drivers/infiniband/ulp/srp/ib_srp.c

   1 /*
   2  * Copyright (c) 2005 Cisco Systems.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #define pr_fmt(fmt) PFX fmt
  34
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/slab.h>
  38 #include <linux/err.h>
  39 #include <linux/string.h>
  40 #include <linux/parser.h>
  41 #include <linux/random.h>
  42 #include <linux/jiffies.h>
  43
  44 #include <linux/atomic.h>
  45
  46 #include <scsi/scsi.h>
  47 #include <scsi/scsi_device.h>
  48 #include <scsi/scsi_dbg.h>
  49 #include <scsi/srp.h>
  50 #include <scsi/scsi_transport_srp.h>
  51
  52 #include "ib_srp.h"
  53
  54 #define DRV_NAME        "ib_srp"
  55 #define PFX             DRV_NAME ": "
  56 #define DRV_VERSION     "0.2"
  57 #define DRV_RELDATE     "November 1, 2005"
  58
  59 MODULE_AUTHOR("Roland Dreier");
  60 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
  61                    "v" DRV_VERSION " (" DRV_RELDATE ")");
  62 MODULE_LICENSE("Dual BSD/GPL");
  63
  64 static unsigned int srp_sg_tablesize;
  65 static unsigned int cmd_sg_entries;
  66 static unsigned int indirect_sg_entries;
  67 static bool allow_ext_sg;
  68 static int topspin_workarounds = 1;
  69
  70 module_param(srp_sg_tablesize, uint, 0444);
  71 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
  72
  73 module_param(cmd_sg_entries, uint, 0444);
  74 MODULE_PARM_DESC(cmd_sg_entries,
  75                  "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
  76
  77 module_param(indirect_sg_entries, uint, 0444);
  78 MODULE_PARM_DESC(indirect_sg_entries,
  79                  "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
  80
  81 module_param(allow_ext_sg, bool, 0444);
  82 MODULE_PARM_DESC(allow_ext_sg,
  83                   "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
  84
  85 module_param(topspin_workarounds, int, 0444);
  86 MODULE_PARM_DESC(topspin_workarounds,
  87                  "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
  88
  89 static void srp_add_one(struct ib_device *device);
  90 static void srp_remove_one(struct ib_device *device);
  91 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
  92 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
  93 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
  94
  95 static struct scsi_transport_template *ib_srp_transport_template;
  96
  97 static struct ib_client srp_client = {
  98         .name   = "srp",
  99         .add    = srp_add_one,
 100         .remove = srp_remove_one
 101 };
 102
 103 static struct ib_sa_client srp_sa_client;
 104
 105 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
 106 {
 107         return (struct srp_target_port *) host->hostdata;
 108 }
 109
 110 static const char *srp_target_info(struct Scsi_Host *host)
 111 {
 112         return host_to_target(host)->target_name;
 113 }
 114
 115 static int srp_target_is_topspin(struct srp_target_port *target)
 116 {
 117         static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
 118         static const u8 cisco_oui[3]   = { 0x00, 0x1b, 0x0d };
 119
 120         return topspin_workarounds &&
 121                 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
 122                  !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
 123 }
 124
 125 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
 126                                    gfp_t gfp_mask,
 127                                    enum dma_data_direction direction)
 128 {
 129         struct srp_iu *iu;
 130
 131         iu = kmalloc(sizeof *iu, gfp_mask);
 132         if (!iu)
 133                 goto out;
 134
 135         iu->buf = kzalloc(size, gfp_mask);
 136         if (!iu->buf)
 137                 goto out_free_iu;
 138
 139         iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
 140                                     direction);
 141         if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
 142                 goto out_free_buf;
 143
 144         iu->size      = size;
 145         iu->direction = direction;
 146
 147         return iu;
 148
 149 out_free_buf:
 150         kfree(iu->buf);
 151 out_free_iu:
 152         kfree(iu);
 153 out:
 154         return NULL;
 155 }
 156
 157 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
 158 {
 159         if (!iu)
 160                 return;
 161
 162         ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
 163                             iu->direction);
 164         kfree(iu->buf);
 165         kfree(iu);
 166 }
 167
 168 static void srp_qp_event(struct ib_event *event, void *context)
 169 {
 170         pr_debug("QP event %d\n", event->event);
 171 }
 172
 173 static int srp_init_qp(struct srp_target_port *target,
 174                        struct ib_qp *qp)
 175 {
 176         struct ib_qp_attr *attr;
 177         int ret;
 178
 179         attr = kmalloc(sizeof *attr, GFP_KERNEL);
 180         if (!attr)
 181                 return -ENOMEM;
 182
 183         ret = ib_find_pkey(target->srp_host->srp_dev->dev,
 184                            target->srp_host->port,
 185                            be16_to_cpu(target->path.pkey),
 186                            &attr->pkey_index);
 187         if (ret)
 188                 goto out;
 189
 190         attr->qp_state        = IB_QPS_INIT;
 191         attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
 192                                     IB_ACCESS_REMOTE_WRITE);
 193         attr->port_num        = target->srp_host->port;
 194
 195         ret = ib_modify_qp(qp, attr,
 196                            IB_QP_STATE          |
 197                            IB_QP_PKEY_INDEX     |
 198                            IB_QP_ACCESS_FLAGS   |
 199                            IB_QP_PORT);
 200
 201 out:
 202         kfree(attr);
 203         return ret;
 204 }
 205
 206 static int srp_new_cm_id(struct srp_target_port *target)
 207 {
 208         struct ib_cm_id *new_cm_id;
 209
 210         new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
 211                                     srp_cm_handler, target);
 212         if (IS_ERR(new_cm_id))
 213                 return PTR_ERR(new_cm_id);
 214
 215         if (target->cm_id)
 216                 ib_destroy_cm_id(target->cm_id);
 217         target->cm_id = new_cm_id;
 218
 219         return 0;
 220 }
 221
 222 static int srp_create_target_ib(struct srp_target_port *target)
 223 {
 224         struct ib_qp_init_attr *init_attr;
 225         struct ib_cq *recv_cq, *send_cq;
 226         struct ib_qp *qp;
 227         int ret;
 228
 229         init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
 230         if (!init_attr)
 231                 return -ENOMEM;
 232
 233         recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
 234                                srp_recv_completion, NULL, target, SRP_RQ_SIZE, 0);
 235         if (IS_ERR(recv_cq)) {
 236                 ret = PTR_ERR(recv_cq);
 237                 goto err;
 238         }
 239
 240         send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
 241                                srp_send_completion, NULL, target, SRP_SQ_SIZE, 0);
 242         if (IS_ERR(send_cq)) {
 243                 ret = PTR_ERR(send_cq);
 244                 goto err_recv_cq;
 245         }
 246
 247         ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
 248
 249         init_attr->event_handler       = srp_qp_event;
 250         init_attr->cap.max_send_wr     = SRP_SQ_SIZE;
 251         init_attr->cap.max_recv_wr     = SRP_RQ_SIZE;
 252         init_attr->cap.max_recv_sge    = 1;
 253         init_attr->cap.max_send_sge    = 1;
 254         init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
 255         init_attr->qp_type             = IB_QPT_RC;
 256         init_attr->send_cq             = send_cq;
 257         init_attr->recv_cq             = recv_cq;
 258
 259         qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
 260         if (IS_ERR(qp)) {
 261                 ret = PTR_ERR(qp);
 262                 goto err_send_cq;
 263         }
 264
 265         ret = srp_init_qp(target, qp);
 266         if (ret)
 267                 goto err_qp;
 268
 269         if (target->qp)
 270                 ib_destroy_qp(target->qp);
 271         if (target->recv_cq)
 272                 ib_destroy_cq(target->recv_cq);
 273         if (target->send_cq)
 274                 ib_destroy_cq(target->send_cq);
 275
 276         target->qp = qp;
 277         target->recv_cq = recv_cq;
 278         target->send_cq = send_cq;
 279
 280         kfree(init_attr);
 281         return 0;
 282
 283 err_qp:
 284         ib_destroy_qp(qp);
 285
 286 err_send_cq:
 287         ib_destroy_cq(send_cq);
 288
 289 err_recv_cq:
 290         ib_destroy_cq(recv_cq);
 291
 292 err:
 293         kfree(init_attr);
 294         return ret;
 295 }
 296
 297 static void srp_free_target_ib(struct srp_target_port *target)
 298 {
 299         int i;
 300
 301         ib_destroy_qp(target->qp);
 302         ib_destroy_cq(target->send_cq);
 303         ib_destroy_cq(target->recv_cq);
 304
 305         target->qp = NULL;
 306         target->send_cq = target->recv_cq = NULL;
 307
 308         for (i = 0; i < SRP_RQ_SIZE; ++i)
 309                 srp_free_iu(target->srp_host, target->rx_ring[i]);
 310         for (i = 0; i < SRP_SQ_SIZE; ++i)
 311                 srp_free_iu(target->srp_host, target->tx_ring[i]);
 312 }
 313
 314 static void srp_path_rec_completion(int status,
 315                                     struct ib_sa_path_rec *pathrec,
 316                                     void *target_ptr)
 317 {
 318         struct srp_target_port *target = target_ptr;
 319
 320         target->status = status;
 321         if (status)
 322                 shost_printk(KERN_ERR, target->scsi_host,
 323                              PFX "Got failed path rec status %d\n", status);
 324         else
 325                 target->path = *pathrec;
 326         complete(&target->done);
 327 }
 328
 329 static int srp_lookup_path(struct srp_target_port *target)
 330 {
 331         target->path.numb_path = 1;
 332
 333         init_completion(&target->done);
 334
 335         target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
 336                                                    target->srp_host->srp_dev->dev,
 337                                                    target->srp_host->port,
 338                                                    &target->path,
 339                                                    IB_SA_PATH_REC_SERVICE_ID    |
 340                                                    IB_SA_PATH_REC_DGID          |
 341                                                    IB_SA_PATH_REC_SGID          |
 342                                                    IB_SA_PATH_REC_NUMB_PATH     |
 343                                                    IB_SA_PATH_REC_PKEY,
 344                                                    SRP_PATH_REC_TIMEOUT_MS,
 345                                                    GFP_KERNEL,
 346                                                    srp_path_rec_completion,
 347                                                    target, &target->path_query);
 348         if (target->path_query_id < 0)
 349                 return target->path_query_id;
 350
 351         wait_for_completion(&target->done);
 352
 353         if (target->status < 0)
 354                 shost_printk(KERN_WARNING, target->scsi_host,
 355                              PFX "Path record query failed\n");
 356
 357         return target->status;
 358 }
 359
 360 static int srp_send_req(struct srp_target_port *target)
 361 {
 362         struct {
 363                 struct ib_cm_req_param param;
 364                 struct srp_login_req   priv;
 365         } *req = NULL;
 366         int status;
 367
 368         req = kzalloc(sizeof *req, GFP_KERNEL);
 369         if (!req)
 370                 return -ENOMEM;
 371
 372         req->param.primary_path               = &target->path;
 373         req->param.alternate_path             = NULL;
 374         req->param.service_id                 = target->service_id;
 375         req->param.qp_num                     = target->qp->qp_num;
 376         req->param.qp_type                    = target->qp->qp_type;
 377         req->param.private_data               = &req->priv;
 378         req->param.private_data_len           = sizeof req->priv;
 379         req->param.flow_control               = 1;
 380
 381         get_random_bytes(&req->param.starting_psn, 4);
 382         req->param.starting_psn              &= 0xffffff;
 383
 384         /*
 385          * Pick some arbitrary defaults here; we could make these
 386          * module parameters if anyone cared about setting them.
