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  * $Id: ib_srp.c 3932 2005-11-01 17:19:29Z roland $
  33  */
  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 <asm/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
  51 #include <rdma/ib_cache.h>
  52
  53 #include "ib_srp.h"
  54
  55 #define DRV_NAME        "ib_srp"
  56 #define PFX             DRV_NAME ": "
  57 #define DRV_VERSION     "0.2"
  58 #define DRV_RELDATE     "November 1, 2005"
  59
  60 MODULE_AUTHOR("Roland Dreier");
  61 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
  62                    "v" DRV_VERSION " (" DRV_RELDATE ")");
  63 MODULE_LICENSE("Dual BSD/GPL");
  64
  65 static int srp_sg_tablesize = SRP_DEF_SG_TABLESIZE;
  66 static int srp_max_iu_len;
  67
  68 module_param(srp_sg_tablesize, int, 0444);
  69 MODULE_PARM_DESC(srp_sg_tablesize,
  70                  "Max number of gather/scatter entries per I/O (default is 12)");
  71
  72 static int topspin_workarounds = 1;
  73
  74 module_param(topspin_workarounds, int, 0444);
  75 MODULE_PARM_DESC(topspin_workarounds,
  76                  "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
  77
  78 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
  79
  80 static int mellanox_workarounds = 1;
  81
  82 module_param(mellanox_workarounds, int, 0444);
  83 MODULE_PARM_DESC(mellanox_workarounds,
  84                  "Enable workarounds for Mellanox SRP target bugs if != 0");
  85
  86 static const u8 mellanox_oui[3] = { 0x00, 0x02, 0xc9 };
  87
  88 static void srp_add_one(struct ib_device *device);
  89 static void srp_remove_one(struct ib_device *device);
  90 static void srp_completion(struct ib_cq *cq, void *target_ptr);
  91 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
  92
  93 static struct ib_client srp_client = {
  94         .name   = "srp",
  95         .add    = srp_add_one,
  96         .remove = srp_remove_one
  97 };
  98
  99 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
 100 {
 101         return (struct srp_target_port *) host->hostdata;
 102 }
 103
 104 static const char *srp_target_info(struct Scsi_Host *host)
 105 {
 106         return host_to_target(host)->target_name;
 107 }
 108
 109 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
 110                                    gfp_t gfp_mask,
 111                                    enum dma_data_direction direction)
 112 {
 113         struct srp_iu *iu;
 114
 115         iu = kmalloc(sizeof *iu, gfp_mask);
 116         if (!iu)
 117                 goto out;
 118
 119         iu->buf = kzalloc(size, gfp_mask);
 120         if (!iu->buf)
 121                 goto out_free_iu;
 122
 123         iu->dma = dma_map_single(host->dev->dev->dma_device,
 124                                  iu->buf, size, direction);
 125         if (dma_mapping_error(iu->dma))
 126                 goto out_free_buf;
 127
 128         iu->size      = size;
 129         iu->direction = direction;
 130
 131         return iu;
 132
 133 out_free_buf:
 134         kfree(iu->buf);
 135 out_free_iu:
 136         kfree(iu);
 137 out:
 138         return NULL;
 139 }
 140
 141 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
 142 {
 143         if (!iu)
 144                 return;
 145
 146         dma_unmap_single(host->dev->dev->dma_device,
 147                          iu->dma, iu->size, iu->direction);
 148         kfree(iu->buf);
 149         kfree(iu);
 150 }
 151
 152 static void srp_qp_event(struct ib_event *event, void *context)
 153 {
 154         printk(KERN_ERR PFX "QP event %d\n", event->event);
 155 }
 156
 157 static int srp_init_qp(struct srp_target_port *target,
 158                        struct ib_qp *qp)
 159 {
 160         struct ib_qp_attr *attr;
 161         int ret;
 162
 163         attr = kmalloc(sizeof *attr, GFP_KERNEL);
 164         if (!attr)
 165                 return -ENOMEM;
 166
 167         ret = ib_find_cached_pkey(target->srp_host->dev->dev,
 168                                   target->srp_host->port,
 169                                   be16_to_cpu(target->path.pkey),
 170                                   &attr->pkey_index);
 171         if (ret)
 172                 goto out;
 173
 174         attr->qp_state        = IB_QPS_INIT;
 175         attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
 176                                     IB_ACCESS_REMOTE_WRITE);
 177         attr->port_num        = target->srp_host->port;
 178
 179         ret = ib_modify_qp(qp, attr,
 180                            IB_QP_STATE          |
 181                            IB_QP_PKEY_INDEX     |
 182                            IB_QP_ACCESS_FLAGS   |
 183                            IB_QP_PORT);
 184
 185 out:
 186         kfree(attr);
 187         return ret;
 188 }
 189
 190 static int srp_create_target_ib(struct srp_target_port *target)
 191 {
 192         struct ib_qp_init_attr *init_attr;
 193         int ret;
 194
 195         init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
 196         if (!init_attr)
 197                 return -ENOMEM;
 198
 199         target->cq = ib_create_cq(target->srp_host->dev->dev, srp_completion,
 200                                   NULL, target, SRP_CQ_SIZE);
 201         if (IS_ERR(target->cq)) {
 202                 ret = PTR_ERR(target->cq);
 203                 goto out;
 204         }
 205
 206         ib_req_notify_cq(target->cq, IB_CQ_NEXT_COMP);
 207
 208         init_attr->event_handler       = srp_qp_event;
 209         init_attr->cap.max_send_wr     = SRP_SQ_SIZE;
 210         init_attr->cap.max_recv_wr     = SRP_RQ_SIZE;
 211         init_attr->cap.max_recv_sge    = 1;
 212         init_attr->cap.max_send_sge    = 1;
 213         init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
 214         init_attr->qp_type             = IB_QPT_RC;
 215         init_attr->send_cq             = target->cq;
 216         init_attr->recv_cq             = target->cq;
 217
 218         target->qp = ib_create_qp(target->srp_host->dev->pd, init_attr);
 219         if (IS_ERR(target->qp)) {
 220                 ret = PTR_ERR(target->qp);
 221                 ib_destroy_cq(target->cq);
 222                 goto out;
 223         }
 224
 225         ret = srp_init_qp(target, target->qp);
 226         if (ret) {
 227                 ib_destroy_qp(target->qp);
 228                 ib_destroy_cq(target->cq);
 229                 goto out;
 230         }
 231
 232 out:
 233         kfree(init_attr);
 234         return ret;
 235 }
 236
 237 static void srp_free_target_ib(struct srp_target_port *target)
 238 {
 239         int i;
 240
 241         ib_destroy_qp(target->qp);
 242         ib_destroy_cq(target->cq);
 243
 244         for (i = 0; i < SRP_RQ_SIZE; ++i)
 245                 srp_free_iu(target->srp_host, target->rx_ring[i]);
 246         for (i = 0; i < SRP_SQ_SIZE + 1; ++i)
 247                 srp_free_iu(target->srp_host, target->tx_ring[i]);
 248 }
 249
 250 static void srp_path_rec_completion(int status,
 251                                     struct ib_sa_path_rec *pathrec,
 252                                     void *target_ptr)
 253 {
 254         struct srp_target_port *target = target_ptr;
 255
 256         target->status = status;
 257         if (status)
 258                 printk(KERN_ERR PFX "Got failed path rec status %d\n", status);
 259         else
 260                 target->path = *pathrec;
 261         complete(&target->done);
 262 }
 263
 264 static int srp_lookup_path(struct srp_target_port *target)
 265 {
 266         target->path.numb_path = 1;
 267
 268         init_completion(&target->done);
 269
 270         target->path_query_id = ib_sa_path_rec_get(target->srp_host->dev->dev,
 271                                                    target->srp_host->port,
 272                                                    &target->path,
 273                                                    IB_SA_PATH_REC_DGID          |
 274                                                    IB_SA_PATH_REC_SGID          |
 275                                                    IB_SA_PATH_REC_NUMB_PATH     |
 276                                                    IB_SA_PATH_REC_PKEY,
 277                                                    SRP_PATH_REC_TIMEOUT_MS,
 278                                                    GFP_KERNEL,
 279                                                    srp_path_rec_completion,
 280                                                    target, &target->path_query);
 281         if (target->path_query_id < 0)
 282                 return target->path_query_id;
 283
 284         wait_for_completion(&target->done);
 285
 286         if (target->status < 0)
 287                 printk(KERN_WARNING PFX "Path record query failed\n");
 288
 289         return target->status;
 290 }
 291
 292 static int srp_send_req(struct srp_target_port *target)
 293 {
 294         struct {
 295                 struct ib_cm_req_param param;
 296                 struct srp_login_req   priv;
 297         } *req = NULL;
 298         int status;
 299
 300         req = kzalloc(sizeof *req, GFP_KERNEL);
 301         if (!req)
 302                 return -ENOMEM;
 303
 304         req->param.primary_path               = &target->path;
 305         req->param.alternate_path             = NULL;
 306         req->param.service_id                 = target->service_id;
 307         req->param.qp_num                     = target->qp->qp_num;
 308         req->param.qp_type                    = target->qp->qp_type;
 309         req->param.private_data               = &req->priv;
 310         req->param.private_data_len           = sizeof req->priv;
 311         req->param.flow_control               = 1;
 312
 313         get_random_bytes(&req->param.starting_psn, 4);
 314         req->param.starting_psn              &= 0xffffff;
 315
 316         /*
 317          * Pick some arbitrary defaults here; we could make these
 318          * module parameters if anyone cared about setting them.
