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