1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0 3 * 4 * Copyright (c) 2004-2007 Voltaire, Inc. All rights reserved. 5 * Copyright (c) 2005 Intel Corporation. All rights reserved. 6 * Copyright (c) 2005 Mellanox Technologies Ltd. All rights reserved. 7 * Copyright (c) 2009 HNR Consulting. All rights reserved. 8 * Copyright (c) 2014 Intel Corporation. All rights reserved. 9 * 10 * This software is available to you under a choice of one of two 11 * licenses. You may choose to be licensed under the terms of the GNU 12 * General Public License (GPL) Version 2, available from the file 13 * COPYING in the main directory of this source tree, or the 14 * OpenIB.org BSD license below: 15 * 16 * Redistribution and use in source and binary forms, with or 17 * without modification, are permitted provided that the following 18 * conditions are met: 19 * 20 * - Redistributions of source code must retain the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer. 23 * 24 * - Redistributions in binary form must reproduce the above 25 * copyright notice, this list of conditions and the following 26 * disclaimer in the documentation and/or other materials 27 * provided with the distribution. 28 * 29 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 30 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 31 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 32 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 33 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 34 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 35 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 36 * SOFTWARE. 37 */ 38 39 #include <sys/cdefs.h> 40 __FBSDID("$FreeBSD$"); 41 42 #define LINUXKPI_PARAM_PREFIX ibcore_ 43 #define KBUILD_MODNAME "ibcore" 44 45 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 46 47 #include <linux/dma-mapping.h> 48 #include <linux/slab.h> 49 #include <linux/module.h> 50 #include <rdma/ib_cache.h> 51 52 #include "mad_priv.h" 53 #include "mad_rmpp.h" 54 #include "smi.h" 55 #include "opa_smi.h" 56 #include "agent.h" 57 #include "core_priv.h" 58 59 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE; 60 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE; 61 62 module_param_named(send_queue_size, mad_sendq_size, int, 0444); 63 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests"); 64 module_param_named(recv_queue_size, mad_recvq_size, int, 0444); 65 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests"); 66 67 static struct list_head ib_mad_port_list; 68 static u32 ib_mad_client_id = 0; 69 70 /* Port list lock */ 71 static DEFINE_SPINLOCK(ib_mad_port_list_lock); 72 73 /* Forward declarations */ 74 static int method_in_use(struct ib_mad_mgmt_method_table **method, 75 struct ib_mad_reg_req *mad_reg_req); 76 static void remove_mad_reg_req(struct ib_mad_agent_private *priv); 77 static struct ib_mad_agent_private *find_mad_agent( 78 struct ib_mad_port_private *port_priv, 79 const struct ib_mad_hdr *mad); 80 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info, 81 struct ib_mad_private *mad); 82 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv); 83 static void timeout_sends(struct work_struct *work); 84 static void local_completions(struct work_struct *work); 85 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req, 86 struct ib_mad_agent_private *agent_priv, 87 u8 mgmt_class); 88 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req, 89 struct ib_mad_agent_private *agent_priv); 90 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv, 91 struct ib_wc *wc); 92 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc); 93 94 /* 95 * Returns a ib_mad_port_private structure or NULL for a device/port 96 * Assumes ib_mad_port_list_lock is being held 97 */ 98 static inline struct ib_mad_port_private * 99 __ib_get_mad_port(struct ib_device *device, int port_num) 100 { 101 struct ib_mad_port_private *entry; 102 103 list_for_each_entry(entry, &ib_mad_port_list, port_list) { 104 if (entry->device == device && entry->port_num == port_num) 105 return entry; 106 } 107 return NULL; 108 } 109 110 /* 111 * Wrapper function to return a ib_mad_port_private structure or NULL 112 * for a device/port 113 */ 114 static inline struct ib_mad_port_private * 115 ib_get_mad_port(struct ib_device *device, int port_num) 116 { 117 struct ib_mad_port_private *entry; 118 unsigned long flags; 119 120 spin_lock_irqsave(&ib_mad_port_list_lock, flags); 121 entry = __ib_get_mad_port(device, port_num); 122 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags); 123 124 return entry; 125 } 126 127 static inline u8 convert_mgmt_class(u8 mgmt_class) 128 { 129 /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */ 130 return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ? 131 0 : mgmt_class; 132 } 133 134 static int get_spl_qp_index(enum ib_qp_type qp_type) 135 { 136 switch (qp_type) 137 { 138 case IB_QPT_SMI: 139 return 0; 140 case IB_QPT_GSI: 141 return 1; 142 default: 143 return -1; 144 } 145 } 146 147 static int vendor_class_index(u8 mgmt_class) 148 { 149 return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START; 150 } 151 152 static int is_vendor_class(u8 mgmt_class) 153 { 154 if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) || 155 (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END)) 156 return 0; 157 return 1; 158 } 159 160 static int is_vendor_oui(char *oui) 161 { 162 if (oui[0] || oui[1] || oui[2]) 163 return 1; 164 return 0; 165 } 166 167 static int is_vendor_method_in_use( 168 struct ib_mad_mgmt_vendor_class *vendor_class, 169 struct ib_mad_reg_req *mad_reg_req) 170 { 171 struct ib_mad_mgmt_method_table *method; 172 int i; 173 174 for (i = 0; i < MAX_MGMT_OUI; i++) { 175 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) { 176 method = vendor_class->method_table[i]; 177 if (method) { 178 if (method_in_use(&method, mad_reg_req)) 179 return 1; 180 else 181 break; 182 } 183 } 184 } 185 return 0; 186 } 187 188 int ib_response_mad(const struct ib_mad_hdr *hdr) 189 { 190 return ((hdr->method & IB_MGMT_METHOD_RESP) || 191 (hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) || 192 ((hdr->mgmt_class == IB_MGMT_CLASS_BM) && 193 (hdr->attr_mod & IB_BM_ATTR_MOD_RESP))); 194 } 195 EXPORT_SYMBOL(ib_response_mad); 196 197 /* 198 * ib_register_mad_agent - Register to send/receive MADs 199 */ 200 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device, 201 u8 port_num, 202 enum ib_qp_type qp_type, 203 struct ib_mad_reg_req *mad_reg_req, 204 u8 rmpp_version, 205 ib_mad_send_handler send_handler, 206 ib_mad_recv_handler recv_handler, 207 void *context, 208 u32 registration_flags) 209 { 210 struct ib_mad_port_private *port_priv; 211 struct ib_mad_agent *ret = ERR_PTR(-EINVAL); 212 struct ib_mad_agent_private *mad_agent_priv; 213 struct ib_mad_reg_req *reg_req = NULL; 214 struct ib_mad_mgmt_class_table *class; 215 struct ib_mad_mgmt_vendor_class_table *vendor; 216 struct ib_mad_mgmt_vendor_class *vendor_class; 217 struct ib_mad_mgmt_method_table *method; 218 int ret2, qpn; 219 unsigned long flags; 220 u8 mgmt_class, vclass; 221 222 if ((qp_type == IB_QPT_SMI && !rdma_cap_ib_smi(device, port_num)) || 223 (qp_type == IB_QPT_GSI && !rdma_cap_ib_cm(device, port_num))) 224 return ERR_PTR(-EPROTONOSUPPORT); 225 226 /* Validate parameters */ 227 qpn = get_spl_qp_index(qp_type); 228 if (qpn == -1) { 229 dev_notice(&device->dev, 230 "ib_register_mad_agent: invalid QP Type %d\n", 231 qp_type); 232 goto error1; 233 } 234 235 if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) { 236 dev_notice(&device->dev, 237 "ib_register_mad_agent: invalid RMPP Version %u\n", 238 rmpp_version); 239 goto error1; 240 } 241 242 /* Validate MAD registration request if supplied */ 243 if (mad_reg_req) { 244 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) { 245 dev_notice(&device->dev, 246 "ib_register_mad_agent: invalid Class Version %u\n", 247 mad_reg_req->mgmt_class_version); 248 goto error1; 249 } 250 if (!recv_handler) { 251 dev_notice(&device->dev, 252 "ib_register_mad_agent: no recv_handler\n"); 253 goto error1; 254 } 255 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) { 256 /* 257 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only 258 * one in this range currently allowed 259 */ 260 if (mad_reg_req->mgmt_class != 261 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) { 262 dev_notice(&device->dev, 263 "ib_register_mad_agent: Invalid Mgmt Class 0x%x\n", 264 mad_reg_req->mgmt_class); 265 goto error1; 266 } 267 } else if (mad_reg_req->mgmt_class == 0) { 268 /* 269 * Class 0 is reserved in IBA and is used for 270 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE 271 */ 272 dev_notice(&device->dev, 273 "ib_register_mad_agent: Invalid Mgmt Class 0\n"); 274 goto error1; 275 } else if (is_vendor_class(mad_reg_req->mgmt_class)) { 276 /* 277 * If class is in "new" vendor range, 278 * ensure supplied OUI is not zero 279 */ 280 if (!is_vendor_oui(mad_reg_req->oui)) { 281 dev_notice(&device->dev, 282 "ib_register_mad_agent: No OUI specified for class 0x%x\n", 283 mad_reg_req->mgmt_class); 284 goto error1; 285 } 286 } 287 /* Make sure class supplied is consistent with RMPP */ 288 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) { 289 if (rmpp_version) { 290 dev_notice(&device->dev, 291 "ib_register_mad_agent: RMPP version for non-RMPP class 0x%x\n", 292 mad_reg_req->mgmt_class); 293 goto error1; 294 } 295 } 296 297 /* Make sure class supplied is consistent with QP type */ 298 if (qp_type == IB_QPT_SMI) { 299 if ((mad_reg_req->mgmt_class != 300 IB_MGMT_CLASS_SUBN_LID_ROUTED) && 301 (mad_reg_req->mgmt_class != 302 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) { 303 dev_notice(&device->dev, 304 "ib_register_mad_agent: Invalid SM QP type: class 0x%x\n", 305 mad_reg_req->mgmt_class); 306 goto error1; 307 } 308 } else { 309 if ((mad_reg_req->mgmt_class == 310 IB_MGMT_CLASS_SUBN_LID_ROUTED) || 311 (mad_reg_req->mgmt_class == 312 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) { 313 dev_notice(&device->dev, 314 "ib_register_mad_agent: Invalid GS QP type: class 0x%x\n", 315 mad_reg_req->mgmt_class); 316 goto error1; 317 } 318 } 319 } else { 320 /* No registration request supplied */ 321 if (!send_handler) 322 goto error1; 323 if (registration_flags & IB_MAD_USER_RMPP) 324 goto error1; 325 } 326 327 /* Validate device and port */ 328 port_priv = ib_get_mad_port(device, port_num); 329 if (!port_priv) { 330 dev_notice(&device->dev, "ib_register_mad_agent: Invalid port\n"); 331 ret = ERR_PTR(-ENODEV); 332 goto error1; 333 } 334 335 /* Verify the QP requested is supported. For example, Ethernet devices 336 * will not have QP0 */ 337 if (!port_priv->qp_info[qpn].qp) { 338 dev_notice(&device->dev, 339 "ib_register_mad_agent: QP %d not supported\n", qpn); 340 ret = ERR_PTR(-EPROTONOSUPPORT); 341 goto error1; 342 } 343 344 /* Allocate structures */ 345 mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL); 346 if (!mad_agent_priv) { 347 ret = ERR_PTR(-ENOMEM); 348 goto error1; 349 } 350 351 if (mad_reg_req) { 352 reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL); 353 if (!reg_req) { 354 ret = ERR_PTR(-ENOMEM); 355 goto error3; 356 } 357 } 358 359 /* Now, fill in the various structures */ 360 mad_agent_priv->qp_info = &port_priv->qp_info[qpn]; 361 mad_agent_priv->reg_req = reg_req; 362 mad_agent_priv->agent.rmpp_version = rmpp_version; 363 mad_agent_priv->agent.device = device; 364 mad_agent_priv->agent.recv_handler = recv_handler; 365 mad_agent_priv->agent.send_handler = send_handler; 366 mad_agent_priv->agent.context = context; 367 mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp; 368 mad_agent_priv->agent.port_num = port_num; 369 mad_agent_priv->agent.flags = registration_flags; 370 spin_lock_init(&mad_agent_priv->lock); 371 INIT_LIST_HEAD(&mad_agent_priv->send_list); 372 INIT_LIST_HEAD(&mad_agent_priv->wait_list); 373 INIT_LIST_HEAD(&mad_agent_priv->done_list); 374 INIT_LIST_HEAD(&mad_agent_priv->rmpp_list); 375 INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends); 376 INIT_LIST_HEAD(&mad_agent_priv->local_list); 377 INIT_WORK(&mad_agent_priv->local_work, local_completions); 378 atomic_set(&mad_agent_priv->refcount, 1); 379 init_completion(&mad_agent_priv->comp); 380 381 spin_lock_irqsave(&port_priv->reg_lock, flags); 382 mad_agent_priv->agent.hi_tid = ++ib_mad_client_id; 383 384 /* 385 * Make sure MAD registration (if supplied) 386 * is non overlapping with any existing ones 387 */ 388 if (mad_reg_req) { 389 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class); 390 if (!is_vendor_class(mgmt_class)) { 391 class = port_priv->version[mad_reg_req-> 392 mgmt_class_version].class; 393 if (class) { 394 method = class->method_table[mgmt_class]; 395 if (method) { 396 if (method_in_use(&method, 397 mad_reg_req)) 398 goto error4; 399 } 400 } 401 ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv, 402 mgmt_class); 403 } else { 404 /* "New" vendor class range */ 405 vendor = port_priv->version[mad_reg_req-> 406 mgmt_class_version].vendor; 407 if (vendor) { 408 vclass = vendor_class_index(mgmt_class); 409 vendor_class = vendor->vendor_class[vclass]; 410 if (vendor_class) { 411 if (is_vendor_method_in_use( 412 vendor_class, 413 mad_reg_req)) 414 goto error4; 415 } 416 } 417 ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv); 418 } 419 if (ret2) { 420 ret = ERR_PTR(ret2); 421 goto error4; 422 } 423 } 424 425 /* Add mad agent into port's agent list */ 426 list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list); 427 spin_unlock_irqrestore(&port_priv->reg_lock, flags); 428 429 return &mad_agent_priv->agent; 430 431 error4: 432 spin_unlock_irqrestore(&port_priv->reg_lock, flags); 433 kfree(reg_req); 434 error3: 435 kfree(mad_agent_priv); 436 error1: 437 return ret; 438 } 439 EXPORT_SYMBOL(ib_register_mad_agent); 440 441 static inline int is_snooping_sends(int mad_snoop_flags) 442 { 443 return (mad_snoop_flags & 444 (/*IB_MAD_SNOOP_POSTED_SENDS | 445 IB_MAD_SNOOP_RMPP_SENDS |*/ 446 IB_MAD_SNOOP_SEND_COMPLETIONS /*| 447 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/)); 448 } 449 450 static inline int is_snooping_recvs(int mad_snoop_flags) 451 { 452 return (mad_snoop_flags & 453 (IB_MAD_SNOOP_RECVS /*| 454 IB_MAD_SNOOP_RMPP_RECVS*/)); 455 } 456 457 static int register_snoop_agent(struct ib_mad_qp_info *qp_info, 458 struct ib_mad_snoop_private *mad_snoop_priv) 459 { 460 struct ib_mad_snoop_private **new_snoop_table; 461 unsigned long flags; 462 int i; 463 464 spin_lock_irqsave(&qp_info->snoop_lock, flags); 465 /* Check for empty slot in array. */ 466 for (i = 0; i < qp_info->snoop_table_size; i++) 467 if (!qp_info->snoop_table[i]) 468 break; 469 470 if (i == qp_info->snoop_table_size) { 471 /* Grow table. */ 472 new_snoop_table = krealloc(qp_info->snoop_table, 473 sizeof mad_snoop_priv * 474 (qp_info->snoop_table_size + 1), 475 GFP_ATOMIC); 476 if (!new_snoop_table) { 477 i = -ENOMEM; 478 goto out; 479 } 480 481 qp_info->snoop_table = new_snoop_table; 482 qp_info->snoop_table_size++; 483 } 484 qp_info->snoop_table[i] = mad_snoop_priv; 485 atomic_inc(&qp_info->snoop_count); 486 out: 487 spin_unlock_irqrestore(&qp_info->snoop_lock, flags); 488 return i; 489 } 490 491 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device, 492 u8 port_num, 493 enum ib_qp_type qp_type, 494 int mad_snoop_flags, 495 ib_mad_snoop_handler snoop_handler, 496 ib_mad_recv_handler recv_handler, 497 void *context) 498 { 499 struct ib_mad_port_private *port_priv; 500 struct ib_mad_agent *ret; 501 struct ib_mad_snoop_private *mad_snoop_priv; 502 int qpn; 503 504 /* Validate parameters */ 505 if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) || 506 (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) { 507 ret = ERR_PTR(-EINVAL); 508 goto error1; 509 } 510 qpn = get_spl_qp_index(qp_type); 511 if (qpn == -1) { 512 ret = ERR_PTR(-EINVAL); 513 goto error1; 514 } 515 port_priv = ib_get_mad_port(device, port_num); 516 if (!port_priv) { 517 ret = ERR_PTR(-ENODEV); 518 goto error1; 519 } 520 /* Allocate structures */ 521 mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL); 522 if (!mad_snoop_priv) { 523 ret = ERR_PTR(-ENOMEM); 524 goto error1; 525 } 526 527 /* Now, fill in the various structures */ 528 mad_snoop_priv->qp_info = &port_priv->qp_info[qpn]; 529 mad_snoop_priv->agent.device = device; 530 mad_snoop_priv->agent.recv_handler = recv_handler; 531 mad_snoop_priv->agent.snoop_handler = snoop_handler; 532 mad_snoop_priv->agent.context = context; 533 mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp; 534 mad_snoop_priv->agent.port_num = port_num; 535 mad_snoop_priv->mad_snoop_flags = mad_snoop_flags; 536 init_completion(&mad_snoop_priv->comp); 537 mad_snoop_priv->snoop_index = register_snoop_agent( 538 &port_priv->qp_info[qpn], 539 mad_snoop_priv); 540 if (mad_snoop_priv->snoop_index < 0) { 541 ret = ERR_PTR(mad_snoop_priv->snoop_index); 542 goto error2; 543 } 544 545 atomic_set(&mad_snoop_priv->refcount, 1); 546 return &mad_snoop_priv->agent; 547 548 error2: 549 kfree(mad_snoop_priv); 550 error1: 551 return ret; 552 } 553 EXPORT_SYMBOL(ib_register_mad_snoop); 554 555 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv) 556 { 557 if (atomic_dec_and_test(&mad_agent_priv->refcount)) 558 complete(&mad_agent_priv->comp); 559 } 560 561 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv) 562 { 563 if (atomic_dec_and_test(&mad_snoop_priv->refcount)) 564 complete(&mad_snoop_priv->comp); 565 } 566 567 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv) 568 { 569 struct ib_mad_port_private *port_priv; 570 unsigned long flags; 571 572 /* Note that we could still be handling received MADs */ 573 574 /* 575 * Canceling all sends results in dropping received response 576 * MADs, preventing us from queuing additional work 577 */ 578 cancel_mads(mad_agent_priv); 579 port_priv = mad_agent_priv->qp_info->port_priv; 580 cancel_delayed_work_sync(&mad_agent_priv->timed_work); 581 582 spin_lock_irqsave(&port_priv->reg_lock, flags); 583 remove_mad_reg_req(mad_agent_priv); 584 list_del(&mad_agent_priv->agent_list); 585 spin_unlock_irqrestore(&port_priv->reg_lock, flags); 586 587 flush_workqueue(port_priv->wq); 588 ib_cancel_rmpp_recvs(mad_agent_priv); 589 590 deref_mad_agent(mad_agent_priv); 591 wait_for_completion(&mad_agent_priv->comp); 592 593 kfree(mad_agent_priv->reg_req); 594 kfree(mad_agent_priv); 595 } 596 597 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv) 598 { 599 struct ib_mad_qp_info *qp_info; 600 unsigned long flags; 601 602 qp_info = mad_snoop_priv->qp_info; 603 spin_lock_irqsave(&qp_info->snoop_lock, flags); 604 qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL; 605 atomic_dec(&qp_info->snoop_count); 606 spin_unlock_irqrestore(&qp_info->snoop_lock, flags); 607 608 deref_snoop_agent(mad_snoop_priv); 609 wait_for_completion(&mad_snoop_priv->comp); 610 611 kfree(mad_snoop_priv); 612 } 613 614 /* 615 * ib_unregister_mad_agent - Unregisters a client from using MAD services 616 */ 617 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent) 618 { 619 struct ib_mad_agent_private *mad_agent_priv; 620 struct ib_mad_snoop_private *mad_snoop_priv; 621 622 /* If the TID is zero, the agent can only snoop. */ 623 if (mad_agent->hi_tid) { 624 mad_agent_priv = container_of(mad_agent, 625 struct ib_mad_agent_private, 626 agent); 627 unregister_mad_agent(mad_agent_priv); 628 } else { 629 mad_snoop_priv = container_of(mad_agent, 630 struct ib_mad_snoop_private, 631 agent); 632 unregister_mad_snoop(mad_snoop_priv); 633 } 634 return 0; 635 } 636 EXPORT_SYMBOL(ib_unregister_mad_agent); 637 638 static void dequeue_mad(struct ib_mad_list_head *mad_list) 639 { 640 struct ib_mad_queue *mad_queue; 641 unsigned long flags; 642 643 BUG_ON(!mad_list->mad_queue); 644 mad_queue = mad_list->mad_queue; 645 spin_lock_irqsave(&mad_queue->lock, flags); 646 list_del(&mad_list->list); 647 mad_queue->count--; 648 spin_unlock_irqrestore(&mad_queue->lock, flags); 649 } 650 651 static void snoop_send(struct ib_mad_qp_info *qp_info, 652 struct ib_mad_send_buf *send_buf, 653 struct ib_mad_send_wc *mad_send_wc, 654 int mad_snoop_flags) 655 { 656 struct ib_mad_snoop_private *mad_snoop_priv; 657 unsigned long flags; 658 int i; 659 660 spin_lock_irqsave(&qp_info->snoop_lock, flags); 661 for (i = 0; i < qp_info->snoop_table_size; i++) { 662 mad_snoop_priv = qp_info->snoop_table[i]; 663 if (!mad_snoop_priv || 664 !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags)) 665 continue; 666 667 atomic_inc(&mad_snoop_priv->refcount); 668 spin_unlock_irqrestore(&qp_info->snoop_lock, flags); 669 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent, 670 send_buf, mad_send_wc); 671 deref_snoop_agent(mad_snoop_priv); 672 spin_lock_irqsave(&qp_info->snoop_lock, flags); 673 } 674 spin_unlock_irqrestore(&qp_info->snoop_lock, flags); 675 } 676 677 static void snoop_recv(struct ib_mad_qp_info *qp_info, 678 struct ib_mad_recv_wc *mad_recv_wc, 679 int mad_snoop_flags) 680 { 681 struct ib_mad_snoop_private *mad_snoop_priv; 682 unsigned long flags; 683 int i; 684 685 spin_lock_irqsave(&qp_info->snoop_lock, flags); 686 for (i = 0; i < qp_info->snoop_table_size; i++) { 687 mad_snoop_priv = qp_info->snoop_table[i]; 688 if (!mad_snoop_priv || 689 !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags)) 690 continue; 691 692 atomic_inc(&mad_snoop_priv->refcount); 693 spin_unlock_irqrestore(&qp_info->snoop_lock, flags); 694 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL, 695 mad_recv_wc); 696 deref_snoop_agent(mad_snoop_priv); 697 spin_lock_irqsave(&qp_info->snoop_lock, flags); 698 } 699 spin_unlock_irqrestore(&qp_info->snoop_lock, flags); 700 } 701 702 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid, 703 u16 pkey_index, u8 port_num, struct ib_wc *wc) 704 { 705 memset(wc, 0, sizeof *wc); 706 wc->wr_cqe = cqe; 707 wc->status = IB_WC_SUCCESS; 708 wc->opcode = IB_WC_RECV; 709 wc->pkey_index = pkey_index; 710 wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh); 711 wc->src_qp = IB_QP0; 712 wc->qp = qp; 713 wc->slid = slid; 714 wc->sl = 0; 715 wc->dlid_path_bits = 0; 716 wc->port_num = port_num; 717 } 718 719 static size_t mad_priv_size(const struct ib_mad_private *mp) 720 { 721 return sizeof(struct ib_mad_private) + mp->mad_size; 722 } 723 724 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags) 725 { 726 size_t size = sizeof(struct ib_mad_private) + mad_size; 727 struct ib_mad_private *ret = kzalloc(size, flags); 728 729 if (ret) 730 ret->mad_size = mad_size; 731 732 return ret; 733 } 734 735 static size_t port_mad_size(const struct ib_mad_port_private *port_priv) 736 { 737 return rdma_max_mad_size(port_priv->device, port_priv->port_num); 738 } 739 740 static size_t mad_priv_dma_size(const struct ib_mad_private *mp) 741 { 742 return sizeof(struct ib_grh) + mp->mad_size; 743 } 744 745 /* 746 * Return 0 if SMP is to be sent 747 * Return 1 if SMP was consumed locally (whether or not solicited) 748 * Return < 0 if error 749 */ 750 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv, 751 struct ib_mad_send_wr_private *mad_send_wr) 752 { 753 int ret = 0; 754 struct ib_smp *smp = mad_send_wr->send_buf.mad; 755 struct opa_smp *opa_smp = (struct opa_smp *)smp; 756 unsigned long flags; 757 struct ib_mad_local_private *local; 758 struct ib_mad_private *mad_priv; 759 struct ib_mad_port_private *port_priv; 760 struct ib_mad_agent_private *recv_mad_agent = NULL; 761 struct ib_device *device = mad_agent_priv->agent.device; 762 u8 port_num; 763 struct ib_wc mad_wc; 764 struct ib_ud_wr *send_wr = &mad_send_wr->send_wr; 765 size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv); 766 u16 out_mad_pkey_index = 0; 767 u16 drslid; 768 bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device, 769 mad_agent_priv->qp_info->port_priv->port_num); 770 771 if (rdma_cap_ib_switch(device) && 772 smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) 773 port_num = send_wr->port_num; 774 else 775 port_num = mad_agent_priv->agent.port_num; 776 777 /* 778 * Directed route handling starts if the initial LID routed part of 779 * a request or the ending LID routed part of a response is empty. 