2 * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 * $Id: mad.c 1389 2004-12-27 22:56:47Z roland $
35 #include <linux/dma-mapping.h>
41 MODULE_LICENSE("Dual BSD/GPL");
42 MODULE_DESCRIPTION("kernel IB MAD API");
43 MODULE_AUTHOR("Hal Rosenstock");
44 MODULE_AUTHOR("Sean Hefty");
47 kmem_cache_t *ib_mad_cache;
48 static struct list_head ib_mad_port_list;
49 static u32 ib_mad_client_id = 0;
52 static spinlock_t ib_mad_port_list_lock;
55 /* Forward declarations */
56 static int method_in_use(struct ib_mad_mgmt_method_table **method,
57 struct ib_mad_reg_req *mad_reg_req);
58 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
59 static struct ib_mad_agent_private *find_mad_agent(
60 struct ib_mad_port_private *port_priv,
61 struct ib_mad *mad, int solicited);
62 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
63 struct ib_mad_private *mad);
64 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
65 static void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
66 struct ib_mad_send_wc *mad_send_wc);
67 static void timeout_sends(void *data);
68 static void cancel_sends(void *data);
69 static void local_completions(void *data);
70 static int solicited_mad(struct ib_mad *mad);
71 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
72 struct ib_mad_agent_private *agent_priv,
74 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
75 struct ib_mad_agent_private *agent_priv);
78 * Returns a ib_mad_port_private structure or NULL for a device/port
79 * Assumes ib_mad_port_list_lock is being held
81 static inline struct ib_mad_port_private *
82 __ib_get_mad_port(struct ib_device *device, int port_num)
84 struct ib_mad_port_private *entry;
86 list_for_each_entry(entry, &ib_mad_port_list, port_list) {
87 if (entry->device == device && entry->port_num == port_num)
94 * Wrapper function to return a ib_mad_port_private structure or NULL
97 static inline struct ib_mad_port_private *
98 ib_get_mad_port(struct ib_device *device, int port_num)
100 struct ib_mad_port_private *entry;
103 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
104 entry = __ib_get_mad_port(device, port_num);
105 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
110 static inline u8 convert_mgmt_class(u8 mgmt_class)
112 /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
113 return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
117 static int get_spl_qp_index(enum ib_qp_type qp_type)
130 static int vendor_class_index(u8 mgmt_class)
132 return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
135 static int is_vendor_class(u8 mgmt_class)
137 if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
138 (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
143 static int is_vendor_oui(char *oui)
145 if (oui[0] || oui[1] || oui[2])
150 static int is_vendor_method_in_use(
151 struct ib_mad_mgmt_vendor_class *vendor_class,
152 struct ib_mad_reg_req *mad_reg_req)
154 struct ib_mad_mgmt_method_table *method;
157 for (i = 0; i < MAX_MGMT_OUI; i++) {
158 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
159 method = vendor_class->method_table[i];
161 if (method_in_use(&method, mad_reg_req))
172 * ib_register_mad_agent - Register to send/receive MADs
174 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
176 enum ib_qp_type qp_type,
177 struct ib_mad_reg_req *mad_reg_req,
179 ib_mad_send_handler send_handler,
180 ib_mad_recv_handler recv_handler,
183 struct ib_mad_port_private *port_priv;
184 struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
185 struct ib_mad_agent_private *mad_agent_priv;
186 struct ib_mad_reg_req *reg_req = NULL;
187 struct ib_mad_mgmt_class_table *class;
188 struct ib_mad_mgmt_vendor_class_table *vendor;
189 struct ib_mad_mgmt_vendor_class *vendor_class;
190 struct ib_mad_mgmt_method_table *method;
193 u8 mgmt_class, vclass;
195 /* Validate parameters */
196 qpn = get_spl_qp_index(qp_type);
201 goto error1; /* XXX: until RMPP implemented */
203 /* Validate MAD registration request if supplied */
205 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
209 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
211 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
212 * one in this range currently allowed
214 if (mad_reg_req->mgmt_class !=
215 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
217 } else if (mad_reg_req->mgmt_class == 0) {
219 * Class 0 is reserved in IBA and is used for
220 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
223 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
225 * If class is in "new" vendor range,
226 * ensure supplied OUI is not zero
228 if (!is_vendor_oui(mad_reg_req->oui))
231 /* Make sure class supplied is consistent with QP type */
232 if (qp_type == IB_QPT_SMI) {
233 if ((mad_reg_req->mgmt_class !=
234 IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
235 (mad_reg_req->mgmt_class !=
236 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
239 if ((mad_reg_req->mgmt_class ==
240 IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
241 (mad_reg_req->mgmt_class ==
242 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
246 /* No registration request supplied */
251 /* Validate device and port */
252 port_priv = ib_get_mad_port(device, port_num);
254 ret = ERR_PTR(-ENODEV);
258 /* Allocate structures */
259 mad_agent_priv = kmalloc(sizeof *mad_agent_priv, GFP_KERNEL);
260 if (!mad_agent_priv) {
261 ret = ERR_PTR(-ENOMEM);
264 memset(mad_agent_priv, 0, sizeof *mad_agent_priv);
266 mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
267 IB_ACCESS_LOCAL_WRITE);
268 if (IS_ERR(mad_agent_priv->agent.mr)) {
269 ret = ERR_PTR(-ENOMEM);
274 reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
276 ret = ERR_PTR(-ENOMEM);
279 /* Make a copy of the MAD registration request */
280 memcpy(reg_req, mad_reg_req, sizeof *reg_req);
283 /* Now, fill in the various structures */
284 mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
285 mad_agent_priv->reg_req = reg_req;
286 mad_agent_priv->rmpp_version = rmpp_version;
287 mad_agent_priv->agent.device = device;
288 mad_agent_priv->agent.recv_handler = recv_handler;
289 mad_agent_priv->agent.send_handler = send_handler;
290 mad_agent_priv->agent.context = context;
291 mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
292 mad_agent_priv->agent.port_num = port_num;
294 spin_lock_irqsave(&port_priv->reg_lock, flags);
295 mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
298 * Make sure MAD registration (if supplied)
299 * is non overlapping with any existing ones
302 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
303 if (!is_vendor_class(mgmt_class)) {
304 class = port_priv->version[mad_reg_req->
305 mgmt_class_version].class;
307 method = class->method_table[mgmt_class];
309 if (method_in_use(&method,
314 ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
317 /* "New" vendor class range */
318 vendor = port_priv->version[mad_reg_req->
319 mgmt_class_version].vendor;
321 vclass = vendor_class_index(mgmt_class);
322 vendor_class = vendor->vendor_class[vclass];
324 if (is_vendor_method_in_use(
330 ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
338 /* Add mad agent into port's agent list */
339 list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
340 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
342 spin_lock_init(&mad_agent_priv->lock);
343 INIT_LIST_HEAD(&mad_agent_priv->send_list);
344 INIT_LIST_HEAD(&mad_agent_priv->wait_list);
345 INIT_WORK(&mad_agent_priv->timed_work, timeout_sends, mad_agent_priv);
346 INIT_LIST_HEAD(&mad_agent_priv->local_list);
347 INIT_WORK(&mad_agent_priv->local_work, local_completions,
349 INIT_LIST_HEAD(&mad_agent_priv->canceled_list);
350 INIT_WORK(&mad_agent_priv->canceled_work, cancel_sends, mad_agent_priv);
351 atomic_set(&mad_agent_priv->refcount, 1);
352 init_waitqueue_head(&mad_agent_priv->wait);
