xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision f8e17c17)
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) 2014,2018 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 <linux/security.h>
44 #include <linux/xarray.h>
45 #include <rdma/ib_cache.h>
46 
47 #include "mad_priv.h"
48 #include "core_priv.h"
49 #include "mad_rmpp.h"
50 #include "smi.h"
51 #include "opa_smi.h"
52 #include "agent.h"
53 
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/ib_mad.h>
56 
57 #ifdef CONFIG_TRACEPOINTS
58 static void create_mad_addr_info(struct ib_mad_send_wr_private *mad_send_wr,
59 			  struct ib_mad_qp_info *qp_info,
60 			  struct trace_event_raw_ib_mad_send_template *entry)
61 {
62 	u16 pkey;
63 	struct ib_device *dev = qp_info->port_priv->device;
64 	u8 pnum = qp_info->port_priv->port_num;
65 	struct ib_ud_wr *wr = &mad_send_wr->send_wr;
66 	struct rdma_ah_attr attr = {};
67 
68 	rdma_query_ah(wr->ah, &attr);
69 
70 	/* These are common */
71 	entry->sl = attr.sl;
72 	ib_query_pkey(dev, pnum, wr->pkey_index, &pkey);
73 	entry->pkey = pkey;
74 	entry->rqpn = wr->remote_qpn;
75 	entry->rqkey = wr->remote_qkey;
76 	entry->dlid = rdma_ah_get_dlid(&attr);
77 }
78 #endif
79 
80 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
81 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
82 
83 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
84 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
85 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
86 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
87 
88 /* Client ID 0 is used for snoop-only clients */
89 static DEFINE_XARRAY_ALLOC1(ib_mad_clients);
90 static u32 ib_mad_client_next;
91 static struct list_head ib_mad_port_list;
92 
93 /* Port list lock */
94 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
95 
96 /* Forward declarations */
97 static int method_in_use(struct ib_mad_mgmt_method_table **method,
98 			 struct ib_mad_reg_req *mad_reg_req);
99 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
100 static struct ib_mad_agent_private *find_mad_agent(
101 					struct ib_mad_port_private *port_priv,
102 					const struct ib_mad_hdr *mad);
103 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
104 				    struct ib_mad_private *mad);
105 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
106 static void timeout_sends(struct work_struct *work);
107 static void local_completions(struct work_struct *work);
108 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
109 			      struct ib_mad_agent_private *agent_priv,
110 			      u8 mgmt_class);
111 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
112 			   struct ib_mad_agent_private *agent_priv);
113 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
114 			      struct ib_wc *wc);
115 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
116 
117 /*
118  * Returns a ib_mad_port_private structure or NULL for a device/port
119  * Assumes ib_mad_port_list_lock is being held
120  */
121 static inline struct ib_mad_port_private *
122 __ib_get_mad_port(struct ib_device *device, int port_num)
123 {
124 	struct ib_mad_port_private *entry;
125 
126 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
127 		if (entry->device == device && entry->port_num == port_num)
128 			return entry;
129 	}
130 	return NULL;
131 }
132 
133 /*
134  * Wrapper function to return a ib_mad_port_private structure or NULL
135  * for a device/port
136  */
137 static inline struct ib_mad_port_private *
138 ib_get_mad_port(struct ib_device *device, int port_num)
139 {
140 	struct ib_mad_port_private *entry;
141 	unsigned long flags;
142 
143 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
144 	entry = __ib_get_mad_port(device, port_num);
145 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
146 
147 	return entry;
148 }
149 
150 static inline u8 convert_mgmt_class(u8 mgmt_class)
151 {
152 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
153 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
154 		0 : mgmt_class;
155 }
156 
157 static int get_spl_qp_index(enum ib_qp_type qp_type)
158 {
159 	switch (qp_type)
160 	{
161 	case IB_QPT_SMI:
162 		return 0;
163 	case IB_QPT_GSI:
164 		return 1;
165 	default:
166 		return -1;
167 	}
168 }
169 
170 static int vendor_class_index(u8 mgmt_class)
171 {
172 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
173 }
174 
175 static int is_vendor_class(u8 mgmt_class)
176 {
177 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
178 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
179 		return 0;
180 	return 1;
181 }
182 
183 static int is_vendor_oui(char *oui)
184 {
185 	if (oui[0] || oui[1] || oui[2])
186 		return 1;
187 	return 0;
188 }
189 
190 static int is_vendor_method_in_use(
191 		struct ib_mad_mgmt_vendor_class *vendor_class,
192 		struct ib_mad_reg_req *mad_reg_req)
193 {
194 	struct ib_mad_mgmt_method_table *method;
195 	int i;
196 
197 	for (i = 0; i < MAX_MGMT_OUI; i++) {
198 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
199 			method = vendor_class->method_table[i];
200 			if (method) {
201 				if (method_in_use(&method, mad_reg_req))
202 					return 1;
203 				else
204 					break;
205 			}
206 		}
207 	}
208 	return 0;
209 }
210 
211 int ib_response_mad(const struct ib_mad_hdr *hdr)
212 {
213 	return ((hdr->method & IB_MGMT_METHOD_RESP) ||
214 		(hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
215 		((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
216 		 (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
217 }
218 EXPORT_SYMBOL(ib_response_mad);
219 
220 /*
221  * ib_register_mad_agent - Register to send/receive MADs
222  *
223  * Context: Process context.
224  */
225 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
226 					   u8 port_num,
227 					   enum ib_qp_type qp_type,
228 					   struct ib_mad_reg_req *mad_reg_req,
229 					   u8 rmpp_version,
230 					   ib_mad_send_handler send_handler,
231 					   ib_mad_recv_handler recv_handler,
232 					   void *context,
233 					   u32 registration_flags)
234 {
235 	struct ib_mad_port_private *port_priv;
236 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
237 	struct ib_mad_agent_private *mad_agent_priv;
238 	struct ib_mad_reg_req *reg_req = NULL;
239 	struct ib_mad_mgmt_class_table *class;
240 	struct ib_mad_mgmt_vendor_class_table *vendor;
241 	struct ib_mad_mgmt_vendor_class *vendor_class;
242 	struct ib_mad_mgmt_method_table *method;
243 	int ret2, qpn;
244 	u8 mgmt_class, vclass;
245 
246 	if ((qp_type == IB_QPT_SMI && !rdma_cap_ib_smi(device, port_num)) ||
247 	    (qp_type == IB_QPT_GSI && !rdma_cap_ib_cm(device, port_num)))
248 		return ERR_PTR(-EPROTONOSUPPORT);
249 
250 	/* Validate parameters */
251 	qpn = get_spl_qp_index(qp_type);
252 	if (qpn == -1) {
253 		dev_dbg_ratelimited(&device->dev, "%s: invalid QP Type %d\n",
254 				    __func__, qp_type);
255 		goto error1;
256 	}
257 
258 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
259 		dev_dbg_ratelimited(&device->dev,
260 				    "%s: invalid RMPP Version %u\n",
261 				    __func__, rmpp_version);
262 		goto error1;
263 	}
264 
265 	/* Validate MAD registration request if supplied */
266 	if (mad_reg_req) {
267 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
268 			dev_dbg_ratelimited(&device->dev,
269 					    "%s: invalid Class Version %u\n",
270 					    __func__,
271 					    mad_reg_req->mgmt_class_version);
272 			goto error1;
273 		}
274 		if (!recv_handler) {
275 			dev_dbg_ratelimited(&device->dev,
276 					    "%s: no recv_handler\n", __func__);
277 			goto error1;
278 		}
279 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
280 			/*
281 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
282 			 * one in this range currently allowed
283 			 */
284 			if (mad_reg_req->mgmt_class !=
285 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
286 				dev_dbg_ratelimited(&device->dev,
287 					"%s: Invalid Mgmt Class 0x%x\n",
288 					__func__, mad_reg_req->mgmt_class);
289 				goto error1;
290 			}
291 		} else if (mad_reg_req->mgmt_class == 0) {
292 			/*
293 			 * Class 0 is reserved in IBA and is used for
294 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
295 			 */
296 			dev_dbg_ratelimited(&device->dev,
297 					    "%s: Invalid Mgmt Class 0\n",
298 					    __func__);
299 			goto error1;
300 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
301 			/*
302 			 * If class is in "new" vendor range,
303 			 * ensure supplied OUI is not zero
304 			 */
305 			if (!is_vendor_oui(mad_reg_req->oui)) {
306 				dev_dbg_ratelimited(&device->dev,
307 					"%s: No OUI specified for class 0x%x\n",
308 					__func__,
309 					mad_reg_req->mgmt_class);
310 				goto error1;
311 			}
312 		}
313 		/* Make sure class supplied is consistent with RMPP */
314 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
315 			if (rmpp_version) {
316 				dev_dbg_ratelimited(&device->dev,
317 					"%s: RMPP version for non-RMPP class 0x%x\n",
318 					__func__, mad_reg_req->mgmt_class);
319 				goto error1;
320 			}
321 		}
322 
323 		/* Make sure class supplied is consistent with QP type */
324 		if (qp_type == IB_QPT_SMI) {
325 			if ((mad_reg_req->mgmt_class !=
326 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
327 			    (mad_reg_req->mgmt_class !=
328 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
329 				dev_dbg_ratelimited(&device->dev,
330 					"%s: Invalid SM QP type: class 0x%x\n",
331 					__func__, mad_reg_req->mgmt_class);
332 				goto error1;
333 			}
334 		} else {
335 			if ((mad_reg_req->mgmt_class ==
336 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
337 			    (mad_reg_req->mgmt_class ==
338 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
339 				dev_dbg_ratelimited(&device->dev,
340 					"%s: Invalid GS QP type: class 0x%x\n",
341 					__func__, mad_reg_req->mgmt_class);
342 				goto error1;
343 			}
344 		}
345 	} else {
346 		/* No registration request supplied */
347 		if (!send_handler)
348 			goto error1;
349 		if (registration_flags & IB_MAD_USER_RMPP)
350 			goto error1;
351 	}
352 
353 	/* Validate device and port */
354 	port_priv = ib_get_mad_port(device, port_num);
355 	if (!port_priv) {
356 		dev_dbg_ratelimited(&device->dev, "%s: Invalid port %d\n",
357 				    __func__, port_num);
358 		ret = ERR_PTR(-ENODEV);
359 		goto error1;
360 	}
361 
362 	/* Verify the QP requested is supported. For example, Ethernet devices
363 	 * will not have QP0.
364 	 */
365 	if (!port_priv->qp_info[qpn].qp) {
366 		dev_dbg_ratelimited(&device->dev, "%s: QP %d not supported\n",
367 				    __func__, qpn);
368 		ret = ERR_PTR(-EPROTONOSUPPORT);
369 		goto error1;
370 	}
371 
372 	/* Allocate structures */
373 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
374 	if (!mad_agent_priv) {
375 		ret = ERR_PTR(-ENOMEM);
376 		goto error1;
377 	}
378 
379 	if (mad_reg_req) {
380 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
381 		if (!reg_req) {
382 			ret = ERR_PTR(-ENOMEM);
383 			goto error3;
384 		}
385 	}
386 
387 	/* Now, fill in the various structures */
388 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
389 	mad_agent_priv->reg_req = reg_req;
390 	mad_agent_priv->agent.rmpp_version = rmpp_version;
391 	mad_agent_priv->agent.device = device;
392 	mad_agent_priv->agent.recv_handler = recv_handler;
393 	mad_agent_priv->agent.send_handler = send_handler;
394 	mad_agent_priv->agent.context = context;
395 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
396 	mad_agent_priv->agent.port_num = port_num;
397 	mad_agent_priv->agent.flags = registration_flags;
398 	spin_lock_init(&mad_agent_priv->lock);
399 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
400 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
401 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
402 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
403 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
404 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
405 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
406 	atomic_set(&mad_agent_priv->refcount, 1);
407 	init_completion(&mad_agent_priv->comp);
408 
409 	ret2 = ib_mad_agent_security_setup(&mad_agent_priv->agent, qp_type);
410 	if (ret2) {
411 		ret = ERR_PTR(ret2);
412 		goto error4;
413 	}
414 
415 	/*
416 	 * The mlx4 driver uses the top byte to distinguish which virtual
417 	 * function generated the MAD, so we must avoid using it.
