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         spin_lock_init(&ioctx->spinlock);
1273         ioctx->state = SRPT_STATE_NEW;
1274         ioctx->n_rbuf = 0;
1275         ioctx->rbufs = NULL;
1276         ioctx->n_rdma = 0;
1277         ioctx->n_rdma_ius = 0;
1278         ioctx->rdma_ius = NULL;
1279         ioctx->mapped_sg_count = 0;
1280         init_completion(&ioctx->tx_done);
1281         ioctx->queue_status_only = false;
1282         /*
1283          * transport_init_se_cmd() does not initialize all fields, so do it
1284          * here.
1285          */
1286         memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1287         memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1288
1289         return ioctx;
1290 }
1291
1292 /**
1293  * srpt_abort_cmd() - Abort a SCSI command.
1294  * @ioctx:   I/O context associated with the SCSI command.
1295  * @context: Preferred execution context.
1296  */
1297 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1298 {
1299         enum srpt_command_state state;
1300         unsigned long flags;
1301
1302         BUG_ON(!ioctx);
1303
1304         /*
1305          * If the command is in a state where the target core is waiting for
1306          * the ib_srpt driver, change the state to the next state. Changing
1307          * the state of the command from SRPT_STATE_NEED_DATA to
1308          * SRPT_STATE_DATA_IN ensures that srpt_xmit_response() will call this
1309          * function a second time.
1310          */
1311
1312         spin_lock_irqsave(&ioctx->spinlock, flags);
1313         state = ioctx->state;
1314         switch (state) {
1315         case SRPT_STATE_NEED_DATA:
1316                 ioctx->state = SRPT_STATE_DATA_IN;
1317                 break;
1318         case SRPT_STATE_DATA_IN:
1319         case SRPT_STATE_CMD_RSP_SENT:
1320         case SRPT_STATE_MGMT_RSP_SENT:
1321                 ioctx->state = SRPT_STATE_DONE;
1322                 break;
1323         default:
1324                 break;
1325         }
1326         spin_unlock_irqrestore(&ioctx->spinlock, flags);
1327
1328         if (state == SRPT_STATE_DONE) {
1329                 struct srpt_rdma_ch *ch = ioctx->ch;
1330
1331                 BUG_ON(ch->sess == NULL);
1332
1333                 target_put_sess_cmd(ch->sess, &ioctx->cmd);
1334                 goto out;
1335         }
1336
1337         pr_debug("Aborting cmd with state %d and tag %lld\n", state,
1338                  ioctx->tag);
1339
1340         switch (state) {
1341         case SRPT_STATE_NEW:
1342         case SRPT_STATE_DATA_IN:
1343         case SRPT_STATE_MGMT:
1344                 /*
1345                  * Do nothing - defer abort processing until
1346                  * srpt_queue_response() is invoked.
1347                  */
1348                 WARN_ON(!transport_check_aborted_status(&ioctx->cmd, false));
1349                 break;
1350         case SRPT_STATE_NEED_DATA:
1351                 /* DMA_TO_DEVICE (write) - RDMA read error. */
1352
1353                 /* XXX(hch): this is a horrible layering violation.. */
1354                 spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1355                 ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1356                 ioctx->cmd.transport_state &= ~CMD_T_ACTIVE;
1357                 spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1358
1359                 complete(&ioctx->cmd.transport_lun_stop_comp);
1360                 break;
1361         case SRPT_STATE_CMD_RSP_SENT:
1362                 /*
1363                  * SRP_RSP sending failed or the SRP_RSP send completion has
1364                  * not been received in time.
1365                  */
1366                 srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1367                 spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1368                 ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1369                 spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1370                 target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1371                 break;
1372         case SRPT_STATE_MGMT_RSP_SENT:
1373                 srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1374                 target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1375                 break;
1376         default:
1377                 WARN(1, "Unexpected command state (%d)", state);
1378                 break;
1379         }
1380
1381 out:
1382         return state;
1383 }
1384
1385 /**
1386  * srpt_handle_send_err_comp() - Process an IB_WC_SEND error completion.
1387  */
1388 static void srpt_handle_send_err_comp(struct srpt_rdma_ch *ch, u64 wr_id)
1389 {
1390         struct srpt_send_ioctx *ioctx;
1391         enum srpt_command_state state;
1392         struct se_cmd *cmd;
1393         u32 index;
1394
1395         atomic_inc(&ch->sq_wr_avail);
1396
1397         index = idx_from_wr_id(wr_id);
1398         ioctx = ch->ioctx_ring[index];
1399         state = srpt_get_cmd_state(ioctx);
1400         cmd = &ioctx->cmd;
1401
1402         WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1403                 && state != SRPT_STATE_MGMT_RSP_SENT
1404                 && state != SRPT_STATE_NEED_DATA
1405                 && state != SRPT_STATE_DONE);
1406
1407         /* If SRP_RSP sending failed, undo the ch->req_lim change. */
1408         if (state == SRPT_STATE_CMD_RSP_SENT
1409             || state == SRPT_STATE_MGMT_RSP_SENT)
1410                 atomic_dec(&ch->req_lim);
1411
1412         srpt_abort_cmd(ioctx);
1413 }
1414
1415 /**
1416  * srpt_handle_send_comp() - Process an IB send completion notification.
1417  */
1418 static void srpt_handle_send_comp(struct srpt_rdma_ch *ch,
1419                                   struct srpt_send_ioctx *ioctx)
1420 {
1421         enum srpt_command_state state;
1422
1423         atomic_inc(&ch->sq_wr_avail);
1424
1425         state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1426
1427         if (WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1428                     && state != SRPT_STATE_MGMT_RSP_SENT
1429                     && state != SRPT_STATE_DONE))
1430                 pr_debug("state = %d\n", state);
1431
1432         if (state != SRPT_STATE_DONE) {
1433                 srpt_unmap_sg_to_ib_sge(ch, ioctx);
1434                 transport_generic_free_cmd(&ioctx->cmd, 0);
1435         } else {
1436                 printk(KERN_ERR "IB completion has been received too late for"
1437                        " wr_id = %u.\n", ioctx->ioctx.index);
1438         }
1439 }
1440
1441 /**
1442  * srpt_handle_rdma_comp() - Process an IB RDMA completion notification.
1443  *
1444  * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping
1445  * the data that has been transferred via IB RDMA had to be postponed until the
1446  * check_stop_free() callback.  None of this is necessary anymore and needs to
1447  * be cleaned up.
1448  */
1449 static void srpt_handle_rdma_comp(struct srpt_rdma_ch *ch,
1450                                   struct srpt_send_ioctx *ioctx,
1451                                   enum srpt_opcode opcode)
1452 {
1453         WARN_ON(ioctx->n_rdma <= 0);
1454         atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1455
1456         if (opcode == SRPT_RDMA_READ_LAST) {
1457                 if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1458                                                 SRPT_STATE_DATA_IN))
1459                         target_execute_cmd(&ioctx->cmd);
1460                 else
1461                         printk(KERN_ERR "%s[%d]: wrong state = %d\n", __func__,
1462                                __LINE__, srpt_get_cmd_state(ioctx));
1463         } else if (opcode == SRPT_RDMA_ABORT) {
1464                 ioctx->rdma_aborted = true;
1465         } else {
1466                 WARN(true, "unexpected opcode %d\n", opcode);
1467         }
1468 }
1469
1470 /**
1471  * srpt_handle_rdma_err_comp() - Process an IB RDMA error completion.
1472  */
1473 static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch,
1474                                       struct srpt_send_ioctx *ioctx,
1475                                       enum srpt_opcode opcode)
1476 {
1477         struct se_cmd *cmd;
1478         enum srpt_command_state state;
1479         unsigned long flags;
1480
1481         cmd = &ioctx->cmd;
1482         state = srpt_get_cmd_state(ioctx);
1483         switch (opcode) {
1484         case SRPT_RDMA_READ_LAST:
1485                 if (ioctx->n_rdma <= 0) {
1486                         printk(KERN_ERR "Received invalid RDMA read"
1487                                " error completion with idx %d\n",
1488                                ioctx->ioctx.index);
1489                         break;
1490                 }
1491                 atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1492                 if (state == SRPT_STATE_NEED_DATA)
1493                         srpt_abort_cmd(ioctx);
1494                 else
1495                         printk(KERN_ERR "%s[%d]: wrong state = %d\n",
1496                                __func__, __LINE__, state);
1497                 break;
1498         case SRPT_RDMA_WRITE_LAST:
1499                 spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1500                 ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1501                 spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1502                 break;
1503         default:
1504                 printk(KERN_ERR "%s[%d]: opcode = %u\n", __func__,
1505                        __LINE__, opcode);
1506                 break;
1507         }
1508 }
1509
1510 /**
1511  * srpt_build_cmd_rsp() - Build an SRP_RSP response.
1512  * @ch: RDMA channel through which the request has been received.
1513  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1514  *   be built in the buffer ioctx->buf points at and hence this function will
1515  *   overwrite the request data.
1516  * @tag: tag of the request for which this response is being generated.
1517  * @status: value for the STATUS field of the SRP_RSP information unit.
1518  *
1519  * Returns the size in bytes of the SRP_RSP response.
1520  *
1521  * An SRP_RSP response contains a SCSI status or service response. See also
1522  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1523  * response. See also SPC-2 for more information about sense data.
1524  */
1525 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1526                               struct srpt_send_ioctx *ioctx, u64 tag,
1527                               int status)
1528 {
1529         struct srp_rsp *srp_rsp;
1530         const u8 *sense_data;
1531         int sense_data_len, max_sense_len;
1532
1533         /*
1534          * The lowest bit of all SAM-3 status codes is zero (see also
1535          * paragraph 5.3 in SAM-3).
1536          */
1537         WARN_ON(status & 1);
1538
1539         srp_rsp = ioctx->ioctx.buf;
1540         BUG_ON(!srp_rsp);
1541
1542         sense_data = ioctx->sense_data;
1543         sense_data_len = ioctx->cmd.scsi_sense_length;
1544         WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1545
1546         memset(srp_rsp, 0, sizeof *srp_rsp);
1547         srp_rsp->opcode = SRP_RSP;
1548         srp_rsp->req_lim_delta =
1549                 __constant_cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1550         srp_rsp->tag = tag;
1551         srp_rsp->status = status;
1552
1553         if (sense_data_len) {
1554                 BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1555                 max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1556                 if (sense_data_len > max_sense_len) {
1557                         printk(KERN_WARNING "truncated sense data from %d to %d"
1558                                " bytes\n", sense_data_len, max_sense_len);
1559                         sense_data_len = max_sense_len;
1560                 }
1561
1562                 srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1563                 srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1564                 memcpy(srp_rsp + 1, sense_data, sense_data_len);
1565         }
1566
1567         return sizeof(*srp_rsp) + sense_data_len;
1568 }
1569
1570 /**
1571  * srpt_build_tskmgmt_rsp() - Build a task management response.
1572  * @ch:       RDMA channel through which the request has been received.
1573  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1574  * @rsp_code: RSP_CODE that will be stored in the response.
1575  * @tag:      Tag of the request for which this response is being generated.
1576  *
1577  * Returns the size in bytes of the SRP_RSP response.
1578  *
1579  * An SRP_RSP response contains a SCSI status or service response. See also
1580  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1581  * response.
1582  */
1583 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1584                                   struct srpt_send_ioctx *ioctx,
1585                                   u8 rsp_code, u64 tag)
1586 {
1587         struct srp_rsp *srp_rsp;
1588         int resp_data_len;
1589         int resp_len;
1590
1591         resp_data_len = (rsp_code == SRP_TSK_MGMT_SUCCESS) ? 0 : 4;
1592         resp_len = sizeof(*srp_rsp) + resp_data_len;
1593
1594         srp_rsp = ioctx->ioctx.buf;
1595         BUG_ON(!srp_rsp);
1596         memset(srp_rsp, 0, sizeof *srp_rsp);
1597
1598         srp_rsp->opcode = SRP_RSP;
1599         srp_rsp->req_lim_delta = __constant_cpu_to_be32(1
1600                                     + atomic_xchg(&ch->req_lim_delta, 0));
1601         srp_rsp->tag = tag;
1602
1603         if (rsp_code != SRP_TSK_MGMT_SUCCESS) {
1604                 srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1605                 srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1606                 srp_rsp->data[3] = rsp_code;
1607         }
1608
1609         return resp_len;
1610 }
1611
1612 #define NO_SUCH_LUN ((uint64_t)-1LL)
1613
1614 /*
1615  * SCSI LUN addressing method. See also SAM-2 and the section about
1616  * eight byte LUNs.
1617  */
1618 enum scsi_lun_addr_method {
1619         SCSI_LUN_ADDR_METHOD_PERIPHERAL   = 0,
1620         SCSI_LUN_ADDR_METHOD_FLAT         = 1,
1621         SCSI_LUN_ADDR_METHOD_LUN          = 2,
1622         SCSI_LUN_ADDR_METHOD_EXTENDED_LUN = 3,
1623 };
1624
1625 /*
1626  * srpt_unpack_lun() - Convert from network LUN to linear LUN.