 387          */
 388         req->param.responder_resources        = 4;
 389         req->param.remote_cm_response_timeout = 20;
 390         req->param.local_cm_response_timeout  = 20;
 391         req->param.retry_count                = 7;
 392         req->param.rnr_retry_count            = 7;
 393         req->param.max_cm_retries             = 15;
 394
 395         req->priv.opcode        = SRP_LOGIN_REQ;
 396         req->priv.tag           = 0;
 397         req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
 398         req->priv.req_buf_fmt   = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
 399                                               SRP_BUF_FORMAT_INDIRECT);
 400         /*
 401          * In the published SRP specification (draft rev. 16a), the
 402          * port identifier format is 8 bytes of ID extension followed
 403          * by 8 bytes of GUID.  Older drafts put the two halves in the
 404          * opposite order, so that the GUID comes first.
 405          *
 406          * Targets conforming to these obsolete drafts can be
 407          * recognized by the I/O Class they report.
 408          */
 409         if (target->io_class == SRP_REV10_IB_IO_CLASS) {
 410                 memcpy(req->priv.initiator_port_id,
 411                        &target->path.sgid.global.interface_id, 8);
 412                 memcpy(req->priv.initiator_port_id + 8,
 413                        &target->initiator_ext, 8);
 414                 memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
 415                 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
 416         } else {
 417                 memcpy(req->priv.initiator_port_id,
 418                        &target->initiator_ext, 8);
 419                 memcpy(req->priv.initiator_port_id + 8,
 420                        &target->path.sgid.global.interface_id, 8);
 421                 memcpy(req->priv.target_port_id,     &target->id_ext, 8);
 422                 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
 423         }
 424
 425         /*
 426          * Topspin/Cisco SRP targets will reject our login unless we
 427          * zero out the first 8 bytes of our initiator port ID and set
 428          * the second 8 bytes to the local node GUID.
 429          */
 430         if (srp_target_is_topspin(target)) {
 431                 shost_printk(KERN_DEBUG, target->scsi_host,
 432                              PFX "Topspin/Cisco initiator port ID workaround "
 433                              "activated for target GUID %016llx\n",
 434                              (unsigned long long) be64_to_cpu(target->ioc_guid));
 435                 memset(req->priv.initiator_port_id, 0, 8);
 436                 memcpy(req->priv.initiator_port_id + 8,
 437                        &target->srp_host->srp_dev->dev->node_guid, 8);
 438         }
 439
 440         status = ib_send_cm_req(target->cm_id, &req->param);
 441
 442         kfree(req);
 443
 444         return status;
 445 }
 446
 447 static bool srp_queue_remove_work(struct srp_target_port *target)
 448 {
 449         bool changed = false;
 450
 451         spin_lock_irq(&target->lock);
 452         if (target->state != SRP_TARGET_REMOVED) {
 453                 target->state = SRP_TARGET_REMOVED;
 454                 changed = true;
 455         }
 456         spin_unlock_irq(&target->lock);
 457
 458         if (changed)
 459                 queue_work(system_long_wq, &target->remove_work);
 460
 461         return changed;
 462 }
 463
 464 static bool srp_change_conn_state(struct srp_target_port *target,
 465                                   bool connected)
 466 {
 467         bool changed = false;
 468
 469         spin_lock_irq(&target->lock);
 470         if (target->connected != connected) {
 471                 target->connected = connected;
 472                 changed = true;
 473         }
 474         spin_unlock_irq(&target->lock);
 475
 476         return changed;
 477 }
 478
 479 static void srp_disconnect_target(struct srp_target_port *target)
 480 {
 481         if (srp_change_conn_state(target, false)) {
 482                 /* XXX should send SRP_I_LOGOUT request */
 483
 484                 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
 485                         shost_printk(KERN_DEBUG, target->scsi_host,
 486                                      PFX "Sending CM DREQ failed\n");
 487                 }
 488         }
 489 }
 490
 491 static void srp_free_req_data(struct srp_target_port *target)
 492 {
 493         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
 494         struct srp_request *req;
 495         int i;
 496
 497         for (i = 0, req = target->req_ring; i < SRP_CMD_SQ_SIZE; ++i, ++req) {
 498                 kfree(req->fmr_list);
 499                 kfree(req->map_page);
 500                 if (req->indirect_dma_addr) {
 501                         ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
 502                                             target->indirect_size,
 503                                             DMA_TO_DEVICE);
 504                 }
 505                 kfree(req->indirect_desc);
 506         }
 507 }
 508
 509 /**
 510  * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
 511  * @shost: SCSI host whose attributes to remove from sysfs.
 512  *
 513  * Note: Any attributes defined in the host template and that did not exist
 514  * before invocation of this function will be ignored.
 515  */
 516 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
 517 {
 518         struct device_attribute **attr;
 519
 520         for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
 521                 device_remove_file(&shost->shost_dev, *attr);
 522 }
 523
 524 static void srp_remove_target(struct srp_target_port *target)
 525 {
 526         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 527
 528         srp_del_scsi_host_attr(target->scsi_host);
 529         srp_remove_host(target->scsi_host);
 530         scsi_remove_host(target->scsi_host);
 531         srp_disconnect_target(target);
 532         ib_destroy_cm_id(target->cm_id);
 533         srp_free_target_ib(target);
 534         srp_free_req_data(target);
 535         scsi_host_put(target->scsi_host);
 536 }
 537
 538 static void srp_remove_work(struct work_struct *work)
 539 {
 540         struct srp_target_port *target =
 541                 container_of(work, struct srp_target_port, remove_work);
 542
 543         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 544
 545         spin_lock(&target->srp_host->target_lock);
 546         list_del(&target->list);
 547         spin_unlock(&target->srp_host->target_lock);
 548
 549         srp_remove_target(target);
 550 }
 551
 552 static void srp_rport_delete(struct srp_rport *rport)
 553 {
 554         struct srp_target_port *target = rport->lld_data;
 555
 556         srp_queue_remove_work(target);
 557 }
 558
 559 static int srp_connect_target(struct srp_target_port *target)
 560 {
 561         int retries = 3;
 562         int ret;
 563
 564         WARN_ON_ONCE(target->connected);
 565
 566         target->qp_in_error = false;
 567
 568         ret = srp_lookup_path(target);
 569         if (ret)
 570                 return ret;
 571
 572         while (1) {
 573                 init_completion(&target->done);
 574                 ret = srp_send_req(target);
 575                 if (ret)
 576                         return ret;
 577                 wait_for_completion(&target->done);
 578
 579                 /*
 580                  * The CM event handling code will set status to
 581                  * SRP_PORT_REDIRECT if we get a port redirect REJ
 582                  * back, or SRP_DLID_REDIRECT if we get a lid/qp
 583                  * redirect REJ back.
 584                  */
 585                 switch (target->status) {
 586                 case 0:
 587                         srp_change_conn_state(target, true);
 588                         return 0;
 589
 590                 case SRP_PORT_REDIRECT:
 591                         ret = srp_lookup_path(target);
 592                         if (ret)
 593                                 return ret;
 594                         break;
 595
 596                 case SRP_DLID_REDIRECT:
 597                         break;
 598
 599                 case SRP_STALE_CONN:
 600                         /* Our current CM id was stale, and is now in timewait.
 601                          * Try to reconnect with a new one.
 602                          */
 603                         if (!retries-- || srp_new_cm_id(target)) {
 604                                 shost_printk(KERN_ERR, target->scsi_host, PFX
 605                                              "giving up on stale connection\n");
 606                                 target->status = -ECONNRESET;
 607                                 return target->status;
 608                         }
 609
 610                         shost_printk(KERN_ERR, target->scsi_host, PFX
 611                                      "retrying stale connection\n");
 612                         break;
 613
 614                 default:
 615                         return target->status;
 616                 }
 617         }
 618 }
 619
 620 static void srp_unmap_data(struct scsi_cmnd *scmnd,
 621                            struct srp_target_port *target,
 622                            struct srp_request *req)
 623 {
 624         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
 625         struct ib_pool_fmr **pfmr;
 626
 627         if (!scsi_sglist(scmnd) ||
 628             (scmnd->sc_data_direction != DMA_TO_DEVICE &&
 629              scmnd->sc_data_direction != DMA_FROM_DEVICE))
 630                 return;
 631
 632         pfmr = req->fmr_list;
 633         while (req->nfmr--)
 634                 ib_fmr_pool_unmap(*pfmr++);
 635
 636         ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
 637                         scmnd->sc_data_direction);
 638 }
 639
 640 /**
 641  * srp_claim_req - Take ownership of the scmnd associated with a request.
 642  * @target: SRP target port.
 643  * @req: SRP request.
 644  * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
 645  *         ownership of @req->scmnd if it equals @scmnd.
 646  *
 647  * Return value:
 648  * Either NULL or a pointer to the SCSI command the caller became owner of.
 649  */
 650 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
 651                                        struct srp_request *req,
 652                                        struct scsi_cmnd *scmnd)
 653 {
 654         unsigned long flags;
 655
 656         spin_lock_irqsave(&target->lock, flags);
 657         if (!scmnd) {
 658                 scmnd = req->scmnd;
 659                 req->scmnd = NULL;
 660         } else if (req->scmnd == scmnd) {
 661                 req->scmnd = NULL;
 662         } else {
 663                 scmnd = NULL;
 664         }
 665         spin_unlock_irqrestore(&target->lock, flags);
 666
 667         return scmnd;
 668 }
 669
 670 /**
 671  * srp_free_req() - Unmap data and add request to the free request list.
