drivers/infiniband/core/sa_query.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Voltaire, Inc.  All rights reserved.
   4  * Copyright (c) 2006 Intel Corporation.  All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/err.h>
  38 #include <linux/random.h>
  39 #include <linux/spinlock.h>
  40 #include <linux/slab.h>
  41 #include <linux/dma-mapping.h>
  42 #include <linux/kref.h>
  43 #include <linux/idr.h>
  44 #include <linux/workqueue.h>
  45 #include <uapi/linux/if_ether.h>
  46 #include <rdma/ib_pack.h>
  47 #include <rdma/ib_cache.h>
  48 #include <rdma/rdma_netlink.h>
  49 #include <net/netlink.h>
  50 #include <uapi/rdma/ib_user_sa.h>
  51 #include <rdma/ib_marshall.h>
  52 #include <rdma/ib_addr.h>
  53 #include "sa.h"
  54 #include "core_priv.h"
  55
  56 #define IB_SA_LOCAL_SVC_TIMEOUT_MIN             100
  57 #define IB_SA_LOCAL_SVC_TIMEOUT_DEFAULT         2000
  58 #define IB_SA_LOCAL_SVC_TIMEOUT_MAX             200000
  59 #define IB_SA_CPI_MAX_RETRY_CNT                 3
  60 #define IB_SA_CPI_RETRY_WAIT                    1000 /*msecs */
  61 static int sa_local_svc_timeout_ms = IB_SA_LOCAL_SVC_TIMEOUT_DEFAULT;
  62
  63 struct ib_sa_sm_ah {
  64         struct ib_ah        *ah;
  65         struct kref          ref;
  66         u16                  pkey_index;
  67         u8                   src_path_mask;
  68 };
  69
  70 enum rdma_class_port_info_type {
  71         RDMA_CLASS_PORT_INFO_IB,
  72         RDMA_CLASS_PORT_INFO_OPA
  73 };
  74
  75 struct rdma_class_port_info {
  76         enum rdma_class_port_info_type type;
  77         union {
  78                 struct ib_class_port_info ib;
  79                 struct opa_class_port_info opa;
  80         };
  81 };
  82
  83 struct ib_sa_classport_cache {
  84         bool valid;
  85         int retry_cnt;
  86         struct rdma_class_port_info data;
  87 };
  88
  89 struct ib_sa_port {
  90         struct ib_mad_agent *agent;
  91         struct ib_sa_sm_ah  *sm_ah;
  92         struct work_struct   update_task;
  93         struct ib_sa_classport_cache classport_info;
  94         struct delayed_work ib_cpi_work;
  95         spinlock_t                   classport_lock; /* protects class port info set */
  96         spinlock_t           ah_lock;
  97         u8                   port_num;
  98 };
  99
 100 struct ib_sa_device {
 101         int                     start_port, end_port;
 102         struct ib_event_handler event_handler;
 103         struct ib_sa_port port[0];
 104 };
 105
 106 struct ib_sa_query {
 107         void (*callback)(struct ib_sa_query *, int, struct ib_sa_mad *);
 108         void (*release)(struct ib_sa_query *);
 109         struct ib_sa_client    *client;
 110         struct ib_sa_port      *port;
 111         struct ib_mad_send_buf *mad_buf;
 112         struct ib_sa_sm_ah     *sm_ah;
 113         int                     id;
 114         u32                     flags;
 115         struct list_head        list; /* Local svc request list */
 116         u32                     seq; /* Local svc request sequence number */
 117         unsigned long           timeout; /* Local svc timeout */
 118         u8                      path_use; /* How will the pathrecord be used */
 119 };
 120
 121 #define IB_SA_ENABLE_LOCAL_SERVICE      0x00000001
 122 #define IB_SA_CANCEL                    0x00000002
 123 #define IB_SA_QUERY_OPA                 0x00000004
 124
 125 struct ib_sa_service_query {
 126         void (*callback)(int, struct ib_sa_service_rec *, void *);
 127         void *context;
 128         struct ib_sa_query sa_query;
 129 };
 130
 131 struct ib_sa_path_query {
 132         void (*callback)(int, struct ib_sa_path_rec *, void *);
 133         void *context;
 134         struct ib_sa_query sa_query;
 135 };
 136
 137 struct ib_sa_guidinfo_query {
 138         void (*callback)(int, struct ib_sa_guidinfo_rec *, void *);
 139         void *context;
 140         struct ib_sa_query sa_query;
 141 };
 142
 143 struct ib_sa_classport_info_query {
 144         void (*callback)(void *);
 145         void *context;
 146         struct ib_sa_query sa_query;
 147 };
 148
 149 struct ib_sa_mcmember_query {
 150         void (*callback)(int, struct ib_sa_mcmember_rec *, void *);
 151         void *context;
 152         struct ib_sa_query sa_query;
 153 };
 154
 155 static LIST_HEAD(ib_nl_request_list);
 156 static DEFINE_SPINLOCK(ib_nl_request_lock);
 157 static atomic_t ib_nl_sa_request_seq;
 158 static struct workqueue_struct *ib_nl_wq;
 159 static struct delayed_work ib_nl_timed_work;
 160 static const struct nla_policy ib_nl_policy[LS_NLA_TYPE_MAX] = {
 161         [LS_NLA_TYPE_PATH_RECORD]       = {.type = NLA_BINARY,
 162                 .len = sizeof(struct ib_path_rec_data)},
 163         [LS_NLA_TYPE_TIMEOUT]           = {.type = NLA_U32},
 164         [LS_NLA_TYPE_SERVICE_ID]        = {.type = NLA_U64},
 165         [LS_NLA_TYPE_DGID]              = {.type = NLA_BINARY,
 166                 .len = sizeof(struct rdma_nla_ls_gid)},
 167         [LS_NLA_TYPE_SGID]              = {.type = NLA_BINARY,
 168                 .len = sizeof(struct rdma_nla_ls_gid)},
 169         [LS_NLA_TYPE_TCLASS]            = {.type = NLA_U8},
 170         [LS_NLA_TYPE_PKEY]              = {.type = NLA_U16},
 171         [LS_NLA_TYPE_QOS_CLASS]         = {.type = NLA_U16},
 172 };
 173
 174
 175 static void ib_sa_add_one(struct ib_device *device);
 176 static void ib_sa_remove_one(struct ib_device *device, void *client_data);
 177
 178 static struct ib_client sa_client = {
 179         .name   = "sa",
 180         .add    = ib_sa_add_one,
 181         .remove = ib_sa_remove_one
 182 };
 183
 184 static DEFINE_SPINLOCK(idr_lock);
 185 static DEFINE_IDR(query_idr);
 186
 187 static DEFINE_SPINLOCK(tid_lock);
 188 static u32 tid;
 189
 190 #define PATH_REC_FIELD(field) \
 191         .struct_offset_bytes = offsetof(struct ib_sa_path_rec, field),          \
 192         .struct_size_bytes   = sizeof ((struct ib_sa_path_rec *) 0)->field,     \
 193         .field_name          = "sa_path_rec:" #field
 194
 195 static const struct ib_field path_rec_table[] = {
 196         { PATH_REC_FIELD(service_id),
 197           .offset_words = 0,
 198           .offset_bits  = 0,
 199           .size_bits    = 64 },
 200         { PATH_REC_FIELD(dgid),
 201           .offset_words = 2,
 202           .offset_bits  = 0,
 203           .size_bits    = 128 },
 204         { PATH_REC_FIELD(sgid),
 205           .offset_words = 6,
 206           .offset_bits  = 0,
 207           .size_bits    = 128 },
 208         { PATH_REC_FIELD(dlid),
 209           .offset_words = 10,
 210           .offset_bits  = 0,
 211           .size_bits    = 16 },
 212         { PATH_REC_FIELD(slid),
 213           .offset_words = 10,
 214           .offset_bits  = 16,
 215           .size_bits    = 16 },
 216         { PATH_REC_FIELD(raw_traffic),
 217           .offset_words = 11,
 218           .offset_bits  = 0,
 219           .size_bits    = 1 },
 220         { RESERVED,
 221           .offset_words = 11,
 222           .offset_bits  = 1,
 223           .size_bits    = 3 },
 224         { PATH_REC_FIELD(flow_label),
 225           .offset_words = 11,
 226           .offset_bits  = 4,
 227           .size_bits    = 20 },
 228         { PATH_REC_FIELD(hop_limit),
 229           .offset_words = 11,
 230           .offset_bits  = 24,
 231           .size_bits    = 8 },
 232         { PATH_REC_FIELD(traffic_class),
 233           .offset_words = 12,
 234           .offset_bits  = 0,
 235           .size_bits    = 8 },
 236         { PATH_REC_FIELD(reversible),
 237           .offset_words = 12,
 238           .offset_bits  = 8,
 239           .size_bits    = 1 },
 240         { PATH_REC_FIELD(numb_path),
 241           .offset_words = 12,
 242           .offset_bits  = 9,
 243           .size_bits    = 7 },
 244         { PATH_REC_FIELD(pkey),
 245           .offset_words = 12,
 246           .offset_bits  = 16,
 247           .size_bits    = 16 },
 248         { PATH_REC_FIELD(qos_class),
 249           .offset_words = 13,
 250           .offset_bits  = 0,
 251           .size_bits    = 12 },
 252         { PATH_REC_FIELD(sl),
 253           .offset_words = 13,
 254           .offset_bits  = 12,
 255           .size_bits    = 4 },
 256         { PATH_REC_FIELD(mtu_selector),
 257           .offset_words = 13,
 258           .offset_bits  = 16,
 259           .size_bits    = 2 },
 260         { PATH_REC_FIELD(mtu),
 261           .offset_words = 13,
 262           .offset_bits  = 18,
 263           .size_bits    = 6 },
 264         { PATH_REC_FIELD(rate_selector),
 265           .offset_words = 13,
 266           .offset_bits  = 24,
 267           .size_bits    = 2 },
 268         { PATH_REC_FIELD(rate),
 269           .offset_words = 13,
 270           .offset_bits  = 26,
 271           .size_bits    = 6 },
 272         { PATH_REC_FIELD(packet_life_time_selector),
 273           .offset_words = 14,
 274           .offset_bits  = 0,
 275           .size_bits    = 2 },
 276         { PATH_REC_FIELD(packet_life_time),
 277           .offset_words = 14,
 278           .offset_bits  = 2,
 279           .size_bits    = 6 },
 280         { PATH_REC_FIELD(preference),
 281           .offset_words = 14,
 282           .offset_bits  = 8,
 283           .size_bits    = 8 },
 284         { RESERVED,
 285           .offset_words = 14,
 286           .offset_bits  = 16,
 287           .size_bits    = 48 },
 288 };
 289
 290 #define MCMEMBER_REC_FIELD(field) \
 291         .struct_offset_bytes = offsetof(struct ib_sa_mcmember_rec, field),      \
 292         .struct_size_bytes   = sizeof ((struct ib_sa_mcmember_rec *) 0)->field, \
 293         .field_name          = "sa_mcmember_rec:" #field
 294
 295 static const struct ib_field mcmember_rec_table[] = {
 296         { MCMEMBER_REC_FIELD(mgid),
 297           .offset_words = 0,
 298           .offset_bits  = 0,
 299           .size_bits    = 128 },
 300         { MCMEMBER_REC_FIELD(port_gid),
 301           .offset_words = 4,
 302           .offset_bits  = 0,
 303           .size_bits    = 128 },
 304         { MCMEMBER_REC_FIELD(qkey),
 305           .offset_words = 8,
 306           .offset_bits  = 0,
 307           .size_bits    = 32 },
 308         { MCMEMBER_REC_FIELD(mlid),
 309           .offset_words = 9,
