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
18 * - Redistributions of source code must retain the above
19 * copyright notice, this list of conditions and the following
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
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 <rdma/ib_cache.h>
51 MODULE_LICENSE("Dual BSD/GPL");
52 MODULE_DESCRIPTION("kernel IB MAD API");
53 MODULE_AUTHOR("Hal Rosenstock");
54 MODULE_AUTHOR("Sean Hefty");
56 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
57 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
59 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
60 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
61 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
62 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
64 static struct list_head ib_mad_port_list;
65 static u32 ib_mad_client_id = 0;
68 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
70 /* Forward declarations */
71 static int method_in_use(struct ib_mad_mgmt_method_table **method,
72 struct ib_mad_reg_req *mad_reg_req);
73 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
74 static struct ib_mad_agent_private *find_mad_agent(
75 struct ib_mad_port_private *port_priv,
76 const struct ib_mad_hdr *mad);
77 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
78 struct ib_mad_private *mad);
79 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
80 static void timeout_sends(struct work_struct *work);
81 static void local_completions(struct work_struct *work);
82 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
83 struct ib_mad_agent_private *agent_priv,
85 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
86 struct ib_mad_agent_private *agent_priv);
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)
105 * Wrapper function to return a ib_mad_port_private structure or NULL
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;
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);
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 ?
128 static int get_spl_qp_index(enum ib_qp_type qp_type)
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))
154 static int is_vendor_oui(char *oui)
156 if (oui[0] || oui[1] || oui[2])
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;
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];
172 if (method_in_use(&method, mad_reg_req))
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,
196 enum ib_qp_type qp_type,
197 struct ib_mad_reg_req *mad_reg_req,
199 ib_mad_send_handler send_handler,
200 ib_mad_recv_handler recv_handler,
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;
214 u8 mgmt_class, vclass;
216 /* Validate parameters */
217 qpn = get_spl_qp_index(qp_type);
219 dev_notice(&device->dev,
220 "ib_register_mad_agent: invalid QP Type %d\n",
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",
232 /* Validate MAD registration request if supplied */
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);
241 dev_notice(&device->dev,
242 "ib_register_mad_agent: no recv_handler\n");
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);
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");
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);
277 /* Make sure class supplied is consistent with RMPP */
278 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
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);
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);
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);
310 /* No registration request supplied */
313 if (registration_flags & IB_MAD_USER_RMPP)
317 /* Validate device and port */
318 port_priv = ib_get_mad_port(device, port_num);
320 dev_notice(&device->dev, "ib_register_mad_agent: Invalid port\n");
321 ret = ERR_PTR(-ENODEV);
325 /* Verify the QP requested is supported. For example, Ethernet devices
326 * will not have QP0 */
327 if (!port_priv->qp_info[qpn].qp) {
328 dev_notice(&device->dev,
329 "ib_register_mad_agent: QP %d not supported\n", qpn);
330 ret = ERR_PTR(-EPROTONOSUPPORT);
334 /* Allocate structures */
335 mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
336 if (!mad_agent_priv) {
337 ret = ERR_PTR(-ENOMEM);
342 reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
344 ret = ERR_PTR(-ENOMEM);
349 /* Now, fill in the various structures */
350 mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
351 mad_agent_priv->reg_req = reg_req;
352 mad_agent_priv->agent.rmpp_version = rmpp_version;
353 mad_agent_priv->agent.device = device;
354 mad_agent_priv->agent.recv_handler = recv_handler;
355 mad_agent_priv->agent.send_handler = send_handler;
356 mad_agent_priv->agent.context = context;
357 mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
358 mad_agent_priv->agent.port_num = port_num;
359 mad_agent_priv->agent.flags = registration_flags;
360 spin_lock_init(&mad_agent_priv->lock);
361 INIT_LIST_HEAD(&mad_agent_priv->send_list);
362 INIT_LIST_HEAD(&mad_agent_priv->wait_list);
363 INIT_LIST_HEAD(&mad_agent_priv->done_list);
364 INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
365 INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
366 INIT_LIST_HEAD(&mad_agent_priv->local_list);
367 INIT_WORK(&mad_agent_priv->local_work, local_completions);
368 atomic_set(&mad_agent_priv->refcount, 1);
369 init_completion(&mad_agent_priv->comp);
371 spin_lock_irqsave(&port_priv->reg_lock, flags);
372 mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
375 * Make sure MAD registration (if supplied)
376 * is non overlapping with any existing ones
379 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
380 if (!is_vendor_class(mgmt_class)) {
381 class = port_priv->version[mad_reg_req->
382 mgmt_class_version].class;
384 method = class->method_table[mgmt_class];
386 if (method_in_use(&method,
391 ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
394 /* "New" vendor class range */
395 vendor = port_priv->version[mad_reg_req->
396 mgmt_class_version].vendor;
398 vclass = vendor_class_index(mgmt_class);
399 vendor_class = vendor->vendor_class[vclass];
401 if (is_vendor_method_in_use(
407 ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
415 /* Add mad agent into port's agent list */
416 list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
417 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
419 return &mad_agent_priv->agent;
422 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
425 kfree(mad_agent_priv);
429 EXPORT_SYMBOL(ib_register_mad_agent);
431 static inline int is_snooping_sends(int mad_snoop_flags)
433 return (mad_snoop_flags &
434 (/*IB_MAD_SNOOP_POSTED_SENDS |
435 IB_MAD_SNOOP_RMPP_SENDS |*/
436 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
437 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
440 static inline int is_snooping_recvs(int mad_snoop_flags)
442 return (mad_snoop_flags &
443 (IB_MAD_SNOOP_RECVS /*|
444 IB_MAD_SNOOP_RMPP_RECVS*/));
447 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
448 struct ib_mad_snoop_private *mad_snoop_priv)
450 struct ib_mad_snoop_private **new_snoop_table;
454 spin_lock_irqsave(&qp_info->snoop_lock, flags);
455 /* Check for empty slot in array. */
456 for (i = 0; i < qp_info->snoop_table_size; i++)
457 if (!qp_info->snoop_table[i])
460 if (i == qp_info->snoop_table_size) {
462 new_snoop_table = krealloc(qp_info->snoop_table,
463 sizeof mad_snoop_priv *
464 (qp_info->snoop_table_size + 1),
466 if (!new_snoop_table) {
471 qp_info->snoop_table = new_snoop_table;
472 qp_info->snoop_table_size++;
474 qp_info->snoop_table[i] = mad_snoop_priv;
475 atomic_inc(&qp_info->snoop_count);
477 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
481 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
483 enum ib_qp_type qp_type,
485 ib_mad_snoop_handler snoop_handler,
486 ib_mad_recv_handler recv_handler,
489 struct ib_mad_port_private *port_priv;
490 struct ib_mad_agent *ret;
491 struct ib_mad_snoop_private *mad_snoop_priv;
494 /* Validate parameters */
495 if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
496 (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
497 ret = ERR_PTR(-EINVAL);
500 qpn = get_spl_qp_index(qp_type);
502 ret = ERR_PTR(-EINVAL);
505 port_priv = ib_get_mad_port(device, port_num);
507 ret = ERR_PTR(-ENODEV);
510 /* Allocate structures */
511 mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
512 if (!mad_snoop_priv) {
513 ret = ERR_PTR(-ENOMEM);
517 /* Now, fill in the various structures */
518 mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
519 mad_snoop_priv->agent.device = device;
520 mad_snoop_priv->agent.recv_handler = recv_handler;
521 mad_snoop_priv->agent.snoop_handler = snoop_handler;
522 mad_snoop_priv->agent.context = context;
523 mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
524 mad_snoop_priv->agent.port_num = port_num;
525 mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
526 init_completion(&mad_snoop_priv->comp);
527 mad_snoop_priv->snoop_index = register_snoop_agent(
528 &port_priv->qp_info[qpn],
530 if (mad_snoop_priv->snoop_index < 0) {
531 ret = ERR_PTR(mad_snoop_priv->snoop_index);
535 atomic_set(&mad_snoop_priv->refcount, 1);
