drivers/target/target_core_user.c

   1 /*
   2  * Copyright (C) 2013 Shaohua Li <shli@kernel.org>
   3  * Copyright (C) 2014 Red Hat, Inc.
   4  * Copyright (C) 2015 Arrikto, Inc.
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms and conditions of the GNU General Public License,
   8  * version 2, as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  13  * more details.
  14  *
  15  * You should have received a copy of the GNU General Public License along with
  16  * this program; if not, write to the Free Software Foundation, Inc.,
  17  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  18  */
  19
  20 #include <linux/spinlock.h>
  21 #include <linux/module.h>
  22 #include <linux/idr.h>
  23 #include <linux/kernel.h>
  24 #include <linux/timer.h>
  25 #include <linux/parser.h>
  26 #include <linux/vmalloc.h>
  27 #include <linux/uio_driver.h>
  28 #include <net/genetlink.h>
  29 #include <scsi/scsi_common.h>
  30 #include <scsi/scsi_proto.h>
  31 #include <target/target_core_base.h>
  32 #include <target/target_core_fabric.h>
  33 #include <target/target_core_backend.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         if (se_cmd->se_cmd_flags & SCF_BIDI) {
 172                 BUG_ON(!(se_cmd->t_bidi_data_sg && se_cmd->t_bidi_data_nents));
 173                 tcmu_cmd->data_length += se_cmd->t_bidi_data_sg->length;
 174         }
 175
 176         tcmu_cmd->deadline = jiffies + msecs_to_jiffies(TCMU_TIME_OUT);
 177
 178         idr_preload(GFP_KERNEL);
 179         spin_lock_irq(&udev->commands_lock);
 180         cmd_id = idr_alloc(&udev->commands, tcmu_cmd, 0,
 181                 USHRT_MAX, GFP_NOWAIT);
 182         spin_unlock_irq(&udev->commands_lock);
 183         idr_preload_end();
 184
 185         if (cmd_id < 0) {
 186                 kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
 187                 return NULL;
 188         }
 189         tcmu_cmd->cmd_id = cmd_id;
 190
 191         return tcmu_cmd;
 192 }
 193
 194 static inline void tcmu_flush_dcache_range(void *vaddr, size_t size)
 195 {
 196         unsigned long offset = (unsigned long) vaddr & ~PAGE_MASK;
 197
 198         size = round_up(size+offset, PAGE_SIZE);
 199         vaddr -= offset;
 200
 201         while (size) {
 202                 flush_dcache_page(virt_to_page(vaddr));
 203                 size -= PAGE_SIZE;
 204         }
 205 }
 206
 207 /*
 208  * Some ring helper functions. We don't assume size is a power of 2 so
 209  * we can't use circ_buf.h.
 210  */
 211 static inline size_t spc_used(size_t head, size_t tail, size_t size)
 212 {
 213         int diff = head - tail;
 214
 215         if (diff >= 0)
 216                 return diff;
 217         else
 218                 return size + diff;
 219 }
 220
 221 static inline size_t spc_free(size_t head, size_t tail, size_t size)
 222 {
 223         /* Keep 1 byte unused or we can't tell full from empty */
 224         return (size - spc_used(head, tail, size) - 1);
 225 }
 226
 227 static inline size_t head_to_end(size_t head, size_t size)
 228 {
 229         return size - head;
 230 }
 231
 232 #define UPDATE_HEAD(head, used, size) smp_store_release(&head, ((head % size) + used) % size)
 233
 234 static void alloc_and_scatter_data_area(struct tcmu_dev *udev,
 235         struct scatterlist *data_sg, unsigned int data_nents,
 236         struct iovec **iov, int *iov_cnt, bool copy_data)
 237 {
 238         int i;
 239         void *from, *to;
 240         size_t copy_bytes;
 241         struct scatterlist *sg;
 242
 243         for_each_sg(data_sg, sg, data_nents, i) {
 244                 copy_bytes = min_t(size_t, sg->length,
 245                                  head_to_end(udev->data_head, udev->data_size));
 246                 from = kmap_atomic(sg_page(sg)) + sg->offset;
 247                 to = (void *) udev->mb_addr + udev->data_off + udev->data_head;
 248
 249                 if (copy_data) {
 250                         memcpy(to, from, copy_bytes);
 251                         tcmu_flush_dcache_range(to, copy_bytes);
 252                 }
 253
 254                 /* Even iov_base is relative to mb_addr */
 255                 (*iov)->iov_len = copy_bytes;
 256                 (*iov)->iov_base = (void __user *) udev->data_off +
 257                                                 udev->data_head;
 258                 (*iov_cnt)++;
 259                 (*iov)++;
 260
 261                 UPDATE_HEAD(udev->data_head, copy_bytes, udev->data_size);
 262
 263                 /* Uh oh, we wrapped the buffer. Must split sg across 2 iovs. */
 264                 if (sg->length != copy_bytes) {
 265                         void *from_skip = from + copy_bytes;
 266
 267                         copy_bytes = sg->length - copy_bytes;
 268
 269                         (*iov)->iov_len = copy_bytes;
 270                         (*iov)->iov_base = (void __user *) udev->data_off +
 271                                                         udev->data_head;
 272
 273                         if (copy_data) {
