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) KBUILD_MODNAME ": " 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 #include <rdma/ib_cache.h>
  44
  45 #include <linux/atomic.h>
  46
  47 #include <scsi/scsi.h>
  48 #include <scsi/scsi_device.h>
  49 #include <scsi/scsi_dbg.h>
  50 #include <scsi/scsi_tcq.h>
  51 #include <scsi/srp.h>
  52 #include <scsi/scsi_transport_srp.h>
  53
  54 #include "ib_srp.h"
  55
  56 #define DRV_NAME        "ib_srp"
  57 #define PFX             DRV_NAME ": "
  58 #define DRV_VERSION     "2.0"
  59 #define DRV_RELDATE     "July 26, 2015"
  60
  61 MODULE_AUTHOR("Roland Dreier");
  62 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator");
  63 MODULE_LICENSE("Dual BSD/GPL");
  64 MODULE_VERSION(DRV_VERSION);
  65 MODULE_INFO(release_date, DRV_RELDATE);
  66
  67 static unsigned int srp_sg_tablesize;
  68 static unsigned int cmd_sg_entries;
  69 static unsigned int indirect_sg_entries;
  70 static bool allow_ext_sg;
  71 static bool prefer_fr = true;
  72 static bool register_always = true;
  73 static int topspin_workarounds = 1;
  74
  75 module_param(srp_sg_tablesize, uint, 0444);
  76 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
  77
  78 module_param(cmd_sg_entries, uint, 0444);
  79 MODULE_PARM_DESC(cmd_sg_entries,
  80                  "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
  81
  82 module_param(indirect_sg_entries, uint, 0444);
  83 MODULE_PARM_DESC(indirect_sg_entries,
  84                  "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
  85
  86 module_param(allow_ext_sg, bool, 0444);
  87 MODULE_PARM_DESC(allow_ext_sg,
  88                   "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
  89
  90 module_param(topspin_workarounds, int, 0444);
  91 MODULE_PARM_DESC(topspin_workarounds,
  92                  "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
  93
  94 module_param(prefer_fr, bool, 0444);
  95 MODULE_PARM_DESC(prefer_fr,
  96 "Whether to use fast registration if both FMR and fast registration are supported");
  97
  98 module_param(register_always, bool, 0444);
  99 MODULE_PARM_DESC(register_always,
 100                  "Use memory registration even for contiguous memory regions");
 101
 102 static const struct kernel_param_ops srp_tmo_ops;
 103
 104 static int srp_reconnect_delay = 10;
 105 module_param_cb(reconnect_delay, &srp_tmo_ops, &srp_reconnect_delay,
 106                 S_IRUGO | S_IWUSR);
 107 MODULE_PARM_DESC(reconnect_delay, "Time between successive reconnect attempts");
 108
 109 static int srp_fast_io_fail_tmo = 15;
 110 module_param_cb(fast_io_fail_tmo, &srp_tmo_ops, &srp_fast_io_fail_tmo,
 111                 S_IRUGO | S_IWUSR);
 112 MODULE_PARM_DESC(fast_io_fail_tmo,
 113                  "Number of seconds between the observation of a transport"
 114                  " layer error and failing all I/O. \"off\" means that this"
 115                  " functionality is disabled.");
 116
 117 static int srp_dev_loss_tmo = 600;
 118 module_param_cb(dev_loss_tmo, &srp_tmo_ops, &srp_dev_loss_tmo,
 119                 S_IRUGO | S_IWUSR);
 120 MODULE_PARM_DESC(dev_loss_tmo,
 121                  "Maximum number of seconds that the SRP transport should"
 122                  " insulate transport layer errors. After this time has been"
 123                  " exceeded the SCSI host is removed. Should be"
 124                  " between 1 and " __stringify(SCSI_DEVICE_BLOCK_MAX_TIMEOUT)
 125                  " if fast_io_fail_tmo has not been set. \"off\" means that"
 126                  " this functionality is disabled.");
 127
 128 static unsigned ch_count;
 129 module_param(ch_count, uint, 0444);
 130 MODULE_PARM_DESC(ch_count,
 131                  "Number of RDMA channels to use for communication with an SRP target. Using more than one channel improves performance if the HCA supports multiple completion vectors. The default value is the minimum of four times the number of online CPU sockets and the number of completion vectors supported by the HCA.");
 132
 133 static void srp_add_one(struct ib_device *device);
 134 static void srp_remove_one(struct ib_device *device, void *client_data);
 135 static void srp_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 136 static void srp_handle_qp_err(struct ib_cq *cq, struct ib_wc *wc,
 137                 const char *opname);
 138 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
 139
 140 static struct scsi_transport_template *ib_srp_transport_template;
 141 static struct workqueue_struct *srp_remove_wq;
 142
 143 static struct ib_client srp_client = {
 144         .name   = "srp",
 145         .add    = srp_add_one,
 146         .remove = srp_remove_one
 147 };
 148
 149 static struct ib_sa_client srp_sa_client;
 150
 151 static int srp_tmo_get(char *buffer, const struct kernel_param *kp)
 152 {
 153         int tmo = *(int *)kp->arg;
 154
 155         if (tmo >= 0)
 156                 return sprintf(buffer, "%d", tmo);
 157         else
 158                 return sprintf(buffer, "off");
 159 }
 160
 161 static int srp_tmo_set(const char *val, const struct kernel_param *kp)
 162 {
 163         int tmo, res;
 164
 165         res = srp_parse_tmo(&tmo, val);
 166         if (res)
 167                 goto out;
 168
 169         if (kp->arg == &srp_reconnect_delay)
 170                 res = srp_tmo_valid(tmo, srp_fast_io_fail_tmo,
 171                                     srp_dev_loss_tmo);
 172         else if (kp->arg == &srp_fast_io_fail_tmo)
 173                 res = srp_tmo_valid(srp_reconnect_delay, tmo, srp_dev_loss_tmo);
 174         else
 175                 res = srp_tmo_valid(srp_reconnect_delay, srp_fast_io_fail_tmo,
 176                                     tmo);
 177         if (res)
 178                 goto out;
 179         *(int *)kp->arg = tmo;
 180
 181 out:
 182         return res;
 183 }
 184
 185 static const struct kernel_param_ops srp_tmo_ops = {
 186         .get = srp_tmo_get,
 187         .set = srp_tmo_set,
 188 };
 189
 190 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
 191 {
 192         return (struct srp_target_port *) host->hostdata;
 193 }
 194
 195 static const char *srp_target_info(struct Scsi_Host *host)
 196 {
 197         return host_to_target(host)->target_name;
 198 }
 199
 200 static int srp_target_is_topspin(struct srp_target_port *target)
 201 {
 202         static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
 203         static const u8 cisco_oui[3]   = { 0x00, 0x1b, 0x0d };
 204
 205         return topspin_workarounds &&
 206                 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
 207                  !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
 208 }
 209
 210 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
 211                                    gfp_t gfp_mask,
 212                                    enum dma_data_direction direction)
 213 {
 214         struct srp_iu *iu;
 215
 216         iu = kmalloc(sizeof *iu, gfp_mask);
 217         if (!iu)
 218                 goto out;
 219
 220         iu->buf = kzalloc(size, gfp_mask);
 221         if (!iu->buf)
 222                 goto out_free_iu;
 223
 224         iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
 225                                     direction);
 226         if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
 227                 goto out_free_buf;
 228
 229         iu->size      = size;
 230         iu->direction = direction;
 231
 232         return iu;
 233
 234 out_free_buf:
 235         kfree(iu->buf);
 236 out_free_iu:
 237         kfree(iu);
 238 out:
 239         return NULL;
 240 }
 241
 242 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
 243 {
 244         if (!iu)
 245                 return;
 246
 247         ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
 248                             iu->direction);
 249         kfree(iu->buf);
 250         kfree(iu);
 251 }
 252
 253 static void srp_qp_event(struct ib_event *event, void *context)
 254 {
 255         pr_debug("QP event %s (%d)\n",
 256                  ib_event_msg(event->event), event->event);
 257 }
 258
 259 static int srp_init_qp(struct srp_target_port *target,
 260                        struct ib_qp *qp)
 261 {
 262         struct ib_qp_attr *attr;
 263         int ret;
 264
 265         attr = kmalloc(sizeof *attr, GFP_KERNEL);
 266         if (!attr)
 267                 return -ENOMEM;
 268
 269         ret = ib_find_cached_pkey(target->srp_host->srp_dev->dev,
 270                                   target->srp_host->port,
 271                                   be16_to_cpu(target->pkey),
 272                                   &attr->pkey_index);
 273         if (ret)
 274                 goto out;
 275
 276         attr->qp_state        = IB_QPS_INIT;
 277         attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
 278                                     IB_ACCESS_REMOTE_WRITE);
 279         attr->port_num        = target->srp_host->port;
 280
 281         ret = ib_modify_qp(qp, attr,
 282                            IB_QP_STATE          |
 283                            IB_QP_PKEY_INDEX     |
 284                            IB_QP_ACCESS_FLAGS   |
 285                            IB_QP_PORT);
 286
 287 out:
 288         kfree(attr);
 289         return ret;
 290 }
 291
 292 static int srp_new_cm_id(struct srp_rdma_ch *ch)
 293 {
 294         struct srp_target_port *target = ch->target;
 295         struct ib_cm_id *new_cm_id;
 296
 297         new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
 298                                     srp_cm_handler, ch);
 299         if (IS_ERR(new_cm_id))
 300                 return PTR_ERR(new_cm_id);
 301
 302         if (ch->cm_id)
 303                 ib_destroy_cm_id(ch->cm_id);
 304         ch->cm_id = new_cm_id;
 305         ch->path.sgid = target->sgid;
 306         ch->path.dgid = target->orig_dgid;
 307         ch->path.pkey = target->pkey;
 308         ch->path.service_id = target->service_id;
 309
 310         return 0;
 311 }
 312
 313 static struct ib_fmr_pool *srp_alloc_fmr_pool(struct srp_target_port *target)
 314 {
 315         struct srp_device *dev = target->srp_host->srp_dev;
 316         struct ib_fmr_pool_param fmr_param;
 317
 318         memset(&fmr_param, 0, sizeof(fmr_param));
 319         fmr_param.pool_size         = target->mr_pool_size;
 320         fmr_param.dirty_watermark   = fmr_param.pool_size / 4;
 321         fmr_param.cache             = 1;
 322         fmr_param.max_pages_per_fmr = dev->max_pages_per_mr;
 323         fmr_param.page_shift        = ilog2(dev->mr_page_size);
 324         fmr_param.access            = (IB_ACCESS_LOCAL_WRITE |
 325                                        IB_ACCESS_REMOTE_WRITE |
 326                                        IB_ACCESS_REMOTE_READ);
 327
 328         return ib_create_fmr_pool(dev->pd, &fmr_param);
 329 }
 330
 331 /**
 332  * srp_destroy_fr_pool() - free the resources owned by a pool
 333  * @pool: Fast registration pool to be destroyed.
 334  */
 335 static void srp_destroy_fr_pool(struct srp_fr_pool *pool)
 336 {
 337         int i;
 338         struct srp_fr_desc *d;
 339
 340         if (!pool)
 341                 return;
 342
 343         for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
 344                 if (d->mr)
 345                         ib_dereg_mr(d->mr);
 346         }
 347         kfree(pool);
 348 }
 349
 350 /**
 351  * srp_create_fr_pool() - allocate and initialize a pool for fast registration
 352  * @device:            IB device to allocate fast registration descriptors for.
 353  * @pd:                Protection domain associated with the FR descriptors.
 354  * @pool_size:         Number of descriptors to allocate.
 355  * @max_page_list_len: Maximum fast registration work request page list length.
 356  */
 357 static struct srp_fr_pool *srp_create_fr_pool(struct ib_device *device,
 358                                               struct ib_pd *pd, int pool_size,
 359                                               int max_page_list_len)
 360 {
 361         struct srp_fr_pool *pool;
 362         struct srp_fr_desc *d;
 363         struct ib_mr *mr;
 364         int i, ret = -EINVAL;
 365
 366         if (pool_size <= 0)
 367                 goto err;
 368         ret = -ENOMEM;
 369         pool = kzalloc(sizeof(struct srp_fr_pool) +
 370                        pool_size * sizeof(struct srp_fr_desc), GFP_KERNEL);
 371         if (!pool)
 372                 goto err;
 373         pool->size = pool_size;
 374         pool->max_page_list_len = max_page_list_len;
 375         spin_lock_init(&pool->lock);
 376         INIT_LIST_HEAD(&pool->free_list);
 377
 378         for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
 379                 mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 380                                  max_page_list_len);
 381                 if (IS_ERR(mr)) {
 382                         ret = PTR_ERR(mr);
 383                         goto destroy_pool;
 384                 }
 385                 d->mr = mr;
 386                 list_add_tail(&d->entry, &pool->free_list);
 387         }
 388
 389 out:
 390         return pool;
 391
 392 destroy_pool:
 393         srp_destroy_fr_pool(pool);
 394
 395 err:
 396         pool = ERR_PTR(ret);
 397         goto out;
 398 }
 399
 400 /**
 401  * srp_fr_pool_get() - obtain a descriptor suitable for fast registration
 402  * @pool: Pool to obtain descriptor from.
 403  */
 404 static struct srp_fr_desc *srp_fr_pool_get(struct srp_fr_pool *pool)
 405 {
 406         struct srp_fr_desc *d = NULL;
 407         unsigned long flags;
 408
 409         spin_lock_irqsave(&pool->lock, flags);
 410         if (!list_empty(&pool->free_list)) {
 411                 d = list_first_entry(&pool->free_list, typeof(*d), entry);
 412                 list_del(&d->entry);
 413         }
 414         spin_unlock_irqrestore(&pool->lock, flags);
 415
 416         return d;
 417 }
 418
 419 /**
 420  * srp_fr_pool_put() - put an FR descriptor back in the free list
 421  * @pool: Pool the descriptor was allocated from.
 422  * @desc: Pointer to an array of fast registration descriptor pointers.
 423  * @n:    Number of descriptors to put back.
 424  *
 425  * Note: The caller must already have queued an invalidation request for
 426  * desc->mr->rkey before calling this function.
 427  */
 428 static void srp_fr_pool_put(struct srp_fr_pool *pool, struct srp_fr_desc **desc,
 429                             int n)
 430 {
 431         unsigned long flags;
 432         int i;
 433
 434         spin_lock_irqsave(&pool->lock, flags);
 435         for (i = 0; i < n; i++)
 436                 list_add(&desc[i]->entry, &pool->free_list);
 437         spin_unlock_irqrestore(&pool->lock, flags);
 438 }
 439
 440 static struct srp_fr_pool *srp_alloc_fr_pool(struct srp_target_port *target)
 441 {
 442         struct srp_device *dev = target->srp_host->srp_dev;
 443
 444         return srp_create_fr_pool(dev->dev, dev->pd, target->mr_pool_size,
 445                                   dev->max_pages_per_mr);
 446 }
 447
 448 /**
 449  * srp_destroy_qp() - destroy an RDMA queue pair
 450  * @ch: SRP RDMA channel.
 451  *
 452  * Drain the qp before destroying it.  This avoids that the receive
 453  * completion handler can access the queue pair while it is
 454  * being destroyed.
