.pkey_index,
2690 				     recv_mad_agent->agent.port_num, &wc);
2691 
2692 			local->mad_priv->header.recv_wc.wc = &wc;
2693 
2694 			base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2695 			if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2696 				local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2697 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2698 			} else {
2699 				local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2700 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2701 			}
2702 
2703 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2704 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2705 				 &local->mad_priv->header.recv_wc.rmpp_list);
2706 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2707 			local->mad_priv->header.recv_wc.recv_buf.mad =
2708 						(struct ib_mad *)local->mad_priv->mad;
2709 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2710 				snoop_recv(recv_mad_agent->qp_info,
2711 					  &local->mad_priv->header.recv_wc,
2712 					   IB_MAD_SNOOP_RECVS);
2713 			recv_mad_agent->agent.recv_handler(
2714 						&recv_mad_agent->agent,
2715 						&local->mad_send_wr->send_buf,
2716 						&local->mad_priv->header.recv_wc);
2717 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2718 			atomic_dec(&recv_mad_agent->refcount);
2719 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2720 		}
2721 
2722 local_send_completion:
2723 		/* Complete send */
2724 		mad_send_wc.status = IB_WC_SUCCESS;
2725 		mad_send_wc.vendor_err = 0;
2726 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2727 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2728 			snoop_send(mad_agent_priv->qp_info,
2729 				   &local->mad_send_wr->send_buf,
2730 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2731 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2732 						   &mad_send_wc);
2733 
2734 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2735 		atomic_dec(&mad_agent_priv->refcount);
2736 		if (free_mad)
2737 			kfree(local->mad_priv);
2738 		kfree(local);
2739 	}
2740 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2741 }
2742 
2743 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2744 {
2745 	int ret;
2746 
2747 	if (!mad_send_wr->retries_left)
2748 		return -ETIMEDOUT;
2749 
2750 	mad_send_wr->retries_left--;
2751 	mad_send_wr->send_buf.retries++;
2752 
2753 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2754 
2755 	if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2756 		ret = ib_retry_rmpp(mad_send_wr);
2757 		switch (ret) {
2758 		case IB_RMPP_RESULT_UNHANDLED:
2759 			ret = ib_send_mad(mad_send_wr);
2760 			break;
2761 		case IB_RMPP_RESULT_CONSUMED:
2762 			ret = 0;
2763 			break;
2764 		default:
2765 			ret = -ECOMM;
2766 			break;
2767 		}
2768 	} else
2769 		ret = ib_send_mad(mad_send_wr);
2770 
2771 	if (!ret) {
2772 		mad_send_wr->refcount++;
2773 		list_add_tail(&mad_send_wr->agent_list,
2774 			      &mad_send_wr->mad_agent_priv->send_list);
2775 	}
2776 	return ret;
2777 }
2778 
2779 static void timeout_sends(struct work_struct *work)
2780 {
2781 	struct ib_mad_agent_private *mad_agent_priv;
2782 	struct ib_mad_send_wr_private *mad_send_wr;
2783 	struct ib_mad_send_wc mad_send_wc;
2784 	unsigned long flags, delay;
2785 
2786 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2787 				      timed_work.work);
2788 	mad_send_wc.vendor_err = 0;
2789 
2790 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2791 	while (!list_empty(&mad_agent_priv->wait_list)) {
2792 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2793 					 struct ib_mad_send_wr_private,
2794 					 agent_list);
2795 
2796 		if (time_after(mad_send_wr->timeout, jiffies)) {
2797 			delay = mad_send_wr->timeout - jiffies;
2798 			if ((long)delay <= 0)
2799 				delay = 1;
2800 			queue_delayed_work(mad_agent_priv->qp_info->
2801 					   port_priv->wq,
