drivers/infiniband/ulp/srp/ib_srp.c

   1 /*
   2  * Copyright (c) 2005 Cisco Systems.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #define pr_fmt(fmt) PFX fmt
  34
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/slab.h>
  38 #include <linux/err.h>
  39 #include <linux/string.h>
  40 #include <linux/parser.h>
  41 #include <linux/random.h>
  42 #include <linux/jiffies.h>
  43
  44 #include <linux/atomic.h>
  45
  46 #include <scsi/scsi.h>
  47 #include <scsi/scsi_device.h>
  48 #include <scsi/scsi_dbg.h>
  49 #include <scsi/srp.h>
  50 #include <scsi/scsi_transport_srp.h>
  51
  52 #include "ib_srp.h"
  53
  54 #define DRV_NAME        "ib_srp"
  55 #define PFX             DRV_NAME ": "
  56 #define DRV_VERSION     "0.2"
  57 #define DRV_RELDATE     "November 1, 2005"
  58
  59 MODULE_AUTHOR("Roland Dreier");
  60 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
  61                    "v" DRV_VERSION " (" DRV_RELDATE ")");
  62 MODULE_LICENSE("Dual BSD/GPL");
  63
  64 static unsigned int srp_sg_tablesize;
  65 static unsigned int cmd_sg_entries;
  66 static unsigned int indirect_sg_entries;
  67 static bool allow_ext_sg;
  68 static int topspin_workarounds = 1;
  69
  70 module_param(srp_sg_tablesize, uint, 0444);
  71 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
  72
  73 module_param(cmd_sg_entries, uint, 0444);
  74 MODULE_PARM_DESC(cmd_sg_entries,
  75                  "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
  76
  77 module_param(indirect_sg_entries, uint, 0444);
  78 MODULE_PARM_DESC(indirect_sg_entries,
  79                  "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
  80
  81 module_param(allow_ext_sg, bool, 0444);
  82 MODULE_PARM_DESC(allow_ext_sg,
  83                   "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
  84
  85 module_param(topspin_workarounds, int, 0444);
  86 MODULE_PARM_DESC(topspin_workarounds,
  87                  "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
  88
  89 static void srp_add_one(struct ib_device *device);
  90 static void srp_remove_one(struct ib_device *device);
  91 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
  92 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
  93 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
  94
  95 static struct scsi_transport_template *ib_srp_transport_template;
  96
  97 static struct ib_client srp_client = {
  98         .name   = "srp",
  99         .add    = srp_add_one,
 100         .remove = srp_remove_one
 101 };
 102
 103 static struct ib_sa_client srp_sa_client;
 104
 105 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
 106 {
 107         return (struct srp_target_port *) host->hostdata;
 108 }
 109
 110 static const char *srp_target_info(struct Scsi_Host *host)
 111 {
 112         return host_to_target(host)->target_name;
 113 }
 114
 115 static int srp_target_is_topspin(struct srp_target_port *target)
 116 {
 117         static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
 118         static const u8 cisco_oui[3]   = { 0x00, 0x1b, 0x0d };
 119
 120         return topspin_workarounds &&
 121                 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
 122                  !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
 123 }
 124
 125 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
 126                                    gfp_t gfp_mask,
 127                                    enum dma_data_direction direction)
 128 {
 129         struct srp_iu *iu;
 130
 131         iu = kmalloc(sizeof *iu, gfp_mask);
 132         if (!iu)
 133                 goto out;
 134
 135         iu->buf = kzalloc(size, gfp_mask);
 136         if (!iu->buf)
 137                 goto out_free_iu;
 138
 139         iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
 140                                     direction);
 141         if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
 142                 goto out_free_buf;
 143
 144         iu->size      = size;
 145         iu->direction = direction;
 146
 147         return iu;
 148
 149 out_free_buf:
 150         kfree(iu->buf);
 151 out_free_iu:
 152         kfree(iu);
 153 out:
 154         return NULL;
 155 }
 156
 157 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
 158 {
 159         if (!iu)
 160                 return;
 161
 162         ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
 163                             iu->direction);
 164         kfree(iu->buf);
 165         kfree(iu);
 166 }
 167
 168 static void srp_qp_event(struct ib_event *event, void *context)
 169 {
 170         pr_debug("QP event %d\n", event->event);
 171 }
 172
 173 static int srp_init_qp(struct srp_target_port *target,
 174                        struct ib_qp *qp)
 175 {
 176         struct ib_qp_attr *attr;
 177         int ret;
 178
 179         attr = kmalloc(sizeof *attr, GFP_KERNEL);
 180         if (!attr)
 181                 return -ENOMEM;
 182
 183         ret = ib_find_pkey(target->srp_host->srp_dev->dev,
 184                            target->srp_host->port,
 185                            be16_to_cpu(target->path.pkey),
 186                            &attr->pkey_index);
 187         if (ret)
 188                 goto out;
 189
 190         attr->qp_state        = IB_QPS_INIT;
 191         attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
 192                                     IB_ACCESS_REMOTE_WRITE);
 193         attr->port_num        = target->srp_host->port;
 194
 195         ret = ib_modify_qp(qp, attr,
 196                            IB_QP_STATE          |
 197                            IB_QP_PKEY_INDEX     |
 198                            IB_QP_ACCESS_FLAGS   |
 199                            IB_QP_PORT);
 200
 201 out:
 202         kfree(attr);
 203         return ret;
 204 }
 205
 206 static int srp_new_cm_id(struct srp_target_port *target)
 207 {
 208         struct ib_cm_id *new_cm_id;
 209
 210         new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
 211                                     srp_cm_handler, target);
 212         if (IS_ERR(new_cm_id))
 213                 return PTR_ERR(new_cm_id);
 214
 215         if (target->cm_id)
 216                 ib_destroy_cm_id(target->cm_id);
 217         target->cm_id = new_cm_id;
 218
 219         return 0;
 220 }
 221
 222 static int srp_create_target_ib(struct srp_target_port *target)
 223 {
 224         struct ib_qp_init_attr *init_attr;
 225         struct ib_cq *recv_cq, *send_cq;
 226         struct ib_qp *qp;
 227         int ret;
 228
 229         init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
 230         if (!init_attr)
 231                 return -ENOMEM;
 232
 233         recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
 234                                srp_recv_completion, NULL, target, SRP_RQ_SIZE, 0);
 235         if (IS_ERR(recv_cq)) {
 236                 ret = PTR_ERR(recv_cq);
 237                 goto err;
 238         }
 239
 240         send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
 241                                srp_send_completion, NULL, target, SRP_SQ_SIZE, 0);
 242         if (IS_ERR(send_cq)) {
 243                 ret = PTR_ERR(send_cq);
 244                 goto err_recv_cq;
 245         }
 246
 247         ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
 248
 249         init_attr->event_handler       = srp_qp_event;
 250         init_attr->cap.max_send_wr     = SRP_SQ_SIZE;
 251         init_attr->cap.max_recv_wr     = SRP_RQ_SIZE;
 252         init_attr->cap.max_recv_sge    = 1;
 253         init_attr->cap.max_send_sge    = 1;
 254         init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
 255         init_attr->qp_type             = IB_QPT_RC;
 256         init_attr->send_cq             = send_cq;
 257         init_attr->recv_cq             = recv_cq;
 258
 259         qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
 260         if (IS_ERR(qp)) {
 261                 ret = PTR_ERR(qp);
 262                 goto err_send_cq;
 263         }
 264
 265         ret = srp_init_qp(target, qp);
 266         if (ret)
 267                 goto err_qp;
 268
 269         if (target->qp)
 270                 ib_destroy_qp(target->qp);
 271         if (target->recv_cq)
 272                 ib_destroy_cq(target->recv_cq);
 273         if (target->send_cq)
 274                 ib_destroy_cq(target->send_cq);
 275
 276         target->qp = qp;
 277         target->recv_cq = recv_cq;
 278         target->send_cq = send_cq;
 279
 280         kfree(init_attr);
 281         return 0;
 282
 283 err_qp:
 284         ib_destroy_qp(qp);
 285
 286 err_send_cq:
 287         ib_destroy_cq(send_cq);
 288
 289 err_recv_cq:
 290         ib_destroy_cq(recv_cq);
 291
 292 err:
 293         kfree(init_attr);
 294         return ret;
 295 }
 296
 297 static void srp_free_target_ib(struct srp_target_port *target)
 298 {
 299         int i;
 300
 301         ib_destroy_qp(target->qp);
 302         ib_destroy_cq(target->send_cq);
 303         ib_destroy_cq(target->recv_cq);
 304
 305         target->qp = NULL;
 306         target->send_cq = target->recv_cq = NULL;
 307
 308         for (i = 0; i < SRP_RQ_SIZE; ++i)
 309                 srp_free_iu(target->srp_host, target->rx_ring[i]);
 310         for (i = 0; i < SRP_SQ_SIZE; ++i)
 311                 srp_free_iu(target->srp_host, target->tx_ring[i]);
 312 }
 313
 314 static void srp_path_rec_completion(int status,
 315                                     struct ib_sa_path_rec *pathrec,
 316                                     void *target_ptr)
 317 {
 318         struct srp_target_port *target = target_ptr;
 319
 320         target->status = status;
 321         if (status)
 322                 shost_printk(KERN_ERR, target->scsi_host,
 323                              PFX "Got failed path rec status %d\n", status);
 324         else
 325                 target->path = *pathrec;
 326         complete(&target->done);
 327 }
 328
 329 static int srp_lookup_path(struct srp_target_port *target)
 330 {
 331         target->path.numb_path = 1;
 332
 333         init_completion(&target->done);
 334
 335         target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
 336                                                    target->srp_host->srp_dev->dev,
 337                                                    target->srp_host->port,
 338                                                    &target->path,
 339                                                    IB_SA_PATH_REC_SERVICE_ID    |
 340                                                    IB_SA_PATH_REC_DGID          |
 341                                                    IB_SA_PATH_REC_SGID          |
 342                                                    IB_SA_PATH_REC_NUMB_PATH     |
 343                                                    IB_SA_PATH_REC_PKEY,
 344                                                    SRP_PATH_REC_TIMEOUT_MS,
 345                                                    GFP_KERNEL,
 346                                                    srp_path_rec_completion,
 347                                                    target, &target->path_query);
 348         if (target->path_query_id < 0)
 349                 return target->path_query_id;
 350
 351         wait_for_completion(&target->done);
 352
 353         if (target->status < 0)
 354                 shost_printk(KERN_WARNING, target->scsi_host,
 355                              PFX "Path record query failed\n");
 356
 357         return target->status;
 358 }
 359
 360 static int srp_send_req(struct srp_target_port *target)
 361 {
 362         struct {
 363                 struct ib_cm_req_param param;
 364                 struct srp_login_req   priv;
 365         } *req = NULL;
 366         int status;
 367
 368         req = kzalloc(sizeof *req, GFP_KERNEL);
 369         if (!req)
 370                 return -ENOMEM;
 371
 372         req->param.primary_path               = &target->path;
 373         req->param.alternate_path             = NULL;
 374         req->param.service_id                 = target->service_id;
 375         req->param.qp_num                     = target->qp->qp_num;
 376         req->param.qp_type                    = target->qp->qp_type;
 377         req->param.private_data               = &req->priv;
 378         req->param.private_data_len           = sizeof req->priv;
 379         req->param.flow_control               = 1;
 380
 381         get_random_bytes(&req->param.starting_psn, 4);
 382         req->param.starting_psn              &= 0xffffff;
 383
 384         /*
 385          * Pick some arbitrary defaults here; we could make these
 386          * module parameters if anyone cared about setting them.
