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