drivers/target/target_core_user.c

   1 /*
   2  * Copyright (C) 2013 Shaohua Li <shli@kernel.org>
   3  * Copyright (C) 2014 Red Hat, Inc.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along with
  15  * this program; if not, write to the Free Software Foundation, Inc.,
  16  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  17  */
  18
  19 #include <linux/spinlock.h>
  20 #include <linux/module.h>
  21 #include <linux/idr.h>
  22 #include <linux/timer.h>
  23 #include <linux/parser.h>
  24 #include <scsi/scsi.h>
  25 #include <scsi/scsi_host.h>
  26 #include <linux/uio_driver.h>
  27 #include <net/genetlink.h>
  28 #include <target/target_core_base.h>
  29 #include <target/target_core_fabric.h>
  30 #include <target/target_core_backend.h>
  31 #include <target/target_core_backend_configfs.h>
  32
  33 #include <linux/target_core_user.h>
  34
  35 /*
  36  * Define a shared-memory interface for LIO to pass SCSI commands and
  37  * data to userspace for processing. This is to allow backends that
  38  * are too complex for in-kernel support to be possible.
  39  *
  40  * It uses the UIO framework to do a lot of the device-creation and
  41  * introspection work for us.
  42  *
  43  * See the .h file for how the ring is laid out. Note that while the
  44  * command ring is defined, the particulars of the data area are
  45  * not. Offset values in the command entry point to other locations
  46  * internal to the mmap()ed area. There is separate space outside the
  47  * command ring for data buffers. This leaves maximum flexibility for
  48  * moving buffer allocations, or even page flipping or other
  49  * allocation techniques, without altering the command ring layout.
  50  *
  51  * SECURITY:
  52  * The user process must be assumed to be malicious. There's no way to
  53  * prevent it breaking the command ring protocol if it wants, but in
  54  * order to prevent other issues we must only ever read *data* from
  55  * the shared memory area, not offsets or sizes. This applies to
  56  * command ring entries as well as the mailbox. Extra code needed for
  57  * this may have a 'UAM' comment.
  58  */
  59
  60
  61 #define TCMU_TIME_OUT (30 * MSEC_PER_SEC)
  62
  63 #define CMDR_SIZE (16 * 4096)
  64 #define DATA_SIZE (257 * 4096)
  65
  66 #define TCMU_RING_SIZE (CMDR_SIZE + DATA_SIZE)
  67
  68 static struct device *tcmu_root_device;
  69
  70 struct tcmu_hba {
  71         u32 host_id;
  72 };
  73
  74 /* User wants all cmds or just some */
  75 enum passthru_level {
  76         TCMU_PASS_ALL = 0,
  77         TCMU_PASS_IO,
  78         TCMU_PASS_INVALID,
  79 };
  80
  81 #define TCMU_CONFIG_LEN 256
  82
  83 struct tcmu_dev {
  84         struct se_device se_dev;
  85
  86         char *name;
  87         struct se_hba *hba;
  88
  89 #define TCMU_DEV_BIT_OPEN 0
  90 #define TCMU_DEV_BIT_BROKEN 1
  91         unsigned long flags;
  92         enum passthru_level pass_level;
  93
  94         struct uio_info uio_info;
  95
  96         struct tcmu_mailbox *mb_addr;
  97         size_t dev_size;
  98         u32 cmdr_size;
  99         u32 cmdr_last_cleaned;
 100         /* Offset of data ring from start of mb */
 101         size_t data_off;
 102         size_t data_size;
 103         /* Ring head + tail values. */
 104         /* Must add data_off and mb_addr to get the address */
 105         size_t data_head;
 106         size_t data_tail;
 107
 108         wait_queue_head_t wait_cmdr;
 109         /* TODO should this be a mutex? */
 110         spinlock_t cmdr_lock;
 111
 112         struct idr commands;
 113         spinlock_t commands_lock;
 114
 115         struct timer_list timeout;
 116
 117         char dev_config[TCMU_CONFIG_LEN];
 118 };
 119
 120 #define TCMU_DEV(_se_dev) container_of(_se_dev, struct tcmu_dev, se_dev)
 121
 122 #define CMDR_OFF sizeof(struct tcmu_mailbox)
 123
 124 struct tcmu_cmd {
 125         struct se_cmd *se_cmd;
 126         struct tcmu_dev *tcmu_dev;
 127
 128         uint16_t cmd_id;
 129
 130         /* Can't use se_cmd->data_length when cleaning up expired cmds, because if
 131            cmd has been completed then accessing se_cmd is off limits */
 132         size_t data_length;
 133
 134         unsigned long deadline;
 135
 136 #define TCMU_CMD_BIT_EXPIRED 0
 137         unsigned long flags;
 138 };
 139
 140 static struct kmem_cache *tcmu_cmd_cache;
 141
 142 /* multicast group */
 143 enum tcmu_multicast_groups {
 144         TCMU_MCGRP_CONFIG,
 145 };
 146
 147 static const struct genl_multicast_group tcmu_mcgrps[] = {
 148         [TCMU_MCGRP_CONFIG] = { .name = "config", },
 149 };
 150
 151 /* Our generic netlink family */
 152 static struct genl_family tcmu_genl_family = {
 153         .id = GENL_ID_GENERATE,
 154         .hdrsize = 0,
 155         .name = "TCM-USER",
 156         .version = 1,
 157         .maxattr = TCMU_ATTR_MAX,
 158         .mcgrps = tcmu_mcgrps,
 159         .n_mcgrps = ARRAY_SIZE(tcmu_mcgrps),
 160 };
 161
 162 static struct tcmu_cmd *tcmu_alloc_cmd(struct se_cmd *se_cmd)
 163 {
 164         struct se_device *se_dev = se_cmd->se_dev;
 165         struct tcmu_dev *udev = TCMU_DEV(se_dev);
 166         struct tcmu_cmd *tcmu_cmd;
 167         int cmd_id;
 168
 169         tcmu_cmd = kmem_cache_zalloc(tcmu_cmd_cache, GFP_KERNEL);
 170         if (!tcmu_cmd)
 171                 return NULL;
 172
 173         tcmu_cmd->se_cmd = se_cmd;
 174         tcmu_cmd->tcmu_dev = udev;
 175         tcmu_cmd->data_length = se_cmd->data_length;
 176
 177         tcmu_cmd->deadline = jiffies + msecs_to_jiffies(TCMU_TIME_OUT);
 178
 179         idr_preload(GFP_KERNEL);
 180         spin_lock_irq(&udev->commands_lock);
 181         cmd_id = idr_alloc(&udev->commands, tcmu_cmd, 0,
 182                 USHRT_MAX, GFP_NOWAIT);
 183         spin_unlock_irq(&udev->commands_lock);
 184         idr_preload_end();
 185
 186         if (cmd_id < 0) {
 187                 kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
 188                 return NULL;
 189         }
 190         tcmu_cmd->cmd_id = cmd_id;
 191
 192         return tcmu_cmd;
 193 }
 194
 195 static inline void tcmu_flush_dcache_range(void *vaddr, size_t size)
 196 {
 197         unsigned long offset = (unsigned long) vaddr & ~PAGE_MASK;
 198
 199         size = round_up(size+offset, PAGE_SIZE);
 200         vaddr -= offset;
 201
 202         while (size) {
 203                 flush_dcache_page(virt_to_page(vaddr));
 204                 size -= PAGE_SIZE;
 205         }
 206 }
 207
 208 /*
 209  * Some ring helper functions. We don't assume size is a power of 2 so
 210  * we can't use circ_buf.h.