354 return &mad_agent_priv->agent;
357 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
360 kfree(mad_agent_priv);
362 ib_dereg_mr(mad_agent_priv->agent.mr);
366 EXPORT_SYMBOL(ib_register_mad_agent);
368 static inline int is_snooping_sends(int mad_snoop_flags)
370 return (mad_snoop_flags &
371 (/*IB_MAD_SNOOP_POSTED_SENDS |
372 IB_MAD_SNOOP_RMPP_SENDS |*/
373 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
374 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
377 static inline int is_snooping_recvs(int mad_snoop_flags)
379 return (mad_snoop_flags &
380 (IB_MAD_SNOOP_RECVS /*|
381 IB_MAD_SNOOP_RMPP_RECVS*/));
384 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
385 struct ib_mad_snoop_private *mad_snoop_priv)
387 struct ib_mad_snoop_private **new_snoop_table;
391 spin_lock_irqsave(&qp_info->snoop_lock, flags);
392 /* Check for empty slot in array. */
393 for (i = 0; i < qp_info->snoop_table_size; i++)
394 if (!qp_info->snoop_table[i])
397 if (i == qp_info->snoop_table_size) {
399 new_snoop_table = kmalloc(sizeof mad_snoop_priv *
400 qp_info->snoop_table_size + 1,
402 if (!new_snoop_table) {
406 if (qp_info->snoop_table) {
407 memcpy(new_snoop_table, qp_info->snoop_table,
408 sizeof mad_snoop_priv *
409 qp_info->snoop_table_size);
410 kfree(qp_info->snoop_table);
412 qp_info->snoop_table = new_snoop_table;
413 qp_info->snoop_table_size++;
415 qp_info->snoop_table[i] = mad_snoop_priv;
416 atomic_inc(&qp_info->snoop_count);
418 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
422 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
424 enum ib_qp_type qp_type,
426 ib_mad_snoop_handler snoop_handler,
427 ib_mad_recv_handler recv_handler,
430 struct ib_mad_port_private *port_priv;
431 struct ib_mad_agent *ret;
432 struct ib_mad_snoop_private *mad_snoop_priv;
435 /* Validate parameters */
436 if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
437 (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
438 ret = ERR_PTR(-EINVAL);
441 qpn = get_spl_qp_index(qp_type);
443 ret = ERR_PTR(-EINVAL);
446 port_priv = ib_get_mad_port(device, port_num);
448 ret = ERR_PTR(-ENODEV);
451 /* Allocate structures */
452 mad_snoop_priv = kmalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
453 if (!mad_snoop_priv) {
454 ret = ERR_PTR(-ENOMEM);
458 /* Now, fill in the various structures */
459 memset(mad_snoop_priv, 0, sizeof *mad_snoop_priv);
460 mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
461 mad_snoop_priv->agent.device = device;
462 mad_snoop_priv->agent.recv_handler = recv_handler;
463 mad_snoop_priv->agent.snoop_handler = snoop_handler;
464 mad_snoop_priv->agent.context = context;
465 mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
466 mad_snoop_priv->agent.port_num = port_num;
467 mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
468 init_waitqueue_head(&mad_snoop_priv->wait);
469 mad_snoop_priv->snoop_index = register_snoop_agent(
470 &port_priv->qp_info[qpn],
472 if (mad_snoop_priv->snoop_index < 0) {
473 ret = ERR_PTR(mad_snoop_priv->snoop_index);
477 atomic_set(&mad_snoop_priv->refcount, 1);
478 return &mad_snoop_priv->agent;
481 kfree(mad_snoop_priv);
485 EXPORT_SYMBOL(ib_register_mad_snoop);
487 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
489 struct ib_mad_port_private *port_priv;
492 /* Note that we could still be handling received MADs */
495 * Canceling all sends results in dropping received response
496 * MADs, preventing us from queuing additional work
498 cancel_mads(mad_agent_priv);
499 port_priv = mad_agent_priv->qp_info->port_priv;
500 cancel_delayed_work(&mad_agent_priv->timed_work);
502 spin_lock_irqsave(&port_priv->reg_lock, flags);
503 remove_mad_reg_req(mad_agent_priv);
504 list_del(&mad_agent_priv->agent_list);
505 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
507 flush_workqueue(port_priv->wq);
509 atomic_dec(&mad_agent_priv->refcount);
510 wait_event(mad_agent_priv->wait,
511 !atomic_read(&mad_agent_priv->refcount));
513 if (mad_agent_priv->reg_req)
514 kfree(mad_agent_priv->reg_req);
515 ib_dereg_mr(mad_agent_priv->agent.mr);
516 kfree(mad_agent_priv);
519 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
521 struct ib_mad_qp_info *qp_info;
524 qp_info = mad_snoop_priv->qp_info;
525 spin_lock_irqsave(&qp_info->snoop_lock, flags);
526 qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
527 atomic_dec(&qp_info->snoop_count);
528 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
530 atomic_dec(&mad_snoop_priv->refcount);
531 wait_event(mad_snoop_priv->wait,
532 !atomic_read(&mad_snoop_priv->refcount));
534 kfree(mad_snoop_priv);
538 * ib_unregister_mad_agent - Unregisters a client from using MAD services
540 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
542 struct ib_mad_agent_private *mad_agent_priv;
543 struct ib_mad_snoop_private *mad_snoop_priv;
545 /* If the TID is zero, the agent can only snoop. */
546 if (mad_agent->hi_tid) {
547 mad_agent_priv = container_of(mad_agent,
548 struct ib_mad_agent_private,
550 unregister_mad_agent(mad_agent_priv);
552 mad_snoop_priv = container_of(mad_agent,
553 struct ib_mad_snoop_private,
555 unregister_mad_snoop(mad_snoop_priv);
559 EXPORT_SYMBOL(ib_unregister_mad_agent);
561 static void dequeue_mad(struct ib_mad_list_head *mad_list)
563 struct ib_mad_queue *mad_queue;
566 BUG_ON(!mad_list->mad_queue);
567 mad_queue = mad_list->mad_queue;
568 spin_lock_irqsave(&mad_queue->lock, flags);
569 list_del(&mad_list->list);
571 spin_unlock_irqrestore(&mad_queue->lock, flags);
574 static void snoop_send(struct ib_mad_qp_info *qp_info,
575 struct ib_send_wr *send_wr,
576 struct ib_mad_send_wc *mad_send_wc,
579 struct ib_mad_snoop_private *mad_snoop_priv;
583 spin_lock_irqsave(&qp_info->snoop_lock, flags);
584 for (i = 0; i < qp_info->snoop_table_size; i++) {
585 mad_snoop_priv = qp_info->snoop_table[i];
586 if (!mad_snoop_priv ||
587 !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
590 atomic_inc(&mad_snoop_priv->refcount);
591 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
592 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
593 send_wr, mad_send_wc);
594 if (atomic_dec_and_test(&mad_snoop_priv->refcount))
595 wake_up(&mad_snoop_priv->wait);
596 spin_lock_irqsave(&qp_info->snoop_lock, flags);
598 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
601 static void snoop_recv(struct ib_mad_qp_info *qp_info,
602 struct ib_mad_recv_wc *mad_recv_wc,
605 struct ib_mad_snoop_private *mad_snoop_priv;
609 spin_lock_irqsave(&qp_info->snoop_lock, flags);
610 for (i = 0; i < qp_info->snoop_table_size; i++) {
611 mad_snoop_priv = qp_info->snoop_table[i];
612 if (!mad_snoop_priv ||
613 !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
616 atomic_inc(&mad_snoop_priv->refcount);
617 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
618 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
620 if (atomic_dec_and_test(&mad_snoop_priv->refcount))
621 wake_up(&mad_snoop_priv->wait);
622 spin_lock_irqsave(&qp_info->snoop_lock, flags);
624 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
627 static void build_smp_wc(u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
630 memset(wc, 0, sizeof *wc);
632 wc->status = IB_WC_SUCCESS;
633 wc->opcode = IB_WC_RECV;
634 wc->pkey_index = pkey_index;
635 wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
640 wc->dlid_path_bits = 0;
641 wc->port_num = port_num;
645 * Return 0 if SMP is to be sent
646 * Return 1 if SMP was consumed locally (whether or not solicited)
647 * Return < 0 if error
649 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
651 struct ib_send_wr *send_wr)
655 struct ib_mad_local_private *local;
656 struct ib_mad_private *mad_priv;
657 struct ib_mad_port_private *port_priv;
658 struct ib_mad_agent_private *recv_mad_agent = NULL;
659 struct ib_device *device = mad_agent_priv->agent.device;
