drivers/infiniband/ulp/srpt/ib_srpt.c

   1 /*
   2  * Copyright (c) 2006 - 2009 Mellanox Technology Inc.  All rights reserved.
   3  * Copyright (C) 2008 - 2011 Bart Van Assche <bvanassche@acm.org>.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  *
  33  */
  34
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/slab.h>
  38 #include <linux/err.h>
  39 #include <linux/ctype.h>
  40 #include <linux/kthread.h>
  41 #include <linux/string.h>
  42 #include <linux/delay.h>
  43 #include <linux/atomic.h>
  44 #include <scsi/scsi_tcq.h>
  45 #include <target/configfs_macros.h>
  46 #include <target/target_core_base.h>
  47 #include <target/target_core_fabric_configfs.h>
  48 #include <target/target_core_fabric.h>
  49 #include <target/target_core_configfs.h>
  50 #include "ib_srpt.h"
  51
  52 /* Name of this kernel module. */
  53 #define DRV_NAME                "ib_srpt"
  54 #define DRV_VERSION             "2.0.0"
  55 #define DRV_RELDATE             "2011-02-14"
  56
  57 #define SRPT_ID_STRING  "Linux SRP target"
  58
  59 #undef pr_fmt
  60 #define pr_fmt(fmt) DRV_NAME " " fmt
  61
  62 MODULE_AUTHOR("Vu Pham and Bart Van Assche");
  63 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol target "
  64                    "v" DRV_VERSION " (" DRV_RELDATE ")");
  65 MODULE_LICENSE("Dual BSD/GPL");
  66
  67 /*
  68  * Global Variables
  69  */
  70
  71 static u64 srpt_service_guid;
  72 static DEFINE_SPINLOCK(srpt_dev_lock);  /* Protects srpt_dev_list. */
  73 static LIST_HEAD(srpt_dev_list);        /* List of srpt_device structures. */
  74
  75 static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE;
  76 module_param(srp_max_req_size, int, 0444);
  77 MODULE_PARM_DESC(srp_max_req_size,
  78                  "Maximum size of SRP request messages in bytes.");
  79
  80 static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE;
  81 module_param(srpt_srq_size, int, 0444);
  82 MODULE_PARM_DESC(srpt_srq_size,
  83                  "Shared receive queue (SRQ) size.");
  84
  85 static int srpt_get_u64_x(char *buffer, struct kernel_param *kp)
  86 {
  87         return sprintf(buffer, "0x%016llx", *(u64 *)kp->arg);
  88 }
  89 module_param_call(srpt_service_guid, NULL, srpt_get_u64_x, &srpt_service_guid,
  90                   0444);
  91 MODULE_PARM_DESC(srpt_service_guid,
  92                  "Using this value for ioc_guid, id_ext, and cm_listen_id"
  93                  " instead of using the node_guid of the first HCA.");
  94
  95 static struct ib_client srpt_client;
  96 static struct target_fabric_configfs *srpt_target;
  97 static void srpt_release_channel(struct srpt_rdma_ch *ch);
  98 static int srpt_queue_status(struct se_cmd *cmd);
  99
 100 /**
 101  * opposite_dma_dir() - Swap DMA_TO_DEVICE and DMA_FROM_DEVICE.
 102  */
 103 static inline
 104 enum dma_data_direction opposite_dma_dir(enum dma_data_direction dir)
 105 {
 106         switch (dir) {
 107         case DMA_TO_DEVICE:     return DMA_FROM_DEVICE;
 108         case DMA_FROM_DEVICE:   return DMA_TO_DEVICE;
 109         default:                return dir;
 110         }
 111 }
 112
 113 /**
 114  * srpt_sdev_name() - Return the name associated with the HCA.
 115  *
 116  * Examples are ib0, ib1, ...
 117  */
 118 static inline const char *srpt_sdev_name(struct srpt_device *sdev)
 119 {
 120         return sdev->device->name;
 121 }
 122
 123 static enum rdma_ch_state srpt_get_ch_state(struct srpt_rdma_ch *ch)
 124 {
 125         unsigned long flags;
 126         enum rdma_ch_state state;
 127
 128         spin_lock_irqsave(&ch->spinlock, flags);
 129         state = ch->state;
 130         spin_unlock_irqrestore(&ch->spinlock, flags);
 131         return state;
 132 }
 133
 134 static enum rdma_ch_state
 135 srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new_state)
 136 {
 137         unsigned long flags;
 138         enum rdma_ch_state prev;
 139
 140         spin_lock_irqsave(&ch->spinlock, flags);
 141         prev = ch->state;
 142         ch->state = new_state;
 143         spin_unlock_irqrestore(&ch->spinlock, flags);
 144         return prev;
 145 }
 146
 147 /**
 148  * srpt_test_and_set_ch_state() - Test and set the channel state.
 149  *
 150  * Returns true if and only if the channel state has been set to the new state.
 151  */
 152 static bool
 153 srpt_test_and_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state old,
 154                            enum rdma_ch_state new)
 155 {
 156         unsigned long flags;
 157         enum rdma_ch_state prev;
 158
 159         spin_lock_irqsave(&ch->spinlock, flags);
 160         prev = ch->state;
 161         if (prev == old)
 162                 ch->state = new;
 163         spin_unlock_irqrestore(&ch->spinlock, flags);
 164         return prev == old;
 165 }
 166
 167 /**
 168  * srpt_event_handler() - Asynchronous IB event callback function.
 169  *
 170  * Callback function called by the InfiniBand core when an asynchronous IB
 171  * event occurs. This callback may occur in interrupt context. See also
 172  * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand
 173  * Architecture Specification.
 174  */
 175 static void srpt_event_handler(struct ib_event_handler *handler,
 176                                struct ib_event *event)
 177 {
 178         struct srpt_device *sdev;
 179         struct srpt_port *sport;
 180
 181         sdev = ib_get_client_data(event->device, &srpt_client);
 182         if (!sdev || sdev->device != event->device)
 183                 return;
 184
 185         pr_debug("ASYNC event= %d on device= %s\n", event->event,
 186                  srpt_sdev_name(sdev));
 187
 188         switch (event->event) {
 189         case IB_EVENT_PORT_ERR:
 190                 if (event->element.port_num <= sdev->device->phys_port_cnt) {
 191                         sport = &sdev->port[event->element.port_num - 1];
 192                         sport->lid = 0;
 193                         sport->sm_lid = 0;
 194                 }
 195                 break;
 196         case IB_EVENT_PORT_ACTIVE:
 197         case IB_EVENT_LID_CHANGE:
 198         case IB_EVENT_PKEY_CHANGE:
 199         case IB_EVENT_SM_CHANGE:
 200         case IB_EVENT_CLIENT_REREGISTER:
 201                 /* Refresh port data asynchronously. */
 202                 if (event->element.port_num <= sdev->device->phys_port_cnt) {
 203                         sport = &sdev->port[event->element.port_num - 1];
 204                         if (!sport->lid && !sport->sm_lid)
 205                                 schedule_work(&sport->work);
 206                 }
 207                 break;
 208         default:
 209                 printk(KERN_ERR "received unrecognized IB event %d\n",
 210                        event->event);
 211                 break;
 212         }
 213 }
 214
 215 /**
 216  * srpt_srq_event() - SRQ event callback function.
 217  */
 218 static void srpt_srq_event(struct ib_event *event, void *ctx)
 219 {
 220         printk(KERN_INFO "SRQ event %d\n", event->event);
 221 }
 222
 223 /**
 224  * srpt_qp_event() - QP event callback function.
 225  */
 226 static void srpt_qp_event(struct ib_event *event, struct srpt_rdma_ch *ch)
 227 {
 228         pr_debug("QP event %d on cm_id=%p sess_name=%s state=%d\n",
 229                  event->event, ch->cm_id, ch->sess_name, srpt_get_ch_state(ch));
 230
 231         switch (event->event) {
 232         case IB_EVENT_COMM_EST:
 233                 ib_cm_notify(ch->cm_id, event->event);
 234                 break;
 235         case IB_EVENT_QP_LAST_WQE_REACHED:
 236                 if (srpt_test_and_set_ch_state(ch, CH_DRAINING,
 237                                                CH_RELEASING))
 238                         srpt_release_channel(ch);
 239                 else
 240                         pr_debug("%s: state %d - ignored LAST_WQE.\n",
 241                                  ch->sess_name, srpt_get_ch_state(ch));
 242                 break;
 243         default:
 244                 printk(KERN_ERR "received unrecognized IB QP event %d\n",
 245                        event->event);
 246                 break;
 247         }
 248 }
 249
 250 /**
 251  * srpt_set_ioc() - Helper function for initializing an IOUnitInfo structure.
 252  *
 253  * @slot: one-based slot number.
 254  * @value: four-bit value.
 255  *
 256  * Copies the lowest four bits of value in element slot of the array of four
 257  * bit elements called c_list (controller list). The index slot is one-based.
 258  */
 259 static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value)
 260 {
 261         u16 id;
 262         u8 tmp;
 263
 264         id = (slot - 1) / 2;
 265         if (slot & 0x1) {
 266                 tmp = c_list[id] & 0xf;
 267                 c_list[id] = (value << 4) | tmp;
 268         } else {
 269                 tmp = c_list[id] & 0xf0;
 270                 c_list[id] = (value & 0xf) | tmp;
 271         }
 272 }
 273
 274 /**
 275  * srpt_get_class_port_info() - Copy ClassPortInfo to a management datagram.
 276  *
 277  * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture
 278  * Specification.
 279  */
 280 static void srpt_get_class_port_info(struct ib_dm_mad *mad)
 281 {
 282         struct ib_class_port_info *cif;
 283
 284         cif = (struct ib_class_port_info *)mad->data;
 285         memset(cif, 0, sizeof *cif);
 286         cif->base_version = 1;
 287         cif->class_version = 1;
 288         cif->resp_time_value = 20;
 289
 290         mad->mad_hdr.status = 0;
 291 }
 292
 293 /**
 294  * srpt_get_iou() - Write IOUnitInfo to a management datagram.
 295  *
 296  * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture
 297  * Specification. See also section B.7, table B.6 in the SRP r16a document.
 298  */
 299 static void srpt_get_iou(struct ib_dm_mad *mad)
 300 {
 301         struct ib_dm_iou_info *ioui;
 302         u8 slot;
 303         int i;
 304
 305         ioui = (struct ib_dm_iou_info *)mad->data;
 306         ioui->change_id = __constant_cpu_to_be16(1);
 307         ioui->max_controllers = 16;
 308
 309         /* set present for slot 1 and empty for the rest */
 310         srpt_set_ioc(ioui->controller_list, 1, 1);
 311         for (i = 1, slot = 2; i < 16; i++, slot++)
 312                 srpt_set_ioc(ioui->controller_list, slot, 0);
 313
 314         mad->mad_hdr.status = 0;
 315 }
 316
 317 /**
 318  * srpt_get_ioc() - Write IOControllerprofile to a management datagram.
 319  *
 320  * See also section 16.3.3.4 IOControllerProfile in the InfiniBand
 321  * Architecture Specification. See also section B.7, table B.7 in the SRP
 322  * r16a document.
 323  */
 324 static void srpt_get_ioc(struct srpt_port *sport, u32 slot,
 325                          struct ib_dm_mad *mad)
 326 {
 327         struct srpt_device *sdev = sport->sdev;
 328         struct ib_dm_ioc_profile *iocp;
 329
 330         iocp = (struct ib_dm_ioc_profile *)mad->data;
 331
 332         if (!slot || slot > 16) {
 333                 mad->mad_hdr.status
 334                         = __constant_cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
 335                 return;
 336         }
 337
 338         if (slot > 2) {
 339                 mad->mad_hdr.status
 340                         = __constant_cpu_to_be16(DM_MAD_STATUS_NO_IOC);
 341                 return;
 342         }
 343
 344         memset(iocp, 0, sizeof *iocp);
 345         strcpy(iocp->id_string, SRPT_ID_STRING);
 346         iocp->guid = cpu_to_be64(srpt_service_guid);
 347         iocp->vendor_id = cpu_to_be32(sdev->dev_attr.vendor_id);
 348         iocp->device_id = cpu_to_be32(sdev->dev_attr.vendor_part_id);
 349         iocp->device_version = cpu_to_be16(sdev->dev_attr.hw_ver);
 350         iocp->subsys_vendor_id = cpu_to_be32(sdev->dev_attr.vendor_id);
 351         iocp->subsys_device_id = 0x0;
 352         iocp->io_class = __constant_cpu_to_be16(SRP_REV16A_IB_IO_CLASS);
 353         iocp->io_subclass = __constant_cpu_to_be16(SRP_IO_SUBCLASS);
 354         iocp->protocol = __constant_cpu_to_be16(SRP_PROTOCOL);
 355         iocp->protocol_version = __constant_cpu_to_be16(SRP_PROTOCOL_VERSION);
 356         iocp->send_queue_depth = cpu_to_be16(sdev->srq_size);
 357         iocp->rdma_read_depth = 4;
 358         iocp->send_size = cpu_to_be32(srp_max_req_size);
 359         iocp->rdma_size = cpu_to_be32(min(sport->port_attrib.srp_max_rdma_size,
 360                                           1U << 24));
 361         iocp->num_svc_entries = 1;
 362         iocp->op_cap_mask = SRP_SEND_TO_IOC | SRP_SEND_FROM_IOC |
 363                 SRP_RDMA_READ_FROM_IOC | SRP_RDMA_WRITE_FROM_IOC;
 364
 365         mad->mad_hdr.status = 0;
 366 }
 367
 368 /**
 369  * srpt_get_svc_entries() - Write ServiceEntries to a management datagram.
 370  *
 371  * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture
 372  * Specification. See also section B.7, table B.8 in the SRP r16a document.
 373  */
 374 static void srpt_get_svc_entries(u64 ioc_guid,
 375                                  u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad)
 376 {
 377         struct ib_dm_svc_entries *svc_entries;
 378
 379         WARN_ON(!ioc_guid);
 380
 381         if (!slot || slot > 16) {
 382                 mad->mad_hdr.status
 383                         = __constant_cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
 384                 return;
 385         }
 386
 387         if (slot > 2 || lo > hi || hi > 1) {
 388                 mad->mad_hdr.status
 389                         = __constant_cpu_to_be16(DM_MAD_STATUS_NO_IOC);
 390                 return;
 391         }
 392
 393         svc_entries = (struct ib_dm_svc_entries *)mad->data;
 394         memset(svc_entries, 0, sizeof *svc_entries);
 395         svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid);
 396         snprintf(svc_entries->service_entries[0].name,
 397                  sizeof(svc_entries->service_entries[0].name),
 398                  "%s%016llx",
 399                  SRP_SERVICE_NAME_PREFIX,
 400                  ioc_guid);
 401
 402         mad->mad_hdr.status = 0;
 403 }
 404
 405 /**
 406  * srpt_mgmt_method_get() - Process a received management datagram.
 407  * @sp:      source port through which the MAD has been received.
 408  * @rq_mad:  received MAD.
 409  * @rsp_mad: response MAD.
 410  */
 411 static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad,
 412                                  struct ib_dm_mad *rsp_mad)
 413 {
 414         u16 attr_id;
 415         u32 slot;
 416         u8 hi, lo;
 417
 418         attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id);
 419         switch (attr_id) {
 420         case DM_ATTR_CLASS_PORT_INFO:
 421                 srpt_get_class_port_info(rsp_mad);
 422                 break;
 423         case DM_ATTR_IOU_INFO:
 424                 srpt_get_iou(rsp_mad);
 425                 break;
 426         case DM_ATTR_IOC_PROFILE:
 427                 slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
 428                 srpt_get_ioc(sp, slot, rsp_mad);
 429                 break;
 430         case DM_ATTR_SVC_ENTRIES:
 431                 slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
 432                 hi = (u8) ((slot >> 8) & 0xff);
 433                 lo = (u8) (slot & 0xff);
 434                 slot = (u16) ((slot >> 16) & 0xffff);
 435                 srpt_get_svc_entries(srpt_service_guid,
 436                                      slot, hi, lo, rsp_mad);
 437                 break;
 438         default:
 439                 rsp_mad->mad_hdr.status =
 440                     __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
 441                 break;
 442         }
 443 }
 444
 445 /**
 446  * srpt_mad_send_handler() - Post MAD-send callback function.
 447  */
 448 static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent,
 449                                   struct ib_mad_send_wc *mad_wc)
 450 {
 451         ib_destroy_ah(mad_wc->send_buf->ah);
 452         ib_free_send_mad(mad_wc->send_buf);
 453 }
 454
 455 /**
 456  * srpt_mad_recv_handler() - MAD reception callback function.
 457  */
 458 static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent,
 459                                   struct ib_mad_recv_wc *mad_wc)
 460 {
 461         struct srpt_port *sport = (struct srpt_port *)mad_agent->context;
 462         struct ib_ah *ah;
 463         struct ib_mad_send_buf *rsp;
 464         struct ib_dm_mad *dm_mad;
 465
 466         if (!mad_wc || !mad_wc->recv_buf.mad)
 467                 return;
 468
 469         ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc,
 470                                   mad_wc->recv_buf.grh, mad_agent->port_num);
 471         if (IS_ERR(ah))
 472                 goto err;
 473
 474         BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR);
 475
 476         rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp,
 477                                  mad_wc->wc->pkey_index, 0,
 478                                  IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA,
 479                                  GFP_KERNEL);
 480         if (IS_ERR(rsp))
 481                 goto err_rsp;
 482
 483         rsp->ah = ah;
 484
 485         dm_mad = rsp->mad;
 486         memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof *dm_mad);
 487         dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
 488         dm_mad->mad_hdr.status = 0;
 489
 490         switch (mad_wc->recv_buf.mad->mad_hdr.method) {
 491         case IB_MGMT_METHOD_GET:
 492                 srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad);
 493                 break;
 494         case IB_MGMT_METHOD_SET:
 495                 dm_mad->mad_hdr.status =
 496                     __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
 497                 break;
 498         default:
 499                 dm_mad->mad_hdr.status =
 500                     __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD);
 501                 break;
 502         }
 503
 504         if (!ib_post_send_mad(rsp, NULL)) {
 505                 ib_free_recv_mad(mad_wc);
 506                 /* will destroy_ah & free_send_mad in send completion */
 507                 return;
 508         }
 509
 510         ib_free_send_mad(rsp);
 511
 512 err_rsp:
 513         ib_destroy_ah(ah);
 514 err:
 515         ib_free_recv_mad(mad_wc);
 516 }
 517
 518 /**
 519  * srpt_refresh_port() - Configure a HCA port.