418 	 */
419 	ret2 = xa_alloc_cyclic(&ib_mad_clients, &mad_agent_priv->agent.hi_tid,
420 			mad_agent_priv, XA_LIMIT(0, (1 << 24) - 1),
421 			&ib_mad_client_next, GFP_KERNEL);
422 	if (ret2 < 0) {
423 		ret = ERR_PTR(ret2);
424 		goto error5;
425 	}
426 
427 	/*
428 	 * Make sure MAD registration (if supplied)
429 	 * is non overlapping with any existing ones
430 	 */
431 	spin_lock_irq(&port_priv->reg_lock);
432 	if (mad_reg_req) {
433 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
434 		if (!is_vendor_class(mgmt_class)) {
435 			class = port_priv->version[mad_reg_req->
436 						   mgmt_class_version].class;
437 			if (class) {
438 				method = class->method_table[mgmt_class];
439 				if (method) {
440 					if (method_in_use(&method,
441 							   mad_reg_req))
442 						goto error6;
443 				}
444 			}
445 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
446 						  mgmt_class);
447 		} else {
448 			/* "New" vendor class range */
449 			vendor = port_priv->version[mad_reg_req->
450 						    mgmt_class_version].vendor;
451 			if (vendor) {
452 				vclass = vendor_class_index(mgmt_class);
453 				vendor_class = vendor->vendor_class[vclass];
454 				if (vendor_class) {
455 					if (is_vendor_method_in_use(
456 							vendor_class,
457 							mad_reg_req))
458 						goto error6;
459 				}
460 			}
461 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
462 		}
463 		if (ret2) {
464 			ret = ERR_PTR(ret2);
465 			goto error6;
466 		}
467 	}
468 	spin_unlock_irq(&port_priv->reg_lock);
469 
470 	trace_ib_mad_create_agent(mad_agent_priv);
471 	return &mad_agent_priv->agent;
472 error6:
473 	spin_unlock_irq(&port_priv->reg_lock);
474 	xa_erase(&ib_mad_clients, mad_agent_priv->agent.hi_tid);
475 error5:
476 	ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
477 error4:
478 	kfree(reg_req);
479 error3:
480 	kfree(mad_agent_priv);
481 error1:
482 	return ret;
483 }
484 EXPORT_SYMBOL(ib_register_mad_agent);
485 
486 static inline int is_snooping_sends(int mad_snoop_flags)
487 {
488 	return (mad_snoop_flags &
489 		(/*IB_MAD_SNOOP_POSTED_SENDS |
490 		 IB_MAD_SNOOP_RMPP_SENDS |*/
491 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
492 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
493 }
494 
495 static inline int is_snooping_recvs(int mad_snoop_flags)
496 {
497 	return (mad_snoop_flags &
498 		(IB_MAD_SNOOP_RECVS /*|
499 		 IB_MAD_SNOOP_RMPP_RECVS*/));
500 }
501 
502 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
503 				struct ib_mad_snoop_private *mad_snoop_priv)
504 {
505 	struct ib_mad_snoop_private **new_snoop_table;
506 	unsigned long flags;
507 	int i;
508 
509 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
510 	/* Check for empty slot in array. */
511 	for (i = 0; i < qp_info->snoop_table_size; i++)
512 		if (!qp_info->snoop_table[i])
513 			break;
514 
515 	if (i == qp_info->snoop_table_size) {
516 		/* Grow table. */
517 		new_snoop_table = krealloc(qp_info->snoop_table,
518 					   sizeof mad_snoop_priv *
519 					   (qp_info->snoop_table_size + 1),
520 					   GFP_ATOMIC);
521 		if (!new_snoop_table) {
522 			i = -ENOMEM;
523 			goto out;
524 		}
525 
526 		qp_info->snoop_table = new_snoop_table;
527 		qp_info->snoop_table_size++;
528 	}
529 	qp_info->snoop_table[i] = mad_snoop_priv;
530 	atomic_inc(&qp_info->snoop_count);
531 out:
532 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
533 	return i;
534 }
535 
536 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
537 					   u8 port_num,
538 					   enum ib_qp_type qp_type,
539 					   int mad_snoop_flags,
540 					   ib_mad_snoop_handler snoop_handler,
541 					   ib_mad_recv_handler recv_handler,
542 					   void *context)
543 {
544 	struct ib_mad_port_private *port_priv;
545 	struct ib_mad_agent *ret;
546 	struct ib_mad_snoop_private *mad_snoop_priv;
547 	int qpn;
548 	int err;
549 
550 	/* Validate parameters */
551 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
552 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
553 		ret = ERR_PTR(-EINVAL);
554 		goto error1;
555 	}
556 	qpn = get_spl_qp_index(qp_type);
557 	if (qpn == -1) {
558 		ret = ERR_PTR(-EINVAL);
559 		goto error1;
560 	}
561 	port_priv = ib_get_mad_port(device, port_num);
562 	if (!port_priv) {
563 		ret = ERR_PTR(-ENODEV);
564 		goto error1;
565 	}
566 	/* Allocate structures */
567 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
568 	if (!mad_snoop_priv) {
569 		ret = ERR_PTR(-ENOMEM);
570 		goto error1;
571 	}
572 
573 	/* Now, fill in the various structures */
574 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
575 	mad_snoop_priv->agent.device = device;
576 	mad_snoop_priv->agent.recv_handler = recv_handler;
577 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
578 	mad_snoop_priv->agent.context = context;
579 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
580 	mad_snoop_priv->agent.port_num = port_num;
581 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
582 	init_completion(&mad_snoop_priv->comp);
583 
584 	err = ib_mad_agent_security_setup(&mad_snoop_priv->agent, qp_type);
585 	if (err) {
586 		ret = ERR_PTR(err);
587 		goto error2;
588 	}
589 
590 	mad_snoop_priv->snoop_index = register_snoop_agent(
591 						&port_priv->qp_info[qpn],
592 						mad_snoop_priv);
593 	if (mad_snoop_priv->snoop_index < 0) {
594 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
595 		goto error3;
596 	}
597 
598 	atomic_set(&mad_snoop_priv->refcount, 1);
599 	return &mad_snoop_priv->agent;
600 error3:
601 	ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
602 error2:
603 	kfree(mad_snoop_priv);
604 error1:
605 	return ret;
606 }
607 EXPORT_SYMBOL(ib_register_mad_snoop);
608 
609 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
610 {
611 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
612 		complete(&mad_agent_priv->comp);
613 }
614 
615 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
616 {
617 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
618 		complete(&mad_snoop_priv->comp);
619 }
620 
621 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
622 {
623 	struct ib_mad_port_private *port_priv;
624 
625 	/* Note that we could still be handling received MADs */
626 	trace_ib_mad_unregister_agent(mad_agent_priv);
627 
628 	/*
629 	 * Canceling all sends results in dropping received response
630 	 * MADs, preventing us from queuing additional work
631 	 */
632 	cancel_mads(mad_agent_priv);
633 	port_priv = mad_agent_priv->qp_info->port_priv;
634 	cancel_delayed_work(&mad_agent_priv->timed_work);
635 
636 	spin_lock_irq(&port_priv->reg_lock);
637 	remove_mad_reg_req(mad_agent_priv);
638 	spin_unlock_irq(&port_priv->reg_lock);
639 	xa_erase(&ib_mad_clients, mad_agent_priv->agent.hi_tid);
640 
641 	flush_workqueue(port_priv->wq);
642 	ib_cancel_rmpp_recvs(mad_agent_priv);
643 
644 	deref_mad_agent(mad_agent_priv);
645 	wait_for_completion(&mad_agent_priv->comp);
646 
647 	ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
648 
649 	kfree(mad_agent_priv->reg_req);
650 	kfree_rcu(mad_agent_priv, rcu);
651 }
652 
653 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
654 {
655 	struct ib_mad_qp_info *qp_info;
656 	unsigned long flags;
657 
658 	qp_info = mad_snoop_priv->qp_info;
659 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
660 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
661 	atomic_dec(&qp_info->snoop_count);
662 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
663 
664 	deref_snoop_agent(mad_snoop_priv);
665 	wait_for_completion(&mad_snoop_priv->comp);
666 
667 	ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
668 
669 	kfree(mad_snoop_priv);
670 }
671 
672 /*
673  * ib_unregister_mad_agent - Unregisters a client from using MAD services
674  *
675  * Context: Process context.
676  */
677 void ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
678 {
679 	struct ib_mad_agent_private *mad_agent_priv;
680 	struct ib_mad_snoop_private *mad_snoop_priv;
681 
682 	/* If the TID is zero, the agent can only snoop. */
683 	if (mad_agent->hi_tid) {
684 		mad_agent_priv = container_of(mad_agent,
685 					      struct ib_mad_agent_private,
686 					      agent);
687 		unregister_mad_agent(mad_agent_priv);
688 	} else {
689 		mad_snoop_priv = container_of(mad_agent,
690 					      struct ib_mad_snoop_private,
691 					      agent);
692 		unregister_mad_snoop(mad_snoop_priv);
693 	}
694 }
695 EXPORT_SYMBOL(ib_unregister_mad_agent);
696 
697 static void dequeue_mad(struct ib_mad_list_head *mad_list)
698 {
699 	struct ib_mad_queue *mad_queue;
700 	unsigned long flags;
701 
702 	mad_queue = mad_list->mad_queue;
703 	spin_lock_irqsave(&mad_queue->lock, flags);
704 	list_del(&mad_list->list);
705 	mad_queue->count--;
706 	spin_unlock_irqrestore(&mad_queue->lock, flags);
707 }
708 
709 static void snoop_send(struct ib_mad_qp_info *qp_info,
710 		       struct ib_mad_send_buf *send_buf,
711 		       struct ib_mad_send_wc *mad_send_wc,
712 		       int mad_snoop_flags)
713 {
714 	struct ib_mad_snoop_private *mad_snoop_priv;
715 	unsigned long flags;
716 	int i;
717 
718 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
719 	for (i = 0; i < qp_info->snoop_table_size; i++) {
720 		mad_snoop_priv = qp_info->snoop_table[i];
721 		if (!mad_snoop_priv ||
722 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
723 			continue;
724 
725 		atomic_inc(&mad_snoop_priv->refcount);
726 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
727 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
728 						    send_buf, mad_send_wc);
729 		deref_snoop_agent(mad_snoop_priv);
730 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
731 	}
732 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
733 }
734 
735 static void snoop_recv(struct ib_mad_qp_info *qp_info,
736 		       struct ib_mad_recv_wc *mad_recv_wc,
737 		       int mad_snoop_flags)
738 {
739 	struct ib_mad_snoop_private *mad_snoop_priv;
740 	unsigned long flags;
741 	int i;
742 
743 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
744 	for (i = 0; i < qp_info->snoop_table_size; i++) {
745 		mad_snoop_priv = qp_info->snoop_table[i];
746 		if (!mad_snoop_priv ||
747 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
748 			continue;
749 
750 		atomic_inc(&mad_snoop_priv->refcount);
751 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
752 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
753 						   mad_recv_wc);
754 		deref_snoop_agent(mad_snoop_priv);
755 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
756 	}
757 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
758 }
759 
760 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
761 		u16 pkey_index, u8 port_num, struct ib_wc *wc)
762 {
763 	memset(wc, 0, sizeof *wc);
764 	wc->wr_cqe = cqe;
765 	wc->status = IB_WC_SUCCESS;
766 	wc->opcode = IB_WC_RECV;
767 	wc->pkey_index = pkey_index;
768 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
769 	wc->src_qp = IB_QP0;
770 	wc->qp = qp;
771 	wc->slid = slid;
772 	wc->sl = 0;
773 	wc->dlid_path_bits = 0;
774 	wc->port_num = port_num;
775 }
776 
777 static size_t mad_priv_size(const struct ib_mad_private *mp)
778 {
779 	return sizeof(struct ib_mad_private) + mp->mad_size;
780 }
781 
782 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
783 {
784 	size_t size = sizeof(struct ib_mad_private) + mad_size;
785 	struct ib_mad_private *ret = kzalloc(size, flags);
786 
787 	if (ret)
788 		ret->mad_size = mad_size;
789 
790 	return ret;
791 }
792 
793 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
794 {
795 	return rdma_max_mad_size(port_priv->device, port_priv->port_num);
796 }
797 
798 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
799 {
800 	return sizeof(struct ib_grh) + mp->mad_size;
801 }
802 
803 /*
804  * Return 0 if SMP is to be sent
805  * Return 1 if SMP was consumed locally (whether or not solicited)
806  * Return < 0 if error
807  */
808 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
809 				  struct ib_mad_send_wr_private *mad_send_wr)
810 {
811 	int ret = 0;
812 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
813 	struct opa_smp *opa_smp = (struct opa_smp *)smp;
814 	unsigned long flags;
815 	struct ib_mad_local_private *local;
816 	struct ib_mad_private *mad_priv;
817 	struct ib_mad_port_private *port_priv;
818 	struct ib_mad_agent_private *recv_mad_agent = NULL;
819 	struct ib_device *device = mad_agent_priv->agent.device;
820 	u8 port_num;
821 	struct ib_wc mad_wc;
822 	struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
823 	size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
824 	u16 out_mad_pkey_index = 0;
825 	u16 drslid;
826 	bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
827 				    mad_agent_priv->qp_info->port_priv->port_num);
828 
829 	if (rdma_cap_ib_switch(device) &&
830 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
831 		port_num = send_wr->port_num;
832 	else
833 		port_num = mad_agent_priv->agent.port_num;
834 
835 	/*
836 	 * Directed route handling starts if the initial LID routed part of
837 	 * a request or the ending LID routed part of a response is empty.