660 u8 port_num = mad_agent_priv->agent.port_num;
663 if (!smi_handle_dr_smp_send(smp, device->node_type, port_num)) {
665 printk(KERN_ERR PFX "Invalid directed route\n");
668 /* Check to post send on QP or process locally */
669 ret = smi_check_local_dr_smp(smp, device, port_num);
670 if (!ret || !device->process_mad)
673 local = kmalloc(sizeof *local, GFP_ATOMIC);
676 printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
679 local->mad_priv = NULL;
680 local->recv_mad_agent = NULL;
681 mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
684 printk(KERN_ERR PFX "No memory for local response MAD\n");
689 build_smp_wc(send_wr->wr_id, smp->dr_slid, send_wr->wr.ud.pkey_index,
690 send_wr->wr.ud.port_num, &mad_wc);
692 /* No GRH for DR SMP */
693 ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
694 (struct ib_mad *)smp,
695 (struct ib_mad *)&mad_priv->mad);
698 case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
700 * See if response is solicited and
701 * there is a recv handler
703 if (solicited_mad(&mad_priv->mad.mad) &&
704 mad_agent_priv->agent.recv_handler) {
705 local->mad_priv = mad_priv;
706 local->recv_mad_agent = mad_agent_priv;
708 * Reference MAD agent until receive
709 * side of local completion handled
711 atomic_inc(&mad_agent_priv->refcount);
713 kmem_cache_free(ib_mad_cache, mad_priv);
715 case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
716 kmem_cache_free(ib_mad_cache, mad_priv);
718 case IB_MAD_RESULT_SUCCESS:
719 /* Treat like an incoming receive MAD */
720 solicited = solicited_mad(&mad_priv->mad.mad);
721 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
722 mad_agent_priv->agent.port_num);
724 mad_priv->mad.mad.mad_hdr.tid =
725 ((struct ib_mad *)smp)->mad_hdr.tid;
726 recv_mad_agent = find_mad_agent(port_priv,
730 if (!port_priv || !recv_mad_agent) {
731 kmem_cache_free(ib_mad_cache, mad_priv);
736 local->mad_priv = mad_priv;
737 local->recv_mad_agent = recv_mad_agent;
740 kmem_cache_free(ib_mad_cache, mad_priv);
746 local->send_wr = *send_wr;
747 local->send_wr.sg_list = local->sg_list;
748 memcpy(local->sg_list, send_wr->sg_list,
749 sizeof *send_wr->sg_list * send_wr->num_sge);
750 local->send_wr.next = NULL;
751 local->tid = send_wr->wr.ud.mad_hdr->tid;
752 local->wr_id = send_wr->wr_id;
753 /* Reference MAD agent until send side of local completion handled */
754 atomic_inc(&mad_agent_priv->refcount);
755 /* Queue local completion to local list */
756 spin_lock_irqsave(&mad_agent_priv->lock, flags);
757 list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
758 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
759 queue_work(mad_agent_priv->qp_info->port_priv->wq,
760 &mad_agent_priv->local_work);
766 static int get_buf_length(int hdr_len, int data_len)
770 seg_size = sizeof(struct ib_mad) - hdr_len;
771 if (data_len && seg_size) {
772 pad = seg_size - data_len % seg_size;
777 return hdr_len + data_len + pad;
780 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
781 u32 remote_qpn, u16 pkey_index,
783 int hdr_len, int data_len,
784 unsigned int __nocast gfp_mask)
786 struct ib_mad_agent_private *mad_agent_priv;
787 struct ib_mad_send_buf *send_buf;
791 mad_agent_priv = container_of(mad_agent,
792 struct ib_mad_agent_private, agent);
793 buf_size = get_buf_length(hdr_len, data_len);
795 buf = kmalloc(sizeof *send_buf + buf_size, gfp_mask);
797 return ERR_PTR(-ENOMEM);
799 send_buf = buf + buf_size;
800 memset(send_buf, 0, sizeof *send_buf);
803 send_buf->sge.addr = dma_map_single(mad_agent->device->dma_device,
804 buf, buf_size, DMA_TO_DEVICE);
805 pci_unmap_addr_set(send_buf, mapping, send_buf->sge.addr);
806 send_buf->sge.length = buf_size;
807 send_buf->sge.lkey = mad_agent->mr->lkey;
809 send_buf->send_wr.wr_id = (unsigned long) send_buf;
810 send_buf->send_wr.sg_list = &send_buf->sge;
811 send_buf->send_wr.num_sge = 1;
812 send_buf->send_wr.opcode = IB_WR_SEND;
813 send_buf->send_wr.send_flags = IB_SEND_SIGNALED;
814 send_buf->send_wr.wr.ud.ah = ah;
815 send_buf->send_wr.wr.ud.mad_hdr = &send_buf->mad->mad_hdr;
816 send_buf->send_wr.wr.ud.remote_qpn = remote_qpn;
817 send_buf->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
818 send_buf->send_wr.wr.ud.pkey_index = pkey_index;
819 send_buf->mad_agent = mad_agent;
820 atomic_inc(&mad_agent_priv->refcount);
823 EXPORT_SYMBOL(ib_create_send_mad);
825 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
827 struct ib_mad_agent_private *mad_agent_priv;
829 mad_agent_priv = container_of(send_buf->mad_agent,
830 struct ib_mad_agent_private, agent);
832 dma_unmap_single(send_buf->mad_agent->device->dma_device,
833 pci_unmap_addr(send_buf, mapping),
834 send_buf->sge.length, DMA_TO_DEVICE);
835 kfree(send_buf->mad);
837 if (atomic_dec_and_test(&mad_agent_priv->refcount))
838 wake_up(&mad_agent_priv->wait);
840 EXPORT_SYMBOL(ib_free_send_mad);
842 static int ib_send_mad(struct ib_mad_agent_private *mad_agent_priv,
843 struct ib_mad_send_wr_private *mad_send_wr)
845 struct ib_mad_qp_info *qp_info;
846 struct ib_send_wr *bad_send_wr;
850 /* Replace user's WR ID with our own to find WR upon completion */
851 qp_info = mad_agent_priv->qp_info;
852 mad_send_wr->wr_id = mad_send_wr->send_wr.wr_id;
853 mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
854 mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
856 spin_lock_irqsave(&qp_info->send_queue.lock, flags);
857 if (qp_info->send_queue.count++ < qp_info->send_queue.max_active) {
858 list_add_tail(&mad_send_wr->mad_list.list,
859 &qp_info->send_queue.list);
860 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
861 ret = ib_post_send(mad_agent_priv->agent.qp,
862 &mad_send_wr->send_wr, &bad_send_wr);
864 printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
865 dequeue_mad(&mad_send_wr->mad_list);
868 list_add_tail(&mad_send_wr->mad_list.list,
869 &qp_info->overflow_list);
870 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
877 * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
878 * with the registered client
880 int ib_post_send_mad(struct ib_mad_agent *mad_agent,
881 struct ib_send_wr *send_wr,
882 struct ib_send_wr **bad_send_wr)
885 struct ib_mad_agent_private *mad_agent_priv;
887 /* Validate supplied parameters */
891 if (!mad_agent || !send_wr)
894 if (!mad_agent->send_handler)
897 mad_agent_priv = container_of(mad_agent,
898 struct ib_mad_agent_private,
901 /* Walk list of send WRs and post each on send list */
904 struct ib_send_wr *next_send_wr;
905 struct ib_mad_send_wr_private *mad_send_wr;
908 /* Validate more parameters */
909 if (send_wr->num_sge > IB_MAD_SEND_REQ_MAX_SG)
912 if (send_wr->wr.ud.timeout_ms && !mad_agent->recv_handler)
915 if (!send_wr->wr.ud.mad_hdr) {
916 printk(KERN_ERR PFX "MAD header must be supplied "
917 "in WR %p\n", send_wr);
922 * Save pointer to next work request to post in case the
923 * current one completes, and the user modifies the work
924 * request associated with the completion
926 next_send_wr = (struct ib_send_wr *)send_wr->next;
928 smp = (struct ib_smp *)send_wr->wr.ud.mad_hdr;
929 if (smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
930 ret = handle_outgoing_dr_smp(mad_agent_priv, smp,
932 if (ret < 0) /* error */
934 else if (ret == 1) /* locally consumed */
938 /* Allocate MAD send WR tracking structure */
939 mad_send_wr = kmalloc(sizeof *mad_send_wr, GFP_ATOMIC);
941 printk(KERN_ERR PFX "No memory for "
942 "ib_mad_send_wr_private\n");
947 mad_send_wr->send_wr = *send_wr;
948 mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
949 memcpy(mad_send_wr->sg_list, send_wr->sg_list,
950 sizeof *send_wr->sg_list * send_wr->num_sge);
951 mad_send_wr->send_wr.next = NULL;
952 mad_send_wr->tid = send_wr->wr.ud.mad_hdr->tid;
953 mad_send_wr->agent = mad_agent;
954 /* Timeout will be updated after send completes */
955 mad_send_wr->timeout = msecs_to_jiffies(send_wr->wr.
