drivers/infiniband/ulp/srp/ib_srp.c

   1 /*
   2  * Copyright (c) 2005 Cisco Systems.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #define pr_fmt(fmt) PFX fmt
  34
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/slab.h>
  38 #include <linux/err.h>
  39 #include <linux/string.h>
  40 #include <linux/parser.h>
  41 #include <linux/random.h>
  42 #include <linux/jiffies.h>
  43
  44 #include <linux/atomic.h>
  45
  46 #include <scsi/scsi.h>
  47 #include <scsi/scsi_device.h>
  48 #include <scsi/scsi_dbg.h>
  49 #include <scsi/srp.h>
  50 #include <scsi/scsi_transport_srp.h>
  51
  52 #include "ib_srp.h"
  53
  54 #define DRV_NAME        "ib_srp"
  55 #define PFX             DRV_NAME ": "
  56 #define DRV_VERSION     "1.0"
  57 #define DRV_RELDATE     "July 1, 2013"
  58
  59 MODULE_AUTHOR("Roland Dreier");
  60 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
  61                    "v" DRV_VERSION " (" DRV_RELDATE ")");
  62 MODULE_LICENSE("Dual BSD/GPL");
  63
  64 static unsigned int srp_sg_tablesize;
  65 static unsigned int cmd_sg_entries;
  66 static unsigned int indirect_sg_entries;
  67 static bool allow_ext_sg;
  68 static int topspin_workarounds = 1;
  69
  70 module_param(srp_sg_tablesize, uint, 0444);
  71 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
  72
  73 module_param(cmd_sg_entries, uint, 0444);
  74 MODULE_PARM_DESC(cmd_sg_entries,
  75                  "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
  76
  77 module_param(indirect_sg_entries, uint, 0444);
  78 MODULE_PARM_DESC(indirect_sg_entries,
  79                  "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
  80
  81 module_param(allow_ext_sg, bool, 0444);
  82 MODULE_PARM_DESC(allow_ext_sg,
  83                   "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
  84
  85 module_param(topspin_workarounds, int, 0444);
  86 MODULE_PARM_DESC(topspin_workarounds,
  87                  "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
  88
  89 static void srp_add_one(struct ib_device *device);
  90 static void srp_remove_one(struct ib_device *device);
  91 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
  92 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
  93 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
  94
  95 static struct scsi_transport_template *ib_srp_transport_template;
  96
  97 static struct ib_client srp_client = {
  98         .name   = "srp",
  99         .add    = srp_add_one,
 100         .remove = srp_remove_one
 101 };
 102
 103 static struct ib_sa_client srp_sa_client;
 104
 105 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
 106 {
 107         return (struct srp_target_port *) host->hostdata;
 108 }
 109
 110 static const char *srp_target_info(struct Scsi_Host *host)
 111 {
 112         return host_to_target(host)->target_name;
 113 }
 114
 115 static int srp_target_is_topspin(struct srp_target_port *target)
 116 {
 117         static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
 118         static const u8 cisco_oui[3]   = { 0x00, 0x1b, 0x0d };
 119
 120         return topspin_workarounds &&
 121                 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
 122                  !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
 123 }
 124
 125 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
 126                                    gfp_t gfp_mask,
 127                                    enum dma_data_direction direction)
 128 {
 129         struct srp_iu *iu;
 130
 131         iu = kmalloc(sizeof *iu, gfp_mask);
 132         if (!iu)
 133                 goto out;
 134
 135         iu->buf = kzalloc(size, gfp_mask);
 136         if (!iu->buf)
 137                 goto out_free_iu;
 138
 139         iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
 140                                     direction);
 141         if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
 142                 goto out_free_buf;
 143
 144         iu->size      = size;
 145         iu->direction = direction;
 146
 147         return iu;
 148
 149 out_free_buf:
 150         kfree(iu->buf);
 151 out_free_iu:
 152         kfree(iu);
 153 out:
 154         return NULL;
 155 }
 156
 157 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
 158 {
 159         if (!iu)
 160                 return;
 161
 162         ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
 163                             iu->direction);
 164         kfree(iu->buf);
 165         kfree(iu);
 166 }
 167
 168 static void srp_qp_event(struct ib_event *event, void *context)
 169 {
 170         pr_debug("QP event %d\n", event->event);
 171 }
 172
 173 static int srp_init_qp(struct srp_target_port *target,
 174                        struct ib_qp *qp)
 175 {
 176         struct ib_qp_attr *attr;
 177         int ret;
 178
 179         attr = kmalloc(sizeof *attr, GFP_KERNEL);
 180         if (!attr)
 181                 return -ENOMEM;
 182
 183         ret = ib_find_pkey(target->srp_host->srp_dev->dev,
 184                            target->srp_host->port,
 185                            be16_to_cpu(target->path.pkey),
 186                            &attr->pkey_index);
 187         if (ret)
 188                 goto out;
 189
 190         attr->qp_state        = IB_QPS_INIT;
 191         attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
 192                                     IB_ACCESS_REMOTE_WRITE);
 193         attr->port_num        = target->srp_host->port;
 194
 195         ret = ib_modify_qp(qp, attr,
 196                            IB_QP_STATE          |
 197                            IB_QP_PKEY_INDEX     |
 198                            IB_QP_ACCESS_FLAGS   |
 199                            IB_QP_PORT);
 200
 201 out:
 202         kfree(attr);
 203         return ret;
 204 }
 205
 206 static int srp_new_cm_id(struct srp_target_port *target)
 207 {
 208         struct ib_cm_id *new_cm_id;
 209
 210         new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
 211                                     srp_cm_handler, target);
 212         if (IS_ERR(new_cm_id))
 213                 return PTR_ERR(new_cm_id);
 214
 215         if (target->cm_id)
 216                 ib_destroy_cm_id(target->cm_id);
 217         target->cm_id = new_cm_id;
 218
 219         return 0;
 220 }
 221
 222 static int srp_create_target_ib(struct srp_target_port *target)
 223 {
 224         struct ib_qp_init_attr *init_attr;
 225         struct ib_cq *recv_cq, *send_cq;
 226         struct ib_qp *qp;
 227         int ret;
 228
 229         init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
 230         if (!init_attr)
 231                 return -ENOMEM;
 232
 233         recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
 234                                srp_recv_completion, NULL, target, SRP_RQ_SIZE,
 235                                target->comp_vector);
 236         if (IS_ERR(recv_cq)) {
 237                 ret = PTR_ERR(recv_cq);
 238                 goto err;
 239         }
 240
 241         send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
 242                                srp_send_completion, NULL, target, SRP_SQ_SIZE,
 243                                target->comp_vector);
 244         if (IS_ERR(send_cq)) {
 245                 ret = PTR_ERR(send_cq);
 246                 goto err_recv_cq;
 247         }
 248
 249         ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
 250
 251         init_attr->event_handler       = srp_qp_event;
 252         init_attr->cap.max_send_wr     = SRP_SQ_SIZE;
 253         init_attr->cap.max_recv_wr     = SRP_RQ_SIZE;
 254         init_attr->cap.max_recv_sge    = 1;
 255         init_attr->cap.max_send_sge    = 1;
 256         init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
 257         init_attr->qp_type             = IB_QPT_RC;
 258         init_attr->send_cq             = send_cq;
 259         init_attr->recv_cq             = recv_cq;
 260
 261         qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
 262         if (IS_ERR(qp)) {
 263                 ret = PTR_ERR(qp);
 264                 goto err_send_cq;
 265         }
 266
 267         ret = srp_init_qp(target, qp);
 268         if (ret)
 269                 goto err_qp;
 270
 271         if (target->qp)
 272                 ib_destroy_qp(target->qp);
 273         if (target->recv_cq)
 274                 ib_destroy_cq(target->recv_cq);
 275         if (target->send_cq)
 276                 ib_destroy_cq(target->send_cq);
 277
 278         target->qp = qp;
 279         target->recv_cq = recv_cq;
 280         target->send_cq = send_cq;
 281
 282         kfree(init_attr);
 283         return 0;
 284
 285 err_qp:
 286         ib_destroy_qp(qp);
 287
 288 err_send_cq:
 289         ib_destroy_cq(send_cq);
 290
 291 err_recv_cq:
 292         ib_destroy_cq(recv_cq);
 293
 294 err:
 295         kfree(init_attr);
 296         return ret;
 297 }
 298
 299 static void srp_free_target_ib(struct srp_target_port *target)
 300 {
 301         int i;
 302
 303         ib_destroy_qp(target->qp);
 304         ib_destroy_cq(target->send_cq);
 305         ib_destroy_cq(target->recv_cq);
 306
 307         target->qp = NULL;
 308         target->send_cq = target->recv_cq = NULL;
 309
 310         for (i = 0; i < SRP_RQ_SIZE; ++i)
 311                 srp_free_iu(target->srp_host, target->rx_ring[i]);
 312         for (i = 0; i < SRP_SQ_SIZE; ++i)
 313                 srp_free_iu(target->srp_host, target->tx_ring[i]);
 314 }
 315
 316 static void srp_path_rec_completion(int status,
 317                                     struct ib_sa_path_rec *pathrec,
 318                                     void *target_ptr)
 319 {
 320         struct srp_target_port *target = target_ptr;
 321
 322         target->status = status;
 323         if (status)
 324                 shost_printk(KERN_ERR, target->scsi_host,
 325                              PFX "Got failed path rec status %d\n", status);
 326         else
 327                 target->path = *pathrec;
 328         complete(&target->done);
 329 }
 330
 331 static int srp_lookup_path(struct srp_target_port *target)
 332 {
 333         target->path.numb_path = 1;
 334
 335         init_completion(&target->done);
 336
 337         target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
 338                                                    target->srp_host->srp_dev->dev,
 339                                                    target->srp_host->port,
 340                                                    &target->path,
 341                                                    IB_SA_PATH_REC_SERVICE_ID    |
 342                                                    IB_SA_PATH_REC_DGID          |
 343                                                    IB_SA_PATH_REC_SGID          |
 344                                                    IB_SA_PATH_REC_NUMB_PATH     |
 345                                                    IB_SA_PATH_REC_PKEY,
 346                                                    SRP_PATH_REC_TIMEOUT_MS,
 347                                                    GFP_KERNEL,
 348                                                    srp_path_rec_completion,
 349                                                    target, &target->path_query);
 350         if (target->path_query_id < 0)
 351                 return target->path_query_id;
 352
 353         wait_for_completion(&target->done);
 354
 355         if (target->status < 0)
 356                 shost_printk(KERN_WARNING, target->scsi_host,
 357                              PFX "Path record query failed\n");
 358
 359         return target->status;
 360 }
 361
 362 static int srp_send_req(struct srp_target_port *target)
 363 {
 364         struct {
 365                 struct ib_cm_req_param param;
 366                 struct srp_login_req   priv;
 367         } *req = NULL;
 368         int status;
 369
 370         req = kzalloc(sizeof *req, GFP_KERNEL);
 371         if (!req)
 372                 return -ENOMEM;
 373
 374         req->param.primary_path               = &target->path;
 375         req->param.alternate_path             = NULL;
 376         req->param.service_id                 = target->service_id;
 377         req->param.qp_num                     = target->qp->qp_num;
 378         req->param.qp_type                    = target->qp->qp_type;
 379         req->param.private_data               = &req->priv;
 380         req->param.private_data_len           = sizeof req->priv;
 381         req->param.flow_control               = 1;
 382
 383         get_random_bytes(&req->param.starting_psn, 4);
 384         req->param.starting_psn              &= 0xffffff;
 385
 386         /*
 387          * Pick some arbitrary defaults here; we could make these
 388          * module parameters if anyone cared about setting them.