536 return &mad_snoop_priv->agent;
539 kfree(mad_snoop_priv);
543 EXPORT_SYMBOL(ib_register_mad_snoop);
545 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
547 if (atomic_dec_and_test(&mad_agent_priv->refcount))
548 complete(&mad_agent_priv->comp);
551 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
553 if (atomic_dec_and_test(&mad_snoop_priv->refcount))
554 complete(&mad_snoop_priv->comp);
557 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
559 struct ib_mad_port_private *port_priv;
562 /* Note that we could still be handling received MADs */
565 * Canceling all sends results in dropping received response
566 * MADs, preventing us from queuing additional work
568 cancel_mads(mad_agent_priv);
569 port_priv = mad_agent_priv->qp_info->port_priv;
570 cancel_delayed_work(&mad_agent_priv->timed_work);
572 spin_lock_irqsave(&port_priv->reg_lock, flags);
573 remove_mad_reg_req(mad_agent_priv);
574 list_del(&mad_agent_priv->agent_list);
575 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
577 flush_workqueue(port_priv->wq);
578 ib_cancel_rmpp_recvs(mad_agent_priv);
580 deref_mad_agent(mad_agent_priv);
581 wait_for_completion(&mad_agent_priv->comp);
583 kfree(mad_agent_priv->reg_req);
584 kfree(mad_agent_priv);
587 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
589 struct ib_mad_qp_info *qp_info;
592 qp_info = mad_snoop_priv->qp_info;
593 spin_lock_irqsave(&qp_info->snoop_lock, flags);
594 qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
595 atomic_dec(&qp_info->snoop_count);
596 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
598 deref_snoop_agent(mad_snoop_priv);
599 wait_for_completion(&mad_snoop_priv->comp);
601 kfree(mad_snoop_priv);
605 * ib_unregister_mad_agent - Unregisters a client from using MAD services
607 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
609 struct ib_mad_agent_private *mad_agent_priv;
610 struct ib_mad_snoop_private *mad_snoop_priv;
612 /* If the TID is zero, the agent can only snoop. */
613 if (mad_agent->hi_tid) {
614 mad_agent_priv = container_of(mad_agent,
615 struct ib_mad_agent_private,
617 unregister_mad_agent(mad_agent_priv);
619 mad_snoop_priv = container_of(mad_agent,
620 struct ib_mad_snoop_private,
622 unregister_mad_snoop(mad_snoop_priv);
626 EXPORT_SYMBOL(ib_unregister_mad_agent);
628 static void dequeue_mad(struct ib_mad_list_head *mad_list)
630 struct ib_mad_queue *mad_queue;
633 BUG_ON(!mad_list->mad_queue);
634 mad_queue = mad_list->mad_queue;
635 spin_lock_irqsave(&mad_queue->lock, flags);
636 list_del(&mad_list->list);
638 spin_unlock_irqrestore(&mad_queue->lock, flags);
641 static void snoop_send(struct ib_mad_qp_info *qp_info,
642 struct ib_mad_send_buf *send_buf,
643 struct ib_mad_send_wc *mad_send_wc,
646 struct ib_mad_snoop_private *mad_snoop_priv;
650 spin_lock_irqsave(&qp_info->snoop_lock, flags);
651 for (i = 0; i < qp_info->snoop_table_size; i++) {
652 mad_snoop_priv = qp_info->snoop_table[i];
653 if (!mad_snoop_priv ||
654 !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
657 atomic_inc(&mad_snoop_priv->refcount);
658 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
659 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
660 send_buf, mad_send_wc);
661 deref_snoop_agent(mad_snoop_priv);
662 spin_lock_irqsave(&qp_info->snoop_lock, flags);
664 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
667 static void snoop_recv(struct ib_mad_qp_info *qp_info,
668 struct ib_mad_recv_wc *mad_recv_wc,
671 struct ib_mad_snoop_private *mad_snoop_priv;
675 spin_lock_irqsave(&qp_info->snoop_lock, flags);
676 for (i = 0; i < qp_info->snoop_table_size; i++) {
677 mad_snoop_priv = qp_info->snoop_table[i];
678 if (!mad_snoop_priv ||
679 !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
682 atomic_inc(&mad_snoop_priv->refcount);
683 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
684 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
686 deref_snoop_agent(mad_snoop_priv);
687 spin_lock_irqsave(&qp_info->snoop_lock, flags);
689 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
692 static void build_smp_wc(struct ib_qp *qp,
693 u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
696 memset(wc, 0, sizeof *wc);
698 wc->status = IB_WC_SUCCESS;
699 wc->opcode = IB_WC_RECV;
700 wc->pkey_index = pkey_index;
701 wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
706 wc->dlid_path_bits = 0;
707 wc->port_num = port_num;
710 static size_t mad_priv_size(const struct ib_mad_private *mp)
712 return sizeof(struct ib_mad_private) + mp->mad_size;
715 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
717 size_t size = sizeof(struct ib_mad_private) + mad_size;
718 struct ib_mad_private *ret = kzalloc(size, flags);
721 ret->mad_size = mad_size;
726 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
728 return rdma_max_mad_size(port_priv->device, port_priv->port_num);
731 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
733 return sizeof(struct ib_grh) + mp->mad_size;
737 * Return 0 if SMP is to be sent
738 * Return 1 if SMP was consumed locally (whether or not solicited)
739 * Return < 0 if error
741 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
742 struct ib_mad_send_wr_private *mad_send_wr)
745 struct ib_smp *smp = mad_send_wr->send_buf.mad;
746 struct opa_smp *opa_smp = (struct opa_smp *)smp;
748 struct ib_mad_local_private *local;
749 struct ib_mad_private *mad_priv;
750 struct ib_mad_port_private *port_priv;
751 struct ib_mad_agent_private *recv_mad_agent = NULL;
752 struct ib_device *device = mad_agent_priv->agent.device;
755 struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
756 size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
757 u16 out_mad_pkey_index = 0;
759 bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
760 mad_agent_priv->qp_info->port_priv->port_num);
762 if (rdma_cap_ib_switch(device) &&
763 smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
764 port_num = send_wr->port_num;
766 port_num = mad_agent_priv->agent.port_num;
769 * Directed route handling starts if the initial LID routed part of
770 * a request or the ending LID routed part of a response is empty.
771 * If we are at the start of the LID routed part, don't update the
772 * hop_ptr or hop_cnt. See section 14.2.2, Vol 1 IB spec.
774 if (opa && smp->class_version == OPA_SMP_CLASS_VERSION) {
777 if ((opa_get_smp_direction(opa_smp)
778 ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
779 OPA_LID_PERMISSIVE &&
780 opa_smi_handle_dr_smp_send(opa_smp,
781 rdma_cap_ib_switch(device),
782 port_num) == IB_SMI_DISCARD) {
784 dev_err(&device->dev, "OPA Invalid directed route\n");
787 opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
788 if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
789 opa_drslid & 0xffff0000) {
791 dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
795 drslid = (u16)(opa_drslid & 0x0000ffff);
797 /* Check to post send on QP or process locally */
798 if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
799 opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
802 if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
804 smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
807 dev_err(&device->dev, "Invalid directed route\n");
810 drslid = be16_to_cpu(smp->dr_slid);
812 /* Check to post send on QP or process locally */
813 if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
814 smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
818 local = kmalloc(sizeof *local, GFP_ATOMIC);
821 dev_err(&device->dev, "No memory for ib_mad_local_private\n");
824 local->mad_priv = NULL;
825 local->recv_mad_agent = NULL;
826 mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
829 dev_err(&device->dev, "No memory for local response MAD\n");
834 build_smp_wc(mad_agent_priv->agent.qp,
835 send_wr->wr.wr_id, drslid,
837 send_wr->port_num, &mad_wc);
839 if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
840 mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
841 + mad_send_wr->send_buf.data_len
842 + sizeof(struct ib_grh);
845 /* No GRH for DR SMP */
846 ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
847 (const struct ib_mad_hdr *)smp, mad_size,
848 (struct ib_mad_hdr *)mad_priv->mad,
849 &mad_size, &out_mad_pkey_index);
852 case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
853 if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
854 mad_agent_priv->agent.recv_handler) {
855 local->mad_priv = mad_priv;
856 local->recv_mad_agent = mad_agent_priv;
858 * Reference MAD agent until receive
859 * side of local completion handled
861 atomic_inc(&mad_agent_priv->refcount);
865 case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
868 case IB_MAD_RESULT_SUCCESS:
869 /* Treat like an incoming receive MAD */
870 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
871 mad_agent_priv->agent.port_num);
873 memcpy(mad_priv->mad, smp, mad_priv->mad_size);
874 recv_mad_agent = find_mad_agent(port_priv,
875 (const struct ib_mad_hdr *)mad_priv->mad);
877 if (!port_priv || !recv_mad_agent) {
879 * No receiving agent so drop packet and
880 * generate send completion.