957 mad_send_wr->retry = 0;
958 /* One reference for each work request to QP + response */
959 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
960 mad_send_wr->status = IB_WC_SUCCESS;
962 /* Reference MAD agent until send completes */
963 atomic_inc(&mad_agent_priv->refcount);
964 spin_lock_irqsave(&mad_agent_priv->lock, flags);
965 list_add_tail(&mad_send_wr->agent_list,
966 &mad_agent_priv->send_list);
967 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
969 ret = ib_send_mad(mad_agent_priv, mad_send_wr);
971 /* Fail send request */
972 spin_lock_irqsave(&mad_agent_priv->lock, flags);
973 list_del(&mad_send_wr->agent_list);
974 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
975 atomic_dec(&mad_agent_priv->refcount);
979 send_wr = next_send_wr;
984 *bad_send_wr = send_wr;
988 EXPORT_SYMBOL(ib_post_send_mad);
991 * ib_free_recv_mad - Returns data buffers used to receive
992 * a MAD to the access layer
994 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
996 struct ib_mad_recv_buf *entry;
997 struct ib_mad_private_header *mad_priv_hdr;
998 struct ib_mad_private *priv;
1000 mad_priv_hdr = container_of(mad_recv_wc,
1001 struct ib_mad_private_header,
1003 priv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1006 * Walk receive buffer list associated with this WC
1007 * No need to remove them from list of receive buffers
1009 list_for_each_entry(entry, &mad_recv_wc->recv_buf.list, list) {
1010 /* Free previous receive buffer */
1011 kmem_cache_free(ib_mad_cache, priv);
1012 mad_priv_hdr = container_of(mad_recv_wc,
1013 struct ib_mad_private_header,
1015 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1019 /* Free last buffer */
1020 kmem_cache_free(ib_mad_cache, priv);
1022 EXPORT_SYMBOL(ib_free_recv_mad);
1024 void ib_coalesce_recv_mad(struct ib_mad_recv_wc *mad_recv_wc,
1027 printk(KERN_ERR PFX "ib_coalesce_recv_mad() not implemented yet\n");
1029 EXPORT_SYMBOL(ib_coalesce_recv_mad);
1031 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1033 ib_mad_send_handler send_handler,
1034 ib_mad_recv_handler recv_handler,
1037 return ERR_PTR(-EINVAL); /* XXX: for now */
1039 EXPORT_SYMBOL(ib_redirect_mad_qp);
1041 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1044 printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1047 EXPORT_SYMBOL(ib_process_mad_wc);
1049 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1050 struct ib_mad_reg_req *mad_reg_req)
1054 for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1055 i < IB_MGMT_MAX_METHODS;
1056 i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1058 if ((*method)->agent[i]) {
1059 printk(KERN_ERR PFX "Method %d already in use\n", i);
1066 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1068 /* Allocate management method table */
1069 *method = kmalloc(sizeof **method, GFP_ATOMIC);
1071 printk(KERN_ERR PFX "No memory for "
1072 "ib_mad_mgmt_method_table\n");
1075 /* Clear management method table */
1076 memset(*method, 0, sizeof **method);
1082 * Check to see if there are any methods still in use
1084 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1088 for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1089 if (method->agent[i])
1095 * Check to see if there are any method tables for this class still in use
1097 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1101 for (i = 0; i < MAX_MGMT_CLASS; i++)
1102 if (class->method_table[i])
1107 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1111 for (i = 0; i < MAX_MGMT_OUI; i++)
1112 if (vendor_class->method_table[i])
1117 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1122 for (i = 0; i < MAX_MGMT_OUI; i++)
1123 /* Is there matching OUI for this vendor class ? */
1124 if (!memcmp(vendor_class->oui[i], oui, 3))
1130 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1134 for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1135 if (vendor->vendor_class[i])
1141 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1142 struct ib_mad_agent_private *agent)
1146 /* Remove any methods for this mad agent */
1147 for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1148 if (method->agent[i] == agent) {
1149 method->agent[i] = NULL;
1154 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1155 struct ib_mad_agent_private *agent_priv,
1158 struct ib_mad_port_private *port_priv;
1159 struct ib_mad_mgmt_class_table **class;
1160 struct ib_mad_mgmt_method_table **method;
1163 port_priv = agent_priv->qp_info->port_priv;
1164 class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1166 /* Allocate management class table for "new" class version */
1167 *class = kmalloc(sizeof **class, GFP_ATOMIC);
1169 printk(KERN_ERR PFX "No memory for "
1170 "ib_mad_mgmt_class_table\n");
1174 /* Clear management class table */
1175 memset(*class, 0, sizeof(**class));
1176 /* Allocate method table for this management class */
1177 method = &(*class)->method_table[mgmt_class];
1178 if ((ret = allocate_method_table(method)))
1181 method = &(*class)->method_table[mgmt_class];
1183 /* Allocate method table for this management class */
1184 if ((ret = allocate_method_table(method)))
1189 /* Now, make sure methods are not already in use */
1190 if (method_in_use(method, mad_reg_req))
1193 /* Finally, add in methods being registered */
1194 for (i = find_first_bit(mad_reg_req->method_mask,
1195 IB_MGMT_MAX_METHODS);
1196 i < IB_MGMT_MAX_METHODS;
1197 i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1199 (*method)->agent[i] = agent_priv;
1204 /* Remove any methods for this mad agent */
1205 remove_methods_mad_agent(*method, agent_priv);
1206 /* Now, check to see if there are any methods in use */
1207 if (!check_method_table(*method)) {
1208 /* If not, release management method table */
1221 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1222 struct ib_mad_agent_private *agent_priv)
1224 struct ib_mad_port_private *port_priv;
1225 struct ib_mad_mgmt_vendor_class_table **vendor_table;
1226 struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1227 struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1228 struct ib_mad_mgmt_method_table **method;
1229 int i, ret = -ENOMEM;
1232 /* "New" vendor (with OUI) class */
1233 vclass = vendor_class_index(mad_reg_req->mgmt_class);
1234 port_priv = agent_priv->qp_info->port_priv;
1235 vendor_table = &port_priv->version[
1236 mad_reg_req->mgmt_class_version].vendor;
1237 if (!*vendor_table) {
1238 /* Allocate mgmt vendor class table for "new" class version */
1239 vendor = kmalloc(sizeof *vendor, GFP_ATOMIC);
1241 printk(KERN_ERR PFX "No memory for "
1242 "ib_mad_mgmt_vendor_class_table\n");
1245 /* Clear management vendor class table */
1246 memset(vendor, 0, sizeof(*vendor));
1247 *vendor_table = vendor;
1249 if (!(*vendor_table)->vendor_class[vclass]) {
1250 /* Allocate table for this management vendor class */
1251 vendor_class = kmalloc(sizeof *vendor_class, GFP_ATOMIC);
1252 if (!vendor_class) {
1253 printk(KERN_ERR PFX "No memory for "
1254 "ib_mad_mgmt_vendor_class\n");
1257 memset(vendor_class, 0, sizeof(*vendor_class));
1258 (*vendor_table)->vendor_class[vclass] = vendor_class;
1260 for (i = 0; i < MAX_MGMT_OUI; i++) {
1261 /* Is there matching OUI for this vendor class ? */
1262 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1263 mad_reg_req->oui, 3)) {
1264 method = &(*vendor_table)->vendor_class[
1265 vclass]->method_table[i];
1270 for (i = 0; i < MAX_MGMT_OUI; i++) {
1271 /* OUI slot available ? */
1272 if (!is_vendor_oui((*vendor_table)->vendor_class[
1274 method = &(*vendor_table)->vendor_class[
1275 vclass]->method_table[i];
1277 /* Allocate method table for this OUI */
1278 if ((ret = allocate_method_table(method)))
1280 memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1281 mad_reg_req->oui, 3);
1285 printk(KERN_ERR PFX "All OUI slots in use\n");
1289 /* Now, make sure methods are not already in use */
1290 if (method_in_use(method, mad_reg_req))
1293 /* Finally, add in methods being registered */
1294 for (i = find_first_bit(mad_reg_req->method_mask,
1295 IB_MGMT_MAX_METHODS);
1296 i < IB_MGMT_MAX_METHODS;
1297 i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1299 (*method)->agent[i] = agent_priv;
1304 /* Remove any methods for this mad agent */
1305 remove_methods_mad_agent(*method, agent_priv);
1306 /* Now, check to see if there are any methods in use */
1307 if (!check_method_table(*method)) {
1308 /* If not, release management method table */
1315 (*vendor_table)->vendor_class[vclass] = NULL;
1316 kfree(vendor_class);
1320 *vendor_table = NULL;
1327 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1329 struct ib_mad_port_private *port_priv;
1330 struct ib_mad_mgmt_class_table *class;
1331 struct ib_mad_mgmt_method_table *method;
1332 struct ib_mad_mgmt_vendor_class_table *vendor;
1333 struct ib_mad_mgmt_vendor_class *vendor_class;
1338 * Was MAD registration request supplied
1339 * with original registration ?
1341 if (!agent_priv->reg_req) {
1345 port_priv = agent_priv->qp_info->port_priv;
1346 mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1347 class = port_priv->version[
1348 agent_priv->reg_req->mgmt_class_version].class;
1352 method = class->method_table[mgmt_class];
1354 /* Remove any methods for this mad agent */
1355 remove_methods_mad_agent(method, agent_priv);
1356 /* Now, check to see if there are any methods still in use */
1357 if (!check_method_table(method)) {
1358 /* If not, release management method table */
1360 class->method_table[mgmt_class] = NULL;
1361 /* Any management classes left ? */
1362 if (!check_class_table(class)) {
1363 /* If not, release management class table */
1366 agent_priv->reg_req->
1367 mgmt_class_version].class = NULL;
1373 if (!is_vendor_class(mgmt_class))
1376 /* normalize mgmt_class to vendor range 2 */
1377 mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1378 vendor = port_priv->version[
1379 agent_priv->reg_req->mgmt_class_version].vendor;
1384 vendor_class = vendor->vendor_class[mgmt_class];
1386 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1389 method = vendor_class->method_table[index];
1391 /* Remove any methods for this mad agent */
1392 remove_methods_mad_agent(method, agent_priv);
1394 * Now, check to see if there are
1395 * any methods still in use
1397 if (!check_method_table(method)) {
1398 /* If not, release management method table */
1400 vendor_class->method_table[index] = NULL;
1401 memset(vendor_class->oui[index], 0, 3);
1402 /* Any OUIs left ? */
1403 if (!check_vendor_class(vendor_class)) {
1404 /* If not, release vendor class table */
1405 kfree(vendor_class);
1406 vendor->vendor_class[mgmt_class] = NULL;
1407 /* Any other vendor classes left ? */
1408 if (!check_vendor_table(vendor)) {
1411 agent_priv->reg_req->
1412 mgmt_class_version].
