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/scsi_tcq.h>
  50 #include <scsi/srp.h>
  51 #include <scsi/scsi_transport_srp.h>
  52
  53 #include "ib_srp.h"
  54
  55 #define DRV_NAME        "ib_srp"
  56 #define PFX             DRV_NAME ": "
  57 #define DRV_VERSION     "1.0"
  58 #define DRV_RELDATE     "July 1, 2013"
  59
  60 MODULE_AUTHOR("Roland Dreier");
  61 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
  62                    "v" DRV_VERSION " (" DRV_RELDATE ")");
  63 MODULE_LICENSE("Dual BSD/GPL");
  64
  65 static unsigned int srp_sg_tablesize;
  66 static unsigned int cmd_sg_entries;
  67 static unsigned int indirect_sg_entries;
  68 static bool allow_ext_sg;
  69 static int topspin_workarounds = 1;
  70
  71 module_param(srp_sg_tablesize, uint, 0444);
  72 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
  73
  74 module_param(cmd_sg_entries, uint, 0444);
  75 MODULE_PARM_DESC(cmd_sg_entries,
  76                  "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
  77
  78 module_param(indirect_sg_entries, uint, 0444);
  79 MODULE_PARM_DESC(indirect_sg_entries,
  80                  "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
  81
  82 module_param(allow_ext_sg, bool, 0444);
  83 MODULE_PARM_DESC(allow_ext_sg,
  84                   "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
  85
  86 module_param(topspin_workarounds, int, 0444);
  87 MODULE_PARM_DESC(topspin_workarounds,
  88                  "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
  89
  90 static struct kernel_param_ops srp_tmo_ops;
  91
  92 static int srp_reconnect_delay = 10;
  93 module_param_cb(reconnect_delay, &srp_tmo_ops, &srp_reconnect_delay,
  94                 S_IRUGO | S_IWUSR);
  95 MODULE_PARM_DESC(reconnect_delay, "Time between successive reconnect attempts");
  96
  97 static int srp_fast_io_fail_tmo = 15;
  98 module_param_cb(fast_io_fail_tmo, &srp_tmo_ops, &srp_fast_io_fail_tmo,
  99                 S_IRUGO | S_IWUSR);
 100 MODULE_PARM_DESC(fast_io_fail_tmo,
 101                  "Number of seconds between the observation of a transport"
 102                  " layer error and failing all I/O. \"off\" means that this"
 103                  " functionality is disabled.");
 104
 105 static int srp_dev_loss_tmo = 600;
 106 module_param_cb(dev_loss_tmo, &srp_tmo_ops, &srp_dev_loss_tmo,
 107                 S_IRUGO | S_IWUSR);
 108 MODULE_PARM_DESC(dev_loss_tmo,
 109                  "Maximum number of seconds that the SRP transport should"
 110                  " insulate transport layer errors. After this time has been"
 111                  " exceeded the SCSI host is removed. Should be"
 112                  " between 1 and " __stringify(SCSI_DEVICE_BLOCK_MAX_TIMEOUT)
 113                  " if fast_io_fail_tmo has not been set. \"off\" means that"
 114                  " this functionality is disabled.");
 115
 116 static void srp_add_one(struct ib_device *device);
 117 static void srp_remove_one(struct ib_device *device);
 118 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
 119 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
 120 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
 121
 122 static struct scsi_transport_template *ib_srp_transport_template;
 123
 124 static struct ib_client srp_client = {
 125         .name   = "srp",
 126         .add    = srp_add_one,
 127         .remove = srp_remove_one
 128 };
 129
 130 static struct ib_sa_client srp_sa_client;
 131
 132 static int srp_tmo_get(char *buffer, const struct kernel_param *kp)
 133 {
 134         int tmo = *(int *)kp->arg;
 135
 136         if (tmo >= 0)
 137                 return sprintf(buffer, "%d", tmo);
 138         else
 139                 return sprintf(buffer, "off");
 140 }
 141
 142 static int srp_tmo_set(const char *val, const struct kernel_param *kp)
 143 {
 144         int tmo, res;
 145
 146         if (strncmp(val, "off", 3) != 0) {
 147                 res = kstrtoint(val, 0, &tmo);
 148                 if (res)
 149                         goto out;
 150         } else {
 151                 tmo = -1;
 152         }
 153         if (kp->arg == &srp_reconnect_delay)
 154                 res = srp_tmo_valid(tmo, srp_fast_io_fail_tmo,
 155                                     srp_dev_loss_tmo);
 156         else if (kp->arg == &srp_fast_io_fail_tmo)
 157                 res = srp_tmo_valid(srp_reconnect_delay, tmo, srp_dev_loss_tmo);
 158         else
 159                 res = srp_tmo_valid(srp_reconnect_delay, srp_fast_io_fail_tmo,
 160                                     tmo);
 161         if (res)
 162                 goto out;
 163         *(int *)kp->arg = tmo;
 164
 165 out:
 166         return res;
 167 }
 168
 169 static struct kernel_param_ops srp_tmo_ops = {
 170         .get = srp_tmo_get,
 171         .set = srp_tmo_set,
 172 };
 173
 174 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
 175 {
 176         return (struct srp_target_port *) host->hostdata;
 177 }
 178
 179 static const char *srp_target_info(struct Scsi_Host *host)
 180 {
 181         return host_to_target(host)->target_name;
 182 }
 183
 184 static int srp_target_is_topspin(struct srp_target_port *target)
 185 {
 186         static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
 187         static const u8 cisco_oui[3]   = { 0x00, 0x1b, 0x0d };
 188
 189         return topspin_workarounds &&
 190                 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
 191                  !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
 192 }
 193
 194 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
 195                                    gfp_t gfp_mask,
 196                                    enum dma_data_direction direction)
 197 {
 198         struct srp_iu *iu;
 199
 200         iu = kmalloc(sizeof *iu, gfp_mask);
 201         if (!iu)
 202                 goto out;
 203
 204         iu->buf = kzalloc(size, gfp_mask);
 205         if (!iu->buf)
 206                 goto out_free_iu;
 207
 208         iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
 209                                     direction);
 210         if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
 211                 goto out_free_buf;
 212
 213         iu->size      = size;
 214         iu->direction = direction;
 215
 216         return iu;
 217
 218 out_free_buf:
 219         kfree(iu->buf);
 220 out_free_iu:
 221         kfree(iu);
 222 out:
 223         return NULL;
 224 }
 225
 226 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
 227 {
 228         if (!iu)
 229                 return;
 230
 231         ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
 232                             iu->direction);
 233         kfree(iu->buf);
 234         kfree(iu);
 235 }
 236
 237 static void srp_qp_event(struct ib_event *event, void *context)
 238 {
 239         pr_debug("QP event %d\n", event->event);
 240 }
 241
 242 static int srp_init_qp(struct srp_target_port *target,
 243                        struct ib_qp *qp)
 244 {
 245         struct ib_qp_attr *attr;
 246         int ret;
 247
 248         attr = kmalloc(sizeof *attr, GFP_KERNEL);
 249         if (!attr)
 250                 return -ENOMEM;
 251
 252         ret = ib_find_pkey(target->srp_host->srp_dev->dev,
 253                            target->srp_host->port,
 254                            be16_to_cpu(target->path.pkey),
 255                            &attr->pkey_index);
 256         if (ret)
 257                 goto out;
 258
 259         attr->qp_state        = IB_QPS_INIT;
 260         attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
 261                                     IB_ACCESS_REMOTE_WRITE);
 262         attr->port_num        = target->srp_host->port;
 263
 264         ret = ib_modify_qp(qp, attr,
 265                            IB_QP_STATE          |
 266                            IB_QP_PKEY_INDEX     |
 267                            IB_QP_ACCESS_FLAGS   |
 268                            IB_QP_PORT);
 269
 270 out:
 271         kfree(attr);
 272         return ret;
 273 }
 274
 275 static int srp_new_cm_id(struct srp_target_port *target)
 276 {
 277         struct ib_cm_id *new_cm_id;
 278
 279         new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
 280                                     srp_cm_handler, target);
 281         if (IS_ERR(new_cm_id))
 282                 return PTR_ERR(new_cm_id);
 283
 284         if (target->cm_id)
 285                 ib_destroy_cm_id(target->cm_id);
 286         target->cm_id = new_cm_id;
 287
 288         return 0;
 289 }
 290
 291 static int srp_create_target_ib(struct srp_target_port *target)
 292 {
 293         struct ib_qp_init_attr *init_attr;
 294         struct ib_cq *recv_cq, *send_cq;
 295         struct ib_qp *qp;
 296         int ret;
 297
 298         init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
 299         if (!init_attr)
 300                 return -ENOMEM;
 301
 302         recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
 303                                srp_recv_completion, NULL, target,
 304                                target->queue_size, target->comp_vector);
 305         if (IS_ERR(recv_cq)) {
 306                 ret = PTR_ERR(recv_cq);
 307                 goto err;
 308         }
 309
 310         send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
 311                                srp_send_completion, NULL, target,
 312                                target->queue_size, target->comp_vector);
 313         if (IS_ERR(send_cq)) {
 314                 ret = PTR_ERR(send_cq);
 315                 goto err_recv_cq;
 316         }
 317
 318         ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
 319
 320         init_attr->event_handler       = srp_qp_event;
 321         init_attr->cap.max_send_wr     = target->queue_size;
 322         init_attr->cap.max_recv_wr     = target->queue_size;
 323         init_attr->cap.max_recv_sge    = 1;
 324         init_attr->cap.max_send_sge    = 1;
 325         init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
 326         init_attr->qp_type             = IB_QPT_RC;
 327         init_attr->send_cq             = send_cq;
 328         init_attr->recv_cq             = recv_cq;
 329
 330         qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
 331         if (IS_ERR(qp)) {
 332                 ret = PTR_ERR(qp);
 333                 goto err_send_cq;
 334         }
 335
 336         ret = srp_init_qp(target, qp);
 337         if (ret)
 338                 goto err_qp;
 339
 340         if (target->qp)
 341                 ib_destroy_qp(target->qp);
 342         if (target->recv_cq)
 343                 ib_destroy_cq(target->recv_cq);
 344         if (target->send_cq)
 345                 ib_destroy_cq(target->send_cq);
 346
 347         target->qp = qp;
 348         target->recv_cq = recv_cq;
 349         target->send_cq = send_cq;
 350
 351         kfree(init_attr);
 352         return 0;
 353
 354 err_qp:
 355         ib_destroy_qp(qp);
 356
 357 err_send_cq:
 358         ib_destroy_cq(send_cq);
 359
 360 err_recv_cq:
 361         ib_destroy_cq(recv_cq);
 362
 363 err:
 364         kfree(init_attr);
 365         return ret;
 366 }
 367
 368 /*
 369  * Note: this function may be called without srp_alloc_iu_bufs() having been
 370  * invoked. Hence the target->[rt]x_ring checks.