 211  */
 212 static inline size_t spc_used(size_t head, size_t tail, size_t size)
 213 {
 214         int diff = head - tail;
 215
 216         if (diff >= 0)
 217                 return diff;
 218         else
 219                 return size + diff;
 220 }
 221
 222 static inline size_t spc_free(size_t head, size_t tail, size_t size)
 223 {
 224         /* Keep 1 byte unused or we can't tell full from empty */
 225         return (size - spc_used(head, tail, size) - 1);
 226 }
 227
 228 static inline size_t head_to_end(size_t head, size_t size)
 229 {
 230         return size - head;
 231 }
 232
 233 #define UPDATE_HEAD(head, used, size) smp_store_release(&head, ((head % size) + used) % size)
 234
 235 /*
 236  * We can't queue a command until we have space available on the cmd ring *and* space
 237  * space avail on the data ring.
 238  *
 239  * Called with ring lock held.
 240  */
 241 static bool is_ring_space_avail(struct tcmu_dev *udev, size_t cmd_size, size_t data_needed)
 242 {
 243         struct tcmu_mailbox *mb = udev->mb_addr;
 244         size_t space;
 245         u32 cmd_head;
 246         size_t cmd_needed;
 247
 248         tcmu_flush_dcache_range(mb, sizeof(*mb));
 249
 250         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 251
 252         /*
 253          * If cmd end-of-ring space is too small then we need space for a NOP plus
 254          * original cmd - cmds are internally contiguous.
 255          */
 256         if (head_to_end(cmd_head, udev->cmdr_size) >= cmd_size)
 257                 cmd_needed = cmd_size;
 258         else
 259                 cmd_needed = cmd_size + head_to_end(cmd_head, udev->cmdr_size);
 260
 261         space = spc_free(cmd_head, udev->cmdr_last_cleaned, udev->cmdr_size);
 262         if (space < cmd_needed) {
 263                 pr_debug("no cmd space: %u %u %u\n", cmd_head,
 264                        udev->cmdr_last_cleaned, udev->cmdr_size);
 265                 return false;
 266         }
 267
 268         space = spc_free(udev->data_head, udev->data_tail, udev->data_size);
 269         if (space < data_needed) {
 270                 pr_debug("no data space: %zu %zu %zu\n", udev->data_head,
 271                        udev->data_tail, udev->data_size);
 272                 return false;
 273         }
 274
 275         return true;
 276 }
 277
 278 static int tcmu_queue_cmd_ring(struct tcmu_cmd *tcmu_cmd)
 279 {
 280         struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
 281         struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
 282         size_t base_command_size, command_size;
 283         struct tcmu_mailbox *mb;
 284         struct tcmu_cmd_entry *entry;
 285         int i;
 286         struct scatterlist *sg;
 287         struct iovec *iov;
 288         int iov_cnt = 0;
 289         uint32_t cmd_head;
 290         uint64_t cdb_off;
 291
 292         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags))
 293                 return -EINVAL;
 294
 295         /*
 296          * Must be a certain minimum size for response sense info, but
 297          * also may be larger if the iov array is large.
 298          *
 299          * iovs = sgl_nents+1, for end-of-ring case, plus another 1
 300          * b/c size == offsetof one-past-element.
