2 * Copyright (c) 2004-2007 Voltaire, Inc. All rights reserved.
3 * Copyright (c) 2005 Intel Corporation. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies Ltd. All rights reserved.
5 * Copyright (c) 2009 HNR Consulting. All rights reserved.
6 * Copyright (c) 2014,2018 Intel Corporation. All rights reserved.
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 <linux/security.h>
44 #include <linux/xarray.h>
45 #include <rdma/ib_cache.h>
48 #include "core_priv.h"
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/ib_mad.h>
57 #ifdef CONFIG_TRACEPOINTS
58 static void create_mad_addr_info(struct ib_mad_send_wr_private *mad_send_wr,
59 struct ib_mad_qp_info *qp_info,
60 struct trace_event_raw_ib_mad_send_template *entry)
63 struct ib_device *dev = qp_info->port_priv->device;
64 u8 pnum = qp_info->port_priv->port_num;
65 struct ib_ud_wr *wr = &mad_send_wr->send_wr;
66 struct rdma_ah_attr attr = {};
68 rdma_query_ah(wr->ah, &attr);
70 /* These are common */
72 ib_query_pkey(dev, pnum, wr->pkey_index, &pkey);
74 entry->rqpn = wr->remote_qpn;
75 entry->rqkey = wr->remote_qkey;
76 entry->dlid = rdma_ah_get_dlid(&attr);
80 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
81 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
83 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
84 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
85 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
86 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
88 static DEFINE_XARRAY_ALLOC1(ib_mad_clients);
89 static u32 ib_mad_client_next;
90 static struct list_head ib_mad_port_list;
93 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
95 /* Forward declarations */
96 static int method_in_use(struct ib_mad_mgmt_method_table **method,
97 struct ib_mad_reg_req *mad_reg_req);
98 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
99 static struct ib_mad_agent_private *find_mad_agent(
100 struct ib_mad_port_private *port_priv,
101 const struct ib_mad_hdr *mad);
102 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
103 struct ib_mad_private *mad);
104 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
105 static void timeout_sends(struct work_struct *work);
106 static void local_completions(struct work_struct *work);
107 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
108 struct ib_mad_agent_private *agent_priv,
110 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
111 struct ib_mad_agent_private *agent_priv);
112 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
114 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
117 * Returns a ib_mad_port_private structure or NULL for a device/port
118 * Assumes ib_mad_port_list_lock is being held
120 static inline struct ib_mad_port_private *
121 __ib_get_mad_port(struct ib_device *device, int port_num)
123 struct ib_mad_port_private *entry;
125 list_for_each_entry(entry, &ib_mad_port_list, port_list) {
126 if (entry->device == device && entry->port_num == port_num)
133 * Wrapper function to return a ib_mad_port_private structure or NULL
136 static inline struct ib_mad_port_private *
137 ib_get_mad_port(struct ib_device *device, int port_num)
139 struct ib_mad_port_private *entry;
142 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
143 entry = __ib_get_mad_port(device, port_num);
144 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
149 static inline u8 convert_mgmt_class(u8 mgmt_class)
151 /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
152 return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
156 static int get_spl_qp_index(enum ib_qp_type qp_type)
169 static int vendor_class_index(u8 mgmt_class)
171 return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
174 static int is_vendor_class(u8 mgmt_class)
176 if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
177 (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
182 static int is_vendor_oui(char *oui)
184 if (oui[0] || oui[1] || oui[2])
189 static int is_vendor_method_in_use(
190 struct ib_mad_mgmt_vendor_class *vendor_class,
191 struct ib_mad_reg_req *mad_reg_req)
193 struct ib_mad_mgmt_method_table *method;
196 for (i = 0; i < MAX_MGMT_OUI; i++) {
197 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
198 method = vendor_class->method_table[i];
200 if (method_in_use(&method, mad_reg_req))
210 int ib_response_mad(const struct ib_mad_hdr *hdr)
212 return ((hdr->method & IB_MGMT_METHOD_RESP) ||
213 (hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
214 ((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
215 (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
217 EXPORT_SYMBOL(ib_response_mad);
220 * ib_register_mad_agent - Register to send/receive MADs
222 * Context: Process context.
224 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
226 enum ib_qp_type qp_type,
227 struct ib_mad_reg_req *mad_reg_req,
229 ib_mad_send_handler send_handler,
230 ib_mad_recv_handler recv_handler,
232 u32 registration_flags)
234 struct ib_mad_port_private *port_priv;
235 struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
236 struct ib_mad_agent_private *mad_agent_priv;
237 struct ib_mad_reg_req *reg_req = NULL;
238 struct ib_mad_mgmt_class_table *class;
239 struct ib_mad_mgmt_vendor_class_table *vendor;
240 struct ib_mad_mgmt_vendor_class *vendor_class;
241 struct ib_mad_mgmt_method_table *method;
243 u8 mgmt_class, vclass;
245 if ((qp_type == IB_QPT_SMI && !rdma_cap_ib_smi(device, port_num)) ||
246 (qp_type == IB_QPT_GSI && !rdma_cap_ib_cm(device, port_num)))
247 return ERR_PTR(-EPROTONOSUPPORT);
249 /* Validate parameters */
250 qpn = get_spl_qp_index(qp_type);
252 dev_dbg_ratelimited(&device->dev, "%s: invalid QP Type %d\n",
257 if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
258 dev_dbg_ratelimited(&device->dev,
259 "%s: invalid RMPP Version %u\n",
260 __func__, rmpp_version);
264 /* Validate MAD registration request if supplied */
266 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
267 dev_dbg_ratelimited(&device->dev,
268 "%s: invalid Class Version %u\n",
270 mad_reg_req->mgmt_class_version);
274 dev_dbg_ratelimited(&device->dev,
275 "%s: no recv_handler\n", __func__);
278 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
280 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
281 * one in this range currently allowed
283 if (mad_reg_req->mgmt_class !=
284 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
285 dev_dbg_ratelimited(&device->dev,
286 "%s: Invalid Mgmt Class 0x%x\n",
287 __func__, mad_reg_req->mgmt_class);
290 } else if (mad_reg_req->mgmt_class == 0) {
292 * Class 0 is reserved in IBA and is used for
293 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
295 dev_dbg_ratelimited(&device->dev,
296 "%s: Invalid Mgmt Class 0\n",
299 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
301 * If class is in "new" vendor range,
302 * ensure supplied OUI is not zero
304 if (!is_vendor_oui(mad_reg_req->oui)) {
305 dev_dbg_ratelimited(&device->dev,
306 "%s: No OUI specified for class 0x%x\n",
308 mad_reg_req->mgmt_class);
312 /* Make sure class supplied is consistent with RMPP */
313 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
315 dev_dbg_ratelimited(&device->dev,
316 "%s: RMPP version for non-RMPP class 0x%x\n",
317 __func__, mad_reg_req->mgmt_class);
322 /* Make sure class supplied is consistent with QP type */
323 if (qp_type == IB_QPT_SMI) {
324 if ((mad_reg_req->mgmt_class !=
325 IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
326 (mad_reg_req->mgmt_class !=
327 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
328 dev_dbg_ratelimited(&device->dev,
329 "%s: Invalid SM QP type: class 0x%x\n",
330 __func__, mad_reg_req->mgmt_class);
334 if ((mad_reg_req->mgmt_class ==
335 IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
336 (mad_reg_req->mgmt_class ==
337 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
338 dev_dbg_ratelimited(&device->dev,
339 "%s: Invalid GS QP type: class 0x%x\n",
340 __func__, mad_reg_req->mgmt_class);
345 /* No registration request supplied */
348 if (registration_flags & IB_MAD_USER_RMPP)
352 /* Validate device and port */
353 port_priv = ib_get_mad_port(device, port_num);
355 dev_dbg_ratelimited(&device->dev, "%s: Invalid port %d\n",
357 ret = ERR_PTR(-ENODEV);
361 /* Verify the QP requested is supported. For example, Ethernet devices
364 if (!port_priv->qp_info[qpn].qp) {
365 dev_dbg_ratelimited(&device->dev, "%s: QP %d not supported\n",
367 ret = ERR_PTR(-EPROTONOSUPPORT);
371 /* Allocate structures */
372 mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
373 if (!mad_agent_priv) {
374 ret = ERR_PTR(-ENOMEM);
379 reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
381 ret = ERR_PTR(-ENOMEM);
386 /* Now, fill in the various structures */
387 mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
388 mad_agent_priv->reg_req = reg_req;
389 mad_agent_priv->agent.rmpp_version = rmpp_version;
390 mad_agent_priv->agent.device = device;
391 mad_agent_priv->agent.recv_handler = recv_handler;
392 mad_agent_priv->agent.send_handler = send_handler;
393 mad_agent_priv->agent.context = context;
394 mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
395 mad_agent_priv->agent.port_num = port_num;
396 mad_agent_priv->agent.flags = registration_flags;
397 spin_lock_init(&mad_agent_priv->lock);
398 INIT_LIST_HEAD(&mad_agent_priv->send_list);
399 INIT_LIST_HEAD(&mad_agent_priv->wait_list);
400 INIT_LIST_HEAD(&mad_agent_priv->done_list);
401 INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
402 INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
403 INIT_LIST_HEAD(&mad_agent_priv->local_list);
404 INIT_WORK(&mad_agent_priv->local_work, local_completions);
405 refcount_set(&mad_agent_priv->refcount, 1);
406 init_completion(&mad_agent_priv->comp);
408 ret2 = ib_mad_agent_security_setup(&mad_agent_priv->agent, qp_type);
415 * The mlx4 driver uses the top byte to distinguish which virtual
416 * function generated the MAD, so we must avoid using it.