838 	 * If we are at the start of the LID routed part, don't update the
839 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
840 	 */
841 	if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
842 		u32 opa_drslid;
843 
844 		trace_ib_mad_handle_out_opa_smi(opa_smp);
845 
846 		if ((opa_get_smp_direction(opa_smp)
847 		     ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
848 		     OPA_LID_PERMISSIVE &&
849 		     opa_smi_handle_dr_smp_send(opa_smp,
850 						rdma_cap_ib_switch(device),
851 						port_num) == IB_SMI_DISCARD) {
852 			ret = -EINVAL;
853 			dev_err(&device->dev, "OPA Invalid directed route\n");
854 			goto out;
855 		}
856 		opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
857 		if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
858 		    opa_drslid & 0xffff0000) {
859 			ret = -EINVAL;
860 			dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
861 			       opa_drslid);
862 			goto out;
863 		}
864 		drslid = (u16)(opa_drslid & 0x0000ffff);
865 
866 		/* Check to post send on QP or process locally */
867 		if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
868 		    opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
869 			goto out;
870 	} else {
871 		trace_ib_mad_handle_out_ib_smi(smp);
872 
873 		if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
874 		     IB_LID_PERMISSIVE &&
875 		     smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
876 		     IB_SMI_DISCARD) {
877 			ret = -EINVAL;
878 			dev_err(&device->dev, "Invalid directed route\n");
879 			goto out;
880 		}
881 		drslid = be16_to_cpu(smp->dr_slid);
882 
883 		/* Check to post send on QP or process locally */
884 		if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
885 		    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
886 			goto out;
887 	}
888 
889 	local = kmalloc(sizeof *local, GFP_ATOMIC);
890 	if (!local) {
891 		ret = -ENOMEM;
892 		goto out;
893 	}
894 	local->mad_priv = NULL;
895 	local->recv_mad_agent = NULL;
896 	mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
897 	if (!mad_priv) {
898 		ret = -ENOMEM;
899 		kfree(local);
900 		goto out;
901 	}
902 
903 	build_smp_wc(mad_agent_priv->agent.qp,
904 		     send_wr->wr.wr_cqe, drslid,
905 		     send_wr->pkey_index,
906 		     send_wr->port_num, &mad_wc);
907 
908 	if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
909 		mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
910 					+ mad_send_wr->send_buf.data_len
911 					+ sizeof(struct ib_grh);
912 	}
913 
914 	/* No GRH for DR SMP */
915 	ret = device->ops.process_mad(device, 0, port_num, &mad_wc, NULL,
916 				      (const struct ib_mad *)smp,
917 				      (struct ib_mad *)mad_priv->mad, &mad_size,
918 				      &out_mad_pkey_index);
919 	switch (ret)
920 	{
921 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
922 		if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
923 		    mad_agent_priv->agent.recv_handler) {
924 			local->mad_priv = mad_priv;
925 			local->recv_mad_agent = mad_agent_priv;
926 			/*
927 			 * Reference MAD agent until receive
928 			 * side of local completion handled
929 			 */
930 			atomic_inc(&mad_agent_priv->refcount);
931 		} else
932 			kfree(mad_priv);
933 		break;
934 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
935 		kfree(mad_priv);
936 		break;
937 	case IB_MAD_RESULT_SUCCESS:
938 		/* Treat like an incoming receive MAD */
939 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
940 					    mad_agent_priv->agent.port_num);
941 		if (port_priv) {
942 			memcpy(mad_priv->mad, smp, mad_priv->mad_size);
943 			recv_mad_agent = find_mad_agent(port_priv,
944 						        (const struct ib_mad_hdr *)mad_priv->mad);
945 		}
946 		if (!port_priv || !recv_mad_agent) {
947 			/*
948 			 * No receiving agent so drop packet and
949 			 * generate send completion.
950 			 */
951 			kfree(mad_priv);
952 			break;
953 		}
954 		local->mad_priv = mad_priv;
955 		local->recv_mad_agent = recv_mad_agent;
956 		break;
957 	default:
958 		kfree(mad_priv);
959 		kfree(local);
960 		ret = -EINVAL;
961 		goto out;
962 	}
963 
964 	local->mad_send_wr = mad_send_wr;
965 	if (opa) {
966 		local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
967 		local->return_wc_byte_len = mad_size;
968 	}
969 	/* Reference MAD agent until send side of local completion handled */
970 	atomic_inc(&mad_agent_priv->refcount);
971 	/* Queue local completion to local list */
972 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
973 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
974 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
975 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
976 		   &mad_agent_priv->local_work);
977 	ret = 1;
978 out:
979 	return ret;
980 }
981 
982 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
983 {
984 	int seg_size, pad;
985 
986 	seg_size = mad_size - hdr_len;
987 	if (data_len && seg_size) {
988 		pad = seg_size - data_len % seg_size;
989 		return pad == seg_size ? 0 : pad;
990 	} else
991 		return seg_size;
992 }
993 
994 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
995 {
996 	struct ib_rmpp_segment *s, *t;
997 
998 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
999 		list_del(&s->list);
1000 		kfree(s);
1001 	}
1002 }
1003 
1004 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
1005 				size_t mad_size, gfp_t gfp_mask)
1006 {
1007 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
1008 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
1009 	struct ib_rmpp_segment *seg = NULL;
1010 	int left, seg_size, pad;
1011 
1012 	send_buf->seg_size = mad_size - send_buf->hdr_len;
1013 	send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
1014 	seg_size = send_buf->seg_size;
1015 	pad = send_wr->pad;
1016 
1017 	/* Allocate data segments. */
1018 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
1019 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
1020 		if (!seg) {
1021 			free_send_rmpp_list(send_wr);
1022 			return -ENOMEM;
1023 		}
1024 		seg->num = ++send_buf->seg_count;
1025 		list_add_tail(&seg->list, &send_wr->rmpp_list);
1026 	}
1027 
1028 	/* Zero any padding */
1029 	if (pad)
1030 		memset(seg->data + seg_size - pad, 0, pad);
1031 
1032 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
1033 					  agent.rmpp_version;
1034 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
1035 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
1036 
1037 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
1038 					struct ib_rmpp_segment, list);
1039 	send_wr->last_ack_seg = send_wr->cur_seg;
1040 	return 0;
1041 }
1042 
1043 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
1044 {
1045 	return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
1046 }
1047 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
1048 
1049 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
1050 					    u32 remote_qpn, u16 pkey_index,
1051 					    int rmpp_active,
1052 					    int hdr_len, int data_len,
1053 					    gfp_t gfp_mask,
1054 					    u8 base_version)
1055 {
1056 	struct ib_mad_agent_private *mad_agent_priv;
1057 	struct ib_mad_send_wr_private *mad_send_wr;
1058 	int pad, message_size, ret, size;
1059 	void *buf;
1060 	size_t mad_size;
1061 	bool opa;
1062 
1063 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
1064 				      agent);
1065 
1066 	opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
1067 
1068 	if (opa && base_version == OPA_MGMT_BASE_VERSION)
1069 		mad_size = sizeof(struct opa_mad);
1070 	else
1071 		mad_size = sizeof(struct ib_mad);
1072 
1073 	pad = get_pad_size(hdr_len, data_len, mad_size);
1074 	message_size = hdr_len + data_len + pad;
1075 
1076 	if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1077 		if (!rmpp_active && message_size > mad_size)
1078 			return ERR_PTR(-EINVAL);
1079 	} else
1080 		if (rmpp_active || message_size > mad_size)
1081 			return ERR_PTR(-EINVAL);
1082 
1083 	size = rmpp_active ? hdr_len : mad_size;
1084 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1085 	if (!buf)
1086 		return ERR_PTR(-ENOMEM);
1087 
1088 	mad_send_wr = buf + size;
1089 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1090 	mad_send_wr->send_buf.mad = buf;
1091 	mad_send_wr->send_buf.hdr_len = hdr_len;
1092 	mad_send_wr->send_buf.data_len = data_len;
1093 	mad_send_wr->pad = pad;
1094 
1095 	mad_send_wr->mad_agent_priv = mad_agent_priv;
1096 	mad_send_wr->sg_list[0].length = hdr_len;
1097 	mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1098 
1099 	/* OPA MADs don't have to be the full 2048 bytes */
1100 	if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1101 	    data_len < mad_size - hdr_len)
1102 		mad_send_wr->sg_list[1].length = data_len;
1103 	else
1104 		mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1105 
1106 	mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1107 
1108 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1109 
1110 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1111 	mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1112 	mad_send_wr->send_wr.wr.num_sge = 2;
1113 	mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1114 	mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1115 	mad_send_wr->send_wr.remote_qpn = remote_qpn;
1116 	mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1117 	mad_send_wr->send_wr.pkey_index = pkey_index;
1118 
1119 	if (rmpp_active) {
1120 		ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1121 		if (ret) {
1122 			kfree(buf);
1123 			return ERR_PTR(ret);
1124 		}
1125 	}
1126 
1127 	mad_send_wr->send_buf.mad_agent = mad_agent;
1128 	atomic_inc(&mad_agent_priv->refcount);
1129 	return &mad_send_wr->send_buf;
1130 }
1131 EXPORT_SYMBOL(ib_create_send_mad);
1132 
1133 int ib_get_mad_data_offset(u8 mgmt_class)
1134 {
1135 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1136 		return IB_MGMT_SA_HDR;
1137 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1138 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1139 		 (mgmt_class == IB_MGMT_CLASS_BIS))
1140 		return IB_MGMT_DEVICE_HDR;
1141 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1142 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1143 		return IB_MGMT_VENDOR_HDR;
1144 	else
1145 		return IB_MGMT_MAD_HDR;
1146 }
1147 EXPORT_SYMBOL(ib_get_mad_data_offset);
1148 
1149 int ib_is_mad_class_rmpp(u8 mgmt_class)
1150 {
1151 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1152 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1153 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1154 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
1155 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1156 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1157 		return 1;
1158 	return 0;
1159 }
1160 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1161 
1162 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1163 {
1164 	struct ib_mad_send_wr_private *mad_send_wr;
1165 	struct list_head *list;
1166 
1167 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1168 				   send_buf);
1169 	list = &mad_send_wr->cur_seg->list;
1170 
1171 	if (mad_send_wr->cur_seg->num < seg_num) {
1172 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
1173 			if (mad_send_wr->cur_seg->num == seg_num)
1174 				break;
1175 	} else if (mad_send_wr->cur_seg->num > seg_num) {
1176 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1177 			if (mad_send_wr->cur_seg->num == seg_num)
1178 				break;
1179 	}
1180 	return mad_send_wr->cur_seg->data;
1181 }
1182 EXPORT_SYMBOL(ib_get_rmpp_segment);
1183 
1184 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1185 {
1186 	if (mad_send_wr->send_buf.seg_count)
1187 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1188 					   mad_send_wr->seg_num);
1189 	else
1190 		return mad_send_wr->send_buf.mad +
1191 		       mad_send_wr->send_buf.hdr_len;
1192 }
1193 
1194 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1195 {
1196 	struct ib_mad_agent_private *mad_agent_priv;
1197 	struct ib_mad_send_wr_private *mad_send_wr;
1198 
1199 	mad_agent_priv = container_of(send_buf->mad_agent,
1200 				      struct ib_mad_agent_private, agent);
1201 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1202 				   send_buf);
1203 
1204 	free_send_rmpp_list(mad_send_wr);
1205 	kfree(send_buf->mad);
1206 	deref_mad_agent(mad_agent_priv);
1207 }
1208 EXPORT_SYMBOL(ib_free_send_mad);
1209 
1210 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1211 {
1212 	struct ib_mad_qp_info *qp_info;
1213 	struct list_head *list;
1214 	struct ib_mad_agent *mad_agent;
1215 	struct ib_sge *sge;
1216 	unsigned long flags;
1217 	int ret;
1218 
1219 	/* Set WR ID to find mad_send_wr upon completion */
1220 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1221 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1222 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1223 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1224 
1225 	mad_agent = mad_send_wr->send_buf.mad_agent;
1226 	sge = mad_send_wr->sg_list;
1227 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1228 					mad_send_wr->send_buf.mad,
1229 					sge[0].length,
1230 					DMA_TO_DEVICE);
1231 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1232 		return -ENOMEM;
1233 
1234 	mad_send_wr->header_mapping = sge[0].addr;
1235 
1236 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1237 					ib_get_payload(mad_send_wr),
1238 					sge[1].length,
1239 					DMA_TO_DEVICE);
1240 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1241 		ib_dma_unmap_single(mad_agent->device,
1242 				    mad_send_wr->header_mapping,
1243 				    sge[0].length, DMA_TO_DEVICE);
1244 		return -ENOMEM;
1245 	}
1246 	mad_send_wr->payload_mapping = sge[1].addr;
1247 
1248 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1249 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1250 		trace_ib_mad_ib_send_mad(mad_send_wr, qp_info);
1251 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1252 				   NULL);
1253 		list = &qp_info->send_queue.list;
1254 	} else {
1255 		ret = 0;
1256 		list = &qp_info->overflow_list;
1257 	}
1258 
1259 	if (!ret) {
1260 		qp_info->send_queue.count++;
1261 		list_add_tail(&mad_send_wr->mad_list.list, list);
1262 	}
1263 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1264 	if (ret) {
1265 		ib_dma_unmap_single(mad_agent->device,
1266 				    mad_send_wr->header_mapping,
1267 				    sge[0].length, DMA_TO_DEVICE);
1268 		ib_dma_unmap_single(mad_agent->device,
1269 				    mad_send_wr->payload_mapping,
1270 				    sge[1].length, DMA_TO_DEVICE);