 520  *
 521  * Enable InfiniBand management datagram processing, update the cached sm_lid,
 522  * lid and gid values, and register a callback function for processing MADs
 523  * on the specified port.
 524  *
 525  * Note: It is safe to call this function more than once for the same port.
 526  */
 527 static int srpt_refresh_port(struct srpt_port *sport)
 528 {
 529         struct ib_mad_reg_req reg_req;
 530         struct ib_port_modify port_modify;
 531         struct ib_port_attr port_attr;
 532         int ret;
 533
 534         memset(&port_modify, 0, sizeof port_modify);
 535         port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
 536         port_modify.clr_port_cap_mask = 0;
 537
 538         ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
 539         if (ret)
 540                 goto err_mod_port;
 541
 542         ret = ib_query_port(sport->sdev->device, sport->port, &port_attr);
 543         if (ret)
 544                 goto err_query_port;
 545
 546         sport->sm_lid = port_attr.sm_lid;
 547         sport->lid = port_attr.lid;
 548
 549         ret = ib_query_gid(sport->sdev->device, sport->port, 0, &sport->gid);
 550         if (ret)
 551                 goto err_query_port;
 552
 553         if (!sport->mad_agent) {
 554                 memset(&reg_req, 0, sizeof reg_req);
 555                 reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT;
 556                 reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION;
 557                 set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask);
 558                 set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask);
 559
 560                 sport->mad_agent = ib_register_mad_agent(sport->sdev->device,
 561                                                          sport->port,
 562                                                          IB_QPT_GSI,
 563                                                          &reg_req, 0,
 564                                                          srpt_mad_send_handler,
 565                                                          srpt_mad_recv_handler,
 566                                                          sport);
 567                 if (IS_ERR(sport->mad_agent)) {
 568                         ret = PTR_ERR(sport->mad_agent);
 569                         sport->mad_agent = NULL;
 570                         goto err_query_port;
 571                 }
 572         }
 573
 574         return 0;
 575
 576 err_query_port:
 577
 578         port_modify.set_port_cap_mask = 0;
 579         port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
 580         ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
 581
 582 err_mod_port:
 583
 584         return ret;
 585 }
 586
 587 /**
 588  * srpt_unregister_mad_agent() - Unregister MAD callback functions.
 589  *
 590  * Note: It is safe to call this function more than once for the same device.
 591  */
 592 static void srpt_unregister_mad_agent(struct srpt_device *sdev)
 593 {
 594         struct ib_port_modify port_modify = {
 595                 .clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP,
 596         };
 597         struct srpt_port *sport;
 598         int i;
 599
 600         for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
 601                 sport = &sdev->port[i - 1];
 602                 WARN_ON(sport->port != i);
 603                 if (ib_modify_port(sdev->device, i, 0, &port_modify) < 0)
 604                         printk(KERN_ERR "disabling MAD processing failed.\n");
 605                 if (sport->mad_agent) {
 606                         ib_unregister_mad_agent(sport->mad_agent);
 607                         sport->mad_agent = NULL;
 608                 }
 609         }
 610 }
 611
 612 /**
 613  * srpt_alloc_ioctx() - Allocate an SRPT I/O context structure.
 614  */
 615 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev,
 616                                            int ioctx_size, int dma_size,
 617                                            enum dma_data_direction dir)
 618 {
 619         struct srpt_ioctx *ioctx;
 620
 621         ioctx = kmalloc(ioctx_size, GFP_KERNEL);
 622         if (!ioctx)
 623                 goto err;
 624
 625         ioctx->buf = kmalloc(dma_size, GFP_KERNEL);
 626         if (!ioctx->buf)
 627                 goto err_free_ioctx;
 628
 629         ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf, dma_size, dir);
 630         if (ib_dma_mapping_error(sdev->device, ioctx->dma))
 631                 goto err_free_buf;
 632
 633         return ioctx;
 634
 635 err_free_buf:
 636         kfree(ioctx->buf);
 637 err_free_ioctx:
 638         kfree(ioctx);
 639 err:
 640         return NULL;
 641 }
 642
 643 /**
 644  * srpt_free_ioctx() - Free an SRPT I/O context structure.
 645  */
 646 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx,
 647                             int dma_size, enum dma_data_direction dir)
 648 {
 649         if (!ioctx)
 650                 return;
 651
 652         ib_dma_unmap_single(sdev->device, ioctx->dma, dma_size, dir);
 653         kfree(ioctx->buf);
 654         kfree(ioctx);
 655 }
 656
 657 /**
 658  * srpt_alloc_ioctx_ring() - Allocate a ring of SRPT I/O context structures.
 659  * @sdev:       Device to allocate the I/O context ring for.
 660  * @ring_size:  Number of elements in the I/O context ring.
 661  * @ioctx_size: I/O context size.
 662  * @dma_size:   DMA buffer size.
 663  * @dir:        DMA data direction.
 664  */
 665 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev,
 666                                 int ring_size, int ioctx_size,
 667                                 int dma_size, enum dma_data_direction dir)
 668 {
 669         struct srpt_ioctx **ring;
 670         int i;
 671
 672         WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx)
 673                 && ioctx_size != sizeof(struct srpt_send_ioctx));
 674
 675         ring = kmalloc(ring_size * sizeof(ring[0]), GFP_KERNEL);
 676         if (!ring)
 677                 goto out;
 678         for (i = 0; i < ring_size; ++i) {
 679                 ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, dma_size, dir);
 680                 if (!ring[i])
 681                         goto err;
 682                 ring[i]->index = i;
 683         }
 684         goto out;
 685
 686 err:
 687         while (--i >= 0)
 688                 srpt_free_ioctx(sdev, ring[i], dma_size, dir);
 689         kfree(ring);
 690         ring = NULL;
 691 out:
 692         return ring;
 693 }
 694
 695 /**
 696  * srpt_free_ioctx_ring() - Free the ring of SRPT I/O context structures.
 697  */
 698 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring,
 699                                  struct srpt_device *sdev, int ring_size,
 700                                  int dma_size, enum dma_data_direction dir)
 701 {
 702         int i;
 703
 704         for (i = 0; i < ring_size; ++i)
 705                 srpt_free_ioctx(sdev, ioctx_ring[i], dma_size, dir);
 706         kfree(ioctx_ring);
 707 }
 708
 709 /**
 710  * srpt_get_cmd_state() - Get the state of a SCSI command.
 711  */
 712 static enum srpt_command_state srpt_get_cmd_state(struct srpt_send_ioctx *ioctx)
 713 {
 714         enum srpt_command_state state;
 715         unsigned long flags;
 716
 717         BUG_ON(!ioctx);
 718
 719         spin_lock_irqsave(&ioctx->spinlock, flags);
 720         state = ioctx->state;
 721         spin_unlock_irqrestore(&ioctx->spinlock, flags);
 722         return state;
 723 }
 724
 725 /**
 726  * srpt_set_cmd_state() - Set the state of a SCSI command.
 727  *
 728  * Does not modify the state of aborted commands. Returns the previous command
 729  * state.
 730  */
 731 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx,
 732                                                   enum srpt_command_state new)
 733 {
 734         enum srpt_command_state previous;
 735         unsigned long flags;
 736
 737         BUG_ON(!ioctx);
 738
 739         spin_lock_irqsave(&ioctx->spinlock, flags);
 740         previous = ioctx->state;
 741         if (previous != SRPT_STATE_DONE)
 742                 ioctx->state = new;
 743         spin_unlock_irqrestore(&ioctx->spinlock, flags);
 744
 745         return previous;
 746 }
 747
 748 /**
 749  * srpt_test_and_set_cmd_state() - Test and set the state of a command.
 750  *
 751  * Returns true if and only if the previous command state was equal to 'old'.
 752  */
 753 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx,
 754                                         enum srpt_command_state old,
 755                                         enum srpt_command_state new)
 756 {
 757         enum srpt_command_state previous;
 758         unsigned long flags;
 759
 760         WARN_ON(!ioctx);
 761         WARN_ON(old == SRPT_STATE_DONE);
 762         WARN_ON(new == SRPT_STATE_NEW);
 763
 764         spin_lock_irqsave(&ioctx->spinlock, flags);
 765         previous = ioctx->state;
 766         if (previous == old)
 767                 ioctx->state = new;
 768         spin_unlock_irqrestore(&ioctx->spinlock, flags);
 769         return previous == old;
 770 }
 771
 772 /**
 773  * srpt_post_recv() - Post an IB receive request.
 774  */
 775 static int srpt_post_recv(struct srpt_device *sdev,
 776                           struct srpt_recv_ioctx *ioctx)
 777 {
 778         struct ib_sge list;
 779         struct ib_recv_wr wr, *bad_wr;
 780
 781         BUG_ON(!sdev);
 782         wr.wr_id = encode_wr_id(SRPT_RECV, ioctx->ioctx.index);
 783
 784         list.addr = ioctx->ioctx.dma;
 785         list.length = srp_max_req_size;
 786         list.lkey = sdev->mr->lkey;
 787
 788         wr.next = NULL;
 789         wr.sg_list = &list;
 790         wr.num_sge = 1;
 791
 792         return ib_post_srq_recv(sdev->srq, &wr, &bad_wr);
 793 }
 794
 795 /**
 796  * srpt_post_send() - Post an IB send request.
 797  *
 798  * Returns zero upon success and a non-zero value upon failure.
 799  */
 800 static int srpt_post_send(struct srpt_rdma_ch *ch,
 801                           struct srpt_send_ioctx *ioctx, int len)
 802 {
 803         struct ib_sge list;
 804         struct ib_send_wr wr, *bad_wr;
 805         struct srpt_device *sdev = ch->sport->sdev;
 806         int ret;
 807
 808         atomic_inc(&ch->req_lim);
 809
 810         ret = -ENOMEM;
 811         if (unlikely(atomic_dec_return(&ch->sq_wr_avail) < 0)) {
 812                 printk(KERN_WARNING "IB send queue full (needed 1)\n");
 813                 goto out;
 814         }
 815
 816         ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, len,
 817                                       DMA_TO_DEVICE);
 818
 819         list.addr = ioctx->ioctx.dma;
 820         list.length = len;
 821         list.lkey = sdev->mr->lkey;
 822
 823         wr.next = NULL;
 824         wr.wr_id = encode_wr_id(SRPT_SEND, ioctx->ioctx.index);
 825         wr.sg_list = &list;
 826         wr.num_sge = 1;
 827         wr.opcode = IB_WR_SEND;
 828         wr.send_flags = IB_SEND_SIGNALED;
 829
 830         ret = ib_post_send(ch->qp, &wr, &bad_wr);
 831
 832 out:
 833         if (ret < 0) {
 834                 atomic_inc(&ch->sq_wr_avail);
 835                 atomic_dec(&ch->req_lim);
 836         }
 837         return ret;
 838 }
 839
 840 /**
 841  * srpt_get_desc_tbl() - Parse the data descriptors of an SRP_CMD request.
 842  * @ioctx: Pointer to the I/O context associated with the request.
 843  * @srp_cmd: Pointer to the SRP_CMD request data.
 844  * @dir: Pointer to the variable to which the transfer direction will be
 845  *   written.
 846  * @data_len: Pointer to the variable to which the total data length of all
 847  *   descriptors in the SRP_CMD request will be written.
 848  *
 849  * This function initializes ioctx->nrbuf and ioctx->r_bufs.
 850  *
 851  * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors;
 852  * -ENOMEM when memory allocation fails and zero upon success.
 853  */
 854 static int srpt_get_desc_tbl(struct srpt_send_ioctx *ioctx,
 855                              struct srp_cmd *srp_cmd,
 856                              enum dma_data_direction *dir, u64 *data_len)
 857 {
 858         struct srp_indirect_buf *idb;
 859         struct srp_direct_buf *db;
 860         unsigned add_cdb_offset;
 861         int ret;
 862
 863         /*
 864          * The pointer computations below will only be compiled correctly
 865          * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check
 866          * whether srp_cmd::add_data has been declared as a byte pointer.
 867          */
 868         BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0)
 869                      && !__same_type(srp_cmd->add_data[0], (u8)0));
 870
 871         BUG_ON(!dir);
 872         BUG_ON(!data_len);
 873
 874         ret = 0;
 875         *data_len = 0;
 876
 877         /*
 878          * The lower four bits of the buffer format field contain the DATA-IN
 879          * buffer descriptor format, and the highest four bits contain the
 880          * DATA-OUT buffer descriptor format.
 881          */
 882         *dir = DMA_NONE;
 883         if (srp_cmd->buf_fmt & 0xf)
 884                 /* DATA-IN: transfer data from target to initiator (read). */
 885                 *dir = DMA_FROM_DEVICE;
 886         else if (srp_cmd->buf_fmt >> 4)
 887                 /* DATA-OUT: transfer data from initiator to target (write). */
 888                 *dir = DMA_TO_DEVICE;
 889
 890         /*
 891          * According to the SRP spec, the lower two bits of the 'ADDITIONAL
 892          * CDB LENGTH' field are reserved and the size in bytes of this field
 893          * is four times the value specified in bits 3..7. Hence the "& ~3".
 894          */
 895         add_cdb_offset = srp_cmd->add_cdb_len & ~3;
 896         if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) ||
 897             ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) {
 898                 ioctx->n_rbuf = 1;
 899                 ioctx->rbufs = &ioctx->single_rbuf;
 900
 901                 db = (struct srp_direct_buf *)(srp_cmd->add_data
 902                                                + add_cdb_offset);
 903                 memcpy(ioctx->rbufs, db, sizeof *db);
 904                 *data_len = be32_to_cpu(db->len);
 905         } else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) ||
 906                    ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) {
 907                 idb = (struct srp_indirect_buf *)(srp_cmd->add_data
 908                                                   + add_cdb_offset);
 909
 910                 ioctx->n_rbuf = be32_to_cpu(idb->table_desc.len) / sizeof *db;
 911
 912                 if (ioctx->n_rbuf >
 913                     (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) {
 914                         printk(KERN_ERR "received unsupported SRP_CMD request"
 915                                " type (%u out + %u in != %u / %zu)\n",
 916                                srp_cmd->data_out_desc_cnt,
 917                                srp_cmd->data_in_desc_cnt,
 918                                be32_to_cpu(idb->table_desc.len),
 919                                sizeof(*db));
 920                         ioctx->n_rbuf = 0;
 921                         ret = -EINVAL;
 922                         goto out;
 923                 }
 924
 925                 if (ioctx->n_rbuf == 1)
 926                         ioctx->rbufs = &ioctx->single_rbuf;
 927                 else {
 928                         ioctx->rbufs =
 929                                 kmalloc(ioctx->n_rbuf * sizeof *db, GFP_ATOMIC);
 930                         if (!ioctx->rbufs) {
 931                                 ioctx->n_rbuf = 0;
 932                                 ret = -ENOMEM;
 933                                 goto out;
 934                         }
 935                 }
 936
 937                 db = idb->desc_list;
 938                 memcpy(ioctx->rbufs, db, ioctx->n_rbuf * sizeof *db);
 939                 *data_len = be32_to_cpu(idb->len);
 940         }
 941 out:
 942         return ret;
 943 }
 944
 945 /**
 946  * srpt_init_ch_qp() - Initialize queue pair attributes.
 947  *
 948  * Initialized the attributes of queue pair 'qp' by allowing local write,
 949  * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT.
 950  */
 951 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp)
 952 {
 953         struct ib_qp_attr *attr;
 954         int ret;
 955
 956         attr = kzalloc(sizeof *attr, GFP_KERNEL);
 957         if (!attr)
 958                 return -ENOMEM;
 959
 960         attr->qp_state = IB_QPS_INIT;
 961         attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
 962             IB_ACCESS_REMOTE_WRITE;
 963         attr->port_num = ch->sport->port;
 964         attr->pkey_index = 0;
 965
 966         ret = ib_modify_qp(qp, attr,
 967                            IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT |
 968                            IB_QP_PKEY_INDEX);
 969
 970         kfree(attr);
 971         return ret;
 972 }
 973
 974 /**
 975  * srpt_ch_qp_rtr() - Change the state of a channel to 'ready to receive' (RTR).
 976  * @ch: channel of the queue pair.
 977  * @qp: queue pair to change the state of.
 978  *
 979  * Returns zero upon success and a negative value upon failure.
 980  *
 981  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
 982  * If this structure ever becomes larger, it might be necessary to allocate
 983  * it dynamically instead of on the stack.
 984  */
 985 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp)
 986 {
 987         struct ib_qp_attr qp_attr;
 988         int attr_mask;
 989         int ret;
 990
 991         qp_attr.qp_state = IB_QPS_RTR;
 992         ret = ib_cm_init_qp_attr(ch->cm_id, &qp_attr, &attr_mask);
 993         if (ret)
 994                 goto out;
 995
 996         qp_attr.max_dest_rd_atomic = 4;
 997
 998         ret = ib_modify_qp(qp, &qp_attr, attr_mask);
 999
1000 out:
1001         return ret;
1002 }
1003
1004 /**
1005  * srpt_ch_qp_rts() - Change the state of a channel to 'ready to send' (RTS).
1006  * @ch: channel of the queue pair.
1007  * @qp: queue pair to change the state of.
1008  *
1009  * Returns zero upon success and a negative value upon failure.
1010  *
1011  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1012  * If this structure ever becomes larger, it might be necessary to allocate
1013  * it dynamically instead of on the stack.
1014  */
1015 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1016 {
1017         struct ib_qp_attr qp_attr;
1018         int attr_mask;
1019         int ret;
1020
1021         qp_attr.qp_state = IB_QPS_RTS;
1022         ret = ib_cm_init_qp_attr(ch->cm_id, &qp_attr, &attr_mask);
1023         if (ret)
1024                 goto out;
1025
1026         qp_attr.max_rd_atomic = 4;
1027
1028         ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1029
1030 out:
1031         return ret;
1032 }
1033
1034 /**
1035  * srpt_ch_qp_err() - Set the channel queue pair state to 'error'.
1036  */
1037 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch)
1038 {
1039         struct ib_qp_attr qp_attr;
1040
1041         qp_attr.qp_state = IB_QPS_ERR;
1042         return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE);
1043 }
1044
1045 /**
1046  * srpt_unmap_sg_to_ib_sge() - Unmap an IB SGE list.