 389          */
 390         req->param.responder_resources        = 4;
 391         req->param.remote_cm_response_timeout = 20;
 392         req->param.local_cm_response_timeout  = 20;
 393         req->param.retry_count                = target->tl_retry_count;
 394         req->param.rnr_retry_count            = 7;
 395         req->param.max_cm_retries             = 15;
 396
 397         req->priv.opcode        = SRP_LOGIN_REQ;
 398         req->priv.tag           = 0;
 399         req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
 400         req->priv.req_buf_fmt   = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
 401                                               SRP_BUF_FORMAT_INDIRECT);
 402         /*
 403          * In the published SRP specification (draft rev. 16a), the
 404          * port identifier format is 8 bytes of ID extension followed
 405          * by 8 bytes of GUID.  Older drafts put the two halves in the
 406          * opposite order, so that the GUID comes first.
 407          *
 408          * Targets conforming to these obsolete drafts can be
 409          * recognized by the I/O Class they report.
 410          */
 411         if (target->io_class == SRP_REV10_IB_IO_CLASS) {
 412                 memcpy(req->priv.initiator_port_id,
 413                        &target->path.sgid.global.interface_id, 8);
 414                 memcpy(req->priv.initiator_port_id + 8,
 415                        &target->initiator_ext, 8);
 416                 memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
 417                 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
 418         } else {
 419                 memcpy(req->priv.initiator_port_id,
 420                        &target->initiator_ext, 8);
 421                 memcpy(req->priv.initiator_port_id + 8,
 422                        &target->path.sgid.global.interface_id, 8);
 423                 memcpy(req->priv.target_port_id,     &target->id_ext, 8);
 424                 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
 425         }
 426
 427         /*
 428          * Topspin/Cisco SRP targets will reject our login unless we
 429          * zero out the first 8 bytes of our initiator port ID and set
 430          * the second 8 bytes to the local node GUID.
 431          */
 432         if (srp_target_is_topspin(target)) {
 433                 shost_printk(KERN_DEBUG, target->scsi_host,
 434                              PFX "Topspin/Cisco initiator port ID workaround "
 435                              "activated for target GUID %016llx\n",
 436                              (unsigned long long) be64_to_cpu(target->ioc_guid));
 437                 memset(req->priv.initiator_port_id, 0, 8);
 438                 memcpy(req->priv.initiator_port_id + 8,
 439                        &target->srp_host->srp_dev->dev->node_guid, 8);
 440         }
 441
 442         status = ib_send_cm_req(target->cm_id, &req->param);
 443
 444         kfree(req);
 445
 446         return status;
 447 }
 448
 449 static bool srp_queue_remove_work(struct srp_target_port *target)
 450 {
 451         bool changed = false;
 452
 453         spin_lock_irq(&target->lock);
 454         if (target->state != SRP_TARGET_REMOVED) {
 455                 target->state = SRP_TARGET_REMOVED;
 456                 changed = true;
 457         }
 458         spin_unlock_irq(&target->lock);
 459
 460         if (changed)
 461                 queue_work(system_long_wq, &target->remove_work);
 462
 463         return changed;
 464 }
 465
 466 static bool srp_change_conn_state(struct srp_target_port *target,
 467                                   bool connected)
 468 {
 469         bool changed = false;
 470
 471         spin_lock_irq(&target->lock);
 472         if (target->connected != connected) {
 473                 target->connected = connected;
 474                 changed = true;
 475         }
 476         spin_unlock_irq(&target->lock);
 477
 478         return changed;
 479 }
 480
 481 static void srp_disconnect_target(struct srp_target_port *target)
 482 {
 483         if (srp_change_conn_state(target, false)) {
 484                 /* XXX should send SRP_I_LOGOUT request */
 485
 486                 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
 487                         shost_printk(KERN_DEBUG, target->scsi_host,
 488                                      PFX "Sending CM DREQ failed\n");
 489                 }
 490         }
 491 }
 492
 493 static void srp_free_req_data(struct srp_target_port *target)
 494 {
 495         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
 496         struct srp_request *req;
 497         int i;
 498
 499         for (i = 0, req = target->req_ring; i < SRP_CMD_SQ_SIZE; ++i, ++req) {
 500                 kfree(req->fmr_list);
 501                 kfree(req->map_page);
 502                 if (req->indirect_dma_addr) {
 503                         ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
 504                                             target->indirect_size,
 505                                             DMA_TO_DEVICE);
 506                 }
 507                 kfree(req->indirect_desc);
 508         }
 509 }
 510
 511 /**
 512  * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
 513  * @shost: SCSI host whose attributes to remove from sysfs.
 514  *
 515  * Note: Any attributes defined in the host template and that did not exist
 516  * before invocation of this function will be ignored.
 517  */
 518 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
 519 {
 520         struct device_attribute **attr;
 521
 522         for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
 523                 device_remove_file(&shost->shost_dev, *attr);
 524 }
 525
 526 static void srp_remove_target(struct srp_target_port *target)
 527 {
 528         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 529
 530         srp_del_scsi_host_attr(target->scsi_host);
 531         srp_rport_get(target->rport);
 532         srp_remove_host(target->scsi_host);
 533         scsi_remove_host(target->scsi_host);
 534         srp_disconnect_target(target);
 535         ib_destroy_cm_id(target->cm_id);
 536         srp_free_target_ib(target);
 537         srp_rport_put(target->rport);
 538         srp_free_req_data(target);
 539         scsi_host_put(target->scsi_host);
 540 }
 541
 542 static void srp_remove_work(struct work_struct *work)
 543 {
 544         struct srp_target_port *target =
 545                 container_of(work, struct srp_target_port, remove_work);
 546
 547         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 548
 549         srp_remove_target(target);
 550
 551         spin_lock(&target->srp_host->target_lock);
 552         list_del(&target->list);
 553         spin_unlock(&target->srp_host->target_lock);
 554 }
 555
 556 static void srp_rport_delete(struct srp_rport *rport)
 557 {
 558         struct srp_target_port *target = rport->lld_data;
 559
 560         srp_queue_remove_work(target);
 561 }
 562
 563 static int srp_connect_target(struct srp_target_port *target)
 564 {
 565         int retries = 3;
 566         int ret;
 567
 568         WARN_ON_ONCE(target->connected);
 569
 570         target->qp_in_error = false;
 571
 572         ret = srp_lookup_path(target);
 573         if (ret)
 574                 return ret;
 575
 576         while (1) {
 577                 init_completion(&target->done);
 578                 ret = srp_send_req(target);
 579                 if (ret)
 580                         return ret;
 581                 wait_for_completion(&target->done);
 582
 583                 /*
 584                  * The CM event handling code will set status to
 585                  * SRP_PORT_REDIRECT if we get a port redirect REJ
 586                  * back, or SRP_DLID_REDIRECT if we get a lid/qp
 587                  * redirect REJ back.
 588                  */
 589                 switch (target->status) {
 590                 case 0:
 591                         srp_change_conn_state(target, true);
 592                         return 0;
 593
 594                 case SRP_PORT_REDIRECT:
 595                         ret = srp_lookup_path(target);
 596                         if (ret)
 597                                 return ret;
 598                         break;
 599
 600                 case SRP_DLID_REDIRECT:
 601                         break;
 602
 603                 case SRP_STALE_CONN:
 604                         /* Our current CM id was stale, and is now in timewait.
 605                          * Try to reconnect with a new one.
 606                          */
 607                         if (!retries-- || srp_new_cm_id(target)) {
 608                                 shost_printk(KERN_ERR, target->scsi_host, PFX
 609                                              "giving up on stale connection\n");
 610                                 target->status = -ECONNRESET;
 611                                 return target->status;
 612                         }
 613
 614                         shost_printk(KERN_ERR, target->scsi_host, PFX
 615                                      "retrying stale connection\n");
 616                         break;
 617
 618                 default:
 619                         return target->status;
 620                 }
 621         }
 622 }
 623
 624 static void srp_unmap_data(struct scsi_cmnd *scmnd,
 625                            struct srp_target_port *target,
 626                            struct srp_request *req)
 627 {
 628         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
 629         struct ib_pool_fmr **pfmr;
 630
 631         if (!scsi_sglist(scmnd) ||
 632             (scmnd->sc_data_direction != DMA_TO_DEVICE &&
 633              scmnd->sc_data_direction != DMA_FROM_DEVICE))
 634                 return;
 635
 636         pfmr = req->fmr_list;
 637         while (req->nfmr--)
 638                 ib_fmr_pool_unmap(*pfmr++);
 639
 640         ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
 641                         scmnd->sc_data_direction);
 642 }
 643
 644 /**
 645  * srp_claim_req - Take ownership of the scmnd associated with a request.
 646  * @target: SRP target port.
 647  * @req: SRP request.
 648  * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
 649  *         ownership of @req->scmnd if it equals @scmnd.
 650  *
 651  * Return value:
 652  * Either NULL or a pointer to the SCSI command the caller became owner of.
 653  */
 654 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
 655                                        struct srp_request *req,
 656                                        struct scsi_cmnd *scmnd)
 657 {
 658         unsigned long flags;
 659
 660         spin_lock_irqsave(&target->lock, flags);
 661         if (!scmnd) {
 662                 scmnd = req->scmnd;
 663                 req->scmnd = NULL;
 664         } else if (req->scmnd == scmnd) {
 665                 req->scmnd = NULL;
 666         } else {
 667                 scmnd = NULL;
 668         }
 669         spin_unlock_irqrestore(&target->lock, flags);
 670
 671         return scmnd;
 672 }
 673
 674 /**
 675  * srp_free_req() - Unmap data and add request to the free request list.
 676  */
 677 static void srp_free_req(struct srp_target_port *target,
 678                          struct srp_request *req, struct scsi_cmnd *scmnd,
 679                          s32 req_lim_delta)
 680 {
 681         unsigned long flags;
 682
 683         srp_unmap_data(scmnd, target, req);
 684
 685         spin_lock_irqsave(&target->lock, flags);
 686         target->req_lim += req_lim_delta;
 687         list_add_tail(&req->list, &target->free_reqs);
 688         spin_unlock_irqrestore(&target->lock, flags);
 689 }
 690
 691 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
 692 {
 693         struct scsi_cmnd *scmnd = srp_claim_req(target, req, NULL);
 694
 695         if (scmnd) {
 696                 srp_free_req(target, req, scmnd, 0);
 697                 scmnd->result = DID_RESET << 16;
 698                 scmnd->scsi_done(scmnd);
 699         }
 700 }
 701
 702 static int srp_reconnect_target(struct srp_target_port *target)
 703 {
 704         struct Scsi_Host *shost = target->scsi_host;
 705         int i, ret;
 706
 707         scsi_target_block(&shost->shost_gendev);
 708
 709         srp_disconnect_target(target);
 710         /*
 711          * Now get a new local CM ID so that we avoid confusing the target in
 712          * case things are really fouled up. Doing so also ensures that all CM
 713          * callbacks will have finished before a new QP is allocated.