 672  */
 673 static void srp_free_req(struct srp_target_port *target,
 674                          struct srp_request *req, struct scsi_cmnd *scmnd,
 675                          s32 req_lim_delta)
 676 {
 677         unsigned long flags;
 678
 679         srp_unmap_data(scmnd, target, req);
 680
 681         spin_lock_irqsave(&target->lock, flags);
 682         target->req_lim += req_lim_delta;
 683         list_add_tail(&req->list, &target->free_reqs);
 684         spin_unlock_irqrestore(&target->lock, flags);
 685 }
 686
 687 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
 688 {
 689         struct scsi_cmnd *scmnd = srp_claim_req(target, req, NULL);
 690
 691         if (scmnd) {
 692                 srp_free_req(target, req, scmnd, 0);
 693                 scmnd->result = DID_RESET << 16;
 694                 scmnd->scsi_done(scmnd);
 695         }
 696 }
 697
 698 static int srp_reconnect_target(struct srp_target_port *target)
 699 {
 700         struct Scsi_Host *shost = target->scsi_host;
 701         int i, ret;
 702
 703         if (target->state != SRP_TARGET_LIVE)
 704                 return -EAGAIN;
 705
 706         scsi_target_block(&shost->shost_gendev);
 707
 708         srp_disconnect_target(target);
 709         /*
 710          * Now get a new local CM ID so that we avoid confusing the
 711          * target in case things are really fouled up.
 712          */
 713         ret = srp_new_cm_id(target);
 714         if (ret)
 715                 goto unblock;
 716
 717         ret = srp_create_target_ib(target);
 718         if (ret)
 719                 goto unblock;
 720
 721         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
 722                 struct srp_request *req = &target->req_ring[i];
 723                 if (req->scmnd)
 724                         srp_reset_req(target, req);
 725         }
 726
 727         INIT_LIST_HEAD(&target->free_tx);
 728         for (i = 0; i < SRP_SQ_SIZE; ++i)
 729                 list_add(&target->tx_ring[i]->list, &target->free_tx);
 730
 731         ret = srp_connect_target(target);
 732
 733 unblock:
 734         scsi_target_unblock(&shost->shost_gendev, ret == 0 ? SDEV_RUNNING :
 735                             SDEV_TRANSPORT_OFFLINE);
 736
 737         if (ret)
 738                 goto err;
 739
 740         shost_printk(KERN_INFO, target->scsi_host, PFX "reconnect succeeded\n");
 741
 742         return ret;
 743
 744 err:
 745         shost_printk(KERN_ERR, target->scsi_host,
 746                      PFX "reconnect failed (%d), removing target port.\n", ret);
 747
 748         /*
 749          * We couldn't reconnect, so kill our target port off.
 750          * However, we have to defer the real removal because we
 751          * are in the context of the SCSI error handler now, which
 752          * will deadlock if we call scsi_remove_host().
 753          */
 754         srp_queue_remove_work(target);
 755
 756         return ret;
 757 }
 758
 759 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
 760                          unsigned int dma_len, u32 rkey)
 761 {
 762         struct srp_direct_buf *desc = state->desc;
 763
 764         desc->va = cpu_to_be64(dma_addr);
 765         desc->key = cpu_to_be32(rkey);
 766         desc->len = cpu_to_be32(dma_len);
 767
 768         state->total_len += dma_len;
 769         state->desc++;
 770         state->ndesc++;
 771 }
 772
 773 static int srp_map_finish_fmr(struct srp_map_state *state,
 774                               struct srp_target_port *target)
 775 {
 776         struct srp_device *dev = target->srp_host->srp_dev;
 777         struct ib_pool_fmr *fmr;
 778         u64 io_addr = 0;
 779
 780         if (!state->npages)
 781                 return 0;
 782
 783         if (state->npages == 1) {
 784                 srp_map_desc(state, state->base_dma_addr, state->fmr_len,
 785                              target->rkey);
 786                 state->npages = state->fmr_len = 0;
 787                 return 0;
 788         }
 789
 790         fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
 791                                    state->npages, io_addr);
 792         if (IS_ERR(fmr))
 793                 return PTR_ERR(fmr);
 794
 795         *state->next_fmr++ = fmr;
 796         state->nfmr++;
 797
 798         srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
 799         state->npages = state->fmr_len = 0;
 800         return 0;
 801 }
 802
 803 static void srp_map_update_start(struct srp_map_state *state,
 804                                  struct scatterlist *sg, int sg_index,
 805                                  dma_addr_t dma_addr)
 806 {
 807         state->unmapped_sg = sg;
 808         state->unmapped_index = sg_index;
 809         state->unmapped_addr = dma_addr;
 810 }
 811
 812 static int srp_map_sg_entry(struct srp_map_state *state,
 813                             struct srp_target_port *target,
 814                             struct scatterlist *sg, int sg_index,
 815                             int use_fmr)
 816 {
 817         struct srp_device *dev = target->srp_host->srp_dev;
 818         struct ib_device *ibdev = dev->dev;
 819         dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
 820         unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
 821         unsigned int len;
 822         int ret;
 823
 824         if (!dma_len)
 825                 return 0;
 826
 827         if (use_fmr == SRP_MAP_NO_FMR) {
 828                 /* Once we're in direct map mode for a request, we don't
 829                  * go back to FMR mode, so no need to update anything
 830                  * other than the descriptor.
 831                  */
 832                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
 833                 return 0;
 834         }
 835
 836         /* If we start at an offset into the FMR page, don't merge into
 837          * the current FMR. Finish it out, and use the kernel's MR for this
 838          * sg entry. This is to avoid potential bugs on some SRP targets
 839          * that were never quite defined, but went away when the initiator
 840          * avoided using FMR on such page fragments.
 841          */
 842         if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
 843                 ret = srp_map_finish_fmr(state, target);
 844                 if (ret)
 845                         return ret;
 846
 847                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
 848                 srp_map_update_start(state, NULL, 0, 0);
 849                 return 0;
 850         }
 851
 852         /* If this is the first sg to go into the FMR, save our position.
 853          * We need to know the first unmapped entry, its index, and the
 854          * first unmapped address within that entry to be able to restart
 855          * mapping after an error.
 856          */
 857         if (!state->unmapped_sg)
 858                 srp_map_update_start(state, sg, sg_index, dma_addr);
 859
 860         while (dma_len) {
 861                 if (state->npages == SRP_FMR_SIZE) {
 862                         ret = srp_map_finish_fmr(state, target);
 863                         if (ret)
 864                                 return ret;
 865
 866                         srp_map_update_start(state, sg, sg_index, dma_addr);
 867                 }
 868
 869                 len = min_t(unsigned int, dma_len, dev->fmr_page_size);
 870
 871                 if (!state->npages)
 872                         state->base_dma_addr = dma_addr;
 873                 state->pages[state->npages++] = dma_addr;
 874                 state->fmr_len += len;
 875                 dma_addr += len;
 876                 dma_len -= len;
 877         }
 878
 879         /* If the last entry of the FMR wasn't a full page, then we need to
 880          * close it out and start a new one -- we can only merge at page
 881          * boundries.
 882          */
 883         ret = 0;
 884         if (len != dev->fmr_page_size) {
 885                 ret = srp_map_finish_fmr(state, target);
 886                 if (!ret)
 887                         srp_map_update_start(state, NULL, 0, 0);
 888         }
 889         return ret;
 890 }
 891
 892 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
 893                         struct srp_request *req)
 894 {
 895         struct scatterlist *scat, *sg;
 896         struct srp_cmd *cmd = req->cmd->buf;
 897         int i, len, nents, count, use_fmr;
 898         struct srp_device *dev;
 899         struct ib_device *ibdev;
 900         struct srp_map_state state;
 901         struct srp_indirect_buf *indirect_hdr;
 902         u32 table_len;
 903         u8 fmt;
 904
 905         if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
 906                 return sizeof (struct srp_cmd);
 907
 908         if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
 909             scmnd->sc_data_direction != DMA_TO_DEVICE) {
 910                 shost_printk(KERN_WARNING, target->scsi_host,
 911                              PFX "Unhandled data direction %d\n",
 912                              scmnd->sc_data_direction);
 913                 return -EINVAL;
 914         }
 915
 916         nents = scsi_sg_count(scmnd);
 917         scat  = scsi_sglist(scmnd);
 918
 919         dev = target->srp_host->srp_dev;
 920         ibdev = dev->dev;
 921
 922         count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
 923         if (unlikely(count == 0))
 924                 return -EIO;
 925
 926         fmt = SRP_DATA_DESC_DIRECT;
 927         len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
 928
 929         if (count == 1) {
 930                 /*
 931                  * The midlayer only generated a single gather/scatter
 932                  * entry, or DMA mapping coalesced everything to a
 933                  * single entry.  So a direct descriptor along with
 934                  * the DMA MR suffices.
 935                  */
 936                 struct srp_direct_buf *buf = (void *) cmd->add_data;
 937
 938                 buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
 939                 buf->key = cpu_to_be32(target->rkey);
 940                 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
 941
 942                 req->nfmr = 0;
 943                 goto map_complete;
 944         }
 945
 946         /* We have more than one scatter/gather entry, so build our indirect
 947          * descriptor table, trying to merge as many entries with FMR as we
 948          * can.
 949          */
 950         indirect_hdr = (void *) cmd->add_data;
 951
 952         ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
 953                                    target->indirect_size, DMA_TO_DEVICE);
 954
 955         memset(&state, 0, sizeof(state));
 956         state.desc      = req->indirect_desc;
 957         state.pages     = req->map_page;
 958         state.next_fmr  = req->fmr_list;
 959
 960         use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
 961
 962         for_each_sg(scat, sg, count, i) {
 963                 if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
 964                         /* FMR mapping failed, so backtrack to the first
 965                          * unmapped entry and continue on without using FMR.
 966                          */
 967                         dma_addr_t dma_addr;
 968                         unsigned int dma_len;
 969
 970 backtrack:
 971                         sg = state.unmapped_sg;
 972                         i = state.unmapped_index;
 973
 974                         dma_addr = ib_sg_dma_address(ibdev, sg);
 975                         dma_len = ib_sg_dma_len(ibdev, sg);
 976                         dma_len -= (state.unmapped_addr - dma_addr);
 977                         dma_addr = state.unmapped_addr;
 978                         use_fmr = SRP_MAP_NO_FMR;
 979                         srp_map_desc(&state, dma_addr, dma_len, target->rkey);
 980                 }
 981         }
 982
 983         if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
 984                 goto backtrack;
 985
 986         /* We've mapped the request, now pull as much of the indirect
 987          * descriptor table as we can into the command buffer. If this
 988          * target is not using an external indirect table, we are
 989          * guaranteed to fit into the command, as the SCSI layer won't
 990          * give us more S/G entries than we allow.
