locking/refcounts: Include fewer headers in <linux/refcount.h>
[linux/fpc-iii.git] / drivers / infiniband / core / mad.c
blobf742ae7a768b2ea4e3b98f665a297c6e5c314a5e
1 /*
2 * Copyright (c) 2004-2007 Voltaire, Inc. All rights reserved.
3 * Copyright (c) 2005 Intel Corporation. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies Ltd. All rights reserved.
5 * Copyright (c) 2009 HNR Consulting. All rights reserved.
6 * Copyright (c) 2014 Intel Corporation. All rights reserved.
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:
14 * Redistribution and use in source and binary forms, with or
15 * without modification, are permitted provided that the following
16 * conditions are met:
18 * - Redistributions of source code must retain the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer.
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.
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.
38 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
40 #include <linux/dma-mapping.h>
41 #include <linux/slab.h>
42 #include <linux/module.h>
43 #include <linux/security.h>
44 #include <rdma/ib_cache.h>
46 #include "mad_priv.h"
47 #include "core_priv.h"
48 #include "mad_rmpp.h"
49 #include "smi.h"
50 #include "opa_smi.h"
51 #include "agent.h"
53 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
54 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
56 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
57 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
58 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
59 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
61 static struct list_head ib_mad_port_list;
62 static atomic_t ib_mad_client_id = ATOMIC_INIT(0);
64 /* Port list lock */
65 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
67 /* Forward declarations */
68 static int method_in_use(struct ib_mad_mgmt_method_table **method,
69 struct ib_mad_reg_req *mad_reg_req);
70 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
71 static struct ib_mad_agent_private *find_mad_agent(
72 struct ib_mad_port_private *port_priv,
73 const struct ib_mad_hdr *mad);
74 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
75 struct ib_mad_private *mad);
76 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
77 static void timeout_sends(struct work_struct *work);
78 static void local_completions(struct work_struct *work);
79 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
80 struct ib_mad_agent_private *agent_priv,
81 u8 mgmt_class);
82 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
83 struct ib_mad_agent_private *agent_priv);
84 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
85 struct ib_wc *wc);
86 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
89 * Returns a ib_mad_port_private structure or NULL for a device/port
90 * Assumes ib_mad_port_list_lock is being held
92 static inline struct ib_mad_port_private *
93 __ib_get_mad_port(struct ib_device *device, int port_num)
95 struct ib_mad_port_private *entry;
97 list_for_each_entry(entry, &ib_mad_port_list, port_list) {
98 if (entry->device == device && entry->port_num == port_num)
99 return entry;
101 return NULL;
105 * Wrapper function to return a ib_mad_port_private structure or NULL
106 * for a device/port
108 static inline struct ib_mad_port_private *
109 ib_get_mad_port(struct ib_device *device, int port_num)
111 struct ib_mad_port_private *entry;
112 unsigned long flags;
114 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
115 entry = __ib_get_mad_port(device, port_num);
116 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
118 return entry;
121 static inline u8 convert_mgmt_class(u8 mgmt_class)
123 /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
124 return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
125 0 : mgmt_class;
128 static int get_spl_qp_index(enum ib_qp_type qp_type)
130 switch (qp_type)
132 case IB_QPT_SMI:
133 return 0;
134 case IB_QPT_GSI:
135 return 1;
136 default:
137 return -1;
141 static int vendor_class_index(u8 mgmt_class)
143 return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
146 static int is_vendor_class(u8 mgmt_class)
148 if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
149 (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
150 return 0;
151 return 1;
154 static int is_vendor_oui(char *oui)
156 if (oui[0] || oui[1] || oui[2])
157 return 1;
158 return 0;
161 static int is_vendor_method_in_use(
162 struct ib_mad_mgmt_vendor_class *vendor_class,
163 struct ib_mad_reg_req *mad_reg_req)
165 struct ib_mad_mgmt_method_table *method;
166 int i;
168 for (i = 0; i < MAX_MGMT_OUI; i++) {
169 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
170 method = vendor_class->method_table[i];
171 if (method) {
172 if (method_in_use(&method, mad_reg_req))
173 return 1;
174 else
175 break;
179 return 0;
182 int ib_response_mad(const struct ib_mad_hdr *hdr)
184 return ((hdr->method & IB_MGMT_METHOD_RESP) ||
185 (hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
186 ((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
187 (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
189 EXPORT_SYMBOL(ib_response_mad);
192 * ib_register_mad_agent - Register to send/receive MADs
194 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
195 u8 port_num,
196 enum ib_qp_type qp_type,
197 struct ib_mad_reg_req *mad_reg_req,
198 u8 rmpp_version,
199 ib_mad_send_handler send_handler,
200 ib_mad_recv_handler recv_handler,
201 void *context,
202 u32 registration_flags)
204 struct ib_mad_port_private *port_priv;
205 struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
206 struct ib_mad_agent_private *mad_agent_priv;
207 struct ib_mad_reg_req *reg_req = NULL;
208 struct ib_mad_mgmt_class_table *class;
209 struct ib_mad_mgmt_vendor_class_table *vendor;
210 struct ib_mad_mgmt_vendor_class *vendor_class;
211 struct ib_mad_mgmt_method_table *method;
212 int ret2, qpn;
213 unsigned long flags;
214 u8 mgmt_class, vclass;
216 /* Validate parameters */
217 qpn = get_spl_qp_index(qp_type);
218 if (qpn == -1) {
219 dev_notice(&device->dev,
220 "ib_register_mad_agent: invalid QP Type %d\n",
221 qp_type);
222 goto error1;
225 if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
226 dev_notice(&device->dev,
227 "ib_register_mad_agent: invalid RMPP Version %u\n",
228 rmpp_version);
229 goto error1;
232 /* Validate MAD registration request if supplied */
233 if (mad_reg_req) {
234 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
235 dev_notice(&device->dev,
236 "ib_register_mad_agent: invalid Class Version %u\n",
237 mad_reg_req->mgmt_class_version);
238 goto error1;
240 if (!recv_handler) {
241 dev_notice(&device->dev,
242 "ib_register_mad_agent: no recv_handler\n");
243 goto error1;
245 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
247 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
248 * one in this range currently allowed
250 if (mad_reg_req->mgmt_class !=
251 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
252 dev_notice(&device->dev,
253 "ib_register_mad_agent: Invalid Mgmt Class 0x%x\n",
254 mad_reg_req->mgmt_class);
255 goto error1;
257 } else if (mad_reg_req->mgmt_class == 0) {
259 * Class 0 is reserved in IBA and is used for
260 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
262 dev_notice(&device->dev,
263 "ib_register_mad_agent: Invalid Mgmt Class 0\n");
264 goto error1;
265 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
267 * If class is in "new" vendor range,
268 * ensure supplied OUI is not zero
270 if (!is_vendor_oui(mad_reg_req->oui)) {
271 dev_notice(&device->dev,
272 "ib_register_mad_agent: No OUI specified for class 0x%x\n",
273 mad_reg_req->mgmt_class);
274 goto error1;
277 /* Make sure class supplied is consistent with RMPP */
278 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
279 if (rmpp_version) {
280 dev_notice(&device->dev,
281 "ib_register_mad_agent: RMPP version for non-RMPP class 0x%x\n",
282 mad_reg_req->mgmt_class);
283 goto error1;
287 /* Make sure class supplied is consistent with QP type */
288 if (qp_type == IB_QPT_SMI) {
289 if ((mad_reg_req->mgmt_class !=
290 IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
291 (mad_reg_req->mgmt_class !=
292 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
293 dev_notice(&device->dev,
294 "ib_register_mad_agent: Invalid SM QP type: class 0x%x\n",
295 mad_reg_req->mgmt_class);
296 goto error1;
298 } else {
299 if ((mad_reg_req->mgmt_class ==
300 IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
301 (mad_reg_req->mgmt_class ==
302 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
303 dev_notice(&device->dev,
304 "ib_register_mad_agent: Invalid GS QP type: class 0x%x\n",
305 mad_reg_req->mgmt_class);
306 goto error1;
309 } else {
310 /* No registration request supplied */
311 if (!send_handler)
312 goto error1;
313 if (registration_flags & IB_MAD_USER_RMPP)
314 goto error1;
317 /* Validate device and port */
318 port_priv = ib_get_mad_port(device, port_num);
319 if (!port_priv) {
320 dev_notice(&device->dev,
321 "ib_register_mad_agent: Invalid port %d\n",
322 port_num);
323 ret = ERR_PTR(-ENODEV);
324 goto error1;
327 /* Verify the QP requested is supported. For example, Ethernet devices
328 * will not have QP0 */
329 if (!port_priv->qp_info[qpn].qp) {
330 dev_notice(&device->dev,
331 "ib_register_mad_agent: QP %d not supported\n", qpn);
332 ret = ERR_PTR(-EPROTONOSUPPORT);
333 goto error1;
336 /* Allocate structures */
337 mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
338 if (!mad_agent_priv) {
339 ret = ERR_PTR(-ENOMEM);
340 goto error1;
343 if (mad_reg_req) {
344 reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
345 if (!reg_req) {
346 ret = ERR_PTR(-ENOMEM);
347 goto error3;
351 /* Now, fill in the various structures */
352 mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
353 mad_agent_priv->reg_req = reg_req;
354 mad_agent_priv->agent.rmpp_version = rmpp_version;
355 mad_agent_priv->agent.device = device;
356 mad_agent_priv->agent.recv_handler = recv_handler;
357 mad_agent_priv->agent.send_handler = send_handler;
358 mad_agent_priv->agent.context = context;
359 mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
360 mad_agent_priv->agent.port_num = port_num;
361 mad_agent_priv->agent.flags = registration_flags;
362 spin_lock_init(&mad_agent_priv->lock);
363 INIT_LIST_HEAD(&mad_agent_priv->send_list);
364 INIT_LIST_HEAD(&mad_agent_priv->wait_list);
365 INIT_LIST_HEAD(&mad_agent_priv->done_list);
366 INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
367 INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
368 INIT_LIST_HEAD(&mad_agent_priv->local_list);
369 INIT_WORK(&mad_agent_priv->local_work, local_completions);
370 atomic_set(&mad_agent_priv->refcount, 1);
371 init_completion(&mad_agent_priv->comp);
373 ret2 = ib_mad_agent_security_setup(&mad_agent_priv->agent, qp_type);
374 if (ret2) {
375 ret = ERR_PTR(ret2);
376 goto error4;
379 spin_lock_irqsave(&port_priv->reg_lock, flags);
380 mad_agent_priv->agent.hi_tid = atomic_inc_return(&ib_mad_client_id);
383 * Make sure MAD registration (if supplied)
384 * is non overlapping with any existing ones
386 if (mad_reg_req) {
387 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
388 if (!is_vendor_class(mgmt_class)) {
389 class = port_priv->version[mad_reg_req->
390 mgmt_class_version].class;
