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     "1.0"
  57 #define DRV_RELDATE     "July 1, 2013"
  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 struct kernel_param_ops srp_tmo_ops;
  90
  91 static int srp_reconnect_delay = 10;
  92 module_param_cb(reconnect_delay, &srp_tmo_ops, &srp_reconnect_delay,
  93                 S_IRUGO | S_IWUSR);
  94 MODULE_PARM_DESC(reconnect_delay, "Time between successive reconnect attempts");
  95
  96 static int srp_fast_io_fail_tmo = 15;
  97 module_param_cb(fast_io_fail_tmo, &srp_tmo_ops, &srp_fast_io_fail_tmo,
  98                 S_IRUGO | S_IWUSR);
  99 MODULE_PARM_DESC(fast_io_fail_tmo,
 100                  "Number of seconds between the observation of a transport"
 101                  " layer error and failing all I/O. \"off\" means that this"
 102                  " functionality is disabled.");
 103
 104 static int srp_dev_loss_tmo = 600;
 105 module_param_cb(dev_loss_tmo, &srp_tmo_ops, &srp_dev_loss_tmo,
 106                 S_IRUGO | S_IWUSR);
 107 MODULE_PARM_DESC(dev_loss_tmo,
 108                  "Maximum number of seconds that the SRP transport should"
 109                  " insulate transport layer errors. After this time has been"
 110                  " exceeded the SCSI host is removed. Should be"
 111                  " between 1 and " __stringify(SCSI_DEVICE_BLOCK_MAX_TIMEOUT)
 112                  " if fast_io_fail_tmo has not been set. \"off\" means that"
 113                  " this functionality is disabled.");
 114
 115 static void srp_add_one(struct ib_device *device);
 116 static void srp_remove_one(struct ib_device *device);
 117 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
 118 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
 119 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
 120
 121 static struct scsi_transport_template *ib_srp_transport_template;
 122
 123 static struct ib_client srp_client = {
 124         .name   = "srp",
 125         .add    = srp_add_one,
 126         .remove = srp_remove_one
 127 };
 128
 129 static struct ib_sa_client srp_sa_client;
 130
 131 static int srp_tmo_get(char *buffer, const struct kernel_param *kp)
 132 {
 133         int tmo = *(int *)kp->arg;
 134
 135         if (tmo >= 0)
 136                 return sprintf(buffer, "%d", tmo);
 137         else
 138                 return sprintf(buffer, "off");
 139 }
 140
 141 static int srp_tmo_set(const char *val, const struct kernel_param *kp)
 142 {
 143         int tmo, res;
 144
 145         if (strncmp(val, "off", 3) != 0) {
 146                 res = kstrtoint(val, 0, &tmo);
 147                 if (res)
 148                         goto out;
 149         } else {
 150                 tmo = -1;
 151         }
 152         if (kp->arg == &srp_reconnect_delay)
 153                 res = srp_tmo_valid(tmo, srp_fast_io_fail_tmo,
 154                                     srp_dev_loss_tmo);
 155         else if (kp->arg == &srp_fast_io_fail_tmo)
 156                 res = srp_tmo_valid(srp_reconnect_delay, tmo, srp_dev_loss_tmo);
 157         else
 158                 res = srp_tmo_valid(srp_reconnect_delay, srp_fast_io_fail_tmo,
 159                                     tmo);
 160         if (res)
 161                 goto out;
 162         *(int *)kp->arg = tmo;
 163
 164 out:
 165         return res;
 166 }
 167
 168 static struct kernel_param_ops srp_tmo_ops = {
 169         .get = srp_tmo_get,
 170         .set = srp_tmo_set,
 171 };
 172
 173 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
 174 {
 175         return (struct srp_target_port *) host->hostdata;
 176 }
 177
 178 static const char *srp_target_info(struct Scsi_Host *host)
 179 {
 180         return host_to_target(host)->target_name;
 181 }
 182
 183 static int srp_target_is_topspin(struct srp_target_port *target)
 184 {
 185         static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
 186         static const u8 cisco_oui[3]   = { 0x00, 0x1b, 0x0d };
 187
 188         return topspin_workarounds &&
 189                 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
 190                  !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
 191 }
 192
 193 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
 194                                    gfp_t gfp_mask,
 195                                    enum dma_data_direction direction)
 196 {
 197         struct srp_iu *iu;
 198
 199         iu = kmalloc(sizeof *iu, gfp_mask);
 200         if (!iu)
 201                 goto out;
 202
 203         iu->buf = kzalloc(size, gfp_mask);
 204         if (!iu->buf)
 205                 goto out_free_iu;
 206
 207         iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
 208                                     direction);
 209         if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
 210                 goto out_free_buf;
 211
 212         iu->size      = size;
 213         iu->direction = direction;
 214
 215         return iu;
 216
 217 out_free_buf:
 218         kfree(iu->buf);
 219 out_free_iu:
 220         kfree(iu);
 221 out:
 222         return NULL;
 223 }
 224
 225 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
 226 {
 227         if (!iu)
 228                 return;
 229
 230         ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
 231                             iu->direction);
 232         kfree(iu->buf);
 233         kfree(iu);
 234 }
 235
 236 static void srp_qp_event(struct ib_event *event, void *context)
 237 {
 238         pr_debug("QP event %d\n", event->event);
 239 }
 240
 241 static int srp_init_qp(struct srp_target_port *target,
 242                        struct ib_qp *qp)
 243 {
 244         struct ib_qp_attr *attr;
 245         int ret;
 246
 247         attr = kmalloc(sizeof *attr, GFP_KERNEL);
 248         if (!attr)
 249                 return -ENOMEM;
 250
 251         ret = ib_find_pkey(target->srp_host->srp_dev->dev,
 252                            target->srp_host->port,
 253                            be16_to_cpu(target->path.pkey),
 254                            &attr->pkey_index);
 255         if (ret)
 256                 goto out;
 257
 258         attr->qp_state        = IB_QPS_INIT;
 259         attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
 260                                     IB_ACCESS_REMOTE_WRITE);
 261         attr->port_num        = target->srp_host->port;
 262
 263         ret = ib_modify_qp(qp, attr,
 264                            IB_QP_STATE          |
 265                            IB_QP_PKEY_INDEX     |
 266                            IB_QP_ACCESS_FLAGS   |
 267                            IB_QP_PORT);
 268
 269 out:
 270         kfree(attr);
 271         return ret;
 272 }
 273
 274 static int srp_new_cm_id(struct srp_target_port *target)
 275 {
 276         struct ib_cm_id *new_cm_id;
 277
 278         new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
 279                                     srp_cm_handler, target);
 280         if (IS_ERR(new_cm_id))
 281                 return PTR_ERR(new_cm_id);
 282
 283         if (target->cm_id)
 284                 ib_destroy_cm_id(target->cm_id);
 285         target->cm_id = new_cm_id;
 286
 287         return 0;
 288 }
 289
 290 static int srp_create_target_ib(struct srp_target_port *target)
 291 {
 292         struct ib_qp_init_attr *init_attr;
 293         struct ib_cq *recv_cq, *send_cq;
 294         struct ib_qp *qp;
 295         int ret;
 296
 297         init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
 298         if (!init_attr)
 299                 return -ENOMEM;
 300
 301         recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
 302                                srp_recv_completion, NULL, target, SRP_RQ_SIZE,
 303                                target->comp_vector);
 304         if (IS_ERR(recv_cq)) {
 305                 ret = PTR_ERR(recv_cq);
 306                 goto err;
 307         }
 308
 309         send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
 310                                srp_send_completion, NULL, target, SRP_SQ_SIZE,
 311                                target->comp_vector);
 312         if (IS_ERR(send_cq)) {
 313                 ret = PTR_ERR(send_cq);
 314                 goto err_recv_cq;
 315         }
 316
 317         ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
 318
 319         init_attr->event_handler       = srp_qp_event;
 320         init_attr->cap.max_send_wr     = SRP_SQ_SIZE;
 321         init_attr->cap.max_recv_wr     = SRP_RQ_SIZE;
 322         init_attr->cap.max_recv_sge    = 1;
 323         init_attr->cap.max_send_sge    = 1;
 324         init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
 325         init_attr->qp_type             = IB_QPT_RC;
 326         init_attr->send_cq             = send_cq;
 327         init_attr->recv_cq             = recv_cq;
 328
 329         qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
 330         if (IS_ERR(qp)) {
 331                 ret = PTR_ERR(qp);
 332                 goto err_send_cq;
 333         }
 334
 335         ret = srp_init_qp(target, qp);
 336         if (ret)
 337                 goto err_qp;
 338
 339         if (target->qp)
 340                 ib_destroy_qp(target->qp);
 341         if (target->recv_cq)
 342                 ib_destroy_cq(target->recv_cq);
 343         if (target->send_cq)
 344                 ib_destroy_cq(target->send_cq);
 345
 346         target->qp = qp;
 347         target->recv_cq = recv_cq;
 348         target->send_cq = send_cq;
 349
 350         kfree(init_attr);
 351         return 0;
 352
 353 err_qp:
 354         ib_destroy_qp(qp);
 355
 356 err_send_cq:
 357         ib_destroy_cq(send_cq);
 358
 359 err_recv_cq:
 360         ib_destroy_cq(recv_cq);
 361
 362 err:
 363         kfree(init_attr);
 364         return ret;
 365 }
 366
 367 static void srp_free_target_ib(struct srp_target_port *target)
 368 {
 369         int i;
 370
 371         ib_destroy_qp(target->qp);
 372         ib_destroy_cq(target->send_cq);
 373         ib_destroy_cq(target->recv_cq);
 374
 375         target->qp = NULL;
 376         target->send_cq = target->recv_cq = NULL;
 377
 378         for (i = 0; i < SRP_RQ_SIZE; ++i)
 379                 srp_free_iu(target->srp_host, target->rx_ring[i]);
 380         for (i = 0; i < SRP_SQ_SIZE; ++i)
 381                 srp_free_iu(target->srp_host, target->tx_ring[i]);
 382 }
 383
 384 static void srp_path_rec_completion(int status,
 385                                     struct ib_sa_path_rec *pathrec,
 386                                     void *target_ptr)
 387 {
 388         struct srp_target_port *target = target_ptr;
 389
 390         target->status = status;
 391         if (status)
 392                 shost_printk(KERN_ERR, target->scsi_host,
 393                              PFX "Got failed path rec status %d\n", status);
 394         else
 395                 target->path = *pathrec;
 396         complete(&target->done);
 397 }
 398
 399 static int srp_lookup_path(struct srp_target_port *target)
 400 {
 401         target->path.numb_path = 1;
 402
 403         init_completion(&target->done);
 404
 405         target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
 406                                                    target->srp_host->srp_dev->dev,
 407                                                    target->srp_host->port,
 408                                                    &target->path,
 409                                                    IB_SA_PATH_REC_SERVICE_ID    |
 410                                                    IB_SA_PATH_REC_DGID          |
 411                                                    IB_SA_PATH_REC_SGID          |
 412                                                    IB_SA_PATH_REC_NUMB_PATH     |
 413                                                    IB_SA_PATH_REC_PKEY,
 414                                                    SRP_PATH_REC_TIMEOUT_MS,
 415                                                    GFP_KERNEL,
 416                                                    srp_path_rec_completion,
 417                                                    target, &target->path_query);
 418         if (target->path_query_id < 0)
 419                 return target->path_query_id;
 420
 421         wait_for_completion(&target->done);
 422
 423         if (target->status < 0)
 424                 shost_printk(KERN_WARNING, target->scsi_host,
 425                              PFX "Path record query failed\n");
 426
 427         return target->status;
 428 }
 429
 430 static int srp_send_req(struct srp_target_port *target)
 431 {
 432         struct {
 433                 struct ib_cm_req_param param;
 434                 struct srp_login_req   priv;
 435         } *req = NULL;
 436         int status;
 437
 438         req = kzalloc(sizeof *req, GFP_KERNEL);
 439         if (!req)
 440                 return -ENOMEM;
 441
 442         req->param.primary_path               = &target->path;
 443         req->param.alternate_path             = NULL;
 444         req->param.service_id                 = target->service_id;
 445         req->param.qp_num                     = target->qp->qp_num;
 446         req->param.qp_type                    = target->qp->qp_type;
 447         req->param.private_data               = &req->priv;
 448         req->param.private_data_len           = sizeof req->priv;
 449         req->param.flow_control               = 1;
 450
 451         get_random_bytes(&req->param.starting_psn, 4);
 452         req->param.starting_psn              &= 0xffffff;
 453
 454         /*
 455          * Pick some arbitrary defaults here; we could make these
 456          * module parameters if anyone cared about setting them.