 371  */
 372 static void srp_free_target_ib(struct srp_target_port *target)
 373 {
 374         int i;
 375
 376         ib_destroy_qp(target->qp);
 377         ib_destroy_cq(target->send_cq);
 378         ib_destroy_cq(target->recv_cq);
 379
 380         target->qp = NULL;
 381         target->send_cq = target->recv_cq = NULL;
 382
 383         if (target->rx_ring) {
 384                 for (i = 0; i < target->queue_size; ++i)
 385                         srp_free_iu(target->srp_host, target->rx_ring[i]);
 386                 kfree(target->rx_ring);
 387                 target->rx_ring = NULL;
 388         }
 389         if (target->tx_ring) {
 390                 for (i = 0; i < target->queue_size; ++i)
 391                         srp_free_iu(target->srp_host, target->tx_ring[i]);
 392                 kfree(target->tx_ring);
 393                 target->tx_ring = NULL;
 394         }
 395 }
 396
 397 static void srp_path_rec_completion(int status,
 398                                     struct ib_sa_path_rec *pathrec,
 399                                     void *target_ptr)
 400 {
 401         struct srp_target_port *target = target_ptr;
 402
 403         target->status = status;
 404         if (status)
 405                 shost_printk(KERN_ERR, target->scsi_host,
 406                              PFX "Got failed path rec status %d\n", status);
 407         else
 408                 target->path = *pathrec;
 409         complete(&target->done);
 410 }
 411
 412 static int srp_lookup_path(struct srp_target_port *target)
 413 {
 414         int ret;
 415
 416         target->path.numb_path = 1;
 417
 418         init_completion(&target->done);
 419
 420         target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
 421                                                    target->srp_host->srp_dev->dev,
 422                                                    target->srp_host->port,
 423                                                    &target->path,
 424                                                    IB_SA_PATH_REC_SERVICE_ID    |
 425                                                    IB_SA_PATH_REC_DGID          |
 426                                                    IB_SA_PATH_REC_SGID          |
 427                                                    IB_SA_PATH_REC_NUMB_PATH     |
 428                                                    IB_SA_PATH_REC_PKEY,
 429                                                    SRP_PATH_REC_TIMEOUT_MS,
 430                                                    GFP_KERNEL,
 431                                                    srp_path_rec_completion,
 432                                                    target, &target->path_query);
 433         if (target->path_query_id < 0)
 434                 return target->path_query_id;
 435
 436         ret = wait_for_completion_interruptible(&target->done);
 437         if (ret < 0)
 438                 return ret;
 439
 440         if (target->status < 0)
 441                 shost_printk(KERN_WARNING, target->scsi_host,
 442                              PFX "Path record query failed\n");
 443
 444         return target->status;
 445 }
 446
 447 static int srp_send_req(struct srp_target_port *target)
 448 {
 449         struct {
 450                 struct ib_cm_req_param param;
 451                 struct srp_login_req   priv;
 452         } *req = NULL;
 453         int status;
 454
 455         req = kzalloc(sizeof *req, GFP_KERNEL);
 456         if (!req)
 457                 return -ENOMEM;
 458
 459         req->param.primary_path               = &target->path;
 460         req->param.alternate_path             = NULL;
 461         req->param.service_id                 = target->service_id;
 462         req->param.qp_num                     = target->qp->qp_num;
 463         req->param.qp_type                    = target->qp->qp_type;
 464         req->param.private_data               = &req->priv;
 465         req->param.private_data_len           = sizeof req->priv;
 466         req->param.flow_control               = 1;
 467
 468         get_random_bytes(&req->param.starting_psn, 4);
 469         req->param.starting_psn              &= 0xffffff;
 470
 471         /*
 472          * Pick some arbitrary defaults here; we could make these
 473          * module parameters if anyone cared about setting them.
 474          */
 475         req->param.responder_resources        = 4;
 476         req->param.remote_cm_response_timeout = 20;
 477         req->param.local_cm_response_timeout  = 20;
 478         req->param.retry_count                = target->tl_retry_count;
 479         req->param.rnr_retry_count            = 7;
 480         req->param.max_cm_retries             = 15;
 481
 482         req->priv.opcode        = SRP_LOGIN_REQ;
 483         req->priv.tag           = 0;
 484         req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
 485         req->priv.req_buf_fmt   = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
 486                                               SRP_BUF_FORMAT_INDIRECT);
 487         /*
 488          * In the published SRP specification (draft rev. 16a), the
 489          * port identifier format is 8 bytes of ID extension followed
 490          * by 8 bytes of GUID.  Older drafts put the two halves in the
 491          * opposite order, so that the GUID comes first.
 492          *
 493          * Targets conforming to these obsolete drafts can be
 494          * recognized by the I/O Class they report.
 495          */
 496         if (target->io_class == SRP_REV10_IB_IO_CLASS) {
 497                 memcpy(req->priv.initiator_port_id,
 498                        &target->path.sgid.global.interface_id, 8);
 499                 memcpy(req->priv.initiator_port_id + 8,
 500                        &target->initiator_ext, 8);
 501                 memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
 502                 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
 503         } else {
 504                 memcpy(req->priv.initiator_port_id,
 505                        &target->initiator_ext, 8);
 506                 memcpy(req->priv.initiator_port_id + 8,
 507                        &target->path.sgid.global.interface_id, 8);
 508                 memcpy(req->priv.target_port_id,     &target->id_ext, 8);
 509                 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
 510         }
 511
 512         /*
 513          * Topspin/Cisco SRP targets will reject our login unless we
 514          * zero out the first 8 bytes of our initiator port ID and set
 515          * the second 8 bytes to the local node GUID.
 516          */
 517         if (srp_target_is_topspin(target)) {
 518                 shost_printk(KERN_DEBUG, target->scsi_host,
 519                              PFX "Topspin/Cisco initiator port ID workaround "
 520                              "activated for target GUID %016llx\n",
 521                              (unsigned long long) be64_to_cpu(target->ioc_guid));
 522                 memset(req->priv.initiator_port_id, 0, 8);
 523                 memcpy(req->priv.initiator_port_id + 8,
 524                        &target->srp_host->srp_dev->dev->node_guid, 8);
 525         }
 526
 527         status = ib_send_cm_req(target->cm_id, &req->param);
 528
 529         kfree(req);
 530
 531         return status;
 532 }
 533
 534 static bool srp_queue_remove_work(struct srp_target_port *target)
 535 {
 536         bool changed = false;
 537
 538         spin_lock_irq(&target->lock);
 539         if (target->state != SRP_TARGET_REMOVED) {
 540                 target->state = SRP_TARGET_REMOVED;
 541                 changed = true;
 542         }
 543         spin_unlock_irq(&target->lock);
 544
 545         if (changed)
 546                 queue_work(system_long_wq, &target->remove_work);
 547
 548         return changed;
 549 }
 550
 551 static bool srp_change_conn_state(struct srp_target_port *target,
 552                                   bool connected)
 553 {
 554         bool changed = false;
 555
 556         spin_lock_irq(&target->lock);
 557         if (target->connected != connected) {
 558                 target->connected = connected;
 559                 changed = true;
 560         }
 561         spin_unlock_irq(&target->lock);
 562
 563         return changed;
 564 }
 565
 566 static void srp_disconnect_target(struct srp_target_port *target)
 567 {
 568         if (srp_change_conn_state(target, false)) {
 569                 /* XXX should send SRP_I_LOGOUT request */
 570
 571                 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
 572                         shost_printk(KERN_DEBUG, target->scsi_host,
 573                                      PFX "Sending CM DREQ failed\n");
 574                 }
 575         }
 576 }
 577
 578 static void srp_free_req_data(struct srp_target_port *target)
 579 {
 580         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
 581         struct srp_request *req;
 582         int i;
 583
 584         if (!target->req_ring)
 585                 return;
 586
 587         for (i = 0; i < target->req_ring_size; ++i) {
 588                 req = &target->req_ring[i];
 589                 kfree(req->fmr_list);
 590                 kfree(req->map_page);
 591                 if (req->indirect_dma_addr) {
 592                         ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
 593                                             target->indirect_size,
 594                                             DMA_TO_DEVICE);
 595                 }
 596                 kfree(req->indirect_desc);
 597         }
 598
 599         kfree(target->req_ring);
 600         target->req_ring = NULL;
 601 }
 602
 603 static int srp_alloc_req_data(struct srp_target_port *target)
 604 {
 605         struct srp_device *srp_dev = target->srp_host->srp_dev;
 606         struct ib_device *ibdev = srp_dev->dev;
 607         struct srp_request *req;
 608         dma_addr_t dma_addr;
 609         int i, ret = -ENOMEM;
 610
 611         INIT_LIST_HEAD(&target->free_reqs);
 612
 613         target->req_ring = kzalloc(target->req_ring_size *
 614                                    sizeof(*target->req_ring), GFP_KERNEL);
 615         if (!target->req_ring)
 616                 goto out;
 617
 618         for (i = 0; i < target->req_ring_size; ++i) {
 619                 req = &target->req_ring[i];
 620                 req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof(void *),
 621                                         GFP_KERNEL);
 622                 req->map_page = kmalloc(SRP_FMR_SIZE * sizeof(void *),
 623                                         GFP_KERNEL);
 624                 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
 625                 if (!req->fmr_list || !req->map_page || !req->indirect_desc)
 626                         goto out;
 627
 628                 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
 629                                              target->indirect_size,
 630                                              DMA_TO_DEVICE);
 631                 if (ib_dma_mapping_error(ibdev, dma_addr))
 632                         goto out;
 633
 634                 req->indirect_dma_addr = dma_addr;
 635                 req->index = i;
 636                 list_add_tail(&req->list, &target->free_reqs);
 637         }
 638         ret = 0;
 639
 640 out:
 641         return ret;
 642 }
 643
 644 /**
 645  * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
 646  * @shost: SCSI host whose attributes to remove from sysfs.
 647  *
 648  * Note: Any attributes defined in the host template and that did not exist
 649  * before invocation of this function will be ignored.
 650  */
 651 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
 652 {
 653         struct device_attribute **attr;
 654
 655         for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
 656                 device_remove_file(&shost->shost_dev, *attr);
 657 }
 658
 659 static void srp_remove_target(struct srp_target_port *target)
 660 {
 661         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 662
 663         srp_del_scsi_host_attr(target->scsi_host);
 664         srp_rport_get(target->rport);
 665         srp_remove_host(target->scsi_host);
 666         scsi_remove_host(target->scsi_host);
 667         srp_stop_rport_timers(target->rport);
 668         srp_disconnect_target(target);
 669         ib_destroy_cm_id(target->cm_id);
 670         srp_free_target_ib(target);
 671         cancel_work_sync(&target->tl_err_work);
 672         srp_rport_put(target->rport);
 673         srp_free_req_data(target);
 674
 675         spin_lock(&target->srp_host->target_lock);
 676         list_del(&target->list);
 677         spin_unlock(&target->srp_host->target_lock);
 678
 679         scsi_host_put(target->scsi_host);
 680 }
 681
 682 static void srp_remove_work(struct work_struct *work)
 683 {
 684         struct srp_target_port *target =
 685                 container_of(work, struct srp_target_port, remove_work);
 686
 687         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 688
 689         srp_remove_target(target);
 690 }
 691
 692 static void srp_rport_delete(struct srp_rport *rport)
 693 {
 694         struct srp_target_port *target = rport->lld_data;
 695
 696         srp_queue_remove_work(target);
 697 }
 698
 699 static int srp_connect_target(struct srp_target_port *target)
 700 {
 701         int retries = 3;
 702         int ret;
 703
 704         WARN_ON_ONCE(target->connected);
 705
 706         target->qp_in_error = false;
 707
 708         ret = srp_lookup_path(target);
 709         if (ret)
 710                 return ret;
 711
 712         while (1) {
 713                 init_completion(&target->done);
 714                 ret = srp_send_req(target);
 715                 if (ret)
 716                         return ret;
 717                 ret = wait_for_completion_interruptible(&target->done);
 718                 if (ret < 0)
 719                         return ret;
 720
 721                 /*
 722                  * The CM event handling code will set status to
 723                  * SRP_PORT_REDIRECT if we get a port redirect REJ
 724                  * back, or SRP_DLID_REDIRECT if we get a lid/qp
 725                  * redirect REJ back.
 726                  */
 727                 switch (target->status) {
 728                 case 0:
 729                         srp_change_conn_state(target, true);
 730                         return 0;
 731
 732                 case SRP_PORT_REDIRECT:
 733                         ret = srp_lookup_path(target);
 734                         if (ret)
 735                                 return ret;
 736                         break;
 737
 738                 case SRP_DLID_REDIRECT:
 739                         break;
 740
 741                 case SRP_STALE_CONN:
 742                         /* Our current CM id was stale, and is now in timewait.
 743                          * Try to reconnect with a new one.
 744                          */
 745                         if (!retries-- || srp_new_cm_id(target)) {
 746                                 shost_printk(KERN_ERR, target->scsi_host, PFX
 747                                              "giving up on stale connection\n");
 748                                 target->status = -ECONNRESET;
 749                                 return target->status;
 750                         }
 751
 752                         shost_printk(KERN_ERR, target->scsi_host, PFX
 753                                      "retrying stale connection\n");
 754                         break;
 755
 756                 default:
 757                         return target->status;
 758                 }
 759         }
 760 }
 761
 762 static void srp_unmap_data(struct scsi_cmnd *scmnd,
 763                            struct srp_target_port *target,
 764                            struct srp_request *req)
 765 {
 766         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
 767         struct ib_pool_fmr **pfmr;
 768
 769         if (!scsi_sglist(scmnd) ||
 770             (scmnd->sc_data_direction != DMA_TO_DEVICE &&
 771              scmnd->sc_data_direction != DMA_FROM_DEVICE))
 772                 return;
 773
 774         pfmr = req->fmr_list;
 775         while (req->nfmr--)
 776                 ib_fmr_pool_unmap(*pfmr++);
 777
 778         ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
 779                         scmnd->sc_data_direction);
 780 }
 781
 782 /**
 783  * srp_claim_req - Take ownership of the scmnd associated with a request.
 784  * @target: SRP target port.
 785  * @req: SRP request.
 786  * @sdev: If not NULL, only take ownership for this SCSI device.