 274                                 to = (void *) udev->mb_addr +
 275                                         udev->data_off + udev->data_head;
 276                                 memcpy(to, from_skip, copy_bytes);
 277                                 tcmu_flush_dcache_range(to, copy_bytes);
 278                         }
 279
 280                         (*iov_cnt)++;
 281                         (*iov)++;
 282
 283                         UPDATE_HEAD(udev->data_head,
 284                                 copy_bytes, udev->data_size);
 285                 }
 286
 287                 kunmap_atomic(from - sg->offset);
 288         }
 289 }
 290
 291 static void gather_and_free_data_area(struct tcmu_dev *udev,
 292         struct scatterlist *data_sg, unsigned int data_nents)
 293 {
 294         int i;
 295         void *from, *to;
 296         size_t copy_bytes;
 297         struct scatterlist *sg;
 298
 299         /* It'd be easier to look at entry's iovec again, but UAM */
 300         for_each_sg(data_sg, sg, data_nents, i) {
 301                 copy_bytes = min_t(size_t, sg->length,
 302                                  head_to_end(udev->data_tail, udev->data_size));
 303
 304                 to = kmap_atomic(sg_page(sg)) + sg->offset;
 305                 WARN_ON(sg->length + sg->offset > PAGE_SIZE);
 306                 from = (void *) udev->mb_addr +
 307                         udev->data_off + udev->data_tail;
 308                 tcmu_flush_dcache_range(from, copy_bytes);
 309                 memcpy(to, from, copy_bytes);
 310
 311                 UPDATE_HEAD(udev->data_tail, copy_bytes, udev->data_size);
 312
 313                 /* Uh oh, wrapped the data buffer for this sg's data */
 314                 if (sg->length != copy_bytes) {
 315                         void *to_skip = to + copy_bytes;
 316
 317                         from = (void *) udev->mb_addr +
 318                                 udev->data_off + udev->data_tail;
 319                         WARN_ON(udev->data_tail);
 320                         copy_bytes = sg->length - copy_bytes;
 321                         tcmu_flush_dcache_range(from, copy_bytes);
 322                         memcpy(to_skip, from, copy_bytes);
 323
 324                         UPDATE_HEAD(udev->data_tail,
 325                                 copy_bytes, udev->data_size);
 326                 }
 327                 kunmap_atomic(to - sg->offset);
 328         }
 329 }
 330
 331 /*
 332  * We can't queue a command until we have space available on the cmd ring *and*
 333  * space available on the data ring.
 334  *
 335  * Called with ring lock held.
 336  */
 337 static bool is_ring_space_avail(struct tcmu_dev *udev, size_t cmd_size, size_t data_needed)
 338 {
 339         struct tcmu_mailbox *mb = udev->mb_addr;
 340         size_t space;
 341         u32 cmd_head;
 342         size_t cmd_needed;
 343
 344         tcmu_flush_dcache_range(mb, sizeof(*mb));
 345
 346         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 347
 348         /*
 349          * If cmd end-of-ring space is too small then we need space for a NOP plus
 350          * original cmd - cmds are internally contiguous.
 351          */
 352         if (head_to_end(cmd_head, udev->cmdr_size) >= cmd_size)
 353                 cmd_needed = cmd_size;
 354         else
 355                 cmd_needed = cmd_size + head_to_end(cmd_head, udev->cmdr_size);
 356
 357         space = spc_free(cmd_head, udev->cmdr_last_cleaned, udev->cmdr_size);
 358         if (space < cmd_needed) {
 359                 pr_debug("no cmd space: %u %u %u\n", cmd_head,
 360                        udev->cmdr_last_cleaned, udev->cmdr_size);
 361                 return false;
 362         }
 363
 364         space = spc_free(udev->data_head, udev->data_tail, udev->data_size);
 365         if (space < data_needed) {
 366                 pr_debug("no data space: %zu %zu %zu\n", udev->data_head,
 367                        udev->data_tail, udev->data_size);
 368                 return false;
 369         }
 370
 371         return true;
 372 }
 373
 374 static int tcmu_queue_cmd_ring(struct tcmu_cmd *tcmu_cmd)
 375 {
 376         struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
 377         struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
 378         size_t base_command_size, command_size;
 379         struct tcmu_mailbox *mb;
 380         struct tcmu_cmd_entry *entry;
 381         struct iovec *iov;
 382         int iov_cnt;
 383         uint32_t cmd_head;
 384         uint64_t cdb_off;
 385         bool copy_to_data_area;
 386
 387         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags))
 388                 return -EINVAL;
 389
 390         /*
 391          * Must be a certain minimum size for response sense info, but
 392          * also may be larger if the iov array is large.
 393          *
 394          * iovs = sgl_nents+1, for end-of-ring case, plus another 1
 395          * b/c size == offsetof one-past-element.
