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 <linux/stringify.h>
  29 #include <linux/bitops.h>
  30 #include <net/genetlink.h>
  31 #include <scsi/scsi_common.h>
  32 #include <scsi/scsi_proto.h>
  33 #include <target/target_core_base.h>
  34 #include <target/target_core_fabric.h>
  35 #include <target/target_core_backend.h>
  36
  37 #include <linux/target_core_user.h>
  38
  39 /*
  40  * Define a shared-memory interface for LIO to pass SCSI commands and
  41  * data to userspace for processing. This is to allow backends that
  42  * are too complex for in-kernel support to be possible.
  43  *
  44  * It uses the UIO framework to do a lot of the device-creation and
  45  * introspection work for us.
  46  *
  47  * See the .h file for how the ring is laid out. Note that while the
  48  * command ring is defined, the particulars of the data area are
  49  * not. Offset values in the command entry point to other locations
  50  * internal to the mmap()ed area. There is separate space outside the
  51  * command ring for data buffers. This leaves maximum flexibility for
  52  * moving buffer allocations, or even page flipping or other
  53  * allocation techniques, without altering the command ring layout.
  54  *
  55  * SECURITY:
  56  * The user process must be assumed to be malicious. There's no way to
  57  * prevent it breaking the command ring protocol if it wants, but in
  58  * order to prevent other issues we must only ever read *data* from
  59  * the shared memory area, not offsets or sizes. This applies to
  60  * command ring entries as well as the mailbox. Extra code needed for
  61  * this may have a 'UAM' comment.
  62  */
  63
  64
  65 #define TCMU_TIME_OUT (30 * MSEC_PER_SEC)
  66
  67 #define DATA_BLOCK_BITS 256
  68 #define DATA_BLOCK_SIZE 4096
  69
  70 #define CMDR_SIZE (16 * 4096)
  71 #define DATA_SIZE (DATA_BLOCK_BITS * DATA_BLOCK_SIZE)
  72
  73 #define TCMU_RING_SIZE (CMDR_SIZE + DATA_SIZE)
  74
  75 static struct device *tcmu_root_device;
  76
  77 struct tcmu_hba {
  78         u32 host_id;
  79 };
  80
  81 #define TCMU_CONFIG_LEN 256
  82
  83 struct tcmu_dev {
  84         struct se_device se_dev;
  85
  86         char *name;
  87         struct se_hba *hba;
  88
  89 #define TCMU_DEV_BIT_OPEN 0
  90 #define TCMU_DEV_BIT_BROKEN 1
  91         unsigned long flags;
  92
  93         struct uio_info uio_info;
  94
  95         struct tcmu_mailbox *mb_addr;
  96         size_t dev_size;
  97         u32 cmdr_size;
  98         u32 cmdr_last_cleaned;
  99         /* Offset of data area from start of mb */
 100         /* Must add data_off and mb_addr to get the address */
 101         size_t data_off;
 102         size_t data_size;
 103
 104         DECLARE_BITMAP(data_bitmap, DATA_BLOCK_BITS);
 105
 106         wait_queue_head_t wait_cmdr;
 107         /* TODO should this be a mutex? */
 108         spinlock_t cmdr_lock;
 109
 110         struct idr commands;
 111         spinlock_t commands_lock;
 112
 113         struct timer_list timeout;
 114
 115         char dev_config[TCMU_CONFIG_LEN];
 116 };
 117
 118 #define TCMU_DEV(_se_dev) container_of(_se_dev, struct tcmu_dev, se_dev)
 119
 120 #define CMDR_OFF sizeof(struct tcmu_mailbox)
 121
 122 struct tcmu_cmd {
 123         struct se_cmd *se_cmd;
 124         struct tcmu_dev *tcmu_dev;
 125
 126         uint16_t cmd_id;
 127
 128         /* Can't use se_cmd when cleaning up expired cmds, because if
 129            cmd has been completed then accessing se_cmd is off limits */
 130         DECLARE_BITMAP(data_bitmap, DATA_BLOCK_BITS);
 131
 132         unsigned long deadline;
 133
 134 #define TCMU_CMD_BIT_EXPIRED 0
 135         unsigned long flags;
 136 };
 137
 138 static struct kmem_cache *tcmu_cmd_cache;
 139
 140 /* multicast group */
 141 enum tcmu_multicast_groups {
 142         TCMU_MCGRP_CONFIG,
 143 };
 144
 145 static const struct genl_multicast_group tcmu_mcgrps[] = {
 146         [TCMU_MCGRP_CONFIG] = { .name = "config", },
 147 };
 148
 149 /* Our generic netlink family */