1627  *
1628  * Convert an 2-byte, 4-byte, 6-byte or 8-byte LUN structure in network byte
1629  * order (big endian) to a linear LUN. Supports three LUN addressing methods:
1630  * peripheral, flat and logical unit. See also SAM-2, section 4.9.4 (page 40).
1631  */
1632 static uint64_t srpt_unpack_lun(const uint8_t *lun, int len)
1633 {
1634         uint64_t res = NO_SUCH_LUN;
1635         int addressing_method;
1636
1637         if (unlikely(len < 2)) {
1638                 printk(KERN_ERR "Illegal LUN length %d, expected 2 bytes or "
1639                        "more", len);
1640                 goto out;
1641         }
1642
1643         switch (len) {
1644         case 8:
1645                 if ((*((__be64 *)lun) &
1646                      __constant_cpu_to_be64(0x0000FFFFFFFFFFFFLL)) != 0)
1647                         goto out_err;
1648                 break;
1649         case 4:
1650                 if (*((__be16 *)&lun[2]) != 0)
1651                         goto out_err;
1652                 break;
1653         case 6:
1654                 if (*((__be32 *)&lun[2]) != 0)
1655                         goto out_err;
1656                 break;
1657         case 2:
1658                 break;
1659         default:
1660                 goto out_err;
1661         }
1662
1663         addressing_method = (*lun) >> 6; /* highest two bits of byte 0 */
1664         switch (addressing_method) {
1665         case SCSI_LUN_ADDR_METHOD_PERIPHERAL:
1666         case SCSI_LUN_ADDR_METHOD_FLAT:
1667         case SCSI_LUN_ADDR_METHOD_LUN:
1668                 res = *(lun + 1) | (((*lun) & 0x3f) << 8);
1669                 break;
1670
1671         case SCSI_LUN_ADDR_METHOD_EXTENDED_LUN:
1672         default:
1673                 printk(KERN_ERR "Unimplemented LUN addressing method %u",
1674                        addressing_method);
1675                 break;
1676         }
1677
1678 out:
1679         return res;
1680
1681 out_err:
1682         printk(KERN_ERR "Support for multi-level LUNs has not yet been"
1683                " implemented");
1684         goto out;
1685 }
1686
1687 static int srpt_check_stop_free(struct se_cmd *cmd)
1688 {
1689         struct srpt_send_ioctx *ioctx = container_of(cmd,
1690                                 struct srpt_send_ioctx, cmd);
1691
1692         return target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1693 }
1694
1695 /**
1696  * srpt_handle_cmd() - Process SRP_CMD.
1697  */
1698 static int srpt_handle_cmd(struct srpt_rdma_ch *ch,
1699                            struct srpt_recv_ioctx *recv_ioctx,
1700                            struct srpt_send_ioctx *send_ioctx)
1701 {
1702         struct se_cmd *cmd;
1703         struct srp_cmd *srp_cmd;
1704         uint64_t unpacked_lun;
1705         u64 data_len;
1706         enum dma_data_direction dir;
1707         sense_reason_t ret;
1708         int rc;
1709
1710         BUG_ON(!send_ioctx);
1711
1712         srp_cmd = recv_ioctx->ioctx.buf;
1713         cmd = &send_ioctx->cmd;
1714         send_ioctx->tag = srp_cmd->tag;
1715
1716         switch (srp_cmd->task_attr) {
1717         case SRP_CMD_SIMPLE_Q:
1718                 cmd->sam_task_attr = MSG_SIMPLE_TAG;
1719                 break;
1720         case SRP_CMD_ORDERED_Q:
1721         default:
1722                 cmd->sam_task_attr = MSG_ORDERED_TAG;
1723                 break;
1724         case SRP_CMD_HEAD_OF_Q:
1725                 cmd->sam_task_attr = MSG_HEAD_TAG;
1726                 break;
1727         case SRP_CMD_ACA:
1728                 cmd->sam_task_attr = MSG_ACA_TAG;
1729                 break;
1730         }
1731
1732         if (srpt_get_desc_tbl(send_ioctx, srp_cmd, &dir, &data_len)) {
1733                 printk(KERN_ERR "0x%llx: parsing SRP descriptor table failed.\n",
1734                        srp_cmd->tag);
1735                 ret = TCM_INVALID_CDB_FIELD;
1736                 goto send_sense;
1737         }
1738
1739         unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_cmd->lun,
1740                                        sizeof(srp_cmd->lun));
1741         rc = target_submit_cmd(cmd, ch->sess, srp_cmd->cdb,
1742                         &send_ioctx->sense_data[0], unpacked_lun, data_len,
1743                         MSG_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF);
1744         if (rc != 0) {
1745                 ret = TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1746                 goto send_sense;
1747         }
1748         return 0;
1749
1750 send_sense:
1751         transport_send_check_condition_and_sense(cmd, ret, 0);
1752         return -1;
1753 }
1754
1755 /**
1756  * srpt_rx_mgmt_fn_tag() - Process a task management function by tag.
1757  * @ch: RDMA channel of the task management request.
1758  * @fn: Task management function to perform.
1759  * @req_tag: Tag of the SRP task management request.
1760  * @mgmt_ioctx: I/O context of the task management request.
1761  *
1762  * Returns zero if the target core will process the task management
1763  * request asynchronously.
1764  *
1765  * Note: It is assumed that the initiator serializes tag-based task management
1766  * requests.
1767  */
1768 static int srpt_rx_mgmt_fn_tag(struct srpt_send_ioctx *ioctx, u64 tag)
1769 {
1770         struct srpt_device *sdev;
1771         struct srpt_rdma_ch *ch;
1772         struct srpt_send_ioctx *target;
1773         int ret, i;
1774
1775         ret = -EINVAL;
1776         ch = ioctx->ch;
1777         BUG_ON(!ch);
1778         BUG_ON(!ch->sport);
1779         sdev = ch->sport->sdev;
1780         BUG_ON(!sdev);
1781         spin_lock_irq(&sdev->spinlock);
1782         for (i = 0; i < ch->rq_size; ++i) {
1783                 target = ch->ioctx_ring[i];
1784                 if (target->cmd.se_lun == ioctx->cmd.se_lun &&
1785                     target->tag == tag &&
1786                     srpt_get_cmd_state(target) != SRPT_STATE_DONE) {
1787                         ret = 0;
1788                         /* now let the target core abort &target->cmd; */
1789                         break;
1790                 }
1791         }
1792         spin_unlock_irq(&sdev->spinlock);
1793         return ret;
1794 }
1795
1796 static int srp_tmr_to_tcm(int fn)
1797 {
1798         switch (fn) {
1799         case SRP_TSK_ABORT_TASK:
1800                 return TMR_ABORT_TASK;
1801         case SRP_TSK_ABORT_TASK_SET:
1802                 return TMR_ABORT_TASK_SET;
1803         case SRP_TSK_CLEAR_TASK_SET:
1804                 return TMR_CLEAR_TASK_SET;
1805         case SRP_TSK_LUN_RESET:
1806                 return TMR_LUN_RESET;
1807         case SRP_TSK_CLEAR_ACA:
1808                 return TMR_CLEAR_ACA;
1809         default:
1810                 return -1;
1811         }
1812 }
1813
1814 /**
1815  * srpt_handle_tsk_mgmt() - Process an SRP_TSK_MGMT information unit.
1816  *
1817  * Returns 0 if and only if the request will be processed by the target core.
1818  *
1819  * For more information about SRP_TSK_MGMT information units, see also section
1820  * 6.7 in the SRP r16a document.
1821  */
1822 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1823                                  struct srpt_recv_ioctx *recv_ioctx,
1824                                  struct srpt_send_ioctx *send_ioctx)
1825 {
1826         struct srp_tsk_mgmt *srp_tsk;
1827         struct se_cmd *cmd;
1828         struct se_session *sess = ch->sess;
1829         uint64_t unpacked_lun;
1830         uint32_t tag = 0;
1831         int tcm_tmr;
1832         int rc;
1833
1834         BUG_ON(!send_ioctx);
1835
1836         srp_tsk = recv_ioctx->ioctx.buf;
1837         cmd = &send_ioctx->cmd;
1838
1839         pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld"
1840                  " cm_id %p sess %p\n", srp_tsk->tsk_mgmt_func,
1841                  srp_tsk->task_tag, srp_tsk->tag, ch->cm_id, ch->sess);
1842
1843         srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1844         send_ioctx->tag = srp_tsk->tag;
1845         tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1846         if (tcm_tmr < 0) {
1847                 send_ioctx->cmd.se_tmr_req->response =
1848                         TMR_TASK_MGMT_FUNCTION_NOT_SUPPORTED;
1849                 goto fail;
1850         }
1851         unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_tsk->lun,
1852                                        sizeof(srp_tsk->lun));
1853
1854         if (srp_tsk->tsk_mgmt_func == SRP_TSK_ABORT_TASK) {
1855                 rc = srpt_rx_mgmt_fn_tag(send_ioctx, srp_tsk->task_tag);
1856                 if (rc < 0) {
1857                         send_ioctx->cmd.se_tmr_req->response =
1858                                         TMR_TASK_DOES_NOT_EXIST;
1859                         goto fail;
1860                 }
1861                 tag = srp_tsk->task_tag;
1862         }
1863         rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL, unpacked_lun,
1864                                 srp_tsk, tcm_tmr, GFP_KERNEL, tag,
1865                                 TARGET_SCF_ACK_KREF);
1866         if (rc != 0) {
1867                 send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1868                 goto fail;
1869         }
1870         return;
1871 fail:
1872         transport_send_check_condition_and_sense(cmd, 0, 0); // XXX:
1873 }
1874
1875 /**
1876  * srpt_handle_new_iu() - Process a newly received information unit.
1877  * @ch:    RDMA channel through which the information unit has been received.
1878  * @ioctx: SRPT I/O context associated with the information unit.
1879  */
1880 static void srpt_handle_new_iu(struct srpt_rdma_ch *ch,
1881                                struct srpt_recv_ioctx *recv_ioctx,
1882                                struct srpt_send_ioctx *send_ioctx)
1883 {
1884         struct srp_cmd *srp_cmd;
1885         enum rdma_ch_state ch_state;
1886
1887         BUG_ON(!ch);
1888         BUG_ON(!recv_ioctx);
1889
1890         ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1891                                    recv_ioctx->ioctx.dma, srp_max_req_size,
1892                                    DMA_FROM_DEVICE);
1893
1894         ch_state = srpt_get_ch_state(ch);
1895         if (unlikely(ch_state == CH_CONNECTING)) {
1896                 list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1897                 goto out;
1898         }
1899
1900         if (unlikely(ch_state != CH_LIVE))
1901                 goto out;
1902
1903         srp_cmd = recv_ioctx->ioctx.buf;
1904         if (srp_cmd->opcode == SRP_CMD || srp_cmd->opcode == SRP_TSK_MGMT) {
1905                 if (!send_ioctx)
1906                         send_ioctx = srpt_get_send_ioctx(ch);
1907                 if (unlikely(!send_ioctx)) {
1908                         list_add_tail(&recv_ioctx->wait_list,
1909                                       &ch->cmd_wait_list);
1910                         goto out;
1911                 }
1912         }
1913
1914         switch (srp_cmd->opcode) {
1915         case SRP_CMD:
1916                 srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1917                 break;
1918         case SRP_TSK_MGMT:
1919                 srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1920                 break;
1921         case SRP_I_LOGOUT:
1922                 printk(KERN_ERR "Not yet implemented: SRP_I_LOGOUT\n");
1923                 break;
1924         case SRP_CRED_RSP:
1925                 pr_debug("received SRP_CRED_RSP\n");
1926                 break;
1927         case SRP_AER_RSP:
1928                 pr_debug("received SRP_AER_RSP\n");
1929                 break;
1930         case SRP_RSP:
1931                 printk(KERN_ERR "Received SRP_RSP\n");
1932                 break;
1933         default:
1934                 printk(KERN_ERR "received IU with unknown opcode 0x%x\n",
1935                        srp_cmd->opcode);
1936                 break;
1937         }
1938
1939         srpt_post_recv(ch->sport->sdev, recv_ioctx);
1940 out:
1941         return;
1942 }
1943
1944 static void srpt_process_rcv_completion(struct ib_cq *cq,
1945                                         struct srpt_rdma_ch *ch,
1946                                         struct ib_wc *wc)
1947 {
1948         struct srpt_device *sdev = ch->sport->sdev;
1949         struct srpt_recv_ioctx *ioctx;
1950         u32 index;
1951
1952         index = idx_from_wr_id(wc->wr_id);
1953         if (wc->status == IB_WC_SUCCESS) {
1954                 int req_lim;
1955
1956                 req_lim = atomic_dec_return(&ch->req_lim);
1957                 if (unlikely(req_lim < 0))
1958                         printk(KERN_ERR "req_lim = %d < 0\n", req_lim);
1959                 ioctx = sdev->ioctx_ring[index];
1960                 srpt_handle_new_iu(ch, ioctx, NULL);
1961         } else {
1962                 printk(KERN_INFO "receiving failed for idx %u with status %d\n",
1963                        index, wc->status);
1964         }
1965 }
1966
1967 /**
1968  * srpt_process_send_completion() - Process an IB send completion.