 310           .offset_bits  = 0,
 311           .size_bits    = 16 },
 312         { MCMEMBER_REC_FIELD(mtu_selector),
 313           .offset_words = 9,
 314           .offset_bits  = 16,
 315           .size_bits    = 2 },
 316         { MCMEMBER_REC_FIELD(mtu),
 317           .offset_words = 9,
 318           .offset_bits  = 18,
 319           .size_bits    = 6 },
 320         { MCMEMBER_REC_FIELD(traffic_class),
 321           .offset_words = 9,
 322           .offset_bits  = 24,
 323           .size_bits    = 8 },
 324         { MCMEMBER_REC_FIELD(pkey),
 325           .offset_words = 10,
 326           .offset_bits  = 0,
 327           .size_bits    = 16 },
 328         { MCMEMBER_REC_FIELD(rate_selector),
 329           .offset_words = 10,
 330           .offset_bits  = 16,
 331           .size_bits    = 2 },
 332         { MCMEMBER_REC_FIELD(rate),
 333           .offset_words = 10,
 334           .offset_bits  = 18,
 335           .size_bits    = 6 },
 336         { MCMEMBER_REC_FIELD(packet_life_time_selector),
 337           .offset_words = 10,
 338           .offset_bits  = 24,
 339           .size_bits    = 2 },
 340         { MCMEMBER_REC_FIELD(packet_life_time),
 341           .offset_words = 10,
 342           .offset_bits  = 26,
 343           .size_bits    = 6 },
 344         { MCMEMBER_REC_FIELD(sl),
 345           .offset_words = 11,
 346           .offset_bits  = 0,
 347           .size_bits    = 4 },
 348         { MCMEMBER_REC_FIELD(flow_label),
 349           .offset_words = 11,
 350           .offset_bits  = 4,
 351           .size_bits    = 20 },
 352         { MCMEMBER_REC_FIELD(hop_limit),
 353           .offset_words = 11,
 354           .offset_bits  = 24,
 355           .size_bits    = 8 },
 356         { MCMEMBER_REC_FIELD(scope),
 357           .offset_words = 12,
 358           .offset_bits  = 0,
 359           .size_bits    = 4 },
 360         { MCMEMBER_REC_FIELD(join_state),
 361           .offset_words = 12,
 362           .offset_bits  = 4,
 363           .size_bits    = 4 },
 364         { MCMEMBER_REC_FIELD(proxy_join),
 365           .offset_words = 12,
 366           .offset_bits  = 8,
 367           .size_bits    = 1 },
 368         { RESERVED,
 369           .offset_words = 12,
 370           .offset_bits  = 9,
 371           .size_bits    = 23 },
 372 };
 373
 374 #define SERVICE_REC_FIELD(field) \
 375         .struct_offset_bytes = offsetof(struct ib_sa_service_rec, field),       \
 376         .struct_size_bytes   = sizeof ((struct ib_sa_service_rec *) 0)->field,  \
 377         .field_name          = "sa_service_rec:" #field
 378
 379 static const struct ib_field service_rec_table[] = {
 380         { SERVICE_REC_FIELD(id),
 381           .offset_words = 0,
 382           .offset_bits  = 0,
 383           .size_bits    = 64 },
 384         { SERVICE_REC_FIELD(gid),
 385           .offset_words = 2,
 386           .offset_bits  = 0,
 387           .size_bits    = 128 },
 388         { SERVICE_REC_FIELD(pkey),
 389           .offset_words = 6,
 390           .offset_bits  = 0,
 391           .size_bits    = 16 },
 392         { SERVICE_REC_FIELD(lease),
 393           .offset_words = 7,
 394           .offset_bits  = 0,
 395           .size_bits    = 32 },
 396         { SERVICE_REC_FIELD(key),
 397           .offset_words = 8,
 398           .offset_bits  = 0,
 399           .size_bits    = 128 },
 400         { SERVICE_REC_FIELD(name),
 401           .offset_words = 12,
 402           .offset_bits  = 0,
 403           .size_bits    = 64*8 },
 404         { SERVICE_REC_FIELD(data8),
 405           .offset_words = 28,
 406           .offset_bits  = 0,
 407           .size_bits    = 16*8 },
 408         { SERVICE_REC_FIELD(data16),
 409           .offset_words = 32,
 410           .offset_bits  = 0,
 411           .size_bits    = 8*16 },
 412         { SERVICE_REC_FIELD(data32),
 413           .offset_words = 36,
 414           .offset_bits  = 0,
 415           .size_bits    = 4*32 },
 416         { SERVICE_REC_FIELD(data64),
 417           .offset_words = 40,
 418           .offset_bits  = 0,
 419           .size_bits    = 2*64 },
 420 };
 421
 422 #define CLASSPORTINFO_REC_FIELD(field) \
 423         .struct_offset_bytes = offsetof(struct ib_class_port_info, field),      \
 424         .struct_size_bytes   = sizeof((struct ib_class_port_info *)0)->field,   \
 425         .field_name          = "ib_class_port_info:" #field
 426
 427 static const struct ib_field ib_classport_info_rec_table[] = {
 428         { CLASSPORTINFO_REC_FIELD(base_version),
 429           .offset_words = 0,
 430           .offset_bits  = 0,
 431           .size_bits    = 8 },
 432         { CLASSPORTINFO_REC_FIELD(class_version),
 433           .offset_words = 0,
 434           .offset_bits  = 8,
 435           .size_bits    = 8 },
 436         { CLASSPORTINFO_REC_FIELD(capability_mask),
 437           .offset_words = 0,
 438           .offset_bits  = 16,
 439           .size_bits    = 16 },
 440         { CLASSPORTINFO_REC_FIELD(cap_mask2_resp_time),
 441           .offset_words = 1,
 442           .offset_bits  = 0,
 443           .size_bits    = 32 },
 444         { CLASSPORTINFO_REC_FIELD(redirect_gid),
 445           .offset_words = 2,
 446           .offset_bits  = 0,
 447           .size_bits    = 128 },
 448         { CLASSPORTINFO_REC_FIELD(redirect_tcslfl),
 449           .offset_words = 6,
 450           .offset_bits  = 0,
 451           .size_bits    = 32 },
 452         { CLASSPORTINFO_REC_FIELD(redirect_lid),
 453           .offset_words = 7,
 454           .offset_bits  = 0,
 455           .size_bits    = 16 },
 456         { CLASSPORTINFO_REC_FIELD(redirect_pkey),
 457           .offset_words = 7,
 458           .offset_bits  = 16,
 459           .size_bits    = 16 },
 460
 461         { CLASSPORTINFO_REC_FIELD(redirect_qp),
 462           .offset_words = 8,
 463           .offset_bits  = 0,
 464           .size_bits    = 32 },
 465         { CLASSPORTINFO_REC_FIELD(redirect_qkey),
 466           .offset_words = 9,
 467           .offset_bits  = 0,
 468           .size_bits    = 32 },
 469
 470         { CLASSPORTINFO_REC_FIELD(trap_gid),
 471           .offset_words = 10,
 472           .offset_bits  = 0,
 473           .size_bits    = 128 },
 474         { CLASSPORTINFO_REC_FIELD(trap_tcslfl),
 475           .offset_words = 14,
 476           .offset_bits  = 0,
 477           .size_bits    = 32 },
 478
 479         { CLASSPORTINFO_REC_FIELD(trap_lid),
 480           .offset_words = 15,
 481           .offset_bits  = 0,
 482           .size_bits    = 16 },
 483         { CLASSPORTINFO_REC_FIELD(trap_pkey),
 484           .offset_words = 15,
 485           .offset_bits  = 16,
 486           .size_bits    = 16 },
 487
 488         { CLASSPORTINFO_REC_FIELD(trap_hlqp),
 489           .offset_words = 16,
 490           .offset_bits  = 0,
 491           .size_bits    = 32 },
 492         { CLASSPORTINFO_REC_FIELD(trap_qkey),
 493           .offset_words = 17,
 494           .offset_bits  = 0,
 495           .size_bits    = 32 },
 496 };
 497
 498 #define OPA_CLASSPORTINFO_REC_FIELD(field) \
 499         .struct_offset_bytes =\
 500                 offsetof(struct opa_class_port_info, field),    \
 501         .struct_size_bytes   = \
 502                 sizeof((struct opa_class_port_info *)0)->field, \
 503         .field_name          = "opa_class_port_info:" #field
 504
 505 static const struct ib_field opa_classport_info_rec_table[] = {
 506         { OPA_CLASSPORTINFO_REC_FIELD(base_version),
 507           .offset_words = 0,
 508           .offset_bits  = 0,
 509           .size_bits    = 8 },
 510         { OPA_CLASSPORTINFO_REC_FIELD(class_version),
 511           .offset_words = 0,
 512           .offset_bits  = 8,
 513           .size_bits    = 8 },
 514         { OPA_CLASSPORTINFO_REC_FIELD(cap_mask),
 515           .offset_words = 0,
 516           .offset_bits  = 16,
 517           .size_bits    = 16 },
 518         { OPA_CLASSPORTINFO_REC_FIELD(cap_mask2_resp_time),
 519           .offset_words = 1,
 520           .offset_bits  = 0,
 521           .size_bits    = 32 },
 522         { OPA_CLASSPORTINFO_REC_FIELD(redirect_gid),
 523           .offset_words = 2,
 524           .offset_bits  = 0,
 525           .size_bits    = 128 },
 526         { OPA_CLASSPORTINFO_REC_FIELD(redirect_tc_fl),
 527           .offset_words = 6,
 528           .offset_bits  = 0,
 529           .size_bits    = 32 },
 530         { OPA_CLASSPORTINFO_REC_FIELD(redirect_lid),
 531           .offset_words = 7,
 532           .offset_bits  = 0,
 533           .size_bits    = 32 },
 534         { OPA_CLASSPORTINFO_REC_FIELD(redirect_sl_qp),
 535           .offset_words = 8,
 536           .offset_bits  = 0,
 537           .size_bits    = 32 },
 538         { OPA_CLASSPORTINFO_REC_FIELD(redirect_qkey),
 539           .offset_words = 9,
 540           .offset_bits  = 0,
 541           .size_bits    = 32 },
 542         { OPA_CLASSPORTINFO_REC_FIELD(trap_gid),
 543           .offset_words = 10,
 544           .offset_bits  = 0,
 545           .size_bits    = 128 },
 546         { OPA_CLASSPORTINFO_REC_FIELD(trap_tc_fl),
 547           .offset_words = 14,
 548           .offset_bits  = 0,
 549           .size_bits    = 32 },
 550         { OPA_CLASSPORTINFO_REC_FIELD(trap_lid),
 551           .offset_words = 15,
 552           .offset_bits  = 0,
 553           .size_bits    = 32 },
 554         { OPA_CLASSPORTINFO_REC_FIELD(trap_hl_qp),
 555           .offset_words = 16,
 556           .offset_bits  = 0,
 557           .size_bits    = 32 },
 558         { OPA_CLASSPORTINFO_REC_FIELD(trap_qkey),
 559           .offset_words = 17,
 560           .offset_bits  = 0,
 561           .size_bits    = 32 },
 562         { OPA_CLASSPORTINFO_REC_FIELD(trap_pkey),
 563           .offset_words = 18,
 564           .offset_bits  = 0,
 565           .size_bits    = 16 },
 566         { OPA_CLASSPORTINFO_REC_FIELD(redirect_pkey),
 567           .offset_words = 18,
 568           .offset_bits  = 16,
 569           .size_bits    = 16 },
 570         { OPA_CLASSPORTINFO_REC_FIELD(trap_sl_rsvd),
 571           .offset_words = 19,
 572           .offset_bits  = 0,
 573           .size_bits    = 8 },
 574         { RESERVED,
 575           .offset_words = 19,
 576           .offset_bits  = 8,
 577           .size_bits    = 24 },