1424 static int response_mad(struct ib_mad *mad)
1426 /* Trap represses are responses although response bit is reset */
1427 return ((mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
1428 (mad->mad_hdr.method & IB_MGMT_METHOD_RESP));
1431 static int solicited_mad(struct ib_mad *mad)
1433 /* CM MADs are never solicited */
1434 if (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CM) {
1438 /* XXX: Determine whether MAD is using RMPP */
1440 /* Not using RMPP */
1441 /* Is this MAD a response to a previous MAD ? */
1442 return response_mad(mad);
1445 static struct ib_mad_agent_private *
1446 find_mad_agent(struct ib_mad_port_private *port_priv,
1450 struct ib_mad_agent_private *mad_agent = NULL;
1451 unsigned long flags;
1453 spin_lock_irqsave(&port_priv->reg_lock, flags);
1456 * Whether MAD was solicited determines type of routing to
1461 struct ib_mad_agent_private *entry;
1464 * Routing is based on high 32 bits of transaction ID
1467 hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1468 list_for_each_entry(entry, &port_priv->agent_list,
1470 if (entry->agent.hi_tid == hi_tid) {
1476 struct ib_mad_mgmt_class_table *class;
1477 struct ib_mad_mgmt_method_table *method;
1478 struct ib_mad_mgmt_vendor_class_table *vendor;
1479 struct ib_mad_mgmt_vendor_class *vendor_class;
1480 struct ib_vendor_mad *vendor_mad;
1484 * Routing is based on version, class, and method
1485 * For "newer" vendor MADs, also based on OUI
1487 if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1489 if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1490 class = port_priv->version[
1491 mad->mad_hdr.class_version].class;
1494 method = class->method_table[convert_mgmt_class(
1495 mad->mad_hdr.mgmt_class)];
1497 mad_agent = method->agent[mad->mad_hdr.method &
1498 ~IB_MGMT_METHOD_RESP];
1500 vendor = port_priv->version[
1501 mad->mad_hdr.class_version].vendor;
1504 vendor_class = vendor->vendor_class[vendor_class_index(
1505 mad->mad_hdr.mgmt_class)];
1508 /* Find matching OUI */
1509 vendor_mad = (struct ib_vendor_mad *)mad;
1510 index = find_vendor_oui(vendor_class, vendor_mad->oui);
1513 method = vendor_class->method_table[index];
1515 mad_agent = method->agent[mad->mad_hdr.method &
1516 ~IB_MGMT_METHOD_RESP];
1522 if (mad_agent->agent.recv_handler)
1523 atomic_inc(&mad_agent->refcount);
1525 printk(KERN_NOTICE PFX "No receive handler for client "
1527 &mad_agent->agent, port_priv->port_num);
1532 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1537 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1541 /* Make sure MAD base version is understood */
1542 if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1543 printk(KERN_ERR PFX "MAD received with unsupported base "
1544 "version %d\n", mad->mad_hdr.base_version);
1548 /* Filter SMI packets sent to other than QP0 */
1549 if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1550 (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1554 /* Filter GSI packets sent to QP0 */
1564 * Return start of fully reassembled MAD, or NULL, if MAD isn't assembled yet
1566 static struct ib_mad_private *
1567 reassemble_recv(struct ib_mad_agent_private *mad_agent_priv,
1568 struct ib_mad_private *recv)
1570 /* Until we have RMPP, all receives are reassembled!... */
1571 INIT_LIST_HEAD(&recv->header.recv_wc.recv_buf.list);
1575 static struct ib_mad_send_wr_private*
1576 find_send_req(struct ib_mad_agent_private *mad_agent_priv,
1579 struct ib_mad_send_wr_private *mad_send_wr;
1581 list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
1583 if (mad_send_wr->tid == tid)
1588 * It's possible to receive the response before we've
1589 * been notified that the send has completed
1591 list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
1593 if (mad_send_wr->tid == tid && mad_send_wr->timeout) {
1594 /* Verify request has not been canceled */
1595 return (mad_send_wr->status == IB_WC_SUCCESS) ?
1602 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1603 struct ib_mad_private *recv,
1606 struct ib_mad_send_wr_private *mad_send_wr;
1607 struct ib_mad_send_wc mad_send_wc;
1608 unsigned long flags;
1610 /* Fully reassemble receive before processing */
1611 recv = reassemble_recv(mad_agent_priv, recv);
1613 if (atomic_dec_and_test(&mad_agent_priv->refcount))
1614 wake_up(&mad_agent_priv->wait);
1618 /* Complete corresponding request */
1620 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1621 mad_send_wr = find_send_req(mad_agent_priv,
1622 recv->mad.mad.mad_hdr.tid);
1624 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1625 ib_free_recv_mad(&recv->header.recv_wc);
1626 if (atomic_dec_and_test(&mad_agent_priv->refcount))
1627 wake_up(&mad_agent_priv->wait);
1630 /* Timeout = 0 means that we won't wait for a response */
1631 mad_send_wr->timeout = 0;
1632 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1634 /* Defined behavior is to complete response before request */
1635 recv->header.recv_wc.wc->wr_id = mad_send_wr->wr_id;
1636 mad_agent_priv->agent.recv_handler(
1637 &mad_agent_priv->agent,
1638 &recv->header.recv_wc);
1639 atomic_dec(&mad_agent_priv->refcount);
1641 mad_send_wc.status = IB_WC_SUCCESS;
1642 mad_send_wc.vendor_err = 0;
1643 mad_send_wc.wr_id = mad_send_wr->wr_id;
1644 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1646 mad_agent_priv->agent.recv_handler(
1647 &mad_agent_priv->agent,
1648 &recv->header.recv_wc);
1649 if (atomic_dec_and_test(&mad_agent_priv->refcount))
1650 wake_up(&mad_agent_priv->wait);
1654 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1657 struct ib_mad_qp_info *qp_info;
1658 struct ib_mad_private_header *mad_priv_hdr;
1659 struct ib_mad_private *recv, *response;
1660 struct ib_mad_list_head *mad_list;
1661 struct ib_mad_agent_private *mad_agent;
1664 response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1666 printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1667 "for response buffer\n");
1669 mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1670 qp_info = mad_list->mad_queue->qp_info;
1671 dequeue_mad(mad_list);
1673 mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1675 recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1676 dma_unmap_single(port_priv->device->dma_device,
1677 pci_unmap_addr(&recv->header, mapping),
1678 sizeof(struct ib_mad_private) -
1679 sizeof(struct ib_mad_private_header),
1682 /* Setup MAD receive work completion from "normal" work completion */
1683 recv->header.wc = *wc;
1684 recv->header.recv_wc.wc = &recv->header.wc;
1685 recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1686 recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1687 recv->header.recv_wc.recv_buf.grh = &recv->grh;
1689 if (atomic_read(&qp_info->snoop_count))
1690 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1693 if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1696 if (recv->mad.mad.mad_hdr.mgmt_class ==
1697 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1698 if (!smi_handle_dr_smp_recv(&recv->mad.smp,
1699 port_priv->device->node_type,
1700 port_priv->port_num,
1701 port_priv->device->phys_port_cnt))
1703 if (!smi_check_forward_dr_smp(&recv->mad.smp))
1705 if (!smi_handle_dr_smp_send(&recv->mad.smp,
1706 port_priv->device->node_type,
1707 port_priv->port_num))
1709 if (!smi_check_local_dr_smp(&recv->mad.smp,
1711 port_priv->port_num))
1716 /* Give driver "right of first refusal" on incoming MAD */
1717 if (port_priv->device->process_mad) {
1721 printk(KERN_ERR PFX "No memory for response MAD\n");
1723 * Is it better to assume that
1724 * it wouldn't be processed ?