 387          */
 388         req->param.responder_resources        = 4;
 389         req->param.remote_cm_response_timeout = 20;
 390         req->param.local_cm_response_timeout  = 20;
 391         req->param.retry_count                = 7;
 392         req->param.rnr_retry_count            = 7;
 393         req->param.max_cm_retries             = 15;
 394
 395         req->priv.opcode        = SRP_LOGIN_REQ;
 396         req->priv.tag           = 0;
 397         req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
 398         req->priv.req_buf_fmt   = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
 399                                               SRP_BUF_FORMAT_INDIRECT);
 400         /*
 401          * In the published SRP specification (draft rev. 16a), the
 402          * port identifier format is 8 bytes of ID extension followed
 403          * by 8 bytes of GUID.  Older drafts put the two halves in the
 404          * opposite order, so that the GUID comes first.
 405          *
 406          * Targets conforming to these obsolete drafts can be
 407          * recognized by the I/O Class they report.
 408          */
 409         if (target->io_class == SRP_REV10_IB_IO_CLASS) {
 410                 memcpy(req->priv.initiator_port_id,
 411                        &target->path.sgid.global.interface_id, 8);
 412                 memcpy(req->priv.initiator_port_id + 8,
 413                        &target->initiator_ext, 8);
 414                 memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
 415                 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
 416         } else {
 417                 memcpy(req->priv.initiator_port_id,
 418                        &target->initiator_ext, 8);
 419                 memcpy(req->priv.initiator_port_id + 8,
 420                        &target->path.sgid.global.interface_id, 8);
 421                 memcpy(req->priv.target_port_id,     &target->id_ext, 8);
 422                 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
 423         }
 424
 425         /*
 426          * Topspin/Cisco SRP targets will reject our login unless we
 427          * zero out the first 8 bytes of our initiator port ID and set
 428          * the second 8 bytes to the local node GUID.
 429          */
 430         if (srp_target_is_topspin(target)) {
 431                 shost_printk(KERN_DEBUG, target->scsi_host,
 432                              PFX "Topspin/Cisco initiator port ID workaround "
 433                              "activated for target GUID %016llx\n",
 434                              (unsigned long long) be64_to_cpu(target->ioc_guid));
 435                 memset(req->priv.initiator_port_id, 0, 8);
 436                 memcpy(req->priv.initiator_port_id + 8,
 437                        &target->srp_host->srp_dev->dev->node_guid, 8);
 438         }
 439
 440         status = ib_send_cm_req(target->cm_id, &req->param);
 441
 442         kfree(req);
 443
 444         return status;
 445 }
 446
 447 static bool srp_queue_remove_work(struct srp_target_port *target)
 448 {
 449         bool changed = false;
 450
 451         spin_lock_irq(&target->lock);
 452         if (target->state != SRP_TARGET_REMOVED) {
 453                 target->state = SRP_TARGET_REMOVED;
 454                 changed = true;
 455         }
 456         spin_unlock_irq(&target->lock);
 457
 458         if (changed)
 459                 queue_work(system_long_wq, &target->remove_work);
 460
 461         return changed;
 462 }
 463
 464 static bool srp_change_conn_state(struct srp_target_port *target,
 465                                   bool connected)
 466 {
 467         bool changed = false;
 468
 469         spin_lock_irq(&target->lock);
 470         if (target->connected != connected) {
 471                 target->connected = connected;
 472                 changed = true;
 473         }
 474         spin_unlock_irq(&target->lock);
 475
 476         return changed;
 477 }
 478
 479 static void srp_disconnect_target(struct srp_target_port *target)
 480 {
 481         if (srp_change_conn_state(target, false)) {
 482                 /* XXX should send SRP_I_LOGOUT request */
 483
 484                 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
 485                         shost_printk(KERN_DEBUG, target->scsi_host,
 486                                      PFX "Sending CM DREQ failed\n");
 487                 }
 488         }
 489 }
 490
 491 static void srp_free_req_data(struct srp_target_port *target)
 492 {
 493         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
 494         struct srp_request *req;
 495         int i;
 496
 497         for (i = 0, req = target->req_ring; i < SRP_CMD_SQ_SIZE; ++i, ++req) {
 498                 kfree(req->fmr_list);
 499                 kfree(req->map_page);
 500                 if (req->indirect_dma_addr) {
 501                         ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
 502                                             target->indirect_size,
 503                                             DMA_TO_DEVICE);
 504                 }
 505                 kfree(req->indirect_desc);
 506         }
 507 }
 508
 509 /**
 510  * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
 511  * @shost: SCSI host whose attributes to remove from sysfs.
 512  *
 513  * Note: Any attributes defined in the host template and that did not exist
 514  * before invocation of this function will be ignored.
 515  */
 516 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
 517 {
 518         struct device_attribute **attr;
 519
 520         for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
 521                 device_remove_file(&shost->shost_dev, *attr);
 522 }
 523
 524 static void srp_remove_target(struct srp_target_port *target)
 525 {
 526         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 527
 528         srp_del_scsi_host_attr(target->scsi_host);
 529         srp_remove_host(target->scsi_host);
 530         scsi_remove_host(target->scsi_host);
 531         srp_disconnect_target(target);
 532         ib_destroy_cm_id(target->cm_id);
 533         srp_free_target_ib(target);
 534         srp_free_req_data(target);
 535         scsi_host_put(target->scsi_host);
 536 }
 537
 538 static void srp_remove_work(struct work_struct *work)
 539 {
 540         struct srp_target_port *target =
 541                 container_of(work, struct srp_target_port, remove_work);
 542
 543         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 544
 545         spin_lock(&target->srp_host->target_lock);
 546         list_del(&target->list);
 547         spin_unlock(&target->srp_host->target_lock);
 548
 549         srp_remove_target(target);
 550 }
 551
 552 static void srp_rport_delete(struct srp_rport *rport)
 553 {
 554         struct srp_target_port *target = rport->lld_data;
 555
 556         srp_queue_remove_work(target);
 557 }
 558
 559 static int srp_connect_target(struct srp_target_port *target)
 560 {
 561         int retries = 3;
 562         int ret;
 563
 564         WARN_ON_ONCE(target->connected);
 565
 566         target->qp_in_error = false;
 567
 568         ret = srp_lookup_path(target);
 569         if (ret)
 570                 return ret;
 571
 572         while (1) {
 573                 init_completion(&target->done);
 574                 ret = srp_send_req(target);
 575                 if (ret)
 576                         return ret;
 577                 wait_for_completion(&target->done);
 578
 579                 /*
 580                  * The CM event handling code will set status to
 581                  * SRP_PORT_REDIRECT if we get a port redirect REJ
 582                  * back, or SRP_DLID_REDIRECT if we get a lid/qp
 583                  * redirect REJ back.
 584                  */
 585                 switch (target->status) {
 586                 case 0:
 587                         srp_change_conn_state(target, true);
 588                         return 0;
 589
 590                 case SRP_PORT_REDIRECT:
 591                         ret = srp_lookup_path(target);
 592                         if (ret)
 593                                 return ret;
 594                         break;
 595
 596                 case SRP_DLID_REDIRECT:
 597                         break;
 598
 599                 case SRP_STALE_CONN:
 600                         /* Our current CM id was stale, and is now in timewait.
 601                          * Try to reconnect with a new one.
 602                          */
 603                         if (!retries-- || srp_new_cm_id(target)) {
 604                                 shost_printk(KERN_ERR, target->scsi_host, PFX
 605                                              "giving up on stale connection\n");
 606                                 target->status = -ECONNRESET;
 607                                 return target->status;
 608                         }
 609
 610                         shost_printk(KERN_ERR, target->scsi_host, PFX
 611                                      "retrying stale connection\n");
 612                         break;
 613
 614                 default:
 615                         return target->status;
 616                 }
 617         }
 618 }
 619
 620 static void srp_unmap_data(struct scsi_cmnd *scmnd,
 621                            struct srp_target_port *target,
 622                            struct srp_request *req)
 623 {
 624         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
 625         struct ib_pool_fmr **pfmr;
 626
 627         if (!scsi_sglist(scmnd) ||
 628             (scmnd->sc_data_direction != DMA_TO_DEVICE &&
 629              scmnd->sc_data_direction != DMA_FROM_DEVICE))
 630                 return;
 631
 632         pfmr = req->fmr_list;
 633         while (req->nfmr--)
 634                 ib_fmr_pool_unmap(*pfmr++);
 635
 636         ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
 637                         scmnd->sc_data_direction);
 638 }
 639
 640 /**
 641  * srp_claim_req - Take ownership of the scmnd associated with a request.
 642  * @target: SRP target port.
 643  * @req: SRP request.
 644  * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
 645  *         ownership of @req->scmnd if it equals @scmnd.
 646  *
 647  * Return value:
 648  * Either NULL or a pointer to the SCSI command the caller became owner of.
 649  */
 650 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
 651                                        struct srp_request *req,
 652                                        struct scsi_cmnd *scmnd)
 653 {
 654         unsigned long flags;
 655
 656         spin_lock_irqsave(&target->lock, flags);
 657         if (!scmnd) {
 658                 scmnd = req->scmnd;
 659                 req->scmnd = NULL;
 660         } else if (req->scmnd == scmnd) {
 661                 req->scmnd = NULL;
 662         } else {
 663                 scmnd = NULL;
 664         }
 665         spin_unlock_irqrestore(&target->lock, flags);
 666
 667         return scmnd;
 668 }
 669
 670 /**
 671  * srp_free_req() - Unmap data and add request to the free request list.