1271 	}
1272 	return ret;
1273 }
1274 
1275 /*
1276  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1277  *  with the registered client
1278  */
1279 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1280 		     struct ib_mad_send_buf **bad_send_buf)
1281 {
1282 	struct ib_mad_agent_private *mad_agent_priv;
1283 	struct ib_mad_send_buf *next_send_buf;
1284 	struct ib_mad_send_wr_private *mad_send_wr;
1285 	unsigned long flags;
1286 	int ret = -EINVAL;
1287 
1288 	/* Walk list of send WRs and post each on send list */
1289 	for (; send_buf; send_buf = next_send_buf) {
1290 		mad_send_wr = container_of(send_buf,
1291 					   struct ib_mad_send_wr_private,
1292 					   send_buf);
1293 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1294 
1295 		ret = ib_mad_enforce_security(mad_agent_priv,
1296 					      mad_send_wr->send_wr.pkey_index);
1297 		if (ret)
1298 			goto error;
1299 
1300 		if (!send_buf->mad_agent->send_handler ||
1301 		    (send_buf->timeout_ms &&
1302 		     !send_buf->mad_agent->recv_handler)) {
1303 			ret = -EINVAL;
1304 			goto error;
1305 		}
1306 
1307 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1308 			if (mad_agent_priv->agent.rmpp_version) {
1309 				ret = -EINVAL;
1310 				goto error;
1311 			}
1312 		}
1313 
1314 		/*
1315 		 * Save pointer to next work request to post in case the
1316 		 * current one completes, and the user modifies the work
1317 		 * request associated with the completion
1318 		 */
1319 		next_send_buf = send_buf->next;
1320 		mad_send_wr->send_wr.ah = send_buf->ah;
1321 
1322 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1323 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1324 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1325 						     mad_send_wr);
1326 			if (ret < 0)		/* error */
1327 				goto error;
1328 			else if (ret == 1)	/* locally consumed */
1329 				continue;
1330 		}
1331 
1332 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1333 		/* Timeout will be updated after send completes */
1334 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1335 		mad_send_wr->max_retries = send_buf->retries;
1336 		mad_send_wr->retries_left = send_buf->retries;
1337 		send_buf->retries = 0;
1338 		/* Reference for work request to QP + response */
1339 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1340 		mad_send_wr->status = IB_WC_SUCCESS;
1341 
1342 		/* Reference MAD agent until send completes */
1343 		atomic_inc(&mad_agent_priv->refcount);
1344 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1345 		list_add_tail(&mad_send_wr->agent_list,
1346 			      &mad_agent_priv->send_list);
1347 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1348 
1349 		if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1350 			ret = ib_send_rmpp_mad(mad_send_wr);
1351 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1352 				ret = ib_send_mad(mad_send_wr);
1353 		} else
1354 			ret = ib_send_mad(mad_send_wr);
1355 		if (ret < 0) {
1356 			/* Fail send request */
1357 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1358 			list_del(&mad_send_wr->agent_list);
1359 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1360 			atomic_dec(&mad_agent_priv->refcount);
1361 			goto error;
1362 		}
1363 	}
1364 	return 0;
1365 error:
1366 	if (bad_send_buf)
1367 		*bad_send_buf = send_buf;
1368 	return ret;
1369 }
1370 EXPORT_SYMBOL(ib_post_send_mad);
1371 
1372 /*
1373  * ib_free_recv_mad - Returns data buffers used to receive
1374  *  a MAD to the access layer
1375  */
1376 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1377 {
1378 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1379 	struct ib_mad_private_header *mad_priv_hdr;
1380 	struct ib_mad_private *priv;
1381 	struct list_head free_list;
1382 
1383 	INIT_LIST_HEAD(&free_list);
1384 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1385 
1386 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1387 					&free_list, list) {
1388 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1389 					   recv_buf);
1390 		mad_priv_hdr = container_of(mad_recv_wc,
1391 					    struct ib_mad_private_header,
1392 					    recv_wc);
1393 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1394 				    header);
1395 		kfree(priv);
1396 	}
1397 }
1398 EXPORT_SYMBOL(ib_free_recv_mad);
1399 
1400 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1401 			 struct ib_mad_reg_req *mad_reg_req)
1402 {
1403 	int i;
1404 
1405 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1406 		if ((*method)->agent[i]) {
1407 			pr_err("Method %d already in use\n", i);
1408 			return -EINVAL;
1409 		}
1410 	}
1411 	return 0;
1412 }
1413 
1414 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1415 {
1416 	/* Allocate management method table */
1417 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1418 	return (*method) ? 0 : (-ENOMEM);
1419 }
1420 
1421 /*
1422  * Check to see if there are any methods still in use
1423  */
1424 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1425 {
1426 	int i;
1427 
1428 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1429 		if (method->agent[i])
1430 			return 1;
1431 	return 0;
1432 }
1433 
1434 /*
1435  * Check to see if there are any method tables for this class still in use
1436  */
1437 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1438 {
1439 	int i;
1440 
1441 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1442 		if (class->method_table[i])
1443 			return 1;
1444 	return 0;
1445 }
1446 
1447 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1448 {
1449 	int i;
1450 
1451 	for (i = 0; i < MAX_MGMT_OUI; i++)
1452 		if (vendor_class->method_table[i])
1453 			return 1;
1454 	return 0;
1455 }
1456 
1457 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1458 			   const char *oui)
1459 {
1460 	int i;
1461 
1462 	for (i = 0; i < MAX_MGMT_OUI; i++)
1463 		/* Is there matching OUI for this vendor class ? */
1464 		if (!memcmp(vendor_class->oui[i], oui, 3))
1465 			return i;
1466 
1467 	return -1;
1468 }
1469 
1470 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1471 {
1472 	int i;
1473 
1474 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1475 		if (vendor->vendor_class[i])
1476 			return 1;
1477 
1478 	return 0;
1479 }
1480 
1481 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1482 				     struct ib_mad_agent_private *agent)
1483 {
1484 	int i;
1485 
1486 	/* Remove any methods for this mad agent */
1487 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1488 		if (method->agent[i] == agent) {
1489 			method->agent[i] = NULL;
1490 		}
1491 	}
1492 }
1493 
1494 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1495 			      struct ib_mad_agent_private *agent_priv,
1496 			      u8 mgmt_class)
1497 {
1498 	struct ib_mad_port_private *port_priv;
1499 	struct ib_mad_mgmt_class_table **class;
1500 	struct ib_mad_mgmt_method_table **method;
1501 	int i, ret;
1502 
1503 	port_priv = agent_priv->qp_info->port_priv;
1504 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1505 	if (!*class) {
1506 		/* Allocate management class table for "new" class version */
1507 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1508 		if (!*class) {
1509 			ret = -ENOMEM;
1510 			goto error1;
1511 		}
1512 
1513 		/* Allocate method table for this management class */
1514 		method = &(*class)->method_table[mgmt_class];
1515 		if ((ret = allocate_method_table(method)))
1516 			goto error2;
1517 	} else {
1518 		method = &(*class)->method_table[mgmt_class];
1519 		if (!*method) {
1520 			/* Allocate method table for this management class */
1521 			if ((ret = allocate_method_table(method)))
1522 				goto error1;
1523 		}
1524 	}
1525 
1526 	/* Now, make sure methods are not already in use */
1527 	if (method_in_use(method, mad_reg_req))
1528 		goto error3;
1529 
1530 	/* Finally, add in methods being registered */
1531 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1532 		(*method)->agent[i] = agent_priv;
1533 
1534 	return 0;
1535 
1536 error3:
1537 	/* Remove any methods for this mad agent */
1538 	remove_methods_mad_agent(*method, agent_priv);
1539 	/* Now, check to see if there are any methods in use */
1540 	if (!check_method_table(*method)) {
1541 		/* If not, release management method table */
1542 		kfree(*method);
1543 		*method = NULL;
1544 	}
1545 	ret = -EINVAL;
1546 	goto error1;
1547 error2:
1548 	kfree(*class);
1549 	*class = NULL;
1550 error1:
1551 	return ret;
1552 }
1553 
1554 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1555 			   struct ib_mad_agent_private *agent_priv)
1556 {
1557 	struct ib_mad_port_private *port_priv;
1558 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1559 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1560 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1561 	struct ib_mad_mgmt_method_table **method;
1562 	int i, ret = -ENOMEM;
1563 	u8 vclass;
1564 
1565 	/* "New" vendor (with OUI) class */
1566 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1567 	port_priv = agent_priv->qp_info->port_priv;
1568 	vendor_table = &port_priv->version[
1569 				mad_reg_req->mgmt_class_version].vendor;
1570 	if (!*vendor_table) {
1571 		/* Allocate mgmt vendor class table for "new" class version */
1572 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1573 		if (!vendor)
1574 			goto error1;
1575 
1576 		*vendor_table = vendor;
1577 	}
1578 	if (!(*vendor_table)->vendor_class[vclass]) {
1579 		/* Allocate table for this management vendor class */
1580 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1581 		if (!vendor_class)
1582 			goto error2;
1583 
1584 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1585 	}
1586 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1587 		/* Is there matching OUI for this vendor class ? */
1588 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1589 			    mad_reg_req->oui, 3)) {
1590 			method = &(*vendor_table)->vendor_class[
1591 						vclass]->method_table[i];
1592 			if (!*method)
1593 				goto error3;
1594 			goto check_in_use;
1595 		}
1596 	}
1597 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1598 		/* OUI slot available ? */
1599 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1600 				vclass]->oui[i])) {
1601 			method = &(*vendor_table)->vendor_class[
1602 				vclass]->method_table[i];
1603 			/* Allocate method table for this OUI */
1604 			if (!*method) {
1605 				ret = allocate_method_table(method);
1606 				if (ret)
1607 					goto error3;
1608 			}
1609 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1610 			       mad_reg_req->oui, 3);
1611 			goto check_in_use;
1612 		}
1613 	}
1614 	dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1615 	goto error3;
1616 
1617 check_in_use:
1618 	/* Now, make sure methods are not already in use */
1619 	if (method_in_use(method, mad_reg_req))
1620 		goto error4;
1621 
1622 	/* Finally, add in methods being registered */
1623 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1624 		(*method)->agent[i] = agent_priv;
1625 
1626 	return 0;
1627 
1628 error4:
1629 	/* Remove any methods for this mad agent */
1630 	remove_methods_mad_agent(*method, agent_priv);
1631 	/* Now, check to see if there are any methods in use */
1632 	if (!check_method_table(*method)) {
1633 		/* If not, release management method table */
1634 		kfree(*method);
1635 		*method = NULL;
1636 	}
1637 	ret = -EINVAL;
1638 error3:
1639 	if (vendor_class) {
1640 		(*vendor_table)->vendor_class[vclass] = NULL;
1641 		kfree(vendor_class);
1642 	}
1643 error2:
1644 	if (vendor) {
1645 		*vendor_table = NULL;
1646 		kfree(vendor);
1647 	}
1648 error1:
1649 	return ret;
1650 }
1651 
1652 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1653 {
1654 	struct ib_mad_port_private *port_priv;
1655 	struct ib_mad_mgmt_class_table *class;
1656 	struct ib_mad_mgmt_method_table *method;
1657 	struct ib_mad_mgmt_vendor_class_table *vendor;
1658 	struct ib_mad_mgmt_vendor_class *vendor_class;
1659 	int index;
1660 	u8 mgmt_class;
1661 
1662 	/*
1663 	 * Was MAD registration request supplied
1664 	 * with original registration ?
1665 	 */
1666 	if (!agent_priv->reg_req) {
1667 		goto out;
1668 	}
1669 
1670 	port_priv = agent_priv->qp_info->port_priv;
1671 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1672 	class = port_priv->version[
1673 			agent_priv->reg_req->mgmt_class_version].class;
1674 	if (!class)
1675 		goto vendor_check;
1676 
1677 	method = class->method_table[mgmt_class];
1678 	if (method) {
1679 		/* Remove any methods for this mad agent */
1680 		remove_methods_mad_agent(method, agent_priv);
1681 		/* Now, check to see if there are any methods still in use */
1682 		if (!check_method_table(method)) {
1683 			/* If not, release management method table */
1684 			kfree(method);
1685 			class->method_table[mgmt_class] = NULL;
1686 			/* Any management classes left ? */
1687 			if (!check_class_table(class)) {
1688 				/* If not, release management class table */
1689 				kfree(class);
1690 				port_priv->version[
1691 					agent_priv->reg_req->
1692 					mgmt_class_version].class = NULL;
1693 			}
1694 		}
1695 	}
1696 
1697 vendor_check:
1698 	if (!is_vendor_class(mgmt_class))
1699 		goto out;
1700 
1701 	/* normalize mgmt_class to vendor range 2 */
1702 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1703 	vendor = port_priv->version[
1704 			agent_priv->reg_req->mgmt_class_version].vendor;
1705 
1706 	if (!vendor)
1707 		goto out;
1708 
1709 	vendor_class = vendor->vendor_class[mgmt_class];
1710 	if (vendor_class) {
1711 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1712 		if (index < 0)
1713 			goto out;
1714 		method = vendor_class->method_table[index];
1715 		if (method) {
1716 			/* Remove any methods for this mad agent */
1717 			remove_methods_mad_agent(method, agent_priv);
1718 			/*
1719 			 * Now, check to see if there are
1720 			 * any methods still in use
1721 			 */
1722 			if (!check_method_table(method)) {
1723 				/* If not, release management method table */
1724 				kfree(method);
1725 				vendor_class->method_table[index] = NULL;
1726 				memset(vendor_class->oui[index], 0, 3);
1727 				/* Any OUIs left ? */
1728 				if (!check_vendor_class(vendor_class)) {
1729 					/* If not, release vendor class table */
1730 					kfree(vendor_class);
1731 					vendor->vendor_class[mgmt_class] = NULL;
1732 					/* Any other vendor classes left ? */
1733 					if (!check_vendor_table(vendor)) {
1734 						kfree(vendor);
1735 						port_priv->version[
1736 							agent_priv->reg_req->
1737 							mgmt_class_version].