 319          */
 320         req->param.responder_resources        = 4;
 321         req->param.remote_cm_response_timeout = 20;
 322         req->param.local_cm_response_timeout  = 20;
 323         req->param.retry_count                = 7;
 324         req->param.rnr_retry_count            = 7;
 325         req->param.max_cm_retries             = 15;
 326
 327         req->priv.opcode        = SRP_LOGIN_REQ;
 328         req->priv.tag           = 0;
 329         req->priv.req_it_iu_len = cpu_to_be32(srp_max_iu_len);
 330         req->priv.req_buf_fmt   = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
 331                                               SRP_BUF_FORMAT_INDIRECT);
 332         /*
 333          * In the published SRP specification (draft rev. 16a), the
 334          * port identifier format is 8 bytes of ID extension followed
 335          * by 8 bytes of GUID.  Older drafts put the two halves in the
 336          * opposite order, so that the GUID comes first.
 337          *
 338          * Targets conforming to these obsolete drafts can be
 339          * recognized by the I/O Class they report.
 340          */
 341         if (target->io_class == SRP_REV10_IB_IO_CLASS) {
 342                 memcpy(req->priv.initiator_port_id,
 343                        target->srp_host->initiator_port_id + 8, 8);
 344                 memcpy(req->priv.initiator_port_id + 8,
 345                        target->srp_host->initiator_port_id, 8);
 346                 memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
 347                 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
 348         } else {
 349                 memcpy(req->priv.initiator_port_id,
 350                        target->srp_host->initiator_port_id, 16);
 351                 memcpy(req->priv.target_port_id,     &target->id_ext, 8);
 352                 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
 353         }
 354
 355         /*
 356          * Topspin/Cisco SRP targets will reject our login unless we
 357          * zero out the first 8 bytes of our initiator port ID.  The
 358          * second 8 bytes must be our local node GUID, but we always
 359          * use that anyway.
 360          */
 361         if (topspin_workarounds && !memcmp(&target->ioc_guid, topspin_oui, 3)) {
 362                 printk(KERN_DEBUG PFX "Topspin/Cisco initiator port ID workaround "
 363                        "activated for target GUID %016llx\n",
 364                        (unsigned long long) be64_to_cpu(target->ioc_guid));
 365                 memset(req->priv.initiator_port_id, 0, 8);
 366         }
 367
 368         status = ib_send_cm_req(target->cm_id, &req->param);
 369
 370         kfree(req);
 371
 372         return status;
 373 }
 374
 375 static void srp_disconnect_target(struct srp_target_port *target)
 376 {
 377         /* XXX should send SRP_I_LOGOUT request */
 378
 379         init_completion(&target->done);
 380         if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
 381                 printk(KERN_DEBUG PFX "Sending CM DREQ failed\n");
 382                 return;
 383         }
 384         wait_for_completion(&target->done);
 385 }
 386
 387 static void srp_remove_work(void *target_ptr)
 388 {
 389         struct srp_target_port *target = target_ptr;
 390
 391         spin_lock_irq(target->scsi_host->host_lock);
 392         if (target->state != SRP_TARGET_DEAD) {
 393                 spin_unlock_irq(target->scsi_host->host_lock);
 394                 return;
 395         }
 396         target->state = SRP_TARGET_REMOVED;
 397         spin_unlock_irq(target->scsi_host->host_lock);
 398
 399         spin_lock(&target->srp_host->target_lock);
 400         list_del(&target->list);
 401         spin_unlock(&target->srp_host->target_lock);
 402
 403         scsi_remove_host(target->scsi_host);
 404         ib_destroy_cm_id(target->cm_id);
 405         srp_free_target_ib(target);
 406         scsi_host_put(target->scsi_host);
 407 }
 408
 409 static int srp_connect_target(struct srp_target_port *target)
 410 {
 411         int ret;
 412
 413         ret = srp_lookup_path(target);
 414         if (ret)
 415                 return ret;
 416
 417         while (1) {
 418                 init_completion(&target->done);
 419                 ret = srp_send_req(target);
 420                 if (ret)
 421                         return ret;
 422                 wait_for_completion(&target->done);
 423
 424                 /*
 425                  * The CM event handling code will set status to
 426                  * SRP_PORT_REDIRECT if we get a port redirect REJ
 427                  * back, or SRP_DLID_REDIRECT if we get a lid/qp
 428                  * redirect REJ back.
 429                  */
 430                 switch (target->status) {
 431                 case 0:
 432                         return 0;
 433
 434                 case SRP_PORT_REDIRECT:
 435                         ret = srp_lookup_path(target);
 436                         if (ret)
 437                                 return ret;
 438                         break;
 439
 440                 case SRP_DLID_REDIRECT:
 441                         break;
 442
 443                 default:
 444                         return target->status;
 445                 }
 446         }
 447 }
 448
 449 static void srp_unmap_data(struct scsi_cmnd *scmnd,
 450                            struct srp_target_port *target,
 451                            struct srp_request *req)
 452 {
 453         struct scatterlist *scat;
 454         int nents;
 455
 456         if (!scmnd->request_buffer ||
 457             (scmnd->sc_data_direction != DMA_TO_DEVICE &&
 458              scmnd->sc_data_direction != DMA_FROM_DEVICE))
 459                 return;
 460
 461         if (req->fmr) {
 462                 ib_fmr_pool_unmap(req->fmr);
 463                 req->fmr = NULL;
 464         }
 465
 466         /*
 467          * This handling of non-SG commands can be killed when the
 468          * SCSI midlayer no longer generates non-SG commands.
 469          */
 470         if (likely(scmnd->use_sg)) {
 471                 nents = scmnd->use_sg;
 472                 scat  = scmnd->request_buffer;
 473         } else {
 474                 nents = 1;
 475                 scat  = &req->fake_sg;
 476         }
 477
 478         dma_unmap_sg(target->srp_host->dev->dev->dma_device, scat, nents,
 479                      scmnd->sc_data_direction);
 480 }
 481
 482 static void srp_remove_req(struct srp_target_port *target, struct srp_request *req)
 483 {
 484         srp_unmap_data(req->scmnd, target, req);
 485         list_move_tail(&req->list, &target->free_reqs);
 486 }
 487
 488 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
 489 {
 490         req->scmnd->result = DID_RESET << 16;
 491         req->scmnd->scsi_done(req->scmnd);
 492         srp_remove_req(target, req);
 493 }
 494
 495 static int srp_reconnect_target(struct srp_target_port *target)
 496 {
 497         struct ib_cm_id *new_cm_id;
 498         struct ib_qp_attr qp_attr;
 499         struct srp_request *req, *tmp;
 500         struct ib_wc wc;
 501         int ret;
 502
 503         spin_lock_irq(target->scsi_host->host_lock);
 504         if (target->state != SRP_TARGET_LIVE) {
 505                 spin_unlock_irq(target->scsi_host->host_lock);
 506                 return -EAGAIN;
 507         }
 508         target->state = SRP_TARGET_CONNECTING;
 509         spin_unlock_irq(target->scsi_host->host_lock);
 510
 511         srp_disconnect_target(target);
 512         /*
 513          * Now get a new local CM ID so that we avoid confusing the
 514          * target in case things are really fouled up.