 672  */
 673 static void srp_free_req(struct srp_target_port *target,
 674                          struct srp_request *req, struct scsi_cmnd *scmnd,
 675                          s32 req_lim_delta)
 676 {
 677         unsigned long flags;
 678
 679         srp_unmap_data(scmnd, target, req);
 680
 681         spin_lock_irqsave(&target->lock, flags);
 682         target->req_lim += req_lim_delta;
 683         list_add_tail(&req->list, &target->free_reqs);
 684         spin_unlock_irqrestore(&target->lock, flags);
 685 }
 686
 687 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
 688 {
 689         struct scsi_cmnd *scmnd = srp_claim_req(target, req, NULL);
 690
 691         if (scmnd) {
 692                 srp_free_req(target, req, scmnd, 0);
 693                 scmnd->result = DID_RESET << 16;
 694                 scmnd->scsi_done(scmnd);
 695         }
 696 }
 697
 698 static int srp_reconnect_target(struct srp_target_port *target)
 699 {
 700         struct Scsi_Host *shost = target->scsi_host;
 701         int i, ret;
 702
 703         scsi_target_block(&shost->shost_gendev);
 704
 705         srp_disconnect_target(target);
 706         /*
 707          * Now get a new local CM ID so that we avoid confusing the target in
 708          * case things are really fouled up. Doing so also ensures that all CM
 709          * callbacks will have finished before a new QP is allocated.
 710          */
 711         ret = srp_new_cm_id(target);
 712         /*
 713          * Whether or not creating a new CM ID succeeded, create a new
 714          * QP. This guarantees that all completion callback function
 715          * invocations have finished before request resetting starts.
 716          */
 717         if (ret == 0)
 718                 ret = srp_create_target_ib(target);
 719         else
 720                 srp_create_target_ib(target);
 721
 722         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
 723                 struct srp_request *req = &target->req_ring[i];
 724                 if (req->scmnd)
 725                         srp_reset_req(target, req);
 726         }
 727
 728         INIT_LIST_HEAD(&target->free_tx);
 729         for (i = 0; i < SRP_SQ_SIZE; ++i)
 730                 list_add(&target->tx_ring[i]->list, &target->free_tx);
 731
 732         if (ret == 0)
 733                 ret = srp_connect_target(target);
 734
 735         scsi_target_unblock(&shost->shost_gendev, ret == 0 ? SDEV_RUNNING :
 736                             SDEV_TRANSPORT_OFFLINE);
 737         target->transport_offline = !!ret;
 738
 739         if (ret)
 740                 goto err;
 741
 742         shost_printk(KERN_INFO, target->scsi_host, PFX "reconnect succeeded\n");
 743
 744         return ret;
 745
 746 err:
 747         shost_printk(KERN_ERR, target->scsi_host,
 748                      PFX "reconnect failed (%d), removing target port.\n", ret);
 749
 750         /*
 751          * We couldn't reconnect, so kill our target port off.
 752          * However, we have to defer the real removal because we
 753          * are in the context of the SCSI error handler now, which
 754          * will deadlock if we call scsi_remove_host().
 755          */
 756         srp_queue_remove_work(target);
 757
 758         return ret;
 759 }
 760
 761 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
 762                          unsigned int dma_len, u32 rkey)
 763 {
 764         struct srp_direct_buf *desc = state->desc;
 765
 766         desc->va = cpu_to_be64(dma_addr);
 767         desc->key = cpu_to_be32(rkey);
 768         desc->len = cpu_to_be32(dma_len);
 769
 770         state->total_len += dma_len;
 771         state->desc++;
 772         state->ndesc++;
 773 }
 774
 775 static int srp_map_finish_fmr(struct srp_map_state *state,
 776                               struct srp_target_port *target)
 777 {
 778         struct srp_device *dev = target->srp_host->srp_dev;
 779         struct ib_pool_fmr *fmr;
 780         u64 io_addr = 0;
 781
 782         if (!state->npages)
 783                 return 0;
 784
 785         if (state->npages == 1) {
 786                 srp_map_desc(state, state->base_dma_addr, state->fmr_len,
 787                              target->rkey);
 788                 state->npages = state->fmr_len = 0;
 789                 return 0;
 790         }
 791
 792         fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
 793                                    state->npages, io_addr);
 794         if (IS_ERR(fmr))
 795                 return PTR_ERR(fmr);
 796
 797         *state->next_fmr++ = fmr;
 798         state->nfmr++;
 799
 800         srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
 801         state->npages = state->fmr_len = 0;
 802         return 0;
 803 }
 804
 805 static void srp_map_update_start(struct srp_map_state *state,
 806                                  struct scatterlist *sg, int sg_index,
 807                                  dma_addr_t dma_addr)
 808 {
 809         state->unmapped_sg = sg;
 810         state->unmapped_index = sg_index;
 811         state->unmapped_addr = dma_addr;
 812 }
 813
 814 static int srp_map_sg_entry(struct srp_map_state *state,
 815                             struct srp_target_port *target,
 816                             struct scatterlist *sg, int sg_index,
 817                             int use_fmr)
 818 {
 819         struct srp_device *dev = target->srp_host->srp_dev;
 820         struct ib_device *ibdev = dev->dev;
 821         dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
 822         unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
 823         unsigned int len;
 824         int ret;
 825
 826         if (!dma_len)
 827                 return 0;
 828
 829         if (use_fmr == SRP_MAP_NO_FMR) {
 830                 /* Once we're in direct map mode for a request, we don't
 831                  * go back to FMR mode, so no need to update anything
 832                  * other than the descriptor.
 833                  */
 834                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
 835                 return 0;
 836         }
 837
 838         /* If we start at an offset into the FMR page, don't merge into
 839          * the current FMR. Finish it out, and use the kernel's MR for this
 840          * sg entry. This is to avoid potential bugs on some SRP targets
 841          * that were never quite defined, but went away when the initiator
 842          * avoided using FMR on such page fragments.
 843          */
 844         if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
 845                 ret = srp_map_finish_fmr(state, target);
 846                 if (ret)
 847                         return ret;
 848
 849                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
 850                 srp_map_update_start(state, NULL, 0, 0);
 851                 return 0;
 852         }
 853
 854         /* If this is the first sg to go into the FMR, save our position.
 855          * We need to know the first unmapped entry, its index, and the
 856          * first unmapped address within that entry to be able to restart
 857          * mapping after an error.
 858          */
 859         if (!state->unmapped_sg)
 860                 srp_map_update_start(state, sg, sg_index, dma_addr);
 861
 862         while (dma_len) {
 863                 if (state->npages == SRP_FMR_SIZE) {
 864                         ret = srp_map_finish_fmr(state, target);
 865                         if (ret)
 866                                 return ret;
 867
 868                         srp_map_update_start(state, sg, sg_index, dma_addr);
 869                 }
 870
 871                 len = min_t(unsigned int, dma_len, dev->fmr_page_size);
 872
 873                 if (!state->npages)
 874                         state->base_dma_addr = dma_addr;
 875                 state->pages[state->npages++] = dma_addr;
 876                 state->fmr_len += len;
 877                 dma_addr += len;
 878                 dma_len -= len;
 879         }
 880
 881         /* If the last entry of the FMR wasn't a full page, then we need to
 882          * close it out and start a new one -- we can only merge at page
 883          * boundries.
 884          */
 885         ret = 0;
 886         if (len != dev->fmr_page_size) {
 887                 ret = srp_map_finish_fmr(state, target);
 888                 if (!ret)
 889                         srp_map_update_start(state, NULL, 0, 0);
 890         }
 891         return ret;
 892 }
 893
 894 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
 895                         struct srp_request *req)
 896 {
 897         struct scatterlist *scat, *sg;
 898         struct srp_cmd *cmd = req->cmd->buf;
 899         int i, len, nents, count, use_fmr;
 900         struct srp_device *dev;
 901         struct ib_device *ibdev;
 902         struct srp_map_state state;
 903         struct srp_indirect_buf *indirect_hdr;
 904         u32 table_len;
 905         u8 fmt;
 906
 907         if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
 908                 return sizeof (struct srp_cmd);
 909
 910         if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
 911             scmnd->sc_data_direction != DMA_TO_DEVICE) {
 912                 shost_printk(KERN_WARNING, target->scsi_host,
 913                              PFX "Unhandled data direction %d\n",
 914                              scmnd->sc_data_direction);
 915                 return -EINVAL;
 916         }
 917
 918         nents = scsi_sg_count(scmnd);
 919         scat  = scsi_sglist(scmnd);
 920
 921         dev = target->srp_host->srp_dev;
 922         ibdev = dev->dev;
 923
 924         count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
 925         if (unlikely(count == 0))
 926                 return -EIO;
 927
 928         fmt = SRP_DATA_DESC_DIRECT;
 929         len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
 930
 931         if (count == 1) {
 932                 /*
 933                  * The midlayer only generated a single gather/scatter
 934                  * entry, or DMA mapping coalesced everything to a
 935                  * single entry.  So a direct descriptor along with
 936                  * the DMA MR suffices.
 937                  */
 938                 struct srp_direct_buf *buf = (void *) cmd->add_data;
 939
 940                 buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
 941                 buf->key = cpu_to_be32(target->rkey);
 942                 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
 943
 944                 req->nfmr = 0;
 945                 goto map_complete;
 946         }
 947
 948         /* We have more than one scatter/gather entry, so build our indirect
 949          * descriptor table, trying to merge as many entries with FMR as we
 950          * can.
 951          */
 952         indirect_hdr = (void *) cmd->add_data;
 953
 954         ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
 955                                    target->indirect_size, DMA_TO_DEVICE);
 956
 957         memset(&state, 0, sizeof(state));
 958         state.desc      = req->indirect_desc;
 959         state.pages     = req->map_page;
 960         state.next_fmr  = req->fmr_list;
 961
 962         use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
 963
 964         for_each_sg(scat, sg, count, i) {
 965                 if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
 966                         /* FMR mapping failed, so backtrack to the first
 967                          * unmapped entry and continue on without using FMR.
 968                          */
 969                         dma_addr_t dma_addr;
 970                         unsigned int dma_len;
 971
 972 backtrack:
 973                         sg = state.unmapped_sg;
 974                         i = state.unmapped_index;
 975
 976                         dma_addr = ib_sg_dma_address(ibdev, sg);
 977                         dma_len = ib_sg_dma_len(ibdev, sg);
 978                         dma_len -= (state.unmapped_addr - dma_addr);
 979                         dma_addr = state.unmapped_addr;
 980                         use_fmr = SRP_MAP_NO_FMR;
 981                         srp_map_desc(&state, dma_addr, dma_len, target->rkey);
 982                 }
 983         }
 984
 985         if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
 986                 goto backtrack;
 987
 988         /* We've mapped the request, now pull as much of the indirect
 989          * descriptor table as we can into the command buffer. If this
 990          * target is not using an external indirect table, we are
 991          * guaranteed to fit into the command, as the SCSI layer won't
 992          * give us more S/G entries than we allow.