 714          */
 715         ret = srp_new_cm_id(target);
 716         /*
 717          * Whether or not creating a new CM ID succeeded, create a new
 718          * QP. This guarantees that all completion callback function
 719          * invocations have finished before request resetting starts.
 720          */
 721         if (ret == 0)
 722                 ret = srp_create_target_ib(target);
 723         else
 724                 srp_create_target_ib(target);
 725
 726         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
 727                 struct srp_request *req = &target->req_ring[i];
 728                 if (req->scmnd)
 729                         srp_reset_req(target, req);
 730         }
 731
 732         INIT_LIST_HEAD(&target->free_tx);
 733         for (i = 0; i < SRP_SQ_SIZE; ++i)
 734                 list_add(&target->tx_ring[i]->list, &target->free_tx);
 735
 736         if (ret == 0)
 737                 ret = srp_connect_target(target);
 738
 739         scsi_target_unblock(&shost->shost_gendev, ret == 0 ? SDEV_RUNNING :
 740                             SDEV_TRANSPORT_OFFLINE);
 741         target->transport_offline = !!ret;
 742
 743         if (ret)
 744                 goto err;
 745
 746         shost_printk(KERN_INFO, target->scsi_host, PFX "reconnect succeeded\n");
 747
 748         return ret;
 749
 750 err:
 751         shost_printk(KERN_ERR, target->scsi_host,
 752                      PFX "reconnect failed (%d), removing target port.\n", ret);
 753
 754         /*
 755          * We couldn't reconnect, so kill our target port off.
 756          * However, we have to defer the real removal because we
 757          * are in the context of the SCSI error handler now, which
 758          * will deadlock if we call scsi_remove_host().
 759          */
 760         srp_queue_remove_work(target);
 761
 762         return ret;
 763 }
 764
 765 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
 766                          unsigned int dma_len, u32 rkey)
 767 {
 768         struct srp_direct_buf *desc = state->desc;
 769
 770         desc->va = cpu_to_be64(dma_addr);
 771         desc->key = cpu_to_be32(rkey);
 772         desc->len = cpu_to_be32(dma_len);
 773
 774         state->total_len += dma_len;
 775         state->desc++;
 776         state->ndesc++;
 777 }
 778
 779 static int srp_map_finish_fmr(struct srp_map_state *state,
 780                               struct srp_target_port *target)
 781 {
 782         struct srp_device *dev = target->srp_host->srp_dev;
 783         struct ib_pool_fmr *fmr;
 784         u64 io_addr = 0;
 785
 786         if (!state->npages)
 787                 return 0;
 788
 789         if (state->npages == 1) {
 790                 srp_map_desc(state, state->base_dma_addr, state->fmr_len,
 791                              target->rkey);
 792                 state->npages = state->fmr_len = 0;
 793                 return 0;
 794         }
 795
 796         fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
 797                                    state->npages, io_addr);
 798         if (IS_ERR(fmr))
 799                 return PTR_ERR(fmr);
 800
 801         *state->next_fmr++ = fmr;
 802         state->nfmr++;
 803
 804         srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
 805         state->npages = state->fmr_len = 0;
 806         return 0;
 807 }
 808
 809 static void srp_map_update_start(struct srp_map_state *state,
 810                                  struct scatterlist *sg, int sg_index,
 811                                  dma_addr_t dma_addr)
 812 {
 813         state->unmapped_sg = sg;
 814         state->unmapped_index = sg_index;
 815         state->unmapped_addr = dma_addr;
 816 }
 817
 818 static int srp_map_sg_entry(struct srp_map_state *state,
 819                             struct srp_target_port *target,
 820                             struct scatterlist *sg, int sg_index,
 821                             int use_fmr)
 822 {
 823         struct srp_device *dev = target->srp_host->srp_dev;
 824         struct ib_device *ibdev = dev->dev;
 825         dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
 826         unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
 827         unsigned int len;
 828         int ret;
 829
 830         if (!dma_len)
 831                 return 0;
 832
 833         if (use_fmr == SRP_MAP_NO_FMR) {
 834                 /* Once we're in direct map mode for a request, we don't
 835                  * go back to FMR mode, so no need to update anything
 836                  * other than the descriptor.
 837                  */
 838                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
 839                 return 0;
 840         }
 841
 842         /* If we start at an offset into the FMR page, don't merge into
 843          * the current FMR. Finish it out, and use the kernel's MR for this
 844          * sg entry. This is to avoid potential bugs on some SRP targets
 845          * that were never quite defined, but went away when the initiator
 846          * avoided using FMR on such page fragments.
 847          */
 848         if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
 849                 ret = srp_map_finish_fmr(state, target);
 850                 if (ret)
 851                         return ret;
 852
 853                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
 854                 srp_map_update_start(state, NULL, 0, 0);
 855                 return 0;
 856         }
 857
 858         /* If this is the first sg to go into the FMR, save our position.
 859          * We need to know the first unmapped entry, its index, and the
 860          * first unmapped address within that entry to be able to restart
 861          * mapping after an error.
 862          */
 863         if (!state->unmapped_sg)
 864                 srp_map_update_start(state, sg, sg_index, dma_addr);
 865
 866         while (dma_len) {
 867                 if (state->npages == SRP_FMR_SIZE) {
 868                         ret = srp_map_finish_fmr(state, target);
 869                         if (ret)
 870                                 return ret;
 871
 872                         srp_map_update_start(state, sg, sg_index, dma_addr);
 873                 }
 874
 875                 len = min_t(unsigned int, dma_len, dev->fmr_page_size);
 876
 877                 if (!state->npages)
 878                         state->base_dma_addr = dma_addr;
 879                 state->pages[state->npages++] = dma_addr;
 880                 state->fmr_len += len;
 881                 dma_addr += len;
 882                 dma_len -= len;
 883         }
 884
 885         /* If the last entry of the FMR wasn't a full page, then we need to
 886          * close it out and start a new one -- we can only merge at page
 887          * boundries.
 888          */
 889         ret = 0;
 890         if (len != dev->fmr_page_size) {
 891                 ret = srp_map_finish_fmr(state, target);
 892                 if (!ret)
 893                         srp_map_update_start(state, NULL, 0, 0);
 894         }
 895         return ret;
 896 }
 897
 898 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
 899                         struct srp_request *req)
 900 {
 901         struct scatterlist *scat, *sg;
 902         struct srp_cmd *cmd = req->cmd->buf;
 903         int i, len, nents, count, use_fmr;
 904         struct srp_device *dev;
 905         struct ib_device *ibdev;
 906         struct srp_map_state state;
 907         struct srp_indirect_buf *indirect_hdr;
 908         u32 table_len;
 909         u8 fmt;
 910
 911         if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
 912                 return sizeof (struct srp_cmd);
 913
 914         if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
 915             scmnd->sc_data_direction != DMA_TO_DEVICE) {
 916                 shost_printk(KERN_WARNING, target->scsi_host,
 917                              PFX "Unhandled data direction %d\n",
 918                              scmnd->sc_data_direction);
 919                 return -EINVAL;
 920         }
 921
 922         nents = scsi_sg_count(scmnd);
 923         scat  = scsi_sglist(scmnd);
 924
 925         dev = target->srp_host->srp_dev;
 926         ibdev = dev->dev;
 927
 928         count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
 929         if (unlikely(count == 0))
 930                 return -EIO;
 931
 932         fmt = SRP_DATA_DESC_DIRECT;
 933         len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
 934
 935         if (count == 1) {
 936                 /*
 937                  * The midlayer only generated a single gather/scatter
 938                  * entry, or DMA mapping coalesced everything to a
 939                  * single entry.  So a direct descriptor along with
 940                  * the DMA MR suffices.
 941                  */
 942                 struct srp_direct_buf *buf = (void *) cmd->add_data;
 943
 944                 buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
 945                 buf->key = cpu_to_be32(target->rkey);
 946                 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
 947
 948                 req->nfmr = 0;
 949                 goto map_complete;
 950         }
 951
 952         /* We have more than one scatter/gather entry, so build our indirect
 953          * descriptor table, trying to merge as many entries with FMR as we
 954          * can.
 955          */
 956         indirect_hdr = (void *) cmd->add_data;
 957
 958         ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
 959                                    target->indirect_size, DMA_TO_DEVICE);
 960
 961         memset(&state, 0, sizeof(state));
 962         state.desc      = req->indirect_desc;
 963         state.pages     = req->map_page;
 964         state.next_fmr  = req->fmr_list;
 965
 966         use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
 967
 968         for_each_sg(scat, sg, count, i) {
 969                 if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
 970                         /* FMR mapping failed, so backtrack to the first
 971                          * unmapped entry and continue on without using FMR.
 972                          */
 973                         dma_addr_t dma_addr;
 974                         unsigned int dma_len;
 975
 976 backtrack:
 977                         sg = state.unmapped_sg;
 978                         i = state.unmapped_index;
 979
 980                         dma_addr = ib_sg_dma_address(ibdev, sg);
 981                         dma_len = ib_sg_dma_len(ibdev, sg);
 982                         dma_len -= (state.unmapped_addr - dma_addr);
 983                         dma_addr = state.unmapped_addr;
 984                         use_fmr = SRP_MAP_NO_FMR;
 985                         srp_map_desc(&state, dma_addr, dma_len, target->rkey);
 986                 }
 987         }
 988
 989         if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
 990                 goto backtrack;
 991
 992         /* We've mapped the request, now pull as much of the indirect
 993          * descriptor table as we can into the command buffer. If this
 994          * target is not using an external indirect table, we are
 995          * guaranteed to fit into the command, as the SCSI layer won't
 996          * give us more S/G entries than we allow.
 997          */
 998         req->nfmr = state.nfmr;
 999         if (state.ndesc == 1) {
1000                 /* FMR mapping was able to collapse this to one entry,
1001                  * so use a direct descriptor.