 150 static struct genl_family tcmu_genl_family = {
 151         .id = GENL_ID_GENERATE,
 152         .hdrsize = 0,
 153         .name = "TCM-USER",
 154         .version = 1,
 155         .maxattr = TCMU_ATTR_MAX,
 156         .mcgrps = tcmu_mcgrps,
 157         .n_mcgrps = ARRAY_SIZE(tcmu_mcgrps),
 158         .netnsok = true,
 159 };
 160
 161 static struct tcmu_cmd *tcmu_alloc_cmd(struct se_cmd *se_cmd)
 162 {
 163         struct se_device *se_dev = se_cmd->se_dev;
 164         struct tcmu_dev *udev = TCMU_DEV(se_dev);
 165         struct tcmu_cmd *tcmu_cmd;
 166         int cmd_id;
 167
 168         tcmu_cmd = kmem_cache_zalloc(tcmu_cmd_cache, GFP_KERNEL);
 169         if (!tcmu_cmd)
 170                 return NULL;
 171
 172         tcmu_cmd->se_cmd = se_cmd;
 173         tcmu_cmd->tcmu_dev = udev;
 174         tcmu_cmd->deadline = jiffies + msecs_to_jiffies(TCMU_TIME_OUT);
 175
 176         idr_preload(GFP_KERNEL);
 177         spin_lock_irq(&udev->commands_lock);
 178         cmd_id = idr_alloc(&udev->commands, tcmu_cmd, 0,
 179                 USHRT_MAX, GFP_NOWAIT);
 180         spin_unlock_irq(&udev->commands_lock);
 181         idr_preload_end();
 182
 183         if (cmd_id < 0) {
 184                 kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
 185                 return NULL;
 186         }
 187         tcmu_cmd->cmd_id = cmd_id;
 188
 189         return tcmu_cmd;
 190 }
 191
 192 static inline void tcmu_flush_dcache_range(void *vaddr, size_t size)
 193 {
 194         unsigned long offset = offset_in_page(vaddr);
 195
 196         size = round_up(size+offset, PAGE_SIZE);
 197         vaddr -= offset;
 198
 199         while (size) {
 200                 flush_dcache_page(virt_to_page(vaddr));
 201                 size -= PAGE_SIZE;
 202         }
 203 }
 204
 205 /*
 206  * Some ring helper functions. We don't assume size is a power of 2 so
 207  * we can't use circ_buf.h.
 208  */
 209 static inline size_t spc_used(size_t head, size_t tail, size_t size)
 210 {
 211         int diff = head - tail;
 212
 213         if (diff >= 0)
 214                 return diff;
 215         else
 216                 return size + diff;
 217 }
 218
 219 static inline size_t spc_free(size_t head, size_t tail, size_t size)
 220 {
 221         /* Keep 1 byte unused or we can't tell full from empty */
 222         return (size - spc_used(head, tail, size) - 1);
 223 }
 224
 225 static inline size_t head_to_end(size_t head, size_t size)
 226 {
 227         return size - head;
 228 }
 229
 230 static inline void new_iov(struct iovec **iov, int *iov_cnt,
 231                            struct tcmu_dev *udev)
 232 {
 233         struct iovec *iovec;
 234
 235         if (*iov_cnt != 0)
 236                 (*iov)++;
 237         (*iov_cnt)++;
 238
 239         iovec = *iov;
 240         memset(iovec, 0, sizeof(struct iovec));
 241 }
 242
 243 #define UPDATE_HEAD(head, used, size) smp_store_release(&head, ((head % size) + used) % size)
 244
 245 /* offset is relative to mb_addr */
 246 static inline size_t get_block_offset(struct tcmu_dev *dev,
 247                 int block, int remaining)
 248 {
 249         return dev->data_off + block * DATA_BLOCK_SIZE +
 250                 DATA_BLOCK_SIZE - remaining;
 251 }
 252
 253 static inline size_t iov_tail(struct tcmu_dev *udev, struct iovec *iov)
 254 {
 255         return (size_t)iov->iov_base + iov->iov_len;
 256 }
 257
 258 static void alloc_and_scatter_data_area(struct tcmu_dev *udev,
 259         struct scatterlist *data_sg, unsigned int data_nents,
 260         struct iovec **iov, int *iov_cnt, bool copy_data)
 261 {
 262         int i, block;
 263         int block_remaining = 0;
 264         void *from, *to;
 265         size_t copy_bytes, to_offset;
 266         struct scatterlist *sg;
 267
 268         for_each_sg(data_sg, sg, data_nents, i) {
 269                 int sg_remaining = sg->length;
 270                 from = kmap_atomic(sg_page(sg)) + sg->offset;
 271                 while (sg_remaining > 0) {