 396         */
 397         base_command_size = max(offsetof(struct tcmu_cmd_entry,
 398                                          req.iov[se_cmd->t_bidi_data_nents +
 399                                                  se_cmd->t_data_nents + 2]),
 400                                 sizeof(struct tcmu_cmd_entry));
 401         command_size = base_command_size
 402                 + round_up(scsi_command_size(se_cmd->t_task_cdb), TCMU_OP_ALIGN_SIZE);
 403
 404         WARN_ON(command_size & (TCMU_OP_ALIGN_SIZE-1));
 405
 406         spin_lock_irq(&udev->cmdr_lock);
 407
 408         mb = udev->mb_addr;
 409         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 410         if ((command_size > (udev->cmdr_size / 2))
 411             || tcmu_cmd->data_length > (udev->data_size - 1))
 412                 pr_warn("TCMU: Request of size %zu/%zu may be too big for %u/%zu "
 413                         "cmd/data ring buffers\n", command_size, tcmu_cmd->data_length,
 414                         udev->cmdr_size, udev->data_size);
 415
 416         while (!is_ring_space_avail(udev, command_size, tcmu_cmd->data_length)) {
 417                 int ret;
 418                 DEFINE_WAIT(__wait);
 419
 420                 prepare_to_wait(&udev->wait_cmdr, &__wait, TASK_INTERRUPTIBLE);
 421
 422                 pr_debug("sleeping for ring space\n");
 423                 spin_unlock_irq(&udev->cmdr_lock);
 424                 ret = schedule_timeout(msecs_to_jiffies(TCMU_TIME_OUT));
 425                 finish_wait(&udev->wait_cmdr, &__wait);
 426                 if (!ret) {
 427                         pr_warn("tcmu: command timed out\n");
 428                         return -ETIMEDOUT;
 429                 }
 430
 431                 spin_lock_irq(&udev->cmdr_lock);
 432
 433                 /* We dropped cmdr_lock, cmd_head is stale */
 434                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 435         }
 436
 437         /* Insert a PAD if end-of-ring space is too small */
 438         if (head_to_end(cmd_head, udev->cmdr_size) < command_size) {
 439                 size_t pad_size = head_to_end(cmd_head, udev->cmdr_size);
 440
 441                 entry = (void *) mb + CMDR_OFF + cmd_head;
 442                 tcmu_flush_dcache_range(entry, sizeof(*entry));
 443                 tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_PAD);
 444                 tcmu_hdr_set_len(&entry->hdr.len_op, pad_size);
 445                 entry->hdr.cmd_id = 0; /* not used for PAD */
 446                 entry->hdr.kflags = 0;
 447                 entry->hdr.uflags = 0;
 448
 449                 UPDATE_HEAD(mb->cmd_head, pad_size, udev->cmdr_size);
 450
 451                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 452                 WARN_ON(cmd_head != 0);
 453         }
 454
 455         entry = (void *) mb + CMDR_OFF + cmd_head;
 456         tcmu_flush_dcache_range(entry, sizeof(*entry));
 457         tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_CMD);
 458         tcmu_hdr_set_len(&entry->hdr.len_op, command_size);
 459         entry->hdr.cmd_id = tcmu_cmd->cmd_id;
 460         entry->hdr.kflags = 0;
 461         entry->hdr.uflags = 0;
 462
 463         /*
 464          * Fix up iovecs, and handle if allocation in data ring wrapped.
 465          */
 466         iov = &entry->req.iov[0];
 467         iov_cnt = 0;
 468         copy_to_data_area = (se_cmd->data_direction == DMA_TO_DEVICE
 469                 || se_cmd->se_cmd_flags & SCF_BIDI);
 470         alloc_and_scatter_data_area(udev, se_cmd->t_data_sg,
 471                 se_cmd->t_data_nents, &iov, &iov_cnt, copy_to_data_area);
 472         entry->req.iov_cnt = iov_cnt;
 473         entry->req.iov_dif_cnt = 0;
 474
 475         /* Handle BIDI commands */
 476         iov_cnt = 0;
 477         alloc_and_scatter_data_area(udev, se_cmd->t_bidi_data_sg,
 478                 se_cmd->t_bidi_data_nents, &iov, &iov_cnt, false);
 479         entry->req.iov_bidi_cnt = iov_cnt;
 480
 481         /* All offsets relative to mb_addr, not start of entry! */
 482         cdb_off = CMDR_OFF + cmd_head + base_command_size;
 483         memcpy((void *) mb + cdb_off, se_cmd->t_task_cdb, scsi_command_size(se_cmd->t_task_cdb));
 484         entry->req.cdb_off = cdb_off;
 485         tcmu_flush_dcache_range(entry, sizeof(*entry));
 486
 487         UPDATE_HEAD(mb->cmd_head, command_size, udev->cmdr_size);
 488         tcmu_flush_dcache_range(mb, sizeof(*mb));
 489
 490         spin_unlock_irq(&udev->cmdr_lock);
 491
 492         /* TODO: only if FLUSH and FUA? */
 493         uio_event_notify(&udev->uio_info);
 494
 495         mod_timer(&udev->timeout,
 496                 round_jiffies_up(jiffies + msecs_to_jiffies(TCMU_TIME_OUT)));
 497
 498         return 0;
 499 }
 500
 501 static int tcmu_queue_cmd(struct se_cmd *se_cmd)