2802 					   &mad_agent_priv->timed_work, delay);
2803 			break;
2804 		}
2805 
2806 		list_del(&mad_send_wr->agent_list);
2807 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2808 		    !retry_send(mad_send_wr))
2809 			continue;
2810 
2811 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2812 
2813 		if (mad_send_wr->status == IB_WC_SUCCESS)
2814 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2815 		else
2816 			mad_send_wc.status = mad_send_wr->status;
2817 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2818 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2819 						   &mad_send_wc);
2820 
2821 		atomic_dec(&mad_agent_priv->refcount);
2822 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2823 	}
2824 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2825 }
2826 
2827 /*
2828  * Allocate receive MADs and post receive WRs for them
2829  */
2830 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2831 				    struct ib_mad_private *mad)
2832 {
2833 	unsigned long flags;
2834 	int post, ret;
2835 	struct ib_mad_private *mad_priv;
2836 	struct ib_sge sg_list;
2837 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2838 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2839 
2840 	/* Initialize common scatter list fields */
2841 	sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2842 
2843 	/* Initialize common receive WR fields */
2844 	recv_wr.next = NULL;
2845 	recv_wr.sg_list = &sg_list;
2846 	recv_wr.num_sge = 1;
2847 
2848 	do {
2849 		/* Allocate and map receive buffer */
2850 		if (mad) {
2851 			mad_priv = mad;
2852 			mad = NULL;
2853 		} else {
2854 			mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2855 						     GFP_ATOMIC);
2856 			if (!mad_priv) {
2857 				ret = -ENOMEM;
2858 				break;
2859 			}
2860 		}
2861 		sg_list.length = mad_priv_dma_size(mad_priv);
2862 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2863 						 &mad_priv->grh,
2864 						 mad_priv_dma_size(mad_priv),
2865 						 DMA_FROM_DEVICE);
2866 		if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2867 						  sg_list.addr))) {
2868 			ret = -ENOMEM;
2869 			break;
2870 		}
2871 		mad_priv->header.mapping = sg_list.addr;
2872 		mad_priv->header.mad_list.mad_queue = recv_queue;
2873 		mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2874 		recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2875 
2876 		/* Post receive WR */
2877 		spin_lock_irqsave(&recv_queue->lock, flags);
2878 		post = (++recv_queue->count < recv_queue->max_active);
2879 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2880 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2881 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2882 		if (ret) {
2883 			spin_lock_irqsave(&recv_queue->lock, flags);
2884 			list_del(&mad_priv->header.mad_list.list);
2885 			recv_queue->count--;
2886 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2887 			ib_dma_unmap_single(qp_info->port_priv->device,
2888 					    mad_priv->header.mapping,
2889 					    mad_priv_dma_size(mad_priv),
2890 					    DMA_FROM_DEVICE);
2891 			kfree(mad_priv);
2892 			dev_err(&qp_info->port_priv->device->dev,
2893 				"ib_post_recv failed: %d\n", ret);
2894 			break;
2895 		}
2896 	} while (post);
2897 
2898 	return ret;
2899 }
2900 
2901 /*
2902  * Return all the posted receive MADs
2903  */
2904 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2905 {
2906 	struct ib_mad_private_header *mad_priv_hdr;
2907 	struct ib_mad_private *recv;
2908 	struct ib_mad_list_head *mad_list;
2909 
2910 	if (!qp_info->qp)
2911 		return;
2912 
2913 	while (!list_empty(&qp_info->recv_queue.list)) {
2914 
2915 		mad_list = list_entry(qp_info->recv_queue.list.next,
2916 				      struct ib_mad_list_head, list);
2917 		mad_priv_hdr = container_of(mad_list,
2918 					    struct ib_mad_private_header,
2919 					    mad_list);