391 if (class) {
392 method = class->method_table[mgmt_class];
393 if (method) {
394 if (method_in_use(&method,
395 mad_reg_req))
396 goto error5;
399 ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
400 mgmt_class);
401 } else {
402 /* "New" vendor class range */
403 vendor = port_priv->version[mad_reg_req->
404 mgmt_class_version].vendor;
405 if (vendor) {
406 vclass = vendor_class_index(mgmt_class);
407 vendor_class = vendor->vendor_class[vclass];
408 if (vendor_class) {
409 if (is_vendor_method_in_use(
410 vendor_class,
411 mad_reg_req))
412 goto error5;
415 ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
417 if (ret2) {
418 ret = ERR_PTR(ret2);
419 goto error5;
423 /* Add mad agent into port's agent list */
424 list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
425 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
427 return &mad_agent_priv->agent;
428 error5:
429 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
430 ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
431 error4:
432 kfree(reg_req);
433 error3:
434 kfree(mad_agent_priv);
435 error1:
436 return ret;
438 EXPORT_SYMBOL(ib_register_mad_agent);
440 static inline int is_snooping_sends(int mad_snoop_flags)
442 return (mad_snoop_flags &
443 (/*IB_MAD_SNOOP_POSTED_SENDS |
444 IB_MAD_SNOOP_RMPP_SENDS |*/
445 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
446 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
449 static inline int is_snooping_recvs(int mad_snoop_flags)
451 return (mad_snoop_flags &
452 (IB_MAD_SNOOP_RECVS /*|
453 IB_MAD_SNOOP_RMPP_RECVS*/));
456 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
457 struct ib_mad_snoop_private *mad_snoop_priv)
459 struct ib_mad_snoop_private **new_snoop_table;
460 unsigned long flags;
461 int i;
463 spin_lock_irqsave(&qp_info->snoop_lock, flags);
464 /* Check for empty slot in array. */
465 for (i = 0; i < qp_info->snoop_table_size; i++)
466 if (!qp_info->snoop_table[i])
467 break;
469 if (i == qp_info->snoop_table_size) {
470 /* Grow table. */
471 new_snoop_table = krealloc(qp_info->snoop_table,
472 sizeof mad_snoop_priv *
473 (qp_info->snoop_table_size + 1),
474 GFP_ATOMIC);
475 if (!new_snoop_table) {
476 i = -ENOMEM;
477 goto out;
480 qp_info->snoop_table = new_snoop_table;
481 qp_info->snoop_table_size++;
483 qp_info->snoop_table[i] = mad_snoop_priv;
484 atomic_inc(&qp_info->snoop_count);
485 out:
486 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
487 return i;
490 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
491 u8 port_num,
492 enum ib_qp_type qp_type,
493 int mad_snoop_flags,
494 ib_mad_snoop_handler snoop_handler,
495 ib_mad_recv_handler recv_handler,
496 void *context)
498 struct ib_mad_port_private *port_priv;
499 struct ib_mad_agent *ret;
500 struct ib_mad_snoop_private *mad_snoop_priv;
501 int qpn;
502 int err;
504 /* Validate parameters */
505 if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
506 (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
507 ret = ERR_PTR(-EINVAL);
508 goto error1;
510 qpn = get_spl_qp_index(qp_type);
511 if (qpn == -1) {
512 ret = ERR_PTR(-EINVAL);
513 goto error1;
515 port_priv = ib_get_mad_port(device, port_num);
516 if (!port_priv) {
517 ret = ERR_PTR(-ENODEV);
518 goto error1;
520 /* Allocate structures */
521 mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
522 if (!mad_snoop_priv) {
523 ret = ERR_PTR(-ENOMEM);
524 goto error1;
527 /* Now, fill in the various structures */
528 mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
529 mad_snoop_priv->agent.device = device;
530 mad_snoop_priv->agent.recv_handler = recv_handler;
531 mad_snoop_priv->agent.snoop_handler = snoop_handler;
532 mad_snoop_priv->agent.context = context;
533 mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
534 mad_snoop_priv->agent.port_num = port_num;
535 mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
536 init_completion(&mad_snoop_priv->comp);
538 err = ib_mad_agent_security_setup(&mad_snoop_priv->agent, qp_type);
539 if (err) {
540 ret = ERR_PTR(err);
541 goto error2;
544 mad_snoop_priv->snoop_index = register_snoop_agent(
545 &port_priv->qp_info[qpn],
546 mad_snoop_priv);
547 if (mad_snoop_priv->snoop_index < 0) {
548 ret = ERR_PTR(mad_snoop_priv->snoop_index);
549 goto error3;
552 atomic_set(&mad_snoop_priv->refcount, 1);
553 return &mad_snoop_priv->agent;
554 error3:
555 ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
556 error2:
557 kfree(mad_snoop_priv);
558 error1:
559 return ret;
561 EXPORT_SYMBOL(ib_register_mad_snoop);
563 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
565 if (atomic_dec_and_test(&mad_agent_priv->refcount))
566 complete(&mad_agent_priv->comp);
569 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
571 if (atomic_dec_and_test(&mad_snoop_priv->refcount))
572 complete(&mad_snoop_priv->comp);
575 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
577 struct ib_mad_port_private *port_priv;
578 unsigned long flags;
580 /* Note that we could still be handling received MADs */
583 * Canceling all sends results in dropping received response
584 * MADs, preventing us from queuing additional work
586 cancel_mads(mad_agent_priv);
587 port_priv = mad_agent_priv->qp_info->port_priv;
588 cancel_delayed_work(&mad_agent_priv->timed_work);
590 spin_lock_irqsave(&port_priv->reg_lock, flags);
591 remove_mad_reg_req(mad_agent_priv);
592 list_del(&mad_agent_priv->agent_list);
593 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
595 flush_workqueue(port_priv->wq);
596 ib_cancel_rmpp_recvs(mad_agent_priv);
598 deref_mad_agent(mad_agent_priv);
599 wait_for_completion(&mad_agent_priv->comp);
601 ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
603 kfree(mad_agent_priv->reg_req);
604 kfree(mad_agent_priv);
607 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
609 struct ib_mad_qp_info *qp_info;
610 unsigned long flags;
612 qp_info = mad_snoop_priv->qp_info;
613 spin_lock_irqsave(&qp_info->snoop_lock, flags);
614 qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
615 atomic_dec(&qp_info->snoop_count);
616 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
618 deref_snoop_agent(mad_snoop_priv);
619 wait_for_completion(&mad_snoop_priv->comp);
621 ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
623 kfree(mad_snoop_priv);
627 * ib_unregister_mad_agent - Unregisters a client from using MAD services
629 void ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
631 struct ib_mad_agent_private *mad_agent_priv;
632 struct ib_mad_snoop_private *mad_snoop_priv;
634 /* If the TID is zero, the agent can only snoop. */
635 if (mad_agent->hi_tid) {
636 mad_agent_priv = container_of(mad_agent,
637 struct ib_mad_agent_private,
638 agent);
639 unregister_mad_agent(mad_agent_priv);
640 } else {
641 mad_snoop_priv = container_of(mad_agent,
642 struct ib_mad_snoop_private,
643 agent);
644 unregister_mad_snoop(mad_snoop_priv);
647 EXPORT_SYMBOL(ib_unregister_mad_agent);
649 static void dequeue_mad(struct ib_mad_list_head *mad_list)
651 struct ib_mad_queue *mad_queue;
652 unsigned long flags;
654 mad_queue = mad_list->mad_queue;
655 spin_lock_irqsave(&mad_queue->lock, flags);
656 list_del(&mad_list->list);
657 mad_queue->count--;
658 spin_unlock_irqrestore(&mad_queue->lock, flags);
661 static void snoop_send(struct ib_mad_qp_info *qp_info,
662 struct ib_mad_send_buf *send_buf,
663 struct ib_mad_send_wc *mad_send_wc,
664 int mad_snoop_flags)
666 struct ib_mad_snoop_private *mad_snoop_priv;
667 unsigned long flags;
668 int i;
670 spin_lock_irqsave(&qp_info->snoop_lock, flags);
671 for (i = 0; i < qp_info->snoop_table_size; i++) {
672 mad_snoop_priv = qp_info->snoop_table[i];
673 if (!mad_snoop_priv ||
674 !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
675 continue;
677 atomic_inc(&mad_snoop_priv->refcount);
678 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
679 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
680 send_buf, mad_send_wc);
681 deref_snoop_agent(mad_snoop_priv);
682 spin_lock_irqsave(&qp_info->snoop_lock, flags);
684 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
687 static void snoop_recv(struct ib_mad_qp_info *qp_info,
688 struct ib_mad_recv_wc *mad_recv_wc,
689 int mad_snoop_flags)
691 struct ib_mad_snoop_private *mad_snoop_priv;
692 unsigned long flags;
693 int i;
695 spin_lock_irqsave(&qp_info->snoop_lock, flags);
696 for (i = 0; i < qp_info->snoop_table_size; i++) {
697 mad_snoop_priv = qp_info->snoop_table[i];
698 if (!mad_snoop_priv ||
699 !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
700 continue;
702 atomic_inc(&mad_snoop_priv->refcount);
703 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
704 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
705 mad_recv_wc);
706 deref_snoop_agent(mad_snoop_priv);
707 spin_lock_irqsave(&qp_info->snoop_lock, flags);
709 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
712 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
713 u16 pkey_index, u8 port_num, struct ib_wc *wc)
715 memset(wc, 0, sizeof *wc);
716 wc->wr_cqe = cqe;
717 wc->status = IB_WC_SUCCESS;
718 wc->opcode = IB_WC_RECV;
719 wc->pkey_index = pkey_index;
720 wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
721 wc->src_qp = IB_QP0;
722 wc->qp = qp;
723 wc->slid = slid;
724 wc->sl = 0;
725 wc->dlid_path_bits = 0;
726 wc->port_num = port_num;
729 static size_t mad_priv_size(const struct ib_mad_private *mp)
731 return sizeof(struct ib_mad_private) + mp->mad_size;
734 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
736 size_t size = sizeof(struct ib_mad_private) + mad_size;
737 struct ib_mad_private *ret = kzalloc(size, flags);
739 if (ret)
740 ret->mad_size = mad_size;
742 return ret;
745 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
747 return rdma_max_mad_size(port_priv->device, port_priv->port_num);
750 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
752 return sizeof(struct ib_grh) + mp->mad_size;
756 * Return 0 if SMP is to be sent
757 * Return 1 if SMP was consumed locally (whether or not solicited)
758 * Return < 0 if error
760 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
761 struct ib_mad_send_wr_private *mad_send_wr)
763 int ret = 0;
764 struct ib_smp *smp = mad_send_wr->send_buf.mad;
765 struct opa_smp *opa_smp = (struct opa_smp *)smp;
766 unsigned long flags;
767 struct ib_mad_local_private *local;
768 struct ib_mad_private *mad_priv;
769 struct ib_mad_port_private *port_priv;
770 struct ib_mad_agent_private *recv_mad_agent = NULL;
771 struct ib_device *device = mad_agent_priv->agent.device;
772 u8 port_num;
773 struct ib_wc mad_wc;
774 struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
775 size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
776 u16 out_mad_pkey_index = 0;
777 u16 drslid;
778 bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
779 mad_agent_priv->qp_info->port_priv->port_num);
781 if (rdma_cap_ib_switch(device) &&
782 smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
783 port_num = send_wr->port_num;
784 else
785 port_num = mad_agent_priv->agent.port_num;
788 * Directed route handling starts if the initial LID routed part of
789 * a request or the ending LID routed part of a response is empty.
790 * If we are at the start of the LID routed part, don't update the
791 * hop_ptr or hop_cnt. See section 14.2.2, Vol 1 IB spec.