 502 {
 503         struct se_device *se_dev = se_cmd->se_dev;
 504         struct tcmu_dev *udev = TCMU_DEV(se_dev);
 505         struct tcmu_cmd *tcmu_cmd;
 506         int ret;
 507
 508         tcmu_cmd = tcmu_alloc_cmd(se_cmd);
 509         if (!tcmu_cmd)
 510                 return -ENOMEM;
 511
 512         ret = tcmu_queue_cmd_ring(tcmu_cmd);
 513         if (ret < 0) {
 514                 pr_err("TCMU: Could not queue command\n");
 515                 spin_lock_irq(&udev->commands_lock);
 516                 idr_remove(&udev->commands, tcmu_cmd->cmd_id);
 517                 spin_unlock_irq(&udev->commands_lock);
 518
 519                 kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
 520         }
 521
 522         return ret;
 523 }
 524
 525 static void tcmu_handle_completion(struct tcmu_cmd *cmd, struct tcmu_cmd_entry *entry)
 526 {
 527         struct se_cmd *se_cmd = cmd->se_cmd;
 528         struct tcmu_dev *udev = cmd->tcmu_dev;
 529
 530         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
 531                 /* cmd has been completed already from timeout, just reclaim data
 532                    ring space */
 533                 UPDATE_HEAD(udev->data_tail, cmd->data_length, udev->data_size);
 534                 return;
 535         }
 536
 537         if (entry->hdr.uflags & TCMU_UFLAG_UNKNOWN_OP) {
 538                 UPDATE_HEAD(udev->data_tail, cmd->data_length, udev->data_size);
 539                 pr_warn("TCMU: Userspace set UNKNOWN_OP flag on se_cmd %p\n",
 540                         cmd->se_cmd);
 541                 transport_generic_request_failure(cmd->se_cmd,
 542                         TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE);
 543                 cmd->se_cmd = NULL;
 544                 kmem_cache_free(tcmu_cmd_cache, cmd);
 545                 return;
 546         }
 547
 548         if (entry->rsp.scsi_status == SAM_STAT_CHECK_CONDITION) {
 549                 memcpy(se_cmd->sense_buffer, entry->rsp.sense_buffer,
 550                                se_cmd->scsi_sense_length);
 551
 552                 UPDATE_HEAD(udev->data_tail, cmd->data_length, udev->data_size);
 553         } else if (se_cmd->se_cmd_flags & SCF_BIDI) {
 554                 /* Discard data_out buffer */
 555                 UPDATE_HEAD(udev->data_tail,
 556                         (size_t)se_cmd->t_data_sg->length, udev->data_size);
 557
 558                 /* Get Data-In buffer */
 559                 gather_and_free_data_area(udev,
 560                         se_cmd->t_bidi_data_sg, se_cmd->t_bidi_data_nents);
 561         } else if (se_cmd->data_direction == DMA_FROM_DEVICE) {
 562                 gather_and_free_data_area(udev,
 563                         se_cmd->t_data_sg, se_cmd->t_data_nents);
 564         } else if (se_cmd->data_direction == DMA_TO_DEVICE) {
 565                 UPDATE_HEAD(udev->data_tail, cmd->data_length, udev->data_size);
 566         } else if (se_cmd->data_direction != DMA_NONE) {
 567                 pr_warn("TCMU: data direction was %d!\n",
 568                         se_cmd->data_direction);
 569         }
 570
 571         target_complete_cmd(cmd->se_cmd, entry->rsp.scsi_status);
 572         cmd->se_cmd = NULL;
 573
 574         kmem_cache_free(tcmu_cmd_cache, cmd);
 575 }
 576
 577 static unsigned int tcmu_handle_completions(struct tcmu_dev *udev)
 578 {
 579         struct tcmu_mailbox *mb;
 580         LIST_HEAD(cpl_cmds);
 581         unsigned long flags;
 582         int handled = 0;
 583
 584         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags)) {
 585                 pr_err("ring broken, not handling completions\n");
 586                 return 0;
 587         }
 588
 589         spin_lock_irqsave(&udev->cmdr_lock, flags);
 590
 591         mb = udev->mb_addr;
 592         tcmu_flush_dcache_range(mb, sizeof(*mb));
 593
 594         while (udev->cmdr_last_cleaned != ACCESS_ONCE(mb->cmd_tail)) {
 595
 596                 struct tcmu_cmd_entry *entry = (void *) mb + CMDR_OFF + udev->cmdr_last_cleaned;
 597                 struct tcmu_cmd *cmd;
 598
 599                 tcmu_flush_dcache_range(entry, sizeof(*entry));
 600
 601                 if (tcmu_hdr_get_op(entry->hdr.len_op) == TCMU_OP_PAD) {
 602                         UPDATE_HEAD(udev->cmdr_last_cleaned,
 603                                     tcmu_hdr_get_len(entry->hdr.len_op),
 604                                     udev->cmdr_size);
 605                         continue;
 606                 }
 607                 WARN_ON(tcmu_hdr_get_op(entry->hdr.len_op) != TCMU_OP_CMD);
 608
 609                 spin_lock(&udev->commands_lock);
 610                 cmd = idr_find(&udev->commands, entry->hdr.cmd_id);
 611                 if (cmd)
 612                         idr_remove(&udev->commands, cmd->cmd_id);