 515          */
 516         new_cm_id = ib_create_cm_id(target->srp_host->dev->dev,
 517                                     srp_cm_handler, target);
 518         if (IS_ERR(new_cm_id)) {
 519                 ret = PTR_ERR(new_cm_id);
 520                 goto err;
 521         }
 522         ib_destroy_cm_id(target->cm_id);
 523         target->cm_id = new_cm_id;
 524
 525         qp_attr.qp_state = IB_QPS_RESET;
 526         ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
 527         if (ret)
 528                 goto err;
 529
 530         ret = srp_init_qp(target, target->qp);
 531         if (ret)
 532                 goto err;
 533
 534         while (ib_poll_cq(target->cq, 1, &wc) > 0)
 535                 ; /* nothing */
 536
 537         spin_lock_irq(target->scsi_host->host_lock);
 538         list_for_each_entry_safe(req, tmp, &target->req_queue, list)
 539                 srp_reset_req(target, req);
 540         spin_unlock_irq(target->scsi_host->host_lock);
 541
 542         target->rx_head  = 0;
 543         target->tx_head  = 0;
 544         target->tx_tail  = 0;
 545
 546         ret = srp_connect_target(target);
 547         if (ret)
 548                 goto err;
 549
 550         spin_lock_irq(target->scsi_host->host_lock);
 551         if (target->state == SRP_TARGET_CONNECTING) {
 552                 ret = 0;
 553                 target->state = SRP_TARGET_LIVE;
 554         } else
 555                 ret = -EAGAIN;
 556         spin_unlock_irq(target->scsi_host->host_lock);
 557
 558         return ret;
 559
 560 err:
 561         printk(KERN_ERR PFX "reconnect failed (%d), removing target port.\n", ret);
 562
 563         /*
 564          * We couldn't reconnect, so kill our target port off.
 565          * However, we have to defer the real removal because we might
 566          * be in the context of the SCSI error handler now, which
 567          * would deadlock if we call scsi_remove_host().
 568          */
 569         spin_lock_irq(target->scsi_host->host_lock);
 570         if (target->state == SRP_TARGET_CONNECTING) {
 571                 target->state = SRP_TARGET_DEAD;
 572                 INIT_WORK(&target->work, srp_remove_work, target);
 573                 schedule_work(&target->work);
 574         }
 575         spin_unlock_irq(target->scsi_host->host_lock);
 576
 577         return ret;
 578 }
 579
 580 static int srp_map_fmr(struct srp_target_port *target, struct scatterlist *scat,
 581                        int sg_cnt, struct srp_request *req,
 582                        struct srp_direct_buf *buf)
 583 {
 584         u64 io_addr = 0;
 585         u64 *dma_pages;
 586         u32 len;
 587         int page_cnt;
 588         int i, j;
 589         int ret;
 590         struct srp_device *dev = target->srp_host->dev;
 591
 592         if (!dev->fmr_pool)
 593                 return -ENODEV;
 594
 595         if ((sg_dma_address(&scat[0]) & ~dev->fmr_page_mask) &&
 596             mellanox_workarounds && !memcmp(&target->ioc_guid, mellanox_oui, 3))
 597                 return -EINVAL;
 598
 599         len = page_cnt = 0;
 600         for (i = 0; i < sg_cnt; ++i) {
 601                 if (sg_dma_address(&scat[i]) & ~dev->fmr_page_mask) {
 602                         if (i > 0)
 603                                 return -EINVAL;
 604                         else
 605                                 ++page_cnt;
 606                 }
 607                 if ((sg_dma_address(&scat[i]) + sg_dma_len(&scat[i])) &
 608                     ~dev->fmr_page_mask) {
 609                         if (i < sg_cnt - 1)
 610                                 return -EINVAL;
 611                         else
 612                                 ++page_cnt;
 613                 }
 614
 615                 len += sg_dma_len(&scat[i]);
 616         }
 617
 618         page_cnt += len >> dev->fmr_page_shift;
 619         if (page_cnt > SRP_FMR_SIZE)
 620                 return -ENOMEM;
 621
 622         dma_pages = kmalloc(sizeof (u64) * page_cnt, GFP_ATOMIC);
 623         if (!dma_pages)
 624                 return -ENOMEM;
 625
 626         page_cnt = 0;
 627         for (i = 0; i < sg_cnt; ++i)
 628                 for (j = 0; j < sg_dma_len(&scat[i]); j += dev->fmr_page_size)
 629                         dma_pages[page_cnt++] =
 630                                 (sg_dma_address(&scat[i]) & dev->fmr_page_mask) + j;
 631
 632         req->fmr = ib_fmr_pool_map_phys(dev->fmr_pool,
 633                                         dma_pages, page_cnt, io_addr);
 634         if (IS_ERR(req->fmr)) {
 635                 ret = PTR_ERR(req->fmr);
 636                 req->fmr = NULL;
 637                 goto out;
 638         }
 639
 640         buf->va  = cpu_to_be64(sg_dma_address(&scat[0]) & ~dev->fmr_page_mask);
 641         buf->key = cpu_to_be32(req->fmr->fmr->rkey);
 642         buf->len = cpu_to_be32(len);
 643
 644         ret = 0;
 645
 646 out:
 647         kfree(dma_pages);
 648
 649         return ret;
 650 }
 651
 652 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
 653                         struct srp_request *req)
 654 {
 655         struct scatterlist *scat;
 656         struct srp_cmd *cmd = req->cmd->buf;
 657         int len, nents, count;
 658         u8 fmt = SRP_DATA_DESC_DIRECT;
 659
 660         if (!scmnd->request_buffer || scmnd->sc_data_direction == DMA_NONE)
 661                 return sizeof (struct srp_cmd);
 662
 663         if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
 664             scmnd->sc_data_direction != DMA_TO_DEVICE) {
 665                 printk(KERN_WARNING PFX "Unhandled data direction %d\n",
 666                        scmnd->sc_data_direction);
 667                 return -EINVAL;
 668         }
 669
 670         /*
 671          * This handling of non-SG commands can be killed when the
 672          * SCSI midlayer no longer generates non-SG commands.
 673          */
 674         if (likely(scmnd->use_sg)) {
 675                 nents = scmnd->use_sg;
 676                 scat  = scmnd->request_buffer;
 677         } else {
 678                 nents = 1;
 679                 scat  = &req->fake_sg;
 680                 sg_init_one(scat, scmnd->request_buffer, scmnd->request_bufflen);
 681         }
 682
 683         count = dma_map_sg(target->srp_host->dev->dev->dma_device,
 684                            scat, nents, scmnd->sc_data_direction);
 685
 686         fmt = SRP_DATA_DESC_DIRECT;
 687         len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
 688
 689         if (count == 1) {
 690                 /*
 691                  * The midlayer only generated a single gather/scatter
 692                  * entry, or DMA mapping coalesced everything to a
 693                  * single entry.  So a direct descriptor along with
 694                  * the DMA MR suffices.
 695                  */
 696                 struct srp_direct_buf *buf = (void *) cmd->add_data;
 697
 698                 buf->va  = cpu_to_be64(sg_dma_address(scat));
 699                 buf->key = cpu_to_be32(target->srp_host->dev->mr->rkey);
 700                 buf->len = cpu_to_be32(sg_dma_len(scat));
 701         } else if (srp_map_fmr(target, scat, count, req,
 702                                (void *) cmd->add_data)) {
 703                 /*
 704                  * FMR mapping failed, and the scatterlist has more
 705                  * than one entry.  Generate an indirect memory
 706                  * descriptor.