 578 };
 579
 580 #define GUIDINFO_REC_FIELD(field) \
 581         .struct_offset_bytes = offsetof(struct ib_sa_guidinfo_rec, field),      \
 582         .struct_size_bytes   = sizeof((struct ib_sa_guidinfo_rec *) 0)->field,  \
 583         .field_name          = "sa_guidinfo_rec:" #field
 584
 585 static const struct ib_field guidinfo_rec_table[] = {
 586         { GUIDINFO_REC_FIELD(lid),
 587           .offset_words = 0,
 588           .offset_bits  = 0,
 589           .size_bits    = 16 },
 590         { GUIDINFO_REC_FIELD(block_num),
 591           .offset_words = 0,
 592           .offset_bits  = 16,
 593           .size_bits    = 8 },
 594         { GUIDINFO_REC_FIELD(res1),
 595           .offset_words = 0,
 596           .offset_bits  = 24,
 597           .size_bits    = 8 },
 598         { GUIDINFO_REC_FIELD(res2),
 599           .offset_words = 1,
 600           .offset_bits  = 0,
 601           .size_bits    = 32 },
 602         { GUIDINFO_REC_FIELD(guid_info_list),
 603           .offset_words = 2,
 604           .offset_bits  = 0,
 605           .size_bits    = 512 },
 606 };
 607
 608 static inline void ib_sa_disable_local_svc(struct ib_sa_query *query)
 609 {
 610         query->flags &= ~IB_SA_ENABLE_LOCAL_SERVICE;
 611 }
 612
 613 static inline int ib_sa_query_cancelled(struct ib_sa_query *query)
 614 {
 615         return (query->flags & IB_SA_CANCEL);
 616 }
 617
 618 static void ib_nl_set_path_rec_attrs(struct sk_buff *skb,
 619                                      struct ib_sa_query *query)
 620 {
 621         struct ib_sa_path_rec *sa_rec = query->mad_buf->context[1];
 622         struct ib_sa_mad *mad = query->mad_buf->mad;
 623         ib_sa_comp_mask comp_mask = mad->sa_hdr.comp_mask;
 624         u16 val16;
 625         u64 val64;
 626         struct rdma_ls_resolve_header *header;
 627
 628         query->mad_buf->context[1] = NULL;
 629
 630         /* Construct the family header first */
 631         header = (struct rdma_ls_resolve_header *)
 632                 skb_put(skb, NLMSG_ALIGN(sizeof(*header)));
 633         memcpy(header->device_name, query->port->agent->device->name,
 634                LS_DEVICE_NAME_MAX);
 635         header->port_num = query->port->port_num;
 636
 637         if ((comp_mask & IB_SA_PATH_REC_REVERSIBLE) &&
 638             sa_rec->reversible != 0)
 639                 query->path_use = LS_RESOLVE_PATH_USE_GMP;
 640         else
 641                 query->path_use = LS_RESOLVE_PATH_USE_UNIDIRECTIONAL;
 642         header->path_use = query->path_use;
 643
 644         /* Now build the attributes */
 645         if (comp_mask & IB_SA_PATH_REC_SERVICE_ID) {
 646                 val64 = be64_to_cpu(sa_rec->service_id);
 647                 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_SERVICE_ID,
 648                         sizeof(val64), &val64);
 649         }
 650         if (comp_mask & IB_SA_PATH_REC_DGID)
 651                 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_DGID,
 652                         sizeof(sa_rec->dgid), &sa_rec->dgid);
 653         if (comp_mask & IB_SA_PATH_REC_SGID)
 654                 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_SGID,
 655                         sizeof(sa_rec->sgid), &sa_rec->sgid);
 656         if (comp_mask & IB_SA_PATH_REC_TRAFFIC_CLASS)
 657                 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_TCLASS,
 658                         sizeof(sa_rec->traffic_class), &sa_rec->traffic_class);
 659
 660         if (comp_mask & IB_SA_PATH_REC_PKEY) {
 661                 val16 = be16_to_cpu(sa_rec->pkey);
 662                 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_PKEY,
 663                         sizeof(val16), &val16);
 664         }
 665         if (comp_mask & IB_SA_PATH_REC_QOS_CLASS) {
 666                 val16 = be16_to_cpu(sa_rec->qos_class);
 667                 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_QOS_CLASS,
 668                         sizeof(val16), &val16);
 669         }
 670 }
 671
 672 static int ib_nl_get_path_rec_attrs_len(ib_sa_comp_mask comp_mask)
 673 {
 674         int len = 0;
 675
 676         if (comp_mask & IB_SA_PATH_REC_SERVICE_ID)
 677                 len += nla_total_size(sizeof(u64));
 678         if (comp_mask & IB_SA_PATH_REC_DGID)
 679                 len += nla_total_size(sizeof(struct rdma_nla_ls_gid));
 680         if (comp_mask & IB_SA_PATH_REC_SGID)
 681                 len += nla_total_size(sizeof(struct rdma_nla_ls_gid));
 682         if (comp_mask & IB_SA_PATH_REC_TRAFFIC_CLASS)
 683                 len += nla_total_size(sizeof(u8));
 684         if (comp_mask & IB_SA_PATH_REC_PKEY)
 685                 len += nla_total_size(sizeof(u16));
 686         if (comp_mask & IB_SA_PATH_REC_QOS_CLASS)
 687                 len += nla_total_size(sizeof(u16));
 688
 689         /*
 690          * Make sure that at least some of the required comp_mask bits are
 691          * set.
 692          */
 693         if (WARN_ON(len == 0))
 694                 return len;
 695
 696         /* Add the family header */
 697         len += NLMSG_ALIGN(sizeof(struct rdma_ls_resolve_header));
 698
 699         return len;
 700 }
 701
 702 static int ib_nl_send_msg(struct ib_sa_query *query, gfp_t gfp_mask)
 703 {
 704         struct sk_buff *skb = NULL;
 705         struct nlmsghdr *nlh;
 706         void *data;
 707         int ret = 0;
 708         struct ib_sa_mad *mad;
 709         int len;
 710
 711         mad = query->mad_buf->mad;
 712         len = ib_nl_get_path_rec_attrs_len(mad->sa_hdr.comp_mask);
 713         if (len <= 0)
 714                 return -EMSGSIZE;
 715
 716         skb = nlmsg_new(len, gfp_mask);
 717         if (!skb)
 718                 return -ENOMEM;
 719
 720         /* Put nlmsg header only for now */
 721         data = ibnl_put_msg(skb, &nlh, query->seq, 0, RDMA_NL_LS,
 722                             RDMA_NL_LS_OP_RESOLVE, NLM_F_REQUEST);
 723         if (!data) {
 724                 nlmsg_free(skb);
 725                 return -EMSGSIZE;
 726         }
 727
 728         /* Add attributes */
 729         ib_nl_set_path_rec_attrs(skb, query);
 730
 731         /* Repair the nlmsg header length */
 732         nlmsg_end(skb, nlh);
 733
 734         ret = ibnl_multicast(skb, nlh, RDMA_NL_GROUP_LS, gfp_mask);
 735         if (!ret)
 736                 ret = len;
 737         else
 738                 ret = 0;
 739
 740         return ret;
 741 }
 742
 743 static int ib_nl_make_request(struct ib_sa_query *query, gfp_t gfp_mask)
 744 {
 745         unsigned long flags;
 746         unsigned long delay;
 747         int ret;
 748
 749         INIT_LIST_HEAD(&query->list);
 750         query->seq = (u32)atomic_inc_return(&ib_nl_sa_request_seq);
 751
 752         /* Put the request on the list first.*/
 753         spin_lock_irqsave(&ib_nl_request_lock, flags);
 754         delay = msecs_to_jiffies(sa_local_svc_timeout_ms);
 755         query->timeout = delay + jiffies;
 756         list_add_tail(&query->list, &ib_nl_request_list);
 757         /* Start the timeout if this is the only request */
 758         if (ib_nl_request_list.next == &query->list)
 759                 queue_delayed_work(ib_nl_wq, &ib_nl_timed_work, delay);
 760         spin_unlock_irqrestore(&ib_nl_request_lock, flags);
 761
 762         ret = ib_nl_send_msg(query, gfp_mask);
 763         if (ret <= 0) {
 764                 ret = -EIO;
 765                 /* Remove the request */
 766                 spin_lock_irqsave(&ib_nl_request_lock, flags);
 767                 list_del(&query->list);
 768                 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
 769         } else {
 770                 ret = 0;
 771         }
 772
 773         return ret;
 774 }
 775
 776 static int ib_nl_cancel_request(struct ib_sa_query *query)
 777 {
 778         unsigned long flags;
 779         struct ib_sa_query *wait_query;
 780         int found = 0;
 781
 782         spin_lock_irqsave(&ib_nl_request_lock, flags);
 783         list_for_each_entry(wait_query, &ib_nl_request_list, list) {
 784                 /* Let the timeout to take care of the callback */
 785                 if (query == wait_query) {
 786                         query->flags |= IB_SA_CANCEL;
 787                         query->timeout = jiffies;
 788                         list_move(&query->list, &ib_nl_request_list);
 789                         found = 1;
 790                         mod_delayed_work(ib_nl_wq, &ib_nl_timed_work, 1);
 791                         break;
 792                 }
 793         }
 794         spin_unlock_irqrestore(&ib_nl_request_lock, flags);
 795
 796         return found;
 797 }
 798
 799 static void send_handler(struct ib_mad_agent *agent,
 800                          struct ib_mad_send_wc *mad_send_wc);
 801
 802 static void ib_nl_process_good_resolve_rsp(struct ib_sa_query *query,
 803                                            const struct nlmsghdr *nlh)
 804 {
 805         struct ib_mad_send_wc mad_send_wc;
 806         struct ib_sa_mad *mad = NULL;
 807         const struct nlattr *head, *curr;
 808         struct ib_path_rec_data  *rec;
 809         int len, rem;
 810         u32 mask = 0;
 811         int status = -EIO;
 812
 813         if (query->callback) {
 814                 head = (const struct nlattr *) nlmsg_data(nlh);
 815                 len = nlmsg_len(nlh);
 816                 switch (query->path_use) {
 817                 case LS_RESOLVE_PATH_USE_UNIDIRECTIONAL:
 818                         mask = IB_PATH_PRIMARY | IB_PATH_OUTBOUND;
 819                         break;
 820
 821                 case LS_RESOLVE_PATH_USE_ALL:
 822                 case LS_RESOLVE_PATH_USE_GMP:
 823                 default:
 824                         mask = IB_PATH_PRIMARY | IB_PATH_GMP |
 825                                 IB_PATH_BIDIRECTIONAL;
 826                         break;
 827                 }
 828                 nla_for_each_attr(curr, head, len, rem) {
 829                         if (curr->nla_type == LS_NLA_TYPE_PATH_RECORD) {
 830                                 rec = nla_data(curr);
 831                                 /*
 832                                  * Get the first one. In the future, we may
 833                                  * need to get up to 6 pathrecords.