 613                 spin_unlock(&udev->commands_lock);
 614
 615                 if (!cmd) {
 616                         pr_err("cmd_id not found, ring is broken\n");
 617                         set_bit(TCMU_DEV_BIT_BROKEN, &udev->flags);
 618                         break;
 619                 }
 620
 621                 tcmu_handle_completion(cmd, entry);
 622
 623                 UPDATE_HEAD(udev->cmdr_last_cleaned,
 624                             tcmu_hdr_get_len(entry->hdr.len_op),
 625                             udev->cmdr_size);
 626
 627                 handled++;
 628         }
 629
 630         if (mb->cmd_tail == mb->cmd_head)
 631                 del_timer(&udev->timeout); /* no more pending cmds */
 632
 633         spin_unlock_irqrestore(&udev->cmdr_lock, flags);
 634
 635         wake_up(&udev->wait_cmdr);
 636
 637         return handled;
 638 }
 639
 640 static int tcmu_check_expired_cmd(int id, void *p, void *data)
 641 {
 642         struct tcmu_cmd *cmd = p;
 643
 644         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
 645                 return 0;
 646
 647         if (!time_after(cmd->deadline, jiffies))
 648                 return 0;
 649
 650         set_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags);
 651         target_complete_cmd(cmd->se_cmd, SAM_STAT_CHECK_CONDITION);
 652         cmd->se_cmd = NULL;
 653
 654         kmem_cache_free(tcmu_cmd_cache, cmd);
 655
 656         return 0;
 657 }
 658
 659 static void tcmu_device_timedout(unsigned long data)
 660 {
 661         struct tcmu_dev *udev = (struct tcmu_dev *)data;
 662         unsigned long flags;
 663         int handled;
 664
 665         handled = tcmu_handle_completions(udev);
 666
 667         pr_warn("%d completions handled from timeout\n", handled);
 668
 669         spin_lock_irqsave(&udev->commands_lock, flags);
 670         idr_for_each(&udev->commands, tcmu_check_expired_cmd, NULL);
 671         spin_unlock_irqrestore(&udev->commands_lock, flags);
 672
 673         /*
 674          * We don't need to wakeup threads on wait_cmdr since they have their
 675          * own timeout.
 676          */
 677 }
 678
 679 static int tcmu_attach_hba(struct se_hba *hba, u32 host_id)
 680 {
 681         struct tcmu_hba *tcmu_hba;
 682
 683         tcmu_hba = kzalloc(sizeof(struct tcmu_hba), GFP_KERNEL);
 684         if (!tcmu_hba)
 685                 return -ENOMEM;
 686
 687         tcmu_hba->host_id = host_id;
 688         hba->hba_ptr = tcmu_hba;
 689
 690         return 0;
 691 }
 692
 693 static void tcmu_detach_hba(struct se_hba *hba)
 694 {
 695         kfree(hba->hba_ptr);
 696         hba->hba_ptr = NULL;
 697 }
 698
 699 static struct se_device *tcmu_alloc_device(struct se_hba *hba, const char *name)
 700 {
 701         struct tcmu_dev *udev;
 702
 703         udev = kzalloc(sizeof(struct tcmu_dev), GFP_KERNEL);
 704         if (!udev)
 705                 return NULL;
 706
 707         udev->name = kstrdup(name, GFP_KERNEL);
 708         if (!udev->name) {
 709                 kfree(udev);
 710                 return NULL;
 711         }
 712
 713         udev->hba = hba;
 714
 715         init_waitqueue_head(&udev->wait_cmdr);
 716         spin_lock_init(&udev->cmdr_lock);
 717
 718         idr_init(&udev->commands);
 719         spin_lock_init(&udev->commands_lock);
 720
 721         setup_timer(&udev->timeout, tcmu_device_timedout,
 722                 (unsigned long)udev);
 723
 724         return &udev->se_dev;
 725 }
 726
 727 static int tcmu_irqcontrol(struct uio_info *info, s32 irq_on)
 728 {
 729         struct tcmu_dev *tcmu_dev = container_of(info, struct tcmu_dev, uio_info);
 730
 731         tcmu_handle_completions(tcmu_dev);
 732
 733         return 0;
 734 }
 735
 736 /*
 737  * mmap code from uio.c. Copied here because we want to hook mmap()
 738  * and this stuff must come along.
 739  */
 740 static int tcmu_find_mem_index(struct vm_area_struct *vma)
 741 {
 742         struct tcmu_dev *udev = vma->vm_private_data;
 743         struct uio_info *info = &udev->uio_info;
 744
 745         if (vma->vm_pgoff < MAX_UIO_MAPS) {
 746                 if (info->mem[vma->vm_pgoff].size == 0)
 747                         return -1;
 748                 return (int)vma->vm_pgoff;
 749         }
 750         return -1;
 751 }
 752
 753 static int tcmu_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 754 {
 755         struct tcmu_dev *udev = vma->vm_private_data;
 756         struct uio_info *info = &udev->uio_info;
 757         struct page *page;
 758         unsigned long offset;
 759         void *addr;
 760
 761         int mi = tcmu_find_mem_index(vma);
 762         if (mi < 0)
 763                 return VM_FAULT_SIGBUS;
 764
 765         /*
 766          * We need to subtract mi because userspace uses offset = N*PAGE_SIZE
 767          * to use mem[N].