1738 							vendor = NULL;
1739 					}
1740 				}
1741 			}
1742 		}
1743 	}
1744 
1745 out:
1746 	return;
1747 }
1748 
1749 static struct ib_mad_agent_private *
1750 find_mad_agent(struct ib_mad_port_private *port_priv,
1751 	       const struct ib_mad_hdr *mad_hdr)
1752 {
1753 	struct ib_mad_agent_private *mad_agent = NULL;
1754 	unsigned long flags;
1755 
1756 	if (ib_response_mad(mad_hdr)) {
1757 		u32 hi_tid;
1758 
1759 		/*
1760 		 * Routing is based on high 32 bits of transaction ID
1761 		 * of MAD.
1762 		 */
1763 		hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1764 		rcu_read_lock();
1765 		mad_agent = xa_load(&ib_mad_clients, hi_tid);
1766 		if (mad_agent && !atomic_inc_not_zero(&mad_agent->refcount))
1767 			mad_agent = NULL;
1768 		rcu_read_unlock();
1769 	} else {
1770 		struct ib_mad_mgmt_class_table *class;
1771 		struct ib_mad_mgmt_method_table *method;
1772 		struct ib_mad_mgmt_vendor_class_table *vendor;
1773 		struct ib_mad_mgmt_vendor_class *vendor_class;
1774 		const struct ib_vendor_mad *vendor_mad;
1775 		int index;
1776 
1777 		spin_lock_irqsave(&port_priv->reg_lock, flags);
1778 		/*
1779 		 * Routing is based on version, class, and method
1780 		 * For "newer" vendor MADs, also based on OUI
1781 		 */
1782 		if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1783 			goto out;
1784 		if (!is_vendor_class(mad_hdr->mgmt_class)) {
1785 			class = port_priv->version[
1786 					mad_hdr->class_version].class;
1787 			if (!class)
1788 				goto out;
1789 			if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1790 			    ARRAY_SIZE(class->method_table))
1791 				goto out;
1792 			method = class->method_table[convert_mgmt_class(
1793 							mad_hdr->mgmt_class)];
1794 			if (method)
1795 				mad_agent = method->agent[mad_hdr->method &
1796 							  ~IB_MGMT_METHOD_RESP];
1797 		} else {
1798 			vendor = port_priv->version[
1799 					mad_hdr->class_version].vendor;
1800 			if (!vendor)
1801 				goto out;
1802 			vendor_class = vendor->vendor_class[vendor_class_index(
1803 						mad_hdr->mgmt_class)];
1804 			if (!vendor_class)
1805 				goto out;
1806 			/* Find matching OUI */
1807 			vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1808 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1809 			if (index == -1)
1810 				goto out;
1811 			method = vendor_class->method_table[index];
1812 			if (method) {
1813 				mad_agent = method->agent[mad_hdr->method &
1814 							  ~IB_MGMT_METHOD_RESP];
1815 			}
1816 		}
1817 		if (mad_agent)
1818 			atomic_inc(&mad_agent->refcount);
1819 out:
1820 		spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1821 	}
1822 
1823 	if (mad_agent && !mad_agent->agent.recv_handler) {
1824 		dev_notice(&port_priv->device->dev,
1825 			   "No receive handler for client %p on port %d\n",
1826 			   &mad_agent->agent, port_priv->port_num);
1827 		deref_mad_agent(mad_agent);
1828 		mad_agent = NULL;
1829 	}
1830 
1831 	return mad_agent;
1832 }
1833 
1834 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1835 			const struct ib_mad_qp_info *qp_info,
1836 			bool opa)
1837 {
1838 	int valid = 0;
1839 	u32 qp_num = qp_info->qp->qp_num;
1840 
1841 	/* Make sure MAD base version is understood */
1842 	if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1843 	    (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1844 		pr_err("MAD received with unsupported base version %d %s\n",
1845 		       mad_hdr->base_version, opa ? "(opa)" : "");
1846 		goto out;
1847 	}
1848 
1849 	/* Filter SMI packets sent to other than QP0 */
1850 	if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1851 	    (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1852 		if (qp_num == 0)
1853 			valid = 1;
1854 	} else {
1855 		/* CM attributes other than ClassPortInfo only use Send method */
1856 		if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1857 		    (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1858 		    (mad_hdr->method != IB_MGMT_METHOD_SEND))
1859 			goto out;
1860 		/* Filter GSI packets sent to QP0 */
1861 		if (qp_num != 0)
1862 			valid = 1;
1863 	}
1864 
1865 out:
1866 	return valid;
1867 }
1868 
1869 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1870 			    const struct ib_mad_hdr *mad_hdr)
1871 {
1872 	struct ib_rmpp_mad *rmpp_mad;
1873 
1874 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1875 	return !mad_agent_priv->agent.rmpp_version ||
1876 		!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1877 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1878 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1879 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1880 }
1881 
1882 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1883 				     const struct ib_mad_recv_wc *rwc)
1884 {
1885 	return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1886 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1887 }
1888 
1889 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1890 				   const struct ib_mad_send_wr_private *wr,
1891 				   const struct ib_mad_recv_wc *rwc )
1892 {
1893 	struct rdma_ah_attr attr;
1894 	u8 send_resp, rcv_resp;
1895 	union ib_gid sgid;
1896 	struct ib_device *device = mad_agent_priv->agent.device;
1897 	u8 port_num = mad_agent_priv->agent.port_num;
1898 	u8 lmc;
1899 	bool has_grh;
1900 
1901 	send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1902 	rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1903 
1904 	if (send_resp == rcv_resp)
1905 		/* both requests, or both responses. GIDs different */
1906 		return 0;
1907 
1908 	if (rdma_query_ah(wr->send_buf.ah, &attr))
1909 		/* Assume not equal, to avoid false positives. */
1910 		return 0;
1911 
1912 	has_grh = !!(rdma_ah_get_ah_flags(&attr) & IB_AH_GRH);
1913 	if (has_grh != !!(rwc->wc->wc_flags & IB_WC_GRH))
1914 		/* one has GID, other does not.  Assume different */
1915 		return 0;
1916 
1917 	if (!send_resp && rcv_resp) {
1918 		/* is request/response. */
1919 		if (!has_grh) {
1920 			if (ib_get_cached_lmc(device, port_num, &lmc))
1921 				return 0;
1922 			return (!lmc || !((rdma_ah_get_path_bits(&attr) ^
1923 					   rwc->wc->dlid_path_bits) &
1924 					  ((1 << lmc) - 1)));
1925 		} else {
1926 			const struct ib_global_route *grh =
1927 					rdma_ah_read_grh(&attr);
1928 
1929 			if (rdma_query_gid(device, port_num,
1930 					   grh->sgid_index, &sgid))
1931 				return 0;
1932 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1933 				       16);
1934 		}
1935 	}
1936 
1937 	if (!has_grh)
1938 		return rdma_ah_get_dlid(&attr) == rwc->wc->slid;
1939 	else
1940 		return !memcmp(rdma_ah_read_grh(&attr)->dgid.raw,
1941 			       rwc->recv_buf.grh->sgid.raw,
1942 			       16);
1943 }
1944 
1945 static inline int is_direct(u8 class)
1946 {
1947 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1948 }
1949 
1950 struct ib_mad_send_wr_private*
1951 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1952 		 const struct ib_mad_recv_wc *wc)
1953 {
1954 	struct ib_mad_send_wr_private *wr;
1955 	const struct ib_mad_hdr *mad_hdr;
1956 
1957 	mad_hdr = &wc->recv_buf.mad->mad_hdr;
1958 
1959 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1960 		if ((wr->tid == mad_hdr->tid) &&
1961 		    rcv_has_same_class(wr, wc) &&
1962 		    /*
1963 		     * Don't check GID for direct routed MADs.
1964 		     * These might have permissive LIDs.
1965 		     */
1966 		    (is_direct(mad_hdr->mgmt_class) ||
1967 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1968 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1969 	}
1970 
1971 	/*
1972 	 * It's possible to receive the response before we've
1973 	 * been notified that the send has completed
1974 	 */
1975 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1976 		if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1977 		    wr->tid == mad_hdr->tid &&
1978 		    wr->timeout &&
1979 		    rcv_has_same_class(wr, wc) &&
1980 		    /*
1981 		     * Don't check GID for direct routed MADs.
1982 		     * These might have permissive LIDs.
1983 		     */
1984 		    (is_direct(mad_hdr->mgmt_class) ||
1985 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1986 			/* Verify request has not been canceled */
1987 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1988 	}
1989 	return NULL;
1990 }
1991 
1992 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1993 {
1994 	mad_send_wr->timeout = 0;
1995 	if (mad_send_wr->refcount == 1)
1996 		list_move_tail(&mad_send_wr->agent_list,
1997 			      &mad_send_wr->mad_agent_priv->done_list);
1998 }
1999 
2000 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
2001 				 struct ib_mad_recv_wc *mad_recv_wc)
2002 {
2003 	struct ib_mad_send_wr_private *mad_send_wr;
2004 	struct ib_mad_send_wc mad_send_wc;
2005 	unsigned long flags;
2006 	int ret;
2007 
2008 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
2009 	ret = ib_mad_enforce_security(mad_agent_priv,
2010 				      mad_recv_wc->wc->pkey_index);
2011 	if (ret) {
2012 		ib_free_recv_mad(mad_recv_wc);
2013 		deref_mad_agent(mad_agent_priv);
2014 		return;
2015 	}
2016 
2017 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
2018 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2019 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
2020 						      mad_recv_wc);
2021 		if (!mad_recv_wc) {
2022 			deref_mad_agent(mad_agent_priv);
2023 			return;
2024 		}
2025 	}
2026 
2027 	/* Complete corresponding request */
2028 	if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
2029 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2030 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
2031 		if (!mad_send_wr) {
2032 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2033 			if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
2034 			   && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
2035 			   && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
2036 					& IB_MGMT_RMPP_FLAG_ACTIVE)) {
2037 				/* user rmpp is in effect
2038 				 * and this is an active RMPP MAD
2039 				 */
2040 				mad_agent_priv->agent.recv_handler(
2041 						&mad_agent_priv->agent, NULL,
2042 						mad_recv_wc);
2043 				atomic_dec(&mad_agent_priv->refcount);
2044 			} else {
2045 				/* not user rmpp, revert to normal behavior and
2046 				 * drop the mad */
2047 				ib_free_recv_mad(mad_recv_wc);
2048 				deref_mad_agent(mad_agent_priv);
2049 				return;
2050 			}
2051 		} else {
2052 			ib_mark_mad_done(mad_send_wr);
2053 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2054 
2055 			/* Defined behavior is to complete response before request */
2056 			mad_agent_priv->agent.recv_handler(
2057 					&mad_agent_priv->agent,
2058 					&mad_send_wr->send_buf,
2059 					mad_recv_wc);
2060 			atomic_dec(&mad_agent_priv->refcount);
2061 
2062 			mad_send_wc.status = IB_WC_SUCCESS;
2063 			mad_send_wc.vendor_err = 0;
2064 			mad_send_wc.send_buf = &mad_send_wr->send_buf;
2065 			ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2066 		}
2067 	} else {
2068 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2069 						   mad_recv_wc);
2070 		deref_mad_agent(mad_agent_priv);
2071 	}
2072 
2073 	return;
2074 }
2075 
2076 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2077 				     const struct ib_mad_qp_info *qp_info,
2078 				     const struct ib_wc *wc,
2079 				     int port_num,
2080 				     struct ib_mad_private *recv,
2081 				     struct ib_mad_private *response)
2082 {
2083 	enum smi_forward_action retsmi;
2084 	struct ib_smp *smp = (struct ib_smp *)recv->mad;
2085 
2086 	trace_ib_mad_handle_ib_smi(smp);
2087 
2088 	if (smi_handle_dr_smp_recv(smp,
2089 				   rdma_cap_ib_switch(port_priv->device),
2090 				   port_num,
2091 				   port_priv->device->phys_port_cnt) ==
2092 				   IB_SMI_DISCARD)
2093 		return IB_SMI_DISCARD;
2094 
2095 	retsmi = smi_check_forward_dr_smp(smp);
2096 	if (retsmi == IB_SMI_LOCAL)
2097 		return IB_SMI_HANDLE;
2098 
2099 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2100 		if (smi_handle_dr_smp_send(smp,
2101 					   rdma_cap_ib_switch(port_priv->device),
2102 					   port_num) == IB_SMI_DISCARD)
2103 			return IB_SMI_DISCARD;
2104 
2105 		if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2106 			return IB_SMI_DISCARD;
2107 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2108 		/* forward case for switches */
2109 		memcpy(response, recv, mad_priv_size(response));
2110 		response->header.recv_wc.wc = &response->header.wc;
2111 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2112 		response->header.recv_wc.recv_buf.grh = &response->grh;
2113 
2114 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2115 				    &response->grh, wc,
2116 				    port_priv->device,
2117 				    smi_get_fwd_port(smp),
2118 				    qp_info->qp->qp_num,
2119 				    response->mad_size,
2120 				    false);
2121 
2122 		return IB_SMI_DISCARD;
2123 	}
2124 	return IB_SMI_HANDLE;
2125 }
2126 
2127 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2128 				    struct ib_mad_private *response,
2129 				    size_t *resp_len, bool opa)
2130 {
2131 	const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2132 	struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2133 
2134 	if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2135 	    recv_hdr->method == IB_MGMT_METHOD_SET) {
2136 		memcpy(response, recv, mad_priv_size(response));
2137 		response->header.recv_wc.wc = &response->header.wc;
2138 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2139 		response->header.recv_wc.recv_buf.grh = &response->grh;