780 * If we are at the start of the LID routed part, don't update the 781 * hop_ptr or hop_cnt. See section 14.2.2, Vol 1 IB spec. 782 */ 783 if (opa && smp->class_version == OPA_SMP_CLASS_VERSION) { 784 u32 opa_drslid; 785 786 if ((opa_get_smp_direction(opa_smp) 787 ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) == 788 OPA_LID_PERMISSIVE && 789 opa_smi_handle_dr_smp_send(opa_smp, 790 rdma_cap_ib_switch(device), 791 port_num) == IB_SMI_DISCARD) { 792 ret = -EINVAL; 793 dev_err(&device->dev, "OPA Invalid directed route\n"); 794 goto out; 795 } 796 opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid); 797 if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) && 798 opa_drslid & 0xffff0000) { 799 ret = -EINVAL; 800 dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n", 801 opa_drslid); 802 goto out; 803 } 804 drslid = (u16)(opa_drslid & 0x0000ffff); 805 806 /* Check to post send on QP or process locally */ 807 if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD && 808 opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD) 809 goto out; 810 } else { 811 if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) == 812 IB_LID_PERMISSIVE && 813 smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) == 814 IB_SMI_DISCARD) { 815 ret = -EINVAL; 816 dev_err(&device->dev, "Invalid directed route\n"); 817 goto out; 818 } 819 drslid = be16_to_cpu(smp->dr_slid); 820 821 /* Check to post send on QP or process locally */ 822 if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD && 823 smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD) 824 goto out; 825 } 826 827 local = kmalloc(sizeof *local, GFP_ATOMIC); 828 if (!local) { 829 ret = -ENOMEM; 830 goto out; 831 } 832 local->mad_priv = NULL; 833 local->recv_mad_agent = NULL; 834 mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC); 835 if (!mad_priv) { 836 ret = -ENOMEM; 837 kfree(local); 838 goto out; 839 } 840 841 build_smp_wc(mad_agent_priv->agent.qp, 842 send_wr->wr.wr_cqe, drslid, 843 send_wr->pkey_index, 844 send_wr->port_num, &mad_wc); 845 846 if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) { 847 mad_wc.byte_len = mad_send_wr->send_buf.hdr_len 848 + mad_send_wr->send_buf.data_len 849 + sizeof(struct ib_grh); 850 } 851 852 /* No GRH for DR SMP */ 853 ret = device->process_mad(device, 0, port_num, &mad_wc, NULL, 854 (const struct ib_mad_hdr *)smp, mad_size, 855 (struct ib_mad_hdr *)mad_priv->mad, 856 &mad_size, &out_mad_pkey_index); 857 switch (ret) 858 { 859 case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY: 860 if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) && 861 mad_agent_priv->agent.recv_handler) { 862 local->mad_priv = mad_priv; 863 local->recv_mad_agent = mad_agent_priv; 864 /* 865 * Reference MAD agent until receive 866 * side of local completion handled 867 */ 868 atomic_inc(&mad_agent_priv->refcount); 869 } else 870 kfree(mad_priv); 871 break; 872 case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED: 873 kfree(mad_priv); 874 break; 875 case IB_MAD_RESULT_SUCCESS: 876 /* Treat like an incoming receive MAD */ 877 port_priv = ib_get_mad_port(mad_agent_priv->agent.device, 878 mad_agent_priv->agent.port_num); 879 if (port_priv) { 880 memcpy(mad_priv->mad, smp, mad_priv->mad_size); 881 recv_mad_agent = find_mad_agent(port_priv, 882 (const struct ib_mad_hdr *)mad_priv->mad); 883 } 884 if (!port_priv || !recv_mad_agent) { 885 /* 886 * No receiving agent so drop packet and 887 * generate send completion. 888 */ 889 kfree(mad_priv); 890 break; 891 } 892 local->mad_priv = mad_priv; 893 local->recv_mad_agent = recv_mad_agent; 894 break; 895 default: 896 kfree(mad_priv); 897 kfree(local); 898 ret = -EINVAL; 899 goto out; 900 } 901 902 local->mad_send_wr = mad_send_wr; 903 if (opa) { 904 local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index; 905 local->return_wc_byte_len = mad_size; 906 } 907 /* Reference MAD agent until send side of local completion handled */ 908 atomic_inc(&mad_agent_priv->refcount); 909 /* Queue local completion to local list */ 910 spin_lock_irqsave(&mad_agent_priv->lock, flags); 911 list_add_tail(&local->completion_list, &mad_agent_priv->local_list); 912 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 913 queue_work(mad_agent_priv->qp_info->port_priv->wq, 914 &mad_agent_priv->local_work); 915 ret = 1; 916 out: 917 return ret; 918 } 919 920 static int get_pad_size(int hdr_len, int data_len, size_t mad_size) 921 { 922 int seg_size, pad; 923 924 seg_size = mad_size - hdr_len; 925 if (data_len && seg_size) { 926 pad = seg_size - data_len % seg_size; 927 return pad == seg_size ? 0 : pad; 928 } else 929 return seg_size; 930 } 931 932 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr) 933 { 934 struct ib_rmpp_segment *s, *t; 935 936 list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) { 937 list_del(&s->list); 938 kfree(s); 939 } 940 } 941 942 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr, 943 size_t mad_size, gfp_t gfp_mask) 944 { 945 struct ib_mad_send_buf *send_buf = &send_wr->send_buf; 946 struct ib_rmpp_mad *rmpp_mad = send_buf->mad; 947 struct ib_rmpp_segment *seg = NULL; 948 int left, seg_size, pad; 949 950 send_buf->seg_size = mad_size - send_buf->hdr_len; 951 send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR; 952 seg_size = send_buf->seg_size; 953 pad = send_wr->pad; 954 955 /* Allocate data segments. */ 956 for (left = send_buf->data_len + pad; left > 0; left -= seg_size) { 957 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask); 958 if (!seg) { 959 free_send_rmpp_list(send_wr); 960 return -ENOMEM; 961 } 962 seg->num = ++send_buf->seg_count; 963 list_add_tail(&seg->list, &send_wr->rmpp_list); 964 } 965 966 /* Zero any padding */ 967 if (pad) 968 memset(seg->data + seg_size - pad, 0, pad); 969 970 rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv-> 971 agent.rmpp_version; 972 rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA; 973 ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE); 974 975 send_wr->cur_seg = container_of(send_wr->rmpp_list.next, 976 struct ib_rmpp_segment, list); 977 send_wr->last_ack_seg = send_wr->cur_seg; 978 return 0; 979 } 980 981 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent) 982 { 983 return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP); 984 } 985 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent); 986 987 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent, 988 u32 remote_qpn, u16 pkey_index, 989 int rmpp_active, 990 int hdr_len, int data_len, 991 gfp_t gfp_mask, 992 u8 base_version) 993 { 994 struct ib_mad_agent_private *mad_agent_priv; 995 struct ib_mad_send_wr_private *mad_send_wr; 996 int pad, message_size, ret, size; 997 void *buf; 998 size_t mad_size; 999 bool opa; 1000 1001 mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private, 1002 agent); 1003 1004 opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num); 1005 1006 if (opa && base_version == OPA_MGMT_BASE_VERSION) 1007 mad_size = sizeof(struct opa_mad); 1008 else 1009 mad_size = sizeof(struct ib_mad); 1010 1011 pad = get_pad_size(hdr_len, data_len, mad_size); 1012 message_size = hdr_len + data_len + pad; 1013 1014 if (ib_mad_kernel_rmpp_agent(mad_agent)) { 1015 if (!rmpp_active && message_size > mad_size) 1016 return ERR_PTR(-EINVAL); 1017 } else 1018 if (rmpp_active || message_size > mad_size) 1019 return ERR_PTR(-EINVAL); 1020 1021 size = rmpp_active ? hdr_len : mad_size; 1022 buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask); 1023 if (!buf) 1024 return ERR_PTR(-ENOMEM); 1025 1026 mad_send_wr = (struct ib_mad_send_wr_private *)((char *)buf + size); 1027 INIT_LIST_HEAD(&mad_send_wr->rmpp_list); 1028 mad_send_wr->send_buf.mad = buf; 1029 mad_send_wr->send_buf.hdr_len = hdr_len; 1030 mad_send_wr->send_buf.data_len = data_len; 1031 mad_send_wr->pad = pad; 1032 1033 mad_send_wr->mad_agent_priv = mad_agent_priv; 1034 mad_send_wr->sg_list[0].length = hdr_len; 1035 mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey; 1036 1037 /* OPA MADs don't have to be the full 2048 bytes */ 1038 if (opa && base_version == OPA_MGMT_BASE_VERSION && 1039 data_len < mad_size - hdr_len) 1040 mad_send_wr->sg_list[1].length = data_len; 1041 else 1042 mad_send_wr->sg_list[1].length = mad_size - hdr_len; 1043 1044 mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey; 1045 1046 mad_send_wr->mad_list.cqe.done = ib_mad_send_done; 1047 1048 mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe; 1049 mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list; 1050 mad_send_wr->send_wr.wr.num_sge = 2; 1051 mad_send_wr->send_wr.wr.opcode = IB_WR_SEND; 1052 mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED; 1053 mad_send_wr->send_wr.remote_qpn = remote_qpn; 1054 mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY; 1055 mad_send_wr->send_wr.pkey_index = pkey_index; 1056 1057 if (rmpp_active) { 1058 ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask); 1059 if (ret) { 1060 kfree(buf); 1061 return ERR_PTR(ret); 1062 } 1063 } 1064 1065 mad_send_wr->send_buf.mad_agent = mad_agent; 1066 atomic_inc(&mad_agent_priv->refcount); 1067 return &mad_send_wr->send_buf; 1068 } 1069 EXPORT_SYMBOL(ib_create_send_mad); 1070 1071 int ib_get_mad_data_offset(u8 mgmt_class) 1072 { 1073 if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM) 1074 return IB_MGMT_SA_HDR; 1075 else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) || 1076 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) || 1077 (mgmt_class == IB_MGMT_CLASS_BIS)) 1078 return IB_MGMT_DEVICE_HDR; 1079 else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) && 1080 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)) 1081 return IB_MGMT_VENDOR_HDR; 1082 else 1083 return IB_MGMT_MAD_HDR; 1084 } 1085 EXPORT_SYMBOL(ib_get_mad_data_offset); 1086 1087 int ib_is_mad_class_rmpp(u8 mgmt_class) 1088 { 1089 if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) || 1090 (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) || 1091 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) || 1092 (mgmt_class == IB_MGMT_CLASS_BIS) || 1093 ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) && 1094 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))) 1095 return 1; 1096 return 0; 1097 } 1098 EXPORT_SYMBOL(ib_is_mad_class_rmpp); 1099 1100 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num) 1101 { 1102 struct ib_mad_send_wr_private *mad_send_wr; 1103 struct list_head *list; 1104 1105 mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private, 1106 send_buf); 1107 list = &mad_send_wr->cur_seg->list; 1108 1109 if (mad_send_wr->cur_seg->num < seg_num) { 1110 list_for_each_entry(mad_send_wr->cur_seg, list, list) 1111 if (mad_send_wr->cur_seg->num == seg_num) 1112 break; 1113 } else if (mad_send_wr->cur_seg->num > seg_num) { 1114 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list) 1115 if (mad_send_wr->cur_seg->num == seg_num) 1116 break; 1117 } 1118 return mad_send_wr->cur_seg->data; 1119 } 1120 EXPORT_SYMBOL(ib_get_rmpp_segment); 1121 1122 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr) 1123 { 1124 if (mad_send_wr->send_buf.seg_count) 1125 return ib_get_rmpp_segment(&mad_send_wr->send_buf, 1126 mad_send_wr->seg_num); 1127 else 1128 return (char *)mad_send_wr->send_buf.mad + 1129 mad_send_wr->send_buf.hdr_len; 1130 } 1131 1132 void ib_free_send_mad(struct ib_mad_send_buf *send_buf) 1133 { 1134 struct ib_mad_agent_private *mad_agent_priv; 1135 struct ib_mad_send_wr_private *mad_send_wr; 1136 1137 mad_agent_priv = container_of(send_buf->mad_agent, 1138 struct ib_mad_agent_private, agent); 1139 mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private, 1140 send_buf); 1141 1142 free_send_rmpp_list(mad_send_wr); 1143 kfree(send_buf->mad); 1144 deref_mad_agent(mad_agent_priv); 1145 } 1146 EXPORT_SYMBOL(ib_free_send_mad); 1147 1148 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr) 1149 { 1150 struct ib_mad_qp_info *qp_info; 1151 struct list_head *list; 1152 const struct ib_send_wr *bad_send_wr; 1153 struct ib_mad_agent *mad_agent; 1154 struct ib_sge *sge; 1155 unsigned long flags; 1156 int ret; 1157 1158 /* Set WR ID to find mad_send_wr upon completion */ 1159 qp_info = mad_send_wr->mad_agent_priv->qp_info; 1160 mad_send_wr->mad_list.mad_queue = &qp_info->send_queue; 1161 mad_send_wr->mad_list.cqe.done = ib_mad_send_done; 1162 mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe; 1163 1164 mad_agent = mad_send_wr->send_buf.mad_agent; 1165 sge = mad_send_wr->sg_list; 1166 sge[0].addr = ib_dma_map_single(mad_agent->device, 1167 mad_send_wr->send_buf.mad, 1168 sge[0].length, 1169 DMA_TO_DEVICE); 1170 if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr))) 1171 return -ENOMEM; 1172 1173 mad_send_wr->header_mapping = sge[0].addr; 1174 1175 sge[1].addr = ib_dma_map_single(mad_agent->device, 1176 ib_get_payload(mad_send_wr), 1177 sge[1].length, 1178 DMA_TO_DEVICE); 1179 if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) { 1180 ib_dma_unmap_single(mad_agent->device, 1181 mad_send_wr->header_mapping, 1182 sge[0].length, DMA_TO_DEVICE); 1183 return -ENOMEM; 1184 } 1185 mad_send_wr->payload_mapping = sge[1].addr; 1186 1187 spin_lock_irqsave(&qp_info->send_queue.lock, flags); 1188 if (qp_info->send_queue.count < qp_info->send_queue.max_active) { 1189 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr, 1190 &bad_send_wr); 1191 list = &qp_info->send_queue.list; 1192 } else { 1193 ret = 0; 1194 list = &qp_info->overflow_list; 1195 } 1196 1197 if (!ret) { 1198 qp_info->send_queue.count++; 1199 list_add_tail(&mad_send_wr->mad_list.list, list); 1200 } 1201 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags); 1202 if (ret) { 1203 ib_dma_unmap_single(mad_agent->device, 1204 mad_send_wr->header_mapping, 1205 sge[0].length, DMA_TO_DEVICE); 1206 ib_dma_unmap_single(mad_agent->device, 1207 mad_send_wr->payload_mapping, 1208 sge[1].length, DMA_TO_DEVICE); 1209 } 1210 return ret; 1211 } 1212 1213 /* 1214 * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated 1215 * with the registered client 1216 */ 1217 int ib_post_send_mad(struct ib_mad_send_buf *send_buf, 1218 struct ib_mad_send_buf **bad_send_buf) 1219 { 1220 struct ib_mad_agent_private *mad_agent_priv; 1221 struct ib_mad_send_buf *next_send_buf; 1222 struct ib_mad_send_wr_private *mad_send_wr; 1223 unsigned long flags; 1224 int ret = -EINVAL; 1225 1226 /* Walk list of send WRs and post each on send list */ 1227 for (; send_buf; send_buf = next_send_buf) { 1228 1229 mad_send_wr = container_of(send_buf, 1230 struct ib_mad_send_wr_private, 1231 send_buf); 1232 mad_agent_priv = mad_send_wr->mad_agent_priv; 1233 1234 if (!send_buf->mad_agent->send_handler || 1235 (send_buf->timeout_ms && 1236 !send_buf->mad_agent->recv_handler)) { 1237 ret = -EINVAL; 1238 goto error; 1239 } 1240 1241 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) { 1242 if (mad_agent_priv->agent.rmpp_version) { 1243 ret = -EINVAL; 1244 goto error; 1245 } 1246 } 1247 1248 /* 1249 * Save pointer to next work request to post in case the 1250 * current one completes, and the user modifies the work 1251 * request associated with the completion 1252 */ 1253 next_send_buf = send_buf->next; 1254 mad_send_wr->send_wr.ah = send_buf->ah; 1255 1256 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class == 1257 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) { 1258 ret = handle_outgoing_dr_smp(mad_agent_priv, 1259 mad_send_wr); 1260 if (ret < 0) /* error */ 1261 goto error; 1262 else if (ret == 1) /* locally consumed */ 1263 continue; 1264 } 1265 1266 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid; 1267 /* Timeout will be updated after send completes */ 1268 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms); 1269 mad_send_wr->max_retries = send_buf->retries; 1270 mad_send_wr->retries_left = send_buf->retries; 1271 send_buf->retries = 0; 1272 /* Reference for work request to QP + response */ 1273 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0); 1274 mad_send_wr->status = IB_WC_SUCCESS; 1275 1276 /* Reference MAD agent until send completes */ 1277 atomic_inc(&mad_agent_priv->refcount); 1278 spin_lock_irqsave(&mad_agent_priv->lock, flags); 1279 list_add_tail(&mad_send_wr->agent_list, 1280 &mad_agent_priv->send_list); 1281 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 1282 1283 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) { 1284 ret = ib_send_rmpp_mad(mad_send_wr); 1285 if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED) 1286 ret = ib_send_mad(mad_send_wr); 1287 } else 1288 ret = ib_send_mad(mad_send_wr); 1289 if (ret < 0) { 1290 /* Fail send request */ 1291 spin_lock_irqsave(&mad_agent_priv->lock, flags); 1292 list_del(&mad_send_wr->agent_list); 1293 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 1294 atomic_dec(&mad_agent_priv->refcount); 1295 goto error; 1296 } 1297 } 1298 return 0; 1299 error: 1300 if (bad_send_buf) 1301 *bad_send_buf = send_buf; 1302 return ret; 1303 } 1304 EXPORT_SYMBOL(ib_post_send_mad); 1305 1306 /* 1307 * ib_free_recv_mad - Returns data buffers used to receive 1308 * a MAD to the access layer 1309 */ 1310 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc) 1311 { 1312 struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf; 1313 struct ib_mad_private_header *mad_priv_hdr; 1314 struct ib_mad_private *priv; 1315 struct list_head free_list; 1316 1317 INIT_LIST_HEAD(&free_list); 1318 list_splice_init(&mad_recv_wc->rmpp_list, &free_list); 1319 1320 list_for_each_entry_safe(mad_recv_buf, temp_recv_buf, 1321 &free_list, list) { 1322 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc, 1323 recv_buf); 1324 mad_priv_hdr = container_of(mad_recv_wc, 1325 struct ib_mad_private_header, 1326 recv_wc); 1327 priv = container_of(mad_priv_hdr, struct ib_mad_private, 1328 header); 1329 kfree(priv); 1330 } 1331 } 1332 EXPORT_SYMBOL(ib_free_recv_mad); 1333 1334 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp, 1335 u8 rmpp_version, 1336 ib_mad_send_handler send_handler, 1337 ib_mad_recv_handler recv_handler, 1338 void *context) 1339 { 1340 return ERR_PTR(-EINVAL); /* XXX: for now */ 1341 } 1342 EXPORT_SYMBOL(ib_redirect_mad_qp); 1343 1344 int ib_process_mad_wc(struct ib_mad_agent *mad_agent, 1345 struct ib_wc *wc) 1346 { 1347 dev_err(&mad_agent->device->dev, 1348 "ib_process_mad_wc() not implemented yet\n"); 1349 return 0; 1350 } 1351 EXPORT_SYMBOL(ib_process_mad_wc); 1352 1353 static int method_in_use(struct ib_mad_mgmt_method_table **method, 1354 struct ib_mad_reg_req *mad_reg_req) 1355 { 1356 int i; 1357 1358 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) { 1359 if ((*method)->agent[i]) { 1360 pr_err("Method %d already in use\n", i); 1361 return -EINVAL; 1362 } 1363 } 1364 return 0; 1365 } 1366 1367 static int allocate_method_table(struct ib_mad_mgmt_method_table **method) 1368 { 1369 /* Allocate management method table */ 1370 *method = kzalloc(sizeof **method, GFP_ATOMIC); 1371 return (*method) ? 0 : (-ENOMEM); 1372 } 1373 1374 /* 1375 * Check to see if there are any methods still in use 1376 */ 1377 static int check_method_table(struct ib_mad_mgmt_method_table *method) 1378 { 1379 int i; 1380 1381 for (i = 0; i < IB_MGMT_MAX_METHODS; i++) 1382 if (method->agent[i]) 1383 return 1; 1384 return 0; 1385 } 1386 1387 /* 1388 * Check to see if there are any method tables for this class still in use 1389 */ 1390 static int check_class_table(struct ib_mad_mgmt_class_table *class) 1391 { 1392 int i; 1393 1394 for (i = 0; i < MAX_MGMT_CLASS; i++) 1395 if (class->method_table[i]) 1396 return 1; 1397 return 0; 1398 } 1399 1400 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class) 1401 { 1402 int i; 1403 1404 for (i = 0; i < MAX_MGMT_OUI; i++) 1405 if (vendor_class->method_table[i]) 1406 return 1; 1407 return 0; 1408 } 1409 1410 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class, 1411 const char *oui) 1412 { 1413 int i; 1414 1415 for (i = 0; i < MAX_MGMT_OUI; i++) 1416 /* Is there matching OUI for this vendor class ? */ 1417 if (!memcmp(vendor_class->oui[i], oui, 3)) 1418 return i; 1419 1420 return -1; 1421 } 1422 1423 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor) 1424 { 1425 int i; 1426 1427 for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++) 1428 if (vendor->vendor_class[i]) 1429 return 1; 1430 1431 return 0; 1432 } 1433 1434 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method, 1435 struct ib_mad_agent_private *agent) 1436 { 1437 int i; 1438 1439 /* Remove any methods for this mad agent */ 1440 for (i = 0; i < IB_MGMT_MAX_METHODS; i++) { 1441 if (method->agent[i] == agent) { 1442 method->agent[i] = NULL; 1443 } 1444 } 1445 } 1446 1447 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req, 1448 struct ib_mad_agent_private *agent_priv, 1449 u8 mgmt_class) 1450 { 1451 struct ib_mad_port_private *port_priv; 1452 struct ib_mad_mgmt_class_table **class; 1453 struct ib_mad_mgmt_method_table **method; 1454 int i, ret; 1455 1456 port_priv = agent_priv->qp_info->port_priv; 1457 class = &port_priv->version[mad_reg_req->mgmt_class_version].class; 1458 if (!*class) { 1459 /* Allocate management class table for "new" class version */ 1460 *class = kzalloc(sizeof **class, GFP_ATOMIC); 1461 if (!*class) { 1462 ret = -ENOMEM; 1463 goto error1; 1464 } 1465 1466 /* Allocate method table for this management class */ 1467 method = &(*class)->method_table[mgmt_class]; 1468 if ((ret = allocate_method_table(method))) 1469 goto error2; 1470 } else { 1471 method = &(*class)->method_table[mgmt_class]; 1472 if (!*method) { 1473 /* Allocate method table for this management class */ 1474 if ((ret = allocate_method_table(method))) 1475 goto error1; 1476 } 1477 } 1478 1479 /* Now, make sure methods are not already in use */ 1480 if (method_in_use(method, mad_reg_req)) 1481 goto error3; 1482 1483 /* Finally, add in methods being registered */ 1484 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) 1485 (*method)->agent[i] = agent_priv; 1486 1487 return 0; 1488 1489 error3: 1490 /* Remove any methods for this mad agent */ 1491 remove_methods_mad_agent(*method, agent_priv); 1492 /* Now, check to see if there are any methods in use */ 1493 if (!check_method_table(*method)) { 1494 /* If not, release management method table */ 1495 kfree(*method); 1496 *method = NULL; 1497 } 1498 ret = -EINVAL; 1499 goto error1; 1500 error2: 1501 kfree(*class); 1502 *class = NULL; 1503 error1: 1504 return ret; 1505 } 1506 1507 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req, 1508 struct ib_mad_agent_private *agent_priv) 1509 { 1510 struct ib_mad_port_private *port_priv; 1511 struct ib_mad_mgmt_vendor_class_table **vendor_table; 1512 struct ib_mad_mgmt_vendor_class_table *vendor = NULL; 1513 struct ib_mad_mgmt_vendor_class *vendor_class = NULL; 1514 struct ib_mad_mgmt_method_table **method; 1515 int i, ret = -ENOMEM; 1516 u8 vclass; 1517 1518 /* "New" vendor (with OUI) class */ 1519 vclass = vendor_class_index(mad_reg_req->mgmt_class); 1520 port_priv = agent_priv->qp_info->port_priv; 1521 vendor_table = &port_priv->version[ 1522 mad_reg_req->mgmt_class_version].vendor; 1523 if (!*vendor_table) { 1524 /* Allocate mgmt vendor class table for "new" class version */ 1525 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC); 1526 if (!vendor) 1527 goto error1; 1528 1529 *vendor_table = vendor; 1530 } 1531 if (!(*vendor_table)->vendor_class[vclass]) { 1532 /* Allocate table for this management vendor class */ 1533 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC); 1534 if (!vendor_class) 1535 goto error2; 1536 1537 (*vendor_table)->vendor_class[vclass] = vendor_class; 1538 } 1539 for (i = 0; i < MAX_MGMT_OUI; i++) { 1540 /* Is there matching OUI for this vendor class ? */ 1541 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i], 1542 mad_reg_req->oui, 3)) { 1543 method = &(*vendor_table)->vendor_class[ 1544 vclass]->method_table[i]; 1545 if (!*method) 1546 goto error3; 1547 goto check_in_use; 1548 } 1549 } 1550 for (i = 0; i < MAX_MGMT_OUI; i++) { 1551 /* OUI slot available ? */ 1552 if (!is_vendor_oui((*vendor_table)->vendor_class[ 1553 vclass]->oui[i])) { 1554 method = &(*vendor_table)->vendor_class[ 1555 vclass]->method_table[i]; 1556 /* Allocate method table for this OUI */ 1557 if (!*method) { 1558 ret = allocate_method_table(method); 1559 if (ret) 1560 goto error3; 1561 } 1562 memcpy((*vendor_table)->vendor_class[vclass]->oui[i], 1563 mad_reg_req->oui, 3); 1564 goto check_in_use; 1565 } 1566 } 1567 dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n"); 1568 goto error3; 1569 1570 check_in_use: 1571 /* Now, make sure methods are not already in use */ 1572 if (method_in_use(method, mad_reg_req)) 1573 goto error4; 1574 1575 /* Finally, add in methods being registered */ 1576 for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) 1577 (*method)->agent[i] = agent_priv; 1578 1579 return 0; 1580 1581 error4: 1582 /* Remove any methods for this mad agent */ 1583 remove_methods_mad_agent(*method, agent_priv); 1584 /* Now, check to see if there are any methods in use */ 1585 if (!check_method_table(*method)) { 1586 /* If not, release management method table */ 1587 kfree(*method); 1588 *method = NULL; 1589 } 1590 ret = -EINVAL; 1591 error3: 1592 if (vendor_class) { 1593 (*vendor_table)->vendor_class[vclass] = NULL; 1594 kfree(vendor_class); 1595 } 1596 error2: 1597 if (vendor) { 1598 *vendor_table = NULL; 1599 kfree(vendor); 1600 } 1601 error1: 1602 return ret; 1603 } 1604 1605 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv) 1606 { 1607 struct ib_mad_port_private *port_priv; 1608 struct ib_mad_mgmt_class_table *class; 1609 struct ib_mad_mgmt_method_table *method; 1610 struct ib_mad_mgmt_vendor_class_table *vendor; 1611 struct ib_mad_mgmt_vendor_class *vendor_class; 1612 int index; 1613 u8 mgmt_class; 1614 1615 /* 1616 * Was MAD registration request supplied 1617 * with original registration ? 1618 */ 1619 if (!agent_priv->reg_req) { 1620 goto out; 1621 } 1622 1623 port_priv = agent_priv->qp_info->port_priv; 1624 mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class); 1625 class = port_priv->version[ 1626 agent_priv->reg_req->mgmt_class_version].class; 1627 if (!class) 1628 goto vendor_check; 1629 1630 method = class->method_table[mgmt_class]; 1631 if (method) { 1632 /* Remove any methods for this mad agent */ 1633 remove_methods_mad_agent(method, agent_priv); 1634 /* Now, check to see if there are any methods still in use */ 1635 if (!check_method_table(method)) { 1636 /* If not, release management method table */ 1637 kfree(method); 1638 class->method_table[mgmt_class] = NULL; 1639 /* Any management classes left ? */ 1640 if (!check_class_table(class)) { 1641 /* If not, release management class table */ 1642 kfree(class); 1643 port_priv->version[ 1644 agent_priv->reg_req-> 1645 mgmt_class_version].class = NULL; 1646 } 1647 } 1648 } 1649 1650 vendor_check: 1651 if (!is_vendor_class(mgmt_class)) 1652 goto out; 1653 1654 /* normalize mgmt_class to vendor range 2 */ 1655 mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class); 1656 vendor = port_priv->version[ 1657 agent_priv->reg_req->mgmt_class_version].vendor; 1658 1659 if (!vendor) 1660 goto out; 1661 1662 vendor_class = vendor->vendor_class[mgmt_class]; 1663 if (vendor_class) { 1664 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui); 1665 if (index < 0) 1666 goto out; 1667 method = vendor_class->method_table[index]; 1668 if (method) { 1669 /* Remove any methods for this mad agent */ 1670 remove_methods_mad_agent(method, agent_priv); 1671 /* 1672 * Now, check to see if there are 1673 * any methods still in use 1674 */ 1675 if (!check_method_table(method)) { 1676 /* If not, release management method table */ 1677 kfree(method); 1678 vendor_class->method_table[index] = NULL; 1679 memset(vendor_class->oui[index], 0, 3); 1680 /* Any OUIs left ? */ 1681 if (!check_vendor_class(vendor_class)) { 1682 /* If not, release vendor class table */ 1683 kfree(vendor_class); 1684 vendor->vendor_class[mgmt_class] = NULL; 1685 /* Any other vendor classes left ? */ 1686 if (!check_vendor_table(vendor)) { 1687 kfree(vendor); 1688 port_priv->version[ 1689 agent_priv->reg_req-> 1690 mgmt_class_version]. 