 768          */
 769         offset = (vmf->pgoff - mi) << PAGE_SHIFT;
 770
 771         addr = (void *)(unsigned long)info->mem[mi].addr + offset;
 772         if (info->mem[mi].memtype == UIO_MEM_LOGICAL)
 773                 page = virt_to_page(addr);
 774         else
 775                 page = vmalloc_to_page(addr);
 776         get_page(page);
 777         vmf->page = page;
 778         return 0;
 779 }
 780
 781 static const struct vm_operations_struct tcmu_vm_ops = {
 782         .fault = tcmu_vma_fault,
 783 };
 784
 785 static int tcmu_mmap(struct uio_info *info, struct vm_area_struct *vma)
 786 {
 787         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 788
 789         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
 790         vma->vm_ops = &tcmu_vm_ops;
 791
 792         vma->vm_private_data = udev;
 793
 794         /* Ensure the mmap is exactly the right size */
 795         if (vma_pages(vma) != (TCMU_RING_SIZE >> PAGE_SHIFT))
 796                 return -EINVAL;
 797
 798         return 0;
 799 }
 800
 801 static int tcmu_open(struct uio_info *info, struct inode *inode)
 802 {
 803         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 804
 805         /* O_EXCL not supported for char devs, so fake it? */
 806         if (test_and_set_bit(TCMU_DEV_BIT_OPEN, &udev->flags))
 807                 return -EBUSY;
 808
 809         pr_debug("open\n");
 810
 811         return 0;
 812 }
 813
 814 static int tcmu_release(struct uio_info *info, struct inode *inode)
 815 {
 816         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 817
 818         clear_bit(TCMU_DEV_BIT_OPEN, &udev->flags);
 819
 820         pr_debug("close\n");
 821
 822         return 0;
 823 }
 824
 825 static int tcmu_netlink_event(enum tcmu_genl_cmd cmd, const char *name, int minor)
 826 {
 827         struct sk_buff *skb;
 828         void *msg_header;
 829         int ret = -ENOMEM;
 830
 831         skb = genlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
 832         if (!skb)
 833                 return ret;
 834
 835         msg_header = genlmsg_put(skb, 0, 0, &tcmu_genl_family, 0, cmd);
 836         if (!msg_header)
 837                 goto free_skb;
 838
 839         ret = nla_put_string(skb, TCMU_ATTR_DEVICE, name);
 840         if (ret < 0)
 841                 goto free_skb;
 842
 843         ret = nla_put_u32(skb, TCMU_ATTR_MINOR, minor);
 844         if (ret < 0)
 845                 goto free_skb;
 846
 847         genlmsg_end(skb, msg_header);
 848
 849         ret = genlmsg_multicast(&tcmu_genl_family, skb, 0,
 850                                 TCMU_MCGRP_CONFIG, GFP_KERNEL);
 851
 852         /* We don't care if no one is listening */
 853         if (ret == -ESRCH)
 854                 ret = 0;
 855
 856         return ret;
 857 free_skb:
 858         nlmsg_free(skb);
 859         return ret;
 860 }
 861
 862 static int tcmu_configure_device(struct se_device *dev)
 863 {
 864         struct tcmu_dev *udev = TCMU_DEV(dev);
 865         struct tcmu_hba *hba = udev->hba->hba_ptr;
 866         struct uio_info *info;
 867         struct tcmu_mailbox *mb;
 868         size_t size;
 869         size_t used;
 870         int ret = 0;
 871         char *str;
 872
 873         info = &udev->uio_info;
 874
 875         size = snprintf(NULL, 0, "tcm-user/%u/%s/%s", hba->host_id, udev->name,
 876                         udev->dev_config);
 877         size += 1; /* for \0 */
 878         str = kmalloc(size, GFP_KERNEL);
 879         if (!str)
 880                 return -ENOMEM;
 881
 882         used = snprintf(str, size, "tcm-user/%u/%s", hba->host_id, udev->name);
 883
 884         if (udev->dev_config[0])
 885                 snprintf(str + used, size - used, "/%s", udev->dev_config);
 886
 887         info->name = str;
 888
 889         udev->mb_addr = vzalloc(TCMU_RING_SIZE);
 890         if (!udev->mb_addr) {
 891                 ret = -ENOMEM;
 892                 goto err_vzalloc;
 893         }
 894
 895         /* mailbox fits in first part of CMDR space */
 896         udev->cmdr_size = CMDR_SIZE - CMDR_OFF;
 897         udev->data_off = CMDR_SIZE;
 898         udev->data_size = TCMU_RING_SIZE - CMDR_SIZE;
 899
 900         mb = udev->mb_addr;
 901         mb->version = TCMU_MAILBOX_VERSION;
 902         mb->cmdr_off = CMDR_OFF;
 903         mb->cmdr_size = udev->cmdr_size;
 904
 905         WARN_ON(!PAGE_ALIGNED(udev->data_off));
 906         WARN_ON(udev->data_size % PAGE_SIZE);
 907
 908         info->version = xstr(TCMU_MAILBOX_VERSION);
 909
 910         info->mem[0].name = "tcm-user command & data buffer";
 911         info->mem[0].addr = (phys_addr_t) udev->mb_addr;
 912         info->mem[0].size = TCMU_RING_SIZE;
 913         info->mem[0].memtype = UIO_MEM_VIRTUAL;
 914