418 ret2 = xa_alloc_cyclic(&ib_mad_clients, &mad_agent_priv->agent.hi_tid,
419 mad_agent_priv, XA_LIMIT(0, (1 << 24) - 1),
420 &ib_mad_client_next, GFP_KERNEL);
427 * Make sure MAD registration (if supplied)
428 * is non overlapping with any existing ones
430 spin_lock_irq(&port_priv->reg_lock);
432 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
433 if (!is_vendor_class(mgmt_class)) {
434 class = port_priv->version[mad_reg_req->
435 mgmt_class_version].class;
437 method = class->method_table[mgmt_class];
439 if (method_in_use(&method,
444 ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
447 /* "New" vendor class range */
448 vendor = port_priv->version[mad_reg_req->
449 mgmt_class_version].vendor;
451 vclass = vendor_class_index(mgmt_class);
452 vendor_class = vendor->vendor_class[vclass];
454 if (is_vendor_method_in_use(
460 ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
467 spin_unlock_irq(&port_priv->reg_lock);
469 trace_ib_mad_create_agent(mad_agent_priv);
470 return &mad_agent_priv->agent;
472 spin_unlock_irq(&port_priv->reg_lock);
473 xa_erase(&ib_mad_clients, mad_agent_priv->agent.hi_tid);
475 ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
479 kfree(mad_agent_priv);
483 EXPORT_SYMBOL(ib_register_mad_agent);
485 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
487 if (refcount_dec_and_test(&mad_agent_priv->refcount))
488 complete(&mad_agent_priv->comp);
491 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
493 struct ib_mad_port_private *port_priv;
495 /* Note that we could still be handling received MADs */
496 trace_ib_mad_unregister_agent(mad_agent_priv);
499 * Canceling all sends results in dropping received response
500 * MADs, preventing us from queuing additional work
502 cancel_mads(mad_agent_priv);
503 port_priv = mad_agent_priv->qp_info->port_priv;
504 cancel_delayed_work(&mad_agent_priv->timed_work);
506 spin_lock_irq(&port_priv->reg_lock);
507 remove_mad_reg_req(mad_agent_priv);
508 spin_unlock_irq(&port_priv->reg_lock);
509 xa_erase(&ib_mad_clients, mad_agent_priv->agent.hi_tid);
511 flush_workqueue(port_priv->wq);
513 deref_mad_agent(mad_agent_priv);
514 wait_for_completion(&mad_agent_priv->comp);
515 ib_cancel_rmpp_recvs(mad_agent_priv);
517 ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
519 kfree(mad_agent_priv->reg_req);
520 kfree_rcu(mad_agent_priv, rcu);
524 * ib_unregister_mad_agent - Unregisters a client from using MAD services
526 * Context: Process context.
528 void ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
530 struct ib_mad_agent_private *mad_agent_priv;
532 mad_agent_priv = container_of(mad_agent,
533 struct ib_mad_agent_private,
535 unregister_mad_agent(mad_agent_priv);
537 EXPORT_SYMBOL(ib_unregister_mad_agent);
539 static void dequeue_mad(struct ib_mad_list_head *mad_list)
541 struct ib_mad_queue *mad_queue;
544 mad_queue = mad_list->mad_queue;
545 spin_lock_irqsave(&mad_queue->lock, flags);
546 list_del(&mad_list->list);
548 spin_unlock_irqrestore(&mad_queue->lock, flags);
551 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
552 u16 pkey_index, u8 port_num, struct ib_wc *wc)
554 memset(wc, 0, sizeof *wc);
556 wc->status = IB_WC_SUCCESS;
557 wc->opcode = IB_WC_RECV;
558 wc->pkey_index = pkey_index;
559 wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
564 wc->dlid_path_bits = 0;
565 wc->port_num = port_num;
568 static size_t mad_priv_size(const struct ib_mad_private *mp)
570 return sizeof(struct ib_mad_private) + mp->mad_size;
573 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
575 size_t size = sizeof(struct ib_mad_private) + mad_size;
576 struct ib_mad_private *ret = kzalloc(size, flags);
579 ret->mad_size = mad_size;
584 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
586 return rdma_max_mad_size(port_priv->device, port_priv->port_num);
589 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
591 return sizeof(struct ib_grh) + mp->mad_size;
595 * Return 0 if SMP is to be sent
596 * Return 1 if SMP was consumed locally (whether or not solicited)
597 * Return < 0 if error
599 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
600 struct ib_mad_send_wr_private *mad_send_wr)
603 struct ib_smp *smp = mad_send_wr->send_buf.mad;
604 struct opa_smp *opa_smp = (struct opa_smp *)smp;
606 struct ib_mad_local_private *local;
607 struct ib_mad_private *mad_priv;
608 struct ib_mad_port_private *port_priv;
609 struct ib_mad_agent_private *recv_mad_agent = NULL;
610 struct ib_device *device = mad_agent_priv->agent.device;
613 struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
614 size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
615 u16 out_mad_pkey_index = 0;
617 bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
618 mad_agent_priv->qp_info->port_priv->port_num);
620 if (rdma_cap_ib_switch(device) &&
621 smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
622 port_num = send_wr->port_num;
624 port_num = mad_agent_priv->agent.port_num;
627 * Directed route handling starts if the initial LID routed part of
628 * a request or the ending LID routed part of a response is empty.
629 * If we are at the start of the LID routed part, don't update the
630 * hop_ptr or hop_cnt. See section 14.2.2, Vol 1 IB spec.
632 if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
635 trace_ib_mad_handle_out_opa_smi(opa_smp);
637 if ((opa_get_smp_direction(opa_smp)
638 ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
639 OPA_LID_PERMISSIVE &&
640 opa_smi_handle_dr_smp_send(opa_smp,
641 rdma_cap_ib_switch(device),
642 port_num) == IB_SMI_DISCARD) {
644 dev_err(&device->dev, "OPA Invalid directed route\n");
647 opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
648 if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
649 opa_drslid & 0xffff0000) {
651 dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
655 drslid = (u16)(opa_drslid & 0x0000ffff);
657 /* Check to post send on QP or process locally */
658 if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
659 opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
662 trace_ib_mad_handle_out_ib_smi(smp);
664 if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
666 smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
669 dev_err(&device->dev, "Invalid directed route\n");
672 drslid = be16_to_cpu(smp->dr_slid);
674 /* Check to post send on QP or process locally */
675 if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
676 smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
680 local = kmalloc(sizeof *local, GFP_ATOMIC);
685 local->mad_priv = NULL;
686 local->recv_mad_agent = NULL;
687 mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
694 build_smp_wc(mad_agent_priv->agent.qp,
695 send_wr->wr.wr_cqe, drslid,
697 send_wr->port_num, &mad_wc);
699 if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
700 mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
701 + mad_send_wr->send_buf.data_len
702 + sizeof(struct ib_grh);
705 /* No GRH for DR SMP */
706 ret = device->ops.process_mad(device, 0, port_num, &mad_wc, NULL,
707 (const struct ib_mad *)smp,
708 (struct ib_mad *)mad_priv->mad, &mad_size,
709 &out_mad_pkey_index);
712 case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
713 if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
714 mad_agent_priv->agent.recv_handler) {
715 local->mad_priv = mad_priv;
716 local->recv_mad_agent = mad_agent_priv;
718 * Reference MAD agent until receive
719 * side of local completion handled
721 refcount_inc(&mad_agent_priv->refcount);
725 case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
728 case IB_MAD_RESULT_SUCCESS:
729 /* Treat like an incoming receive MAD */
730 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
731 mad_agent_priv->agent.port_num);
733 memcpy(mad_priv->mad, smp, mad_priv->mad_size);
734 recv_mad_agent = find_mad_agent(port_priv,
735 (const struct ib_mad_hdr *)mad_priv->mad);
737 if (!port_priv || !recv_mad_agent) {
739 * No receiving agent so drop packet and
740 * generate send completion.