 455  */
 456 static void srp_destroy_qp(struct srp_rdma_ch *ch)
 457 {
 458         ib_drain_rq(ch->qp);
 459         ib_destroy_qp(ch->qp);
 460 }
 461
 462 static int srp_create_ch_ib(struct srp_rdma_ch *ch)
 463 {
 464         struct srp_target_port *target = ch->target;
 465         struct srp_device *dev = target->srp_host->srp_dev;
 466         struct ib_qp_init_attr *init_attr;
 467         struct ib_cq *recv_cq, *send_cq;
 468         struct ib_qp *qp;
 469         struct ib_fmr_pool *fmr_pool = NULL;
 470         struct srp_fr_pool *fr_pool = NULL;
 471         const int m = dev->use_fast_reg ? 3 : 1;
 472         int ret;
 473
 474         init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
 475         if (!init_attr)
 476                 return -ENOMEM;
 477
 478         /* queue_size + 1 for ib_drain_rq() */
 479         recv_cq = ib_alloc_cq(dev->dev, ch, target->queue_size + 1,
 480                                 ch->comp_vector, IB_POLL_SOFTIRQ);
 481         if (IS_ERR(recv_cq)) {
 482                 ret = PTR_ERR(recv_cq);
 483                 goto err;
 484         }
 485
 486         send_cq = ib_alloc_cq(dev->dev, ch, m * target->queue_size,
 487                                 ch->comp_vector, IB_POLL_DIRECT);
 488         if (IS_ERR(send_cq)) {
 489                 ret = PTR_ERR(send_cq);
 490                 goto err_recv_cq;
 491         }
 492
 493         init_attr->event_handler       = srp_qp_event;
 494         init_attr->cap.max_send_wr     = m * target->queue_size;
 495         init_attr->cap.max_recv_wr     = target->queue_size + 1;
 496         init_attr->cap.max_recv_sge    = 1;
 497         init_attr->cap.max_send_sge    = 1;
 498         init_attr->sq_sig_type         = IB_SIGNAL_REQ_WR;
 499         init_attr->qp_type             = IB_QPT_RC;
 500         init_attr->send_cq             = send_cq;
 501         init_attr->recv_cq             = recv_cq;
 502
 503         qp = ib_create_qp(dev->pd, init_attr);
 504         if (IS_ERR(qp)) {
 505                 ret = PTR_ERR(qp);
 506                 goto err_send_cq;
 507         }
 508
 509         ret = srp_init_qp(target, qp);
 510         if (ret)
 511                 goto err_qp;
 512
 513         if (dev->use_fast_reg) {
 514                 fr_pool = srp_alloc_fr_pool(target);
 515                 if (IS_ERR(fr_pool)) {
 516                         ret = PTR_ERR(fr_pool);
 517                         shost_printk(KERN_WARNING, target->scsi_host, PFX
 518                                      "FR pool allocation failed (%d)\n", ret);
 519                         goto err_qp;
 520                 }
 521         } else if (dev->use_fmr) {
 522                 fmr_pool = srp_alloc_fmr_pool(target);
 523                 if (IS_ERR(fmr_pool)) {
 524                         ret = PTR_ERR(fmr_pool);
 525                         shost_printk(KERN_WARNING, target->scsi_host, PFX
 526                                      "FMR pool allocation failed (%d)\n", ret);
 527                         goto err_qp;
 528                 }
 529         }
 530
 531         if (ch->qp)
 532                 srp_destroy_qp(ch);
 533         if (ch->recv_cq)
 534                 ib_free_cq(ch->recv_cq);
 535         if (ch->send_cq)
 536                 ib_free_cq(ch->send_cq);
 537
 538         ch->qp = qp;
 539         ch->recv_cq = recv_cq;
 540         ch->send_cq = send_cq;
 541
 542         if (dev->use_fast_reg) {
 543                 if (ch->fr_pool)
 544                         srp_destroy_fr_pool(ch->fr_pool);
 545                 ch->fr_pool = fr_pool;
 546         } else if (dev->use_fmr) {
 547                 if (ch->fmr_pool)
 548                         ib_destroy_fmr_pool(ch->fmr_pool);
 549                 ch->fmr_pool = fmr_pool;
 550         }
 551
 552         kfree(init_attr);
 553         return 0;
 554
 555 err_qp:
 556         srp_destroy_qp(ch);
 557
 558 err_send_cq:
 559         ib_free_cq(send_cq);
 560
 561 err_recv_cq:
 562         ib_free_cq(recv_cq);
 563
 564 err:
 565         kfree(init_attr);
 566         return ret;
 567 }
 568
 569 /*
 570  * Note: this function may be called without srp_alloc_iu_bufs() having been
 571  * invoked. Hence the ch->[rt]x_ring checks.
 572  */
 573 static void srp_free_ch_ib(struct srp_target_port *target,
 574                            struct srp_rdma_ch *ch)
 575 {
 576         struct srp_device *dev = target->srp_host->srp_dev;
 577         int i;
 578
 579         if (!ch->target)
 580                 return;
 581
 582         if (ch->cm_id) {
 583                 ib_destroy_cm_id(ch->cm_id);
 584                 ch->cm_id = NULL;
 585         }
 586
 587         /* If srp_new_cm_id() succeeded but srp_create_ch_ib() not, return. */
 588         if (!ch->qp)
 589                 return;
 590
 591         if (dev->use_fast_reg) {
 592                 if (ch->fr_pool)
 593                         srp_destroy_fr_pool(ch->fr_pool);
 594         } else if (dev->use_fmr) {
 595                 if (ch->fmr_pool)
 596                         ib_destroy_fmr_pool(ch->fmr_pool);
 597         }
 598
 599         srp_destroy_qp(ch);
 600         ib_free_cq(ch->send_cq);
 601         ib_free_cq(ch->recv_cq);
 602
 603         /*
 604          * Avoid that the SCSI error handler tries to use this channel after
 605          * it has been freed. The SCSI error handler can namely continue
 606          * trying to perform recovery actions after scsi_remove_host()
 607          * returned.
 608          */
 609         ch->target = NULL;
 610
 611         ch->qp = NULL;
 612         ch->send_cq = ch->recv_cq = NULL;
 613
 614         if (ch->rx_ring) {
 615                 for (i = 0; i < target->queue_size; ++i)
 616                         srp_free_iu(target->srp_host, ch->rx_ring[i]);
 617                 kfree(ch->rx_ring);
 618                 ch->rx_ring = NULL;
 619         }
 620         if (ch->tx_ring) {
 621                 for (i = 0; i < target->queue_size; ++i)
 622                         srp_free_iu(target->srp_host, ch->tx_ring[i]);
 623                 kfree(ch->tx_ring);
 624                 ch->tx_ring = NULL;
 625         }
 626 }
 627
 628 static void srp_path_rec_completion(int status,
 629                                     struct ib_sa_path_rec *pathrec,
 630                                     void *ch_ptr)
 631 {
 632         struct srp_rdma_ch *ch = ch_ptr;
 633         struct srp_target_port *target = ch->target;
 634
 635         ch->status = status;
 636         if (status)
 637                 shost_printk(KERN_ERR, target->scsi_host,
 638                              PFX "Got failed path rec status %d\n", status);
 639         else
 640                 ch->path = *pathrec;
 641         complete(&ch->done);
 642 }
 643
 644 static int srp_lookup_path(struct srp_rdma_ch *ch)
 645 {
 646         struct srp_target_port *target = ch->target;
 647         int ret;
 648
 649         ch->path.numb_path = 1;
 650
 651         init_completion(&ch->done);
 652
 653         ch->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
 654                                                target->srp_host->srp_dev->dev,
 655                                                target->srp_host->port,
 656                                                &ch->path,
 657                                                IB_SA_PATH_REC_SERVICE_ID |
 658                                                IB_SA_PATH_REC_DGID       |
 659                                                IB_SA_PATH_REC_SGID       |
 660                                                IB_SA_PATH_REC_NUMB_PATH  |
 661                                                IB_SA_PATH_REC_PKEY,
 662                                                SRP_PATH_REC_TIMEOUT_MS,
 663                                                GFP_KERNEL,
 664                                                srp_path_rec_completion,
 665                                                ch, &ch->path_query);
 666         if (ch->path_query_id < 0)
 667                 return ch->path_query_id;
 668
 669         ret = wait_for_completion_interruptible(&ch->done);
 670         if (ret < 0)
 671                 return ret;
 672
 673         if (ch->status < 0)
 674                 shost_printk(KERN_WARNING, target->scsi_host,
 675                              PFX "Path record query failed\n");
 676
 677         return ch->status;
 678 }
 679
 680 static int srp_send_req(struct srp_rdma_ch *ch, bool multich)
 681 {
 682         struct srp_target_port *target = ch->target;
 683         struct {
 684                 struct ib_cm_req_param param;
 685                 struct srp_login_req   priv;
 686         } *req = NULL;
 687         int status;
 688
 689         req = kzalloc(sizeof *req, GFP_KERNEL);
 690         if (!req)
 691                 return -ENOMEM;
 692
 693         req->param.primary_path               = &ch->path;
 694         req->param.alternate_path             = NULL;
 695         req->param.service_id                 = target->service_id;
 696         req->param.qp_num                     = ch->qp->qp_num;
 697         req->param.qp_type                    = ch->qp->qp_type;
 698         req->param.private_data               = &req->priv;
 699         req->param.private_data_len           = sizeof req->priv;
 700         req->param.flow_control               = 1;
 701
 702         get_random_bytes(&req->param.starting_psn, 4);
 703         req->param.starting_psn              &= 0xffffff;
 704
 705         /*
 706          * Pick some arbitrary defaults here; we could make these
 707          * module parameters if anyone cared about setting them.
 708          */
 709         req->param.responder_resources        = 4;
 710         req->param.remote_cm_response_timeout = 20;
 711         req->param.local_cm_response_timeout  = 20;
 712         req->param.retry_count                = target->tl_retry_count;
 713         req->param.rnr_retry_count            = 7;
 714         req->param.max_cm_retries             = 15;
 715
 716         req->priv.opcode        = SRP_LOGIN_REQ;
 717         req->priv.tag           = 0;
 718         req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
 719         req->priv.req_buf_fmt   = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
 720                                               SRP_BUF_FORMAT_INDIRECT);
 721         req->priv.req_flags     = (multich ? SRP_MULTICHAN_MULTI :
 722                                    SRP_MULTICHAN_SINGLE);
 723         /*
 724          * In the published SRP specification (draft rev. 16a), the
 725          * port identifier format is 8 bytes of ID extension followed
 726          * by 8 bytes of GUID.  Older drafts put the two halves in the
 727          * opposite order, so that the GUID comes first.
 728          *
 729          * Targets conforming to these obsolete drafts can be
 730          * recognized by the I/O Class they report.
 731          */
 732         if (target->io_class == SRP_REV10_IB_IO_CLASS) {
 733                 memcpy(req->priv.initiator_port_id,
 734                        &target->sgid.global.interface_id, 8);
 735                 memcpy(req->priv.initiator_port_id + 8,
 736                        &target->initiator_ext, 8);
 737                 memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
 738                 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
 739         } else {
 740                 memcpy(req->priv.initiator_port_id,
 741                        &target->initiator_ext, 8);
 742                 memcpy(req->priv.initiator_port_id + 8,
 743                        &target->sgid.global.interface_id, 8);
 744                 memcpy(req->priv.target_port_id,     &target->id_ext, 8);
 745                 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
 746         }
 747
 748         /*
 749          * Topspin/Cisco SRP targets will reject our login unless we
 750          * zero out the first 8 bytes of our initiator port ID and set
 751          * the second 8 bytes to the local node GUID.
 752          */
 753         if (srp_target_is_topspin(target)) {
 754                 shost_printk(KERN_DEBUG, target->scsi_host,
 755                              PFX "Topspin/Cisco initiator port ID workaround "
 756                              "activated for target GUID %016llx\n",
 757                              be64_to_cpu(target->ioc_guid));
 758                 memset(req->priv.initiator_port_id, 0, 8);
 759                 memcpy(req->priv.initiator_port_id + 8,
 760                        &target->srp_host->srp_dev->dev->node_guid, 8);
 761         }
 762
 763         status = ib_send_cm_req(ch->cm_id, &req->param);
 764
 765         kfree(req);
 766
 767         return status;
 768 }
 769
 770 static bool srp_queue_remove_work(struct srp_target_port *target)
 771 {
 772         bool changed = false;
 773
 774         spin_lock_irq(&target->lock);
 775         if (target->state != SRP_TARGET_REMOVED) {
 776                 target->state = SRP_TARGET_REMOVED;
 777                 changed = true;
 778         }
 779         spin_unlock_irq(&target->lock);
 780
 781         if (changed)
 782                 queue_work(srp_remove_wq, &target->remove_work);
 783
 784         return changed;
 785 }
 786
 787 static void srp_disconnect_target(struct srp_target_port *target)
 788 {
 789         struct srp_rdma_ch *ch;
 790         int i;
 791
 792         /* XXX should send SRP_I_LOGOUT request */
 793
 794         for (i = 0; i < target->ch_count; i++) {
 795                 ch = &target->ch[i];
 796                 ch->connected = false;
 797                 if (ch->cm_id && ib_send_cm_dreq(ch->cm_id, NULL, 0)) {
 798                         shost_printk(KERN_DEBUG, target->scsi_host,
 799                                      PFX "Sending CM DREQ failed\n");
 800                 }
 801         }
 802 }
 803
 804 static void srp_free_req_data(struct srp_target_port *target,
 805                               struct srp_rdma_ch *ch)
 806 {
 807         struct srp_device *dev = target->srp_host->srp_dev;
 808         struct ib_device *ibdev = dev->dev;
 809         struct srp_request *req;
 810         int i;
 811
 812         if (!ch->req_ring)
 813                 return;
 814
 815         for (i = 0; i < target->req_ring_size; ++i) {
 816                 req = &ch->req_ring[i];
 817                 if (dev->use_fast_reg) {
 818                         kfree(req->fr_list);
 819                 } else {
 820                         kfree(req->fmr_list);
 821                         kfree(req->map_page);
 822                 }
 823                 if (req->indirect_dma_addr) {
 824                         ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
 825                                             target->indirect_size,
 826                                             DMA_TO_DEVICE);
 827                 }
 828                 kfree(req->indirect_desc);
 829         }
 830
 831         kfree(ch->req_ring);
 832         ch->req_ring = NULL;
 833 }
 834
 835 static int srp_alloc_req_data(struct srp_rdma_ch *ch)
 836 {
 837         struct srp_target_port *target = ch->target;
 838         struct srp_device *srp_dev = target->srp_host->srp_dev;
 839         struct ib_device *ibdev = srp_dev->dev;
 840         struct srp_request *req;
 841         void *mr_list;
 842         dma_addr_t dma_addr;
 843         int i, ret = -ENOMEM;
 844
 845         ch->req_ring = kcalloc(target->req_ring_size, sizeof(*ch->req_ring),
 846                                GFP_KERNEL);
 847         if (!ch->req_ring)
 848                 goto out;
 849
 850         for (i = 0; i < target->req_ring_size; ++i) {
 851                 req = &ch->req_ring[i];
 852                 mr_list = kmalloc(target->cmd_sg_cnt * sizeof(void *),
 853                                   GFP_KERNEL);
 854                 if (!mr_list)
 855                         goto out;
 856                 if (srp_dev->use_fast_reg) {
 857                         req->fr_list = mr_list;
 858                 } else {
 859                         req->fmr_list = mr_list;
 860                         req->map_page = kmalloc(srp_dev->max_pages_per_mr *
 861                                                 sizeof(void *), GFP_KERNEL);
 862                         if (!req->map_page)
 863                                 goto out;
 864                 }
 865                 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
 866                 if (!req->indirect_desc)
 867                         goto out;
 868
 869                 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
 870                                              target->indirect_size,
 871                                              DMA_TO_DEVICE);
 872                 if (ib_dma_mapping_error(ibdev, dma_addr))
 873                         goto out;
 874
 875                 req->indirect_dma_addr = dma_addr;
 876         }
 877         ret = 0;
 878
 879 out:
 880         return ret;
 881 }
 882
 883 /**
 884  * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
 885  * @shost: SCSI host whose attributes to remove from sysfs.
 886  *
 887  * Note: Any attributes defined in the host template and that did not exist
 888  * before invocation of this function will be ignored.
 889  */
 890 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
 891 {
 892         struct device_attribute **attr;
 893
 894         for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
 895                 device_remove_file(&shost->shost_dev, *attr);
 896 }
 897
 898 static void srp_remove_target(struct srp_target_port *target)
 899 {
 900         struct srp_rdma_ch *ch;
 901         int i;
 902
 903         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 904
 905         srp_del_scsi_host_attr(target->scsi_host);
 906         srp_rport_get(target->rport);
 907         srp_remove_host(target->scsi_host);
 908         scsi_remove_host(target->scsi_host);
 909         srp_stop_rport_timers(target->rport);
 910         srp_disconnect_target(target);
 911         for (i = 0; i < target->ch_count; i++) {
 912                 ch = &target->ch[i];
 913                 srp_free_ch_ib(target, ch);
 914         }
 915         cancel_work_sync(&target->tl_err_work);
 916         srp_rport_put(target->rport);
 917         for (i = 0; i < target->ch_count; i++) {
 918                 ch = &target->ch[i];
 919                 srp_free_req_data(target, ch);
 920         }
 921         kfree(target->ch);
 922         target->ch = NULL;
 923
 924         spin_lock(&target->srp_host->target_lock);
 925         list_del(&target->list);
 926         spin_unlock(&target->srp_host->target_lock);
 927
 928         scsi_host_put(target->scsi_host);
 929 }
 930
 931 static void srp_remove_work(struct work_struct *work)
 932 {
 933         struct srp_target_port *target =
 934                 container_of(work, struct srp_target_port, remove_work);
 935
 936         WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
 937
 938         srp_remove_target(target);
 939 }
 940
 941 static void srp_rport_delete(struct srp_rport *rport)
 942 {
 943         struct srp_target_port *target = rport->lld_data;
 944
 945         srp_queue_remove_work(target);
 946 }
 947
 948 /**
 949  * srp_connected_ch() - number of connected channels
 950  * @target: SRP target port.