 993          */
 994         req->nfmr = state.nfmr;
 995         if (state.ndesc == 1) {
 996                 /* FMR mapping was able to collapse this to one entry,
 997                  * so use a direct descriptor.
 998                  */
 999                 struct srp_direct_buf *buf = (void *) cmd->add_data;
1000
1001                 *buf = req->indirect_desc[0];
1002                 goto map_complete;
1003         }
1004
1005         if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1006                                                 !target->allow_ext_sg)) {
1007                 shost_printk(KERN_ERR, target->scsi_host,
1008                              "Could not fit S/G list into SRP_CMD\n");
1009                 return -EIO;
1010         }
1011
1012         count = min(state.ndesc, target->cmd_sg_cnt);
1013         table_len = state.ndesc * sizeof (struct srp_direct_buf);
1014
1015         fmt = SRP_DATA_DESC_INDIRECT;
1016         len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1017         len += count * sizeof (struct srp_direct_buf);
1018
1019         memcpy(indirect_hdr->desc_list, req->indirect_desc,
1020                count * sizeof (struct srp_direct_buf));
1021
1022         indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1023         indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1024         indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1025         indirect_hdr->len = cpu_to_be32(state.total_len);
1026
1027         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1028                 cmd->data_out_desc_cnt = count;
1029         else
1030                 cmd->data_in_desc_cnt = count;
1031
1032         ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1033                                       DMA_TO_DEVICE);
1034
1035 map_complete:
1036         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1037                 cmd->buf_fmt = fmt << 4;
1038         else
1039                 cmd->buf_fmt = fmt;
1040
1041         return len;
1042 }
1043
1044 /*
1045  * Return an IU and possible credit to the free pool
1046  */
1047 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1048                           enum srp_iu_type iu_type)
1049 {
1050         unsigned long flags;
1051
1052         spin_lock_irqsave(&target->lock, flags);
1053         list_add(&iu->list, &target->free_tx);
1054         if (iu_type != SRP_IU_RSP)
1055                 ++target->req_lim;
1056         spin_unlock_irqrestore(&target->lock, flags);
1057 }
1058
1059 /*
1060  * Must be called with target->lock held to protect req_lim and free_tx.
1061  * If IU is not sent, it must be returned using srp_put_tx_iu().
1062  *
1063  * Note:
1064  * An upper limit for the number of allocated information units for each
1065  * request type is:
1066  * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1067  *   more than Scsi_Host.can_queue requests.
1068  * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1069  * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1070  *   one unanswered SRP request to an initiator.
1071  */
1072 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1073                                       enum srp_iu_type iu_type)
1074 {
1075         s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1076         struct srp_iu *iu;
1077
1078         srp_send_completion(target->send_cq, target);
1079
1080         if (list_empty(&target->free_tx))
1081                 return NULL;
1082
1083         /* Initiator responses to target requests do not consume credits */
1084         if (iu_type != SRP_IU_RSP) {
1085                 if (target->req_lim <= rsv) {
1086                         ++target->zero_req_lim;
1087                         return NULL;
1088                 }
1089
1090                 --target->req_lim;
1091         }
1092
1093         iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1094         list_del(&iu->list);
1095         return iu;
1096 }
1097
1098 static int srp_post_send(struct srp_target_port *target,
1099                          struct srp_iu *iu, int len)
1100 {
1101         struct ib_sge list;
1102         struct ib_send_wr wr, *bad_wr;
1103
1104         list.addr   = iu->dma;
1105         list.length = len;
1106         list.lkey   = target->lkey;
1107
1108         wr.next       = NULL;
1109         wr.wr_id      = (uintptr_t) iu;
1110         wr.sg_list    = &list;
1111         wr.num_sge    = 1;
1112         wr.opcode     = IB_WR_SEND;
1113         wr.send_flags = IB_SEND_SIGNALED;
1114
1115         return ib_post_send(target->qp, &wr, &bad_wr);
1116 }
1117
1118 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1119 {
1120         struct ib_recv_wr wr, *bad_wr;
1121         struct ib_sge list;
1122
1123         list.addr   = iu->dma;
1124         list.length = iu->size;
1125         list.lkey   = target->lkey;
1126
1127         wr.next     = NULL;
1128         wr.wr_id    = (uintptr_t) iu;
1129         wr.sg_list  = &list;
1130         wr.num_sge  = 1;
1131
1132         return ib_post_recv(target->qp, &wr, &bad_wr);
1133 }
1134
1135 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1136 {
1137         struct srp_request *req;
1138         struct scsi_cmnd *scmnd;
1139         unsigned long flags;
1140
1141         if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1142                 spin_lock_irqsave(&target->lock, flags);
1143                 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1144                 spin_unlock_irqrestore(&target->lock, flags);
1145
1146                 target->tsk_mgmt_status = -1;
1147                 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1148                         target->tsk_mgmt_status = rsp->data[3];
1149                 complete(&target->tsk_mgmt_done);
1150         } else {
1151                 req = &target->req_ring[rsp->tag];
1152                 scmnd = srp_claim_req(target, req, NULL);
1153                 if (!scmnd) {
1154                         shost_printk(KERN_ERR, target->scsi_host,
1155                                      "Null scmnd for RSP w/tag %016llx\n",
1156                                      (unsigned long long) rsp->tag);
1157
1158                         spin_lock_irqsave(&target->lock, flags);
1159                         target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1160                         spin_unlock_irqrestore(&target->lock, flags);
1161
1162                         return;
1163                 }
1164                 scmnd->result = rsp->status;
1165
1166                 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1167                         memcpy(scmnd->sense_buffer, rsp->data +
1168                                be32_to_cpu(rsp->resp_data_len),
1169                                min_t(int, be32_to_cpu(rsp->sense_data_len),
1170                                      SCSI_SENSE_BUFFERSIZE));
1171                 }
1172
1173                 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1174                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1175                 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1176                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1177
1178                 srp_free_req(target, req, scmnd,
1179                              be32_to_cpu(rsp->req_lim_delta));
1180
1181                 scmnd->host_scribble = NULL;
1182                 scmnd->scsi_done(scmnd);
1183         }
1184 }
1185
1186 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1187                                void *rsp, int len)
1188 {
1189         struct ib_device *dev = target->srp_host->srp_dev->dev;
1190         unsigned long flags;
1191         struct srp_iu *iu;
1192         int err;
1193
1194         spin_lock_irqsave(&target->lock, flags);
1195         target->req_lim += req_delta;
1196         iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1197         spin_unlock_irqrestore(&target->lock, flags);
1198
1199         if (!iu) {
1200                 shost_printk(KERN_ERR, target->scsi_host, PFX
1201                              "no IU available to send response\n");
1202                 return 1;
1203         }
1204
1205         ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1206         memcpy(iu->buf, rsp, len);
1207         ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1208
1209         err = srp_post_send(target, iu, len);
1210         if (err) {
1211                 shost_printk(KERN_ERR, target->scsi_host, PFX
1212                              "unable to post response: %d\n", err);
1213                 srp_put_tx_iu(target, iu, SRP_IU_RSP);
1214         }
1215
1216         return err;
1217 }
1218
1219 static void srp_process_cred_req(struct srp_target_port *target,
1220                                  struct srp_cred_req *req)
1221 {
1222         struct srp_cred_rsp rsp = {
1223                 .opcode = SRP_CRED_RSP,
1224                 .tag = req->tag,
1225         };
1226         s32 delta = be32_to_cpu(req->req_lim_delta);
1227
1228         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1229                 shost_printk(KERN_ERR, target->scsi_host, PFX
1230                              "problems processing SRP_CRED_REQ\n");
1231 }
1232
1233 static void srp_process_aer_req(struct srp_target_port *target,
1234                                 struct srp_aer_req *req)
1235 {
1236         struct srp_aer_rsp rsp = {
1237                 .opcode = SRP_AER_RSP,
1238                 .tag = req->tag,
1239         };
1240         s32 delta = be32_to_cpu(req->req_lim_delta);
1241
1242         shost_printk(KERN_ERR, target->scsi_host, PFX
1243                      "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1244
1245         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1246                 shost_printk(KERN_ERR, target->scsi_host, PFX
1247                              "problems processing SRP_AER_REQ\n");
1248 }
1249
1250 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1251 {
1252         struct ib_device *dev = target->srp_host->srp_dev->dev;
1253         struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1254         int res;
1255         u8 opcode;
1256
1257         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1258                                    DMA_FROM_DEVICE);
1259
1260         opcode = *(u8 *) iu->buf;
1261
1262         if (0) {
1263                 shost_printk(KERN_ERR, target->scsi_host,
1264                              PFX "recv completion, opcode 0x%02x\n", opcode);
1265                 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1266                                iu->buf, wc->byte_len, true);
1267         }
1268
1269         switch (opcode) {
1270         case SRP_RSP:
1271                 srp_process_rsp(target, iu->buf);
1272                 break;
1273
1274         case SRP_CRED_REQ:
1275                 srp_process_cred_req(target, iu->buf);
1276                 break;
1277
1278         case SRP_AER_REQ:
1279                 srp_process_aer_req(target, iu->buf);
1280                 break;
1281
1282         case SRP_T_LOGOUT:
1283                 /* XXX Handle target logout */
1284                 shost_printk(KERN_WARNING, target->scsi_host,
1285                              PFX "Got target logout request\n");
1286                 break;
1287
1288         default:
1289                 shost_printk(KERN_WARNING, target->scsi_host,
1290                              PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1291                 break;
1292         }
1293
1294         ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1295                                       DMA_FROM_DEVICE);
1296
1297         res = srp_post_recv(target, iu);
1298         if (res != 0)
1299                 shost_printk(KERN_ERR, target->scsi_host,
1300                              PFX "Recv failed with error code %d\n", res);
1301 }
1302
1303 static void srp_handle_qp_err(enum ib_wc_status wc_status, bool send_err,
1304                               struct srp_target_port *target)
1305 {
1306         if (target->connected && !target->qp_in_error) {
1307                 shost_printk(KERN_ERR, target->scsi_host,
1308                              PFX "failed %s status %d\n",
1309                              send_err ? "send" : "receive",
1310                              wc_status);
1311         }
1312         target->qp_in_error = true;
1313 }
1314
1315 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1316 {
1317         struct srp_target_port *target = target_ptr;
1318         struct ib_wc wc;
1319
1320         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1321         while (ib_poll_cq(cq, 1, &wc) > 0) {
1322                 if (likely(wc.status == IB_WC_SUCCESS)) {
1323                         srp_handle_recv(target, &wc);
1324                 } else {
1325                         srp_handle_qp_err(wc.status, false, target);
1326                 }
1327         }
1328 }
1329
1330 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1331 {
1332         struct srp_target_port *target = target_ptr;
1333         struct ib_wc wc;
1334         struct srp_iu *iu;
1335
1336         while (ib_poll_cq(cq, 1, &wc) > 0) {
1337                 if (likely(wc.status == IB_WC_SUCCESS)) {
1338                         iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1339                         list_add(&iu->list, &target->free_tx);
1340                 } else {
1341                         srp_handle_qp_err(wc.status, true, target);
1342                 }
1343         }
1344 }
1345
1346 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1347 {
1348         struct srp_target_port *target = host_to_target(shost);
1349         struct srp_request *req;
1350         struct srp_iu *iu;
1351         struct srp_cmd *cmd;
1352         struct ib_device *dev;
1353         unsigned long flags;
1354         int len;
1355
1356         if (unlikely(target->transport_offline)) {
1357                 scmnd->result = DID_NO_CONNECT << 16;
1358                 scmnd->scsi_done(scmnd);
1359                 return 0;
1360         }
1361
1362         spin_lock_irqsave(&target->lock, flags);
1363         iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1364         if (!iu)
1365                 goto err_unlock;
1366
1367         req = list_first_entry(&target->free_reqs, struct srp_request, list);
1368         list_del(&req->list);
1369         spin_unlock_irqrestore(&target->lock, flags);
1370
1371         dev = target->srp_host->srp_dev->dev;
1372         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1373                                    DMA_TO_DEVICE);
1374
1375         scmnd->result        = 0;
1376         scmnd->host_scribble = (void *) req;
1377
1378         cmd = iu->buf;
1379         memset(cmd, 0, sizeof *cmd);
1380
1381         cmd->opcode = SRP_CMD;
1382         cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
1383         cmd->tag    = req->index;
1384         memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1385
1386         req->scmnd    = scmnd;
1387         req->cmd      = iu;
1388
1389         len = srp_map_data(scmnd, target, req);
1390         if (len < 0) {
1391                 shost_printk(KERN_ERR, target->scsi_host,
1392                              PFX "Failed to map data\n");
1393                 goto err_iu;
1394         }
1395
1396         ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1397                                       DMA_TO_DEVICE);
1398
1399         if (srp_post_send(target, iu, len)) {
1400                 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1401                 goto err_unmap;
1402         }
1403
1404         return 0;
1405
1406 err_unmap:
1407         srp_unmap_data(scmnd, target, req);
1408
1409 err_iu:
1410         srp_put_tx_iu(target, iu, SRP_IU_CMD);
1411
1412         /*
1413          * Avoid that the loops that iterate over the request ring can
1414          * encounter a dangling SCSI command pointer.