 787  * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
 788  *         ownership of @req->scmnd if it equals @scmnd.
 789  *
 790  * Return value:
 791  * Either NULL or a pointer to the SCSI command the caller became owner of.
 792  */
 793 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
 794                                        struct srp_request *req,
 795                                        struct scsi_device *sdev,
 796                                        struct scsi_cmnd *scmnd)
 797 {
 798         unsigned long flags;
 799
 800         spin_lock_irqsave(&target->lock, flags);
 801         if (req->scmnd &&
 802             (!sdev || req->scmnd->device == sdev) &&
 803             (!scmnd || req->scmnd == scmnd)) {
 804                 scmnd = req->scmnd;
 805                 req->scmnd = NULL;
 806         } else {
 807                 scmnd = NULL;
 808         }
 809         spin_unlock_irqrestore(&target->lock, flags);
 810
 811         return scmnd;
 812 }
 813
 814 /**
 815  * srp_free_req() - Unmap data and add request to the free request list.
 816  */
 817 static void srp_free_req(struct srp_target_port *target,
 818                          struct srp_request *req, struct scsi_cmnd *scmnd,
 819                          s32 req_lim_delta)
 820 {
 821         unsigned long flags;
 822
 823         srp_unmap_data(scmnd, target, req);
 824
 825         spin_lock_irqsave(&target->lock, flags);
 826         target->req_lim += req_lim_delta;
 827         list_add_tail(&req->list, &target->free_reqs);
 828         spin_unlock_irqrestore(&target->lock, flags);
 829 }
 830
 831 static void srp_finish_req(struct srp_target_port *target,
 832                            struct srp_request *req, struct scsi_device *sdev,
 833                            int result)
 834 {
 835         struct scsi_cmnd *scmnd = srp_claim_req(target, req, sdev, NULL);
 836
 837         if (scmnd) {
 838                 srp_free_req(target, req, scmnd, 0);
 839                 scmnd->result = result;
 840                 scmnd->scsi_done(scmnd);
 841         }
 842 }
 843
 844 static void srp_terminate_io(struct srp_rport *rport)
 845 {
 846         struct srp_target_port *target = rport->lld_data;
 847         struct Scsi_Host *shost = target->scsi_host;
 848         struct scsi_device *sdev;
 849         int i;
 850
 851         /*
 852          * Invoking srp_terminate_io() while srp_queuecommand() is running
 853          * is not safe. Hence the warning statement below.
 854          */
 855         shost_for_each_device(sdev, shost)
 856                 WARN_ON_ONCE(sdev->request_queue->request_fn_active);
 857
 858         for (i = 0; i < target->req_ring_size; ++i) {
 859                 struct srp_request *req = &target->req_ring[i];
 860                 srp_finish_req(target, req, NULL, DID_TRANSPORT_FAILFAST << 16);
 861         }
 862 }
 863
 864 /*
 865  * It is up to the caller to ensure that srp_rport_reconnect() calls are
 866  * serialized and that no concurrent srp_queuecommand(), srp_abort(),
 867  * srp_reset_device() or srp_reset_host() calls will occur while this function
 868  * is in progress. One way to realize that is not to call this function
 869  * directly but to call srp_reconnect_rport() instead since that last function
 870  * serializes calls of this function via rport->mutex and also blocks
 871  * srp_queuecommand() calls before invoking this function.
 872  */
 873 static int srp_rport_reconnect(struct srp_rport *rport)
 874 {
 875         struct srp_target_port *target = rport->lld_data;
 876         int i, ret;
 877
 878         srp_disconnect_target(target);
 879         /*
 880          * Now get a new local CM ID so that we avoid confusing the target in
 881          * case things are really fouled up. Doing so also ensures that all CM
 882          * callbacks will have finished before a new QP is allocated.
 883          */
 884         ret = srp_new_cm_id(target);
 885         /*
 886          * Whether or not creating a new CM ID succeeded, create a new
 887          * QP. This guarantees that all completion callback function
 888          * invocations have finished before request resetting starts.
 889          */
 890         if (ret == 0)
 891                 ret = srp_create_target_ib(target);
 892         else
 893                 srp_create_target_ib(target);
 894
 895         for (i = 0; i < target->req_ring_size; ++i) {
 896                 struct srp_request *req = &target->req_ring[i];
 897                 srp_finish_req(target, req, NULL, DID_RESET << 16);
 898         }
 899
 900         INIT_LIST_HEAD(&target->free_tx);
 901         for (i = 0; i < target->queue_size; ++i)
 902                 list_add(&target->tx_ring[i]->list, &target->free_tx);
 903
 904         if (ret == 0)
 905                 ret = srp_connect_target(target);
 906
 907         if (ret == 0)
 908                 shost_printk(KERN_INFO, target->scsi_host,
 909                              PFX "reconnect succeeded\n");
 910
 911         return ret;
 912 }
 913
 914 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
 915                          unsigned int dma_len, u32 rkey)
 916 {
 917         struct srp_direct_buf *desc = state->desc;
 918
 919         desc->va = cpu_to_be64(dma_addr);
 920         desc->key = cpu_to_be32(rkey);
 921         desc->len = cpu_to_be32(dma_len);
 922
 923         state->total_len += dma_len;
 924         state->desc++;
 925         state->ndesc++;
 926 }
 927
 928 static int srp_map_finish_fmr(struct srp_map_state *state,
 929                               struct srp_target_port *target)
 930 {
 931         struct srp_device *dev = target->srp_host->srp_dev;
 932         struct ib_pool_fmr *fmr;
 933         u64 io_addr = 0;
 934
 935         if (!state->npages)
 936                 return 0;
 937
 938         if (state->npages == 1) {
 939                 srp_map_desc(state, state->base_dma_addr, state->fmr_len,
 940                              target->rkey);
 941                 state->npages = state->fmr_len = 0;
 942                 return 0;
 943         }
 944
 945         fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
 946                                    state->npages, io_addr);
 947         if (IS_ERR(fmr))
 948                 return PTR_ERR(fmr);
 949
 950         *state->next_fmr++ = fmr;
 951         state->nfmr++;
 952
 953         srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
 954         state->npages = state->fmr_len = 0;
 955         return 0;
 956 }
 957
 958 static void srp_map_update_start(struct srp_map_state *state,
 959                                  struct scatterlist *sg, int sg_index,
 960                                  dma_addr_t dma_addr)
 961 {
 962         state->unmapped_sg = sg;
 963         state->unmapped_index = sg_index;
 964         state->unmapped_addr = dma_addr;
 965 }
 966
 967 static int srp_map_sg_entry(struct srp_map_state *state,
 968                             struct srp_target_port *target,
 969                             struct scatterlist *sg, int sg_index,
 970                             int use_fmr)
 971 {
 972         struct srp_device *dev = target->srp_host->srp_dev;
 973         struct ib_device *ibdev = dev->dev;
 974         dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
 975         unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
 976         unsigned int len;
 977         int ret;
 978
 979         if (!dma_len)
 980                 return 0;
 981
 982         if (use_fmr == SRP_MAP_NO_FMR) {
 983                 /* Once we're in direct map mode for a request, we don't
 984                  * go back to FMR mode, so no need to update anything
 985                  * other than the descriptor.
 986                  */
 987                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
 988                 return 0;
 989         }
 990
 991         /* If we start at an offset into the FMR page, don't merge into
 992          * the current FMR. Finish it out, and use the kernel's MR for this
 993          * sg entry. This is to avoid potential bugs on some SRP targets
 994          * that were never quite defined, but went away when the initiator
 995          * avoided using FMR on such page fragments.
 996          */
 997         if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
 998                 ret = srp_map_finish_fmr(state, target);
 999                 if (ret)
1000                         return ret;
1001
1002                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
1003                 srp_map_update_start(state, NULL, 0, 0);
1004                 return 0;
1005         }
1006
1007         /* If this is the first sg to go into the FMR, save our position.
1008          * We need to know the first unmapped entry, its index, and the
1009          * first unmapped address within that entry to be able to restart
1010          * mapping after an error.
1011          */
1012         if (!state->unmapped_sg)
1013                 srp_map_update_start(state, sg, sg_index, dma_addr);
1014
1015         while (dma_len) {
1016                 if (state->npages == SRP_FMR_SIZE) {
1017                         ret = srp_map_finish_fmr(state, target);
1018                         if (ret)
1019                                 return ret;
1020
1021                         srp_map_update_start(state, sg, sg_index, dma_addr);
1022                 }
1023
1024                 len = min_t(unsigned int, dma_len, dev->fmr_page_size);
1025
1026                 if (!state->npages)
1027                         state->base_dma_addr = dma_addr;
1028                 state->pages[state->npages++] = dma_addr;
1029                 state->fmr_len += len;
1030                 dma_addr += len;
1031                 dma_len -= len;
1032         }
1033
1034         /* If the last entry of the FMR wasn't a full page, then we need to
1035          * close it out and start a new one -- we can only merge at page
1036          * boundries.
1037          */
1038         ret = 0;
1039         if (len != dev->fmr_page_size) {
1040                 ret = srp_map_finish_fmr(state, target);
1041                 if (!ret)
1042                         srp_map_update_start(state, NULL, 0, 0);
1043         }
1044         return ret;
1045 }
1046
1047 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
1048                         struct srp_request *req)
1049 {
1050         struct scatterlist *scat, *sg;
1051         struct srp_cmd *cmd = req->cmd->buf;
1052         int i, len, nents, count, use_fmr;
1053         struct srp_device *dev;
1054         struct ib_device *ibdev;
1055         struct srp_map_state state;
1056         struct srp_indirect_buf *indirect_hdr;
1057         u32 table_len;
1058         u8 fmt;
1059
1060         if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
1061                 return sizeof (struct srp_cmd);
1062
1063         if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
1064             scmnd->sc_data_direction != DMA_TO_DEVICE) {
1065                 shost_printk(KERN_WARNING, target->scsi_host,
1066                              PFX "Unhandled data direction %d\n",
1067                              scmnd->sc_data_direction);
1068                 return -EINVAL;
1069         }
1070
1071         nents = scsi_sg_count(scmnd);
1072         scat  = scsi_sglist(scmnd);
1073
1074         dev = target->srp_host->srp_dev;
1075         ibdev = dev->dev;
1076
1077         count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
1078         if (unlikely(count == 0))
1079                 return -EIO;
1080
1081         fmt = SRP_DATA_DESC_DIRECT;
1082         len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
1083
1084         if (count == 1) {
1085                 /*
1086                  * The midlayer only generated a single gather/scatter
1087                  * entry, or DMA mapping coalesced everything to a
1088                  * single entry.  So a direct descriptor along with
1089                  * the DMA MR suffices.
1090                  */
1091                 struct srp_direct_buf *buf = (void *) cmd->add_data;
1092
1093                 buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
1094                 buf->key = cpu_to_be32(target->rkey);
1095                 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
1096
1097                 req->nfmr = 0;
1098                 goto map_complete;
1099         }
1100
1101         /* We have more than one scatter/gather entry, so build our indirect
1102          * descriptor table, trying to merge as many entries with FMR as we
1103          * can.
1104          */
1105         indirect_hdr = (void *) cmd->add_data;
1106
1107         ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
1108                                    target->indirect_size, DMA_TO_DEVICE);
1109
1110         memset(&state, 0, sizeof(state));
1111         state.desc      = req->indirect_desc;
1112         state.pages     = req->map_page;
1113         state.next_fmr  = req->fmr_list;
1114
1115         use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
1116
1117         for_each_sg(scat, sg, count, i) {
1118                 if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
1119                         /* FMR mapping failed, so backtrack to the first
1120                          * unmapped entry and continue on without using FMR.
1121                          */
1122                         dma_addr_t dma_addr;
1123                         unsigned int dma_len;
1124
1125 backtrack:
1126                         sg = state.unmapped_sg;
1127                         i = state.unmapped_index;
1128
1129                         dma_addr = ib_sg_dma_address(ibdev, sg);
1130                         dma_len = ib_sg_dma_len(ibdev, sg);
1131                         dma_len -= (state.unmapped_addr - dma_addr);
1132                         dma_addr = state.unmapped_addr;
1133                         use_fmr = SRP_MAP_NO_FMR;
1134                         srp_map_desc(&state, dma_addr, dma_len, target->rkey);
1135                 }
1136         }
1137
1138         if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
1139                 goto backtrack;
1140
1141         /* We've mapped the request, now pull as much of the indirect
1142          * descriptor table as we can into the command buffer. If this
1143          * target is not using an external indirect table, we are
1144          * guaranteed to fit into the command, as the SCSI layer won't
1145          * give us more S/G entries than we allow.