 991          */
 992         req->nfmr = state.nfmr;
 993         if (state.ndesc == 1) {
 994                 /* FMR mapping was able to collapse this to one entry,
 995                  * so use a direct descriptor.
 996                  */
 997                 struct srp_direct_buf *buf = (void *) cmd->add_data;
 998
 999                 *buf = req->indirect_desc[0];
1000                 goto map_complete;
1001         }
1002
1003         if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1004                                                 !target->allow_ext_sg)) {
1005                 shost_printk(KERN_ERR, target->scsi_host,
1006                              "Could not fit S/G list into SRP_CMD\n");
1007                 return -EIO;
1008         }
1009
1010         count = min(state.ndesc, target->cmd_sg_cnt);
1011         table_len = state.ndesc * sizeof (struct srp_direct_buf);
1012
1013         fmt = SRP_DATA_DESC_INDIRECT;
1014         len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1015         len += count * sizeof (struct srp_direct_buf);
1016
1017         memcpy(indirect_hdr->desc_list, req->indirect_desc,
1018                count * sizeof (struct srp_direct_buf));
1019
1020         indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1021         indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1022         indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1023         indirect_hdr->len = cpu_to_be32(state.total_len);
1024
1025         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1026                 cmd->data_out_desc_cnt = count;
1027         else
1028                 cmd->data_in_desc_cnt = count;
1029
1030         ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1031                                       DMA_TO_DEVICE);
1032
1033 map_complete:
1034         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1035                 cmd->buf_fmt = fmt << 4;
1036         else
1037                 cmd->buf_fmt = fmt;
1038
1039         return len;
1040 }
1041
1042 /*
1043  * Return an IU and possible credit to the free pool
1044  */
1045 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1046                           enum srp_iu_type iu_type)
1047 {
1048         unsigned long flags;
1049
1050         spin_lock_irqsave(&target->lock, flags);
1051         list_add(&iu->list, &target->free_tx);
1052         if (iu_type != SRP_IU_RSP)
1053                 ++target->req_lim;
1054         spin_unlock_irqrestore(&target->lock, flags);
1055 }
1056
1057 /*
1058  * Must be called with target->lock held to protect req_lim and free_tx.
1059  * If IU is not sent, it must be returned using srp_put_tx_iu().
1060  *
1061  * Note:
1062  * An upper limit for the number of allocated information units for each
1063  * request type is:
1064  * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1065  *   more than Scsi_Host.can_queue requests.
1066  * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1067  * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1068  *   one unanswered SRP request to an initiator.
1069  */
1070 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1071                                       enum srp_iu_type iu_type)
1072 {
1073         s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1074         struct srp_iu *iu;
1075
1076         srp_send_completion(target->send_cq, target);
1077
1078         if (list_empty(&target->free_tx))
1079                 return NULL;
1080
1081         /* Initiator responses to target requests do not consume credits */
1082         if (iu_type != SRP_IU_RSP) {
1083                 if (target->req_lim <= rsv) {
1084                         ++target->zero_req_lim;
1085                         return NULL;
1086                 }
1087
1088                 --target->req_lim;
1089         }
1090
1091         iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1092         list_del(&iu->list);
1093         return iu;
1094 }
1095
1096 static int srp_post_send(struct srp_target_port *target,
1097                          struct srp_iu *iu, int len)
1098 {
1099         struct ib_sge list;
1100         struct ib_send_wr wr, *bad_wr;
1101
1102         list.addr   = iu->dma;
1103         list.length = len;
1104         list.lkey   = target->lkey;
1105
1106         wr.next       = NULL;
1107         wr.wr_id      = (uintptr_t) iu;
1108         wr.sg_list    = &list;
1109         wr.num_sge    = 1;
1110         wr.opcode     = IB_WR_SEND;
1111         wr.send_flags = IB_SEND_SIGNALED;
1112
1113         return ib_post_send(target->qp, &wr, &bad_wr);
1114 }
1115
1116 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1117 {
1118         struct ib_recv_wr wr, *bad_wr;
1119         struct ib_sge list;
1120
1121         list.addr   = iu->dma;
1122         list.length = iu->size;
1123         list.lkey   = target->lkey;
1124
1125         wr.next     = NULL;
1126         wr.wr_id    = (uintptr_t) iu;
1127         wr.sg_list  = &list;
1128         wr.num_sge  = 1;
1129
1130         return ib_post_recv(target->qp, &wr, &bad_wr);
1131 }
1132
1133 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1134 {
1135         struct srp_request *req;
1136         struct scsi_cmnd *scmnd;
1137         unsigned long flags;
1138
1139         if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1140                 spin_lock_irqsave(&target->lock, flags);
1141                 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1142                 spin_unlock_irqrestore(&target->lock, flags);
1143
1144                 target->tsk_mgmt_status = -1;
1145                 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1146                         target->tsk_mgmt_status = rsp->data[3];
1147                 complete(&target->tsk_mgmt_done);
1148         } else {
1149                 req = &target->req_ring[rsp->tag];
1150                 scmnd = srp_claim_req(target, req, NULL);
1151                 if (!scmnd) {
1152                         shost_printk(KERN_ERR, target->scsi_host,
1153                                      "Null scmnd for RSP w/tag %016llx\n",
1154                                      (unsigned long long) rsp->tag);
1155
1156                         spin_lock_irqsave(&target->lock, flags);
1157                         target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1158                         spin_unlock_irqrestore(&target->lock, flags);
1159
1160                         return;
1161                 }
1162                 scmnd->result = rsp->status;
1163
1164                 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1165                         memcpy(scmnd->sense_buffer, rsp->data +
1166                                be32_to_cpu(rsp->resp_data_len),
1167                                min_t(int, be32_to_cpu(rsp->sense_data_len),
1168                                      SCSI_SENSE_BUFFERSIZE));
1169                 }
1170
1171                 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1172                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1173                 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1174                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1175
1176                 srp_free_req(target, req, scmnd,
1177                              be32_to_cpu(rsp->req_lim_delta));
1178
1179                 scmnd->host_scribble = NULL;
1180                 scmnd->scsi_done(scmnd);
1181         }
1182 }
1183
1184 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1185                                void *rsp, int len)
1186 {
1187         struct ib_device *dev = target->srp_host->srp_dev->dev;
1188         unsigned long flags;
1189         struct srp_iu *iu;
1190         int err;
1191
1192         spin_lock_irqsave(&target->lock, flags);
1193         target->req_lim += req_delta;
1194         iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1195         spin_unlock_irqrestore(&target->lock, flags);
1196
1197         if (!iu) {
1198                 shost_printk(KERN_ERR, target->scsi_host, PFX
1199                              "no IU available to send response\n");
1200                 return 1;
1201         }
1202
1203         ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1204         memcpy(iu->buf, rsp, len);
1205         ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1206
1207         err = srp_post_send(target, iu, len);
1208         if (err) {
1209                 shost_printk(KERN_ERR, target->scsi_host, PFX
1210                              "unable to post response: %d\n", err);
1211                 srp_put_tx_iu(target, iu, SRP_IU_RSP);
1212         }
1213
1214         return err;
1215 }
1216
1217 static void srp_process_cred_req(struct srp_target_port *target,
1218                                  struct srp_cred_req *req)
1219 {
1220         struct srp_cred_rsp rsp = {
1221                 .opcode = SRP_CRED_RSP,
1222                 .tag = req->tag,
1223         };
1224         s32 delta = be32_to_cpu(req->req_lim_delta);
1225
1226         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1227                 shost_printk(KERN_ERR, target->scsi_host, PFX
1228                              "problems processing SRP_CRED_REQ\n");
1229 }
1230
1231 static void srp_process_aer_req(struct srp_target_port *target,
1232                                 struct srp_aer_req *req)
1233 {
1234         struct srp_aer_rsp rsp = {
1235                 .opcode = SRP_AER_RSP,
1236                 .tag = req->tag,
1237         };
1238         s32 delta = be32_to_cpu(req->req_lim_delta);
1239
1240         shost_printk(KERN_ERR, target->scsi_host, PFX
1241                      "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1242
1243         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1244                 shost_printk(KERN_ERR, target->scsi_host, PFX
1245                              "problems processing SRP_AER_REQ\n");
1246 }
1247
1248 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1249 {
1250         struct ib_device *dev = target->srp_host->srp_dev->dev;
1251         struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1252         int res;
1253         u8 opcode;
1254
1255         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1256                                    DMA_FROM_DEVICE);
1257
1258         opcode = *(u8 *) iu->buf;
1259
1260         if (0) {
1261                 shost_printk(KERN_ERR, target->scsi_host,
1262                              PFX "recv completion, opcode 0x%02x\n", opcode);
1263                 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1264                                iu->buf, wc->byte_len, true);
1265         }
1266
1267         switch (opcode) {
1268         case SRP_RSP:
1269                 srp_process_rsp(target, iu->buf);
1270                 break;
1271
1272         case SRP_CRED_REQ:
1273                 srp_process_cred_req(target, iu->buf);
1274                 break;
1275
1276         case SRP_AER_REQ:
1277                 srp_process_aer_req(target, iu->buf);
1278                 break;
1279
1280         case SRP_T_LOGOUT:
1281                 /* XXX Handle target logout */
1282                 shost_printk(KERN_WARNING, target->scsi_host,
1283                              PFX "Got target logout request\n");