 457          */
 458         req->param.responder_resources        = 4;
 459         req->param.remote_cm_response_timeout = 20;
 460         req->param.local_cm_response_timeout  = 20;
 461         req->param.retry_count                = target->tl_retry_count;
 462         req->param.rnr_retry_count            = 7;
 463         req->param.max_cm_retries             = 15;
 464
 465         req->priv.opcode        = SRP_LOGIN_REQ;
 466         req->priv.tag           = 0;
 467         req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
 468         req->priv.req_buf_fmt   = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
 469                                               SRP_BUF_FORMAT_INDIRECT);
 470         /*
 471          * In the published SRP specification (draft rev. 16a), the
 472          * port identifier format is 8 bytes of ID extension followed
 473          * by 8 bytes of GUID.  Older drafts put the two halves in the
 474          * opposite order, so that the GUID comes first.
 475          *
 476          * Targets conforming to these obsolete drafts can be
 477          * recognized by the I/O Class they report.
 478          */
 479         if (target->io_class == SRP_REV10_IB_IO_CLASS) {
 480                 memcpy(req->priv.initiator_port_id,
 481                        &target->path.sgid.global.interface_id, 8);
 482                 memcpy(req->priv.initiator_port_id + 8,
 483                        &target->initiator_ext, 8);
 484                 memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
 485                 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
 486         } else {
 487                 memcpy(req->priv.initiator_port_id,
 488                        &target->initiator_ext, 8);
 489                 memcpy(req->priv.initiator_port_id + 8,
 490                        &target->path.sgid.global.interface_id, 8);
 491                 memcpy(req->priv.target_port_id,     &target->id_ext, 8);
 492                 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
 493         }
 494
 495         /*
 496          * Topspin/Cisco SRP targets will reject our login unless we
 497          * zero out the first 8 bytes of our initiator port ID and set
 498          * the second 8 bytes to the local node GUID.
 499          */
 500         if (srp_target_is_topspin(target)) {
 501                 shost_printk(KERN_DEBUG, target->scsi_host,
 502                              PFX "Topspin/Cisco initiator port ID workaround "
 503                              "activated for target GUID %016llx\n",
 504                              (unsigned long long) be64_to_cpu(target->ioc_guid));
 505                 memset(req->priv.initiator_port_id, 0, 8);
 506                 memcpy(req->priv.initiator_port_id + 8,
 507                        &target->srp_host->srp_dev->dev->node_guid, 8);
 508         }
 509
 510         status = ib_send_cm_req(target->cm_id, &req->param);
 511
 512         kfree(req);
 513
 514         return status;
 515 }
 516
 517 static bool srp_queue_remove_work(struct srp_target_port *target)
 518 {
 519         bool changed = false;
 520
 521         spin_lock_irq(&target->lock);
 522         if (target->state != SRP_TARGET_REMOVED) {
 523                 target->state = SRP_TARGET_REMOVED;
 524                 changed = true;
 525         }
 526         spin_unlock_irq(&target->lock);
 527
 528         if (changed)
 529                 queue_work(system_long_wq, &target->remove_work);
 530
 531         return changed;
 532 }
 533
 534 static bool srp_change_conn_state(struct srp_target_port *target,
 535                                   bool connected)
 536 {
 537         bool changed = false;
 538
 539         spin_lock_irq(&target->lock);
 540         if (target->connected != connected) {
 541                 target->connected = connected;
 542                 changed = true;
 543         }
 544         spin_unlock_irq(&target->lock);
 545
 546         return changed;
 547 }
 548
 549 static void srp_disconnect_target(struct srp_target_port *target)
 550 {
 551         if (srp_change_conn_state(target, false)) {
 552                 /* XXX should send SRP_I_LOGOUT request */
 553
 554                 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
 555                         shost_printk(KERN_DEBUG, target->scsi_host,
 556                                      PFX "Sending CM DREQ failed\n");
 557                 }
 558         }
 559 }
 560
 561 static void srp_free_req_data(struct srp_target_port *target)
 562 {
 563         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
 564         struct srp_request *req;
 565         int i;
 566
 567         for (i = 0, req = target->req_ring; i < SRP_CMD_SQ_SIZE; ++i, ++req) {
 568                 kfree(req->fmr_list);
 569                 kfree(req->map_page);
 570                 if (req->indirect_dma_addr) {
 571                         ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
 572                                             target->indirect_size,
 573                                             DMA_TO_DEVICE);
 574                 }
 575                 kfree(req->indirect_desc);
 576         }
 577 }
 578
 579 static int srp_alloc_req_data(struct srp_target_port *target)
 580 {
 581         struct srp_device *srp_dev = target->srp_host->srp_dev;
 582         struct ib_device *ibdev = srp_dev->dev;
 583         struct srp_request *req;
 584         dma_addr_t dma_addr;
 585         int i, ret = -ENOMEM;
 586
 587         INIT_LIST_HEAD(&target->free_reqs);
 588
 589         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
 590                 req = &target->req_ring[i];
 591                 req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof(void *),
 592                                         GFP_KERNEL);
 593                 req->map_page = kmalloc(SRP_FMR_SIZE * sizeof(void *),
 594                                         GFP_KERNEL);
 595                 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
 596                 if (!req->fmr_list || !req->map_page || !req->indirect_desc)
 597                         goto out;
 598
 599                 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
 600                                              target->indirect_size,
 601                                              DMA_TO_DEVICE);
 602                 if (ib_dma_mapping_error(ibdev, dma_addr))
 603                         goto out;
 604
 605                 req->indirect_dma_addr = dma_addr;
 606                 req->index = i;
 607                 list_add_tail(&req->list, &target->free_reqs);
 608         }
 609         ret = 0;
 610
 611 out:
 612         return ret;
 613 }
 614
 615 /**
 616  * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
 617  * @shost: SCSI host whose attributes to remove from sysfs.
 618  *
 619  * Note: Any attributes defined in the host template and that did not exist
 620  * before invocation of this function will be ignored.
 621  */
 622 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
 623 {
 624         struct device_attribute **attr;
 625
 626         for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
 627                 device_remove_file(&shost->shost_dev, *attr);
 628 }
 629
 630 static void srp_remove_target(struct srp_target_port *target)
 631 {
 632         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 633
 634         srp_del_scsi_host_attr(target->scsi_host);
 635         srp_rport_get(target->rport);
 636         srp_remove_host(target->scsi_host);
 637         scsi_remove_host(target->scsi_host);
 638         srp_disconnect_target(target);
 639         ib_destroy_cm_id(target->cm_id);
 640         srp_free_target_ib(target);
 641         cancel_work_sync(&target->tl_err_work);
 642         srp_rport_put(target->rport);
 643         srp_free_req_data(target);
 644         scsi_host_put(target->scsi_host);
 645 }
 646
 647 static void srp_remove_work(struct work_struct *work)
 648 {
 649         struct srp_target_port *target =
 650                 container_of(work, struct srp_target_port, remove_work);
 651
 652         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 653
 654         srp_remove_target(target);
 655
 656         spin_lock(&target->srp_host->target_lock);
 657         list_del(&target->list);
 658         spin_unlock(&target->srp_host->target_lock);
 659 }
 660
 661 static void srp_rport_delete(struct srp_rport *rport)
 662 {
 663         struct srp_target_port *target = rport->lld_data;
 664
 665         srp_queue_remove_work(target);
 666 }
 667
 668 static int srp_connect_target(struct srp_target_port *target)
 669 {
 670         int retries = 3;
 671         int ret;
 672
 673         WARN_ON_ONCE(target->connected);
 674
 675         target->qp_in_error = false;
 676
 677         ret = srp_lookup_path(target);
 678         if (ret)
 679                 return ret;
 680
 681         while (1) {
 682                 init_completion(&target->done);
 683                 ret = srp_send_req(target);
 684                 if (ret)
 685                         return ret;
 686                 wait_for_completion(&target->done);
 687
 688                 /*
 689                  * The CM event handling code will set status to
 690                  * SRP_PORT_REDIRECT if we get a port redirect REJ
 691                  * back, or SRP_DLID_REDIRECT if we get a lid/qp
 692                  * redirect REJ back.
 693                  */
 694                 switch (target->status) {
 695                 case 0:
 696                         srp_change_conn_state(target, true);
 697                         return 0;
 698
 699                 case SRP_PORT_REDIRECT:
 700                         ret = srp_lookup_path(target);
 701                         if (ret)
 702                                 return ret;
 703                         break;
 704
 705                 case SRP_DLID_REDIRECT:
 706                         break;
 707
 708                 case SRP_STALE_CONN:
 709                         /* Our current CM id was stale, and is now in timewait.
 710                          * Try to reconnect with a new one.
 711                          */
 712                         if (!retries-- || srp_new_cm_id(target)) {
 713                                 shost_printk(KERN_ERR, target->scsi_host, PFX
 714                                              "giving up on stale connection\n");
 715                                 target->status = -ECONNRESET;
 716                                 return target->status;
 717                         }
 718
 719                         shost_printk(KERN_ERR, target->scsi_host, PFX
 720                                      "retrying stale connection\n");
 721                         break;
 722
 723                 default:
 724                         return target->status;
 725                 }
 726         }
 727 }
 728
 729 static void srp_unmap_data(struct scsi_cmnd *scmnd,
 730                            struct srp_target_port *target,
 731                            struct srp_request *req)
 732 {
 733         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
 734         struct ib_pool_fmr **pfmr;
 735
 736         if (!scsi_sglist(scmnd) ||
 737             (scmnd->sc_data_direction != DMA_TO_DEVICE &&
 738              scmnd->sc_data_direction != DMA_FROM_DEVICE))
 739                 return;
 740
 741         pfmr = req->fmr_list;
 742         while (req->nfmr--)
 743                 ib_fmr_pool_unmap(*pfmr++);
 744
 745         ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
 746                         scmnd->sc_data_direction);
 747 }
 748
 749 /**
 750  * srp_claim_req - Take ownership of the scmnd associated with a request.
 751  * @target: SRP target port.
 752  * @req: SRP request.
 753  * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
 754  *         ownership of @req->scmnd if it equals @scmnd.
 755  *
 756  * Return value:
 757  * Either NULL or a pointer to the SCSI command the caller became owner of.