1146          */
1147         req->nfmr = state.nfmr;
1148         if (state.ndesc == 1) {
1149                 /* FMR mapping was able to collapse this to one entry,
1150                  * so use a direct descriptor.
1151                  */
1152                 struct srp_direct_buf *buf = (void *) cmd->add_data;
1153
1154                 *buf = req->indirect_desc[0];
1155                 goto map_complete;
1156         }
1157
1158         if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1159                                                 !target->allow_ext_sg)) {
1160                 shost_printk(KERN_ERR, target->scsi_host,
1161                              "Could not fit S/G list into SRP_CMD\n");
1162                 return -EIO;
1163         }
1164
1165         count = min(state.ndesc, target->cmd_sg_cnt);
1166         table_len = state.ndesc * sizeof (struct srp_direct_buf);
1167
1168         fmt = SRP_DATA_DESC_INDIRECT;
1169         len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1170         len += count * sizeof (struct srp_direct_buf);
1171
1172         memcpy(indirect_hdr->desc_list, req->indirect_desc,
1173                count * sizeof (struct srp_direct_buf));
1174
1175         indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1176         indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1177         indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1178         indirect_hdr->len = cpu_to_be32(state.total_len);
1179
1180         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1181                 cmd->data_out_desc_cnt = count;
1182         else
1183                 cmd->data_in_desc_cnt = count;
1184
1185         ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1186                                       DMA_TO_DEVICE);
1187
1188 map_complete:
1189         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1190                 cmd->buf_fmt = fmt << 4;
1191         else
1192                 cmd->buf_fmt = fmt;
1193
1194         return len;
1195 }
1196
1197 /*
1198  * Return an IU and possible credit to the free pool
1199  */
1200 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1201                           enum srp_iu_type iu_type)
1202 {
1203         unsigned long flags;
1204
1205         spin_lock_irqsave(&target->lock, flags);
1206         list_add(&iu->list, &target->free_tx);
1207         if (iu_type != SRP_IU_RSP)
1208                 ++target->req_lim;
1209         spin_unlock_irqrestore(&target->lock, flags);
1210 }
1211
1212 /*
1213  * Must be called with target->lock held to protect req_lim and free_tx.
1214  * If IU is not sent, it must be returned using srp_put_tx_iu().
1215  *
1216  * Note:
1217  * An upper limit for the number of allocated information units for each
1218  * request type is:
1219  * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1220  *   more than Scsi_Host.can_queue requests.
1221  * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1222  * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1223  *   one unanswered SRP request to an initiator.
1224  */
1225 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1226                                       enum srp_iu_type iu_type)
1227 {
1228         s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1229         struct srp_iu *iu;
1230
1231         srp_send_completion(target->send_cq, target);
1232
1233         if (list_empty(&target->free_tx))
1234                 return NULL;
1235
1236         /* Initiator responses to target requests do not consume credits */
1237         if (iu_type != SRP_IU_RSP) {
1238                 if (target->req_lim <= rsv) {
1239                         ++target->zero_req_lim;
1240                         return NULL;
1241                 }
1242
1243                 --target->req_lim;
1244         }
1245
1246         iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1247         list_del(&iu->list);
1248         return iu;
1249 }
1250
1251 static int srp_post_send(struct srp_target_port *target,
1252                          struct srp_iu *iu, int len)
1253 {
1254         struct ib_sge list;
1255         struct ib_send_wr wr, *bad_wr;
1256
1257         list.addr   = iu->dma;
1258         list.length = len;
1259         list.lkey   = target->lkey;
1260
1261         wr.next       = NULL;
1262         wr.wr_id      = (uintptr_t) iu;
1263         wr.sg_list    = &list;
1264         wr.num_sge    = 1;
1265         wr.opcode     = IB_WR_SEND;
1266         wr.send_flags = IB_SEND_SIGNALED;
1267
1268         return ib_post_send(target->qp, &wr, &bad_wr);
1269 }
1270
1271 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1272 {
1273         struct ib_recv_wr wr, *bad_wr;
1274         struct ib_sge list;
1275
1276         list.addr   = iu->dma;
1277         list.length = iu->size;
1278         list.lkey   = target->lkey;
1279
1280         wr.next     = NULL;
1281         wr.wr_id    = (uintptr_t) iu;
1282         wr.sg_list  = &list;
1283         wr.num_sge  = 1;
1284
1285         return ib_post_recv(target->qp, &wr, &bad_wr);
1286 }
1287
1288 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1289 {
1290         struct srp_request *req;
1291         struct scsi_cmnd *scmnd;
1292         unsigned long flags;
1293
1294         if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1295                 spin_lock_irqsave(&target->lock, flags);
1296                 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1297                 spin_unlock_irqrestore(&target->lock, flags);
1298
1299                 target->tsk_mgmt_status = -1;
1300                 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1301                         target->tsk_mgmt_status = rsp->data[3];
1302                 complete(&target->tsk_mgmt_done);
1303         } else {
1304                 req = &target->req_ring[rsp->tag];
1305                 scmnd = srp_claim_req(target, req, NULL, NULL);
1306                 if (!scmnd) {
1307                         shost_printk(KERN_ERR, target->scsi_host,
1308                                      "Null scmnd for RSP w/tag %016llx\n",
1309                                      (unsigned long long) rsp->tag);
1310
1311                         spin_lock_irqsave(&target->lock, flags);
1312                         target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1313                         spin_unlock_irqrestore(&target->lock, flags);
1314
1315                         return;
1316                 }
1317                 scmnd->result = rsp->status;
1318
1319                 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1320                         memcpy(scmnd->sense_buffer, rsp->data +
1321                                be32_to_cpu(rsp->resp_data_len),
1322                                min_t(int, be32_to_cpu(rsp->sense_data_len),
1323                                      SCSI_SENSE_BUFFERSIZE));
1324                 }
1325
1326                 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1327                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1328                 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1329                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1330
1331                 srp_free_req(target, req, scmnd,
1332                              be32_to_cpu(rsp->req_lim_delta));
1333
1334                 scmnd->host_scribble = NULL;
1335                 scmnd->scsi_done(scmnd);
1336         }
1337 }
1338
1339 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1340                                void *rsp, int len)
1341 {
1342         struct ib_device *dev = target->srp_host->srp_dev->dev;
1343         unsigned long flags;
1344         struct srp_iu *iu;
1345         int err;
1346
1347         spin_lock_irqsave(&target->lock, flags);
1348         target->req_lim += req_delta;
1349         iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1350         spin_unlock_irqrestore(&target->lock, flags);
1351
1352         if (!iu) {
1353                 shost_printk(KERN_ERR, target->scsi_host, PFX
1354                              "no IU available to send response\n");
1355                 return 1;
1356         }
1357
1358         ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1359         memcpy(iu->buf, rsp, len);
1360         ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1361
1362         err = srp_post_send(target, iu, len);
1363         if (err) {
1364                 shost_printk(KERN_ERR, target->scsi_host, PFX
1365                              "unable to post response: %d\n", err);
1366                 srp_put_tx_iu(target, iu, SRP_IU_RSP);
1367         }
1368
1369         return err;
1370 }
1371
1372 static void srp_process_cred_req(struct srp_target_port *target,
1373                                  struct srp_cred_req *req)
1374 {
1375         struct srp_cred_rsp rsp = {
1376                 .opcode = SRP_CRED_RSP,
1377                 .tag = req->tag,
1378         };
1379         s32 delta = be32_to_cpu(req->req_lim_delta);
1380
1381         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1382                 shost_printk(KERN_ERR, target->scsi_host, PFX
1383                              "problems processing SRP_CRED_REQ\n");
1384 }
1385
1386 static void srp_process_aer_req(struct srp_target_port *target,
1387                                 struct srp_aer_req *req)
1388 {
1389         struct srp_aer_rsp rsp = {
1390                 .opcode = SRP_AER_RSP,
1391                 .tag = req->tag,
1392         };
1393         s32 delta = be32_to_cpu(req->req_lim_delta);
1394
1395         shost_printk(KERN_ERR, target->scsi_host, PFX
1396                      "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1397
1398         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1399                 shost_printk(KERN_ERR, target->scsi_host, PFX
1400                              "problems processing SRP_AER_REQ\n");
1401 }
1402
1403 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1404 {
1405         struct ib_device *dev = target->srp_host->srp_dev->dev;
1406         struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1407         int res;
1408         u8 opcode;
1409
1410         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1411                                    DMA_FROM_DEVICE);
1412
1413         opcode = *(u8 *) iu->buf;
1414
1415         if (0) {
1416                 shost_printk(KERN_ERR, target->scsi_host,
1417                              PFX "recv completion, opcode 0x%02x\n", opcode);
1418                 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1419                                iu->buf, wc->byte_len, true);
1420         }
1421
1422         switch (opcode) {
1423         case SRP_RSP:
1424                 srp_process_rsp(target, iu->buf);
1425                 break;
1426
1427         case SRP_CRED_REQ:
1428                 srp_process_cred_req(target, iu->buf);
1429                 break;
1430
1431         case SRP_AER_REQ:
1432                 srp_process_aer_req(target, iu->buf);
1433                 break;
1434
1435         case SRP_T_LOGOUT:
1436                 /* XXX Handle target logout */
1437                 shost_printk(KERN_WARNING, target->scsi_host,
1438                              PFX "Got target logout request\n");
1439                 break;
1440
1441         default:
1442                 shost_printk(KERN_WARNING, target->scsi_host,
1443                              PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1444                 break;
1445         }
1446
1447         ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1448                                       DMA_FROM_DEVICE);
1449
1450         res = srp_post_recv(target, iu);
1451         if (res != 0)
1452                 shost_printk(KERN_ERR, target->scsi_host,
1453                              PFX "Recv failed with error code %d\n", res);
1454 }
1455
1456 /**
1457  * srp_tl_err_work() - handle a transport layer error
1458  *
1459  * Note: This function may get invoked before the rport has been created,
1460  * hence the target->rport test.
1461  */
1462 static void srp_tl_err_work(struct work_struct *work)
1463 {
1464         struct srp_target_port *target;
1465
1466         target = container_of(work, struct srp_target_port, tl_err_work);
1467         if (target->rport)
1468                 srp_start_tl_fail_timers(target->rport);
1469 }
1470
1471 static void srp_handle_qp_err(enum ib_wc_status wc_status, bool send_err,
1472                               struct srp_target_port *target)
1473 {
1474         if (target->connected && !target->qp_in_error) {
1475                 shost_printk(KERN_ERR, target->scsi_host,
1476                              PFX "failed %s status %d\n",
1477                              send_err ? "send" : "receive",
1478                              wc_status);
1479                 queue_work(system_long_wq, &target->tl_err_work);
1480         }
1481         target->qp_in_error = true;
1482 }
1483
1484 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1485 {
1486         struct srp_target_port *target = target_ptr;
1487         struct ib_wc wc;
1488
1489         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1490         while (ib_poll_cq(cq, 1, &wc) > 0) {
1491                 if (likely(wc.status == IB_WC_SUCCESS)) {
1492                         srp_handle_recv(target, &wc);
1493                 } else {
1494                         srp_handle_qp_err(wc.status, false, target);
1495                 }
1496         }
1497 }
1498
1499 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1500 {
1501         struct srp_target_port *target = target_ptr;
1502         struct ib_wc wc;
1503         struct srp_iu *iu;
1504
1505         while (ib_poll_cq(cq, 1, &wc) > 0) {
1506                 if (likely(wc.status == IB_WC_SUCCESS)) {
1507                         iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1508                         list_add(&iu->list, &target->free_tx);
1509                 } else {
1510                         srp_handle_qp_err(wc.status, true, target);
1511                 }
1512         }
1513 }
1514
1515 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1516 {
1517         struct srp_target_port *target = host_to_target(shost);
1518         struct srp_rport *rport = target->rport;
1519         struct srp_request *req;
1520         struct srp_iu *iu;
1521         struct srp_cmd *cmd;
1522         struct ib_device *dev;
1523         unsigned long flags;
1524         int len, result;
1525         const bool in_scsi_eh = !in_interrupt() && current == shost->ehandler;
1526
1527         /*
1528          * The SCSI EH thread is the only context from which srp_queuecommand()
1529          * can get invoked for blocked devices (SDEV_BLOCK /
1530          * SDEV_CREATED_BLOCK). Avoid racing with srp_reconnect_rport() by
1531          * locking the rport mutex if invoked from inside the SCSI EH.