793 if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
794 u32 opa_drslid;
796 if ((opa_get_smp_direction(opa_smp)
797 ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
798 OPA_LID_PERMISSIVE &&
799 opa_smi_handle_dr_smp_send(opa_smp,
800 rdma_cap_ib_switch(device),
801 port_num) == IB_SMI_DISCARD) {
802 ret = -EINVAL;
803 dev_err(&device->dev, "OPA Invalid directed route\n");
804 goto out;
806 opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
807 if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
808 opa_drslid & 0xffff0000) {
809 ret = -EINVAL;
810 dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
811 opa_drslid);
812 goto out;
814 drslid = (u16)(opa_drslid & 0x0000ffff);
816 /* Check to post send on QP or process locally */
817 if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
818 opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
819 goto out;
820 } else {
821 if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
822 IB_LID_PERMISSIVE &&
823 smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
824 IB_SMI_DISCARD) {
825 ret = -EINVAL;
826 dev_err(&device->dev, "Invalid directed route\n");
827 goto out;
829 drslid = be16_to_cpu(smp->dr_slid);
831 /* Check to post send on QP or process locally */
832 if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
833 smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
834 goto out;
837 local = kmalloc(sizeof *local, GFP_ATOMIC);
838 if (!local) {
839 ret = -ENOMEM;
840 goto out;
842 local->mad_priv = NULL;
843 local->recv_mad_agent = NULL;
844 mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
845 if (!mad_priv) {
846 ret = -ENOMEM;
847 kfree(local);
848 goto out;
851 build_smp_wc(mad_agent_priv->agent.qp,
852 send_wr->wr.wr_cqe, drslid,
853 send_wr->pkey_index,
854 send_wr->port_num, &mad_wc);
856 if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
857 mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
858 + mad_send_wr->send_buf.data_len
859 + sizeof(struct ib_grh);
862 /* No GRH for DR SMP */
863 ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
864 (const struct ib_mad_hdr *)smp, mad_size,
865 (struct ib_mad_hdr *)mad_priv->mad,
866 &mad_size, &out_mad_pkey_index);
867 switch (ret)
869 case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
870 if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
871 mad_agent_priv->agent.recv_handler) {
872 local->mad_priv = mad_priv;
873 local->recv_mad_agent = mad_agent_priv;
875 * Reference MAD agent until receive
876 * side of local completion handled
878 atomic_inc(&mad_agent_priv->refcount);
879 } else
880 kfree(mad_priv);
881 break;
882 case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
883 kfree(mad_priv);
884 break;
885 case IB_MAD_RESULT_SUCCESS:
886 /* Treat like an incoming receive MAD */
887 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
888 mad_agent_priv->agent.port_num);
889 if (port_priv) {
890 memcpy(mad_priv->mad, smp, mad_priv->mad_size);
891 recv_mad_agent = find_mad_agent(port_priv,
892 (const struct ib_mad_hdr *)mad_priv->mad);
894 if (!port_priv || !recv_mad_agent) {
896 * No receiving agent so drop packet and
897 * generate send completion.
899 kfree(mad_priv);
900 break;
902 local->mad_priv = mad_priv;
903 local->recv_mad_agent = recv_mad_agent;
904 break;
905 default:
906 kfree(mad_priv);
907 kfree(local);
908 ret = -EINVAL;
909 goto out;
912 local->mad_send_wr = mad_send_wr;
913 if (opa) {
914 local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
915 local->return_wc_byte_len = mad_size;
917 /* Reference MAD agent until send side of local completion handled */
918 atomic_inc(&mad_agent_priv->refcount);
919 /* Queue local completion to local list */
920 spin_lock_irqsave(&mad_agent_priv->lock, flags);
921 list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
922 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
923 queue_work(mad_agent_priv->qp_info->port_priv->wq,
924 &mad_agent_priv->local_work);
925 ret = 1;
926 out:
927 return ret;
930 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
932 int seg_size, pad;
934 seg_size = mad_size - hdr_len;
935 if (data_len && seg_size) {
936 pad = seg_size - data_len % seg_size;
937 return pad == seg_size ? 0 : pad;
938 } else
939 return seg_size;
942 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
944 struct ib_rmpp_segment *s, *t;
946 list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
947 list_del(&s->list);
948 kfree(s);
952 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
953 size_t mad_size, gfp_t gfp_mask)
955 struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
956 struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
957 struct ib_rmpp_segment *seg = NULL;
958 int left, seg_size, pad;
960 send_buf->seg_size = mad_size - send_buf->hdr_len;
961 send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
962 seg_size = send_buf->seg_size;
963 pad = send_wr->pad;
965 /* Allocate data segments. */
966 for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
967 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
968 if (!seg) {
969 free_send_rmpp_list(send_wr);
970 return -ENOMEM;
972 seg->num = ++send_buf->seg_count;
973 list_add_tail(&seg->list, &send_wr->rmpp_list);
976 /* Zero any padding */
977 if (pad)
978 memset(seg->data + seg_size - pad, 0, pad);
980 rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
981 agent.rmpp_version;
982 rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
983 ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
985 send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
986 struct ib_rmpp_segment, list);
987 send_wr->last_ack_seg = send_wr->cur_seg;
988 return 0;
991 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
993 return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
995 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
997 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
998 u32 remote_qpn, u16 pkey_index,
999 int rmpp_active,
1000 int hdr_len, int data_len,
1001 gfp_t gfp_mask,
1002 u8 base_version)
1004 struct ib_mad_agent_private *mad_agent_priv;
1005 struct ib_mad_send_wr_private *mad_send_wr;
1006 int pad, message_size, ret, size;
1007 void *buf;
1008 size_t mad_size;
1009 bool opa;
1011 mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
1012 agent);
1014 opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
1016 if (opa && base_version == OPA_MGMT_BASE_VERSION)
1017 mad_size = sizeof(struct opa_mad);
1018 else
1019 mad_size = sizeof(struct ib_mad);
1021 pad = get_pad_size(hdr_len, data_len, mad_size);
1022 message_size = hdr_len + data_len + pad;
1024 if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1025 if (!rmpp_active && message_size > mad_size)
1026 return ERR_PTR(-EINVAL);
1027 } else
1028 if (rmpp_active || message_size > mad_size)
1029 return ERR_PTR(-EINVAL);
1031 size = rmpp_active ? hdr_len : mad_size;
1032 buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1033 if (!buf)
1034 return ERR_PTR(-ENOMEM);
1036 mad_send_wr = buf + size;
1037 INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1038 mad_send_wr->send_buf.mad = buf;
1039 mad_send_wr->send_buf.hdr_len = hdr_len;
1040 mad_send_wr->send_buf.data_len = data_len;
1041 mad_send_wr->pad = pad;
1043 mad_send_wr->mad_agent_priv = mad_agent_priv;
1044 mad_send_wr->sg_list[0].length = hdr_len;
1045 mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1047 /* OPA MADs don't have to be the full 2048 bytes */
1048 if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1049 data_len < mad_size - hdr_len)
1050 mad_send_wr->sg_list[1].length = data_len;
1051 else
1052 mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1054 mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1056 mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1058 mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1059 mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1060 mad_send_wr->send_wr.wr.num_sge = 2;
1061 mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1062 mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1063 mad_send_wr->send_wr.remote_qpn = remote_qpn;
1064 mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1065 mad_send_wr->send_wr.pkey_index = pkey_index;
1067 if (rmpp_active) {
1068 ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1069 if (ret) {
1070 kfree(buf);
1071 return ERR_PTR(ret);
1075 mad_send_wr->send_buf.mad_agent = mad_agent;
1076 atomic_inc(&mad_agent_priv->refcount);
1077 return &mad_send_wr->send_buf;
1079 EXPORT_SYMBOL(ib_create_send_mad);
1081 int ib_get_mad_data_offset(u8 mgmt_class)
1083 if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1084 return IB_MGMT_SA_HDR;
1085 else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1086 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1087 (mgmt_class == IB_MGMT_CLASS_BIS))
1088 return IB_MGMT_DEVICE_HDR;
1089 else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1090 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1091 return IB_MGMT_VENDOR_HDR;
1092 else
1093 return IB_MGMT_MAD_HDR;
1095 EXPORT_SYMBOL(ib_get_mad_data_offset);
1097 int ib_is_mad_class_rmpp(u8 mgmt_class)
1099 if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1100 (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1101 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1102 (mgmt_class == IB_MGMT_CLASS_BIS) ||
1103 ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1104 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1105 return 1;
1106 return 0;
1108 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1110 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1112 struct ib_mad_send_wr_private *mad_send_wr;
1113 struct list_head *list;
1115 mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1116 send_buf);
1117 list = &mad_send_wr->cur_seg->list;
1119 if (mad_send_wr->cur_seg->num < seg_num) {
1120 list_for_each_entry(mad_send_wr->cur_seg, list, list)
1121 if (mad_send_wr->cur_seg->num == seg_num)
1122 break;
1123 } else if (mad_send_wr->cur_seg->num > seg_num) {
1124 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1125 if (mad_send_wr->cur_seg->num == seg_num)
1126 break;
1128 return mad_send_wr->cur_seg->data;
1130 EXPORT_SYMBOL(ib_get_rmpp_segment);
1132 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1134 if (mad_send_wr->send_buf.seg_count)
1135 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1136 mad_send_wr->seg_num);
1137 else
1138 return mad_send_wr->send_buf.mad +
1139 mad_send_wr->send_buf.hdr_len;
1142 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1144 struct ib_mad_agent_private *mad_agent_priv;
1145 struct ib_mad_send_wr_private *mad_send_wr;
1147 mad_agent_priv = container_of(send_buf->mad_agent,
1148 struct ib_mad_agent_private, agent);
1149 mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1150 send_buf);
1152 free_send_rmpp_list(mad_send_wr);
1153 kfree(send_buf->mad);
1154 deref_mad_agent(mad_agent_priv);
1156 EXPORT_SYMBOL(ib_free_send_mad);
1158 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1160 struct ib_mad_qp_info *qp_info;
1161 struct list_head *list;
1162 struct ib_send_wr *bad_send_wr;
1163 struct ib_mad_agent *mad_agent;
1164 struct ib_sge *sge;
1165 unsigned long flags;
1166 int ret;
1168 /* Set WR ID to find mad_send_wr upon completion */
1169 qp_info = mad_send_wr->mad_agent_priv->qp_info;
1170 mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1171 mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1172 mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1174 mad_agent = mad_send_wr->send_buf.mad_agent;
1175 sge = mad_send_wr->sg_list;
1176 sge[0].addr = ib_dma_map_single(mad_agent->device,
1177 mad_send_wr->send_buf.mad,
1178 sge[0].length,
1179 DMA_TO_DEVICE);
1180 if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1181 return -ENOMEM;
1183 mad_send_wr->header_mapping = sge[0].addr;
1185 sge[1].addr = ib_dma_map_single(mad_agent->device,
1186 ib_get_payload(mad_send_wr),
1187 sge[1].length,
1188 DMA_TO_DEVICE);
1189 if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1190 ib_dma_unmap_single(mad_agent->device,
1191 mad_send_wr->header_mapping,
1192 sge[0].length, DMA_TO_DEVICE);
1193 return -ENOMEM;
1195 mad_send_wr->payload_mapping = sge[1].addr;
1197 spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1198 if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1199 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1200 &bad_send_wr);
1201 list = &qp_info->send_queue.list;
1202 } else {
1203 ret = 0;
1204 list = &qp_info->overflow_list;
1207 if (!ret) {
1208 qp_info->send_queue.count++;
1209 list_add_tail(&mad_send_wr->mad_list.list, list);
1211 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1212 if (ret) {
1213 ib_dma_unmap_single(mad_agent->device,
1214 mad_send_wr->header_mapping,
1215 sge[0].length, DMA_TO_DEVICE);
1216 ib_dma_unmap_single(mad_agent->device,
1217 mad_send_wr->payload_mapping,
1218 sge[1].length, DMA_TO_DEVICE);
1220 return ret;
1224 * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1225 * with the registered client
1227 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1228 struct ib_mad_send_buf **bad_send_buf)