745 local->mad_priv = mad_priv;
746 local->recv_mad_agent = recv_mad_agent;
755 local->mad_send_wr = mad_send_wr;
757 local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
758 local->return_wc_byte_len = mad_size;
760 /* Reference MAD agent until send side of local completion handled */
761 refcount_inc(&mad_agent_priv->refcount);
762 /* Queue local completion to local list */
763 spin_lock_irqsave(&mad_agent_priv->lock, flags);
764 list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
765 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
766 queue_work(mad_agent_priv->qp_info->port_priv->wq,
767 &mad_agent_priv->local_work);
773 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
777 seg_size = mad_size - hdr_len;
778 if (data_len && seg_size) {
779 pad = seg_size - data_len % seg_size;
780 return pad == seg_size ? 0 : pad;
785 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
787 struct ib_rmpp_segment *s, *t;
789 list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
795 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
796 size_t mad_size, gfp_t gfp_mask)
798 struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
799 struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
800 struct ib_rmpp_segment *seg = NULL;
801 int left, seg_size, pad;
803 send_buf->seg_size = mad_size - send_buf->hdr_len;
804 send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
805 seg_size = send_buf->seg_size;
808 /* Allocate data segments. */
809 for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
810 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
812 free_send_rmpp_list(send_wr);
815 seg->num = ++send_buf->seg_count;
816 list_add_tail(&seg->list, &send_wr->rmpp_list);
819 /* Zero any padding */
821 memset(seg->data + seg_size - pad, 0, pad);
823 rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
825 rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
826 ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
828 send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
829 struct ib_rmpp_segment, list);
830 send_wr->last_ack_seg = send_wr->cur_seg;
834 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
836 return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
838 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
840 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
841 u32 remote_qpn, u16 pkey_index,
843 int hdr_len, int data_len,
847 struct ib_mad_agent_private *mad_agent_priv;
848 struct ib_mad_send_wr_private *mad_send_wr;
849 int pad, message_size, ret, size;
854 mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
857 opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
859 if (opa && base_version == OPA_MGMT_BASE_VERSION)
860 mad_size = sizeof(struct opa_mad);
862 mad_size = sizeof(struct ib_mad);
864 pad = get_pad_size(hdr_len, data_len, mad_size);
865 message_size = hdr_len + data_len + pad;
867 if (ib_mad_kernel_rmpp_agent(mad_agent)) {
868 if (!rmpp_active && message_size > mad_size)
869 return ERR_PTR(-EINVAL);
871 if (rmpp_active || message_size > mad_size)
872 return ERR_PTR(-EINVAL);
874 size = rmpp_active ? hdr_len : mad_size;
875 buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
877 return ERR_PTR(-ENOMEM);
879 mad_send_wr = buf + size;
880 INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
881 mad_send_wr->send_buf.mad = buf;
882 mad_send_wr->send_buf.hdr_len = hdr_len;
883 mad_send_wr->send_buf.data_len = data_len;
884 mad_send_wr->pad = pad;
886 mad_send_wr->mad_agent_priv = mad_agent_priv;
887 mad_send_wr->sg_list[0].length = hdr_len;
888 mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
890 /* OPA MADs don't have to be the full 2048 bytes */
891 if (opa && base_version == OPA_MGMT_BASE_VERSION &&
892 data_len < mad_size - hdr_len)
893 mad_send_wr->sg_list[1].length = data_len;
895 mad_send_wr->sg_list[1].length = mad_size - hdr_len;
897 mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
899 mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
901 mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
902 mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
903 mad_send_wr->send_wr.wr.num_sge = 2;
904 mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
905 mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
906 mad_send_wr->send_wr.remote_qpn = remote_qpn;
907 mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
908 mad_send_wr->send_wr.pkey_index = pkey_index;
911 ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
918 mad_send_wr->send_buf.mad_agent = mad_agent;
919 refcount_inc(&mad_agent_priv->refcount);
920 return &mad_send_wr->send_buf;
922 EXPORT_SYMBOL(ib_create_send_mad);
924 int ib_get_mad_data_offset(u8 mgmt_class)
926 if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
927 return IB_MGMT_SA_HDR;
928 else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
929 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
930 (mgmt_class == IB_MGMT_CLASS_BIS))
931 return IB_MGMT_DEVICE_HDR;
932 else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
933 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
934 return IB_MGMT_VENDOR_HDR;
936 return IB_MGMT_MAD_HDR;
938 EXPORT_SYMBOL(ib_get_mad_data_offset);
940 int ib_is_mad_class_rmpp(u8 mgmt_class)
942 if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
943 (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
944 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
945 (mgmt_class == IB_MGMT_CLASS_BIS) ||
946 ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
947 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
951 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
953 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
955 struct ib_mad_send_wr_private *mad_send_wr;
956 struct list_head *list;
958 mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
960 list = &mad_send_wr->cur_seg->list;
962 if (mad_send_wr->cur_seg->num < seg_num) {
963 list_for_each_entry(mad_send_wr->cur_seg, list, list)
964 if (mad_send_wr->cur_seg->num == seg_num)
966 } else if (mad_send_wr->cur_seg->num > seg_num) {
967 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
968 if (mad_send_wr->cur_seg->num == seg_num)
971 return mad_send_wr->cur_seg->data;
973 EXPORT_SYMBOL(ib_get_rmpp_segment);
975 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
977 if (mad_send_wr->send_buf.seg_count)
978 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
979 mad_send_wr->seg_num);
981 return mad_send_wr->send_buf.mad +
982 mad_send_wr->send_buf.hdr_len;
985 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
987 struct ib_mad_agent_private *mad_agent_priv;
988 struct ib_mad_send_wr_private *mad_send_wr;
990 mad_agent_priv = container_of(send_buf->mad_agent,
991 struct ib_mad_agent_private, agent);
992 mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
995 free_send_rmpp_list(mad_send_wr);
996 kfree(send_buf->mad);
997 deref_mad_agent(mad_agent_priv);
999 EXPORT_SYMBOL(ib_free_send_mad);
1001 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1003 struct ib_mad_qp_info *qp_info;
1004 struct list_head *list;
1005 struct ib_mad_agent *mad_agent;
1007 unsigned long flags;
1010 /* Set WR ID to find mad_send_wr upon completion */
1011 qp_info = mad_send_wr->mad_agent_priv->qp_info;
1012 mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1013 mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1014 mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1016 mad_agent = mad_send_wr->send_buf.mad_agent;
1017 sge = mad_send_wr->sg_list;
1018 sge[0].addr = ib_dma_map_single(mad_agent->device,
1019 mad_send_wr->send_buf.mad,
1022 if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1025 mad_send_wr->header_mapping = sge[0].addr;
1027 sge[1].addr = ib_dma_map_single(mad_agent->device,
1028 ib_get_payload(mad_send_wr),
1031 if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1032 ib_dma_unmap_single(mad_agent->device,
1033 mad_send_wr->header_mapping,
1034 sge[0].length, DMA_TO_DEVICE);
1037 mad_send_wr->payload_mapping = sge[1].addr;
1039 spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1040 if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1041 trace_ib_mad_ib_send_mad(mad_send_wr, qp_info);
1042 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1044 list = &qp_info->send_queue.list;
1047 list = &qp_info->overflow_list;
1051 qp_info->send_queue.count++;
1052 list_add_tail(&mad_send_wr->mad_list.list, list);
1054 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1056 ib_dma_unmap_single(mad_agent->device,
1057 mad_send_wr->header_mapping,
1058 sge[0].length, DMA_TO_DEVICE);
1059 ib_dma_unmap_single(mad_agent->device,
1060 mad_send_wr->payload_mapping,
1061 sge[1].length, DMA_TO_DEVICE);
1067 * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1068 * with the registered client
1070 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1071 struct ib_mad_send_buf **bad_send_buf)
1073 struct ib_mad_agent_private *mad_agent_priv;
1074 struct ib_mad_send_buf *next_send_buf;
1075 struct ib_mad_send_wr_private *mad_send_wr;
1076 unsigned long flags;
1079 /* Walk list of send WRs and post each on send list */
1080 for (; send_buf; send_buf = next_send_buf) {
1081 mad_send_wr = container_of(send_buf,
1082 struct ib_mad_send_wr_private,
1084 mad_agent_priv = mad_send_wr->mad_agent_priv;
1086 ret = ib_mad_enforce_security(mad_agent_priv,
1087 mad_send_wr->send_wr.pkey_index);
1091 if (!send_buf->mad_agent->send_handler ||
1092 (send_buf->timeout_ms &&
1093 !send_buf->mad_agent->recv_handler)) {
1098 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1099 if (mad_agent_priv->agent.rmpp_version) {
1106 * Save pointer to next work request to post in case the
1107 * current one completes, and the user modifies the work
1108 * request associated with the completion