885 local->mad_priv = mad_priv;
886 local->recv_mad_agent = recv_mad_agent;
895 local->mad_send_wr = mad_send_wr;
897 local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
898 local->return_wc_byte_len = mad_size;
900 /* Reference MAD agent until send side of local completion handled */
901 atomic_inc(&mad_agent_priv->refcount);
902 /* Queue local completion to local list */
903 spin_lock_irqsave(&mad_agent_priv->lock, flags);
904 list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
905 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
906 queue_work(mad_agent_priv->qp_info->port_priv->wq,
907 &mad_agent_priv->local_work);
913 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
917 seg_size = mad_size - hdr_len;
918 if (data_len && seg_size) {
919 pad = seg_size - data_len % seg_size;
920 return pad == seg_size ? 0 : pad;
925 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
927 struct ib_rmpp_segment *s, *t;
929 list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
935 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
936 size_t mad_size, gfp_t gfp_mask)
938 struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
939 struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
940 struct ib_rmpp_segment *seg = NULL;
941 int left, seg_size, pad;
943 send_buf->seg_size = mad_size - send_buf->hdr_len;
944 send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
945 seg_size = send_buf->seg_size;
948 /* Allocate data segments. */
949 for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
950 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
952 dev_err(&send_buf->mad_agent->device->dev,
953 "alloc_send_rmpp_segs: RMPP mem alloc failed for len %zd, gfp %#x\n",
954 sizeof (*seg) + seg_size, gfp_mask);
955 free_send_rmpp_list(send_wr);
958 seg->num = ++send_buf->seg_count;
959 list_add_tail(&seg->list, &send_wr->rmpp_list);
962 /* Zero any padding */
964 memset(seg->data + seg_size - pad, 0, pad);
966 rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
968 rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
969 ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
971 send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
972 struct ib_rmpp_segment, list);
973 send_wr->last_ack_seg = send_wr->cur_seg;
977 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
979 return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
981 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
983 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
984 u32 remote_qpn, u16 pkey_index,
986 int hdr_len, int data_len,
990 struct ib_mad_agent_private *mad_agent_priv;
991 struct ib_mad_send_wr_private *mad_send_wr;
992 int pad, message_size, ret, size;
997 mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
1000 opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
1002 if (opa && base_version == OPA_MGMT_BASE_VERSION)
1003 mad_size = sizeof(struct opa_mad);
1005 mad_size = sizeof(struct ib_mad);
1007 pad = get_pad_size(hdr_len, data_len, mad_size);
1008 message_size = hdr_len + data_len + pad;
1010 if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1011 if (!rmpp_active && message_size > mad_size)
1012 return ERR_PTR(-EINVAL);
1014 if (rmpp_active || message_size > mad_size)
1015 return ERR_PTR(-EINVAL);
1017 size = rmpp_active ? hdr_len : mad_size;
1018 buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1020 return ERR_PTR(-ENOMEM);
1022 mad_send_wr = buf + size;
1023 INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1024 mad_send_wr->send_buf.mad = buf;
1025 mad_send_wr->send_buf.hdr_len = hdr_len;
1026 mad_send_wr->send_buf.data_len = data_len;
1027 mad_send_wr->pad = pad;
1029 mad_send_wr->mad_agent_priv = mad_agent_priv;
1030 mad_send_wr->sg_list[0].length = hdr_len;
1031 mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1033 /* OPA MADs don't have to be the full 2048 bytes */
1034 if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1035 data_len < mad_size - hdr_len)
1036 mad_send_wr->sg_list[1].length = data_len;
1038 mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1040 mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1042 mad_send_wr->send_wr.wr.wr_id = (unsigned long) mad_send_wr;
1043 mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1044 mad_send_wr->send_wr.wr.num_sge = 2;
1045 mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1046 mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1047 mad_send_wr->send_wr.remote_qpn = remote_qpn;
1048 mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1049 mad_send_wr->send_wr.pkey_index = pkey_index;
1052 ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1055 return ERR_PTR(ret);
1059 mad_send_wr->send_buf.mad_agent = mad_agent;
1060 atomic_inc(&mad_agent_priv->refcount);
1061 return &mad_send_wr->send_buf;
1063 EXPORT_SYMBOL(ib_create_send_mad);
1065 int ib_get_mad_data_offset(u8 mgmt_class)
1067 if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1068 return IB_MGMT_SA_HDR;
1069 else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1070 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1071 (mgmt_class == IB_MGMT_CLASS_BIS))
1072 return IB_MGMT_DEVICE_HDR;
1073 else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1074 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1075 return IB_MGMT_VENDOR_HDR;
1077 return IB_MGMT_MAD_HDR;
1079 EXPORT_SYMBOL(ib_get_mad_data_offset);
1081 int ib_is_mad_class_rmpp(u8 mgmt_class)
1083 if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1084 (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1085 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1086 (mgmt_class == IB_MGMT_CLASS_BIS) ||
1087 ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1088 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1092 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1094 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1096 struct ib_mad_send_wr_private *mad_send_wr;
1097 struct list_head *list;
1099 mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1101 list = &mad_send_wr->cur_seg->list;
1103 if (mad_send_wr->cur_seg->num < seg_num) {
1104 list_for_each_entry(mad_send_wr->cur_seg, list, list)
1105 if (mad_send_wr->cur_seg->num == seg_num)
1107 } else if (mad_send_wr->cur_seg->num > seg_num) {
1108 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1109 if (mad_send_wr->cur_seg->num == seg_num)
1112 return mad_send_wr->cur_seg->data;
1114 EXPORT_SYMBOL(ib_get_rmpp_segment);
1116 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1118 if (mad_send_wr->send_buf.seg_count)
1119 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1120 mad_send_wr->seg_num);
1122 return mad_send_wr->send_buf.mad +
1123 mad_send_wr->send_buf.hdr_len;
1126 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1128 struct ib_mad_agent_private *mad_agent_priv;
1129 struct ib_mad_send_wr_private *mad_send_wr;
1131 mad_agent_priv = container_of(send_buf->mad_agent,
1132 struct ib_mad_agent_private, agent);
1133 mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1136 free_send_rmpp_list(mad_send_wr);
1137 kfree(send_buf->mad);
1138 deref_mad_agent(mad_agent_priv);
1140 EXPORT_SYMBOL(ib_free_send_mad);
1142 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1144 struct ib_mad_qp_info *qp_info;
1145 struct list_head *list;
1146 struct ib_send_wr *bad_send_wr;
1147 struct ib_mad_agent *mad_agent;
1149 unsigned long flags;
1152 /* Set WR ID to find mad_send_wr upon completion */
1153 qp_info = mad_send_wr->mad_agent_priv->qp_info;
1154 mad_send_wr->send_wr.wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
1155 mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1157 mad_agent = mad_send_wr->send_buf.mad_agent;
1158 sge = mad_send_wr->sg_list;
1159 sge[0].addr = ib_dma_map_single(mad_agent->device,
1160 mad_send_wr->send_buf.mad,
1163 if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1166 mad_send_wr->header_mapping = sge[0].addr;
1168 sge[1].addr = ib_dma_map_single(mad_agent->device,
1169 ib_get_payload(mad_send_wr),
1172 if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1173 ib_dma_unmap_single(mad_agent->device,
1174 mad_send_wr->header_mapping,
1175 sge[0].length, DMA_TO_DEVICE);
1178 mad_send_wr->payload_mapping = sge[1].addr;
1180 spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1181 if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1182 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1184 list = &qp_info->send_queue.list;
1187 list = &qp_info->overflow_list;
1191 qp_info->send_queue.count++;
1192 list_add_tail(&mad_send_wr->mad_list.list, list);
1194 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1196 ib_dma_unmap_single(mad_agent->device,
1197 mad_send_wr->header_mapping,
1198 sge[0].length, DMA_TO_DEVICE);
1199 ib_dma_unmap_single(mad_agent->device,
1200 mad_send_wr->payload_mapping,
1201 sge[1].length, DMA_TO_DEVICE);
1207 * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1208 * with the registered client
1210 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1211 struct ib_mad_send_buf **bad_send_buf)
1213 struct ib_mad_agent_private *mad_agent_priv;
1214 struct ib_mad_send_buf *next_send_buf;
1215 struct ib_mad_send_wr_private *mad_send_wr;
1216 unsigned long flags;
1219 /* Walk list of send WRs and post each on send list */
1220 for (; send_buf; send_buf = next_send_buf) {
1222 mad_send_wr = container_of(send_buf,
1223 struct ib_mad_send_wr_private,
1225 mad_agent_priv = mad_send_wr->mad_agent_priv;
1227 if (!send_buf->mad_agent->send_handler ||
1228 (send_buf->timeout_ms &&
1229 !send_buf->mad_agent->recv_handler)) {
1234 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1235 if (mad_agent_priv->agent.rmpp_version) {
1242 * Save pointer to next work request to post in case the
1243 * current one completes, and the user modifies the work
1244 * request associated with the completion
1246 next_send_buf = send_buf->next;
1247 mad_send_wr->send_wr.ah = send_buf->ah;
1249 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1250 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1251 ret = handle_outgoing_dr_smp(mad_agent_priv,
1253 if (ret < 0) /* error */
1255 else if (ret == 1) /* locally consumed */
1259 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1260 /* Timeout will be updated after send completes */
1261 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1262 mad_send_wr->max_retries = send_buf->retries;
1263 mad_send_wr->retries_left = send_buf->retries;
1264 send_buf->retries = 0;
1265 /* Reference for work request to QP + response */
1266 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1267 mad_send_wr->status = IB_WC_SUCCESS;
1269 /* Reference MAD agent until send completes */
1270 atomic_inc(&mad_agent_priv->refcount);
1271 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1272 list_add_tail(&mad_send_wr->agent_list,
1273 &mad_agent_priv->send_list);
1274 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1276 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1277 ret = ib_send_rmpp_mad(mad_send_wr);
1278 if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1279 ret = ib_send_mad(mad_send_wr);
1281 ret = ib_send_mad(mad_send_wr);
1283 /* Fail send request */
1284 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1285 list_del(&mad_send_wr->agent_list);
1286 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1287 atomic_dec(&mad_agent_priv->refcount);
1294 *bad_send_buf = send_buf;
1297 EXPORT_SYMBOL(ib_post_send_mad);
1300 * ib_free_recv_mad - Returns data buffers used to receive
1301 * a MAD to the access layer
1303 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1305 struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1306 struct ib_mad_private_header *mad_priv_hdr;
1307 struct ib_mad_private *priv;
1308 struct list_head free_list;
1310 INIT_LIST_HEAD(&free_list);
1311 list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1313 list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1315 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1317 mad_priv_hdr = container_of(mad_recv_wc,
1318 struct ib_mad_private_header,
1320 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1325 EXPORT_SYMBOL(ib_free_recv_mad);
1327 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1329 ib_mad_send_handler send_handler,
1330 ib_mad_recv_handler recv_handler,
1333 return ERR_PTR(-EINVAL); /* XXX: for now */
1335 EXPORT_SYMBOL(ib_redirect_mad_qp);
1337 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1340 dev_err(&mad_agent->device->dev,
1341 "ib_process_mad_wc() not implemented yet\n");
1344 EXPORT_SYMBOL(ib_process_mad_wc);
1346 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1347 struct ib_mad_reg_req *mad_reg_req)
1351 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1352 if ((*method)->agent[i]) {
1353 pr_err("Method %d already in use\n", i);
1360 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1362 /* Allocate management method table */