1729 ret = port_priv->device->process_mad(port_priv->device, 0,
1730 port_priv->port_num,
1733 &response->mad.mad);
1734 if (ret & IB_MAD_RESULT_SUCCESS) {
1735 if (ret & IB_MAD_RESULT_CONSUMED)
1737 if (ret & IB_MAD_RESULT_REPLY) {
1739 if (!agent_send(response, &recv->grh, wc,
1741 port_priv->port_num))
1748 /* Determine corresponding MAD agent for incoming receive MAD */
1749 solicited = solicited_mad(&recv->mad.mad);
1750 mad_agent = find_mad_agent(port_priv, &recv->mad.mad, solicited);
1752 ib_mad_complete_recv(mad_agent, recv, solicited);
1754 * recv is freed up in error cases in ib_mad_complete_recv
1755 * or via recv_handler in ib_mad_complete_recv()
1761 /* Post another receive request for this QP */
1763 ib_mad_post_receive_mads(qp_info, response);
1765 kmem_cache_free(ib_mad_cache, recv);
1767 ib_mad_post_receive_mads(qp_info, recv);
1770 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1772 struct ib_mad_send_wr_private *mad_send_wr;
1773 unsigned long delay;
1775 if (list_empty(&mad_agent_priv->wait_list)) {
1776 cancel_delayed_work(&mad_agent_priv->timed_work);
1778 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1779 struct ib_mad_send_wr_private,
1782 if (time_after(mad_agent_priv->timeout,
1783 mad_send_wr->timeout)) {
1784 mad_agent_priv->timeout = mad_send_wr->timeout;
1785 cancel_delayed_work(&mad_agent_priv->timed_work);
1786 delay = mad_send_wr->timeout - jiffies;
1787 if ((long)delay <= 0)
1789 queue_delayed_work(mad_agent_priv->qp_info->
1791 &mad_agent_priv->timed_work, delay);
1796 static void wait_for_response(struct ib_mad_agent_private *mad_agent_priv,
1797 struct ib_mad_send_wr_private *mad_send_wr )
1799 struct ib_mad_send_wr_private *temp_mad_send_wr;
1800 struct list_head *list_item;
1801 unsigned long delay;
1803 list_del(&mad_send_wr->agent_list);
1805 delay = mad_send_wr->timeout;
1806 mad_send_wr->timeout += jiffies;
1808 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1809 temp_mad_send_wr = list_entry(list_item,
1810 struct ib_mad_send_wr_private,
1812 if (time_after(mad_send_wr->timeout,
1813 temp_mad_send_wr->timeout))
1816 list_add(&mad_send_wr->agent_list, list_item);
1818 /* Reschedule a work item if we have a shorter timeout */
1819 if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
1820 cancel_delayed_work(&mad_agent_priv->timed_work);
1821 queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
1822 &mad_agent_priv->timed_work, delay);
1827 * Process a send work completion
1829 static void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
1830 struct ib_mad_send_wc *mad_send_wc)
1832 struct ib_mad_agent_private *mad_agent_priv;
1833 unsigned long flags;
1835 mad_agent_priv = container_of(mad_send_wr->agent,
1836 struct ib_mad_agent_private, agent);
1838 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1839 if (mad_send_wc->status != IB_WC_SUCCESS &&
1840 mad_send_wr->status == IB_WC_SUCCESS) {
1841 mad_send_wr->status = mad_send_wc->status;
1842 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
1845 if (--mad_send_wr->refcount > 0) {
1846 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
1847 mad_send_wr->status == IB_WC_SUCCESS) {
1848 wait_for_response(mad_agent_priv, mad_send_wr);
1850 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1854 /* Remove send from MAD agent and notify client of completion */
1855 list_del(&mad_send_wr->agent_list);
1856 adjust_timeout(mad_agent_priv);
1857 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1859 if (mad_send_wr->status != IB_WC_SUCCESS )
1860 mad_send_wc->status = mad_send_wr->status;
1861 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
1864 /* Release reference on agent taken when sending */
1865 if (atomic_dec_and_test(&mad_agent_priv->refcount))
1866 wake_up(&mad_agent_priv->wait);
1871 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
1874 struct ib_mad_send_wr_private *mad_send_wr, *queued_send_wr;
1875 struct ib_mad_list_head *mad_list;
1876 struct ib_mad_qp_info *qp_info;
1877 struct ib_mad_queue *send_queue;
1878 struct ib_send_wr *bad_send_wr;
1879 unsigned long flags;
1882 mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1883 mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1885 send_queue = mad_list->mad_queue;
1886 qp_info = send_queue->qp_info;
1889 queued_send_wr = NULL;
1890 spin_lock_irqsave(&send_queue->lock, flags);
1891 list_del(&mad_list->list);
1893 /* Move queued send to the send queue */
1894 if (send_queue->count-- > send_queue->max_active) {
1895 mad_list = container_of(qp_info->overflow_list.next,
1896 struct ib_mad_list_head, list);
1897 queued_send_wr = container_of(mad_list,
1898 struct ib_mad_send_wr_private,
1900 list_del(&mad_list->list);
1901 list_add_tail(&mad_list->list, &send_queue->list);
1903 spin_unlock_irqrestore(&send_queue->lock, flags);
1905 /* Restore client wr_id in WC and complete send */
1906 wc->wr_id = mad_send_wr->wr_id;
1907 if (atomic_read(&qp_info->snoop_count))
1908 snoop_send(qp_info, &mad_send_wr->send_wr,
1909 (struct ib_mad_send_wc *)wc,
1910 IB_MAD_SNOOP_SEND_COMPLETIONS);
1911 ib_mad_complete_send_wr(mad_send_wr, (struct ib_mad_send_wc *)wc);
1913 if (queued_send_wr) {
1914 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
1917 printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
1918 mad_send_wr = queued_send_wr;
1919 wc->status = IB_WC_LOC_QP_OP_ERR;
1925 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
1927 struct ib_mad_send_wr_private *mad_send_wr;
1928 struct ib_mad_list_head *mad_list;
1929 unsigned long flags;
1931 spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1932 list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
1933 mad_send_wr = container_of(mad_list,
1934 struct ib_mad_send_wr_private,
1936 mad_send_wr->retry = 1;
1938 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1941 static void mad_error_handler(struct ib_mad_port_private *port_priv,
1944 struct ib_mad_list_head *mad_list;
1945 struct ib_mad_qp_info *qp_info;
1946 struct ib_mad_send_wr_private *mad_send_wr;
1949 /* Determine if failure was a send or receive */
1950 mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1951 qp_info = mad_list->mad_queue->qp_info;
1952 if (mad_list->mad_queue == &qp_info->recv_queue)
1954 * Receive errors indicate that the QP has entered the error
1955 * state - error handling/shutdown code will cleanup
1960 * Send errors will transition the QP to SQE - move
1961 * QP to RTS and repost flushed work requests
1963 mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1965 if (wc->status == IB_WC_WR_FLUSH_ERR) {
1966 if (mad_send_wr->retry) {
1968 struct ib_send_wr *bad_send_wr;
1970 mad_send_wr->retry = 0;
1971 ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
1974 ib_mad_send_done_handler(port_priv, wc);
1976 ib_mad_send_done_handler(port_priv, wc);
1978 struct ib_qp_attr *attr;
1980 /* Transition QP to RTS and fail offending send */
1981 attr = kmalloc(sizeof *attr, GFP_KERNEL);
1983 attr->qp_state = IB_QPS_RTS;
1984 attr->cur_qp_state = IB_QPS_SQE;
1985 ret = ib_modify_qp(qp_info->qp, attr,
1986 IB_QP_STATE | IB_QP_CUR_STATE);
1989 printk(KERN_ERR PFX "mad_error_handler - "
1990 "ib_modify_qp to RTS : %d\n", ret);
1992 mark_sends_for_retry(qp_info);
1994 ib_mad_send_done_handler(port_priv, wc);
1999 * IB MAD completion callback
2001 static void ib_mad_completion_handler(void *data)
2003 struct ib_mad_port_private *port_priv;
2006 port_priv = (struct ib_mad_port_private *)data;
2007 ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2009 while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2010 if (wc.status == IB_WC_SUCCESS) {
2011 switch (wc.opcode) {
2013 ib_mad_send_done_handler(port_priv, &wc);
2016 ib_mad_recv_done_handler(port_priv, &wc);
2023 mad_error_handler(port_priv, &wc);
2027 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2029 unsigned long flags;
2030 struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2031 struct ib_mad_send_wc mad_send_wc;
2032 struct list_head cancel_list;
2034 INIT_LIST_HEAD(&cancel_list);
2036 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2037 list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2038 &mad_agent_priv->send_list, agent_list) {
2039 if (mad_send_wr->status == IB_WC_SUCCESS) {
2040 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2041 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2045 /* Empty wait list to prevent receives from finding a request */
2046 list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2047 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2049 /* Report all cancelled requests */
2050 mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2051 mad_send_wc.vendor_err = 0;
2053 list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2054 &cancel_list, agent_list) {
2055 mad_send_wc.wr_id = mad_send_wr->wr_id;
2056 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2059 list_del(&mad_send_wr->agent_list);
2061 atomic_dec(&mad_agent_priv->refcount);
2065 static struct ib_mad_send_wr_private*
2066 find_send_by_wr_id(struct ib_mad_agent_private *mad_agent_priv,