1110 next_send_buf = send_buf->next;
1111 mad_send_wr->send_wr.ah = send_buf->ah;
1113 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1114 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1115 ret = handle_outgoing_dr_smp(mad_agent_priv,
1117 if (ret < 0) /* error */
1119 else if (ret == 1) /* locally consumed */
1123 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1124 /* Timeout will be updated after send completes */
1125 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1126 mad_send_wr->max_retries = send_buf->retries;
1127 mad_send_wr->retries_left = send_buf->retries;
1128 send_buf->retries = 0;
1129 /* Reference for work request to QP + response */
1130 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1131 mad_send_wr->status = IB_WC_SUCCESS;
1133 /* Reference MAD agent until send completes */
1134 refcount_inc(&mad_agent_priv->refcount);
1135 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1136 list_add_tail(&mad_send_wr->agent_list,
1137 &mad_agent_priv->send_list);
1138 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1140 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1141 ret = ib_send_rmpp_mad(mad_send_wr);
1142 if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1143 ret = ib_send_mad(mad_send_wr);
1145 ret = ib_send_mad(mad_send_wr);
1147 /* Fail send request */
1148 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1149 list_del(&mad_send_wr->agent_list);
1150 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1151 deref_mad_agent(mad_agent_priv);
1158 *bad_send_buf = send_buf;
1161 EXPORT_SYMBOL(ib_post_send_mad);
1164 * ib_free_recv_mad - Returns data buffers used to receive
1165 * a MAD to the access layer
1167 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1169 struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1170 struct ib_mad_private_header *mad_priv_hdr;
1171 struct ib_mad_private *priv;
1172 struct list_head free_list;
1174 INIT_LIST_HEAD(&free_list);
1175 list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1177 list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1179 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1181 mad_priv_hdr = container_of(mad_recv_wc,
1182 struct ib_mad_private_header,
1184 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1189 EXPORT_SYMBOL(ib_free_recv_mad);
1191 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1192 struct ib_mad_reg_req *mad_reg_req)
1196 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1197 if ((*method)->agent[i]) {
1198 pr_err("Method %d already in use\n", i);
1205 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1207 /* Allocate management method table */
1208 *method = kzalloc(sizeof **method, GFP_ATOMIC);
1209 return (*method) ? 0 : (-ENOMEM);
1213 * Check to see if there are any methods still in use
1215 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1219 for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1220 if (method->agent[i])
1226 * Check to see if there are any method tables for this class still in use
1228 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1232 for (i = 0; i < MAX_MGMT_CLASS; i++)
1233 if (class->method_table[i])
1238 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1242 for (i = 0; i < MAX_MGMT_OUI; i++)
1243 if (vendor_class->method_table[i])
1248 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1253 for (i = 0; i < MAX_MGMT_OUI; i++)
1254 /* Is there matching OUI for this vendor class ? */
1255 if (!memcmp(vendor_class->oui[i], oui, 3))
1261 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1265 for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1266 if (vendor->vendor_class[i])
1272 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1273 struct ib_mad_agent_private *agent)
1277 /* Remove any methods for this mad agent */
1278 for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1279 if (method->agent[i] == agent) {
1280 method->agent[i] = NULL;
1285 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1286 struct ib_mad_agent_private *agent_priv,
1289 struct ib_mad_port_private *port_priv;
1290 struct ib_mad_mgmt_class_table **class;
1291 struct ib_mad_mgmt_method_table **method;
1294 port_priv = agent_priv->qp_info->port_priv;
1295 class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1297 /* Allocate management class table for "new" class version */
1298 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1304 /* Allocate method table for this management class */
1305 method = &(*class)->method_table[mgmt_class];
1306 if ((ret = allocate_method_table(method)))
1309 method = &(*class)->method_table[mgmt_class];
1311 /* Allocate method table for this management class */
1312 if ((ret = allocate_method_table(method)))
1317 /* Now, make sure methods are not already in use */
1318 if (method_in_use(method, mad_reg_req))
1321 /* Finally, add in methods being registered */
1322 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1323 (*method)->agent[i] = agent_priv;
1328 /* Remove any methods for this mad agent */
1329 remove_methods_mad_agent(*method, agent_priv);
1330 /* Now, check to see if there are any methods in use */
1331 if (!check_method_table(*method)) {
1332 /* If not, release management method table */
1345 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1346 struct ib_mad_agent_private *agent_priv)
1348 struct ib_mad_port_private *port_priv;
1349 struct ib_mad_mgmt_vendor_class_table **vendor_table;
1350 struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1351 struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1352 struct ib_mad_mgmt_method_table **method;
1353 int i, ret = -ENOMEM;
1356 /* "New" vendor (with OUI) class */
1357 vclass = vendor_class_index(mad_reg_req->mgmt_class);
1358 port_priv = agent_priv->qp_info->port_priv;
1359 vendor_table = &port_priv->version[
1360 mad_reg_req->mgmt_class_version].vendor;
1361 if (!*vendor_table) {
1362 /* Allocate mgmt vendor class table for "new" class version */
1363 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1367 *vendor_table = vendor;
1369 if (!(*vendor_table)->vendor_class[vclass]) {
1370 /* Allocate table for this management vendor class */
1371 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1375 (*vendor_table)->vendor_class[vclass] = vendor_class;
1377 for (i = 0; i < MAX_MGMT_OUI; i++) {
1378 /* Is there matching OUI for this vendor class ? */
1379 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1380 mad_reg_req->oui, 3)) {
1381 method = &(*vendor_table)->vendor_class[
1382 vclass]->method_table[i];
1388 for (i = 0; i < MAX_MGMT_OUI; i++) {
1389 /* OUI slot available ? */
1390 if (!is_vendor_oui((*vendor_table)->vendor_class[
1392 method = &(*vendor_table)->vendor_class[
1393 vclass]->method_table[i];
1394 /* Allocate method table for this OUI */
1396 ret = allocate_method_table(method);
1400 memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1401 mad_reg_req->oui, 3);
1405 dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1409 /* Now, make sure methods are not already in use */
1410 if (method_in_use(method, mad_reg_req))
1413 /* Finally, add in methods being registered */
1414 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1415 (*method)->agent[i] = agent_priv;
1420 /* Remove any methods for this mad agent */
1421 remove_methods_mad_agent(*method, agent_priv);
1422 /* Now, check to see if there are any methods in use */
1423 if (!check_method_table(*method)) {
1424 /* If not, release management method table */
1431 (*vendor_table)->vendor_class[vclass] = NULL;
1432 kfree(vendor_class);
1436 *vendor_table = NULL;
1443 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1445 struct ib_mad_port_private *port_priv;
1446 struct ib_mad_mgmt_class_table *class;
1447 struct ib_mad_mgmt_method_table *method;
1448 struct ib_mad_mgmt_vendor_class_table *vendor;
1449 struct ib_mad_mgmt_vendor_class *vendor_class;
1454 * Was MAD registration request supplied
1455 * with original registration ?
1457 if (!agent_priv->reg_req) {
1461 port_priv = agent_priv->qp_info->port_priv;
1462 mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1463 class = port_priv->version[
1464 agent_priv->reg_req->mgmt_class_version].class;
1468 method = class->method_table[mgmt_class];
1470 /* Remove any methods for this mad agent */
1471 remove_methods_mad_agent(method, agent_priv);
1472 /* Now, check to see if there are any methods still in use */
1473 if (!check_method_table(method)) {
1474 /* If not, release management method table */
1476 class->method_table[mgmt_class] = NULL;
1477 /* Any management classes left ? */
1478 if (!check_class_table(class)) {
1479 /* If not, release management class table */
1482 agent_priv->reg_req->
1483 mgmt_class_version].class = NULL;
1489 if (!is_vendor_class(mgmt_class))
1492 /* normalize mgmt_class to vendor range 2 */
1493 mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1494 vendor = port_priv->version[
1495 agent_priv->reg_req->mgmt_class_version].vendor;
1500 vendor_class = vendor->vendor_class[mgmt_class];
1502 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1505 method = vendor_class->method_table[index];
1507 /* Remove any methods for this mad agent */
1508 remove_methods_mad_agent(method, agent_priv);
1510 * Now, check to see if there are
1511 * any methods still in use
1513 if (!check_method_table(method)) {
1514 /* If not, release management method table */
1516 vendor_class->method_table[index] = NULL;
1517 memset(vendor_class->oui[index], 0, 3);
1518 /* Any OUIs left ? */
1519 if (!check_vendor_class(vendor_class)) {
1520 /* If not, release vendor class table */
1521 kfree(vendor_class);
1522 vendor->vendor_class[mgmt_class] = NULL;
1523 /* Any other vendor classes left ? */
1524 if (!check_vendor_table(vendor)) {
1527 agent_priv->reg_req->
1528 mgmt_class_version].