1230 struct ib_mad_agent_private *mad_agent_priv;
1231 struct ib_mad_send_buf *next_send_buf;
1232 struct ib_mad_send_wr_private *mad_send_wr;
1233 unsigned long flags;
1234 int ret = -EINVAL;
1236 /* Walk list of send WRs and post each on send list */
1237 for (; send_buf; send_buf = next_send_buf) {
1238 mad_send_wr = container_of(send_buf,
1239 struct ib_mad_send_wr_private,
1240 send_buf);
1241 mad_agent_priv = mad_send_wr->mad_agent_priv;
1243 ret = ib_mad_enforce_security(mad_agent_priv,
1244 mad_send_wr->send_wr.pkey_index);
1245 if (ret)
1246 goto error;
1248 if (!send_buf->mad_agent->send_handler ||
1249 (send_buf->timeout_ms &&
1250 !send_buf->mad_agent->recv_handler)) {
1251 ret = -EINVAL;
1252 goto error;
1255 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1256 if (mad_agent_priv->agent.rmpp_version) {
1257 ret = -EINVAL;
1258 goto error;
1263 * Save pointer to next work request to post in case the
1264 * current one completes, and the user modifies the work
1265 * request associated with the completion
1267 next_send_buf = send_buf->next;
1268 mad_send_wr->send_wr.ah = send_buf->ah;
1270 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1271 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1272 ret = handle_outgoing_dr_smp(mad_agent_priv,
1273 mad_send_wr);
1274 if (ret < 0) /* error */
1275 goto error;
1276 else if (ret == 1) /* locally consumed */
1277 continue;
1280 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1281 /* Timeout will be updated after send completes */
1282 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1283 mad_send_wr->max_retries = send_buf->retries;
1284 mad_send_wr->retries_left = send_buf->retries;
1285 send_buf->retries = 0;
1286 /* Reference for work request to QP + response */
1287 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1288 mad_send_wr->status = IB_WC_SUCCESS;
1290 /* Reference MAD agent until send completes */
1291 atomic_inc(&mad_agent_priv->refcount);
1292 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1293 list_add_tail(&mad_send_wr->agent_list,
1294 &mad_agent_priv->send_list);
1295 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1297 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1298 ret = ib_send_rmpp_mad(mad_send_wr);
1299 if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1300 ret = ib_send_mad(mad_send_wr);
1301 } else
1302 ret = ib_send_mad(mad_send_wr);
1303 if (ret < 0) {
1304 /* Fail send request */
1305 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1306 list_del(&mad_send_wr->agent_list);
1307 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1308 atomic_dec(&mad_agent_priv->refcount);
1309 goto error;
1312 return 0;
1313 error:
1314 if (bad_send_buf)
1315 *bad_send_buf = send_buf;
1316 return ret;
1318 EXPORT_SYMBOL(ib_post_send_mad);
1321 * ib_free_recv_mad - Returns data buffers used to receive
1322 * a MAD to the access layer
1324 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1326 struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1327 struct ib_mad_private_header *mad_priv_hdr;
1328 struct ib_mad_private *priv;
1329 struct list_head free_list;
1331 INIT_LIST_HEAD(&free_list);
1332 list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1334 list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1335 &free_list, list) {
1336 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1337 recv_buf);
1338 mad_priv_hdr = container_of(mad_recv_wc,
1339 struct ib_mad_private_header,
1340 recv_wc);
1341 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1342 header);
1343 kfree(priv);
1346 EXPORT_SYMBOL(ib_free_recv_mad);
1348 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1349 u8 rmpp_version,
1350 ib_mad_send_handler send_handler,
1351 ib_mad_recv_handler recv_handler,
1352 void *context)
1354 return ERR_PTR(-EINVAL); /* XXX: for now */
1356 EXPORT_SYMBOL(ib_redirect_mad_qp);
1358 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1359 struct ib_wc *wc)
1361 dev_err(&mad_agent->device->dev,
1362 "ib_process_mad_wc() not implemented yet\n");
1363 return 0;
1365 EXPORT_SYMBOL(ib_process_mad_wc);
1367 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1368 struct ib_mad_reg_req *mad_reg_req)
1370 int i;
1372 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1373 if ((*method)->agent[i]) {
1374 pr_err("Method %d already in use\n", i);
1375 return -EINVAL;
1378 return 0;
1381 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1383 /* Allocate management method table */
1384 *method = kzalloc(sizeof **method, GFP_ATOMIC);
1385 return (*method) ? 0 : (-ENOMEM);
1389 * Check to see if there are any methods still in use
1391 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1393 int i;
1395 for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1396 if (method->agent[i])
1397 return 1;
1398 return 0;
1402 * Check to see if there are any method tables for this class still in use
1404 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1406 int i;
1408 for (i = 0; i < MAX_MGMT_CLASS; i++)
1409 if (class->method_table[i])
1410 return 1;
1411 return 0;
1414 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1416 int i;
1418 for (i = 0; i < MAX_MGMT_OUI; i++)
1419 if (vendor_class->method_table[i])
1420 return 1;
1421 return 0;
1424 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1425 const char *oui)
1427 int i;
1429 for (i = 0; i < MAX_MGMT_OUI; i++)
1430 /* Is there matching OUI for this vendor class ? */
1431 if (!memcmp(vendor_class->oui[i], oui, 3))
1432 return i;
1434 return -1;
1437 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1439 int i;
1441 for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1442 if (vendor->vendor_class[i])
1443 return 1;
1445 return 0;
1448 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1449 struct ib_mad_agent_private *agent)
1451 int i;
1453 /* Remove any methods for this mad agent */
1454 for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1455 if (method->agent[i] == agent) {
1456 method->agent[i] = NULL;
1461 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1462 struct ib_mad_agent_private *agent_priv,
1463 u8 mgmt_class)
1465 struct ib_mad_port_private *port_priv;
1466 struct ib_mad_mgmt_class_table **class;
1467 struct ib_mad_mgmt_method_table **method;
1468 int i, ret;
1470 port_priv = agent_priv->qp_info->port_priv;
1471 class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1472 if (!*class) {
1473 /* Allocate management class table for "new" class version */
1474 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1475 if (!*class) {
1476 ret = -ENOMEM;
1477 goto error1;
1480 /* Allocate method table for this management class */
1481 method = &(*class)->method_table[mgmt_class];
1482 if ((ret = allocate_method_table(method)))
1483 goto error2;
1484 } else {
1485 method = &(*class)->method_table[mgmt_class];
1486 if (!*method) {
1487 /* Allocate method table for this management class */
1488 if ((ret = allocate_method_table(method)))
1489 goto error1;
1493 /* Now, make sure methods are not already in use */
1494 if (method_in_use(method, mad_reg_req))
1495 goto error3;
1497 /* Finally, add in methods being registered */
1498 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1499 (*method)->agent[i] = agent_priv;
1501 return 0;
1503 error3:
1504 /* Remove any methods for this mad agent */
1505 remove_methods_mad_agent(*method, agent_priv);
1506 /* Now, check to see if there are any methods in use */
1507 if (!check_method_table(*method)) {
1508 /* If not, release management method table */
1509 kfree(*method);
1510 *method = NULL;
1512 ret = -EINVAL;
1513 goto error1;
1514 error2:
1515 kfree(*class);
1516 *class = NULL;
1517 error1:
1518 return ret;
1521 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1522 struct ib_mad_agent_private *agent_priv)
1524 struct ib_mad_port_private *port_priv;
1525 struct ib_mad_mgmt_vendor_class_table **vendor_table;
1526 struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1527 struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1528 struct ib_mad_mgmt_method_table **method;
1529 int i, ret = -ENOMEM;
1530 u8 vclass;
1532 /* "New" vendor (with OUI) class */
1533 vclass = vendor_class_index(mad_reg_req->mgmt_class);
1534 port_priv = agent_priv->qp_info->port_priv;
1535 vendor_table = &port_priv->version[
1536 mad_reg_req->mgmt_class_version].vendor;
1537 if (!*vendor_table) {
1538 /* Allocate mgmt vendor class table for "new" class version */
1539 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1540 if (!vendor)
1541 goto error1;
1543 *vendor_table = vendor;
1545 if (!(*vendor_table)->vendor_class[vclass]) {
1546 /* Allocate table for this management vendor class */
1547 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1548 if (!vendor_class)
1549 goto error2;
1551 (*vendor_table)->vendor_class[vclass] = vendor_class;
1553 for (i = 0; i < MAX_MGMT_OUI; i++) {
1554 /* Is there matching OUI for this vendor class ? */
1555 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1556 mad_reg_req->oui, 3)) {
1557 method = &(*vendor_table)->vendor_class[
1558 vclass]->method_table[i];
1559 if (!*method)
1560 goto error3;
1561 goto check_in_use;
1564 for (i = 0; i < MAX_MGMT_OUI; i++) {
1565 /* OUI slot available ? */
1566 if (!is_vendor_oui((*vendor_table)->vendor_class[
1567 vclass]->oui[i])) {
1568 method = &(*vendor_table)->vendor_class[
1569 vclass]->method_table[i];
1570 /* Allocate method table for this OUI */
1571 if (!*method) {
1572 ret = allocate_method_table(method);
1573 if (ret)
1574 goto error3;
1576 memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1577 mad_reg_req->oui, 3);
1578 goto check_in_use;
1581 dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1582 goto error3;
1584 check_in_use:
1585 /* Now, make sure methods are not already in use */
1586 if (method_in_use(method, mad_reg_req))
1587 goto error4;
1589 /* Finally, add in methods being registered */
1590 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1591 (*method)->agent[i] = agent_priv;
1593 return 0;
1595 error4:
1596 /* Remove any methods for this mad agent */
1597 remove_methods_mad_agent(*method, agent_priv);
1598 /* Now, check to see if there are any methods in use */
1599 if (!check_method_table(*method)) {
1600 /* If not, release management method table */
1601 kfree(*method);
1602 *method = NULL;
1604 ret = -EINVAL;
1605 error3:
1606 if (vendor_class) {
1607 (*vendor_table)->vendor_class[vclass] = NULL;
1608 kfree(vendor_class);
1610 error2:
1611 if (vendor) {
1612 *vendor_table = NULL;
1613 kfree(vendor);
1615 error1:
1616 return ret;
1619 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1621 struct ib_mad_port_private *port_priv;
1622 struct ib_mad_mgmt_class_table *class;
1623 struct ib_mad_mgmt_method_table *method;
1624 struct ib_mad_mgmt_vendor_class_table *vendor;
1625 struct ib_mad_mgmt_vendor_class *vendor_class;
1626 int index;
1627 u8 mgmt_class;
1630 * Was MAD registration request supplied
1631 * with original registration ?
1633 if (!agent_priv->reg_req) {
1634 goto out;
1637 port_priv = agent_priv->qp_info->port_priv;
1638 mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1639 class = port_priv->version[
1640 agent_priv->reg_req->mgmt_class_version].class;
1641 if (!class)
1642 goto vendor_check;
1644 method = class->method_table[mgmt_class];
1645 if (method) {
1646 /* Remove any methods for this mad agent */
1647 remove_methods_mad_agent(method, agent_priv);
1648 /* Now, check to see if there are any methods still in use */
1649 if (!check_method_table(method)) {
1650 /* If not, release management method table */
1651 kfree(method);
1652 class->method_table[mgmt_class] = NULL;
1653 /* Any management classes left ? */
1654 if (!check_class_table(class)) {
1655 /* If not, release management class table */
1656 kfree(class);
1657 port_priv->version[
1658 agent_priv->reg_req->
1659 mgmt_class_version].class = NULL;
1664 vendor_check:
1665 if (!is_vendor_class(mgmt_class))
1666 goto out;
1668 /* normalize mgmt_class to vendor range 2 */
1669 mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1670 vendor = port_priv->version[
1671 agent_priv->reg_req->mgmt_class_version].vendor;
1673 if (!vendor)
1674 goto out;
1676 vendor_class = vendor->vendor_class[mgmt_class];
1677 if (vendor_class) {
1678 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1679 if (index < 0)
1680 goto out;
1681 method = vendor_class->method_table[index];
1682 if (method) {
1683 /* Remove any methods for this mad agent */
1684 remove_methods_mad_agent(method, agent_priv);
1686 * Now, check to see if there are
1687 * any methods still in use
1689 if (!check_method_table(method)) {
1690 /* If not, release management method table */
1691 kfree(method);
1692 vendor_class->method_table[index] = NULL;
1693 memset(vendor_class->oui[index], 0, 3);
1694 /* Any OUIs left ? */
1695 if (!check_vendor_class(vendor_class)) {
1696 /* If not, release vendor class table */
1697 kfree(vendor_class);
1698 vendor->vendor_class[mgmt_class] = NULL;
1699 /* Any other vendor classes left ? */
1700 if (!check_vendor_table(vendor)) {
1701 kfree(vendor);
1702 port_priv->version[
1703 agent_priv->reg_req->
1704 mgmt_class_version].