 951  */
 952 static int srp_connected_ch(struct srp_target_port *target)
 953 {
 954         int i, c = 0;
 955
 956         for (i = 0; i < target->ch_count; i++)
 957                 c += target->ch[i].connected;
 958
 959         return c;
 960 }
 961
 962 static int srp_connect_ch(struct srp_rdma_ch *ch, bool multich)
 963 {
 964         struct srp_target_port *target = ch->target;
 965         int ret;
 966
 967         WARN_ON_ONCE(!multich && srp_connected_ch(target) > 0);
 968
 969         ret = srp_lookup_path(ch);
 970         if (ret)
 971                 goto out;
 972
 973         while (1) {
 974                 init_completion(&ch->done);
 975                 ret = srp_send_req(ch, multich);
 976                 if (ret)
 977                         goto out;
 978                 ret = wait_for_completion_interruptible(&ch->done);
 979                 if (ret < 0)
 980                         goto out;
 981
 982                 /*
 983                  * The CM event handling code will set status to
 984                  * SRP_PORT_REDIRECT if we get a port redirect REJ
 985                  * back, or SRP_DLID_REDIRECT if we get a lid/qp
 986                  * redirect REJ back.
 987                  */
 988                 ret = ch->status;
 989                 switch (ret) {
 990                 case 0:
 991                         ch->connected = true;
 992                         goto out;
 993
 994                 case SRP_PORT_REDIRECT:
 995                         ret = srp_lookup_path(ch);
 996                         if (ret)
 997                                 goto out;
 998                         break;
 999
1000                 case SRP_DLID_REDIRECT:
1001                         break;
1002
1003                 case SRP_STALE_CONN:
1004                         shost_printk(KERN_ERR, target->scsi_host, PFX
1005                                      "giving up on stale connection\n");
1006                         ret = -ECONNRESET;
1007                         goto out;
1008
1009                 default:
1010                         goto out;
1011                 }
1012         }
1013
1014 out:
1015         return ret <= 0 ? ret : -ENODEV;
1016 }
1017
1018 static void srp_inv_rkey_err_done(struct ib_cq *cq, struct ib_wc *wc)
1019 {
1020         srp_handle_qp_err(cq, wc, "INV RKEY");
1021 }
1022
1023 static int srp_inv_rkey(struct srp_request *req, struct srp_rdma_ch *ch,
1024                 u32 rkey)
1025 {
1026         struct ib_send_wr *bad_wr;
1027         struct ib_send_wr wr = {
1028                 .opcode             = IB_WR_LOCAL_INV,
1029                 .next               = NULL,
1030                 .num_sge            = 0,
1031                 .send_flags         = 0,
1032                 .ex.invalidate_rkey = rkey,
1033         };
1034
1035         wr.wr_cqe = &req->reg_cqe;
1036         req->reg_cqe.done = srp_inv_rkey_err_done;
1037         return ib_post_send(ch->qp, &wr, &bad_wr);
1038 }
1039
1040 static void srp_unmap_data(struct scsi_cmnd *scmnd,
1041                            struct srp_rdma_ch *ch,
1042                            struct srp_request *req)
1043 {
1044         struct srp_target_port *target = ch->target;
1045         struct srp_device *dev = target->srp_host->srp_dev;
1046         struct ib_device *ibdev = dev->dev;
1047         int i, res;
1048
1049         if (!scsi_sglist(scmnd) ||
1050             (scmnd->sc_data_direction != DMA_TO_DEVICE &&
1051              scmnd->sc_data_direction != DMA_FROM_DEVICE))
1052                 return;
1053
1054         if (dev->use_fast_reg) {
1055                 struct srp_fr_desc **pfr;
1056
1057                 for (i = req->nmdesc, pfr = req->fr_list; i > 0; i--, pfr++) {
1058                         res = srp_inv_rkey(req, ch, (*pfr)->mr->rkey);
1059                         if (res < 0) {
1060                                 shost_printk(KERN_ERR, target->scsi_host, PFX
1061                                   "Queueing INV WR for rkey %#x failed (%d)\n",
1062                                   (*pfr)->mr->rkey, res);
1063                                 queue_work(system_long_wq,
1064                                            &target->tl_err_work);
1065                         }
1066                 }
1067                 if (req->nmdesc)
1068                         srp_fr_pool_put(ch->fr_pool, req->fr_list,
1069                                         req->nmdesc);
1070         } else if (dev->use_fmr) {
1071                 struct ib_pool_fmr **pfmr;
1072
1073                 for (i = req->nmdesc, pfmr = req->fmr_list; i > 0; i--, pfmr++)
1074                         ib_fmr_pool_unmap(*pfmr);
1075         }
1076
1077         ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
1078                         scmnd->sc_data_direction);
1079 }
1080
1081 /**
1082  * srp_claim_req - Take ownership of the scmnd associated with a request.
1083  * @ch: SRP RDMA channel.
1084  * @req: SRP request.
1085  * @sdev: If not NULL, only take ownership for this SCSI device.
1086  * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
1087  *         ownership of @req->scmnd if it equals @scmnd.
1088  *
1089  * Return value:
1090  * Either NULL or a pointer to the SCSI command the caller became owner of.
1091  */
1092 static struct scsi_cmnd *srp_claim_req(struct srp_rdma_ch *ch,
1093                                        struct srp_request *req,
1094                                        struct scsi_device *sdev,
1095                                        struct scsi_cmnd *scmnd)
1096 {
1097         unsigned long flags;
1098
1099         spin_lock_irqsave(&ch->lock, flags);
1100         if (req->scmnd &&
1101             (!sdev || req->scmnd->device == sdev) &&
1102             (!scmnd || req->scmnd == scmnd)) {
1103                 scmnd = req->scmnd;
1104                 req->scmnd = NULL;
1105         } else {
1106                 scmnd = NULL;
1107         }
1108         spin_unlock_irqrestore(&ch->lock, flags);
1109
1110         return scmnd;
1111 }
1112
1113 /**
1114  * srp_free_req() - Unmap data and adjust ch->req_lim.
1115  * @ch:     SRP RDMA channel.
1116  * @req:    Request to be freed.
1117  * @scmnd:  SCSI command associated with @req.
1118  * @req_lim_delta: Amount to be added to @target->req_lim.
1119  */
1120 static void srp_free_req(struct srp_rdma_ch *ch, struct srp_request *req,
1121                          struct scsi_cmnd *scmnd, s32 req_lim_delta)
1122 {
1123         unsigned long flags;
1124
1125         srp_unmap_data(scmnd, ch, req);
1126
1127         spin_lock_irqsave(&ch->lock, flags);
1128         ch->req_lim += req_lim_delta;
1129         spin_unlock_irqrestore(&ch->lock, flags);
1130 }
1131
1132 static void srp_finish_req(struct srp_rdma_ch *ch, struct srp_request *req,
1133                            struct scsi_device *sdev, int result)
1134 {
1135         struct scsi_cmnd *scmnd = srp_claim_req(ch, req, sdev, NULL);
1136
1137         if (scmnd) {
1138                 srp_free_req(ch, req, scmnd, 0);
1139                 scmnd->result = result;
1140                 scmnd->scsi_done(scmnd);
1141         }
1142 }
1143
1144 static void srp_terminate_io(struct srp_rport *rport)
1145 {
1146         struct srp_target_port *target = rport->lld_data;
1147         struct srp_rdma_ch *ch;
1148         struct Scsi_Host *shost = target->scsi_host;
1149         struct scsi_device *sdev;
1150         int i, j;
1151
1152         /*
1153          * Invoking srp_terminate_io() while srp_queuecommand() is running
1154          * is not safe. Hence the warning statement below.
1155          */
1156         shost_for_each_device(sdev, shost)
1157                 WARN_ON_ONCE(sdev->request_queue->request_fn_active);
1158
1159         for (i = 0; i < target->ch_count; i++) {
1160                 ch = &target->ch[i];
1161
1162                 for (j = 0; j < target->req_ring_size; ++j) {
1163                         struct srp_request *req = &ch->req_ring[j];
1164
1165                         srp_finish_req(ch, req, NULL,
1166                                        DID_TRANSPORT_FAILFAST << 16);
1167                 }
1168         }
1169 }
1170
1171 /*
1172  * It is up to the caller to ensure that srp_rport_reconnect() calls are
1173  * serialized and that no concurrent srp_queuecommand(), srp_abort(),
1174  * srp_reset_device() or srp_reset_host() calls will occur while this function
1175  * is in progress. One way to realize that is not to call this function
1176  * directly but to call srp_reconnect_rport() instead since that last function
1177  * serializes calls of this function via rport->mutex and also blocks
1178  * srp_queuecommand() calls before invoking this function.
1179  */
1180 static int srp_rport_reconnect(struct srp_rport *rport)
1181 {
1182         struct srp_target_port *target = rport->lld_data;
1183         struct srp_rdma_ch *ch;
1184         int i, j, ret = 0;
1185         bool multich = false;
1186
1187         srp_disconnect_target(target);
1188
1189         if (target->state == SRP_TARGET_SCANNING)
1190                 return -ENODEV;
1191
1192         /*
1193          * Now get a new local CM ID so that we avoid confusing the target in
1194          * case things are really fouled up. Doing so also ensures that all CM
1195          * callbacks will have finished before a new QP is allocated.
1196          */
1197         for (i = 0; i < target->ch_count; i++) {
1198                 ch = &target->ch[i];
1199                 ret += srp_new_cm_id(ch);
1200         }
1201         for (i = 0; i < target->ch_count; i++) {
1202                 ch = &target->ch[i];
1203                 for (j = 0; j < target->req_ring_size; ++j) {
1204                         struct srp_request *req = &ch->req_ring[j];
1205
1206                         srp_finish_req(ch, req, NULL, DID_RESET << 16);
1207                 }
1208         }
1209         for (i = 0; i < target->ch_count; i++) {
1210                 ch = &target->ch[i];
1211                 /*
1212                  * Whether or not creating a new CM ID succeeded, create a new
1213                  * QP. This guarantees that all completion callback function
1214                  * invocations have finished before request resetting starts.
1215                  */
1216                 ret += srp_create_ch_ib(ch);
1217
1218                 INIT_LIST_HEAD(&ch->free_tx);
1219                 for (j = 0; j < target->queue_size; ++j)
1220                         list_add(&ch->tx_ring[j]->list, &ch->free_tx);
1221         }
1222
1223         target->qp_in_error = false;
1224
1225         for (i = 0; i < target->ch_count; i++) {
1226                 ch = &target->ch[i];
1227                 if (ret)
1228                         break;
1229                 ret = srp_connect_ch(ch, multich);
1230                 multich = true;
1231         }
1232
1233         if (ret == 0)
1234                 shost_printk(KERN_INFO, target->scsi_host,
1235                              PFX "reconnect succeeded\n");
1236
1237         return ret;
1238 }
1239
1240 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
1241                          unsigned int dma_len, u32 rkey)
1242 {
1243         struct srp_direct_buf *desc = state->desc;
1244
1245         WARN_ON_ONCE(!dma_len);
1246
1247         desc->va = cpu_to_be64(dma_addr);
1248         desc->key = cpu_to_be32(rkey);
1249         desc->len = cpu_to_be32(dma_len);
1250
1251         state->total_len += dma_len;
1252         state->desc++;
1253         state->ndesc++;
1254 }
1255
1256 static int srp_map_finish_fmr(struct srp_map_state *state,
1257                               struct srp_rdma_ch *ch)
1258 {
1259         struct srp_target_port *target = ch->target;
1260         struct srp_device *dev = target->srp_host->srp_dev;
1261         struct ib_pool_fmr *fmr;
1262         u64 io_addr = 0;
1263
1264         if (state->fmr.next >= state->fmr.end)
1265                 return -ENOMEM;
1266
1267         WARN_ON_ONCE(!dev->use_fmr);
1268
1269         if (state->npages == 0)
1270                 return 0;
1271
1272         if (state->npages == 1 && target->global_mr) {
1273                 srp_map_desc(state, state->base_dma_addr, state->dma_len,
1274                              target->global_mr->rkey);
1275                 goto reset_state;
1276         }
1277
1278         fmr = ib_fmr_pool_map_phys(ch->fmr_pool, state->pages,
1279                                    state->npages, io_addr);
1280         if (IS_ERR(fmr))
1281                 return PTR_ERR(fmr);
1282
1283         *state->fmr.next++ = fmr;
1284         state->nmdesc++;
1285
1286         srp_map_desc(state, state->base_dma_addr & ~dev->mr_page_mask,
1287                      state->dma_len, fmr->fmr->rkey);
1288
1289 reset_state:
1290         state->npages = 0;
1291         state->dma_len = 0;
1292
1293         return 0;
1294 }
1295
1296 static void srp_reg_mr_err_done(struct ib_cq *cq, struct ib_wc *wc)
1297 {
1298         srp_handle_qp_err(cq, wc, "FAST REG");
1299 }
1300
1301 static int srp_map_finish_fr(struct srp_map_state *state,
1302                              struct srp_request *req,
1303                              struct srp_rdma_ch *ch, int sg_nents)
1304 {
1305         struct srp_target_port *target = ch->target;
1306         struct srp_device *dev = target->srp_host->srp_dev;
1307         struct ib_send_wr *bad_wr;
1308         struct ib_reg_wr wr;
1309         struct srp_fr_desc *desc;
1310         u32 rkey;
1311         int n, err;
1312
1313         if (state->fr.next >= state->fr.end)
1314                 return -ENOMEM;
1315
1316         WARN_ON_ONCE(!dev->use_fast_reg);
1317
1318         if (sg_nents == 1 && target->global_mr) {
1319                 srp_map_desc(state, sg_dma_address(state->sg),
1320                              sg_dma_len(state->sg),
1321                              target->global_mr->rkey);
1322                 return 1;
1323         }
1324
1325         desc = srp_fr_pool_get(ch->fr_pool);
1326         if (!desc)
1327                 return -ENOMEM;
1328
1329         rkey = ib_inc_rkey(desc->mr->rkey);
1330         ib_update_fast_reg_key(desc->mr, rkey);
1331
1332         n = ib_map_mr_sg(desc->mr, state->sg, sg_nents, 0, dev->mr_page_size);
1333         if (unlikely(n < 0))
1334                 return n;
1335
1336         req->reg_cqe.done = srp_reg_mr_err_done;
1337
1338         wr.wr.next = NULL;
1339         wr.wr.opcode = IB_WR_REG_MR;
1340         wr.wr.wr_cqe = &req->reg_cqe;
1341         wr.wr.num_sge = 0;
1342         wr.wr.send_flags = 0;
1343         wr.mr = desc->mr;
1344         wr.key = desc->mr->rkey;
1345         wr.access = (IB_ACCESS_LOCAL_WRITE |
1346                      IB_ACCESS_REMOTE_READ |
1347                      IB_ACCESS_REMOTE_WRITE);
1348
1349         *state->fr.next++ = desc;
1350         state->nmdesc++;
1351
1352         srp_map_desc(state, desc->mr->iova,
1353                      desc->mr->length, desc->mr->rkey);
1354
1355         err = ib_post_send(ch->qp, &wr.wr, &bad_wr);
1356         if (unlikely(err))
1357                 return err;
1358
1359         return n;
1360 }
1361
1362 static int srp_map_sg_entry(struct srp_map_state *state,
1363                             struct srp_rdma_ch *ch,
1364                             struct scatterlist *sg, int sg_index)
1365 {
1366         struct srp_target_port *target = ch->target;
1367         struct srp_device *dev = target->srp_host->srp_dev;
1368         struct ib_device *ibdev = dev->dev;
1369         dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
1370         unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
1371         unsigned int len = 0;
1372         int ret;
1373
1374         WARN_ON_ONCE(!dma_len);
1375
1376         while (dma_len) {
1377                 unsigned offset = dma_addr & ~dev->mr_page_mask;
1378                 if (state->npages == dev->max_pages_per_mr || offset != 0) {
1379                         ret = srp_map_finish_fmr(state, ch);
1380                         if (ret)
1381                                 return ret;
1382                 }
1383
1384                 len = min_t(unsigned int, dma_len, dev->mr_page_size - offset);
1385
1386                 if (!state->npages)
1387                         state->base_dma_addr = dma_addr;
1388                 state->pages[state->npages++] = dma_addr & dev->mr_page_mask;
1389                 state->dma_len += len;
1390                 dma_addr += len;
1391                 dma_len -= len;
1392         }
1393
1394         /*
1395          * If the last entry of the MR wasn't a full page, then we need to
1396          * close it out and start a new one -- we can only merge at page
1397          * boundaries.