1363 *method = kzalloc(sizeof **method, GFP_ATOMIC);
1365 pr_err("No memory for ib_mad_mgmt_method_table\n");
1373 * Check to see if there are any methods still in use
1375 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1379 for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1380 if (method->agent[i])
1386 * Check to see if there are any method tables for this class still in use
1388 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1392 for (i = 0; i < MAX_MGMT_CLASS; i++)
1393 if (class->method_table[i])
1398 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1402 for (i = 0; i < MAX_MGMT_OUI; i++)
1403 if (vendor_class->method_table[i])
1408 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1413 for (i = 0; i < MAX_MGMT_OUI; i++)
1414 /* Is there matching OUI for this vendor class ? */
1415 if (!memcmp(vendor_class->oui[i], oui, 3))
1421 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1425 for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1426 if (vendor->vendor_class[i])
1432 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1433 struct ib_mad_agent_private *agent)
1437 /* Remove any methods for this mad agent */
1438 for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1439 if (method->agent[i] == agent) {
1440 method->agent[i] = NULL;
1445 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1446 struct ib_mad_agent_private *agent_priv,
1449 struct ib_mad_port_private *port_priv;
1450 struct ib_mad_mgmt_class_table **class;
1451 struct ib_mad_mgmt_method_table **method;
1454 port_priv = agent_priv->qp_info->port_priv;
1455 class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1457 /* Allocate management class table for "new" class version */
1458 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1460 dev_err(&agent_priv->agent.device->dev,
1461 "No memory for ib_mad_mgmt_class_table\n");
1466 /* Allocate method table for this management class */
1467 method = &(*class)->method_table[mgmt_class];
1468 if ((ret = allocate_method_table(method)))
1471 method = &(*class)->method_table[mgmt_class];
1473 /* Allocate method table for this management class */
1474 if ((ret = allocate_method_table(method)))
1479 /* Now, make sure methods are not already in use */
1480 if (method_in_use(method, mad_reg_req))
1483 /* Finally, add in methods being registered */
1484 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1485 (*method)->agent[i] = agent_priv;
1490 /* Remove any methods for this mad agent */
1491 remove_methods_mad_agent(*method, agent_priv);
1492 /* Now, check to see if there are any methods in use */
1493 if (!check_method_table(*method)) {
1494 /* If not, release management method table */
1507 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1508 struct ib_mad_agent_private *agent_priv)
1510 struct ib_mad_port_private *port_priv;
1511 struct ib_mad_mgmt_vendor_class_table **vendor_table;
1512 struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1513 struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1514 struct ib_mad_mgmt_method_table **method;
1515 int i, ret = -ENOMEM;
1518 /* "New" vendor (with OUI) class */
1519 vclass = vendor_class_index(mad_reg_req->mgmt_class);
1520 port_priv = agent_priv->qp_info->port_priv;
1521 vendor_table = &port_priv->version[
1522 mad_reg_req->mgmt_class_version].vendor;
1523 if (!*vendor_table) {
1524 /* Allocate mgmt vendor class table for "new" class version */
1525 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1527 dev_err(&agent_priv->agent.device->dev,
1528 "No memory for ib_mad_mgmt_vendor_class_table\n");
1532 *vendor_table = vendor;
1534 if (!(*vendor_table)->vendor_class[vclass]) {
1535 /* Allocate table for this management vendor class */
1536 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1537 if (!vendor_class) {
1538 dev_err(&agent_priv->agent.device->dev,
1539 "No memory for ib_mad_mgmt_vendor_class\n");
1543 (*vendor_table)->vendor_class[vclass] = vendor_class;
1545 for (i = 0; i < MAX_MGMT_OUI; i++) {
1546 /* Is there matching OUI for this vendor class ? */
1547 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1548 mad_reg_req->oui, 3)) {
1549 method = &(*vendor_table)->vendor_class[
1550 vclass]->method_table[i];
1556 for (i = 0; i < MAX_MGMT_OUI; i++) {
1557 /* OUI slot available ? */
1558 if (!is_vendor_oui((*vendor_table)->vendor_class[
1560 method = &(*vendor_table)->vendor_class[
1561 vclass]->method_table[i];
1562 /* Allocate method table for this OUI */
1564 ret = allocate_method_table(method);
1568 memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1569 mad_reg_req->oui, 3);
1573 dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1577 /* Now, make sure methods are not already in use */
1578 if (method_in_use(method, mad_reg_req))
1581 /* Finally, add in methods being registered */
1582 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1583 (*method)->agent[i] = agent_priv;
1588 /* Remove any methods for this mad agent */
1589 remove_methods_mad_agent(*method, agent_priv);
1590 /* Now, check to see if there are any methods in use */
1591 if (!check_method_table(*method)) {
1592 /* If not, release management method table */
1599 (*vendor_table)->vendor_class[vclass] = NULL;
1600 kfree(vendor_class);
1604 *vendor_table = NULL;
1611 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1613 struct ib_mad_port_private *port_priv;
1614 struct ib_mad_mgmt_class_table *class;
1615 struct ib_mad_mgmt_method_table *method;
1616 struct ib_mad_mgmt_vendor_class_table *vendor;
1617 struct ib_mad_mgmt_vendor_class *vendor_class;
1622 * Was MAD registration request supplied
1623 * with original registration ?
1625 if (!agent_priv->reg_req) {
1629 port_priv = agent_priv->qp_info->port_priv;
1630 mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1631 class = port_priv->version[
1632 agent_priv->reg_req->mgmt_class_version].class;
1636 method = class->method_table[mgmt_class];
1638 /* Remove any methods for this mad agent */
1639 remove_methods_mad_agent(method, agent_priv);
1640 /* Now, check to see if there are any methods still in use */
1641 if (!check_method_table(method)) {
1642 /* If not, release management method table */
1644 class->method_table[mgmt_class] = NULL;
1645 /* Any management classes left ? */
1646 if (!check_class_table(class)) {
1647 /* If not, release management class table */
1650 agent_priv->reg_req->
1651 mgmt_class_version].class = NULL;
1657 if (!is_vendor_class(mgmt_class))
1660 /* normalize mgmt_class to vendor range 2 */
1661 mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1662 vendor = port_priv->version[
1663 agent_priv->reg_req->mgmt_class_version].vendor;
1668 vendor_class = vendor->vendor_class[mgmt_class];
1670 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1673 method = vendor_class->method_table[index];
1675 /* Remove any methods for this mad agent */
1676 remove_methods_mad_agent(method, agent_priv);
1678 * Now, check to see if there are
1679 * any methods still in use
1681 if (!check_method_table(method)) {
1682 /* If not, release management method table */
1684 vendor_class->method_table[index] = NULL;
1685 memset(vendor_class->oui[index], 0, 3);
1686 /* Any OUIs left ? */
1687 if (!check_vendor_class(vendor_class)) {
1688 /* If not, release vendor class table */
1689 kfree(vendor_class);
1690 vendor->vendor_class[mgmt_class] = NULL;
1691 /* Any other vendor classes left ? */
1692 if (!check_vendor_table(vendor)) {
1695 agent_priv->reg_req->
1696 mgmt_class_version].
1708 static struct ib_mad_agent_private *
1709 find_mad_agent(struct ib_mad_port_private *port_priv,
1710 const struct ib_mad_hdr *mad_hdr)
1712 struct ib_mad_agent_private *mad_agent = NULL;
1713 unsigned long flags;
1715 spin_lock_irqsave(&port_priv->reg_lock, flags);
1716 if (ib_response_mad(mad_hdr)) {
1718 struct ib_mad_agent_private *entry;
1721 * Routing is based on high 32 bits of transaction ID
1724 hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1725 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1726 if (entry->agent.hi_tid == hi_tid) {
1732 struct ib_mad_mgmt_class_table *class;
1733 struct ib_mad_mgmt_method_table *method;
1734 struct ib_mad_mgmt_vendor_class_table *vendor;
1735 struct ib_mad_mgmt_vendor_class *vendor_class;
1736 const struct ib_vendor_mad *vendor_mad;
1740 * Routing is based on version, class, and method
1741 * For "newer" vendor MADs, also based on OUI
1743 if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1745 if (!is_vendor_class(mad_hdr->mgmt_class)) {
1746 class = port_priv->version[
1747 mad_hdr->class_version].class;
1750 if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1751 ARRAY_SIZE(class->method_table))
1753 method = class->method_table[convert_mgmt_class(
1754 mad_hdr->mgmt_class)];
1756 mad_agent = method->agent[mad_hdr->method &
1757 ~IB_MGMT_METHOD_RESP];
1759 vendor = port_priv->version[
1760 mad_hdr->class_version].vendor;
1763 vendor_class = vendor->vendor_class[vendor_class_index(
1764 mad_hdr->mgmt_class)];
1767 /* Find matching OUI */
1768 vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1769 index = find_vendor_oui(vendor_class, vendor_mad->oui);
1772 method = vendor_class->method_table[index];
1774 mad_agent = method->agent[mad_hdr->method &
1775 ~IB_MGMT_METHOD_RESP];
1781 if (mad_agent->agent.recv_handler)
1782 atomic_inc(&mad_agent->refcount);
1784 dev_notice(&port_priv->device->dev,
1785 "No receive handler for client %p on port %d\n",
1786 &mad_agent->agent, port_priv->port_num);
1791 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1796 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1797 const struct ib_mad_qp_info *qp_info,
1801 u32 qp_num = qp_info->qp->qp_num;
1803 /* Make sure MAD base version is understood */
1804 if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1805 (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1806 pr_err("MAD received with unsupported base version %d %s\n",
1807 mad_hdr->base_version, opa ? "(opa)" : "");
1811 /* Filter SMI packets sent to other than QP0 */
1812 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1813 (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1817 /* CM attributes other than ClassPortInfo only use Send method */
1818 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1819 (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1820 (mad_hdr->method != IB_MGMT_METHOD_SEND))
1822 /* Filter GSI packets sent to QP0 */
1831 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1832 const struct ib_mad_hdr *mad_hdr)
1834 struct ib_rmpp_mad *rmpp_mad;
1836 rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1837 return !mad_agent_priv->agent.rmpp_version ||
1838 !ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1839 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1840 IB_MGMT_RMPP_FLAG_ACTIVE) ||
1841 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1844 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1845 const struct ib_mad_recv_wc *rwc)
1847 return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1848 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1851 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1852 const struct ib_mad_send_wr_private *wr,
1853 const struct ib_mad_recv_wc *rwc )
1855 struct ib_ah_attr attr;
1856 u8 send_resp, rcv_resp;
1858 struct ib_device *device = mad_agent_priv->agent.device;
1859 u8 port_num = mad_agent_priv->agent.port_num;
1862 send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1863 rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1865 if (send_resp == rcv_resp)
1866 /* both requests, or both responses. GIDs different */
1869 if (ib_query_ah(wr->send_buf.ah, &attr))
1870 /* Assume not equal, to avoid false positives. */
1873 if (!!(attr.ah_flags & IB_AH_GRH) !=
1874 !!(rwc->wc->wc_flags & IB_WC_GRH))
1875 /* one has GID, other does not. Assume different */
1878 if (!send_resp && rcv_resp) {
1879 /* is request/response. */
1880 if (!(attr.ah_flags & IB_AH_GRH)) {
1881 if (ib_get_cached_lmc(device, port_num, &lmc))
1883 return (!lmc || !((attr.src_path_bits ^
1884 rwc->wc->dlid_path_bits) &
1887 if (ib_get_cached_gid(device, port_num,
1888 attr.grh.sgid_index, &sgid, NULL))
1890 return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1895 if (!(attr.ah_flags & IB_AH_GRH))
1896 return attr.dlid == rwc->wc->slid;
1898 return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1902 static inline int is_direct(u8 class)
1904 return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1907 struct ib_mad_send_wr_private*
1908 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1909 const struct ib_mad_recv_wc *wc)
1911 struct ib_mad_send_wr_private *wr;
1912 const struct ib_mad_hdr *mad_hdr;
1914 mad_hdr = &wc->recv_buf.mad->mad_hdr;
1916 list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1917 if ((wr->tid == mad_hdr->tid) &&
1918 rcv_has_same_class(wr, wc) &&
1920 * Don't check GID for direct routed MADs.