2920 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2921 				    header);
2922 
2923 		/* Remove from posted receive MAD list */
2924 		list_del(&mad_list->list);
2925 
2926 		ib_dma_unmap_single(qp_info->port_priv->device,
2927 				    recv->header.mapping,
2928 				    mad_priv_dma_size(recv),
2929 				    DMA_FROM_DEVICE);
2930 		kfree(recv);
2931 	}
2932 
2933 	qp_info->recv_queue.count = 0;
2934 }
2935 
2936 /*
2937  * Start the port
2938  */
2939 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2940 {
2941 	int ret, i;
2942 	struct ib_qp_attr *attr;
2943 	struct ib_qp *qp;
2944 	u16 pkey_index;
2945 
2946 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2947 	if (!attr)
2948 		return -ENOMEM;
2949 
2950 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2951 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
2952 	if (ret)
2953 		pkey_index = 0;
2954 
2955 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2956 		qp = port_priv->qp_info[i].qp;
2957 		if (!qp)
2958 			continue;
2959 
2960 		/*
2961 		 * PKey index for QP1 is irrelevant but
2962 		 * one is needed for the Reset to Init transition
2963 		 */
2964 		attr->qp_state = IB_QPS_INIT;
2965 		attr->pkey_index = pkey_index;
2966 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2967 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2968 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2969 		if (ret) {
2970 			dev_err(&port_priv->device->dev,
2971 				"Couldn't change QP%d state to INIT: %d\n",
2972 				i, ret);
2973 			goto out;
2974 		}
2975 
2976 		attr->qp_state = IB_QPS_RTR;
2977 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2978 		if (ret) {
2979 			dev_err(&port_priv->device->dev,
2980 				"Couldn't change QP%d state to RTR: %d\n",
2981 				i, ret);
2982 			goto out;
2983 		}
2984 
2985 		attr->qp_state = IB_QPS_RTS;
2986 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2987 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2988 		if (ret) {
2989 			dev_err(&port_priv->device->dev,
2990 				"Couldn't change QP%d state to RTS: %d\n",
2991 				i, ret);
2992 			goto out;
2993 		}
2994 	}
2995 
2996 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2997 	if (ret) {
2998 		dev_err(&port_priv->device->dev,
2999 			"Failed to request completion notification: %d\n",
3000 			ret);
3001 		goto out;
3002 	}
3003 
3004 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3005 		if (!port_priv->qp_info[i].qp)
3006 			continue;
3007 
3008 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3009 		if (ret) {
3010 			dev_err(&port_priv->device->dev,
3011 				"Couldn't post receive WRs\n");
3012 			goto out;
3013 		}
3014 	}
3015 out:
3016 	kfree(attr);
3017 	return ret;
3018 }
3019 
3020 static void qp_event_handler(struct ib_event *event, void *qp_context)
3021 {
3022 	struct ib_mad_qp_info	*qp_info = qp_context;
3023 
3024 	/* It's worse than that! He's dead, Jim! */
3025 	dev_err(&qp_info->port_priv->device->dev,
3026 		"Fatal error (%d) on MAD QP (%d)\n",
3027 		event->event, qp_info->qp->qp_num);
3028 }
3029 
3030 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3031 			   struct ib_mad_queue *mad_queue)
3032 {
3033 	mad_queue->qp_info = qp_info;
3034 	mad_queue->count = 0;
3035 	spin_lock_init(&mad_queue->lock);
3036 	INIT_LIST_HEAD(&mad_queue->list);
3037 }
3038 
3039 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3040 			struct ib_mad_qp_info *qp_info)
3041 {
3042 	qp_info->port_priv = port_priv;
3043 	init_mad_queue(qp_info, &qp_info->send_queue);
3044 	init_mad_queue(qp_info, &qp_info->recv_queue);
3045 	INIT_LIST_HEAD(&qp_info->overflow_list);
3046 	spin_lock_init(&qp_info->snoop_lock);
3047 	qp_info->snoop_table = NULL;
3048 	qp_info->snoop_table_size = 0;
3049 	atomic_set(&qp_info->snoop_count, 0);
3050 }
3051 
3052 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3053 			 enum ib_qp_type qp_type)
3054 {
3055 	struct ib_qp_init_attr	qp_init_attr;