1398          */
1399         ret = 0;
1400         if (len != dev->mr_page_size)
1401                 ret = srp_map_finish_fmr(state, ch);
1402         return ret;
1403 }
1404
1405 static int srp_map_sg_fmr(struct srp_map_state *state, struct srp_rdma_ch *ch,
1406                           struct srp_request *req, struct scatterlist *scat,
1407                           int count)
1408 {
1409         struct scatterlist *sg;
1410         int i, ret;
1411
1412         state->pages = req->map_page;
1413         state->fmr.next = req->fmr_list;
1414         state->fmr.end = req->fmr_list + ch->target->cmd_sg_cnt;
1415
1416         for_each_sg(scat, sg, count, i) {
1417                 ret = srp_map_sg_entry(state, ch, sg, i);
1418                 if (ret)
1419                         return ret;
1420         }
1421
1422         ret = srp_map_finish_fmr(state, ch);
1423         if (ret)
1424                 return ret;
1425
1426         return 0;
1427 }
1428
1429 static int srp_map_sg_fr(struct srp_map_state *state, struct srp_rdma_ch *ch,
1430                          struct srp_request *req, struct scatterlist *scat,
1431                          int count)
1432 {
1433         state->desc = req->indirect_desc;
1434         state->fr.next = req->fr_list;
1435         state->fr.end = req->fr_list + ch->target->cmd_sg_cnt;
1436         state->sg = scat;
1437
1438         if (count == 0)
1439                 return 0;
1440
1441         while (count) {
1442                 int i, n;
1443
1444                 n = srp_map_finish_fr(state, req, ch, count);
1445                 if (unlikely(n < 0))
1446                         return n;
1447
1448                 count -= n;
1449                 for (i = 0; i < n; i++)
1450                         state->sg = sg_next(state->sg);
1451         }
1452
1453         return 0;
1454 }
1455
1456 static int srp_map_sg_dma(struct srp_map_state *state, struct srp_rdma_ch *ch,
1457                           struct srp_request *req, struct scatterlist *scat,
1458                           int count)
1459 {
1460         struct srp_target_port *target = ch->target;
1461         struct srp_device *dev = target->srp_host->srp_dev;
1462         struct scatterlist *sg;
1463         int i;
1464
1465         state->desc = req->indirect_desc;
1466         for_each_sg(scat, sg, count, i) {
1467                 srp_map_desc(state, ib_sg_dma_address(dev->dev, sg),
1468                              ib_sg_dma_len(dev->dev, sg),
1469                              target->global_mr->rkey);
1470         }
1471
1472         return 0;
1473 }
1474
1475 /*
1476  * Register the indirect data buffer descriptor with the HCA.
1477  *
1478  * Note: since the indirect data buffer descriptor has been allocated with
1479  * kmalloc() it is guaranteed that this buffer is a physically contiguous
1480  * memory buffer.
1481  */
1482 static int srp_map_idb(struct srp_rdma_ch *ch, struct srp_request *req,
1483                        void **next_mr, void **end_mr, u32 idb_len,
1484                        __be32 *idb_rkey)
1485 {
1486         struct srp_target_port *target = ch->target;
1487         struct srp_device *dev = target->srp_host->srp_dev;
1488         struct srp_map_state state;
1489         struct srp_direct_buf idb_desc;
1490         u64 idb_pages[1];
1491         struct scatterlist idb_sg[1];
1492         int ret;
1493
1494         memset(&state, 0, sizeof(state));
1495         memset(&idb_desc, 0, sizeof(idb_desc));
1496         state.gen.next = next_mr;
1497         state.gen.end = end_mr;
1498         state.desc = &idb_desc;
1499         state.base_dma_addr = req->indirect_dma_addr;
1500         state.dma_len = idb_len;
1501
1502         if (dev->use_fast_reg) {
1503                 state.sg = idb_sg;
1504                 sg_set_buf(idb_sg, req->indirect_desc, idb_len);
1505                 idb_sg->dma_address = req->indirect_dma_addr; /* hack! */
1506 #ifdef CONFIG_NEED_SG_DMA_LENGTH
1507                 idb_sg->dma_length = idb_sg->length;          /* hack^2 */
1508 #endif
1509                 ret = srp_map_finish_fr(&state, req, ch, 1);
1510                 if (ret < 0)
1511                         return ret;
1512         } else if (dev->use_fmr) {
1513                 state.pages = idb_pages;
1514                 state.pages[0] = (req->indirect_dma_addr &
1515                                   dev->mr_page_mask);
1516                 state.npages = 1;
1517                 ret = srp_map_finish_fmr(&state, ch);
1518                 if (ret < 0)
1519                         return ret;
1520         } else {
1521                 return -EINVAL;
1522         }
1523
1524         *idb_rkey = idb_desc.key;
1525
1526         return 0;
1527 }
1528
1529 /**
1530  * srp_map_data() - map SCSI data buffer onto an SRP request
1531  * @scmnd: SCSI command to map
1532  * @ch: SRP RDMA channel
1533  * @req: SRP request
1534  *
1535  * Returns the length in bytes of the SRP_CMD IU or a negative value if
1536  * mapping failed.
1537  */
1538 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_rdma_ch *ch,
1539                         struct srp_request *req)
1540 {
1541         struct srp_target_port *target = ch->target;
1542         struct scatterlist *scat;
1543         struct srp_cmd *cmd = req->cmd->buf;
1544         int len, nents, count, ret;
1545         struct srp_device *dev;
1546         struct ib_device *ibdev;
1547         struct srp_map_state state;
1548         struct srp_indirect_buf *indirect_hdr;
1549         u32 idb_len, table_len;
1550         __be32 idb_rkey;
1551         u8 fmt;
1552
1553         if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
1554                 return sizeof (struct srp_cmd);
1555
1556         if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
1557             scmnd->sc_data_direction != DMA_TO_DEVICE) {
1558                 shost_printk(KERN_WARNING, target->scsi_host,
1559                              PFX "Unhandled data direction %d\n",
1560                              scmnd->sc_data_direction);
1561                 return -EINVAL;
1562         }
1563
1564         nents = scsi_sg_count(scmnd);
1565         scat  = scsi_sglist(scmnd);
1566
1567         dev = target->srp_host->srp_dev;
1568         ibdev = dev->dev;
1569
1570         count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
1571         if (unlikely(count == 0))
1572                 return -EIO;
1573
1574         fmt = SRP_DATA_DESC_DIRECT;
1575         len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
1576
1577         if (count == 1 && target->global_mr) {
1578                 /*
1579                  * The midlayer only generated a single gather/scatter
1580                  * entry, or DMA mapping coalesced everything to a
1581                  * single entry.  So a direct descriptor along with
1582                  * the DMA MR suffices.
1583                  */
1584                 struct srp_direct_buf *buf = (void *) cmd->add_data;
1585
1586                 buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
1587                 buf->key = cpu_to_be32(target->global_mr->rkey);
1588                 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
1589
1590                 req->nmdesc = 0;
1591                 goto map_complete;
1592         }
1593
1594         /*
1595          * We have more than one scatter/gather entry, so build our indirect
1596          * descriptor table, trying to merge as many entries as we can.
1597          */
1598         indirect_hdr = (void *) cmd->add_data;
1599
1600         ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
1601                                    target->indirect_size, DMA_TO_DEVICE);
1602
1603         memset(&state, 0, sizeof(state));
1604         if (dev->use_fast_reg)
1605                 ret = srp_map_sg_fr(&state, ch, req, scat, count);
1606         else if (dev->use_fmr)
1607                 ret = srp_map_sg_fmr(&state, ch, req, scat, count);
1608         else
1609                 ret = srp_map_sg_dma(&state, ch, req, scat, count);
1610         req->nmdesc = state.nmdesc;
1611         if (ret < 0)
1612                 goto unmap;
1613
1614         /* We've mapped the request, now pull as much of the indirect
1615          * descriptor table as we can into the command buffer. If this
1616          * target is not using an external indirect table, we are
1617          * guaranteed to fit into the command, as the SCSI layer won't
1618          * give us more S/G entries than we allow.
1619          */
1620         if (state.ndesc == 1) {
1621                 /*
1622                  * Memory registration collapsed the sg-list into one entry,
1623                  * so use a direct descriptor.
1624                  */
1625                 struct srp_direct_buf *buf = (void *) cmd->add_data;
1626
1627                 *buf = req->indirect_desc[0];
1628                 goto map_complete;
1629         }
1630
1631         if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1632                                                 !target->allow_ext_sg)) {
1633                 shost_printk(KERN_ERR, target->scsi_host,
1634                              "Could not fit S/G list into SRP_CMD\n");
1635                 ret = -EIO;
1636                 goto unmap;
1637         }
1638
1639         count = min(state.ndesc, target->cmd_sg_cnt);
1640         table_len = state.ndesc * sizeof (struct srp_direct_buf);
1641         idb_len = sizeof(struct srp_indirect_buf) + table_len;
1642
1643         fmt = SRP_DATA_DESC_INDIRECT;
1644         len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1645         len += count * sizeof (struct srp_direct_buf);
1646
1647         memcpy(indirect_hdr->desc_list, req->indirect_desc,
1648                count * sizeof (struct srp_direct_buf));
1649
1650         if (!target->global_mr) {
1651                 ret = srp_map_idb(ch, req, state.gen.next, state.gen.end,
1652                                   idb_len, &idb_rkey);
1653                 if (ret < 0)
1654                         goto unmap;
1655                 req->nmdesc++;
1656         } else {
1657                 idb_rkey = cpu_to_be32(target->global_mr->rkey);
1658         }
1659
1660         indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1661         indirect_hdr->table_desc.key = idb_rkey;
1662         indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1663         indirect_hdr->len = cpu_to_be32(state.total_len);
1664
1665         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1666                 cmd->data_out_desc_cnt = count;
1667         else
1668                 cmd->data_in_desc_cnt = count;
1669
1670         ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1671                                       DMA_TO_DEVICE);
1672
1673 map_complete:
1674         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1675                 cmd->buf_fmt = fmt << 4;
1676         else
1677                 cmd->buf_fmt = fmt;
1678
1679         return len;
1680
1681 unmap:
1682         srp_unmap_data(scmnd, ch, req);
1683         if (ret == -ENOMEM && req->nmdesc >= target->mr_pool_size)
1684                 ret = -E2BIG;
1685         return ret;
1686 }
1687
1688 /*
1689  * Return an IU and possible credit to the free pool
1690  */
1691 static void srp_put_tx_iu(struct srp_rdma_ch *ch, struct srp_iu *iu,
1692                           enum srp_iu_type iu_type)
1693 {
1694         unsigned long flags;
1695
1696         spin_lock_irqsave(&ch->lock, flags);
1697         list_add(&iu->list, &ch->free_tx);
1698         if (iu_type != SRP_IU_RSP)
1699                 ++ch->req_lim;
1700         spin_unlock_irqrestore(&ch->lock, flags);
1701 }
1702
1703 /*
1704  * Must be called with ch->lock held to protect req_lim and free_tx.
1705  * If IU is not sent, it must be returned using srp_put_tx_iu().
1706  *
1707  * Note:
1708  * An upper limit for the number of allocated information units for each
1709  * request type is:
1710  * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1711  *   more than Scsi_Host.can_queue requests.
1712  * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1713  * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1714  *   one unanswered SRP request to an initiator.