1921 * These might have permissive LIDs.
1923 (is_direct(mad_hdr->mgmt_class) ||
1924 rcv_has_same_gid(mad_agent_priv, wr, wc)))
1925 return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1929 * It's possible to receive the response before we've
1930 * been notified that the send has completed
1932 list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1933 if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1934 wr->tid == mad_hdr->tid &&
1936 rcv_has_same_class(wr, wc) &&
1938 * Don't check GID for direct routed MADs.
1939 * These might have permissive LIDs.
1941 (is_direct(mad_hdr->mgmt_class) ||
1942 rcv_has_same_gid(mad_agent_priv, wr, wc)))
1943 /* Verify request has not been canceled */
1944 return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1949 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1951 mad_send_wr->timeout = 0;
1952 if (mad_send_wr->refcount == 1)
1953 list_move_tail(&mad_send_wr->agent_list,
1954 &mad_send_wr->mad_agent_priv->done_list);
1957 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1958 struct ib_mad_recv_wc *mad_recv_wc)
1960 struct ib_mad_send_wr_private *mad_send_wr;
1961 struct ib_mad_send_wc mad_send_wc;
1962 unsigned long flags;
1964 INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1965 list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1966 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1967 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1970 deref_mad_agent(mad_agent_priv);
1975 /* Complete corresponding request */
1976 if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1977 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1978 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1980 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1981 if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
1982 && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
1983 && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
1984 & IB_MGMT_RMPP_FLAG_ACTIVE)) {
1985 /* user rmpp is in effect
1986 * and this is an active RMPP MAD
1988 mad_recv_wc->wc->wr_id = 0;
1989 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1991 atomic_dec(&mad_agent_priv->refcount);
1993 /* not user rmpp, revert to normal behavior and
1995 ib_free_recv_mad(mad_recv_wc);
1996 deref_mad_agent(mad_agent_priv);
2000 ib_mark_mad_done(mad_send_wr);
2001 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2003 /* Defined behavior is to complete response before request */
2004 mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
2005 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
2007 atomic_dec(&mad_agent_priv->refcount);
2009 mad_send_wc.status = IB_WC_SUCCESS;
2010 mad_send_wc.vendor_err = 0;
2011 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2012 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2015 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
2017 deref_mad_agent(mad_agent_priv);
2021 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2022 const struct ib_mad_qp_info *qp_info,
2023 const struct ib_wc *wc,
2025 struct ib_mad_private *recv,
2026 struct ib_mad_private *response)
2028 enum smi_forward_action retsmi;
2029 struct ib_smp *smp = (struct ib_smp *)recv->mad;
2031 if (smi_handle_dr_smp_recv(smp,
2032 rdma_cap_ib_switch(port_priv->device),
2034 port_priv->device->phys_port_cnt) ==
2036 return IB_SMI_DISCARD;
2038 retsmi = smi_check_forward_dr_smp(smp);
2039 if (retsmi == IB_SMI_LOCAL)
2040 return IB_SMI_HANDLE;
2042 if (retsmi == IB_SMI_SEND) { /* don't forward */
2043 if (smi_handle_dr_smp_send(smp,
2044 rdma_cap_ib_switch(port_priv->device),
2045 port_num) == IB_SMI_DISCARD)
2046 return IB_SMI_DISCARD;
2048 if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2049 return IB_SMI_DISCARD;
2050 } else if (rdma_cap_ib_switch(port_priv->device)) {
2051 /* forward case for switches */
2052 memcpy(response, recv, mad_priv_size(response));
2053 response->header.recv_wc.wc = &response->header.wc;
2054 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2055 response->header.recv_wc.recv_buf.grh = &response->grh;
2057 agent_send_response((const struct ib_mad_hdr *)response->mad,
2060 smi_get_fwd_port(smp),
2061 qp_info->qp->qp_num,
2065 return IB_SMI_DISCARD;
2067 return IB_SMI_HANDLE;
2070 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2071 struct ib_mad_private *response,
2072 size_t *resp_len, bool opa)
2074 const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2075 struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2077 if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2078 recv_hdr->method == IB_MGMT_METHOD_SET) {
2079 memcpy(response, recv, mad_priv_size(response));
2080 response->header.recv_wc.wc = &response->header.wc;
2081 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2082 response->header.recv_wc.recv_buf.grh = &response->grh;
2083 resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2084 resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2085 if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2086 resp_hdr->status |= IB_SMP_DIRECTION;
2088 if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2089 if (recv_hdr->mgmt_class ==
2090 IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2091 recv_hdr->mgmt_class ==
2092 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2093 *resp_len = opa_get_smp_header_size(
2094 (struct opa_smp *)recv->mad);
2096 *resp_len = sizeof(struct ib_mad_hdr);
2105 static enum smi_action
2106 handle_opa_smi(struct ib_mad_port_private *port_priv,
2107 struct ib_mad_qp_info *qp_info,
2110 struct ib_mad_private *recv,
2111 struct ib_mad_private *response)
2113 enum smi_forward_action retsmi;
2114 struct opa_smp *smp = (struct opa_smp *)recv->mad;
2116 if (opa_smi_handle_dr_smp_recv(smp,
2117 rdma_cap_ib_switch(port_priv->device),
2119 port_priv->device->phys_port_cnt) ==
2121 return IB_SMI_DISCARD;
2123 retsmi = opa_smi_check_forward_dr_smp(smp);
2124 if (retsmi == IB_SMI_LOCAL)
2125 return IB_SMI_HANDLE;
2127 if (retsmi == IB_SMI_SEND) { /* don't forward */
2128 if (opa_smi_handle_dr_smp_send(smp,
2129 rdma_cap_ib_switch(port_priv->device),
2130 port_num) == IB_SMI_DISCARD)
2131 return IB_SMI_DISCARD;
2133 if (opa_smi_check_local_smp(smp, port_priv->device) ==
2135 return IB_SMI_DISCARD;
2137 } else if (rdma_cap_ib_switch(port_priv->device)) {
2138 /* forward case for switches */
2139 memcpy(response, recv, mad_priv_size(response));
2140 response->header.recv_wc.wc = &response->header.wc;
2141 response->header.recv_wc.recv_buf.opa_mad =
2142 (struct opa_mad *)response->mad;
2143 response->header.recv_wc.recv_buf.grh = &response->grh;
2145 agent_send_response((const struct ib_mad_hdr *)response->mad,
2148 opa_smi_get_fwd_port(smp),
2149 qp_info->qp->qp_num,
2150 recv->header.wc.byte_len,
2153 return IB_SMI_DISCARD;
2156 return IB_SMI_HANDLE;
2159 static enum smi_action
2160 handle_smi(struct ib_mad_port_private *port_priv,
2161 struct ib_mad_qp_info *qp_info,
2164 struct ib_mad_private *recv,
2165 struct ib_mad_private *response,
2168 struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2170 if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2171 mad_hdr->class_version == OPA_SMI_CLASS_VERSION)
2172 return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2175 return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2178 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
2181 struct ib_mad_qp_info *qp_info;
2182 struct ib_mad_private_header *mad_priv_hdr;
2183 struct ib_mad_private *recv, *response = NULL;
2184 struct ib_mad_list_head *mad_list;
2185 struct ib_mad_agent_private *mad_agent;
2187 int ret = IB_MAD_RESULT_SUCCESS;
2189 u16 resp_mad_pkey_index = 0;
2192 mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2193 qp_info = mad_list->mad_queue->qp_info;
2194 dequeue_mad(mad_list);
2196 opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2197 qp_info->port_priv->port_num);
2199 mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2201 recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2202 ib_dma_unmap_single(port_priv->device,
2203 recv->header.mapping,
2204 mad_priv_dma_size(recv),
2207 /* Setup MAD receive work completion from "normal" work completion */
2208 recv->header.wc = *wc;
2209 recv->header.recv_wc.wc = &recv->header.wc;
2211 if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2212 recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2213 recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2215 recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2216 recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2219 recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2220 recv->header.recv_wc.recv_buf.grh = &recv->grh;
2222 if (atomic_read(&qp_info->snoop_count))
2223 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2226 if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2229 mad_size = recv->mad_size;
2230 response = alloc_mad_private(mad_size, GFP_KERNEL);
2232 dev_err(&port_priv->device->dev,
2233 "ib_mad_recv_done_handler no memory for response buffer\n");
2237 if (rdma_cap_ib_switch(port_priv->device))
2238 port_num = wc->port_num;
2240 port_num = port_priv->port_num;
2242 if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2243 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2244 if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2250 /* Give driver "right of first refusal" on incoming MAD */
2251 if (port_priv->device->process_mad) {
2252 ret = port_priv->device->process_mad(port_priv->device, 0,