 834                                  */
 835                                 if ((rec->flags & mask) == mask) {
 836                                         mad = query->mad_buf->mad;
 837                                         mad->mad_hdr.method |=
 838                                                 IB_MGMT_METHOD_RESP;
 839                                         memcpy(mad->data, rec->path_rec,
 840                                                sizeof(rec->path_rec));
 841                                         status = 0;
 842                                         break;
 843                                 }
 844                         }
 845                 }
 846                 query->callback(query, status, mad);
 847         }
 848
 849         mad_send_wc.send_buf = query->mad_buf;
 850         mad_send_wc.status = IB_WC_SUCCESS;
 851         send_handler(query->mad_buf->mad_agent, &mad_send_wc);
 852 }
 853
 854 static void ib_nl_request_timeout(struct work_struct *work)
 855 {
 856         unsigned long flags;
 857         struct ib_sa_query *query;
 858         unsigned long delay;
 859         struct ib_mad_send_wc mad_send_wc;
 860         int ret;
 861
 862         spin_lock_irqsave(&ib_nl_request_lock, flags);
 863         while (!list_empty(&ib_nl_request_list)) {
 864                 query = list_entry(ib_nl_request_list.next,
 865                                    struct ib_sa_query, list);
 866
 867                 if (time_after(query->timeout, jiffies)) {
 868                         delay = query->timeout - jiffies;
 869                         if ((long)delay <= 0)
 870                                 delay = 1;
 871                         queue_delayed_work(ib_nl_wq, &ib_nl_timed_work, delay);
 872                         break;
 873                 }
 874
 875                 list_del(&query->list);
 876                 ib_sa_disable_local_svc(query);
 877                 /* Hold the lock to protect against query cancellation */
 878                 if (ib_sa_query_cancelled(query))
 879                         ret = -1;
 880                 else
 881                         ret = ib_post_send_mad(query->mad_buf, NULL);
 882                 if (ret) {
 883                         mad_send_wc.send_buf = query->mad_buf;
 884                         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
 885                         spin_unlock_irqrestore(&ib_nl_request_lock, flags);
 886                         send_handler(query->port->agent, &mad_send_wc);
 887                         spin_lock_irqsave(&ib_nl_request_lock, flags);
 888                 }
 889         }
 890         spin_unlock_irqrestore(&ib_nl_request_lock, flags);
 891 }
 892
 893 int ib_nl_handle_set_timeout(struct sk_buff *skb,
 894                              struct netlink_callback *cb)
 895 {
 896         const struct nlmsghdr *nlh = (struct nlmsghdr *)cb->nlh;
 897         int timeout, delta, abs_delta;
 898         const struct nlattr *attr;
 899         unsigned long flags;
 900         struct ib_sa_query *query;
 901         long delay = 0;
 902         struct nlattr *tb[LS_NLA_TYPE_MAX];
 903         int ret;
 904
 905         if (!(nlh->nlmsg_flags & NLM_F_REQUEST) ||
 906             !(NETLINK_CB(skb).sk) ||
 907             !netlink_capable(skb, CAP_NET_ADMIN))
 908                 return -EPERM;
 909
 910         ret = nla_parse(tb, LS_NLA_TYPE_MAX - 1, nlmsg_data(nlh),
 911                         nlmsg_len(nlh), ib_nl_policy, NULL);
 912         attr = (const struct nlattr *)tb[LS_NLA_TYPE_TIMEOUT];
 913         if (ret || !attr)
 914                 goto settimeout_out;
 915
 916         timeout = *(int *) nla_data(attr);
 917         if (timeout < IB_SA_LOCAL_SVC_TIMEOUT_MIN)
 918                 timeout = IB_SA_LOCAL_SVC_TIMEOUT_MIN;
 919         if (timeout > IB_SA_LOCAL_SVC_TIMEOUT_MAX)
 920                 timeout = IB_SA_LOCAL_SVC_TIMEOUT_MAX;
 921
 922         delta = timeout - sa_local_svc_timeout_ms;
 923         if (delta < 0)
 924                 abs_delta = -delta;
 925         else
 926                 abs_delta = delta;
 927
 928         if (delta != 0) {
 929                 spin_lock_irqsave(&ib_nl_request_lock, flags);
 930                 sa_local_svc_timeout_ms = timeout;
 931                 list_for_each_entry(query, &ib_nl_request_list, list) {
 932                         if (delta < 0 && abs_delta > query->timeout)
 933                                 query->timeout = 0;
 934                         else
 935                                 query->timeout += delta;
 936
 937                         /* Get the new delay from the first entry */
 938                         if (!delay) {
 939                                 delay = query->timeout - jiffies;
 940                                 if (delay <= 0)
 941                                         delay = 1;
 942                         }
 943                 }
 944                 if (delay)
 945                         mod_delayed_work(ib_nl_wq, &ib_nl_timed_work,
 946                                          (unsigned long)delay);
 947                 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
 948         }
 949
 950 settimeout_out:
 951         return skb->len;
 952 }
 953
 954 static inline int ib_nl_is_good_resolve_resp(const struct nlmsghdr *nlh)
 955 {
 956         struct nlattr *tb[LS_NLA_TYPE_MAX];
 957         int ret;
 958
 959         if (nlh->nlmsg_flags & RDMA_NL_LS_F_ERR)
 960                 return 0;
 961
 962         ret = nla_parse(tb, LS_NLA_TYPE_MAX - 1, nlmsg_data(nlh),
 963                         nlmsg_len(nlh), ib_nl_policy, NULL);
 964         if (ret)
 965                 return 0;
 966
 967         return 1;
 968 }
 969
 970 int ib_nl_handle_resolve_resp(struct sk_buff *skb,
 971                               struct netlink_callback *cb)
 972 {
 973         const struct nlmsghdr *nlh = (struct nlmsghdr *)cb->nlh;
 974         unsigned long flags;
 975         struct ib_sa_query *query;
 976         struct ib_mad_send_buf *send_buf;
 977         struct ib_mad_send_wc mad_send_wc;
 978         int found = 0;
 979         int ret;
 980
 981         if ((nlh->nlmsg_flags & NLM_F_REQUEST) ||
 982             !(NETLINK_CB(skb).sk) ||
 983             !netlink_capable(skb, CAP_NET_ADMIN))
 984                 return -EPERM;
 985
 986         spin_lock_irqsave(&ib_nl_request_lock, flags);
 987         list_for_each_entry(query, &ib_nl_request_list, list) {
 988                 /*
 989                  * If the query is cancelled, let the timeout routine
 990                  * take care of it.
 991                  */
 992                 if (nlh->nlmsg_seq == query->seq) {
 993                         found = !ib_sa_query_cancelled(query);
 994                         if (found)
 995                                 list_del(&query->list);
 996                         break;
 997                 }
 998         }
 999
1000         if (!found) {
1001                 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1002                 goto resp_out;
1003         }
1004
1005         send_buf = query->mad_buf;
1006
1007         if (!ib_nl_is_good_resolve_resp(nlh)) {
1008                 /* if the result is a failure, send out the packet via IB */
1009                 ib_sa_disable_local_svc(query);
1010                 ret = ib_post_send_mad(query->mad_buf, NULL);
1011                 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1012                 if (ret) {
1013                         mad_send_wc.send_buf = send_buf;
1014                         mad_send_wc.status = IB_WC_GENERAL_ERR;
1015                         send_handler(query->port->agent, &mad_send_wc);
1016                 }
1017         } else {
1018                 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1019                 ib_nl_process_good_resolve_rsp(query, nlh);
1020         }
1021
1022 resp_out:
1023         return skb->len;
1024 }
1025
1026 static void free_sm_ah(struct kref *kref)
1027 {
1028         struct ib_sa_sm_ah *sm_ah = container_of(kref, struct ib_sa_sm_ah, ref);
1029
1030         rdma_destroy_ah(sm_ah->ah);
1031         kfree(sm_ah);
1032 }
1033
1034 void ib_sa_register_client(struct ib_sa_client *client)
1035 {
1036         atomic_set(&client->users, 1);
1037         init_completion(&client->comp);
1038 }
1039 EXPORT_SYMBOL(ib_sa_register_client);
1040
1041 void ib_sa_unregister_client(struct ib_sa_client *client)
1042 {
1043         ib_sa_client_put(client);
1044         wait_for_completion(&client->comp);
1045 }
1046 EXPORT_SYMBOL(ib_sa_unregister_client);
1047
1048 /**
1049  * ib_sa_cancel_query - try to cancel an SA query
1050  * @id:ID of query to cancel
1051  * @query:query pointer to cancel
1052  *
1053  * Try to cancel an SA query.  If the id and query don't match up or
1054  * the query has already completed, nothing is done.  Otherwise the
1055  * query is canceled and will complete with a status of -EINTR.
1056  */
1057 void ib_sa_cancel_query(int id, struct ib_sa_query *query)
1058 {
1059         unsigned long flags;
1060         struct ib_mad_agent *agent;
1061         struct ib_mad_send_buf *mad_buf;
1062
1063         spin_lock_irqsave(&idr_lock, flags);
1064         if (idr_find(&query_idr, id) != query) {
1065                 spin_unlock_irqrestore(&idr_lock, flags);
1066                 return;
1067         }
1068         agent = query->port->agent;
1069         mad_buf = query->mad_buf;
1070         spin_unlock_irqrestore(&idr_lock, flags);
1071
1072         /*
1073          * If the query is still on the netlink request list, schedule
1074          * it to be cancelled by the timeout routine. Otherwise, it has been
1075          * sent to the MAD layer and has to be cancelled from there.