1715  */
1716 static struct srp_iu *__srp_get_tx_iu(struct srp_rdma_ch *ch,
1717                                       enum srp_iu_type iu_type)
1718 {
1719         struct srp_target_port *target = ch->target;
1720         s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1721         struct srp_iu *iu;
1722
1723         ib_process_cq_direct(ch->send_cq, -1);
1724
1725         if (list_empty(&ch->free_tx))
1726                 return NULL;
1727
1728         /* Initiator responses to target requests do not consume credits */
1729         if (iu_type != SRP_IU_RSP) {
1730                 if (ch->req_lim <= rsv) {
1731                         ++target->zero_req_lim;
1732                         return NULL;
1733                 }
1734
1735                 --ch->req_lim;
1736         }
1737
1738         iu = list_first_entry(&ch->free_tx, struct srp_iu, list);
1739         list_del(&iu->list);
1740         return iu;
1741 }
1742
1743 static void srp_send_done(struct ib_cq *cq, struct ib_wc *wc)
1744 {
1745         struct srp_iu *iu = container_of(wc->wr_cqe, struct srp_iu, cqe);
1746         struct srp_rdma_ch *ch = cq->cq_context;
1747
1748         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1749                 srp_handle_qp_err(cq, wc, "SEND");
1750                 return;
1751         }
1752
1753         list_add(&iu->list, &ch->free_tx);
1754 }
1755
1756 static int srp_post_send(struct srp_rdma_ch *ch, struct srp_iu *iu, int len)
1757 {
1758         struct srp_target_port *target = ch->target;
1759         struct ib_sge list;
1760         struct ib_send_wr wr, *bad_wr;
1761
1762         list.addr   = iu->dma;
1763         list.length = len;
1764         list.lkey   = target->lkey;
1765
1766         iu->cqe.done = srp_send_done;
1767
1768         wr.next       = NULL;
1769         wr.wr_cqe     = &iu->cqe;
1770         wr.sg_list    = &list;
1771         wr.num_sge    = 1;
1772         wr.opcode     = IB_WR_SEND;
1773         wr.send_flags = IB_SEND_SIGNALED;
1774
1775         return ib_post_send(ch->qp, &wr, &bad_wr);
1776 }
1777
1778 static int srp_post_recv(struct srp_rdma_ch *ch, struct srp_iu *iu)
1779 {
1780         struct srp_target_port *target = ch->target;
1781         struct ib_recv_wr wr, *bad_wr;
1782         struct ib_sge list;
1783
1784         list.addr   = iu->dma;
1785         list.length = iu->size;
1786         list.lkey   = target->lkey;
1787
1788         iu->cqe.done = srp_recv_done;
1789
1790         wr.next     = NULL;
1791         wr.wr_cqe   = &iu->cqe;
1792         wr.sg_list  = &list;
1793         wr.num_sge  = 1;
1794
1795         return ib_post_recv(ch->qp, &wr, &bad_wr);
1796 }
1797
1798 static void srp_process_rsp(struct srp_rdma_ch *ch, struct srp_rsp *rsp)
1799 {
1800         struct srp_target_port *target = ch->target;
1801         struct srp_request *req;
1802         struct scsi_cmnd *scmnd;
1803         unsigned long flags;
1804
1805         if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1806                 spin_lock_irqsave(&ch->lock, flags);
1807                 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1808                 spin_unlock_irqrestore(&ch->lock, flags);
1809
1810                 ch->tsk_mgmt_status = -1;
1811                 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1812                         ch->tsk_mgmt_status = rsp->data[3];
1813                 complete(&ch->tsk_mgmt_done);
1814         } else {
1815                 scmnd = scsi_host_find_tag(target->scsi_host, rsp->tag);
1816                 if (scmnd) {
1817                         req = (void *)scmnd->host_scribble;
1818                         scmnd = srp_claim_req(ch, req, NULL, scmnd);
1819                 }
1820                 if (!scmnd) {
1821                         shost_printk(KERN_ERR, target->scsi_host,
1822                                      "Null scmnd for RSP w/tag %#016llx received on ch %td / QP %#x\n",
1823                                      rsp->tag, ch - target->ch, ch->qp->qp_num);
1824
1825                         spin_lock_irqsave(&ch->lock, flags);
1826                         ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1827                         spin_unlock_irqrestore(&ch->lock, flags);
1828
1829                         return;
1830                 }
1831                 scmnd->result = rsp->status;
1832
1833                 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1834                         memcpy(scmnd->sense_buffer, rsp->data +
1835                                be32_to_cpu(rsp->resp_data_len),
1836                                min_t(int, be32_to_cpu(rsp->sense_data_len),
1837                                      SCSI_SENSE_BUFFERSIZE));
1838                 }
1839
1840                 if (unlikely(rsp->flags & SRP_RSP_FLAG_DIUNDER))
1841                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1842                 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DIOVER))
1843                         scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_in_res_cnt));
1844                 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOUNDER))
1845                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1846                 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOOVER))
1847                         scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_out_res_cnt));
1848
1849                 srp_free_req(ch, req, scmnd,
1850                              be32_to_cpu(rsp->req_lim_delta));
1851
1852                 scmnd->host_scribble = NULL;
1853                 scmnd->scsi_done(scmnd);
1854         }
1855 }
1856
1857 static int srp_response_common(struct srp_rdma_ch *ch, s32 req_delta,
1858                                void *rsp, int len)
1859 {
1860         struct srp_target_port *target = ch->target;
1861         struct ib_device *dev = target->srp_host->srp_dev->dev;
1862         unsigned long flags;
1863         struct srp_iu *iu;
1864         int err;
1865
1866         spin_lock_irqsave(&ch->lock, flags);
1867         ch->req_lim += req_delta;
1868         iu = __srp_get_tx_iu(ch, SRP_IU_RSP);
1869         spin_unlock_irqrestore(&ch->lock, flags);
1870
1871         if (!iu) {
1872                 shost_printk(KERN_ERR, target->scsi_host, PFX
1873                              "no IU available to send response\n");
1874                 return 1;
1875         }
1876
1877         ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1878         memcpy(iu->buf, rsp, len);
1879         ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1880
1881         err = srp_post_send(ch, iu, len);
1882         if (err) {
1883                 shost_printk(KERN_ERR, target->scsi_host, PFX
1884                              "unable to post response: %d\n", err);
1885                 srp_put_tx_iu(ch, iu, SRP_IU_RSP);
1886         }
1887
1888         return err;
1889 }
1890
1891 static void srp_process_cred_req(struct srp_rdma_ch *ch,
1892                                  struct srp_cred_req *req)
1893 {
1894         struct srp_cred_rsp rsp = {
1895                 .opcode = SRP_CRED_RSP,
1896                 .tag = req->tag,
1897         };
1898         s32 delta = be32_to_cpu(req->req_lim_delta);
1899
1900         if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1901                 shost_printk(KERN_ERR, ch->target->scsi_host, PFX
1902                              "problems processing SRP_CRED_REQ\n");
1903 }
1904
1905 static void srp_process_aer_req(struct srp_rdma_ch *ch,
1906                                 struct srp_aer_req *req)
1907 {
1908         struct srp_target_port *target = ch->target;
1909         struct srp_aer_rsp rsp = {
1910                 .opcode = SRP_AER_RSP,
1911                 .tag = req->tag,
1912         };
1913         s32 delta = be32_to_cpu(req->req_lim_delta);
1914
1915         shost_printk(KERN_ERR, target->scsi_host, PFX
1916                      "ignoring AER for LUN %llu\n", scsilun_to_int(&req->lun));
1917
1918         if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1919                 shost_printk(KERN_ERR, target->scsi_host, PFX
1920                              "problems processing SRP_AER_REQ\n");
1921 }
1922
1923 static void srp_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1924 {
1925         struct srp_iu *iu = container_of(wc->wr_cqe, struct srp_iu, cqe);
1926         struct srp_rdma_ch *ch = cq->cq_context;
1927         struct srp_target_port *target = ch->target;
1928         struct ib_device *dev = target->srp_host->srp_dev->dev;
1929         int res;
1930         u8 opcode;
1931
1932         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1933                 srp_handle_qp_err(cq, wc, "RECV");
1934                 return;
1935         }
1936
1937         ib_dma_sync_single_for_cpu(dev, iu->dma, ch->max_ti_iu_len,
1938                                    DMA_FROM_DEVICE);
1939
1940         opcode = *(u8 *) iu->buf;
1941
1942         if (0) {
1943                 shost_printk(KERN_ERR, target->scsi_host,
1944                              PFX "recv completion, opcode 0x%02x\n", opcode);
1945                 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1946                                iu->buf, wc->byte_len, true);
1947         }
1948
1949         switch (opcode) {
1950         case SRP_RSP:
1951                 srp_process_rsp(ch, iu->buf);
1952                 break;
1953
1954         case SRP_CRED_REQ:
1955                 srp_process_cred_req(ch, iu->buf);
1956                 break;
1957
1958         case SRP_AER_REQ:
1959                 srp_process_aer_req(ch, iu->buf);
1960                 break;
1961
1962         case SRP_T_LOGOUT:
1963                 /* XXX Handle target logout */
1964                 shost_printk(KERN_WARNING, target->scsi_host,
1965                              PFX "Got target logout request\n");
1966                 break;
1967
1968         default:
1969                 shost_printk(KERN_WARNING, target->scsi_host,
1970                              PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1971                 break;
1972         }
1973
1974         ib_dma_sync_single_for_device(dev, iu->dma, ch->max_ti_iu_len,
1975                                       DMA_FROM_DEVICE);
1976
1977         res = srp_post_recv(ch, iu);
1978         if (res != 0)
1979                 shost_printk(KERN_ERR, target->scsi_host,
1980                              PFX "Recv failed with error code %d\n", res);
1981 }
1982
1983 /**
1984  * srp_tl_err_work() - handle a transport layer error
1985  * @work: Work structure embedded in an SRP target port.
1986  *
1987  * Note: This function may get invoked before the rport has been created,
1988  * hence the target->rport test.
1989  */
1990 static void srp_tl_err_work(struct work_struct *work)
1991 {
1992         struct srp_target_port *target;
1993
1994         target = container_of(work, struct srp_target_port, tl_err_work);
1995         if (target->rport)
1996                 srp_start_tl_fail_timers(target->rport);
1997 }
1998
1999 static void srp_handle_qp_err(struct ib_cq *cq, struct ib_wc *wc,
2000                 const char *opname)
2001 {
2002         struct srp_rdma_ch *ch = cq->cq_context;
2003         struct srp_target_port *target = ch->target;
2004
2005         if (ch->connected && !target->qp_in_error) {
2006                 shost_printk(KERN_ERR, target->scsi_host,
2007                              PFX "failed %s status %s (%d) for CQE %p\n",
2008                              opname, ib_wc_status_msg(wc->status), wc->status,
2009                              wc->wr_cqe);
2010                 queue_work(system_long_wq, &target->tl_err_work);
2011         }
2012         target->qp_in_error = true;
2013 }
2014
2015 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
2016 {
2017         struct srp_target_port *target = host_to_target(shost);
2018         struct srp_rport *rport = target->rport;
2019         struct srp_rdma_ch *ch;
2020         struct srp_request *req;
2021         struct srp_iu *iu;
2022         struct srp_cmd *cmd;
2023         struct ib_device *dev;
2024         unsigned long flags;
2025         u32 tag;
2026         u16 idx;
2027         int len, ret;
2028         const bool in_scsi_eh = !in_interrupt() && current == shost->ehandler;
2029
2030         /*
2031          * The SCSI EH thread is the only context from which srp_queuecommand()
2032          * can get invoked for blocked devices (SDEV_BLOCK /
2033          * SDEV_CREATED_BLOCK). Avoid racing with srp_reconnect_rport() by
2034          * locking the rport mutex if invoked from inside the SCSI EH.
2035          */
2036         if (in_scsi_eh)
2037                 mutex_lock(&rport->mutex);
2038
2039         scmnd->result = srp_chkready(target->rport);
2040         if (unlikely(scmnd->result))
2041                 goto err;
2042
2043         WARN_ON_ONCE(scmnd->request->tag < 0);
2044         tag = blk_mq_unique_tag(scmnd->request);
2045         ch = &target->ch[blk_mq_unique_tag_to_hwq(tag)];
2046         idx = blk_mq_unique_tag_to_tag(tag);
2047         WARN_ONCE(idx >= target->req_ring_size, "%s: tag %#x: idx %d >= %d\n",
2048                   dev_name(&shost->shost_gendev), tag, idx,
2049                   target->req_ring_size);
2050
2051         spin_lock_irqsave(&ch->lock, flags);
2052         iu = __srp_get_tx_iu(ch, SRP_IU_CMD);
2053         spin_unlock_irqrestore(&ch->lock, flags);
2054
2055         if (!iu)
2056                 goto err;
2057
2058         req = &ch->req_ring[idx];
2059         dev = target->srp_host->srp_dev->dev;
2060         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
2061                                    DMA_TO_DEVICE);
2062
2063         scmnd->host_scribble = (void *) req;
2064
2065         cmd = iu->buf;
2066         memset(cmd, 0, sizeof *cmd);
2067
2068         cmd->opcode = SRP_CMD;
2069         int_to_scsilun(scmnd->device->lun, &cmd->lun);
2070         cmd->tag    = tag;
2071         memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
2072
2073         req->scmnd    = scmnd;
2074         req->cmd      = iu;
2075
2076         len = srp_map_data(scmnd, ch, req);
2077         if (len < 0) {
2078                 shost_printk(KERN_ERR, target->scsi_host,
2079                              PFX "Failed to map data (%d)\n", len);
2080                 /*
2081                  * If we ran out of memory descriptors (-ENOMEM) because an
2082                  * application is queuing many requests with more than
2083                  * max_pages_per_mr sg-list elements, tell the SCSI mid-layer
2084                  * to reduce queue depth temporarily.
2085                  */
2086                 scmnd->result = len == -ENOMEM ?
2087                         DID_OK << 16 | QUEUE_FULL << 1 : DID_ERROR << 16;
2088                 goto err_iu;
2089         }
2090
2091         ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
2092                                       DMA_TO_DEVICE);
2093
2094         if (srp_post_send(ch, iu, len)) {
2095                 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
2096                 goto err_unmap;
2097         }
2098
2099         ret = 0;
2100
2101 unlock_rport:
2102         if (in_scsi_eh)
2103                 mutex_unlock(&rport->mutex);
2104
2105         return ret;
2106
2107 err_unmap:
2108         srp_unmap_data(scmnd, ch, req);
2109
2110 err_iu:
2111         srp_put_tx_iu(ch, iu, SRP_IU_CMD);
2112
2113         /*
2114          * Avoid that the loops that iterate over the request ring can
2115          * encounter a dangling SCSI command pointer.
2116          */
2117         req->scmnd = NULL;
2118
2119 err:
2120         if (scmnd->result) {
2121                 scmnd->scsi_done(scmnd);
2122                 ret = 0;
2123         } else {
2124                 ret = SCSI_MLQUEUE_HOST_BUSY;
2125         }
2126
2127         goto unlock_rport;
2128 }
2129
2130 /*
2131  * Note: the resources allocated in this function are freed in
2132  * srp_free_ch_ib().
2133  */
2134 static int srp_alloc_iu_bufs(struct srp_rdma_ch *ch)
2135 {
2136         struct srp_target_port *target = ch->target;
2137         int i;
2138
2139         ch->rx_ring = kcalloc(target->queue_size, sizeof(*ch->rx_ring),
2140                               GFP_KERNEL);
2141         if (!ch->rx_ring)
2142                 goto err_no_ring;
2143         ch->tx_ring = kcalloc(target->queue_size, sizeof(*ch->tx_ring),
2144                               GFP_KERNEL);
2145         if (!ch->tx_ring)
2146                 goto err_no_ring;
2147
2148         for (i = 0; i < target->queue_size; ++i) {
2149                 ch->rx_ring[i] = srp_alloc_iu(target->srp_host,
2150                                               ch->max_ti_iu_len,
2151                                               GFP_KERNEL, DMA_FROM_DEVICE);
2152                 if (!ch->rx_ring[i])
2153                         goto err;
2154         }
2155
2156         for (i = 0; i < target->queue_size; ++i) {
2157                 ch->tx_ring[i] = srp_alloc_iu(target->srp_host,
2158                                               target->max_iu_len,
2159                                               GFP_KERNEL, DMA_TO_DEVICE);
2160                 if (!ch->tx_ring[i])
2161                         goto err;
2162
2163                 list_add(&ch->tx_ring[i]->list, &ch->free_tx);
2164         }
2165
2166         return 0;
2167
2168 err:
2169         for (i = 0; i < target->queue_size; ++i) {
2170                 srp_free_iu(target->srp_host, ch->rx_ring[i]);
2171                 srp_free_iu(target->srp_host, ch->tx_ring[i]);
2172         }
2173
2174
2175 err_no_ring:
2176         kfree(ch->tx_ring);
2177         ch->tx_ring = NULL;
2178         kfree(ch->rx_ring);
2179         ch->rx_ring = NULL;
2180
2181         return -ENOMEM;
2182 }
2183
2184 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
2185 {
2186         uint64_t T_tr_ns, max_compl_time_ms;
2187         uint32_t rq_tmo_jiffies;
2188
2189         /*
2190          * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
2191          * table 91), both the QP timeout and the retry count have to be set
2192          * for RC QP's during the RTR to RTS transition.
2193          */
2194         WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
2195                      (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
2196
2197         /*
2198          * Set target->rq_tmo_jiffies to one second more than the largest time
2199          * it can take before an error completion is generated. See also
2200          * C9-140..142 in the IBTA spec for more information about how to
2201          * convert the QP Local ACK Timeout value to nanoseconds.
2202          */
2203         T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
2204         max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
2205         do_div(max_compl_time_ms, NSEC_PER_MSEC);
2206         rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
2207
2208         return rq_tmo_jiffies;
2209 }
2210
2211 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
2212                                const struct srp_login_rsp *lrsp,
2213                                struct srp_rdma_ch *ch)
2214 {
2215         struct srp_target_port *target = ch->target;
2216         struct ib_qp_attr *qp_attr = NULL;
2217         int attr_mask = 0;
2218         int ret;
2219         int i;
2220
2221         if (lrsp->opcode == SRP_LOGIN_RSP) {
2222                 ch->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
2223                 ch->req_lim       = be32_to_cpu(lrsp->req_lim_delta);
2224
2225                 /*
2226                  * Reserve credits for task management so we don't
2227                  * bounce requests back to the SCSI mid-layer.
2228                  */
2229                 target->scsi_host->can_queue
2230                         = min(ch->req_lim - SRP_TSK_MGMT_SQ_SIZE,
2231                               target->scsi_host->can_queue);
2232                 target->scsi_host->cmd_per_lun
2233                         = min_t(int, target->scsi_host->can_queue,
2234                                 target->scsi_host->cmd_per_lun);
2235         } else {
2236                 shost_printk(KERN_WARNING, target->scsi_host,
2237                              PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
2238                 ret = -ECONNRESET;
2239                 goto error;
2240         }
2241
2242         if (!ch->rx_ring) {
2243                 ret = srp_alloc_iu_bufs(ch);
2244                 if (ret)
2245                         goto error;
2246         }
2247
2248         ret = -ENOMEM;
2249         qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
2250         if (!qp_attr)
2251                 goto error;
2252
2253         qp_attr->qp_state = IB_QPS_RTR;
2254         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2255         if (ret)
2256                 goto error_free;
2257
2258         ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2259         if (ret)
2260                 goto error_free;
2261
2262         for (i = 0; i < target->queue_size; i++) {
2263                 struct srp_iu *iu = ch->rx_ring[i];
2264
2265                 ret = srp_post_recv(ch, iu);
2266                 if (ret)
2267                         goto error_free;
2268         }
2269
2270         qp_attr->qp_state = IB_QPS_RTS;
2271         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2272         if (ret)
2273                 goto error_free;
2274
2275         target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
2276
2277         ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2278         if (ret)
2279                 goto error_free;
2280
2281         ret = ib_send_cm_rtu(cm_id, NULL, 0);
2282
2283 error_free:
2284         kfree(qp_attr);
2285
2286 error:
2287         ch->status = ret;
2288 }
2289
2290 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
2291                                struct ib_cm_event *event,
2292                                struct srp_rdma_ch *ch)
2293 {
2294         struct srp_target_port *target = ch->target;
2295         struct Scsi_Host *shost = target->scsi_host;
2296         struct ib_class_port_info *cpi;
2297         int opcode;
2298
2299         switch (event->param.rej_rcvd.reason) {
2300         case IB_CM_REJ_PORT_CM_REDIRECT:
2301                 cpi = event->param.rej_rcvd.ari;
2302                 ch->path.dlid = cpi->redirect_lid;
2303                 ch->path.pkey = cpi->redirect_pkey;
2304                 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
2305                 memcpy(ch->path.dgid.raw, cpi->redirect_gid, 16);
2306
2307                 ch->status = ch->path.dlid ?