 301         */
 302         base_command_size = max(offsetof(struct tcmu_cmd_entry,
 303                                          req.iov[se_cmd->t_data_nents + 2]),
 304                                 sizeof(struct tcmu_cmd_entry));
 305         command_size = base_command_size
 306                 + round_up(scsi_command_size(se_cmd->t_task_cdb), TCMU_OP_ALIGN_SIZE);
 307
 308         WARN_ON(command_size & (TCMU_OP_ALIGN_SIZE-1));
 309
 310         spin_lock_irq(&udev->cmdr_lock);
 311
 312         mb = udev->mb_addr;
 313         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 314         if ((command_size > (udev->cmdr_size / 2))
 315             || tcmu_cmd->data_length > (udev->data_size - 1))
 316                 pr_warn("TCMU: Request of size %zu/%zu may be too big for %u/%zu "
 317                         "cmd/data ring buffers\n", command_size, tcmu_cmd->data_length,
 318                         udev->cmdr_size, udev->data_size);
 319
 320         while (!is_ring_space_avail(udev, command_size, tcmu_cmd->data_length)) {
 321                 int ret;
 322                 DEFINE_WAIT(__wait);
 323
 324                 prepare_to_wait(&udev->wait_cmdr, &__wait, TASK_INTERRUPTIBLE);
 325
 326                 pr_debug("sleeping for ring space\n");
 327                 spin_unlock_irq(&udev->cmdr_lock);
 328                 ret = schedule_timeout(msecs_to_jiffies(TCMU_TIME_OUT));
 329                 finish_wait(&udev->wait_cmdr, &__wait);
 330                 if (!ret) {
 331                         pr_warn("tcmu: command timed out\n");
 332                         return -ETIMEDOUT;
 333                 }
 334
 335                 spin_lock_irq(&udev->cmdr_lock);
 336
 337                 /* We dropped cmdr_lock, cmd_head is stale */
 338                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 339         }
 340
 341         /* Insert a PAD if end-of-ring space is too small */
 342         if (head_to_end(cmd_head, udev->cmdr_size) < command_size) {
 343                 size_t pad_size = head_to_end(cmd_head, udev->cmdr_size);
 344
 345                 entry = (void *) mb + CMDR_OFF + cmd_head;
 346                 tcmu_flush_dcache_range(entry, sizeof(*entry));
 347                 tcmu_hdr_set_op(&entry->hdr, TCMU_OP_PAD);
 348                 tcmu_hdr_set_len(&entry->hdr, pad_size);
 349
 350                 UPDATE_HEAD(mb->cmd_head, pad_size, udev->cmdr_size);
 351
 352                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 353                 WARN_ON(cmd_head != 0);
 354         }
 355
 356         entry = (void *) mb + CMDR_OFF + cmd_head;
 357         tcmu_flush_dcache_range(entry, sizeof(*entry));
 358         tcmu_hdr_set_op(&entry->hdr, TCMU_OP_CMD);
 359         tcmu_hdr_set_len(&entry->hdr, command_size);
 360         entry->cmd_id = tcmu_cmd->cmd_id;
 361
 362         /*
 363          * Fix up iovecs, and handle if allocation in data ring wrapped.
 364          */
 365         iov = &entry->req.iov[0];
 366         for_each_sg(se_cmd->t_data_sg, sg, se_cmd->t_data_nents, i) {
 367                 size_t copy_bytes = min((size_t)sg->length,
 368                                      head_to_end(udev->data_head, udev->data_size));
 369                 void *from = kmap_atomic(sg_page(sg)) + sg->offset;
 370                 void *to = (void *) mb + udev->data_off + udev->data_head;
 371
 372                 if (tcmu_cmd->se_cmd->data_direction == DMA_TO_DEVICE) {
 373                         memcpy(to, from, copy_bytes);
 374                         tcmu_flush_dcache_range(to, copy_bytes);
 375                 }
 376
 377                 /* Even iov_base is relative to mb_addr */
 378                 iov->iov_len = copy_bytes;
 379                 iov->iov_base = (void *) udev->data_off + udev->data_head;
 380                 iov_cnt++;
 381                 iov++;
 382
 383                 UPDATE_HEAD(udev->data_head, copy_bytes, udev->data_size);
 384
 385                 /* Uh oh, we wrapped the buffer. Must split sg across 2 iovs. */
 386                 if (sg->length != copy_bytes) {
 387                         from += copy_bytes;
 388                         copy_bytes = sg->length - copy_bytes;
 389
 390                         iov->iov_len = copy_bytes;
 391                         iov->iov_base = (void *) udev->data_off + udev->data_head;
 392
 393                         if (se_cmd->data_direction == DMA_TO_DEVICE) {
 394                                 to = (void *) mb + udev->data_off + udev->data_head;
 395                                 memcpy(to, from, copy_bytes);
 396                                 tcmu_flush_dcache_range(to, copy_bytes);
 397                         }
 398
 399                         iov_cnt++;
 400                         iov++;
 401
 402                         UPDATE_HEAD(udev->data_head, copy_bytes, udev->data_size);
 403                 }
 404
 405                 kunmap_atomic(from);
 406         }
 407         entry->req.iov_cnt = iov_cnt;
 408
 409         /* All offsets relative to mb_addr, not start of entry! */
 410         cdb_off = CMDR_OFF + cmd_head + base_command_size;
 411         memcpy((void *) mb + cdb_off, se_cmd->t_task_cdb, scsi_command_size(se_cmd->t_task_cdb));
 412         entry->req.cdb_off = cdb_off;
 413         tcmu_flush_dcache_range(entry, sizeof(*entry));
 414
 415         UPDATE_HEAD(mb->cmd_head, command_size, udev->cmdr_size);
 416         tcmu_flush_dcache_range(mb, sizeof(*mb));
 417
 418         spin_unlock_irq(&udev->cmdr_lock);
 419
 420         /* TODO: only if FLUSH and FUA? */
 421         uio_event_notify(&udev->uio_info);
 422
 423         mod_timer(&udev->timeout,
 424                 round_jiffies_up(jiffies + msecs_to_jiffies(TCMU_TIME_OUT)));
 425
 426         return 0;
 427 }
 428
 429 static int tcmu_queue_cmd(struct se_cmd *se_cmd)
 430 {
 431         struct se_device *se_dev = se_cmd->se_dev;
 432         struct tcmu_dev *udev = TCMU_DEV(se_dev);
 433         struct tcmu_cmd *tcmu_cmd;
 434         int ret;
 435
 436         tcmu_cmd = tcmu_alloc_cmd(se_cmd);
 437         if (!tcmu_cmd)
 438                 return -ENOMEM;
 439
 440         ret = tcmu_queue_cmd_ring(tcmu_cmd);
 441         if (ret < 0) {
 442                 pr_err("TCMU: Could not queue command\n");
 443                 spin_lock_irq(&udev->commands_lock);
 444                 idr_remove(&udev->commands, tcmu_cmd->cmd_id);
 445                 spin_unlock_irq(&udev->commands_lock);
 446
 447                 kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
 448         }
 449
 450         return ret;
 451 }
 452