1076          */
1077         if (!ib_nl_cancel_request(query))
1078                 ib_cancel_mad(agent, mad_buf);
1079 }
1080 EXPORT_SYMBOL(ib_sa_cancel_query);
1081
1082 static u8 get_src_path_mask(struct ib_device *device, u8 port_num)
1083 {
1084         struct ib_sa_device *sa_dev;
1085         struct ib_sa_port   *port;
1086         unsigned long flags;
1087         u8 src_path_mask;
1088
1089         sa_dev = ib_get_client_data(device, &sa_client);
1090         if (!sa_dev)
1091                 return 0x7f;
1092
1093         port  = &sa_dev->port[port_num - sa_dev->start_port];
1094         spin_lock_irqsave(&port->ah_lock, flags);
1095         src_path_mask = port->sm_ah ? port->sm_ah->src_path_mask : 0x7f;
1096         spin_unlock_irqrestore(&port->ah_lock, flags);
1097
1098         return src_path_mask;
1099 }
1100
1101 int ib_init_ah_from_path(struct ib_device *device, u8 port_num,
1102                          struct ib_sa_path_rec *rec,
1103                          struct rdma_ah_attr *ah_attr)
1104 {
1105         int ret;
1106         u16 gid_index;
1107         int use_roce;
1108         struct net_device *ndev = NULL;
1109
1110         memset(ah_attr, 0, sizeof *ah_attr);
1111         ah_attr->type = rdma_ah_find_type(device, port_num);
1112
1113         rdma_ah_set_dlid(ah_attr, be16_to_cpu(rec->dlid));
1114         rdma_ah_set_sl(ah_attr, rec->sl);
1115         rdma_ah_set_path_bits(ah_attr, be16_to_cpu(rec->slid) &
1116                               get_src_path_mask(device, port_num));
1117         rdma_ah_set_port_num(ah_attr, port_num);
1118         rdma_ah_set_static_rate(ah_attr, rec->rate);
1119         use_roce = rdma_cap_eth_ah(device, port_num);
1120
1121         if (use_roce) {
1122                 struct net_device *idev;
1123                 struct net_device *resolved_dev;
1124                 struct rdma_dev_addr dev_addr = {.bound_dev_if = rec->ifindex,
1125                                                  .net = rec->net ? rec->net :
1126                                                          &init_net};
1127                 union {
1128                         struct sockaddr     _sockaddr;
1129                         struct sockaddr_in  _sockaddr_in;
1130                         struct sockaddr_in6 _sockaddr_in6;
1131                 } sgid_addr, dgid_addr;
1132
1133                 if (!device->get_netdev)
1134                         return -EOPNOTSUPP;
1135
1136                 rdma_gid2ip(&sgid_addr._sockaddr, &rec->sgid);
1137                 rdma_gid2ip(&dgid_addr._sockaddr, &rec->dgid);
1138
1139                 /* validate the route */
1140                 ret = rdma_resolve_ip_route(&sgid_addr._sockaddr,
1141                                             &dgid_addr._sockaddr, &dev_addr);
1142                 if (ret)
1143                         return ret;
1144
1145                 if ((dev_addr.network == RDMA_NETWORK_IPV4 ||
1146                      dev_addr.network == RDMA_NETWORK_IPV6) &&
1147                     rec->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
1148                         return -EINVAL;
1149
1150                 idev = device->get_netdev(device, port_num);
1151                 if (!idev)
1152                         return -ENODEV;
1153
1154                 resolved_dev = dev_get_by_index(dev_addr.net,
1155                                                 dev_addr.bound_dev_if);
1156                 if (resolved_dev->flags & IFF_LOOPBACK) {
1157                         dev_put(resolved_dev);
1158                         resolved_dev = idev;
1159                         dev_hold(resolved_dev);
1160                 }
1161                 ndev = ib_get_ndev_from_path(rec);
1162                 rcu_read_lock();
1163                 if ((ndev && ndev != resolved_dev) ||
1164                     (resolved_dev != idev &&
1165                      !rdma_is_upper_dev_rcu(idev, resolved_dev)))
1166                         ret = -EHOSTUNREACH;
1167                 rcu_read_unlock();
1168                 dev_put(idev);
1169                 dev_put(resolved_dev);
1170                 if (ret) {
1171                         if (ndev)
1172                                 dev_put(ndev);
1173                         return ret;
1174                 }
1175         }
1176
1177         if (rec->hop_limit > 0 || use_roce) {
1178                 ret = ib_find_cached_gid_by_port(device, &rec->sgid,
1179                                                  rec->gid_type, port_num, ndev,
1180                                                  &gid_index);
1181                 if (ret) {
1182                         if (ndev)
1183                                 dev_put(ndev);
1184                         return ret;
1185                 }
1186
1187                 rdma_ah_set_grh(ah_attr, &rec->dgid,
1188                                 be32_to_cpu(rec->flow_label),
1189                                 gid_index, rec->hop_limit,
1190                                 rec->traffic_class);
1191                 if (ndev)
1192                         dev_put(ndev);
1193         }
1194
1195         if (use_roce)
1196                 memcpy(ah_attr->roce.dmac, rec->dmac, ETH_ALEN);
1197
1198         return 0;
1199 }
1200 EXPORT_SYMBOL(ib_init_ah_from_path);
1201
1202 static int alloc_mad(struct ib_sa_query *query, gfp_t gfp_mask)
1203 {
1204         unsigned long flags;
1205
1206         spin_lock_irqsave(&query->port->ah_lock, flags);
1207         if (!query->port->sm_ah) {
1208                 spin_unlock_irqrestore(&query->port->ah_lock, flags);
1209                 return -EAGAIN;
1210         }
1211         kref_get(&query->port->sm_ah->ref);
1212         query->sm_ah = query->port->sm_ah;
1213         spin_unlock_irqrestore(&query->port->ah_lock, flags);
1214
1215         query->mad_buf = ib_create_send_mad(query->port->agent, 1,
1216                                             query->sm_ah->pkey_index,
1217                                             0, IB_MGMT_SA_HDR, IB_MGMT_SA_DATA,
1218                                             gfp_mask,
1219                                             ((query->flags & IB_SA_QUERY_OPA) ?
1220                                              OPA_MGMT_BASE_VERSION :
1221                                              IB_MGMT_BASE_VERSION));
1222         if (IS_ERR(query->mad_buf)) {
1223                 kref_put(&query->sm_ah->ref, free_sm_ah);
1224                 return -ENOMEM;
1225         }
1226
1227         query->mad_buf->ah = query->sm_ah->ah;
1228
1229         return 0;
1230 }
1231
1232 static void free_mad(struct ib_sa_query *query)
1233 {
1234         ib_free_send_mad(query->mad_buf);
1235         kref_put(&query->sm_ah->ref, free_sm_ah);
1236 }
1237
1238 static void init_mad(struct ib_sa_query *query, struct ib_mad_agent *agent)
1239 {
1240         struct ib_sa_mad *mad = query->mad_buf->mad;
1241         unsigned long flags;
1242
1243         memset(mad, 0, sizeof *mad);
1244
1245         if (query->flags & IB_SA_QUERY_OPA) {
1246                 mad->mad_hdr.base_version  = OPA_MGMT_BASE_VERSION;
1247                 mad->mad_hdr.class_version = OPA_SA_CLASS_VERSION;
1248         } else {
1249                 mad->mad_hdr.base_version  = IB_MGMT_BASE_VERSION;
1250                 mad->mad_hdr.class_version = IB_SA_CLASS_VERSION;
1251         }
1252         mad->mad_hdr.mgmt_class    = IB_MGMT_CLASS_SUBN_ADM;
1253         spin_lock_irqsave(&tid_lock, flags);
1254         mad->mad_hdr.tid           =
1255                 cpu_to_be64(((u64) agent->hi_tid) << 32 | tid++);
1256         spin_unlock_irqrestore(&tid_lock, flags);
1257 }
1258
1259 static int send_mad(struct ib_sa_query *query, int timeout_ms, gfp_t gfp_mask)
1260 {
1261         bool preload = gfpflags_allow_blocking(gfp_mask);
1262         unsigned long flags;
1263         int ret, id;
1264
1265         if (preload)
1266                 idr_preload(gfp_mask);
1267         spin_lock_irqsave(&idr_lock, flags);
1268
1269         id = idr_alloc(&query_idr, query, 0, 0, GFP_NOWAIT);
1270
1271         spin_unlock_irqrestore(&idr_lock, flags);
1272         if (preload)
1273                 idr_preload_end();
1274         if (id < 0)
1275                 return id;
1276
1277         query->mad_buf->timeout_ms  = timeout_ms;
1278         query->mad_buf->context[0] = query;
1279         query->id = id;
1280
1281         if (query->flags & IB_SA_ENABLE_LOCAL_SERVICE) {
1282                 if (!ibnl_chk_listeners(RDMA_NL_GROUP_LS)) {
1283                         if (!ib_nl_make_request(query, gfp_mask))
1284                                 return id;
1285                 }
1286                 ib_sa_disable_local_svc(query);
1287         }
1288
1289         ret = ib_post_send_mad(query->mad_buf, NULL);
1290         if (ret) {
1291                 spin_lock_irqsave(&idr_lock, flags);
1292                 idr_remove(&query_idr, id);
1293                 spin_unlock_irqrestore(&idr_lock, flags);
1294         }
1295
1296         /*
1297          * It's not safe to dereference query any more, because the
1298          * send may already have completed and freed the query in
1299          * another context.
1300          */
1301         return ret ? ret : id;
1302 }
1303
1304 void ib_sa_unpack_path(void *attribute, struct ib_sa_path_rec *rec)
1305 {
1306         ib_unpack(path_rec_table, ARRAY_SIZE(path_rec_table), attribute, rec);
1307 }
1308 EXPORT_SYMBOL(ib_sa_unpack_path);
1309
1310 void ib_sa_pack_path(struct ib_sa_path_rec *rec, void *attribute)
1311 {
1312         ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table), rec, attribute);
1313 }
1314 EXPORT_SYMBOL(ib_sa_pack_path);
1315
1316 static void ib_sa_path_rec_callback(struct ib_sa_query *sa_query,
1317                                     int status,
1318                                     struct ib_sa_mad *mad)
1319 {
1320         struct ib_sa_path_query *query =
1321                 container_of(sa_query, struct ib_sa_path_query, sa_query);
1322
1323         if (mad) {
1324                 struct ib_sa_path_rec rec;
1325
1326                 ib_unpack(path_rec_table, ARRAY_SIZE(path_rec_table),
1327                           mad->data, &rec);
1328                 rec.net = NULL;
1329                 rec.ifindex = 0;
1330                 rec.gid_type = IB_GID_TYPE_IB;
1331                 eth_zero_addr(rec.dmac);
1332                 query->callback(status, &rec, query->context);
1333         } else
1334                 query->callback(status, NULL, query->context);
1335 }
1336
1337 static void ib_sa_path_rec_release(struct ib_sa_query *sa_query)
1338 {
1339         kfree(container_of(sa_query, struct ib_sa_path_query, sa_query));
1340 }
1341
1342 /**
1343  * ib_sa_path_rec_get - Start a Path get query
1344  * @client:SA client
1345  * @device:device to send query on
1346  * @port_num: port number to send query on
1347  * @rec:Path Record to send in query
1348  * @comp_mask:component mask to send in query
1349  * @timeout_ms:time to wait for response
1350  * @gfp_mask:GFP mask to use for internal allocations
1351  * @callback:function called when query completes, times out or is
1352  * canceled
1353  * @context:opaque user context passed to callback
1354  * @sa_query:query context, used to cancel query
1355  *
1356  * Send a Path Record Get query to the SA to look up a path.  The
1357  * callback function will be called when the query completes (or
1358  * fails); status is 0 for a successful response, -EINTR if the query
1359  * is canceled, -ETIMEDOUT is the query timed out, or -EIO if an error
1360  * occurred sending the query.  The resp parameter of the callback is
1361  * only valid if status is 0.