1047  */
1048 static void srpt_unmap_sg_to_ib_sge(struct srpt_rdma_ch *ch,
1049                                     struct srpt_send_ioctx *ioctx)
1050 {
1051         struct scatterlist *sg;
1052         enum dma_data_direction dir;
1053
1054         BUG_ON(!ch);
1055         BUG_ON(!ioctx);
1056         BUG_ON(ioctx->n_rdma && !ioctx->rdma_ius);
1057
1058         while (ioctx->n_rdma)
1059                 kfree(ioctx->rdma_ius[--ioctx->n_rdma].sge);
1060
1061         kfree(ioctx->rdma_ius);
1062         ioctx->rdma_ius = NULL;
1063
1064         if (ioctx->mapped_sg_count) {
1065                 sg = ioctx->sg;
1066                 WARN_ON(!sg);
1067                 dir = ioctx->cmd.data_direction;
1068                 BUG_ON(dir == DMA_NONE);
1069                 ib_dma_unmap_sg(ch->sport->sdev->device, sg, ioctx->sg_cnt,
1070                                 opposite_dma_dir(dir));
1071                 ioctx->mapped_sg_count = 0;
1072         }
1073 }
1074
1075 /**
1076  * srpt_map_sg_to_ib_sge() - Map an SG list to an IB SGE list.
1077  */
1078 static int srpt_map_sg_to_ib_sge(struct srpt_rdma_ch *ch,
1079                                  struct srpt_send_ioctx *ioctx)
1080 {
1081         struct se_cmd *cmd;
1082         struct scatterlist *sg, *sg_orig;
1083         int sg_cnt;
1084         enum dma_data_direction dir;
1085         struct rdma_iu *riu;
1086         struct srp_direct_buf *db;
1087         dma_addr_t dma_addr;
1088         struct ib_sge *sge;
1089         u64 raddr;
1090         u32 rsize;
1091         u32 tsize;
1092         u32 dma_len;
1093         int count, nrdma;
1094         int i, j, k;
1095
1096         BUG_ON(!ch);
1097         BUG_ON(!ioctx);
1098         cmd = &ioctx->cmd;
1099         dir = cmd->data_direction;
1100         BUG_ON(dir == DMA_NONE);
1101
1102         ioctx->sg = sg = sg_orig = cmd->t_data_sg;
1103         ioctx->sg_cnt = sg_cnt = cmd->t_data_nents;
1104
1105         count = ib_dma_map_sg(ch->sport->sdev->device, sg, sg_cnt,
1106                               opposite_dma_dir(dir));
1107         if (unlikely(!count))
1108                 return -EAGAIN;
1109
1110         ioctx->mapped_sg_count = count;
1111
1112         if (ioctx->rdma_ius && ioctx->n_rdma_ius)
1113                 nrdma = ioctx->n_rdma_ius;
1114         else {
1115                 nrdma = (count + SRPT_DEF_SG_PER_WQE - 1) / SRPT_DEF_SG_PER_WQE
1116                         + ioctx->n_rbuf;
1117
1118                 ioctx->rdma_ius = kzalloc(nrdma * sizeof *riu, GFP_KERNEL);
1119                 if (!ioctx->rdma_ius)
1120                         goto free_mem;
1121
1122                 ioctx->n_rdma_ius = nrdma;
1123         }
1124
1125         db = ioctx->rbufs;
1126         tsize = cmd->data_length;
1127         dma_len = sg_dma_len(&sg[0]);
1128         riu = ioctx->rdma_ius;
1129
1130         /*
1131          * For each remote desc - calculate the #ib_sge.
1132          * If #ib_sge < SRPT_DEF_SG_PER_WQE per rdma operation then
1133          *      each remote desc rdma_iu is required a rdma wr;
1134          * else
1135          *      we need to allocate extra rdma_iu to carry extra #ib_sge in
1136          *      another rdma wr
1137          */
1138         for (i = 0, j = 0;
1139              j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
1140                 rsize = be32_to_cpu(db->len);
1141                 raddr = be64_to_cpu(db->va);
1142                 riu->raddr = raddr;
1143                 riu->rkey = be32_to_cpu(db->key);
1144                 riu->sge_cnt = 0;
1145
1146                 /* calculate how many sge required for this remote_buf */
1147                 while (rsize > 0 && tsize > 0) {
1148
1149                         if (rsize >= dma_len) {
1150                                 tsize -= dma_len;
1151                                 rsize -= dma_len;
1152                                 raddr += dma_len;
1153
1154                                 if (tsize > 0) {
1155                                         ++j;
1156                                         if (j < count) {
1157                                                 sg = sg_next(sg);
1158                                                 dma_len = sg_dma_len(sg);
1159                                         }
1160                                 }
1161                         } else {
1162                                 tsize -= rsize;
1163                                 dma_len -= rsize;
1164                                 rsize = 0;
1165                         }
1166
1167                         ++riu->sge_cnt;
1168
1169                         if (rsize > 0 && riu->sge_cnt == SRPT_DEF_SG_PER_WQE) {
1170                                 ++ioctx->n_rdma;
1171                                 riu->sge =
1172                                     kmalloc(riu->sge_cnt * sizeof *riu->sge,
1173                                             GFP_KERNEL);
1174                                 if (!riu->sge)
1175                                         goto free_mem;
1176
1177                                 ++riu;
1178                                 riu->sge_cnt = 0;
1179                                 riu->raddr = raddr;
1180                                 riu->rkey = be32_to_cpu(db->key);
1181                         }
1182                 }
1183
1184                 ++ioctx->n_rdma;
1185                 riu->sge = kmalloc(riu->sge_cnt * sizeof *riu->sge,
1186                                    GFP_KERNEL);
1187                 if (!riu->sge)
1188                         goto free_mem;
1189         }
1190
1191         db = ioctx->rbufs;
1192         tsize = cmd->data_length;
1193         riu = ioctx->rdma_ius;
1194         sg = sg_orig;
1195         dma_len = sg_dma_len(&sg[0]);
1196         dma_addr = sg_dma_address(&sg[0]);
1197
1198         /* this second loop is really mapped sg_addres to rdma_iu->ib_sge */
1199         for (i = 0, j = 0;
1200              j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
1201                 rsize = be32_to_cpu(db->len);
1202                 sge = riu->sge;
1203                 k = 0;
1204
1205                 while (rsize > 0 && tsize > 0) {
1206                         sge->addr = dma_addr;
1207                         sge->lkey = ch->sport->sdev->mr->lkey;
1208
1209                         if (rsize >= dma_len) {
1210                                 sge->length =
1211                                         (tsize < dma_len) ? tsize : dma_len;
1212                                 tsize -= dma_len;
1213                                 rsize -= dma_len;
1214
1215                                 if (tsize > 0) {
1216                                         ++j;
1217                                         if (j < count) {
1218                                                 sg = sg_next(sg);
1219                                                 dma_len = sg_dma_len(sg);
1220                                                 dma_addr = sg_dma_address(sg);
1221                                         }
1222                                 }
1223                         } else {
1224                                 sge->length = (tsize < rsize) ? tsize : rsize;
1225                                 tsize -= rsize;
1226                                 dma_len -= rsize;
1227                                 dma_addr += rsize;
1228                                 rsize = 0;
1229                         }
1230
1231                         ++k;
1232                         if (k == riu->sge_cnt && rsize > 0 && tsize > 0) {
1233                                 ++riu;
1234                                 sge = riu->sge;
1235                                 k = 0;
1236                         } else if (rsize > 0 && tsize > 0)
1237                                 ++sge;
1238                 }
1239         }
1240
1241         return 0;
1242
1243 free_mem:
1244         srpt_unmap_sg_to_ib_sge(ch, ioctx);
1245
1246         return -ENOMEM;
1247 }
1248
1249 /**
1250  * srpt_get_send_ioctx() - Obtain an I/O context for sending to the initiator.
1251  */
1252 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch)
1253 {
1254         struct srpt_send_ioctx *ioctx;
1255         unsigned long flags;
1256
1257         BUG_ON(!ch);
1258
1259         ioctx = NULL;
1260         spin_lock_irqsave(&ch->spinlock, flags);
1261         if (!list_empty(&ch->free_list)) {
1262                 ioctx = list_first_entry(&ch->free_list,
1263                                          struct srpt_send_ioctx, free_list);
1264                 list_del(&ioctx->free_list);
1265         }
1266         spin_unlock_irqrestore(&ch->spinlock, flags);
1267
1268         if (!ioctx)
1269                 return ioctx;
1270
1271         BUG_ON(ioctx->ch != ch);
1272         kref_init(&ioctx->kref);
1273         spin_lock_init(&ioctx->spinlock);
1274         ioctx->state = SRPT_STATE_NEW;
1275         ioctx->n_rbuf = 0;
1276         ioctx->rbufs = NULL;
1277         ioctx->n_rdma = 0;
1278         ioctx->n_rdma_ius = 0;
1279         ioctx->rdma_ius = NULL;
1280         ioctx->mapped_sg_count = 0;
1281         init_completion(&ioctx->tx_done);
1282         ioctx->queue_status_only = false;
1283         /*
1284          * transport_init_se_cmd() does not initialize all fields, so do it
1285          * here.
1286          */
1287         memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1288         memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1289
1290         return ioctx;
1291 }
1292
1293 /**
1294  * srpt_put_send_ioctx() - Free up resources.
1295  */
1296 static void srpt_put_send_ioctx(struct srpt_send_ioctx *ioctx)
1297 {
1298         struct srpt_rdma_ch *ch;
1299         unsigned long flags;
1300
1301         BUG_ON(!ioctx);
1302         ch = ioctx->ch;
1303         BUG_ON(!ch);
1304
1305         WARN_ON(srpt_get_cmd_state(ioctx) != SRPT_STATE_DONE);
1306
1307         srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1308         transport_generic_free_cmd(&ioctx->cmd, 0);
1309
1310         if (ioctx->n_rbuf > 1) {
1311                 kfree(ioctx->rbufs);
1312                 ioctx->rbufs = NULL;
1313                 ioctx->n_rbuf = 0;
1314         }
1315
1316         spin_lock_irqsave(&ch->spinlock, flags);
1317         list_add(&ioctx->free_list, &ch->free_list);
1318         spin_unlock_irqrestore(&ch->spinlock, flags);
1319 }
1320
1321 static void srpt_put_send_ioctx_kref(struct kref *kref)
1322 {
1323         srpt_put_send_ioctx(container_of(kref, struct srpt_send_ioctx, kref));
1324 }
1325
1326 /**
1327  * srpt_abort_cmd() - Abort a SCSI command.
1328  * @ioctx:   I/O context associated with the SCSI command.
1329  * @context: Preferred execution context.
1330  */
1331 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1332 {
1333         enum srpt_command_state state;
1334         unsigned long flags;
1335
1336         BUG_ON(!ioctx);
1337
1338         /*
1339          * If the command is in a state where the target core is waiting for
1340          * the ib_srpt driver, change the state to the next state. Changing
1341          * the state of the command from SRPT_STATE_NEED_DATA to
1342          * SRPT_STATE_DATA_IN ensures that srpt_xmit_response() will call this
1343          * function a second time.
1344          */
1345
1346         spin_lock_irqsave(&ioctx->spinlock, flags);
1347         state = ioctx->state;
1348         switch (state) {
1349         case SRPT_STATE_NEED_DATA:
1350                 ioctx->state = SRPT_STATE_DATA_IN;
1351                 break;
1352         case SRPT_STATE_DATA_IN:
1353         case SRPT_STATE_CMD_RSP_SENT:
1354         case SRPT_STATE_MGMT_RSP_SENT:
1355                 ioctx->state = SRPT_STATE_DONE;
1356                 break;
1357         default:
1358                 break;
1359         }
1360         spin_unlock_irqrestore(&ioctx->spinlock, flags);
1361
1362         if (state == SRPT_STATE_DONE)
1363                 goto out;
1364
1365         pr_debug("Aborting cmd with state %d and tag %lld\n", state,
1366                  ioctx->tag);
1367
1368         switch (state) {
1369         case SRPT_STATE_NEW:
1370         case SRPT_STATE_DATA_IN:
1371         case SRPT_STATE_MGMT:
1372                 /*
1373                  * Do nothing - defer abort processing until
1374                  * srpt_queue_response() is invoked.
1375                  */
1376                 WARN_ON(!transport_check_aborted_status(&ioctx->cmd, false));
1377                 break;
1378         case SRPT_STATE_NEED_DATA:
1379                 /* DMA_TO_DEVICE (write) - RDMA read error. */
1380
1381                 /* XXX(hch): this is a horrible layering violation.. */
1382                 spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1383                 ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1384                 ioctx->cmd.transport_state &= ~CMD_T_ACTIVE;
1385                 spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1386
1387                 complete(&ioctx->cmd.transport_lun_stop_comp);
1388                 break;
1389         case SRPT_STATE_CMD_RSP_SENT:
1390                 /*
1391                  * SRP_RSP sending failed or the SRP_RSP send completion has
1392                  * not been received in time.
1393                  */
1394                 srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1395                 spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1396                 ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1397                 spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1398                 kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1399                 break;
1400         case SRPT_STATE_MGMT_RSP_SENT:
1401                 srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1402                 kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1403                 break;
1404         default:
1405                 WARN_ON("ERROR: unexpected command state");
1406                 break;
1407         }
1408
1409 out:
1410         return state;
1411 }
1412
1413 /**
1414  * srpt_handle_send_err_comp() - Process an IB_WC_SEND error completion.
1415  */
1416 static void srpt_handle_send_err_comp(struct srpt_rdma_ch *ch, u64 wr_id)
1417 {
1418         struct srpt_send_ioctx *ioctx;
1419         enum srpt_command_state state;
1420         struct se_cmd *cmd;
1421         u32 index;
1422
1423         atomic_inc(&ch->sq_wr_avail);
1424
1425         index = idx_from_wr_id(wr_id);
1426         ioctx = ch->ioctx_ring[index];
1427         state = srpt_get_cmd_state(ioctx);
1428         cmd = &ioctx->cmd;
1429
1430         WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1431                 && state != SRPT_STATE_MGMT_RSP_SENT
1432                 && state != SRPT_STATE_NEED_DATA
1433                 && state != SRPT_STATE_DONE);
1434
1435         /* If SRP_RSP sending failed, undo the ch->req_lim change. */
1436         if (state == SRPT_STATE_CMD_RSP_SENT
1437             || state == SRPT_STATE_MGMT_RSP_SENT)
1438                 atomic_dec(&ch->req_lim);
1439
1440         srpt_abort_cmd(ioctx);
1441 }
1442
1443 /**
1444  * srpt_handle_send_comp() - Process an IB send completion notification.
1445  */
1446 static void srpt_handle_send_comp(struct srpt_rdma_ch *ch,
1447                                   struct srpt_send_ioctx *ioctx)
1448 {
1449         enum srpt_command_state state;
1450
1451         atomic_inc(&ch->sq_wr_avail);
1452
1453         state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1454
1455         if (WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1456                     && state != SRPT_STATE_MGMT_RSP_SENT
1457                     && state != SRPT_STATE_DONE))
1458                 pr_debug("state = %d\n", state);
1459
1460         if (state != SRPT_STATE_DONE)
1461                 kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1462         else
1463                 printk(KERN_ERR "IB completion has been received too late for"
1464                        " wr_id = %u.\n", ioctx->ioctx.index);
1465 }
1466
1467 /**
1468  * srpt_handle_rdma_comp() - Process an IB RDMA completion notification.
1469  *
1470  * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping
1471  * the data that has been transferred via IB RDMA had to be postponed until the
1472  * check_stop_free() callback.  None of this is necessary anymore and needs to
1473  * be cleaned up.
1474  */
1475 static void srpt_handle_rdma_comp(struct srpt_rdma_ch *ch,
1476                                   struct srpt_send_ioctx *ioctx,
1477                                   enum srpt_opcode opcode)
1478 {
1479         WARN_ON(ioctx->n_rdma <= 0);
1480         atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1481
1482         if (opcode == SRPT_RDMA_READ_LAST) {
1483                 if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1484                                                 SRPT_STATE_DATA_IN))
1485                         target_execute_cmd(&ioctx->cmd);
1486                 else
1487                         printk(KERN_ERR "%s[%d]: wrong state = %d\n", __func__,
1488                                __LINE__, srpt_get_cmd_state(ioctx));
1489         } else if (opcode == SRPT_RDMA_ABORT) {
1490                 ioctx->rdma_aborted = true;
1491         } else {
1492                 WARN(true, "unexpected opcode %d\n", opcode);
1493         }
1494 }
1495
1496 /**
1497  * srpt_handle_rdma_err_comp() - Process an IB RDMA error completion.
1498  */
1499 static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch,
1500                                       struct srpt_send_ioctx *ioctx,
1501                                       enum srpt_opcode opcode)
1502 {
1503         struct se_cmd *cmd;
1504         enum srpt_command_state state;
1505         unsigned long flags;
1506
1507         cmd = &ioctx->cmd;
1508         state = srpt_get_cmd_state(ioctx);
1509         switch (opcode) {
1510         case SRPT_RDMA_READ_LAST:
1511                 if (ioctx->n_rdma <= 0) {
1512                         printk(KERN_ERR "Received invalid RDMA read"
1513                                " error completion with idx %d\n",
1514                                ioctx->ioctx.index);
1515                         break;
1516                 }
1517                 atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1518                 if (state == SRPT_STATE_NEED_DATA)
1519                         srpt_abort_cmd(ioctx);
1520                 else
1521                         printk(KERN_ERR "%s[%d]: wrong state = %d\n",
1522                                __func__, __LINE__, state);
1523                 break;
1524         case SRPT_RDMA_WRITE_LAST:
1525                 spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1526                 ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1527                 spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1528                 break;
1529         default:
1530                 printk(KERN_ERR "%s[%d]: opcode = %u\n", __func__,
1531                        __LINE__, opcode);
1532                 break;
1533         }
1534 }
1535
1536 /**
1537  * srpt_build_cmd_rsp() - Build an SRP_RSP response.
1538  * @ch: RDMA channel through which the request has been received.
1539  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1540  *   be built in the buffer ioctx->buf points at and hence this function will
1541  *   overwrite the request data.
1542  * @tag: tag of the request for which this response is being generated.
1543  * @status: value for the STATUS field of the SRP_RSP information unit.
1544  *
1545  * Returns the size in bytes of the SRP_RSP response.
1546  *
1547  * An SRP_RSP response contains a SCSI status or service response. See also
1548  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1549  * response. See also SPC-2 for more information about sense data.
1550  */
1551 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1552                               struct srpt_send_ioctx *ioctx, u64 tag,
1553                               int status)
1554 {
1555         struct srp_rsp *srp_rsp;
1556         const u8 *sense_data;
1557         int sense_data_len, max_sense_len;
1558
1559         /*
1560          * The lowest bit of all SAM-3 status codes is zero (see also
1561          * paragraph 5.3 in SAM-3).