 453 static void tcmu_handle_completion(struct tcmu_cmd *cmd, struct tcmu_cmd_entry *entry)
 454 {
 455         struct se_cmd *se_cmd = cmd->se_cmd;
 456         struct tcmu_dev *udev = cmd->tcmu_dev;
 457
 458         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
 459                 /* cmd has been completed already from timeout, just reclaim data
 460                    ring space */
 461                 UPDATE_HEAD(udev->data_tail, cmd->data_length, udev->data_size);
 462                 return;
 463         }
 464
 465         if (entry->rsp.scsi_status == SAM_STAT_CHECK_CONDITION) {
 466                 memcpy(se_cmd->sense_buffer, entry->rsp.sense_buffer,
 467                                se_cmd->scsi_sense_length);
 468
 469                 UPDATE_HEAD(udev->data_tail, cmd->data_length, udev->data_size);
 470         }
 471         else if (se_cmd->data_direction == DMA_FROM_DEVICE) {
 472                 struct scatterlist *sg;
 473                 int i;
 474
 475                 /* It'd be easier to look at entry's iovec again, but UAM */
 476                 for_each_sg(se_cmd->t_data_sg, sg, se_cmd->t_data_nents, i) {
 477                         size_t copy_bytes;
 478                         void *to;
 479                         void *from;
 480
 481                         copy_bytes = min((size_t)sg->length,
 482                                          head_to_end(udev->data_tail, udev->data_size));
 483
 484                         to = kmap_atomic(sg_page(sg)) + sg->offset;
 485                         WARN_ON(sg->length + sg->offset > PAGE_SIZE);
 486                         from = (void *) udev->mb_addr + udev->data_off + udev->data_tail;
 487                         tcmu_flush_dcache_range(from, copy_bytes);
 488                         memcpy(to, from, copy_bytes);
 489
 490                         UPDATE_HEAD(udev->data_tail, copy_bytes, udev->data_size);
 491
 492                         /* Uh oh, wrapped the data buffer for this sg's data */
 493                         if (sg->length != copy_bytes) {
 494                                 from = (void *) udev->mb_addr + udev->data_off + udev->data_tail;
 495                                 WARN_ON(udev->data_tail);
 496                                 to += copy_bytes;
 497                                 copy_bytes = sg->length - copy_bytes;
 498                                 tcmu_flush_dcache_range(from, copy_bytes);
 499                                 memcpy(to, from, copy_bytes);
 500
 501                                 UPDATE_HEAD(udev->data_tail, copy_bytes, udev->data_size);
 502                         }
 503
 504                         kunmap_atomic(to);
 505                 }
 506
 507         } else if (se_cmd->data_direction == DMA_TO_DEVICE) {
 508                 UPDATE_HEAD(udev->data_tail, cmd->data_length, udev->data_size);
 509         } else {
 510                 pr_warn("TCMU: data direction was %d!\n", se_cmd->data_direction);
 511         }
 512
 513         target_complete_cmd(cmd->se_cmd, entry->rsp.scsi_status);
 514         cmd->se_cmd = NULL;
 515
 516         kmem_cache_free(tcmu_cmd_cache, cmd);
 517 }
 518
 519 static unsigned int tcmu_handle_completions(struct tcmu_dev *udev)
 520 {
 521         struct tcmu_mailbox *mb;
 522         LIST_HEAD(cpl_cmds);
 523         unsigned long flags;
 524         int handled = 0;
 525
 526         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags)) {
 527                 pr_err("ring broken, not handling completions\n");
 528                 return 0;
 529         }
 530
 531         spin_lock_irqsave(&udev->cmdr_lock, flags);
 532
 533         mb = udev->mb_addr;
 534         tcmu_flush_dcache_range(mb, sizeof(*mb));
 535
 536         while (udev->cmdr_last_cleaned != ACCESS_ONCE(mb->cmd_tail)) {
 537
 538                 struct tcmu_cmd_entry *entry = (void *) mb + CMDR_OFF + udev->cmdr_last_cleaned;
 539                 struct tcmu_cmd *cmd;
 540
 541                 tcmu_flush_dcache_range(entry, sizeof(*entry));
 542
 543                 if (tcmu_hdr_get_op(&entry->hdr) == TCMU_OP_PAD) {
 544                         UPDATE_HEAD(udev->cmdr_last_cleaned, tcmu_hdr_get_len(&entry->hdr), udev->cmdr_size);
 545                         continue;
 546                 }
 547                 WARN_ON(tcmu_hdr_get_op(&entry->hdr) != TCMU_OP_CMD);
 548
 549                 spin_lock(&udev->commands_lock);
 550                 cmd = idr_find(&udev->commands, entry->cmd_id);
 551                 if (cmd)
 552                         idr_remove(&udev->commands, cmd->cmd_id);
 553                 spin_unlock(&udev->commands_lock);
 554
 555                 if (!cmd) {
 556                         pr_err("cmd_id not found, ring is broken\n");
 557                         set_bit(TCMU_DEV_BIT_BROKEN, &udev->flags);
 558                         break;
 559                 }
 560
 561                 tcmu_handle_completion(cmd, entry);
 562
 563                 UPDATE_HEAD(udev->cmdr_last_cleaned, tcmu_hdr_get_len(&entry->hdr), udev->cmdr_size);
 564
 565                 handled++;
 566         }
 567
 568         if (mb->cmd_tail == mb->cmd_head)
 569                 del_timer(&udev->timeout); /* no more pending cmds */
 570
 571         spin_unlock_irqrestore(&udev->cmdr_lock, flags);
 572
 573         wake_up(&udev->wait_cmdr);
 574
 575         return handled;
 576 }
 577
 578 static int tcmu_check_expired_cmd(int id, void *p, void *data)
 579 {
 580         struct tcmu_cmd *cmd = p;
 581
 582         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
 583                 return 0;
 584
 585         if (!time_after(cmd->deadline, jiffies))
 586                 return 0;
 587
 588         set_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags);
 589         target_complete_cmd(cmd->se_cmd, SAM_STAT_CHECK_CONDITION);
 590         cmd->se_cmd = NULL;
 591
 592         kmem_cache_free(tcmu_cmd_cache, cmd);
 593
 594         return 0;
 595 }
 596
 597 static void tcmu_device_timedout(unsigned long data)
 598 {
 599         struct tcmu_dev *udev = (struct tcmu_dev *)data;
 600         unsigned long flags;
 601         int handled;
 602
 603         handled = tcmu_handle_completions(udev);
 604
 605         pr_warn("%d completions handled from timeout\n", handled);
 606
 607         spin_lock_irqsave(&udev->commands_lock, flags);
 608         idr_for_each(&udev->commands, tcmu_check_expired_cmd, NULL);
 609         spin_unlock_irqrestore(&udev->commands_lock, flags);
 610
 611         /*
 612          * We don't need to wakeup threads on wait_cmdr since they have their
 613          * own timeout.
