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