1362  *
1363  * If the return value of ib_sa_path_rec_get() is negative, it is an
1364  * error code.  Otherwise it is a query ID that can be used to cancel
1365  * the query.
1366  */
1367 int ib_sa_path_rec_get(struct ib_sa_client *client,
1368                        struct ib_device *device, u8 port_num,
1369                        struct ib_sa_path_rec *rec,
1370                        ib_sa_comp_mask comp_mask,
1371                        int timeout_ms, gfp_t gfp_mask,
1372                        void (*callback)(int status,
1373                                         struct ib_sa_path_rec *resp,
1374                                         void *context),
1375                        void *context,
1376                        struct ib_sa_query **sa_query)
1377 {
1378         struct ib_sa_path_query *query;
1379         struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1380         struct ib_sa_port   *port;
1381         struct ib_mad_agent *agent;
1382         struct ib_sa_mad *mad;
1383         int ret;
1384
1385         if (!sa_dev)
1386                 return -ENODEV;
1387
1388         port  = &sa_dev->port[port_num - sa_dev->start_port];
1389         agent = port->agent;
1390
1391         query = kzalloc(sizeof(*query), gfp_mask);
1392         if (!query)
1393                 return -ENOMEM;
1394
1395         query->sa_query.port     = port;
1396         ret = alloc_mad(&query->sa_query, gfp_mask);
1397         if (ret)
1398                 goto err1;
1399
1400         ib_sa_client_get(client);
1401         query->sa_query.client = client;
1402         query->callback        = callback;
1403         query->context         = context;
1404
1405         mad = query->sa_query.mad_buf->mad;
1406         init_mad(&query->sa_query, agent);
1407
1408         query->sa_query.callback = callback ? ib_sa_path_rec_callback : NULL;
1409         query->sa_query.release  = ib_sa_path_rec_release;
1410         mad->mad_hdr.method      = IB_MGMT_METHOD_GET;
1411         mad->mad_hdr.attr_id     = cpu_to_be16(IB_SA_ATTR_PATH_REC);
1412         mad->sa_hdr.comp_mask    = comp_mask;
1413
1414         ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table), rec, mad->data);
1415
1416         *sa_query = &query->sa_query;
1417
1418         query->sa_query.flags |= IB_SA_ENABLE_LOCAL_SERVICE;
1419         query->sa_query.mad_buf->context[1] = rec;
1420
1421         ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1422         if (ret < 0)
1423                 goto err2;
1424
1425         return ret;
1426
1427 err2:
1428         *sa_query = NULL;
1429         ib_sa_client_put(query->sa_query.client);
1430         free_mad(&query->sa_query);
1431
1432 err1:
1433         kfree(query);
1434         return ret;
1435 }
1436 EXPORT_SYMBOL(ib_sa_path_rec_get);
1437
1438 static void ib_sa_service_rec_callback(struct ib_sa_query *sa_query,
1439                                     int status,
1440                                     struct ib_sa_mad *mad)
1441 {
1442         struct ib_sa_service_query *query =
1443                 container_of(sa_query, struct ib_sa_service_query, sa_query);
1444
1445         if (mad) {
1446                 struct ib_sa_service_rec rec;
1447
1448                 ib_unpack(service_rec_table, ARRAY_SIZE(service_rec_table),
1449                           mad->data, &rec);
1450                 query->callback(status, &rec, query->context);
1451         } else
1452                 query->callback(status, NULL, query->context);
1453 }
1454
1455 static void ib_sa_service_rec_release(struct ib_sa_query *sa_query)
1456 {
1457         kfree(container_of(sa_query, struct ib_sa_service_query, sa_query));
1458 }
1459
1460 /**
1461  * ib_sa_service_rec_query - Start Service Record operation
1462  * @client:SA client
1463  * @device:device to send request on
1464  * @port_num: port number to send request on
1465  * @method:SA method - should be get, set, or delete
1466  * @rec:Service Record to send in request
1467  * @comp_mask:component mask to send in request
1468  * @timeout_ms:time to wait for response
1469  * @gfp_mask:GFP mask to use for internal allocations
1470  * @callback:function called when request completes, times out or is
1471  * canceled
1472  * @context:opaque user context passed to callback
1473  * @sa_query:request context, used to cancel request
1474  *
1475  * Send a Service Record set/get/delete to the SA to register,
1476  * unregister or query a service record.
1477  * The callback function will be called when the request completes (or
1478  * fails); status is 0 for a successful response, -EINTR if the query
1479  * is canceled, -ETIMEDOUT is the query timed out, or -EIO if an error
1480  * occurred sending the query.  The resp parameter of the callback is
1481  * only valid if status is 0.
1482  *
1483  * If the return value of ib_sa_service_rec_query() is negative, it is an
1484  * error code.  Otherwise it is a request ID that can be used to cancel
1485  * the query.
1486  */
1487 int ib_sa_service_rec_query(struct ib_sa_client *client,
1488                             struct ib_device *device, u8 port_num, u8 method,
1489                             struct ib_sa_service_rec *rec,
1490                             ib_sa_comp_mask comp_mask,
1491                             int timeout_ms, gfp_t gfp_mask,
1492                             void (*callback)(int status,
1493                                              struct ib_sa_service_rec *resp,
1494                                              void *context),
1495                             void *context,
1496                             struct ib_sa_query **sa_query)
1497 {
1498         struct ib_sa_service_query *query;
1499         struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1500         struct ib_sa_port   *port;
1501         struct ib_mad_agent *agent;
1502         struct ib_sa_mad *mad;
1503         int ret;
1504
1505         if (!sa_dev)
1506                 return -ENODEV;
1507
1508         port  = &sa_dev->port[port_num - sa_dev->start_port];
1509         agent = port->agent;
1510
1511         if (method != IB_MGMT_METHOD_GET &&
1512             method != IB_MGMT_METHOD_SET &&
1513             method != IB_SA_METHOD_DELETE)
1514                 return -EINVAL;
1515
1516         query = kzalloc(sizeof(*query), gfp_mask);
1517         if (!query)
1518                 return -ENOMEM;
1519
1520         query->sa_query.port     = port;
1521         ret = alloc_mad(&query->sa_query, gfp_mask);
1522         if (ret)
1523                 goto err1;
1524
1525         ib_sa_client_get(client);
1526         query->sa_query.client = client;
1527         query->callback        = callback;
1528         query->context         = context;
1529
1530         mad = query->sa_query.mad_buf->mad;
1531         init_mad(&query->sa_query, agent);
1532
1533         query->sa_query.callback = callback ? ib_sa_service_rec_callback : NULL;
1534         query->sa_query.release  = ib_sa_service_rec_release;
1535         mad->mad_hdr.method      = method;
1536         mad->mad_hdr.attr_id     = cpu_to_be16(IB_SA_ATTR_SERVICE_REC);
1537         mad->sa_hdr.comp_mask    = comp_mask;
1538
1539         ib_pack(service_rec_table, ARRAY_SIZE(service_rec_table),
1540                 rec, mad->data);
1541
1542         *sa_query = &query->sa_query;
1543
1544         ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1545         if (ret < 0)
1546                 goto err2;
1547
1548         return ret;
1549
1550 err2:
1551         *sa_query = NULL;
1552         ib_sa_client_put(query->sa_query.client);
1553         free_mad(&query->sa_query);
1554
1555 err1:
1556         kfree(query);
1557         return ret;
1558 }
1559 EXPORT_SYMBOL(ib_sa_service_rec_query);
1560
1561 static void ib_sa_mcmember_rec_callback(struct ib_sa_query *sa_query,
1562                                         int status,
1563                                         struct ib_sa_mad *mad)
1564 {
1565         struct ib_sa_mcmember_query *query =
1566                 container_of(sa_query, struct ib_sa_mcmember_query, sa_query);
1567
1568         if (mad) {
1569                 struct ib_sa_mcmember_rec rec;
1570
1571                 ib_unpack(mcmember_rec_table, ARRAY_SIZE(mcmember_rec_table),
1572                           mad->data, &rec);
1573                 query->callback(status, &rec, query->context);
1574         } else
1575                 query->callback(status, NULL, query->context);
1576 }
1577
1578 static void ib_sa_mcmember_rec_release(struct ib_sa_query *sa_query)
1579 {
1580         kfree(container_of(sa_query, struct ib_sa_mcmember_query, sa_query));
1581 }
1582
1583 int ib_sa_mcmember_rec_query(struct ib_sa_client *client,
1584                              struct ib_device *device, u8 port_num,
1585                              u8 method,
1586                              struct ib_sa_mcmember_rec *rec,
1587                              ib_sa_comp_mask comp_mask,
1588                              int timeout_ms, gfp_t gfp_mask,
1589                              void (*callback)(int status,
1590                                               struct ib_sa_mcmember_rec *resp,
1591                                               void *context),
1592                              void *context,
1593                              struct ib_sa_query **sa_query)
1594 {
1595         struct ib_sa_mcmember_query *query;
1596         struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1597         struct ib_sa_port   *port;
1598         struct ib_mad_agent *agent;
1599         struct ib_sa_mad *mad;
1600         int ret;
1601
1602         if (!sa_dev)
1603                 return -ENODEV;
1604
1605         port  = &sa_dev->port[port_num - sa_dev->start_port];
1606         agent = port->agent;
1607
1608         query = kzalloc(sizeof(*query), gfp_mask);
1609         if (!query)
1610                 return -ENOMEM;
1611
1612         query->sa_query.port     = port;
1613         ret = alloc_mad(&query->sa_query, gfp_mask);
1614         if (ret)
1615                 goto err1;
1616
1617         ib_sa_client_get(client);
1618         query->sa_query.client = client;
1619         query->callback        = callback;
1620         query->context         = context;
1621
1622         mad = query->sa_query.mad_buf->mad;
1623         init_mad(&query->sa_query, agent);
1624
1625         query->sa_query.callback = callback ? ib_sa_mcmember_rec_callback : NULL;
1626         query->sa_query.release  = ib_sa_mcmember_rec_release;
1627         mad->mad_hdr.method      = method;
1628         mad->mad_hdr.attr_id     = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
1629         mad->sa_hdr.comp_mask    = comp_mask;
1630
1631         ib_pack(mcmember_rec_table, ARRAY_SIZE(mcmember_rec_table),
1632                 rec, mad->data);
1633
1634         *sa_query = &query->sa_query;
1635
1636         ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1637         if (ret < 0)
1638                 goto err2;
1639
1640         return ret;
1641
1642 err2:
1643         *sa_query = NULL;
1644         ib_sa_client_put(query->sa_query.client);
1645         free_mad(&query->sa_query);
1646
1647 err1:
1648         kfree(query);
1649         return ret;
1650 }
1651
1652 /* Support GuidInfoRecord */
1653 static void ib_sa_guidinfo_rec_callback(struct ib_sa_query *sa_query,
1654                                         int status,
1655                                         struct ib_sa_mad *mad)
1656 {
1657         struct ib_sa_guidinfo_query *query =
1658                 container_of(sa_query, struct ib_sa_guidinfo_query, sa_query);
1659
1660         if (mad) {
1661                 struct ib_sa_guidinfo_rec rec;
1662
1663                 ib_unpack(guidinfo_rec_table, ARRAY_SIZE(guidinfo_rec_table),
1664                           mad->data, &rec);
1665                 query->callback(status, &rec, query->context);
1666         } else
1667                 query->callback(status, NULL, query->context);
1668 }
1669
1670 static void ib_sa_guidinfo_rec_release(struct ib_sa_query *sa_query)
1671 {
1672         kfree(container_of(sa_query, struct ib_sa_guidinfo_query, sa_query));
1673 }
1674
1675 int ib_sa_guid_info_rec_query(struct ib_sa_client *client,
1676                               struct ib_device *device, u8 port_num,
1677                               struct ib_sa_guidinfo_rec *rec,
1678                               ib_sa_comp_mask comp_mask, u8 method,
1679                               int timeout_ms, gfp_t gfp_mask,