2308                         SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
2309                 break;
2310
2311         case IB_CM_REJ_PORT_REDIRECT:
2312                 if (srp_target_is_topspin(target)) {
2313                         /*
2314                          * Topspin/Cisco SRP gateways incorrectly send
2315                          * reject reason code 25 when they mean 24
2316                          * (port redirect).
2317                          */
2318                         memcpy(ch->path.dgid.raw,
2319                                event->param.rej_rcvd.ari, 16);
2320
2321                         shost_printk(KERN_DEBUG, shost,
2322                                      PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
2323                                      be64_to_cpu(ch->path.dgid.global.subnet_prefix),
2324                                      be64_to_cpu(ch->path.dgid.global.interface_id));
2325
2326                         ch->status = SRP_PORT_REDIRECT;
2327                 } else {
2328                         shost_printk(KERN_WARNING, shost,
2329                                      "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
2330                         ch->status = -ECONNRESET;
2331                 }
2332                 break;
2333
2334         case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
2335                 shost_printk(KERN_WARNING, shost,
2336                             "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
2337                 ch->status = -ECONNRESET;
2338                 break;
2339
2340         case IB_CM_REJ_CONSUMER_DEFINED:
2341                 opcode = *(u8 *) event->private_data;
2342                 if (opcode == SRP_LOGIN_REJ) {
2343                         struct srp_login_rej *rej = event->private_data;
2344                         u32 reason = be32_to_cpu(rej->reason);
2345
2346                         if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
2347                                 shost_printk(KERN_WARNING, shost,
2348                                              PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
2349                         else
2350                                 shost_printk(KERN_WARNING, shost, PFX
2351                                              "SRP LOGIN from %pI6 to %pI6 REJECTED, reason 0x%08x\n",
2352                                              target->sgid.raw,
2353                                              target->orig_dgid.raw, reason);
2354                 } else
2355                         shost_printk(KERN_WARNING, shost,
2356                                      "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
2357                                      " opcode 0x%02x\n", opcode);
2358                 ch->status = -ECONNRESET;
2359                 break;
2360
2361         case IB_CM_REJ_STALE_CONN:
2362                 shost_printk(KERN_WARNING, shost, "  REJ reason: stale connection\n");
2363                 ch->status = SRP_STALE_CONN;
2364                 break;
2365
2366         default:
2367                 shost_printk(KERN_WARNING, shost, "  REJ reason 0x%x\n",
2368                              event->param.rej_rcvd.reason);
2369                 ch->status = -ECONNRESET;
2370         }
2371 }
2372
2373 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2374 {
2375         struct srp_rdma_ch *ch = cm_id->context;
2376         struct srp_target_port *target = ch->target;
2377         int comp = 0;
2378
2379         switch (event->event) {
2380         case IB_CM_REQ_ERROR:
2381                 shost_printk(KERN_DEBUG, target->scsi_host,
2382                              PFX "Sending CM REQ failed\n");
2383                 comp = 1;
2384                 ch->status = -ECONNRESET;
2385                 break;
2386
2387         case IB_CM_REP_RECEIVED:
2388                 comp = 1;
2389                 srp_cm_rep_handler(cm_id, event->private_data, ch);
2390                 break;
2391
2392         case IB_CM_REJ_RECEIVED:
2393                 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
2394                 comp = 1;
2395
2396                 srp_cm_rej_handler(cm_id, event, ch);
2397                 break;
2398
2399         case IB_CM_DREQ_RECEIVED:
2400                 shost_printk(KERN_WARNING, target->scsi_host,
2401                              PFX "DREQ received - connection closed\n");
2402                 ch->connected = false;
2403                 if (ib_send_cm_drep(cm_id, NULL, 0))
2404                         shost_printk(KERN_ERR, target->scsi_host,
2405                                      PFX "Sending CM DREP failed\n");
2406                 queue_work(system_long_wq, &target->tl_err_work);
2407                 break;
2408
2409         case IB_CM_TIMEWAIT_EXIT:
2410                 shost_printk(KERN_ERR, target->scsi_host,
2411                              PFX "connection closed\n");
2412                 comp = 1;
2413
2414                 ch->status = 0;
2415                 break;
2416
2417         case IB_CM_MRA_RECEIVED:
2418         case IB_CM_DREQ_ERROR:
2419         case IB_CM_DREP_RECEIVED:
2420                 break;
2421
2422         default:
2423                 shost_printk(KERN_WARNING, target->scsi_host,
2424                              PFX "Unhandled CM event %d\n", event->event);
2425                 break;
2426         }
2427
2428         if (comp)
2429                 complete(&ch->done);
2430
2431         return 0;
2432 }
2433
2434 /**
2435  * srp_change_queue_depth - setting device queue depth
2436  * @sdev: scsi device struct
2437  * @qdepth: requested queue depth
2438  *
2439  * Returns queue depth.
2440  */
2441 static int
2442 srp_change_queue_depth(struct scsi_device *sdev, int qdepth)
2443 {
2444         if (!sdev->tagged_supported)
2445                 qdepth = 1;
2446         return scsi_change_queue_depth(sdev, qdepth);
2447 }
2448
2449 static int srp_send_tsk_mgmt(struct srp_rdma_ch *ch, u64 req_tag, u64 lun,
2450                              u8 func)
2451 {
2452         struct srp_target_port *target = ch->target;
2453         struct srp_rport *rport = target->rport;
2454         struct ib_device *dev = target->srp_host->srp_dev->dev;
2455         struct srp_iu *iu;
2456         struct srp_tsk_mgmt *tsk_mgmt;
2457
2458         if (!ch->connected || target->qp_in_error)
2459                 return -1;
2460
2461         init_completion(&ch->tsk_mgmt_done);
2462
2463         /*
2464          * Lock the rport mutex to avoid that srp_create_ch_ib() is
2465          * invoked while a task management function is being sent.
2466          */
2467         mutex_lock(&rport->mutex);
2468         spin_lock_irq(&ch->lock);
2469         iu = __srp_get_tx_iu(ch, SRP_IU_TSK_MGMT);
2470         spin_unlock_irq(&ch->lock);
2471
2472         if (!iu) {
2473                 mutex_unlock(&rport->mutex);
2474
2475                 return -1;
2476         }
2477
2478         ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
2479                                    DMA_TO_DEVICE);
2480         tsk_mgmt = iu->buf;
2481         memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
2482
2483         tsk_mgmt->opcode        = SRP_TSK_MGMT;
2484         int_to_scsilun(lun, &tsk_mgmt->lun);
2485         tsk_mgmt->tag           = req_tag | SRP_TAG_TSK_MGMT;
2486         tsk_mgmt->tsk_mgmt_func = func;
2487         tsk_mgmt->task_tag      = req_tag;
2488
2489         ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
2490                                       DMA_TO_DEVICE);
2491         if (srp_post_send(ch, iu, sizeof(*tsk_mgmt))) {
2492                 srp_put_tx_iu(ch, iu, SRP_IU_TSK_MGMT);
2493                 mutex_unlock(&rport->mutex);
2494
2495                 return -1;
2496         }
2497         mutex_unlock(&rport->mutex);
2498
2499         if (!wait_for_completion_timeout(&ch->tsk_mgmt_done,
2500                                          msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
2501                 return -1;
2502
2503         return 0;
2504 }
2505
2506 static int srp_abort(struct scsi_cmnd *scmnd)
2507 {
2508         struct srp_target_port *target = host_to_target(scmnd->device->host);
2509         struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
2510         u32 tag;
2511         u16 ch_idx;
2512         struct srp_rdma_ch *ch;
2513         int ret;
2514
2515         shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
2516
2517         if (!req)
2518                 return SUCCESS;
2519         tag = blk_mq_unique_tag(scmnd->request);
2520         ch_idx = blk_mq_unique_tag_to_hwq(tag);
2521         if (WARN_ON_ONCE(ch_idx >= target->ch_count))
2522                 return SUCCESS;
2523         ch = &target->ch[ch_idx];
2524         if (!srp_claim_req(ch, req, NULL, scmnd))
2525                 return SUCCESS;
2526         shost_printk(KERN_ERR, target->scsi_host,
2527                      "Sending SRP abort for tag %#x\n", tag);
2528         if (srp_send_tsk_mgmt(ch, tag, scmnd->device->lun,
2529                               SRP_TSK_ABORT_TASK) == 0)
2530                 ret = SUCCESS;
2531         else if (target->rport->state == SRP_RPORT_LOST)
2532                 ret = FAST_IO_FAIL;
2533         else
2534                 ret = FAILED;
2535         srp_free_req(ch, req, scmnd, 0);
2536         scmnd->result = DID_ABORT << 16;
2537         scmnd->scsi_done(scmnd);
2538
2539         return ret;
2540 }
2541
2542 static int srp_reset_device(struct scsi_cmnd *scmnd)
2543 {
2544         struct srp_target_port *target = host_to_target(scmnd->device->host);
2545         struct srp_rdma_ch *ch;
2546         int i;
2547
2548         shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
2549
2550         ch = &target->ch[0];
2551         if (srp_send_tsk_mgmt(ch, SRP_TAG_NO_REQ, scmnd->device->lun,
2552                               SRP_TSK_LUN_RESET))
2553                 return FAILED;
2554         if (ch->tsk_mgmt_status)
2555                 return FAILED;
2556
2557         for (i = 0; i < target->ch_count; i++) {
2558                 ch = &target->ch[i];
2559                 for (i = 0; i < target->req_ring_size; ++i) {
2560                         struct srp_request *req = &ch->req_ring[i];
2561
2562                         srp_finish_req(ch, req, scmnd->device, DID_RESET << 16);
2563                 }
2564         }
2565
2566         return SUCCESS;
2567 }
2568
2569 static int srp_reset_host(struct scsi_cmnd *scmnd)
2570 {
2571         struct srp_target_port *target = host_to_target(scmnd->device->host);
2572
2573         shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
2574
2575         return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
2576 }
2577
2578 static int srp_slave_configure(struct scsi_device *sdev)
2579 {
2580         struct Scsi_Host *shost = sdev->host;
2581         struct srp_target_port *target = host_to_target(shost);
2582         struct request_queue *q = sdev->request_queue;
2583         unsigned long timeout;
2584
2585         if (sdev->type == TYPE_DISK) {
2586                 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
2587                 blk_queue_rq_timeout(q, timeout);
2588         }
2589
2590         return 0;
2591 }
2592
2593 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
2594                            char *buf)
2595 {
2596         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2597
2598         return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->id_ext));
2599 }
2600
2601 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
2602                              char *buf)
2603 {
2604         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2605
2606         return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->ioc_guid));
2607 }
2608
2609 static ssize_t show_service_id(struct device *dev,
2610                                struct device_attribute *attr, char *buf)
2611 {
2612         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2613
2614         return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->service_id));
2615 }
2616
2617 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
2618                          char *buf)
2619 {
2620         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2621
2622         return sprintf(buf, "0x%04x\n", be16_to_cpu(target->pkey));
2623 }
2624
2625 static ssize_t show_sgid(struct device *dev, struct device_attribute *attr,
2626                          char *buf)
2627 {
2628         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2629
2630         return sprintf(buf, "%pI6\n", target->sgid.raw);
2631 }
2632
2633 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
2634                          char *buf)
2635 {
2636         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2637         struct srp_rdma_ch *ch = &target->ch[0];
2638
2639         return sprintf(buf, "%pI6\n", ch->path.dgid.raw);
2640 }
2641
2642 static ssize_t show_orig_dgid(struct device *dev,
2643                               struct device_attribute *attr, char *buf)
2644 {
2645         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2646
2647         return sprintf(buf, "%pI6\n", target->orig_dgid.raw);
2648 }
2649
2650 static ssize_t show_req_lim(struct device *dev,
2651                             struct device_attribute *attr, char *buf)
2652 {
2653         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2654         struct srp_rdma_ch *ch;
2655         int i, req_lim = INT_MAX;
2656
2657         for (i = 0; i < target->ch_count; i++) {
2658                 ch = &target->ch[i];
2659                 req_lim = min(req_lim, ch->req_lim);
2660         }
2661         return sprintf(buf, "%d\n", req_lim);
2662 }
2663
2664 static ssize_t show_zero_req_lim(struct device *dev,
2665                                  struct device_attribute *attr, char *buf)
2666 {
2667         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2668
2669         return sprintf(buf, "%d\n", target->zero_req_lim);
2670 }
2671
2672 static ssize_t show_local_ib_port(struct device *dev,
2673                                   struct device_attribute *attr, char *buf)
2674 {
2675         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2676
2677         return sprintf(buf, "%d\n", target->srp_host->port);
2678 }
2679
2680 static ssize_t show_local_ib_device(struct device *dev,
2681                                     struct device_attribute *attr, char *buf)
2682 {
2683         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2684
2685         return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
2686 }
2687
2688 static ssize_t show_ch_count(struct device *dev, struct device_attribute *attr,
2689                              char *buf)
2690 {
2691         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2692
2693         return sprintf(buf, "%d\n", target->ch_count);
2694 }
2695
2696 static ssize_t show_comp_vector(struct device *dev,
2697                                 struct device_attribute *attr, char *buf)
2698 {
2699         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2700
2701         return sprintf(buf, "%d\n", target->comp_vector);
2702 }
2703
2704 static ssize_t show_tl_retry_count(struct device *dev,
2705                                    struct device_attribute *attr, char *buf)
2706 {
2707         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2708
2709         return sprintf(buf, "%d\n", target->tl_retry_count);
2710 }
2711
2712 static ssize_t show_cmd_sg_entries(struct device *dev,
2713                                    struct device_attribute *attr, char *buf)
2714 {
2715         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2716
2717         return sprintf(buf, "%u\n", target->cmd_sg_cnt);
2718 }
2719
2720 static ssize_t show_allow_ext_sg(struct device *dev,
2721                                  struct device_attribute *attr, char *buf)
2722 {
2723         struct srp_target_port *target = host_to_target(class_to_shost(dev));
2724
2725         return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
2726 }
2727
2728 static DEVICE_ATTR(id_ext,          S_IRUGO, show_id_ext,          NULL);
2729 static DEVICE_ATTR(ioc_guid,        S_IRUGO, show_ioc_guid,        NULL);
2730 static DEVICE_ATTR(service_id,      S_IRUGO, show_service_id,      NULL);
2731 static DEVICE_ATTR(pkey,            S_IRUGO, show_pkey,            NULL);
2732 static DEVICE_ATTR(sgid,            S_IRUGO, show_sgid,            NULL);
2733 static DEVICE_ATTR(dgid,            S_IRUGO, show_dgid,            NULL);
2734 static DEVICE_ATTR(orig_dgid,       S_IRUGO, show_orig_dgid,       NULL);
2735 static DEVICE_ATTR(req_lim,         S_IRUGO, show_req_lim,         NULL);
2736 static DEVICE_ATTR(zero_req_lim,    S_IRUGO, show_zero_req_lim,    NULL);
2737 static DEVICE_ATTR(local_ib_port,   S_IRUGO, show_local_ib_port,   NULL);
2738 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
2739 static DEVICE_ATTR(ch_count,        S_IRUGO, show_ch_count,        NULL);
2740 static DEVICE_ATTR(comp_vector,     S_IRUGO, show_comp_vector,     NULL);
2741 static DEVICE_ATTR(tl_retry_count,  S_IRUGO, show_tl_retry_count,  NULL);
2742 static DEVICE_ATTR(cmd_sg_entries,  S_IRUGO, show_cmd_sg_entries,  NULL);
2743 static DEVICE_ATTR(allow_ext_sg,    S_IRUGO, show_allow_ext_sg,    NULL);
2744
2745 static struct device_attribute *srp_host_attrs[] = {
2746         &dev_attr_id_ext,
2747         &dev_attr_ioc_guid,
2748         &dev_attr_service_id,
2749         &dev_attr_pkey,
2750         &dev_attr_sgid,
2751         &dev_attr_dgid,
2752         &dev_attr_orig_dgid,
2753         &dev_attr_req_lim,
2754         &dev_attr_zero_req_lim,
2755         &dev_attr_local_ib_port,
2756         &dev_attr_local_ib_device,
2757         &dev_attr_ch_count,
2758         &dev_attr_comp_vector,
2759         &dev_attr_tl_retry_count,
2760         &dev_attr_cmd_sg_entries,
2761         &dev_attr_allow_ext_sg,
2762         NULL
2763 };
2764
2765 static struct scsi_host_template srp_template = {
2766         .module                         = THIS_MODULE,
2767         .name                           = "InfiniBand SRP initiator",
2768         .proc_name                      = DRV_NAME,
2769         .slave_configure                = srp_slave_configure,
2770         .info                           = srp_target_info,
2771         .queuecommand                   = srp_queuecommand,
2772         .change_queue_depth             = srp_change_queue_depth,
2773         .eh_abort_handler               = srp_abort,
2774         .eh_device_reset_handler        = srp_reset_device,
2775         .eh_host_reset_handler          = srp_reset_host,
2776         .skip_settle_delay              = true,
2777         .sg_tablesize                   = SRP_DEF_SG_TABLESIZE,
2778         .can_queue                      = SRP_DEFAULT_CMD_SQ_SIZE,
2779         .this_id                        = -1,
2780         .cmd_per_lun                    = SRP_DEFAULT_CMD_SQ_SIZE,
2781         .use_clustering                 = ENABLE_CLUSTERING,
2782         .shost_attrs                    = srp_host_attrs,
2783         .track_queue_depth              = 1,
2784 };
2785
2786 static int srp_sdev_count(struct Scsi_Host *host)
2787 {
2788         struct scsi_device *sdev;
2789         int c = 0;
2790
2791         shost_for_each_device(sdev, host)
2792                 c++;
2793
2794         return c;
2795 }
2796
2797 /*
2798  * Return values:
2799  * < 0 upon failure. Caller is responsible for SRP target port cleanup.