 614          */
 615 }
 616
 617 static int tcmu_attach_hba(struct se_hba *hba, u32 host_id)
 618 {
 619         struct tcmu_hba *tcmu_hba;
 620
 621         tcmu_hba = kzalloc(sizeof(struct tcmu_hba), GFP_KERNEL);
 622         if (!tcmu_hba)
 623                 return -ENOMEM;
 624
 625         tcmu_hba->host_id = host_id;
 626         hba->hba_ptr = tcmu_hba;
 627
 628         return 0;
 629 }
 630
 631 static void tcmu_detach_hba(struct se_hba *hba)
 632 {
 633         kfree(hba->hba_ptr);
 634         hba->hba_ptr = NULL;
 635 }
 636
 637 static struct se_device *tcmu_alloc_device(struct se_hba *hba, const char *name)
 638 {
 639         struct tcmu_dev *udev;
 640
 641         udev = kzalloc(sizeof(struct tcmu_dev), GFP_KERNEL);
 642         if (!udev)
 643                 return NULL;
 644
 645         udev->name = kstrdup(name, GFP_KERNEL);
 646         if (!udev->name) {
 647                 kfree(udev);
 648                 return NULL;
 649         }
 650
 651         udev->hba = hba;
 652
 653         init_waitqueue_head(&udev->wait_cmdr);
 654         spin_lock_init(&udev->cmdr_lock);
 655
 656         idr_init(&udev->commands);
 657         spin_lock_init(&udev->commands_lock);
 658
 659         setup_timer(&udev->timeout, tcmu_device_timedout,
 660                 (unsigned long)udev);
 661
 662         udev->pass_level = TCMU_PASS_ALL;
 663
 664         return &udev->se_dev;
 665 }
 666
 667 static int tcmu_irqcontrol(struct uio_info *info, s32 irq_on)
 668 {
 669         struct tcmu_dev *tcmu_dev = container_of(info, struct tcmu_dev, uio_info);
 670
 671         tcmu_handle_completions(tcmu_dev);
 672
 673         return 0;
 674 }
 675
 676 /*
 677  * mmap code from uio.c. Copied here because we want to hook mmap()
 678  * and this stuff must come along.
 679  */
 680 static int tcmu_find_mem_index(struct vm_area_struct *vma)
 681 {
 682         struct tcmu_dev *udev = vma->vm_private_data;
 683         struct uio_info *info = &udev->uio_info;
 684
 685         if (vma->vm_pgoff < MAX_UIO_MAPS) {
 686                 if (info->mem[vma->vm_pgoff].size == 0)
 687                         return -1;
 688                 return (int)vma->vm_pgoff;
 689         }
 690         return -1;
 691 }
 692
 693 static int tcmu_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 694 {
 695         struct tcmu_dev *udev = vma->vm_private_data;
 696         struct uio_info *info = &udev->uio_info;
 697         struct page *page;
 698         unsigned long offset;
 699         void *addr;
 700
 701         int mi = tcmu_find_mem_index(vma);
 702         if (mi < 0)
 703                 return VM_FAULT_SIGBUS;
 704
 705         /*
 706          * We need to subtract mi because userspace uses offset = N*PAGE_SIZE
 707          * to use mem[N].
 708          */
 709         offset = (vmf->pgoff - mi) << PAGE_SHIFT;
 710
 711         addr = (void *)(unsigned long)info->mem[mi].addr + offset;
 712         if (info->mem[mi].memtype == UIO_MEM_LOGICAL)
 713                 page = virt_to_page(addr);
 714         else
 715                 page = vmalloc_to_page(addr);
 716         get_page(page);
 717         vmf->page = page;
 718         return 0;
 719 }
 720
 721 static const struct vm_operations_struct tcmu_vm_ops = {
 722         .fault = tcmu_vma_fault,
 723 };
 724
 725 static int tcmu_mmap(struct uio_info *info, struct vm_area_struct *vma)
 726 {
 727         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 728
 729         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
 730         vma->vm_ops = &tcmu_vm_ops;
 731
 732         vma->vm_private_data = udev;
 733
 734         /* Ensure the mmap is exactly the right size */
 735         if (vma_pages(vma) != (TCMU_RING_SIZE >> PAGE_SHIFT))
 736                 return -EINVAL;
 737
 738         return 0;
 739 }
 740
 741 static int tcmu_open(struct uio_info *info, struct inode *inode)
 742 {
 743         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 744
 745         /* O_EXCL not supported for char devs, so fake it? */
 746         if (test_and_set_bit(TCMU_DEV_BIT_OPEN, &udev->flags))
 747                 return -EBUSY;
 748
 749         pr_debug("open\n");
 750
 751         return 0;
 752 }
 753
 754 static int tcmu_release(struct uio_info *info, struct inode *inode)
 755 {
 756         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 757
 758         clear_bit(TCMU_DEV_BIT_OPEN, &udev->flags);
 759
 760         pr_debug("close\n");
 761
 762         return 0;
 763 }
 764
 765 static int tcmu_netlink_event(enum tcmu_genl_cmd cmd, const char *name, int minor)
 766 {
 767         struct sk_buff *skb;
 768         void *msg_header;
 769         int ret = -ENOMEM;
 770
 771         skb = genlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
 772         if (!skb)
 773                 return ret;
 774
 775         msg_header = genlmsg_put(skb, 0, 0, &tcmu_genl_family, 0, cmd);
 776         if (!msg_header)