1680                               void (*callback)(int status,
1681                                                struct ib_sa_guidinfo_rec *resp,
1682                                                void *context),
1683                               void *context,
1684                               struct ib_sa_query **sa_query)
1685 {
1686         struct ib_sa_guidinfo_query *query;
1687         struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1688         struct ib_sa_port *port;
1689         struct ib_mad_agent *agent;
1690         struct ib_sa_mad *mad;
1691         int ret;
1692
1693         if (!sa_dev)
1694                 return -ENODEV;
1695
1696         if (method != IB_MGMT_METHOD_GET &&
1697             method != IB_MGMT_METHOD_SET &&
1698             method != IB_SA_METHOD_DELETE) {
1699                 return -EINVAL;
1700         }
1701
1702         port  = &sa_dev->port[port_num - sa_dev->start_port];
1703         agent = port->agent;
1704
1705         query = kzalloc(sizeof(*query), gfp_mask);
1706         if (!query)
1707                 return -ENOMEM;
1708
1709         query->sa_query.port = port;
1710         ret = alloc_mad(&query->sa_query, gfp_mask);
1711         if (ret)
1712                 goto err1;
1713
1714         ib_sa_client_get(client);
1715         query->sa_query.client = client;
1716         query->callback        = callback;
1717         query->context         = context;
1718
1719         mad = query->sa_query.mad_buf->mad;
1720         init_mad(&query->sa_query, agent);
1721
1722         query->sa_query.callback = callback ? ib_sa_guidinfo_rec_callback : NULL;
1723         query->sa_query.release  = ib_sa_guidinfo_rec_release;
1724
1725         mad->mad_hdr.method      = method;
1726         mad->mad_hdr.attr_id     = cpu_to_be16(IB_SA_ATTR_GUID_INFO_REC);
1727         mad->sa_hdr.comp_mask    = comp_mask;
1728
1729         ib_pack(guidinfo_rec_table, ARRAY_SIZE(guidinfo_rec_table), rec,
1730                 mad->data);
1731
1732         *sa_query = &query->sa_query;
1733
1734         ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1735         if (ret < 0)
1736                 goto err2;
1737
1738         return ret;
1739
1740 err2:
1741         *sa_query = NULL;
1742         ib_sa_client_put(query->sa_query.client);
1743         free_mad(&query->sa_query);
1744
1745 err1:
1746         kfree(query);
1747         return ret;
1748 }
1749 EXPORT_SYMBOL(ib_sa_guid_info_rec_query);
1750
1751 bool ib_sa_sendonly_fullmem_support(struct ib_sa_client *client,
1752                                     struct ib_device *device,
1753                                     u8 port_num)
1754 {
1755         struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1756         struct ib_sa_port *port;
1757         bool ret = false;
1758         unsigned long flags;
1759
1760         if (!sa_dev)
1761                 return ret;
1762
1763         port  = &sa_dev->port[port_num - sa_dev->start_port];
1764
1765         spin_lock_irqsave(&port->classport_lock, flags);
1766         if ((port->classport_info.valid) &&
1767             (port->classport_info.data.type == RDMA_CLASS_PORT_INFO_IB))
1768                 ret = ib_get_cpi_capmask2(&port->classport_info.data.ib)
1769                         & IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT;
1770         spin_unlock_irqrestore(&port->classport_lock, flags);
1771         return ret;
1772 }
1773 EXPORT_SYMBOL(ib_sa_sendonly_fullmem_support);
1774
1775 struct ib_classport_info_context {
1776         struct completion       done;
1777         struct ib_sa_query      *sa_query;
1778 };
1779
1780 static void ib_classportinfo_cb(void *context)
1781 {
1782         struct ib_classport_info_context *cb_ctx = context;
1783
1784         complete(&cb_ctx->done);
1785 }
1786
1787 static void ib_sa_classport_info_rec_callback(struct ib_sa_query *sa_query,
1788                                               int status,
1789                                               struct ib_sa_mad *mad)
1790 {
1791         unsigned long flags;
1792         struct ib_sa_classport_info_query *query =
1793                 container_of(sa_query, struct ib_sa_classport_info_query, sa_query);
1794         struct ib_sa_classport_cache *info = &sa_query->port->classport_info;
1795
1796         if (mad) {
1797                 if (sa_query->flags & IB_SA_QUERY_OPA) {
1798                         struct opa_class_port_info rec;
1799
1800                         ib_unpack(opa_classport_info_rec_table,
1801                                   ARRAY_SIZE(opa_classport_info_rec_table),
1802                                   mad->data, &rec);
1803
1804                         spin_lock_irqsave(&sa_query->port->classport_lock,
1805                                           flags);
1806                         if (!status && !info->valid) {
1807                                 memcpy(&info->data.opa, &rec,
1808                                        sizeof(info->data.opa));
1809
1810                                 info->valid = true;
1811                                 info->data.type = RDMA_CLASS_PORT_INFO_OPA;
1812                         }
1813                         spin_unlock_irqrestore(&sa_query->port->classport_lock,
1814                                                flags);
1815
1816                 } else {
1817                         struct ib_class_port_info rec;
1818
1819                         ib_unpack(ib_classport_info_rec_table,
1820                                   ARRAY_SIZE(ib_classport_info_rec_table),
1821                                   mad->data, &rec);
1822
1823                         spin_lock_irqsave(&sa_query->port->classport_lock,
1824                                           flags);
1825                         if (!status && !info->valid) {
1826                                 memcpy(&info->data.ib, &rec,
1827                                        sizeof(info->data.ib));
1828
1829                                 info->valid = true;
1830                                 info->data.type = RDMA_CLASS_PORT_INFO_IB;
1831                         }
1832                         spin_unlock_irqrestore(&sa_query->port->classport_lock,
1833                                                flags);
1834                 }
1835         }
1836         query->callback(query->context);
1837 }
1838
1839 static void ib_sa_classport_info_rec_release(struct ib_sa_query *sa_query)
1840 {
1841         kfree(container_of(sa_query, struct ib_sa_classport_info_query,
1842                            sa_query));
1843 }
1844
1845 static int ib_sa_classport_info_rec_query(struct ib_sa_port *port,
1846                                           int timeout_ms,
1847                                           void (*callback)(void *context),
1848                                           void *context,
1849                                           struct ib_sa_query **sa_query)
1850 {
1851         struct ib_mad_agent *agent;
1852         struct ib_sa_classport_info_query *query;
1853         struct ib_sa_mad *mad;
1854         gfp_t gfp_mask = GFP_KERNEL;
1855         int ret;
1856
1857         agent = port->agent;
1858
1859         query = kzalloc(sizeof(*query), gfp_mask);
1860         if (!query)
1861                 return -ENOMEM;
1862
1863         query->sa_query.port = port;
1864         query->sa_query.flags |= rdma_cap_opa_ah(port->agent->device,
1865                                                  port->port_num) ?
1866                                  IB_SA_QUERY_OPA : 0;
1867         ret = alloc_mad(&query->sa_query, gfp_mask);
1868         if (ret)
1869                 goto err_free;
1870
1871         query->callback = callback;
1872         query->context = context;
1873
1874         mad = query->sa_query.mad_buf->mad;
1875         init_mad(&query->sa_query, agent);
1876
1877         query->sa_query.callback = ib_sa_classport_info_rec_callback;
1878         query->sa_query.release  = ib_sa_classport_info_rec_release;
1879         mad->mad_hdr.method      = IB_MGMT_METHOD_GET;
1880         mad->mad_hdr.attr_id     = cpu_to_be16(IB_SA_ATTR_CLASS_PORTINFO);
1881         mad->sa_hdr.comp_mask    = 0;
1882         *sa_query = &query->sa_query;
1883
1884         ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1885         if (ret < 0)
1886                 goto err_free_mad;
1887
1888         return ret;
1889
1890 err_free_mad:
1891         *sa_query = NULL;
1892         free_mad(&query->sa_query);
1893
1894 err_free:
1895         kfree(query);
1896         return ret;
1897 }
1898
1899 static void update_ib_cpi(struct work_struct *work)
1900 {
1901         struct ib_sa_port *port =
1902                 container_of(work, struct ib_sa_port, ib_cpi_work.work);
1903         struct ib_classport_info_context *cb_context;
1904         unsigned long flags;
1905         int ret;
1906
1907         /* If the classport info is valid, nothing
1908          * to do here.
1909          */
1910         spin_lock_irqsave(&port->classport_lock, flags);
1911         if (port->classport_info.valid) {
1912                 spin_unlock_irqrestore(&port->classport_lock, flags);
1913                 return;
1914         }
1915         spin_unlock_irqrestore(&port->classport_lock, flags);
1916
1917         cb_context = kmalloc(sizeof(*cb_context), GFP_KERNEL);
1918         if (!cb_context)
1919                 goto err_nomem;
1920
1921         init_completion(&cb_context->done);
1922
1923         ret = ib_sa_classport_info_rec_query(port, 3000,
1924                                              ib_classportinfo_cb, cb_context,
1925                                              &cb_context->sa_query);
1926         if (ret < 0)
1927                 goto free_cb_err;
1928         wait_for_completion(&cb_context->done);
1929 free_cb_err:
1930         kfree(cb_context);
1931         spin_lock_irqsave(&port->classport_lock, flags);
1932
1933         /* If the classport info is still not valid, the query should have
1934          * failed for some reason. Retry issuing the query
1935          */
1936         if (!port->classport_info.valid) {
1937                 port->classport_info.retry_cnt++;
1938                 if (port->classport_info.retry_cnt <=
1939                     IB_SA_CPI_MAX_RETRY_CNT) {
1940                         unsigned long delay =
1941                                 msecs_to_jiffies(IB_SA_CPI_RETRY_WAIT);
1942
1943                         queue_delayed_work(ib_wq, &port->ib_cpi_work, delay);
1944                 }
1945         }
1946         spin_unlock_irqrestore(&port->classport_lock, flags);
1947
1948 err_nomem:
1949         return;
1950 }
1951
1952 static void send_handler(struct ib_mad_agent *agent,
1953                          struct ib_mad_send_wc *mad_send_wc)
1954 {
1955         struct ib_sa_query *query = mad_send_wc->send_buf->context[0];
1956         unsigned long flags;
1957
1958         if (query->callback)
1959                 switch (mad_send_wc->status) {
1960                 case IB_WC_SUCCESS:
1961                         /* No callback -- already got recv */
1962                         break;
1963                 case IB_WC_RESP_TIMEOUT_ERR:
1964                         query->callback(query, -ETIMEDOUT, NULL);
1965                         break;
1966                 case IB_WC_WR_FLUSH_ERR:
1967                         query->callback(query, -EINTR, NULL);
1968                         break;
1969                 default:
1970                         query->callback(query, -EIO, NULL);
1971                         break;
1972                 }
1973
1974         spin_lock_irqsave(&idr_lock, flags);
1975         idr_remove(&query_idr, query->id);
1976         spin_unlock_irqrestore(&idr_lock, flags);
1977
1978         free_mad(query);
1979         if (query->client)
1980                 ib_sa_client_put(query->client);
1981         query->release(query);
1982 }
1983
1984 static void recv_handler(struct ib_mad_agent *mad_agent,
1985                          struct ib_mad_send_buf *send_buf,
1986                          struct ib_mad_recv_wc *mad_recv_wc)
1987 {
1988         struct ib_sa_query *query;
1989
1990         if (!send_buf)
1991                 return;
1992
1993         query = send_buf->context[0];
1994         if (query->callback) {
1995                 if (mad_recv_wc->wc->status == IB_WC_SUCCESS)
1996                         query->callback(query,
1997                                         mad_recv_wc->recv_buf.mad->mad_hdr.status ?