1705 vendor = NULL;
1712 out:
1713 return;
1716 static struct ib_mad_agent_private *
1717 find_mad_agent(struct ib_mad_port_private *port_priv,
1718 const struct ib_mad_hdr *mad_hdr)
1720 struct ib_mad_agent_private *mad_agent = NULL;
1721 unsigned long flags;
1723 spin_lock_irqsave(&port_priv->reg_lock, flags);
1724 if (ib_response_mad(mad_hdr)) {
1725 u32 hi_tid;
1726 struct ib_mad_agent_private *entry;
1729 * Routing is based on high 32 bits of transaction ID
1730 * of MAD.
1732 hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1733 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1734 if (entry->agent.hi_tid == hi_tid) {
1735 mad_agent = entry;
1736 break;
1739 } else {
1740 struct ib_mad_mgmt_class_table *class;
1741 struct ib_mad_mgmt_method_table *method;
1742 struct ib_mad_mgmt_vendor_class_table *vendor;
1743 struct ib_mad_mgmt_vendor_class *vendor_class;
1744 const struct ib_vendor_mad *vendor_mad;
1745 int index;
1748 * Routing is based on version, class, and method
1749 * For "newer" vendor MADs, also based on OUI
1751 if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1752 goto out;
1753 if (!is_vendor_class(mad_hdr->mgmt_class)) {
1754 class = port_priv->version[
1755 mad_hdr->class_version].class;
1756 if (!class)
1757 goto out;
1758 if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1759 ARRAY_SIZE(class->method_table))
1760 goto out;
1761 method = class->method_table[convert_mgmt_class(
1762 mad_hdr->mgmt_class)];
1763 if (method)
1764 mad_agent = method->agent[mad_hdr->method &
1765 ~IB_MGMT_METHOD_RESP];
1766 } else {
1767 vendor = port_priv->version[
1768 mad_hdr->class_version].vendor;
1769 if (!vendor)
1770 goto out;
1771 vendor_class = vendor->vendor_class[vendor_class_index(
1772 mad_hdr->mgmt_class)];
1773 if (!vendor_class)
1774 goto out;
1775 /* Find matching OUI */
1776 vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1777 index = find_vendor_oui(vendor_class, vendor_mad->oui);
1778 if (index == -1)
1779 goto out;
1780 method = vendor_class->method_table[index];
1781 if (method) {
1782 mad_agent = method->agent[mad_hdr->method &
1783 ~IB_MGMT_METHOD_RESP];
1788 if (mad_agent) {
1789 if (mad_agent->agent.recv_handler)
1790 atomic_inc(&mad_agent->refcount);
1791 else {
1792 dev_notice(&port_priv->device->dev,
1793 "No receive handler for client %p on port %d\n",
1794 &mad_agent->agent, port_priv->port_num);
1795 mad_agent = NULL;
1798 out:
1799 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1801 return mad_agent;
1804 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1805 const struct ib_mad_qp_info *qp_info,
1806 bool opa)
1808 int valid = 0;
1809 u32 qp_num = qp_info->qp->qp_num;
1811 /* Make sure MAD base version is understood */
1812 if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1813 (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1814 pr_err("MAD received with unsupported base version %d %s\n",
1815 mad_hdr->base_version, opa ? "(opa)" : "");
1816 goto out;
1819 /* Filter SMI packets sent to other than QP0 */
1820 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1821 (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1822 if (qp_num == 0)
1823 valid = 1;
1824 } else {
1825 /* CM attributes other than ClassPortInfo only use Send method */
1826 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1827 (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1828 (mad_hdr->method != IB_MGMT_METHOD_SEND))
1829 goto out;
1830 /* Filter GSI packets sent to QP0 */
1831 if (qp_num != 0)
1832 valid = 1;
1835 out:
1836 return valid;
1839 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1840 const struct ib_mad_hdr *mad_hdr)
1842 struct ib_rmpp_mad *rmpp_mad;
1844 rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1845 return !mad_agent_priv->agent.rmpp_version ||
1846 !ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1847 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1848 IB_MGMT_RMPP_FLAG_ACTIVE) ||
1849 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1852 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1853 const struct ib_mad_recv_wc *rwc)
1855 return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1856 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1859 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1860 const struct ib_mad_send_wr_private *wr,
1861 const struct ib_mad_recv_wc *rwc )
1863 struct rdma_ah_attr attr;
1864 u8 send_resp, rcv_resp;
1865 union ib_gid sgid;
1866 struct ib_device *device = mad_agent_priv->agent.device;
1867 u8 port_num = mad_agent_priv->agent.port_num;
1868 u8 lmc;
1869 bool has_grh;
1871 send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1872 rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1874 if (send_resp == rcv_resp)
1875 /* both requests, or both responses. GIDs different */
1876 return 0;
1878 if (rdma_query_ah(wr->send_buf.ah, &attr))
1879 /* Assume not equal, to avoid false positives. */
1880 return 0;
1882 has_grh = !!(rdma_ah_get_ah_flags(&attr) & IB_AH_GRH);
1883 if (has_grh != !!(rwc->wc->wc_flags & IB_WC_GRH))
1884 /* one has GID, other does not. Assume different */
1885 return 0;
1887 if (!send_resp && rcv_resp) {
1888 /* is request/response. */
1889 if (!has_grh) {
1890 if (ib_get_cached_lmc(device, port_num, &lmc))
1891 return 0;
1892 return (!lmc || !((rdma_ah_get_path_bits(&attr) ^
1893 rwc->wc->dlid_path_bits) &
1894 ((1 << lmc) - 1)));
1895 } else {
1896 const struct ib_global_route *grh =
1897 rdma_ah_read_grh(&attr);
1899 if (ib_get_cached_gid(device, port_num,
1900 grh->sgid_index, &sgid, NULL))
1901 return 0;
1902 return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1903 16);
1907 if (!has_grh)
1908 return rdma_ah_get_dlid(&attr) == rwc->wc->slid;
1909 else
1910 return !memcmp(rdma_ah_read_grh(&attr)->dgid.raw,
1911 rwc->recv_buf.grh->sgid.raw,
1912 16);
1915 static inline int is_direct(u8 class)
1917 return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1920 struct ib_mad_send_wr_private*
1921 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1922 const struct ib_mad_recv_wc *wc)
1924 struct ib_mad_send_wr_private *wr;
1925 const struct ib_mad_hdr *mad_hdr;
1927 mad_hdr = &wc->recv_buf.mad->mad_hdr;
1929 list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1930 if ((wr->tid == mad_hdr->tid) &&
1931 rcv_has_same_class(wr, wc) &&
1933 * Don't check GID for direct routed MADs.
1934 * These might have permissive LIDs.
1936 (is_direct(mad_hdr->mgmt_class) ||
1937 rcv_has_same_gid(mad_agent_priv, wr, wc)))
1938 return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1942 * It's possible to receive the response before we've
1943 * been notified that the send has completed
1945 list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1946 if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1947 wr->tid == mad_hdr->tid &&
1948 wr->timeout &&
1949 rcv_has_same_class(wr, wc) &&
1951 * Don't check GID for direct routed MADs.
1952 * These might have permissive LIDs.
1954 (is_direct(mad_hdr->mgmt_class) ||
1955 rcv_has_same_gid(mad_agent_priv, wr, wc)))
1956 /* Verify request has not been canceled */
1957 return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1959 return NULL;
1962 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1964 mad_send_wr->timeout = 0;
1965 if (mad_send_wr->refcount == 1)
1966 list_move_tail(&mad_send_wr->agent_list,
1967 &mad_send_wr->mad_agent_priv->done_list);
1970 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1971 struct ib_mad_recv_wc *mad_recv_wc)
1973 struct ib_mad_send_wr_private *mad_send_wr;
1974 struct ib_mad_send_wc mad_send_wc;
1975 unsigned long flags;
1976 int ret;
1978 INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1979 ret = ib_mad_enforce_security(mad_agent_priv,
1980 mad_recv_wc->wc->pkey_index);
1981 if (ret) {
1982 ib_free_recv_mad(mad_recv_wc);
1983 deref_mad_agent(mad_agent_priv);
1984 return;
1987 list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1988 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1989 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1990 mad_recv_wc);
1991 if (!mad_recv_wc) {
1992 deref_mad_agent(mad_agent_priv);
1993 return;
1997 /* Complete corresponding request */
1998 if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1999 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2000 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
2001 if (!mad_send_wr) {
2002 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2003 if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
2004 && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
2005 && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
2006 & IB_MGMT_RMPP_FLAG_ACTIVE)) {
2007 /* user rmpp is in effect
2008 * and this is an active RMPP MAD
2010 mad_agent_priv->agent.recv_handler(
2011 &mad_agent_priv->agent, NULL,
2012 mad_recv_wc);
2013 atomic_dec(&mad_agent_priv->refcount);
2014 } else {
2015 /* not user rmpp, revert to normal behavior and
2016 * drop the mad */
2017 ib_free_recv_mad(mad_recv_wc);
2018 deref_mad_agent(mad_agent_priv);
2019 return;
2021 } else {
2022 ib_mark_mad_done(mad_send_wr);
2023 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2025 /* Defined behavior is to complete response before request */
2026 mad_agent_priv->agent.recv_handler(
2027 &mad_agent_priv->agent,
2028 &mad_send_wr->send_buf,
2029 mad_recv_wc);
2030 atomic_dec(&mad_agent_priv->refcount);
2032 mad_send_wc.status = IB_WC_SUCCESS;
2033 mad_send_wc.vendor_err = 0;
2034 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2035 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2037 } else {
2038 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2039 mad_recv_wc);
2040 deref_mad_agent(mad_agent_priv);
2043 return;
2046 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2047 const struct ib_mad_qp_info *qp_info,
2048 const struct ib_wc *wc,
2049 int port_num,
2050 struct ib_mad_private *recv,
2051 struct ib_mad_private *response)
2053 enum smi_forward_action retsmi;
2054 struct ib_smp *smp = (struct ib_smp *)recv->mad;
2056 if (smi_handle_dr_smp_recv(smp,
2057 rdma_cap_ib_switch(port_priv->device),
2058 port_num,
2059 port_priv->device->phys_port_cnt) ==
2060 IB_SMI_DISCARD)
2061 return IB_SMI_DISCARD;
2063 retsmi = smi_check_forward_dr_smp(smp);
2064 if (retsmi == IB_SMI_LOCAL)
2065 return IB_SMI_HANDLE;
2067 if (retsmi == IB_SMI_SEND) { /* don't forward */
2068 if (smi_handle_dr_smp_send(smp,
2069 rdma_cap_ib_switch(port_priv->device),
2070 port_num) == IB_SMI_DISCARD)
2071 return IB_SMI_DISCARD;
2073 if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2074 return IB_SMI_DISCARD;
2075 } else if (rdma_cap_ib_switch(port_priv->device)) {
2076 /* forward case for switches */
2077 memcpy(response, recv, mad_priv_size(response));
2078 response->header.recv_wc.wc = &response->header.wc;
2079 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2080 response->header.recv_wc.recv_buf.grh = &response->grh;
2082 agent_send_response((const struct ib_mad_hdr *)response->mad,
2083 &response->grh, wc,
2084 port_priv->device,
2085 smi_get_fwd_port(smp),
2086 qp_info->qp->qp_num,
2087 response->mad_size,
2088 false);
2090 return IB_SMI_DISCARD;
2092 return IB_SMI_HANDLE;
2095 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2096 struct ib_mad_private *response,
2097 size_t *resp_len, bool opa)
2099 const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2100 struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2102 if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2103 recv_hdr->method == IB_MGMT_METHOD_SET) {
2104 memcpy(response, recv, mad_priv_size(response));
2105 response->header.recv_wc.wc = &response->header.wc;
2106 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2107 response->header.recv_wc.recv_buf.grh = &response->grh;
2108 resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2109 resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2110 if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2111 resp_hdr->status |= IB_SMP_DIRECTION;