1284                 break;
1285
1286         default:
1287                 shost_printk(KERN_WARNING, target->scsi_host,
1288                              PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1289                 break;
1290         }
1291
1292         ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1293                                       DMA_FROM_DEVICE);
1294
1295         res = srp_post_recv(target, iu);
1296         if (res != 0)
1297                 shost_printk(KERN_ERR, target->scsi_host,
1298                              PFX "Recv failed with error code %d\n", res);
1299 }
1300
1301 static void srp_handle_qp_err(enum ib_wc_status wc_status,
1302                               enum ib_wc_opcode wc_opcode,
1303                               struct srp_target_port *target)
1304 {
1305         if (target->connected && !target->qp_in_error) {
1306                 shost_printk(KERN_ERR, target->scsi_host,
1307                              PFX "failed %s status %d\n",
1308                              wc_opcode & IB_WC_RECV ? "receive" : "send",
1309                              wc_status);
1310         }
1311         target->qp_in_error = true;
1312 }
1313
1314 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1315 {
1316         struct srp_target_port *target = target_ptr;
1317         struct ib_wc wc;
1318
1319         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1320         while (ib_poll_cq(cq, 1, &wc) > 0) {
1321                 if (likely(wc.status == IB_WC_SUCCESS)) {
1322                         srp_handle_recv(target, &wc);
1323                 } else {
1324                         srp_handle_qp_err(wc.status, wc.opcode, target);
1325                 }
1326         }
1327 }
1328
1329 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1330 {
1331         struct srp_target_port *target = target_ptr;
1332         struct ib_wc wc;
1333         struct srp_iu *iu;
1334
1335         while (ib_poll_cq(cq, 1, &wc) > 0) {
1336                 if (likely(wc.status == IB_WC_SUCCESS)) {
1337                         iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1338                         list_add(&iu->list, &target->free_tx);
1339                 } else {
1340                         srp_handle_qp_err(wc.status, wc.opcode, target);
1341                 }
1342         }
1343 }
1344
1345 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1346 {
1347         struct srp_target_port *target = host_to_target(shost);
1348         struct srp_request *req;
1349         struct srp_iu *iu;
1350         struct srp_cmd *cmd;
1351         struct ib_device *dev;
1352         unsigned long flags;
1353         int len;
1354
1355         spin_lock_irqsave(&target->lock, flags);
1356         iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1357         if (!iu)
1358                 goto err_unlock;
1359
1360         req = list_first_entry(&target->free_reqs, struct srp_request, list);
1361         list_del(&req->list);
1362         spin_unlock_irqrestore(&target->lock, flags);
1363
1364         dev = target->srp_host->srp_dev->dev;
1365         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1366                                    DMA_TO_DEVICE);
1367
1368         scmnd->result        = 0;
1369         scmnd->host_scribble = (void *) req;
1370
1371         cmd = iu->buf;
1372         memset(cmd, 0, sizeof *cmd);
1373
1374         cmd->opcode = SRP_CMD;
1375         cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
1376         cmd->tag    = req->index;
1377         memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1378
1379         req->scmnd    = scmnd;
1380         req->cmd      = iu;
1381
1382         len = srp_map_data(scmnd, target, req);
1383         if (len < 0) {
1384                 shost_printk(KERN_ERR, target->scsi_host,
1385                              PFX "Failed to map data\n");
1386                 goto err_iu;
1387         }
1388
1389         ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1390                                       DMA_TO_DEVICE);
1391
1392         if (srp_post_send(target, iu, len)) {
1393                 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1394                 goto err_unmap;
1395         }
1396
1397         return 0;
1398
1399 err_unmap:
1400         srp_unmap_data(scmnd, target, req);
1401
1402 err_iu:
1403         srp_put_tx_iu(target, iu, SRP_IU_CMD);
1404
1405         spin_lock_irqsave(&target->lock, flags);
1406         list_add(&req->list, &target->free_reqs);
1407
1408 err_unlock:
1409         spin_unlock_irqrestore(&target->lock, flags);
1410
1411         return SCSI_MLQUEUE_HOST_BUSY;
1412 }
1413
1414 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1415 {
1416         int i;
1417
1418         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1419                 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1420                                                   target->max_ti_iu_len,
1421                                                   GFP_KERNEL, DMA_FROM_DEVICE);
1422                 if (!target->rx_ring[i])
1423                         goto err;
1424         }
1425
1426         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1427                 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1428                                                   target->max_iu_len,
1429                                                   GFP_KERNEL, DMA_TO_DEVICE);
1430                 if (!target->tx_ring[i])
1431                         goto err;
1432
1433                 list_add(&target->tx_ring[i]->list, &target->free_tx);
1434         }
1435
1436         return 0;
1437
1438 err:
1439         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1440                 srp_free_iu(target->srp_host, target->rx_ring[i]);
1441                 target->rx_ring[i] = NULL;
1442         }
1443
1444         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1445                 srp_free_iu(target->srp_host, target->tx_ring[i]);
1446                 target->tx_ring[i] = NULL;
1447         }
1448
1449         return -ENOMEM;
1450 }
1451
1452 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
1453 {
1454         uint64_t T_tr_ns, max_compl_time_ms;
1455         uint32_t rq_tmo_jiffies;
1456
1457         /*
1458          * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
1459          * table 91), both the QP timeout and the retry count have to be set
1460          * for RC QP's during the RTR to RTS transition.
1461          */
1462         WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
1463                      (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
1464
1465         /*
1466          * Set target->rq_tmo_jiffies to one second more than the largest time
1467          * it can take before an error completion is generated. See also
1468          * C9-140..142 in the IBTA spec for more information about how to
1469          * convert the QP Local ACK Timeout value to nanoseconds.
1470          */
1471         T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
1472         max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
1473         do_div(max_compl_time_ms, NSEC_PER_MSEC);
1474         rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
1475
1476         return rq_tmo_jiffies;
1477 }
1478
1479 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1480                                struct srp_login_rsp *lrsp,
1481                                struct srp_target_port *target)
1482 {
1483         struct ib_qp_attr *qp_attr = NULL;
1484         int attr_mask = 0;
1485         int ret;
1486         int i;
1487
1488         if (lrsp->opcode == SRP_LOGIN_RSP) {
1489                 target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1490                 target->req_lim       = be32_to_cpu(lrsp->req_lim_delta);
1491
1492                 /*
1493                  * Reserve credits for task management so we don't
1494                  * bounce requests back to the SCSI mid-layer.
1495                  */
1496                 target->scsi_host->can_queue
1497                         = min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1498                               target->scsi_host->can_queue);
1499         } else {
1500                 shost_printk(KERN_WARNING, target->scsi_host,
1501                              PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1502                 ret = -ECONNRESET;
1503                 goto error;
1504         }
1505
1506         if (!target->rx_ring[0]) {
1507                 ret = srp_alloc_iu_bufs(target);
1508                 if (ret)
1509                         goto error;
1510         }
1511
1512         ret = -ENOMEM;
1513         qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1514         if (!qp_attr)
1515                 goto error;
1516
1517         qp_attr->qp_state = IB_QPS_RTR;
1518         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1519         if (ret)
1520                 goto error_free;
1521
1522         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1523         if (ret)
1524                 goto error_free;
1525
1526         for (i = 0; i < SRP_RQ_SIZE; i++) {
1527                 struct srp_iu *iu = target->rx_ring[i];
1528                 ret = srp_post_recv(target, iu);
1529                 if (ret)
1530                         goto error_free;
1531         }
1532
1533         qp_attr->qp_state = IB_QPS_RTS;
1534         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1535         if (ret)
1536                 goto error_free;
1537
1538         target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
1539
1540         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1541         if (ret)
1542                 goto error_free;
1543
1544         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1545
1546 error_free:
1547         kfree(qp_attr);
1548
1549 error:
1550         target->status = ret;
1551 }
1552
1553 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1554                                struct ib_cm_event *event,
1555                                struct srp_target_port *target)
1556 {
1557         struct Scsi_Host *shost = target->scsi_host;
1558         struct ib_class_port_info *cpi;
1559         int opcode;
1560
1561         switch (event->param.rej_rcvd.reason) {
1562         case IB_CM_REJ_PORT_CM_REDIRECT:
1563                 cpi = event->param.rej_rcvd.ari;
1564                 target->path.dlid = cpi->redirect_lid;
1565                 target->path.pkey = cpi->redirect_pkey;
1566                 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1567                 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1568
1569                 target->status = target->path.dlid ?
1570                         SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1571                 break;
1572
1573         case IB_CM_REJ_PORT_REDIRECT:
1574                 if (srp_target_is_topspin(target)) {
1575                         /*
1576                          * Topspin/Cisco SRP gateways incorrectly send
1577                          * reject reason code 25 when they mean 24
1578                          * (port redirect).