2140 		resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2141 		resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2142 		if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2143 			resp_hdr->status |= IB_SMP_DIRECTION;
2144 
2145 		if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2146 			if (recv_hdr->mgmt_class ==
2147 			    IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2148 			    recv_hdr->mgmt_class ==
2149 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2150 				*resp_len = opa_get_smp_header_size(
2151 							(struct opa_smp *)recv->mad);
2152 			else
2153 				*resp_len = sizeof(struct ib_mad_hdr);
2154 		}
2155 
2156 		return true;
2157 	} else {
2158 		return false;
2159 	}
2160 }
2161 
2162 static enum smi_action
2163 handle_opa_smi(struct ib_mad_port_private *port_priv,
2164 	       struct ib_mad_qp_info *qp_info,
2165 	       struct ib_wc *wc,
2166 	       int port_num,
2167 	       struct ib_mad_private *recv,
2168 	       struct ib_mad_private *response)
2169 {
2170 	enum smi_forward_action retsmi;
2171 	struct opa_smp *smp = (struct opa_smp *)recv->mad;
2172 
2173 	trace_ib_mad_handle_opa_smi(smp);
2174 
2175 	if (opa_smi_handle_dr_smp_recv(smp,
2176 				   rdma_cap_ib_switch(port_priv->device),
2177 				   port_num,
2178 				   port_priv->device->phys_port_cnt) ==
2179 				   IB_SMI_DISCARD)
2180 		return IB_SMI_DISCARD;
2181 
2182 	retsmi = opa_smi_check_forward_dr_smp(smp);
2183 	if (retsmi == IB_SMI_LOCAL)
2184 		return IB_SMI_HANDLE;
2185 
2186 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2187 		if (opa_smi_handle_dr_smp_send(smp,
2188 					   rdma_cap_ib_switch(port_priv->device),
2189 					   port_num) == IB_SMI_DISCARD)
2190 			return IB_SMI_DISCARD;
2191 
2192 		if (opa_smi_check_local_smp(smp, port_priv->device) ==
2193 		    IB_SMI_DISCARD)
2194 			return IB_SMI_DISCARD;
2195 
2196 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2197 		/* forward case for switches */
2198 		memcpy(response, recv, mad_priv_size(response));
2199 		response->header.recv_wc.wc = &response->header.wc;
2200 		response->header.recv_wc.recv_buf.opa_mad =
2201 				(struct opa_mad *)response->mad;
2202 		response->header.recv_wc.recv_buf.grh = &response->grh;
2203 
2204 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2205 				    &response->grh, wc,
2206 				    port_priv->device,
2207 				    opa_smi_get_fwd_port(smp),
2208 				    qp_info->qp->qp_num,
2209 				    recv->header.wc.byte_len,
2210 				    true);
2211 
2212 		return IB_SMI_DISCARD;
2213 	}
2214 
2215 	return IB_SMI_HANDLE;
2216 }
2217 
2218 static enum smi_action
2219 handle_smi(struct ib_mad_port_private *port_priv,
2220 	   struct ib_mad_qp_info *qp_info,
2221 	   struct ib_wc *wc,
2222 	   int port_num,
2223 	   struct ib_mad_private *recv,
2224 	   struct ib_mad_private *response,
2225 	   bool opa)
2226 {
2227 	struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2228 
2229 	if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2230 	    mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2231 		return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2232 				      response);
2233 
2234 	return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2235 }
2236 
2237 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2238 {
2239 	struct ib_mad_port_private *port_priv = cq->cq_context;
2240 	struct ib_mad_list_head *mad_list =
2241 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2242 	struct ib_mad_qp_info *qp_info;
2243 	struct ib_mad_private_header *mad_priv_hdr;
2244 	struct ib_mad_private *recv, *response = NULL;
2245 	struct ib_mad_agent_private *mad_agent;
2246 	int port_num;
2247 	int ret = IB_MAD_RESULT_SUCCESS;
2248 	size_t mad_size;
2249 	u16 resp_mad_pkey_index = 0;
2250 	bool opa;
2251 
2252 	if (list_empty_careful(&port_priv->port_list))
2253 		return;
2254 
2255 	if (wc->status != IB_WC_SUCCESS) {
2256 		/*
2257 		 * Receive errors indicate that the QP has entered the error
2258 		 * state - error handling/shutdown code will cleanup
2259 		 */
2260 		return;
2261 	}
2262 
2263 	qp_info = mad_list->mad_queue->qp_info;
2264 	dequeue_mad(mad_list);
2265 
2266 	opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2267 			       qp_info->port_priv->port_num);
2268 
2269 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2270 				    mad_list);
2271 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2272 	ib_dma_unmap_single(port_priv->device,
2273 			    recv->header.mapping,
2274 			    mad_priv_dma_size(recv),
2275 			    DMA_FROM_DEVICE);
2276 
2277 	/* Setup MAD receive work completion from "normal" work completion */
2278 	recv->header.wc = *wc;
2279 	recv->header.recv_wc.wc = &recv->header.wc;
2280 
2281 	if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2282 		recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2283 		recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2284 	} else {
2285 		recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2286 		recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2287 	}
2288 
2289 	recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2290 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
2291 
2292 	if (atomic_read(&qp_info->snoop_count))
2293 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2294 
2295 	/* Validate MAD */
2296 	if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2297 		goto out;
2298 
2299 	trace_ib_mad_recv_done_handler(qp_info, wc,
2300 				       (struct ib_mad_hdr *)recv->mad);
2301 
2302 	mad_size = recv->mad_size;
2303 	response = alloc_mad_private(mad_size, GFP_KERNEL);
2304 	if (!response)
2305 		goto out;
2306 
2307 	if (rdma_cap_ib_switch(port_priv->device))
2308 		port_num = wc->port_num;
2309 	else
2310 		port_num = port_priv->port_num;
2311 
2312 	if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2313 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2314 		if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2315 			       response, opa)
2316 		    == IB_SMI_DISCARD)
2317 			goto out;
2318 	}
2319 
2320 	/* Give driver "right of first refusal" on incoming MAD */
2321 	if (port_priv->device->ops.process_mad) {
2322 		ret = port_priv->device->ops.process_mad(
2323 			port_priv->device, 0, port_priv->port_num, wc,
2324 			&recv->grh, (const struct ib_mad *)recv->mad,
2325 			(struct ib_mad *)response->mad, &mad_size,
2326 			&resp_mad_pkey_index);
2327 
2328 		if (opa)
2329 			wc->pkey_index = resp_mad_pkey_index;
2330 
2331 		if (ret & IB_MAD_RESULT_SUCCESS) {
2332 			if (ret & IB_MAD_RESULT_CONSUMED)
2333 				goto out;
2334 			if (ret & IB_MAD_RESULT_REPLY) {
2335 				agent_send_response((const struct ib_mad_hdr *)response->mad,
2336 						    &recv->grh, wc,
2337 						    port_priv->device,
2338 						    port_num,
2339 						    qp_info->qp->qp_num,
2340 						    mad_size, opa);
2341 				goto out;
2342 			}
2343 		}
2344 	}
2345 
2346 	mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2347 	if (mad_agent) {
2348 		trace_ib_mad_recv_done_agent(mad_agent);
2349 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2350 		/*
2351 		 * recv is freed up in error cases in ib_mad_complete_recv
2352 		 * or via recv_handler in ib_mad_complete_recv()
2353 		 */
2354 		recv = NULL;
2355 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2356 		   generate_unmatched_resp(recv, response, &mad_size, opa)) {
2357 		agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2358 				    port_priv->device, port_num,
2359 				    qp_info->qp->qp_num, mad_size, opa);
2360 	}
2361 
2362 out:
2363 	/* Post another receive request for this QP */
2364 	if (response) {
2365 		ib_mad_post_receive_mads(qp_info, response);
2366 		kfree(recv);
2367 	} else
2368 		ib_mad_post_receive_mads(qp_info, recv);
2369 }
2370 
2371 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2372 {
2373 	struct ib_mad_send_wr_private *mad_send_wr;
2374 	unsigned long delay;
2375 
2376 	if (list_empty(&mad_agent_priv->wait_list)) {
2377 		cancel_delayed_work(&mad_agent_priv->timed_work);
2378 	} else {
2379 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2380 					 struct ib_mad_send_wr_private,
2381 					 agent_list);
2382 
2383 		if (time_after(mad_agent_priv->timeout,
2384 			       mad_send_wr->timeout)) {
2385 			mad_agent_priv->timeout = mad_send_wr->timeout;
2386 			delay = mad_send_wr->timeout - jiffies;
2387 			if ((long)delay <= 0)
2388 				delay = 1;
2389 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2390 					 &mad_agent_priv->timed_work, delay);
2391 		}
2392 	}
2393 }
2394 
2395 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2396 {
2397 	struct ib_mad_agent_private *mad_agent_priv;
2398 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2399 	struct list_head *list_item;
2400 	unsigned long delay;
2401 
2402 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2403 	list_del(&mad_send_wr->agent_list);
2404 
2405 	delay = mad_send_wr->timeout;
2406 	mad_send_wr->timeout += jiffies;
2407 
2408 	if (delay) {
2409 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2410 			temp_mad_send_wr = list_entry(list_item,
2411 						struct ib_mad_send_wr_private,
2412 						agent_list);
2413 			if (time_after(mad_send_wr->timeout,
2414 				       temp_mad_send_wr->timeout))
2415 				break;
2416 		}
2417 	}
2418 	else
2419 		list_item = &mad_agent_priv->wait_list;
2420 	list_add(&mad_send_wr->agent_list, list_item);
2421 
2422 	/* Reschedule a work item if we have a shorter timeout */
2423 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2424 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2425 				 &mad_agent_priv->timed_work, delay);
2426 }
2427 
2428 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2429 			  unsigned long timeout_ms)
2430 {
2431 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2432 	wait_for_response(mad_send_wr);
2433 }
2434 
2435 /*
2436  * Process a send work completion
2437  */
2438 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2439 			     struct ib_mad_send_wc *mad_send_wc)
2440 {
2441 	struct ib_mad_agent_private	*mad_agent_priv;
2442 	unsigned long			flags;
2443 	int				ret;
2444 
2445 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2446 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2447 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2448 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2449 		if (ret == IB_RMPP_RESULT_CONSUMED)
2450 			goto done;
2451 	} else
2452 		ret = IB_RMPP_RESULT_UNHANDLED;
2453 
2454 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2455 	    mad_send_wr->status == IB_WC_SUCCESS) {
2456 		mad_send_wr->status = mad_send_wc->status;
2457 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2458 	}
2459 
2460 	if (--mad_send_wr->refcount > 0) {
2461 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2462 		    mad_send_wr->status == IB_WC_SUCCESS) {
2463 			wait_for_response(mad_send_wr);
2464 		}
2465 		goto done;
2466 	}
2467 
2468 	/* Remove send from MAD agent and notify client of completion */
2469 	list_del(&mad_send_wr->agent_list);
2470 	adjust_timeout(mad_agent_priv);
2471 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2472 
2473 	if (mad_send_wr->status != IB_WC_SUCCESS )
2474 		mad_send_wc->status = mad_send_wr->status;
2475 	if (ret == IB_RMPP_RESULT_INTERNAL)
2476 		ib_rmpp_send_handler(mad_send_wc);
2477 	else
2478 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2479 						   mad_send_wc);
2480 
2481 	/* Release reference on agent taken when sending */
2482 	deref_mad_agent(mad_agent_priv);
2483 	return;
2484 done:
2485 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2486 }
2487 
2488 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2489 {
2490 	struct ib_mad_port_private *port_priv = cq->cq_context;
2491 	struct ib_mad_list_head *mad_list =
2492 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2493 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2494 	struct ib_mad_qp_info		*qp_info;
2495 	struct ib_mad_queue		*send_queue;
2496 	struct ib_mad_send_wc		mad_send_wc;
2497 	unsigned long flags;
2498 	int ret;
2499 
2500 	if (list_empty_careful(&port_priv->port_list))
2501 		return;
2502 
2503 	if (wc->status != IB_WC_SUCCESS) {
2504 		if (!ib_mad_send_error(port_priv, wc))
2505 			return;
2506 	}
2507 
2508 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2509 				   mad_list);
2510 	send_queue = mad_list->mad_queue;
2511 	qp_info = send_queue->qp_info;
2512 
2513 	trace_ib_mad_send_done_agent(mad_send_wr->mad_agent_priv);
2514 	trace_ib_mad_send_done_handler(mad_send_wr, wc);
2515 
2516 retry:
2517 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2518 			    mad_send_wr->header_mapping,
2519 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2520 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2521 			    mad_send_wr->payload_mapping,
2522 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2523 	queued_send_wr = NULL;
2524 	spin_lock_irqsave(&send_queue->lock, flags);
2525 	list_del(&mad_list->list);
2526 
2527 	/* Move queued send to the send queue */
2528 	if (send_queue->count-- > send_queue->max_active) {
2529 		mad_list = container_of(qp_info->overflow_list.next,
2530 					struct ib_mad_list_head, list);
2531 		queued_send_wr = container_of(mad_list,
2532 					struct ib_mad_send_wr_private,
2533 					mad_list);
2534 		list_move_tail(&mad_list->list, &send_queue->list);
2535 	}
2536 	spin_unlock_irqrestore(&send_queue->lock, flags);
2537 