 707                  */
 708                 struct srp_indirect_buf *buf = (void *) cmd->add_data;
 709                 u32 datalen = 0;
 710                 int i;
 711
 712                 fmt = SRP_DATA_DESC_INDIRECT;
 713                 len = sizeof (struct srp_cmd) +
 714                         sizeof (struct srp_indirect_buf) +
 715                         count * sizeof (struct srp_direct_buf);
 716
 717                 for (i = 0; i < count; ++i) {
 718                         buf->desc_list[i].va  =
 719                                 cpu_to_be64(sg_dma_address(&scat[i]));
 720                         buf->desc_list[i].key =
 721                                 cpu_to_be32(target->srp_host->dev->mr->rkey);
 722                         buf->desc_list[i].len =
 723                                 cpu_to_be32(sg_dma_len(&scat[i]));
 724                         datalen += sg_dma_len(&scat[i]);
 725                 }
 726
 727                 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
 728                         cmd->data_out_desc_cnt = count;
 729                 else
 730                         cmd->data_in_desc_cnt = count;
 731
 732                 buf->table_desc.va  =
 733                         cpu_to_be64(req->cmd->dma + sizeof *cmd + sizeof *buf);
 734                 buf->table_desc.key =
 735                         cpu_to_be32(target->srp_host->dev->mr->rkey);
 736                 buf->table_desc.len =
 737                         cpu_to_be32(count * sizeof (struct srp_direct_buf));
 738
 739                 buf->len = cpu_to_be32(datalen);
 740         }
 741
 742         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
 743                 cmd->buf_fmt = fmt << 4;
 744         else
 745                 cmd->buf_fmt = fmt;
 746
 747         return len;
 748 }
 749
 750 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
 751 {
 752         struct srp_request *req;
 753         struct scsi_cmnd *scmnd;
 754         unsigned long flags;
 755         s32 delta;
 756
 757         delta = (s32) be32_to_cpu(rsp->req_lim_delta);
 758
 759         spin_lock_irqsave(target->scsi_host->host_lock, flags);
 760
 761         target->req_lim += delta;
 762
 763         req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
 764
 765         if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
 766                 if (be32_to_cpu(rsp->resp_data_len) < 4)
 767                         req->tsk_status = -1;
 768                 else
 769                         req->tsk_status = rsp->data[3];
 770                 complete(&req->done);
 771         } else {
 772                 scmnd = req->scmnd;
 773                 if (!scmnd)
 774                         printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
 775                                (unsigned long long) rsp->tag);
 776                 scmnd->result = rsp->status;
 777
 778                 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
 779                         memcpy(scmnd->sense_buffer, rsp->data +
 780                                be32_to_cpu(rsp->resp_data_len),
 781                                min_t(int, be32_to_cpu(rsp->sense_data_len),
 782                                      SCSI_SENSE_BUFFERSIZE));
 783                 }
 784
 785                 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
 786                         scmnd->resid = be32_to_cpu(rsp->data_out_res_cnt);
 787                 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
 788                         scmnd->resid = be32_to_cpu(rsp->data_in_res_cnt);
 789
 790                 if (!req->tsk_mgmt) {
 791                         scmnd->host_scribble = (void *) -1L;
 792                         scmnd->scsi_done(scmnd);
 793
 794                         srp_remove_req(target, req);
 795                 } else
 796                         req->cmd_done = 1;
 797         }
 798
 799         spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
 800 }
 801
 802 static void srp_reconnect_work(void *target_ptr)
 803 {
 804         struct srp_target_port *target = target_ptr;
 805
 806         srp_reconnect_target(target);
 807 }
 808
 809 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
 810 {
 811         struct srp_iu *iu;
 812         u8 opcode;
 813
 814         iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
 815
 816         dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
 817                                 target->max_ti_iu_len, DMA_FROM_DEVICE);
 818
 819         opcode = *(u8 *) iu->buf;
 820
 821         if (0) {
 822                 int i;
 823
 824                 printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
 825
 826                 for (i = 0; i < wc->byte_len; ++i) {
 827                         if (i % 8 == 0)
 828                                 printk(KERN_ERR "  [%02x] ", i);
 829                         printk(" %02x", ((u8 *) iu->buf)[i]);
 830                         if ((i + 1) % 8 == 0)
 831                                 printk("\n");
 832                 }
 833
 834                 if (wc->byte_len % 8)
 835                         printk("\n");
 836         }
 837
 838         switch (opcode) {
 839         case SRP_RSP:
 840                 srp_process_rsp(target, iu->buf);
 841                 break;
 842
 843         case SRP_T_LOGOUT:
 844                 /* XXX Handle target logout */
 845                 printk(KERN_WARNING PFX "Got target logout request\n");
 846                 break;
 847
 848         default:
 849                 printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
 850                 break;
 851         }
 852
 853         dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
 854                                    target->max_ti_iu_len, DMA_FROM_DEVICE);
 855 }
 856
 857 static void srp_completion(struct ib_cq *cq, void *target_ptr)
 858 {
 859         struct srp_target_port *target = target_ptr;
 860         struct ib_wc wc;
 861         unsigned long flags;
 862
 863         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 864         while (ib_poll_cq(cq, 1, &wc) > 0) {
 865                 if (wc.status) {
 866                         printk(KERN_ERR PFX "failed %s status %d\n",
 867                                wc.wr_id & SRP_OP_RECV ? "receive" : "send",
 868                                wc.status);
 869                         spin_lock_irqsave(target->scsi_host->host_lock, flags);
 870                         if (target->state == SRP_TARGET_LIVE)
 871                                 schedule_work(&target->work);
 872                         spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
 873                         break;
 874                 }
 875
 876                 if (wc.wr_id & SRP_OP_RECV)
 877                         srp_handle_recv(target, &wc);
 878                 else
 879                         ++target->tx_tail;
 880         }
 881 }
 882
 883 static int __srp_post_recv(struct srp_target_port *target)
 884 {
 885         struct srp_iu *iu;
 886         struct ib_sge list;
 887         struct ib_recv_wr wr, *bad_wr;
 888         unsigned int next;
 889         int ret;
 890
 891         next     = target->rx_head & (SRP_RQ_SIZE - 1);
 892         wr.wr_id = next | SRP_OP_RECV;
 893         iu       = target->rx_ring[next];
 894
 895         list.addr   = iu->dma;
 896         list.length = iu->size;
 897         list.lkey   = target->srp_host->dev->mr->lkey;
 898
 899         wr.next     = NULL;
 900         wr.sg_list  = &list;
 901         wr.num_sge  = 1;
 902
 903         ret = ib_post_recv(target->qp, &wr, &bad_wr);
 904         if (!ret)
 905                 ++target->rx_head;
 906
 907         return ret;
 908 }
 909
 910 static int srp_post_recv(struct srp_target_port *target)
 911 {
 912         unsigned long flags;
 913         int ret;
 914
 915         spin_lock_irqsave(target->scsi_host->host_lock, flags);
 916         ret = __srp_post_recv(target);
 917         spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
 918
 919         return ret;
 920 }
 921
 922 /*
 923  * Must be called with target->scsi_host->host_lock held to protect
 924  * req_lim and tx_head.  Lock cannot be dropped between call here and
 925  * call to __srp_post_send().
 926  */
 927 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
 928 {
 929         if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
 930                 return NULL;
 931
 932         if (unlikely(target->req_lim < 1))
 933                 ++target->zero_req_lim;
 934
 935         return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
 936 }
 937
 938 /*
 939  * Must be called with target->scsi_host->host_lock held to protect
 940  * req_lim and tx_head.