 777                 goto free_skb;
 778
 779         ret = nla_put_string(skb, TCMU_ATTR_DEVICE, name);
 780         if (ret < 0)
 781                 goto free_skb;
 782
 783         ret = nla_put_u32(skb, TCMU_ATTR_MINOR, minor);
 784         if (ret < 0)
 785                 goto free_skb;
 786
 787         ret = genlmsg_end(skb, msg_header);
 788         if (ret < 0)
 789                 goto free_skb;
 790
 791         ret = genlmsg_multicast(&tcmu_genl_family, skb, 0,
 792                                 TCMU_MCGRP_CONFIG, GFP_KERNEL);
 793
 794         /* We don't care if no one is listening */
 795         if (ret == -ESRCH)
 796                 ret = 0;
 797
 798         return ret;
 799 free_skb:
 800         nlmsg_free(skb);
 801         return ret;
 802 }
 803
 804 static int tcmu_configure_device(struct se_device *dev)
 805 {
 806         struct tcmu_dev *udev = TCMU_DEV(dev);
 807         struct tcmu_hba *hba = udev->hba->hba_ptr;
 808         struct uio_info *info;
 809         struct tcmu_mailbox *mb;
 810         size_t size;
 811         size_t used;
 812         int ret = 0;
 813         char *str;
 814
 815         info = &udev->uio_info;
 816
 817         size = snprintf(NULL, 0, "tcm-user/%u/%s/%s", hba->host_id, udev->name,
 818                         udev->dev_config);
 819         size += 1; /* for \0 */
 820         str = kmalloc(size, GFP_KERNEL);
 821         if (!str)
 822                 return -ENOMEM;
 823
 824         used = snprintf(str, size, "tcm-user/%u/%s", hba->host_id, udev->name);
 825
 826         if (udev->dev_config[0])
 827                 snprintf(str + used, size - used, "/%s", udev->dev_config);
 828
 829         info->name = str;
 830
 831         udev->mb_addr = vzalloc(TCMU_RING_SIZE);
 832         if (!udev->mb_addr) {
 833                 ret = -ENOMEM;
 834                 goto err_vzalloc;
 835         }
 836
 837         /* mailbox fits in first part of CMDR space */
 838         udev->cmdr_size = CMDR_SIZE - CMDR_OFF;
 839         udev->data_off = CMDR_SIZE;
 840         udev->data_size = TCMU_RING_SIZE - CMDR_SIZE;
 841
 842         mb = udev->mb_addr;
 843         mb->version = 1;
 844         mb->cmdr_off = CMDR_OFF;
 845         mb->cmdr_size = udev->cmdr_size;
 846
 847         WARN_ON(!PAGE_ALIGNED(udev->data_off));
 848         WARN_ON(udev->data_size % PAGE_SIZE);
 849
 850         info->version = "1";
 851
 852         info->mem[0].name = "tcm-user command & data buffer";
 853         info->mem[0].addr = (phys_addr_t) udev->mb_addr;
 854         info->mem[0].size = TCMU_RING_SIZE;
 855         info->mem[0].memtype = UIO_MEM_VIRTUAL;
 856
 857         info->irqcontrol = tcmu_irqcontrol;
 858         info->irq = UIO_IRQ_CUSTOM;
 859
 860         info->mmap = tcmu_mmap;
 861         info->open = tcmu_open;
 862         info->release = tcmu_release;
 863
 864         ret = uio_register_device(tcmu_root_device, info);
 865         if (ret)
 866                 goto err_register;
 867
 868         /* Other attributes can be configured in userspace */
 869         dev->dev_attrib.hw_block_size = 512;
 870         dev->dev_attrib.hw_max_sectors = 128;
 871         dev->dev_attrib.hw_queue_depth = 128;
 872
 873         ret = tcmu_netlink_event(TCMU_CMD_ADDED_DEVICE, udev->uio_info.name,
 874                                  udev->uio_info.uio_dev->minor);
 875         if (ret)
 876                 goto err_netlink;
 877
 878         return 0;
 879
 880 err_netlink:
 881         uio_unregister_device(&udev->uio_info);
 882 err_register:
 883         vfree(udev->mb_addr);
 884 err_vzalloc:
 885         kfree(info->name);
 886
 887         return ret;
 888 }
 889
 890 static int tcmu_check_pending_cmd(int id, void *p, void *data)
 891 {
 892         struct tcmu_cmd *cmd = p;
 893
 894         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
 895                 return 0;
 896         return -EINVAL;
 897 }
 898
 899 static void tcmu_free_device(struct se_device *dev)
 900 {
 901         struct tcmu_dev *udev = TCMU_DEV(dev);
 902         int i;
 903
 904         del_timer_sync(&udev->timeout);
 905
 906         vfree(udev->mb_addr);
 907
 908         /* Upper layer should drain all requests before calling this */
 909         spin_lock_irq(&udev->commands_lock);
 910         i = idr_for_each(&udev->commands, tcmu_check_pending_cmd, NULL);
 911         idr_destroy(&udev->commands);
 912         spin_unlock_irq(&udev->commands_lock);
 913         WARN_ON(i);
 914
 915         /* Device was configured */
 916         if (udev->uio_info.uio_dev) {
 917                 tcmu_netlink_event(TCMU_CMD_REMOVED_DEVICE, udev->uio_info.name,
 918                                    udev->uio_info.uio_dev->minor);
 919
 920                 uio_unregister_device(&udev->uio_info);
 921                 kfree(udev->uio_info.name);
 922                 kfree(udev->name);
 923         }
 924
 925         kfree(udev);
 926 }
 927
 928 enum {