1562          */
1563         WARN_ON(status & 1);
1564
1565         srp_rsp = ioctx->ioctx.buf;
1566         BUG_ON(!srp_rsp);
1567
1568         sense_data = ioctx->sense_data;
1569         sense_data_len = ioctx->cmd.scsi_sense_length;
1570         WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1571
1572         memset(srp_rsp, 0, sizeof *srp_rsp);
1573         srp_rsp->opcode = SRP_RSP;
1574         srp_rsp->req_lim_delta =
1575                 __constant_cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1576         srp_rsp->tag = tag;
1577         srp_rsp->status = status;
1578
1579         if (sense_data_len) {
1580                 BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1581                 max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1582                 if (sense_data_len > max_sense_len) {
1583                         printk(KERN_WARNING "truncated sense data from %d to %d"
1584                                " bytes\n", sense_data_len, max_sense_len);
1585                         sense_data_len = max_sense_len;
1586                 }
1587
1588                 srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1589                 srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1590                 memcpy(srp_rsp + 1, sense_data, sense_data_len);
1591         }
1592
1593         return sizeof(*srp_rsp) + sense_data_len;
1594 }
1595
1596 /**
1597  * srpt_build_tskmgmt_rsp() - Build a task management response.
1598  * @ch:       RDMA channel through which the request has been received.
1599  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1600  * @rsp_code: RSP_CODE that will be stored in the response.
1601  * @tag:      Tag of the request for which this response is being generated.
1602  *
1603  * Returns the size in bytes of the SRP_RSP response.
1604  *
1605  * An SRP_RSP response contains a SCSI status or service response. See also
1606  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1607  * response.
1608  */
1609 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1610                                   struct srpt_send_ioctx *ioctx,
1611                                   u8 rsp_code, u64 tag)
1612 {
1613         struct srp_rsp *srp_rsp;
1614         int resp_data_len;
1615         int resp_len;
1616
1617         resp_data_len = (rsp_code == SRP_TSK_MGMT_SUCCESS) ? 0 : 4;
1618         resp_len = sizeof(*srp_rsp) + resp_data_len;
1619
1620         srp_rsp = ioctx->ioctx.buf;
1621         BUG_ON(!srp_rsp);
1622         memset(srp_rsp, 0, sizeof *srp_rsp);
1623
1624         srp_rsp->opcode = SRP_RSP;
1625         srp_rsp->req_lim_delta = __constant_cpu_to_be32(1
1626                                     + atomic_xchg(&ch->req_lim_delta, 0));
1627         srp_rsp->tag = tag;
1628
1629         if (rsp_code != SRP_TSK_MGMT_SUCCESS) {
1630                 srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1631                 srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1632                 srp_rsp->data[3] = rsp_code;
1633         }
1634
1635         return resp_len;
1636 }
1637
1638 #define NO_SUCH_LUN ((uint64_t)-1LL)
1639
1640 /*
1641  * SCSI LUN addressing method. See also SAM-2 and the section about
1642  * eight byte LUNs.
1643  */
1644 enum scsi_lun_addr_method {
1645         SCSI_LUN_ADDR_METHOD_PERIPHERAL   = 0,
1646         SCSI_LUN_ADDR_METHOD_FLAT         = 1,
1647         SCSI_LUN_ADDR_METHOD_LUN          = 2,
1648         SCSI_LUN_ADDR_METHOD_EXTENDED_LUN = 3,
1649 };
1650
1651 /*
1652  * srpt_unpack_lun() - Convert from network LUN to linear LUN.
1653  *
1654  * Convert an 2-byte, 4-byte, 6-byte or 8-byte LUN structure in network byte
1655  * order (big endian) to a linear LUN. Supports three LUN addressing methods:
1656  * peripheral, flat and logical unit. See also SAM-2, section 4.9.4 (page 40).
1657  */
1658 static uint64_t srpt_unpack_lun(const uint8_t *lun, int len)
1659 {
1660         uint64_t res = NO_SUCH_LUN;
1661         int addressing_method;
1662
1663         if (unlikely(len < 2)) {
1664                 printk(KERN_ERR "Illegal LUN length %d, expected 2 bytes or "
1665                        "more", len);
1666                 goto out;
1667         }
1668
1669         switch (len) {
1670         case 8:
1671                 if ((*((__be64 *)lun) &
1672                      __constant_cpu_to_be64(0x0000FFFFFFFFFFFFLL)) != 0)
1673                         goto out_err;
1674                 break;
1675         case 4:
1676                 if (*((__be16 *)&lun[2]) != 0)
1677                         goto out_err;
1678                 break;
1679         case 6:
1680                 if (*((__be32 *)&lun[2]) != 0)
1681                         goto out_err;
1682                 break;
1683         case 2:
1684                 break;
1685         default:
1686                 goto out_err;
1687         }
1688
1689         addressing_method = (*lun) >> 6; /* highest two bits of byte 0 */
1690         switch (addressing_method) {
1691         case SCSI_LUN_ADDR_METHOD_PERIPHERAL:
1692         case SCSI_LUN_ADDR_METHOD_FLAT:
1693         case SCSI_LUN_ADDR_METHOD_LUN:
1694                 res = *(lun + 1) | (((*lun) & 0x3f) << 8);
1695                 break;
1696
1697         case SCSI_LUN_ADDR_METHOD_EXTENDED_LUN:
1698         default:
1699                 printk(KERN_ERR "Unimplemented LUN addressing method %u",
1700                        addressing_method);
1701                 break;
1702         }
1703
1704 out:
1705         return res;
1706
1707 out_err:
1708         printk(KERN_ERR "Support for multi-level LUNs has not yet been"
1709                " implemented");
1710         goto out;
1711 }
1712
1713 static int srpt_check_stop_free(struct se_cmd *cmd)
1714 {
1715         struct srpt_send_ioctx *ioctx;
1716
1717         ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
1718         return kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1719 }
1720
1721 /**
1722  * srpt_handle_cmd() - Process SRP_CMD.
1723  */
1724 static int srpt_handle_cmd(struct srpt_rdma_ch *ch,
1725                            struct srpt_recv_ioctx *recv_ioctx,
1726                            struct srpt_send_ioctx *send_ioctx)
1727 {
1728         struct se_cmd *cmd;
1729         struct srp_cmd *srp_cmd;
1730         uint64_t unpacked_lun;
1731         u64 data_len;
1732         enum dma_data_direction dir;
1733         int ret;
1734
1735         BUG_ON(!send_ioctx);
1736
1737         srp_cmd = recv_ioctx->ioctx.buf;
1738         kref_get(&send_ioctx->kref);
1739         cmd = &send_ioctx->cmd;
1740         send_ioctx->tag = srp_cmd->tag;
1741
1742         switch (srp_cmd->task_attr) {
1743         case SRP_CMD_SIMPLE_Q:
1744                 cmd->sam_task_attr = MSG_SIMPLE_TAG;
1745                 break;
1746         case SRP_CMD_ORDERED_Q:
1747         default:
1748                 cmd->sam_task_attr = MSG_ORDERED_TAG;
1749                 break;
1750         case SRP_CMD_HEAD_OF_Q:
1751                 cmd->sam_task_attr = MSG_HEAD_TAG;
1752                 break;
1753         case SRP_CMD_ACA:
1754                 cmd->sam_task_attr = MSG_ACA_TAG;
1755                 break;
1756         }
1757
1758         ret = srpt_get_desc_tbl(send_ioctx, srp_cmd, &dir, &data_len);
1759         if (ret) {
1760                 printk(KERN_ERR "0x%llx: parsing SRP descriptor table failed.\n",
1761                        srp_cmd->tag);
1762                 cmd->se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1763                 cmd->scsi_sense_reason = TCM_INVALID_CDB_FIELD;
1764                 kref_put(&send_ioctx->kref, srpt_put_send_ioctx_kref);
1765                 goto send_sense;
1766         }
1767
1768         cmd->data_length = data_len;
1769         cmd->data_direction = dir;
1770         unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_cmd->lun,
1771                                        sizeof(srp_cmd->lun));
1772         if (transport_lookup_cmd_lun(cmd, unpacked_lun) < 0) {
1773                 kref_put(&send_ioctx->kref, srpt_put_send_ioctx_kref);
1774                 goto send_sense;
1775         }
1776         ret = target_setup_cmd_from_cdb(cmd, srp_cmd->cdb);
1777         if (ret < 0) {
1778                 kref_put(&send_ioctx->kref, srpt_put_send_ioctx_kref);
1779                 if (cmd->se_cmd_flags & SCF_SCSI_RESERVATION_CONFLICT) {
1780                         srpt_queue_status(cmd);
1781                         return 0;
1782                 } else
1783                         goto send_sense;
1784         }
1785
1786         transport_handle_cdb_direct(cmd);
1787         return 0;
1788
1789 send_sense:
1790         transport_send_check_condition_and_sense(cmd, cmd->scsi_sense_reason,
1791                                                  0);
1792         return -1;
1793 }
1794
1795 /**
1796  * srpt_rx_mgmt_fn_tag() - Process a task management function by tag.
1797  * @ch: RDMA channel of the task management request.
1798  * @fn: Task management function to perform.
1799  * @req_tag: Tag of the SRP task management request.
1800  * @mgmt_ioctx: I/O context of the task management request.
1801  *
1802  * Returns zero if the target core will process the task management
1803  * request asynchronously.
1804  *
1805  * Note: It is assumed that the initiator serializes tag-based task management
1806  * requests.
1807  */
1808 static int srpt_rx_mgmt_fn_tag(struct srpt_send_ioctx *ioctx, u64 tag)
1809 {
1810         struct srpt_device *sdev;
1811         struct srpt_rdma_ch *ch;
1812         struct srpt_send_ioctx *target;
1813         int ret, i;
1814
1815         ret = -EINVAL;
1816         ch = ioctx->ch;
1817         BUG_ON(!ch);
1818         BUG_ON(!ch->sport);
1819         sdev = ch->sport->sdev;
1820         BUG_ON(!sdev);
1821         spin_lock_irq(&sdev->spinlock);
1822         for (i = 0; i < ch->rq_size; ++i) {
1823                 target = ch->ioctx_ring[i];
1824                 if (target->cmd.se_lun == ioctx->cmd.se_lun &&
1825                     target->tag == tag &&
1826                     srpt_get_cmd_state(target) != SRPT_STATE_DONE) {
1827                         ret = 0;
1828                         /* now let the target core abort &target->cmd; */
1829                         break;
1830                 }
1831         }
1832         spin_unlock_irq(&sdev->spinlock);
1833         return ret;
1834 }
1835
1836 static int srp_tmr_to_tcm(int fn)
1837 {
1838         switch (fn) {
1839         case SRP_TSK_ABORT_TASK:
1840                 return TMR_ABORT_TASK;
1841         case SRP_TSK_ABORT_TASK_SET:
1842                 return TMR_ABORT_TASK_SET;
1843         case SRP_TSK_CLEAR_TASK_SET:
1844                 return TMR_CLEAR_TASK_SET;
1845         case SRP_TSK_LUN_RESET:
1846                 return TMR_LUN_RESET;
1847         case SRP_TSK_CLEAR_ACA:
1848                 return TMR_CLEAR_ACA;
1849         default:
1850                 return -1;
1851         }
1852 }
1853
1854 /**
1855  * srpt_handle_tsk_mgmt() - Process an SRP_TSK_MGMT information unit.
1856  *
1857  * Returns 0 if and only if the request will be processed by the target core.
1858  *
1859  * For more information about SRP_TSK_MGMT information units, see also section
1860  * 6.7 in the SRP r16a document.
1861  */
1862 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1863                                  struct srpt_recv_ioctx *recv_ioctx,
1864                                  struct srpt_send_ioctx *send_ioctx)
1865 {
1866         struct srp_tsk_mgmt *srp_tsk;
1867         struct se_cmd *cmd;
1868         uint64_t unpacked_lun;
1869         int tcm_tmr;
1870         int res;
1871
1872         BUG_ON(!send_ioctx);
1873
1874         srp_tsk = recv_ioctx->ioctx.buf;
1875         cmd = &send_ioctx->cmd;
1876
1877         pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld"
1878                  " cm_id %p sess %p\n", srp_tsk->tsk_mgmt_func,
1879                  srp_tsk->task_tag, srp_tsk->tag, ch->cm_id, ch->sess);
1880
1881         srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1882         send_ioctx->tag = srp_tsk->tag;
1883         tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1884         if (tcm_tmr < 0) {
1885                 send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1886                 send_ioctx->cmd.se_tmr_req->response =
1887                         TMR_TASK_MGMT_FUNCTION_NOT_SUPPORTED;
1888                 goto process_tmr;
1889         }
1890         res = core_tmr_alloc_req(cmd, NULL, tcm_tmr, GFP_KERNEL);
1891         if (res < 0) {
1892                 send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1893                 send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1894                 goto process_tmr;
1895         }
1896
1897         unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_tsk->lun,
1898                                        sizeof(srp_tsk->lun));
1899         res = transport_lookup_tmr_lun(&send_ioctx->cmd, unpacked_lun);
1900         if (res) {
1901                 pr_debug("rejecting TMR for LUN %lld\n", unpacked_lun);
1902                 send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1903                 send_ioctx->cmd.se_tmr_req->response = TMR_LUN_DOES_NOT_EXIST;
1904                 goto process_tmr;
1905         }
1906
1907         if (srp_tsk->tsk_mgmt_func == SRP_TSK_ABORT_TASK)
1908                 srpt_rx_mgmt_fn_tag(send_ioctx, srp_tsk->task_tag);
1909
1910 process_tmr:
1911         kref_get(&send_ioctx->kref);
1912         if (!(send_ioctx->cmd.se_cmd_flags & SCF_SCSI_CDB_EXCEPTION))
1913                 transport_generic_handle_tmr(&send_ioctx->cmd);
1914         else
1915                 transport_send_check_condition_and_sense(cmd,
1916                                                 cmd->scsi_sense_reason, 0);
1917
1918 }
1919
1920 /**
1921  * srpt_handle_new_iu() - Process a newly received information unit.
1922  * @ch:    RDMA channel through which the information unit has been received.
1923  * @ioctx: SRPT I/O context associated with the information unit.
1924  */
1925 static void srpt_handle_new_iu(struct srpt_rdma_ch *ch,
1926                                struct srpt_recv_ioctx *recv_ioctx,
1927                                struct srpt_send_ioctx *send_ioctx)
1928 {
1929         struct srp_cmd *srp_cmd;
1930         enum rdma_ch_state ch_state;
1931
1932         BUG_ON(!ch);
1933         BUG_ON(!recv_ioctx);
1934
1935         ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1936                                    recv_ioctx->ioctx.dma, srp_max_req_size,
1937                                    DMA_FROM_DEVICE);
1938
1939         ch_state = srpt_get_ch_state(ch);
1940         if (unlikely(ch_state == CH_CONNECTING)) {
1941                 list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1942                 goto out;
1943         }
1944
1945         if (unlikely(ch_state != CH_LIVE))
1946                 goto out;
1947
1948         srp_cmd = recv_ioctx->ioctx.buf;
1949         if (srp_cmd->opcode == SRP_CMD || srp_cmd->opcode == SRP_TSK_MGMT) {
1950                 if (!send_ioctx)
1951                         send_ioctx = srpt_get_send_ioctx(ch);
1952                 if (unlikely(!send_ioctx)) {
1953                         list_add_tail(&recv_ioctx->wait_list,
1954                                       &ch->cmd_wait_list);
1955                         goto out;
1956                 }
1957         }
1958
1959         transport_init_se_cmd(&send_ioctx->cmd, &srpt_target->tf_ops, ch->sess,
1960                               0, DMA_NONE, MSG_SIMPLE_TAG,
1961                               send_ioctx->sense_data);
1962
1963         switch (srp_cmd->opcode) {
1964         case SRP_CMD:
1965                 srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1966                 break;
1967         case SRP_TSK_MGMT:
1968                 srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1969                 break;
1970         case SRP_I_LOGOUT:
1971                 printk(KERN_ERR "Not yet implemented: SRP_I_LOGOUT\n");
1972                 break;
1973         case SRP_CRED_RSP:
1974                 pr_debug("received SRP_CRED_RSP\n");
1975                 break;
1976         case SRP_AER_RSP:
1977                 pr_debug("received SRP_AER_RSP\n");
1978                 break;
1979         case SRP_RSP:
1980                 printk(KERN_ERR "Received SRP_RSP\n");
1981                 break;
1982         default:
1983                 printk(KERN_ERR "received IU with unknown opcode 0x%x\n",
1984                        srp_cmd->opcode);
1985                 break;
1986         }
1987
1988         srpt_post_recv(ch->sport->sdev, recv_ioctx);
1989 out:
1990         return;
1991 }
1992
1993 static void srpt_process_rcv_completion(struct ib_cq *cq,
1994                                         struct srpt_rdma_ch *ch,
1995                                         struct ib_wc *wc)
1996 {
1997         struct srpt_device *sdev = ch->sport->sdev;
1998         struct srpt_recv_ioctx *ioctx;
1999         u32 index;
2000
2001         index = idx_from_wr_id(wc->wr_id);
2002         if (wc->status == IB_WC_SUCCESS) {
2003                 int req_lim;
2004
2005                 req_lim = atomic_dec_return(&ch->req_lim);
2006                 if (unlikely(req_lim < 0))
2007                         printk(KERN_ERR "req_lim = %d < 0\n", req_lim);
2008                 ioctx = sdev->ioctx_ring[index];
2009                 srpt_handle_new_iu(ch, ioctx, NULL);
2010         } else {
2011                 printk(KERN_INFO "receiving failed for idx %u with status %d\n",
2012                        index, wc->status);
2013         }
2014 }
2015
2016 /**
2017  * srpt_process_send_completion() - Process an IB send completion.
2018  *
2019  * Note: Although this has not yet been observed during tests, at least in
2020  * theory it is possible that the srpt_get_send_ioctx() call invoked by
2021  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
2022  * value in each response is set to one, and it is possible that this response
2023  * makes the initiator send a new request before the send completion for that
2024  * response has been processed. This could e.g. happen if the call to
2025  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
2026  * if IB retransmission causes generation of the send completion to be
2027  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
2028  * are queued on cmd_wait_list. The code below processes these delayed
2029  * requests one at a time.