2113 if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2114 if (recv_hdr->mgmt_class ==
2115 IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2116 recv_hdr->mgmt_class ==
2117 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2118 *resp_len = opa_get_smp_header_size(
2119 (struct opa_smp *)recv->mad);
2120 else
2121 *resp_len = sizeof(struct ib_mad_hdr);
2124 return true;
2125 } else {
2126 return false;
2130 static enum smi_action
2131 handle_opa_smi(struct ib_mad_port_private *port_priv,
2132 struct ib_mad_qp_info *qp_info,
2133 struct ib_wc *wc,
2134 int port_num,
2135 struct ib_mad_private *recv,
2136 struct ib_mad_private *response)
2138 enum smi_forward_action retsmi;
2139 struct opa_smp *smp = (struct opa_smp *)recv->mad;
2141 if (opa_smi_handle_dr_smp_recv(smp,
2142 rdma_cap_ib_switch(port_priv->device),
2143 port_num,
2144 port_priv->device->phys_port_cnt) ==
2145 IB_SMI_DISCARD)
2146 return IB_SMI_DISCARD;
2148 retsmi = opa_smi_check_forward_dr_smp(smp);
2149 if (retsmi == IB_SMI_LOCAL)
2150 return IB_SMI_HANDLE;
2152 if (retsmi == IB_SMI_SEND) { /* don't forward */
2153 if (opa_smi_handle_dr_smp_send(smp,
2154 rdma_cap_ib_switch(port_priv->device),
2155 port_num) == IB_SMI_DISCARD)
2156 return IB_SMI_DISCARD;
2158 if (opa_smi_check_local_smp(smp, port_priv->device) ==
2159 IB_SMI_DISCARD)
2160 return IB_SMI_DISCARD;
2162 } else if (rdma_cap_ib_switch(port_priv->device)) {
2163 /* forward case for switches */
2164 memcpy(response, recv, mad_priv_size(response));
2165 response->header.recv_wc.wc = &response->header.wc;
2166 response->header.recv_wc.recv_buf.opa_mad =
2167 (struct opa_mad *)response->mad;
2168 response->header.recv_wc.recv_buf.grh = &response->grh;
2170 agent_send_response((const struct ib_mad_hdr *)response->mad,
2171 &response->grh, wc,
2172 port_priv->device,
2173 opa_smi_get_fwd_port(smp),
2174 qp_info->qp->qp_num,
2175 recv->header.wc.byte_len,
2176 true);
2178 return IB_SMI_DISCARD;
2181 return IB_SMI_HANDLE;
2184 static enum smi_action
2185 handle_smi(struct ib_mad_port_private *port_priv,
2186 struct ib_mad_qp_info *qp_info,
2187 struct ib_wc *wc,
2188 int port_num,
2189 struct ib_mad_private *recv,
2190 struct ib_mad_private *response,
2191 bool opa)
2193 struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2195 if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2196 mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2197 return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2198 response);
2200 return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2203 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2205 struct ib_mad_port_private *port_priv = cq->cq_context;
2206 struct ib_mad_list_head *mad_list =
2207 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2208 struct ib_mad_qp_info *qp_info;
2209 struct ib_mad_private_header *mad_priv_hdr;
2210 struct ib_mad_private *recv, *response = NULL;
2211 struct ib_mad_agent_private *mad_agent;
2212 int port_num;
2213 int ret = IB_MAD_RESULT_SUCCESS;
2214 size_t mad_size;
2215 u16 resp_mad_pkey_index = 0;
2216 bool opa;
2218 if (list_empty_careful(&port_priv->port_list))
2219 return;
2221 if (wc->status != IB_WC_SUCCESS) {
2223 * Receive errors indicate that the QP has entered the error
2224 * state - error handling/shutdown code will cleanup
2226 return;
2229 qp_info = mad_list->mad_queue->qp_info;
2230 dequeue_mad(mad_list);
2232 opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2233 qp_info->port_priv->port_num);
2235 mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2236 mad_list);
2237 recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2238 ib_dma_unmap_single(port_priv->device,
2239 recv->header.mapping,
2240 mad_priv_dma_size(recv),
2241 DMA_FROM_DEVICE);
2243 /* Setup MAD receive work completion from "normal" work completion */
2244 recv->header.wc = *wc;
2245 recv->header.recv_wc.wc = &recv->header.wc;
2247 if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2248 recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2249 recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2250 } else {
2251 recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2252 recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2255 recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2256 recv->header.recv_wc.recv_buf.grh = &recv->grh;
2258 if (atomic_read(&qp_info->snoop_count))
2259 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2261 /* Validate MAD */
2262 if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2263 goto out;
2265 mad_size = recv->mad_size;
2266 response = alloc_mad_private(mad_size, GFP_KERNEL);
2267 if (!response)
2268 goto out;
2270 if (rdma_cap_ib_switch(port_priv->device))
2271 port_num = wc->port_num;
2272 else
2273 port_num = port_priv->port_num;
2275 if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2276 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2277 if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2278 response, opa)
2279 == IB_SMI_DISCARD)
2280 goto out;
2283 /* Give driver "right of first refusal" on incoming MAD */
2284 if (port_priv->device->process_mad) {
2285 ret = port_priv->device->process_mad(port_priv->device, 0,
2286 port_priv->port_num,
2287 wc, &recv->grh,
2288 (const struct ib_mad_hdr *)recv->mad,
2289 recv->mad_size,
2290 (struct ib_mad_hdr *)response->mad,
2291 &mad_size, &resp_mad_pkey_index);
2293 if (opa)
2294 wc->pkey_index = resp_mad_pkey_index;
2296 if (ret & IB_MAD_RESULT_SUCCESS) {
2297 if (ret & IB_MAD_RESULT_CONSUMED)
2298 goto out;
2299 if (ret & IB_MAD_RESULT_REPLY) {
2300 agent_send_response((const struct ib_mad_hdr *)response->mad,
2301 &recv->grh, wc,
2302 port_priv->device,
2303 port_num,
2304 qp_info->qp->qp_num,
2305 mad_size, opa);
2306 goto out;
2311 mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2312 if (mad_agent) {
2313 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2315 * recv is freed up in error cases in ib_mad_complete_recv
2316 * or via recv_handler in ib_mad_complete_recv()
2318 recv = NULL;
2319 } else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2320 generate_unmatched_resp(recv, response, &mad_size, opa)) {
2321 agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2322 port_priv->device, port_num,
2323 qp_info->qp->qp_num, mad_size, opa);
2326 out:
2327 /* Post another receive request for this QP */
2328 if (response) {
2329 ib_mad_post_receive_mads(qp_info, response);
2330 kfree(recv);
2331 } else
2332 ib_mad_post_receive_mads(qp_info, recv);
2335 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2337 struct ib_mad_send_wr_private *mad_send_wr;
2338 unsigned long delay;
2340 if (list_empty(&mad_agent_priv->wait_list)) {
2341 cancel_delayed_work(&mad_agent_priv->timed_work);
2342 } else {
2343 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2344 struct ib_mad_send_wr_private,
2345 agent_list);
2347 if (time_after(mad_agent_priv->timeout,
2348 mad_send_wr->timeout)) {
2349 mad_agent_priv->timeout = mad_send_wr->timeout;
2350 delay = mad_send_wr->timeout - jiffies;
2351 if ((long)delay <= 0)
2352 delay = 1;
2353 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2354 &mad_agent_priv->timed_work, delay);
2359 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2361 struct ib_mad_agent_private *mad_agent_priv;
2362 struct ib_mad_send_wr_private *temp_mad_send_wr;
2363 struct list_head *list_item;
2364 unsigned long delay;
2366 mad_agent_priv = mad_send_wr->mad_agent_priv;
2367 list_del(&mad_send_wr->agent_list);
2369 delay = mad_send_wr->timeout;
2370 mad_send_wr->timeout += jiffies;
2372 if (delay) {
2373 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2374 temp_mad_send_wr = list_entry(list_item,
2375 struct ib_mad_send_wr_private,
2376 agent_list);
2377 if (time_after(mad_send_wr->timeout,
2378 temp_mad_send_wr->timeout))
2379 break;
2382 else
2383 list_item = &mad_agent_priv->wait_list;
2384 list_add(&mad_send_wr->agent_list, list_item);
2386 /* Reschedule a work item if we have a shorter timeout */
2387 if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2388 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2389 &mad_agent_priv->timed_work, delay);
2392 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2393 int timeout_ms)
2395 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2396 wait_for_response(mad_send_wr);
2400 * Process a send work completion
2402 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2403 struct ib_mad_send_wc *mad_send_wc)
2405 struct ib_mad_agent_private *mad_agent_priv;
2406 unsigned long flags;
2407 int ret;
2409 mad_agent_priv = mad_send_wr->mad_agent_priv;
2410 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2411 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2412 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2413 if (ret == IB_RMPP_RESULT_CONSUMED)
2414 goto done;
2415 } else
2416 ret = IB_RMPP_RESULT_UNHANDLED;
2418 if (mad_send_wc->status != IB_WC_SUCCESS &&
2419 mad_send_wr->status == IB_WC_SUCCESS) {
2420 mad_send_wr->status = mad_send_wc->status;
2421 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2424 if (--mad_send_wr->refcount > 0) {
2425 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2426 mad_send_wr->status == IB_WC_SUCCESS) {
2427 wait_for_response(mad_send_wr);
2429 goto done;
2432 /* Remove send from MAD agent and notify client of completion */
2433 list_del(&mad_send_wr->agent_list);
2434 adjust_timeout(mad_agent_priv);
2435 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2437 if (mad_send_wr->status != IB_WC_SUCCESS )
2438 mad_send_wc->status = mad_send_wr->status;
2439 if (ret == IB_RMPP_RESULT_INTERNAL)
2440 ib_rmpp_send_handler(mad_send_wc);
2441 else
2442 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2443 mad_send_wc);
2445 /* Release reference on agent taken when sending */
2446 deref_mad_agent(mad_agent_priv);
2447 return;
2448 done:
2449 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2452 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2454 struct ib_mad_port_private *port_priv = cq->cq_context;
2455 struct ib_mad_list_head *mad_list =
2456 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2457 struct ib_mad_send_wr_private *mad_send_wr, *queued_send_wr;
2458 struct ib_mad_qp_info *qp_info;
2459 struct ib_mad_queue *send_queue;
2460 struct ib_send_wr *bad_send_wr;
2461 struct ib_mad_send_wc mad_send_wc;
2462 unsigned long flags;
2463 int ret;
2465 if (list_empty_careful(&port_priv->port_list))
2466 return;
2468 if (wc->status != IB_WC_SUCCESS) {
2469 if (!ib_mad_send_error(port_priv, wc))
2470 return;
2473 mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2474 mad_list);
2475 send_queue = mad_list->mad_queue;
2476 qp_info = send_queue->qp_info;
2478 retry:
2479 ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2480 mad_send_wr->header_mapping,
2481 mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2482 ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2483 mad_send_wr->payload_mapping,
2484 mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2485 queued_send_wr = NULL;
2486 spin_lock_irqsave(&send_queue->lock, flags);
2487 list_del(&mad_list->list);
2489 /* Move queued send to the send queue */
2490 if (send_queue->count-- > send_queue->max_active) {
2491 mad_list = container_of(qp_info->overflow_list.next,
2492 struct ib_mad_list_head, list);
2493 queued_send_wr = container_of(mad_list,
2494 struct ib_mad_send_wr_private,
2495 mad_list);
2496 list_move_tail(&mad_list->list, &send_queue->list);
2498 spin_unlock_irqrestore(&send_queue->lock, flags);
2500 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2501 mad_send_wc.status = wc->status;
2502 mad_send_wc.vendor_err = wc->vendor_err;
2503 if (atomic_read(&qp_info->snoop_count))
2504 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2505 IB_MAD_SNOOP_SEND_COMPLETIONS);