 941  */
 942 static int __srp_post_send(struct srp_target_port *target,
 943                            struct srp_iu *iu, int len)
 944 {
 945         struct ib_sge list;
 946         struct ib_send_wr wr, *bad_wr;
 947         int ret = 0;
 948
 949         list.addr   = iu->dma;
 950         list.length = len;
 951         list.lkey   = target->srp_host->dev->mr->lkey;
 952
 953         wr.next       = NULL;
 954         wr.wr_id      = target->tx_head & SRP_SQ_SIZE;
 955         wr.sg_list    = &list;
 956         wr.num_sge    = 1;
 957         wr.opcode     = IB_WR_SEND;
 958         wr.send_flags = IB_SEND_SIGNALED;
 959
 960         ret = ib_post_send(target->qp, &wr, &bad_wr);
 961
 962         if (!ret) {
 963                 ++target->tx_head;
 964                 --target->req_lim;
 965         }
 966
 967         return ret;
 968 }
 969
 970 static int srp_queuecommand(struct scsi_cmnd *scmnd,
 971                             void (*done)(struct scsi_cmnd *))
 972 {
 973         struct srp_target_port *target = host_to_target(scmnd->device->host);
 974         struct srp_request *req;
 975         struct srp_iu *iu;
 976         struct srp_cmd *cmd;
 977         int len;
 978
 979         if (target->state == SRP_TARGET_CONNECTING)
 980                 goto err;
 981
 982         if (target->state == SRP_TARGET_DEAD ||
 983             target->state == SRP_TARGET_REMOVED) {
 984                 scmnd->result = DID_BAD_TARGET << 16;
 985                 done(scmnd);
 986                 return 0;
 987         }
 988
 989         iu = __srp_get_tx_iu(target);
 990         if (!iu)
 991                 goto err;
 992
 993         dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
 994                                 srp_max_iu_len, DMA_TO_DEVICE);
 995
 996         req = list_entry(target->free_reqs.next, struct srp_request, list);
 997
 998         scmnd->scsi_done     = done;
 999         scmnd->result        = 0;
1000         scmnd->host_scribble = (void *) (long) req->index;
1001
1002         cmd = iu->buf;
1003         memset(cmd, 0, sizeof *cmd);
1004
1005         cmd->opcode = SRP_CMD;
1006         cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
1007         cmd->tag    = req->index;
1008         memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1009
1010         req->scmnd    = scmnd;
1011         req->cmd      = iu;
1012         req->cmd_done = 0;
1013         req->tsk_mgmt = NULL;
1014
1015         len = srp_map_data(scmnd, target, req);
1016         if (len < 0) {
1017                 printk(KERN_ERR PFX "Failed to map data\n");
1018                 goto err;
1019         }
1020
1021         if (__srp_post_recv(target)) {
1022                 printk(KERN_ERR PFX "Recv failed\n");
1023                 goto err_unmap;
1024         }
1025
1026         dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
1027                                    srp_max_iu_len, DMA_TO_DEVICE);
1028
1029         if (__srp_post_send(target, iu, len)) {
1030                 printk(KERN_ERR PFX "Send failed\n");
1031                 goto err_unmap;
1032         }
1033
1034         list_move_tail(&req->list, &target->req_queue);
1035
1036         return 0;
1037
1038 err_unmap:
1039         srp_unmap_data(scmnd, target, req);
1040
1041 err:
1042         return SCSI_MLQUEUE_HOST_BUSY;
1043 }
1044
1045 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1046 {
1047         int i;
1048
1049         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1050                 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1051                                                   target->max_ti_iu_len,
1052                                                   GFP_KERNEL, DMA_FROM_DEVICE);
1053                 if (!target->rx_ring[i])
1054                         goto err;
1055         }
1056
1057         for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1058                 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1059                                                   srp_max_iu_len,
1060                                                   GFP_KERNEL, DMA_TO_DEVICE);
1061                 if (!target->tx_ring[i])
1062                         goto err;
1063         }
1064
1065         return 0;
1066
1067 err:
1068         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1069                 srp_free_iu(target->srp_host, target->rx_ring[i]);
1070                 target->rx_ring[i] = NULL;
1071         }
1072
1073         for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1074                 srp_free_iu(target->srp_host, target->tx_ring[i]);
1075                 target->tx_ring[i] = NULL;
1076         }
1077
1078         return -ENOMEM;
1079 }
1080
1081 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1082                                struct ib_cm_event *event,
1083                                struct srp_target_port *target)
1084 {
1085         struct ib_class_port_info *cpi;
1086         int opcode;
1087
1088         switch (event->param.rej_rcvd.reason) {
1089         case IB_CM_REJ_PORT_CM_REDIRECT:
1090                 cpi = event->param.rej_rcvd.ari;
1091                 target->path.dlid = cpi->redirect_lid;
1092                 target->path.pkey = cpi->redirect_pkey;
1093                 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1094                 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1095
1096                 target->status = target->path.dlid ?
1097                         SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1098                 break;
1099
1100         case IB_CM_REJ_PORT_REDIRECT:
1101                 if (topspin_workarounds &&
1102                     !memcmp(&target->ioc_guid, topspin_oui, 3)) {
1103                         /*
1104                          * Topspin/Cisco SRP gateways incorrectly send
1105                          * reject reason code 25 when they mean 24
1106                          * (port redirect).
1107                          */
1108                         memcpy(target->path.dgid.raw,
1109                                event->param.rej_rcvd.ari, 16);
1110
1111                         printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1112                                (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1113                                (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1114
1115                         target->status = SRP_PORT_REDIRECT;
1116                 } else {
1117                         printk(KERN_WARNING "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1118                         target->status = -ECONNRESET;
1119                 }
1120                 break;
1121
1122         case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1123                 printk(KERN_WARNING "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1124                 target->status = -ECONNRESET;
1125                 break;
1126
1127         case IB_CM_REJ_CONSUMER_DEFINED:
1128                 opcode = *(u8 *) event->private_data;
1129                 if (opcode == SRP_LOGIN_REJ) {
1130                         struct srp_login_rej *rej = event->private_data;
1131                         u32 reason = be32_to_cpu(rej->reason);
1132
1133                         if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1134                                 printk(KERN_WARNING PFX
1135                                        "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1136                         else
1137                                 printk(KERN_WARNING PFX
1138                                        "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1139                 } else
1140                         printk(KERN_WARNING "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1141                                " opcode 0x%02x\n", opcode);
1142                 target->status = -ECONNRESET;
1143                 break;
1144
1145         default:
1146                 printk(KERN_WARNING "  REJ reason 0x%x\n",
1147                        event->param.rej_rcvd.reason);
1148                 target->status = -ECONNRESET;
1149         }
1150 }
1151
1152 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1153 {
1154         struct srp_target_port *target = cm_id->context;
1155         struct ib_qp_attr *qp_attr = NULL;
1156         int attr_mask = 0;
1157         int comp = 0;
1158         int opcode = 0;
1159
1160         switch (event->event) {
1161         case IB_CM_REQ_ERROR:
1162                 printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1163                 comp = 1;
1164                 target->status = -ECONNRESET;
1165                 break;
1166
1167         case IB_CM_REP_RECEIVED:
1168                 comp = 1;
1169                 opcode = *(u8 *) event->private_data;
1170
1171                 if (opcode == SRP_LOGIN_RSP) {
1172                         struct srp_login_rsp *rsp = event->private_data;
1173
1174                         target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1175                         target->req_lim       = be32_to_cpu(rsp->req_lim_delta);
1176
1177                         target->scsi_host->can_queue = min(target->req_lim,
1178                                                            target->scsi_host->can_queue);
1179                 } else {
1180                         printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1181                         target->status = -ECONNRESET;
1182                         break;
1183                 }
1184
1185                 target->status = srp_alloc_iu_bufs(target);
1186                 if (target->status)
1187                         break;
1188
1189                 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1190                 if (!qp_attr) {
1191                         target->status = -ENOMEM;
1192                         break;
1193                 }
1194
1195                 qp_attr->qp_state = IB_QPS_RTR;
1196                 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1197                 if (target->status)
1198                         break;
1199
1200                 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1201                 if (target->status)
1202                         break;
1203
1204                 target->status = srp_post_recv(target);
1205                 if (target->status)
1206                         break;
1207
1208                 qp_attr->qp_state = IB_QPS_RTS;
1209                 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1210                 if (target->status)
1211                         break;
1212
1213                 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1214                 if (target->status)
1215                         break;
1216
1217                 target->status = ib_send_cm_rtu(cm_id, NULL, 0);
1218                 if (target->status)
1219                         break;
1220
1221                 break;
1222
1223         case IB_CM_REJ_RECEIVED:
1224                 printk(KERN_DEBUG PFX "REJ received\n");
1225                 comp = 1;
1226
1227                 srp_cm_rej_handler(cm_id, event, target);