2030  */
2031 static void srpt_process_send_completion(struct ib_cq *cq,
2032                                          struct srpt_rdma_ch *ch,
2033                                          struct ib_wc *wc)
2034 {
2035         struct srpt_send_ioctx *send_ioctx;
2036         uint32_t index;
2037         enum srpt_opcode opcode;
2038
2039         index = idx_from_wr_id(wc->wr_id);
2040         opcode = opcode_from_wr_id(wc->wr_id);
2041         send_ioctx = ch->ioctx_ring[index];
2042         if (wc->status == IB_WC_SUCCESS) {
2043                 if (opcode == SRPT_SEND)
2044                         srpt_handle_send_comp(ch, send_ioctx);
2045                 else {
2046                         WARN_ON(opcode != SRPT_RDMA_ABORT &&
2047                                 wc->opcode != IB_WC_RDMA_READ);
2048                         srpt_handle_rdma_comp(ch, send_ioctx, opcode);
2049                 }
2050         } else {
2051                 if (opcode == SRPT_SEND) {
2052                         printk(KERN_INFO "sending response for idx %u failed"
2053                                " with status %d\n", index, wc->status);
2054                         srpt_handle_send_err_comp(ch, wc->wr_id);
2055                 } else if (opcode != SRPT_RDMA_MID) {
2056                         printk(KERN_INFO "RDMA t %d for idx %u failed with"
2057                                 " status %d", opcode, index, wc->status);
2058                         srpt_handle_rdma_err_comp(ch, send_ioctx, opcode);
2059                 }
2060         }
2061
2062         while (unlikely(opcode == SRPT_SEND
2063                         && !list_empty(&ch->cmd_wait_list)
2064                         && srpt_get_ch_state(ch) == CH_LIVE
2065                         && (send_ioctx = srpt_get_send_ioctx(ch)) != NULL)) {
2066                 struct srpt_recv_ioctx *recv_ioctx;
2067
2068                 recv_ioctx = list_first_entry(&ch->cmd_wait_list,
2069                                               struct srpt_recv_ioctx,
2070                                               wait_list);
2071                 list_del(&recv_ioctx->wait_list);
2072                 srpt_handle_new_iu(ch, recv_ioctx, send_ioctx);
2073         }
2074 }
2075
2076 static void srpt_process_completion(struct ib_cq *cq, struct srpt_rdma_ch *ch)
2077 {
2078         struct ib_wc *const wc = ch->wc;
2079         int i, n;
2080
2081         WARN_ON(cq != ch->cq);
2082
2083         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
2084         while ((n = ib_poll_cq(cq, ARRAY_SIZE(ch->wc), wc)) > 0) {
2085                 for (i = 0; i < n; i++) {
2086                         if (opcode_from_wr_id(wc[i].wr_id) == SRPT_RECV)
2087                                 srpt_process_rcv_completion(cq, ch, &wc[i]);
2088                         else
2089                                 srpt_process_send_completion(cq, ch, &wc[i]);
2090                 }
2091         }
2092 }
2093
2094 /**
2095  * srpt_completion() - IB completion queue callback function.
2096  *
2097  * Notes:
2098  * - It is guaranteed that a completion handler will never be invoked
2099  *   concurrently on two different CPUs for the same completion queue. See also
2100  *   Documentation/infiniband/core_locking.txt and the implementation of
2101  *   handle_edge_irq() in kernel/irq/chip.c.
2102  * - When threaded IRQs are enabled, completion handlers are invoked in thread
2103  *   context instead of interrupt context.
2104  */
2105 static void srpt_completion(struct ib_cq *cq, void *ctx)
2106 {
2107         struct srpt_rdma_ch *ch = ctx;
2108
2109         wake_up_interruptible(&ch->wait_queue);
2110 }
2111
2112 static int srpt_compl_thread(void *arg)
2113 {
2114         struct srpt_rdma_ch *ch;
2115
2116         /* Hibernation / freezing of the SRPT kernel thread is not supported. */
2117         current->flags |= PF_NOFREEZE;
2118
2119         ch = arg;
2120         BUG_ON(!ch);
2121         printk(KERN_INFO "Session %s: kernel thread %s (PID %d) started\n",
2122                ch->sess_name, ch->thread->comm, current->pid);
2123         while (!kthread_should_stop()) {
2124                 wait_event_interruptible(ch->wait_queue,
2125                         (srpt_process_completion(ch->cq, ch),
2126                          kthread_should_stop()));
2127         }
2128         printk(KERN_INFO "Session %s: kernel thread %s (PID %d) stopped\n",
2129                ch->sess_name, ch->thread->comm, current->pid);
2130         return 0;
2131 }
2132
2133 /**
2134  * srpt_create_ch_ib() - Create receive and send completion queues.
2135  */
2136 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
2137 {
2138         struct ib_qp_init_attr *qp_init;
2139         struct srpt_port *sport = ch->sport;
2140         struct srpt_device *sdev = sport->sdev;
2141         u32 srp_sq_size = sport->port_attrib.srp_sq_size;
2142         int ret;
2143
2144         WARN_ON(ch->rq_size < 1);
2145
2146         ret = -ENOMEM;
2147         qp_init = kzalloc(sizeof *qp_init, GFP_KERNEL);
2148         if (!qp_init)
2149                 goto out;
2150
2151         ch->cq = ib_create_cq(sdev->device, srpt_completion, NULL, ch,
2152                               ch->rq_size + srp_sq_size, 0);
2153         if (IS_ERR(ch->cq)) {
2154                 ret = PTR_ERR(ch->cq);
2155                 printk(KERN_ERR "failed to create CQ cqe= %d ret= %d\n",
2156                        ch->rq_size + srp_sq_size, ret);
2157                 goto out;
2158         }
2159
2160         qp_init->qp_context = (void *)ch;
2161         qp_init->event_handler
2162                 = (void(*)(struct ib_event *, void*))srpt_qp_event;
2163         qp_init->send_cq = ch->cq;
2164         qp_init->recv_cq = ch->cq;
2165         qp_init->srq = sdev->srq;
2166         qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
2167         qp_init->qp_type = IB_QPT_RC;
2168         qp_init->cap.max_send_wr = srp_sq_size;
2169         qp_init->cap.max_send_sge = SRPT_DEF_SG_PER_WQE;
2170
2171         ch->qp = ib_create_qp(sdev->pd, qp_init);
2172         if (IS_ERR(ch->qp)) {
2173                 ret = PTR_ERR(ch->qp);
2174                 printk(KERN_ERR "failed to create_qp ret= %d\n", ret);
2175                 goto err_destroy_cq;
2176         }
2177
2178         atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
2179
2180         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
2181                  __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
2182                  qp_init->cap.max_send_wr, ch->cm_id);
2183
2184         ret = srpt_init_ch_qp(ch, ch->qp);
2185         if (ret)
2186                 goto err_destroy_qp;
2187
2188         init_waitqueue_head(&ch->wait_queue);
2189
2190         pr_debug("creating thread for session %s\n", ch->sess_name);
2191
2192         ch->thread = kthread_run(srpt_compl_thread, ch, "ib_srpt_compl");
2193         if (IS_ERR(ch->thread)) {
2194                 printk(KERN_ERR "failed to create kernel thread %ld\n",
2195                        PTR_ERR(ch->thread));
2196                 ch->thread = NULL;
2197                 goto err_destroy_qp;
2198         }
2199
2200 out:
2201         kfree(qp_init);
2202         return ret;
2203
2204 err_destroy_qp:
2205         ib_destroy_qp(ch->qp);
2206 err_destroy_cq:
2207         ib_destroy_cq(ch->cq);
2208         goto out;
2209 }
2210
2211 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
2212 {
2213         if (ch->thread)
2214                 kthread_stop(ch->thread);
2215
2216         ib_destroy_qp(ch->qp);
2217         ib_destroy_cq(ch->cq);
2218 }
2219
2220 /**
2221  * __srpt_close_ch() - Close an RDMA channel by setting the QP error state.
2222  *
2223  * Reset the QP and make sure all resources associated with the channel will
2224  * be deallocated at an appropriate time.
2225  *
2226  * Note: The caller must hold ch->sport->sdev->spinlock.
2227  */
2228 static void __srpt_close_ch(struct srpt_rdma_ch *ch)
2229 {
2230         struct srpt_device *sdev;
2231         enum rdma_ch_state prev_state;
2232         unsigned long flags;
2233
2234         sdev = ch->sport->sdev;
2235
2236         spin_lock_irqsave(&ch->spinlock, flags);
2237         prev_state = ch->state;
2238         switch (prev_state) {
2239         case CH_CONNECTING:
2240         case CH_LIVE:
2241                 ch->state = CH_DISCONNECTING;
2242                 break;
2243         default:
2244                 break;
2245         }
2246         spin_unlock_irqrestore(&ch->spinlock, flags);
2247
2248         switch (prev_state) {
2249         case CH_CONNECTING:
2250                 ib_send_cm_rej(ch->cm_id, IB_CM_REJ_NO_RESOURCES, NULL, 0,
2251                                NULL, 0);
2252                 /* fall through */
2253         case CH_LIVE:
2254                 if (ib_send_cm_dreq(ch->cm_id, NULL, 0) < 0)
2255                         printk(KERN_ERR "sending CM DREQ failed.\n");
2256                 break;
2257         case CH_DISCONNECTING:
2258                 break;
2259         case CH_DRAINING:
2260         case CH_RELEASING:
2261                 break;
2262         }
2263 }
2264
2265 /**
2266  * srpt_close_ch() - Close an RDMA channel.
2267  */
2268 static void srpt_close_ch(struct srpt_rdma_ch *ch)
2269 {
2270         struct srpt_device *sdev;
2271
2272         sdev = ch->sport->sdev;
2273         spin_lock_irq(&sdev->spinlock);
2274         __srpt_close_ch(ch);
2275         spin_unlock_irq(&sdev->spinlock);
2276 }
2277
2278 /**
2279  * srpt_drain_channel() - Drain a channel by resetting the IB queue pair.
2280  * @cm_id: Pointer to the CM ID of the channel to be drained.
2281  *
2282  * Note: Must be called from inside srpt_cm_handler to avoid a race between
2283  * accessing sdev->spinlock and the call to kfree(sdev) in srpt_remove_one()
2284  * (the caller of srpt_cm_handler holds the cm_id spinlock; srpt_remove_one()
2285  * waits until all target sessions for the associated IB device have been
2286  * unregistered and target session registration involves a call to
2287  * ib_destroy_cm_id(), which locks the cm_id spinlock and hence waits until
2288  * this function has finished).
2289  */
2290 static void srpt_drain_channel(struct ib_cm_id *cm_id)
2291 {
2292         struct srpt_device *sdev;
2293         struct srpt_rdma_ch *ch;
2294         int ret;
2295         bool do_reset = false;
2296
2297         WARN_ON_ONCE(irqs_disabled());
2298
2299         sdev = cm_id->context;
2300         BUG_ON(!sdev);
2301         spin_lock_irq(&sdev->spinlock);
2302         list_for_each_entry(ch, &sdev->rch_list, list) {
2303                 if (ch->cm_id == cm_id) {
2304                         do_reset = srpt_test_and_set_ch_state(ch,
2305                                         CH_CONNECTING, CH_DRAINING) ||
2306                                    srpt_test_and_set_ch_state(ch,
2307                                         CH_LIVE, CH_DRAINING) ||
2308                                    srpt_test_and_set_ch_state(ch,
2309                                         CH_DISCONNECTING, CH_DRAINING);
2310                         break;
2311                 }
2312         }
2313         spin_unlock_irq(&sdev->spinlock);
2314
2315         if (do_reset) {
2316                 ret = srpt_ch_qp_err(ch);
2317                 if (ret < 0)
2318                         printk(KERN_ERR "Setting queue pair in error state"
2319                                " failed: %d\n", ret);
2320         }
2321 }
2322
2323 /**
2324  * srpt_find_channel() - Look up an RDMA channel.
2325  * @cm_id: Pointer to the CM ID of the channel to be looked up.
2326  *
2327  * Return NULL if no matching RDMA channel has been found.
2328  */
2329 static struct srpt_rdma_ch *srpt_find_channel(struct srpt_device *sdev,
2330                                               struct ib_cm_id *cm_id)
2331 {
2332         struct srpt_rdma_ch *ch;
2333         bool found;
2334
2335         WARN_ON_ONCE(irqs_disabled());
2336         BUG_ON(!sdev);
2337
2338         found = false;
2339         spin_lock_irq(&sdev->spinlock);
2340         list_for_each_entry(ch, &sdev->rch_list, list) {
2341                 if (ch->cm_id == cm_id) {
2342                         found = true;
2343                         break;
2344                 }
2345         }
2346         spin_unlock_irq(&sdev->spinlock);
2347
2348         return found ? ch : NULL;
2349 }
2350
2351 /**
2352  * srpt_release_channel() - Release channel resources.
2353  *
2354  * Schedules the actual release because:
2355  * - Calling the ib_destroy_cm_id() call from inside an IB CM callback would
2356  *   trigger a deadlock.
2357  * - It is not safe to call TCM transport_* functions from interrupt context.
2358  */
2359 static void srpt_release_channel(struct srpt_rdma_ch *ch)
2360 {
2361         schedule_work(&ch->release_work);
2362 }
2363
2364 static void srpt_release_channel_work(struct work_struct *w)
2365 {
2366         struct srpt_rdma_ch *ch;
2367         struct srpt_device *sdev;
2368
2369         ch = container_of(w, struct srpt_rdma_ch, release_work);
2370         pr_debug("ch = %p; ch->sess = %p; release_done = %p\n", ch, ch->sess,
2371                  ch->release_done);
2372
2373         sdev = ch->sport->sdev;
2374         BUG_ON(!sdev);
2375
2376         transport_deregister_session_configfs(ch->sess);
2377         transport_deregister_session(ch->sess);
2378         ch->sess = NULL;
2379
2380         srpt_destroy_ch_ib(ch);
2381
2382         srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2383                              ch->sport->sdev, ch->rq_size,
2384                              ch->rsp_size, DMA_TO_DEVICE);
2385
2386         spin_lock_irq(&sdev->spinlock);
2387         list_del(&ch->list);
2388         spin_unlock_irq(&sdev->spinlock);
2389
2390         ib_destroy_cm_id(ch->cm_id);
2391
2392         if (ch->release_done)
2393                 complete(ch->release_done);
2394
2395         wake_up(&sdev->ch_releaseQ);
2396
2397         kfree(ch);
2398 }
2399
2400 static struct srpt_node_acl *__srpt_lookup_acl(struct srpt_port *sport,
2401                                                u8 i_port_id[16])
2402 {
2403         struct srpt_node_acl *nacl;
2404
2405         list_for_each_entry(nacl, &sport->port_acl_list, list)
2406                 if (memcmp(nacl->i_port_id, i_port_id,
2407                            sizeof(nacl->i_port_id)) == 0)
2408                         return nacl;
2409
2410         return NULL;
2411 }
2412
2413 static struct srpt_node_acl *srpt_lookup_acl(struct srpt_port *sport,
2414                                              u8 i_port_id[16])
2415 {
2416         struct srpt_node_acl *nacl;
2417
2418         spin_lock_irq(&sport->port_acl_lock);
2419         nacl = __srpt_lookup_acl(sport, i_port_id);
2420         spin_unlock_irq(&sport->port_acl_lock);
2421
2422         return nacl;
2423 }
2424
2425 /**
2426  * srpt_cm_req_recv() - Process the event IB_CM_REQ_RECEIVED.
2427  *
2428  * Ownership of the cm_id is transferred to the target session if this
2429  * functions returns zero. Otherwise the caller remains the owner of cm_id.
2430  */
2431 static int srpt_cm_req_recv(struct ib_cm_id *cm_id,
2432                             struct ib_cm_req_event_param *param,
2433                             void *private_data)
2434 {
2435         struct srpt_device *sdev = cm_id->context;
2436         struct srpt_port *sport = &sdev->port[param->port - 1];
2437         struct srp_login_req *req;
2438         struct srp_login_rsp *rsp;
2439         struct srp_login_rej *rej;
2440         struct ib_cm_rep_param *rep_param;
2441         struct srpt_rdma_ch *ch, *tmp_ch;
2442         struct srpt_node_acl *nacl;
2443         u32 it_iu_len;
2444         int i;
2445         int ret = 0;
2446
2447         WARN_ON_ONCE(irqs_disabled());
2448
2449         if (WARN_ON(!sdev || !private_data))
2450                 return -EINVAL;
2451
2452         req = (struct srp_login_req *)private_data;
2453
2454         it_iu_len = be32_to_cpu(req->req_it_iu_len);
2455
2456         printk(KERN_INFO "Received SRP_LOGIN_REQ with i_port_id 0x%llx:0x%llx,"
2457                " t_port_id 0x%llx:0x%llx and it_iu_len %d on port %d"
2458                " (guid=0x%llx:0x%llx)\n",
2459                be64_to_cpu(*(__be64 *)&req->initiator_port_id[0]),
2460                be64_to_cpu(*(__be64 *)&req->initiator_port_id[8]),
2461                be64_to_cpu(*(__be64 *)&req->target_port_id[0]),
2462                be64_to_cpu(*(__be64 *)&req->target_port_id[8]),
2463                it_iu_len,
2464                param->port,
2465                be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[0]),
2466                be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[8]));
2467
2468         rsp = kzalloc(sizeof *rsp, GFP_KERNEL);
2469         rej = kzalloc(sizeof *rej, GFP_KERNEL);
2470         rep_param = kzalloc(sizeof *rep_param, GFP_KERNEL);
2471
2472         if (!rsp || !rej || !rep_param) {
2473                 ret = -ENOMEM;
2474                 goto out;
2475         }
2476
2477         if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2478                 rej->reason = __constant_cpu_to_be32(
2479                                 SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2480                 ret = -EINVAL;
2481                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because its"
2482                        " length (%d bytes) is out of range (%d .. %d)\n",
2483                        it_iu_len, 64, srp_max_req_size);
2484                 goto reject;
2485         }
2486
2487         if (!sport->enabled) {
2488                 rej->reason = __constant_cpu_to_be32(
2489                              SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2490                 ret = -EINVAL;
2491                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because the target port"
2492                        " has not yet been enabled\n");
2493                 goto reject;
2494         }
2495
2496         if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2497                 rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_NO_CHAN;
2498
2499                 spin_lock_irq(&sdev->spinlock);
2500
2501                 list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list) {
2502                         if (!memcmp(ch->i_port_id, req->initiator_port_id, 16)
2503                             && !memcmp(ch->t_port_id, req->target_port_id, 16)
2504                             && param->port == ch->sport->port
2505                             && param->listen_id == ch->sport->sdev->cm_id
2506                             && ch->cm_id) {
2507                                 enum rdma_ch_state ch_state;
2508
2509                                 ch_state = srpt_get_ch_state(ch);
2510                                 if (ch_state != CH_CONNECTING
2511                                     && ch_state != CH_LIVE)
2512                                         continue;
2513
2514                                 /* found an existing channel */
2515                                 pr_debug("Found existing channel %s"
2516                                          " cm_id= %p state= %d\n",
2517                                          ch->sess_name, ch->cm_id, ch_state);
2518
2519                                 __srpt_close_ch(ch);
2520
2521                                 rsp->rsp_flags =
2522                                         SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2523                         }
2524                 }
2525
2526                 spin_unlock_irq(&sdev->spinlock);
2527
2528         } else
2529                 rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2530
2531         if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2532             || *(__be64 *)(req->target_port_id + 8) !=
2533                cpu_to_be64(srpt_service_guid)) {
2534                 rej->reason = __constant_cpu_to_be32(
2535                                 SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2536                 ret = -ENOMEM;
2537                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because it"
2538                        " has an invalid target port identifier.\n");
2539                 goto reject;
2540         }
2541
2542         ch = kzalloc(sizeof *ch, GFP_KERNEL);
2543         if (!ch) {
2544                 rej->reason = __constant_cpu_to_be32(
2545                                         SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2546                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because no memory.\n");
2547                 ret = -ENOMEM;
2548                 goto reject;
2549         }
2550
2551         INIT_WORK(&ch->release_work, srpt_release_channel_work);
2552         memcpy(ch->i_port_id, req->initiator_port_id, 16);
2553         memcpy(ch->t_port_id, req->target_port_id, 16);
2554         ch->sport = &sdev->port[param->port - 1];
2555         ch->cm_id = cm_id;
2556         /*
2557          * Avoid QUEUE_FULL conditions by limiting the number of buffers used
2558          * for the SRP protocol to the command queue size.