1579                          */
1580                         memcpy(target->path.dgid.raw,
1581                                event->param.rej_rcvd.ari, 16);
1582
1583                         shost_printk(KERN_DEBUG, shost,
1584                                      PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1585                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1586                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1587
1588                         target->status = SRP_PORT_REDIRECT;
1589                 } else {
1590                         shost_printk(KERN_WARNING, shost,
1591                                      "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1592                         target->status = -ECONNRESET;
1593                 }
1594                 break;
1595
1596         case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1597                 shost_printk(KERN_WARNING, shost,
1598                             "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1599                 target->status = -ECONNRESET;
1600                 break;
1601
1602         case IB_CM_REJ_CONSUMER_DEFINED:
1603                 opcode = *(u8 *) event->private_data;
1604                 if (opcode == SRP_LOGIN_REJ) {
1605                         struct srp_login_rej *rej = event->private_data;
1606                         u32 reason = be32_to_cpu(rej->reason);
1607
1608                         if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1609                                 shost_printk(KERN_WARNING, shost,
1610                                              PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1611                         else
1612                                 shost_printk(KERN_WARNING, shost,
1613                                             PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1614                 } else
1615                         shost_printk(KERN_WARNING, shost,
1616                                      "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1617                                      " opcode 0x%02x\n", opcode);
1618                 target->status = -ECONNRESET;
1619                 break;
1620
1621         case IB_CM_REJ_STALE_CONN:
1622                 shost_printk(KERN_WARNING, shost, "  REJ reason: stale connection\n");
1623                 target->status = SRP_STALE_CONN;
1624                 break;
1625
1626         default:
1627                 shost_printk(KERN_WARNING, shost, "  REJ reason 0x%x\n",
1628                              event->param.rej_rcvd.reason);
1629                 target->status = -ECONNRESET;
1630         }
1631 }
1632
1633 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1634 {
1635         struct srp_target_port *target = cm_id->context;
1636         int comp = 0;
1637
1638         switch (event->event) {
1639         case IB_CM_REQ_ERROR:
1640                 shost_printk(KERN_DEBUG, target->scsi_host,
1641                              PFX "Sending CM REQ failed\n");
1642                 comp = 1;
1643                 target->status = -ECONNRESET;
1644                 break;
1645
1646         case IB_CM_REP_RECEIVED:
1647                 comp = 1;
1648                 srp_cm_rep_handler(cm_id, event->private_data, target);
1649                 break;
1650
1651         case IB_CM_REJ_RECEIVED:
1652                 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1653                 comp = 1;
1654
1655                 srp_cm_rej_handler(cm_id, event, target);
1656                 break;
1657
1658         case IB_CM_DREQ_RECEIVED:
1659                 shost_printk(KERN_WARNING, target->scsi_host,
1660                              PFX "DREQ received - connection closed\n");
1661                 srp_change_conn_state(target, false);
1662                 if (ib_send_cm_drep(cm_id, NULL, 0))
1663                         shost_printk(KERN_ERR, target->scsi_host,
1664                                      PFX "Sending CM DREP failed\n");
1665                 break;
1666
1667         case IB_CM_TIMEWAIT_EXIT:
1668                 shost_printk(KERN_ERR, target->scsi_host,
1669                              PFX "connection closed\n");
1670
1671                 target->status = 0;
1672                 break;
1673
1674         case IB_CM_MRA_RECEIVED:
1675         case IB_CM_DREQ_ERROR:
1676         case IB_CM_DREP_RECEIVED:
1677                 break;
1678
1679         default:
1680                 shost_printk(KERN_WARNING, target->scsi_host,
1681                              PFX "Unhandled CM event %d\n", event->event);
1682                 break;
1683         }
1684
1685         if (comp)
1686                 complete(&target->done);
1687
1688         return 0;
1689 }
1690
1691 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1692                              u64 req_tag, unsigned int lun, u8 func)
1693 {
1694         struct ib_device *dev = target->srp_host->srp_dev->dev;
1695         struct srp_iu *iu;
1696         struct srp_tsk_mgmt *tsk_mgmt;
1697
1698         init_completion(&target->tsk_mgmt_done);
1699
1700         spin_lock_irq(&target->lock);
1701         iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1702         spin_unlock_irq(&target->lock);
1703
1704         if (!iu)
1705                 return -1;
1706
1707         ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1708                                    DMA_TO_DEVICE);
1709         tsk_mgmt = iu->buf;
1710         memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1711
1712         tsk_mgmt->opcode        = SRP_TSK_MGMT;
1713         tsk_mgmt->lun           = cpu_to_be64((u64) lun << 48);
1714         tsk_mgmt->tag           = req_tag | SRP_TAG_TSK_MGMT;
1715         tsk_mgmt->tsk_mgmt_func = func;
1716         tsk_mgmt->task_tag      = req_tag;
1717
1718         ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1719                                       DMA_TO_DEVICE);
1720         if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1721                 srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1722                 return -1;
1723         }
1724
1725         if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1726                                          msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1727                 return -1;
1728
1729         return 0;
1730 }
1731
1732 static int srp_abort(struct scsi_cmnd *scmnd)
1733 {
1734         struct srp_target_port *target = host_to_target(scmnd->device->host);
1735         struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1736
1737         shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
1738
1739         if (!req || target->qp_in_error || !srp_claim_req(target, req, scmnd))
1740                 return FAILED;
1741         srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
1742                           SRP_TSK_ABORT_TASK);
1743         srp_free_req(target, req, scmnd, 0);
1744         scmnd->result = DID_ABORT << 16;
1745         scmnd->scsi_done(scmnd);
1746
1747         return SUCCESS;
1748 }
1749
1750 static int srp_reset_device(struct scsi_cmnd *scmnd)
1751 {
1752         struct srp_target_port *target = host_to_target(scmnd->device->host);
1753         int i;
1754
1755         shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
1756
1757         if (target->qp_in_error)
1758                 return FAILED;
1759         if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
1760                               SRP_TSK_LUN_RESET))
1761                 return FAILED;
1762         if (target->tsk_mgmt_status)
1763                 return FAILED;
1764
1765         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1766                 struct srp_request *req = &target->req_ring[i];
1767                 if (req->scmnd && req->scmnd->device == scmnd->device)
1768                         srp_reset_req(target, req);
1769         }
1770
1771         return SUCCESS;
1772 }
1773
1774 static int srp_reset_host(struct scsi_cmnd *scmnd)
1775 {
1776         struct srp_target_port *target = host_to_target(scmnd->device->host);
1777         int ret = FAILED;
1778
1779         shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
1780
1781         if (!srp_reconnect_target(target))
1782                 ret = SUCCESS;
1783
1784         return ret;
1785 }
1786
1787 static int srp_slave_configure(struct scsi_device *sdev)
1788 {
1789         struct Scsi_Host *shost = sdev->host;
1790         struct srp_target_port *target = host_to_target(shost);
1791         struct request_queue *q = sdev->request_queue;
1792         unsigned long timeout;
1793
1794         if (sdev->type == TYPE_DISK) {
1795                 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
1796                 blk_queue_rq_timeout(q, timeout);
1797         }
1798
1799         return 0;
1800 }
1801
1802 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
1803                            char *buf)
1804 {
1805         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1806
1807         return sprintf(buf, "0x%016llx\n",
1808                        (unsigned long long) be64_to_cpu(target->id_ext));
1809 }
1810
1811 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
1812                              char *buf)
1813 {
1814         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1815
1816         return sprintf(buf, "0x%016llx\n",
1817                        (unsigned long long) be64_to_cpu(target->ioc_guid));
1818 }
1819
1820 static ssize_t show_service_id(struct device *dev,
1821                                struct device_attribute *attr, char *buf)
1822 {
1823         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1824
1825         return sprintf(buf, "0x%016llx\n",
1826                        (unsigned long long) be64_to_cpu(target->service_id));
1827 }
1828
1829 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
1830                          char *buf)
1831 {
1832         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1833
1834         return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1835 }
1836
1837 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
1838                          char *buf)
1839 {
1840         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1841
1842         return sprintf(buf, "%pI6\n", target->path.dgid.raw);
1843 }
1844
1845 static ssize_t show_orig_dgid(struct device *dev,
1846                               struct device_attribute *attr, char *buf)
1847 {
1848         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1849
1850         return sprintf(buf, "%pI6\n", target->orig_dgid);
1851 }
1852
1853 static ssize_t show_req_lim(struct device *dev,
1854                             struct device_attribute *attr, char *buf)
1855 {
1856         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1857
1858         return sprintf(buf, "%d\n", target->req_lim);
1859 }
1860
1861 static ssize_t show_zero_req_lim(struct device *dev,
1862                                  struct device_attribute *attr, char *buf)
1863 {
1864         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1865
1866         return sprintf(buf, "%d\n", target->zero_req_lim);
1867 }
1868
1869 static ssize_t show_local_ib_port(struct device *dev,
1870                                   struct device_attribute *attr, char *buf)
1871 {
1872         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1873
1874         return sprintf(buf, "%d\n", target->srp_host->port);
1875 }
1876
1877 static ssize_t show_local_ib_device(struct device *dev,
1878                                     struct device_attribute *attr, char *buf)
1879 {
1880         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1881
1882         return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
1883 }
1884
1885 static ssize_t show_cmd_sg_entries(struct device *dev,
1886                                    struct device_attribute *attr, char *buf)
1887 {
1888         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1889
1890         return sprintf(buf, "%u\n", target->cmd_sg_cnt);
1891 }
1892
1893 static ssize_t show_allow_ext_sg(struct device *dev,
1894                                  struct device_attribute *attr, char *buf)
1895 {
1896         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1897
1898         return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
1899 }
1900
1901 static DEVICE_ATTR(id_ext,          S_IRUGO, show_id_ext,          NULL);
1902 static DEVICE_ATTR(ioc_guid,        S_IRUGO, show_ioc_guid,        NULL);
1903 static DEVICE_ATTR(service_id,      S_IRUGO, show_service_id,      NULL);
1904 static DEVICE_ATTR(pkey,            S_IRUGO, show_pkey,            NULL);
1905 static DEVICE_ATTR(dgid,            S_IRUGO, show_dgid,            NULL);
1906 static DEVICE_ATTR(orig_dgid,       S_IRUGO, show_orig_dgid,       NULL);
1907 static DEVICE_ATTR(req_lim,         S_IRUGO, show_req_lim,         NULL);
1908 static DEVICE_ATTR(zero_req_lim,    S_IRUGO, show_zero_req_lim,    NULL);
1909 static DEVICE_ATTR(local_ib_port,   S_IRUGO, show_local_ib_port,   NULL);
1910 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1911 static DEVICE_ATTR(cmd_sg_entries,  S_IRUGO, show_cmd_sg_entries,  NULL);
1912 static DEVICE_ATTR(allow_ext_sg,    S_IRUGO, show_allow_ext_sg,    NULL);
1913
1914 static struct device_attribute *srp_host_attrs[] = {
1915         &dev_attr_id_ext,
1916         &dev_attr_ioc_guid,
1917         &dev_attr_service_id,
1918         &dev_attr_pkey,
1919         &dev_attr_dgid,
1920         &dev_attr_orig_dgid,
1921         &dev_attr_req_lim,
1922         &dev_attr_zero_req_lim,
1923         &dev_attr_local_ib_port,
1924         &dev_attr_local_ib_device,
1925         &dev_attr_cmd_sg_entries,
1926         &dev_attr_allow_ext_sg,
1927         NULL
1928 };
1929
1930 static struct scsi_host_template srp_template = {
1931         .module                         = THIS_MODULE,
1932         .name                           = "InfiniBand SRP initiator",
1933         .proc_name                      = DRV_NAME,
1934         .slave_configure                = srp_slave_configure,
1935         .info                           = srp_target_info,
1936         .queuecommand                   = srp_queuecommand,
1937         .eh_abort_handler               = srp_abort,
1938         .eh_device_reset_handler        = srp_reset_device,
1939         .eh_host_reset_handler          = srp_reset_host,
1940         .sg_tablesize                   = SRP_DEF_SG_TABLESIZE,
1941         .can_queue                      = SRP_CMD_SQ_SIZE,
1942         .this_id                        = -1,
1943         .cmd_per_lun                    = SRP_CMD_SQ_SIZE,
1944         .use_clustering                 = ENABLE_CLUSTERING,
1945         .shost_attrs                    = srp_host_attrs
1946 };
1947
1948 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1949 {
1950         struct srp_rport_identifiers ids;
1951         struct srp_rport *rport;
1952
1953         sprintf(target->target_name, "SRP.T10:%016llX",
1954                  (unsigned long long) be64_to_cpu(target->id_ext));
1955
1956         if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
1957                 return -ENODEV;
1958
1959         memcpy(ids.port_id, &target->id_ext, 8);
1960         memcpy(ids.port_id + 8, &target->ioc_guid, 8);
1961         ids.roles = SRP_RPORT_ROLE_TARGET;
1962         rport = srp_rport_add(target->scsi_host, &ids);
1963         if (IS_ERR(rport)) {
1964                 scsi_remove_host(target->scsi_host);
1965                 return PTR_ERR(rport);
1966         }
1967
1968         rport->lld_data = target;
1969
1970         spin_lock(&host->target_lock);
1971         list_add_tail(&target->list, &host->target_list);
1972         spin_unlock(&host->target_lock);
1973
1974         target->state = SRP_TARGET_LIVE;
1975         target->connected = false;
1976
1977         scsi_scan_target(&target->scsi_host->shost_gendev,
1978                          0, target->scsi_id, SCAN_WILD_CARD, 0);
1979
1980         return 0;
1981 }
1982
1983 static void srp_release_dev(struct device *dev)
1984 {
1985         struct srp_host *host =
1986                 container_of(dev, struct srp_host, dev);
1987
1988         complete(&host->released);
1989 }
1990
1991 static struct class srp_class = {
1992         .name    = "infiniband_srp",
1993         .dev_release = srp_release_dev
1994 };
1995
1996 /*
1997  * Target ports are added by writing
1998  *
1999  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2000  *     pkey=<P_Key>,service_id=<service ID>
2001  *
2002  * to the add_target sysfs attribute.