1969  *
1970  * Note: Although this has not yet been observed during tests, at least in
1971  * theory it is possible that the srpt_get_send_ioctx() call invoked by
1972  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
1973  * value in each response is set to one, and it is possible that this response
1974  * makes the initiator send a new request before the send completion for that
1975  * response has been processed. This could e.g. happen if the call to
1976  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
1977  * if IB retransmission causes generation of the send completion to be
1978  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
1979  * are queued on cmd_wait_list. The code below processes these delayed
1980  * requests one at a time.
1981  */
1982 static void srpt_process_send_completion(struct ib_cq *cq,
1983                                          struct srpt_rdma_ch *ch,
1984                                          struct ib_wc *wc)
1985 {
1986         struct srpt_send_ioctx *send_ioctx;
1987         uint32_t index;
1988         enum srpt_opcode opcode;
1989
1990         index = idx_from_wr_id(wc->wr_id);
1991         opcode = opcode_from_wr_id(wc->wr_id);
1992         send_ioctx = ch->ioctx_ring[index];
1993         if (wc->status == IB_WC_SUCCESS) {
1994                 if (opcode == SRPT_SEND)
1995                         srpt_handle_send_comp(ch, send_ioctx);
1996                 else {
1997                         WARN_ON(opcode != SRPT_RDMA_ABORT &&
1998                                 wc->opcode != IB_WC_RDMA_READ);
1999                         srpt_handle_rdma_comp(ch, send_ioctx, opcode);
2000                 }
2001         } else {
2002                 if (opcode == SRPT_SEND) {
2003                         printk(KERN_INFO "sending response for idx %u failed"
2004                                " with status %d\n", index, wc->status);
2005                         srpt_handle_send_err_comp(ch, wc->wr_id);
2006                 } else if (opcode != SRPT_RDMA_MID) {
2007                         printk(KERN_INFO "RDMA t %d for idx %u failed with"
2008                                 " status %d", opcode, index, wc->status);
2009                         srpt_handle_rdma_err_comp(ch, send_ioctx, opcode);
2010                 }
2011         }
2012
2013         while (unlikely(opcode == SRPT_SEND
2014                         && !list_empty(&ch->cmd_wait_list)
2015                         && srpt_get_ch_state(ch) == CH_LIVE
2016                         && (send_ioctx = srpt_get_send_ioctx(ch)) != NULL)) {
2017                 struct srpt_recv_ioctx *recv_ioctx;
2018
2019                 recv_ioctx = list_first_entry(&ch->cmd_wait_list,
2020                                               struct srpt_recv_ioctx,
2021                                               wait_list);
2022                 list_del(&recv_ioctx->wait_list);
2023                 srpt_handle_new_iu(ch, recv_ioctx, send_ioctx);
2024         }
2025 }
2026
2027 static void srpt_process_completion(struct ib_cq *cq, struct srpt_rdma_ch *ch)
2028 {
2029         struct ib_wc *const wc = ch->wc;
2030         int i, n;
2031
2032         WARN_ON(cq != ch->cq);
2033
2034         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
2035         while ((n = ib_poll_cq(cq, ARRAY_SIZE(ch->wc), wc)) > 0) {
2036                 for (i = 0; i < n; i++) {
2037                         if (opcode_from_wr_id(wc[i].wr_id) == SRPT_RECV)
2038                                 srpt_process_rcv_completion(cq, ch, &wc[i]);
2039                         else
2040                                 srpt_process_send_completion(cq, ch, &wc[i]);
2041                 }
2042         }
2043 }
2044
2045 /**
2046  * srpt_completion() - IB completion queue callback function.
2047  *
2048  * Notes:
2049  * - It is guaranteed that a completion handler will never be invoked
2050  *   concurrently on two different CPUs for the same completion queue. See also
2051  *   Documentation/infiniband/core_locking.txt and the implementation of
2052  *   handle_edge_irq() in kernel/irq/chip.c.
2053  * - When threaded IRQs are enabled, completion handlers are invoked in thread
2054  *   context instead of interrupt context.
2055  */
2056 static void srpt_completion(struct ib_cq *cq, void *ctx)
2057 {
2058         struct srpt_rdma_ch *ch = ctx;
2059
2060         wake_up_interruptible(&ch->wait_queue);
2061 }
2062
2063 static int srpt_compl_thread(void *arg)
2064 {
2065         struct srpt_rdma_ch *ch;
2066
2067         /* Hibernation / freezing of the SRPT kernel thread is not supported. */
2068         current->flags |= PF_NOFREEZE;
2069
2070         ch = arg;
2071         BUG_ON(!ch);
2072         printk(KERN_INFO "Session %s: kernel thread %s (PID %d) started\n",
2073                ch->sess_name, ch->thread->comm, current->pid);
2074         while (!kthread_should_stop()) {
2075                 wait_event_interruptible(ch->wait_queue,
2076                         (srpt_process_completion(ch->cq, ch),
2077                          kthread_should_stop()));
2078         }
2079         printk(KERN_INFO "Session %s: kernel thread %s (PID %d) stopped\n",
2080                ch->sess_name, ch->thread->comm, current->pid);
2081         return 0;
2082 }
2083
2084 /**
2085  * srpt_create_ch_ib() - Create receive and send completion queues.
2086  */
2087 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
2088 {
2089         struct ib_qp_init_attr *qp_init;
2090         struct srpt_port *sport = ch->sport;
2091         struct srpt_device *sdev = sport->sdev;
2092         u32 srp_sq_size = sport->port_attrib.srp_sq_size;
2093         int ret;
2094
2095         WARN_ON(ch->rq_size < 1);
2096
2097         ret = -ENOMEM;
2098         qp_init = kzalloc(sizeof *qp_init, GFP_KERNEL);
2099         if (!qp_init)
2100                 goto out;
2101
2102         ch->cq = ib_create_cq(sdev->device, srpt_completion, NULL, ch,
2103                               ch->rq_size + srp_sq_size, 0);
2104         if (IS_ERR(ch->cq)) {
2105                 ret = PTR_ERR(ch->cq);
2106                 printk(KERN_ERR "failed to create CQ cqe= %d ret= %d\n",
2107                        ch->rq_size + srp_sq_size, ret);
2108                 goto out;
2109         }
2110
2111         qp_init->qp_context = (void *)ch;
2112         qp_init->event_handler
2113                 = (void(*)(struct ib_event *, void*))srpt_qp_event;
2114         qp_init->send_cq = ch->cq;
2115         qp_init->recv_cq = ch->cq;
2116         qp_init->srq = sdev->srq;
2117         qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
2118         qp_init->qp_type = IB_QPT_RC;
2119         qp_init->cap.max_send_wr = srp_sq_size;
2120         qp_init->cap.max_send_sge = SRPT_DEF_SG_PER_WQE;
2121
2122         ch->qp = ib_create_qp(sdev->pd, qp_init);
2123         if (IS_ERR(ch->qp)) {
2124                 ret = PTR_ERR(ch->qp);
2125                 printk(KERN_ERR "failed to create_qp ret= %d\n", ret);
2126                 goto err_destroy_cq;
2127         }
2128
2129         atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
2130
2131         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
2132                  __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
2133                  qp_init->cap.max_send_wr, ch->cm_id);
2134
2135         ret = srpt_init_ch_qp(ch, ch->qp);
2136         if (ret)
2137                 goto err_destroy_qp;
2138
2139         init_waitqueue_head(&ch->wait_queue);
2140
2141         pr_debug("creating thread for session %s\n", ch->sess_name);
2142
2143         ch->thread = kthread_run(srpt_compl_thread, ch, "ib_srpt_compl");
2144         if (IS_ERR(ch->thread)) {
2145                 printk(KERN_ERR "failed to create kernel thread %ld\n",
2146                        PTR_ERR(ch->thread));
2147                 ch->thread = NULL;
2148                 goto err_destroy_qp;
2149         }
2150
2151 out:
2152         kfree(qp_init);
2153         return ret;
2154
2155 err_destroy_qp:
2156         ib_destroy_qp(ch->qp);
2157 err_destroy_cq:
2158         ib_destroy_cq(ch->cq);
2159         goto out;
2160 }
2161
2162 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
2163 {
2164         if (ch->thread)
2165                 kthread_stop(ch->thread);
2166
2167         ib_destroy_qp(ch->qp);
2168         ib_destroy_cq(ch->cq);
2169 }
2170
2171 /**
2172  * __srpt_close_ch() - Close an RDMA channel by setting the QP error state.
2173  *
2174  * Reset the QP and make sure all resources associated with the channel will
2175  * be deallocated at an appropriate time.
2176  *
2177  * Note: The caller must hold ch->sport->sdev->spinlock.
2178  */
2179 static void __srpt_close_ch(struct srpt_rdma_ch *ch)
2180 {
2181         struct srpt_device *sdev;
2182         enum rdma_ch_state prev_state;
2183         unsigned long flags;
2184
2185         sdev = ch->sport->sdev;
2186
2187         spin_lock_irqsave(&ch->spinlock, flags);
2188         prev_state = ch->state;
2189         switch (prev_state) {
2190         case CH_CONNECTING:
2191         case CH_LIVE:
2192                 ch->state = CH_DISCONNECTING;
2193                 break;
2194         default:
2195                 break;
2196         }
2197         spin_unlock_irqrestore(&ch->spinlock, flags);
2198
2199         switch (prev_state) {
2200         case CH_CONNECTING:
2201                 ib_send_cm_rej(ch->cm_id, IB_CM_REJ_NO_RESOURCES, NULL, 0,
2202                                NULL, 0);
2203                 /* fall through */
2204         case CH_LIVE:
2205                 if (ib_send_cm_dreq(ch->cm_id, NULL, 0) < 0)
2206                         printk(KERN_ERR "sending CM DREQ failed.\n");
2207                 break;
2208         case CH_DISCONNECTING:
2209                 break;
2210         case CH_DRAINING:
2211         case CH_RELEASING:
2212                 break;
2213         }
2214 }
2215
2216 /**
2217  * srpt_close_ch() - Close an RDMA channel.
2218  */
2219 static void srpt_close_ch(struct srpt_rdma_ch *ch)
2220 {
2221         struct srpt_device *sdev;
2222
2223         sdev = ch->sport->sdev;
2224         spin_lock_irq(&sdev->spinlock);
2225         __srpt_close_ch(ch);
2226         spin_unlock_irq(&sdev->spinlock);
2227 }
2228
2229 /**
2230  * srpt_drain_channel() - Drain a channel by resetting the IB queue pair.
2231  * @cm_id: Pointer to the CM ID of the channel to be drained.
2232  *
2233  * Note: Must be called from inside srpt_cm_handler to avoid a race between
2234  * accessing sdev->spinlock and the call to kfree(sdev) in srpt_remove_one()
2235  * (the caller of srpt_cm_handler holds the cm_id spinlock; srpt_remove_one()
2236  * waits until all target sessions for the associated IB device have been
2237  * unregistered and target session registration involves a call to
2238  * ib_destroy_cm_id(), which locks the cm_id spinlock and hence waits until
2239  * this function has finished).
2240  */
2241 static void srpt_drain_channel(struct ib_cm_id *cm_id)
2242 {
2243         struct srpt_device *sdev;
2244         struct srpt_rdma_ch *ch;
2245         int ret;
2246         bool do_reset = false;
2247
2248         WARN_ON_ONCE(irqs_disabled());
2249
2250         sdev = cm_id->context;
2251         BUG_ON(!sdev);
2252         spin_lock_irq(&sdev->spinlock);
2253         list_for_each_entry(ch, &sdev->rch_list, list) {
2254                 if (ch->cm_id == cm_id) {
2255                         do_reset = srpt_test_and_set_ch_state(ch,
2256                                         CH_CONNECTING, CH_DRAINING) ||
2257                                    srpt_test_and_set_ch_state(ch,
2258                                         CH_LIVE, CH_DRAINING) ||
2259                                    srpt_test_and_set_ch_state(ch,
2260                                         CH_DISCONNECTING, CH_DRAINING);
2261                         break;
2262                 }
2263         }
2264         spin_unlock_irq(&sdev->spinlock);
2265
2266         if (do_reset) {
2267                 ret = srpt_ch_qp_err(ch);
2268                 if (ret < 0)
2269                         printk(KERN_ERR "Setting queue pair in error state"
2270                                " failed: %d\n", ret);
2271         }
2272 }
2273
2274 /**
2275  * srpt_find_channel() - Look up an RDMA channel.