1415          */
1416         req->scmnd = NULL;
1417
1418         spin_lock_irqsave(&target->lock, flags);
1419         list_add(&req->list, &target->free_reqs);
1420
1421 err_unlock:
1422         spin_unlock_irqrestore(&target->lock, flags);
1423
1424         return SCSI_MLQUEUE_HOST_BUSY;
1425 }
1426
1427 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1428 {
1429         int i;
1430
1431         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1432                 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1433                                                   target->max_ti_iu_len,
1434                                                   GFP_KERNEL, DMA_FROM_DEVICE);
1435                 if (!target->rx_ring[i])
1436                         goto err;
1437         }
1438
1439         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1440                 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1441                                                   target->max_iu_len,
1442                                                   GFP_KERNEL, DMA_TO_DEVICE);
1443                 if (!target->tx_ring[i])
1444                         goto err;
1445
1446                 list_add(&target->tx_ring[i]->list, &target->free_tx);
1447         }
1448
1449         return 0;
1450
1451 err:
1452         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1453                 srp_free_iu(target->srp_host, target->rx_ring[i]);
1454                 target->rx_ring[i] = NULL;
1455         }
1456
1457         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1458                 srp_free_iu(target->srp_host, target->tx_ring[i]);
1459                 target->tx_ring[i] = NULL;
1460         }
1461
1462         return -ENOMEM;
1463 }
1464
1465 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
1466 {
1467         uint64_t T_tr_ns, max_compl_time_ms;
1468         uint32_t rq_tmo_jiffies;
1469
1470         /*
1471          * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
1472          * table 91), both the QP timeout and the retry count have to be set
1473          * for RC QP's during the RTR to RTS transition.
1474          */
1475         WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
1476                      (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
1477
1478         /*
1479          * Set target->rq_tmo_jiffies to one second more than the largest time
1480          * it can take before an error completion is generated. See also
1481          * C9-140..142 in the IBTA spec for more information about how to
1482          * convert the QP Local ACK Timeout value to nanoseconds.
1483          */
1484         T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
1485         max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
1486         do_div(max_compl_time_ms, NSEC_PER_MSEC);
1487         rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
1488
1489         return rq_tmo_jiffies;
1490 }
1491
1492 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1493                                struct srp_login_rsp *lrsp,
1494                                struct srp_target_port *target)
1495 {
1496         struct ib_qp_attr *qp_attr = NULL;
1497         int attr_mask = 0;
1498         int ret;
1499         int i;
1500
1501         if (lrsp->opcode == SRP_LOGIN_RSP) {
1502                 target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1503                 target->req_lim       = be32_to_cpu(lrsp->req_lim_delta);
1504
1505                 /*
1506                  * Reserve credits for task management so we don't
1507                  * bounce requests back to the SCSI mid-layer.
1508                  */
1509                 target->scsi_host->can_queue
1510                         = min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1511                               target->scsi_host->can_queue);
1512         } else {
1513                 shost_printk(KERN_WARNING, target->scsi_host,
1514                              PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1515                 ret = -ECONNRESET;
1516                 goto error;
1517         }
1518
1519         if (!target->rx_ring[0]) {
1520                 ret = srp_alloc_iu_bufs(target);
1521                 if (ret)
1522                         goto error;
1523         }
1524
1525         ret = -ENOMEM;
1526         qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1527         if (!qp_attr)
1528                 goto error;
1529
1530         qp_attr->qp_state = IB_QPS_RTR;
1531         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1532         if (ret)
1533                 goto error_free;
1534
1535         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1536         if (ret)
1537                 goto error_free;
1538
1539         for (i = 0; i < SRP_RQ_SIZE; i++) {
1540                 struct srp_iu *iu = target->rx_ring[i];
1541                 ret = srp_post_recv(target, iu);
1542                 if (ret)
1543                         goto error_free;
1544         }
1545
1546         qp_attr->qp_state = IB_QPS_RTS;
1547         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1548         if (ret)
1549                 goto error_free;
1550
1551         target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
1552
1553         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1554         if (ret)
1555                 goto error_free;
1556
1557         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1558
1559 error_free:
1560         kfree(qp_attr);
1561
1562 error:
1563         target->status = ret;
1564 }
1565
1566 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1567                                struct ib_cm_event *event,
1568                                struct srp_target_port *target)
1569 {
1570         struct Scsi_Host *shost = target->scsi_host;
1571         struct ib_class_port_info *cpi;
1572         int opcode;
1573
1574         switch (event->param.rej_rcvd.reason) {
1575         case IB_CM_REJ_PORT_CM_REDIRECT:
1576                 cpi = event->param.rej_rcvd.ari;
1577                 target->path.dlid = cpi->redirect_lid;
1578                 target->path.pkey = cpi->redirect_pkey;
1579                 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1580                 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1581
1582                 target->status = target->path.dlid ?
1583                         SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1584                 break;
1585
1586         case IB_CM_REJ_PORT_REDIRECT:
1587                 if (srp_target_is_topspin(target)) {
1588                         /*
1589                          * Topspin/Cisco SRP gateways incorrectly send
1590                          * reject reason code 25 when they mean 24
1591                          * (port redirect).
