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