1002                  */
1003                 struct srp_direct_buf *buf = (void *) cmd->add_data;
1004
1005                 *buf = req->indirect_desc[0];
1006                 goto map_complete;
1007         }
1008
1009         if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1010                                                 !target->allow_ext_sg)) {
1011                 shost_printk(KERN_ERR, target->scsi_host,
1012                              "Could not fit S/G list into SRP_CMD\n");
1013                 return -EIO;
1014         }
1015
1016         count = min(state.ndesc, target->cmd_sg_cnt);
1017         table_len = state.ndesc * sizeof (struct srp_direct_buf);
1018
1019         fmt = SRP_DATA_DESC_INDIRECT;
1020         len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1021         len += count * sizeof (struct srp_direct_buf);
1022
1023         memcpy(indirect_hdr->desc_list, req->indirect_desc,
1024                count * sizeof (struct srp_direct_buf));
1025
1026         indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1027         indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1028         indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1029         indirect_hdr->len = cpu_to_be32(state.total_len);
1030
1031         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1032                 cmd->data_out_desc_cnt = count;
1033         else
1034                 cmd->data_in_desc_cnt = count;
1035
1036         ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1037                                       DMA_TO_DEVICE);
1038
1039 map_complete:
1040         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1041                 cmd->buf_fmt = fmt << 4;
1042         else
1043                 cmd->buf_fmt = fmt;
1044
1045         return len;
1046 }
1047
1048 /*
1049  * Return an IU and possible credit to the free pool
1050  */
1051 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1052                           enum srp_iu_type iu_type)
1053 {
1054         unsigned long flags;
1055
1056         spin_lock_irqsave(&target->lock, flags);
1057         list_add(&iu->list, &target->free_tx);
1058         if (iu_type != SRP_IU_RSP)
1059                 ++target->req_lim;
1060         spin_unlock_irqrestore(&target->lock, flags);
1061 }
1062
1063 /*
1064  * Must be called with target->lock held to protect req_lim and free_tx.
1065  * If IU is not sent, it must be returned using srp_put_tx_iu().
1066  *
1067  * Note:
1068  * An upper limit for the number of allocated information units for each
1069  * request type is:
1070  * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1071  *   more than Scsi_Host.can_queue requests.
1072  * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1073  * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1074  *   one unanswered SRP request to an initiator.
1075  */
1076 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1077                                       enum srp_iu_type iu_type)
1078 {
1079         s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1080         struct srp_iu *iu;
1081
1082         srp_send_completion(target->send_cq, target);
1083
1084         if (list_empty(&target->free_tx))
1085                 return NULL;
1086
1087         /* Initiator responses to target requests do not consume credits */
1088         if (iu_type != SRP_IU_RSP) {
1089                 if (target->req_lim <= rsv) {
1090                         ++target->zero_req_lim;
1091                         return NULL;
1092                 }
1093
1094                 --target->req_lim;
1095         }
1096
1097         iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1098         list_del(&iu->list);
1099         return iu;
1100 }
1101
1102 static int srp_post_send(struct srp_target_port *target,
1103                          struct srp_iu *iu, int len)
1104 {
1105         struct ib_sge list;
1106         struct ib_send_wr wr, *bad_wr;
1107
1108         list.addr   = iu->dma;
1109         list.length = len;
1110         list.lkey   = target->lkey;
1111
1112         wr.next       = NULL;
1113         wr.wr_id      = (uintptr_t) iu;
1114         wr.sg_list    = &list;
1115         wr.num_sge    = 1;
1116         wr.opcode     = IB_WR_SEND;
1117         wr.send_flags = IB_SEND_SIGNALED;
1118
1119         return ib_post_send(target->qp, &wr, &bad_wr);
1120 }
1121
1122 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1123 {
1124         struct ib_recv_wr wr, *bad_wr;
1125         struct ib_sge list;
1126
1127         list.addr   = iu->dma;
1128         list.length = iu->size;
1129         list.lkey   = target->lkey;
1130
1131         wr.next     = NULL;
1132         wr.wr_id    = (uintptr_t) iu;
1133         wr.sg_list  = &list;
1134         wr.num_sge  = 1;
1135
1136         return ib_post_recv(target->qp, &wr, &bad_wr);
1137 }
1138
1139 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1140 {
1141         struct srp_request *req;
1142         struct scsi_cmnd *scmnd;
1143         unsigned long flags;
1144
1145         if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1146                 spin_lock_irqsave(&target->lock, flags);
1147                 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1148                 spin_unlock_irqrestore(&target->lock, flags);
1149
1150                 target->tsk_mgmt_status = -1;
1151                 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1152                         target->tsk_mgmt_status = rsp->data[3];
1153                 complete(&target->tsk_mgmt_done);
1154         } else {
1155                 req = &target->req_ring[rsp->tag];
1156                 scmnd = srp_claim_req(target, req, NULL);
1157                 if (!scmnd) {
1158                         shost_printk(KERN_ERR, target->scsi_host,
1159                                      "Null scmnd for RSP w/tag %016llx\n",
1160                                      (unsigned long long) rsp->tag);
1161
1162                         spin_lock_irqsave(&target->lock, flags);
1163                         target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1164                         spin_unlock_irqrestore(&target->lock, flags);
1165
1166                         return;
1167                 }
1168                 scmnd->result = rsp->status;
1169
1170                 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1171                         memcpy(scmnd->sense_buffer, rsp->data +
1172                                be32_to_cpu(rsp->resp_data_len),
1173                                min_t(int, be32_to_cpu(rsp->sense_data_len),
1174                                      SCSI_SENSE_BUFFERSIZE));
1175                 }
1176
1177                 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1178                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1179                 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1180                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1181
1182                 srp_free_req(target, req, scmnd,
1183                              be32_to_cpu(rsp->req_lim_delta));
1184
1185                 scmnd->host_scribble = NULL;
1186                 scmnd->scsi_done(scmnd);
1187         }
1188 }
1189
1190 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1191                                void *rsp, int len)
1192 {
1193         struct ib_device *dev = target->srp_host->srp_dev->dev;
1194         unsigned long flags;
1195         struct srp_iu *iu;
1196         int err;
1197
1198         spin_lock_irqsave(&target->lock, flags);
1199         target->req_lim += req_delta;
1200         iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1201         spin_unlock_irqrestore(&target->lock, flags);
1202
1203         if (!iu) {
1204                 shost_printk(KERN_ERR, target->scsi_host, PFX
1205                              "no IU available to send response\n");
1206                 return 1;
1207         }
1208
1209         ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1210         memcpy(iu->buf, rsp, len);
1211         ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1212
1213         err = srp_post_send(target, iu, len);
1214         if (err) {
1215                 shost_printk(KERN_ERR, target->scsi_host, PFX
1216                              "unable to post response: %d\n", err);
1217                 srp_put_tx_iu(target, iu, SRP_IU_RSP);
1218         }
1219
1220         return err;
1221 }
1222
1223 static void srp_process_cred_req(struct srp_target_port *target,
1224                                  struct srp_cred_req *req)
1225 {
1226         struct srp_cred_rsp rsp = {
1227                 .opcode = SRP_CRED_RSP,
1228                 .tag = req->tag,
1229         };
1230         s32 delta = be32_to_cpu(req->req_lim_delta);
1231
1232         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1233                 shost_printk(KERN_ERR, target->scsi_host, PFX
1234                              "problems processing SRP_CRED_REQ\n");
1235 }
1236
1237 static void srp_process_aer_req(struct srp_target_port *target,
1238                                 struct srp_aer_req *req)
1239 {
1240         struct srp_aer_rsp rsp = {
1241                 .opcode = SRP_AER_RSP,
1242                 .tag = req->tag,
1243         };
1244         s32 delta = be32_to_cpu(req->req_lim_delta);
1245
1246         shost_printk(KERN_ERR, target->scsi_host, PFX
1247                      "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1248
1249         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1250                 shost_printk(KERN_ERR, target->scsi_host, PFX
1251                              "problems processing SRP_AER_REQ\n");
1252 }
1253
1254 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1255 {
1256         struct ib_device *dev = target->srp_host->srp_dev->dev;
1257         struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1258         int res;
1259         u8 opcode;
1260
1261         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1262                                    DMA_FROM_DEVICE);
1263
1264         opcode = *(u8 *) iu->buf;
1265
1266         if (0) {
1267                 shost_printk(KERN_ERR, target->scsi_host,
1268                              PFX "recv completion, opcode 0x%02x\n", opcode);
1269                 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1270                                iu->buf, wc->byte_len, true);
1271         }
1272
1273         switch (opcode) {
1274         case SRP_RSP:
1275                 srp_process_rsp(target, iu->buf);
1276                 break;
1277
1278         case SRP_CRED_REQ:
1279                 srp_process_cred_req(target, iu->buf);
1280                 break;
1281
1282         case SRP_AER_REQ:
1283                 srp_process_aer_req(target, iu->buf);
1284                 break;
1285
1286         case SRP_T_LOGOUT:
1287                 /* XXX Handle target logout */
1288                 shost_printk(KERN_WARNING, target->scsi_host,
1289                              PFX "Got target logout request\n");
1290                 break;
1291
1292         default:
1293                 shost_printk(KERN_WARNING, target->scsi_host,
1294                              PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1295                 break;
1296         }
1297
1298         ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1299                                       DMA_FROM_DEVICE);
1300
1301         res = srp_post_recv(target, iu);
1302         if (res != 0)
1303                 shost_printk(KERN_ERR, target->scsi_host,
1304                              PFX "Recv failed with error code %d\n", res);
1305 }
1306
1307 static void srp_handle_qp_err(enum ib_wc_status wc_status,
1308                               enum ib_wc_opcode wc_opcode,
1309                               struct srp_target_port *target)
1310 {
1311         if (target->connected && !target->qp_in_error) {
1312                 shost_printk(KERN_ERR, target->scsi_host,
1313                              PFX "failed %s status %d\n",
1314                              wc_opcode & IB_WC_RECV ? "receive" : "send",
1315                              wc_status);
1316         }
1317         target->qp_in_error = true;
1318 }
1319
1320 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1321 {
1322         struct srp_target_port *target = target_ptr;
1323         struct ib_wc wc;
1324
1325         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1326         while (ib_poll_cq(cq, 1, &wc) > 0) {
1327                 if (likely(wc.status == IB_WC_SUCCESS)) {
1328                         srp_handle_recv(target, &wc);
1329                 } else {
1330                         srp_handle_qp_err(wc.status, wc.opcode, target);
1331                 }
1332         }
1333 }
1334
1335 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1336 {
1337         struct srp_target_port *target = target_ptr;
1338         struct ib_wc wc;
1339         struct srp_iu *iu;
1340
1341         while (ib_poll_cq(cq, 1, &wc) > 0) {
1342                 if (likely(wc.status == IB_WC_SUCCESS)) {
1343                         iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1344                         list_add(&iu->list, &target->free_tx);
1345                 } else {
1346                         srp_handle_qp_err(wc.status, wc.opcode, target);
1347                 }
1348         }
1349 }
1350
1351 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1352 {
1353         struct srp_target_port *target = host_to_target(shost);
1354         struct srp_request *req;
1355         struct srp_iu *iu;
1356         struct srp_cmd *cmd;
1357         struct ib_device *dev;
1358         unsigned long flags;
1359         int len;
1360
1361         if (unlikely(target->transport_offline)) {
1362                 scmnd->result = DID_NO_CONNECT << 16;
1363                 scmnd->scsi_done(scmnd);
1364                 return 0;
1365         }
1366
1367         spin_lock_irqsave(&target->lock, flags);
1368         iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1369         if (!iu)
1370                 goto err_unlock;
1371
1372         req = list_first_entry(&target->free_reqs, struct srp_request, list);
1373         list_del(&req->list);
1374         spin_unlock_irqrestore(&target->lock, flags);
1375
1376         dev = target->srp_host->srp_dev->dev;
1377         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1378                                    DMA_TO_DEVICE);
1379
1380         scmnd->result        = 0;
1381         scmnd->host_scribble = (void *) req;
1382
1383         cmd = iu->buf;
1384         memset(cmd, 0, sizeof *cmd);
1385
1386         cmd->opcode = SRP_CMD;
1387         cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
1388         cmd->tag    = req->index;
1389         memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1390
1391         req->scmnd    = scmnd;
1392         req->cmd      = iu;
1393
1394         len = srp_map_data(scmnd, target, req);
1395         if (len < 0) {
1396                 shost_printk(KERN_ERR, target->scsi_host,
1397                              PFX "Failed to map data\n");
1398                 goto err_iu;
1399         }
1400
1401         ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1402                                       DMA_TO_DEVICE);
1403
1404         if (srp_post_send(target, iu, len)) {
1405                 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1406                 goto err_unmap;
1407         }
1408
1409         return 0;
1410
1411 err_unmap:
1412         srp_unmap_data(scmnd, target, req);
1413
1414 err_iu:
1415         srp_put_tx_iu(target, iu, SRP_IU_CMD);
1416
1417         spin_lock_irqsave(&target->lock, flags);
1418         list_add(&req->list, &target->free_reqs);
1419
1420 err_unlock:
1421         spin_unlock_irqrestore(&target->lock, flags);
1422
1423         return SCSI_MLQUEUE_HOST_BUSY;
1424 }
1425
1426 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1427 {
1428         int i;
1429
1430         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1431                 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1432                                                   target->max_ti_iu_len,
1433                                                   GFP_KERNEL, DMA_FROM_DEVICE);
1434                 if (!target->rx_ring[i])
1435                         goto err;
1436         }
1437
1438         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1439                 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1440                                                   target->max_iu_len,
1441                                                   GFP_KERNEL, DMA_TO_DEVICE);
1442                 if (!target->tx_ring[i])
1443                         goto err;
1444
1445                 list_add(&target->tx_ring[i]->list, &target->free_tx);
1446         }
1447
1448         return 0;
1449
1450 err:
1451         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1452                 srp_free_iu(target->srp_host, target->rx_ring[i]);
1453                 target->rx_ring[i] = NULL;
1454         }
1455
1456         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1457                 srp_free_iu(target->srp_host, target->tx_ring[i]);
1458                 target->tx_ring[i] = NULL;
1459         }
1460
1461         return -ENOMEM;
1462 }
1463
1464 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
1465 {
1466         uint64_t T_tr_ns, max_compl_time_ms;
1467         uint32_t rq_tmo_jiffies;
1468
1469         /*
1470          * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
1471          * table 91), both the QP timeout and the retry count have to be set
1472          * for RC QP's during the RTR to RTS transition.