2538 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2539 	mad_send_wc.status = wc->status;
2540 	mad_send_wc.vendor_err = wc->vendor_err;
2541 	if (atomic_read(&qp_info->snoop_count))
2542 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2543 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2544 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2545 
2546 	if (queued_send_wr) {
2547 		trace_ib_mad_send_done_resend(queued_send_wr, qp_info);
2548 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2549 				   NULL);
2550 		if (ret) {
2551 			dev_err(&port_priv->device->dev,
2552 				"ib_post_send failed: %d\n", ret);
2553 			mad_send_wr = queued_send_wr;
2554 			wc->status = IB_WC_LOC_QP_OP_ERR;
2555 			goto retry;
2556 		}
2557 	}
2558 }
2559 
2560 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2561 {
2562 	struct ib_mad_send_wr_private *mad_send_wr;
2563 	struct ib_mad_list_head *mad_list;
2564 	unsigned long flags;
2565 
2566 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2567 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2568 		mad_send_wr = container_of(mad_list,
2569 					   struct ib_mad_send_wr_private,
2570 					   mad_list);
2571 		mad_send_wr->retry = 1;
2572 	}
2573 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2574 }
2575 
2576 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2577 		struct ib_wc *wc)
2578 {
2579 	struct ib_mad_list_head *mad_list =
2580 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2581 	struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2582 	struct ib_mad_send_wr_private *mad_send_wr;
2583 	int ret;
2584 
2585 	/*
2586 	 * Send errors will transition the QP to SQE - move
2587 	 * QP to RTS and repost flushed work requests
2588 	 */
2589 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2590 				   mad_list);
2591 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2592 		if (mad_send_wr->retry) {
2593 			/* Repost send */
2594 			mad_send_wr->retry = 0;
2595 			trace_ib_mad_error_handler(mad_send_wr, qp_info);
2596 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2597 					   NULL);
2598 			if (!ret)
2599 				return false;
2600 		}
2601 	} else {
2602 		struct ib_qp_attr *attr;
2603 
2604 		/* Transition QP to RTS and fail offending send */
2605 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2606 		if (attr) {
2607 			attr->qp_state = IB_QPS_RTS;
2608 			attr->cur_qp_state = IB_QPS_SQE;
2609 			ret = ib_modify_qp(qp_info->qp, attr,
2610 					   IB_QP_STATE | IB_QP_CUR_STATE);
2611 			kfree(attr);
2612 			if (ret)
2613 				dev_err(&port_priv->device->dev,
2614 					"%s - ib_modify_qp to RTS: %d\n",
2615 					__func__, ret);
2616 			else
2617 				mark_sends_for_retry(qp_info);
2618 		}
2619 	}
2620 
2621 	return true;
2622 }
2623 
2624 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2625 {
2626 	unsigned long flags;
2627 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2628 	struct ib_mad_send_wc mad_send_wc;
2629 	struct list_head cancel_list;
2630 
2631 	INIT_LIST_HEAD(&cancel_list);
2632 
2633 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2634 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2635 				 &mad_agent_priv->send_list, agent_list) {
2636 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2637 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2638 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2639 		}
2640 	}
2641 
2642 	/* Empty wait list to prevent receives from finding a request */
2643 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2644 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2645 
2646 	/* Report all cancelled requests */
2647 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2648 	mad_send_wc.vendor_err = 0;
2649 
2650 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2651 				 &cancel_list, agent_list) {
2652 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2653 		list_del(&mad_send_wr->agent_list);
2654 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2655 						   &mad_send_wc);
2656 		atomic_dec(&mad_agent_priv->refcount);
2657 	}
2658 }
2659 
2660 static struct ib_mad_send_wr_private*
2661 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2662 	     struct ib_mad_send_buf *send_buf)
2663 {
2664 	struct ib_mad_send_wr_private *mad_send_wr;
2665 
2666 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2667 			    agent_list) {
2668 		if (&mad_send_wr->send_buf == send_buf)
2669 			return mad_send_wr;
2670 	}
2671 
2672 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2673 			    agent_list) {
2674 		if (is_rmpp_data_mad(mad_agent_priv,
2675 				     mad_send_wr->send_buf.mad) &&
2676 		    &mad_send_wr->send_buf == send_buf)
2677 			return mad_send_wr;
2678 	}
2679 	return NULL;
2680 }
2681 
2682 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2683 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2684 {
2685 	struct ib_mad_agent_private *mad_agent_priv;
2686 	struct ib_mad_send_wr_private *mad_send_wr;
2687 	unsigned long flags;
2688 	int active;
2689 
2690 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2691 				      agent);
2692 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2693 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2694 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2695 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2696 		return -EINVAL;
2697 	}
2698 
2699 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2700 	if (!timeout_ms) {
2701 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2702 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2703 	}
2704 
2705 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2706 	if (active)
2707 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2708 	else
2709 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2710 
2711 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2712 	return 0;
2713 }
2714 EXPORT_SYMBOL(ib_modify_mad);
2715 
2716 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2717 		   struct ib_mad_send_buf *send_buf)
2718 {
2719 	ib_modify_mad(mad_agent, send_buf, 0);
2720 }
2721 EXPORT_SYMBOL(ib_cancel_mad);
2722 
2723 static void local_completions(struct work_struct *work)
2724 {
2725 	struct ib_mad_agent_private *mad_agent_priv;
2726 	struct ib_mad_local_private *local;
2727 	struct ib_mad_agent_private *recv_mad_agent;
2728 	unsigned long flags;
2729 	int free_mad;
2730 	struct ib_wc wc;
2731 	struct ib_mad_send_wc mad_send_wc;
2732 	bool opa;
2733 
2734 	mad_agent_priv =
2735 		container_of(work, struct ib_mad_agent_private, local_work);
2736 
2737 	opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2738 			       mad_agent_priv->qp_info->port_priv->port_num);
2739 
2740 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2741 	while (!list_empty(&mad_agent_priv->local_list)) {
2742 		local = list_entry(mad_agent_priv->local_list.next,
2743 				   struct ib_mad_local_private,
2744 				   completion_list);
2745 		list_del(&local->completion_list);
2746 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2747 		free_mad = 0;
2748 		if (local->mad_priv) {
2749 			u8 base_version;
2750 			recv_mad_agent = local->recv_mad_agent;
2751 			if (!recv_mad_agent) {
2752 				dev_err(&mad_agent_priv->agent.device->dev,
2753 					"No receive MAD agent for local completion\n");
2754 				free_mad = 1;
2755 				goto local_send_completion;
2756 			}
2757 
2758 			/*
2759 			 * Defined behavior is to complete response
2760 			 * before request
2761 			 */
2762 			build_smp_wc(recv_mad_agent->agent.qp,
2763 				     local->mad_send_wr->send_wr.wr.wr_cqe,
2764 				     be16_to_cpu(IB_LID_PERMISSIVE),
2765 				     local->mad_send_wr->send_wr.pkey_index,
2766 				     recv_mad_agent->agent.port_num, &wc);
2767 
2768 			local->mad_priv->header.recv_wc.wc = &wc;
2769 
2770 			base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2771 			if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2772 				local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2773 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2774 			} else {
2775 				local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2776 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2777 			}
2778 
2779 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2780 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2781 				 &local->mad_priv->header.recv_wc.rmpp_list);
2782 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2783 			local->mad_priv->header.recv_wc.recv_buf.mad =
2784 						(struct ib_mad *)local->mad_priv->mad;
2785 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2786 				snoop_recv(recv_mad_agent->qp_info,
2787 					  &local->mad_priv->header.recv_wc,
2788 					   IB_MAD_SNOOP_RECVS);
2789 			recv_mad_agent->agent.recv_handler(
2790 						&recv_mad_agent->agent,
2791 						&local->mad_send_wr->send_buf,
2792 						&local->mad_priv->header.recv_wc);
2793 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2794 			atomic_dec(&recv_mad_agent->refcount);
2795 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2796 		}
2797 
2798 local_send_completion:
2799 		/* Complete send */
2800 		mad_send_wc.status = IB_WC_SUCCESS;
2801 		mad_send_wc.vendor_err = 0;
2802 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2803 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2804 			snoop_send(mad_agent_priv->qp_info,
2805 				   &local->mad_send_wr->send_buf,
2806 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2807 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2808 						   &mad_send_wc);
2809 
2810 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2811 		atomic_dec(&mad_agent_priv->refcount);
2812 		if (free_mad)
2813 			kfree(local->mad_priv);
2814 		kfree(local);
2815 	}
2816 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2817 }
2818 
2819 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2820 {
2821 	int ret;
2822 
2823 	if (!mad_send_wr->retries_left)
2824 		return -ETIMEDOUT;
2825 
2826 	mad_send_wr->retries_left--;
2827 	mad_send_wr->send_buf.retries++;
2828 
2829 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2830 
2831 	if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2832 		ret = ib_retry_rmpp(mad_send_wr);
2833 		switch (ret) {
2834 		case IB_RMPP_RESULT_UNHANDLED:
2835 			ret = ib_send_mad(mad_send_wr);
2836 			break;
2837 		case IB_RMPP_RESULT_CONSUMED:
2838 			ret = 0;
2839 			break;
2840 		default:
2841 			ret = -ECOMM;
2842 			break;
2843 		}
2844 	} else
2845 		ret = ib_send_mad(mad_send_wr);
2846 
2847 	if (!ret) {
2848 		mad_send_wr->refcount++;
2849 		list_add_tail(&mad_send_wr->agent_list,
2850 			      &mad_send_wr->mad_agent_priv->send_list);
2851 	}
2852 	return ret;
2853 }
2854 
2855 static void timeout_sends(struct work_struct *work)
2856 {
2857 	struct ib_mad_agent_private *mad_agent_priv;
2858 	struct ib_mad_send_wr_private *mad_send_wr;
2859 	struct ib_mad_send_wc mad_send_wc;
2860 	unsigned long flags, delay;
2861 
2862 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2863 				      timed_work.work);
2864 	mad_send_wc.vendor_err = 0;
2865 
2866 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2867 	while (!list_empty(&mad_agent_priv->wait_list)) {
2868 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2869 					 struct ib_mad_send_wr_private,
2870 					 agent_list);
2871 
2872 		if (time_after(mad_send_wr->timeout, jiffies)) {
2873 			delay = mad_send_wr->timeout - jiffies;
2874 			if ((long)delay <= 0)
2875 				delay = 1;
2876 			queue_delayed_work(mad_agent_priv->qp_info->
2877 					   port_priv->wq,
2878 					   &mad_agent_priv->timed_work, delay);
2879 			break;
2880 		}
2881 
2882 		list_del(&mad_send_wr->agent_list);
2883 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2884 		    !retry_send(mad_send_wr))
2885 			continue;
2886 
2887 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2888 
2889 		if (mad_send_wr->status == IB_WC_SUCCESS)
2890 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2891 		else
2892 			mad_send_wc.status = mad_send_wr->status;
2893 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2894 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2895 						   &mad_send_wc);
2896 
2897 		atomic_dec(&mad_agent_priv->refcount);
2898 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2899 	}
2900 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2901 }
2902 
2903 /*
2904  * Allocate receive MADs and post receive WRs for them
2905  */
2906 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2907 				    struct ib_mad_private *mad)
2908 {
2909 	unsigned long flags;
2910 	int post, ret;
2911 	struct ib_mad_private *mad_priv;
2912 	struct ib_sge sg_list;
2913 	struct ib_recv_wr recv_wr;
2914 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2915 
2916 	/* Initialize common scatter list fields */
2917 	sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2918 
2919 	/* Initialize common receive WR fields */
2920 	recv_wr.next = NULL;
2921 	recv_wr.sg_list = &sg_list;
2922 	recv_wr.num_sge = 1;
2923 
2924 	do {
2925 		/* Allocate and map receive buffer */
2926 		if (mad) {
2927 			mad_priv = mad;
2928 			mad = NULL;
2929 		} else {
2930 			mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2931 						     GFP_ATOMIC);
2932 			if (!mad_priv) {
2933 				ret = -ENOMEM;
2934 				break;
2935 			}
2936 		}
2937 		sg_list.length = mad_priv_dma_size(mad_priv);
2938 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2939 						 &mad_priv->grh,
2940 						 mad_priv_dma_size(mad_priv),
2941 						 DMA_FROM_DEVICE);