1998                                         -EINVAL : 0,
1999                                         (struct ib_sa_mad *) mad_recv_wc->recv_buf.mad);
2000                 else
2001                         query->callback(query, -EIO, NULL);
2002         }
2003
2004         ib_free_recv_mad(mad_recv_wc);
2005 }
2006
2007 static void update_sm_ah(struct work_struct *work)
2008 {
2009         struct ib_sa_port *port =
2010                 container_of(work, struct ib_sa_port, update_task);
2011         struct ib_sa_sm_ah *new_ah;
2012         struct ib_port_attr port_attr;
2013         struct rdma_ah_attr   ah_attr;
2014
2015         if (ib_query_port(port->agent->device, port->port_num, &port_attr)) {
2016                 pr_warn("Couldn't query port\n");
2017                 return;
2018         }
2019
2020         new_ah = kmalloc(sizeof(*new_ah), GFP_KERNEL);
2021         if (!new_ah)
2022                 return;
2023
2024         kref_init(&new_ah->ref);
2025         new_ah->src_path_mask = (1 << port_attr.lmc) - 1;
2026
2027         new_ah->pkey_index = 0;
2028         if (ib_find_pkey(port->agent->device, port->port_num,
2029                          IB_DEFAULT_PKEY_FULL, &new_ah->pkey_index))
2030                 pr_err("Couldn't find index for default PKey\n");
2031
2032         memset(&ah_attr, 0, sizeof(ah_attr));
2033         ah_attr.type = rdma_ah_find_type(port->agent->device,
2034                                          port->port_num);
2035         rdma_ah_set_dlid(&ah_attr, port_attr.sm_lid);
2036         rdma_ah_set_sl(&ah_attr, port_attr.sm_sl);
2037         rdma_ah_set_port_num(&ah_attr, port->port_num);
2038         if (port_attr.grh_required) {
2039                 rdma_ah_set_ah_flags(&ah_attr, IB_AH_GRH);
2040
2041                 rdma_ah_set_subnet_prefix(&ah_attr,
2042                                           cpu_to_be64(port_attr.subnet_prefix));
2043                 rdma_ah_set_interface_id(&ah_attr,
2044                                          cpu_to_be64(IB_SA_WELL_KNOWN_GUID));
2045         }
2046
2047         new_ah->ah = rdma_create_ah(port->agent->qp->pd, &ah_attr);
2048         if (IS_ERR(new_ah->ah)) {
2049                 pr_warn("Couldn't create new SM AH\n");
2050                 kfree(new_ah);
2051                 return;
2052         }
2053
2054         spin_lock_irq(&port->ah_lock);
2055         if (port->sm_ah)
2056                 kref_put(&port->sm_ah->ref, free_sm_ah);
2057         port->sm_ah = new_ah;
2058         spin_unlock_irq(&port->ah_lock);
2059 }
2060
2061 static void ib_sa_event(struct ib_event_handler *handler,
2062                         struct ib_event *event)
2063 {
2064         if (event->event == IB_EVENT_PORT_ERR    ||
2065             event->event == IB_EVENT_PORT_ACTIVE ||
2066             event->event == IB_EVENT_LID_CHANGE  ||
2067             event->event == IB_EVENT_PKEY_CHANGE ||
2068             event->event == IB_EVENT_SM_CHANGE   ||
2069             event->event == IB_EVENT_CLIENT_REREGISTER) {
2070                 unsigned long flags;
2071                 struct ib_sa_device *sa_dev =
2072                         container_of(handler, typeof(*sa_dev), event_handler);
2073                 u8 port_num = event->element.port_num - sa_dev->start_port;
2074                 struct ib_sa_port *port = &sa_dev->port[port_num];
2075
2076                 if (!rdma_cap_ib_sa(handler->device, port->port_num))
2077                         return;
2078
2079                 spin_lock_irqsave(&port->ah_lock, flags);
2080                 if (port->sm_ah)
2081                         kref_put(&port->sm_ah->ref, free_sm_ah);
2082                 port->sm_ah = NULL;
2083                 spin_unlock_irqrestore(&port->ah_lock, flags);
2084
2085                 if (event->event == IB_EVENT_SM_CHANGE ||
2086                     event->event == IB_EVENT_CLIENT_REREGISTER ||
2087                     event->event == IB_EVENT_LID_CHANGE ||
2088                     event->event == IB_EVENT_PORT_ACTIVE) {
2089                         unsigned long delay =
2090                                 msecs_to_jiffies(IB_SA_CPI_RETRY_WAIT);
2091
2092                         spin_lock_irqsave(&port->classport_lock, flags);
2093                         port->classport_info.valid = false;
2094                         port->classport_info.retry_cnt = 0;
2095                         spin_unlock_irqrestore(&port->classport_lock, flags);
2096                         queue_delayed_work(ib_wq,
2097                                            &port->ib_cpi_work, delay);
2098                 }
2099                 queue_work(ib_wq, &sa_dev->port[port_num].update_task);
2100         }
2101 }
2102
2103 static void ib_sa_add_one(struct ib_device *device)
2104 {
2105         struct ib_sa_device *sa_dev;
2106         int s, e, i;
2107         int count = 0;
2108
2109         s = rdma_start_port(device);
2110         e = rdma_end_port(device);
2111
2112         sa_dev = kzalloc(sizeof *sa_dev +
2113                          (e - s + 1) * sizeof (struct ib_sa_port),
2114                          GFP_KERNEL);
2115         if (!sa_dev)
2116                 return;
2117
2118         sa_dev->start_port = s;
2119         sa_dev->end_port   = e;
2120
2121         for (i = 0; i <= e - s; ++i) {
2122                 spin_lock_init(&sa_dev->port[i].ah_lock);
2123                 if (!rdma_cap_ib_sa(device, i + 1))
2124                         continue;
2125
2126                 sa_dev->port[i].sm_ah    = NULL;
2127                 sa_dev->port[i].port_num = i + s;
2128
2129                 spin_lock_init(&sa_dev->port[i].classport_lock);
2130                 sa_dev->port[i].classport_info.valid = false;
2131
2132                 sa_dev->port[i].agent =
2133                         ib_register_mad_agent(device, i + s, IB_QPT_GSI,
2134                                               NULL, 0, send_handler,
2135                                               recv_handler, sa_dev, 0);
2136                 if (IS_ERR(sa_dev->port[i].agent))
2137                         goto err;
2138
2139                 INIT_WORK(&sa_dev->port[i].update_task, update_sm_ah);
2140                 INIT_DELAYED_WORK(&sa_dev->port[i].ib_cpi_work,
2141                                   update_ib_cpi);
2142
2143                 count++;
2144         }
2145
2146         if (!count)
2147                 goto free;
2148
2149         ib_set_client_data(device, &sa_client, sa_dev);
2150
2151         /*
2152          * We register our event handler after everything is set up,
2153          * and then update our cached info after the event handler is
2154          * registered to avoid any problems if a port changes state
2155          * during our initialization.
2156          */
2157
2158         INIT_IB_EVENT_HANDLER(&sa_dev->event_handler, device, ib_sa_event);
2159         if (ib_register_event_handler(&sa_dev->event_handler))
2160                 goto err;
2161
2162         for (i = 0; i <= e - s; ++i) {
2163                 if (rdma_cap_ib_sa(device, i + 1))
2164                         update_sm_ah(&sa_dev->port[i].update_task);
2165         }
2166
2167         return;
2168
2169 err:
2170         while (--i >= 0) {
2171                 if (rdma_cap_ib_sa(device, i + 1))
2172                         ib_unregister_mad_agent(sa_dev->port[i].agent);
2173         }
2174 free:
2175         kfree(sa_dev);
2176         return;
2177 }
2178
2179 static void ib_sa_remove_one(struct ib_device *device, void *client_data)
2180 {
2181         struct ib_sa_device *sa_dev = client_data;
2182         int i;
2183
2184         if (!sa_dev)
2185                 return;
2186
2187         ib_unregister_event_handler(&sa_dev->event_handler);
2188         flush_workqueue(ib_wq);
2189
2190         for (i = 0; i <= sa_dev->end_port - sa_dev->start_port; ++i) {
2191                 if (rdma_cap_ib_sa(device, i + 1)) {
2192                         cancel_delayed_work_sync(&sa_dev->port[i].ib_cpi_work);
2193                         ib_unregister_mad_agent(sa_dev->port[i].agent);
2194                         if (sa_dev->port[i].sm_ah)
2195                                 kref_put(&sa_dev->port[i].sm_ah->ref, free_sm_ah);
2196                 }
2197
2198         }
2199
2200         kfree(sa_dev);
2201 }
2202
2203 int ib_sa_init(void)
2204 {
2205         int ret;
2206
2207         get_random_bytes(&tid, sizeof tid);
2208
2209         atomic_set(&ib_nl_sa_request_seq, 0);
2210
2211         ret = ib_register_client(&sa_client);
2212         if (ret) {
2213                 pr_err("Couldn't register ib_sa client\n");
2214                 goto err1;
2215         }
2216
2217         ret = mcast_init();
2218         if (ret) {
2219                 pr_err("Couldn't initialize multicast handling\n");
2220                 goto err2;
2221         }
2222
2223         ib_nl_wq = alloc_ordered_workqueue("ib_nl_sa_wq", WQ_MEM_RECLAIM);
2224         if (!ib_nl_wq) {
2225                 ret = -ENOMEM;
2226                 goto err3;
2227         }
2228
2229         INIT_DELAYED_WORK(&ib_nl_timed_work, ib_nl_request_timeout);
2230
2231         return 0;
2232
2233 err3:
2234         mcast_cleanup();
2235 err2:
2236         ib_unregister_client(&sa_client);
2237 err1:
2238         return ret;
2239 }
2240
2241 void ib_sa_cleanup(void)
2242 {
2243         cancel_delayed_work(&ib_nl_timed_work);
2244         flush_workqueue(ib_nl_wq);
2245         destroy_workqueue(ib_nl_wq);
2246         mcast_cleanup();
2247         ib_unregister_client(&sa_client);
2248         idr_destroy(&query_idr);
2249 }