1532          */
1533         if (in_scsi_eh)
1534                 mutex_lock(&rport->mutex);
1535
1536         result = srp_chkready(target->rport);
1537         if (unlikely(result)) {
1538                 scmnd->result = result;
1539                 scmnd->scsi_done(scmnd);
1540                 goto unlock_rport;
1541         }
1542
1543         spin_lock_irqsave(&target->lock, flags);
1544         iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1545         if (!iu)
1546                 goto err_unlock;
1547
1548         req = list_first_entry(&target->free_reqs, struct srp_request, list);
1549         list_del(&req->list);
1550         spin_unlock_irqrestore(&target->lock, flags);
1551
1552         dev = target->srp_host->srp_dev->dev;
1553         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1554                                    DMA_TO_DEVICE);
1555
1556         scmnd->result        = 0;
1557         scmnd->host_scribble = (void *) req;
1558
1559         cmd = iu->buf;
1560         memset(cmd, 0, sizeof *cmd);
1561
1562         cmd->opcode = SRP_CMD;
1563         cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
1564         cmd->tag    = req->index;
1565         memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1566
1567         req->scmnd    = scmnd;
1568         req->cmd      = iu;
1569
1570         len = srp_map_data(scmnd, target, req);
1571         if (len < 0) {
1572                 shost_printk(KERN_ERR, target->scsi_host,
1573                              PFX "Failed to map data\n");
1574                 goto err_iu;
1575         }
1576
1577         ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1578                                       DMA_TO_DEVICE);
1579
1580         if (srp_post_send(target, iu, len)) {
1581                 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1582                 goto err_unmap;
1583         }
1584
1585 unlock_rport:
1586         if (in_scsi_eh)
1587                 mutex_unlock(&rport->mutex);
1588
1589         return 0;
1590
1591 err_unmap:
1592         srp_unmap_data(scmnd, target, req);
1593
1594 err_iu:
1595         srp_put_tx_iu(target, iu, SRP_IU_CMD);
1596
1597         spin_lock_irqsave(&target->lock, flags);
1598         list_add(&req->list, &target->free_reqs);
1599
1600 err_unlock:
1601         spin_unlock_irqrestore(&target->lock, flags);
1602
1603         if (in_scsi_eh)
1604                 mutex_unlock(&rport->mutex);
1605
1606         return SCSI_MLQUEUE_HOST_BUSY;
1607 }
1608
1609 /*
1610  * Note: the resources allocated in this function are freed in
1611  * srp_free_target_ib().
1612  */
1613 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1614 {
1615         int i;
1616
1617         target->rx_ring = kzalloc(target->queue_size * sizeof(*target->rx_ring),
1618                                   GFP_KERNEL);
1619         if (!target->rx_ring)
1620                 goto err_no_ring;
1621         target->tx_ring = kzalloc(target->queue_size * sizeof(*target->tx_ring),
1622                                   GFP_KERNEL);
1623         if (!target->tx_ring)
1624                 goto err_no_ring;
1625
1626         for (i = 0; i < target->queue_size; ++i) {
1627                 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1628                                                   target->max_ti_iu_len,
1629                                                   GFP_KERNEL, DMA_FROM_DEVICE);
1630                 if (!target->rx_ring[i])
1631                         goto err;
1632         }
1633
1634         for (i = 0; i < target->queue_size; ++i) {
1635                 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1636                                                   target->max_iu_len,
1637                                                   GFP_KERNEL, DMA_TO_DEVICE);
1638                 if (!target->tx_ring[i])
1639                         goto err;
1640
1641                 list_add(&target->tx_ring[i]->list, &target->free_tx);
1642         }
1643
1644         return 0;
1645
1646 err:
1647         for (i = 0; i < target->queue_size; ++i) {
1648                 srp_free_iu(target->srp_host, target->rx_ring[i]);
1649                 srp_free_iu(target->srp_host, target->tx_ring[i]);
1650         }
1651
1652
1653 err_no_ring:
1654         kfree(target->tx_ring);
1655         target->tx_ring = NULL;
1656         kfree(target->rx_ring);
1657         target->rx_ring = NULL;
1658
1659         return -ENOMEM;
1660 }
1661
1662 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
1663 {
1664         uint64_t T_tr_ns, max_compl_time_ms;
1665         uint32_t rq_tmo_jiffies;
1666
1667         /*
1668          * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
1669          * table 91), both the QP timeout and the retry count have to be set
1670          * for RC QP's during the RTR to RTS transition.
1671          */
1672         WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
1673                      (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
1674
1675         /*
1676          * Set target->rq_tmo_jiffies to one second more than the largest time
1677          * it can take before an error completion is generated. See also
1678          * C9-140..142 in the IBTA spec for more information about how to
1679          * convert the QP Local ACK Timeout value to nanoseconds.
1680          */
1681         T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
1682         max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
1683         do_div(max_compl_time_ms, NSEC_PER_MSEC);
1684         rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
1685
1686         return rq_tmo_jiffies;
1687 }
1688
1689 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1690                                struct srp_login_rsp *lrsp,
1691                                struct srp_target_port *target)
1692 {
1693         struct ib_qp_attr *qp_attr = NULL;
1694         int attr_mask = 0;
1695         int ret;
1696         int i;
1697
1698         if (lrsp->opcode == SRP_LOGIN_RSP) {
1699                 target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1700                 target->req_lim       = be32_to_cpu(lrsp->req_lim_delta);
1701
1702                 /*
1703                  * Reserve credits for task management so we don't
1704                  * bounce requests back to the SCSI mid-layer.
1705                  */
1706                 target->scsi_host->can_queue
1707                         = min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1708                               target->scsi_host->can_queue);
1709                 target->scsi_host->cmd_per_lun
1710                         = min_t(int, target->scsi_host->can_queue,
1711                                 target->scsi_host->cmd_per_lun);
1712         } else {
1713                 shost_printk(KERN_WARNING, target->scsi_host,
1714                              PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1715                 ret = -ECONNRESET;
1716                 goto error;
1717         }
1718
1719         if (!target->rx_ring) {
1720                 ret = srp_alloc_iu_bufs(target);
1721                 if (ret)
1722                         goto error;
1723         }
1724
1725         ret = -ENOMEM;
1726         qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1727         if (!qp_attr)
1728                 goto error;
1729
1730         qp_attr->qp_state = IB_QPS_RTR;
1731         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1732         if (ret)
1733                 goto error_free;
1734
1735         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1736         if (ret)
1737                 goto error_free;
1738
1739         for (i = 0; i < target->queue_size; i++) {
1740                 struct srp_iu *iu = target->rx_ring[i];
1741                 ret = srp_post_recv(target, iu);
1742                 if (ret)
1743                         goto error_free;
1744         }
1745
1746         qp_attr->qp_state = IB_QPS_RTS;
1747         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1748         if (ret)
1749                 goto error_free;
1750
1751         target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
1752
1753         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1754         if (ret)
1755                 goto error_free;
1756
1757         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1758
1759 error_free:
1760         kfree(qp_attr);
1761
1762 error:
1763         target->status = ret;
1764 }
1765
1766 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1767                                struct ib_cm_event *event,
1768                                struct srp_target_port *target)
1769 {
1770         struct Scsi_Host *shost = target->scsi_host;
1771         struct ib_class_port_info *cpi;
1772         int opcode;
1773
1774         switch (event->param.rej_rcvd.reason) {
1775         case IB_CM_REJ_PORT_CM_REDIRECT:
1776                 cpi = event->param.rej_rcvd.ari;
1777                 target->path.dlid = cpi->redirect_lid;
1778                 target->path.pkey = cpi->redirect_pkey;
1779                 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1780                 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1781
1782                 target->status = target->path.dlid ?
1783                         SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1784                 break;
1785
1786         case IB_CM_REJ_PORT_REDIRECT:
1787                 if (srp_target_is_topspin(target)) {
1788                         /*
1789                          * Topspin/Cisco SRP gateways incorrectly send
1790                          * reject reason code 25 when they mean 24
1791                          * (port redirect).
1792                          */
1793                         memcpy(target->path.dgid.raw,
1794                                event->param.rej_rcvd.ari, 16);
1795
1796                         shost_printk(KERN_DEBUG, shost,
1797                                      PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1798                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1799                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1800
1801                         target->status = SRP_PORT_REDIRECT;
1802                 } else {
1803                         shost_printk(KERN_WARNING, shost,
1804                                      "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1805                         target->status = -ECONNRESET;
1806                 }
1807                 break;
1808
1809         case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1810                 shost_printk(KERN_WARNING, shost,
1811                             "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1812                 target->status = -ECONNRESET;
1813                 break;
1814
1815         case IB_CM_REJ_CONSUMER_DEFINED:
1816                 opcode = *(u8 *) event->private_data;
1817                 if (opcode == SRP_LOGIN_REJ) {
1818                         struct srp_login_rej *rej = event->private_data;
1819                         u32 reason = be32_to_cpu(rej->reason);
1820
1821                         if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1822                                 shost_printk(KERN_WARNING, shost,
1823                                              PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1824                         else
1825                                 shost_printk(KERN_WARNING, shost, PFX
1826                                              "SRP LOGIN from %pI6 to %pI6 REJECTED, reason 0x%08x\n",
1827                                              target->path.sgid.raw,
1828                                              target->orig_dgid, reason);
1829                 } else
1830                         shost_printk(KERN_WARNING, shost,
1831                                      "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1832                                      " opcode 0x%02x\n", opcode);
1833                 target->status = -ECONNRESET;
1834                 break;
1835
1836         case IB_CM_REJ_STALE_CONN:
1837                 shost_printk(KERN_WARNING, shost, "  REJ reason: stale connection\n");
1838                 target->status = SRP_STALE_CONN;
1839                 break;
1840
1841         default:
1842                 shost_printk(KERN_WARNING, shost, "  REJ reason 0x%x\n",
1843                              event->param.rej_rcvd.reason);
1844                 target->status = -ECONNRESET;
1845         }
1846 }
1847
1848 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1849 {
1850         struct srp_target_port *target = cm_id->context;
1851         int comp = 0;
1852
1853         switch (event->event) {
1854         case IB_CM_REQ_ERROR:
1855                 shost_printk(KERN_DEBUG, target->scsi_host,
1856                              PFX "Sending CM REQ failed\n");
1857                 comp = 1;
1858                 target->status = -ECONNRESET;
1859                 break;
1860
1861         case IB_CM_REP_RECEIVED:
1862                 comp = 1;
1863                 srp_cm_rep_handler(cm_id, event->private_data, target);
1864                 break;
1865
1866         case IB_CM_REJ_RECEIVED:
1867                 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1868                 comp = 1;
1869
1870                 srp_cm_rej_handler(cm_id, event, target);
1871                 break;
1872
1873         case IB_CM_DREQ_RECEIVED:
1874                 shost_printk(KERN_WARNING, target->scsi_host,
1875                              PFX "DREQ received - connection closed\n");
1876                 srp_change_conn_state(target, false);
1877                 if (ib_send_cm_drep(cm_id, NULL, 0))
1878                         shost_printk(KERN_ERR, target->scsi_host,
1879                                      PFX "Sending CM DREP failed\n");
1880                 queue_work(system_long_wq, &target->tl_err_work);
1881                 break;
1882
1883         case IB_CM_TIMEWAIT_EXIT:
1884                 shost_printk(KERN_ERR, target->scsi_host,
1885                              PFX "connection closed\n");
1886                 comp = 1;
1887
1888                 target->status = 0;
1889                 break;
1890
1891         case IB_CM_MRA_RECEIVED:
1892         case IB_CM_DREQ_ERROR:
1893         case IB_CM_DREP_RECEIVED:
1894                 break;
1895
1896         default:
1897                 shost_printk(KERN_WARNING, target->scsi_host,
1898                              PFX "Unhandled CM event %d\n", event->event);
1899                 break;
1900         }
1901
1902         if (comp)
1903                 complete(&target->done);
1904
1905         return 0;
1906 }
1907
1908 /**
1909  * srp_change_queue_type - changing device queue tag type
1910  * @sdev: scsi device struct
1911  * @tag_type: requested tag type
1912  *
1913  * Returns queue tag type.