2559          */
2560         ch->rq_size = SRPT_RQ_SIZE;
2561         spin_lock_init(&ch->spinlock);
2562         ch->state = CH_CONNECTING;
2563         INIT_LIST_HEAD(&ch->cmd_wait_list);
2564         ch->rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2565
2566         ch->ioctx_ring = (struct srpt_send_ioctx **)
2567                 srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2568                                       sizeof(*ch->ioctx_ring[0]),
2569                                       ch->rsp_size, DMA_TO_DEVICE);
2570         if (!ch->ioctx_ring)
2571                 goto free_ch;
2572
2573         INIT_LIST_HEAD(&ch->free_list);
2574         for (i = 0; i < ch->rq_size; i++) {
2575                 ch->ioctx_ring[i]->ch = ch;
2576                 list_add_tail(&ch->ioctx_ring[i]->free_list, &ch->free_list);
2577         }
2578
2579         ret = srpt_create_ch_ib(ch);
2580         if (ret) {
2581                 rej->reason = __constant_cpu_to_be32(
2582                                 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2583                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because creating"
2584                        " a new RDMA channel failed.\n");
2585                 goto free_ring;
2586         }
2587
2588         ret = srpt_ch_qp_rtr(ch, ch->qp);
2589         if (ret) {
2590                 rej->reason = __constant_cpu_to_be32(
2591                                 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2592                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because enabling"
2593                        " RTR failed (error code = %d)\n", ret);
2594                 goto destroy_ib;
2595         }
2596         /*
2597          * Use the initator port identifier as the session name.
2598          */
2599         snprintf(ch->sess_name, sizeof(ch->sess_name), "0x%016llx%016llx",
2600                         be64_to_cpu(*(__be64 *)ch->i_port_id),
2601                         be64_to_cpu(*(__be64 *)(ch->i_port_id + 8)));
2602
2603         pr_debug("registering session %s\n", ch->sess_name);
2604
2605         nacl = srpt_lookup_acl(sport, ch->i_port_id);
2606         if (!nacl) {
2607                 printk(KERN_INFO "Rejected login because no ACL has been"
2608                        " configured yet for initiator %s.\n", ch->sess_name);
2609                 rej->reason = __constant_cpu_to_be32(
2610                                 SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2611                 goto destroy_ib;
2612         }
2613
2614         ch->sess = transport_init_session();
2615         if (IS_ERR(ch->sess)) {
2616                 rej->reason = __constant_cpu_to_be32(
2617                                 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2618                 pr_debug("Failed to create session\n");
2619                 goto deregister_session;
2620         }
2621         ch->sess->se_node_acl = &nacl->nacl;
2622         transport_register_session(&sport->port_tpg_1, &nacl->nacl, ch->sess, ch);
2623
2624         pr_debug("Establish connection sess=%p name=%s cm_id=%p\n", ch->sess,
2625                  ch->sess_name, ch->cm_id);
2626
2627         /* create srp_login_response */
2628         rsp->opcode = SRP_LOGIN_RSP;
2629         rsp->tag = req->tag;
2630         rsp->max_it_iu_len = req->req_it_iu_len;
2631         rsp->max_ti_iu_len = req->req_it_iu_len;
2632         ch->max_ti_iu_len = it_iu_len;
2633         rsp->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2634                                               | SRP_BUF_FORMAT_INDIRECT);
2635         rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2636         atomic_set(&ch->req_lim, ch->rq_size);
2637         atomic_set(&ch->req_lim_delta, 0);
2638
2639         /* create cm reply */
2640         rep_param->qp_num = ch->qp->qp_num;
2641         rep_param->private_data = (void *)rsp;
2642         rep_param->private_data_len = sizeof *rsp;
2643         rep_param->rnr_retry_count = 7;
2644         rep_param->flow_control = 1;
2645         rep_param->failover_accepted = 0;
2646         rep_param->srq = 1;
2647         rep_param->responder_resources = 4;
2648         rep_param->initiator_depth = 4;
2649
2650         ret = ib_send_cm_rep(cm_id, rep_param);
2651         if (ret) {
2652                 printk(KERN_ERR "sending SRP_LOGIN_REQ response failed"
2653                        " (error code = %d)\n", ret);
2654                 goto release_channel;
2655         }
2656
2657         spin_lock_irq(&sdev->spinlock);
2658         list_add_tail(&ch->list, &sdev->rch_list);
2659         spin_unlock_irq(&sdev->spinlock);
2660
2661         goto out;
2662
2663 release_channel:
2664         srpt_set_ch_state(ch, CH_RELEASING);
2665         transport_deregister_session_configfs(ch->sess);
2666
2667 deregister_session:
2668         transport_deregister_session(ch->sess);
2669         ch->sess = NULL;
2670
2671 destroy_ib:
2672         srpt_destroy_ch_ib(ch);
2673
2674 free_ring:
2675         srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2676                              ch->sport->sdev, ch->rq_size,
2677                              ch->rsp_size, DMA_TO_DEVICE);
2678 free_ch:
2679         kfree(ch);
2680
2681 reject:
2682         rej->opcode = SRP_LOGIN_REJ;
2683         rej->tag = req->tag;
2684         rej->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2685                                               | SRP_BUF_FORMAT_INDIRECT);
2686
2687         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2688                              (void *)rej, sizeof *rej);
2689
2690 out:
2691         kfree(rep_param);
2692         kfree(rsp);
2693         kfree(rej);
2694
2695         return ret;
2696 }
2697
2698 static void srpt_cm_rej_recv(struct ib_cm_id *cm_id)
2699 {
2700         printk(KERN_INFO "Received IB REJ for cm_id %p.\n", cm_id);
2701         srpt_drain_channel(cm_id);
2702 }
2703
2704 /**
2705  * srpt_cm_rtu_recv() - Process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event.
2706  *
2707  * An IB_CM_RTU_RECEIVED message indicates that the connection is established
2708  * and that the recipient may begin transmitting (RTU = ready to use).
2709  */
2710 static void srpt_cm_rtu_recv(struct ib_cm_id *cm_id)
2711 {
2712         struct srpt_rdma_ch *ch;
2713         int ret;
2714
2715         ch = srpt_find_channel(cm_id->context, cm_id);
2716         BUG_ON(!ch);
2717
2718         if (srpt_test_and_set_ch_state(ch, CH_CONNECTING, CH_LIVE)) {
2719                 struct srpt_recv_ioctx *ioctx, *ioctx_tmp;
2720
2721                 ret = srpt_ch_qp_rts(ch, ch->qp);
2722
2723                 list_for_each_entry_safe(ioctx, ioctx_tmp, &ch->cmd_wait_list,
2724                                          wait_list) {
2725                         list_del(&ioctx->wait_list);
2726                         srpt_handle_new_iu(ch, ioctx, NULL);
2727                 }
2728                 if (ret)
2729                         srpt_close_ch(ch);
2730         }
2731 }
2732
2733 static void srpt_cm_timewait_exit(struct ib_cm_id *cm_id)
2734 {
2735         printk(KERN_INFO "Received IB TimeWait exit for cm_id %p.\n", cm_id);
2736         srpt_drain_channel(cm_id);
2737 }
2738
2739 static void srpt_cm_rep_error(struct ib_cm_id *cm_id)
2740 {
2741         printk(KERN_INFO "Received IB REP error for cm_id %p.\n", cm_id);
2742         srpt_drain_channel(cm_id);
2743 }
2744
2745 /**
2746  * srpt_cm_dreq_recv() - Process reception of a DREQ message.
2747  */
2748 static void srpt_cm_dreq_recv(struct ib_cm_id *cm_id)
2749 {
2750         struct srpt_rdma_ch *ch;
2751         unsigned long flags;
2752         bool send_drep = false;
2753
2754         ch = srpt_find_channel(cm_id->context, cm_id);
2755         BUG_ON(!ch);
2756
2757         pr_debug("cm_id= %p ch->state= %d\n", cm_id, srpt_get_ch_state(ch));
2758
2759         spin_lock_irqsave(&ch->spinlock, flags);
2760         switch (ch->state) {
2761         case CH_CONNECTING:
2762         case CH_LIVE:
2763                 send_drep = true;
2764                 ch->state = CH_DISCONNECTING;
2765                 break;
2766         case CH_DISCONNECTING:
2767         case CH_DRAINING:
2768         case CH_RELEASING:
2769                 WARN(true, "unexpected channel state %d\n", ch->state);
2770                 break;
2771         }
2772         spin_unlock_irqrestore(&ch->spinlock, flags);
2773
2774         if (send_drep) {
2775                 if (ib_send_cm_drep(ch->cm_id, NULL, 0) < 0)
2776                         printk(KERN_ERR "Sending IB DREP failed.\n");
2777                 printk(KERN_INFO "Received DREQ and sent DREP for session %s.\n",
2778                        ch->sess_name);
2779         }
2780 }
2781
2782 /**
2783  * srpt_cm_drep_recv() - Process reception of a DREP message.
2784  */
2785 static void srpt_cm_drep_recv(struct ib_cm_id *cm_id)
2786 {
2787         printk(KERN_INFO "Received InfiniBand DREP message for cm_id %p.\n",
2788                cm_id);
2789         srpt_drain_channel(cm_id);
2790 }
2791
2792 /**
2793  * srpt_cm_handler() - IB connection manager callback function.
2794  *
2795  * A non-zero return value will cause the caller destroy the CM ID.
2796  *
2797  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2798  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2799  * a non-zero value in any other case will trigger a race with the
2800  * ib_destroy_cm_id() call in srpt_release_channel().
2801  */
2802 static int srpt_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2803 {
2804         int ret;
2805
2806         ret = 0;
2807         switch (event->event) {
2808         case IB_CM_REQ_RECEIVED:
2809                 ret = srpt_cm_req_recv(cm_id, &event->param.req_rcvd,
2810                                        event->private_data);
2811                 break;
2812         case IB_CM_REJ_RECEIVED:
2813                 srpt_cm_rej_recv(cm_id);
2814                 break;
2815         case IB_CM_RTU_RECEIVED:
2816         case IB_CM_USER_ESTABLISHED:
2817                 srpt_cm_rtu_recv(cm_id);
2818                 break;
2819         case IB_CM_DREQ_RECEIVED:
2820                 srpt_cm_dreq_recv(cm_id);
2821                 break;
2822         case IB_CM_DREP_RECEIVED:
2823                 srpt_cm_drep_recv(cm_id);
2824                 break;
2825         case IB_CM_TIMEWAIT_EXIT:
2826                 srpt_cm_timewait_exit(cm_id);
2827                 break;
2828         case IB_CM_REP_ERROR:
2829                 srpt_cm_rep_error(cm_id);
2830                 break;
2831         case IB_CM_DREQ_ERROR:
2832                 printk(KERN_INFO "Received IB DREQ ERROR event.\n");
2833                 break;
2834         case IB_CM_MRA_RECEIVED:
2835                 printk(KERN_INFO "Received IB MRA event\n");
2836                 break;
2837         default:
2838                 printk(KERN_ERR "received unrecognized IB CM event %d\n",
2839                        event->event);
2840                 break;
2841         }
2842
2843         return ret;
2844 }
2845
2846 /**
2847  * srpt_perform_rdmas() - Perform IB RDMA.
2848  *
2849  * Returns zero upon success or a negative number upon failure.
2850  */
2851 static int srpt_perform_rdmas(struct srpt_rdma_ch *ch,
2852                               struct srpt_send_ioctx *ioctx)
2853 {
2854         struct ib_send_wr wr;
2855         struct ib_send_wr *bad_wr;
2856         struct rdma_iu *riu;
2857         int i;
2858         int ret;
2859         int sq_wr_avail;
2860         enum dma_data_direction dir;
2861         const int n_rdma = ioctx->n_rdma;
2862
2863         dir = ioctx->cmd.data_direction;
2864         if (dir == DMA_TO_DEVICE) {
2865                 /* write */
2866                 ret = -ENOMEM;
2867                 sq_wr_avail = atomic_sub_return(n_rdma, &ch->sq_wr_avail);
2868                 if (sq_wr_avail < 0) {
2869                         printk(KERN_WARNING "IB send queue full (needed %d)\n",
2870                                n_rdma);
2871                         goto out;
2872                 }
2873         }
2874
2875         ioctx->rdma_aborted = false;
2876         ret = 0;
2877         riu = ioctx->rdma_ius;
2878         memset(&wr, 0, sizeof wr);
2879
2880         for (i = 0; i < n_rdma; ++i, ++riu) {
2881                 if (dir == DMA_FROM_DEVICE) {
2882                         wr.opcode = IB_WR_RDMA_WRITE;
2883                         wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2884                                                 SRPT_RDMA_WRITE_LAST :
2885                                                 SRPT_RDMA_MID,
2886                                                 ioctx->ioctx.index);
2887                 } else {
2888                         wr.opcode = IB_WR_RDMA_READ;
2889                         wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2890                                                 SRPT_RDMA_READ_LAST :
2891                                                 SRPT_RDMA_MID,
2892                                                 ioctx->ioctx.index);
2893                 }
2894                 wr.next = NULL;
2895                 wr.wr.rdma.remote_addr = riu->raddr;
2896                 wr.wr.rdma.rkey = riu->rkey;
2897                 wr.num_sge = riu->sge_cnt;
2898                 wr.sg_list = riu->sge;
2899
2900                 /* only get completion event for the last rdma write */
2901                 if (i == (n_rdma - 1) && dir == DMA_TO_DEVICE)
2902                         wr.send_flags = IB_SEND_SIGNALED;
2903
2904                 ret = ib_post_send(ch->qp, &wr, &bad_wr);
2905                 if (ret)
2906                         break;
2907         }
2908
2909         if (ret)
2910                 printk(KERN_ERR "%s[%d]: ib_post_send() returned %d for %d/%d",
2911                                  __func__, __LINE__, ret, i, n_rdma);
2912         if (ret && i > 0) {
2913                 wr.num_sge = 0;
2914                 wr.wr_id = encode_wr_id(SRPT_RDMA_ABORT, ioctx->ioctx.index);
2915                 wr.send_flags = IB_SEND_SIGNALED;
2916                 while (ch->state == CH_LIVE &&
2917                         ib_post_send(ch->qp, &wr, &bad_wr) != 0) {
2918                         printk(KERN_INFO "Trying to abort failed RDMA transfer [%d]",
2919                                 ioctx->ioctx.index);
2920                         msleep(1000);
2921                 }
2922                 while (ch->state != CH_RELEASING && !ioctx->rdma_aborted) {
2923                         printk(KERN_INFO "Waiting until RDMA abort finished [%d]",
2924                                 ioctx->ioctx.index);
2925                         msleep(1000);
2926                 }
2927         }
2928 out:
2929         if (unlikely(dir == DMA_TO_DEVICE && ret < 0))
2930                 atomic_add(n_rdma, &ch->sq_wr_avail);
2931         return ret;
2932 }
2933
2934 /**
2935  * srpt_xfer_data() - Start data transfer from initiator to target.
2936  */
2937 static int srpt_xfer_data(struct srpt_rdma_ch *ch,
2938                           struct srpt_send_ioctx *ioctx)
2939 {
2940         int ret;
2941
2942         ret = srpt_map_sg_to_ib_sge(ch, ioctx);
2943         if (ret) {
2944                 printk(KERN_ERR "%s[%d] ret=%d\n", __func__, __LINE__, ret);
2945                 goto out;
2946         }
2947
2948         ret = srpt_perform_rdmas(ch, ioctx);
2949         if (ret) {
2950                 if (ret == -EAGAIN || ret == -ENOMEM)
2951                         printk(KERN_INFO "%s[%d] queue full -- ret=%d\n",
2952                                    __func__, __LINE__, ret);
2953                 else
2954                         printk(KERN_ERR "%s[%d] fatal error -- ret=%d\n",
2955                                __func__, __LINE__, ret);
2956                 goto out_unmap;
2957         }
2958
2959 out:
2960         return ret;
2961 out_unmap:
2962         srpt_unmap_sg_to_ib_sge(ch, ioctx);
2963         goto out;
2964 }
2965
2966 static int srpt_write_pending_status(struct se_cmd *se_cmd)
2967 {
2968         struct srpt_send_ioctx *ioctx;
2969
2970         ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2971         return srpt_get_cmd_state(ioctx) == SRPT_STATE_NEED_DATA;
2972 }
2973
2974 /*
2975  * srpt_write_pending() - Start data transfer from initiator to target (write).
2976  */
2977 static int srpt_write_pending(struct se_cmd *se_cmd)
2978 {
2979         struct srpt_rdma_ch *ch;
2980         struct srpt_send_ioctx *ioctx;
2981         enum srpt_command_state new_state;
2982         enum rdma_ch_state ch_state;
2983         int ret;
2984
2985         ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2986
2987         new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2988         WARN_ON(new_state == SRPT_STATE_DONE);
2989
2990         ch = ioctx->ch;
2991         BUG_ON(!ch);
2992
2993         ch_state = srpt_get_ch_state(ch);
2994         switch (ch_state) {
2995         case CH_CONNECTING:
2996                 WARN(true, "unexpected channel state %d\n", ch_state);
2997                 ret = -EINVAL;
2998                 goto out;
2999         case CH_LIVE:
3000                 break;
3001         case CH_DISCONNECTING:
3002         case CH_DRAINING:
3003         case CH_RELEASING:
3004                 pr_debug("cmd with tag %lld: channel disconnecting\n",
3005                          ioctx->tag);
3006                 srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
3007                 ret = -EINVAL;
3008                 goto out;
3009         }
3010         ret = srpt_xfer_data(ch, ioctx);
3011
3012 out:
3013         return ret;
3014 }
3015
3016 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
3017 {
3018         switch (tcm_mgmt_status) {
3019         case TMR_FUNCTION_COMPLETE:
3020                 return SRP_TSK_MGMT_SUCCESS;
3021         case TMR_FUNCTION_REJECTED:
3022                 return SRP_TSK_MGMT_FUNC_NOT_SUPP;
3023         }
3024         return SRP_TSK_MGMT_FAILED;
3025 }
3026
3027 /**
3028  * srpt_queue_response() - Transmits the response to a SCSI command.