2003  */
2004 enum {
2005         SRP_OPT_ERR             = 0,
2006         SRP_OPT_ID_EXT          = 1 << 0,
2007         SRP_OPT_IOC_GUID        = 1 << 1,
2008         SRP_OPT_DGID            = 1 << 2,
2009         SRP_OPT_PKEY            = 1 << 3,
2010         SRP_OPT_SERVICE_ID      = 1 << 4,
2011         SRP_OPT_MAX_SECT        = 1 << 5,
2012         SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2013         SRP_OPT_IO_CLASS        = 1 << 7,
2014         SRP_OPT_INITIATOR_EXT   = 1 << 8,
2015         SRP_OPT_CMD_SG_ENTRIES  = 1 << 9,
2016         SRP_OPT_ALLOW_EXT_SG    = 1 << 10,
2017         SRP_OPT_SG_TABLESIZE    = 1 << 11,
2018         SRP_OPT_ALL             = (SRP_OPT_ID_EXT       |
2019                                    SRP_OPT_IOC_GUID     |
2020                                    SRP_OPT_DGID         |
2021                                    SRP_OPT_PKEY         |
2022                                    SRP_OPT_SERVICE_ID),
2023 };
2024
2025 static const match_table_t srp_opt_tokens = {
2026         { SRP_OPT_ID_EXT,               "id_ext=%s"             },
2027         { SRP_OPT_IOC_GUID,             "ioc_guid=%s"           },
2028         { SRP_OPT_DGID,                 "dgid=%s"               },
2029         { SRP_OPT_PKEY,                 "pkey=%x"               },
2030         { SRP_OPT_SERVICE_ID,           "service_id=%s"         },
2031         { SRP_OPT_MAX_SECT,             "max_sect=%d"           },
2032         { SRP_OPT_MAX_CMD_PER_LUN,      "max_cmd_per_lun=%d"    },
2033         { SRP_OPT_IO_CLASS,             "io_class=%x"           },
2034         { SRP_OPT_INITIATOR_EXT,        "initiator_ext=%s"      },
2035         { SRP_OPT_CMD_SG_ENTRIES,       "cmd_sg_entries=%u"     },
2036         { SRP_OPT_ALLOW_EXT_SG,         "allow_ext_sg=%u"       },
2037         { SRP_OPT_SG_TABLESIZE,         "sg_tablesize=%u"       },
2038         { SRP_OPT_ERR,                  NULL                    }
2039 };
2040
2041 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2042 {
2043         char *options, *sep_opt;
2044         char *p;
2045         char dgid[3];
2046         substring_t args[MAX_OPT_ARGS];
2047         int opt_mask = 0;
2048         int token;
2049         int ret = -EINVAL;
2050         int i;
2051
2052         options = kstrdup(buf, GFP_KERNEL);
2053         if (!options)
2054                 return -ENOMEM;
2055
2056         sep_opt = options;
2057         while ((p = strsep(&sep_opt, ",")) != NULL) {
2058                 if (!*p)
2059                         continue;
2060
2061                 token = match_token(p, srp_opt_tokens, args);
2062                 opt_mask |= token;
2063
2064                 switch (token) {
2065                 case SRP_OPT_ID_EXT:
2066                         p = match_strdup(args);
2067                         if (!p) {
2068                                 ret = -ENOMEM;
2069                                 goto out;
2070                         }
2071                         target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2072                         kfree(p);
2073                         break;
2074
2075                 case SRP_OPT_IOC_GUID:
2076                         p = match_strdup(args);
2077                         if (!p) {
2078                                 ret = -ENOMEM;
2079                                 goto out;
2080                         }
2081                         target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2082                         kfree(p);
2083                         break;
2084
2085                 case SRP_OPT_DGID:
2086                         p = match_strdup(args);
2087                         if (!p) {
2088                                 ret = -ENOMEM;
2089                                 goto out;
2090                         }
2091                         if (strlen(p) != 32) {
2092                                 pr_warn("bad dest GID parameter '%s'\n", p);
2093                                 kfree(p);
2094                                 goto out;
2095                         }
2096
2097                         for (i = 0; i < 16; ++i) {
2098                                 strlcpy(dgid, p + i * 2, 3);
2099                                 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2100                         }
2101                         kfree(p);
2102                         memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2103                         break;
2104
2105                 case SRP_OPT_PKEY:
2106                         if (match_hex(args, &token)) {
2107                                 pr_warn("bad P_Key parameter '%s'\n", p);
2108                                 goto out;
2109                         }
2110                         target->path.pkey = cpu_to_be16(token);
2111                         break;
2112
2113                 case SRP_OPT_SERVICE_ID:
2114                         p = match_strdup(args);
2115                         if (!p) {
2116                                 ret = -ENOMEM;
2117                                 goto out;
2118                         }
2119                         target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2120                         target->path.service_id = target->service_id;
2121                         kfree(p);
2122                         break;
2123
2124                 case SRP_OPT_MAX_SECT:
2125                         if (match_int(args, &token)) {
2126                                 pr_warn("bad max sect parameter '%s'\n", p);
2127                                 goto out;
2128                         }
2129                         target->scsi_host->max_sectors = token;
2130                         break;
2131
2132                 case SRP_OPT_MAX_CMD_PER_LUN:
2133                         if (match_int(args, &token)) {
2134                                 pr_warn("bad max cmd_per_lun parameter '%s'\n",
2135                                         p);
2136                                 goto out;
2137                         }
2138                         target->scsi_host->cmd_per_lun = min(token, SRP_CMD_SQ_SIZE);
2139                         break;
2140
2141                 case SRP_OPT_IO_CLASS:
2142                         if (match_hex(args, &token)) {
2143                                 pr_warn("bad IO class parameter '%s'\n", p);
2144                                 goto out;
2145                         }
2146                         if (token != SRP_REV10_IB_IO_CLASS &&
2147                             token != SRP_REV16A_IB_IO_CLASS) {
2148                                 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2149                                         token, SRP_REV10_IB_IO_CLASS,
2150                                         SRP_REV16A_IB_IO_CLASS);
2151                                 goto out;
2152                         }
2153                         target->io_class = token;
2154                         break;
2155
2156                 case SRP_OPT_INITIATOR_EXT:
2157                         p = match_strdup(args);
2158                         if (!p) {
2159                                 ret = -ENOMEM;
2160                                 goto out;
2161                         }
2162                         target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2163                         kfree(p);
2164                         break;
2165
2166                 case SRP_OPT_CMD_SG_ENTRIES:
2167                         if (match_int(args, &token) || token < 1 || token > 255) {
2168                                 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2169                                         p);
2170                                 goto out;
2171                         }
2172                         target->cmd_sg_cnt = token;
2173                         break;
2174
2175                 case SRP_OPT_ALLOW_EXT_SG:
2176                         if (match_int(args, &token)) {
2177                                 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2178                                 goto out;
2179                         }
2180                         target->allow_ext_sg = !!token;
2181                         break;
2182
2183                 case SRP_OPT_SG_TABLESIZE:
2184                         if (match_int(args, &token) || token < 1 ||
2185                                         token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2186                                 pr_warn("bad max sg_tablesize parameter '%s'\n",
2187                                         p);
2188                                 goto out;
2189                         }
2190                         target->sg_tablesize = token;
2191                         break;
2192
2193                 default:
2194                         pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2195                                 p);
2196                         goto out;
2197                 }
2198         }
2199
2200         if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2201                 ret = 0;
2202         else
2203                 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2204                         if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2205                             !(srp_opt_tokens[i].token & opt_mask))
2206                                 pr_warn("target creation request is missing parameter '%s'\n",
2207                                         srp_opt_tokens[i].pattern);
2208
2209 out:
2210         kfree(options);
2211         return ret;
2212 }
2213
2214 static ssize_t srp_create_target(struct device *dev,
2215                                  struct device_attribute *attr,
2216                                  const char *buf, size_t count)
2217 {
2218         struct srp_host *host =
2219                 container_of(dev, struct srp_host, dev);
2220         struct Scsi_Host *target_host;
2221         struct srp_target_port *target;
2222         struct ib_device *ibdev = host->srp_dev->dev;
2223         dma_addr_t dma_addr;
2224         int i, ret;
2225
2226         target_host = scsi_host_alloc(&srp_template,
2227                                       sizeof (struct srp_target_port));
2228         if (!target_host)
2229                 return -ENOMEM;
2230
2231         target_host->transportt  = ib_srp_transport_template;
2232         target_host->max_channel = 0;
2233         target_host->max_id      = 1;
2234         target_host->max_lun     = SRP_MAX_LUN;
2235         target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2236
2237         target = host_to_target(target_host);
2238
2239         target->io_class        = SRP_REV16A_IB_IO_CLASS;
2240         target->scsi_host       = target_host;
2241         target->srp_host        = host;
2242         target->lkey            = host->srp_dev->mr->lkey;
2243         target->rkey            = host->srp_dev->mr->rkey;
2244         target->cmd_sg_cnt      = cmd_sg_entries;
2245         target->sg_tablesize    = indirect_sg_entries ? : cmd_sg_entries;
2246         target->allow_ext_sg    = allow_ext_sg;
2247
2248         ret = srp_parse_options(buf, target);
2249         if (ret)
2250                 goto err;
2251
2252         if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2253                                 target->cmd_sg_cnt < target->sg_tablesize) {
2254                 pr_warn("No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2255                 target->sg_tablesize = target->cmd_sg_cnt;
2256         }
2257
2258         target_host->sg_tablesize = target->sg_tablesize;
2259         target->indirect_size = target->sg_tablesize *
2260                                 sizeof (struct srp_direct_buf);
2261         target->max_iu_len = sizeof (struct srp_cmd) +
2262                              sizeof (struct srp_indirect_buf) +
2263                              target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2264
2265         INIT_WORK(&target->remove_work, srp_remove_work);
2266         spin_lock_init(&target->lock);
2267         INIT_LIST_HEAD(&target->free_tx);
2268         INIT_LIST_HEAD(&target->free_reqs);