 272                         if (block_remaining == 0) {
 273                                 block = find_first_zero_bit(udev->data_bitmap,
 274                                                 DATA_BLOCK_BITS);
 275                                 block_remaining = DATA_BLOCK_SIZE;
 276                                 set_bit(block, udev->data_bitmap);
 277                         }
 278                         copy_bytes = min_t(size_t, sg_remaining,
 279                                         block_remaining);
 280                         to_offset = get_block_offset(udev, block,
 281                                         block_remaining);
 282                         to = (void *)udev->mb_addr + to_offset;
 283                         if (*iov_cnt != 0 &&
 284                             to_offset == iov_tail(udev, *iov)) {
 285                                 (*iov)->iov_len += copy_bytes;
 286                         } else {
 287                                 new_iov(iov, iov_cnt, udev);
 288                                 (*iov)->iov_base = (void __user *) to_offset;
 289                                 (*iov)->iov_len = copy_bytes;
 290                         }
 291                         if (copy_data) {
 292                                 memcpy(to, from + sg->length - sg_remaining,
 293                                         copy_bytes);
 294                                 tcmu_flush_dcache_range(to, copy_bytes);
 295                         }
 296                         sg_remaining -= copy_bytes;
 297                         block_remaining -= copy_bytes;
 298                 }
 299                 kunmap_atomic(from - sg->offset);
 300         }
 301 }
 302
 303 static void free_data_area(struct tcmu_dev *udev, struct tcmu_cmd *cmd)
 304 {
 305         bitmap_xor(udev->data_bitmap, udev->data_bitmap, cmd->data_bitmap,
 306                    DATA_BLOCK_BITS);
 307 }
 308
 309 static void gather_data_area(struct tcmu_dev *udev, unsigned long *cmd_bitmap,
 310                 struct scatterlist *data_sg, unsigned int data_nents)
 311 {
 312         int i, block;
 313         int block_remaining = 0;
 314         void *from, *to;
 315         size_t copy_bytes, from_offset;
 316         struct scatterlist *sg;
 317
 318         for_each_sg(data_sg, sg, data_nents, i) {
 319                 int sg_remaining = sg->length;
 320                 to = kmap_atomic(sg_page(sg)) + sg->offset;
 321                 while (sg_remaining > 0) {
 322                         if (block_remaining == 0) {
 323                                 block = find_first_bit(cmd_bitmap,
 324                                                 DATA_BLOCK_BITS);
 325                                 block_remaining = DATA_BLOCK_SIZE;
 326                                 clear_bit(block, cmd_bitmap);
 327                         }
 328                         copy_bytes = min_t(size_t, sg_remaining,
 329                                         block_remaining);
 330                         from_offset = get_block_offset(udev, block,
 331                                         block_remaining);
 332                         from = (void *) udev->mb_addr + from_offset;
 333                         tcmu_flush_dcache_range(from, copy_bytes);
 334                         memcpy(to + sg->length - sg_remaining, from,
 335                                         copy_bytes);
 336
 337                         sg_remaining -= copy_bytes;
 338                         block_remaining -= copy_bytes;
 339                 }
 340                 kunmap_atomic(to - sg->offset);
 341         }
 342 }
 343
 344 static inline size_t spc_bitmap_free(unsigned long *bitmap)
 345 {
 346         return DATA_BLOCK_SIZE * (DATA_BLOCK_BITS -
 347                         bitmap_weight(bitmap, DATA_BLOCK_BITS));
 348 }
 349
 350 /*
 351  * We can't queue a command until we have space available on the cmd ring *and*
 352  * space available on the data area.
 353  *
 354  * Called with ring lock held.