1914  */
1915 static int
1916 srp_change_queue_type(struct scsi_device *sdev, int tag_type)
1917 {
1918         if (sdev->tagged_supported) {
1919                 scsi_set_tag_type(sdev, tag_type);
1920                 if (tag_type)
1921                         scsi_activate_tcq(sdev, sdev->queue_depth);
1922                 else
1923                         scsi_deactivate_tcq(sdev, sdev->queue_depth);
1924         } else
1925                 tag_type = 0;
1926
1927         return tag_type;
1928 }
1929
1930 /**
1931  * srp_change_queue_depth - setting device queue depth
1932  * @sdev: scsi device struct
1933  * @qdepth: requested queue depth
1934  * @reason: SCSI_QDEPTH_DEFAULT/SCSI_QDEPTH_QFULL/SCSI_QDEPTH_RAMP_UP
1935  * (see include/scsi/scsi_host.h for definition)
1936  *
1937  * Returns queue depth.
1938  */
1939 static int
1940 srp_change_queue_depth(struct scsi_device *sdev, int qdepth, int reason)
1941 {
1942         struct Scsi_Host *shost = sdev->host;
1943         int max_depth;
1944         if (reason == SCSI_QDEPTH_DEFAULT || reason == SCSI_QDEPTH_RAMP_UP) {
1945                 max_depth = shost->can_queue;
1946                 if (!sdev->tagged_supported)
1947                         max_depth = 1;
1948                 if (qdepth > max_depth)
1949                         qdepth = max_depth;
1950                 scsi_adjust_queue_depth(sdev, scsi_get_tag_type(sdev), qdepth);
1951         } else if (reason == SCSI_QDEPTH_QFULL)
1952                 scsi_track_queue_full(sdev, qdepth);
1953         else
1954                 return -EOPNOTSUPP;
1955
1956         return sdev->queue_depth;
1957 }
1958
1959 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1960                              u64 req_tag, unsigned int lun, u8 func)
1961 {
1962         struct srp_rport *rport = target->rport;
1963         struct ib_device *dev = target->srp_host->srp_dev->dev;
1964         struct srp_iu *iu;
1965         struct srp_tsk_mgmt *tsk_mgmt;
1966
1967         if (!target->connected || target->qp_in_error)
1968                 return -1;
1969
1970         init_completion(&target->tsk_mgmt_done);
1971
1972         /*
1973          * Lock the rport mutex to avoid that srp_create_target_ib() is
1974          * invoked while a task management function is being sent.
1975          */
1976         mutex_lock(&rport->mutex);
1977         spin_lock_irq(&target->lock);
1978         iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1979         spin_unlock_irq(&target->lock);
1980
1981         if (!iu) {
1982                 mutex_unlock(&rport->mutex);
1983
1984                 return -1;
1985         }
1986
1987         ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1988                                    DMA_TO_DEVICE);
1989         tsk_mgmt = iu->buf;
1990         memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1991
1992         tsk_mgmt->opcode        = SRP_TSK_MGMT;
1993         tsk_mgmt->lun           = cpu_to_be64((u64) lun << 48);
1994         tsk_mgmt->tag           = req_tag | SRP_TAG_TSK_MGMT;
1995         tsk_mgmt->tsk_mgmt_func = func;
1996         tsk_mgmt->task_tag      = req_tag;
1997
1998         ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1999                                       DMA_TO_DEVICE);
2000         if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
2001                 srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
2002                 mutex_unlock(&rport->mutex);
2003
2004                 return -1;
2005         }
2006         mutex_unlock(&rport->mutex);
2007
2008         if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
2009                                          msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
2010                 return -1;
2011
2012         return 0;
2013 }
2014
2015 static int srp_abort(struct scsi_cmnd *scmnd)
2016 {
2017         struct srp_target_port *target = host_to_target(scmnd->device->host);
2018         struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
2019         int ret;
2020
2021         shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
2022
2023         if (!req || !srp_claim_req(target, req, NULL, scmnd))
2024                 return SUCCESS;
2025         if (srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
2026                               SRP_TSK_ABORT_TASK) == 0)
2027                 ret = SUCCESS;
2028         else if (target->rport->state == SRP_RPORT_LOST)
2029                 ret = FAST_IO_FAIL;
2030         else
2031                 ret = FAILED;
2032         srp_free_req(target, req, scmnd, 0);
2033         scmnd->result = DID_ABORT << 16;
2034         scmnd->scsi_done(scmnd);
2035
2036         return ret;
2037 }
2038
2039 static int srp_reset_device(struct scsi_cmnd *scmnd)
2040 {
2041         struct srp_target_port *target = host_to_target(scmnd->device->host);
2042         int i;
2043
2044         shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
2045
2046         if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
2047                               SRP_TSK_LUN_RESET))
2048                 return FAILED;
2049         if (target->tsk_mgmt_status)
2050                 return FAILED;
2051
2052         for (i = 0; i < target->req_ring_size; ++i) {
2053                 struct srp_request *req = &target->req_ring[i];
2054                 srp_finish_req(target, req, scmnd->device, DID_RESET << 16);
2055         }
2056
2057         return SUCCESS;
2058 }
2059
2060 static int srp_reset_host(struct scsi_cmnd *scmnd)
2061 {
2062         struct srp_target_port *target = host_to_target(scmnd->device->host);
2063
2064         shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
2065
2066         return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
2067 }
2068
2069 static int srp_slave_configure(struct scsi_device *sdev)
2070 {
2071         struct Scsi_Host *shost = sdev->host;
2072         struct srp_target_port *target = host_to_target(shost);
2073         struct request_queue *q = sdev->request_queue;
2074         unsigned long timeout;
2075
2076         if (sdev->type == TYPE_DISK) {
2077                 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
2078                 blk_queue_rq_timeout(q, timeout);
2079         }
2080
2081         return 0;
2082 }
2083
2084 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
2085                            char *buf)
2086 {
2087         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2088
2089         return sprintf(buf, "0x%016llx\n",
2090                        (unsigned long long) be64_to_cpu(target->id_ext));
2091 }
2092
2093 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
2094                              char *buf)
2095 {
2096         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2097
2098         return sprintf(buf, "0x%016llx\n",
2099                        (unsigned long long) be64_to_cpu(target->ioc_guid));
2100 }
2101
2102 static ssize_t show_service_id(struct device *dev,
2103                                struct device_attribute *attr, char *buf)
2104 {
2105         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2106
2107         return sprintf(buf, "0x%016llx\n",
2108                        (unsigned long long) be64_to_cpu(target->service_id));
2109 }
2110
2111 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
2112                          char *buf)
2113 {
2114         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2115
2116         return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
2117 }
2118
2119 static ssize_t show_sgid(struct device *dev, struct device_attribute *attr,
2120                          char *buf)
2121 {
2122         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2123
2124         return sprintf(buf, "%pI6\n", target->path.sgid.raw);
2125 }
2126
2127 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
2128                          char *buf)
2129 {
2130         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2131
2132         return sprintf(buf, "%pI6\n", target->path.dgid.raw);
2133 }
2134
2135 static ssize_t show_orig_dgid(struct device *dev,
2136                               struct device_attribute *attr, char *buf)
2137 {
2138         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2139
2140         return sprintf(buf, "%pI6\n", target->orig_dgid);
2141 }
2142
2143 static ssize_t show_req_lim(struct device *dev,
2144                             struct device_attribute *attr, char *buf)
2145 {
2146         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2147
2148         return sprintf(buf, "%d\n", target->req_lim);
2149 }
2150
2151 static ssize_t show_zero_req_lim(struct device *dev,
2152                                  struct device_attribute *attr, char *buf)
2153 {
2154         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2155
2156         return sprintf(buf, "%d\n", target->zero_req_lim);
2157 }
2158
2159 static ssize_t show_local_ib_port(struct device *dev,
2160                                   struct device_attribute *attr, char *buf)
2161 {
2162         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2163
2164         return sprintf(buf, "%d\n", target->srp_host->port);
2165 }
2166
2167 static ssize_t show_local_ib_device(struct device *dev,
2168                                     struct device_attribute *attr, char *buf)
2169 {
2170         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2171
2172         return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
2173 }
2174
2175 static ssize_t show_comp_vector(struct device *dev,
2176                                 struct device_attribute *attr, char *buf)
2177 {
2178         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2179
2180         return sprintf(buf, "%d\n", target->comp_vector);
2181 }
2182
2183 static ssize_t show_tl_retry_count(struct device *dev,
2184                                    struct device_attribute *attr, char *buf)
2185 {
2186         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2187
2188         return sprintf(buf, "%d\n", target->tl_retry_count);
2189 }
2190
2191 static ssize_t show_cmd_sg_entries(struct device *dev,
2192                                    struct device_attribute *attr, char *buf)
2193 {
2194         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2195
2196         return sprintf(buf, "%u\n", target->cmd_sg_cnt);
2197 }
2198
2199 static ssize_t show_allow_ext_sg(struct device *dev,
2200                                  struct device_attribute *attr, char *buf)
2201 {
2202         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2203
2204         return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
2205 }
2206
2207 static DEVICE_ATTR(id_ext,          S_IRUGO, show_id_ext,          NULL);
2208 static DEVICE_ATTR(ioc_guid,        S_IRUGO, show_ioc_guid,        NULL);
2209 static DEVICE_ATTR(service_id,      S_IRUGO, show_service_id,      NULL);
2210 static DEVICE_ATTR(pkey,            S_IRUGO, show_pkey,            NULL);
2211 static DEVICE_ATTR(sgid,            S_IRUGO, show_sgid,            NULL);
2212 static DEVICE_ATTR(dgid,            S_IRUGO, show_dgid,            NULL);
2213 static DEVICE_ATTR(orig_dgid,       S_IRUGO, show_orig_dgid,       NULL);
2214 static DEVICE_ATTR(req_lim,         S_IRUGO, show_req_lim,         NULL);
2215 static DEVICE_ATTR(zero_req_lim,    S_IRUGO, show_zero_req_lim,    NULL);
2216 static DEVICE_ATTR(local_ib_port,   S_IRUGO, show_local_ib_port,   NULL);
2217 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
2218 static DEVICE_ATTR(comp_vector,     S_IRUGO, show_comp_vector,     NULL);
2219 static DEVICE_ATTR(tl_retry_count,  S_IRUGO, show_tl_retry_count,  NULL);
2220 static DEVICE_ATTR(cmd_sg_entries,  S_IRUGO, show_cmd_sg_entries,  NULL);
2221 static DEVICE_ATTR(allow_ext_sg,    S_IRUGO, show_allow_ext_sg,    NULL);
2222
2223 static struct device_attribute *srp_host_attrs[] = {
2224         &dev_attr_id_ext,
2225         &dev_attr_ioc_guid,
2226         &dev_attr_service_id,
2227         &dev_attr_pkey,
2228         &dev_attr_sgid,
2229         &dev_attr_dgid,
2230         &dev_attr_orig_dgid,
2231         &dev_attr_req_lim,
2232         &dev_attr_zero_req_lim,
2233         &dev_attr_local_ib_port,
2234         &dev_attr_local_ib_device,
2235         &dev_attr_comp_vector,
2236         &dev_attr_tl_retry_count,
2237         &dev_attr_cmd_sg_entries,
2238         &dev_attr_allow_ext_sg,
2239         NULL
2240 };
2241
2242 static struct scsi_host_template srp_template = {
2243         .module                         = THIS_MODULE,
2244         .name                           = "InfiniBand SRP initiator",
2245         .proc_name                      = DRV_NAME,
2246         .slave_configure                = srp_slave_configure,
2247         .info                           = srp_target_info,
2248         .queuecommand                   = srp_queuecommand,
2249         .change_queue_depth             = srp_change_queue_depth,
2250         .change_queue_type              = srp_change_queue_type,
2251         .eh_abort_handler               = srp_abort,
2252         .eh_device_reset_handler        = srp_reset_device,
2253         .eh_host_reset_handler          = srp_reset_host,
2254         .skip_settle_delay              = true,
2255         .sg_tablesize                   = SRP_DEF_SG_TABLESIZE,
2256         .can_queue                      = SRP_DEFAULT_CMD_SQ_SIZE,
2257         .this_id                        = -1,
2258         .cmd_per_lun                    = SRP_DEFAULT_CMD_SQ_SIZE,
2259         .use_clustering                 = ENABLE_CLUSTERING,
2260         .shost_attrs                    = srp_host_attrs
2261 };
2262
2263 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
2264 {
2265         struct srp_rport_identifiers ids;
2266         struct srp_rport *rport;
2267
2268         sprintf(target->target_name, "SRP.T10:%016llX",
2269                  (unsigned long long) be64_to_cpu(target->id_ext));
2270
2271         if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
2272                 return -ENODEV;
2273
2274         memcpy(ids.port_id, &target->id_ext, 8);
2275         memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2276         ids.roles = SRP_RPORT_ROLE_TARGET;
2277         rport = srp_rport_add(target->scsi_host, &ids);
2278         if (IS_ERR(rport)) {
2279                 scsi_remove_host(target->scsi_host);
2280                 return PTR_ERR(rport);
2281         }
2282
2283         rport->lld_data = target;
2284         target->rport = rport;
2285
2286         spin_lock(&host->target_lock);
2287         list_add_tail(&target->list, &host->target_list);
2288         spin_unlock(&host->target_lock);
2289
2290         target->state = SRP_TARGET_LIVE;
2291
2292         scsi_scan_target(&target->scsi_host->shost_gendev,
2293                          0, target->scsi_id, SCAN_WILD_CARD, 0);
2294
2295         return 0;
2296 }
2297
2298 static void srp_release_dev(struct device *dev)
2299 {
2300         struct srp_host *host =
2301                 container_of(dev, struct srp_host, dev);
2302
2303         complete(&host->released);
2304 }
2305
2306 static struct class srp_class = {
2307         .name    = "infiniband_srp",
2308         .dev_release = srp_release_dev
2309 };
2310
2311 /**
2312  * srp_conn_unique() - check whether the connection to a target is unique
2313  */
2314 static bool srp_conn_unique(struct srp_host *host,
2315                             struct srp_target_port *target)
2316 {
2317         struct srp_target_port *t;
2318         bool ret = false;
2319
2320         if (target->state == SRP_TARGET_REMOVED)
2321                 goto out;
2322
2323         ret = true;
2324
2325         spin_lock(&host->target_lock);
2326         list_for_each_entry(t, &host->target_list, list) {
2327                 if (t != target &&
2328                     target->id_ext == t->id_ext &&
2329                     target->ioc_guid == t->ioc_guid &&
2330                     target->initiator_ext == t->initiator_ext) {
2331                         ret = false;
2332                         break;
2333                 }
2334         }
2335         spin_unlock(&host->target_lock);
2336
2337 out:
2338         return ret;
2339 }
2340
2341 /*
2342  * Target ports are added by writing
2343  *
2344  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2345  *     pkey=<P_Key>,service_id=<service ID>
2346  *
2347  * to the add_target sysfs attribute.