1473          */
1474         WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
1475                      (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
1476
1477         /*
1478          * Set target->rq_tmo_jiffies to one second more than the largest time
1479          * it can take before an error completion is generated. See also
1480          * C9-140..142 in the IBTA spec for more information about how to
1481          * convert the QP Local ACK Timeout value to nanoseconds.
1482          */
1483         T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
1484         max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
1485         do_div(max_compl_time_ms, NSEC_PER_MSEC);
1486         rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
1487
1488         return rq_tmo_jiffies;
1489 }
1490
1491 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1492                                struct srp_login_rsp *lrsp,
1493                                struct srp_target_port *target)
1494 {
1495         struct ib_qp_attr *qp_attr = NULL;
1496         int attr_mask = 0;
1497         int ret;
1498         int i;
1499
1500         if (lrsp->opcode == SRP_LOGIN_RSP) {
1501                 target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1502                 target->req_lim       = be32_to_cpu(lrsp->req_lim_delta);
1503
1504                 /*
1505                  * Reserve credits for task management so we don't
1506                  * bounce requests back to the SCSI mid-layer.
1507                  */
1508                 target->scsi_host->can_queue
1509                         = min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1510                               target->scsi_host->can_queue);
1511         } else {
1512                 shost_printk(KERN_WARNING, target->scsi_host,
1513                              PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1514                 ret = -ECONNRESET;
1515                 goto error;
1516         }
1517
1518         if (!target->rx_ring[0]) {
1519                 ret = srp_alloc_iu_bufs(target);
1520                 if (ret)
1521                         goto error;
1522         }
1523
1524         ret = -ENOMEM;
1525         qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1526         if (!qp_attr)
1527                 goto error;
1528
1529         qp_attr->qp_state = IB_QPS_RTR;
1530         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1531         if (ret)
1532                 goto error_free;
1533
1534         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1535         if (ret)
1536                 goto error_free;
1537
1538         for (i = 0; i < SRP_RQ_SIZE; i++) {
1539                 struct srp_iu *iu = target->rx_ring[i];
1540                 ret = srp_post_recv(target, iu);
1541                 if (ret)
1542                         goto error_free;
1543         }
1544
1545         qp_attr->qp_state = IB_QPS_RTS;
1546         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1547         if (ret)
1548                 goto error_free;
1549
1550         target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
1551
1552         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1553         if (ret)
1554                 goto error_free;
1555
1556         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1557
1558 error_free:
1559         kfree(qp_attr);
1560
1561 error:
1562         target->status = ret;
1563 }
1564
1565 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1566                                struct ib_cm_event *event,
1567                                struct srp_target_port *target)
1568 {
1569         struct Scsi_Host *shost = target->scsi_host;
1570         struct ib_class_port_info *cpi;
1571         int opcode;
1572
1573         switch (event->param.rej_rcvd.reason) {
1574         case IB_CM_REJ_PORT_CM_REDIRECT:
1575                 cpi = event->param.rej_rcvd.ari;
1576                 target->path.dlid = cpi->redirect_lid;
1577                 target->path.pkey = cpi->redirect_pkey;
1578                 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1579                 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1580
1581                 target->status = target->path.dlid ?
1582                         SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1583                 break;
1584
1585         case IB_CM_REJ_PORT_REDIRECT:
1586                 if (srp_target_is_topspin(target)) {
1587                         /*
1588                          * Topspin/Cisco SRP gateways incorrectly send
1589                          * reject reason code 25 when they mean 24
1590                          * (port redirect).
1591                          */
1592                         memcpy(target->path.dgid.raw,
1593                                event->param.rej_rcvd.ari, 16);
1594
1595                         shost_printk(KERN_DEBUG, shost,
1596                                      PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1597                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1598                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1599
1600                         target->status = SRP_PORT_REDIRECT;
1601                 } else {
1602                         shost_printk(KERN_WARNING, shost,
1603                                      "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1604                         target->status = -ECONNRESET;
1605                 }
1606                 break;
1607
1608         case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1609                 shost_printk(KERN_WARNING, shost,
1610                             "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1611                 target->status = -ECONNRESET;
1612                 break;
1613
1614         case IB_CM_REJ_CONSUMER_DEFINED:
1615                 opcode = *(u8 *) event->private_data;
1616                 if (opcode == SRP_LOGIN_REJ) {
1617                         struct srp_login_rej *rej = event->private_data;
1618                         u32 reason = be32_to_cpu(rej->reason);
1619
1620                         if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1621                                 shost_printk(KERN_WARNING, shost,
1622                                              PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1623                         else
1624                                 shost_printk(KERN_WARNING, shost,
1625                                             PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1626                 } else
1627                         shost_printk(KERN_WARNING, shost,
1628                                      "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1629                                      " opcode 0x%02x\n", opcode);
1630                 target->status = -ECONNRESET;
1631                 break;
1632
1633         case IB_CM_REJ_STALE_CONN:
1634                 shost_printk(KERN_WARNING, shost, "  REJ reason: stale connection\n");
1635                 target->status = SRP_STALE_CONN;
1636                 break;
1637
1638         default:
1639                 shost_printk(KERN_WARNING, shost, "  REJ reason 0x%x\n",
1640                              event->param.rej_rcvd.reason);
1641                 target->status = -ECONNRESET;
1642         }
1643 }
1644
1645 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1646 {
1647         struct srp_target_port *target = cm_id->context;
1648         int comp = 0;
1649
1650         switch (event->event) {
1651         case IB_CM_REQ_ERROR:
1652                 shost_printk(KERN_DEBUG, target->scsi_host,
1653                              PFX "Sending CM REQ failed\n");
1654                 comp = 1;
1655                 target->status = -ECONNRESET;
1656                 break;
1657
1658         case IB_CM_REP_RECEIVED:
1659                 comp = 1;
1660                 srp_cm_rep_handler(cm_id, event->private_data, target);
1661                 break;
1662
1663         case IB_CM_REJ_RECEIVED:
1664                 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1665                 comp = 1;
1666
1667                 srp_cm_rej_handler(cm_id, event, target);
1668                 break;
1669
1670         case IB_CM_DREQ_RECEIVED:
1671                 shost_printk(KERN_WARNING, target->scsi_host,
1672                              PFX "DREQ received - connection closed\n");
1673                 srp_change_conn_state(target, false);
1674                 if (ib_send_cm_drep(cm_id, NULL, 0))
1675                         shost_printk(KERN_ERR, target->scsi_host,
1676                                      PFX "Sending CM DREP failed\n");
1677                 break;
1678
1679         case IB_CM_TIMEWAIT_EXIT:
1680                 shost_printk(KERN_ERR, target->scsi_host,
1681                              PFX "connection closed\n");
1682
1683                 target->status = 0;
1684                 break;
1685
1686         case IB_CM_MRA_RECEIVED:
1687         case IB_CM_DREQ_ERROR:
1688         case IB_CM_DREP_RECEIVED:
1689                 break;
1690
1691         default:
1692                 shost_printk(KERN_WARNING, target->scsi_host,
1693                              PFX "Unhandled CM event %d\n", event->event);
1694                 break;
1695         }
1696
1697         if (comp)
1698                 complete(&target->done);
1699
1700         return 0;
1701 }
1702
1703 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1704                              u64 req_tag, unsigned int lun, u8 func)
1705 {
1706         struct ib_device *dev = target->srp_host->srp_dev->dev;
1707         struct srp_iu *iu;
1708         struct srp_tsk_mgmt *tsk_mgmt;
1709
1710         if (!target->connected || target->qp_in_error)
1711                 return -1;
1712
1713         init_completion(&target->tsk_mgmt_done);
1714
1715         spin_lock_irq(&target->lock);
1716         iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1717         spin_unlock_irq(&target->lock);
1718
1719         if (!iu)
1720                 return -1;
1721
1722         ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1723                                    DMA_TO_DEVICE);
1724         tsk_mgmt = iu->buf;
1725         memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1726
1727         tsk_mgmt->opcode        = SRP_TSK_MGMT;
1728         tsk_mgmt->lun           = cpu_to_be64((u64) lun << 48);
1729         tsk_mgmt->tag           = req_tag | SRP_TAG_TSK_MGMT;
1730         tsk_mgmt->tsk_mgmt_func = func;
1731         tsk_mgmt->task_tag      = req_tag;
1732
1733         ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1734                                       DMA_TO_DEVICE);
1735         if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1736                 srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1737                 return -1;
1738         }
1739
1740         if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1741                                          msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1742                 return -1;
1743
1744         return 0;
1745 }
1746
1747 static int srp_abort(struct scsi_cmnd *scmnd)
1748 {
1749         struct srp_target_port *target = host_to_target(scmnd->device->host);
1750         struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1751         int ret;
1752
1753         shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
1754
1755         if (!req || !srp_claim_req(target, req, scmnd))
1756                 return FAILED;
1757         if (srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