1691 vendor = NULL; 1692 } 1693 } 1694 } 1695 } 1696 } 1697 1698 out: 1699 return; 1700 } 1701 1702 static struct ib_mad_agent_private * 1703 find_mad_agent(struct ib_mad_port_private *port_priv, 1704 const struct ib_mad_hdr *mad_hdr) 1705 { 1706 struct ib_mad_agent_private *mad_agent = NULL; 1707 unsigned long flags; 1708 1709 spin_lock_irqsave(&port_priv->reg_lock, flags); 1710 if (ib_response_mad(mad_hdr)) { 1711 u32 hi_tid; 1712 struct ib_mad_agent_private *entry; 1713 1714 /* 1715 * Routing is based on high 32 bits of transaction ID 1716 * of MAD. 1717 */ 1718 hi_tid = be64_to_cpu(mad_hdr->tid) >> 32; 1719 list_for_each_entry(entry, &port_priv->agent_list, agent_list) { 1720 if (entry->agent.hi_tid == hi_tid) { 1721 mad_agent = entry; 1722 break; 1723 } 1724 } 1725 } else { 1726 struct ib_mad_mgmt_class_table *class; 1727 struct ib_mad_mgmt_method_table *method; 1728 struct ib_mad_mgmt_vendor_class_table *vendor; 1729 struct ib_mad_mgmt_vendor_class *vendor_class; 1730 const struct ib_vendor_mad *vendor_mad; 1731 int index; 1732 1733 /* 1734 * Routing is based on version, class, and method 1735 * For "newer" vendor MADs, also based on OUI 1736 */ 1737 if (mad_hdr->class_version >= MAX_MGMT_VERSION) 1738 goto out; 1739 if (!is_vendor_class(mad_hdr->mgmt_class)) { 1740 class = port_priv->version[ 1741 mad_hdr->class_version].class; 1742 if (!class) 1743 goto out; 1744 if (convert_mgmt_class(mad_hdr->mgmt_class) >= 1745 ARRAY_SIZE(class->method_table)) 1746 goto out; 1747 method = class->method_table[convert_mgmt_class( 1748 mad_hdr->mgmt_class)]; 1749 if (method) 1750 mad_agent = method->agent[mad_hdr->method & 1751 ~IB_MGMT_METHOD_RESP]; 1752 } else { 1753 vendor = port_priv->version[ 1754 mad_hdr->class_version].vendor; 1755 if (!vendor) 1756 goto out; 1757 vendor_class = vendor->vendor_class[vendor_class_index( 1758 mad_hdr->mgmt_class)]; 1759 if (!vendor_class) 1760 goto out; 1761 /* Find matching OUI */ 1762 vendor_mad = (const struct ib_vendor_mad *)mad_hdr; 1763 index = find_vendor_oui(vendor_class, vendor_mad->oui); 1764 if (index == -1) 1765 goto out; 1766 method = vendor_class->method_table[index]; 1767 if (method) { 1768 mad_agent = method->agent[mad_hdr->method & 1769 ~IB_MGMT_METHOD_RESP]; 1770 } 1771 } 1772 } 1773 1774 if (mad_agent) { 1775 if (mad_agent->agent.recv_handler) 1776 atomic_inc(&mad_agent->refcount); 1777 else { 1778 dev_notice(&port_priv->device->dev, 1779 "No receive handler for client %p on port %d\n", 1780 &mad_agent->agent, port_priv->port_num); 1781 mad_agent = NULL; 1782 } 1783 } 1784 out: 1785 spin_unlock_irqrestore(&port_priv->reg_lock, flags); 1786 1787 return mad_agent; 1788 } 1789 1790 static int validate_mad(const struct ib_mad_hdr *mad_hdr, 1791 const struct ib_mad_qp_info *qp_info, 1792 bool opa) 1793 { 1794 int valid = 0; 1795 u32 qp_num = qp_info->qp->qp_num; 1796 1797 /* Make sure MAD base version is understood */ 1798 if (mad_hdr->base_version != IB_MGMT_BASE_VERSION && 1799 (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) { 1800 pr_err("MAD received with unsupported base version %d %s\n", 1801 mad_hdr->base_version, opa ? "(opa)" : ""); 1802 goto out; 1803 } 1804 1805 /* Filter SMI packets sent to other than QP0 */ 1806 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) || 1807 (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) { 1808 if (qp_num == 0) 1809 valid = 1; 1810 } else { 1811 /* CM attributes other than ClassPortInfo only use Send method */ 1812 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) && 1813 (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) && 1814 (mad_hdr->method != IB_MGMT_METHOD_SEND)) 1815 goto out; 1816 /* Filter GSI packets sent to QP0 */ 1817 if (qp_num != 0) 1818 valid = 1; 1819 } 1820 1821 out: 1822 return valid; 1823 } 1824 1825 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv, 1826 const struct ib_mad_hdr *mad_hdr) 1827 { 1828 const struct ib_rmpp_mad *rmpp_mad; 1829 1830 rmpp_mad = (const struct ib_rmpp_mad *)mad_hdr; 1831 return !mad_agent_priv->agent.rmpp_version || 1832 !ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) || 1833 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) & 1834 IB_MGMT_RMPP_FLAG_ACTIVE) || 1835 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA); 1836 } 1837 1838 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr, 1839 const struct ib_mad_recv_wc *rwc) 1840 { 1841 return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class == 1842 rwc->recv_buf.mad->mad_hdr.mgmt_class; 1843 } 1844 1845 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv, 1846 const struct ib_mad_send_wr_private *wr, 1847 const struct ib_mad_recv_wc *rwc ) 1848 { 1849 struct ib_ah_attr attr; 1850 u8 send_resp, rcv_resp; 1851 union ib_gid sgid; 1852 struct ib_device *device = mad_agent_priv->agent.device; 1853 u8 port_num = mad_agent_priv->agent.port_num; 1854 u8 lmc; 1855 1856 send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad); 1857 rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr); 1858 1859 if (send_resp == rcv_resp) 1860 /* both requests, or both responses. GIDs different */ 1861 return 0; 1862 1863 if (ib_query_ah(wr->send_buf.ah, &attr)) 1864 /* Assume not equal, to avoid false positives. */ 1865 return 0; 1866 1867 if (!!(attr.ah_flags & IB_AH_GRH) != 1868 !!(rwc->wc->wc_flags & IB_WC_GRH)) 1869 /* one has GID, other does not. Assume different */ 1870 return 0; 1871 1872 if (!send_resp && rcv_resp) { 1873 /* is request/response. */ 1874 if (!(attr.ah_flags & IB_AH_GRH)) { 1875 if (ib_get_cached_lmc(device, port_num, &lmc)) 1876 return 0; 1877 return (!lmc || !((attr.src_path_bits ^ 1878 rwc->wc->dlid_path_bits) & 1879 ((1 << lmc) - 1))); 1880 } else { 1881 if (ib_get_cached_gid(device, port_num, 1882 attr.grh.sgid_index, &sgid, NULL)) 1883 return 0; 1884 return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw, 1885 16); 1886 } 1887 } 1888 1889 if (!(attr.ah_flags & IB_AH_GRH)) 1890 return attr.dlid == rwc->wc->slid; 1891 else 1892 return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw, 1893 16); 1894 } 1895 1896 static inline int is_direct(u8 class) 1897 { 1898 return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE); 1899 } 1900 1901 struct ib_mad_send_wr_private* 1902 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv, 1903 const struct ib_mad_recv_wc *wc) 1904 { 1905 struct ib_mad_send_wr_private *wr; 1906 const struct ib_mad_hdr *mad_hdr; 1907 1908 mad_hdr = &wc->recv_buf.mad->mad_hdr; 1909 1910 list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) { 1911 if ((wr->tid == mad_hdr->tid) && 1912 rcv_has_same_class(wr, wc) && 1913 /* 1914 * Don't check GID for direct routed MADs. 1915 * These might have permissive LIDs. 1916 */ 1917 (is_direct(mad_hdr->mgmt_class) || 1918 rcv_has_same_gid(mad_agent_priv, wr, wc))) 1919 return (wr->status == IB_WC_SUCCESS) ? wr : NULL; 1920 } 1921 1922 /* 1923 * It's possible to receive the response before we've 1924 * been notified that the send has completed 1925 */ 1926 list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) { 1927 if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) && 1928 wr->tid == mad_hdr->tid && 1929 wr->timeout && 1930 rcv_has_same_class(wr, wc) && 1931 /* 1932 * Don't check GID for direct routed MADs. 1933 * These might have permissive LIDs. 1934 */ 1935 (is_direct(mad_hdr->mgmt_class) || 1936 rcv_has_same_gid(mad_agent_priv, wr, wc))) 1937 /* Verify request has not been canceled */ 1938 return (wr->status == IB_WC_SUCCESS) ? wr : NULL; 1939 } 1940 return NULL; 1941 } 1942 1943 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr) 1944 { 1945 mad_send_wr->timeout = 0; 1946 if (mad_send_wr->refcount == 1) 1947 list_move_tail(&mad_send_wr->agent_list, 1948 &mad_send_wr->mad_agent_priv->done_list); 1949 } 1950 1951 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv, 1952 struct ib_mad_recv_wc *mad_recv_wc) 1953 { 1954 struct ib_mad_send_wr_private *mad_send_wr; 1955 struct ib_mad_send_wc mad_send_wc; 1956 unsigned long flags; 1957 1958 INIT_LIST_HEAD(&mad_recv_wc->rmpp_list); 1959 list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list); 1960 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) { 1961 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv, 1962 mad_recv_wc); 1963 if (!mad_recv_wc) { 1964 deref_mad_agent(mad_agent_priv); 1965 return; 1966 } 1967 } 1968 1969 /* Complete corresponding request */ 1970 if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) { 1971 spin_lock_irqsave(&mad_agent_priv->lock, flags); 1972 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc); 1973 if (!mad_send_wr) { 1974 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 1975 if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) 1976 && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class) 1977 && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr) 1978 & IB_MGMT_RMPP_FLAG_ACTIVE)) { 1979 /* user rmpp is in effect 1980 * and this is an active RMPP MAD 1981 */ 1982 mad_agent_priv->agent.recv_handler( 1983 &mad_agent_priv->agent, NULL, 1984 mad_recv_wc); 1985 atomic_dec(&mad_agent_priv->refcount); 1986 } else { 1987 /* not user rmpp, revert to normal behavior and 1988 * drop the mad */ 1989 ib_free_recv_mad(mad_recv_wc); 1990 deref_mad_agent(mad_agent_priv); 1991 return; 1992 } 1993 } else { 1994 ib_mark_mad_done(mad_send_wr); 1995 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 1996 1997 /* Defined behavior is to complete response before request */ 1998 mad_agent_priv->agent.recv_handler( 1999 &mad_agent_priv->agent, 2000 &mad_send_wr->send_buf, 2001 mad_recv_wc); 2002 atomic_dec(&mad_agent_priv->refcount); 2003 2004 mad_send_wc.status = IB_WC_SUCCESS; 2005 mad_send_wc.vendor_err = 0; 2006 mad_send_wc.send_buf = &mad_send_wr->send_buf; 2007 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc); 2008 } 2009 } else { 2010 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL, 2011 mad_recv_wc); 2012 deref_mad_agent(mad_agent_priv); 2013 } 2014 } 2015 2016 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv, 2017 const struct ib_mad_qp_info *qp_info, 2018 const struct ib_wc *wc, 2019 int port_num, 2020 struct ib_mad_private *recv, 2021 struct ib_mad_private *response) 2022 { 2023 enum smi_forward_action retsmi; 2024 struct ib_smp *smp = (struct ib_smp *)recv->mad; 2025 2026 if (smi_handle_dr_smp_recv(smp, 2027 rdma_cap_ib_switch(port_priv->device), 2028 port_num, 2029 port_priv->device->phys_port_cnt) == 2030 IB_SMI_DISCARD) 2031 return IB_SMI_DISCARD; 2032 2033 retsmi = smi_check_forward_dr_smp(smp); 2034 if (retsmi == IB_SMI_LOCAL) 2035 return IB_SMI_HANDLE; 2036 2037 if (retsmi == IB_SMI_SEND) { /* don't forward */ 2038 if (smi_handle_dr_smp_send(smp, 2039 rdma_cap_ib_switch(port_priv->device), 2040 port_num) == IB_SMI_DISCARD) 2041 return IB_SMI_DISCARD; 2042 2043 if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD) 2044 return IB_SMI_DISCARD; 2045 } else if (rdma_cap_ib_switch(port_priv->device)) { 2046 /* forward case for switches */ 2047 memcpy(response, recv, mad_priv_size(response)); 2048 response->header.recv_wc.wc = &response->header.wc; 2049 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad; 2050 response->header.recv_wc.recv_buf.grh = &response->grh; 2051 2052 agent_send_response((const struct ib_mad_hdr *)response->mad, 2053 &response->grh, wc, 2054 port_priv->device, 2055 smi_get_fwd_port(smp), 2056 qp_info->qp->qp_num, 2057 response->mad_size, 2058 false); 2059 2060 return IB_SMI_DISCARD; 2061 } 2062 return IB_SMI_HANDLE; 2063 } 2064 2065 static bool generate_unmatched_resp(const struct ib_mad_private *recv, 2066 struct ib_mad_private *response, 2067 size_t *resp_len, bool opa) 2068 { 2069 const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad; 2070 struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad; 2071 2072 if (recv_hdr->method == IB_MGMT_METHOD_GET || 2073 recv_hdr->method == IB_MGMT_METHOD_SET) { 2074 memcpy(response, recv, mad_priv_size(response)); 