1228                 break;
1229
1230         case IB_CM_DREQ_RECEIVED:
1231                 printk(KERN_WARNING PFX "DREQ received - connection closed\n");
1232                 if (ib_send_cm_drep(cm_id, NULL, 0))
1233                         printk(KERN_ERR PFX "Sending CM DREP failed\n");
1234                 break;
1235
1236         case IB_CM_TIMEWAIT_EXIT:
1237                 printk(KERN_ERR PFX "connection closed\n");
1238
1239                 comp = 1;
1240                 target->status = 0;
1241                 break;
1242
1243         case IB_CM_MRA_RECEIVED:
1244         case IB_CM_DREQ_ERROR:
1245         case IB_CM_DREP_RECEIVED:
1246                 break;
1247
1248         default:
1249                 printk(KERN_WARNING PFX "Unhandled CM event %d\n", event->event);
1250                 break;
1251         }
1252
1253         if (comp)
1254                 complete(&target->done);
1255
1256         kfree(qp_attr);
1257
1258         return 0;
1259 }
1260
1261 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1262                              struct srp_request *req, u8 func)
1263 {
1264         struct srp_iu *iu;
1265         struct srp_tsk_mgmt *tsk_mgmt;
1266
1267         spin_lock_irq(target->scsi_host->host_lock);
1268
1269         if (target->state == SRP_TARGET_DEAD ||
1270             target->state == SRP_TARGET_REMOVED) {
1271                 req->scmnd->result = DID_BAD_TARGET << 16;
1272                 goto out;
1273         }
1274
1275         init_completion(&req->done);
1276
1277         iu = __srp_get_tx_iu(target);
1278         if (!iu)
1279                 goto out;
1280
1281         tsk_mgmt = iu->buf;
1282         memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1283
1284         tsk_mgmt->opcode        = SRP_TSK_MGMT;
1285         tsk_mgmt->lun           = cpu_to_be64((u64) req->scmnd->device->lun << 48);
1286         tsk_mgmt->tag           = req->index | SRP_TAG_TSK_MGMT;
1287         tsk_mgmt->tsk_mgmt_func = func;
1288         tsk_mgmt->task_tag      = req->index;
1289
1290         if (__srp_post_send(target, iu, sizeof *tsk_mgmt))
1291                 goto out;
1292
1293         req->tsk_mgmt = iu;
1294
1295         spin_unlock_irq(target->scsi_host->host_lock);
1296
1297         if (!wait_for_completion_timeout(&req->done,
1298                                          msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1299                 return -1;
1300
1301         return 0;
1302
1303 out:
1304         spin_unlock_irq(target->scsi_host->host_lock);
1305         return -1;
1306 }
1307
1308 static int srp_find_req(struct srp_target_port *target,
1309                         struct scsi_cmnd *scmnd,
1310                         struct srp_request **req)
1311 {
1312         if (scmnd->host_scribble == (void *) -1L)
1313                 return -1;
1314
1315         *req = &target->req_ring[(long) scmnd->host_scribble];
1316
1317         return 0;
1318 }
1319
1320 static int srp_abort(struct scsi_cmnd *scmnd)
1321 {
1322         struct srp_target_port *target = host_to_target(scmnd->device->host);
1323         struct srp_request *req;
1324         int ret = SUCCESS;
1325
1326         printk(KERN_ERR "SRP abort called\n");
1327
1328         if (srp_find_req(target, scmnd, &req))
1329                 return FAILED;
1330         if (srp_send_tsk_mgmt(target, req, SRP_TSK_ABORT_TASK))
1331                 return FAILED;
1332
1333         spin_lock_irq(target->scsi_host->host_lock);
1334
1335         if (req->cmd_done) {
1336                 srp_remove_req(target, req);
1337                 scmnd->scsi_done(scmnd);
1338         } else if (!req->tsk_status) {
1339                 srp_remove_req(target, req);
1340                 scmnd->result = DID_ABORT << 16;
1341         } else
1342                 ret = FAILED;
1343
1344         spin_unlock_irq(target->scsi_host->host_lock);
1345
1346         return ret;
1347 }
1348
1349 static int srp_reset_device(struct scsi_cmnd *scmnd)
1350 {
1351         struct srp_target_port *target = host_to_target(scmnd->device->host);
1352         struct srp_request *req, *tmp;
1353
1354         printk(KERN_ERR "SRP reset_device called\n");
1355
1356         if (srp_find_req(target, scmnd, &req))
1357                 return FAILED;
1358         if (srp_send_tsk_mgmt(target, req, SRP_TSK_LUN_RESET))
1359                 return FAILED;
1360         if (req->tsk_status)
1361                 return FAILED;
1362
1363         spin_lock_irq(target->scsi_host->host_lock);
1364
1365         list_for_each_entry_safe(req, tmp, &target->req_queue, list)
1366                 if (req->scmnd->device == scmnd->device)
1367                         srp_reset_req(target, req);
1368
1369         spin_unlock_irq(target->scsi_host->host_lock);
1370
1371         return SUCCESS;
1372 }
1373
1374 static int srp_reset_host(struct scsi_cmnd *scmnd)
1375 {
1376         struct srp_target_port *target = host_to_target(scmnd->device->host);
1377         int ret = FAILED;
1378
1379         printk(KERN_ERR PFX "SRP reset_host called\n");
1380
1381         if (!srp_reconnect_target(target))
1382                 ret = SUCCESS;
1383
1384         return ret;
1385 }
1386
1387 static ssize_t show_id_ext(struct class_device *cdev, char *buf)
1388 {
1389         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1390
1391         if (target->state == SRP_TARGET_DEAD ||
1392             target->state == SRP_TARGET_REMOVED)
1393                 return -ENODEV;
1394
1395         return sprintf(buf, "0x%016llx\n",
1396                        (unsigned long long) be64_to_cpu(target->id_ext));
1397 }
1398
1399 static ssize_t show_ioc_guid(struct class_device *cdev, char *buf)
1400 {
1401         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1402
1403         if (target->state == SRP_TARGET_DEAD ||
1404             target->state == SRP_TARGET_REMOVED)
1405                 return -ENODEV;
1406
1407         return sprintf(buf, "0x%016llx\n",
1408                        (unsigned long long) be64_to_cpu(target->ioc_guid));
1409 }
1410
1411 static ssize_t show_service_id(struct class_device *cdev, char *buf)
1412 {
1413         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1414
1415         if (target->state == SRP_TARGET_DEAD ||
1416             target->state == SRP_TARGET_REMOVED)
1417                 return -ENODEV;
1418
1419         return sprintf(buf, "0x%016llx\n",
1420                        (unsigned long long) be64_to_cpu(target->service_id));
1421 }
1422
1423 static ssize_t show_pkey(struct class_device *cdev, char *buf)
1424 {
1425         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1426
1427         if (target->state == SRP_TARGET_DEAD ||
1428             target->state == SRP_TARGET_REMOVED)
1429                 return -ENODEV;
1430
1431         return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1432 }
1433
1434 static ssize_t show_dgid(struct class_device *cdev, char *buf)
1435 {
1436         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1437
1438         if (target->state == SRP_TARGET_DEAD ||
1439             target->state == SRP_TARGET_REMOVED)
1440                 return -ENODEV;
1441
1442         return sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1443                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[0]),
1444                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[1]),
1445                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[2]),
1446                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[3]),
1447                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[4]),
1448                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[5]),
1449                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[6]),
1450                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[7]));
1451 }
1452
1453 static ssize_t show_zero_req_lim(struct class_device *cdev, char *buf)
1454 {
1455         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1456
1457         if (target->state == SRP_TARGET_DEAD ||
1458             target->state == SRP_TARGET_REMOVED)
1459                 return -ENODEV;
1460
1461         return sprintf(buf, "%d\n", target->zero_req_lim);
1462 }
1463
1464 static CLASS_DEVICE_ATTR(id_ext,        S_IRUGO, show_id_ext,           NULL);
1465 static CLASS_DEVICE_ATTR(ioc_guid,      S_IRUGO, show_ioc_guid,         NULL);
1466 static CLASS_DEVICE_ATTR(service_id,    S_IRUGO, show_service_id,       NULL);
1467 static CLASS_DEVICE_ATTR(pkey,          S_IRUGO, show_pkey,             NULL);
1468 static CLASS_DEVICE_ATTR(dgid,          S_IRUGO, show_dgid,             NULL);
1469 static CLASS_DEVICE_ATTR(zero_req_lim,  S_IRUGO, show_zero_req_lim,     NULL);
1470
1471 static struct class_device_attribute *srp_host_attrs[] = {
1472         &class_device_attr_id_ext,
1473         &class_device_attr_ioc_guid,
1474         &class_device_attr_service_id,
1475         &class_device_attr_pkey,
1476         &class_device_attr_dgid,
1477         &class_device_attr_zero_req_lim,
1478         NULL
1479 };
1480
1481 static struct scsi_host_template srp_template = {
1482         .module                         = THIS_MODULE,
1483         .name                           = DRV_NAME,
1484         .info                           = srp_target_info,
1485         .queuecommand                   = srp_queuecommand,
1486         .eh_abort_handler               = srp_abort,
1487         .eh_device_reset_handler        = srp_reset_device,
1488         .eh_host_reset_handler          = srp_reset_host,
1489         .can_queue                      = SRP_SQ_SIZE,
1490         .this_id                        = -1,
1491         .cmd_per_lun                    = SRP_SQ_SIZE,
1492         .use_clustering                 = ENABLE_CLUSTERING,
1493         .shost_attrs                    = srp_host_attrs
1494 };
1495
1496 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1497 {
1498         sprintf(target->target_name, "SRP.T10:%016llX",
1499                  (unsigned long long) be64_to_cpu(target->id_ext));
1500
1501         if (scsi_add_host(target->scsi_host, host->dev->dev->dma_device))
1502                 return -ENODEV;
1503
1504         spin_lock(&host->target_lock);
1505         list_add_tail(&target->list, &host->target_list);
1506         spin_unlock(&host->target_lock);
1507
1508         target->state = SRP_TARGET_LIVE;
1509
1510         scsi_scan_target(&target->scsi_host->shost_gendev,
1511                          0, target->scsi_id, SCAN_WILD_CARD, 0);
1512
1513         return 0;
1514 }
1515
1516 static void srp_release_class_dev(struct class_device *class_dev)
1517 {
1518         struct srp_host *host =
1519                 container_of(class_dev, struct srp_host, class_dev);
1520
1521         complete(&host->released);
1522 }
1523
1524 static struct class srp_class = {
1525         .name    = "infiniband_srp",
1526         .release = srp_release_class_dev
1527 };
1528
1529 /*
1530  * Target ports are added by writing
1531  *
1532  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1533  *     pkey=<P_Key>,service_id=<service ID>
1534  *
1535  * to the add_target sysfs attribute.