2253 port_priv->port_num,
2255 (const struct ib_mad_hdr *)recv->mad,
2257 (struct ib_mad_hdr *)response->mad,
2258 &mad_size, &resp_mad_pkey_index);
2261 wc->pkey_index = resp_mad_pkey_index;
2263 if (ret & IB_MAD_RESULT_SUCCESS) {
2264 if (ret & IB_MAD_RESULT_CONSUMED)
2266 if (ret & IB_MAD_RESULT_REPLY) {
2267 agent_send_response((const struct ib_mad_hdr *)response->mad,
2271 qp_info->qp->qp_num,
2278 mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2280 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2282 * recv is freed up in error cases in ib_mad_complete_recv
2283 * or via recv_handler in ib_mad_complete_recv()
2286 } else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2287 generate_unmatched_resp(recv, response, &mad_size, opa)) {
2288 agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2289 port_priv->device, port_num,
2290 qp_info->qp->qp_num, mad_size, opa);
2294 /* Post another receive request for this QP */
2296 ib_mad_post_receive_mads(qp_info, response);
2299 ib_mad_post_receive_mads(qp_info, recv);
2302 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2304 struct ib_mad_send_wr_private *mad_send_wr;
2305 unsigned long delay;
2307 if (list_empty(&mad_agent_priv->wait_list)) {
2308 cancel_delayed_work(&mad_agent_priv->timed_work);
2310 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2311 struct ib_mad_send_wr_private,
2314 if (time_after(mad_agent_priv->timeout,
2315 mad_send_wr->timeout)) {
2316 mad_agent_priv->timeout = mad_send_wr->timeout;
2317 delay = mad_send_wr->timeout - jiffies;
2318 if ((long)delay <= 0)
2320 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2321 &mad_agent_priv->timed_work, delay);
2326 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2328 struct ib_mad_agent_private *mad_agent_priv;
2329 struct ib_mad_send_wr_private *temp_mad_send_wr;
2330 struct list_head *list_item;
2331 unsigned long delay;
2333 mad_agent_priv = mad_send_wr->mad_agent_priv;
2334 list_del(&mad_send_wr->agent_list);
2336 delay = mad_send_wr->timeout;
2337 mad_send_wr->timeout += jiffies;
2340 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2341 temp_mad_send_wr = list_entry(list_item,
2342 struct ib_mad_send_wr_private,
2344 if (time_after(mad_send_wr->timeout,
2345 temp_mad_send_wr->timeout))
2350 list_item = &mad_agent_priv->wait_list;
2351 list_add(&mad_send_wr->agent_list, list_item);
2353 /* Reschedule a work item if we have a shorter timeout */
2354 if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2355 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2356 &mad_agent_priv->timed_work, delay);
2359 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2362 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2363 wait_for_response(mad_send_wr);
2367 * Process a send work completion
2369 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2370 struct ib_mad_send_wc *mad_send_wc)
2372 struct ib_mad_agent_private *mad_agent_priv;
2373 unsigned long flags;
2376 mad_agent_priv = mad_send_wr->mad_agent_priv;
2377 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2378 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2379 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2380 if (ret == IB_RMPP_RESULT_CONSUMED)
2383 ret = IB_RMPP_RESULT_UNHANDLED;
2385 if (mad_send_wc->status != IB_WC_SUCCESS &&
2386 mad_send_wr->status == IB_WC_SUCCESS) {
2387 mad_send_wr->status = mad_send_wc->status;
2388 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2391 if (--mad_send_wr->refcount > 0) {
2392 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2393 mad_send_wr->status == IB_WC_SUCCESS) {
2394 wait_for_response(mad_send_wr);
2399 /* Remove send from MAD agent and notify client of completion */
2400 list_del(&mad_send_wr->agent_list);
2401 adjust_timeout(mad_agent_priv);
2402 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2404 if (mad_send_wr->status != IB_WC_SUCCESS )
2405 mad_send_wc->status = mad_send_wr->status;
2406 if (ret == IB_RMPP_RESULT_INTERNAL)
2407 ib_rmpp_send_handler(mad_send_wc);
2409 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2412 /* Release reference on agent taken when sending */
2413 deref_mad_agent(mad_agent_priv);
2416 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2419 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2422 struct ib_mad_send_wr_private *mad_send_wr, *queued_send_wr;
2423 struct ib_mad_list_head *mad_list;
2424 struct ib_mad_qp_info *qp_info;
2425 struct ib_mad_queue *send_queue;
2426 struct ib_send_wr *bad_send_wr;
2427 struct ib_mad_send_wc mad_send_wc;
2428 unsigned long flags;
2431 mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2432 mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2434 send_queue = mad_list->mad_queue;
2435 qp_info = send_queue->qp_info;
2438 ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2439 mad_send_wr->header_mapping,
2440 mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2441 ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2442 mad_send_wr->payload_mapping,
2443 mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2444 queued_send_wr = NULL;
2445 spin_lock_irqsave(&send_queue->lock, flags);
2446 list_del(&mad_list->list);
2448 /* Move queued send to the send queue */
2449 if (send_queue->count-- > send_queue->max_active) {
2450 mad_list = container_of(qp_info->overflow_list.next,
2451 struct ib_mad_list_head, list);
2452 queued_send_wr = container_of(mad_list,
2453 struct ib_mad_send_wr_private,
2455 list_move_tail(&mad_list->list, &send_queue->list);
2457 spin_unlock_irqrestore(&send_queue->lock, flags);
2459 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2460 mad_send_wc.status = wc->status;
2461 mad_send_wc.vendor_err = wc->vendor_err;
2462 if (atomic_read(&qp_info->snoop_count))
2463 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2464 IB_MAD_SNOOP_SEND_COMPLETIONS);
2465 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2467 if (queued_send_wr) {
2468 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2471 dev_err(&port_priv->device->dev,
2472 "ib_post_send failed: %d\n", ret);
2473 mad_send_wr = queued_send_wr;
2474 wc->status = IB_WC_LOC_QP_OP_ERR;
2480 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2482 struct ib_mad_send_wr_private *mad_send_wr;
2483 struct ib_mad_list_head *mad_list;
2484 unsigned long flags;
2486 spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2487 list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2488 mad_send_wr = container_of(mad_list,
2489 struct ib_mad_send_wr_private,
2491 mad_send_wr->retry = 1;
2493 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2496 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2499 struct ib_mad_list_head *mad_list;
2500 struct ib_mad_qp_info *qp_info;
2501 struct ib_mad_send_wr_private *mad_send_wr;
2504 /* Determine if failure was a send or receive */
2505 mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2506 qp_info = mad_list->mad_queue->qp_info;
2507 if (mad_list->mad_queue == &qp_info->recv_queue)
2509 * Receive errors indicate that the QP has entered the error
2510 * state - error handling/shutdown code will cleanup
2515 * Send errors will transition the QP to SQE - move
2516 * QP to RTS and repost flushed work requests
2518 mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2520 if (wc->status == IB_WC_WR_FLUSH_ERR) {
2521 if (mad_send_wr->retry) {
2523 struct ib_send_wr *bad_send_wr;
2525 mad_send_wr->retry = 0;
2526 ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2529 ib_mad_send_done_handler(port_priv, wc);
2531 ib_mad_send_done_handler(port_priv, wc);
2533 struct ib_qp_attr *attr;
2535 /* Transition QP to RTS and fail offending send */
2536 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2538 attr->qp_state = IB_QPS_RTS;
2539 attr->cur_qp_state = IB_QPS_SQE;
2540 ret = ib_modify_qp(qp_info->qp, attr,
2541 IB_QP_STATE | IB_QP_CUR_STATE);
2544 dev_err(&port_priv->device->dev,
2545 "mad_error_handler - ib_modify_qp to RTS : %d\n",
2548 mark_sends_for_retry(qp_info);
2550 ib_mad_send_done_handler(port_priv, wc);
2555 * IB MAD completion callback
2557 static void ib_mad_completion_handler(struct work_struct *work)
2559 struct ib_mad_port_private *port_priv;
2562 port_priv = container_of(work, struct ib_mad_port_private, work);
2563 ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2565 while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2566 if (wc.status == IB_WC_SUCCESS) {
2567 switch (wc.opcode) {
2569 ib_mad_send_done_handler(port_priv, &wc);
2572 ib_mad_recv_done_handler(port_priv, &wc);
2579 mad_error_handler(port_priv, &wc);
2583 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2585 unsigned long flags;
2586 struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2587 struct ib_mad_send_wc mad_send_wc;
2588 struct list_head cancel_list;
2590 INIT_LIST_HEAD(&cancel_list);
2592 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2593 list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2594 &mad_agent_priv->send_list, agent_list) {
2595 if (mad_send_wr->status == IB_WC_SUCCESS) {
2596 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2597 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2601 /* Empty wait list to prevent receives from finding a request */
2602 list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2603 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2605 /* Report all cancelled requests */
2606 mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2607 mad_send_wc.vendor_err = 0;