 915         info->irqcontrol = tcmu_irqcontrol;
 916         info->irq = UIO_IRQ_CUSTOM;
 917
 918         info->mmap = tcmu_mmap;
 919         info->open = tcmu_open;
 920         info->release = tcmu_release;
 921
 922         ret = uio_register_device(tcmu_root_device, info);
 923         if (ret)
 924                 goto err_register;
 925
 926         /* Other attributes can be configured in userspace */
 927         dev->dev_attrib.hw_block_size = 512;
 928         dev->dev_attrib.hw_max_sectors = 128;
 929         dev->dev_attrib.hw_queue_depth = 128;
 930
 931         ret = tcmu_netlink_event(TCMU_CMD_ADDED_DEVICE, udev->uio_info.name,
 932                                  udev->uio_info.uio_dev->minor);
 933         if (ret)
 934                 goto err_netlink;
 935
 936         return 0;
 937
 938 err_netlink:
 939         uio_unregister_device(&udev->uio_info);
 940 err_register:
 941         vfree(udev->mb_addr);
 942 err_vzalloc:
 943         kfree(info->name);
 944
 945         return ret;
 946 }
 947
 948 static int tcmu_check_pending_cmd(int id, void *p, void *data)
 949 {
 950         struct tcmu_cmd *cmd = p;
 951
 952         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
 953                 return 0;
 954         return -EINVAL;
 955 }
 956
 957 static void tcmu_dev_call_rcu(struct rcu_head *p)
 958 {
 959         struct se_device *dev = container_of(p, struct se_device, rcu_head);
 960         struct tcmu_dev *udev = TCMU_DEV(dev);
 961
 962         kfree(udev);
 963 }
 964
 965 static void tcmu_free_device(struct se_device *dev)
 966 {
 967         struct tcmu_dev *udev = TCMU_DEV(dev);
 968         int i;
 969
 970         del_timer_sync(&udev->timeout);
 971
 972         vfree(udev->mb_addr);
 973
 974         /* Upper layer should drain all requests before calling this */
 975         spin_lock_irq(&udev->commands_lock);
 976         i = idr_for_each(&udev->commands, tcmu_check_pending_cmd, NULL);
 977         idr_destroy(&udev->commands);
 978         spin_unlock_irq(&udev->commands_lock);
 979         WARN_ON(i);
 980
 981         /* Device was configured */
 982         if (udev->uio_info.uio_dev) {
 983                 tcmu_netlink_event(TCMU_CMD_REMOVED_DEVICE, udev->uio_info.name,
 984                                    udev->uio_info.uio_dev->minor);
 985
 986                 uio_unregister_device(&udev->uio_info);
 987                 kfree(udev->uio_info.name);
 988                 kfree(udev->name);
 989         }
 990         call_rcu(&dev->rcu_head, tcmu_dev_call_rcu);
 991 }
 992
 993 enum {
 994         Opt_dev_config, Opt_dev_size, Opt_hw_block_size, Opt_err,
 995 };
 996
 997 static match_table_t tokens = {
 998         {Opt_dev_config, "dev_config=%s"},
 999         {Opt_dev_size, "dev_size=%u"},
1000         {Opt_hw_block_size, "hw_block_size=%u"},
1001         {Opt_err, NULL}
1002 };
1003
1004 static ssize_t tcmu_set_configfs_dev_params(struct se_device *dev,
1005                 const char *page, ssize_t count)
1006 {
1007         struct tcmu_dev *udev = TCMU_DEV(dev);
1008         char *orig, *ptr, *opts, *arg_p;
1009         substring_t args[MAX_OPT_ARGS];
1010         int ret = 0, token;
1011         unsigned long tmp_ul;
1012
1013         opts = kstrdup(page, GFP_KERNEL);
1014         if (!opts)
1015                 return -ENOMEM;
1016
1017         orig = opts;
1018
1019         while ((ptr = strsep(&opts, ",\n")) != NULL) {
1020                 if (!*ptr)
1021                         continue;
1022
1023                 token = match_token(ptr, tokens, args);
1024                 switch (token) {
1025                 case Opt_dev_config:
1026                         if (match_strlcpy(udev->dev_config, &args[0],
1027                                           TCMU_CONFIG_LEN) == 0) {
1028                                 ret = -EINVAL;
1029                                 break;
1030                         }
1031                         pr_debug("TCMU: Referencing Path: %s\n", udev->dev_config);
1032                         break;
1033                 case Opt_dev_size:
1034                         arg_p = match_strdup(&args[0]);
1035                         if (!arg_p) {
1036                                 ret = -ENOMEM;
1037                                 break;
1038                         }
1039                         ret = kstrtoul(arg_p, 0, (unsigned long *) &udev->dev_size);
1040                         kfree(arg_p);
1041                         if (ret < 0)
1042                                 pr_err("kstrtoul() failed for dev_size=\n");
1043                         break;
1044                 case Opt_hw_block_size:
1045                         arg_p = match_strdup(&args[0]);
1046                         if (!arg_p) {
1047                                 ret = -ENOMEM;