2276  * @cm_id: Pointer to the CM ID of the channel to be looked up.
2277  *
2278  * Return NULL if no matching RDMA channel has been found.
2279  */
2280 static struct srpt_rdma_ch *srpt_find_channel(struct srpt_device *sdev,
2281                                               struct ib_cm_id *cm_id)
2282 {
2283         struct srpt_rdma_ch *ch;
2284         bool found;
2285
2286         WARN_ON_ONCE(irqs_disabled());
2287         BUG_ON(!sdev);
2288
2289         found = false;
2290         spin_lock_irq(&sdev->spinlock);
2291         list_for_each_entry(ch, &sdev->rch_list, list) {
2292                 if (ch->cm_id == cm_id) {
2293                         found = true;
2294                         break;
2295                 }
2296         }
2297         spin_unlock_irq(&sdev->spinlock);
2298
2299         return found ? ch : NULL;
2300 }
2301
2302 /**
2303  * srpt_release_channel() - Release channel resources.
2304  *
2305  * Schedules the actual release because:
2306  * - Calling the ib_destroy_cm_id() call from inside an IB CM callback would
2307  *   trigger a deadlock.
2308  * - It is not safe to call TCM transport_* functions from interrupt context.
2309  */
2310 static void srpt_release_channel(struct srpt_rdma_ch *ch)
2311 {
2312         schedule_work(&ch->release_work);
2313 }
2314
2315 static void srpt_release_channel_work(struct work_struct *w)
2316 {
2317         struct srpt_rdma_ch *ch;
2318         struct srpt_device *sdev;
2319         struct se_session *se_sess;
2320
2321         ch = container_of(w, struct srpt_rdma_ch, release_work);
2322         pr_debug("ch = %p; ch->sess = %p; release_done = %p\n", ch, ch->sess,
2323                  ch->release_done);
2324
2325         sdev = ch->sport->sdev;
2326         BUG_ON(!sdev);
2327
2328         se_sess = ch->sess;
2329         BUG_ON(!se_sess);
2330
2331         target_wait_for_sess_cmds(se_sess, 0);
2332
2333         transport_deregister_session_configfs(se_sess);
2334         transport_deregister_session(se_sess);
2335         ch->sess = NULL;
2336
2337         srpt_destroy_ch_ib(ch);
2338
2339         srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2340                              ch->sport->sdev, ch->rq_size,
2341                              ch->rsp_size, DMA_TO_DEVICE);
2342
2343         spin_lock_irq(&sdev->spinlock);
2344         list_del(&ch->list);
2345         spin_unlock_irq(&sdev->spinlock);
2346
2347         ib_destroy_cm_id(ch->cm_id);
2348
2349         if (ch->release_done)
2350                 complete(ch->release_done);
2351
2352         wake_up(&sdev->ch_releaseQ);
2353
2354         kfree(ch);
2355 }
2356
2357 static struct srpt_node_acl *__srpt_lookup_acl(struct srpt_port *sport,
2358                                                u8 i_port_id[16])
2359 {
2360         struct srpt_node_acl *nacl;
2361
2362         list_for_each_entry(nacl, &sport->port_acl_list, list)
2363                 if (memcmp(nacl->i_port_id, i_port_id,
2364                            sizeof(nacl->i_port_id)) == 0)
2365                         return nacl;
2366
2367         return NULL;
2368 }
2369
2370 static struct srpt_node_acl *srpt_lookup_acl(struct srpt_port *sport,
2371                                              u8 i_port_id[16])
2372 {
2373         struct srpt_node_acl *nacl;
2374
2375         spin_lock_irq(&sport->port_acl_lock);
2376         nacl = __srpt_lookup_acl(sport, i_port_id);
2377         spin_unlock_irq(&sport->port_acl_lock);
2378
2379         return nacl;
2380 }
2381
2382 /**
2383  * srpt_cm_req_recv() - Process the event IB_CM_REQ_RECEIVED.
2384  *
2385  * Ownership of the cm_id is transferred to the target session if this
2386  * functions returns zero. Otherwise the caller remains the owner of cm_id.
2387  */
2388 static int srpt_cm_req_recv(struct ib_cm_id *cm_id,
2389                             struct ib_cm_req_event_param *param,
2390                             void *private_data)
2391 {
2392         struct srpt_device *sdev = cm_id->context;
2393         struct srpt_port *sport = &sdev->port[param->port - 1];
2394         struct srp_login_req *req;
2395         struct srp_login_rsp *rsp;
2396         struct srp_login_rej *rej;
2397         struct ib_cm_rep_param *rep_param;
2398         struct srpt_rdma_ch *ch, *tmp_ch;
2399         struct srpt_node_acl *nacl;
2400         u32 it_iu_len;
2401         int i;
2402         int ret = 0;
2403
2404         WARN_ON_ONCE(irqs_disabled());
2405
2406         if (WARN_ON(!sdev || !private_data))
2407                 return -EINVAL;
2408
2409         req = (struct srp_login_req *)private_data;
2410
2411         it_iu_len = be32_to_cpu(req->req_it_iu_len);
2412
2413         printk(KERN_INFO "Received SRP_LOGIN_REQ with i_port_id 0x%llx:0x%llx,"
2414                " t_port_id 0x%llx:0x%llx and it_iu_len %d on port %d"
2415                " (guid=0x%llx:0x%llx)\n",
2416                be64_to_cpu(*(__be64 *)&req->initiator_port_id[0]),
2417                be64_to_cpu(*(__be64 *)&req->initiator_port_id[8]),
2418                be64_to_cpu(*(__be64 *)&req->target_port_id[0]),
2419                be64_to_cpu(*(__be64 *)&req->target_port_id[8]),
2420                it_iu_len,
2421                param->port,
2422                be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[0]),
2423                be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[8]));
2424
2425         rsp = kzalloc(sizeof *rsp, GFP_KERNEL);
2426         rej = kzalloc(sizeof *rej, GFP_KERNEL);
2427         rep_param = kzalloc(sizeof *rep_param, GFP_KERNEL);
2428
2429         if (!rsp || !rej || !rep_param) {
2430                 ret = -ENOMEM;
2431                 goto out;
2432         }
2433
2434         if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2435                 rej->reason = __constant_cpu_to_be32(
2436                                 SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2437                 ret = -EINVAL;
2438                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because its"
2439                        " length (%d bytes) is out of range (%d .. %d)\n",
2440                        it_iu_len, 64, srp_max_req_size);
2441                 goto reject;
2442         }
2443
2444         if (!sport->enabled) {
2445                 rej->reason = __constant_cpu_to_be32(
2446                              SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2447                 ret = -EINVAL;
2448                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because the target port"
2449                        " has not yet been enabled\n");
2450                 goto reject;
2451         }
2452
2453         if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2454                 rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_NO_CHAN;
2455
2456                 spin_lock_irq(&sdev->spinlock);
2457
2458                 list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list) {
2459                         if (!memcmp(ch->i_port_id, req->initiator_port_id, 16)
2460                             && !memcmp(ch->t_port_id, req->target_port_id, 16)
2461                             && param->port == ch->sport->port
2462                             && param->listen_id == ch->sport->sdev->cm_id
2463                             && ch->cm_id) {
2464                                 enum rdma_ch_state ch_state;
2465
2466                                 ch_state = srpt_get_ch_state(ch);
2467                                 if (ch_state != CH_CONNECTING
2468                                     && ch_state != CH_LIVE)
2469                                         continue;
2470
2471                                 /* found an existing channel */
2472                                 pr_debug("Found existing channel %s"
2473                                          " cm_id= %p state= %d\n",
2474                                          ch->sess_name, ch->cm_id, ch_state);
2475
2476                                 __srpt_close_ch(ch);
2477
2478                                 rsp->rsp_flags =
2479                                         SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2480                         }
2481                 }
2482
2483                 spin_unlock_irq(&sdev->spinlock);
2484
2485         } else
2486                 rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2487
2488         if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2489             || *(__be64 *)(req->target_port_id + 8) !=
2490                cpu_to_be64(srpt_service_guid)) {
2491                 rej->reason = __constant_cpu_to_be32(
2492                                 SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2493                 ret = -ENOMEM;
2494                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because it"
2495                        " has an invalid target port identifier.\n");
2496                 goto reject;
2497         }
2498
2499         ch = kzalloc(sizeof *ch, GFP_KERNEL);
2500         if (!ch) {
2501                 rej->reason = __constant_cpu_to_be32(
2502                                         SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2503                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because no memory.\n");
2504                 ret = -ENOMEM;
2505                 goto reject;
2506         }
2507
2508         INIT_WORK(&ch->release_work, srpt_release_channel_work);
2509         memcpy(ch->i_port_id, req->initiator_port_id, 16);
2510         memcpy(ch->t_port_id, req->target_port_id, 16);
2511         ch->sport = &sdev->port[param->port - 1];
2512         ch->cm_id = cm_id;
2513         /*
2514          * Avoid QUEUE_FULL conditions by limiting the number of buffers used
2515          * for the SRP protocol to the command queue size.
2516          */
2517         ch->rq_size = SRPT_RQ_SIZE;
2518         spin_lock_init(&ch->spinlock);
2519         ch->state = CH_CONNECTING;
2520         INIT_LIST_HEAD(&ch->cmd_wait_list);
2521         ch->rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2522
2523         ch->ioctx_ring = (struct srpt_send_ioctx **)
2524                 srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2525                                       sizeof(*ch->ioctx_ring[0]),
2526                                       ch->rsp_size, DMA_TO_DEVICE);
2527         if (!ch->ioctx_ring)
2528                 goto free_ch;
2529
2530         INIT_LIST_HEAD(&ch->free_list);
2531         for (i = 0; i < ch->rq_size; i++) {
2532                 ch->ioctx_ring[i]->ch = ch;
2533                 list_add_tail(&ch->ioctx_ring[i]->free_list, &ch->free_list);
2534         }
2535
2536         ret = srpt_create_ch_ib(ch);
2537         if (ret) {
2538                 rej->reason = __constant_cpu_to_be32(
2539                                 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2540                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because creating"
2541                        " a new RDMA channel failed.\n");
2542                 goto free_ring;
2543         }
2544
2545         ret = srpt_ch_qp_rtr(ch, ch->qp);
2546         if (ret) {
2547                 rej->reason = __constant_cpu_to_be32(
2548                                 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2549                 printk(KERN_ERR "rejected SRP_LOGIN_REQ because enabling"
2550                        " RTR failed (error code = %d)\n", ret);
2551                 goto destroy_ib;
2552         }
2553         /*
2554          * Use the initator port identifier as the session name.
2555          */
2556         snprintf(ch->sess_name, sizeof(ch->sess_name), "0x%016llx%016llx",
2557                         be64_to_cpu(*(__be64 *)ch->i_port_id),
2558                         be64_to_cpu(*(__be64 *)(ch->i_port_id + 8)));
2559
2560         pr_debug("registering session %s\n", ch->sess_name);
2561
2562         nacl = srpt_lookup_acl(sport, ch->i_port_id);
2563         if (!nacl) {
2564                 printk(KERN_INFO "Rejected login because no ACL has been"
2565                        " configured yet for initiator %s.\n", ch->sess_name);
2566                 rej->reason = __constant_cpu_to_be32(
2567                                 SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2568                 goto destroy_ib;
2569         }
2570
2571         ch->sess = transport_init_session();
2572         if (IS_ERR(ch->sess)) {
2573                 rej->reason = __constant_cpu_to_be32(
2574                                 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2575                 pr_debug("Failed to create session\n");
2576                 goto deregister_session;
2577         }
2578         ch->sess->se_node_acl = &nacl->nacl;
2579         transport_register_session(&sport->port_tpg_1, &nacl->nacl, ch->sess, ch);
2580
2581         pr_debug("Establish connection sess=%p name=%s cm_id=%p\n", ch->sess,
2582                  ch->sess_name, ch->cm_id);
2583
2584         /* create srp_login_response */
2585         rsp->opcode = SRP_LOGIN_RSP;
2586         rsp->tag = req->tag;
2587         rsp->max_it_iu_len = req->req_it_iu_len;
2588         rsp->max_ti_iu_len = req->req_it_iu_len;
2589         ch->max_ti_iu_len = it_iu_len;
2590         rsp->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2591                                               | SRP_BUF_FORMAT_INDIRECT);
2592         rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2593         atomic_set(&ch->req_lim, ch->rq_size);
2594         atomic_set(&ch->req_lim_delta, 0);
2595
2596         /* create cm reply */
2597         rep_param->qp_num = ch->qp->qp_num;
2598         rep_param->private_data = (void *)rsp;
2599         rep_param->private_data_len = sizeof *rsp;
2600         rep_param->rnr_retry_count = 7;
2601         rep_param->flow_control = 1;
2602         rep_param->failover_accepted = 0;
2603         rep_param->srq = 1;
2604         rep_param->responder_resources = 4;
2605         rep_param->initiator_depth = 4;
2606
2607         ret = ib_send_cm_rep(cm_id, rep_param);
2608         if (ret) {
2609                 printk(KERN_ERR "sending SRP_LOGIN_REQ response failed"
2610                        " (error code = %d)\n", ret);
2611                 goto release_channel;
2612         }
2613
2614         spin_lock_irq(&sdev->spinlock);
2615         list_add_tail(&ch->list, &sdev->rch_list);
2616         spin_unlock_irq(&sdev->spinlock);
2617
2618         goto out;
2619
2620 release_channel:
2621         srpt_set_ch_state(ch, CH_RELEASING);
2622         transport_deregister_session_configfs(ch->sess);
2623
2624 deregister_session:
2625         transport_deregister_session(ch->sess);
2626         ch->sess = NULL;
2627
2628 destroy_ib:
2629         srpt_destroy_ch_ib(ch);
2630
2631 free_ring:
2632         srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2633                              ch->sport->sdev, ch->rq_size,
2634                              ch->rsp_size, DMA_TO_DEVICE);
2635 free_ch:
2636         kfree(ch);
2637
2638 reject:
2639         rej->opcode = SRP_LOGIN_REJ;
2640         rej->tag = req->tag;
2641         rej->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2642                                               | SRP_BUF_FORMAT_INDIRECT);
2643
2644         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2645                              (void *)rej, sizeof *rej);
2646
2647 out:
2648         kfree(rep_param);
2649         kfree(rsp);
2650         kfree(rej);
2651
2652         return ret;
2653 }
2654
2655 static void srpt_cm_rej_recv(struct ib_cm_id *cm_id)
2656 {
2657         printk(KERN_INFO "Received IB REJ for cm_id %p.\n", cm_id);
2658         srpt_drain_channel(cm_id);
2659 }
2660
2661 /**
2662  * srpt_cm_rtu_recv() - Process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event.