2075 response->header.recv_wc.wc = &response->header.wc; 2076 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad; 2077 response->header.recv_wc.recv_buf.grh = &response->grh; 2078 resp_hdr->method = IB_MGMT_METHOD_GET_RESP; 2079 resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB); 2080 if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) 2081 resp_hdr->status |= IB_SMP_DIRECTION; 2082 2083 if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) { 2084 if (recv_hdr->mgmt_class == 2085 IB_MGMT_CLASS_SUBN_LID_ROUTED || 2086 recv_hdr->mgmt_class == 2087 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) 2088 *resp_len = opa_get_smp_header_size( 2089 (const struct opa_smp *)recv->mad); 2090 else 2091 *resp_len = sizeof(struct ib_mad_hdr); 2092 } 2093 2094 return true; 2095 } else { 2096 return false; 2097 } 2098 } 2099 2100 static enum smi_action 2101 handle_opa_smi(struct ib_mad_port_private *port_priv, 2102 struct ib_mad_qp_info *qp_info, 2103 struct ib_wc *wc, 2104 int port_num, 2105 struct ib_mad_private *recv, 2106 struct ib_mad_private *response) 2107 { 2108 enum smi_forward_action retsmi; 2109 struct opa_smp *smp = (struct opa_smp *)recv->mad; 2110 2111 if (opa_smi_handle_dr_smp_recv(smp, 2112 rdma_cap_ib_switch(port_priv->device), 2113 port_num, 2114 port_priv->device->phys_port_cnt) == 2115 IB_SMI_DISCARD) 2116 return IB_SMI_DISCARD; 2117 2118 retsmi = opa_smi_check_forward_dr_smp(smp); 2119 if (retsmi == IB_SMI_LOCAL) 2120 return IB_SMI_HANDLE; 2121 2122 if (retsmi == IB_SMI_SEND) { /* don't forward */ 2123 if (opa_smi_handle_dr_smp_send(smp, 2124 rdma_cap_ib_switch(port_priv->device), 2125 port_num) == IB_SMI_DISCARD) 2126 return IB_SMI_DISCARD; 2127 2128 if (opa_smi_check_local_smp(smp, port_priv->device) == 2129 IB_SMI_DISCARD) 2130 return IB_SMI_DISCARD; 2131 2132 } else if (rdma_cap_ib_switch(port_priv->device)) { 2133 /* forward case for switches */ 2134 memcpy(response, recv, mad_priv_size(response)); 2135 response->header.recv_wc.wc = &response->header.wc; 2136 response->header.recv_wc.recv_buf.opa_mad = 2137 (struct opa_mad *)response->mad; 2138 response->header.recv_wc.recv_buf.grh = &response->grh; 2139 2140 agent_send_response((const struct ib_mad_hdr *)response->mad, 2141 &response->grh, wc, 2142 port_priv->device, 2143 opa_smi_get_fwd_port(smp), 2144 qp_info->qp->qp_num, 2145 recv->header.wc.byte_len, 2146 true); 2147 2148 return IB_SMI_DISCARD; 2149 } 2150 2151 return IB_SMI_HANDLE; 2152 } 2153 2154 static enum smi_action 2155 handle_smi(struct ib_mad_port_private *port_priv, 2156 struct ib_mad_qp_info *qp_info, 2157 struct ib_wc *wc, 2158 int port_num, 2159 struct ib_mad_private *recv, 2160 struct ib_mad_private *response, 2161 bool opa) 2162 { 2163 struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad; 2164 2165 if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION && 2166 mad_hdr->class_version == OPA_SMI_CLASS_VERSION) 2167 return handle_opa_smi(port_priv, qp_info, wc, port_num, recv, 2168 response); 2169 2170 return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response); 2171 } 2172 2173 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc) 2174 { 2175 struct ib_mad_port_private *port_priv = cq->cq_context; 2176 struct ib_mad_list_head *mad_list = 2177 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe); 2178 struct ib_mad_qp_info *qp_info; 2179 struct ib_mad_private_header *mad_priv_hdr; 2180 struct ib_mad_private *recv, *response = NULL; 2181 struct ib_mad_agent_private *mad_agent; 2182 int port_num; 2183 int ret = IB_MAD_RESULT_SUCCESS; 2184 size_t mad_size; 2185 u16 resp_mad_pkey_index = 0; 2186 bool opa; 2187 2188 if (list_empty_careful(&port_priv->port_list)) 2189 return; 2190 2191 if (wc->status != IB_WC_SUCCESS) { 2192 /* 2193 * Receive errors indicate that the QP has entered the error 2194 * state - error handling/shutdown code will cleanup 2195 */ 2196 return; 2197 } 2198 2199 qp_info = mad_list->mad_queue->qp_info; 2200 dequeue_mad(mad_list); 2201 2202 opa = rdma_cap_opa_mad(qp_info->port_priv->device, 2203 qp_info->port_priv->port_num); 2204 2205 mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header, 2206 mad_list); 2207 recv = container_of(mad_priv_hdr, struct ib_mad_private, header); 2208 ib_dma_unmap_single(port_priv->device, 2209 recv->header.mapping, 2210 mad_priv_dma_size(recv), 2211 DMA_FROM_DEVICE); 2212 2213 /* Setup MAD receive work completion from "normal" work completion */ 2214 recv->header.wc = *wc; 2215 recv->header.recv_wc.wc = &recv->header.wc; 2216 2217 if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) { 2218 recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh); 2219 recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad); 2220 } else { 2221 recv->header.recv_wc.mad_len = sizeof(struct ib_mad); 2222 recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad); 2223 } 2224 2225 recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad; 2226 recv->header.recv_wc.recv_buf.grh = &recv->grh; 2227 2228 if (atomic_read(&qp_info->snoop_count)) 2229 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS); 2230 2231 /* Validate MAD */ 2232 if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa)) 2233 goto out; 2234 2235 mad_size = recv->mad_size; 2236 response = alloc_mad_private(mad_size, GFP_KERNEL); 2237 if (!response) 2238 goto out; 2239 2240 if (rdma_cap_ib_switch(port_priv->device)) 2241 port_num = wc->port_num; 2242 else 2243 port_num = port_priv->port_num; 2244 2245 if (((struct ib_mad_hdr *)recv->mad)->mgmt_class == 2246 IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) { 2247 if (handle_smi(port_priv, qp_info, wc, port_num, recv, 2248 response, opa) 2249 == IB_SMI_DISCARD) 2250 goto out; 2251 } 2252 2253 /* Give driver "right of first refusal" on incoming MAD */ 2254 if (port_priv->device->process_mad) { 2255 ret = port_priv->device->process_mad(port_priv->device, 0, 2256 port_priv->port_num, 2257 wc, &recv->grh, 2258 (const struct ib_mad_hdr *)recv->mad, 2259 recv->mad_size, 2260 (struct ib_mad_hdr *)response->mad, 2261 &mad_size, &resp_mad_pkey_index); 2262 2263 if (opa) 2264 wc->pkey_index = resp_mad_pkey_index; 2265 2266 if (ret & IB_MAD_RESULT_SUCCESS) { 2267 if (ret & IB_MAD_RESULT_CONSUMED) 2268 goto out; 2269 if (ret & IB_MAD_RESULT_REPLY) { 2270 agent_send_response((const struct ib_mad_hdr *)response->mad, 2271 &recv->grh, wc, 2272 port_priv->device, 2273 port_num, 2274 qp_info->qp->qp_num, 2275 mad_size, opa); 2276 goto out; 2277 } 2278 } 2279 } 2280 2281 mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad); 2282 if (mad_agent) { 2283 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc); 2284 /* 2285 * recv is freed up in error cases in ib_mad_complete_recv 2286 * or via recv_handler in ib_mad_complete_recv() 2287 */ 2288 recv = NULL; 2289 } else if ((ret & IB_MAD_RESULT_SUCCESS) && 2290 generate_unmatched_resp(recv, response, &mad_size, opa)) { 2291 agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc, 2292 port_priv->device, port_num, 2293 qp_info->qp->qp_num, mad_size, opa); 2294 } 2295 2296 out: 2297 /* Post another receive request for this QP */ 2298 if (response) { 2299 ib_mad_post_receive_mads(qp_info, response); 2300 kfree(recv); 2301 } else 2302 ib_mad_post_receive_mads(qp_info, recv); 2303 } 2304 2305 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv) 2306 { 2307 struct ib_mad_send_wr_private *mad_send_wr; 2308 unsigned long delay; 2309 2310 if (list_empty(&mad_agent_priv->wait_list)) { 2311 cancel_delayed_work(&mad_agent_priv->timed_work); 2312 } else { 2313 mad_send_wr = list_entry(mad_agent_priv->wait_list.next, 2314 struct ib_mad_send_wr_private, 2315 agent_list); 2316 2317 if (time_after(mad_agent_priv->timeout, 2318 mad_send_wr->timeout)) { 2319 mad_agent_priv->timeout = mad_send_wr->timeout; 2320 delay = mad_send_wr->timeout - jiffies; 2321 if ((long)delay <= 0) 2322 delay = 1; 2323 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq, 2324 &mad_agent_priv->timed_work, delay); 2325 } 2326 } 2327 } 2328 2329 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr) 2330 { 2331 struct ib_mad_agent_private *mad_agent_priv; 2332 struct ib_mad_send_wr_private *temp_mad_send_wr; 2333 struct list_head *list_item; 2334 unsigned long delay; 2335 2336 mad_agent_priv = mad_send_wr->mad_agent_priv; 2337 list_del(&mad_send_wr->agent_list); 2338 2339 delay = mad_send_wr->timeout; 2340 mad_send_wr->timeout += jiffies; 2341 2342 if (delay) { 2343 list_for_each_prev(list_item, &mad_agent_priv->wait_list) { 2344 temp_mad_send_wr = list_entry(list_item, 2345 struct ib_mad_send_wr_private, 2346 agent_list); 2347 if (time_after(mad_send_wr->timeout, 2348 temp_mad_send_wr->timeout)) 2349 break; 2350 } 2351 } 2352 else 2353 list_item = &mad_agent_priv->wait_list; 2354 list_add(&mad_send_wr->agent_list, list_item); 2355 2356 /* Reschedule a work item if we have a shorter timeout */ 2357 if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) 2358 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq, 2359 &mad_agent_priv->timed_work, delay); 2360 } 2361 2362 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr, 2363 int timeout_ms) 2364 { 2365 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms); 2366 wait_for_response(mad_send_wr); 2367 } 2368 2369 /* 2370 * Process a send work completion 2371 */ 2372 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr, 2373 struct ib_mad_send_wc *mad_send_wc) 2374 { 2375 struct ib_mad_agent_private *mad_agent_priv; 2376 unsigned long flags; 2377 int ret; 2378 2379 mad_agent_priv = mad_send_wr->mad_agent_priv; 2380 spin_lock_irqsave(&mad_agent_priv->lock, flags); 2381 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) { 2382 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc); 2383 if (ret == IB_RMPP_RESULT_CONSUMED) 2384 goto done; 2385 } else 2386 ret = IB_RMPP_RESULT_UNHANDLED; 2387 2388 if (mad_send_wc->status != IB_WC_SUCCESS && 2389 mad_send_wr->status == IB_WC_SUCCESS) { 2390 mad_send_wr->status = mad_send_wc->status; 2391 mad_send_wr->refcount -= (mad_send_wr->timeout > 0); 2392 } 2393 2394 if (--mad_send_wr->refcount > 0) { 2395 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout && 2396 mad_send_wr->status == IB_WC_SUCCESS) { 2397 wait_for_response(mad_send_wr); 2398 } 2399 goto done; 2400 } 2401 2402 /* Remove send from MAD agent and notify client of completion */ 2403 list_del(&mad_send_wr->agent_list); 2404 adjust_timeout(mad_agent_priv); 2405 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 2406 2407 if (mad_send_wr->status != IB_WC_SUCCESS ) 2408 mad_send_wc->status = mad_send_wr->status; 2409 if (ret == IB_RMPP_RESULT_INTERNAL) 2410 ib_rmpp_send_handler(mad_send_wc); 2411 else 2412 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent, 2413 mad_send_wc); 2414 2415 /* Release reference on agent taken when sending */ 2416 deref_mad_agent(mad_agent_priv); 2417 return; 2418 done: 2419 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 2420 } 2421 2422 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc) 2423 { 2424 struct ib_mad_port_private *port_priv = cq->cq_context; 2425 struct ib_mad_list_head *mad_list = 2426 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe); 2427 struct ib_mad_send_wr_private *mad_send_wr, *queued_send_wr; 2428 struct ib_mad_qp_info *qp_info; 2429 struct ib_mad_queue *send_queue; 2430 const struct ib_send_wr *bad_send_wr; 2431 struct ib_mad_send_wc mad_send_wc; 2432 unsigned long flags; 2433 int ret; 2434 2435 if (list_empty_careful(&port_priv->port_list)) 2436 return; 2437 2438 if (wc->status != IB_WC_SUCCESS) { 2439 if (!ib_mad_send_error(port_priv, wc)) 2440 return; 2441 } 2442 2443 mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private, 2444 mad_list); 2445 send_queue = mad_list->mad_queue; 2446 qp_info = send_queue->qp_info; 2447 2448 retry: 2449 ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device, 2450 mad_send_wr->header_mapping, 2451 mad_send_wr->sg_list[0].length, DMA_TO_DEVICE); 2452 ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device, 2453 mad_send_wr->payload_mapping, 2454 mad_send_wr->sg_list[1].length, DMA_TO_DEVICE); 2455 queued_send_wr = NULL; 2456 spin_lock_irqsave(&send_queue->lock, flags); 2457 list_del(&mad_list->list); 2458 2459 /* Move queued send to the send queue */ 2460 if (send_queue->count-- > send_queue->max_active) { 2461 mad_list = container_of(qp_info->overflow_list.next, 2462 struct ib_mad_list_head, list); 2463 queued_send_wr = container_of(mad_list, 2464 struct ib_mad_send_wr_private, 2465 mad_list); 2466 list_move_tail(&mad_list->list, &send_queue->list); 2467 } 2468 spin_unlock_irqrestore(&send_queue->lock, flags); 2469 2470 mad_send_wc.send_buf = &mad_send_wr->send_buf; 2471 mad_send_wc.status = wc->status; 2472 mad_send_wc.vendor_err = wc->vendor_err; 2473 if (atomic_read(&qp_info->snoop_count)) 