2506 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2508 if (queued_send_wr) {
2509 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2510 &bad_send_wr);
2511 if (ret) {
2512 dev_err(&port_priv->device->dev,
2513 "ib_post_send failed: %d\n", ret);
2514 mad_send_wr = queued_send_wr;
2515 wc->status = IB_WC_LOC_QP_OP_ERR;
2516 goto retry;
2521 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2523 struct ib_mad_send_wr_private *mad_send_wr;
2524 struct ib_mad_list_head *mad_list;
2525 unsigned long flags;
2527 spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2528 list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2529 mad_send_wr = container_of(mad_list,
2530 struct ib_mad_send_wr_private,
2531 mad_list);
2532 mad_send_wr->retry = 1;
2534 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2537 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2538 struct ib_wc *wc)
2540 struct ib_mad_list_head *mad_list =
2541 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2542 struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2543 struct ib_mad_send_wr_private *mad_send_wr;
2544 int ret;
2547 * Send errors will transition the QP to SQE - move
2548 * QP to RTS and repost flushed work requests
2550 mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2551 mad_list);
2552 if (wc->status == IB_WC_WR_FLUSH_ERR) {
2553 if (mad_send_wr->retry) {
2554 /* Repost send */
2555 struct ib_send_wr *bad_send_wr;
2557 mad_send_wr->retry = 0;
2558 ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2559 &bad_send_wr);
2560 if (!ret)
2561 return false;
2563 } else {
2564 struct ib_qp_attr *attr;
2566 /* Transition QP to RTS and fail offending send */
2567 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2568 if (attr) {
2569 attr->qp_state = IB_QPS_RTS;
2570 attr->cur_qp_state = IB_QPS_SQE;
2571 ret = ib_modify_qp(qp_info->qp, attr,
2572 IB_QP_STATE | IB_QP_CUR_STATE);
2573 kfree(attr);
2574 if (ret)
2575 dev_err(&port_priv->device->dev,
2576 "%s - ib_modify_qp to RTS: %d\n",
2577 __func__, ret);
2578 else
2579 mark_sends_for_retry(qp_info);
2583 return true;
2586 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2588 unsigned long flags;
2589 struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2590 struct ib_mad_send_wc mad_send_wc;
2591 struct list_head cancel_list;
2593 INIT_LIST_HEAD(&cancel_list);
2595 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2596 list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2597 &mad_agent_priv->send_list, agent_list) {
2598 if (mad_send_wr->status == IB_WC_SUCCESS) {
2599 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2600 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2604 /* Empty wait list to prevent receives from finding a request */
2605 list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2606 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2608 /* Report all cancelled requests */
2609 mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2610 mad_send_wc.vendor_err = 0;
2612 list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2613 &cancel_list, agent_list) {
2614 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2615 list_del(&mad_send_wr->agent_list);
2616 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2617 &mad_send_wc);
2618 atomic_dec(&mad_agent_priv->refcount);
2622 static struct ib_mad_send_wr_private*
2623 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2624 struct ib_mad_send_buf *send_buf)
2626 struct ib_mad_send_wr_private *mad_send_wr;
2628 list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2629 agent_list) {
2630 if (&mad_send_wr->send_buf == send_buf)
2631 return mad_send_wr;
2634 list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2635 agent_list) {
2636 if (is_rmpp_data_mad(mad_agent_priv,
2637 mad_send_wr->send_buf.mad) &&
2638 &mad_send_wr->send_buf == send_buf)
2639 return mad_send_wr;
2641 return NULL;
2644 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2645 struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2647 struct ib_mad_agent_private *mad_agent_priv;
2648 struct ib_mad_send_wr_private *mad_send_wr;
2649 unsigned long flags;
2650 int active;
2652 mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2653 agent);
2654 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2655 mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2656 if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2657 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2658 return -EINVAL;
2661 active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2662 if (!timeout_ms) {
2663 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2664 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2667 mad_send_wr->send_buf.timeout_ms = timeout_ms;
2668 if (active)
2669 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2670 else
2671 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2673 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2674 return 0;
2676 EXPORT_SYMBOL(ib_modify_mad);
2678 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2679 struct ib_mad_send_buf *send_buf)
2681 ib_modify_mad(mad_agent, send_buf, 0);
2683 EXPORT_SYMBOL(ib_cancel_mad);
2685 static void local_completions(struct work_struct *work)
2687 struct ib_mad_agent_private *mad_agent_priv;
2688 struct ib_mad_local_private *local;
2689 struct ib_mad_agent_private *recv_mad_agent;
2690 unsigned long flags;
2691 int free_mad;
2692 struct ib_wc wc;
2693 struct ib_mad_send_wc mad_send_wc;
2694 bool opa;
2696 mad_agent_priv =
2697 container_of(work, struct ib_mad_agent_private, local_work);
2699 opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2700 mad_agent_priv->qp_info->port_priv->port_num);
2702 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2703 while (!list_empty(&mad_agent_priv->local_list)) {
2704 local = list_entry(mad_agent_priv->local_list.next,
2705 struct ib_mad_local_private,
2706 completion_list);
2707 list_del(&local->completion_list);
2708 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2709 free_mad = 0;
2710 if (local->mad_priv) {
2711 u8 base_version;
2712 recv_mad_agent = local->recv_mad_agent;
2713 if (!recv_mad_agent) {
2714 dev_err(&mad_agent_priv->agent.device->dev,
2715 "No receive MAD agent for local completion\n");
2716 free_mad = 1;
2717 goto local_send_completion;
2721 * Defined behavior is to complete response
2722 * before request
2724 build_smp_wc(recv_mad_agent->agent.qp,
2725 local->mad_send_wr->send_wr.wr.wr_cqe,
2726 be16_to_cpu(IB_LID_PERMISSIVE),
2727 local->mad_send_wr->send_wr.pkey_index,
2728 recv_mad_agent->agent.port_num, &wc);
2730 local->mad_priv->header.recv_wc.wc = &wc;
2732 base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2733 if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2734 local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2735 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2736 } else {
2737 local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2738 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2741 INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2742 list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2743 &local->mad_priv->header.recv_wc.rmpp_list);
2744 local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2745 local->mad_priv->header.recv_wc.recv_buf.mad =
2746 (struct ib_mad *)local->mad_priv->mad;
2747 if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2748 snoop_recv(recv_mad_agent->qp_info,
2749 &local->mad_priv->header.recv_wc,
2750 IB_MAD_SNOOP_RECVS);
2751 recv_mad_agent->agent.recv_handler(
2752 &recv_mad_agent->agent,
2753 &local->mad_send_wr->send_buf,
2754 &local->mad_priv->header.recv_wc);
2755 spin_lock_irqsave(&recv_mad_agent->lock, flags);
2756 atomic_dec(&recv_mad_agent->refcount);
2757 spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2760 local_send_completion:
2761 /* Complete send */
2762 mad_send_wc.status = IB_WC_SUCCESS;
2763 mad_send_wc.vendor_err = 0;
2764 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2765 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2766 snoop_send(mad_agent_priv->qp_info,
2767 &local->mad_send_wr->send_buf,
2768 &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2769 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2770 &mad_send_wc);
2772 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2773 atomic_dec(&mad_agent_priv->refcount);
2774 if (free_mad)
2775 kfree(local->mad_priv);
2776 kfree(local);
2778 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2781 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2783 int ret;
2785 if (!mad_send_wr->retries_left)
2786 return -ETIMEDOUT;
2788 mad_send_wr->retries_left--;
2789 mad_send_wr->send_buf.retries++;
2791 mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2793 if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2794 ret = ib_retry_rmpp(mad_send_wr);
2795 switch (ret) {
2796 case IB_RMPP_RESULT_UNHANDLED:
2797 ret = ib_send_mad(mad_send_wr);
2798 break;
2799 case IB_RMPP_RESULT_CONSUMED:
2800 ret = 0;
2801 break;
2802 default:
2803 ret = -ECOMM;
2804 break;
2806 } else
2807 ret = ib_send_mad(mad_send_wr);
2809 if (!ret) {
2810 mad_send_wr->refcount++;
2811 list_add_tail(&mad_send_wr->agent_list,
2812 &mad_send_wr->mad_agent_priv->send_list);
2814 return ret;
2817 static void timeout_sends(struct work_struct *work)
2819 struct ib_mad_agent_private *mad_agent_priv;
2820 struct ib_mad_send_wr_private *mad_send_wr;
2821 struct ib_mad_send_wc mad_send_wc;
2822 unsigned long flags, delay;
2824 mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2825 timed_work.work);
2826 mad_send_wc.vendor_err = 0;
2828 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2829 while (!list_empty(&mad_agent_priv->wait_list)) {
2830 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2831 struct ib_mad_send_wr_private,
2832 agent_list);
2834 if (time_after(mad_send_wr->timeout, jiffies)) {
2835 delay = mad_send_wr->timeout - jiffies;
2836 if ((long)delay <= 0)
2837 delay = 1;
2838 queue_delayed_work(mad_agent_priv->qp_info->
2839 port_priv->wq,
2840 &mad_agent_priv->timed_work, delay);
2841 break;
2844 list_del(&mad_send_wr->agent_list);
2845 if (mad_send_wr->status == IB_WC_SUCCESS &&
2846 !retry_send(mad_send_wr))
2847 continue;
2849 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2851 if (mad_send_wr->status == IB_WC_SUCCESS)
2852 mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2853 else
2854 mad_send_wc.status = mad_send_wr->status;
2855 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2856 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2857 &mad_send_wc);
2859 atomic_dec(&mad_agent_priv->refcount);
2860 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2862 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2866 * Allocate receive MADs and post receive WRs for them
2868 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2869 struct ib_mad_private *mad)
2871 unsigned long flags;
2872 int post, ret;
2873 struct ib_mad_private *mad_priv;
2874 struct ib_sge sg_list;
2875 struct ib_recv_wr recv_wr, *bad_recv_wr;
2876 struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2878 /* Initialize common scatter list fields */
2879 sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2881 /* Initialize common receive WR fields */
2882 recv_wr.next = NULL;
2883 recv_wr.sg_list = &sg_list;
2884 recv_wr.num_sge = 1;
2886 do {
2887 /* Allocate and map receive buffer */
2888 if (mad) {
2889 mad_priv = mad;
2890 mad = NULL;
2891 } else {
2892 mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2893 GFP_ATOMIC);
2894 if (!mad_priv) {
2895 ret = -ENOMEM;
2896 break;
2899 sg_list.length = mad_priv_dma_size(mad_priv);
2900 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2901 &mad_priv->grh,
2902 mad_priv_dma_size(mad_priv),
2903 DMA_FROM_DEVICE);
2904 if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2905 sg_list.addr))) {
2906 ret = -ENOMEM;
2907 break;
2909 mad_priv->header.mapping = sg_list.addr;