2663  *
2664  * An IB_CM_RTU_RECEIVED message indicates that the connection is established
2665  * and that the recipient may begin transmitting (RTU = ready to use).
2666  */
2667 static void srpt_cm_rtu_recv(struct ib_cm_id *cm_id)
2668 {
2669         struct srpt_rdma_ch *ch;
2670         int ret;
2671
2672         ch = srpt_find_channel(cm_id->context, cm_id);
2673         BUG_ON(!ch);
2674
2675         if (srpt_test_and_set_ch_state(ch, CH_CONNECTING, CH_LIVE)) {
2676                 struct srpt_recv_ioctx *ioctx, *ioctx_tmp;
2677
2678                 ret = srpt_ch_qp_rts(ch, ch->qp);
2679
2680                 list_for_each_entry_safe(ioctx, ioctx_tmp, &ch->cmd_wait_list,
2681                                          wait_list) {
2682                         list_del(&ioctx->wait_list);
2683                         srpt_handle_new_iu(ch, ioctx, NULL);
2684                 }
2685                 if (ret)
2686                         srpt_close_ch(ch);
2687         }
2688 }
2689
2690 static void srpt_cm_timewait_exit(struct ib_cm_id *cm_id)
2691 {
2692         printk(KERN_INFO "Received IB TimeWait exit for cm_id %p.\n", cm_id);
2693         srpt_drain_channel(cm_id);
2694 }
2695
2696 static void srpt_cm_rep_error(struct ib_cm_id *cm_id)
2697 {
2698         printk(KERN_INFO "Received IB REP error for cm_id %p.\n", cm_id);
2699         srpt_drain_channel(cm_id);
2700 }
2701
2702 /**
2703  * srpt_cm_dreq_recv() - Process reception of a DREQ message.
2704  */
2705 static void srpt_cm_dreq_recv(struct ib_cm_id *cm_id)
2706 {
2707         struct srpt_rdma_ch *ch;
2708         unsigned long flags;
2709         bool send_drep = false;
2710
2711         ch = srpt_find_channel(cm_id->context, cm_id);
2712         BUG_ON(!ch);
2713
2714         pr_debug("cm_id= %p ch->state= %d\n", cm_id, srpt_get_ch_state(ch));
2715
2716         spin_lock_irqsave(&ch->spinlock, flags);
2717         switch (ch->state) {
2718         case CH_CONNECTING:
2719         case CH_LIVE:
2720                 send_drep = true;
2721                 ch->state = CH_DISCONNECTING;
2722                 break;
2723         case CH_DISCONNECTING:
2724         case CH_DRAINING:
2725         case CH_RELEASING:
2726                 WARN(true, "unexpected channel state %d\n", ch->state);
2727                 break;
2728         }
2729         spin_unlock_irqrestore(&ch->spinlock, flags);
2730
2731         if (send_drep) {
2732                 if (ib_send_cm_drep(ch->cm_id, NULL, 0) < 0)
2733                         printk(KERN_ERR "Sending IB DREP failed.\n");
2734                 printk(KERN_INFO "Received DREQ and sent DREP for session %s.\n",
2735                        ch->sess_name);
2736         }
2737 }
2738
2739 /**
2740  * srpt_cm_drep_recv() - Process reception of a DREP message.
2741  */
2742 static void srpt_cm_drep_recv(struct ib_cm_id *cm_id)
2743 {
2744         printk(KERN_INFO "Received InfiniBand DREP message for cm_id %p.\n",
2745                cm_id);
2746         srpt_drain_channel(cm_id);
2747 }
2748
2749 /**
2750  * srpt_cm_handler() - IB connection manager callback function.
2751  *
2752  * A non-zero return value will cause the caller destroy the CM ID.
2753  *
2754  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2755  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2756  * a non-zero value in any other case will trigger a race with the
2757  * ib_destroy_cm_id() call in srpt_release_channel().
2758  */
2759 static int srpt_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2760 {
2761         int ret;
2762
2763         ret = 0;
2764         switch (event->event) {
2765         case IB_CM_REQ_RECEIVED:
2766                 ret = srpt_cm_req_recv(cm_id, &event->param.req_rcvd,
2767                                        event->private_data);
2768                 break;
2769         case IB_CM_REJ_RECEIVED:
2770                 srpt_cm_rej_recv(cm_id);
2771                 break;
2772         case IB_CM_RTU_RECEIVED:
2773         case IB_CM_USER_ESTABLISHED:
2774                 srpt_cm_rtu_recv(cm_id);
2775                 break;
2776         case IB_CM_DREQ_RECEIVED:
2777                 srpt_cm_dreq_recv(cm_id);
2778                 break;
2779         case IB_CM_DREP_RECEIVED:
2780                 srpt_cm_drep_recv(cm_id);
2781                 break;
2782         case IB_CM_TIMEWAIT_EXIT:
2783                 srpt_cm_timewait_exit(cm_id);
2784                 break;
2785         case IB_CM_REP_ERROR:
2786                 srpt_cm_rep_error(cm_id);
2787                 break;
2788         case IB_CM_DREQ_ERROR:
2789                 printk(KERN_INFO "Received IB DREQ ERROR event.\n");
2790                 break;
2791         case IB_CM_MRA_RECEIVED:
2792                 printk(KERN_INFO "Received IB MRA event\n");
2793                 break;
2794         default:
2795                 printk(KERN_ERR "received unrecognized IB CM event %d\n",
2796                        event->event);
2797                 break;
2798         }
2799
2800         return ret;
2801 }
2802
2803 /**
2804  * srpt_perform_rdmas() - Perform IB RDMA.
2805  *
2806  * Returns zero upon success or a negative number upon failure.
2807  */
2808 static int srpt_perform_rdmas(struct srpt_rdma_ch *ch,
2809                               struct srpt_send_ioctx *ioctx)
2810 {
2811         struct ib_send_wr wr;
2812         struct ib_send_wr *bad_wr;
2813         struct rdma_iu *riu;
2814         int i;
2815         int ret;
2816         int sq_wr_avail;
2817         enum dma_data_direction dir;
2818         const int n_rdma = ioctx->n_rdma;
2819
2820         dir = ioctx->cmd.data_direction;
2821         if (dir == DMA_TO_DEVICE) {
2822                 /* write */
2823                 ret = -ENOMEM;
2824                 sq_wr_avail = atomic_sub_return(n_rdma, &ch->sq_wr_avail);
2825                 if (sq_wr_avail < 0) {
2826                         printk(KERN_WARNING "IB send queue full (needed %d)\n",
2827                                n_rdma);
2828                         goto out;
2829                 }
2830         }
2831
2832         ioctx->rdma_aborted = false;
2833         ret = 0;
2834         riu = ioctx->rdma_ius;
2835         memset(&wr, 0, sizeof wr);
2836
2837         for (i = 0; i < n_rdma; ++i, ++riu) {
2838                 if (dir == DMA_FROM_DEVICE) {
2839                         wr.opcode = IB_WR_RDMA_WRITE;
2840                         wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2841                                                 SRPT_RDMA_WRITE_LAST :
2842                                                 SRPT_RDMA_MID,
2843                                                 ioctx->ioctx.index);
2844                 } else {
2845                         wr.opcode = IB_WR_RDMA_READ;
2846                         wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2847                                                 SRPT_RDMA_READ_LAST :
2848                                                 SRPT_RDMA_MID,
2849                                                 ioctx->ioctx.index);
2850                 }
2851                 wr.next = NULL;
2852                 wr.wr.rdma.remote_addr = riu->raddr;
2853                 wr.wr.rdma.rkey = riu->rkey;
2854                 wr.num_sge = riu->sge_cnt;
2855                 wr.sg_list = riu->sge;
2856
2857                 /* only get completion event for the last rdma write */
2858                 if (i == (n_rdma - 1) && dir == DMA_TO_DEVICE)
2859                         wr.send_flags = IB_SEND_SIGNALED;
2860
2861                 ret = ib_post_send(ch->qp, &wr, &bad_wr);
2862                 if (ret)
2863                         break;
2864         }
2865
2866         if (ret)
2867                 printk(KERN_ERR "%s[%d]: ib_post_send() returned %d for %d/%d",
2868                                  __func__, __LINE__, ret, i, n_rdma);
2869         if (ret && i > 0) {
2870                 wr.num_sge = 0;
2871                 wr.wr_id = encode_wr_id(SRPT_RDMA_ABORT, ioctx->ioctx.index);
2872                 wr.send_flags = IB_SEND_SIGNALED;
2873                 while (ch->state == CH_LIVE &&
2874                         ib_post_send(ch->qp, &wr, &bad_wr) != 0) {
2875                         printk(KERN_INFO "Trying to abort failed RDMA transfer [%d]",
2876                                 ioctx->ioctx.index);
2877                         msleep(1000);
2878                 }
2879                 while (ch->state != CH_RELEASING && !ioctx->rdma_aborted) {
2880                         printk(KERN_INFO "Waiting until RDMA abort finished [%d]",
2881                                 ioctx->ioctx.index);
2882                         msleep(1000);
2883                 }
2884         }
2885 out:
2886         if (unlikely(dir == DMA_TO_DEVICE && ret < 0))
2887                 atomic_add(n_rdma, &ch->sq_wr_avail);
2888         return ret;
2889 }
2890
2891 /**
2892  * srpt_xfer_data() - Start data transfer from initiator to target.
2893  */
2894 static int srpt_xfer_data(struct srpt_rdma_ch *ch,
2895                           struct srpt_send_ioctx *ioctx)
2896 {
2897         int ret;
2898
2899         ret = srpt_map_sg_to_ib_sge(ch, ioctx);
2900         if (ret) {
2901                 printk(KERN_ERR "%s[%d] ret=%d\n", __func__, __LINE__, ret);
2902                 goto out;
2903         }
2904
2905         ret = srpt_perform_rdmas(ch, ioctx);
2906         if (ret) {
2907                 if (ret == -EAGAIN || ret == -ENOMEM)
2908                         printk(KERN_INFO "%s[%d] queue full -- ret=%d\n",
2909                                    __func__, __LINE__, ret);
2910                 else
2911                         printk(KERN_ERR "%s[%d] fatal error -- ret=%d\n",
2912                                __func__, __LINE__, ret);
2913                 goto out_unmap;
2914         }
2915
2916 out:
2917         return ret;
2918 out_unmap:
2919         srpt_unmap_sg_to_ib_sge(ch, ioctx);
2920         goto out;
2921 }
2922
2923 static int srpt_write_pending_status(struct se_cmd *se_cmd)
2924 {
2925         struct srpt_send_ioctx *ioctx;
2926
2927         ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2928         return srpt_get_cmd_state(ioctx) == SRPT_STATE_NEED_DATA;
2929 }
2930
2931 /*
2932  * srpt_write_pending() - Start data transfer from initiator to target (write).