1536  */
1537 enum {
1538         SRP_OPT_ERR             = 0,
1539         SRP_OPT_ID_EXT          = 1 << 0,
1540         SRP_OPT_IOC_GUID        = 1 << 1,
1541         SRP_OPT_DGID            = 1 << 2,
1542         SRP_OPT_PKEY            = 1 << 3,
1543         SRP_OPT_SERVICE_ID      = 1 << 4,
1544         SRP_OPT_MAX_SECT        = 1 << 5,
1545         SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
1546         SRP_OPT_IO_CLASS        = 1 << 7,
1547         SRP_OPT_ALL             = (SRP_OPT_ID_EXT       |
1548                                    SRP_OPT_IOC_GUID     |
1549                                    SRP_OPT_DGID         |
1550                                    SRP_OPT_PKEY         |
1551                                    SRP_OPT_SERVICE_ID),
1552 };
1553
1554 static match_table_t srp_opt_tokens = {
1555         { SRP_OPT_ID_EXT,               "id_ext=%s"             },
1556         { SRP_OPT_IOC_GUID,             "ioc_guid=%s"           },
1557         { SRP_OPT_DGID,                 "dgid=%s"               },
1558         { SRP_OPT_PKEY,                 "pkey=%x"               },
1559         { SRP_OPT_SERVICE_ID,           "service_id=%s"         },
1560         { SRP_OPT_MAX_SECT,             "max_sect=%d"           },
1561         { SRP_OPT_MAX_CMD_PER_LUN,      "max_cmd_per_lun=%d"    },
1562         { SRP_OPT_IO_CLASS,             "io_class=%x"           },
1563         { SRP_OPT_ERR,                  NULL                    }
1564 };
1565
1566 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1567 {
1568         char *options, *sep_opt;
1569         char *p;
1570         char dgid[3];
1571         substring_t args[MAX_OPT_ARGS];
1572         int opt_mask = 0;
1573         int token;
1574         int ret = -EINVAL;
1575         int i;
1576
1577         options = kstrdup(buf, GFP_KERNEL);
1578         if (!options)
1579                 return -ENOMEM;
1580
1581         sep_opt = options;
1582         while ((p = strsep(&sep_opt, ",")) != NULL) {
1583                 if (!*p)
1584                         continue;
1585
1586                 token = match_token(p, srp_opt_tokens, args);
1587                 opt_mask |= token;
1588
1589                 switch (token) {
1590                 case SRP_OPT_ID_EXT:
1591                         p = match_strdup(args);
1592                         target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1593                         kfree(p);
1594                         break;
1595
1596                 case SRP_OPT_IOC_GUID:
1597                         p = match_strdup(args);
1598                         target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1599                         kfree(p);
1600                         break;
1601
1602                 case SRP_OPT_DGID:
1603                         p = match_strdup(args);
1604                         if (strlen(p) != 32) {
1605                                 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1606                                 kfree(p);
1607                                 goto out;
1608                         }
1609
1610                         for (i = 0; i < 16; ++i) {
1611                                 strlcpy(dgid, p + i * 2, 3);
1612                                 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1613                         }
1614                         kfree(p);
1615                         break;
1616
1617                 case SRP_OPT_PKEY:
1618                         if (match_hex(args, &token)) {
1619                                 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1620                                 goto out;
1621                         }
1622                         target->path.pkey = cpu_to_be16(token);
1623                         break;
1624
1625                 case SRP_OPT_SERVICE_ID:
1626                         p = match_strdup(args);
1627                         target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1628                         kfree(p);
1629                         break;
1630
1631                 case SRP_OPT_MAX_SECT:
1632                         if (match_int(args, &token)) {
1633                                 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1634                                 goto out;
1635                         }
1636                         target->scsi_host->max_sectors = token;
1637                         break;
1638
1639                 case SRP_OPT_MAX_CMD_PER_LUN:
1640                         if (match_int(args, &token)) {
1641                                 printk(KERN_WARNING PFX "bad max cmd_per_lun parameter '%s'\n", p);
1642                                 goto out;
1643                         }
1644                         target->scsi_host->cmd_per_lun = min(token, SRP_SQ_SIZE);
1645                         break;
1646
1647                 case SRP_OPT_IO_CLASS:
1648                         if (match_hex(args, &token)) {
1649                                 printk(KERN_WARNING PFX "bad  IO class parameter '%s' \n", p);
1650                                 goto out;
1651                         }
1652                         if (token != SRP_REV10_IB_IO_CLASS &&
1653                             token != SRP_REV16A_IB_IO_CLASS) {
1654                                 printk(KERN_WARNING PFX "unknown IO class parameter value"
1655                                        " %x specified (use %x or %x).\n",
1656                                        token, SRP_REV10_IB_IO_CLASS, SRP_REV16A_IB_IO_CLASS);
1657                                 goto out;
1658                         }
1659                         target->io_class = token;
1660                         break;
1661
1662                 default:
1663                         printk(KERN_WARNING PFX "unknown parameter or missing value "
1664                                "'%s' in target creation request\n", p);
1665                         goto out;
1666                 }
1667         }
1668
1669         if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1670                 ret = 0;
1671         else
1672                 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1673                         if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1674                             !(srp_opt_tokens[i].token & opt_mask))
1675                                 printk(KERN_WARNING PFX "target creation request is "
1676                                        "missing parameter '%s'\n",
1677                                        srp_opt_tokens[i].pattern);
1678
1679 out:
1680         kfree(options);
1681         return ret;
1682 }
1683
1684 static ssize_t srp_create_target(struct class_device *class_dev,
1685                                  const char *buf, size_t count)
1686 {
1687         struct srp_host *host =
1688                 container_of(class_dev, struct srp_host, class_dev);
1689         struct Scsi_Host *target_host;
1690         struct srp_target_port *target;
1691         int ret;
1692         int i;
1693
1694         target_host = scsi_host_alloc(&srp_template,
1695                                       sizeof (struct srp_target_port));
1696         if (!target_host)
1697                 return -ENOMEM;
1698
1699         target_host->max_lun = SRP_MAX_LUN;
1700
1701         target = host_to_target(target_host);
1702         memset(target, 0, sizeof *target);
1703
1704         target->io_class   = SRP_REV16A_IB_IO_CLASS;
1705         target->scsi_host  = target_host;
1706         target->srp_host   = host;
1707
1708         INIT_WORK(&target->work, srp_reconnect_work, target);
1709
1710         INIT_LIST_HEAD(&target->free_reqs);
1711         INIT_LIST_HEAD(&target->req_queue);
1712         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1713                 target->req_ring[i].index = i;
1714                 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
1715         }
1716
1717         ret = srp_parse_options(buf, target);
1718         if (ret)
1719                 goto err;
1720
1721         ib_get_cached_gid(host->dev->dev, host->port, 0, &target->path.sgid);
1722
1723         printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1724                "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1725                (unsigned long long) be64_to_cpu(target->id_ext),
1726                (unsigned long long) be64_to_cpu(target->ioc_guid),
1727                be16_to_cpu(target->path.pkey),
1728                (unsigned long long) be64_to_cpu(target->service_id),
1729                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1730                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1731                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1732                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1733                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1734                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1735                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1736                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1737
1738         ret = srp_create_target_ib(target);
1739         if (ret)
1740                 goto err;
1741
1742         target->cm_id = ib_create_cm_id(host->dev->dev, srp_cm_handler, target);
1743         if (IS_ERR(target->cm_id)) {
1744                 ret = PTR_ERR(target->cm_id);