 758  */
 759 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
 760                                        struct srp_request *req,
 761                                        struct scsi_cmnd *scmnd)
 762 {
 763         unsigned long flags;
 764
 765         spin_lock_irqsave(&target->lock, flags);
 766         if (!scmnd) {
 767                 scmnd = req->scmnd;
 768                 req->scmnd = NULL;
 769         } else if (req->scmnd == scmnd) {
 770                 req->scmnd = NULL;
 771         } else {
 772                 scmnd = NULL;
 773         }
 774         spin_unlock_irqrestore(&target->lock, flags);
 775
 776         return scmnd;
 777 }
 778
 779 /**
 780  * srp_free_req() - Unmap data and add request to the free request list.
 781  */
 782 static void srp_free_req(struct srp_target_port *target,
 783                          struct srp_request *req, struct scsi_cmnd *scmnd,
 784                          s32 req_lim_delta)
 785 {
 786         unsigned long flags;
 787
 788         srp_unmap_data(scmnd, target, req);
 789
 790         spin_lock_irqsave(&target->lock, flags);
 791         target->req_lim += req_lim_delta;
 792         list_add_tail(&req->list, &target->free_reqs);
 793         spin_unlock_irqrestore(&target->lock, flags);
 794 }
 795
 796 static void srp_finish_req(struct srp_target_port *target,
 797                            struct srp_request *req, int result)
 798 {
 799         struct scsi_cmnd *scmnd = srp_claim_req(target, req, NULL);
 800
 801         if (scmnd) {
 802                 srp_free_req(target, req, scmnd, 0);
 803                 scmnd->result = result;
 804                 scmnd->scsi_done(scmnd);
 805         }
 806 }
 807
 808 static void srp_terminate_io(struct srp_rport *rport)
 809 {
 810         struct srp_target_port *target = rport->lld_data;
 811         int i;
 812
 813         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
 814                 struct srp_request *req = &target->req_ring[i];
 815                 srp_finish_req(target, req, DID_TRANSPORT_FAILFAST << 16);
 816         }
 817 }
 818
 819 /*
 820  * It is up to the caller to ensure that srp_rport_reconnect() calls are
 821  * serialized and that no concurrent srp_queuecommand(), srp_abort(),
 822  * srp_reset_device() or srp_reset_host() calls will occur while this function
 823  * is in progress. One way to realize that is not to call this function
 824  * directly but to call srp_reconnect_rport() instead since that last function
 825  * serializes calls of this function via rport->mutex and also blocks
 826  * srp_queuecommand() calls before invoking this function.
 827  */
 828 static int srp_rport_reconnect(struct srp_rport *rport)
 829 {
 830         struct srp_target_port *target = rport->lld_data;
 831         int i, ret;
 832
 833         srp_disconnect_target(target);
 834         /*
 835          * Now get a new local CM ID so that we avoid confusing the target in
 836          * case things are really fouled up. Doing so also ensures that all CM
 837          * callbacks will have finished before a new QP is allocated.
 838          */
 839         ret = srp_new_cm_id(target);
 840         /*
 841          * Whether or not creating a new CM ID succeeded, create a new
 842          * QP. This guarantees that all completion callback function
 843          * invocations have finished before request resetting starts.
 844          */
 845         if (ret == 0)
 846                 ret = srp_create_target_ib(target);
 847         else
 848                 srp_create_target_ib(target);
 849
 850         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
 851                 struct srp_request *req = &target->req_ring[i];
 852                 srp_finish_req(target, req, DID_RESET << 16);
 853         }
 854
 855         INIT_LIST_HEAD(&target->free_tx);
 856         for (i = 0; i < SRP_SQ_SIZE; ++i)
 857                 list_add(&target->tx_ring[i]->list, &target->free_tx);
 858
 859         if (ret == 0)
 860                 ret = srp_connect_target(target);
 861
 862         if (ret == 0)
 863                 shost_printk(KERN_INFO, target->scsi_host,
 864                              PFX "reconnect succeeded\n");
 865
 866         return ret;
 867 }
 868
 869 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
 870                          unsigned int dma_len, u32 rkey)
 871 {
 872         struct srp_direct_buf *desc = state->desc;
 873
 874         desc->va = cpu_to_be64(dma_addr);
 875         desc->key = cpu_to_be32(rkey);
 876         desc->len = cpu_to_be32(dma_len);
 877
 878         state->total_len += dma_len;
 879         state->desc++;
 880         state->ndesc++;
 881 }
 882
 883 static int srp_map_finish_fmr(struct srp_map_state *state,
 884                               struct srp_target_port *target)
 885 {
 886         struct srp_device *dev = target->srp_host->srp_dev;
 887         struct ib_pool_fmr *fmr;
 888         u64 io_addr = 0;
 889
 890         if (!state->npages)
 891                 return 0;
 892
 893         if (state->npages == 1) {
 894                 srp_map_desc(state, state->base_dma_addr, state->fmr_len,
 895                              target->rkey);
 896                 state->npages = state->fmr_len = 0;
 897                 return 0;
 898         }
 899
 900         fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
 901                                    state->npages, io_addr);
 902         if (IS_ERR(fmr))
 903                 return PTR_ERR(fmr);
 904
 905         *state->next_fmr++ = fmr;
 906         state->nfmr++;
 907
 908         srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
 909         state->npages = state->fmr_len = 0;
 910         return 0;
 911 }
 912
 913 static void srp_map_update_start(struct srp_map_state *state,
 914                                  struct scatterlist *sg, int sg_index,
 915                                  dma_addr_t dma_addr)
 916 {
 917         state->unmapped_sg = sg;
 918         state->unmapped_index = sg_index;
 919         state->unmapped_addr = dma_addr;
 920 }
 921
 922 static int srp_map_sg_entry(struct srp_map_state *state,
 923                             struct srp_target_port *target,
 924                             struct scatterlist *sg, int sg_index,
 925                             int use_fmr)
 926 {
 927         struct srp_device *dev = target->srp_host->srp_dev;
 928         struct ib_device *ibdev = dev->dev;
 929         dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
 930         unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
 931         unsigned int len;
 932         int ret;
 933
 934         if (!dma_len)
 935                 return 0;
 936
 937         if (use_fmr == SRP_MAP_NO_FMR) {
 938                 /* Once we're in direct map mode for a request, we don't
 939                  * go back to FMR mode, so no need to update anything
 940                  * other than the descriptor.
 941                  */
 942                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
 943                 return 0;
 944         }
 945
 946         /* If we start at an offset into the FMR page, don't merge into
 947          * the current FMR. Finish it out, and use the kernel's MR for this
 948          * sg entry. This is to avoid potential bugs on some SRP targets
 949          * that were never quite defined, but went away when the initiator
 950          * avoided using FMR on such page fragments.
 951          */
 952         if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
 953                 ret = srp_map_finish_fmr(state, target);
 954                 if (ret)
 955                         return ret;
 956
 957                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
 958                 srp_map_update_start(state, NULL, 0, 0);
 959                 return 0;
 960         }
 961
 962         /* If this is the first sg to go into the FMR, save our position.
 963          * We need to know the first unmapped entry, its index, and the
 964          * first unmapped address within that entry to be able to restart
 965          * mapping after an error.
 966          */
 967         if (!state->unmapped_sg)
 968                 srp_map_update_start(state, sg, sg_index, dma_addr);
 969
 970         while (dma_len) {
 971                 if (state->npages == SRP_FMR_SIZE) {
 972                         ret = srp_map_finish_fmr(state, target);
 973                         if (ret)
 974                                 return ret;
 975
 976                         srp_map_update_start(state, sg, sg_index, dma_addr);
 977                 }
 978
 979                 len = min_t(unsigned int, dma_len, dev->fmr_page_size);
 980
 981                 if (!state->npages)
 982                         state->base_dma_addr = dma_addr;
 983                 state->pages[state->npages++] = dma_addr;
 984                 state->fmr_len += len;
 985                 dma_addr += len;
 986                 dma_len -= len;
 987         }
 988
 989         /* If the last entry of the FMR wasn't a full page, then we need to
 990          * close it out and start a new one -- we can only merge at page
 991          * boundries.
 992          */
 993         ret = 0;
 994         if (len != dev->fmr_page_size) {
 995                 ret = srp_map_finish_fmr(state, target);
 996                 if (!ret)
 997                         srp_map_update_start(state, NULL, 0, 0);
 998         }
 999         return ret;
1000 }
1001
1002 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
1003                         struct srp_request *req)
1004 {
1005         struct scatterlist *scat, *sg;
1006         struct srp_cmd *cmd = req->cmd->buf;
1007         int i, len, nents, count, use_fmr;
1008         struct srp_device *dev;
1009         struct ib_device *ibdev;
1010         struct srp_map_state state;
1011         struct srp_indirect_buf *indirect_hdr;
1012         u32 table_len;
1013         u8 fmt;
1014
1015         if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
1016                 return sizeof (struct srp_cmd);
1017
1018         if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
1019             scmnd->sc_data_direction != DMA_TO_DEVICE) {
1020                 shost_printk(KERN_WARNING, target->scsi_host,
1021                              PFX "Unhandled data direction %d\n",
1022                              scmnd->sc_data_direction);
1023                 return -EINVAL;
1024         }
1025
1026         nents = scsi_sg_count(scmnd);
1027         scat  = scsi_sglist(scmnd);
1028
1029         dev = target->srp_host->srp_dev;
1030         ibdev = dev->dev;
1031
1032         count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
1033         if (unlikely(count == 0))
1034                 return -EIO;
1035
1036         fmt = SRP_DATA_DESC_DIRECT;
1037         len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
1038
1039         if (count == 1) {
1040                 /*
1041                  * The midlayer only generated a single gather/scatter
1042                  * entry, or DMA mapping coalesced everything to a
1043                  * single entry.  So a direct descriptor along with
1044                  * the DMA MR suffices.
1045                  */
1046                 struct srp_direct_buf *buf = (void *) cmd->add_data;
1047
1048                 buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
1049                 buf->key = cpu_to_be32(target->rkey);
1050                 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
1051
1052                 req->nfmr = 0;
1053                 goto map_complete;
1054         }
1055
1056         /* We have more than one scatter/gather entry, so build our indirect
1057          * descriptor table, trying to merge as many entries with FMR as we
1058          * can.
1059          */
1060         indirect_hdr = (void *) cmd->add_data;
1061
1062         ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
1063                                    target->indirect_size, DMA_TO_DEVICE);
1064
1065         memset(&state, 0, sizeof(state));
1066         state.desc      = req->indirect_desc;
1067         state.pages     = req->map_page;
1068         state.next_fmr  = req->fmr_list;
1069
1070         use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
1071
1072         for_each_sg(scat, sg, count, i) {
1073                 if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
1074                         /* FMR mapping failed, so backtrack to the first
1075                          * unmapped entry and continue on without using FMR.
1076                          */
1077                         dma_addr_t dma_addr;
1078                         unsigned int dma_len;
1079
1080 backtrack:
1081                         sg = state.unmapped_sg;
1082                         i = state.unmapped_index;
1083
1084                         dma_addr = ib_sg_dma_address(ibdev, sg);
1085                         dma_len = ib_sg_dma_len(ibdev, sg);
1086                         dma_len -= (state.unmapped_addr - dma_addr);
1087                         dma_addr = state.unmapped_addr;
1088                         use_fmr = SRP_MAP_NO_FMR;
1089                         srp_map_desc(&state, dma_addr, dma_len, target->rkey);
1090                 }
1091         }
1092
1093         if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
1094                 goto backtrack;
1095
1096         /* We've mapped the request, now pull as much of the indirect
1097          * descriptor table as we can into the command buffer. If this
1098          * target is not using an external indirect table, we are
1099          * guaranteed to fit into the command, as the SCSI layer won't
1100          * give us more S/G entries than we allow.