2933  */
2934 static int srpt_write_pending(struct se_cmd *se_cmd)
2935 {
2936         struct srpt_rdma_ch *ch;
2937         struct srpt_send_ioctx *ioctx;
2938         enum srpt_command_state new_state;
2939         enum rdma_ch_state ch_state;
2940         int ret;
2941
2942         ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2943
2944         new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2945         WARN_ON(new_state == SRPT_STATE_DONE);
2946
2947         ch = ioctx->ch;
2948         BUG_ON(!ch);
2949
2950         ch_state = srpt_get_ch_state(ch);
2951         switch (ch_state) {
2952         case CH_CONNECTING:
2953                 WARN(true, "unexpected channel state %d\n", ch_state);
2954                 ret = -EINVAL;
2955                 goto out;
2956         case CH_LIVE:
2957                 break;
2958         case CH_DISCONNECTING:
2959         case CH_DRAINING:
2960         case CH_RELEASING:
2961                 pr_debug("cmd with tag %lld: channel disconnecting\n",
2962                          ioctx->tag);
2963                 srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
2964                 ret = -EINVAL;
2965                 goto out;
2966         }
2967         ret = srpt_xfer_data(ch, ioctx);
2968
2969 out:
2970         return ret;
2971 }
2972
2973 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
2974 {
2975         switch (tcm_mgmt_status) {
2976         case TMR_FUNCTION_COMPLETE:
2977                 return SRP_TSK_MGMT_SUCCESS;
2978         case TMR_FUNCTION_REJECTED:
2979                 return SRP_TSK_MGMT_FUNC_NOT_SUPP;
2980         }
2981         return SRP_TSK_MGMT_FAILED;
2982 }
2983
2984 /**
2985  * srpt_queue_response() - Transmits the response to a SCSI command.
2986  *
2987  * Callback function called by the TCM core. Must not block since it can be
2988  * invoked on the context of the IB completion handler.
2989  */
2990 static int srpt_queue_response(struct se_cmd *cmd)
2991 {
2992         struct srpt_rdma_ch *ch;
2993         struct srpt_send_ioctx *ioctx;
2994         enum srpt_command_state state;
2995         unsigned long flags;
2996         int ret;
2997         enum dma_data_direction dir;
2998         int resp_len;
2999         u8 srp_tm_status;
3000
3001         ret = 0;
3002
3003         ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3004         ch = ioctx->ch;
3005         BUG_ON(!ch);
3006
3007         spin_lock_irqsave(&ioctx->spinlock, flags);
3008         state = ioctx->state;
3009         switch (state) {
3010         case SRPT_STATE_NEW:
3011         case SRPT_STATE_DATA_IN:
3012                 ioctx->state = SRPT_STATE_CMD_RSP_SENT;
3013                 break;
3014         case SRPT_STATE_MGMT:
3015                 ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
3016                 break;
3017         default:
3018                 WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
3019                         ch, ioctx->ioctx.index, ioctx->state);
3020                 break;
3021         }
3022         spin_unlock_irqrestore(&ioctx->spinlock, flags);
3023
3024         if (unlikely(transport_check_aborted_status(&ioctx->cmd, false)
3025                      || WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))) {
3026                 atomic_inc(&ch->req_lim_delta);
3027                 srpt_abort_cmd(ioctx);
3028                 goto out;
3029         }
3030
3031         dir = ioctx->cmd.data_direction;
3032
3033         /* For read commands, transfer the data to the initiator. */
3034         if (dir == DMA_FROM_DEVICE && ioctx->cmd.data_length &&
3035             !ioctx->queue_status_only) {
3036                 ret = srpt_xfer_data(ch, ioctx);
3037                 if (ret) {
3038                         printk(KERN_ERR "xfer_data failed for tag %llu\n",
3039                                ioctx->tag);
3040                         goto out;
3041                 }
3042         }
3043
3044         if (state != SRPT_STATE_MGMT)
3045                 resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->tag,
3046                                               cmd->scsi_status);
3047         else {
3048                 srp_tm_status
3049                         = tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
3050                 resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
3051                                                  ioctx->tag);
3052         }
3053         ret = srpt_post_send(ch, ioctx, resp_len);
3054         if (ret) {
3055                 printk(KERN_ERR "sending cmd response failed for tag %llu\n",
3056                        ioctx->tag);
3057                 srpt_unmap_sg_to_ib_sge(ch, ioctx);
3058                 srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
3059                 target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
3060         }
3061
3062 out:
3063         return ret;
3064 }
3065
3066 static int srpt_queue_status(struct se_cmd *cmd)
3067 {
3068         struct srpt_send_ioctx *ioctx;
3069
3070         ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3071         BUG_ON(ioctx->sense_data != cmd->sense_buffer);
3072         if (cmd->se_cmd_flags &
3073             (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
3074                 WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
3075         ioctx->queue_status_only = true;
3076         return srpt_queue_response(cmd);
3077 }
3078
3079 static void srpt_refresh_port_work(struct work_struct *work)
3080 {
3081         struct srpt_port *sport = container_of(work, struct srpt_port, work);
3082
3083         srpt_refresh_port(sport);
3084 }
3085
3086 static int srpt_ch_list_empty(struct srpt_device *sdev)
3087 {
3088         int res;
3089
3090         spin_lock_irq(&sdev->spinlock);
3091         res = list_empty(&sdev->rch_list);
3092         spin_unlock_irq(&sdev->spinlock);
3093
3094         return res;
3095 }
3096
3097 /**
3098  * srpt_release_sdev() - Free the channel resources associated with a target.
3099  */
3100 static int srpt_release_sdev(struct srpt_device *sdev)
3101 {
3102         struct srpt_rdma_ch *ch, *tmp_ch;
3103         int res;
3104
3105         WARN_ON_ONCE(irqs_disabled());
3106
3107         BUG_ON(!sdev);
3108
3109         spin_lock_irq(&sdev->spinlock);
3110         list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list)
3111                 __srpt_close_ch(ch);
3112         spin_unlock_irq(&sdev->spinlock);
3113
3114         res = wait_event_interruptible(sdev->ch_releaseQ,
3115                                        srpt_ch_list_empty(sdev));
3116         if (res)
3117                 printk(KERN_ERR "%s: interrupted.\n", __func__);
3118
3119         return 0;
3120 }
3121
3122 static struct srpt_port *__srpt_lookup_port(const char *name)
3123 {
3124         struct ib_device *dev;
3125         struct srpt_device *sdev;
3126         struct srpt_port *sport;
3127         int i;
3128
3129         list_for_each_entry(sdev, &srpt_dev_list, list) {
3130                 dev = sdev->device;
3131                 if (!dev)
3132                         continue;
3133
3134                 for (i = 0; i < dev->phys_port_cnt; i++) {
3135                         sport = &sdev->port[i];
3136
3137                         if (!strcmp(sport->port_guid, name))
3138                                 return sport;
3139                 }
3140         }
3141
3142         return NULL;
3143 }
3144
3145 static struct srpt_port *srpt_lookup_port(const char *name)
3146 {
3147         struct srpt_port *sport;
3148
3149         spin_lock(&srpt_dev_lock);
3150         sport = __srpt_lookup_port(name);
3151         spin_unlock(&srpt_dev_lock);
3152
3153         return sport;
3154 }
3155
3156 /**
3157  * srpt_add_one() - Infiniband device addition callback function.
3158  */
3159 static void srpt_add_one(struct ib_device *device)
3160 {
3161         struct srpt_device *sdev;
3162         struct srpt_port *sport;
3163         struct ib_srq_init_attr srq_attr;
3164         int i;
3165
3166         pr_debug("device = %p, device->dma_ops = %p\n", device,
3167                  device->dma_ops);
3168
3169         sdev = kzalloc(sizeof *sdev, GFP_KERNEL);
3170         if (!sdev)
3171                 goto err;
3172
3173         sdev->device = device;
3174         INIT_LIST_HEAD(&sdev->rch_list);
3175         init_waitqueue_head(&sdev->ch_releaseQ);
3176         spin_lock_init(&sdev->spinlock);
3177
3178         if (ib_query_device(device, &sdev->dev_attr))
3179                 goto free_dev;
3180
3181         sdev->pd = ib_alloc_pd(device);
3182         if (IS_ERR(sdev->pd))
3183                 goto free_dev;
3184
3185         sdev->mr = ib_get_dma_mr(sdev->pd, IB_ACCESS_LOCAL_WRITE);
3186         if (IS_ERR(sdev->mr))
3187                 goto err_pd;
3188
3189         sdev->srq_size = min(srpt_srq_size, sdev->dev_attr.max_srq_wr);
3190
3191         srq_attr.event_handler = srpt_srq_event;
3192         srq_attr.srq_context = (void *)sdev;
3193         srq_attr.attr.max_wr = sdev->srq_size;
3194         srq_attr.attr.max_sge = 1;
3195         srq_attr.attr.srq_limit = 0;
3196         srq_attr.srq_type = IB_SRQT_BASIC;
3197
3198         sdev->srq = ib_create_srq(sdev->pd, &srq_attr);
3199         if (IS_ERR(sdev->srq))
3200                 goto err_mr;
3201
3202         pr_debug("%s: create SRQ #wr= %d max_allow=%d dev= %s\n",
3203                  __func__, sdev->srq_size, sdev->dev_attr.max_srq_wr,
3204                  device->name);
3205
3206         if (!srpt_service_guid)
3207                 srpt_service_guid = be64_to_cpu(device->node_guid);
3208
3209         sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3210         if (IS_ERR(sdev->cm_id))
3211                 goto err_srq;
3212
3213         /* print out target login information */
3214         pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,"
3215                  "pkey=ffff,service_id=%016llx\n", srpt_service_guid,
3216                  srpt_service_guid, srpt_service_guid);
3217
3218         /*
3219          * We do not have a consistent service_id (ie. also id_ext of target_id)
3220          * to identify this target. We currently use the guid of the first HCA
3221          * in the system as service_id; therefore, the target_id will change
3222          * if this HCA is gone bad and replaced by different HCA
3223          */
3224         if (ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0, NULL))
3225                 goto err_cm;
3226
3227         INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3228                               srpt_event_handler);
3229         if (ib_register_event_handler(&sdev->event_handler))
3230                 goto err_cm;
3231
3232         sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3233                 srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3234                                       sizeof(*sdev->ioctx_ring[0]),
3235                                       srp_max_req_size, DMA_FROM_DEVICE);
3236         if (!sdev->ioctx_ring)
3237                 goto err_event;
3238
3239         for (i = 0; i < sdev->srq_size; ++i)
3240                 srpt_post_recv(sdev, sdev->ioctx_ring[i]);
3241
3242         WARN_ON(sdev->device->phys_port_cnt > ARRAY_SIZE(sdev->port));
3243
3244         for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3245                 sport = &sdev->port[i - 1];
3246                 sport->sdev = sdev;
3247                 sport->port = i;
3248                 sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3249                 sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3250                 sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3251                 INIT_WORK(&sport->work, srpt_refresh_port_work);
3252                 INIT_LIST_HEAD(&sport->port_acl_list);
3253                 spin_lock_init(&sport->port_acl_lock);
3254
3255                 if (srpt_refresh_port(sport)) {
3256                         printk(KERN_ERR "MAD registration failed for %s-%d.\n",
3257                                srpt_sdev_name(sdev), i);
3258                         goto err_ring;
3259                 }
3260                 snprintf(sport->port_guid, sizeof(sport->port_guid),
3261                         "0x%016llx%016llx",
3262                         be64_to_cpu(sport->gid.global.subnet_prefix),
3263                         be64_to_cpu(sport->gid.global.interface_id));
3264         }
3265
3266         spin_lock(&srpt_dev_lock);
3267         list_add_tail(&sdev->list, &srpt_dev_list);
3268         spin_unlock(&srpt_dev_lock);
3269
3270 out:
3271         ib_set_client_data(device, &srpt_client, sdev);
3272         pr_debug("added %s.\n", device->name);
3273         return;
3274
3275 err_ring:
3276         srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3277                              sdev->srq_size, srp_max_req_size,
3278                              DMA_FROM_DEVICE);
3279 err_event:
3280         ib_unregister_event_handler(&sdev->event_handler);
3281 err_cm:
3282         ib_destroy_cm_id(sdev->cm_id);
3283 err_srq:
3284         ib_destroy_srq(sdev->srq);
3285 err_mr:
3286         ib_dereg_mr(sdev->mr);
3287 err_pd:
3288         ib_dealloc_pd(sdev->pd);
3289 free_dev:
3290         kfree(sdev);
3291 err:
3292         sdev = NULL;
3293         printk(KERN_INFO "%s(%s) failed.\n", __func__, device->name);
3294         goto out;
3295 }
3296
3297 /**
3298  * srpt_remove_one() - InfiniBand device removal callback function.