2800  * 0 and target->state == SRP_TARGET_REMOVED if asynchronous target port
2801  *    removal has been scheduled.
2802  * 0 and target->state != SRP_TARGET_REMOVED upon success.
2803  */
2804 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
2805 {
2806         struct srp_rport_identifiers ids;
2807         struct srp_rport *rport;
2808
2809         target->state = SRP_TARGET_SCANNING;
2810         sprintf(target->target_name, "SRP.T10:%016llX",
2811                 be64_to_cpu(target->id_ext));
2812
2813         if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
2814                 return -ENODEV;
2815
2816         memcpy(ids.port_id, &target->id_ext, 8);
2817         memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2818         ids.roles = SRP_RPORT_ROLE_TARGET;
2819         rport = srp_rport_add(target->scsi_host, &ids);
2820         if (IS_ERR(rport)) {
2821                 scsi_remove_host(target->scsi_host);
2822                 return PTR_ERR(rport);
2823         }
2824
2825         rport->lld_data = target;
2826         target->rport = rport;
2827
2828         spin_lock(&host->target_lock);
2829         list_add_tail(&target->list, &host->target_list);
2830         spin_unlock(&host->target_lock);
2831
2832         scsi_scan_target(&target->scsi_host->shost_gendev,
2833                          0, target->scsi_id, SCAN_WILD_CARD, 0);
2834
2835         if (srp_connected_ch(target) < target->ch_count ||
2836             target->qp_in_error) {
2837                 shost_printk(KERN_INFO, target->scsi_host,
2838                              PFX "SCSI scan failed - removing SCSI host\n");
2839                 srp_queue_remove_work(target);
2840                 goto out;
2841         }
2842
2843         pr_debug(PFX "%s: SCSI scan succeeded - detected %d LUNs\n",
2844                  dev_name(&target->scsi_host->shost_gendev),
2845                  srp_sdev_count(target->scsi_host));
2846
2847         spin_lock_irq(&target->lock);
2848         if (target->state == SRP_TARGET_SCANNING)
2849                 target->state = SRP_TARGET_LIVE;
2850         spin_unlock_irq(&target->lock);
2851
2852 out:
2853         return 0;
2854 }
2855
2856 static void srp_release_dev(struct device *dev)
2857 {
2858         struct srp_host *host =
2859                 container_of(dev, struct srp_host, dev);
2860
2861         complete(&host->released);
2862 }
2863
2864 static struct class srp_class = {
2865         .name    = "infiniband_srp",
2866         .dev_release = srp_release_dev
2867 };
2868
2869 /**
2870  * srp_conn_unique() - check whether the connection to a target is unique
2871  * @host:   SRP host.
2872  * @target: SRP target port.
2873  */
2874 static bool srp_conn_unique(struct srp_host *host,
2875                             struct srp_target_port *target)
2876 {
2877         struct srp_target_port *t;
2878         bool ret = false;
2879
2880         if (target->state == SRP_TARGET_REMOVED)
2881                 goto out;
2882
2883         ret = true;
2884
2885         spin_lock(&host->target_lock);
2886         list_for_each_entry(t, &host->target_list, list) {
2887                 if (t != target &&
2888                     target->id_ext == t->id_ext &&
2889                     target->ioc_guid == t->ioc_guid &&
2890                     target->initiator_ext == t->initiator_ext) {
2891                         ret = false;
2892                         break;
2893                 }
2894         }
2895         spin_unlock(&host->target_lock);
2896
2897 out:
2898         return ret;
2899 }
2900
2901 /*
2902  * Target ports are added by writing
2903  *
2904  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2905  *     pkey=<P_Key>,service_id=<service ID>
2906  *
2907  * to the add_target sysfs attribute.
2908  */
2909 enum {
2910         SRP_OPT_ERR             = 0,
2911         SRP_OPT_ID_EXT          = 1 << 0,
2912         SRP_OPT_IOC_GUID        = 1 << 1,
2913         SRP_OPT_DGID            = 1 << 2,
2914         SRP_OPT_PKEY            = 1 << 3,
2915         SRP_OPT_SERVICE_ID      = 1 << 4,
2916         SRP_OPT_MAX_SECT        = 1 << 5,
2917         SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2918         SRP_OPT_IO_CLASS        = 1 << 7,
2919         SRP_OPT_INITIATOR_EXT   = 1 << 8,
2920         SRP_OPT_CMD_SG_ENTRIES  = 1 << 9,
2921         SRP_OPT_ALLOW_EXT_SG    = 1 << 10,
2922         SRP_OPT_SG_TABLESIZE    = 1 << 11,
2923         SRP_OPT_COMP_VECTOR     = 1 << 12,
2924         SRP_OPT_TL_RETRY_COUNT  = 1 << 13,
2925         SRP_OPT_QUEUE_SIZE      = 1 << 14,
2926         SRP_OPT_ALL             = (SRP_OPT_ID_EXT       |
2927                                    SRP_OPT_IOC_GUID     |
2928                                    SRP_OPT_DGID         |
2929                                    SRP_OPT_PKEY         |
2930                                    SRP_OPT_SERVICE_ID),
2931 };
2932
2933 static const match_table_t srp_opt_tokens = {
2934         { SRP_OPT_ID_EXT,               "id_ext=%s"             },
2935         { SRP_OPT_IOC_GUID,             "ioc_guid=%s"           },
2936         { SRP_OPT_DGID,                 "dgid=%s"               },
2937         { SRP_OPT_PKEY,                 "pkey=%x"               },
2938         { SRP_OPT_SERVICE_ID,           "service_id=%s"         },
2939         { SRP_OPT_MAX_SECT,             "max_sect=%d"           },
2940         { SRP_OPT_MAX_CMD_PER_LUN,      "max_cmd_per_lun=%d"    },
2941         { SRP_OPT_IO_CLASS,             "io_class=%x"           },
2942         { SRP_OPT_INITIATOR_EXT,        "initiator_ext=%s"      },
2943         { SRP_OPT_CMD_SG_ENTRIES,       "cmd_sg_entries=%u"     },
2944         { SRP_OPT_ALLOW_EXT_SG,         "allow_ext_sg=%u"       },
2945         { SRP_OPT_SG_TABLESIZE,         "sg_tablesize=%u"       },
2946         { SRP_OPT_COMP_VECTOR,          "comp_vector=%u"        },
2947         { SRP_OPT_TL_RETRY_COUNT,       "tl_retry_count=%u"     },
2948         { SRP_OPT_QUEUE_SIZE,           "queue_size=%d"         },
2949         { SRP_OPT_ERR,                  NULL                    }
2950 };
2951
2952 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2953 {
2954         char *options, *sep_opt;
2955         char *p;
2956         char dgid[3];
2957         substring_t args[MAX_OPT_ARGS];
2958         int opt_mask = 0;
2959         int token;
2960         int ret = -EINVAL;
2961         int i;
2962
2963         options = kstrdup(buf, GFP_KERNEL);
2964         if (!options)
2965                 return -ENOMEM;
2966
2967         sep_opt = options;
2968         while ((p = strsep(&sep_opt, ",\n")) != NULL) {
2969                 if (!*p)
2970                         continue;
2971
2972                 token = match_token(p, srp_opt_tokens, args);
2973                 opt_mask |= token;
2974
2975                 switch (token) {
2976                 case SRP_OPT_ID_EXT:
2977                         p = match_strdup(args);
2978                         if (!p) {
2979                                 ret = -ENOMEM;
2980                                 goto out;
2981                         }
2982                         target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2983                         kfree(p);
2984                         break;
2985
2986                 case SRP_OPT_IOC_GUID:
2987                         p = match_strdup(args);
2988                         if (!p) {
2989                                 ret = -ENOMEM;
2990                                 goto out;
2991                         }
2992                         target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2993                         kfree(p);
2994                         break;
2995
2996                 case SRP_OPT_DGID:
2997                         p = match_strdup(args);
2998                         if (!p) {
2999                                 ret = -ENOMEM;
3000                                 goto out;
3001                         }
3002                         if (strlen(p) != 32) {
3003                                 pr_warn("bad dest GID parameter '%s'\n", p);
3004                                 kfree(p);
3005                                 goto out;
3006                         }
3007
3008                         for (i = 0; i < 16; ++i) {
3009                                 strlcpy(dgid, p + i * 2, sizeof(dgid));
3010                                 if (sscanf(dgid, "%hhx",
3011                                            &target->orig_dgid.raw[i]) < 1) {
3012                                         ret = -EINVAL;
3013                                         kfree(p);
3014                                         goto out;
3015                                 }
3016                         }
3017                         kfree(p);
3018                         break;
3019
3020                 case SRP_OPT_PKEY:
3021                         if (match_hex(args, &token)) {
3022                                 pr_warn("bad P_Key parameter '%s'\n", p);
3023                                 goto out;
3024                         }
3025                         target->pkey = cpu_to_be16(token);
3026                         break;
3027
3028                 case SRP_OPT_SERVICE_ID:
3029                         p = match_strdup(args);
3030                         if (!p) {
3031                                 ret = -ENOMEM;
3032                                 goto out;
3033                         }
3034                         target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
3035                         kfree(p);
3036                         break;
3037
3038                 case SRP_OPT_MAX_SECT:
3039                         if (match_int(args, &token)) {
3040                                 pr_warn("bad max sect parameter '%s'\n", p);
3041                                 goto out;
3042                         }
3043                         target->scsi_host->max_sectors = token;
3044                         break;
3045
3046                 case SRP_OPT_QUEUE_SIZE:
3047                         if (match_int(args, &token) || token < 1) {
3048                                 pr_warn("bad queue_size parameter '%s'\n", p);
3049                                 goto out;
3050                         }
3051                         target->scsi_host->can_queue = token;
3052                         target->queue_size = token + SRP_RSP_SQ_SIZE +
3053                                              SRP_TSK_MGMT_SQ_SIZE;
3054                         if (!(opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3055                                 target->scsi_host->cmd_per_lun = token;
3056                         break;
3057
3058                 case SRP_OPT_MAX_CMD_PER_LUN:
3059                         if (match_int(args, &token) || token < 1) {
3060                                 pr_warn("bad max cmd_per_lun parameter '%s'\n",
3061                                         p);
3062                                 goto out;
3063                         }
3064                         target->scsi_host->cmd_per_lun = token;
3065                         break;
3066
3067                 case SRP_OPT_IO_CLASS:
3068                         if (match_hex(args, &token)) {
3069                                 pr_warn("bad IO class parameter '%s'\n", p);
3070                                 goto out;
3071                         }
3072                         if (token != SRP_REV10_IB_IO_CLASS &&
3073                             token != SRP_REV16A_IB_IO_CLASS) {
3074                                 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
3075                                         token, SRP_REV10_IB_IO_CLASS,
3076                                         SRP_REV16A_IB_IO_CLASS);
3077                                 goto out;
3078                         }
3079                         target->io_class = token;
3080                         break;
3081
3082                 case SRP_OPT_INITIATOR_EXT:
3083                         p = match_strdup(args);
3084                         if (!p) {
3085                                 ret = -ENOMEM;
3086                                 goto out;
3087                         }
3088                         target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
3089                         kfree(p);
3090                         break;
3091
3092                 case SRP_OPT_CMD_SG_ENTRIES:
3093                         if (match_int(args, &token) || token < 1 || token > 255) {
3094                                 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
3095                                         p);
3096                                 goto out;
3097                         }
3098                         target->cmd_sg_cnt = token;
3099                         break;
3100
3101                 case SRP_OPT_ALLOW_EXT_SG:
3102                         if (match_int(args, &token)) {
3103                                 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
3104                                 goto out;
3105                         }
3106                         target->allow_ext_sg = !!token;
3107                         break;
3108
3109                 case SRP_OPT_SG_TABLESIZE:
3110                         if (match_int(args, &token) || token < 1 ||
3111                                         token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
3112                                 pr_warn("bad max sg_tablesize parameter '%s'\n",
3113                                         p);
3114                                 goto out;
3115                         }
3116                         target->sg_tablesize = token;
3117                         break;
3118
3119                 case SRP_OPT_COMP_VECTOR:
3120                         if (match_int(args, &token) || token < 0) {
3121                                 pr_warn("bad comp_vector parameter '%s'\n", p);
3122                                 goto out;
3123                         }
3124                         target->comp_vector = token;
3125                         break;
3126
3127                 case SRP_OPT_TL_RETRY_COUNT:
3128                         if (match_int(args, &token) || token < 2 || token > 7) {
3129                                 pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
3130                                         p);
3131                                 goto out;
3132                         }
3133                         target->tl_retry_count = token;
3134                         break;
3135
3136                 default:
3137                         pr_warn("unknown parameter or missing value '%s' in target creation request\n",
3138                                 p);
3139                         goto out;
3140                 }
3141         }
3142
3143         if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
3144                 ret = 0;
3145         else
3146                 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
3147                         if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
3148                             !(srp_opt_tokens[i].token & opt_mask))
3149                                 pr_warn("target creation request is missing parameter '%s'\n",
3150                                         srp_opt_tokens[i].pattern);
3151
3152         if (target->scsi_host->cmd_per_lun > target->scsi_host->can_queue
3153             && (opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3154                 pr_warn("cmd_per_lun = %d > queue_size = %d\n",
3155                         target->scsi_host->cmd_per_lun,
3156                         target->scsi_host->can_queue);
3157
3158 out:
3159         kfree(options);
3160         return ret;
3161 }
3162
3163 static ssize_t srp_create_target(struct device *dev,
3164                                  struct device_attribute *attr,
3165                                  const char *buf, size_t count)
3166 {
3167         struct srp_host *host =
3168                 container_of(dev, struct srp_host, dev);
3169         struct Scsi_Host *target_host;
3170         struct srp_target_port *target;
3171         struct srp_rdma_ch *ch;
3172         struct srp_device *srp_dev = host->srp_dev;
3173         struct ib_device *ibdev = srp_dev->dev;
3174         int ret, node_idx, node, cpu, i;
3175         bool multich = false;
3176
3177         target_host = scsi_host_alloc(&srp_template,
3178                                       sizeof (struct srp_target_port));
3179         if (!target_host)
3180                 return -ENOMEM;
3181
3182         target_host->transportt  = ib_srp_transport_template;
3183         target_host->max_channel = 0;
3184         target_host->max_id      = 1;
3185         target_host->max_lun     = -1LL;
3186         target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
3187
3188         target = host_to_target(target_host);
3189
3190         target->io_class        = SRP_REV16A_IB_IO_CLASS;
3191         target->scsi_host       = target_host;
3192         target->srp_host        = host;
3193         target->lkey            = host->srp_dev->pd->local_dma_lkey;
3194         target->global_mr       = host->srp_dev->global_mr;
3195         target->cmd_sg_cnt      = cmd_sg_entries;
3196         target->sg_tablesize    = indirect_sg_entries ? : cmd_sg_entries;
3197         target->allow_ext_sg    = allow_ext_sg;
3198         target->tl_retry_count  = 7;
3199         target->queue_size      = SRP_DEFAULT_QUEUE_SIZE;
3200
3201         /*
3202          * Avoid that the SCSI host can be removed by srp_remove_target()
3203          * before this function returns.