1101          */
1102         req->nfmr = state.nfmr;
1103         if (state.ndesc == 1) {
1104                 /* FMR mapping was able to collapse this to one entry,
1105                  * so use a direct descriptor.
1106                  */
1107                 struct srp_direct_buf *buf = (void *) cmd->add_data;
1108
1109                 *buf = req->indirect_desc[0];
1110                 goto map_complete;
1111         }
1112
1113         if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1114                                                 !target->allow_ext_sg)) {
1115                 shost_printk(KERN_ERR, target->scsi_host,
1116                              "Could not fit S/G list into SRP_CMD\n");
1117                 return -EIO;
1118         }
1119
1120         count = min(state.ndesc, target->cmd_sg_cnt);
1121         table_len = state.ndesc * sizeof (struct srp_direct_buf);
1122
1123         fmt = SRP_DATA_DESC_INDIRECT;
1124         len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1125         len += count * sizeof (struct srp_direct_buf);
1126
1127         memcpy(indirect_hdr->desc_list, req->indirect_desc,
1128                count * sizeof (struct srp_direct_buf));
1129
1130         indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1131         indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1132         indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1133         indirect_hdr->len = cpu_to_be32(state.total_len);
1134
1135         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1136                 cmd->data_out_desc_cnt = count;
1137         else
1138                 cmd->data_in_desc_cnt = count;
1139
1140         ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1141                                       DMA_TO_DEVICE);
1142
1143 map_complete:
1144         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1145                 cmd->buf_fmt = fmt << 4;
1146         else
1147                 cmd->buf_fmt = fmt;
1148
1149         return len;
1150 }
1151
1152 /*
1153  * Return an IU and possible credit to the free pool
1154  */
1155 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1156                           enum srp_iu_type iu_type)
1157 {
1158         unsigned long flags;
1159
1160         spin_lock_irqsave(&target->lock, flags);
1161         list_add(&iu->list, &target->free_tx);
1162         if (iu_type != SRP_IU_RSP)
1163                 ++target->req_lim;
1164         spin_unlock_irqrestore(&target->lock, flags);
1165 }
1166
1167 /*
1168  * Must be called with target->lock held to protect req_lim and free_tx.
1169  * If IU is not sent, it must be returned using srp_put_tx_iu().
1170  *
1171  * Note:
1172  * An upper limit for the number of allocated information units for each
1173  * request type is:
1174  * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1175  *   more than Scsi_Host.can_queue requests.
1176  * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1177  * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1178  *   one unanswered SRP request to an initiator.
1179  */
1180 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1181                                       enum srp_iu_type iu_type)
1182 {
1183         s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1184         struct srp_iu *iu;
1185
1186         srp_send_completion(target->send_cq, target);
1187
1188         if (list_empty(&target->free_tx))
1189                 return NULL;
1190
1191         /* Initiator responses to target requests do not consume credits */
1192         if (iu_type != SRP_IU_RSP) {
1193                 if (target->req_lim <= rsv) {
1194                         ++target->zero_req_lim;
1195                         return NULL;
1196                 }
1197
1198                 --target->req_lim;
1199         }
1200
1201         iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1202         list_del(&iu->list);
1203         return iu;
1204 }
1205
1206 static int srp_post_send(struct srp_target_port *target,
1207                          struct srp_iu *iu, int len)
1208 {
1209         struct ib_sge list;
1210         struct ib_send_wr wr, *bad_wr;
1211
1212         list.addr   = iu->dma;
1213         list.length = len;
1214         list.lkey   = target->lkey;
1215
1216         wr.next       = NULL;
1217         wr.wr_id      = (uintptr_t) iu;
1218         wr.sg_list    = &list;
1219         wr.num_sge    = 1;
1220         wr.opcode     = IB_WR_SEND;
1221         wr.send_flags = IB_SEND_SIGNALED;
1222
1223         return ib_post_send(target->qp, &wr, &bad_wr);
1224 }
1225
1226 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1227 {
1228         struct ib_recv_wr wr, *bad_wr;
1229         struct ib_sge list;
1230
1231         list.addr   = iu->dma;
1232         list.length = iu->size;
1233         list.lkey   = target->lkey;
1234
1235         wr.next     = NULL;
1236         wr.wr_id    = (uintptr_t) iu;
1237         wr.sg_list  = &list;
1238         wr.num_sge  = 1;
1239
1240         return ib_post_recv(target->qp, &wr, &bad_wr);
1241 }
1242
1243 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1244 {
1245         struct srp_request *req;
1246         struct scsi_cmnd *scmnd;
1247         unsigned long flags;
1248
1249         if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1250                 spin_lock_irqsave(&target->lock, flags);
1251                 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1252                 spin_unlock_irqrestore(&target->lock, flags);
1253
1254                 target->tsk_mgmt_status = -1;
1255                 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1256                         target->tsk_mgmt_status = rsp->data[3];
1257                 complete(&target->tsk_mgmt_done);
1258         } else {
1259                 req = &target->req_ring[rsp->tag];
1260                 scmnd = srp_claim_req(target, req, NULL);
1261                 if (!scmnd) {
1262                         shost_printk(KERN_ERR, target->scsi_host,
1263                                      "Null scmnd for RSP w/tag %016llx\n",
1264                                      (unsigned long long) rsp->tag);
1265
1266                         spin_lock_irqsave(&target->lock, flags);
1267                         target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1268                         spin_unlock_irqrestore(&target->lock, flags);
1269
1270                         return;
1271                 }
1272                 scmnd->result = rsp->status;
1273
1274                 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1275                         memcpy(scmnd->sense_buffer, rsp->data +
1276                                be32_to_cpu(rsp->resp_data_len),
1277                                min_t(int, be32_to_cpu(rsp->sense_data_len),
1278                                      SCSI_SENSE_BUFFERSIZE));
1279                 }
1280
1281                 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1282                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1283                 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1284                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1285
1286                 srp_free_req(target, req, scmnd,
1287                              be32_to_cpu(rsp->req_lim_delta));
1288
1289                 scmnd->host_scribble = NULL;
1290                 scmnd->scsi_done(scmnd);
1291         }
1292 }
1293
1294 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1295                                void *rsp, int len)
1296 {
1297         struct ib_device *dev = target->srp_host->srp_dev->dev;
1298         unsigned long flags;
1299         struct srp_iu *iu;
1300         int err;
1301
1302         spin_lock_irqsave(&target->lock, flags);
1303         target->req_lim += req_delta;
1304         iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1305         spin_unlock_irqrestore(&target->lock, flags);
1306
1307         if (!iu) {
1308                 shost_printk(KERN_ERR, target->scsi_host, PFX
1309                              "no IU available to send response\n");
1310                 return 1;
1311         }
1312
1313         ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1314         memcpy(iu->buf, rsp, len);
1315         ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1316
1317         err = srp_post_send(target, iu, len);
1318         if (err) {
1319                 shost_printk(KERN_ERR, target->scsi_host, PFX
1320                              "unable to post response: %d\n", err);
1321                 srp_put_tx_iu(target, iu, SRP_IU_RSP);
1322         }
1323
1324         return err;
1325 }
1326
1327 static void srp_process_cred_req(struct srp_target_port *target,
1328                                  struct srp_cred_req *req)
1329 {
1330         struct srp_cred_rsp rsp = {
1331                 .opcode = SRP_CRED_RSP,
1332                 .tag = req->tag,
1333         };
1334         s32 delta = be32_to_cpu(req->req_lim_delta);
1335
1336         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1337                 shost_printk(KERN_ERR, target->scsi_host, PFX
1338                              "problems processing SRP_CRED_REQ\n");
1339 }
1340
1341 static void srp_process_aer_req(struct srp_target_port *target,
1342                                 struct srp_aer_req *req)
1343 {
1344         struct srp_aer_rsp rsp = {
1345                 .opcode = SRP_AER_RSP,
1346                 .tag = req->tag,
1347         };
1348         s32 delta = be32_to_cpu(req->req_lim_delta);
1349
1350         shost_printk(KERN_ERR, target->scsi_host, PFX
1351                      "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1352
1353         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1354                 shost_printk(KERN_ERR, target->scsi_host, PFX
1355                              "problems processing SRP_AER_REQ\n");
1356 }
1357
1358 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1359 {
1360         struct ib_device *dev = target->srp_host->srp_dev->dev;
1361         struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1362         int res;
1363         u8 opcode;
1364
1365         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1366                                    DMA_FROM_DEVICE);
1367
1368         opcode = *(u8 *) iu->buf;
1369
1370         if (0) {
1371                 shost_printk(KERN_ERR, target->scsi_host,
1372                              PFX "recv completion, opcode 0x%02x\n", opcode);
1373                 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1374                                iu->buf, wc->byte_len, true);
1375         }
1376
1377         switch (opcode) {
1378         case SRP_RSP:
1379                 srp_process_rsp(target, iu->buf);
1380                 break;
1381
1382         case SRP_CRED_REQ:
1383                 srp_process_cred_req(target, iu->buf);
1384                 break;
1385
1386         case SRP_AER_REQ:
1387                 srp_process_aer_req(target, iu->buf);
1388                 break;
1389
1390         case SRP_T_LOGOUT:
1391                 /* XXX Handle target logout */
1392                 shost_printk(KERN_WARNING, target->scsi_host,
1393                              PFX "Got target logout request\n");
1394                 break;
1395
1396         default:
1397                 shost_printk(KERN_WARNING, target->scsi_host,
1398                              PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1399                 break;
1400         }
1401
1402         ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1403                                       DMA_FROM_DEVICE);
1404
1405         res = srp_post_recv(target, iu);
1406         if (res != 0)
1407                 shost_printk(KERN_ERR, target->scsi_host,
1408                              PFX "Recv failed with error code %d\n", res);
1409 }
1410
1411 /**
1412  * srp_tl_err_work() - handle a transport layer error
1413  *
1414  * Note: This function may get invoked before the rport has been created,
1415  * hence the target->rport test.
1416  */
1417 static void srp_tl_err_work(struct work_struct *work)
1418 {
1419         struct srp_target_port *target;
1420
1421         target = container_of(work, struct srp_target_port, tl_err_work);
1422         if (target->rport)
1423                 srp_start_tl_fail_timers(target->rport);
1424 }
1425
1426 static void srp_handle_qp_err(enum ib_wc_status wc_status,
1427                               enum ib_wc_opcode wc_opcode,
1428                               struct srp_target_port *target)
1429 {
1430         if (target->connected && !target->qp_in_error) {
1431                 shost_printk(KERN_ERR, target->scsi_host,
1432                              PFX "failed %s status %d\n",
1433                              wc_opcode & IB_WC_RECV ? "receive" : "send",
1434                              wc_status);
1435                 queue_work(system_long_wq, &target->tl_err_work);
1436         }
1437         target->qp_in_error = true;
1438 }
1439
1440 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1441 {
1442         struct srp_target_port *target = target_ptr;
1443         struct ib_wc wc;
1444
1445         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1446         while (ib_poll_cq(cq, 1, &wc) > 0) {
1447                 if (likely(wc.status == IB_WC_SUCCESS)) {
1448                         srp_handle_recv(target, &wc);
1449                 } else {
1450                         srp_handle_qp_err(wc.status, wc.opcode, target);
1451                 }
1452         }
1453 }
1454
1455 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1456 {
1457         struct srp_target_port *target = target_ptr;
1458         struct ib_wc wc;
1459         struct srp_iu *iu;
1460
1461         while (ib_poll_cq(cq, 1, &wc) > 0) {
1462                 if (likely(wc.status == IB_WC_SUCCESS)) {
1463                         iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1464                         list_add(&iu->list, &target->free_tx);
1465                 } else {
1466                         srp_handle_qp_err(wc.status, wc.opcode, target);
1467                 }
1468         }
1469 }
1470
1471 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1472 {
1473         struct srp_target_port *target = host_to_target(shost);
1474         struct srp_rport *rport = target->rport;
1475         struct srp_request *req;
1476         struct srp_iu *iu;
1477         struct srp_cmd *cmd;
1478         struct ib_device *dev;
1479         unsigned long flags;
1480         int len, result;
1481         const bool in_scsi_eh = !in_interrupt() && current == shost->ehandler;
1482
1483         /*
1484          * The SCSI EH thread is the only context from which srp_queuecommand()
1485          * can get invoked for blocked devices (SDEV_BLOCK /
1486          * SDEV_CREATED_BLOCK). Avoid racing with srp_reconnect_rport() by
1487          * locking the rport mutex if invoked from inside the SCSI EH.