2474 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc, 2475 IB_MAD_SNOOP_SEND_COMPLETIONS); 2476 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc); 2477 2478 if (queued_send_wr) { 2479 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr, 2480 &bad_send_wr); 2481 if (ret) { 2482 dev_err(&port_priv->device->dev, 2483 "ib_post_send failed: %d\n", ret); 2484 mad_send_wr = queued_send_wr; 2485 wc->status = IB_WC_LOC_QP_OP_ERR; 2486 goto retry; 2487 } 2488 } 2489 } 2490 2491 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info) 2492 { 2493 struct ib_mad_send_wr_private *mad_send_wr; 2494 struct ib_mad_list_head *mad_list; 2495 unsigned long flags; 2496 2497 spin_lock_irqsave(&qp_info->send_queue.lock, flags); 2498 list_for_each_entry(mad_list, &qp_info->send_queue.list, list) { 2499 mad_send_wr = container_of(mad_list, 2500 struct ib_mad_send_wr_private, 2501 mad_list); 2502 mad_send_wr->retry = 1; 2503 } 2504 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags); 2505 } 2506 2507 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv, 2508 struct ib_wc *wc) 2509 { 2510 struct ib_mad_list_head *mad_list = 2511 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe); 2512 struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info; 2513 struct ib_mad_send_wr_private *mad_send_wr; 2514 int ret; 2515 2516 /* 2517 * Send errors will transition the QP to SQE - move 2518 * QP to RTS and repost flushed work requests 2519 */ 2520 mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private, 2521 mad_list); 2522 if (wc->status == IB_WC_WR_FLUSH_ERR) { 2523 if (mad_send_wr->retry) { 2524 /* Repost send */ 2525 const struct ib_send_wr *bad_send_wr; 2526 2527 mad_send_wr->retry = 0; 2528 ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr, 2529 &bad_send_wr); 2530 if (!ret) 2531 return false; 2532 } 2533 } else { 2534 struct ib_qp_attr *attr; 2535 2536 /* Transition QP to RTS and fail offending send */ 2537 attr = kmalloc(sizeof *attr, GFP_KERNEL); 2538 if (attr) { 2539 attr->qp_state = IB_QPS_RTS; 2540 attr->cur_qp_state = IB_QPS_SQE; 2541 ret = ib_modify_qp(qp_info->qp, attr, 2542 IB_QP_STATE | IB_QP_CUR_STATE); 2543 kfree(attr); 2544 if (ret) 2545 dev_err(&port_priv->device->dev, 2546 "%s - ib_modify_qp to RTS: %d\n", 2547 __func__, ret); 2548 else 2549 mark_sends_for_retry(qp_info); 2550 } 2551 } 2552 2553 return true; 2554 } 2555 2556 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv) 2557 { 2558 unsigned long flags; 2559 struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr; 2560 struct ib_mad_send_wc mad_send_wc; 2561 struct list_head cancel_list; 2562 2563 INIT_LIST_HEAD(&cancel_list); 2564 2565 spin_lock_irqsave(&mad_agent_priv->lock, flags); 2566 list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr, 2567 &mad_agent_priv->send_list, agent_list) { 2568 if (mad_send_wr->status == IB_WC_SUCCESS) { 2569 mad_send_wr->status = IB_WC_WR_FLUSH_ERR; 2570 mad_send_wr->refcount -= (mad_send_wr->timeout > 0); 2571 } 2572 } 2573 2574 /* Empty wait list to prevent receives from finding a request */ 2575 list_splice_init(&mad_agent_priv->wait_list, &cancel_list); 2576 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 2577 2578 /* Report all cancelled requests */ 2579 mad_send_wc.status = IB_WC_WR_FLUSH_ERR; 2580 mad_send_wc.vendor_err = 0; 2581 2582 list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr, 2583 &cancel_list, agent_list) { 2584 mad_send_wc.send_buf = &mad_send_wr->send_buf; 2585 list_del(&mad_send_wr->agent_list); 2586 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent, 2587 &mad_send_wc); 2588 atomic_dec(&mad_agent_priv->refcount); 2589 } 2590 } 2591 2592 static struct ib_mad_send_wr_private* 2593 find_send_wr(struct ib_mad_agent_private *mad_agent_priv, 2594 struct ib_mad_send_buf *send_buf) 2595 { 2596 struct ib_mad_send_wr_private *mad_send_wr; 2597 2598 list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list, 2599 agent_list) { 2600 if (&mad_send_wr->send_buf == send_buf) 2601 return mad_send_wr; 2602 } 2603 2604 list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list, 2605 agent_list) { 2606 if (is_rmpp_data_mad(mad_agent_priv, 2607 mad_send_wr->send_buf.mad) && 2608 &mad_send_wr->send_buf == send_buf) 2609 return mad_send_wr; 2610 } 2611 return NULL; 2612 } 2613 2614 int ib_modify_mad(struct ib_mad_agent *mad_agent, 2615 struct ib_mad_send_buf *send_buf, u32 timeout_ms) 2616 { 2617 struct ib_mad_agent_private *mad_agent_priv; 2618 struct ib_mad_send_wr_private *mad_send_wr; 2619 unsigned long flags; 2620 int active; 2621 2622 mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private, 2623 agent); 2624 spin_lock_irqsave(&mad_agent_priv->lock, flags); 2625 mad_send_wr = find_send_wr(mad_agent_priv, send_buf); 2626 if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) { 2627 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 2628 return -EINVAL; 2629 } 2630 2631 active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1); 2632 if (!timeout_ms) { 2633 mad_send_wr->status = IB_WC_WR_FLUSH_ERR; 2634 mad_send_wr->refcount -= (mad_send_wr->timeout > 0); 2635 } 2636 2637 mad_send_wr->send_buf.timeout_ms = timeout_ms; 2638 if (active) 2639 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms); 2640 else 2641 ib_reset_mad_timeout(mad_send_wr, timeout_ms); 2642 2643 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 2644 return 0; 2645 } 2646 EXPORT_SYMBOL(ib_modify_mad); 2647 2648 void ib_cancel_mad(struct ib_mad_agent *mad_agent, 2649 struct ib_mad_send_buf *send_buf) 2650 { 2651 ib_modify_mad(mad_agent, send_buf, 0); 2652 } 2653 EXPORT_SYMBOL(ib_cancel_mad); 2654 2655 static void local_completions(struct work_struct *work) 2656 { 2657 struct ib_mad_agent_private *mad_agent_priv; 2658 struct ib_mad_local_private *local; 2659 struct ib_mad_agent_private *recv_mad_agent; 2660 unsigned long flags; 2661 int free_mad; 2662 struct ib_wc wc; 2663 struct ib_mad_send_wc mad_send_wc; 2664 bool opa; 2665 2666 mad_agent_priv = 2667 container_of(work, struct ib_mad_agent_private, local_work); 2668 2669 opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device, 2670 mad_agent_priv->qp_info->port_priv->port_num); 2671 2672 spin_lock_irqsave(&mad_agent_priv->lock, flags); 2673 while (!list_empty(&mad_agent_priv->local_list)) { 2674 local = list_entry(mad_agent_priv->local_list.next, 2675 struct ib_mad_local_private, 2676 completion_list); 2677 list_del(&local->completion_list); 2678 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 2679 free_mad = 0; 2680 if (local->mad_priv) { 2681 u8 base_version; 2682 recv_mad_agent = local->recv_mad_agent; 2683 if (!recv_mad_agent) { 2684 dev_err(&mad_agent_priv->agent.device->dev, 2685 "No receive MAD agent for local completion\n"); 2686 free_mad = 1; 2687 goto local_send_completion; 2688 } 2689 2690 /* 2691 * Defined behavior is to complete response 2692 * before request 2693 */ 2694 build_smp_wc(recv_mad_agent->agent.qp, 2695 local->mad_send_wr->send_wr.wr.wr_cqe, 2696 be16_to_cpu(IB_LID_PERMISSIVE), 2697 local->mad_send_wr->send_wr.pkey_index, 2698 recv_mad_agent->agent.port_num, &wc); 2699 2700 local->mad_priv->header.recv_wc.wc = &wc; 2701 2702 base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version; 2703 if (opa && base_version == OPA_MGMT_BASE_VERSION) { 2704 local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len; 2705 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad); 2706 } else { 2707 local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad); 2708 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad); 2709 } 2710 2711 INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list); 2712 list_add(&local->mad_priv->header.recv_wc.recv_buf.list, 2713 &local->mad_priv->header.recv_wc.rmpp_list); 2714 local->mad_priv->header.recv_wc.recv_buf.grh = NULL; 2715 local->mad_priv->header.recv_wc.recv_buf.mad = 2716 (struct ib_mad *)local->mad_priv->mad; 2717 if (atomic_read(&recv_mad_agent->qp_info->snoop_count)) 2718 snoop_recv(recv_mad_agent->qp_info, 2719 &local->mad_priv->header.recv_wc, 2720 IB_MAD_SNOOP_RECVS); 2721 recv_mad_agent->agent.recv_handler( 2722 &recv_mad_agent->agent, 2723 &local->mad_send_wr->send_buf, 2724 &local->mad_priv->header.recv_wc); 2725 spin_lock_irqsave(&recv_mad_agent->lock, flags); 2726 atomic_dec(&recv_mad_agent->refcount); 2727 spin_unlock_irqrestore(&recv_mad_agent->lock, flags); 2728 } 2729 2730 local_send_completion: 2731 /* Complete send */ 2732 mad_send_wc.status = IB_WC_SUCCESS; 2733 mad_send_wc.vendor_err = 0; 2734 mad_send_wc.send_buf = &local->mad_send_wr->send_buf; 2735 if (atomic_read(&mad_agent_priv->qp_info->snoop_count)) 2736 snoop_send(mad_agent_priv->qp_info, 2737 &local->mad_send_wr->send_buf, 2738 &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS); 2739 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent, 2740 &mad_send_wc); 2741 2742 spin_lock_irqsave(&mad_agent_priv->lock, flags); 2743 atomic_dec(&mad_agent_priv->refcount); 2744 if (free_mad) 2745 kfree(local->mad_priv); 2746 kfree(local); 2747 } 2748 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 2749 } 2750 2751 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr) 2752 { 2753 int ret; 2754 2755 if (!mad_send_wr->retries_left) 2756 return -ETIMEDOUT; 2757 2758 mad_send_wr->retries_left--; 2759 mad_send_wr->send_buf.retries++; 2760 2761 mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms); 2762 2763 if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) { 2764 ret = ib_retry_rmpp(mad_send_wr); 2765 switch (ret) { 2766 case IB_RMPP_RESULT_UNHANDLED: 2767 ret = ib_send_mad(mad_send_wr); 2768 break; 2769 case IB_RMPP_RESULT_CONSUMED: 2770 ret = 0; 2771 break; 2772 default: 2773 ret = -ECOMM; 2774 break; 2775 } 2776 } else 2777 ret = ib_send_mad(mad_send_wr); 2778 2779 if (!ret) { 2780 mad_send_wr->refcount++; 2781 list_add_tail(&mad_send_wr->agent_list, 2782 &mad_send_wr->mad_agent_priv->send_list); 2783 } 2784 return ret; 2785 } 2786 2787 static void timeout_sends(struct work_struct *work) 2788 { 2789 struct ib_mad_agent_private *mad_agent_priv; 2790 struct ib_mad_send_wr_private *mad_send_wr; 2791 struct ib_mad_send_wc mad_send_wc; 2792 unsigned long flags, delay; 2793 2794 mad_agent_priv = container_of(work, struct ib_mad_agent_private, 2795 timed_work.work); 2796 mad_send_wc.vendor_err = 0; 2797 2798 spin_lock_irqsave(&mad_agent_priv->lock, flags); 2799 while (!list_empty(&mad_agent_priv->wait_list)) { 2800 mad_send_wr = list_entry(mad_agent_priv->wait_list.next, 2801 struct ib_mad_send_wr_private, 2802 agent_list); 2803 2804 if (time_after(mad_send_wr->timeout, jiffies)) { 2805 delay = mad_send_wr->timeout - jiffies; 2806 if ((long)delay <= 0) 2807 delay = 1; 2808 queue_delayed_work(mad_agent_priv->qp_info-> 2809 port_priv->wq, 2810 &mad_agent_priv->timed_work, delay); 2811 break; 2812 } 2813 2814 list_del(&mad_send_wr->agent_list); 2815 if (mad_send_wr->status == IB_WC_SUCCESS && 2816 !retry_send(mad_send_wr)) 2817 continue; 2818 2819 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 2820 2821 if (mad_send_wr->status == IB_WC_SUCCESS) 2822 mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR; 2823 else 2824 mad_send_wc.status = mad_send_wr->status; 2825 mad_send_wc.send_buf = &mad_send_wr->send_buf; 2826 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent, 2827 &mad_send_wc); 2828 2829 atomic_dec(&mad_agent_priv->refcount); 2830 spin_lock_irqsave(&mad_agent_priv->lock, flags); 2831 } 2832 spin_unlock_irqrestore(&mad_agent_priv->lock, flags); 2833 } 2834 2835 /* 2836 * Allocate receive MADs and post receive WRs for them 2837 */ 2838 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info, 2839 struct ib_mad_private *mad) 2840 { 2841 unsigned long flags; 2842 int post, ret; 2843 struct ib_mad_private *mad_priv; 2844 struct ib_sge sg_list; 2845 struct ib_recv_wr recv_wr; 2846 const struct ib_recv_wr *bad_recv_wr; 2847 struct ib_mad_queue *recv_queue = &qp_info->recv_queue; 2848 2849 /* Initialize common scatter list fields */ 2850 sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey; 2851 2852 /* Initialize common receive WR fields */ 2853 recv_wr.next = NULL; 2854 recv_wr.sg_list = &sg_list; 2855 recv_wr.num_sge = 1; 2856 2857 do { 2858 /* Allocate and map receive buffer */ 2859 if (mad) { 2860 mad_priv = mad; 2861 mad = NULL; 2862 } else { 2863 mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv), 2864 GFP_ATOMIC); 2865 if (!mad_priv) { 2866 ret = -ENOMEM; 2867 break; 2868 } 2869 } 2870 sg_list.length = mad_priv_dma_size(mad_priv); 2871 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device, 2872 &mad_priv->grh, 2873 mad_priv_dma_size(mad_priv), 2874 DMA_FROM_DEVICE); 2875 if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device, 2876 sg_list.addr))) { 2877 kfree(mad_priv); 2878 ret = -ENOMEM; 2879 break; 2880 } 2881 