 929         Opt_dev_config, Opt_dev_size, Opt_err, Opt_pass_level,
 930 };
 931
 932 static match_table_t tokens = {
 933         {Opt_dev_config, "dev_config=%s"},
 934         {Opt_dev_size, "dev_size=%u"},
 935         {Opt_pass_level, "pass_level=%u"},
 936         {Opt_err, NULL}
 937 };
 938
 939 static ssize_t tcmu_set_configfs_dev_params(struct se_device *dev,
 940                 const char *page, ssize_t count)
 941 {
 942         struct tcmu_dev *udev = TCMU_DEV(dev);
 943         char *orig, *ptr, *opts, *arg_p;
 944         substring_t args[MAX_OPT_ARGS];
 945         int ret = 0, token;
 946         int arg;
 947
 948         opts = kstrdup(page, GFP_KERNEL);
 949         if (!opts)
 950                 return -ENOMEM;
 951
 952         orig = opts;
 953
 954         while ((ptr = strsep(&opts, ",\n")) != NULL) {
 955                 if (!*ptr)
 956                         continue;
 957
 958                 token = match_token(ptr, tokens, args);
 959                 switch (token) {
 960                 case Opt_dev_config:
 961                         if (match_strlcpy(udev->dev_config, &args[0],
 962                                           TCMU_CONFIG_LEN) == 0) {
 963                                 ret = -EINVAL;
 964                                 break;
 965                         }
 966                         pr_debug("TCMU: Referencing Path: %s\n", udev->dev_config);
 967                         break;
 968                 case Opt_dev_size:
 969                         arg_p = match_strdup(&args[0]);
 970                         if (!arg_p) {
 971                                 ret = -ENOMEM;
 972                                 break;
 973                         }
 974                         ret = kstrtoul(arg_p, 0, (unsigned long *) &udev->dev_size);
 975                         kfree(arg_p);
 976                         if (ret < 0)
 977                                 pr_err("kstrtoul() failed for dev_size=\n");
 978                         break;
 979                 case Opt_pass_level:
 980                         match_int(args, &arg);
 981                         if (arg >= TCMU_PASS_INVALID) {
 982                                 pr_warn("TCMU: Invalid pass_level: %d\n", arg);
 983                                 break;
 984                         }
 985
 986                         pr_debug("TCMU: Setting pass_level to %d\n", arg);
 987                         udev->pass_level = arg;
 988                         break;
 989                 default:
 990                         break;
 991                 }
 992         }
 993
 994         kfree(orig);
 995         return (!ret) ? count : ret;
 996 }
 997
 998 static ssize_t tcmu_show_configfs_dev_params(struct se_device *dev, char *b)
 999 {
1000         struct tcmu_dev *udev = TCMU_DEV(dev);
1001         ssize_t bl = 0;
1002
1003         bl = sprintf(b + bl, "Config: %s ",
1004                      udev->dev_config[0] ? udev->dev_config : "NULL");
1005         bl += sprintf(b + bl, "Size: %zu PassLevel: %u\n",
1006                       udev->dev_size, udev->pass_level);
1007
1008         return bl;
1009 }
1010
1011 static sector_t tcmu_get_blocks(struct se_device *dev)
1012 {
1013         struct tcmu_dev *udev = TCMU_DEV(dev);
1014
1015         return div_u64(udev->dev_size - dev->dev_attrib.block_size,
1016                        dev->dev_attrib.block_size);
1017 }
1018
1019 static sense_reason_t
1020 tcmu_execute_rw(struct se_cmd *se_cmd, struct scatterlist *sgl, u32 sgl_nents,
1021                 enum dma_data_direction data_direction)
1022 {
1023         int ret;
1024
1025         ret = tcmu_queue_cmd(se_cmd);
1026
1027         if (ret != 0)
1028                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1029         else
1030                 return TCM_NO_SENSE;
1031 }
1032
1033 static sense_reason_t
1034 tcmu_pass_op(struct se_cmd *se_cmd)
1035 {
1036         int ret = tcmu_queue_cmd(se_cmd);
1037
1038         if (ret != 0)
1039                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1040         else
1041                 return TCM_NO_SENSE;
1042 }
1043
1044 static struct sbc_ops tcmu_sbc_ops = {
1045         .execute_rw = tcmu_execute_rw,
1046         .execute_sync_cache     = tcmu_pass_op,
1047         .execute_write_same     = tcmu_pass_op,
1048         .execute_write_same_unmap = tcmu_pass_op,
1049         .execute_unmap          = tcmu_pass_op,
1050 };
1051
1052 static sense_reason_t
1053 tcmu_parse_cdb(struct se_cmd *cmd)
1054 {
1055         unsigned char *cdb = cmd->t_task_cdb;
1056         struct tcmu_dev *udev = TCMU_DEV(cmd->se_dev);
1057         sense_reason_t ret;
1058
1059         switch (udev->pass_level) {
1060         case TCMU_PASS_ALL:
1061                 /* We're just like pscsi, then */
1062                 /*
1063                  * For REPORT LUNS we always need to emulate the response, for everything
1064                  * else, pass it up.