1758                               SRP_TSK_ABORT_TASK) == 0)
1759                 ret = SUCCESS;
1760         else if (target->transport_offline)
1761                 ret = FAST_IO_FAIL;
1762         else
1763                 ret = FAILED;
1764         srp_free_req(target, req, scmnd, 0);
1765         scmnd->result = DID_ABORT << 16;
1766         scmnd->scsi_done(scmnd);
1767
1768         return ret;
1769 }
1770
1771 static int srp_reset_device(struct scsi_cmnd *scmnd)
1772 {
1773         struct srp_target_port *target = host_to_target(scmnd->device->host);
1774         int i;
1775
1776         shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
1777
1778         if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
1779                               SRP_TSK_LUN_RESET))
1780                 return FAILED;
1781         if (target->tsk_mgmt_status)
1782                 return FAILED;
1783
1784         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1785                 struct srp_request *req = &target->req_ring[i];
1786                 if (req->scmnd && req->scmnd->device == scmnd->device)
1787                         srp_reset_req(target, req);
1788         }
1789
1790         return SUCCESS;
1791 }
1792
1793 static int srp_reset_host(struct scsi_cmnd *scmnd)
1794 {
1795         struct srp_target_port *target = host_to_target(scmnd->device->host);
1796         int ret = FAILED;
1797
1798         shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
1799
1800         if (!srp_reconnect_target(target))
1801                 ret = SUCCESS;
1802
1803         return ret;
1804 }
1805
1806 static int srp_slave_configure(struct scsi_device *sdev)
1807 {
1808         struct Scsi_Host *shost = sdev->host;
1809         struct srp_target_port *target = host_to_target(shost);
1810         struct request_queue *q = sdev->request_queue;
1811         unsigned long timeout;
1812
1813         if (sdev->type == TYPE_DISK) {
1814                 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
1815                 blk_queue_rq_timeout(q, timeout);
1816         }
1817
1818         return 0;
1819 }
1820
1821 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
1822                            char *buf)
1823 {
1824         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1825
1826         return sprintf(buf, "0x%016llx\n",
1827                        (unsigned long long) be64_to_cpu(target->id_ext));
1828 }
1829
1830 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
1831                              char *buf)
1832 {
1833         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1834
1835         return sprintf(buf, "0x%016llx\n",
1836                        (unsigned long long) be64_to_cpu(target->ioc_guid));
1837 }
1838
1839 static ssize_t show_service_id(struct device *dev,
1840                                struct device_attribute *attr, char *buf)
1841 {
1842         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1843
1844         return sprintf(buf, "0x%016llx\n",
1845                        (unsigned long long) be64_to_cpu(target->service_id));
1846 }
1847
1848 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
1849                          char *buf)
1850 {
1851         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1852
1853         return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1854 }
1855
1856 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
1857                          char *buf)
1858 {
1859         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1860
1861         return sprintf(buf, "%pI6\n", target->path.dgid.raw);
1862 }
1863
1864 static ssize_t show_orig_dgid(struct device *dev,
1865                               struct device_attribute *attr, char *buf)
1866 {
1867         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1868
1869         return sprintf(buf, "%pI6\n", target->orig_dgid);
1870 }
1871
1872 static ssize_t show_req_lim(struct device *dev,
1873                             struct device_attribute *attr, char *buf)
1874 {
1875         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1876
1877         return sprintf(buf, "%d\n", target->req_lim);
1878 }
1879
1880 static ssize_t show_zero_req_lim(struct device *dev,
1881                                  struct device_attribute *attr, char *buf)
1882 {
1883         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1884
1885         return sprintf(buf, "%d\n", target->zero_req_lim);
1886 }
1887
1888 static ssize_t show_local_ib_port(struct device *dev,
1889                                   struct device_attribute *attr, char *buf)
1890 {
1891         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1892
1893         return sprintf(buf, "%d\n", target->srp_host->port);
1894 }
1895
1896 static ssize_t show_local_ib_device(struct device *dev,
1897                                     struct device_attribute *attr, char *buf)
1898 {
1899         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1900
1901         return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
1902 }
1903
1904 static ssize_t show_comp_vector(struct device *dev,
1905                                 struct device_attribute *attr, char *buf)
1906 {
1907         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1908
1909         return sprintf(buf, "%d\n", target->comp_vector);
1910 }
1911
1912 static ssize_t show_tl_retry_count(struct device *dev,
1913                                    struct device_attribute *attr, char *buf)
1914 {
1915         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1916
1917         return sprintf(buf, "%d\n", target->tl_retry_count);
1918 }
1919
1920 static ssize_t show_cmd_sg_entries(struct device *dev,
1921                                    struct device_attribute *attr, char *buf)
1922 {
1923         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1924
1925         return sprintf(buf, "%u\n", target->cmd_sg_cnt);
1926 }
1927
1928 static ssize_t show_allow_ext_sg(struct device *dev,
1929                                  struct device_attribute *attr, char *buf)
1930 {
1931         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1932
1933         return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
1934 }
1935
1936 static DEVICE_ATTR(id_ext,          S_IRUGO, show_id_ext,          NULL);
1937 static DEVICE_ATTR(ioc_guid,        S_IRUGO, show_ioc_guid,        NULL);
1938 static DEVICE_ATTR(service_id,      S_IRUGO, show_service_id,      NULL);
1939 static DEVICE_ATTR(pkey,            S_IRUGO, show_pkey,            NULL);
1940 static DEVICE_ATTR(dgid,            S_IRUGO, show_dgid,            NULL);
1941 static DEVICE_ATTR(orig_dgid,       S_IRUGO, show_orig_dgid,       NULL);
1942 static DEVICE_ATTR(req_lim,         S_IRUGO, show_req_lim,         NULL);
1943 static DEVICE_ATTR(zero_req_lim,    S_IRUGO, show_zero_req_lim,    NULL);
1944 static DEVICE_ATTR(local_ib_port,   S_IRUGO, show_local_ib_port,   NULL);
1945 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1946 static DEVICE_ATTR(comp_vector,     S_IRUGO, show_comp_vector,     NULL);
1947 static DEVICE_ATTR(tl_retry_count,  S_IRUGO, show_tl_retry_count,  NULL);
1948 static DEVICE_ATTR(cmd_sg_entries,  S_IRUGO, show_cmd_sg_entries,  NULL);
1949 static DEVICE_ATTR(allow_ext_sg,    S_IRUGO, show_allow_ext_sg,    NULL);
1950
1951 static struct device_attribute *srp_host_attrs[] = {
1952         &dev_attr_id_ext,
1953         &dev_attr_ioc_guid,
1954         &dev_attr_service_id,
1955         &dev_attr_pkey,
1956         &dev_attr_dgid,
1957         &dev_attr_orig_dgid,
1958         &dev_attr_req_lim,
1959         &dev_attr_zero_req_lim,
1960         &dev_attr_local_ib_port,
1961         &dev_attr_local_ib_device,
1962         &dev_attr_comp_vector,
1963         &dev_attr_tl_retry_count,
1964         &dev_attr_cmd_sg_entries,
1965         &dev_attr_allow_ext_sg,
1966         NULL
1967 };
1968
1969 static struct scsi_host_template srp_template = {
1970         .module                         = THIS_MODULE,
1971         .name                           = "InfiniBand SRP initiator",
1972         .proc_name                      = DRV_NAME,
1973         .slave_configure                = srp_slave_configure,
1974         .info                           = srp_target_info,
1975         .queuecommand                   = srp_queuecommand,
1976         .eh_abort_handler               = srp_abort,
1977         .eh_device_reset_handler        = srp_reset_device,
1978         .eh_host_reset_handler          = srp_reset_host,
1979         .skip_settle_delay              = true,
1980         .sg_tablesize                   = SRP_DEF_SG_TABLESIZE,
1981         .can_queue                      = SRP_CMD_SQ_SIZE,
1982         .this_id                        = -1,
1983         .cmd_per_lun                    = SRP_CMD_SQ_SIZE,
1984         .use_clustering                 = ENABLE_CLUSTERING,
1985         .shost_attrs                    = srp_host_attrs
1986 };
1987
1988 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1989 {
1990         struct srp_rport_identifiers ids;
1991         struct srp_rport *rport;
1992
1993         sprintf(target->target_name, "SRP.T10:%016llX",
1994                  (unsigned long long) be64_to_cpu(target->id_ext));
1995
1996         if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
1997                 return -ENODEV;
1998
1999         memcpy(ids.port_id, &target->id_ext, 8);
2000         memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2001         ids.roles = SRP_RPORT_ROLE_TARGET;
2002         rport = srp_rport_add(target->scsi_host, &ids);
2003         if (IS_ERR(rport)) {
2004                 scsi_remove_host(target->scsi_host);
2005                 return PTR_ERR(rport);
2006         }
2007
2008         rport->lld_data = target;
2009         target->rport = rport;
2010
2011         spin_lock(&host->target_lock);
2012         list_add_tail(&target->list, &host->target_list);
2013         spin_unlock(&host->target_lock);
2014
2015         target->state = SRP_TARGET_LIVE;
2016
2017         scsi_scan_target(&target->scsi_host->shost_gendev,
2018                          0, target->scsi_id, SCAN_WILD_CARD, 0);
2019
2020         return 0;
2021 }
2022
2023 static void srp_release_dev(struct device *dev)
2024 {
2025         struct srp_host *host =
2026                 container_of(dev, struct srp_host, dev);
2027
2028         complete(&host->released);
2029 }
2030
2031 static struct class srp_class = {
2032         .name    = "infiniband_srp",
2033         .dev_release = srp_release_dev
2034 };
2035
2036 /**
2037  * srp_conn_unique() - check whether the connection to a target is unique
2038  */
2039 static bool srp_conn_unique(struct srp_host *host,
2040                             struct srp_target_port *target)
2041 {
2042         struct srp_target_port *t;
2043         bool ret = false;
2044
2045         if (target->state == SRP_TARGET_REMOVED)
2046                 goto out;
2047
2048         ret = true;
2049
2050         spin_lock(&host->target_lock);
2051         list_for_each_entry(t, &host->target_list, list) {
2052                 if (t != target &&
2053                     target->id_ext == t->id_ext &&
2054                     target->ioc_guid == t->ioc_guid &&
2055                     target->initiator_ext == t->initiator_ext) {
2056                         ret = false;
2057                         break;
2058                 }
2059         }
2060         spin_unlock(&host->target_lock);
2061
2062 out:
2063         return ret;
2064 }
2065
2066 /*
2067  * Target ports are added by writing
2068  *
2069  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2070  *     pkey=<P_Key>,service_id=<service ID>
2071  *
2072  * to the add_target sysfs attribute.