2609 list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2610 &cancel_list, agent_list) {
2611 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2612 list_del(&mad_send_wr->agent_list);
2613 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2615 atomic_dec(&mad_agent_priv->refcount);
2619 static struct ib_mad_send_wr_private*
2620 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2621 struct ib_mad_send_buf *send_buf)
2623 struct ib_mad_send_wr_private *mad_send_wr;
2625 list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2627 if (&mad_send_wr->send_buf == send_buf)
2631 list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2633 if (is_rmpp_data_mad(mad_agent_priv,
2634 mad_send_wr->send_buf.mad) &&
2635 &mad_send_wr->send_buf == send_buf)
2641 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2642 struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2644 struct ib_mad_agent_private *mad_agent_priv;
2645 struct ib_mad_send_wr_private *mad_send_wr;
2646 unsigned long flags;
2649 mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2651 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2652 mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2653 if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2654 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2658 active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2660 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2661 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2664 mad_send_wr->send_buf.timeout_ms = timeout_ms;
2666 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2668 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2670 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2673 EXPORT_SYMBOL(ib_modify_mad);
2675 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2676 struct ib_mad_send_buf *send_buf)
2678 ib_modify_mad(mad_agent, send_buf, 0);
2680 EXPORT_SYMBOL(ib_cancel_mad);
2682 static void local_completions(struct work_struct *work)
2684 struct ib_mad_agent_private *mad_agent_priv;
2685 struct ib_mad_local_private *local;
2686 struct ib_mad_agent_private *recv_mad_agent;
2687 unsigned long flags;
2690 struct ib_mad_send_wc mad_send_wc;
2694 container_of(work, struct ib_mad_agent_private, local_work);
2696 opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2697 mad_agent_priv->qp_info->port_priv->port_num);
2699 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2700 while (!list_empty(&mad_agent_priv->local_list)) {
2701 local = list_entry(mad_agent_priv->local_list.next,
2702 struct ib_mad_local_private,
2704 list_del(&local->completion_list);
2705 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2707 if (local->mad_priv) {
2709 recv_mad_agent = local->recv_mad_agent;
2710 if (!recv_mad_agent) {
2711 dev_err(&mad_agent_priv->agent.device->dev,
2712 "No receive MAD agent for local completion\n");
2714 goto local_send_completion;
2718 * Defined behavior is to complete response
2721 build_smp_wc(recv_mad_agent->agent.qp,
2722 (unsigned long) local->mad_send_wr,
2723 be16_to_cpu(IB_LID_PERMISSIVE),
2724 local->mad_send_wr->send_wr.pkey_index,
2725 recv_mad_agent->agent.port_num, &wc);
2727 local->mad_priv->header.recv_wc.wc = &wc;
2729 base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2730 if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2731 local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2732 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2734 local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2735 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2738 INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2739 list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2740 &local->mad_priv->header.recv_wc.rmpp_list);
2741 local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2742 local->mad_priv->header.recv_wc.recv_buf.mad =
2743 (struct ib_mad *)local->mad_priv->mad;
2744 if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2745 snoop_recv(recv_mad_agent->qp_info,
2746 &local->mad_priv->header.recv_wc,
2747 IB_MAD_SNOOP_RECVS);
2748 recv_mad_agent->agent.recv_handler(
2749 &recv_mad_agent->agent,
2750 &local->mad_priv->header.recv_wc);
2751 spin_lock_irqsave(&recv_mad_agent->lock, flags);
2752 atomic_dec(&recv_mad_agent->refcount);
2753 spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2756 local_send_completion:
2758 mad_send_wc.status = IB_WC_SUCCESS;
2759 mad_send_wc.vendor_err = 0;
2760 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2761 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2762 snoop_send(mad_agent_priv->qp_info,
2763 &local->mad_send_wr->send_buf,
2764 &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2765 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2768 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2769 atomic_dec(&mad_agent_priv->refcount);
2771 kfree(local->mad_priv);
2774 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2777 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2781 if (!mad_send_wr->retries_left)
2784 mad_send_wr->retries_left--;
2785 mad_send_wr->send_buf.retries++;
2787 mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2789 if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2790 ret = ib_retry_rmpp(mad_send_wr);
2792 case IB_RMPP_RESULT_UNHANDLED:
2793 ret = ib_send_mad(mad_send_wr);
2795 case IB_RMPP_RESULT_CONSUMED:
2803 ret = ib_send_mad(mad_send_wr);
2806 mad_send_wr->refcount++;
2807 list_add_tail(&mad_send_wr->agent_list,
2808 &mad_send_wr->mad_agent_priv->send_list);
2813 static void timeout_sends(struct work_struct *work)
2815 struct ib_mad_agent_private *mad_agent_priv;
2816 struct ib_mad_send_wr_private *mad_send_wr;
2817 struct ib_mad_send_wc mad_send_wc;
2818 unsigned long flags, delay;
2820 mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2822 mad_send_wc.vendor_err = 0;
2824 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2825 while (!list_empty(&mad_agent_priv->wait_list)) {
2826 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2827 struct ib_mad_send_wr_private,
2830 if (time_after(mad_send_wr->timeout, jiffies)) {
2831 delay = mad_send_wr->timeout - jiffies;
2832 if ((long)delay <= 0)
2834 queue_delayed_work(mad_agent_priv->qp_info->
2836 &mad_agent_priv->timed_work, delay);
2840 list_del(&mad_send_wr->agent_list);
2841 if (mad_send_wr->status == IB_WC_SUCCESS &&
2842 !retry_send(mad_send_wr))
2845 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2847 if (mad_send_wr->status == IB_WC_SUCCESS)
2848 mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2850 mad_send_wc.status = mad_send_wr->status;
2851 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2852 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2855 atomic_dec(&mad_agent_priv->refcount);
2856 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2858 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2861 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2863 struct ib_mad_port_private *port_priv = cq->cq_context;
2864 unsigned long flags;
2866 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2867 if (!list_empty(&port_priv->port_list))
2868 queue_work(port_priv->wq, &port_priv->work);
2869 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2873 * Allocate receive MADs and post receive WRs for them
2875 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2876 struct ib_mad_private *mad)
2878 unsigned long flags;
2880 struct ib_mad_private *mad_priv;
2881 struct ib_sge sg_list;
2882 struct ib_recv_wr recv_wr, *bad_recv_wr;
2883 struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2885 /* Initialize common scatter list fields */
2886 sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2888 /* Initialize common receive WR fields */
2889 recv_wr.next = NULL;
2890 recv_wr.sg_list = &sg_list;
2891 recv_wr.num_sge = 1;
2894 /* Allocate and map receive buffer */
2899 mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2902 dev_err(&qp_info->port_priv->device->dev,
2903 "No memory for receive buffer\n");
2908 sg_list.length = mad_priv_dma_size(mad_priv);
2909 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2911 mad_priv_dma_size(mad_priv),
2913 if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2919 mad_priv->header.mapping = sg_list.addr;
2920 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2921 mad_priv->header.mad_list.mad_queue = recv_queue;
2923 /* Post receive WR */
2924 spin_lock_irqsave(&recv_queue->lock, flags);
2925 post = (++recv_queue->count < recv_queue->max_active);
2926 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2927 spin_unlock_irqrestore(&recv_queue->lock, flags);
2928 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2930 spin_lock_irqsave(&recv_queue->lock, flags);
2931 list_del(&mad_priv->header.mad_list.list);
2932 recv_queue->count--;
2933 spin_unlock_irqrestore(&recv_queue->lock, flags);
2934 ib_dma_unmap_single(qp_info->port_priv->device,
2935 mad_priv->header.mapping,
2936 mad_priv_dma_size(mad_priv),
2939 dev_err(&qp_info->port_priv->device->dev,
2940 "ib_post_recv failed: %d\n", ret);
2949 * Return all the posted receive MADs
2951 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2953 struct ib_mad_private_header *mad_priv_hdr;
2954 struct ib_mad_private *recv;
2955 struct ib_mad_list_head *mad_list;
2960 while (!list_empty(&qp_info->recv_queue.list)) {
2962 mad_list = list_entry(qp_info->recv_queue.list.next,
2963 struct ib_mad_list_head, list);
2964 mad_priv_hdr = container_of(mad_list,