1592                          */
1593                         memcpy(target->path.dgid.raw,
1594                                event->param.rej_rcvd.ari, 16);
1595
1596                         shost_printk(KERN_DEBUG, shost,
1597                                      PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1598                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1599                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1600
1601                         target->status = SRP_PORT_REDIRECT;
1602                 } else {
1603                         shost_printk(KERN_WARNING, shost,
1604                                      "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1605                         target->status = -ECONNRESET;
1606                 }
1607                 break;
1608
1609         case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1610                 shost_printk(KERN_WARNING, shost,
1611                             "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1612                 target->status = -ECONNRESET;
1613                 break;
1614
1615         case IB_CM_REJ_CONSUMER_DEFINED:
1616                 opcode = *(u8 *) event->private_data;
1617                 if (opcode == SRP_LOGIN_REJ) {
1618                         struct srp_login_rej *rej = event->private_data;
1619                         u32 reason = be32_to_cpu(rej->reason);
1620
1621                         if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1622                                 shost_printk(KERN_WARNING, shost,
1623                                              PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1624                         else
1625                                 shost_printk(KERN_WARNING, shost,
1626                                             PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1627                 } else
1628                         shost_printk(KERN_WARNING, shost,
1629                                      "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1630                                      " opcode 0x%02x\n", opcode);
1631                 target->status = -ECONNRESET;
1632                 break;
1633
1634         case IB_CM_REJ_STALE_CONN:
1635                 shost_printk(KERN_WARNING, shost, "  REJ reason: stale connection\n");
1636                 target->status = SRP_STALE_CONN;
1637                 break;
1638
1639         default:
1640                 shost_printk(KERN_WARNING, shost, "  REJ reason 0x%x\n",
1641                              event->param.rej_rcvd.reason);
1642                 target->status = -ECONNRESET;
1643         }
1644 }
1645
1646 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1647 {
1648         struct srp_target_port *target = cm_id->context;
1649         int comp = 0;
1650
1651         switch (event->event) {
1652         case IB_CM_REQ_ERROR:
1653                 shost_printk(KERN_DEBUG, target->scsi_host,
1654                              PFX "Sending CM REQ failed\n");
1655                 comp = 1;
1656                 target->status = -ECONNRESET;
1657                 break;
1658
1659         case IB_CM_REP_RECEIVED:
1660                 comp = 1;
1661                 srp_cm_rep_handler(cm_id, event->private_data, target);
1662                 break;
1663
1664         case IB_CM_REJ_RECEIVED:
1665                 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1666                 comp = 1;
1667
1668                 srp_cm_rej_handler(cm_id, event, target);
1669                 break;
1670
1671         case IB_CM_DREQ_RECEIVED:
1672                 shost_printk(KERN_WARNING, target->scsi_host,
1673                              PFX "DREQ received - connection closed\n");
1674                 srp_change_conn_state(target, false);
1675                 if (ib_send_cm_drep(cm_id, NULL, 0))
1676                         shost_printk(KERN_ERR, target->scsi_host,
1677                                      PFX "Sending CM DREP failed\n");
1678                 break;
1679
1680         case IB_CM_TIMEWAIT_EXIT:
1681                 shost_printk(KERN_ERR, target->scsi_host,
1682                              PFX "connection closed\n");
1683
1684                 target->status = 0;
1685                 break;
1686
1687         case IB_CM_MRA_RECEIVED:
1688         case IB_CM_DREQ_ERROR:
1689         case IB_CM_DREP_RECEIVED:
1690                 break;
1691
1692         default:
1693                 shost_printk(KERN_WARNING, target->scsi_host,
1694                              PFX "Unhandled CM event %d\n", event->event);
1695                 break;
1696         }
1697
1698         if (comp)
1699                 complete(&target->done);
1700
1701         return 0;
1702 }
1703
1704 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1705                              u64 req_tag, unsigned int lun, u8 func)
1706 {
1707         struct ib_device *dev = target->srp_host->srp_dev->dev;
1708         struct srp_iu *iu;
1709         struct srp_tsk_mgmt *tsk_mgmt;
1710
1711         if (!target->connected || target->qp_in_error)
1712                 return -1;
1713
1714         init_completion(&target->tsk_mgmt_done);
1715
1716         spin_lock_irq(&target->lock);
1717         iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1718         spin_unlock_irq(&target->lock);
1719
1720         if (!iu)
1721                 return -1;
1722
1723         ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1724                                    DMA_TO_DEVICE);
1725         tsk_mgmt = iu->buf;
1726         memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1727
1728         tsk_mgmt->opcode        = SRP_TSK_MGMT;
1729         tsk_mgmt->lun           = cpu_to_be64((u64) lun << 48);
1730         tsk_mgmt->tag           = req_tag | SRP_TAG_TSK_MGMT;
1731         tsk_mgmt->tsk_mgmt_func = func;
1732         tsk_mgmt->task_tag      = req_tag;
1733
1734         ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1735                                       DMA_TO_DEVICE);
1736         if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1737                 srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1738                 return -1;
1739         }
1740
1741         if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1742                                          msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1743                 return -1;
1744
1745         return 0;
1746 }
1747
1748 static int srp_abort(struct scsi_cmnd *scmnd)
1749 {
1750         struct srp_target_port *target = host_to_target(scmnd->device->host);
1751         struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1752
1753         shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
1754
1755         if (!req || !srp_claim_req(target, req, scmnd))
1756                 return FAILED;
1757         srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
1758                           SRP_TSK_ABORT_TASK);
1759         srp_free_req(target, req, scmnd, 0);
1760         scmnd->result = DID_ABORT << 16;
1761         scmnd->scsi_done(scmnd);
1762
1763         return SUCCESS;
1764 }
1765
1766 static int srp_reset_device(struct scsi_cmnd *scmnd)
1767 {
1768         struct srp_target_port *target = host_to_target(scmnd->device->host);
1769         int i;
1770
1771         shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
1772
1773         if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
1774                               SRP_TSK_LUN_RESET))
1775                 return FAILED;
1776         if (target->tsk_mgmt_status)
1777                 return FAILED;
1778
1779         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1780                 struct srp_request *req = &target->req_ring[i];
1781                 if (req->scmnd && req->scmnd->device == scmnd->device)
1782                         srp_reset_req(target, req);
1783         }
1784
1785         return SUCCESS;
1786 }
1787
1788 static int srp_reset_host(struct scsi_cmnd *scmnd)
1789 {
1790         struct srp_target_port *target = host_to_target(scmnd->device->host);
1791         int ret = FAILED;
1792
1793         shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
1794
1795         if (!srp_reconnect_target(target))
1796                 ret = SUCCESS;
1797
1798         return ret;
1799 }
1800
1801 static int srp_slave_configure(struct scsi_device *sdev)
1802 {
1803         struct Scsi_Host *shost = sdev->host;
1804         struct srp_target_port *target = host_to_target(shost);
1805         struct request_queue *q = sdev->request_queue;
1806         unsigned long timeout;
1807
1808         if (sdev->type == TYPE_DISK) {
1809                 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
1810                 blk_queue_rq_timeout(q, timeout);
1811         }
1812
1813         return 0;
1814 }
1815
1816 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
1817                            char *buf)
1818 {
1819         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1820
1821         return sprintf(buf, "0x%016llx\n",
1822                        (unsigned long long) be64_to_cpu(target->id_ext));
1823 }
1824
1825 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
1826                              char *buf)
1827 {
1828         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1829
1830         return sprintf(buf, "0x%016llx\n",
1831                        (unsigned long long) be64_to_cpu(target->ioc_guid));
1832 }
1833
1834 static ssize_t show_service_id(struct device *dev,
1835                                struct device_attribute *attr, char *buf)
1836 {
1837         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1838
1839         return sprintf(buf, "0x%016llx\n",
1840                        (unsigned long long) be64_to_cpu(target->service_id));
1841 }
1842
1843 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
1844                          char *buf)
1845 {
1846         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1847
1848         return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1849 }
1850
1851 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
1852                          char *buf)
1853 {
1854         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1855
1856         return sprintf(buf, "%pI6\n", target->path.dgid.raw);
1857 }
1858
1859 static ssize_t show_orig_dgid(struct device *dev,
1860                               struct device_attribute *attr, char *buf)
1861 {
1862         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1863
1864         return sprintf(buf, "%pI6\n", target->orig_dgid);
1865 }
1866
1867 static ssize_t show_req_lim(struct device *dev,
1868                             struct device_attribute *attr, char *buf)
1869 {
1870         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1871
1872         return sprintf(buf, "%d\n", target->req_lim);
1873 }
1874
1875 static ssize_t show_zero_req_lim(struct device *dev,
1876                                  struct device_attribute *attr, char *buf)
1877 {
1878         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1879
1880         return sprintf(buf, "%d\n", target->zero_req_lim);
1881 }
1882
1883 static ssize_t show_local_ib_port(struct device *dev,
1884                                   struct device_attribute *attr, char *buf)
1885 {
1886         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1887
1888         return sprintf(buf, "%d\n", target->srp_host->port);
1889 }
1890
1891 static ssize_t show_local_ib_device(struct device *dev,
1892                                     struct device_attribute *attr, char *buf)
1893 {
1894         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1895
1896         return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
1897 }
1898
1899 static ssize_t show_cmd_sg_entries(struct device *dev,
1900                                    struct device_attribute *attr, char *buf)
1901 {
1902         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1903
1904         return sprintf(buf, "%u\n", target->cmd_sg_cnt);
1905 }
1906
1907 static ssize_t show_allow_ext_sg(struct device *dev,
1908                                  struct device_attribute *attr, char *buf)
1909 {
1910         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1911
1912         return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
1913 }
1914
1915 static DEVICE_ATTR(id_ext,          S_IRUGO, show_id_ext,          NULL);
1916 static DEVICE_ATTR(ioc_guid,        S_IRUGO, show_ioc_guid,        NULL);
1917 static DEVICE_ATTR(service_id,      S_IRUGO, show_service_id,      NULL);
1918 static DEVICE_ATTR(pkey,            S_IRUGO, show_pkey,            NULL);
1919 static DEVICE_ATTR(dgid,            S_IRUGO, show_dgid,            NULL);
1920 static DEVICE_ATTR(orig_dgid,       S_IRUGO, show_orig_dgid,       NULL);
1921 static DEVICE_ATTR(req_lim,         S_IRUGO, show_req_lim,         NULL);
1922 static DEVICE_ATTR(zero_req_lim,    S_IRUGO, show_zero_req_lim,    NULL);
1923 static DEVICE_ATTR(local_ib_port,   S_IRUGO, show_local_ib_port,   NULL);
1924 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1925 static DEVICE_ATTR(cmd_sg_entries,  S_IRUGO, show_cmd_sg_entries,  NULL);
1926 static DEVICE_ATTR(allow_ext_sg,    S_IRUGO, show_allow_ext_sg,    NULL);
1927
1928 static struct device_attribute *srp_host_attrs[] = {
1929         &dev_attr_id_ext,
1930         &dev_attr_ioc_guid,
1931         &dev_attr_service_id,
1932         &dev_attr_pkey,
1933         &dev_attr_dgid,
1934         &dev_attr_orig_dgid,
1935         &dev_attr_req_lim,
1936         &dev_attr_zero_req_lim,
1937         &dev_attr_local_ib_port,
1938         &dev_attr_local_ib_device,
1939         &dev_attr_cmd_sg_entries,
1940         &dev_attr_allow_ext_sg,
1941         NULL
1942 };
1943
1944 static struct scsi_host_template srp_template = {
1945         .module                         = THIS_MODULE,
1946         .name                           = "InfiniBand SRP initiator",
1947         .proc_name                      = DRV_NAME,
1948         .slave_configure                = srp_slave_configure,
1949         .info                           = srp_target_info,
1950         .queuecommand                   = srp_queuecommand,
1951         .eh_abort_handler               = srp_abort,
1952         .eh_device_reset_handler        = srp_reset_device,
1953         .eh_host_reset_handler          = srp_reset_host,
1954         .sg_tablesize                   = SRP_DEF_SG_TABLESIZE,
1955         .can_queue                      = SRP_CMD_SQ_SIZE,
1956         .this_id                        = -1,
1957         .cmd_per_lun                    = SRP_CMD_SQ_SIZE,
1958         .use_clustering                 = ENABLE_CLUSTERING,
1959         .shost_attrs                    = srp_host_attrs
1960 };
1961
1962 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1963 {
1964         struct srp_rport_identifiers ids;
1965         struct srp_rport *rport;
1966
1967         sprintf(target->target_name, "SRP.T10:%016llX",
1968                  (unsigned long long) be64_to_cpu(target->id_ext));
1969
1970         if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
1971                 return -ENODEV;
1972
1973         memcpy(ids.port_id, &target->id_ext, 8);
1974         memcpy(ids.port_id + 8, &target->ioc_guid, 8);
1975         ids.roles = SRP_RPORT_ROLE_TARGET;
1976         rport = srp_rport_add(target->scsi_host, &ids);
1977         if (IS_ERR(rport)) {
1978                 scsi_remove_host(target->scsi_host);
1979                 return PTR_ERR(rport);
1980         }
1981
1982         rport->lld_data = target;
1983
1984         spin_lock(&host->target_lock);
1985         list_add_tail(&target->list, &host->target_list);
1986         spin_unlock(&host->target_lock);
1987
1988         target->state = SRP_TARGET_LIVE;
1989
1990         scsi_scan_target(&target->scsi_host->shost_gendev,
1991                          0, target->scsi_id, SCAN_WILD_CARD, 0);
1992
1993         return 0;
1994 }
1995
1996 static void srp_release_dev(struct device *dev)
1997 {
1998         struct srp_host *host =
1999                 container_of(dev, struct srp_host, dev);
2000
2001         complete(&host->released);
2002 }
2003
2004 static struct class srp_class = {
2005         .name    = "infiniband_srp",
2006         .dev_release = srp_release_dev
2007 };
2008
2009 /*
2010  * Target ports are added by writing
2011  *
2012  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2013  *     pkey=<P_Key>,service_id=<service ID>
2014  *
2015  * to the add_target sysfs attribute.
