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