1048                                 break;
1049                         }
1050                         ret = kstrtoul(arg_p, 0, &tmp_ul);
1051                         kfree(arg_p);
1052                         if (ret < 0) {
1053                                 pr_err("kstrtoul() failed for hw_block_size=\n");
1054                                 break;
1055                         }
1056                         if (!tmp_ul) {
1057                                 pr_err("hw_block_size must be nonzero\n");
1058                                 break;
1059                         }
1060                         dev->dev_attrib.hw_block_size = tmp_ul;
1061                         break;
1062                 default:
1063                         break;
1064                 }
1065         }
1066
1067         kfree(orig);
1068         return (!ret) ? count : ret;
1069 }
1070
1071 static ssize_t tcmu_show_configfs_dev_params(struct se_device *dev, char *b)
1072 {
1073         struct tcmu_dev *udev = TCMU_DEV(dev);
1074         ssize_t bl = 0;
1075
1076         bl = sprintf(b + bl, "Config: %s ",
1077                      udev->dev_config[0] ? udev->dev_config : "NULL");
1078         bl += sprintf(b + bl, "Size: %zu\n", udev->dev_size);
1079
1080         return bl;
1081 }
1082
1083 static sector_t tcmu_get_blocks(struct se_device *dev)
1084 {
1085         struct tcmu_dev *udev = TCMU_DEV(dev);
1086
1087         return div_u64(udev->dev_size - dev->dev_attrib.block_size,
1088                        dev->dev_attrib.block_size);
1089 }
1090
1091 static sense_reason_t
1092 tcmu_pass_op(struct se_cmd *se_cmd)
1093 {
1094         int ret = tcmu_queue_cmd(se_cmd);
1095
1096         if (ret != 0)
1097                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1098         else
1099                 return TCM_NO_SENSE;
1100 }
1101
1102 static sense_reason_t
1103 tcmu_parse_cdb(struct se_cmd *cmd)
1104 {
1105         return passthrough_parse_cdb(cmd, tcmu_pass_op);
1106 }
1107
1108 static const struct target_backend_ops tcmu_ops = {
1109         .name                   = "user",
1110         .inquiry_prod           = "USER",
1111         .inquiry_rev            = TCMU_VERSION,
1112         .owner                  = THIS_MODULE,
1113         .transport_flags        = TRANSPORT_FLAG_PASSTHROUGH,
1114         .attach_hba             = tcmu_attach_hba,
1115         .detach_hba             = tcmu_detach_hba,
1116         .alloc_device           = tcmu_alloc_device,
1117         .configure_device       = tcmu_configure_device,
1118         .free_device            = tcmu_free_device,
1119         .parse_cdb              = tcmu_parse_cdb,
1120         .set_configfs_dev_params = tcmu_set_configfs_dev_params,
1121         .show_configfs_dev_params = tcmu_show_configfs_dev_params,
1122         .get_device_type        = sbc_get_device_type,
1123         .get_blocks             = tcmu_get_blocks,
1124         .tb_dev_attrib_attrs    = passthrough_attrib_attrs,
1125 };
1126
1127 static int __init tcmu_module_init(void)
1128 {
1129         int ret;
1130
1131         BUILD_BUG_ON((sizeof(struct tcmu_cmd_entry) % TCMU_OP_ALIGN_SIZE) != 0);
1132
1133         tcmu_cmd_cache = kmem_cache_create("tcmu_cmd_cache",
1134                                 sizeof(struct tcmu_cmd),
1135                                 __alignof__(struct tcmu_cmd),
1136                                 0, NULL);
1137         if (!tcmu_cmd_cache)
1138                 return -ENOMEM;
1139
1140         tcmu_root_device = root_device_register("tcm_user");
1141         if (IS_ERR(tcmu_root_device)) {
1142                 ret = PTR_ERR(tcmu_root_device);
1143                 goto out_free_cache;
1144         }
1145
1146         ret = genl_register_family(&tcmu_genl_family);
1147         if (ret < 0) {
1148                 goto out_unreg_device;
1149         }
1150
1151         ret = transport_backend_register(&tcmu_ops);
1152         if (ret)
1153                 goto out_unreg_genl;
1154
1155         return 0;
1156
1157 out_unreg_genl:
1158         genl_unregister_family(&tcmu_genl_family);
1159 out_unreg_device:
1160         root_device_unregister(tcmu_root_device);
1161 out_free_cache:
1162         kmem_cache_destroy(tcmu_cmd_cache);
1163
1164         return ret;
1165 }
1166
1167 static void __exit tcmu_module_exit(void)
1168 {
1169         target_backend_unregister(&tcmu_ops);
1170         genl_unregister_family(&tcmu_genl_family);
1171         root_device_unregister(tcmu_root_device);
1172         kmem_cache_destroy(tcmu_cmd_cache);
1173 }
1174
1175 MODULE_DESCRIPTION("TCM USER subsystem plugin");
1176 MODULE_AUTHOR("Shaohua Li <shli@kernel.org>");
1177 MODULE_AUTHOR("Andy Grover <agrover@redhat.com>");
1178 MODULE_LICENSE("GPL");
1179
1180 module_init(tcmu_module_init);
1181 module_exit(tcmu_module_exit);