2073  */
2074 enum {
2075         SRP_OPT_ERR             = 0,
2076         SRP_OPT_ID_EXT          = 1 << 0,
2077         SRP_OPT_IOC_GUID        = 1 << 1,
2078         SRP_OPT_DGID            = 1 << 2,
2079         SRP_OPT_PKEY            = 1 << 3,
2080         SRP_OPT_SERVICE_ID      = 1 << 4,
2081         SRP_OPT_MAX_SECT        = 1 << 5,
2082         SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2083         SRP_OPT_IO_CLASS        = 1 << 7,
2084         SRP_OPT_INITIATOR_EXT   = 1 << 8,
2085         SRP_OPT_CMD_SG_ENTRIES  = 1 << 9,
2086         SRP_OPT_ALLOW_EXT_SG    = 1 << 10,
2087         SRP_OPT_SG_TABLESIZE    = 1 << 11,
2088         SRP_OPT_COMP_VECTOR     = 1 << 12,
2089         SRP_OPT_TL_RETRY_COUNT  = 1 << 13,
2090         SRP_OPT_ALL             = (SRP_OPT_ID_EXT       |
2091                                    SRP_OPT_IOC_GUID     |
2092                                    SRP_OPT_DGID         |
2093                                    SRP_OPT_PKEY         |
2094                                    SRP_OPT_SERVICE_ID),
2095 };
2096
2097 static const match_table_t srp_opt_tokens = {
2098         { SRP_OPT_ID_EXT,               "id_ext=%s"             },
2099         { SRP_OPT_IOC_GUID,             "ioc_guid=%s"           },
2100         { SRP_OPT_DGID,                 "dgid=%s"               },
2101         { SRP_OPT_PKEY,                 "pkey=%x"               },
2102         { SRP_OPT_SERVICE_ID,           "service_id=%s"         },
2103         { SRP_OPT_MAX_SECT,             "max_sect=%d"           },
2104         { SRP_OPT_MAX_CMD_PER_LUN,      "max_cmd_per_lun=%d"    },
2105         { SRP_OPT_IO_CLASS,             "io_class=%x"           },
2106         { SRP_OPT_INITIATOR_EXT,        "initiator_ext=%s"      },
2107         { SRP_OPT_CMD_SG_ENTRIES,       "cmd_sg_entries=%u"     },
2108         { SRP_OPT_ALLOW_EXT_SG,         "allow_ext_sg=%u"       },
2109         { SRP_OPT_SG_TABLESIZE,         "sg_tablesize=%u"       },
2110         { SRP_OPT_COMP_VECTOR,          "comp_vector=%u"        },
2111         { SRP_OPT_TL_RETRY_COUNT,       "tl_retry_count=%u"     },
2112         { SRP_OPT_ERR,                  NULL                    }
2113 };
2114
2115 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2116 {
2117         char *options, *sep_opt;
2118         char *p;
2119         char dgid[3];
2120         substring_t args[MAX_OPT_ARGS];
2121         int opt_mask = 0;
2122         int token;
2123         int ret = -EINVAL;
2124         int i;
2125
2126         options = kstrdup(buf, GFP_KERNEL);
2127         if (!options)
2128                 return -ENOMEM;
2129
2130         sep_opt = options;
2131         while ((p = strsep(&sep_opt, ",")) != NULL) {
2132                 if (!*p)
2133                         continue;
2134
2135                 token = match_token(p, srp_opt_tokens, args);
2136                 opt_mask |= token;
2137
2138                 switch (token) {
2139                 case SRP_OPT_ID_EXT:
2140                         p = match_strdup(args);
2141                         if (!p) {
2142                                 ret = -ENOMEM;
2143                                 goto out;
2144                         }
2145                         target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2146                         kfree(p);
2147                         break;
2148
2149                 case SRP_OPT_IOC_GUID:
2150                         p = match_strdup(args);
2151                         if (!p) {
2152                                 ret = -ENOMEM;
2153                                 goto out;
2154                         }
2155                         target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2156                         kfree(p);
2157                         break;
2158
2159                 case SRP_OPT_DGID:
2160                         p = match_strdup(args);
2161                         if (!p) {
2162                                 ret = -ENOMEM;
2163                                 goto out;
2164                         }
2165                         if (strlen(p) != 32) {
2166                                 pr_warn("bad dest GID parameter '%s'\n", p);
2167                                 kfree(p);
2168                                 goto out;
2169                         }
2170
2171                         for (i = 0; i < 16; ++i) {
2172                                 strlcpy(dgid, p + i * 2, 3);
2173                                 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2174                         }
2175                         kfree(p);
2176                         memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2177                         break;
2178
2179                 case SRP_OPT_PKEY:
2180                         if (match_hex(args, &token)) {
2181                                 pr_warn("bad P_Key parameter '%s'\n", p);
2182                                 goto out;
2183                         }
2184                         target->path.pkey = cpu_to_be16(token);
2185                         break;
2186
2187                 case SRP_OPT_SERVICE_ID:
2188                         p = match_strdup(args);
2189                         if (!p) {
2190                                 ret = -ENOMEM;
2191                                 goto out;
2192                         }
2193                         target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2194                         target->path.service_id = target->service_id;
2195                         kfree(p);
2196                         break;
2197
2198                 case SRP_OPT_MAX_SECT:
2199                         if (match_int(args, &token)) {
2200                                 pr_warn("bad max sect parameter '%s'\n", p);
2201                                 goto out;
2202                         }
2203                         target->scsi_host->max_sectors = token;
2204                         break;
2205
2206                 case SRP_OPT_MAX_CMD_PER_LUN:
2207                         if (match_int(args, &token)) {
2208                                 pr_warn("bad max cmd_per_lun parameter '%s'\n",
2209                                         p);
2210                                 goto out;
2211                         }
2212                         target->scsi_host->cmd_per_lun = min(token, SRP_CMD_SQ_SIZE);
2213                         break;
2214
2215                 case SRP_OPT_IO_CLASS:
2216                         if (match_hex(args, &token)) {
2217                                 pr_warn("bad IO class parameter '%s'\n", p);
2218                                 goto out;
2219                         }
2220                         if (token != SRP_REV10_IB_IO_CLASS &&
2221                             token != SRP_REV16A_IB_IO_CLASS) {
2222                                 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2223                                         token, SRP_REV10_IB_IO_CLASS,
2224                                         SRP_REV16A_IB_IO_CLASS);
2225                                 goto out;
2226                         }
2227                         target->io_class = token;
2228                         break;
2229
2230                 case SRP_OPT_INITIATOR_EXT:
2231                         p = match_strdup(args);
2232                         if (!p) {
2233                                 ret = -ENOMEM;
2234                                 goto out;
2235                         }
2236                         target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2237                         kfree(p);
2238                         break;
2239
2240                 case SRP_OPT_CMD_SG_ENTRIES:
2241                         if (match_int(args, &token) || token < 1 || token > 255) {
2242                                 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2243                                         p);
2244                                 goto out;
2245                         }
2246                         target->cmd_sg_cnt = token;
2247                         break;
2248
2249                 case SRP_OPT_ALLOW_EXT_SG:
2250                         if (match_int(args, &token)) {
2251                                 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2252                                 goto out;
2253                         }
2254                         target->allow_ext_sg = !!token;
2255                         break;
2256
2257                 case SRP_OPT_SG_TABLESIZE:
2258                         if (match_int(args, &token) || token < 1 ||
2259                                         token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2260                                 pr_warn("bad max sg_tablesize parameter '%s'\n",
2261                                         p);
2262                                 goto out;
2263                         }
2264                         target->sg_tablesize = token;
2265                         break;
2266
2267                 case SRP_OPT_COMP_VECTOR:
2268                         if (match_int(args, &token) || token < 0) {
2269                                 pr_warn("bad comp_vector parameter '%s'\n", p);
2270                                 goto out;
2271                         }
2272                         target->comp_vector = token;
2273                         break;
2274
2275                 case SRP_OPT_TL_RETRY_COUNT:
2276                         if (match_int(args, &token) || token < 2 || token > 7) {
2277                                 pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
2278                                         p);
2279                                 goto out;
2280                         }
2281                         target->tl_retry_count = token;
2282                         break;
2283
2284                 default:
2285                         pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2286                                 p);
2287                         goto out;
2288                 }
2289         }
2290
2291         if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2292                 ret = 0;
2293         else
2294                 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2295                         if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2296                             !(srp_opt_tokens[i].token & opt_mask))
2297                                 pr_warn("target creation request is missing parameter '%s'\n",
2298                                         srp_opt_tokens[i].pattern);
2299
2300 out:
2301         kfree(options);
2302         return ret;
2303 }
2304
2305 static ssize_t srp_create_target(struct device *dev,
2306                                  struct device_attribute *attr,
2307                                  const char *buf, size_t count)
2308 {
2309         struct srp_host *host =
2310                 container_of(dev, struct srp_host, dev);
2311         struct Scsi_Host *target_host;
2312         struct srp_target_port *target;
2313         struct ib_device *ibdev = host->srp_dev->dev;
2314         dma_addr_t dma_addr;
2315         int i, ret;
2316
2317         target_host = scsi_host_alloc(&srp_template,
2318                                       sizeof (struct srp_target_port));
2319         if (!target_host)
2320                 return -ENOMEM;
2321
2322         target_host->transportt  = ib_srp_transport_template;
2323         target_host->max_channel = 0;
2324         target_host->max_id      = 1;
2325         target_host->max_lun     = SRP_MAX_LUN;
2326         target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2327
2328         target = host_to_target(target_host);
2329
2330         target->io_class        = SRP_REV16A_IB_IO_CLASS;
2331         target->scsi_host       = target_host;
2332         target->srp_host        = host;
2333         target->lkey            = host->srp_dev->mr->lkey;
2334         target->rkey            = host->srp_dev->mr->rkey;
2335         target->cmd_sg_cnt      = cmd_sg_entries;
2336         target->sg_tablesize    = indirect_sg_entries ? : cmd_sg_entries;
2337         target->allow_ext_sg    = allow_ext_sg;
2338         target->tl_retry_count  = 7;
2339
2340         ret = srp_parse_options(buf, target);
2341         if (ret)
2342                 goto err;
2343
2344         if (!srp_conn_unique(target->srp_host, target)) {
2345                 shost_printk(KERN_INFO, target->scsi_host,
2346                              PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
2347                              be64_to_cpu(target->id_ext),
2348                              be64_to_cpu(target->ioc_guid),
2349                              be64_to_cpu(target->initiator_ext));
2350                 ret = -EEXIST;
2351                 goto err;
2352         }
2353
2354         if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2355                                 target->cmd_sg_cnt < target->sg_tablesize) {