2069 struct ib_mad_send_wr_private *mad_send_wr;
2071 list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2073 if (mad_send_wr->wr_id == wr_id)
2077 list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2079 if (mad_send_wr->wr_id == wr_id)
2085 void cancel_sends(void *data)
2087 struct ib_mad_agent_private *mad_agent_priv;
2088 struct ib_mad_send_wr_private *mad_send_wr;
2089 struct ib_mad_send_wc mad_send_wc;
2090 unsigned long flags;
2092 mad_agent_priv = data;
2094 mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2095 mad_send_wc.vendor_err = 0;
2097 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2098 while (!list_empty(&mad_agent_priv->canceled_list)) {
2099 mad_send_wr = list_entry(mad_agent_priv->canceled_list.next,
2100 struct ib_mad_send_wr_private,
2103 list_del(&mad_send_wr->agent_list);
2104 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2106 mad_send_wc.wr_id = mad_send_wr->wr_id;
2107 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2111 if (atomic_dec_and_test(&mad_agent_priv->refcount))
2112 wake_up(&mad_agent_priv->wait);
2113 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2115 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2118 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2121 struct ib_mad_agent_private *mad_agent_priv;
2122 struct ib_mad_send_wr_private *mad_send_wr;
2123 unsigned long flags;
2125 mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2127 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2128 mad_send_wr = find_send_by_wr_id(mad_agent_priv, wr_id);
2130 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2134 if (mad_send_wr->status == IB_WC_SUCCESS)
2135 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2137 if (mad_send_wr->refcount != 0) {
2138 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2139 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2143 list_del(&mad_send_wr->agent_list);
2144 list_add_tail(&mad_send_wr->agent_list, &mad_agent_priv->canceled_list);
2145 adjust_timeout(mad_agent_priv);
2146 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2148 queue_work(mad_agent_priv->qp_info->port_priv->wq,
2149 &mad_agent_priv->canceled_work);
2153 EXPORT_SYMBOL(ib_cancel_mad);
2155 static void local_completions(void *data)
2157 struct ib_mad_agent_private *mad_agent_priv;
2158 struct ib_mad_local_private *local;
2159 struct ib_mad_agent_private *recv_mad_agent;
2160 unsigned long flags;
2162 struct ib_mad_send_wc mad_send_wc;
2164 mad_agent_priv = (struct ib_mad_agent_private *)data;
2166 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2167 while (!list_empty(&mad_agent_priv->local_list)) {
2168 local = list_entry(mad_agent_priv->local_list.next,
2169 struct ib_mad_local_private,
2171 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2172 if (local->mad_priv) {
2173 recv_mad_agent = local->recv_mad_agent;
2174 if (!recv_mad_agent) {
2175 printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2176 kmem_cache_free(ib_mad_cache, local->mad_priv);
2177 goto local_send_completion;
2181 * Defined behavior is to complete response
2184 build_smp_wc(local->wr_id, IB_LID_PERMISSIVE,
2186 recv_mad_agent->agent.port_num, &wc);
2188 local->mad_priv->header.recv_wc.wc = &wc;
2189 local->mad_priv->header.recv_wc.mad_len =
2190 sizeof(struct ib_mad);
2191 INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.recv_buf.list);
2192 local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2193 local->mad_priv->header.recv_wc.recv_buf.mad =
2194 &local->mad_priv->mad.mad;
2195 if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2196 snoop_recv(recv_mad_agent->qp_info,
2197 &local->mad_priv->header.recv_wc,
2198 IB_MAD_SNOOP_RECVS);
2199 recv_mad_agent->agent.recv_handler(
2200 &recv_mad_agent->agent,
2201 &local->mad_priv->header.recv_wc);
2202 spin_lock_irqsave(&recv_mad_agent->lock, flags);
2203 atomic_dec(&recv_mad_agent->refcount);
2204 spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2207 local_send_completion:
2209 mad_send_wc.status = IB_WC_SUCCESS;
2210 mad_send_wc.vendor_err = 0;
2211 mad_send_wc.wr_id = local->wr_id;
2212 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2213 snoop_send(mad_agent_priv->qp_info, &local->send_wr,
2215 IB_MAD_SNOOP_SEND_COMPLETIONS);
2216 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2219 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2220 list_del(&local->completion_list);
2221 atomic_dec(&mad_agent_priv->refcount);
2224 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2227 static void timeout_sends(void *data)
2229 struct ib_mad_agent_private *mad_agent_priv;
2230 struct ib_mad_send_wr_private *mad_send_wr;
2231 struct ib_mad_send_wc mad_send_wc;
2232 unsigned long flags, delay;
2234 mad_agent_priv = (struct ib_mad_agent_private *)data;
2236 mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2237 mad_send_wc.vendor_err = 0;
2239 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2240 while (!list_empty(&mad_agent_priv->wait_list)) {
2241 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2242 struct ib_mad_send_wr_private,
2245 if (time_after(mad_send_wr->timeout, jiffies)) {
2246 delay = mad_send_wr->timeout - jiffies;
2247 if ((long)delay <= 0)
2249 queue_delayed_work(mad_agent_priv->qp_info->
2251 &mad_agent_priv->timed_work, delay);
2255 list_del(&mad_send_wr->agent_list);
2256 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2258 mad_send_wc.wr_id = mad_send_wr->wr_id;
2259 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2263 atomic_dec(&mad_agent_priv->refcount);
2264 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2266 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2269 static void ib_mad_thread_completion_handler(struct ib_cq *cq)
2271 struct ib_mad_port_private *port_priv = cq->cq_context;
2273 queue_work(port_priv->wq, &port_priv->work);
2277 * Allocate receive MADs and post receive WRs for them
2279 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2280 struct ib_mad_private *mad)
2282 unsigned long flags;
2284 struct ib_mad_private *mad_priv;
2285 struct ib_sge sg_list;
2286 struct ib_recv_wr recv_wr, *bad_recv_wr;
2287 struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2289 /* Initialize common scatter list fields */
2290 sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2291 sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2293 /* Initialize common receive WR fields */
2294 recv_wr.next = NULL;
2295 recv_wr.sg_list = &sg_list;
2296 recv_wr.num_sge = 1;
2299 /* Allocate and map receive buffer */
2304 mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2306 printk(KERN_ERR PFX "No memory for receive buffer\n");
2311 sg_list.addr = dma_map_single(qp_info->port_priv->
2315 sizeof mad_priv->header,
2317 pci_unmap_addr_set(&mad_priv->header, mapping, sg_list.addr);
2318 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2319 mad_priv->header.mad_list.mad_queue = recv_queue;
2321 /* Post receive WR */
2322 spin_lock_irqsave(&recv_queue->lock, flags);
2323 post = (++recv_queue->count < recv_queue->max_active);
2324 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2325 spin_unlock_irqrestore(&recv_queue->lock, flags);
2326 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2328 spin_lock_irqsave(&recv_queue->lock, flags);
2329 list_del(&mad_priv->header.mad_list.list);
2330 recv_queue->count--;
2331 spin_unlock_irqrestore(&recv_queue->lock, flags);
2332 dma_unmap_single(qp_info->port_priv->device->dma_device,
2333 pci_unmap_addr(&mad_priv->header,
2336 sizeof mad_priv->header,
2338 kmem_cache_free(ib_mad_cache, mad_priv);
2339 printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2348 * Return all the posted receive MADs
2350 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2352 struct ib_mad_private_header *mad_priv_hdr;
2353 struct ib_mad_private *recv;
2354 struct ib_mad_list_head *mad_list;
2356 while (!list_empty(&qp_info->recv_queue.list)) {
2358 mad_list = list_entry(qp_info->recv_queue.list.next,
2359 struct ib_mad_list_head, list);
2360 mad_priv_hdr = container_of(mad_list,
2361 struct ib_mad_private_header,
2363 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2366 /* Remove from posted receive MAD list */
2367 list_del(&mad_list->list);
2369 dma_unmap_single(qp_info->port_priv->device->dma_device,
2370 pci_unmap_addr(&recv->header, mapping),
2371 sizeof(struct ib_mad_private) -
2372 sizeof(struct ib_mad_private_header),
2374 kmem_cache_free(ib_mad_cache, recv);
2377 qp_info->recv_queue.count = 0;
2383 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2386 struct ib_qp_attr *attr;
2389 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2391 printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2395 for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2396 qp = port_priv->qp_info[i].qp;
2398 * PKey index for QP1 is irrelevant but
2399 * one is needed for the Reset to Init transition
2401 attr->qp_state = IB_QPS_INIT;
2402 attr->pkey_index = 0;