1065                  */
1066                 switch (cdb[0]) {
1067                 case REPORT_LUNS:
1068                         cmd->execute_cmd = spc_emulate_report_luns;
1069                         break;
1070                 case READ_6:
1071                 case READ_10:
1072                 case READ_12:
1073                 case READ_16:
1074                 case WRITE_6:
1075                 case WRITE_10:
1076                 case WRITE_12:
1077                 case WRITE_16:
1078                 case WRITE_VERIFY:
1079                         cmd->se_cmd_flags |= SCF_SCSI_DATA_CDB;
1080                         /* FALLTHROUGH */
1081                 default:
1082                         cmd->execute_cmd = tcmu_pass_op;
1083                 }
1084                 ret = TCM_NO_SENSE;
1085                 break;
1086         case TCMU_PASS_IO:
1087                 ret = sbc_parse_cdb(cmd, &tcmu_sbc_ops);
1088                 break;
1089         default:
1090                 pr_err("Unknown tcm-user pass level %d\n", udev->pass_level);
1091                 ret = TCM_CHECK_CONDITION_ABORT_CMD;
1092         }
1093
1094         return ret;
1095 }
1096
1097 DEF_TB_DEFAULT_ATTRIBS(tcmu);
1098
1099 static struct configfs_attribute *tcmu_backend_dev_attrs[] = {
1100         &tcmu_dev_attrib_emulate_model_alias.attr,
1101         &tcmu_dev_attrib_emulate_dpo.attr,
1102         &tcmu_dev_attrib_emulate_fua_write.attr,
1103         &tcmu_dev_attrib_emulate_fua_read.attr,
1104         &tcmu_dev_attrib_emulate_write_cache.attr,
1105         &tcmu_dev_attrib_emulate_ua_intlck_ctrl.attr,
1106         &tcmu_dev_attrib_emulate_tas.attr,
1107         &tcmu_dev_attrib_emulate_tpu.attr,
1108         &tcmu_dev_attrib_emulate_tpws.attr,
1109         &tcmu_dev_attrib_emulate_caw.attr,
1110         &tcmu_dev_attrib_emulate_3pc.attr,
1111         &tcmu_dev_attrib_pi_prot_type.attr,
1112         &tcmu_dev_attrib_hw_pi_prot_type.attr,
1113         &tcmu_dev_attrib_pi_prot_format.attr,
1114         &tcmu_dev_attrib_enforce_pr_isids.attr,
1115         &tcmu_dev_attrib_is_nonrot.attr,
1116         &tcmu_dev_attrib_emulate_rest_reord.attr,
1117         &tcmu_dev_attrib_force_pr_aptpl.attr,
1118         &tcmu_dev_attrib_hw_block_size.attr,
1119         &tcmu_dev_attrib_block_size.attr,
1120         &tcmu_dev_attrib_hw_max_sectors.attr,
1121         &tcmu_dev_attrib_optimal_sectors.attr,
1122         &tcmu_dev_attrib_hw_queue_depth.attr,
1123         &tcmu_dev_attrib_queue_depth.attr,
1124         &tcmu_dev_attrib_max_unmap_lba_count.attr,
1125         &tcmu_dev_attrib_max_unmap_block_desc_count.attr,
1126         &tcmu_dev_attrib_unmap_granularity.attr,
1127         &tcmu_dev_attrib_unmap_granularity_alignment.attr,
1128         &tcmu_dev_attrib_max_write_same_len.attr,
1129         NULL,
1130 };
1131
1132 static struct se_subsystem_api tcmu_template = {
1133         .name                   = "user",
1134         .inquiry_prod           = "USER",
1135         .inquiry_rev            = TCMU_VERSION,
1136         .owner                  = THIS_MODULE,
1137         .transport_type         = TRANSPORT_PLUGIN_VHBA_PDEV,
1138         .attach_hba             = tcmu_attach_hba,
1139         .detach_hba             = tcmu_detach_hba,
1140         .alloc_device           = tcmu_alloc_device,
1141         .configure_device       = tcmu_configure_device,
1142         .free_device            = tcmu_free_device,
1143         .parse_cdb              = tcmu_parse_cdb,
1144         .set_configfs_dev_params = tcmu_set_configfs_dev_params,
1145         .show_configfs_dev_params = tcmu_show_configfs_dev_params,
1146         .get_device_type        = sbc_get_device_type,
1147         .get_blocks             = tcmu_get_blocks,
1148 };
1149
1150 static int __init tcmu_module_init(void)
1151 {
1152         struct target_backend_cits *tbc = &tcmu_template.tb_cits;
1153         int ret;
1154
1155         BUILD_BUG_ON((sizeof(struct tcmu_cmd_entry) % TCMU_OP_ALIGN_SIZE) != 0);
1156
1157         tcmu_cmd_cache = kmem_cache_create("tcmu_cmd_cache",
1158                                 sizeof(struct tcmu_cmd),
1159                                 __alignof__(struct tcmu_cmd),
1160                                 0, NULL);
1161         if (!tcmu_cmd_cache)
1162                 return -ENOMEM;
1163
1164         tcmu_root_device = root_device_register("tcm_user");
1165         if (IS_ERR(tcmu_root_device)) {
1166                 ret = PTR_ERR(tcmu_root_device);
1167                 goto out_free_cache;
1168         }
1169
1170         ret = genl_register_family(&tcmu_genl_family);
1171         if (ret < 0) {
1172                 goto out_unreg_device;
1173         }
1174
1175         target_core_setup_sub_cits(&tcmu_template);
1176         tbc->tb_dev_attrib_cit.ct_attrs = tcmu_backend_dev_attrs;
1177
1178         ret = transport_subsystem_register(&tcmu_template);
1179         if (ret)
1180                 goto out_unreg_genl;
1181
1182         return 0;
1183
1184 out_unreg_genl:
1185         genl_unregister_family(&tcmu_genl_family);
1186 out_unreg_device:
1187         root_device_unregister(tcmu_root_device);
1188 out_free_cache:
1189         kmem_cache_destroy(tcmu_cmd_cache);
1190
1191         return ret;
1192 }
1193
1194 static void __exit tcmu_module_exit(void)
1195 {
1196         transport_subsystem_release(&tcmu_template);
1197         genl_unregister_family(&tcmu_genl_family);
1198         root_device_unregister(tcmu_root_device);
1199         kmem_cache_destroy(tcmu_cmd_cache);
1200 }
1201
1202 MODULE_DESCRIPTION("TCM USER subsystem plugin");
1203 MODULE_AUTHOR("Shaohua Li <shli@kernel.org>");
1204 MODULE_AUTHOR("Andy Grover <agrover@redhat.com>");
1205 MODULE_LICENSE("GPL");
1206
1207 module_init(tcmu_module_init);
1208 module_exit(tcmu_module_exit);