3204          */
3205         scsi_host_get(target->scsi_host);
3206
3207         mutex_lock(&host->add_target_mutex);
3208
3209         ret = srp_parse_options(buf, target);
3210         if (ret)
3211                 goto out;
3212
3213         target->req_ring_size = target->queue_size - SRP_TSK_MGMT_SQ_SIZE;
3214
3215         if (!srp_conn_unique(target->srp_host, target)) {
3216                 shost_printk(KERN_INFO, target->scsi_host,
3217                              PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
3218                              be64_to_cpu(target->id_ext),
3219                              be64_to_cpu(target->ioc_guid),
3220                              be64_to_cpu(target->initiator_ext));
3221                 ret = -EEXIST;
3222                 goto out;
3223         }
3224
3225         if (!srp_dev->has_fmr && !srp_dev->has_fr && !target->allow_ext_sg &&
3226             target->cmd_sg_cnt < target->sg_tablesize) {
3227                 pr_warn("No MR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
3228                 target->sg_tablesize = target->cmd_sg_cnt;
3229         }
3230
3231         target_host->sg_tablesize = target->sg_tablesize;
3232         target->mr_pool_size = target->scsi_host->can_queue;
3233         target->indirect_size = target->sg_tablesize *
3234                                 sizeof (struct srp_direct_buf);
3235         target->max_iu_len = sizeof (struct srp_cmd) +
3236                              sizeof (struct srp_indirect_buf) +
3237                              target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
3238
3239         INIT_WORK(&target->tl_err_work, srp_tl_err_work);
3240         INIT_WORK(&target->remove_work, srp_remove_work);
3241         spin_lock_init(&target->lock);
3242         ret = ib_query_gid(ibdev, host->port, 0, &target->sgid, NULL);
3243         if (ret)
3244                 goto out;
3245
3246         ret = -ENOMEM;
3247         target->ch_count = max_t(unsigned, num_online_nodes(),
3248                                  min(ch_count ? :
3249                                      min(4 * num_online_nodes(),
3250                                          ibdev->num_comp_vectors),
3251                                      num_online_cpus()));
3252         target->ch = kcalloc(target->ch_count, sizeof(*target->ch),
3253                              GFP_KERNEL);
3254         if (!target->ch)
3255                 goto out;
3256
3257         node_idx = 0;
3258         for_each_online_node(node) {
3259                 const int ch_start = (node_idx * target->ch_count /
3260                                       num_online_nodes());
3261                 const int ch_end = ((node_idx + 1) * target->ch_count /
3262                                     num_online_nodes());
3263                 const int cv_start = (node_idx * ibdev->num_comp_vectors /
3264                                       num_online_nodes() + target->comp_vector)
3265                                      % ibdev->num_comp_vectors;
3266                 const int cv_end = ((node_idx + 1) * ibdev->num_comp_vectors /
3267                                     num_online_nodes() + target->comp_vector)
3268                                    % ibdev->num_comp_vectors;
3269                 int cpu_idx = 0;
3270
3271                 for_each_online_cpu(cpu) {
3272                         if (cpu_to_node(cpu) != node)
3273                                 continue;
3274                         if (ch_start + cpu_idx >= ch_end)
3275                                 continue;
3276                         ch = &target->ch[ch_start + cpu_idx];
3277                         ch->target = target;
3278                         ch->comp_vector = cv_start == cv_end ? cv_start :
3279                                 cv_start + cpu_idx % (cv_end - cv_start);
3280                         spin_lock_init(&ch->lock);
3281                         INIT_LIST_HEAD(&ch->free_tx);
3282                         ret = srp_new_cm_id(ch);
3283                         if (ret)
3284                                 goto err_disconnect;
3285
3286                         ret = srp_create_ch_ib(ch);
3287                         if (ret)
3288                                 goto err_disconnect;
3289
3290                         ret = srp_alloc_req_data(ch);
3291                         if (ret)
3292                                 goto err_disconnect;
3293
3294                         ret = srp_connect_ch(ch, multich);
3295                         if (ret) {
3296                                 shost_printk(KERN_ERR, target->scsi_host,
3297                                              PFX "Connection %d/%d failed\n",
3298                                              ch_start + cpu_idx,
3299                                              target->ch_count);
3300                                 if (node_idx == 0 && cpu_idx == 0) {
3301                                         goto err_disconnect;
3302                                 } else {
3303                                         srp_free_ch_ib(target, ch);
3304                                         srp_free_req_data(target, ch);
3305                                         target->ch_count = ch - target->ch;
3306                                         goto connected;
3307                                 }
3308                         }
3309
3310                         multich = true;
3311                         cpu_idx++;
3312                 }
3313                 node_idx++;
3314         }
3315
3316 connected:
3317         target->scsi_host->nr_hw_queues = target->ch_count;
3318
3319         ret = srp_add_target(host, target);
3320         if (ret)
3321                 goto err_disconnect;
3322
3323         if (target->state != SRP_TARGET_REMOVED) {
3324                 shost_printk(KERN_DEBUG, target->scsi_host, PFX
3325                              "new target: id_ext %016llx ioc_guid %016llx pkey %04x service_id %016llx sgid %pI6 dgid %pI6\n",
3326                              be64_to_cpu(target->id_ext),
3327                              be64_to_cpu(target->ioc_guid),
3328                              be16_to_cpu(target->pkey),
3329                              be64_to_cpu(target->service_id),
3330                              target->sgid.raw, target->orig_dgid.raw);
3331         }
3332
3333         ret = count;
3334
3335 out:
3336         mutex_unlock(&host->add_target_mutex);
3337
3338         scsi_host_put(target->scsi_host);
3339         if (ret < 0)
3340                 scsi_host_put(target->scsi_host);
3341
3342         return ret;
3343
3344 err_disconnect:
3345         srp_disconnect_target(target);
3346
3347         for (i = 0; i < target->ch_count; i++) {
3348                 ch = &target->ch[i];
3349                 srp_free_ch_ib(target, ch);
3350                 srp_free_req_data(target, ch);
3351         }
3352
3353         kfree(target->ch);
3354         goto out;
3355 }
3356
3357 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
3358
3359 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
3360                           char *buf)
3361 {
3362         struct srp_host *host = container_of(dev, struct srp_host, dev);
3363
3364         return sprintf(buf, "%s\n", host->srp_dev->dev->name);
3365 }
3366
3367 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
3368
3369 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
3370                          char *buf)
3371 {
3372         struct srp_host *host = container_of(dev, struct srp_host, dev);
3373
3374         return sprintf(buf, "%d\n", host->port);
3375 }
3376
3377 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
3378
3379 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
3380 {
3381         struct srp_host *host;
3382
3383         host = kzalloc(sizeof *host, GFP_KERNEL);
3384         if (!host)
3385                 return NULL;
3386
3387         INIT_LIST_HEAD(&host->target_list);
3388         spin_lock_init(&host->target_lock);
3389         init_completion(&host->released);
3390         mutex_init(&host->add_target_mutex);
3391         host->srp_dev = device;
3392         host->port = port;
3393
3394         host->dev.class = &srp_class;
3395         host->dev.parent = device->dev->dma_device;
3396         dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
3397
3398         if (device_register(&host->dev))
3399                 goto free_host;
3400         if (device_create_file(&host->dev, &dev_attr_add_target))
3401                 goto err_class;
3402         if (device_create_file(&host->dev, &dev_attr_ibdev))
3403                 goto err_class;
3404         if (device_create_file(&host->dev, &dev_attr_port))
3405                 goto err_class;
3406
3407         return host;
3408
3409 err_class:
3410         device_unregister(&host->dev);
3411
3412 free_host:
3413         kfree(host);
3414
3415         return NULL;
3416 }
3417
3418 static void srp_add_one(struct ib_device *device)
3419 {
3420         struct srp_device *srp_dev;
3421         struct srp_host *host;
3422         int mr_page_shift, p;
3423         u64 max_pages_per_mr;
3424
3425         srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
3426         if (!srp_dev)
3427                 return;
3428
3429         srp_dev->has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
3430                             device->map_phys_fmr && device->unmap_fmr);
3431         srp_dev->has_fr = (device->attrs.device_cap_flags &
3432                            IB_DEVICE_MEM_MGT_EXTENSIONS);
3433         if (!srp_dev->has_fmr && !srp_dev->has_fr)
3434                 dev_warn(&device->dev, "neither FMR nor FR is supported\n");
3435
3436         srp_dev->use_fast_reg = (srp_dev->has_fr &&
3437                                  (!srp_dev->has_fmr || prefer_fr));
3438         srp_dev->use_fmr = !srp_dev->use_fast_reg && srp_dev->has_fmr;
3439
3440         /*
3441          * Use the smallest page size supported by the HCA, down to a
3442          * minimum of 4096 bytes. We're unlikely to build large sglists
3443          * out of smaller entries.
3444          */
3445         mr_page_shift           = max(12, ffs(device->attrs.page_size_cap) - 1);
3446         srp_dev->mr_page_size   = 1 << mr_page_shift;
3447         srp_dev->mr_page_mask   = ~((u64) srp_dev->mr_page_size - 1);
3448         max_pages_per_mr        = device->attrs.max_mr_size;
3449         do_div(max_pages_per_mr, srp_dev->mr_page_size);
3450         srp_dev->max_pages_per_mr = min_t(u64, SRP_MAX_PAGES_PER_MR,
3451                                           max_pages_per_mr);
3452         if (srp_dev->use_fast_reg) {
3453                 srp_dev->max_pages_per_mr =
3454                         min_t(u32, srp_dev->max_pages_per_mr,
3455                               device->attrs.max_fast_reg_page_list_len);
3456         }
3457         srp_dev->mr_max_size    = srp_dev->mr_page_size *
3458                                    srp_dev->max_pages_per_mr;
3459         pr_debug("%s: mr_page_shift = %d, device->max_mr_size = %#llx, device->max_fast_reg_page_list_len = %u, max_pages_per_mr = %d, mr_max_size = %#x\n",
3460                  device->name, mr_page_shift, device->attrs.max_mr_size,
3461                  device->attrs.max_fast_reg_page_list_len,
3462                  srp_dev->max_pages_per_mr, srp_dev->mr_max_size);
3463
3464         INIT_LIST_HEAD(&srp_dev->dev_list);
3465
3466         srp_dev->dev = device;
3467         srp_dev->pd  = ib_alloc_pd(device);
3468         if (IS_ERR(srp_dev->pd))
3469                 goto free_dev;
3470
3471         if (!register_always || (!srp_dev->has_fmr && !srp_dev->has_fr)) {
3472                 srp_dev->global_mr = ib_get_dma_mr(srp_dev->pd,
3473                                                    IB_ACCESS_LOCAL_WRITE |
3474                                                    IB_ACCESS_REMOTE_READ |
3475                                                    IB_ACCESS_REMOTE_WRITE);
3476                 if (IS_ERR(srp_dev->global_mr))
3477                         goto err_pd;
3478         } else {
3479                 srp_dev->global_mr = NULL;
3480         }
3481
3482         for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
3483                 host = srp_add_port(srp_dev, p);
3484                 if (host)
3485                         list_add_tail(&host->list, &srp_dev->dev_list);
3486         }
3487
3488         ib_set_client_data(device, &srp_client, srp_dev);
3489         return;
3490
3491 err_pd:
3492         ib_dealloc_pd(srp_dev->pd);
3493
3494 free_dev:
3495         kfree(srp_dev);
3496 }
3497
3498 static void srp_remove_one(struct ib_device *device, void *client_data)
3499 {
3500         struct srp_device *srp_dev;
3501         struct srp_host *host, *tmp_host;
3502         struct srp_target_port *target;
3503
3504         srp_dev = client_data;
3505         if (!srp_dev)
3506                 return;
3507
3508         list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
3509                 device_unregister(&host->dev);
3510                 /*
3511                  * Wait for the sysfs entry to go away, so that no new
3512                  * target ports can be created.
3513                  */
3514                 wait_for_completion(&host->released);
3515
3516                 /*
3517                  * Remove all target ports.
3518                  */
3519                 spin_lock(&host->target_lock);
3520                 list_for_each_entry(target, &host->target_list, list)
3521                         srp_queue_remove_work(target);
3522                 spin_unlock(&host->target_lock);
3523
3524                 /*
3525                  * Wait for tl_err and target port removal tasks.
3526                  */
3527                 flush_workqueue(system_long_wq);
3528                 flush_workqueue(srp_remove_wq);
3529
3530                 kfree(host);
3531         }
3532
3533         if (srp_dev->global_mr)
3534                 ib_dereg_mr(srp_dev->global_mr);
3535         ib_dealloc_pd(srp_dev->pd);
3536
3537         kfree(srp_dev);
3538 }
3539
3540 static struct srp_function_template ib_srp_transport_functions = {
3541         .has_rport_state         = true,
3542         .reset_timer_if_blocked  = true,
3543         .reconnect_delay         = &srp_reconnect_delay,
3544         .fast_io_fail_tmo        = &srp_fast_io_fail_tmo,
3545         .dev_loss_tmo            = &srp_dev_loss_tmo,
3546         .reconnect               = srp_rport_reconnect,
3547         .rport_delete            = srp_rport_delete,
3548         .terminate_rport_io      = srp_terminate_io,
3549 };
3550
3551 static int __init srp_init_module(void)
3552 {
3553         int ret;
3554
3555         if (srp_sg_tablesize) {
3556                 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
3557                 if (!cmd_sg_entries)
3558                         cmd_sg_entries = srp_sg_tablesize;
3559         }
3560
3561         if (!cmd_sg_entries)
3562                 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
3563
3564         if (cmd_sg_entries > 255) {
3565                 pr_warn("Clamping cmd_sg_entries to 255\n");
3566                 cmd_sg_entries = 255;
3567         }
3568
3569         if (!indirect_sg_entries)
3570                 indirect_sg_entries = cmd_sg_entries;
3571         else if (indirect_sg_entries < cmd_sg_entries) {
3572                 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
3573                         cmd_sg_entries);
3574                 indirect_sg_entries = cmd_sg_entries;
3575         }
3576
3577         srp_remove_wq = create_workqueue("srp_remove");
3578         if (!srp_remove_wq) {
3579                 ret = -ENOMEM;
3580                 goto out;
3581         }
3582
3583         ret = -ENOMEM;
3584         ib_srp_transport_template =
3585                 srp_attach_transport(&ib_srp_transport_functions);
3586         if (!ib_srp_transport_template)
3587                 goto destroy_wq;
3588
3589         ret = class_register(&srp_class);
3590         if (ret) {
3591                 pr_err("couldn't register class infiniband_srp\n");
3592                 goto release_tr;
3593         }
3594
3595         ib_sa_register_client(&srp_sa_client);
3596
3597         ret = ib_register_client(&srp_client);
3598         if (ret) {
3599                 pr_err("couldn't register IB client\n");
3600                 goto unreg_sa;
3601         }
3602
3603 out:
3604         return ret;
3605
3606 unreg_sa:
3607         ib_sa_unregister_client(&srp_sa_client);
3608         class_unregister(&srp_class);
3609
3610 release_tr:
3611         srp_release_transport(ib_srp_transport_template);
3612
3613 destroy_wq:
3614         destroy_workqueue(srp_remove_wq);
3615         goto out;
3616 }
3617
3618 static void __exit srp_cleanup_module(void)
3619 {
3620         ib_unregister_client(&srp_client);
3621         ib_sa_unregister_client(&srp_sa_client);
3622         class_unregister(&srp_class);
3623         srp_release_transport(ib_srp_transport_template);
3624         destroy_workqueue(srp_remove_wq);
3625 }
3626
3627 module_init(srp_init_module);
3628 module_exit(srp_cleanup_module);