2910 mad_priv->header.mad_list.mad_queue = recv_queue;
2911 mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2912 recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2914 /* Post receive WR */
2915 spin_lock_irqsave(&recv_queue->lock, flags);
2916 post = (++recv_queue->count < recv_queue->max_active);
2917 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2918 spin_unlock_irqrestore(&recv_queue->lock, flags);
2919 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2920 if (ret) {
2921 spin_lock_irqsave(&recv_queue->lock, flags);
2922 list_del(&mad_priv->header.mad_list.list);
2923 recv_queue->count--;
2924 spin_unlock_irqrestore(&recv_queue->lock, flags);
2925 ib_dma_unmap_single(qp_info->port_priv->device,
2926 mad_priv->header.mapping,
2927 mad_priv_dma_size(mad_priv),
2928 DMA_FROM_DEVICE);
2929 kfree(mad_priv);
2930 dev_err(&qp_info->port_priv->device->dev,
2931 "ib_post_recv failed: %d\n", ret);
2932 break;
2934 } while (post);
2936 return ret;
2940 * Return all the posted receive MADs
2942 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2944 struct ib_mad_private_header *mad_priv_hdr;
2945 struct ib_mad_private *recv;
2946 struct ib_mad_list_head *mad_list;
2948 if (!qp_info->qp)
2949 return;
2951 while (!list_empty(&qp_info->recv_queue.list)) {
2953 mad_list = list_entry(qp_info->recv_queue.list.next,
2954 struct ib_mad_list_head, list);
2955 mad_priv_hdr = container_of(mad_list,
2956 struct ib_mad_private_header,
2957 mad_list);
2958 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2959 header);
2961 /* Remove from posted receive MAD list */
2962 list_del(&mad_list->list);
2964 ib_dma_unmap_single(qp_info->port_priv->device,
2965 recv->header.mapping,
2966 mad_priv_dma_size(recv),
2967 DMA_FROM_DEVICE);
2968 kfree(recv);
2971 qp_info->recv_queue.count = 0;
2975 * Start the port
2977 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2979 int ret, i;
2980 struct ib_qp_attr *attr;
2981 struct ib_qp *qp;
2982 u16 pkey_index;
2984 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2985 if (!attr)
2986 return -ENOMEM;
2988 ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2989 IB_DEFAULT_PKEY_FULL, &pkey_index);
2990 if (ret)
2991 pkey_index = 0;
2993 for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2994 qp = port_priv->qp_info[i].qp;
2995 if (!qp)
2996 continue;
2999 * PKey index for QP1 is irrelevant but
3000 * one is needed for the Reset to Init transition
3002 attr->qp_state = IB_QPS_INIT;
3003 attr->pkey_index = pkey_index;
3004 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
3005 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
3006 IB_QP_PKEY_INDEX | IB_QP_QKEY);
3007 if (ret) {
3008 dev_err(&port_priv->device->dev,
3009 "Couldn't change QP%d state to INIT: %d\n",
3010 i, ret);
3011 goto out;
3014 attr->qp_state = IB_QPS_RTR;
3015 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
3016 if (ret) {
3017 dev_err(&port_priv->device->dev,
3018 "Couldn't change QP%d state to RTR: %d\n",
3019 i, ret);
3020 goto out;
3023 attr->qp_state = IB_QPS_RTS;
3024 attr->sq_psn = IB_MAD_SEND_Q_PSN;
3025 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3026 if (ret) {
3027 dev_err(&port_priv->device->dev,
3028 "Couldn't change QP%d state to RTS: %d\n",
3029 i, ret);
3030 goto out;
3034 ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3035 if (ret) {
3036 dev_err(&port_priv->device->dev,
3037 "Failed to request completion notification: %d\n",
3038 ret);
3039 goto out;
3042 for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3043 if (!port_priv->qp_info[i].qp)
3044 continue;
3046 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3047 if (ret) {
3048 dev_err(&port_priv->device->dev,
3049 "Couldn't post receive WRs\n");
3050 goto out;
3053 out:
3054 kfree(attr);
3055 return ret;
3058 static void qp_event_handler(struct ib_event *event, void *qp_context)
3060 struct ib_mad_qp_info *qp_info = qp_context;
3062 /* It's worse than that! He's dead, Jim! */
3063 dev_err(&qp_info->port_priv->device->dev,
3064 "Fatal error (%d) on MAD QP (%d)\n",
3065 event->event, qp_info->qp->qp_num);
3068 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3069 struct ib_mad_queue *mad_queue)
3071 mad_queue->qp_info = qp_info;
3072 mad_queue->count = 0;
3073 spin_lock_init(&mad_queue->lock);
3074 INIT_LIST_HEAD(&mad_queue->list);
3077 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3078 struct ib_mad_qp_info *qp_info)
3080 qp_info->port_priv = port_priv;
3081 init_mad_queue(qp_info, &qp_info->send_queue);
3082 init_mad_queue(qp_info, &qp_info->recv_queue);
3083 INIT_LIST_HEAD(&qp_info->overflow_list);
3084 spin_lock_init(&qp_info->snoop_lock);
3085 qp_info->snoop_table = NULL;
3086 qp_info->snoop_table_size = 0;
3087 atomic_set(&qp_info->snoop_count, 0);
3090 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3091 enum ib_qp_type qp_type)
3093 struct ib_qp_init_attr qp_init_attr;
3094 int ret;
3096 memset(&qp_init_attr, 0, sizeof qp_init_attr);
3097 qp_init_attr.send_cq = qp_info->port_priv->cq;
3098 qp_init_attr.recv_cq = qp_info->port_priv->cq;
3099 qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3100 qp_init_attr.cap.max_send_wr = mad_sendq_size;
3101 qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3102 qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3103 qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3104 qp_init_attr.qp_type = qp_type;
3105 qp_init_attr.port_num = qp_info->port_priv->port_num;
3106 qp_init_attr.qp_context = qp_info;
3107 qp_init_attr.event_handler = qp_event_handler;
3108 qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3109 if (IS_ERR(qp_info->qp)) {
3110 dev_err(&qp_info->port_priv->device->dev,
3111 "Couldn't create ib_mad QP%d\n",
3112 get_spl_qp_index(qp_type));
3113 ret = PTR_ERR(qp_info->qp);
3114 goto error;
3116 /* Use minimum queue sizes unless the CQ is resized */
3117 qp_info->send_queue.max_active = mad_sendq_size;
3118 qp_info->recv_queue.max_active = mad_recvq_size;
3119 return 0;
3121 error:
3122 return ret;
3125 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3127 if (!qp_info->qp)
3128 return;
3130 ib_destroy_qp(qp_info->qp);
3131 kfree(qp_info->snoop_table);
3135 * Open the port
3136 * Create the QP, PD, MR, and CQ if needed
3138 static int ib_mad_port_open(struct ib_device *device,
3139 int port_num)
3141 int ret, cq_size;
3142 struct ib_mad_port_private *port_priv;
3143 unsigned long flags;
3144 char name[sizeof "ib_mad123"];
3145 int has_smi;
3147 if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3148 return -EFAULT;
3150 if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3151 rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3152 return -EFAULT;
3154 /* Create new device info */
3155 port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3156 if (!port_priv)
3157 return -ENOMEM;
3159 port_priv->device = device;
3160 port_priv->port_num = port_num;
3161 spin_lock_init(&port_priv->reg_lock);
3162 INIT_LIST_HEAD(&port_priv->agent_list);
3163 init_mad_qp(port_priv, &port_priv->qp_info[0]);
3164 init_mad_qp(port_priv, &port_priv->qp_info[1]);
3166 cq_size = mad_sendq_size + mad_recvq_size;
3167 has_smi = rdma_cap_ib_smi(device, port_num);
3168 if (has_smi)
3169 cq_size *= 2;
3171 port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3172 IB_POLL_WORKQUEUE);
3173 if (IS_ERR(port_priv->cq)) {
3174 dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3175 ret = PTR_ERR(port_priv->cq);
3176 goto error3;
3179 port_priv->pd = ib_alloc_pd(device, 0);
3180 if (IS_ERR(port_priv->pd)) {
3181 dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3182 ret = PTR_ERR(port_priv->pd);
3183 goto error4;
3186 if (has_smi) {
3187 ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3188 if (ret)
3189 goto error6;
3191 ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3192 if (ret)
3193 goto error7;
3195 snprintf(name, sizeof name, "ib_mad%d", port_num);
3196 port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3197 if (!port_priv->wq) {
3198 ret = -ENOMEM;
3199 goto error8;
3202 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3203 list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3204 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3206 ret = ib_mad_port_start(port_priv);
3207 if (ret) {
3208 dev_err(&device->dev, "Couldn't start port\n");
3209 goto error9;
3212 return 0;
3214 error9:
3215 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3216 list_del_init(&port_priv->port_list);
3217 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3219 destroy_workqueue(port_priv->wq);
3220 error8:
3221 destroy_mad_qp(&port_priv->qp_info[1]);
3222 error7:
3223 destroy_mad_qp(&port_priv->qp_info[0]);
3224 error6:
3225 ib_dealloc_pd(port_priv->pd);
3226 error4:
3227 ib_free_cq(port_priv->cq);
3228 cleanup_recv_queue(&port_priv->qp_info[1]);
3229 cleanup_recv_queue(&port_priv->qp_info[0]);
3230 error3:
3231 kfree(port_priv);
3233 return ret;
3237 * Close the port
3238 * If there are no classes using the port, free the port
3239 * resources (CQ, MR, PD, QP) and remove the port's info structure
3241 static int ib_mad_port_close(struct ib_device *device, int port_num)
3243 struct ib_mad_port_private *port_priv;
3244 unsigned long flags;
3246 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3247 port_priv = __ib_get_mad_port(device, port_num);
3248 if (port_priv == NULL) {
3249 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3250 dev_err(&device->dev, "Port %d not found\n", port_num);
3251 return -ENODEV;
3253 list_del_init(&port_priv->port_list);
3254 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3256 destroy_workqueue(port_priv->wq);
3257 destroy_mad_qp(&port_priv->qp_info[1]);
3258 destroy_mad_qp(&port_priv->qp_info[0]);
3259 ib_dealloc_pd(port_priv->pd);
3260 ib_free_cq(port_priv->cq);
3261 cleanup_recv_queue(&port_priv->qp_info[1]);
3262 cleanup_recv_queue(&port_priv->qp_info[0]);
3263 /* XXX: Handle deallocation of MAD registration tables */
3265 kfree(port_priv);
3267 return 0;
3270 static void ib_mad_init_device(struct ib_device *device)
3272 int start, i;
3274 start = rdma_start_port(device);
3276 for (i = start; i <= rdma_end_port(device); i++) {
3277 if (!rdma_cap_ib_mad(device, i))
3278 continue;
3280 if (ib_mad_port_open(device, i)) {
3281 dev_err(&device->dev, "Couldn't open port %d\n", i);
3282 goto error;
3284 if (ib_agent_port_open(device, i)) {
3285 dev_err(&device->dev,
3286 "Couldn't open port %d for agents\n", i);
3287 goto error_agent;
3290 return;
3292 error_agent:
3293 if (ib_mad_port_close(device, i))
3294 dev_err(&device->dev, "Couldn't close port %d\n", i);
3296 error:
3297 while (--i >= start) {
3298 if (!rdma_cap_ib_mad(device, i))
3299 continue;
3301 if (ib_agent_port_close(device, i))
3302 dev_err(&device->dev,
3303 "Couldn't close port %d for agents\n", i);
3304 if (ib_mad_port_close(device, i))
3305 dev_err(&device->dev, "Couldn't close port %d\n", i);
3309 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3311 int i;
3313 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3314 if (!rdma_cap_ib_mad(device, i))
3315 continue;
3317 if (ib_agent_port_close(device, i))
3318 dev_err(&device->dev,
3319 "Couldn't close port %d for agents\n", i);
3320 if (ib_mad_port_close(device, i))
3321 dev_err(&device->dev, "Couldn't close port %d\n", i);
3325 static struct ib_client mad_client = {
3326 .name = "mad",
3327 .add = ib_mad_init_device,
3328 .remove = ib_mad_remove_device
3331 int ib_mad_init(void)
3333 mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3334 mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3336 mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3337 mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3339 INIT_LIST_HEAD(&ib_mad_port_list);
3341 if (ib_register_client(&mad_client)) {
3342 pr_err("Couldn't register ib_mad client\n");
3343 return -EINVAL;
3346 return 0;
3349 void ib_mad_cleanup(void)
3351 ib_unregister_client(&mad_client);