3299  */
3300 static void srpt_remove_one(struct ib_device *device)
3301 {
3302         struct srpt_device *sdev;
3303         int i;
3304
3305         sdev = ib_get_client_data(device, &srpt_client);
3306         if (!sdev) {
3307                 printk(KERN_INFO "%s(%s): nothing to do.\n", __func__,
3308                        device->name);
3309                 return;
3310         }
3311
3312         srpt_unregister_mad_agent(sdev);
3313
3314         ib_unregister_event_handler(&sdev->event_handler);
3315
3316         /* Cancel any work queued by the just unregistered IB event handler. */
3317         for (i = 0; i < sdev->device->phys_port_cnt; i++)
3318                 cancel_work_sync(&sdev->port[i].work);
3319
3320         ib_destroy_cm_id(sdev->cm_id);
3321
3322         /*
3323          * Unregistering a target must happen after destroying sdev->cm_id
3324          * such that no new SRP_LOGIN_REQ information units can arrive while
3325          * destroying the target.
3326          */
3327         spin_lock(&srpt_dev_lock);
3328         list_del(&sdev->list);
3329         spin_unlock(&srpt_dev_lock);
3330         srpt_release_sdev(sdev);
3331
3332         ib_destroy_srq(sdev->srq);
3333         ib_dereg_mr(sdev->mr);
3334         ib_dealloc_pd(sdev->pd);
3335
3336         srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3337                              sdev->srq_size, srp_max_req_size, DMA_FROM_DEVICE);
3338         sdev->ioctx_ring = NULL;
3339         kfree(sdev);
3340 }
3341
3342 static struct ib_client srpt_client = {
3343         .name = DRV_NAME,
3344         .add = srpt_add_one,
3345         .remove = srpt_remove_one
3346 };
3347
3348 static int srpt_check_true(struct se_portal_group *se_tpg)
3349 {
3350         return 1;
3351 }
3352
3353 static int srpt_check_false(struct se_portal_group *se_tpg)
3354 {
3355         return 0;
3356 }
3357
3358 static char *srpt_get_fabric_name(void)
3359 {
3360         return "srpt";
3361 }
3362
3363 static u8 srpt_get_fabric_proto_ident(struct se_portal_group *se_tpg)
3364 {
3365         return SCSI_TRANSPORTID_PROTOCOLID_SRP;
3366 }
3367
3368 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3369 {
3370         struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3371
3372         return sport->port_guid;
3373 }
3374
3375 static u16 srpt_get_tag(struct se_portal_group *tpg)
3376 {
3377         return 1;
3378 }
3379
3380 static u32 srpt_get_default_depth(struct se_portal_group *se_tpg)
3381 {
3382         return 1;
3383 }
3384
3385 static u32 srpt_get_pr_transport_id(struct se_portal_group *se_tpg,
3386                                     struct se_node_acl *se_nacl,
3387                                     struct t10_pr_registration *pr_reg,
3388                                     int *format_code, unsigned char *buf)
3389 {
3390         struct srpt_node_acl *nacl;
3391         struct spc_rdma_transport_id *tr_id;
3392
3393         nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3394         tr_id = (void *)buf;
3395         tr_id->protocol_identifier = SCSI_TRANSPORTID_PROTOCOLID_SRP;
3396         memcpy(tr_id->i_port_id, nacl->i_port_id, sizeof(tr_id->i_port_id));
3397         return sizeof(*tr_id);
3398 }
3399
3400 static u32 srpt_get_pr_transport_id_len(struct se_portal_group *se_tpg,
3401                                         struct se_node_acl *se_nacl,
3402                                         struct t10_pr_registration *pr_reg,
3403                                         int *format_code)
3404 {
3405         *format_code = 0;
3406         return sizeof(struct spc_rdma_transport_id);
3407 }
3408
3409 static char *srpt_parse_pr_out_transport_id(struct se_portal_group *se_tpg,
3410                                             const char *buf, u32 *out_tid_len,
3411                                             char **port_nexus_ptr)
3412 {
3413         struct spc_rdma_transport_id *tr_id;
3414
3415         *port_nexus_ptr = NULL;
3416         *out_tid_len = sizeof(struct spc_rdma_transport_id);
3417         tr_id = (void *)buf;
3418         return (char *)tr_id->i_port_id;
3419 }
3420
3421 static struct se_node_acl *srpt_alloc_fabric_acl(struct se_portal_group *se_tpg)
3422 {
3423         struct srpt_node_acl *nacl;
3424
3425         nacl = kzalloc(sizeof(struct srpt_node_acl), GFP_KERNEL);
3426         if (!nacl) {
3427                 printk(KERN_ERR "Unable to allocate struct srpt_node_acl\n");
3428                 return NULL;
3429         }
3430
3431         return &nacl->nacl;
3432 }
3433
3434 static void srpt_release_fabric_acl(struct se_portal_group *se_tpg,
3435                                     struct se_node_acl *se_nacl)
3436 {
3437         struct srpt_node_acl *nacl;
3438
3439         nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3440         kfree(nacl);
3441 }
3442
3443 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3444 {
3445         return 1;
3446 }
3447
3448 static void srpt_release_cmd(struct se_cmd *se_cmd)
3449 {
3450         struct srpt_send_ioctx *ioctx = container_of(se_cmd,
3451                                 struct srpt_send_ioctx, cmd);
3452         struct srpt_rdma_ch *ch = ioctx->ch;
3453         unsigned long flags;
3454
3455         WARN_ON(ioctx->state != SRPT_STATE_DONE);
3456         WARN_ON(ioctx->mapped_sg_count != 0);
3457
3458         if (ioctx->n_rbuf > 1) {
3459                 kfree(ioctx->rbufs);
3460                 ioctx->rbufs = NULL;
3461                 ioctx->n_rbuf = 0;
3462         }
3463
3464         spin_lock_irqsave(&ch->spinlock, flags);
3465         list_add(&ioctx->free_list, &ch->free_list);
3466         spin_unlock_irqrestore(&ch->spinlock, flags);
3467 }
3468
3469 /**
3470  * srpt_shutdown_session() - Whether or not a session may be shut down.
3471  */
3472 static int srpt_shutdown_session(struct se_session *se_sess)
3473 {
3474         return true;
3475 }
3476
3477 /**
3478  * srpt_close_session() - Forcibly close a session.
3479  *
3480  * Callback function invoked by the TCM core to clean up sessions associated
3481  * with a node ACL when the user invokes
3482  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3483  */
3484 static void srpt_close_session(struct se_session *se_sess)
3485 {
3486         DECLARE_COMPLETION_ONSTACK(release_done);
3487         struct srpt_rdma_ch *ch;
3488         struct srpt_device *sdev;
3489         int res;
3490
3491         ch = se_sess->fabric_sess_ptr;
3492         WARN_ON(ch->sess != se_sess);
3493
3494         pr_debug("ch %p state %d\n", ch, srpt_get_ch_state(ch));
3495
3496         sdev = ch->sport->sdev;
3497         spin_lock_irq(&sdev->spinlock);
3498         BUG_ON(ch->release_done);
3499         ch->release_done = &release_done;
3500         __srpt_close_ch(ch);
3501         spin_unlock_irq(&sdev->spinlock);
3502
3503         res = wait_for_completion_timeout(&release_done, 60 * HZ);
3504         WARN_ON(res <= 0);
3505 }
3506
3507 /**
3508  * srpt_sess_get_index() - Return the value of scsiAttIntrPortIndex (SCSI-MIB).
3509  *
3510  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3511  * This object represents an arbitrary integer used to uniquely identify a
3512  * particular attached remote initiator port to a particular SCSI target port
3513  * within a particular SCSI target device within a particular SCSI instance.
3514  */
3515 static u32 srpt_sess_get_index(struct se_session *se_sess)
3516 {
3517         return 0;
3518 }
3519
3520 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3521 {
3522 }
3523
3524 static u32 srpt_get_task_tag(struct se_cmd *se_cmd)
3525 {
3526         struct srpt_send_ioctx *ioctx;
3527
3528         ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3529         return ioctx->tag;
3530 }
3531
3532 /* Note: only used from inside debug printk's by the TCM core. */
3533 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3534 {
3535         struct srpt_send_ioctx *ioctx;
3536
3537         ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3538         return srpt_get_cmd_state(ioctx);
3539 }
3540
3541 /**
3542  * srpt_parse_i_port_id() - Parse an initiator port ID.
3543  * @name: ASCII representation of a 128-bit initiator port ID.
3544  * @i_port_id: Binary 128-bit port ID.
3545  */
3546 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3547 {
3548         const char *p;
3549         unsigned len, count, leading_zero_bytes;
3550         int ret, rc;
3551
3552         p = name;
3553         if (strnicmp(p, "0x", 2) == 0)
3554                 p += 2;
3555         ret = -EINVAL;
3556         len = strlen(p);
3557         if (len % 2)
3558                 goto out;
3559         count = min(len / 2, 16U);
3560         leading_zero_bytes = 16 - count;
3561         memset(i_port_id, 0, leading_zero_bytes);
3562         rc = hex2bin(i_port_id + leading_zero_bytes, p, count);
3563         if (rc < 0)
3564                 pr_debug("hex2bin failed for srpt_parse_i_port_id: %d\n", rc);
3565         ret = 0;
3566 out:
3567         return ret;
3568 }
3569
3570 /*
3571  * configfs callback function invoked for
3572  * mkdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3573  */
3574 static struct se_node_acl *srpt_make_nodeacl(struct se_portal_group *tpg,
3575                                              struct config_group *group,
3576                                              const char *name)
3577 {
3578         struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3579         struct se_node_acl *se_nacl, *se_nacl_new;
3580         struct srpt_node_acl *nacl;
3581         int ret = 0;
3582         u32 nexus_depth = 1;
3583         u8 i_port_id[16];
3584
3585         if (srpt_parse_i_port_id(i_port_id, name) < 0) {
3586                 printk(KERN_ERR "invalid initiator port ID %s\n", name);
3587                 ret = -EINVAL;
3588                 goto err;
3589         }
3590
3591         se_nacl_new = srpt_alloc_fabric_acl(tpg);
3592         if (!se_nacl_new) {
3593                 ret = -ENOMEM;
3594                 goto err;
3595         }
3596         /*
3597          * nacl_new may be released by core_tpg_add_initiator_node_acl()
3598          * when converting a node ACL from demo mode to explict
3599          */
3600         se_nacl = core_tpg_add_initiator_node_acl(tpg, se_nacl_new, name,
3601                                                   nexus_depth);
3602         if (IS_ERR(se_nacl)) {
3603                 ret = PTR_ERR(se_nacl);
3604                 goto err;
3605         }
3606         /* Locate our struct srpt_node_acl and set sdev and i_port_id. */
3607         nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3608         memcpy(&nacl->i_port_id[0], &i_port_id[0], 16);
3609         nacl->sport = sport;
3610
3611         spin_lock_irq(&sport->port_acl_lock);
3612         list_add_tail(&nacl->list, &sport->port_acl_list);
3613         spin_unlock_irq(&sport->port_acl_lock);
3614
3615         return se_nacl;
3616 err:
3617         return ERR_PTR(ret);
3618 }
3619
3620 /*
3621  * configfs callback function invoked for
3622  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3623  */
3624 static void srpt_drop_nodeacl(struct se_node_acl *se_nacl)
3625 {
3626         struct srpt_node_acl *nacl;
3627         struct srpt_device *sdev;
3628         struct srpt_port *sport;
3629
3630         nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3631         sport = nacl->sport;
3632         sdev = sport->sdev;
3633         spin_lock_irq(&sport->port_acl_lock);
3634         list_del(&nacl->list);
3635         spin_unlock_irq(&sport->port_acl_lock);
3636         core_tpg_del_initiator_node_acl(&sport->port_tpg_1, se_nacl, 1);
3637         srpt_release_fabric_acl(NULL, se_nacl);
3638 }
3639
3640 static ssize_t srpt_tpg_attrib_show_srp_max_rdma_size(
3641         struct se_portal_group *se_tpg,
3642         char *page)
3643 {
3644         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3645
3646         return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3647 }
3648
3649 static ssize_t srpt_tpg_attrib_store_srp_max_rdma_size(
3650         struct se_portal_group *se_tpg,
3651         const char *page,
3652         size_t count)
3653 {
3654         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3655         unsigned long val;
3656         int ret;
3657
3658         ret = strict_strtoul(page, 0, &val);
3659         if (ret < 0) {
3660                 pr_err("strict_strtoul() failed with ret: %d\n", ret);
3661                 return -EINVAL;
3662         }
3663         if (val > MAX_SRPT_RDMA_SIZE) {
3664                 pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3665                         MAX_SRPT_RDMA_SIZE);
3666                 return -EINVAL;
3667         }
3668         if (val < DEFAULT_MAX_RDMA_SIZE) {
3669                 pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3670                         val, DEFAULT_MAX_RDMA_SIZE);
3671                 return -EINVAL;
3672         }
3673         sport->port_attrib.srp_max_rdma_size = val;
3674
3675         return count;
3676 }
3677
3678 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rdma_size, S_IRUGO | S_IWUSR);
3679