 355  */
 356 static bool is_ring_space_avail(struct tcmu_dev *udev, size_t cmd_size, size_t data_needed)
 357 {
 358         struct tcmu_mailbox *mb = udev->mb_addr;
 359         size_t space, cmd_needed;
 360         u32 cmd_head;
 361
 362         tcmu_flush_dcache_range(mb, sizeof(*mb));
 363
 364         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 365
 366         /*
 367          * If cmd end-of-ring space is too small then we need space for a NOP plus
 368          * original cmd - cmds are internally contiguous.
 369          */
 370         if (head_to_end(cmd_head, udev->cmdr_size) >= cmd_size)
 371                 cmd_needed = cmd_size;
 372         else
 373                 cmd_needed = cmd_size + head_to_end(cmd_head, udev->cmdr_size);
 374
 375         space = spc_free(cmd_head, udev->cmdr_last_cleaned, udev->cmdr_size);
 376         if (space < cmd_needed) {
 377                 pr_debug("no cmd space: %u %u %u\n", cmd_head,
 378                        udev->cmdr_last_cleaned, udev->cmdr_size);
 379                 return false;
 380         }
 381
 382         space = spc_bitmap_free(udev->data_bitmap);
 383         if (space < data_needed) {
 384                 pr_debug("no data space: only %zu available, but ask for %zu\n",
 385                                 space, data_needed);
 386                 return false;
 387         }
 388
 389         return true;
 390 }
 391
 392 static sense_reason_t
 393 tcmu_queue_cmd_ring(struct tcmu_cmd *tcmu_cmd)
 394 {
 395         struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
 396         struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
 397         size_t base_command_size, command_size;
 398         struct tcmu_mailbox *mb;
 399         struct tcmu_cmd_entry *entry;
 400         struct iovec *iov;
 401         int iov_cnt;
 402         uint32_t cmd_head;
 403         uint64_t cdb_off;
 404         bool copy_to_data_area;
 405         size_t data_length;
 406         DECLARE_BITMAP(old_bitmap, DATA_BLOCK_BITS);
 407
 408         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags))
 409                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
 410
 411         /*
 412          * Must be a certain minimum size for response sense info, but
 413          * also may be larger if the iov array is large.
 414          *
 415          * We prepare way too many iovs for potential uses here, because it's
 416          * expensive to tell how many regions are freed in the bitmap
 417         */
 418         base_command_size = max(offsetof(struct tcmu_cmd_entry,
 419                                 req.iov[se_cmd->t_bidi_data_nents +
 420                                         se_cmd->t_data_nents]),
 421                                 sizeof(struct tcmu_cmd_entry));
 422         command_size = base_command_size
 423                 + round_up(scsi_command_size(se_cmd->t_task_cdb), TCMU_OP_ALIGN_SIZE);
 424
 425         WARN_ON(command_size & (TCMU_OP_ALIGN_SIZE-1));
 426
 427         spin_lock_irq(&udev->cmdr_lock);
 428
 429         mb = udev->mb_addr;
 430         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 431         data_length = se_cmd->data_length;
 432         if (se_cmd->se_cmd_flags & SCF_BIDI) {
 433                 BUG_ON(!(se_cmd->t_bidi_data_sg && se_cmd->t_bidi_data_nents));
 434                 data_length += se_cmd->t_bidi_data_sg->length;
 435         }
 436         if ((command_size > (udev->cmdr_size / 2)) ||
 437             data_length > udev->data_size) {
 438                 pr_warn("TCMU: Request of size %zu/%zu is too big for %u/%zu "
 439                         "cmd ring/data area\n", command_size, data_length,
 440                         udev->cmdr_size, udev->data_size);
 441                 spin_unlock_irq(&udev->cmdr_lock);
 442                 return TCM_INVALID_CDB_FIELD;
 443         }
 444
 445         while (!is_ring_space_avail(udev, command_size, data_length)) {
 446                 int ret;
 447                 DEFINE_WAIT(__wait);
 448
 449                 prepare_to_wait(&udev->wait_cmdr, &__wait, TASK_INTERRUPTIBLE);