2965 struct ib_mad_private_header,
2967 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2970 /* Remove from posted receive MAD list */
2971 list_del(&mad_list->list);
2973 ib_dma_unmap_single(qp_info->port_priv->device,
2974 recv->header.mapping,
2975 mad_priv_dma_size(recv),
2980 qp_info->recv_queue.count = 0;
2986 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2989 struct ib_qp_attr *attr;
2993 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2995 dev_err(&port_priv->device->dev,
2996 "Couldn't kmalloc ib_qp_attr\n");
3000 ret = ib_find_pkey(port_priv->device, port_priv->port_num,
3001 IB_DEFAULT_PKEY_FULL, &pkey_index);
3005 for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3006 qp = port_priv->qp_info[i].qp;
3011 * PKey index for QP1 is irrelevant but
3012 * one is needed for the Reset to Init transition
3014 attr->qp_state = IB_QPS_INIT;
3015 attr->pkey_index = pkey_index;
3016 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
3017 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
3018 IB_QP_PKEY_INDEX | IB_QP_QKEY);
3020 dev_err(&port_priv->device->dev,
3021 "Couldn't change QP%d state to INIT: %d\n",
3026 attr->qp_state = IB_QPS_RTR;
3027 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
3029 dev_err(&port_priv->device->dev,
3030 "Couldn't change QP%d state to RTR: %d\n",
3035 attr->qp_state = IB_QPS_RTS;
3036 attr->sq_psn = IB_MAD_SEND_Q_PSN;
3037 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3039 dev_err(&port_priv->device->dev,
3040 "Couldn't change QP%d state to RTS: %d\n",
3046 ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3048 dev_err(&port_priv->device->dev,
3049 "Failed to request completion notification: %d\n",
3054 for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3055 if (!port_priv->qp_info[i].qp)
3058 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3060 dev_err(&port_priv->device->dev,
3061 "Couldn't post receive WRs\n");
3070 static void qp_event_handler(struct ib_event *event, void *qp_context)
3072 struct ib_mad_qp_info *qp_info = qp_context;
3074 /* It's worse than that! He's dead, Jim! */
3075 dev_err(&qp_info->port_priv->device->dev,
3076 "Fatal error (%d) on MAD QP (%d)\n",
3077 event->event, qp_info->qp->qp_num);
3080 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3081 struct ib_mad_queue *mad_queue)
3083 mad_queue->qp_info = qp_info;
3084 mad_queue->count = 0;
3085 spin_lock_init(&mad_queue->lock);
3086 INIT_LIST_HEAD(&mad_queue->list);
3089 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3090 struct ib_mad_qp_info *qp_info)
3092 qp_info->port_priv = port_priv;
3093 init_mad_queue(qp_info, &qp_info->send_queue);
3094 init_mad_queue(qp_info, &qp_info->recv_queue);
3095 INIT_LIST_HEAD(&qp_info->overflow_list);
3096 spin_lock_init(&qp_info->snoop_lock);
3097 qp_info->snoop_table = NULL;
3098 qp_info->snoop_table_size = 0;
3099 atomic_set(&qp_info->snoop_count, 0);
3102 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3103 enum ib_qp_type qp_type)
3105 struct ib_qp_init_attr qp_init_attr;
3108 memset(&qp_init_attr, 0, sizeof qp_init_attr);
3109 qp_init_attr.send_cq = qp_info->port_priv->cq;
3110 qp_init_attr.recv_cq = qp_info->port_priv->cq;
3111 qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3112 qp_init_attr.cap.max_send_wr = mad_sendq_size;
3113 qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3114 qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3115 qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3116 qp_init_attr.qp_type = qp_type;
3117 qp_init_attr.port_num = qp_info->port_priv->port_num;
3118 qp_init_attr.qp_context = qp_info;
3119 qp_init_attr.event_handler = qp_event_handler;
3120 qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3121 if (IS_ERR(qp_info->qp)) {
3122 dev_err(&qp_info->port_priv->device->dev,
3123 "Couldn't create ib_mad QP%d\n",
3124 get_spl_qp_index(qp_type));
3125 ret = PTR_ERR(qp_info->qp);
3128 /* Use minimum queue sizes unless the CQ is resized */
3129 qp_info->send_queue.max_active = mad_sendq_size;
3130 qp_info->recv_queue.max_active = mad_recvq_size;
3137 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3142 ib_destroy_qp(qp_info->qp);
3143 kfree(qp_info->snoop_table);
3148 * Create the QP, PD, MR, and CQ if needed
3150 static int ib_mad_port_open(struct ib_device *device,
3154 struct ib_mad_port_private *port_priv;
3155 unsigned long flags;
3156 char name[sizeof "ib_mad123"];
3158 struct ib_cq_init_attr cq_attr = {};
3160 if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3163 if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3164 rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3167 /* Create new device info */
3168 port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3170 dev_err(&device->dev, "No memory for ib_mad_port_private\n");
3174 port_priv->device = device;
3175 port_priv->port_num = port_num;
3176 spin_lock_init(&port_priv->reg_lock);
3177 INIT_LIST_HEAD(&port_priv->agent_list);
3178 init_mad_qp(port_priv, &port_priv->qp_info[0]);
3179 init_mad_qp(port_priv, &port_priv->qp_info[1]);
3181 cq_size = mad_sendq_size + mad_recvq_size;
3182 has_smi = rdma_cap_ib_smi(device, port_num);
3186 cq_attr.cqe = cq_size;
3187 port_priv->cq = ib_create_cq(port_priv->device,
3188 ib_mad_thread_completion_handler,
3189 NULL, port_priv, &cq_attr);
3190 if (IS_ERR(port_priv->cq)) {
3191 dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3192 ret = PTR_ERR(port_priv->cq);
3196 port_priv->pd = ib_alloc_pd(device);
3197 if (IS_ERR(port_priv->pd)) {
3198 dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3199 ret = PTR_ERR(port_priv->pd);
3204 ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3208 ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3212 snprintf(name, sizeof name, "ib_mad%d", port_num);
3213 port_priv->wq = create_singlethread_workqueue(name);
3214 if (!port_priv->wq) {
3218 INIT_WORK(&port_priv->work, ib_mad_completion_handler);
3220 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3221 list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3222 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3224 ret = ib_mad_port_start(port_priv);
3226 dev_err(&device->dev, "Couldn't start port\n");
3233 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3234 list_del_init(&port_priv->port_list);
3235 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3237 destroy_workqueue(port_priv->wq);
3239 destroy_mad_qp(&port_priv->qp_info[1]);
3241 destroy_mad_qp(&port_priv->qp_info[0]);
3243 ib_dealloc_pd(port_priv->pd);
3245 ib_destroy_cq(port_priv->cq);
3246 cleanup_recv_queue(&port_priv->qp_info[1]);
3247 cleanup_recv_queue(&port_priv->qp_info[0]);
3256 * If there are no classes using the port, free the port
3257 * resources (CQ, MR, PD, QP) and remove the port's info structure
3259 static int ib_mad_port_close(struct ib_device *device, int port_num)
3261 struct ib_mad_port_private *port_priv;
3262 unsigned long flags;
3264 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3265 port_priv = __ib_get_mad_port(device, port_num);
3266 if (port_priv == NULL) {
3267 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3268 dev_err(&device->dev, "Port %d not found\n", port_num);
3271 list_del_init(&port_priv->port_list);
3272 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3274 destroy_workqueue(port_priv->wq);
3275 destroy_mad_qp(&port_priv->qp_info[1]);
3276 destroy_mad_qp(&port_priv->qp_info[0]);
3277 ib_dealloc_pd(port_priv->pd);
3278 ib_destroy_cq(port_priv->cq);
3279 cleanup_recv_queue(&port_priv->qp_info[1]);
3280 cleanup_recv_queue(&port_priv->qp_info[0]);
3281 /* XXX: Handle deallocation of MAD registration tables */
3288 static void ib_mad_init_device(struct ib_device *device)
3292 start = rdma_start_port(device);
3294 for (i = start; i <= rdma_end_port(device); i++) {
3295 if (!rdma_cap_ib_mad(device, i))
3298 if (ib_mad_port_open(device, i)) {
3299 dev_err(&device->dev, "Couldn't open port %d\n", i);
3302 if (ib_agent_port_open(device, i)) {
3303 dev_err(&device->dev,
3304 "Couldn't open port %d for agents\n", i);
3311 if (ib_mad_port_close(device, i))
3312 dev_err(&device->dev, "Couldn't close port %d\n", i);
3315 while (--i >= start) {
3316 if (!rdma_cap_ib_mad(device, i))
3319 if (ib_agent_port_close(device, i))
3320 dev_err(&device->dev,
3321 "Couldn't close port %d for agents\n", i);
3322 if (ib_mad_port_close(device, i))
3323 dev_err(&device->dev, "Couldn't close port %d\n", i);
3327 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3331 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3332 if (!rdma_cap_ib_mad(device, i))
3335 if (ib_agent_port_close(device, i))
3336 dev_err(&device->dev,
3337 "Couldn't close port %d for agents\n", i);
3338 if (ib_mad_port_close(device, i))
3339 dev_err(&device->dev, "Couldn't close port %d\n", i);
3343 static struct ib_client mad_client = {
3345 .add = ib_mad_init_device,
3346 .remove = ib_mad_remove_device
3349 static int __init ib_mad_init_module(void)
3351 mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3352 mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3354 mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3355 mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3357 INIT_LIST_HEAD(&ib_mad_port_list);
3359 if (ib_register_client(&mad_client)) {
3360 pr_err("Couldn't register ib_mad client\n");
3367 static void __exit ib_mad_cleanup_module(void)
3369 ib_unregister_client(&mad_client);
3372 module_init(ib_mad_init_module);
3373 module_exit(ib_mad_cleanup_module);