2356                 pr_warn("No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2357                 target->sg_tablesize = target->cmd_sg_cnt;
2358         }
2359
2360         target_host->sg_tablesize = target->sg_tablesize;
2361         target->indirect_size = target->sg_tablesize *
2362                                 sizeof (struct srp_direct_buf);
2363         target->max_iu_len = sizeof (struct srp_cmd) +
2364                              sizeof (struct srp_indirect_buf) +
2365                              target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2366
2367         INIT_WORK(&target->remove_work, srp_remove_work);
2368         spin_lock_init(&target->lock);
2369         INIT_LIST_HEAD(&target->free_tx);
2370         INIT_LIST_HEAD(&target->free_reqs);
2371         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
2372                 struct srp_request *req = &target->req_ring[i];
2373
2374                 req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof (void *),
2375                                         GFP_KERNEL);
2376                 req->map_page = kmalloc(SRP_FMR_SIZE * sizeof (void *),
2377                                         GFP_KERNEL);
2378                 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
2379                 if (!req->fmr_list || !req->map_page || !req->indirect_desc)
2380                         goto err_free_mem;
2381
2382                 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
2383                                              target->indirect_size,
2384                                              DMA_TO_DEVICE);
2385                 if (ib_dma_mapping_error(ibdev, dma_addr))
2386                         goto err_free_mem;
2387
2388                 req->indirect_dma_addr = dma_addr;
2389                 req->index = i;
2390                 list_add_tail(&req->list, &target->free_reqs);
2391         }
2392
2393         ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2394
2395         shost_printk(KERN_DEBUG, target->scsi_host, PFX
2396                      "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
2397                      "service_id %016llx dgid %pI6\n",
2398                (unsigned long long) be64_to_cpu(target->id_ext),
2399                (unsigned long long) be64_to_cpu(target->ioc_guid),
2400                be16_to_cpu(target->path.pkey),
2401                (unsigned long long) be64_to_cpu(target->service_id),
2402                target->path.dgid.raw);
2403
2404         ret = srp_create_target_ib(target);
2405         if (ret)
2406                 goto err_free_mem;
2407
2408         ret = srp_new_cm_id(target);
2409         if (ret)
2410                 goto err_free_ib;
2411
2412         ret = srp_connect_target(target);
2413         if (ret) {
2414                 shost_printk(KERN_ERR, target->scsi_host,
2415                              PFX "Connection failed\n");
2416                 goto err_cm_id;
2417         }
2418
2419         ret = srp_add_target(host, target);
2420         if (ret)
2421                 goto err_disconnect;
2422
2423         return count;
2424
2425 err_disconnect:
2426         srp_disconnect_target(target);
2427
2428 err_cm_id:
2429         ib_destroy_cm_id(target->cm_id);
2430
2431 err_free_ib:
2432         srp_free_target_ib(target);
2433
2434 err_free_mem:
2435         srp_free_req_data(target);
2436
2437 err:
2438         scsi_host_put(target_host);
2439
2440         return ret;
2441 }
2442
2443 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2444
2445 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2446                           char *buf)
2447 {
2448         struct srp_host *host = container_of(dev, struct srp_host, dev);
2449
2450         return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2451 }
2452
2453 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2454
2455 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2456                          char *buf)
2457 {
2458         struct srp_host *host = container_of(dev, struct srp_host, dev);
2459
2460         return sprintf(buf, "%d\n", host->port);
2461 }
2462
2463 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2464
2465 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2466 {
2467         struct srp_host *host;
2468
2469         host = kzalloc(sizeof *host, GFP_KERNEL);
2470         if (!host)
2471                 return NULL;
2472
2473         INIT_LIST_HEAD(&host->target_list);
2474         spin_lock_init(&host->target_lock);
2475         init_completion(&host->released);
2476         host->srp_dev = device;
2477         host->port = port;
2478
2479         host->dev.class = &srp_class;
2480         host->dev.parent = device->dev->dma_device;
2481         dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2482
2483         if (device_register(&host->dev))
2484                 goto free_host;
2485         if (device_create_file(&host->dev, &dev_attr_add_target))
2486                 goto err_class;
2487         if (device_create_file(&host->dev, &dev_attr_ibdev))
2488                 goto err_class;
2489         if (device_create_file(&host->dev, &dev_attr_port))
2490                 goto err_class;
2491
2492         return host;
2493
2494 err_class:
2495         device_unregister(&host->dev);
2496
2497 free_host:
2498         kfree(host);
2499
2500         return NULL;
2501 }
2502
2503 static void srp_add_one(struct ib_device *device)
2504 {
2505         struct srp_device *srp_dev;
2506         struct ib_device_attr *dev_attr;
2507         struct ib_fmr_pool_param fmr_param;
2508         struct srp_host *host;
2509         int max_pages_per_fmr, fmr_page_shift, s, e, p;
2510
2511         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2512         if (!dev_attr)
2513                 return;
2514
2515         if (ib_query_device(device, dev_attr)) {
2516                 pr_warn("Query device failed for %s\n", device->name);
2517                 goto free_attr;
2518         }
2519
2520         srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2521         if (!srp_dev)
2522                 goto free_attr;
2523
2524         /*
2525          * Use the smallest page size supported by the HCA, down to a
2526          * minimum of 4096 bytes. We're unlikely to build large sglists
2527          * out of smaller entries.
2528          */
2529         fmr_page_shift          = max(12, ffs(dev_attr->page_size_cap) - 1);
2530         srp_dev->fmr_page_size  = 1 << fmr_page_shift;
2531         srp_dev->fmr_page_mask  = ~((u64) srp_dev->fmr_page_size - 1);
2532         srp_dev->fmr_max_size   = srp_dev->fmr_page_size * SRP_FMR_SIZE;
2533
2534         INIT_LIST_HEAD(&srp_dev->dev_list);
2535
2536         srp_dev->dev = device;
2537         srp_dev->pd  = ib_alloc_pd(device);
2538         if (IS_ERR(srp_dev->pd))
2539                 goto free_dev;
2540
2541         srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2542                                     IB_ACCESS_LOCAL_WRITE |
2543                                     IB_ACCESS_REMOTE_READ |
2544                                     IB_ACCESS_REMOTE_WRITE);
2545         if (IS_ERR(srp_dev->mr))
2546                 goto err_pd;
2547
2548         for (max_pages_per_fmr = SRP_FMR_SIZE;
2549                         max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2550                         max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2551                 memset(&fmr_param, 0, sizeof fmr_param);
2552                 fmr_param.pool_size         = SRP_FMR_POOL_SIZE;
2553                 fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
2554                 fmr_param.cache             = 1;
2555                 fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2556                 fmr_param.page_shift        = fmr_page_shift;
2557                 fmr_param.access            = (IB_ACCESS_LOCAL_WRITE |
2558                                                IB_ACCESS_REMOTE_WRITE |
2559                                                IB_ACCESS_REMOTE_READ);
2560
2561                 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2562                 if (!IS_ERR(srp_dev->fmr_pool))
2563                         break;
2564         }
2565
2566         if (IS_ERR(srp_dev->fmr_pool))
2567                 srp_dev->fmr_pool = NULL;
2568
2569         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2570                 s = 0;
2571                 e = 0;
2572         } else {
2573                 s = 1;
2574                 e = device->phys_port_cnt;
2575         }
2576
2577         for (p = s; p <= e; ++p) {
2578                 host = srp_add_port(srp_dev, p);
2579                 if (host)
2580                         list_add_tail(&host->list, &srp_dev->dev_list);
2581         }
2582
2583         ib_set_client_data(device, &srp_client, srp_dev);
2584
2585         goto free_attr;
2586
2587 err_pd:
2588         ib_dealloc_pd(srp_dev->pd);
2589
2590 free_dev:
2591         kfree(srp_dev);
2592
2593 free_attr:
2594         kfree(dev_attr);
2595 }
2596
2597 static void srp_remove_one(struct ib_device *device)
2598 {
2599         struct srp_device *srp_dev;
2600         struct srp_host *host, *tmp_host;
2601         struct srp_target_port *target;
2602
2603         srp_dev = ib_get_client_data(device, &srp_client);
2604         if (!srp_dev)
2605                 return;
2606
2607         list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2608                 device_unregister(&host->dev);
2609                 /*
2610                  * Wait for the sysfs entry to go away, so that no new
2611                  * target ports can be created.
2612                  */
2613                 wait_for_completion(&host->released);
2614
2615                 /*
2616                  * Remove all target ports.
2617                  */
2618                 spin_lock(&host->target_lock);
2619                 list_for_each_entry(target, &host->target_list, list)
2620                         srp_queue_remove_work(target);
2621                 spin_unlock(&host->target_lock);
2622
2623                 /*
2624                  * Wait for target port removal tasks.
2625                  */
2626                 flush_workqueue(system_long_wq);
2627
2628                 kfree(host);
2629         }
2630
2631         if (srp_dev->fmr_pool)
2632                 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2633         ib_dereg_mr(srp_dev->mr);
2634         ib_dealloc_pd(srp_dev->pd);
2635
2636         kfree(srp_dev);
2637 }
2638
2639 static struct srp_function_template ib_srp_transport_functions = {
2640         .rport_delete            = srp_rport_delete,
2641 };
2642
2643 static int __init srp_init_module(void)
2644 {
2645         int ret;
2646
2647         BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2648
2649         if (srp_sg_tablesize) {
2650                 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2651                 if (!cmd_sg_entries)
2652                         cmd_sg_entries = srp_sg_tablesize;
2653         }
2654
2655         if (!cmd_sg_entries)
2656                 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2657
2658         if (cmd_sg_entries > 255) {
2659                 pr_warn("Clamping cmd_sg_entries to 255\n");
2660                 cmd_sg_entries = 255;
2661         }
2662
2663         if (!indirect_sg_entries)
2664                 indirect_sg_entries = cmd_sg_entries;
2665         else if (indirect_sg_entries < cmd_sg_entries) {
2666                 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
2667                         cmd_sg_entries);
2668                 indirect_sg_entries = cmd_sg_entries;
2669         }
2670
2671         ib_srp_transport_template =
2672                 srp_attach_transport(&ib_srp_transport_functions);
2673         if (!ib_srp_transport_template)
2674                 return -ENOMEM;
2675
2676         ret = class_register(&srp_class);
2677         if (ret) {
2678                 pr_err("couldn't register class infiniband_srp\n");
2679                 srp_release_transport(ib_srp_transport_template);
2680                 return ret;
2681         }
2682
2683         ib_sa_register_client(&srp_sa_client);
2684
2685         ret = ib_register_client(&srp_client);
2686         if (ret) {
2687                 pr_err("couldn't register IB client\n");
2688                 srp_release_transport(ib_srp_transport_template);
2689                 ib_sa_unregister_client(&srp_sa_client);
2690                 class_unregister(&srp_class);
2691                 return ret;
2692         }
2693
2694         return 0;
2695 }
2696
2697 static void __exit srp_cleanup_module(void)
2698 {
2699         ib_unregister_client(&srp_client);
2700         ib_sa_unregister_client(&srp_sa_client);
2701         class_unregister(&srp_class);
2702         srp_release_transport(ib_srp_transport_template);
2703 }
2704
2705 module_init(srp_init_module);
2706 module_exit(srp_cleanup_module);