1488          */
1489         if (in_scsi_eh)
1490                 mutex_lock(&rport->mutex);
1491
1492         result = srp_chkready(target->rport);
1493         if (unlikely(result)) {
1494                 scmnd->result = result;
1495                 scmnd->scsi_done(scmnd);
1496                 goto unlock_rport;
1497         }
1498
1499         spin_lock_irqsave(&target->lock, flags);
1500         iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1501         if (!iu)
1502                 goto err_unlock;
1503
1504         req = list_first_entry(&target->free_reqs, struct srp_request, list);
1505         list_del(&req->list);
1506         spin_unlock_irqrestore(&target->lock, flags);
1507
1508         dev = target->srp_host->srp_dev->dev;
1509         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1510                                    DMA_TO_DEVICE);
1511
1512         scmnd->result        = 0;
1513         scmnd->host_scribble = (void *) req;
1514
1515         cmd = iu->buf;
1516         memset(cmd, 0, sizeof *cmd);
1517
1518         cmd->opcode = SRP_CMD;
1519         cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
1520         cmd->tag    = req->index;
1521         memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1522
1523         req->scmnd    = scmnd;
1524         req->cmd      = iu;
1525
1526         len = srp_map_data(scmnd, target, req);
1527         if (len < 0) {
1528                 shost_printk(KERN_ERR, target->scsi_host,
1529                              PFX "Failed to map data\n");
1530                 goto err_iu;
1531         }
1532
1533         ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1534                                       DMA_TO_DEVICE);
1535
1536         if (srp_post_send(target, iu, len)) {
1537                 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1538                 goto err_unmap;
1539         }
1540
1541 unlock_rport:
1542         if (in_scsi_eh)
1543                 mutex_unlock(&rport->mutex);
1544
1545         return 0;
1546
1547 err_unmap:
1548         srp_unmap_data(scmnd, target, req);
1549
1550 err_iu:
1551         srp_put_tx_iu(target, iu, SRP_IU_CMD);
1552
1553         spin_lock_irqsave(&target->lock, flags);
1554         list_add(&req->list, &target->free_reqs);
1555
1556 err_unlock:
1557         spin_unlock_irqrestore(&target->lock, flags);
1558
1559         if (in_scsi_eh)
1560                 mutex_unlock(&rport->mutex);
1561
1562         return SCSI_MLQUEUE_HOST_BUSY;
1563 }
1564
1565 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1566 {
1567         int i;
1568
1569         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1570                 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1571                                                   target->max_ti_iu_len,
1572                                                   GFP_KERNEL, DMA_FROM_DEVICE);
1573                 if (!target->rx_ring[i])
1574                         goto err;
1575         }
1576
1577         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1578                 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1579                                                   target->max_iu_len,
1580                                                   GFP_KERNEL, DMA_TO_DEVICE);
1581                 if (!target->tx_ring[i])
1582                         goto err;
1583
1584                 list_add(&target->tx_ring[i]->list, &target->free_tx);
1585         }
1586
1587         return 0;
1588
1589 err:
1590         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1591                 srp_free_iu(target->srp_host, target->rx_ring[i]);
1592                 target->rx_ring[i] = NULL;
1593         }
1594
1595         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1596                 srp_free_iu(target->srp_host, target->tx_ring[i]);
1597                 target->tx_ring[i] = NULL;
1598         }
1599
1600         return -ENOMEM;
1601 }
1602
1603 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
1604 {
1605         uint64_t T_tr_ns, max_compl_time_ms;
1606         uint32_t rq_tmo_jiffies;
1607
1608         /*
1609          * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
1610          * table 91), both the QP timeout and the retry count have to be set
1611          * for RC QP's during the RTR to RTS transition.
1612          */
1613         WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
1614                      (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
1615
1616         /*
1617          * Set target->rq_tmo_jiffies to one second more than the largest time
1618          * it can take before an error completion is generated. See also
1619          * C9-140..142 in the IBTA spec for more information about how to
1620          * convert the QP Local ACK Timeout value to nanoseconds.
1621          */
1622         T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
1623         max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
1624         do_div(max_compl_time_ms, NSEC_PER_MSEC);
1625         rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
1626
1627         return rq_tmo_jiffies;
1628 }
1629
1630 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1631                                struct srp_login_rsp *lrsp,
1632                                struct srp_target_port *target)
1633 {
1634         struct ib_qp_attr *qp_attr = NULL;
1635         int attr_mask = 0;
1636         int ret;
1637         int i;
1638
1639         if (lrsp->opcode == SRP_LOGIN_RSP) {
1640                 target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1641                 target->req_lim       = be32_to_cpu(lrsp->req_lim_delta);
1642
1643                 /*
1644                  * Reserve credits for task management so we don't
1645                  * bounce requests back to the SCSI mid-layer.
1646                  */
1647                 target->scsi_host->can_queue
1648                         = min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1649                               target->scsi_host->can_queue);
1650         } else {
1651                 shost_printk(KERN_WARNING, target->scsi_host,
1652                              PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1653                 ret = -ECONNRESET;
1654                 goto error;
1655         }
1656
1657         if (!target->rx_ring[0]) {
1658                 ret = srp_alloc_iu_bufs(target);
1659                 if (ret)
1660                         goto error;
1661         }
1662
1663         ret = -ENOMEM;
1664         qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1665         if (!qp_attr)
1666                 goto error;
1667
1668         qp_attr->qp_state = IB_QPS_RTR;
1669         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1670         if (ret)
1671                 goto error_free;
1672
1673         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1674         if (ret)
1675                 goto error_free;
1676
1677         for (i = 0; i < SRP_RQ_SIZE; i++) {
1678                 struct srp_iu *iu = target->rx_ring[i];
1679                 ret = srp_post_recv(target, iu);
1680                 if (ret)
1681                         goto error_free;
1682         }
1683
1684         qp_attr->qp_state = IB_QPS_RTS;
1685         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1686         if (ret)
1687                 goto error_free;
1688
1689         target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
1690
1691         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1692         if (ret)
1693                 goto error_free;
1694
1695         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1696
1697 error_free:
1698         kfree(qp_attr);
1699
1700 error:
1701         target->status = ret;
1702 }
1703
1704 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1705                                struct ib_cm_event *event,
1706                                struct srp_target_port *target)
1707 {
1708         struct Scsi_Host *shost = target->scsi_host;
1709         struct ib_class_port_info *cpi;
1710         int opcode;
1711
1712         switch (event->param.rej_rcvd.reason) {
1713         case IB_CM_REJ_PORT_CM_REDIRECT:
1714                 cpi = event->param.rej_rcvd.ari;
1715                 target->path.dlid = cpi->redirect_lid;
1716                 target->path.pkey = cpi->redirect_pkey;
1717                 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1718                 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1719
1720                 target->status = target->path.dlid ?
1721                         SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1722                 break;
1723
1724         case IB_CM_REJ_PORT_REDIRECT:
1725                 if (srp_target_is_topspin(target)) {
1726                         /*
1727                          * Topspin/Cisco SRP gateways incorrectly send
1728                          * reject reason code 25 when they mean 24
1729                          * (port redirect).
1730                          */
1731                         memcpy(target->path.dgid.raw,
1732                                event->param.rej_rcvd.ari, 16);
1733
1734                         shost_printk(KERN_DEBUG, shost,
1735                                      PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1736                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1737                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1738
1739                         target->status = SRP_PORT_REDIRECT;
1740                 } else {
1741                         shost_printk(KERN_WARNING, shost,
1742                                      "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1743                         target->status = -ECONNRESET;
1744                 }
1745                 break;
1746
1747         case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1748                 shost_printk(KERN_WARNING, shost,
1749                             "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1750                 target->status = -ECONNRESET;
1751                 break;
1752
1753         case IB_CM_REJ_CONSUMER_DEFINED:
1754                 opcode = *(u8 *) event->private_data;
1755                 if (opcode == SRP_LOGIN_REJ) {
1756                         struct srp_login_rej *rej = event->private_data;
1757                         u32 reason = be32_to_cpu(rej->reason);
1758
1759                         if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1760                                 shost_printk(KERN_WARNING, shost,
1761                                              PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1762                         else
1763                                 shost_printk(KERN_WARNING, shost,
1764                                             PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1765                 } else
1766                         shost_printk(KERN_WARNING, shost,
1767                                      "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1768                                      " opcode 0x%02x\n", opcode);
1769                 target->status = -ECONNRESET;
1770                 break;
1771
1772         case IB_CM_REJ_STALE_CONN:
1773                 shost_printk(KERN_WARNING, shost, "  REJ reason: stale connection\n");
1774                 target->status = SRP_STALE_CONN;
1775                 break;
1776
1777         default:
1778                 shost_printk(KERN_WARNING, shost, "  REJ reason 0x%x\n",
1779                              event->param.rej_rcvd.reason);
1780                 target->status = -ECONNRESET;
1781         }
1782 }
1783
1784 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1785 {
1786         struct srp_target_port *target = cm_id->context;
1787         int comp = 0;
1788
1789         switch (event->event) {
1790         case IB_CM_REQ_ERROR:
1791                 shost_printk(KERN_DEBUG, target->scsi_host,
1792                              PFX "Sending CM REQ failed\n");
1793                 comp = 1;
1794                 target->status = -ECONNRESET;
1795                 break;
1796
1797         case IB_CM_REP_RECEIVED:
1798                 comp = 1;
1799                 srp_cm_rep_handler(cm_id, event->private_data, target);
1800                 break;
1801
1802         case IB_CM_REJ_RECEIVED:
1803                 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1804                 comp = 1;
1805
1806                 srp_cm_rej_handler(cm_id, event, target);
1807                 break;
1808
1809         case IB_CM_DREQ_RECEIVED:
1810                 shost_printk(KERN_WARNING, target->scsi_host,
1811                              PFX "DREQ received - connection closed\n");
1812                 srp_change_conn_state(target, false);
1813                 if (ib_send_cm_drep(cm_id, NULL, 0))
1814                         shost_printk(KERN_ERR, target->scsi_host,
1815                                      PFX "Sending CM DREP failed\n");
1816                 queue_work(system_long_wq, &target->tl_err_work);
1817                 break;
1818
1819         case IB_CM_TIMEWAIT_EXIT:
1820                 shost_printk(KERN_ERR, target->scsi_host,
1821                              PFX "connection closed\n");
1822
1823                 target->status = 0;
1824                 break;
1825
1826         case IB_CM_MRA_RECEIVED:
1827         case IB_CM_DREQ_ERROR:
1828         case IB_CM_DREP_RECEIVED:
1829                 break;
1830
1831         default:
1832                 shost_printk(KERN_WARNING, target->scsi_host,
1833                              PFX "Unhandled CM event %d\n", event->event);
1834                 break;
1835         }
1836
1837         if (comp)
1838                 complete(&target->done);
1839
1840         return 0;
1841 }
1842
1843 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1844                              u64 req_tag, unsigned int lun, u8 func)
1845 {
1846         struct srp_rport *rport = target->rport;
1847         struct ib_device *dev = target->srp_host->srp_dev->dev;
1848         struct srp_iu *iu;
1849         struct srp_tsk_mgmt *tsk_mgmt;
1850
1851         if (!target->connected || target->qp_in_error)
1852                 return -1;
1853
1854         init_completion(&target->tsk_mgmt_done);
1855
1856         /*
1857          * Lock the rport mutex to avoid that srp_create_target_ib() is
1858          * invoked while a task management function is being sent.