1540 static struct ib_mad_agent_private *
1541 find_mad_agent(struct ib_mad_port_private *port_priv,
1542 const struct ib_mad_hdr *mad_hdr)
1544 struct ib_mad_agent_private *mad_agent = NULL;
1545 unsigned long flags;
1547 if (ib_response_mad(mad_hdr)) {
1551 * Routing is based on high 32 bits of transaction ID
1554 hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1556 mad_agent = xa_load(&ib_mad_clients, hi_tid);
1557 if (mad_agent && !refcount_inc_not_zero(&mad_agent->refcount))
1561 struct ib_mad_mgmt_class_table *class;
1562 struct ib_mad_mgmt_method_table *method;
1563 struct ib_mad_mgmt_vendor_class_table *vendor;
1564 struct ib_mad_mgmt_vendor_class *vendor_class;
1565 const struct ib_vendor_mad *vendor_mad;
1568 spin_lock_irqsave(&port_priv->reg_lock, flags);
1570 * Routing is based on version, class, and method
1571 * For "newer" vendor MADs, also based on OUI
1573 if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1575 if (!is_vendor_class(mad_hdr->mgmt_class)) {
1576 class = port_priv->version[
1577 mad_hdr->class_version].class;
1580 if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1581 ARRAY_SIZE(class->method_table))
1583 method = class->method_table[convert_mgmt_class(
1584 mad_hdr->mgmt_class)];
1586 mad_agent = method->agent[mad_hdr->method &
1587 ~IB_MGMT_METHOD_RESP];
1589 vendor = port_priv->version[
1590 mad_hdr->class_version].vendor;
1593 vendor_class = vendor->vendor_class[vendor_class_index(
1594 mad_hdr->mgmt_class)];
1597 /* Find matching OUI */
1598 vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1599 index = find_vendor_oui(vendor_class, vendor_mad->oui);
1602 method = vendor_class->method_table[index];
1604 mad_agent = method->agent[mad_hdr->method &
1605 ~IB_MGMT_METHOD_RESP];
1609 refcount_inc(&mad_agent->refcount);
1611 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1614 if (mad_agent && !mad_agent->agent.recv_handler) {
1615 dev_notice(&port_priv->device->dev,
1616 "No receive handler for client %p on port %d\n",
1617 &mad_agent->agent, port_priv->port_num);
1618 deref_mad_agent(mad_agent);
1625 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1626 const struct ib_mad_qp_info *qp_info,
1630 u32 qp_num = qp_info->qp->qp_num;
1632 /* Make sure MAD base version is understood */
1633 if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1634 (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1635 pr_err("MAD received with unsupported base version %d %s\n",
1636 mad_hdr->base_version, opa ? "(opa)" : "");
1640 /* Filter SMI packets sent to other than QP0 */
1641 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1642 (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1646 /* CM attributes other than ClassPortInfo only use Send method */
1647 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1648 (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1649 (mad_hdr->method != IB_MGMT_METHOD_SEND))
1651 /* Filter GSI packets sent to QP0 */
1660 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1661 const struct ib_mad_hdr *mad_hdr)
1663 struct ib_rmpp_mad *rmpp_mad;
1665 rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1666 return !mad_agent_priv->agent.rmpp_version ||
1667 !ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1668 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1669 IB_MGMT_RMPP_FLAG_ACTIVE) ||
1670 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1673 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1674 const struct ib_mad_recv_wc *rwc)
1676 return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1677 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1680 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1681 const struct ib_mad_send_wr_private *wr,
1682 const struct ib_mad_recv_wc *rwc )
1684 struct rdma_ah_attr attr;
1685 u8 send_resp, rcv_resp;
1687 struct ib_device *device = mad_agent_priv->agent.device;
1688 u8 port_num = mad_agent_priv->agent.port_num;
1692 send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1693 rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1695 if (send_resp == rcv_resp)
1696 /* both requests, or both responses. GIDs different */
1699 if (rdma_query_ah(wr->send_buf.ah, &attr))
1700 /* Assume not equal, to avoid false positives. */
1703 has_grh = !!(rdma_ah_get_ah_flags(&attr) & IB_AH_GRH);
1704 if (has_grh != !!(rwc->wc->wc_flags & IB_WC_GRH))
1705 /* one has GID, other does not. Assume different */
1708 if (!send_resp && rcv_resp) {
1709 /* is request/response. */
1711 if (ib_get_cached_lmc(device, port_num, &lmc))
1713 return (!lmc || !((rdma_ah_get_path_bits(&attr) ^
1714 rwc->wc->dlid_path_bits) &
1717 const struct ib_global_route *grh =
1718 rdma_ah_read_grh(&attr);
1720 if (rdma_query_gid(device, port_num,
1721 grh->sgid_index, &sgid))
1723 return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1729 return rdma_ah_get_dlid(&attr) == rwc->wc->slid;
1731 return !memcmp(rdma_ah_read_grh(&attr)->dgid.raw,
1732 rwc->recv_buf.grh->sgid.raw,
1736 static inline int is_direct(u8 class)
1738 return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1741 struct ib_mad_send_wr_private*
1742 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1743 const struct ib_mad_recv_wc *wc)
1745 struct ib_mad_send_wr_private *wr;
1746 const struct ib_mad_hdr *mad_hdr;
1748 mad_hdr = &wc->recv_buf.mad->mad_hdr;
1750 list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1751 if ((wr->tid == mad_hdr->tid) &&
1752 rcv_has_same_class(wr, wc) &&
1754 * Don't check GID for direct routed MADs.
1755 * These might have permissive LIDs.
1757 (is_direct(mad_hdr->mgmt_class) ||
1758 rcv_has_same_gid(mad_agent_priv, wr, wc)))
1759 return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1763 * It's possible to receive the response before we've
1764 * been notified that the send has completed
1766 list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1767 if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1768 wr->tid == mad_hdr->tid &&
1770 rcv_has_same_class(wr, wc) &&
1772 * Don't check GID for direct routed MADs.
1773 * These might have permissive LIDs.
1775 (is_direct(mad_hdr->mgmt_class) ||
1776 rcv_has_same_gid(mad_agent_priv, wr, wc)))
1777 /* Verify request has not been canceled */
1778 return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1783 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1785 mad_send_wr->timeout = 0;
1786 if (mad_send_wr->refcount == 1)
1787 list_move_tail(&mad_send_wr->agent_list,
1788 &mad_send_wr->mad_agent_priv->done_list);
1791 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1792 struct ib_mad_recv_wc *mad_recv_wc)
1794 struct ib_mad_send_wr_private *mad_send_wr;
1795 struct ib_mad_send_wc mad_send_wc;
1796 unsigned long flags;
1799 INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1800 ret = ib_mad_enforce_security(mad_agent_priv,
1801 mad_recv_wc->wc->pkey_index);
1803 ib_free_recv_mad(mad_recv_wc);
1804 deref_mad_agent(mad_agent_priv);
1808 list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1809 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1810 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1813 deref_mad_agent(mad_agent_priv);
1818 /* Complete corresponding request */
1819 if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1820 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1821 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1823 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1824 if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
1825 && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
1826 && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
1827 & IB_MGMT_RMPP_FLAG_ACTIVE)) {
1828 /* user rmpp is in effect
1829 * and this is an active RMPP MAD
1831 mad_agent_priv->agent.recv_handler(
1832 &mad_agent_priv->agent, NULL,
1834 deref_mad_agent(mad_agent_priv);
1836 /* not user rmpp, revert to normal behavior and
1838 ib_free_recv_mad(mad_recv_wc);
1839 deref_mad_agent(mad_agent_priv);
1843 ib_mark_mad_done(mad_send_wr);
1844 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1846 /* Defined behavior is to complete response before request */
1847 mad_agent_priv->agent.recv_handler(
1848 &mad_agent_priv->agent,
1849 &mad_send_wr->send_buf,
1851 deref_mad_agent(mad_agent_priv);
1853 mad_send_wc.status = IB_WC_SUCCESS;
1854 mad_send_wc.vendor_err = 0;
1855 mad_send_wc.send_buf = &mad_send_wr->send_buf;
1856 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1859 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
1861 deref_mad_agent(mad_agent_priv);
1867 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
1868 const struct ib_mad_qp_info *qp_info,
1869 const struct ib_wc *wc,
1871 struct ib_mad_private *recv,
1872 struct ib_mad_private *response)
1874 enum smi_forward_action retsmi;
1875 struct ib_smp *smp = (struct ib_smp *)recv->mad;
1877 trace_ib_mad_handle_ib_smi(smp);
1879 if (smi_handle_dr_smp_recv(smp,
1880 rdma_cap_ib_switch(port_priv->device),
1882 port_priv->device->phys_port_cnt) ==
1884 return IB_SMI_DISCARD;
1886 retsmi = smi_check_forward_dr_smp(smp);
1887 if (retsmi == IB_SMI_LOCAL)
1888 return IB_SMI_HANDLE;
1890 if (retsmi == IB_SMI_SEND) { /* don't forward */
1891 if (smi_handle_dr_smp_send(smp,
1892 rdma_cap_ib_switch(port_priv->device),
1893 port_num) == IB_SMI_DISCARD)
1894 return IB_SMI_DISCARD;
1896 if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
1897 return IB_SMI_DISCARD;
1898 } else if (rdma_cap_ib_switch(port_priv->device)) {
1899 /* forward case for switches */
1900 memcpy(response, recv, mad_priv_size(response));
1901 response->header.recv_wc.wc = &response->header.wc;
1902 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
1903 response->header.recv_wc.recv_buf.grh = &response->grh;
1905 agent_send_response((const struct ib_mad_hdr *)response->mad,
1908 smi_get_fwd_port(smp),
1909 qp_info->qp->qp_num,
1913 return IB_SMI_DISCARD;
1915 return IB_SMI_HANDLE;
1918 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
1919 struct ib_mad_private *response,
1920 size_t *resp_len, bool opa)
1922 const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
1923 struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
1925 if (recv_hdr->method == IB_MGMT_METHOD_GET ||
1926 recv_hdr->method == IB_MGMT_METHOD_SET) {
1927 memcpy(response, recv, mad_priv_size(response));
1928 response->header.recv_wc.wc = &response->header.wc;
1929 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