2942 		if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2943 						  sg_list.addr))) {
2944 			ret = -ENOMEM;
2945 			break;
2946 		}
2947 		mad_priv->header.mapping = sg_list.addr;
2948 		mad_priv->header.mad_list.mad_queue = recv_queue;
2949 		mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2950 		recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2951 
2952 		/* Post receive WR */
2953 		spin_lock_irqsave(&recv_queue->lock, flags);
2954 		post = (++recv_queue->count < recv_queue->max_active);
2955 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2956 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2957 		ret = ib_post_recv(qp_info->qp, &recv_wr, NULL);
2958 		if (ret) {
2959 			spin_lock_irqsave(&recv_queue->lock, flags);
2960 			list_del(&mad_priv->header.mad_list.list);
2961 			recv_queue->count--;
2962 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2963 			ib_dma_unmap_single(qp_info->port_priv->device,
2964 					    mad_priv->header.mapping,
2965 					    mad_priv_dma_size(mad_priv),
2966 					    DMA_FROM_DEVICE);
2967 			kfree(mad_priv);
2968 			dev_err(&qp_info->port_priv->device->dev,
2969 				"ib_post_recv failed: %d\n", ret);
2970 			break;
2971 		}
2972 	} while (post);
2973 
2974 	return ret;
2975 }
2976 
2977 /*
2978  * Return all the posted receive MADs
2979  */
2980 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2981 {
2982 	struct ib_mad_private_header *mad_priv_hdr;
2983 	struct ib_mad_private *recv;
2984 	struct ib_mad_list_head *mad_list;
2985 
2986 	if (!qp_info->qp)
2987 		return;
2988 
2989 	while (!list_empty(&qp_info->recv_queue.list)) {
2990 
2991 		mad_list = list_entry(qp_info->recv_queue.list.next,
2992 				      struct ib_mad_list_head, list);
2993 		mad_priv_hdr = container_of(mad_list,
2994 					    struct ib_mad_private_header,
2995 					    mad_list);
2996 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2997 				    header);
2998 
2999 		/* Remove from posted receive MAD list */
3000 		list_del(&mad_list->list);
3001 
3002 		ib_dma_unmap_single(qp_info->port_priv->device,
3003 				    recv->header.mapping,
3004 				    mad_priv_dma_size(recv),
3005 				    DMA_FROM_DEVICE);
3006 		kfree(recv);
3007 	}
3008 
3009 	qp_info->recv_queue.count = 0;
3010 }
3011 
3012 /*
3013  * Start the port
3014  */
3015 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
3016 {
3017 	int ret, i;
3018 	struct ib_qp_attr *attr;
3019 	struct ib_qp *qp;
3020 	u16 pkey_index;
3021 
3022 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
3023 	if (!attr)
3024 		return -ENOMEM;
3025 
3026 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
3027 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
3028 	if (ret)
3029 		pkey_index = 0;
3030 
3031 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3032 		qp = port_priv->qp_info[i].qp;
3033 		if (!qp)
3034 			continue;
3035 
3036 		/*
3037 		 * PKey index for QP1 is irrelevant but
3038 		 * one is needed for the Reset to Init transition
3039 		 */
3040 		attr->qp_state = IB_QPS_INIT;
3041 		attr->pkey_index = pkey_index;
3042 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
3043 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
3044 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
3045 		if (ret) {
3046 			dev_err(&port_priv->device->dev,
3047 				"Couldn't change QP%d state to INIT: %d\n",
3048 				i, ret);
3049 			goto out;
3050 		}
3051 
3052 		attr->qp_state = IB_QPS_RTR;
3053 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
3054 		if (ret) {
3055 			dev_err(&port_priv->device->dev,
3056 				"Couldn't change QP%d state to RTR: %d\n",
3057 				i, ret);
3058 			goto out;
3059 		}
3060 
3061 		attr->qp_state = IB_QPS_RTS;
3062 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
3063 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3064 		if (ret) {
3065 			dev_err(&port_priv->device->dev,
3066 				"Couldn't change QP%d state to RTS: %d\n",
3067 				i, ret);
3068 			goto out;
3069 		}
3070 	}
3071 
3072 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3073 	if (ret) {
3074 		dev_err(&port_priv->device->dev,
3075 			"Failed to request completion notification: %d\n",
3076 			ret);
3077 		goto out;
3078 	}
3079 
3080 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3081 		if (!port_priv->qp_info[i].qp)
3082 			continue;
3083 
3084 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3085 		if (ret) {
3086 			dev_err(&port_priv->device->dev,
3087 				"Couldn't post receive WRs\n");
3088 			goto out;
3089 		}
3090 	}
3091 out:
3092 	kfree(attr);
3093 	return ret;
3094 }
3095 
3096 static void qp_event_handler(struct ib_event *event, void *qp_context)
3097 {
3098 	struct ib_mad_qp_info	*qp_info = qp_context;
3099 
3100 	/* It's worse than that! He's dead, Jim! */
3101 	dev_err(&qp_info->port_priv->device->dev,
3102 		"Fatal error (%d) on MAD QP (%d)\n",
3103 		event->event, qp_info->qp->qp_num);
3104 }
3105 
3106 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3107 			   struct ib_mad_queue *mad_queue)
3108 {
3109 	mad_queue->qp_info = qp_info;
3110 	mad_queue->count = 0;
3111 	spin_lock_init(&mad_queue->lock);
3112 	INIT_LIST_HEAD(&mad_queue->list);
3113 }
3114 
3115 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3116 			struct ib_mad_qp_info *qp_info)
3117 {
3118 	qp_info->port_priv = port_priv;
3119 	init_mad_queue(qp_info, &qp_info->send_queue);
3120 	init_mad_queue(qp_info, &qp_info->recv_queue);
3121 	INIT_LIST_HEAD(&qp_info->overflow_list);
3122 	spin_lock_init(&qp_info->snoop_lock);
3123 	qp_info->snoop_table = NULL;
3124 	qp_info->snoop_table_size = 0;
3125 	atomic_set(&qp_info->snoop_count, 0);
3126 }
3127 
3128 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3129 			 enum ib_qp_type qp_type)
3130 {
3131 	struct ib_qp_init_attr	qp_init_attr;
3132 	int ret;
3133 
3134 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
3135 	qp_init_attr.send_cq = qp_info->port_priv->cq;
3136 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
3137 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3138 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
3139 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3140 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3141 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3142 	qp_init_attr.qp_type = qp_type;
3143 	qp_init_attr.port_num = qp_info->port_priv->port_num;
3144 	qp_init_attr.qp_context = qp_info;
3145 	qp_init_attr.event_handler = qp_event_handler;
3146 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3147 	if (IS_ERR(qp_info->qp)) {
3148 		dev_err(&qp_info->port_priv->device->dev,
3149 			"Couldn't create ib_mad QP%d\n",
3150 			get_spl_qp_index(qp_type));
3151 		ret = PTR_ERR(qp_info->qp);
3152 		goto error;
3153 	}
3154 	/* Use minimum queue sizes unless the CQ is resized */
3155 	qp_info->send_queue.max_active = mad_sendq_size;
3156 	qp_info->recv_queue.max_active = mad_recvq_size;
3157 	return 0;
3158 
3159 error:
3160 	return ret;
3161 }
3162 
3163 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3164 {
3165 	if (!qp_info->qp)
3166 		return;
3167 
3168 	ib_destroy_qp(qp_info->qp);
3169 	kfree(qp_info->snoop_table);
3170 }
3171 
3172 /*
3173  * Open the port
3174  * Create the QP, PD, MR, and CQ if needed
3175  */
3176 static int ib_mad_port_open(struct ib_device *device,
3177 			    int port_num)
3178 {
3179 	int ret, cq_size;
3180 	struct ib_mad_port_private *port_priv;
3181 	unsigned long flags;
3182 	char name[sizeof "ib_mad123"];
3183 	int has_smi;
3184 
3185 	if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3186 		return -EFAULT;
3187 
3188 	if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3189 		    rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3190 		return -EFAULT;
3191 
3192 	/* Create new device info */
3193 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3194 	if (!port_priv)
3195 		return -ENOMEM;
3196 
3197 	port_priv->device = device;
3198 	port_priv->port_num = port_num;
3199 	spin_lock_init(&port_priv->reg_lock);
3200 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
3201 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
3202 
3203 	cq_size = mad_sendq_size + mad_recvq_size;
3204 	has_smi = rdma_cap_ib_smi(device, port_num);
3205 	if (has_smi)
3206 		cq_size *= 2;
3207 
3208 	port_priv->pd = ib_alloc_pd(device, 0);
3209 	if (IS_ERR(port_priv->pd)) {
3210 		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3211 		ret = PTR_ERR(port_priv->pd);
3212 		goto error3;
3213 	}
3214 
3215 	port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3216 			IB_POLL_UNBOUND_WORKQUEUE);
3217 	if (IS_ERR(port_priv->cq)) {
3218 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3219 		ret = PTR_ERR(port_priv->cq);
3220 		goto error4;
3221 	}
3222 
3223 	if (has_smi) {
3224 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3225 		if (ret)
3226 			goto error6;
3227 	}
3228 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3229 	if (ret)
3230 		goto error7;
3231 
3232 	snprintf(name, sizeof name, "ib_mad%d", port_num);
3233 	port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3234 	if (!port_priv->wq) {
3235 		ret = -ENOMEM;
3236 		goto error8;
3237 	}
3238 
3239 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3240 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3241 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3242 
3243 	ret = ib_mad_port_start(port_priv);
3244 	if (ret) {
3245 		dev_err(&device->dev, "Couldn't start port\n");
3246 		goto error9;
3247 	}
3248 
3249 	return 0;
3250 
3251 error9:
3252 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3253 	list_del_init(&port_priv->port_list);
3254 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3255 
3256 	destroy_workqueue(port_priv->wq);
3257 error8:
3258 	destroy_mad_qp(&port_priv->qp_info[1]);
3259 error7:
3260 	destroy_mad_qp(&port_priv->qp_info[0]);
3261 error6:
3262 	ib_free_cq(port_priv->cq);
3263 	cleanup_recv_queue(&port_priv->qp_info[1]);
3264 	cleanup_recv_queue(&port_priv->qp_info[0]);
3265 error4:
3266 	ib_dealloc_pd(port_priv->pd);
3267 error3:
3268 	kfree(port_priv);
3269 
3270 	return ret;
3271 }
3272 
3273 /*
3274  * Close the port
3275  * If there are no classes using the port, free the port
3276  * resources (CQ, MR, PD, QP) and remove the port's info structure
3277  */
3278 static int ib_mad_port_close(struct ib_device *device, int port_num)
3279 {
3280 	struct ib_mad_port_private *port_priv;
3281 	unsigned long flags;
3282 
3283 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3284 	port_priv = __ib_get_mad_port(device, port_num);
3285 	if (port_priv == NULL) {
3286 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3287 		dev_err(&device->dev, "Port %d not found\n", port_num);
3288 		return -ENODEV;
3289 	}
3290 	list_del_init(&port_priv->port_list);
3291 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3292 
3293 	destroy_workqueue(port_priv->wq);
3294 	destroy_mad_qp(&port_priv->qp_info[1]);
3295 	destroy_mad_qp(&port_priv->qp_info[0]);
3296 	ib_free_cq(port_priv->cq);
3297 	ib_dealloc_pd(port_priv->pd);
3298 	cleanup_recv_queue(&port_priv->qp_info[1]);
3299 	cleanup_recv_queue(&port_priv->qp_info[0]);
3300 	/* XXX: Handle deallocation of MAD registration tables */
3301 
3302 	kfree(port_priv);
3303 
3304 	return 0;
3305 }
3306 
3307 static void ib_mad_init_device(struct ib_device *device)
3308 {
3309 	int start, i;
3310 
3311 	start = rdma_start_port(device);
3312 
3313 	for (i = start; i <= rdma_end_port(device); i++) {
3314 		if (!rdma_cap_ib_mad(device, i))
3315 			continue;
3316 
3317 		if (ib_mad_port_open(device, i)) {
3318 			dev_err(&device->dev, "Couldn't open port %d\n", i);
3319 			goto error;
3320 		}
3321 		if (ib_agent_port_open(device, i)) {
3322 			dev_err(&device->dev,
3323 				"Couldn't open port %d for agents\n", i);
3324 			goto error_agent;
3325 		}
3326 	}
3327 	return;
3328 
3329 error_agent:
3330 	if (ib_mad_port_close(device, i))
3331 		dev_err(&device->dev, "Couldn't close port %d\n", i);
3332 
3333 error:
3334 	while (--i >= start) {
3335 		if (!rdma_cap_ib_mad(device, i))
3336 			continue;
3337 
3338 		if (ib_agent_port_close(device, i))
3339 			dev_err(&device->dev,
3340 				"Couldn't close port %d for agents\n", i);
3341 		if (ib_mad_port_close(device, i))
3342 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3343 	}
3344 }
3345 
3346 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3347 {
3348 	unsigned int i;
3349 
3350 	rdma_for_each_port (device, i) {
3351 		if (!rdma_cap_ib_mad(device, i))
3352 			continue;
3353 
3354 		if (ib_agent_port_close(device, i))
3355 			dev_err(&device->dev,
3356 				"Couldn't close port %d for agents\n", i);
3357 		if (ib_mad_port_close(device, i))
3358 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3359 	}
3360 }
3361 
3362 static struct ib_client mad_client = {
3363 	.name   = "mad",
3364 	.add = ib_mad_init_device,
3365 	.remove = ib_mad_remove_device
3366 };
3367 
3368 int ib_mad_init(void)
3369 {
3370 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3371 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3372 
3373 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3374 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3375 
3376 	INIT_LIST_HEAD(&ib_mad_port_list);
3377 
3378 	if (ib_register_client(&mad_client)) {
3379 		pr_err("Couldn't register ib_mad client\n");
3380 		return -EINVAL;
3381 	}
3382 
3383 	return 0;
3384 }
3385 
3386 void ib_mad_cleanup(void)
3387 {
3388 	ib_unregister_client(&mad_client);
3389 }
3390