2403 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2404 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2405 IB_QP_PKEY_INDEX | IB_QP_QKEY);
2407 printk(KERN_ERR PFX "Couldn't change QP%d state to "
2408 "INIT: %d\n", i, ret);
2412 attr->qp_state = IB_QPS_RTR;
2413 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2415 printk(KERN_ERR PFX "Couldn't change QP%d state to "
2416 "RTR: %d\n", i, ret);
2420 attr->qp_state = IB_QPS_RTS;
2421 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2422 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2424 printk(KERN_ERR PFX "Couldn't change QP%d state to "
2425 "RTS: %d\n", i, ret);
2430 ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2432 printk(KERN_ERR PFX "Failed to request completion "
2433 "notification: %d\n", ret);
2437 for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2438 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2440 printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2449 static void qp_event_handler(struct ib_event *event, void *qp_context)
2451 struct ib_mad_qp_info *qp_info = qp_context;
2453 /* It's worse than that! He's dead, Jim! */
2454 printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2455 event->event, qp_info->qp->qp_num);
2458 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2459 struct ib_mad_queue *mad_queue)
2461 mad_queue->qp_info = qp_info;
2462 mad_queue->count = 0;
2463 spin_lock_init(&mad_queue->lock);
2464 INIT_LIST_HEAD(&mad_queue->list);
2467 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2468 struct ib_mad_qp_info *qp_info)
2470 qp_info->port_priv = port_priv;
2471 init_mad_queue(qp_info, &qp_info->send_queue);
2472 init_mad_queue(qp_info, &qp_info->recv_queue);
2473 INIT_LIST_HEAD(&qp_info->overflow_list);
2474 spin_lock_init(&qp_info->snoop_lock);
2475 qp_info->snoop_table = NULL;
2476 qp_info->snoop_table_size = 0;
2477 atomic_set(&qp_info->snoop_count, 0);
2480 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2481 enum ib_qp_type qp_type)
2483 struct ib_qp_init_attr qp_init_attr;
2486 memset(&qp_init_attr, 0, sizeof qp_init_attr);
2487 qp_init_attr.send_cq = qp_info->port_priv->cq;
2488 qp_init_attr.recv_cq = qp_info->port_priv->cq;
2489 qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2490 qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2491 qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2492 qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2493 qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2494 qp_init_attr.qp_type = qp_type;
2495 qp_init_attr.port_num = qp_info->port_priv->port_num;
2496 qp_init_attr.qp_context = qp_info;
2497 qp_init_attr.event_handler = qp_event_handler;
2498 qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2499 if (IS_ERR(qp_info->qp)) {
2500 printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2501 get_spl_qp_index(qp_type));
2502 ret = PTR_ERR(qp_info->qp);
2505 /* Use minimum queue sizes unless the CQ is resized */
2506 qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2507 qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2514 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2516 ib_destroy_qp(qp_info->qp);
2517 if (qp_info->snoop_table)
2518 kfree(qp_info->snoop_table);
2523 * Create the QP, PD, MR, and CQ if needed
2525 static int ib_mad_port_open(struct ib_device *device,
2529 struct ib_mad_port_private *port_priv;
2530 unsigned long flags;
2531 char name[sizeof "ib_mad123"];
2533 /* First, check if port already open at MAD layer */
2534 port_priv = ib_get_mad_port(device, port_num);
2536 printk(KERN_DEBUG PFX "%s port %d already open\n",
2537 device->name, port_num);
2541 /* Create new device info */
2542 port_priv = kmalloc(sizeof *port_priv, GFP_KERNEL);
2544 printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2547 memset(port_priv, 0, sizeof *port_priv);
2548 port_priv->device = device;
2549 port_priv->port_num = port_num;
2550 spin_lock_init(&port_priv->reg_lock);
2551 INIT_LIST_HEAD(&port_priv->agent_list);
2552 init_mad_qp(port_priv, &port_priv->qp_info[0]);
2553 init_mad_qp(port_priv, &port_priv->qp_info[1]);
2555 cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2556 port_priv->cq = ib_create_cq(port_priv->device,
2558 ib_mad_thread_completion_handler,
2559 NULL, port_priv, cq_size);
2560 if (IS_ERR(port_priv->cq)) {
2561 printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2562 ret = PTR_ERR(port_priv->cq);
2566 port_priv->pd = ib_alloc_pd(device);
2567 if (IS_ERR(port_priv->pd)) {
2568 printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2569 ret = PTR_ERR(port_priv->pd);
2573 port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2574 if (IS_ERR(port_priv->mr)) {
2575 printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2576 ret = PTR_ERR(port_priv->mr);
2580 ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2583 ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2587 snprintf(name, sizeof name, "ib_mad%d", port_num);
2588 port_priv->wq = create_singlethread_workqueue(name);
2589 if (!port_priv->wq) {
2593 INIT_WORK(&port_priv->work, ib_mad_completion_handler, port_priv);
2595 ret = ib_mad_port_start(port_priv);
2597 printk(KERN_ERR PFX "Couldn't start port\n");
2601 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2602 list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2603 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2607 destroy_workqueue(port_priv->wq);
2609 destroy_mad_qp(&port_priv->qp_info[1]);
2611 destroy_mad_qp(&port_priv->qp_info[0]);
2613 ib_dereg_mr(port_priv->mr);
2615 ib_dealloc_pd(port_priv->pd);
2617 ib_destroy_cq(port_priv->cq);
2618 cleanup_recv_queue(&port_priv->qp_info[1]);
2619 cleanup_recv_queue(&port_priv->qp_info[0]);
2628 * If there are no classes using the port, free the port
2629 * resources (CQ, MR, PD, QP) and remove the port's info structure
2631 static int ib_mad_port_close(struct ib_device *device, int port_num)
2633 struct ib_mad_port_private *port_priv;
2634 unsigned long flags;
2636 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2637 port_priv = __ib_get_mad_port(device, port_num);
2638 if (port_priv == NULL) {
2639 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2640 printk(KERN_ERR PFX "Port %d not found\n", port_num);
2643 list_del(&port_priv->port_list);
2644 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2646 /* Stop processing completions. */
2647 flush_workqueue(port_priv->wq);
2648 destroy_workqueue(port_priv->wq);
2649 destroy_mad_qp(&port_priv->qp_info[1]);
2650 destroy_mad_qp(&port_priv->qp_info[0]);
2651 ib_dereg_mr(port_priv->mr);
2652 ib_dealloc_pd(port_priv->pd);
2653 ib_destroy_cq(port_priv->cq);
2654 cleanup_recv_queue(&port_priv->qp_info[1]);
2655 cleanup_recv_queue(&port_priv->qp_info[0]);
2656 /* XXX: Handle deallocation of MAD registration tables */
2663 static void ib_mad_init_device(struct ib_device *device)
2665 int ret, num_ports, cur_port, i, ret2;
2667 if (device->node_type == IB_NODE_SWITCH) {
2671 num_ports = device->phys_port_cnt;
2674 for (i = 0; i < num_ports; i++, cur_port++) {
2675 ret = ib_mad_port_open(device, cur_port);
2677 printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2678 device->name, cur_port);
2679 goto error_device_open;
2681 ret = ib_agent_port_open(device, cur_port);
2683 printk(KERN_ERR PFX "Couldn't open %s port %d "
2685 device->name, cur_port);
2686 goto error_device_open;
2690 goto error_device_query;
2695 ret2 = ib_agent_port_close(device, cur_port);
2697 printk(KERN_ERR PFX "Couldn't close %s port %d "
2699 device->name, cur_port);
2701 ret2 = ib_mad_port_close(device, cur_port);
2703 printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2704 device->name, cur_port);
2713 static void ib_mad_remove_device(struct ib_device *device)
2715 int ret = 0, i, num_ports, cur_port, ret2;
2717 if (device->node_type == IB_NODE_SWITCH) {
2721 num_ports = device->phys_port_cnt;
2724 for (i = 0; i < num_ports; i++, cur_port++) {
2725 ret2 = ib_agent_port_close(device, cur_port);
2727 printk(KERN_ERR PFX "Couldn't close %s port %d "
2729 device->name, cur_port);
2733 ret2 = ib_mad_port_close(device, cur_port);
2735 printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2736 device->name, cur_port);
2743 static struct ib_client mad_client = {
2745 .add = ib_mad_init_device,
2746 .remove = ib_mad_remove_device
2749 static int __init ib_mad_init_module(void)
2753 spin_lock_init(&ib_mad_port_list_lock);
2754 spin_lock_init(&ib_agent_port_list_lock);
2756 ib_mad_cache = kmem_cache_create("ib_mad",
2757 sizeof(struct ib_mad_private),
2762 if (!ib_mad_cache) {
2763 printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2768 INIT_LIST_HEAD(&ib_mad_port_list);
2770 if (ib_register_client(&mad_client)) {
2771 printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2779 kmem_cache_destroy(ib_mad_cache);
2784 static void __exit ib_mad_cleanup_module(void)
2786 ib_unregister_client(&mad_client);
2788 if (kmem_cache_destroy(ib_mad_cache)) {
2789 printk(KERN_DEBUG PFX "Failed to destroy ib_mad cache\n");
2793 module_init(ib_mad_init_module);
2794 module_exit(ib_mad_cleanup_module);