1930 response->header.recv_wc.recv_buf.grh = &response->grh;
1931 resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
1932 resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
1933 if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
1934 resp_hdr->status |= IB_SMP_DIRECTION;
1936 if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
1937 if (recv_hdr->mgmt_class ==
1938 IB_MGMT_CLASS_SUBN_LID_ROUTED ||
1939 recv_hdr->mgmt_class ==
1940 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
1941 *resp_len = opa_get_smp_header_size(
1942 (struct opa_smp *)recv->mad);
1944 *resp_len = sizeof(struct ib_mad_hdr);
1953 static enum smi_action
1954 handle_opa_smi(struct ib_mad_port_private *port_priv,
1955 struct ib_mad_qp_info *qp_info,
1958 struct ib_mad_private *recv,
1959 struct ib_mad_private *response)
1961 enum smi_forward_action retsmi;
1962 struct opa_smp *smp = (struct opa_smp *)recv->mad;
1964 trace_ib_mad_handle_opa_smi(smp);
1966 if (opa_smi_handle_dr_smp_recv(smp,
1967 rdma_cap_ib_switch(port_priv->device),
1969 port_priv->device->phys_port_cnt) ==
1971 return IB_SMI_DISCARD;
1973 retsmi = opa_smi_check_forward_dr_smp(smp);
1974 if (retsmi == IB_SMI_LOCAL)
1975 return IB_SMI_HANDLE;
1977 if (retsmi == IB_SMI_SEND) { /* don't forward */
1978 if (opa_smi_handle_dr_smp_send(smp,
1979 rdma_cap_ib_switch(port_priv->device),
1980 port_num) == IB_SMI_DISCARD)
1981 return IB_SMI_DISCARD;
1983 if (opa_smi_check_local_smp(smp, port_priv->device) ==
1985 return IB_SMI_DISCARD;
1987 } else if (rdma_cap_ib_switch(port_priv->device)) {
1988 /* forward case for switches */
1989 memcpy(response, recv, mad_priv_size(response));
1990 response->header.recv_wc.wc = &response->header.wc;
1991 response->header.recv_wc.recv_buf.opa_mad =
1992 (struct opa_mad *)response->mad;
1993 response->header.recv_wc.recv_buf.grh = &response->grh;
1995 agent_send_response((const struct ib_mad_hdr *)response->mad,
1998 opa_smi_get_fwd_port(smp),
1999 qp_info->qp->qp_num,
2000 recv->header.wc.byte_len,
2003 return IB_SMI_DISCARD;
2006 return IB_SMI_HANDLE;
2009 static enum smi_action
2010 handle_smi(struct ib_mad_port_private *port_priv,
2011 struct ib_mad_qp_info *qp_info,
2014 struct ib_mad_private *recv,
2015 struct ib_mad_private *response,
2018 struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2020 if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2021 mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2022 return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2025 return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2028 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2030 struct ib_mad_port_private *port_priv = cq->cq_context;
2031 struct ib_mad_list_head *mad_list =
2032 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2033 struct ib_mad_qp_info *qp_info;
2034 struct ib_mad_private_header *mad_priv_hdr;
2035 struct ib_mad_private *recv, *response = NULL;
2036 struct ib_mad_agent_private *mad_agent;
2038 int ret = IB_MAD_RESULT_SUCCESS;
2040 u16 resp_mad_pkey_index = 0;
2043 if (list_empty_careful(&port_priv->port_list))
2046 if (wc->status != IB_WC_SUCCESS) {
2048 * Receive errors indicate that the QP has entered the error
2049 * state - error handling/shutdown code will cleanup
2054 qp_info = mad_list->mad_queue->qp_info;
2055 dequeue_mad(mad_list);
2057 opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2058 qp_info->port_priv->port_num);
2060 mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2062 recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2063 ib_dma_unmap_single(port_priv->device,
2064 recv->header.mapping,
2065 mad_priv_dma_size(recv),
2068 /* Setup MAD receive work completion from "normal" work completion */
2069 recv->header.wc = *wc;
2070 recv->header.recv_wc.wc = &recv->header.wc;
2072 if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2073 recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2074 recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2076 recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2077 recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2080 recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2081 recv->header.recv_wc.recv_buf.grh = &recv->grh;
2084 if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2087 trace_ib_mad_recv_done_handler(qp_info, wc,
2088 (struct ib_mad_hdr *)recv->mad);
2090 mad_size = recv->mad_size;
2091 response = alloc_mad_private(mad_size, GFP_KERNEL);
2095 if (rdma_cap_ib_switch(port_priv->device))
2096 port_num = wc->port_num;
2098 port_num = port_priv->port_num;
2100 if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2101 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2102 if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2108 /* Give driver "right of first refusal" on incoming MAD */
2109 if (port_priv->device->ops.process_mad) {
2110 ret = port_priv->device->ops.process_mad(
2111 port_priv->device, 0, port_priv->port_num, wc,
2112 &recv->grh, (const struct ib_mad *)recv->mad,
2113 (struct ib_mad *)response->mad, &mad_size,
2114 &resp_mad_pkey_index);
2117 wc->pkey_index = resp_mad_pkey_index;
2119 if (ret & IB_MAD_RESULT_SUCCESS) {
2120 if (ret & IB_MAD_RESULT_CONSUMED)
2122 if (ret & IB_MAD_RESULT_REPLY) {
2123 agent_send_response((const struct ib_mad_hdr *)response->mad,
2127 qp_info->qp->qp_num,
2134 mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2136 trace_ib_mad_recv_done_agent(mad_agent);
2137 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2139 * recv is freed up in error cases in ib_mad_complete_recv
2140 * or via recv_handler in ib_mad_complete_recv()
2143 } else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2144 generate_unmatched_resp(recv, response, &mad_size, opa)) {
2145 agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2146 port_priv->device, port_num,
2147 qp_info->qp->qp_num, mad_size, opa);
2151 /* Post another receive request for this QP */
2153 ib_mad_post_receive_mads(qp_info, response);
2156 ib_mad_post_receive_mads(qp_info, recv);
2159 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2161 struct ib_mad_send_wr_private *mad_send_wr;
2162 unsigned long delay;
2164 if (list_empty(&mad_agent_priv->wait_list)) {
2165 cancel_delayed_work(&mad_agent_priv->timed_work);
2167 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2168 struct ib_mad_send_wr_private,
2171 if (time_after(mad_agent_priv->timeout,
2172 mad_send_wr->timeout)) {
2173 mad_agent_priv->timeout = mad_send_wr->timeout;
2174 delay = mad_send_wr->timeout - jiffies;
2175 if ((long)delay <= 0)
2177 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2178 &mad_agent_priv->timed_work, delay);
2183 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2185 struct ib_mad_agent_private *mad_agent_priv;
2186 struct ib_mad_send_wr_private *temp_mad_send_wr;
2187 struct list_head *list_item;
2188 unsigned long delay;
2190 mad_agent_priv = mad_send_wr->mad_agent_priv;
2191 list_del(&mad_send_wr->agent_list);
2193 delay = mad_send_wr->timeout;
2194 mad_send_wr->timeout += jiffies;
2197 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2198 temp_mad_send_wr = list_entry(list_item,
2199 struct ib_mad_send_wr_private,
2201 if (time_after(mad_send_wr->timeout,
2202 temp_mad_send_wr->timeout))
2207 list_item = &mad_agent_priv->wait_list;
2208 list_add(&mad_send_wr->agent_list, list_item);
2210 /* Reschedule a work item if we have a shorter timeout */
2211 if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2212 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2213 &mad_agent_priv->timed_work, delay);
2216 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2217 unsigned long timeout_ms)
2219 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2220 wait_for_response(mad_send_wr);
2224 * Process a send work completion
2226 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2227 struct ib_mad_send_wc *mad_send_wc)
2229 struct ib_mad_agent_private *mad_agent_priv;
2230 unsigned long flags;
2233 mad_agent_priv = mad_send_wr->mad_agent_priv;
2234 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2235 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2236 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2237 if (ret == IB_RMPP_RESULT_CONSUMED)
2240 ret = IB_RMPP_RESULT_UNHANDLED;
2242 if (mad_send_wc->status != IB_WC_SUCCESS &&
2243 mad_send_wr->status == IB_WC_SUCCESS) {
2244 mad_send_wr->status = mad_send_wc->status;
2245 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2248 if (--mad_send_wr->refcount > 0) {
2249 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2250 mad_send_wr->status == IB_WC_SUCCESS) {
2251 wait_for_response(mad_send_wr);
2256 /* Remove send from MAD agent and notify client of completion */
2257 list_del(&mad_send_wr->agent_list);
2258 adjust_timeout(mad_agent_priv);
2259 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2261 if (mad_send_wr->status != IB_WC_SUCCESS )
2262 mad_send_wc->status = mad_send_wr->status;
2263 if (ret == IB_RMPP_RESULT_INTERNAL)
2264 ib_rmpp_send_handler(mad_send_wc);
2266 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2269 /* Release reference on agent taken when sending */
2270 deref_mad_agent(mad_agent_priv);
2273 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2276 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2278 struct ib_mad_port_private *port_priv = cq->cq_context;
2279 struct ib_mad_list_head *mad_list =
2280 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2281 struct ib_mad_send_wr_private *mad_send_wr, *queued_send_wr;
2282 struct ib_mad_qp_info *qp_info;
2283 struct ib_mad_queue *send_queue;
2284 struct ib_mad_send_wc mad_send_wc;
2285 unsigned long flags;
2288 if (list_empty_careful(&port_priv->port_list))
2291 if (wc->status != IB_WC_SUCCESS) {
2292 if (!ib_mad_send_error(port_priv, wc))
2296 mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2298 send_queue = mad_list->mad_queue;
2299 qp_info = send_queue->qp_info;
2301 trace_ib_mad_send_done_agent(mad_send_wr->mad_agent_priv);
2302 trace_ib_mad_send_done_handler(mad_send_wr, wc);
2305 ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2306 mad_send_wr->header_mapping,
2307 mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2308 ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2309 mad_send_wr->payload_mapping,
2310 mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2311 queued_send_wr = NULL;
2312 spin_lock_irqsave(&send_queue->lock, flags);
2313 list_del(&mad_list->list);