2016  */
2017 enum {
2018         SRP_OPT_ERR             = 0,
2019         SRP_OPT_ID_EXT          = 1 << 0,
2020         SRP_OPT_IOC_GUID        = 1 << 1,
2021         SRP_OPT_DGID            = 1 << 2,
2022         SRP_OPT_PKEY            = 1 << 3,
2023         SRP_OPT_SERVICE_ID      = 1 << 4,
2024         SRP_OPT_MAX_SECT        = 1 << 5,
2025         SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2026         SRP_OPT_IO_CLASS        = 1 << 7,
2027         SRP_OPT_INITIATOR_EXT   = 1 << 8,
2028         SRP_OPT_CMD_SG_ENTRIES  = 1 << 9,
2029         SRP_OPT_ALLOW_EXT_SG    = 1 << 10,
2030         SRP_OPT_SG_TABLESIZE    = 1 << 11,
2031         SRP_OPT_ALL             = (SRP_OPT_ID_EXT       |
2032                                    SRP_OPT_IOC_GUID     |
2033                                    SRP_OPT_DGID         |
2034                                    SRP_OPT_PKEY         |
2035                                    SRP_OPT_SERVICE_ID),
2036 };
2037
2038 static const match_table_t srp_opt_tokens = {
2039         { SRP_OPT_ID_EXT,               "id_ext=%s"             },
2040         { SRP_OPT_IOC_GUID,             "ioc_guid=%s"           },
2041         { SRP_OPT_DGID,                 "dgid=%s"               },
2042         { SRP_OPT_PKEY,                 "pkey=%x"               },
2043         { SRP_OPT_SERVICE_ID,           "service_id=%s"         },
2044         { SRP_OPT_MAX_SECT,             "max_sect=%d"           },
2045         { SRP_OPT_MAX_CMD_PER_LUN,      "max_cmd_per_lun=%d"    },
2046         { SRP_OPT_IO_CLASS,             "io_class=%x"           },
2047         { SRP_OPT_INITIATOR_EXT,        "initiator_ext=%s"      },
2048         { SRP_OPT_CMD_SG_ENTRIES,       "cmd_sg_entries=%u"     },
2049         { SRP_OPT_ALLOW_EXT_SG,         "allow_ext_sg=%u"       },
2050         { SRP_OPT_SG_TABLESIZE,         "sg_tablesize=%u"       },
2051         { SRP_OPT_ERR,                  NULL                    }
2052 };
2053
2054 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2055 {
2056         char *options, *sep_opt;
2057         char *p;
2058         char dgid[3];
2059         substring_t args[MAX_OPT_ARGS];
2060         int opt_mask = 0;
2061         int token;
2062         int ret = -EINVAL;
2063         int i;
2064
2065         options = kstrdup(buf, GFP_KERNEL);
2066         if (!options)
2067                 return -ENOMEM;
2068
2069         sep_opt = options;
2070         while ((p = strsep(&sep_opt, ",")) != NULL) {
2071                 if (!*p)
2072                         continue;
2073
2074                 token = match_token(p, srp_opt_tokens, args);
2075                 opt_mask |= token;
2076
2077                 switch (token) {
2078                 case SRP_OPT_ID_EXT:
2079                         p = match_strdup(args);
2080                         if (!p) {
2081                                 ret = -ENOMEM;
2082                                 goto out;
2083                         }
2084                         target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2085                         kfree(p);
2086                         break;
2087
2088                 case SRP_OPT_IOC_GUID:
2089                         p = match_strdup(args);
2090                         if (!p) {
2091                                 ret = -ENOMEM;
2092                                 goto out;
2093                         }
2094                         target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2095                         kfree(p);
2096                         break;
2097
2098                 case SRP_OPT_DGID:
2099                         p = match_strdup(args);
2100                         if (!p) {
2101                                 ret = -ENOMEM;
2102                                 goto out;
2103                         }
2104                         if (strlen(p) != 32) {
2105                                 pr_warn("bad dest GID parameter '%s'\n", p);
2106                                 kfree(p);
2107                                 goto out;
2108                         }
2109
2110                         for (i = 0; i < 16; ++i) {
2111                                 strlcpy(dgid, p + i * 2, 3);
2112                                 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2113                         }
2114                         kfree(p);
2115                         memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2116                         break;
2117
2118                 case SRP_OPT_PKEY:
2119                         if (match_hex(args, &token)) {
2120                                 pr_warn("bad P_Key parameter '%s'\n", p);
2121                                 goto out;
2122                         }
2123                         target->path.pkey = cpu_to_be16(token);
2124                         break;
2125
2126                 case SRP_OPT_SERVICE_ID:
2127                         p = match_strdup(args);
2128                         if (!p) {
2129                                 ret = -ENOMEM;
2130                                 goto out;
2131                         }
2132                         target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2133                         target->path.service_id = target->service_id;
2134                         kfree(p);
2135                         break;
2136
2137                 case SRP_OPT_MAX_SECT:
2138                         if (match_int(args, &token)) {
2139                                 pr_warn("bad max sect parameter '%s'\n", p);
2140                                 goto out;
2141                         }
2142                         target->scsi_host->max_sectors = token;
2143                         break;
2144
2145                 case SRP_OPT_MAX_CMD_PER_LUN:
2146                         if (match_int(args, &token)) {
2147                                 pr_warn("bad max cmd_per_lun parameter '%s'\n",
2148                                         p);
2149                                 goto out;
2150                         }
2151                         target->scsi_host->cmd_per_lun = min(token, SRP_CMD_SQ_SIZE);
2152                         break;
2153
2154                 case SRP_OPT_IO_CLASS:
2155                         if (match_hex(args, &token)) {
2156                                 pr_warn("bad IO class parameter '%s'\n", p);
2157                                 goto out;
2158                         }
2159                         if (token != SRP_REV10_IB_IO_CLASS &&
2160                             token != SRP_REV16A_IB_IO_CLASS) {
2161                                 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2162                                         token, SRP_REV10_IB_IO_CLASS,
2163                                         SRP_REV16A_IB_IO_CLASS);
2164                                 goto out;
2165                         }
2166                         target->io_class = token;
2167                         break;
2168
2169                 case SRP_OPT_INITIATOR_EXT:
2170                         p = match_strdup(args);
2171                         if (!p) {
2172                                 ret = -ENOMEM;
2173                                 goto out;
2174                         }
2175                         target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2176                         kfree(p);
2177                         break;
2178
2179                 case SRP_OPT_CMD_SG_ENTRIES:
2180                         if (match_int(args, &token) || token < 1 || token > 255) {
2181                                 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2182                                         p);
2183                                 goto out;
2184                         }
2185                         target->cmd_sg_cnt = token;
2186                         break;
2187
2188                 case SRP_OPT_ALLOW_EXT_SG:
2189                         if (match_int(args, &token)) {
2190                                 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2191                                 goto out;
2192                         }
2193                         target->allow_ext_sg = !!token;
2194                         break;
2195
2196                 case SRP_OPT_SG_TABLESIZE:
2197                         if (match_int(args, &token) || token < 1 ||
2198                                         token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2199                                 pr_warn("bad max sg_tablesize parameter '%s'\n",
2200                                         p);
2201                                 goto out;
2202                         }
2203                         target->sg_tablesize = token;
2204                         break;
2205
2206                 default:
2207                         pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2208                                 p);
2209                         goto out;
2210                 }
2211         }
2212
2213         if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2214                 ret = 0;
2215         else
2216                 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2217                         if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2218                             !(srp_opt_tokens[i].token & opt_mask))
2219                                 pr_warn("target creation request is missing parameter '%s'\n",
2220                                         srp_opt_tokens[i].pattern);
2221
2222 out:
2223         kfree(options);
2224         return ret;
2225 }
2226
2227 static ssize_t srp_create_target(struct device *dev,
2228                                  struct device_attribute *attr,
2229                                  const char *buf, size_t count)
2230 {
2231         struct srp_host *host =
2232                 container_of(dev, struct srp_host, dev);
2233         struct Scsi_Host *target_host;
2234         struct srp_target_port *target;
2235         struct ib_device *ibdev = host->srp_dev->dev;
2236         dma_addr_t dma_addr;
2237         int i, ret;
2238
2239         target_host = scsi_host_alloc(&srp_template,
2240                                       sizeof (struct srp_target_port));
2241         if (!target_host)
2242                 return -ENOMEM;
2243
2244         target_host->transportt  = ib_srp_transport_template;
2245         target_host->max_channel = 0;
2246         target_host->max_id      = 1;
2247         target_host->max_lun     = SRP_MAX_LUN;
2248         target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2249
2250         target = host_to_target(target_host);
2251
2252         target->io_class        = SRP_REV16A_IB_IO_CLASS;
2253         target->scsi_host       = target_host;
2254         target->srp_host        = host;
2255         target->lkey            = host->srp_dev->mr->lkey;
2256         target->rkey            = host->srp_dev->mr->rkey;
2257         target->cmd_sg_cnt      = cmd_sg_entries;
2258         target->sg_tablesize    = indirect_sg_entries ? : cmd_sg_entries;
2259         target->allow_ext_sg    = allow_ext_sg;
2260
2261         ret = srp_parse_options(buf, target);
2262         if (ret)
2263                 goto err;
2264
2265         if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2266                                 target->cmd_sg_cnt < target->sg_tablesize) {
2267                 pr_warn("No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2268                 target->sg_tablesize = target->cmd_sg_cnt;
2269         }
2270
2271         target_host->sg_tablesize = target->sg_tablesize;
2272         target->indirect_size = target->sg_tablesize *
2273                                 sizeof (struct srp_direct_buf);
2274         target->max_iu_len = sizeof (struct srp_cmd) +
2275                              sizeof (struct srp_indirect_buf) +
2276                              target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2277
2278         INIT_WORK(&target->remove_work, srp_remove_work);
2279         spin_lock_init(&target->lock);
2280         INIT_LIST_HEAD(&target->free_tx);
2281         INIT_LIST_HEAD(&target->free_reqs);