1859          */
1860         mutex_lock(&rport->mutex);
1861         spin_lock_irq(&target->lock);
1862         iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1863         spin_unlock_irq(&target->lock);
1864
1865         if (!iu) {
1866                 mutex_unlock(&rport->mutex);
1867
1868                 return -1;
1869         }
1870
1871         ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1872                                    DMA_TO_DEVICE);
1873         tsk_mgmt = iu->buf;
1874         memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1875
1876         tsk_mgmt->opcode        = SRP_TSK_MGMT;
1877         tsk_mgmt->lun           = cpu_to_be64((u64) lun << 48);
1878         tsk_mgmt->tag           = req_tag | SRP_TAG_TSK_MGMT;
1879         tsk_mgmt->tsk_mgmt_func = func;
1880         tsk_mgmt->task_tag      = req_tag;
1881
1882         ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1883                                       DMA_TO_DEVICE);
1884         if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1885                 srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1886                 mutex_unlock(&rport->mutex);
1887
1888                 return -1;
1889         }
1890         mutex_unlock(&rport->mutex);
1891
1892         if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1893                                          msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1894                 return -1;
1895
1896         return 0;
1897 }
1898
1899 static int srp_abort(struct scsi_cmnd *scmnd)
1900 {
1901         struct srp_target_port *target = host_to_target(scmnd->device->host);
1902         struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1903         int ret;
1904
1905         shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
1906
1907         if (!req || !srp_claim_req(target, req, scmnd))
1908                 return FAILED;
1909         if (srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
1910                               SRP_TSK_ABORT_TASK) == 0)
1911                 ret = SUCCESS;
1912         else if (target->rport->state == SRP_RPORT_LOST)
1913                 ret = FAST_IO_FAIL;
1914         else
1915                 ret = FAILED;
1916         srp_free_req(target, req, scmnd, 0);
1917         scmnd->result = DID_ABORT << 16;
1918         scmnd->scsi_done(scmnd);
1919
1920         return ret;
1921 }
1922
1923 static int srp_reset_device(struct scsi_cmnd *scmnd)
1924 {
1925         struct srp_target_port *target = host_to_target(scmnd->device->host);
1926         int i;
1927
1928         shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
1929
1930         if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
1931                               SRP_TSK_LUN_RESET))
1932                 return FAILED;
1933         if (target->tsk_mgmt_status)
1934                 return FAILED;
1935
1936         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1937                 struct srp_request *req = &target->req_ring[i];
1938                 if (req->scmnd && req->scmnd->device == scmnd->device)
1939                         srp_finish_req(target, req, DID_RESET << 16);
1940         }
1941
1942         return SUCCESS;
1943 }
1944
1945 static int srp_reset_host(struct scsi_cmnd *scmnd)
1946 {
1947         struct srp_target_port *target = host_to_target(scmnd->device->host);
1948
1949         shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
1950
1951         return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
1952 }
1953
1954 static int srp_slave_configure(struct scsi_device *sdev)
1955 {
1956         struct Scsi_Host *shost = sdev->host;
1957         struct srp_target_port *target = host_to_target(shost);
1958         struct request_queue *q = sdev->request_queue;
1959         unsigned long timeout;
1960
1961         if (sdev->type == TYPE_DISK) {
1962                 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
1963                 blk_queue_rq_timeout(q, timeout);
1964         }
1965
1966         return 0;
1967 }
1968
1969 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
1970                            char *buf)
1971 {
1972         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1973
1974         return sprintf(buf, "0x%016llx\n",
1975                        (unsigned long long) be64_to_cpu(target->id_ext));
1976 }
1977
1978 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
1979                              char *buf)
1980 {
1981         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1982
1983         return sprintf(buf, "0x%016llx\n",
1984                        (unsigned long long) be64_to_cpu(target->ioc_guid));
1985 }
1986
1987 static ssize_t show_service_id(struct device *dev,
1988                                struct device_attribute *attr, char *buf)
1989 {
1990         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1991
1992         return sprintf(buf, "0x%016llx\n",
1993                        (unsigned long long) be64_to_cpu(target->service_id));
1994 }
1995
1996 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
1997                          char *buf)
1998 {
1999         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2000
2001         return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
2002 }
2003
2004 static ssize_t show_sgid(struct device *dev, struct device_attribute *attr,
2005                          char *buf)
2006 {
2007         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2008
2009         return sprintf(buf, "%pI6\n", target->path.sgid.raw);
2010 }
2011
2012 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
2013                          char *buf)
2014 {
2015         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2016
2017         return sprintf(buf, "%pI6\n", target->path.dgid.raw);
2018 }
2019
2020 static ssize_t show_orig_dgid(struct device *dev,
2021                               struct device_attribute *attr, char *buf)
2022 {
2023         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2024
2025         return sprintf(buf, "%pI6\n", target->orig_dgid);
2026 }
2027
2028 static ssize_t show_req_lim(struct device *dev,
2029                             struct device_attribute *attr, char *buf)
2030 {
2031         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2032
2033         return sprintf(buf, "%d\n", target->req_lim);
2034 }
2035
2036 static ssize_t show_zero_req_lim(struct device *dev,
2037                                  struct device_attribute *attr, char *buf)
2038 {
2039         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2040
2041         return sprintf(buf, "%d\n", target->zero_req_lim);
2042 }
2043
2044 static ssize_t show_local_ib_port(struct device *dev,
2045                                   struct device_attribute *attr, char *buf)
2046 {
2047         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2048
2049         return sprintf(buf, "%d\n", target->srp_host->port);
2050 }
2051
2052 static ssize_t show_local_ib_device(struct device *dev,
2053                                     struct device_attribute *attr, char *buf)
2054 {
2055         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2056
2057         return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
2058 }
2059
2060 static ssize_t show_comp_vector(struct device *dev,
2061                                 struct device_attribute *attr, char *buf)
2062 {
2063         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2064
2065         return sprintf(buf, "%d\n", target->comp_vector);
2066 }
2067
2068 static ssize_t show_tl_retry_count(struct device *dev,
2069                                    struct device_attribute *attr, char *buf)
2070 {
2071         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2072
2073         return sprintf(buf, "%d\n", target->tl_retry_count);
2074 }
2075
2076 static ssize_t show_cmd_sg_entries(struct device *dev,
2077                                    struct device_attribute *attr, char *buf)
2078 {
2079         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2080
2081         return sprintf(buf, "%u\n", target->cmd_sg_cnt);
2082 }
2083
2084 static ssize_t show_allow_ext_sg(struct device *dev,
2085                                  struct device_attribute *attr, char *buf)
2086 {
2087         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2088
2089         return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
2090 }
2091
2092 static DEVICE_ATTR(id_ext,          S_IRUGO, show_id_ext,          NULL);
2093 static DEVICE_ATTR(ioc_guid,        S_IRUGO, show_ioc_guid,        NULL);
2094 static DEVICE_ATTR(service_id,      S_IRUGO, show_service_id,      NULL);
2095 static DEVICE_ATTR(pkey,            S_IRUGO, show_pkey,            NULL);
2096 static DEVICE_ATTR(sgid,            S_IRUGO, show_sgid,            NULL);
2097 static DEVICE_ATTR(dgid,            S_IRUGO, show_dgid,            NULL);
2098 static DEVICE_ATTR(orig_dgid,       S_IRUGO, show_orig_dgid,       NULL);
2099 static DEVICE_ATTR(req_lim,         S_IRUGO, show_req_lim,         NULL);
2100 static DEVICE_ATTR(zero_req_lim,    S_IRUGO, show_zero_req_lim,    NULL);
2101 static DEVICE_ATTR(local_ib_port,   S_IRUGO, show_local_ib_port,   NULL);
2102 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
2103 static DEVICE_ATTR(comp_vector,     S_IRUGO, show_comp_vector,     NULL);
2104 static DEVICE_ATTR(tl_retry_count,  S_IRUGO, show_tl_retry_count,  NULL);
2105 static DEVICE_ATTR(cmd_sg_entries,  S_IRUGO, show_cmd_sg_entries,  NULL);
2106 static DEVICE_ATTR(allow_ext_sg,    S_IRUGO, show_allow_ext_sg,    NULL);
2107
2108 static struct device_attribute *srp_host_attrs[] = {
2109         &dev_attr_id_ext,
2110         &dev_attr_ioc_guid,
2111         &dev_attr_service_id,
2112         &dev_attr_pkey,
2113         &dev_attr_sgid,
2114         &dev_attr_dgid,
2115         &dev_attr_orig_dgid,
2116         &dev_attr_req_lim,
2117         &dev_attr_zero_req_lim,
2118         &dev_attr_local_ib_port,
2119         &dev_attr_local_ib_device,
2120         &dev_attr_comp_vector,
2121         &dev_attr_tl_retry_count,
2122         &dev_attr_cmd_sg_entries,
2123         &dev_attr_allow_ext_sg,
2124         NULL
2125 };
2126
2127 static struct scsi_host_template srp_template = {
2128         .module                         = THIS_MODULE,
2129         .name                           = "InfiniBand SRP initiator",
2130         .proc_name                      = DRV_NAME,
2131         .slave_configure                = srp_slave_configure,
2132         .info                           = srp_target_info,
2133         .queuecommand                   = srp_queuecommand,
2134         .eh_abort_handler               = srp_abort,
2135         .eh_device_reset_handler        = srp_reset_device,
2136         .eh_host_reset_handler          = srp_reset_host,
2137         .skip_settle_delay              = true,
2138         .sg_tablesize                   = SRP_DEF_SG_TABLESIZE,
2139         .can_queue                      = SRP_CMD_SQ_SIZE,
2140         .this_id                        = -1,
2141         .cmd_per_lun                    = SRP_CMD_SQ_SIZE,
2142         .use_clustering                 = ENABLE_CLUSTERING,
2143         .shost_attrs                    = srp_host_attrs
2144 };
2145
2146 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
2147 {
2148         struct srp_rport_identifiers ids;
2149         struct srp_rport *rport;
2150
2151         sprintf(target->target_name, "SRP.T10:%016llX",
2152                  (unsigned long long) be64_to_cpu(target->id_ext));
2153
2154         if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
2155                 return -ENODEV;
2156
2157         memcpy(ids.port_id, &target->id_ext, 8);
2158         memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2159         ids.roles = SRP_RPORT_ROLE_TARGET;
2160         rport = srp_rport_add(target->scsi_host, &ids);
2161         if (IS_ERR(rport)) {
2162                 scsi_remove_host(target->scsi_host);
2163                 return PTR_ERR(rport);
2164         }
2165
2166         rport->lld_data = target;
2167         target->rport = rport;
2168
2169         spin_lock(&host->target_lock);
2170         list_add_tail(&target->list, &host->target_list);
2171         spin_unlock(&host->target_lock);
2172
2173         target->state = SRP_TARGET_LIVE;
2174
2175         scsi_scan_target(&target->scsi_host->shost_gendev,
2176                          0, target->scsi_id, SCAN_WILD_CARD, 0);
2177
2178         return 0;
2179 }
2180
2181 static void srp_release_dev(struct device *dev)
2182 {
2183         struct srp_host *host =
2184                 container_of(dev, struct srp_host, dev);
2185
2186         complete(&host->released);
2187 }
2188
2189 static struct class srp_class = {
2190         .name    = "infiniband_srp",
2191         .dev_release = srp_release_dev
2192 };
2193
2194 /**
2195  * srp_conn_unique() - check whether the connection to a target is unique
2196  */
2197 static bool srp_conn_unique(struct srp_host *host,
2198                             struct srp_target_port *target)
2199 {
2200         struct srp_target_port *t;
2201         bool ret = false;
2202
2203         if (target->state == SRP_TARGET_REMOVED)
2204                 goto out;
2205
2206         ret = true;
2207
2208         spin_lock(&host->target_lock);
2209         list_for_each_entry(t, &host->target_list, list) {
2210                 if (t != target &&
2211                     target->id_ext == t->id_ext &&
2212                     target->ioc_guid == t->ioc_guid &&
2213                     target->initiator_ext == t->initiator_ext) {
2214                         ret = false;
2215                         break;
2216                 }
2217         }
2218         spin_unlock(&host->target_lock);
2219
2220 out:
2221         return ret;
2222 }
2223
2224 /*
2225  * Target ports are added by writing
2226  *
2227  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2228  *     pkey=<P_Key>,service_id=<service ID>
2229  *
2230  * to the add_target sysfs attribute.