2315 /* Move queued send to the send queue */
2316 if (send_queue->count-- > send_queue->max_active) {
2317 mad_list = container_of(qp_info->overflow_list.next,
2318 struct ib_mad_list_head, list);
2319 queued_send_wr = container_of(mad_list,
2320 struct ib_mad_send_wr_private,
2322 list_move_tail(&mad_list->list, &send_queue->list);
2324 spin_unlock_irqrestore(&send_queue->lock, flags);
2326 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2327 mad_send_wc.status = wc->status;
2328 mad_send_wc.vendor_err = wc->vendor_err;
2329 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2331 if (queued_send_wr) {
2332 trace_ib_mad_send_done_resend(queued_send_wr, qp_info);
2333 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2336 dev_err(&port_priv->device->dev,
2337 "ib_post_send failed: %d\n", ret);
2338 mad_send_wr = queued_send_wr;
2339 wc->status = IB_WC_LOC_QP_OP_ERR;
2345 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2347 struct ib_mad_send_wr_private *mad_send_wr;
2348 struct ib_mad_list_head *mad_list;
2349 unsigned long flags;
2351 spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2352 list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2353 mad_send_wr = container_of(mad_list,
2354 struct ib_mad_send_wr_private,
2356 mad_send_wr->retry = 1;
2358 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2361 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2364 struct ib_mad_list_head *mad_list =
2365 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2366 struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2367 struct ib_mad_send_wr_private *mad_send_wr;
2371 * Send errors will transition the QP to SQE - move
2372 * QP to RTS and repost flushed work requests
2374 mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2376 if (wc->status == IB_WC_WR_FLUSH_ERR) {
2377 if (mad_send_wr->retry) {
2379 mad_send_wr->retry = 0;
2380 trace_ib_mad_error_handler(mad_send_wr, qp_info);
2381 ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2387 struct ib_qp_attr *attr;
2389 /* Transition QP to RTS and fail offending send */
2390 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2392 attr->qp_state = IB_QPS_RTS;
2393 attr->cur_qp_state = IB_QPS_SQE;
2394 ret = ib_modify_qp(qp_info->qp, attr,
2395 IB_QP_STATE | IB_QP_CUR_STATE);
2398 dev_err(&port_priv->device->dev,
2399 "%s - ib_modify_qp to RTS: %d\n",
2402 mark_sends_for_retry(qp_info);
2409 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2411 unsigned long flags;
2412 struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2413 struct ib_mad_send_wc mad_send_wc;
2414 struct list_head cancel_list;
2416 INIT_LIST_HEAD(&cancel_list);
2418 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2419 list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2420 &mad_agent_priv->send_list, agent_list) {
2421 if (mad_send_wr->status == IB_WC_SUCCESS) {
2422 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2423 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2427 /* Empty wait list to prevent receives from finding a request */
2428 list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2429 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2431 /* Report all cancelled requests */
2432 mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2433 mad_send_wc.vendor_err = 0;
2435 list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2436 &cancel_list, agent_list) {
2437 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2438 list_del(&mad_send_wr->agent_list);
2439 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2441 deref_mad_agent(mad_agent_priv);
2445 static struct ib_mad_send_wr_private*
2446 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2447 struct ib_mad_send_buf *send_buf)
2449 struct ib_mad_send_wr_private *mad_send_wr;
2451 list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2453 if (&mad_send_wr->send_buf == send_buf)
2457 list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2459 if (is_rmpp_data_mad(mad_agent_priv,
2460 mad_send_wr->send_buf.mad) &&
2461 &mad_send_wr->send_buf == send_buf)
2467 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2468 struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2470 struct ib_mad_agent_private *mad_agent_priv;
2471 struct ib_mad_send_wr_private *mad_send_wr;
2472 unsigned long flags;
2475 mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2477 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2478 mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2479 if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2480 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2484 active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2486 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2487 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2490 mad_send_wr->send_buf.timeout_ms = timeout_ms;
2492 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2494 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2496 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2499 EXPORT_SYMBOL(ib_modify_mad);
2501 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2502 struct ib_mad_send_buf *send_buf)
2504 ib_modify_mad(mad_agent, send_buf, 0);
2506 EXPORT_SYMBOL(ib_cancel_mad);
2508 static void local_completions(struct work_struct *work)
2510 struct ib_mad_agent_private *mad_agent_priv;
2511 struct ib_mad_local_private *local;
2512 struct ib_mad_agent_private *recv_mad_agent;
2513 unsigned long flags;
2516 struct ib_mad_send_wc mad_send_wc;
2520 container_of(work, struct ib_mad_agent_private, local_work);
2522 opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2523 mad_agent_priv->qp_info->port_priv->port_num);
2525 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2526 while (!list_empty(&mad_agent_priv->local_list)) {
2527 local = list_entry(mad_agent_priv->local_list.next,
2528 struct ib_mad_local_private,
2530 list_del(&local->completion_list);
2531 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2533 if (local->mad_priv) {
2535 recv_mad_agent = local->recv_mad_agent;
2536 if (!recv_mad_agent) {
2537 dev_err(&mad_agent_priv->agent.device->dev,
2538 "No receive MAD agent for local completion\n");
2540 goto local_send_completion;
2544 * Defined behavior is to complete response
2547 build_smp_wc(recv_mad_agent->agent.qp,
2548 local->mad_send_wr->send_wr.wr.wr_cqe,
2549 be16_to_cpu(IB_LID_PERMISSIVE),
2550 local->mad_send_wr->send_wr.pkey_index,
2551 recv_mad_agent->agent.port_num, &wc);
2553 local->mad_priv->header.recv_wc.wc = &wc;
2555 base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2556 if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2557 local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2558 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2560 local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2561 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2564 INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2565 list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2566 &local->mad_priv->header.recv_wc.rmpp_list);
2567 local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2568 local->mad_priv->header.recv_wc.recv_buf.mad =
2569 (struct ib_mad *)local->mad_priv->mad;
2570 recv_mad_agent->agent.recv_handler(
2571 &recv_mad_agent->agent,
2572 &local->mad_send_wr->send_buf,
2573 &local->mad_priv->header.recv_wc);
2574 spin_lock_irqsave(&recv_mad_agent->lock, flags);
2575 deref_mad_agent(recv_mad_agent);
2576 spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2579 local_send_completion:
2581 mad_send_wc.status = IB_WC_SUCCESS;
2582 mad_send_wc.vendor_err = 0;
2583 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2584 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2587 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2588 deref_mad_agent(mad_agent_priv);
2590 kfree(local->mad_priv);
2593 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2596 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2600 if (!mad_send_wr->retries_left)
2603 mad_send_wr->retries_left--;
2604 mad_send_wr->send_buf.retries++;
2606 mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2608 if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2609 ret = ib_retry_rmpp(mad_send_wr);
2611 case IB_RMPP_RESULT_UNHANDLED:
2612 ret = ib_send_mad(mad_send_wr);
2614 case IB_RMPP_RESULT_CONSUMED:
2622 ret = ib_send_mad(mad_send_wr);
2625 mad_send_wr->refcount++;
2626 list_add_tail(&mad_send_wr->agent_list,
2627 &mad_send_wr->mad_agent_priv->send_list);
2632 static void timeout_sends(struct work_struct *work)
2634 struct ib_mad_agent_private *mad_agent_priv;
2635 struct ib_mad_send_wr_private *mad_send_wr;
2636 struct ib_mad_send_wc mad_send_wc;
2637 unsigned long flags, delay;
2639 mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2641 mad_send_wc.vendor_err = 0;
2643 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2644 while (!list_empty(&mad_agent_priv->wait_list)) {
2645 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2646 struct ib_mad_send_wr_private,
2649 if (time_after(mad_send_wr->timeout, jiffies)) {
2650 delay = mad_send_wr->timeout - jiffies;
2651 if ((long)delay <= 0)
2653 queue_delayed_work(mad_agent_priv->qp_info->
2655 &mad_agent_priv->timed_work, delay);
2659 list_del(&mad_send_wr->agent_list);
2660 if (mad_send_wr->status == IB_WC_SUCCESS &&
2661 !retry_send(mad_send_wr))
2664 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2666 if (mad_send_wr->status == IB_WC_SUCCESS)
2667 mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2669 mad_send_wc.status = mad_send_wr->status;
2670 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2671 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2674 deref_mad_agent(mad_agent_priv);
2675 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2677 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2681 * Allocate receive MADs and post receive WRs for them
2683 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2684 struct ib_mad_private *mad)
2686 unsigned long flags;
2688 struct ib_mad_private *mad_priv;
2689 struct ib_sge sg_list;
2690 struct ib_recv_wr recv_wr;
2691 struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2693 /* Initialize common scatter list fields */
2694 sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2696 /* Initialize common receive WR fields */
2697 recv_wr.next = NULL;
2698 recv_wr.sg_list = &sg_list;
2699 recv_wr.num_sge = 1;
2702 /* Allocate and map receive buffer */