3056 	int ret;
3057 
3058 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
3059 	qp_init_attr.send_cq = qp_info->port_priv->cq;
3060 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
3061 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3062 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
3063 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3064 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3065 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3066 	qp_init_attr.qp_type = qp_type;
3067 	qp_init_attr.port_num = qp_info->port_priv->port_num;
3068 	qp_init_attr.qp_context = qp_info;
3069 	qp_init_attr.event_handler = qp_event_handler;
3070 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3071 	if (IS_ERR(qp_info->qp)) {
3072 		dev_err(&qp_info->port_priv->device->dev,
3073 			"Couldn't create ib_mad QP%d\n",
3074 			get_spl_qp_index(qp_type));
3075 		ret = PTR_ERR(qp_info->qp);
3076 		goto error;
3077 	}
3078 	/* Use minimum queue sizes unless the CQ is resized */
3079 	qp_info->send_queue.max_active = mad_sendq_size;
3080 	qp_info->recv_queue.max_active = mad_recvq_size;
3081 	return 0;
3082 
3083 error:
3084 	return ret;
3085 }
3086 
3087 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3088 {
3089 	if (!qp_info->qp)
3090 		return;
3091 
3092 	ib_destroy_qp(qp_info->qp);
3093 	kfree(qp_info->snoop_table);
3094 }
3095 
3096 /*
3097  * Open the port
3098  * Create the QP, PD, MR, and CQ if needed
3099  */
3100 static int ib_mad_port_open(struct ib_device *device,
3101 			    int port_num)
3102 {
3103 	int ret, cq_size;
3104 	struct ib_mad_port_private *port_priv;
3105 	unsigned long flags;
3106 	char name[sizeof "ib_mad123"];
3107 	int has_smi;
3108 
3109 	if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3110 		return -EFAULT;
3111 
3112 	if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3113 		    rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3114 		return -EFAULT;
3115 
3116 	/* Create new device info */
3117 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3118 	if (!port_priv)
3119 		return -ENOMEM;
3120 
3121 	port_priv->device = device;
3122 	port_priv->port_num = port_num;
3123 	spin_lock_init(&port_priv->reg_lock);
3124 	INIT_LIST_HEAD(&port_priv->agent_list);
3125 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
3126 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
3127 
3128 	cq_size = mad_sendq_size + mad_recvq_size;
3129 	has_smi = rdma_cap_ib_smi(device, port_num);
3130 	if (has_smi)
3131 		cq_size *= 2;
3132 
3133 	port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3134 			IB_POLL_WORKQUEUE);
3135 	if (IS_ERR(port_priv->cq)) {
3136 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3137 		ret = PTR_ERR(port_priv->cq);
3138 		goto error3;
3139 	}
3140 
3141 	port_priv->pd = ib_alloc_pd(device, 0);
3142 	if (IS_ERR(port_priv->pd)) {
3143 		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3144 		ret = PTR_ERR(port_priv->pd);
3145 		goto error4;
3146 	}
3147 
3148 	if (has_smi) {
3149 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3150 		if (ret)
3151 			goto error6;
3152 	}
3153 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3154 	if (ret)
3155 		goto error7;
3156 
3157 	snprintf(name, sizeof name, "ib_mad%d", port_num);
3158 	port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3159 	if (!port_priv->wq) {
3160 		ret = -ENOMEM;
3161 		goto error8;
3162 	}
3163 
3164 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3165 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3166 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3167 
3168 	ret = ib_mad_port_start(port_priv);
3169 	if (ret) {
3170 		dev_err(&device->dev, "Couldn't start port\n");
3171 		goto error9;
3172 	}
3173 
3174 	return 0;
3175 
3176 error9:
3177 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3178 	list_del_init(&port_priv->port_list);