3029  *
3030  * Callback function called by the TCM core. Must not block since it can be
3031  * invoked on the context of the IB completion handler.
3032  */
3033 static int srpt_queue_response(struct se_cmd *cmd)
3034 {
3035         struct srpt_rdma_ch *ch;
3036         struct srpt_send_ioctx *ioctx;
3037         enum srpt_command_state state;
3038         unsigned long flags;
3039         int ret;
3040         enum dma_data_direction dir;
3041         int resp_len;
3042         u8 srp_tm_status;
3043
3044         ret = 0;
3045
3046         ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3047         ch = ioctx->ch;
3048         BUG_ON(!ch);
3049
3050         spin_lock_irqsave(&ioctx->spinlock, flags);
3051         state = ioctx->state;
3052         switch (state) {
3053         case SRPT_STATE_NEW:
3054         case SRPT_STATE_DATA_IN:
3055                 ioctx->state = SRPT_STATE_CMD_RSP_SENT;
3056                 break;
3057         case SRPT_STATE_MGMT:
3058                 ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
3059                 break;
3060         default:
3061                 WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
3062                         ch, ioctx->ioctx.index, ioctx->state);
3063                 break;
3064         }
3065         spin_unlock_irqrestore(&ioctx->spinlock, flags);
3066
3067         if (unlikely(transport_check_aborted_status(&ioctx->cmd, false)
3068                      || WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))) {
3069                 atomic_inc(&ch->req_lim_delta);
3070                 srpt_abort_cmd(ioctx);
3071                 goto out;
3072         }
3073
3074         dir = ioctx->cmd.data_direction;
3075
3076         /* For read commands, transfer the data to the initiator. */
3077         if (dir == DMA_FROM_DEVICE && ioctx->cmd.data_length &&
3078             !ioctx->queue_status_only) {
3079                 ret = srpt_xfer_data(ch, ioctx);
3080                 if (ret) {
3081                         printk(KERN_ERR "xfer_data failed for tag %llu\n",
3082                                ioctx->tag);
3083                         goto out;
3084                 }
3085         }
3086
3087         if (state != SRPT_STATE_MGMT)
3088                 resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->tag,
3089                                               cmd->scsi_status);
3090         else {
3091                 srp_tm_status
3092                         = tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
3093                 resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
3094                                                  ioctx->tag);
3095         }
3096         ret = srpt_post_send(ch, ioctx, resp_len);
3097         if (ret) {
3098                 printk(KERN_ERR "sending cmd response failed for tag %llu\n",
3099                        ioctx->tag);
3100                 srpt_unmap_sg_to_ib_sge(ch, ioctx);
3101                 srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
3102                 kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
3103         }
3104
3105 out:
3106         return ret;
3107 }
3108
3109 static int srpt_queue_status(struct se_cmd *cmd)
3110 {
3111         struct srpt_send_ioctx *ioctx;
3112
3113         ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3114         BUG_ON(ioctx->sense_data != cmd->sense_buffer);
3115         if (cmd->se_cmd_flags &
3116             (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
3117                 WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
3118         ioctx->queue_status_only = true;
3119         return srpt_queue_response(cmd);
3120 }
3121
3122 static void srpt_refresh_port_work(struct work_struct *work)
3123 {
3124         struct srpt_port *sport = container_of(work, struct srpt_port, work);
3125
3126         srpt_refresh_port(sport);
3127 }
3128
3129 static int srpt_ch_list_empty(struct srpt_device *sdev)
3130 {
3131         int res;
3132
3133         spin_lock_irq(&sdev->spinlock);
3134         res = list_empty(&sdev->rch_list);
3135         spin_unlock_irq(&sdev->spinlock);
3136
3137         return res;
3138 }
3139
3140 /**
3141  * srpt_release_sdev() - Free the channel resources associated with a target.
3142  */
3143 static int srpt_release_sdev(struct srpt_device *sdev)
3144 {
3145         struct srpt_rdma_ch *ch, *tmp_ch;
3146         int res;
3147
3148         WARN_ON_ONCE(irqs_disabled());
3149
3150         BUG_ON(!sdev);
3151
3152         spin_lock_irq(&sdev->spinlock);
3153         list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list)
3154                 __srpt_close_ch(ch);
3155         spin_unlock_irq(&sdev->spinlock);
3156
3157         res = wait_event_interruptible(sdev->ch_releaseQ,
3158                                        srpt_ch_list_empty(sdev));
3159         if (res)
3160                 printk(KERN_ERR "%s: interrupted.\n", __func__);
3161
3162         return 0;
3163 }
3164
3165 static struct srpt_port *__srpt_lookup_port(const char *name)
3166 {
3167         struct ib_device *dev;
3168         struct srpt_device *sdev;
3169         struct srpt_port *sport;
3170         int i;
3171
3172         list_for_each_entry(sdev, &srpt_dev_list, list) {
3173                 dev = sdev->device;
3174                 if (!dev)
3175                         continue;
3176
3177                 for (i = 0; i < dev->phys_port_cnt; i++) {
3178                         sport = &sdev->port[i];
3179
3180                         if (!strcmp(sport->port_guid, name))
3181                                 return sport;
3182                 }
3183         }
3184
3185         return NULL;
3186 }
3187
3188 static struct srpt_port *srpt_lookup_port(const char *name)
3189 {
3190         struct srpt_port *sport;
3191
3192         spin_lock(&srpt_dev_lock);
3193         sport = __srpt_lookup_port(name);
3194         spin_unlock(&srpt_dev_lock);
3195
3196         return sport;
3197 }
3198
3199 /**
3200  * srpt_add_one() - Infiniband device addition callback function.
3201  */
3202 static void srpt_add_one(struct ib_device *device)
3203 {
3204         struct srpt_device *sdev;
3205         struct srpt_port *sport;
3206         struct ib_srq_init_attr srq_attr;
3207         int i;
3208
3209         pr_debug("device = %p, device->dma_ops = %p\n", device,
3210                  device->dma_ops);
3211
3212         sdev = kzalloc(sizeof *sdev, GFP_KERNEL);
3213         if (!sdev)
3214                 goto err;
3215
3216         sdev->device = device;
3217         INIT_LIST_HEAD(&sdev->rch_list);
3218         init_waitqueue_head(&sdev->ch_releaseQ);
3219         spin_lock_init(&sdev->spinlock);
3220
3221         if (ib_query_device(device, &sdev->dev_attr))
3222                 goto free_dev;
3223
3224         sdev->pd = ib_alloc_pd(device);
3225         if (IS_ERR(sdev->pd))
3226                 goto free_dev;
3227
3228         sdev->mr = ib_get_dma_mr(sdev->pd, IB_ACCESS_LOCAL_WRITE);
3229         if (IS_ERR(sdev->mr))
3230                 goto err_pd;
3231
3232         sdev->srq_size = min(srpt_srq_size, sdev->dev_attr.max_srq_wr);
3233
3234         srq_attr.event_handler = srpt_srq_event;
3235         srq_attr.srq_context = (void *)sdev;
3236         srq_attr.attr.max_wr = sdev->srq_size;
3237         srq_attr.attr.max_sge = 1;
3238         srq_attr.attr.srq_limit = 0;
3239         srq_attr.srq_type = IB_SRQT_BASIC;
3240
3241         sdev->srq = ib_create_srq(sdev->pd, &srq_attr);
3242         if (IS_ERR(sdev->srq))
3243                 goto err_mr;
3244
3245         pr_debug("%s: create SRQ #wr= %d max_allow=%d dev= %s\n",
3246                  __func__, sdev->srq_size, sdev->dev_attr.max_srq_wr,
3247                  device->name);
3248
3249         if (!srpt_service_guid)
3250                 srpt_service_guid = be64_to_cpu(device->node_guid);
3251
3252         sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3253         if (IS_ERR(sdev->cm_id))
3254                 goto err_srq;
3255
3256         /* print out target login information */
3257         pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,"
3258                  "pkey=ffff,service_id=%016llx\n", srpt_service_guid,
3259                  srpt_service_guid, srpt_service_guid);
3260
3261         /*
3262          * We do not have a consistent service_id (ie. also id_ext of target_id)
3263          * to identify this target. We currently use the guid of the first HCA
3264          * in the system as service_id; therefore, the target_id will change
3265          * if this HCA is gone bad and replaced by different HCA
3266          */
3267         if (ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0, NULL))
3268                 goto err_cm;
3269
3270         INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3271                               srpt_event_handler);
3272         if (ib_register_event_handler(&sdev->event_handler))
3273                 goto err_cm;
3274
3275         sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3276                 srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3277                                       sizeof(*sdev->ioctx_ring[0]),
3278                                       srp_max_req_size, DMA_FROM_DEVICE);
3279         if (!sdev->ioctx_ring)
3280                 goto err_event;
3281
3282         for (i = 0; i < sdev->srq_size; ++i)
3283                 srpt_post_recv(sdev, sdev->ioctx_ring[i]);
3284
3285         WARN_ON(sdev->device->phys_port_cnt > ARRAY_SIZE(sdev->port));
3286
3287         for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3288                 sport = &sdev->port[i - 1];
3289                 sport->sdev = sdev;
3290                 sport->port = i;
3291                 sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3292                 sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3293                 sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3294                 INIT_WORK(&sport->work, srpt_refresh_port_work);
3295                 INIT_LIST_HEAD(&sport->port_acl_list);
3296                 spin_lock_init(&sport->port_acl_lock);
3297
3298                 if (srpt_refresh_port(sport)) {
3299                         printk(KERN_ERR "MAD registration failed for %s-%d.\n",
3300                                srpt_sdev_name(sdev), i);
3301                         goto err_ring;
3302                 }
3303                 snprintf(sport->port_guid, sizeof(sport->port_guid),
3304                         "0x%016llx%016llx",
3305                         be64_to_cpu(sport->gid.global.subnet_prefix),
3306                         be64_to_cpu(sport->gid.global.interface_id));
3307         }
3308
3309         spin_lock(&srpt_dev_lock);
3310         list_add_tail(&sdev->list, &srpt_dev_list);
3311         spin_unlock(&srpt_dev_lock);
3312
3313 out:
3314         ib_set_client_data(device, &srpt_client, sdev);
3315         pr_debug("added %s.\n", device->name);
3316         return;
3317
3318 err_ring:
3319         srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3320                              sdev->srq_size, srp_max_req_size,
3321                              DMA_FROM_DEVICE);
3322 err_event:
3323         ib_unregister_event_handler(&sdev->event_handler);
3324 err_cm:
3325         ib_destroy_cm_id(sdev->cm_id);
3326 err_srq:
3327         ib_destroy_srq(sdev->srq);
3328 err_mr:
3329         ib_dereg_mr(sdev->mr);
3330 err_pd:
3331         ib_dealloc_pd(sdev->pd);
3332 free_dev:
3333         kfree(sdev);
3334 err:
3335         sdev = NULL;
3336         printk(KERN_INFO "%s(%s) failed.\n", __func__, device->name);
3337         goto out;
3338 }
3339
3340 /**
3341  * srpt_remove_one() - InfiniBand device removal callback function.
3342  */
3343 static void srpt_remove_one(struct ib_device *device)
3344 {
3345         struct srpt_device *sdev;
3346         int i;
3347
3348         sdev = ib_get_client_data(device, &srpt_client);
3349         if (!sdev) {
3350                 printk(KERN_INFO "%s(%s): nothing to do.\n", __func__,
3351                        device->name);
3352                 return;
3353         }
3354
3355         srpt_unregister_mad_agent(sdev);
3356
3357         ib_unregister_event_handler(&sdev->event_handler);
3358
3359         /* Cancel any work queued by the just unregistered IB event handler. */
3360         for (i = 0; i < sdev->device->phys_port_cnt; i++)
3361                 cancel_work_sync(&sdev->port[i].work);
3362
3363         ib_destroy_cm_id(sdev->cm_id);
3364
3365         /*
3366          * Unregistering a target must happen after destroying sdev->cm_id
3367          * such that no new SRP_LOGIN_REQ information units can arrive while
3368          * destroying the target.
3369          */
3370         spin_lock(&srpt_dev_lock);
3371         list_del(&sdev->list);
3372         spin_unlock(&srpt_dev_lock);
3373         srpt_release_sdev(sdev);
3374
3375         ib_destroy_srq(sdev->srq);
3376         ib_dereg_mr(sdev->mr);
3377         ib_dealloc_pd(sdev->pd);
3378
3379         srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3380                              sdev->srq_size, srp_max_req_size, DMA_FROM_DEVICE);
3381         sdev->ioctx_ring = NULL;
3382         kfree(sdev);
3383 }
3384
3385 static struct ib_client srpt_client = {
3386         .name = DRV_NAME,
3387         .add = srpt_add_one,
3388         .remove = srpt_remove_one
3389 };
3390
3391 static int srpt_check_true(struct se_portal_group *se_tpg)
3392 {
3393         return 1;
3394 }
3395
3396 static int srpt_check_false(struct se_portal_group *se_tpg)
3397 {
3398         return 0;
3399 }
3400
3401 static char *srpt_get_fabric_name(void)
3402 {
3403         return "srpt";
3404 }
3405
3406 static u8 srpt_get_fabric_proto_ident(struct se_portal_group *se_tpg)
3407 {
3408         return SCSI_TRANSPORTID_PROTOCOLID_SRP;
3409 }
3410
3411 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3412 {
3413         struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3414
3415         return sport->port_guid;
3416 }
3417
3418 static u16 srpt_get_tag(struct se_portal_group *tpg)
3419 {
3420         return 1;
3421 }
3422
3423 static u32 srpt_get_default_depth(struct se_portal_group *se_tpg)
3424 {
3425         return 1;
3426 }
3427
3428 static u32 srpt_get_pr_transport_id(struct se_portal_group *se_tpg,
3429                                     struct se_node_acl *se_nacl,
3430                                     struct t10_pr_registration *pr_reg,
3431                                     int *format_code, unsigned char *buf)
3432 {
3433         struct srpt_node_acl *nacl;
3434         struct spc_rdma_transport_id *tr_id;
3435
3436         nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3437         tr_id = (void *)buf;
3438         tr_id->protocol_identifier = SCSI_TRANSPORTID_PROTOCOLID_SRP;
3439         memcpy(tr_id->i_port_id, nacl->i_port_id, sizeof(tr_id->i_port_id));
3440         return sizeof(*tr_id);
3441 }
3442
3443 static u32 srpt_get_pr_transport_id_len(struct se_portal_group *se_tpg,
3444                                         struct se_node_acl *se_nacl,
3445                                         struct t10_pr_registration *pr_reg,
3446                                         int *format_code)
3447 {
3448         *format_code = 0;
3449         return sizeof(struct spc_rdma_transport_id);
3450 }
3451
3452 static char *srpt_parse_pr_out_transport_id(struct se_portal_group *se_tpg,
3453                                             const char *buf, u32 *out_tid_len,
3454                                             char **port_nexus_ptr)
3455 {
3456         struct spc_rdma_transport_id *tr_id;
3457
3458         *port_nexus_ptr = NULL;
3459         *out_tid_len = sizeof(struct spc_rdma_transport_id);
3460         tr_id = (void *)buf;
3461         return (char *)tr_id->i_port_id;
3462 }
3463
3464 static struct se_node_acl *srpt_alloc_fabric_acl(struct se_portal_group *se_tpg)
3465 {
3466         struct srpt_node_acl *nacl;
3467
3468         nacl = kzalloc(sizeof(struct srpt_node_acl), GFP_KERNEL);
3469         if (!nacl) {
3470                 printk(KERN_ERR "Unable to allocate struct srpt_node_acl\n");
3471                 return NULL;
3472         }
3473
3474         return &nacl->nacl;
3475 }
3476
3477 static void srpt_release_fabric_acl(struct se_portal_group *se_tpg,
3478                                     struct se_node_acl *se_nacl)
3479 {
3480         struct srpt_node_acl *nacl;
3481
3482         nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3483         kfree(nacl);
3484 }
3485
3486 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3487 {
3488         return 1;
3489 }
3490
3491 static void srpt_release_cmd(struct se_cmd *se_cmd)
3492 {
3493 }
3494
3495 /**
3496  * srpt_shutdown_session() - Whether or not a session may be shut down.
3497  */
3498 static int srpt_shutdown_session(struct se_session *se_sess)
3499 {
3500         return true;
3501 }
3502
3503 /**
3504  * srpt_close_session() - Forcibly close a session.
3505  *
3506  * Callback function invoked by the TCM core to clean up sessions associated
3507  * with a node ACL when the user invokes
3508  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3509  */
3510 static void srpt_close_session(struct se_session *se_sess)
3511 {
3512         DECLARE_COMPLETION_ONSTACK(release_done);
3513         struct srpt_rdma_ch *ch;
3514         struct srpt_device *sdev;
3515         int res;
3516
3517         ch = se_sess->fabric_sess_ptr;
3518         WARN_ON(ch->sess != se_sess);
3519
3520         pr_debug("ch %p state %d\n", ch, srpt_get_ch_state(ch));
3521
3522         sdev = ch->sport->sdev;
3523         spin_lock_irq(&sdev->spinlock);
3524         BUG_ON(ch->release_done);
3525         ch->release_done = &release_done;
3526         __srpt_close_ch(ch);
3527         spin_unlock_irq(&sdev->spinlock);
3528
3529         res = wait_for_completion_timeout(&release_done, 60 * HZ);
3530         WARN_ON(res <= 0);
3531 }
3532
3533 /**
3534  * srpt_sess_get_index() - Return the value of scsiAttIntrPortIndex (SCSI-MIB).
3535  *
3536  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3537  * This object represents an arbitrary integer used to uniquely identify a
3538  * particular attached remote initiator port to a particular SCSI target port
3539  * within a particular SCSI target device within a particular SCSI instance.