2348  */
2349 enum {
2350         SRP_OPT_ERR             = 0,
2351         SRP_OPT_ID_EXT          = 1 << 0,
2352         SRP_OPT_IOC_GUID        = 1 << 1,
2353         SRP_OPT_DGID            = 1 << 2,
2354         SRP_OPT_PKEY            = 1 << 3,
2355         SRP_OPT_SERVICE_ID      = 1 << 4,
2356         SRP_OPT_MAX_SECT        = 1 << 5,
2357         SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2358         SRP_OPT_IO_CLASS        = 1 << 7,
2359         SRP_OPT_INITIATOR_EXT   = 1 << 8,
2360         SRP_OPT_CMD_SG_ENTRIES  = 1 << 9,
2361         SRP_OPT_ALLOW_EXT_SG    = 1 << 10,
2362         SRP_OPT_SG_TABLESIZE    = 1 << 11,
2363         SRP_OPT_COMP_VECTOR     = 1 << 12,
2364         SRP_OPT_TL_RETRY_COUNT  = 1 << 13,
2365         SRP_OPT_QUEUE_SIZE      = 1 << 14,
2366         SRP_OPT_ALL             = (SRP_OPT_ID_EXT       |
2367                                    SRP_OPT_IOC_GUID     |
2368                                    SRP_OPT_DGID         |
2369                                    SRP_OPT_PKEY         |
2370                                    SRP_OPT_SERVICE_ID),
2371 };
2372
2373 static const match_table_t srp_opt_tokens = {
2374         { SRP_OPT_ID_EXT,               "id_ext=%s"             },
2375         { SRP_OPT_IOC_GUID,             "ioc_guid=%s"           },
2376         { SRP_OPT_DGID,                 "dgid=%s"               },
2377         { SRP_OPT_PKEY,                 "pkey=%x"               },
2378         { SRP_OPT_SERVICE_ID,           "service_id=%s"         },
2379         { SRP_OPT_MAX_SECT,             "max_sect=%d"           },
2380         { SRP_OPT_MAX_CMD_PER_LUN,      "max_cmd_per_lun=%d"    },
2381         { SRP_OPT_IO_CLASS,             "io_class=%x"           },
2382         { SRP_OPT_INITIATOR_EXT,        "initiator_ext=%s"      },
2383         { SRP_OPT_CMD_SG_ENTRIES,       "cmd_sg_entries=%u"     },
2384         { SRP_OPT_ALLOW_EXT_SG,         "allow_ext_sg=%u"       },
2385         { SRP_OPT_SG_TABLESIZE,         "sg_tablesize=%u"       },
2386         { SRP_OPT_COMP_VECTOR,          "comp_vector=%u"        },
2387         { SRP_OPT_TL_RETRY_COUNT,       "tl_retry_count=%u"     },
2388         { SRP_OPT_QUEUE_SIZE,           "queue_size=%d"         },
2389         { SRP_OPT_ERR,                  NULL                    }
2390 };
2391
2392 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2393 {
2394         char *options, *sep_opt;
2395         char *p;
2396         char dgid[3];
2397         substring_t args[MAX_OPT_ARGS];
2398         int opt_mask = 0;
2399         int token;
2400         int ret = -EINVAL;
2401         int i;
2402
2403         options = kstrdup(buf, GFP_KERNEL);
2404         if (!options)
2405                 return -ENOMEM;
2406
2407         sep_opt = options;
2408         while ((p = strsep(&sep_opt, ",")) != NULL) {
2409                 if (!*p)
2410                         continue;
2411
2412                 token = match_token(p, srp_opt_tokens, args);
2413                 opt_mask |= token;
2414
2415                 switch (token) {
2416                 case SRP_OPT_ID_EXT:
2417                         p = match_strdup(args);
2418                         if (!p) {
2419                                 ret = -ENOMEM;
2420                                 goto out;
2421                         }
2422                         target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2423                         kfree(p);
2424                         break;
2425
2426                 case SRP_OPT_IOC_GUID:
2427                         p = match_strdup(args);
2428                         if (!p) {
2429                                 ret = -ENOMEM;
2430                                 goto out;
2431                         }
2432                         target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2433                         kfree(p);
2434                         break;
2435
2436                 case SRP_OPT_DGID:
2437                         p = match_strdup(args);
2438                         if (!p) {
2439                                 ret = -ENOMEM;
2440                                 goto out;
2441                         }
2442                         if (strlen(p) != 32) {
2443                                 pr_warn("bad dest GID parameter '%s'\n", p);
2444                                 kfree(p);
2445                                 goto out;
2446                         }
2447
2448                         for (i = 0; i < 16; ++i) {
2449                                 strlcpy(dgid, p + i * 2, 3);
2450                                 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2451                         }
2452                         kfree(p);
2453                         memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2454                         break;
2455
2456                 case SRP_OPT_PKEY:
2457                         if (match_hex(args, &token)) {
2458                                 pr_warn("bad P_Key parameter '%s'\n", p);
2459                                 goto out;
2460                         }
2461                         target->path.pkey = cpu_to_be16(token);
2462                         break;
2463
2464                 case SRP_OPT_SERVICE_ID:
2465                         p = match_strdup(args);
2466                         if (!p) {
2467                                 ret = -ENOMEM;
2468                                 goto out;
2469                         }
2470                         target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2471                         target->path.service_id = target->service_id;
2472                         kfree(p);
2473                         break;
2474
2475                 case SRP_OPT_MAX_SECT:
2476                         if (match_int(args, &token)) {
2477                                 pr_warn("bad max sect parameter '%s'\n", p);
2478                                 goto out;
2479                         }
2480                         target->scsi_host->max_sectors = token;
2481                         break;
2482
2483                 case SRP_OPT_QUEUE_SIZE:
2484                         if (match_int(args, &token) || token < 1) {
2485                                 pr_warn("bad queue_size parameter '%s'\n", p);
2486                                 goto out;
2487                         }
2488                         target->scsi_host->can_queue = token;
2489                         target->queue_size = token + SRP_RSP_SQ_SIZE +
2490                                              SRP_TSK_MGMT_SQ_SIZE;
2491                         if (!(opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
2492                                 target->scsi_host->cmd_per_lun = token;
2493                         break;
2494
2495                 case SRP_OPT_MAX_CMD_PER_LUN:
2496                         if (match_int(args, &token) || token < 1) {
2497                                 pr_warn("bad max cmd_per_lun parameter '%s'\n",
2498                                         p);
2499                                 goto out;
2500                         }
2501                         target->scsi_host->cmd_per_lun = token;
2502                         break;
2503
2504                 case SRP_OPT_IO_CLASS:
2505                         if (match_hex(args, &token)) {
2506                                 pr_warn("bad IO class parameter '%s'\n", p);
2507                                 goto out;
2508                         }
2509                         if (token != SRP_REV10_IB_IO_CLASS &&
2510                             token != SRP_REV16A_IB_IO_CLASS) {
2511                                 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2512                                         token, SRP_REV10_IB_IO_CLASS,
2513                                         SRP_REV16A_IB_IO_CLASS);
2514                                 goto out;
2515                         }
2516                         target->io_class = token;
2517                         break;
2518
2519                 case SRP_OPT_INITIATOR_EXT:
2520                         p = match_strdup(args);
2521                         if (!p) {
2522                                 ret = -ENOMEM;
2523                                 goto out;
2524                         }
2525                         target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2526                         kfree(p);
2527                         break;
2528
2529                 case SRP_OPT_CMD_SG_ENTRIES:
2530                         if (match_int(args, &token) || token < 1 || token > 255) {
2531                                 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2532                                         p);
2533                                 goto out;
2534                         }
2535                         target->cmd_sg_cnt = token;
2536                         break;
2537
2538                 case SRP_OPT_ALLOW_EXT_SG:
2539                         if (match_int(args, &token)) {
2540                                 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2541                                 goto out;
2542                         }
2543                         target->allow_ext_sg = !!token;
2544                         break;
2545
2546                 case SRP_OPT_SG_TABLESIZE:
2547                         if (match_int(args, &token) || token < 1 ||
2548                                         token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2549                                 pr_warn("bad max sg_tablesize parameter '%s'\n",
2550                                         p);
2551                                 goto out;
2552                         }
2553                         target->sg_tablesize = token;
2554                         break;
2555
2556                 case SRP_OPT_COMP_VECTOR:
2557                         if (match_int(args, &token) || token < 0) {
2558                                 pr_warn("bad comp_vector parameter '%s'\n", p);
2559                                 goto out;
2560                         }
2561                         target->comp_vector = token;
2562                         break;
2563
2564                 case SRP_OPT_TL_RETRY_COUNT:
2565                         if (match_int(args, &token) || token < 2 || token > 7) {
2566                                 pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
2567                                         p);
2568                                 goto out;
2569                         }
2570                         target->tl_retry_count = token;
2571                         break;
2572
2573                 default:
2574                         pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2575                                 p);
2576                         goto out;
2577                 }
2578         }
2579
2580         if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2581                 ret = 0;
2582         else
2583                 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2584                         if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2585                             !(srp_opt_tokens[i].token & opt_mask))
2586                                 pr_warn("target creation request is missing parameter '%s'\n",
2587                                         srp_opt_tokens[i].pattern);
2588
2589         if (target->scsi_host->cmd_per_lun > target->scsi_host->can_queue
2590             && (opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
2591                 pr_warn("cmd_per_lun = %d > queue_size = %d\n",
2592                         target->scsi_host->cmd_per_lun,
2593                         target->scsi_host->can_queue);
2594
2595 out:
2596         kfree(options);
2597         return ret;
2598 }
2599
2600 static ssize_t srp_create_target(struct device *dev,
2601                                  struct device_attribute *attr,
2602                                  const char *buf, size_t count)
2603 {
2604         struct srp_host *host =
2605                 container_of(dev, struct srp_host, dev);
2606         struct Scsi_Host *target_host;
2607         struct srp_target_port *target;
2608         struct ib_device *ibdev = host->srp_dev->dev;
2609         int ret;
2610
2611         target_host = scsi_host_alloc(&srp_template,
2612                                       sizeof (struct srp_target_port));
2613         if (!target_host)
2614                 return -ENOMEM;
2615
2616         target_host->transportt  = ib_srp_transport_template;
2617         target_host->max_channel = 0;
2618         target_host->max_id      = 1;
2619         target_host->max_lun     = SRP_MAX_LUN;
2620         target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2621
2622         target = host_to_target(target_host);
2623
2624         target->io_class        = SRP_REV16A_IB_IO_CLASS;
2625         target->scsi_host       = target_host;
2626         target->srp_host        = host;
2627         target->lkey            = host->srp_dev->mr->lkey;
2628         target->rkey            = host->srp_dev->mr->rkey;
2629         target->cmd_sg_cnt      = cmd_sg_entries;
2630         target->sg_tablesize    = indirect_sg_entries ? : cmd_sg_entries;
2631         target->allow_ext_sg    = allow_ext_sg;
2632         target->tl_retry_count  = 7;
2633         target->queue_size      = SRP_DEFAULT_QUEUE_SIZE;
2634