2707 mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2714 sg_list.length = mad_priv_dma_size(mad_priv);
2715 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2717 mad_priv_dma_size(mad_priv),
2719 if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2725 mad_priv->header.mapping = sg_list.addr;
2726 mad_priv->header.mad_list.mad_queue = recv_queue;
2727 mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2728 recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2730 /* Post receive WR */
2731 spin_lock_irqsave(&recv_queue->lock, flags);
2732 post = (++recv_queue->count < recv_queue->max_active);
2733 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2734 spin_unlock_irqrestore(&recv_queue->lock, flags);
2735 ret = ib_post_recv(qp_info->qp, &recv_wr, NULL);
2737 spin_lock_irqsave(&recv_queue->lock, flags);
2738 list_del(&mad_priv->header.mad_list.list);
2739 recv_queue->count--;
2740 spin_unlock_irqrestore(&recv_queue->lock, flags);
2741 ib_dma_unmap_single(qp_info->port_priv->device,
2742 mad_priv->header.mapping,
2743 mad_priv_dma_size(mad_priv),
2746 dev_err(&qp_info->port_priv->device->dev,
2747 "ib_post_recv failed: %d\n", ret);
2756 * Return all the posted receive MADs
2758 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2760 struct ib_mad_private_header *mad_priv_hdr;
2761 struct ib_mad_private *recv;
2762 struct ib_mad_list_head *mad_list;
2767 while (!list_empty(&qp_info->recv_queue.list)) {
2769 mad_list = list_entry(qp_info->recv_queue.list.next,
2770 struct ib_mad_list_head, list);
2771 mad_priv_hdr = container_of(mad_list,
2772 struct ib_mad_private_header,
2774 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2777 /* Remove from posted receive MAD list */
2778 list_del(&mad_list->list);
2780 ib_dma_unmap_single(qp_info->port_priv->device,
2781 recv->header.mapping,
2782 mad_priv_dma_size(recv),
2787 qp_info->recv_queue.count = 0;
2793 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2796 struct ib_qp_attr *attr;
2800 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2804 ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2805 IB_DEFAULT_PKEY_FULL, &pkey_index);
2809 for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2810 qp = port_priv->qp_info[i].qp;
2815 * PKey index for QP1 is irrelevant but
2816 * one is needed for the Reset to Init transition
2818 attr->qp_state = IB_QPS_INIT;
2819 attr->pkey_index = pkey_index;
2820 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2821 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2822 IB_QP_PKEY_INDEX | IB_QP_QKEY);
2824 dev_err(&port_priv->device->dev,
2825 "Couldn't change QP%d state to INIT: %d\n",
2830 attr->qp_state = IB_QPS_RTR;
2831 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2833 dev_err(&port_priv->device->dev,
2834 "Couldn't change QP%d state to RTR: %d\n",
2839 attr->qp_state = IB_QPS_RTS;
2840 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2841 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2843 dev_err(&port_priv->device->dev,
2844 "Couldn't change QP%d state to RTS: %d\n",
2850 ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2852 dev_err(&port_priv->device->dev,
2853 "Failed to request completion notification: %d\n",
2858 for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2859 if (!port_priv->qp_info[i].qp)
2862 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2864 dev_err(&port_priv->device->dev,
2865 "Couldn't post receive WRs\n");
2874 static void qp_event_handler(struct ib_event *event, void *qp_context)
2876 struct ib_mad_qp_info *qp_info = qp_context;
2878 /* It's worse than that! He's dead, Jim! */
2879 dev_err(&qp_info->port_priv->device->dev,
2880 "Fatal error (%d) on MAD QP (%d)\n",
2881 event->event, qp_info->qp->qp_num);
2884 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2885 struct ib_mad_queue *mad_queue)
2887 mad_queue->qp_info = qp_info;
2888 mad_queue->count = 0;
2889 spin_lock_init(&mad_queue->lock);
2890 INIT_LIST_HEAD(&mad_queue->list);
2893 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2894 struct ib_mad_qp_info *qp_info)
2896 qp_info->port_priv = port_priv;
2897 init_mad_queue(qp_info, &qp_info->send_queue);
2898 init_mad_queue(qp_info, &qp_info->recv_queue);
2899 INIT_LIST_HEAD(&qp_info->overflow_list);
2902 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2903 enum ib_qp_type qp_type)
2905 struct ib_qp_init_attr qp_init_attr;
2908 memset(&qp_init_attr, 0, sizeof qp_init_attr);
2909 qp_init_attr.send_cq = qp_info->port_priv->cq;
2910 qp_init_attr.recv_cq = qp_info->port_priv->cq;
2911 qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2912 qp_init_attr.cap.max_send_wr = mad_sendq_size;
2913 qp_init_attr.cap.max_recv_wr = mad_recvq_size;
2914 qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2915 qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2916 qp_init_attr.qp_type = qp_type;
2917 qp_init_attr.port_num = qp_info->port_priv->port_num;
2918 qp_init_attr.qp_context = qp_info;
2919 qp_init_attr.event_handler = qp_event_handler;
2920 qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2921 if (IS_ERR(qp_info->qp)) {
2922 dev_err(&qp_info->port_priv->device->dev,
2923 "Couldn't create ib_mad QP%d\n",
2924 get_spl_qp_index(qp_type));
2925 ret = PTR_ERR(qp_info->qp);
2928 /* Use minimum queue sizes unless the CQ is resized */
2929 qp_info->send_queue.max_active = mad_sendq_size;
2930 qp_info->recv_queue.max_active = mad_recvq_size;
2937 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2942 ib_destroy_qp(qp_info->qp);
2947 * Create the QP, PD, MR, and CQ if needed
2949 static int ib_mad_port_open(struct ib_device *device,
2953 struct ib_mad_port_private *port_priv;
2954 unsigned long flags;
2955 char name[sizeof "ib_mad123"];
2958 if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
2961 if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
2962 rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
2965 /* Create new device info */
2966 port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2970 port_priv->device = device;
2971 port_priv->port_num = port_num;
2972 spin_lock_init(&port_priv->reg_lock);
2973 init_mad_qp(port_priv, &port_priv->qp_info[0]);
2974 init_mad_qp(port_priv, &port_priv->qp_info[1]);
2976 cq_size = mad_sendq_size + mad_recvq_size;
2977 has_smi = rdma_cap_ib_smi(device, port_num);
2981 port_priv->pd = ib_alloc_pd(device, 0);
2982 if (IS_ERR(port_priv->pd)) {
2983 dev_err(&device->dev, "Couldn't create ib_mad PD\n");
2984 ret = PTR_ERR(port_priv->pd);
2988 port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
2989 IB_POLL_UNBOUND_WORKQUEUE);
2990 if (IS_ERR(port_priv->cq)) {
2991 dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
2992 ret = PTR_ERR(port_priv->cq);
2997 ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3001 ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3005 snprintf(name, sizeof name, "ib_mad%d", port_num);
3006 port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3007 if (!port_priv->wq) {
3012 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3013 list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3014 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3016 ret = ib_mad_port_start(port_priv);
3018 dev_err(&device->dev, "Couldn't start port\n");
3025 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3026 list_del_init(&port_priv->port_list);
3027 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3029 destroy_workqueue(port_priv->wq);
3031 destroy_mad_qp(&port_priv->qp_info[1]);
3033 destroy_mad_qp(&port_priv->qp_info[0]);
3035 ib_free_cq(port_priv->cq);
3036 cleanup_recv_queue(&port_priv->qp_info[1]);
3037 cleanup_recv_queue(&port_priv->qp_info[0]);
3039 ib_dealloc_pd(port_priv->pd);
3048 * If there are no classes using the port, free the port
3049 * resources (CQ, MR, PD, QP) and remove the port's info structure
3051 static int ib_mad_port_close(struct ib_device *device, int port_num)
3053 struct ib_mad_port_private *port_priv;
3054 unsigned long flags;
3056 spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3057 port_priv = __ib_get_mad_port(device, port_num);
3058 if (port_priv == NULL) {
3059 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3060 dev_err(&device->dev, "Port %d not found\n", port_num);
3063 list_del_init(&port_priv->port_list);
3064 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3066 destroy_workqueue(port_priv->wq);
3067 destroy_mad_qp(&port_priv->qp_info[1]);
3068 destroy_mad_qp(&port_priv->qp_info[0]);
3069 ib_free_cq(port_priv->cq);
3070 ib_dealloc_pd(port_priv->pd);
3071 cleanup_recv_queue(&port_priv->qp_info[1]);
3072 cleanup_recv_queue(&port_priv->qp_info[0]);
3073 /* XXX: Handle deallocation of MAD registration tables */
3080 static int ib_mad_init_device(struct ib_device *device)
3083 unsigned int count = 0;
3086 start = rdma_start_port(device);
3088 for (i = start; i <= rdma_end_port(device); i++) {
3089 if (!rdma_cap_ib_mad(device, i))
3092 ret = ib_mad_port_open(device, i);
3094 dev_err(&device->dev, "Couldn't open port %d\n", i);
3097 ret = ib_agent_port_open(device, i);
3099 dev_err(&device->dev,
3100 "Couldn't open port %d for agents\n", i);
3111 if (ib_mad_port_close(device, i))
3112 dev_err(&device->dev, "Couldn't close port %d\n", i);
3115 while (--i >= start) {
3116 if (!rdma_cap_ib_mad(device, i))
3119 if (ib_agent_port_close(device, i))
3120 dev_err(&device->dev,
3121 "Couldn't close port %d for agents\n", i);
3122 if (ib_mad_port_close(device, i))
3123 dev_err(&device->dev, "Couldn't close port %d\n", i);
3128 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3132 rdma_for_each_port (device, i) {
3133 if (!rdma_cap_ib_mad(device, i))
3136 if (ib_agent_port_close(device, i))
3137 dev_err(&device->dev,
3138 "Couldn't close port %d for agents\n", i);
3139 if (ib_mad_port_close(device, i))
3140 dev_err(&device->dev, "Couldn't close port %d\n", i);
3144 static struct ib_client mad_client = {
3146 .add = ib_mad_init_device,
3147 .remove = ib_mad_remove_device
3150 int ib_mad_init(void)
3152 mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3153 mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3155 mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3156 mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3158 INIT_LIST_HEAD(&ib_mad_port_list);
3160 if (ib_register_client(&mad_client)) {
3161 pr_err("Couldn't register ib_mad client\n");
3168 void ib_mad_cleanup(void)
3170 ib_unregister_client(&mad_client);