mad_priv->header.mapping = sg_list.addr; 2882 mad_priv->header.mad_list.mad_queue = recv_queue; 2883 mad_priv->header.mad_list.cqe.done = ib_mad_recv_done; 2884 recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe; 2885 2886 /* Post receive WR */ 2887 spin_lock_irqsave(&recv_queue->lock, flags); 2888 post = (++recv_queue->count < recv_queue->max_active); 2889 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list); 2890 spin_unlock_irqrestore(&recv_queue->lock, flags); 2891 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr); 2892 if (ret) { 2893 spin_lock_irqsave(&recv_queue->lock, flags); 2894 list_del(&mad_priv->header.mad_list.list); 2895 recv_queue->count--; 2896 spin_unlock_irqrestore(&recv_queue->lock, flags); 2897 ib_dma_unmap_single(qp_info->port_priv->device, 2898 mad_priv->header.mapping, 2899 mad_priv_dma_size(mad_priv), 2900 DMA_FROM_DEVICE); 2901 kfree(mad_priv); 2902 dev_err(&qp_info->port_priv->device->dev, 2903 "ib_post_recv failed: %d\n", ret); 2904 break; 2905 } 2906 } while (post); 2907 2908 return ret; 2909 } 2910 2911 /* 2912 * Return all the posted receive MADs 2913 */ 2914 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info) 2915 { 2916 struct ib_mad_private_header *mad_priv_hdr; 2917 struct ib_mad_private *recv; 2918 struct ib_mad_list_head *mad_list; 2919 2920 if (!qp_info->qp) 2921 return; 2922 2923 while (!list_empty(&qp_info->recv_queue.list)) { 2924 2925 mad_list = list_entry(qp_info->recv_queue.list.next, 2926 struct ib_mad_list_head, list); 2927 mad_priv_hdr = container_of(mad_list, 2928 struct ib_mad_private_header, 2929 mad_list); 2930 recv = container_of(mad_priv_hdr, struct ib_mad_private, 2931 header); 2932 2933 /* Remove from posted receive MAD list */ 2934 list_del(&mad_list->list); 2935 2936 ib_dma_unmap_single(qp_info->port_priv->device, 2937 recv->header.mapping, 2938 mad_priv_dma_size(recv), 2939 DMA_FROM_DEVICE); 2940 kfree(recv); 2941 } 2942 2943 qp_info->recv_queue.count = 0; 2944 } 2945 2946 /* 2947 * Start the port 2948 */ 2949 static int ib_mad_port_start(struct ib_mad_port_private *port_priv) 2950 { 2951 int ret, i; 2952 struct ib_qp_attr *attr; 2953 struct ib_qp *qp; 2954 u16 pkey_index; 2955 2956 attr = kmalloc(sizeof *attr, GFP_KERNEL); 2957 if (!attr) 2958 return -ENOMEM; 2959 2960 ret = ib_find_pkey(port_priv->device, port_priv->port_num, 2961 IB_DEFAULT_PKEY_FULL, &pkey_index); 2962 if (ret) 2963 pkey_index = 0; 2964 2965 for (i = 0; i < IB_MAD_QPS_CORE; i++) { 2966 qp = port_priv->qp_info[i].qp; 2967 if (!qp) 2968 continue; 2969 2970 /* 2971 * PKey index for QP1 is irrelevant but 2972 * one is needed for the Reset to Init transition 2973 */ 2974 attr->qp_state = IB_QPS_INIT; 2975 attr->pkey_index = pkey_index; 2976 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY; 2977 ret = ib_modify_qp(qp, attr, IB_QP_STATE | 2978 IB_QP_PKEY_INDEX | IB_QP_QKEY); 2979 if (ret) { 2980 dev_err(&port_priv->device->dev, 2981 "Couldn't change QP%d state to INIT: %d\n", 2982 i, ret); 2983 goto out; 2984 } 2985 2986 attr->qp_state = IB_QPS_RTR; 2987 ret = ib_modify_qp(qp, attr, IB_QP_STATE); 2988 if (ret) { 2989 dev_err(&port_priv->device->dev, 2990 "Couldn't change QP%d state to RTR: %d\n", 2991 i, ret); 2992 goto out; 2993 } 2994 2995 attr->qp_state = IB_QPS_RTS; 2996 attr->sq_psn = IB_MAD_SEND_Q_PSN; 2997 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN); 2998 if (ret) { 2999 dev_err(&port_priv->device->dev, 3000 "Couldn't change QP%d state to RTS: %d\n", 3001 i, ret); 3002 goto out; 3003 } 3004 } 3005 3006 ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP); 3007 if (ret) { 3008 dev_err(&port_priv->device->dev, 3009 "Failed to request completion notification: %d\n", 3010 ret); 3011 goto out; 3012 } 3013 3014 for (i = 0; i < IB_MAD_QPS_CORE; i++) { 3015 if (!port_priv->qp_info[i].qp) 3016 continue; 3017 3018 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL); 3019 if (ret) { 3020 dev_err(&port_priv->device->dev, 3021 "Couldn't post receive WRs\n"); 3022 goto out; 3023 } 3024 } 3025 out: 3026 kfree(attr); 3027 return ret; 3028 } 3029 3030 static void qp_event_handler(struct ib_event *event, void *qp_context) 3031 { 3032 struct ib_mad_qp_info *qp_info = qp_context; 3033 3034 /* It's worse than that! He's dead, Jim! */ 3035 dev_err(&qp_info->port_priv->device->dev, 3036 "Fatal error (%d) on MAD QP (%d)\n", 3037 event->event, qp_info->qp->qp_num); 3038 } 3039 3040 static void init_mad_queue(struct ib_mad_qp_info *qp_info, 3041 struct ib_mad_queue *mad_queue) 3042 { 3043 mad_queue->qp_info = qp_info; 3044 mad_queue->count = 0; 3045 spin_lock_init(&mad_queue->lock); 3046 INIT_LIST_HEAD(&mad_queue->list); 3047 } 3048 3049 static void init_mad_qp(struct ib_mad_port_private *port_priv, 3050 struct ib_mad_qp_info *qp_info) 3051 { 3052 qp_info->port_priv = port_priv; 3053 init_mad_queue(qp_info, &qp_info->send_queue); 3054 init_mad_queue(qp_info, &qp_info->recv_queue); 3055 INIT_LIST_HEAD(&qp_info->overflow_list); 3056 spin_lock_init(&qp_info->snoop_lock); 3057 qp_info->snoop_table = NULL; 3058 qp_info->snoop_table_size = 0; 3059 atomic_set(&qp_info->snoop_count, 0); 3060 } 3061 3062 static int create_mad_qp(struct ib_mad_qp_info *qp_info, 3063 enum ib_qp_type qp_type) 3064 { 3065 struct ib_qp_init_attr qp_init_attr; 3066 int ret; 3067 3068 memset(&qp_init_attr, 0, sizeof qp_init_attr); 3069 qp_init_attr.send_cq = qp_info->port_priv->cq; 3070 qp_init_attr.recv_cq = qp_info->port_priv->cq; 3071 qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR; 3072 qp_init_attr.cap.max_send_wr = mad_sendq_size; 3073 qp_init_attr.cap.max_recv_wr = mad_recvq_size; 3074 qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG; 3075 qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG; 3076 qp_init_attr.qp_type = qp_type; 3077 qp_init_attr.port_num = qp_info->port_priv->port_num; 3078 qp_init_attr.qp_context = qp_info; 3079 qp_init_attr.event_handler = qp_event_handler; 3080 qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr); 3081 if (IS_ERR(qp_info->qp)) { 3082 dev_err(&qp_info->port_priv->device->dev, 3083 "Couldn't create ib_mad QP%d\n", 3084 get_spl_qp_index(qp_type)); 3085 ret = PTR_ERR(qp_info->qp); 3086 goto error; 3087 } 3088 /* Use minimum queue sizes unless the CQ is resized */ 3089 qp_info->send_queue.max_active = mad_sendq_size; 3090 qp_info->recv_queue.max_active = mad_recvq_size; 3091 return 0; 3092 3093 error: 3094 return ret; 3095 } 3096 3097 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info) 3098 { 3099 if (!qp_info->qp) 3100 return; 3101 3102 ib_destroy_qp(qp_info->qp); 3103 kfree(qp_info->snoop_table); 3104 } 3105 3106 /* 3107 * Open the port 3108 * Create the QP, PD, MR, and CQ if needed 3109 */ 3110 static int ib_mad_port_open(struct ib_device *device, 3111 int port_num) 3112 { 3113 int ret, cq_size; 3114 struct ib_mad_port_private *port_priv; 3115 unsigned long flags; 3116 char name[sizeof "ib_mad123"]; 3117 int has_smi; 3118 3119 if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE)) 3120 return -EFAULT; 3121 3122 if (WARN_ON(rdma_cap_opa_mad(device, port_num) && 3123 rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE)) 3124 return -EFAULT; 3125 3126 /* Create new device info */ 3127 port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL); 3128 if (!port_priv) 3129 return -ENOMEM; 3130 3131 port_priv->device = device; 3132 port_priv->port_num = port_num; 3133 spin_lock_init(&port_priv->reg_lock); 3134 INIT_LIST_HEAD(&port_priv->agent_list); 3135 init_mad_qp(port_priv, &port_priv->qp_info[0]); 3136 init_mad_qp(port_priv, &port_priv->qp_info[1]); 3137 3138 cq_size = mad_sendq_size + mad_recvq_size; 3139 has_smi = rdma_cap_ib_smi(device, port_num); 3140 if (has_smi) 3141 cq_size *= 2; 3142 3143 port_priv->pd = ib_alloc_pd(device, 0); 3144 if (IS_ERR(port_priv->pd)) { 3145 dev_err(&device->dev, "Couldn't create ib_mad PD\n"); 3146 ret = PTR_ERR(port_priv->pd); 3147 goto error3; 3148 } 3149 3150 port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0, 3151 IB_POLL_WORKQUEUE); 3152 if (IS_ERR(port_priv->cq)) { 3153 dev_err(&device->dev, "Couldn't create ib_mad CQ\n"); 3154 ret = PTR_ERR(port_priv->cq); 3155 goto error4; 3156 } 3157 3158 if (has_smi) { 3159 ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI); 3160 if (ret) 3161 goto error6; 3162 } 3163 ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI); 3164 if (ret) 3165 goto error7; 3166 3167 snprintf(name, sizeof name, "ib_mad%d", port_num); 3168 port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM); 3169 if (!port_priv->wq) { 3170 ret = -ENOMEM; 3171 goto error8; 3172 } 3173 3174 spin_lock_irqsave(&ib_mad_port_list_lock, flags); 3175 list_add_tail(&port_priv->port_list, &ib_mad_port_list); 3176 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags); 3177 3178 ret = ib_mad_port_start(port_priv); 3179 if (ret) { 3180 dev_err(&device->dev, "Couldn't start port\n"); 3181 goto error9; 3182 } 3183 3184 return 0; 3185 3186 error9: 3187 spin_lock_irqsave(&ib_mad_port_list_lock, flags); 3188 list_del_init(&port_priv->port_list); 3189 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags); 3190 3191 destroy_workqueue(port_priv->wq); 3192 error8: 3193 destroy_mad_qp(&port_priv->qp_info[1]); 3194 error7: 3195 destroy_mad_qp(&port_priv->qp_info[0]); 3196 error6: 3197 ib_free_cq(port_priv->cq); 3198 cleanup_recv_queue(&port_priv->qp_info[1]); 3199 cleanup_recv_queue(&port_priv->qp_info[0]); 3200 error4: 3201 ib_dealloc_pd(port_priv->pd); 3202 error3: 3203 kfree(port_priv); 3204 3205 return ret; 3206 } 3207 3208 /* 3209 * Close the port 3210 * If there are no classes using the port, free the port 3211 * resources (CQ, MR, PD, QP) and remove the port's info structure 3212 */ 3213 static int ib_mad_port_close(struct ib_device *device, int port_num) 3214 { 3215 struct ib_mad_port_private *port_priv; 3216 unsigned long flags; 3217 3218 spin_lock_irqsave(&ib_mad_port_list_lock, flags); 3219 port_priv = __ib_get_mad_port(device, port_num); 3220 if (port_priv == NULL) { 3221 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags); 3222 dev_err(&device->dev, "Port %d not found\n", port_num); 3223 return -ENODEV; 3224 } 3225 list_del_init(&port_priv->port_list); 3226 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags); 3227 3228 destroy_workqueue(port_priv->wq); 3229 destroy_mad_qp(&port_priv->qp_info[1]); 3230 destroy_mad_qp(&port_priv->qp_info[0]); 3231 ib_free_cq(port_priv->cq); 3232 ib_dealloc_pd(port_priv->pd); 3233 cleanup_recv_queue(&port_priv->qp_info[1]); 3234 cleanup_recv_queue(&port_priv->qp_info[0]); 3235 /* XXX: Handle deallocation of MAD registration tables */ 3236 3237 kfree(port_priv); 3238 3239 return 0; 3240 } 3241 3242 static void ib_mad_init_device(struct ib_device *device) 3243 { 3244 int start, i; 3245 3246 start = rdma_start_port(device); 3247 3248 for (i = start; i <= rdma_end_port(device); i++) { 3249 if (!rdma_cap_ib_mad(device, i)) 3250 continue; 3251 3252 if (ib_mad_port_open(device, i)) { 3253 dev_err(&device->dev, "Couldn't open port %d\n", i); 3254 goto error; 3255 } 3256 if (ib_agent_port_open(device, i)) { 3257 dev_err(&device->dev, 3258 "Couldn't open port %d for agents\n", i); 3259 goto error_agent; 3260 } 3261 } 3262 return; 3263 3264 error_agent: 3265 if (ib_mad_port_close(device, i)) 3266 dev_err(&device->dev, "Couldn't close port %d\n", i); 3267 3268 error: 3269 while (--i >= start) { 3270 if (!rdma_cap_ib_mad(device, i)) 3271 continue; 3272 3273 if (ib_agent_port_close(device, i)) 3274 dev_err(&device->dev, 3275 "Couldn't close port %d for agents\n", i); 3276 if (ib_mad_port_close(device, i)) 3277 dev_err(&device->dev, "Couldn't close port %d\n", i); 3278 } 3279 } 3280 3281 static void ib_mad_remove_device(struct ib_device *device, void *client_data) 3282 { 3283 int i; 3284 3285 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) { 3286 if (!rdma_cap_ib_mad(device, i)) 3287 continue; 3288 3289 if (ib_agent_port_close(device, i)) 3290 dev_err(&device->dev, 3291 "Couldn't close port %d for agents\n", i); 3292 if (ib_mad_port_close(device, i)) 3293 dev_err(&device->dev, "Couldn't close port %d\n", i); 3294 } 3295 } 3296 3297 static struct ib_client mad_client = { 3298 .name = "mad", 3299 .add = ib_mad_init_device, 3300 .remove = ib_mad_remove_device 3301 }; 3302 3303 int ib_mad_init(void) 3304 { 3305 mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE); 3306 mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE); 3307 3308 mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE); 3309 mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE); 3310 3311 INIT_LIST_HEAD(&ib_mad_port_list); 3312 3313 if (ib_register_client(&mad_client)) { 3314 pr_err("Couldn't register ib_mad client\n"); 3315 return -EINVAL; 3316 } 3317 3318 return 0; 3319 } 3320 3321 void ib_mad_cleanup(void) 3322 { 3323 ib_unregister_client(&mad_client); 3324 } 3325