2635         mutex_lock(&host->add_target_mutex);
2636
2637         ret = srp_parse_options(buf, target);
2638         if (ret)
2639                 goto err;
2640
2641         target->req_ring_size = target->queue_size - SRP_TSK_MGMT_SQ_SIZE;
2642
2643         if (!srp_conn_unique(target->srp_host, target)) {
2644                 shost_printk(KERN_INFO, target->scsi_host,
2645                              PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
2646                              be64_to_cpu(target->id_ext),
2647                              be64_to_cpu(target->ioc_guid),
2648                              be64_to_cpu(target->initiator_ext));
2649                 ret = -EEXIST;
2650                 goto err;
2651         }
2652
2653         if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2654                                 target->cmd_sg_cnt < target->sg_tablesize) {
2655                 pr_warn("No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2656                 target->sg_tablesize = target->cmd_sg_cnt;
2657         }
2658
2659         target_host->sg_tablesize = target->sg_tablesize;
2660         target->indirect_size = target->sg_tablesize *
2661                                 sizeof (struct srp_direct_buf);
2662         target->max_iu_len = sizeof (struct srp_cmd) +
2663                              sizeof (struct srp_indirect_buf) +
2664                              target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2665
2666         INIT_WORK(&target->tl_err_work, srp_tl_err_work);
2667         INIT_WORK(&target->remove_work, srp_remove_work);
2668         spin_lock_init(&target->lock);
2669         INIT_LIST_HEAD(&target->free_tx);
2670         ret = srp_alloc_req_data(target);
2671         if (ret)
2672                 goto err_free_mem;
2673
2674         ret = ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2675         if (ret)
2676                 goto err_free_mem;
2677
2678         ret = srp_create_target_ib(target);
2679         if (ret)
2680                 goto err_free_mem;
2681
2682         ret = srp_new_cm_id(target);
2683         if (ret)
2684                 goto err_free_ib;
2685
2686         ret = srp_connect_target(target);
2687         if (ret) {
2688                 shost_printk(KERN_ERR, target->scsi_host,
2689                              PFX "Connection failed\n");
2690                 goto err_cm_id;
2691         }
2692
2693         ret = srp_add_target(host, target);
2694         if (ret)
2695                 goto err_disconnect;
2696
2697         shost_printk(KERN_DEBUG, target->scsi_host, PFX
2698                      "new target: id_ext %016llx ioc_guid %016llx pkey %04x service_id %016llx sgid %pI6 dgid %pI6\n",
2699                      be64_to_cpu(target->id_ext),
2700                      be64_to_cpu(target->ioc_guid),
2701                      be16_to_cpu(target->path.pkey),
2702                      be64_to_cpu(target->service_id),
2703                      target->path.sgid.raw, target->path.dgid.raw);
2704
2705         ret = count;
2706
2707 out:
2708         mutex_unlock(&host->add_target_mutex);
2709         return ret;
2710
2711 err_disconnect:
2712         srp_disconnect_target(target);
2713
2714 err_cm_id:
2715         ib_destroy_cm_id(target->cm_id);
2716
2717 err_free_ib:
2718         srp_free_target_ib(target);
2719
2720 err_free_mem:
2721         srp_free_req_data(target);
2722
2723 err:
2724         scsi_host_put(target_host);
2725         goto out;
2726 }
2727
2728 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2729
2730 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2731                           char *buf)
2732 {
2733         struct srp_host *host = container_of(dev, struct srp_host, dev);
2734
2735         return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2736 }
2737
2738 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2739
2740 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2741                          char *buf)
2742 {
2743         struct srp_host *host = container_of(dev, struct srp_host, dev);
2744
2745         return sprintf(buf, "%d\n", host->port);
2746 }
2747
2748 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2749
2750 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2751 {
2752         struct srp_host *host;
2753
2754         host = kzalloc(sizeof *host, GFP_KERNEL);
2755         if (!host)
2756                 return NULL;
2757
2758         INIT_LIST_HEAD(&host->target_list);
2759         spin_lock_init(&host->target_lock);
2760         init_completion(&host->released);
2761         mutex_init(&host->add_target_mutex);
2762         host->srp_dev = device;
2763         host->port = port;
2764
2765         host->dev.class = &srp_class;
2766         host->dev.parent = device->dev->dma_device;
2767         dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2768
2769         if (device_register(&host->dev))
2770                 goto free_host;
2771         if (device_create_file(&host->dev, &dev_attr_add_target))
2772                 goto err_class;
2773         if (device_create_file(&host->dev, &dev_attr_ibdev))
2774                 goto err_class;
2775         if (device_create_file(&host->dev, &dev_attr_port))
2776                 goto err_class;
2777
2778         return host;
2779
2780 err_class:
2781         device_unregister(&host->dev);
2782
2783 free_host:
2784         kfree(host);
2785
2786         return NULL;
2787 }
2788
2789 static void srp_add_one(struct ib_device *device)
2790 {
2791         struct srp_device *srp_dev;
2792         struct ib_device_attr *dev_attr;
2793         struct ib_fmr_pool_param fmr_param;
2794         struct srp_host *host;
2795         int max_pages_per_fmr, fmr_page_shift, s, e, p;
2796
2797         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2798         if (!dev_attr)
2799                 return;
2800
2801         if (ib_query_device(device, dev_attr)) {
2802                 pr_warn("Query device failed for %s\n", device->name);
2803                 goto free_attr;
2804         }
2805
2806         srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2807         if (!srp_dev)
2808                 goto free_attr;
2809
2810         /*
2811          * Use the smallest page size supported by the HCA, down to a
2812          * minimum of 4096 bytes. We're unlikely to build large sglists
2813          * out of smaller entries.
2814          */
2815         fmr_page_shift          = max(12, ffs(dev_attr->page_size_cap) - 1);
2816         srp_dev->fmr_page_size  = 1 << fmr_page_shift;
2817         srp_dev->fmr_page_mask  = ~((u64) srp_dev->fmr_page_size - 1);
2818         srp_dev->fmr_max_size   = srp_dev->fmr_page_size * SRP_FMR_SIZE;
2819
2820         INIT_LIST_HEAD(&srp_dev->dev_list);
2821
2822         srp_dev->dev = device;
2823         srp_dev->pd  = ib_alloc_pd(device);
2824         if (IS_ERR(srp_dev->pd))
2825                 goto free_dev;
2826
2827         srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2828                                     IB_ACCESS_LOCAL_WRITE |
2829                                     IB_ACCESS_REMOTE_READ |
2830                                     IB_ACCESS_REMOTE_WRITE);
2831         if (IS_ERR(srp_dev->mr))
2832                 goto err_pd;
2833
2834         for (max_pages_per_fmr = SRP_FMR_SIZE;
2835                         max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2836                         max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2837                 memset(&fmr_param, 0, sizeof fmr_param);
2838                 fmr_param.pool_size         = SRP_FMR_POOL_SIZE;
2839                 fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
2840                 fmr_param.cache             = 1;
2841                 fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2842                 fmr_param.page_shift        = fmr_page_shift;
2843                 fmr_param.access            = (IB_ACCESS_LOCAL_WRITE |
2844                                                IB_ACCESS_REMOTE_WRITE |
2845                                                IB_ACCESS_REMOTE_READ);
2846
2847                 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2848                 if (!IS_ERR(srp_dev->fmr_pool))
2849                         break;
2850         }
2851
2852         if (IS_ERR(srp_dev->fmr_pool))
2853                 srp_dev->fmr_pool = NULL;
2854
2855         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2856                 s = 0;
2857                 e = 0;
2858         } else {
2859                 s = 1;
2860                 e = device->phys_port_cnt;
2861         }
2862
2863         for (p = s; p <= e; ++p) {
2864                 host = srp_add_port(srp_dev, p);
2865                 if (host)
2866                         list_add_tail(&host->list, &srp_dev->dev_list);
2867         }
2868
2869         ib_set_client_data(device, &srp_client, srp_dev);
2870
2871         goto free_attr;
2872
2873 err_pd:
2874         ib_dealloc_pd(srp_dev->pd);
2875
2876 free_dev:
2877         kfree(srp_dev);
2878
2879 free_attr:
2880         kfree(dev_attr);
2881 }
2882
2883 static void srp_remove_one(struct ib_device *device)
2884 {
2885         struct srp_device *srp_dev;
2886         struct srp_host *host, *tmp_host;
2887         struct srp_target_port *target;
2888
2889         srp_dev = ib_get_client_data(device, &srp_client);
2890         if (!srp_dev)
2891                 return;
2892
2893         list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2894                 device_unregister(&host->dev);
2895                 /*
2896                  * Wait for the sysfs entry to go away, so that no new
2897                  * target ports can be created.
2898                  */
2899                 wait_for_completion(&host->released);
2900
2901                 /*
2902                  * Remove all target ports.
2903                  */
2904                 spin_lock(&host->target_lock);
2905                 list_for_each_entry(target, &host->target_list, list)
2906                         srp_queue_remove_work(target);
2907                 spin_unlock(&host->target_lock);
2908
2909                 /*
2910                  * Wait for target port removal tasks.
2911                  */
2912                 flush_workqueue(system_long_wq);
2913
2914                 kfree(host);
2915         }
2916
2917         if (srp_dev->fmr_pool)
2918                 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2919         ib_dereg_mr(srp_dev->mr);
2920         ib_dealloc_pd(srp_dev->pd);
2921
2922         kfree(srp_dev);
2923 }
2924
2925 static struct srp_function_template ib_srp_transport_functions = {
2926         .has_rport_state         = true,
2927         .reset_timer_if_blocked  = true,
2928         .reconnect_delay         = &srp_reconnect_delay,
2929         .fast_io_fail_tmo        = &srp_fast_io_fail_tmo,
2930         .dev_loss_tmo            = &srp_dev_loss_tmo,
2931         .reconnect               = srp_rport_reconnect,
2932         .rport_delete            = srp_rport_delete,
2933         .terminate_rport_io      = srp_terminate_io,
2934 };
2935
2936 static int __init srp_init_module(void)
2937 {
2938         int ret;
2939
2940         BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2941
2942         if (srp_sg_tablesize) {
2943                 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2944                 if (!cmd_sg_entries)
2945                         cmd_sg_entries = srp_sg_tablesize;
2946         }
2947
2948         if (!cmd_sg_entries)
2949                 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2950
2951         if (cmd_sg_entries > 255) {
2952                 pr_warn("Clamping cmd_sg_entries to 255\n");
2953                 cmd_sg_entries = 255;
2954         }
2955
2956         if (!indirect_sg_entries)
2957                 indirect_sg_entries = cmd_sg_entries;
2958         else if (indirect_sg_entries < cmd_sg_entries) {
2959                 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
2960                         cmd_sg_entries);
2961                 indirect_sg_entries = cmd_sg_entries;
2962         }
2963
2964         ib_srp_transport_template =
2965                 srp_attach_transport(&ib_srp_transport_functions);
2966         if (!ib_srp_transport_template)
2967                 return -ENOMEM;
2968
2969         ret = class_register(&srp_class);
2970         if (ret) {
2971                 pr_err("couldn't register class infiniband_srp\n");
2972                 srp_release_transport(ib_srp_transport_template);
2973                 return ret;
2974         }
2975
2976         ib_sa_register_client(&srp_sa_client);
2977
2978         ret = ib_register_client(&srp_client);
2979         if (ret) {
2980                 pr_err("couldn't register IB client\n");
2981                 srp_release_transport(ib_srp_transport_template);
2982                 ib_sa_unregister_client(&srp_sa_client);
2983                 class_unregister(&srp_class);
2984                 return ret;
2985         }
2986
2987         return 0;
2988 }
2989
2990 static void __exit srp_cleanup_module(void)
2991 {
2992         ib_unregister_client(&srp_client);
2993         ib_sa_unregister_client(&srp_sa_client);
2994         class_unregister(&srp_class);
2995         srp_release_transport(ib_srp_transport_template);
2996 }
2997
2998 module_init(srp_init_module);
2999 module_exit(srp_cleanup_module);