2282         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
2283                 struct srp_request *req = &target->req_ring[i];
2284
2285                 req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof (void *),
2286                                         GFP_KERNEL);
2287                 req->map_page = kmalloc(SRP_FMR_SIZE * sizeof (void *),
2288                                         GFP_KERNEL);
2289                 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
2290                 if (!req->fmr_list || !req->map_page || !req->indirect_desc)
2291                         goto err_free_mem;
2292
2293                 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
2294                                              target->indirect_size,
2295                                              DMA_TO_DEVICE);
2296                 if (ib_dma_mapping_error(ibdev, dma_addr))
2297                         goto err_free_mem;
2298
2299                 req->indirect_dma_addr = dma_addr;
2300                 req->index = i;
2301                 list_add_tail(&req->list, &target->free_reqs);
2302         }
2303
2304         ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2305
2306         shost_printk(KERN_DEBUG, target->scsi_host, PFX
2307                      "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
2308                      "service_id %016llx dgid %pI6\n",
2309                (unsigned long long) be64_to_cpu(target->id_ext),
2310                (unsigned long long) be64_to_cpu(target->ioc_guid),
2311                be16_to_cpu(target->path.pkey),
2312                (unsigned long long) be64_to_cpu(target->service_id),
2313                target->path.dgid.raw);
2314
2315         ret = srp_create_target_ib(target);
2316         if (ret)
2317                 goto err_free_mem;
2318
2319         ret = srp_new_cm_id(target);
2320         if (ret)
2321                 goto err_free_ib;
2322
2323         ret = srp_connect_target(target);
2324         if (ret) {
2325                 shost_printk(KERN_ERR, target->scsi_host,
2326                              PFX "Connection failed\n");
2327                 goto err_cm_id;
2328         }
2329
2330         ret = srp_add_target(host, target);
2331         if (ret)
2332                 goto err_disconnect;
2333
2334         return count;
2335
2336 err_disconnect:
2337         srp_disconnect_target(target);
2338
2339 err_cm_id:
2340         ib_destroy_cm_id(target->cm_id);
2341
2342 err_free_ib:
2343         srp_free_target_ib(target);
2344
2345 err_free_mem:
2346         srp_free_req_data(target);
2347
2348 err:
2349         scsi_host_put(target_host);
2350
2351         return ret;
2352 }
2353
2354 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2355
2356 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2357                           char *buf)
2358 {
2359         struct srp_host *host = container_of(dev, struct srp_host, dev);
2360
2361         return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2362 }
2363
2364 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2365
2366 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2367                          char *buf)
2368 {
2369         struct srp_host *host = container_of(dev, struct srp_host, dev);
2370
2371         return sprintf(buf, "%d\n", host->port);
2372 }
2373
2374 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2375
2376 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2377 {
2378         struct srp_host *host;
2379
2380         host = kzalloc(sizeof *host, GFP_KERNEL);
2381         if (!host)
2382                 return NULL;
2383
2384         INIT_LIST_HEAD(&host->target_list);
2385         spin_lock_init(&host->target_lock);
2386         init_completion(&host->released);
2387         host->srp_dev = device;
2388         host->port = port;
2389
2390         host->dev.class = &srp_class;
2391         host->dev.parent = device->dev->dma_device;
2392         dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2393
2394         if (device_register(&host->dev))
2395                 goto free_host;
2396         if (device_create_file(&host->dev, &dev_attr_add_target))
2397                 goto err_class;
2398         if (device_create_file(&host->dev, &dev_attr_ibdev))
2399                 goto err_class;
2400         if (device_create_file(&host->dev, &dev_attr_port))
2401                 goto err_class;
2402
2403         return host;
2404
2405 err_class:
2406         device_unregister(&host->dev);
2407
2408 free_host:
2409         kfree(host);
2410
2411         return NULL;
2412 }
2413
2414 static void srp_add_one(struct ib_device *device)
2415 {
2416         struct srp_device *srp_dev;
2417         struct ib_device_attr *dev_attr;
2418         struct ib_fmr_pool_param fmr_param;
2419         struct srp_host *host;
2420         int max_pages_per_fmr, fmr_page_shift, s, e, p;
2421
2422         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2423         if (!dev_attr)
2424                 return;
2425
2426         if (ib_query_device(device, dev_attr)) {
2427                 pr_warn("Query device failed for %s\n", device->name);
2428                 goto free_attr;
2429         }
2430
2431         srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2432         if (!srp_dev)
2433                 goto free_attr;
2434
2435         /*
2436          * Use the smallest page size supported by the HCA, down to a
2437          * minimum of 4096 bytes. We're unlikely to build large sglists
2438          * out of smaller entries.
2439          */
2440         fmr_page_shift          = max(12, ffs(dev_attr->page_size_cap) - 1);
2441         srp_dev->fmr_page_size  = 1 << fmr_page_shift;
2442         srp_dev->fmr_page_mask  = ~((u64) srp_dev->fmr_page_size - 1);
2443         srp_dev->fmr_max_size   = srp_dev->fmr_page_size * SRP_FMR_SIZE;
2444
2445         INIT_LIST_HEAD(&srp_dev->dev_list);
2446
2447         srp_dev->dev = device;
2448         srp_dev->pd  = ib_alloc_pd(device);
2449         if (IS_ERR(srp_dev->pd))
2450                 goto free_dev;
2451
2452         srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2453                                     IB_ACCESS_LOCAL_WRITE |
2454                                     IB_ACCESS_REMOTE_READ |
2455                                     IB_ACCESS_REMOTE_WRITE);
2456         if (IS_ERR(srp_dev->mr))
2457                 goto err_pd;
2458
2459         for (max_pages_per_fmr = SRP_FMR_SIZE;
2460                         max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2461                         max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2462                 memset(&fmr_param, 0, sizeof fmr_param);
2463                 fmr_param.pool_size         = SRP_FMR_POOL_SIZE;
2464                 fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
2465                 fmr_param.cache             = 1;
2466                 fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2467                 fmr_param.page_shift        = fmr_page_shift;
2468                 fmr_param.access            = (IB_ACCESS_LOCAL_WRITE |
2469                                                IB_ACCESS_REMOTE_WRITE |
2470                                                IB_ACCESS_REMOTE_READ);
2471
2472                 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2473                 if (!IS_ERR(srp_dev->fmr_pool))
2474                         break;
2475         }
2476
2477         if (IS_ERR(srp_dev->fmr_pool))
2478                 srp_dev->fmr_pool = NULL;
2479
2480         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2481                 s = 0;
2482                 e = 0;
2483         } else {
2484                 s = 1;
2485                 e = device->phys_port_cnt;
2486         }
2487
2488         for (p = s; p <= e; ++p) {
2489                 host = srp_add_port(srp_dev, p);
2490                 if (host)
2491                         list_add_tail(&host->list, &srp_dev->dev_list);
2492         }
2493
2494         ib_set_client_data(device, &srp_client, srp_dev);
2495
2496         goto free_attr;
2497
2498 err_pd:
2499         ib_dealloc_pd(srp_dev->pd);
2500
2501 free_dev:
2502         kfree(srp_dev);
2503
2504 free_attr:
2505         kfree(dev_attr);
2506 }
2507
2508 static void srp_remove_one(struct ib_device *device)
2509 {
2510         struct srp_device *srp_dev;
2511         struct srp_host *host, *tmp_host;
2512         struct srp_target_port *target;
2513
2514         srp_dev = ib_get_client_data(device, &srp_client);
2515
2516         list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2517                 device_unregister(&host->dev);
2518                 /*
2519                  * Wait for the sysfs entry to go away, so that no new
2520                  * target ports can be created.
2521                  */
2522                 wait_for_completion(&host->released);
2523
2524                 /*
2525                  * Remove all target ports.
2526                  */
2527                 spin_lock(&host->target_lock);
2528                 list_for_each_entry(target, &host->target_list, list)
2529                         srp_queue_remove_work(target);
2530                 spin_unlock(&host->target_lock);
2531
2532                 /*
2533                  * Wait for target port removal tasks.
2534                  */
2535                 flush_workqueue(system_long_wq);
2536
2537                 kfree(host);
2538         }
2539
2540         if (srp_dev->fmr_pool)
2541                 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2542         ib_dereg_mr(srp_dev->mr);
2543         ib_dealloc_pd(srp_dev->pd);
2544
2545         kfree(srp_dev);
2546 }
2547
2548 static struct srp_function_template ib_srp_transport_functions = {
2549         .rport_delete            = srp_rport_delete,
2550 };
2551
2552 static int __init srp_init_module(void)
2553 {
2554         int ret;
2555
2556         BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2557
2558         if (srp_sg_tablesize) {
2559                 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2560                 if (!cmd_sg_entries)
2561                         cmd_sg_entries = srp_sg_tablesize;
2562         }
2563
2564         if (!cmd_sg_entries)
2565                 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2566
2567         if (cmd_sg_entries > 255) {
2568                 pr_warn("Clamping cmd_sg_entries to 255\n");
2569                 cmd_sg_entries = 255;
2570         }
2571
2572         if (!indirect_sg_entries)
2573                 indirect_sg_entries = cmd_sg_entries;
2574         else if (indirect_sg_entries < cmd_sg_entries) {
2575                 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
2576                         cmd_sg_entries);
2577                 indirect_sg_entries = cmd_sg_entries;
2578         }
2579
2580         ib_srp_transport_template =
2581                 srp_attach_transport(&ib_srp_transport_functions);
2582         if (!ib_srp_transport_template)
2583                 return -ENOMEM;
2584
2585         ret = class_register(&srp_class);
2586         if (ret) {
2587                 pr_err("couldn't register class infiniband_srp\n");
2588                 srp_release_transport(ib_srp_transport_template);
2589                 return ret;
2590         }
2591
2592         ib_sa_register_client(&srp_sa_client);
2593
2594         ret = ib_register_client(&srp_client);
2595         if (ret) {
2596                 pr_err("couldn't register IB client\n");
2597                 srp_release_transport(ib_srp_transport_template);
2598                 ib_sa_unregister_client(&srp_sa_client);
2599                 class_unregister(&srp_class);
2600                 return ret;
2601         }
2602
2603         return 0;
2604 }
2605
2606 static void __exit srp_cleanup_module(void)
2607 {
2608         ib_unregister_client(&srp_client);
2609         ib_sa_unregister_client(&srp_sa_client);
2610         class_unregister(&srp_class);
2611         srp_release_transport(ib_srp_transport_template);
2612 }
2613
2614 module_init(srp_init_module);
2615 module_exit(srp_cleanup_module);