3179 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3180 
3181 	destroy_workqueue(port_priv->wq);
3182 error8:
3183 	destroy_mad_qp(&port_priv->qp_info[1]);
3184 error7:
3185 	destroy_mad_qp(&port_priv->qp_info[0]);
3186 error6:
3187 	ib_dealloc_pd(port_priv->pd);
3188 error4:
3189 	ib_free_cq(port_priv->cq);
3190 	cleanup_recv_queue(&port_priv->qp_info[1]);
3191 	cleanup_recv_queue(&port_priv->qp_info[0]);
3192 error3:
3193 	kfree(port_priv);
3194 
3195 	return ret;
3196 }
3197 
3198 /*
3199  * Close the port
3200  * If there are no classes using the port, free the port
3201  * resources (CQ, MR, PD, QP) and remove the port's info structure
3202  */
3203 static int ib_mad_port_close(struct ib_device *device, int port_num)
3204 {
3205 	struct ib_mad_port_private *port_priv;
3206 	unsigned long flags;
3207 
3208 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3209 	port_priv = __ib_get_mad_port(device, port_num);
3210 	if (port_priv == NULL) {
3211 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3212 		dev_err(&device->dev, "Port %d not found\n", port_num);
3213 		return -ENODEV;
3214 	}
3215 	list_del_init(&port_priv->port_list);
3216 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3217 
3218 	destroy_workqueue(port_priv->wq);
3219 	destroy_mad_qp(&port_priv->qp_info[1]);
3220 	destroy_mad_qp(&port_priv->qp_info[0]);
3221 	ib_dealloc_pd(port_priv->pd);
3222 	ib_free_cq(port_priv->cq);
3223 	cleanup_recv_queue(&port_priv->qp_info[1]);
3224 	cleanup_recv_queue(&port_priv->qp_info[0]);
3225 	/* XXX: Handle deallocation of MAD registration tables */
3226 
3227 	kfree(port_priv);
3228 
3229 	return 0;
3230 }
3231 
3232 static void ib_mad_init_device(struct ib_device *device)
3233 {
3234 	int start, i;
3235 
3236 	start = rdma_start_port(device);
3237 
3238 	for (i = start; i <= rdma_end_port(device); i++) {
3239 		if (!rdma_cap_ib_mad(device, i))
3240 			continue;
3241 
3242 		if (ib_mad_port_open(device, i)) {
3243 			dev_err(&device->dev, "Couldn't open port %d\n", i);
3244 			goto error;
3245 		}
3246 		if (ib_agent_port_open(device, i)) {
3247 			dev_err(&device->dev,
3248 				"Couldn't open port %d for agents\n", i);
3249 			goto error_agent;
3250 		}
3251 	}
3252 	return;
3253 
3254 error_agent:
3255 	if (ib_mad_port_close(device, i))
3256 		dev_err(&device->dev, "Couldn't close port %d\n", i);
3257 
3258 error:
3259 	while (--i >= start) {
3260 		if (!rdma_cap_ib_mad(device, i))
3261 			continue;
3262 
3263 		if (ib_agent_port_close(device, i))
3264 			dev_err(&device->dev,
3265 				"Couldn't close port %d for agents\n", i);
3266 		if (ib_mad_port_close(device, i))
3267 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3268 	}
3269 }
3270 
3271 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3272 {
3273 	int i;
3274 
3275 	for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3276 		if (!rdma_cap_ib_mad(device, i))
3277 			continue;
3278 
3279 		if (ib_agent_port_close(device, i))
3280 			dev_err(&device->dev,
3281 				"Couldn't close port %d for agents\n", i);
3282 		if (ib_mad_port_close(device, i))
3283 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3284 	}
3285 }
3286 
3287 static struct ib_client mad_client = {
3288 	.name   = "mad",
3289 	.add = ib_mad_init_device,
3290 	.remove = ib_mad_remove_device
3291 };
3292 
3293 int ib_mad_init(void)
3294 {
3295 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3296 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3297 
3298 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3299 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3300 
3301 	INIT_LIST_HEAD(&ib_mad_port_list);
3302 
3303 	if (ib_register_client(&mad_client)) {
3304 		pr_err("Couldn't register ib_mad client\n");
3305 		return -EINVAL;
3306 	}
3307 
3308 	return 0;
3309 }
3310 
3311 void ib_mad_cleanup(void)
3312 {
3313 	ib_unregister_client(&mad_client);
3314 }
3315