1745                 goto err_free;
1746         }
1747
1748         ret = srp_connect_target(target);
1749         if (ret) {
1750                 printk(KERN_ERR PFX "Connection failed\n");
1751                 goto err_cm_id;
1752         }
1753
1754         ret = srp_add_target(host, target);
1755         if (ret)
1756                 goto err_disconnect;
1757
1758         return count;
1759
1760 err_disconnect:
1761         srp_disconnect_target(target);
1762
1763 err_cm_id:
1764         ib_destroy_cm_id(target->cm_id);
1765
1766 err_free:
1767         srp_free_target_ib(target);
1768
1769 err:
1770         scsi_host_put(target_host);
1771
1772         return ret;
1773 }
1774
1775 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1776
1777 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1778 {
1779         struct srp_host *host =
1780                 container_of(class_dev, struct srp_host, class_dev);
1781
1782         return sprintf(buf, "%s\n", host->dev->dev->name);
1783 }
1784
1785 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1786
1787 static ssize_t show_port(struct class_device *class_dev, char *buf)
1788 {
1789         struct srp_host *host =
1790                 container_of(class_dev, struct srp_host, class_dev);
1791
1792         return sprintf(buf, "%d\n", host->port);
1793 }
1794
1795 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1796
1797 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
1798 {
1799         struct srp_host *host;
1800
1801         host = kzalloc(sizeof *host, GFP_KERNEL);
1802         if (!host)
1803                 return NULL;
1804
1805         INIT_LIST_HEAD(&host->target_list);
1806         spin_lock_init(&host->target_lock);
1807         init_completion(&host->released);
1808         host->dev  = device;
1809         host->port = port;
1810
1811         host->initiator_port_id[7] = port;
1812         memcpy(host->initiator_port_id + 8, &device->dev->node_guid, 8);
1813
1814         host->class_dev.class = &srp_class;
1815         host->class_dev.dev   = device->dev->dma_device;
1816         snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1817                  device->dev->name, port);
1818
1819         if (class_device_register(&host->class_dev))
1820                 goto free_host;
1821         if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1822                 goto err_class;
1823         if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1824                 goto err_class;
1825         if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1826                 goto err_class;
1827
1828         return host;
1829
1830 err_class:
1831         class_device_unregister(&host->class_dev);
1832
1833 free_host:
1834         kfree(host);
1835
1836         return NULL;
1837 }
1838
1839 static void srp_add_one(struct ib_device *device)
1840 {
1841         struct srp_device *srp_dev;
1842         struct ib_device_attr *dev_attr;
1843         struct ib_fmr_pool_param fmr_param;
1844         struct srp_host *host;
1845         int s, e, p;
1846
1847         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
1848         if (!dev_attr)
1849                 return;
1850
1851         if (ib_query_device(device, dev_attr)) {
1852                 printk(KERN_WARNING PFX "Query device failed for %s\n",
1853                        device->name);
1854                 goto free_attr;
1855         }
1856
1857         srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
1858         if (!srp_dev)
1859                 goto free_attr;
1860
1861         /*
1862          * Use the smallest page size supported by the HCA, down to a
1863          * minimum of 512 bytes (which is the smallest sector that a
1864          * SCSI command will ever carry).
1865          */
1866         srp_dev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
1867         srp_dev->fmr_page_size  = 1 << srp_dev->fmr_page_shift;
1868         srp_dev->fmr_page_mask  = ~((unsigned long) srp_dev->fmr_page_size - 1);
1869
1870         INIT_LIST_HEAD(&srp_dev->dev_list);
1871
1872         srp_dev->dev = device;
1873         srp_dev->pd  = ib_alloc_pd(device);
1874         if (IS_ERR(srp_dev->pd))
1875                 goto free_dev;
1876
1877         srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
1878                                     IB_ACCESS_LOCAL_WRITE |
1879                                     IB_ACCESS_REMOTE_READ |
1880                                     IB_ACCESS_REMOTE_WRITE);
1881         if (IS_ERR(srp_dev->mr))
1882                 goto err_pd;
1883
1884         memset(&fmr_param, 0, sizeof fmr_param);
1885         fmr_param.pool_size         = SRP_FMR_POOL_SIZE;
1886         fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
1887         fmr_param.cache             = 1;
1888         fmr_param.max_pages_per_fmr = SRP_FMR_SIZE;
1889         fmr_param.page_shift        = srp_dev->fmr_page_shift;
1890         fmr_param.access            = (IB_ACCESS_LOCAL_WRITE |
1891                                        IB_ACCESS_REMOTE_WRITE |
1892                                        IB_ACCESS_REMOTE_READ);
1893
1894         srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
1895         if (IS_ERR(srp_dev->fmr_pool))
1896                 srp_dev->fmr_pool = NULL;
1897
1898         if (device->node_type == IB_NODE_SWITCH) {
1899                 s = 0;
1900                 e = 0;
1901         } else {
1902                 s = 1;
1903                 e = device->phys_port_cnt;
1904         }
1905
1906         for (p = s; p <= e; ++p) {
1907                 host = srp_add_port(srp_dev, p);
1908                 if (host)
1909                         list_add_tail(&host->list, &srp_dev->dev_list);
1910         }
1911
1912         ib_set_client_data(device, &srp_client, srp_dev);
1913
1914         goto free_attr;
1915
1916 err_pd:
1917         ib_dealloc_pd(srp_dev->pd);
1918
1919 free_dev:
1920         kfree(srp_dev);
1921
1922 free_attr:
1923         kfree(dev_attr);
1924 }
1925
1926 static void srp_remove_one(struct ib_device *device)
1927 {
1928         struct srp_device *srp_dev;
1929         struct srp_host *host, *tmp_host;
1930         LIST_HEAD(target_list);
1931         struct srp_target_port *target, *tmp_target;
1932
1933         srp_dev = ib_get_client_data(device, &srp_client);
1934
1935         list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
1936                 class_device_unregister(&host->class_dev);
1937                 /*
1938                  * Wait for the sysfs entry to go away, so that no new
1939                  * target ports can be created.
1940                  */
1941                 wait_for_completion(&host->released);
1942
1943                 /*
1944                  * Mark all target ports as removed, so we stop queueing
1945                  * commands and don't try to reconnect.
1946                  */
1947                 spin_lock(&host->target_lock);
1948                 list_for_each_entry(target, &host->target_list, list) {
1949                         spin_lock_irq(target->scsi_host->host_lock);
1950                         target->state = SRP_TARGET_REMOVED;
1951                         spin_unlock_irq(target->scsi_host->host_lock);
1952                 }
1953                 spin_unlock(&host->target_lock);
1954
1955                 /*
1956                  * Wait for any reconnection tasks that may have
1957                  * started before we marked our target ports as
1958                  * removed, and any target port removal tasks.
1959                  */
1960                 flush_scheduled_work();
1961
1962                 list_for_each_entry_safe(target, tmp_target,
1963                                          &host->target_list, list) {
1964                         scsi_remove_host(target->scsi_host);
1965                         srp_disconnect_target(target);
1966                         ib_destroy_cm_id(target->cm_id);
1967                         srp_free_target_ib(target);
1968                         scsi_host_put(target->scsi_host);
1969                 }
1970
1971                 kfree(host);
1972         }
1973
1974         if (srp_dev->fmr_pool)
1975                 ib_destroy_fmr_pool(srp_dev->fmr_pool);
1976         ib_dereg_mr(srp_dev->mr);
1977         ib_dealloc_pd(srp_dev->pd);
1978
1979         kfree(srp_dev);
1980 }
1981
1982 static int __init srp_init_module(void)
1983 {
1984         int ret;
1985
1986         srp_template.sg_tablesize = srp_sg_tablesize;
1987         srp_max_iu_len = (sizeof (struct srp_cmd) +
1988                           sizeof (struct srp_indirect_buf) +
1989                           srp_sg_tablesize * 16);
1990
1991         ret = class_register(&srp_class);
1992         if (ret) {
1993                 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
1994                 return ret;
1995         }
1996
1997         ret = ib_register_client(&srp_client);
1998         if (ret) {
1999                 printk(KERN_ERR PFX "couldn't register IB client\n");
2000                 class_unregister(&srp_class);
2001                 return ret;
2002         }
2003
2004         return 0;
2005 }
2006
2007 static void __exit srp_cleanup_module(void)
2008 {
2009         ib_unregister_client(&srp_client);
2010         class_unregister(&srp_class);
2011 }
2012
2013 module_init(srp_init_module);
2014 module_exit(srp_cleanup_module);