2231  */
2232 enum {
2233         SRP_OPT_ERR             = 0,
2234         SRP_OPT_ID_EXT          = 1 << 0,
2235         SRP_OPT_IOC_GUID        = 1 << 1,
2236         SRP_OPT_DGID            = 1 << 2,
2237         SRP_OPT_PKEY            = 1 << 3,
2238         SRP_OPT_SERVICE_ID      = 1 << 4,
2239         SRP_OPT_MAX_SECT        = 1 << 5,
2240         SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2241         SRP_OPT_IO_CLASS        = 1 << 7,
2242         SRP_OPT_INITIATOR_EXT   = 1 << 8,
2243         SRP_OPT_CMD_SG_ENTRIES  = 1 << 9,
2244         SRP_OPT_ALLOW_EXT_SG    = 1 << 10,
2245         SRP_OPT_SG_TABLESIZE    = 1 << 11,
2246         SRP_OPT_COMP_VECTOR     = 1 << 12,
2247         SRP_OPT_TL_RETRY_COUNT  = 1 << 13,
2248         SRP_OPT_ALL             = (SRP_OPT_ID_EXT       |
2249                                    SRP_OPT_IOC_GUID     |
2250                                    SRP_OPT_DGID         |
2251                                    SRP_OPT_PKEY         |
2252                                    SRP_OPT_SERVICE_ID),
2253 };
2254
2255 static const match_table_t srp_opt_tokens = {
2256         { SRP_OPT_ID_EXT,               "id_ext=%s"             },
2257         { SRP_OPT_IOC_GUID,             "ioc_guid=%s"           },
2258         { SRP_OPT_DGID,                 "dgid=%s"               },
2259         { SRP_OPT_PKEY,                 "pkey=%x"               },
2260         { SRP_OPT_SERVICE_ID,           "service_id=%s"         },
2261         { SRP_OPT_MAX_SECT,             "max_sect=%d"           },
2262         { SRP_OPT_MAX_CMD_PER_LUN,      "max_cmd_per_lun=%d"    },
2263         { SRP_OPT_IO_CLASS,             "io_class=%x"           },
2264         { SRP_OPT_INITIATOR_EXT,        "initiator_ext=%s"      },
2265         { SRP_OPT_CMD_SG_ENTRIES,       "cmd_sg_entries=%u"     },
2266         { SRP_OPT_ALLOW_EXT_SG,         "allow_ext_sg=%u"       },
2267         { SRP_OPT_SG_TABLESIZE,         "sg_tablesize=%u"       },
2268         { SRP_OPT_COMP_VECTOR,          "comp_vector=%u"        },
2269         { SRP_OPT_TL_RETRY_COUNT,       "tl_retry_count=%u"     },
2270         { SRP_OPT_ERR,                  NULL                    }
2271 };
2272
2273 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2274 {
2275         char *options, *sep_opt;
2276         char *p;
2277         char dgid[3];
2278         substring_t args[MAX_OPT_ARGS];
2279         int opt_mask = 0;
2280         int token;
2281         int ret = -EINVAL;
2282         int i;
2283
2284         options = kstrdup(buf, GFP_KERNEL);
2285         if (!options)
2286                 return -ENOMEM;
2287
2288         sep_opt = options;
2289         while ((p = strsep(&sep_opt, ",")) != NULL) {
2290                 if (!*p)
2291                         continue;
2292
2293                 token = match_token(p, srp_opt_tokens, args);
2294                 opt_mask |= token;
2295
2296                 switch (token) {
2297                 case SRP_OPT_ID_EXT:
2298                         p = match_strdup(args);
2299                         if (!p) {
2300                                 ret = -ENOMEM;
2301                                 goto out;
2302                         }
2303                         target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2304                         kfree(p);
2305                         break;
2306
2307                 case SRP_OPT_IOC_GUID:
2308                         p = match_strdup(args);
2309                         if (!p) {
2310                                 ret = -ENOMEM;
2311                                 goto out;
2312                         }
2313                         target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2314                         kfree(p);
2315                         break;
2316
2317                 case SRP_OPT_DGID:
2318                         p = match_strdup(args);
2319                         if (!p) {
2320                                 ret = -ENOMEM;
2321                                 goto out;
2322                         }
2323                         if (strlen(p) != 32) {
2324                                 pr_warn("bad dest GID parameter '%s'\n", p);
2325                                 kfree(p);
2326                                 goto out;
2327                         }
2328
2329                         for (i = 0; i < 16; ++i) {
2330                                 strlcpy(dgid, p + i * 2, 3);
2331                                 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2332                         }
2333                         kfree(p);
2334                         memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2335                         break;
2336
2337                 case SRP_OPT_PKEY:
2338                         if (match_hex(args, &token)) {
2339                                 pr_warn("bad P_Key parameter '%s'\n", p);
2340                                 goto out;
2341                         }
2342                         target->path.pkey = cpu_to_be16(token);
2343                         break;
2344
2345                 case SRP_OPT_SERVICE_ID:
2346                         p = match_strdup(args);
2347                         if (!p) {
2348                                 ret = -ENOMEM;
2349                                 goto out;
2350                         }
2351                         target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2352                         target->path.service_id = target->service_id;
2353                         kfree(p);
2354                         break;
2355
2356                 case SRP_OPT_MAX_SECT:
2357                         if (match_int(args, &token)) {
2358                                 pr_warn("bad max sect parameter '%s'\n", p);
2359                                 goto out;
2360                         }
2361                         target->scsi_host->max_sectors = token;
2362                         break;
2363
2364                 case SRP_OPT_MAX_CMD_PER_LUN:
2365                         if (match_int(args, &token)) {
2366                                 pr_warn("bad max cmd_per_lun parameter '%s'\n",
2367                                         p);
2368                                 goto out;
2369                         }
2370                         target->scsi_host->cmd_per_lun = min(token, SRP_CMD_SQ_SIZE);
2371                         break;
2372
2373                 case SRP_OPT_IO_CLASS:
2374                         if (match_hex(args, &token)) {
2375                                 pr_warn("bad IO class parameter '%s'\n", p);
2376                                 goto out;
2377                         }
2378                         if (token != SRP_REV10_IB_IO_CLASS &&
2379                             token != SRP_REV16A_IB_IO_CLASS) {
2380                                 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2381                                         token, SRP_REV10_IB_IO_CLASS,
2382                                         SRP_REV16A_IB_IO_CLASS);
2383                                 goto out;
2384                         }
2385                         target->io_class = token;
2386                         break;
2387
2388                 case SRP_OPT_INITIATOR_EXT:
2389                         p = match_strdup(args);
2390                         if (!p) {
2391                                 ret = -ENOMEM;
2392                                 goto out;
2393                         }
2394                         target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2395                         kfree(p);
2396                         break;
2397
2398                 case SRP_OPT_CMD_SG_ENTRIES:
2399                         if (match_int(args, &token) || token < 1 || token > 255) {
2400                                 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2401                                         p);
2402                                 goto out;
2403                         }
2404                         target->cmd_sg_cnt = token;
2405                         break;
2406
2407                 case SRP_OPT_ALLOW_EXT_SG:
2408                         if (match_int(args, &token)) {
2409                                 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2410                                 goto out;
2411                         }
2412                         target->allow_ext_sg = !!token;
2413                         break;
2414
2415                 case SRP_OPT_SG_TABLESIZE:
2416                         if (match_int(args, &token) || token < 1 ||
2417                                         token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2418                                 pr_warn("bad max sg_tablesize parameter '%s'\n",
2419                                         p);
2420                                 goto out;
2421                         }
2422                         target->sg_tablesize = token;
2423                         break;
2424
2425                 case SRP_OPT_COMP_VECTOR:
2426                         if (match_int(args, &token) || token < 0) {
2427                                 pr_warn("bad comp_vector parameter '%s'\n", p);
2428                                 goto out;
2429                         }
2430                         target->comp_vector = token;
2431                         break;
2432
2433                 case SRP_OPT_TL_RETRY_COUNT:
2434                         if (match_int(args, &token) || token < 2 || token > 7) {
2435                                 pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
2436                                         p);
2437                                 goto out;
2438                         }
2439                         target->tl_retry_count = token;
2440                         break;
2441
2442                 default:
2443                         pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2444                                 p);
2445                         goto out;
2446                 }
2447         }
2448
2449         if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2450                 ret = 0;
2451         else
2452                 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2453                         if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2454                             !(srp_opt_tokens[i].token & opt_mask))
2455                                 pr_warn("target creation request is missing parameter '%s'\n",
2456                                         srp_opt_tokens[i].pattern);
2457
2458 out:
2459         kfree(options);
2460         return ret;
2461 }
2462
2463 static ssize_t srp_create_target(struct device *dev,
2464                                  struct device_attribute *attr,
2465                                  const char *buf, size_t count)
2466 {
2467         struct srp_host *host =
2468                 container_of(dev, struct srp_host, dev);
2469         struct Scsi_Host *target_host;
2470         struct srp_target_port *target;
2471         struct ib_device *ibdev = host->srp_dev->dev;
2472         int ret;
2473
2474         target_host = scsi_host_alloc(&srp_template,
2475                                       sizeof (struct srp_target_port));
2476         if (!target_host)
2477                 return -ENOMEM;
2478
2479         target_host->transportt  = ib_srp_transport_template;
2480         target_host->max_channel = 0;
2481         target_host->max_id      = 1;
2482         target_host->max_lun     = SRP_MAX_LUN;
2483         target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2484
2485         target = host_to_target(target_host);
2486
2487         target->io_class        = SRP_REV16A_IB_IO_CLASS;
2488         target->scsi_host       = target_host;
2489         target->srp_host        = host;
2490         target->lkey            = host->srp_dev->mr->lkey;
2491         target->rkey            = host->srp_dev->mr->rkey;
2492         target->cmd_sg_cnt      = cmd_sg_entries;
2493         target->sg_tablesize    = indirect_sg_entries ? : cmd_sg_entries;
2494         target->allow_ext_sg    = allow_ext_sg;
2495         target->tl_retry_count  = 7;
2496
2497         ret = srp_parse_options(buf, target);
2498         if (ret)
2499                 goto err;
2500
2501         if (!srp_conn_unique(target->srp_host, target)) {
2502                 shost_printk(KERN_INFO, target->scsi_host,
2503                              PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
2504                              be64_to_cpu(target->id_ext),
2505                              be64_to_cpu(target->ioc_guid),