2269         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
2270                 struct srp_request *req = &target->req_ring[i];
2271
2272                 req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof (void *),
2273                                         GFP_KERNEL);
2274                 req->map_page = kmalloc(SRP_FMR_SIZE * sizeof (void *),
2275                                         GFP_KERNEL);
2276                 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
2277                 if (!req->fmr_list || !req->map_page || !req->indirect_desc)
2278                         goto err_free_mem;
2279
2280                 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
2281                                              target->indirect_size,
2282                                              DMA_TO_DEVICE);
2283                 if (ib_dma_mapping_error(ibdev, dma_addr))
2284                         goto err_free_mem;
2285
2286                 req->indirect_dma_addr = dma_addr;
2287                 req->index = i;
2288                 list_add_tail(&req->list, &target->free_reqs);
2289         }
2290
2291         ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2292
2293         shost_printk(KERN_DEBUG, target->scsi_host, PFX
2294                      "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
2295                      "service_id %016llx dgid %pI6\n",
2296                (unsigned long long) be64_to_cpu(target->id_ext),
2297                (unsigned long long) be64_to_cpu(target->ioc_guid),
2298                be16_to_cpu(target->path.pkey),
2299                (unsigned long long) be64_to_cpu(target->service_id),
2300                target->path.dgid.raw);
2301
2302         ret = srp_create_target_ib(target);
2303         if (ret)
2304                 goto err_free_mem;
2305
2306         ret = srp_new_cm_id(target);
2307         if (ret)
2308                 goto err_free_ib;
2309
2310         ret = srp_connect_target(target);
2311         if (ret) {
2312                 shost_printk(KERN_ERR, target->scsi_host,
2313                              PFX "Connection failed\n");
2314                 goto err_cm_id;
2315         }
2316
2317         ret = srp_add_target(host, target);
2318         if (ret)
2319                 goto err_disconnect;
2320
2321         return count;
2322
2323 err_disconnect:
2324         srp_disconnect_target(target);
2325
2326 err_cm_id:
2327         ib_destroy_cm_id(target->cm_id);
2328
2329 err_free_ib:
2330         srp_free_target_ib(target);
2331
2332 err_free_mem:
2333         srp_free_req_data(target);
2334
2335 err:
2336         scsi_host_put(target_host);
2337
2338         return ret;
2339 }
2340
2341 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2342
2343 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2344                           char *buf)
2345 {
2346         struct srp_host *host = container_of(dev, struct srp_host, dev);
2347
2348         return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2349 }
2350
2351 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2352
2353 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2354                          char *buf)
2355 {
2356         struct srp_host *host = container_of(dev, struct srp_host, dev);
2357
2358         return sprintf(buf, "%d\n", host->port);
2359 }
2360
2361 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2362
2363 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2364 {
2365         struct srp_host *host;
2366
2367         host = kzalloc(sizeof *host, GFP_KERNEL);
2368         if (!host)
2369                 return NULL;
2370
2371         INIT_LIST_HEAD(&host->target_list);
2372         spin_lock_init(&host->target_lock);
2373         init_completion(&host->released);
2374         host->srp_dev = device;
2375         host->port = port;
2376
2377         host->dev.class = &srp_class;
2378         host->dev.parent = device->dev->dma_device;
2379         dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2380
2381         if (device_register(&host->dev))
2382                 goto free_host;
2383         if (device_create_file(&host->dev, &dev_attr_add_target))
2384                 goto err_class;
2385         if (device_create_file(&host->dev, &dev_attr_ibdev))
2386                 goto err_class;
2387         if (device_create_file(&host->dev, &dev_attr_port))
2388                 goto err_class;
2389
2390         return host;
2391
2392 err_class:
2393         device_unregister(&host->dev);
2394
2395 free_host:
2396         kfree(host);
2397
2398         return NULL;
2399 }
2400
2401 static void srp_add_one(struct ib_device *device)
2402 {
2403         struct srp_device *srp_dev;
2404         struct ib_device_attr *dev_attr;
2405         struct ib_fmr_pool_param fmr_param;
2406         struct srp_host *host;
2407         int max_pages_per_fmr, fmr_page_shift, s, e, p;
2408
2409         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2410         if (!dev_attr)
2411                 return;
2412
2413         if (ib_query_device(device, dev_attr)) {
2414                 pr_warn("Query device failed for %s\n", device->name);
2415                 goto free_attr;
2416         }
2417
2418         srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2419         if (!srp_dev)
2420                 goto free_attr;
2421
2422         /*
2423          * Use the smallest page size supported by the HCA, down to a
2424          * minimum of 4096 bytes. We're unlikely to build large sglists
2425          * out of smaller entries.
2426          */
2427         fmr_page_shift          = max(12, ffs(dev_attr->page_size_cap) - 1);
2428         srp_dev->fmr_page_size  = 1 << fmr_page_shift;
2429         srp_dev->fmr_page_mask  = ~((u64) srp_dev->fmr_page_size - 1);
2430         srp_dev->fmr_max_size   = srp_dev->fmr_page_size * SRP_FMR_SIZE;
2431
2432         INIT_LIST_HEAD(&srp_dev->dev_list);
2433
2434         srp_dev->dev = device;
2435         srp_dev->pd  = ib_alloc_pd(device);
2436         if (IS_ERR(srp_dev->pd))
2437                 goto free_dev;
2438
2439         srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2440                                     IB_ACCESS_LOCAL_WRITE |
2441                                     IB_ACCESS_REMOTE_READ |
2442                                     IB_ACCESS_REMOTE_WRITE);
2443         if (IS_ERR(srp_dev->mr))
2444                 goto err_pd;
2445
2446         for (max_pages_per_fmr = SRP_FMR_SIZE;
2447                         max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2448                         max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2449                 memset(&fmr_param, 0, sizeof fmr_param);
2450                 fmr_param.pool_size         = SRP_FMR_POOL_SIZE;
2451                 fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
2452                 fmr_param.cache             = 1;
2453                 fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2454                 fmr_param.page_shift        = fmr_page_shift;
2455                 fmr_param.access            = (IB_ACCESS_LOCAL_WRITE |
2456                                                IB_ACCESS_REMOTE_WRITE |
2457                                                IB_ACCESS_REMOTE_READ);
2458
2459                 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2460                 if (!IS_ERR(srp_dev->fmr_pool))
2461                         break;
2462         }
2463
2464         if (IS_ERR(srp_dev->fmr_pool))
2465                 srp_dev->fmr_pool = NULL;
2466
2467         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2468                 s = 0;
2469                 e = 0;
2470         } else {
2471                 s = 1;
2472                 e = device->phys_port_cnt;
2473         }
2474
2475         for (p = s; p <= e; ++p) {
2476                 host = srp_add_port(srp_dev, p);
2477                 if (host)
2478                         list_add_tail(&host->list, &srp_dev->dev_list);
2479         }
2480
2481         ib_set_client_data(device, &srp_client, srp_dev);
2482
2483         goto free_attr;
2484
2485 err_pd:
2486         ib_dealloc_pd(srp_dev->pd);
2487
2488 free_dev:
2489         kfree(srp_dev);
2490
2491 free_attr:
2492         kfree(dev_attr);
2493 }
2494
2495 static void srp_remove_one(struct ib_device *device)
2496 {
2497         struct srp_device *srp_dev;
2498         struct srp_host *host, *tmp_host;
2499         struct srp_target_port *target;
2500
2501         srp_dev = ib_get_client_data(device, &srp_client);
2502
2503         list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2504                 device_unregister(&host->dev);
2505                 /*
2506                  * Wait for the sysfs entry to go away, so that no new
2507                  * target ports can be created.
2508                  */
2509                 wait_for_completion(&host->released);
2510
2511                 /*
2512                  * Remove all target ports.
2513                  */
2514                 spin_lock(&host->target_lock);
2515                 list_for_each_entry(target, &host->target_list, list)
2516                         srp_queue_remove_work(target);
2517                 spin_unlock(&host->target_lock);
2518
2519                 /*
2520                  * Wait for target port removal tasks.
2521                  */
2522                 flush_workqueue(system_long_wq);
2523
2524                 kfree(host);
2525         }
2526
2527         if (srp_dev->fmr_pool)
2528                 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2529         ib_dereg_mr(srp_dev->mr);
2530         ib_dealloc_pd(srp_dev->pd);
2531
2532         kfree(srp_dev);
2533 }
2534
2535 static struct srp_function_template ib_srp_transport_functions = {
2536         .rport_delete            = srp_rport_delete,
2537 };
2538
2539 static int __init srp_init_module(void)
2540 {
2541         int ret;
2542
2543         BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2544
2545         if (srp_sg_tablesize) {
2546                 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2547                 if (!cmd_sg_entries)
2548                         cmd_sg_entries = srp_sg_tablesize;
2549         }
2550
2551         if (!cmd_sg_entries)
2552                 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2553
2554         if (cmd_sg_entries > 255) {
2555                 pr_warn("Clamping cmd_sg_entries to 255\n");
2556                 cmd_sg_entries = 255;
2557         }
2558
2559         if (!indirect_sg_entries)
2560                 indirect_sg_entries = cmd_sg_entries;
2561         else if (indirect_sg_entries < cmd_sg_entries) {
2562                 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
2563                         cmd_sg_entries);
2564                 indirect_sg_entries = cmd_sg_entries;
2565         }
2566
2567         ib_srp_transport_template =
2568                 srp_attach_transport(&ib_srp_transport_functions);
2569         if (!ib_srp_transport_template)
2570                 return -ENOMEM;
2571
2572         ret = class_register(&srp_class);
2573         if (ret) {
2574                 pr_err("couldn't register class infiniband_srp\n");
2575                 srp_release_transport(ib_srp_transport_template);
2576                 return ret;
2577         }
2578
2579         ib_sa_register_client(&srp_sa_client);
2580
2581         ret = ib_register_client(&srp_client);
2582         if (ret) {
2583                 pr_err("couldn't register IB client\n");
2584                 srp_release_transport(ib_srp_transport_template);
2585                 ib_sa_unregister_client(&srp_sa_client);
2586                 class_unregister(&srp_class);
2587                 return ret;
2588         }
2589
2590         return 0;
2591 }
2592
2593 static void __exit srp_cleanup_module(void)
2594 {
2595         ib_unregister_client(&srp_client);
2596         ib_sa_unregister_client(&srp_sa_client);
2597         class_unregister(&srp_class);
2598         srp_release_transport(ib_srp_transport_template);
2599 }
2600
2601 module_init(srp_init_module);
2602 module_exit(srp_cleanup_module);