3680 static ssize_t srpt_tpg_attrib_show_srp_max_rsp_size(
3681         struct se_portal_group *se_tpg,
3682         char *page)
3683 {
3684         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3685
3686         return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3687 }
3688
3689 static ssize_t srpt_tpg_attrib_store_srp_max_rsp_size(
3690         struct se_portal_group *se_tpg,
3691         const char *page,
3692         size_t count)
3693 {
3694         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3695         unsigned long val;
3696         int ret;
3697
3698         ret = strict_strtoul(page, 0, &val);
3699         if (ret < 0) {
3700                 pr_err("strict_strtoul() failed with ret: %d\n", ret);
3701                 return -EINVAL;
3702         }
3703         if (val > MAX_SRPT_RSP_SIZE) {
3704                 pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3705                         MAX_SRPT_RSP_SIZE);
3706                 return -EINVAL;
3707         }
3708         if (val < MIN_MAX_RSP_SIZE) {
3709                 pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3710                         MIN_MAX_RSP_SIZE);
3711                 return -EINVAL;
3712         }
3713         sport->port_attrib.srp_max_rsp_size = val;
3714
3715         return count;
3716 }
3717
3718 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rsp_size, S_IRUGO | S_IWUSR);
3719
3720 static ssize_t srpt_tpg_attrib_show_srp_sq_size(
3721         struct se_portal_group *se_tpg,
3722         char *page)
3723 {
3724         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3725
3726         return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size);
3727 }
3728
3729 static ssize_t srpt_tpg_attrib_store_srp_sq_size(
3730         struct se_portal_group *se_tpg,
3731         const char *page,
3732         size_t count)
3733 {
3734         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3735         unsigned long val;
3736         int ret;
3737
3738         ret = strict_strtoul(page, 0, &val);
3739         if (ret < 0) {
3740                 pr_err("strict_strtoul() failed with ret: %d\n", ret);
3741                 return -EINVAL;
3742         }
3743         if (val > MAX_SRPT_SRQ_SIZE) {
3744                 pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3745                         MAX_SRPT_SRQ_SIZE);
3746                 return -EINVAL;
3747         }
3748         if (val < MIN_SRPT_SRQ_SIZE) {
3749                 pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3750                         MIN_SRPT_SRQ_SIZE);
3751                 return -EINVAL;
3752         }
3753         sport->port_attrib.srp_sq_size = val;
3754
3755         return count;
3756 }
3757
3758 TF_TPG_ATTRIB_ATTR(srpt, srp_sq_size, S_IRUGO | S_IWUSR);
3759
3760 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3761         &srpt_tpg_attrib_srp_max_rdma_size.attr,
3762         &srpt_tpg_attrib_srp_max_rsp_size.attr,
3763         &srpt_tpg_attrib_srp_sq_size.attr,
3764         NULL,
3765 };
3766
3767 static ssize_t srpt_tpg_show_enable(
3768         struct se_portal_group *se_tpg,
3769         char *page)
3770 {
3771         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3772
3773         return snprintf(page, PAGE_SIZE, "%d\n", (sport->enabled) ? 1: 0);
3774 }
3775
3776 static ssize_t srpt_tpg_store_enable(
3777         struct se_portal_group *se_tpg,
3778         const char *page,
3779         size_t count)
3780 {
3781         struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3782         unsigned long tmp;
3783         int ret;
3784
3785         ret = strict_strtoul(page, 0, &tmp);
3786         if (ret < 0) {
3787                 printk(KERN_ERR "Unable to extract srpt_tpg_store_enable\n");
3788                 return -EINVAL;
3789         }
3790
3791         if ((tmp != 0) && (tmp != 1)) {
3792                 printk(KERN_ERR "Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
3793                 return -EINVAL;
3794         }
3795         if (tmp == 1)
3796                 sport->enabled = true;
3797         else
3798                 sport->enabled = false;
3799
3800         return count;
3801 }
3802
3803 TF_TPG_BASE_ATTR(srpt, enable, S_IRUGO | S_IWUSR);
3804
3805 static struct configfs_attribute *srpt_tpg_attrs[] = {
3806         &srpt_tpg_enable.attr,
3807         NULL,
3808 };
3809
3810 /**
3811  * configfs callback invoked for
3812  * mkdir /sys/kernel/config/target/$driver/$port/$tpg
3813  */
3814 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3815                                              struct config_group *group,
3816                                              const char *name)
3817 {
3818         struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3819         int res;
3820
3821         /* Initialize sport->port_wwn and sport->port_tpg_1 */
3822         res = core_tpg_register(&srpt_target->tf_ops, &sport->port_wwn,
3823                         &sport->port_tpg_1, sport, TRANSPORT_TPG_TYPE_NORMAL);
3824         if (res)
3825                 return ERR_PTR(res);
3826
3827         return &sport->port_tpg_1;
3828 }
3829
3830 /**
3831  * configfs callback invoked for
3832  * rmdir /sys/kernel/config/target/$driver/$port/$tpg
3833  */
3834 static void srpt_drop_tpg(struct se_portal_group *tpg)
3835 {
3836         struct srpt_port *sport = container_of(tpg,
3837                                 struct srpt_port, port_tpg_1);
3838
3839         sport->enabled = false;
3840         core_tpg_deregister(&sport->port_tpg_1);
3841 }
3842
3843 /**
3844  * configfs callback invoked for
3845  * mkdir /sys/kernel/config/target/$driver/$port
3846  */
3847 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3848                                       struct config_group *group,
3849                                       const char *name)
3850 {
3851         struct srpt_port *sport;
3852         int ret;
3853
3854         sport = srpt_lookup_port(name);
3855         pr_debug("make_tport(%s)\n", name);
3856         ret = -EINVAL;
3857         if (!sport)
3858                 goto err;
3859
3860         return &sport->port_wwn;
3861
3862 err:
3863         return ERR_PTR(ret);
3864 }
3865
3866 /**
3867  * configfs callback invoked for
3868  * rmdir /sys/kernel/config/target/$driver/$port
3869  */
3870 static void srpt_drop_tport(struct se_wwn *wwn)
3871 {
3872         struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3873
3874         pr_debug("drop_tport(%s\n", config_item_name(&sport->port_wwn.wwn_group.cg_item));
3875 }
3876
3877 static ssize_t srpt_wwn_show_attr_version(struct target_fabric_configfs *tf,
3878                                               char *buf)
3879 {
3880         return scnprintf(buf, PAGE_SIZE, "%s\n", DRV_VERSION);
3881 }
3882
3883 TF_WWN_ATTR_RO(srpt, version);
3884
3885 static struct configfs_attribute *srpt_wwn_attrs[] = {
3886         &srpt_wwn_version.attr,
3887         NULL,
3888 };
3889
3890 static struct target_core_fabric_ops srpt_template = {
3891         .get_fabric_name                = srpt_get_fabric_name,
3892         .get_fabric_proto_ident         = srpt_get_fabric_proto_ident,
3893         .tpg_get_wwn                    = srpt_get_fabric_wwn,
3894         .tpg_get_tag                    = srpt_get_tag,
3895         .tpg_get_default_depth          = srpt_get_default_depth,
3896         .tpg_get_pr_transport_id        = srpt_get_pr_transport_id,
3897         .tpg_get_pr_transport_id_len    = srpt_get_pr_transport_id_len,
3898         .tpg_parse_pr_out_transport_id  = srpt_parse_pr_out_transport_id,
3899         .tpg_check_demo_mode            = srpt_check_false,
3900         .tpg_check_demo_mode_cache      = srpt_check_true,
3901         .tpg_check_demo_mode_write_protect = srpt_check_true,
3902         .tpg_check_prod_mode_write_protect = srpt_check_false,
3903         .tpg_alloc_fabric_acl           = srpt_alloc_fabric_acl,
3904         .tpg_release_fabric_acl         = srpt_release_fabric_acl,
3905         .tpg_get_inst_index             = srpt_tpg_get_inst_index,
3906         .release_cmd                    = srpt_release_cmd,
3907         .check_stop_free                = srpt_check_stop_free,
3908         .shutdown_session               = srpt_shutdown_session,
3909         .close_session                  = srpt_close_session,
3910         .sess_get_index                 = srpt_sess_get_index,
3911         .sess_get_initiator_sid         = NULL,
3912         .write_pending                  = srpt_write_pending,
3913         .write_pending_status           = srpt_write_pending_status,
3914         .set_default_node_attributes    = srpt_set_default_node_attrs,
3915         .get_task_tag                   = srpt_get_task_tag,
3916         .get_cmd_state                  = srpt_get_tcm_cmd_state,
3917         .queue_data_in                  = srpt_queue_response,
3918         .queue_status                   = srpt_queue_status,
3919         .queue_tm_rsp                   = srpt_queue_response,
3920         /*
3921          * Setup function pointers for generic logic in
3922          * target_core_fabric_configfs.c
3923          */
3924         .fabric_make_wwn                = srpt_make_tport,
3925         .fabric_drop_wwn                = srpt_drop_tport,
3926         .fabric_make_tpg                = srpt_make_tpg,
3927         .fabric_drop_tpg                = srpt_drop_tpg,
3928         .fabric_post_link               = NULL,
3929         .fabric_pre_unlink              = NULL,
3930         .fabric_make_np                 = NULL,
3931         .fabric_drop_np                 = NULL,
3932         .fabric_make_nodeacl            = srpt_make_nodeacl,
3933         .fabric_drop_nodeacl            = srpt_drop_nodeacl,
3934 };
3935
3936 /**
3937  * srpt_init_module() - Kernel module initialization.
3938  *
3939  * Note: Since ib_register_client() registers callback functions, and since at
3940  * least one of these callback functions (srpt_add_one()) calls target core
3941  * functions, this driver must be registered with the target core before
3942  * ib_register_client() is called.
3943  */
3944 static int __init srpt_init_module(void)
3945 {
3946         int ret;
3947
3948         ret = -EINVAL;
3949         if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3950                 printk(KERN_ERR "invalid value %d for kernel module parameter"
3951                        " srp_max_req_size -- must be at least %d.\n",
3952                        srp_max_req_size, MIN_MAX_REQ_SIZE);
3953                 goto out;
3954         }
3955
3956         if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3957             || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3958                 printk(KERN_ERR "invalid value %d for kernel module parameter"
3959                        " srpt_srq_size -- must be in the range [%d..%d].\n",
3960                        srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3961                 goto out;
3962         }
3963
3964         srpt_target = target_fabric_configfs_init(THIS_MODULE, "srpt");
3965         if (IS_ERR(srpt_target)) {
3966                 printk(KERN_ERR "couldn't register\n");
3967                 ret = PTR_ERR(srpt_target);
3968                 goto out;
3969         }
3970
3971         srpt_target->tf_ops = srpt_template;
3972
3973         /*
3974          * Set up default attribute lists.
3975          */
3976         srpt_target->tf_cit_tmpl.tfc_wwn_cit.ct_attrs = srpt_wwn_attrs;
3977         srpt_target->tf_cit_tmpl.tfc_tpg_base_cit.ct_attrs = srpt_tpg_attrs;
3978         srpt_target->tf_cit_tmpl.tfc_tpg_attrib_cit.ct_attrs = srpt_tpg_attrib_attrs;
3979         srpt_target->tf_cit_tmpl.tfc_tpg_param_cit.ct_attrs = NULL;
3980         srpt_target->tf_cit_tmpl.tfc_tpg_np_base_cit.ct_attrs = NULL;
3981         srpt_target->tf_cit_tmpl.tfc_tpg_nacl_base_cit.ct_attrs = NULL;
3982         srpt_target->tf_cit_tmpl.tfc_tpg_nacl_attrib_cit.ct_attrs = NULL;
3983         srpt_target->tf_cit_tmpl.tfc_tpg_nacl_auth_cit.ct_attrs = NULL;
3984         srpt_target->tf_cit_tmpl.tfc_tpg_nacl_param_cit.ct_attrs = NULL;
3985
3986         ret = target_fabric_configfs_register(srpt_target);
3987         if (ret < 0) {
3988                 printk(KERN_ERR "couldn't register\n");
3989                 goto out_free_target;
3990         }
3991
3992         ret = ib_register_client(&srpt_client);
3993         if (ret) {
3994                 printk(KERN_ERR "couldn't register IB client\n");
3995                 goto out_unregister_target;
3996         }
3997
3998         return 0;
3999
4000 out_unregister_target:
4001         target_fabric_configfs_deregister(srpt_target);
4002         srpt_target = NULL;
4003 out_free_target:
4004         if (srpt_target)
4005                 target_fabric_configfs_free(srpt_target);
4006 out:
4007         return ret;
4008 }
4009
4010 static void __exit srpt_cleanup_module(void)
4011 {
4012         ib_unregister_client(&srpt_client);
4013         target_fabric_configfs_deregister(srpt_target);
4014         srpt_target = NULL;
4015 }
4016
4017 module_init(srpt_init_module);
4018 module_exit(srpt_cleanup_module);