2506                              be64_to_cpu(target->initiator_ext));
2507                 ret = -EEXIST;
2508                 goto err;
2509         }
2510
2511         if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2512                                 target->cmd_sg_cnt < target->sg_tablesize) {
2513                 pr_warn("No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2514                 target->sg_tablesize = target->cmd_sg_cnt;
2515         }
2516
2517         target_host->sg_tablesize = target->sg_tablesize;
2518         target->indirect_size = target->sg_tablesize *
2519                                 sizeof (struct srp_direct_buf);
2520         target->max_iu_len = sizeof (struct srp_cmd) +
2521                              sizeof (struct srp_indirect_buf) +
2522                              target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2523
2524         INIT_WORK(&target->tl_err_work, srp_tl_err_work);
2525         INIT_WORK(&target->remove_work, srp_remove_work);
2526         spin_lock_init(&target->lock);
2527         INIT_LIST_HEAD(&target->free_tx);
2528         ret = srp_alloc_req_data(target);
2529         if (ret)
2530                 goto err_free_mem;
2531
2532         ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2533
2534         shost_printk(KERN_DEBUG, target->scsi_host, PFX
2535                      "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
2536                      "service_id %016llx dgid %pI6\n",
2537                (unsigned long long) be64_to_cpu(target->id_ext),
2538                (unsigned long long) be64_to_cpu(target->ioc_guid),
2539                be16_to_cpu(target->path.pkey),
2540                (unsigned long long) be64_to_cpu(target->service_id),
2541                target->path.dgid.raw);
2542
2543         ret = srp_create_target_ib(target);
2544         if (ret)
2545                 goto err_free_mem;
2546
2547         ret = srp_new_cm_id(target);
2548         if (ret)
2549                 goto err_free_ib;
2550
2551         ret = srp_connect_target(target);
2552         if (ret) {
2553                 shost_printk(KERN_ERR, target->scsi_host,
2554                              PFX "Connection failed\n");
2555                 goto err_cm_id;
2556         }
2557
2558         ret = srp_add_target(host, target);
2559         if (ret)
2560                 goto err_disconnect;
2561
2562         return count;
2563
2564 err_disconnect:
2565         srp_disconnect_target(target);
2566
2567 err_cm_id:
2568         ib_destroy_cm_id(target->cm_id);
2569
2570 err_free_ib:
2571         srp_free_target_ib(target);
2572
2573 err_free_mem:
2574         srp_free_req_data(target);
2575
2576 err:
2577         scsi_host_put(target_host);
2578
2579         return ret;
2580 }
2581
2582 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2583
2584 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2585                           char *buf)
2586 {
2587         struct srp_host *host = container_of(dev, struct srp_host, dev);
2588
2589         return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2590 }
2591
2592 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2593
2594 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2595                          char *buf)
2596 {
2597         struct srp_host *host = container_of(dev, struct srp_host, dev);
2598
2599         return sprintf(buf, "%d\n", host->port);
2600 }
2601
2602 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2603
2604 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2605 {
2606         struct srp_host *host;
2607
2608         host = kzalloc(sizeof *host, GFP_KERNEL);
2609         if (!host)
2610                 return NULL;
2611
2612         INIT_LIST_HEAD(&host->target_list);
2613         spin_lock_init(&host->target_lock);
2614         init_completion(&host->released);
2615         host->srp_dev = device;
2616         host->port = port;
2617
2618         host->dev.class = &srp_class;
2619         host->dev.parent = device->dev->dma_device;
2620         dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2621
2622         if (device_register(&host->dev))
2623                 goto free_host;
2624         if (device_create_file(&host->dev, &dev_attr_add_target))
2625                 goto err_class;
2626         if (device_create_file(&host->dev, &dev_attr_ibdev))
2627                 goto err_class;
2628         if (device_create_file(&host->dev, &dev_attr_port))
2629                 goto err_class;
2630
2631         return host;
2632
2633 err_class:
2634         device_unregister(&host->dev);
2635
2636 free_host:
2637         kfree(host);
2638
2639         return NULL;
2640 }
2641
2642 static void srp_add_one(struct ib_device *device)
2643 {
2644         struct srp_device *srp_dev;
2645         struct ib_device_attr *dev_attr;
2646         struct ib_fmr_pool_param fmr_param;
2647         struct srp_host *host;
2648         int max_pages_per_fmr, fmr_page_shift, s, e, p;
2649
2650         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2651         if (!dev_attr)
2652                 return;
2653
2654         if (ib_query_device(device, dev_attr)) {
2655                 pr_warn("Query device failed for %s\n", device->name);
2656                 goto free_attr;
2657         }
2658
2659         srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2660         if (!srp_dev)
2661                 goto free_attr;
2662
2663         /*
2664          * Use the smallest page size supported by the HCA, down to a
2665          * minimum of 4096 bytes. We're unlikely to build large sglists
2666          * out of smaller entries.
2667          */
2668         fmr_page_shift          = max(12, ffs(dev_attr->page_size_cap) - 1);
2669         srp_dev->fmr_page_size  = 1 << fmr_page_shift;
2670         srp_dev->fmr_page_mask  = ~((u64) srp_dev->fmr_page_size - 1);
2671         srp_dev->fmr_max_size   = srp_dev->fmr_page_size * SRP_FMR_SIZE;
2672
2673         INIT_LIST_HEAD(&srp_dev->dev_list);
2674
2675         srp_dev->dev = device;
2676         srp_dev->pd  = ib_alloc_pd(device);
2677         if (IS_ERR(srp_dev->pd))
2678                 goto free_dev;
2679
2680         srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2681                                     IB_ACCESS_LOCAL_WRITE |
2682                                     IB_ACCESS_REMOTE_READ |
2683                                     IB_ACCESS_REMOTE_WRITE);
2684         if (IS_ERR(srp_dev->mr))
2685                 goto err_pd;
2686
2687         for (max_pages_per_fmr = SRP_FMR_SIZE;
2688                         max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2689                         max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2690                 memset(&fmr_param, 0, sizeof fmr_param);
2691                 fmr_param.pool_size         = SRP_FMR_POOL_SIZE;
2692                 fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
2693                 fmr_param.cache             = 1;
2694                 fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2695                 fmr_param.page_shift        = fmr_page_shift;
2696                 fmr_param.access            = (IB_ACCESS_LOCAL_WRITE |
2697                                                IB_ACCESS_REMOTE_WRITE |
2698                                                IB_ACCESS_REMOTE_READ);
2699
2700                 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2701                 if (!IS_ERR(srp_dev->fmr_pool))
2702                         break;
2703         }
2704
2705         if (IS_ERR(srp_dev->fmr_pool))
2706                 srp_dev->fmr_pool = NULL;
2707
2708         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2709                 s = 0;
2710                 e = 0;
2711         } else {
2712                 s = 1;
2713                 e = device->phys_port_cnt;
2714         }
2715
2716         for (p = s; p <= e; ++p) {
2717                 host = srp_add_port(srp_dev, p);
2718                 if (host)
2719                         list_add_tail(&host->list, &srp_dev->dev_list);
2720         }
2721
2722         ib_set_client_data(device, &srp_client, srp_dev);
2723
2724         goto free_attr;
2725
2726 err_pd:
2727         ib_dealloc_pd(srp_dev->pd);
2728
2729 free_dev:
2730         kfree(srp_dev);
2731
2732 free_attr:
2733         kfree(dev_attr);
2734 }
2735
2736 static void srp_remove_one(struct ib_device *device)
2737 {
2738         struct srp_device *srp_dev;
2739         struct srp_host *host, *tmp_host;
2740         struct srp_target_port *target;
2741
2742         srp_dev = ib_get_client_data(device, &srp_client);
2743         if (!srp_dev)
2744                 return;
2745
2746         list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2747                 device_unregister(&host->dev);
2748                 /*
2749                  * Wait for the sysfs entry to go away, so that no new
2750                  * target ports can be created.
2751                  */
2752                 wait_for_completion(&host->released);
2753
2754                 /*
2755                  * Remove all target ports.
2756                  */
2757                 spin_lock(&host->target_lock);
2758                 list_for_each_entry(target, &host->target_list, list)
2759                         srp_queue_remove_work(target);
2760                 spin_unlock(&host->target_lock);
2761
2762                 /*
2763                  * Wait for target port removal tasks.
2764                  */
2765                 flush_workqueue(system_long_wq);
2766
2767                 kfree(host);
2768         }
2769
2770         if (srp_dev->fmr_pool)
2771                 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2772         ib_dereg_mr(srp_dev->mr);
2773         ib_dealloc_pd(srp_dev->pd);
2774
2775         kfree(srp_dev);
2776 }
2777
2778 static struct srp_function_template ib_srp_transport_functions = {
2779         .has_rport_state         = true,
2780         .reset_timer_if_blocked  = true,
2781         .reconnect_delay         = &srp_reconnect_delay,
2782         .fast_io_fail_tmo        = &srp_fast_io_fail_tmo,
2783         .dev_loss_tmo            = &srp_dev_loss_tmo,
2784         .reconnect               = srp_rport_reconnect,
2785         .rport_delete            = srp_rport_delete,
2786         .terminate_rport_io      = srp_terminate_io,
2787 };
2788
2789 static int __init srp_init_module(void)
2790 {
2791         int ret;
2792
2793         BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2794
2795         if (srp_sg_tablesize) {
2796                 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2797                 if (!cmd_sg_entries)
2798                         cmd_sg_entries = srp_sg_tablesize;
2799         }
2800
2801         if (!cmd_sg_entries)
2802                 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2803
2804         if (cmd_sg_entries > 255) {
2805                 pr_warn("Clamping cmd_sg_entries to 255\n");
2806                 cmd_sg_entries = 255;
2807         }
2808
2809         if (!indirect_sg_entries)
2810                 indirect_sg_entries = cmd_sg_entries;
2811         else if (indirect_sg_entries < cmd_sg_entries) {
2812                 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
2813                         cmd_sg_entries);
2814                 indirect_sg_entries = cmd_sg_entries;
2815         }
2816
2817         ib_srp_transport_template =
2818                 srp_attach_transport(&ib_srp_transport_functions);
2819         if (!ib_srp_transport_template)
2820                 return -ENOMEM;
2821
2822         ret = class_register(&srp_class);
2823         if (ret) {
2824                 pr_err("couldn't register class infiniband_srp\n");
2825                 srp_release_transport(ib_srp_transport_template);
2826                 return ret;
2827         }
2828
2829         ib_sa_register_client(&srp_sa_client);
2830
2831         ret = ib_register_client(&srp_client);
2832         if (ret) {
2833                 pr_err("couldn't register IB client\n");
2834                 srp_release_transport(ib_srp_transport_template);
2835                 ib_sa_unregister_client(&srp_sa_client);
2836                 class_unregister(&srp_class);
2837                 return ret;
2838         }
2839
2840         return 0;
2841 }
2842
2843 static void __exit srp_cleanup_module(void)
2844 {
2845         ib_unregister_client(&srp_client);
2846         ib_sa_unregister_client(&srp_sa_client);
2847         class_unregister(&srp_class);
2848         srp_release_transport(ib_srp_transport_template);
2849 }
2850
2851 module_init(srp_init_module);
2852 module_exit(srp_cleanup_module);