3540  */
3541 static u32 srpt_sess_get_index(struct se_session *se_sess)
3542 {
3543         return 0;
3544 }
3545
3546 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3547 {
3548 }
3549
3550 static u32 srpt_get_task_tag(struct se_cmd *se_cmd)
3551 {
3552         struct srpt_send_ioctx *ioctx;
3553
3554         ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3555         return ioctx->tag;
3556 }
3557
3558 /* Note: only used from inside debug printk's by the TCM core. */
3559 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3560 {
3561         struct srpt_send_ioctx *ioctx;
3562
3563         ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3564         return srpt_get_cmd_state(ioctx);
3565 }
3566
3567 static u16 srpt_set_fabric_sense_len(struct se_cmd *cmd, u32 sense_length)
3568 {
3569         return 0;
3570 }
3571
3572 static u16 srpt_get_fabric_sense_len(void)
3573 {
3574         return 0;
3575 }
3576
3577 /**
3578  * srpt_parse_i_port_id() - Parse an initiator port ID.
3579  * @name: ASCII representation of a 128-bit initiator port ID.
3580  * @i_port_id: Binary 128-bit port ID.
3581  */
3582 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3583 {
3584         const char *p;
3585         unsigned len, count, leading_zero_bytes;
3586         int ret, rc;
3587
3588         p = name;
3589         if (strnicmp(p, "0x", 2) == 0)
3590                 p += 2;
3591         ret = -EINVAL;
3592         len = strlen(p);
3593         if (len % 2)
3594                 goto out;
3595         count = min(len / 2, 16U);
3596         leading_zero_bytes = 16 - count;
3597         memset(i_port_id, 0, leading_zero_bytes);
3598         rc = hex2bin(i_port_id + leading_zero_bytes, p, count);
3599         if (rc < 0)
3600                 pr_debug("hex2bin failed for srpt_parse_i_port_id: %d\n", rc);
3601         ret = 0;
3602 out:
3603         return ret;
3604 }
3605
3606 /*
3607  * configfs callback function invoked for
3608  * mkdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3609  */
3610 static struct se_node_acl *srpt_make_nodeacl(struct se_portal_group *tpg,
3611                                              struct config_group *group,
3612                                              const char *name)
3613 {
3614         struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3615         struct se_node_acl *se_nacl, *se_nacl_new;
3616         struct srpt_node_acl *nacl;
3617         int ret = 0;
3618         u32 nexus_depth = 1;
3619         u8 i_port_id[16];
3620
3621         if (srpt_parse_i_port_id(i_port_id, name) < 0) {
3622                 printk(KERN_ERR "invalid initiator port ID %s\n", name);
3623                 ret = -EINVAL;
3624                 goto err;
3625         }
3626
3627         se_nacl_new = srpt_alloc_fabric_acl(tpg);
3628         if (!se_nacl_new) {
3629                 ret = -ENOMEM;
3630                 goto err;
3631         }
3632         /*
3633          * nacl_new may be released by core_tpg_add_initiator_node_acl()
3634          * when converting a node ACL from demo mode to explict
3635          */
3636         se_nacl = core_tpg_add_initiator_node_acl(tpg, se_nacl_new, name,
3637                                                   nexus_depth);
3638         if (IS_ERR(se_nacl)) {
3639                 ret = PTR_ERR(se_nacl);
3640                 goto err;
3641         }
3642         /* Locate our struct srpt_node_acl and set sdev and i_port_id. */
3643         nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3644         memcpy(&nacl->i_port_id[0], &i_port_id[0], 16);
3645         nacl->sport = sport;
3646
3647         spin_lock_irq(&sport->port_acl_lock);
3648         list_add_tail(&nacl->list, &sport->port_acl_list);
3649         spin_unlock_irq(&sport->port_acl_lock);
3650
3651         return se_nacl;
3652 err:
3653         return ERR_PTR(ret);
3654 }
3655
3656 /*
3657  * configfs callback function invoked for
3658  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3659  */
3660 static void srpt_drop_nodeacl(struct se_node_acl *se_nacl)
3661 {
3662         struct srpt_node_acl *nacl;
3663         struct srpt_device *sdev;
3664         struct srpt_port *sport;
3665
3666         nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3667         sport = nacl->sport;
3668         sdev = sport->sdev;
3669         spin_lock_irq(&sport->port_acl_lock);
3670         list_del(&nacl->list);
3671         spin_unlock_irq(&sport->port_acl_lock);
3672         core_tpg_del_initiator_node_acl(&sport->port_tpg_1, se_nacl, 1);
3673         srpt_release_fabric_acl(NULL, se_nacl);
3674 }
3675
3676 static ssize_t srpt_tpg_attrib_show_srp_max_rdma_size(
3677         struct se_portal_group *se_tpg,
3678         char *page)
3679 {
3680         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3681
3682         return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3683 }
3684
3685 static ssize_t srpt_tpg_attrib_store_srp_max_rdma_size(
3686         struct se_portal_group *se_tpg,
3687         const char *page,
3688         size_t count)
3689 {
3690         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3691         unsigned long val;
3692         int ret;
3693
3694         ret = strict_strtoul(page, 0, &val);
3695         if (ret < 0) {
3696                 pr_err("strict_strtoul() failed with ret: %d\n", ret);
3697                 return -EINVAL;
3698         }
3699         if (val > MAX_SRPT_RDMA_SIZE) {
3700                 pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3701                         MAX_SRPT_RDMA_SIZE);
3702                 return -EINVAL;
3703         }
3704         if (val < DEFAULT_MAX_RDMA_SIZE) {
3705                 pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3706                         val, DEFAULT_MAX_RDMA_SIZE);
3707                 return -EINVAL;
3708         }
3709         sport->port_attrib.srp_max_rdma_size = val;
3710
3711         return count;
3712 }
3713
3714 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rdma_size, S_IRUGO | S_IWUSR);
3715
3716 static ssize_t srpt_tpg_attrib_show_srp_max_rsp_size(
3717         struct se_portal_group *se_tpg,
3718         char *page)
3719 {
3720         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3721
3722         return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3723 }
3724
3725 static ssize_t srpt_tpg_attrib_store_srp_max_rsp_size(
3726         struct se_portal_group *se_tpg,
3727         const char *page,
3728         size_t count)
3729 {
3730         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3731         unsigned long val;
3732         int ret;
3733
3734         ret = strict_strtoul(page, 0, &val);
3735         if (ret < 0) {
3736                 pr_err("strict_strtoul() failed with ret: %d\n", ret);
3737                 return -EINVAL;
3738         }
3739         if (val > MAX_SRPT_RSP_SIZE) {
3740                 pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3741                         MAX_SRPT_RSP_SIZE);
3742                 return -EINVAL;
3743         }
3744         if (val < MIN_MAX_RSP_SIZE) {
3745                 pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3746                         MIN_MAX_RSP_SIZE);
3747                 return -EINVAL;
3748         }
3749         sport->port_attrib.srp_max_rsp_size = val;
3750
3751         return count;
3752 }
3753
3754 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rsp_size, S_IRUGO | S_IWUSR);
3755
3756 static ssize_t srpt_tpg_attrib_show_srp_sq_size(
3757         struct se_portal_group *se_tpg,
3758         char *page)
3759 {
3760         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3761
3762         return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size);
3763 }
3764
3765 static ssize_t srpt_tpg_attrib_store_srp_sq_size(
3766         struct se_portal_group *se_tpg,
3767         const char *page,
3768         size_t count)
3769 {
3770         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3771         unsigned long val;
3772         int ret;
3773
3774         ret = strict_strtoul(page, 0, &val);
3775         if (ret < 0) {
3776                 pr_err("strict_strtoul() failed with ret: %d\n", ret);
3777                 return -EINVAL;
3778         }
3779         if (val > MAX_SRPT_SRQ_SIZE) {
3780                 pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3781                         MAX_SRPT_SRQ_SIZE);
3782                 return -EINVAL;
3783         }
3784         if (val < MIN_SRPT_SRQ_SIZE) {
3785                 pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3786                         MIN_SRPT_SRQ_SIZE);
3787                 return -EINVAL;
3788         }
3789         sport->port_attrib.srp_sq_size = val;
3790
3791         return count;
3792 }
3793
3794 TF_TPG_ATTRIB_ATTR(srpt, srp_sq_size, S_IRUGO | S_IWUSR);
3795
3796 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3797         &srpt_tpg_attrib_srp_max_rdma_size.attr,
3798         &srpt_tpg_attrib_srp_max_rsp_size.attr,
3799         &srpt_tpg_attrib_srp_sq_size.attr,
3800         NULL,
3801 };
3802
3803 static ssize_t srpt_tpg_show_enable(
3804         struct se_portal_group *se_tpg,
3805         char *page)
3806 {
3807         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3808
3809         return snprintf(page, PAGE_SIZE, "%d\n", (sport->enabled) ? 1: 0);
3810 }
3811
3812 static ssize_t srpt_tpg_store_enable(
3813         struct se_portal_group *se_tpg,
3814         const char *page,
3815         size_t count)
3816 {
3817         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3818         unsigned long tmp;
3819         int ret;
3820
3821         ret = strict_strtoul(page, 0, &tmp);
3822         if (ret < 0) {
3823                 printk(KERN_ERR "Unable to extract srpt_tpg_store_enable\n");
3824                 return -EINVAL;
3825         }
3826
3827         if ((tmp != 0) && (tmp != 1)) {
3828                 printk(KERN_ERR "Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
3829                 return -EINVAL;
3830         }
3831         if (tmp == 1)
3832                 sport->enabled = true;
3833         else
3834                 sport->enabled = false;
3835
3836         return count;
3837 }
3838
3839 TF_TPG_BASE_ATTR(srpt, enable, S_IRUGO | S_IWUSR);
3840
3841 static struct configfs_attribute *srpt_tpg_attrs[] = {
3842         &srpt_tpg_enable.attr,
3843         NULL,
3844 };
3845
3846 /**
3847  * configfs callback invoked for
3848  * mkdir /sys/kernel/config/target/$driver/$port/$tpg
3849  */
3850 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3851                                              struct config_group *group,
3852                                              const char *name)
3853 {
3854         struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3855         int res;
3856
3857         /* Initialize sport->port_wwn and sport->port_tpg_1 */
3858         res = core_tpg_register(&srpt_target->tf_ops, &sport->port_wwn,
3859                         &sport->port_tpg_1, sport, TRANSPORT_TPG_TYPE_NORMAL);
3860         if (res)
3861                 return ERR_PTR(res);
3862
3863         return &sport->port_tpg_1;
3864 }
3865
3866 /**
3867  * configfs callback invoked for
3868  * rmdir /sys/kernel/config/target/$driver/$port/$tpg
3869  */
3870 static void srpt_drop_tpg(struct se_portal_group *tpg)
3871 {
3872         struct srpt_port *sport = container_of(tpg,
3873                                 struct srpt_port, port_tpg_1);
3874
3875         sport->enabled = false;
3876         core_tpg_deregister(&sport->port_tpg_1);
3877 }
3878
3879 /**
3880  * configfs callback invoked for
3881  * mkdir /sys/kernel/config/target/$driver/$port
3882  */
3883 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3884                                       struct config_group *group,
3885                                       const char *name)
3886 {
3887         struct srpt_port *sport;
3888         int ret;
3889
3890         sport = srpt_lookup_port(name);
3891         pr_debug("make_tport(%s)\n", name);
3892         ret = -EINVAL;
3893         if (!sport)
3894                 goto err;
3895
3896         return &sport->port_wwn;
3897
3898 err:
3899         return ERR_PTR(ret);
3900 }
3901
3902 /**
3903  * configfs callback invoked for
3904  * rmdir /sys/kernel/config/target/$driver/$port
3905  */
3906 static void srpt_drop_tport(struct se_wwn *wwn)
3907 {
3908         struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3909
3910         pr_debug("drop_tport(%s\n", config_item_name(&sport->port_wwn.wwn_group.cg_item));
3911 }
3912
3913 static ssize_t srpt_wwn_show_attr_version(struct target_fabric_configfs *tf,
3914                                               char *buf)
3915 {
3916         return scnprintf(buf, PAGE_SIZE, "%s\n", DRV_VERSION);
3917 }
3918
3919 TF_WWN_ATTR_RO(srpt, version);
3920
3921 static struct configfs_attribute *srpt_wwn_attrs[] = {
3922         &srpt_wwn_version.attr,
3923         NULL,
3924 };
3925
3926 static struct target_core_fabric_ops srpt_template = {
3927         .get_fabric_name                = srpt_get_fabric_name,
3928         .get_fabric_proto_ident         = srpt_get_fabric_proto_ident,
3929         .tpg_get_wwn                    = srpt_get_fabric_wwn,
3930         .tpg_get_tag                    = srpt_get_tag,
3931         .tpg_get_default_depth          = srpt_get_default_depth,
3932         .tpg_get_pr_transport_id        = srpt_get_pr_transport_id,
3933         .tpg_get_pr_transport_id_len    = srpt_get_pr_transport_id_len,
3934         .tpg_parse_pr_out_transport_id  = srpt_parse_pr_out_transport_id,
3935         .tpg_check_demo_mode            = srpt_check_false,
3936         .tpg_check_demo_mode_cache      = srpt_check_true,
3937         .tpg_check_demo_mode_write_protect = srpt_check_true,
3938         .tpg_check_prod_mode_write_protect = srpt_check_false,
3939         .tpg_alloc_fabric_acl           = srpt_alloc_fabric_acl,
3940         .tpg_release_fabric_acl         = srpt_release_fabric_acl,
3941         .tpg_get_inst_index             = srpt_tpg_get_inst_index,
3942         .release_cmd                    = srpt_release_cmd,
3943         .check_stop_free                = srpt_check_stop_free,
3944         .shutdown_session               = srpt_shutdown_session,
3945         .close_session                  = srpt_close_session,
3946         .sess_get_index                 = srpt_sess_get_index,
3947         .sess_get_initiator_sid         = NULL,
3948         .write_pending                  = srpt_write_pending,
3949         .write_pending_status           = srpt_write_pending_status,
3950         .set_default_node_attributes    = srpt_set_default_node_attrs,
3951         .get_task_tag                   = srpt_get_task_tag,
3952         .get_cmd_state                  = srpt_get_tcm_cmd_state,
3953         .queue_data_in                  = srpt_queue_response,
3954         .queue_status                   = srpt_queue_status,
3955         .queue_tm_rsp                   = srpt_queue_response,
3956         .get_fabric_sense_len           = srpt_get_fabric_sense_len,
3957         .set_fabric_sense_len           = srpt_set_fabric_sense_len,
3958         /*
3959          * Setup function pointers for generic logic in
3960          * target_core_fabric_configfs.c
3961          */
3962         .fabric_make_wwn                = srpt_make_tport,
3963         .fabric_drop_wwn                = srpt_drop_tport,
3964         .fabric_make_tpg                = srpt_make_tpg,
3965         .fabric_drop_tpg                = srpt_drop_tpg,
3966         .fabric_post_link               = NULL,
3967         .fabric_pre_unlink              = NULL,
3968         .fabric_make_np                 = NULL,
3969         .fabric_drop_np                 = NULL,
3970         .fabric_make_nodeacl            = srpt_make_nodeacl,
3971         .fabric_drop_nodeacl            = srpt_drop_nodeacl,
3972 };
3973
3974 /**
3975  * srpt_init_module() - Kernel module initialization.
3976  *
3977  * Note: Since ib_register_client() registers callback functions, and since at
3978  * least one of these callback functions (srpt_add_one()) calls target core
3979  * functions, this driver must be registered with the target core before
3980  * ib_register_client() is called.
3981  */
3982 static int __init srpt_init_module(void)
3983 {
3984         int ret;
3985
3986         ret = -EINVAL;
3987         if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3988                 printk(KERN_ERR "invalid value %d for kernel module parameter"
3989                        " srp_max_req_size -- must be at least %d.\n",
3990                        srp_max_req_size, MIN_MAX_REQ_SIZE);
3991                 goto out;
3992         }
3993
3994         if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3995             || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3996                 printk(KERN_ERR "invalid value %d for kernel module parameter"
3997                        " srpt_srq_size -- must be in the range [%d..%d].\n",
3998                        srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3999                 goto out;
4000         }
4001
4002         srpt_target = target_fabric_configfs_init(THIS_MODULE, "srpt");
4003         if (IS_ERR(srpt_target)) {
4004                 printk(KERN_ERR "couldn't register\n");
4005                 ret = PTR_ERR(srpt_target);
4006                 goto out;
4007         }
4008
4009         srpt_target->tf_ops = srpt_template;
4010
4011         /*
4012          * Set up default attribute lists.
4013          */
4014         srpt_target->tf_cit_tmpl.tfc_wwn_cit.ct_attrs = srpt_wwn_attrs;
4015         srpt_target->tf_cit_tmpl.tfc_tpg_base_cit.ct_attrs = srpt_tpg_attrs;
4016         srpt_target->tf_cit_tmpl.tfc_tpg_attrib_cit.ct_attrs = srpt_tpg_attrib_attrs;
4017         srpt_target->tf_cit_tmpl.tfc_tpg_param_cit.ct_attrs = NULL;
4018         srpt_target->tf_cit_tmpl.tfc_tpg_np_base_cit.ct_attrs = NULL;
4019         srpt_target->tf_cit_tmpl.tfc_tpg_nacl_base_cit.ct_attrs = NULL;
4020         srpt_target->tf_cit_tmpl.tfc_tpg_nacl_attrib_cit.ct_attrs = NULL;
4021         srpt_target->tf_cit_tmpl.tfc_tpg_nacl_auth_cit.ct_attrs = NULL;
4022         srpt_target->tf_cit_tmpl.tfc_tpg_nacl_param_cit.ct_attrs = NULL;
4023
4024         ret = target_fabric_configfs_register(srpt_target);
4025         if (ret < 0) {
4026                 printk(KERN_ERR "couldn't register\n");
4027                 goto out_free_target;
4028         }
4029
4030         ret = ib_register_client(&srpt_client);
4031         if (ret) {
4032                 printk(KERN_ERR "couldn't register IB client\n");
4033                 goto out_unregister_target;
4034         }
4035
4036         return 0;
4037
4038 out_unregister_target:
4039         target_fabric_configfs_deregister(srpt_target);
4040         srpt_target = NULL;
4041 out_free_target:
4042         if (srpt_target)
4043                 target_fabric_configfs_free(srpt_target);
4044 out:
4045         return ret;
4046 }
4047
4048 static void __exit srpt_cleanup_module(void)
4049 {
4050         ib_unregister_client(&srpt_client);
4051         target_fabric_configfs_deregister(srpt_target);
4052         srpt_target = NULL;
4053 }
4054
4055 module_init(srpt_init_module);
4056 module_exit(srpt_cleanup_module);