 450
 451                 pr_debug("sleeping for ring space\n");
 452                 spin_unlock_irq(&udev->cmdr_lock);
 453                 ret = schedule_timeout(msecs_to_jiffies(TCMU_TIME_OUT));
 454                 finish_wait(&udev->wait_cmdr, &__wait);
 455                 if (!ret) {
 456                         pr_warn("tcmu: command timed out\n");
 457                         return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
 458                 }
 459
 460                 spin_lock_irq(&udev->cmdr_lock);
 461
 462                 /* We dropped cmdr_lock, cmd_head is stale */
 463                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 464         }
 465
 466         /* Insert a PAD if end-of-ring space is too small */
 467         if (head_to_end(cmd_head, udev->cmdr_size) < command_size) {
 468                 size_t pad_size = head_to_end(cmd_head, udev->cmdr_size);
 469
 470                 entry = (void *) mb + CMDR_OFF + cmd_head;
 471                 tcmu_flush_dcache_range(entry, sizeof(*entry));
 472                 tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_PAD);
 473                 tcmu_hdr_set_len(&entry->hdr.len_op, pad_size);
 474                 entry->hdr.cmd_id = 0; /* not used for PAD */
 475                 entry->hdr.kflags = 0;
 476                 entry->hdr.uflags = 0;
 477
 478                 UPDATE_HEAD(mb->cmd_head, pad_size, udev->cmdr_size);
 479
 480                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 481                 WARN_ON(cmd_head != 0);
 482         }
 483
 484         entry = (void *) mb + CMDR_OFF + cmd_head;
 485         tcmu_flush_dcache_range(entry, sizeof(*entry));
 486         tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_CMD);
 487         tcmu_hdr_set_len(&entry->hdr.len_op, command_size);
 488         entry->hdr.cmd_id = tcmu_cmd->cmd_id;
 489         entry->hdr.kflags = 0;
 490         entry->hdr.uflags = 0;
 491
 492         bitmap_copy(old_bitmap, udev->data_bitmap, DATA_BLOCK_BITS);
 493
 494         /* Handle allocating space from the data area */
 495         iov = &entry->req.iov[0];
 496         iov_cnt = 0;
 497         copy_to_data_area = (se_cmd->data_direction == DMA_TO_DEVICE
 498                 || se_cmd->se_cmd_flags & SCF_BIDI);
 499         alloc_and_scatter_data_area(udev, se_cmd->t_data_sg,
 500                 se_cmd->t_data_nents, &iov, &iov_cnt, copy_to_data_area);
 501         entry->req.iov_cnt = iov_cnt;
 502         entry->req.iov_dif_cnt = 0;
 503
 504         /* Handle BIDI commands */
 505         iov_cnt = 0;
 506         alloc_and_scatter_data_area(udev, se_cmd->t_bidi_data_sg,
 507                 se_cmd->t_bidi_data_nents, &iov, &iov_cnt, false);
 508         entry->req.iov_bidi_cnt = iov_cnt;
 509
 510         /* cmd's data_bitmap is what changed in process */
 511         bitmap_xor(tcmu_cmd->data_bitmap, old_bitmap, udev->data_bitmap,
 512                         DATA_BLOCK_BITS);
 513
 514         /* All offsets relative to mb_addr, not start of entry! */
 515         cdb_off = CMDR_OFF + cmd_head + base_command_size;
 516         memcpy((void *) mb + cdb_off, se_cmd->t_task_cdb, scsi_command_size(se_cmd->t_task_cdb));
 517         entry->req.cdb_off = cdb_off;
 518         tcmu_flush_dcache_range(entry, sizeof(*entry));
 519
 520         UPDATE_HEAD(mb->cmd_head, command_size, udev->cmdr_size);
 521         tcmu_flush_dcache_range(mb, sizeof(*mb));
 522
 523         spin_unlock_irq(&udev->cmdr_lock);
 524
 525         /* TODO: only if FLUSH and FUA? */
 526         uio_event_notify(&udev->uio_info);
 527
 528         mod_timer(&udev->timeout,
 529                 round_jiffies_up(jiffies + msecs_to_jiffies(TCMU_TIME_OUT)));
 530
 531         return TCM_NO_SENSE;
 532 }
 533
 534 static sense_reason_t
 535 tcmu_queue_cmd(struct se_cmd *se_cmd)
 536 {
 537         struct se_device *se_dev = se_cmd->se_dev;
 538         struct tcmu_dev *udev = TCMU_DEV(se_dev);
 539         struct tcmu_cmd *tcmu_cmd;
 540         int ret;
 541
 542         tcmu_cmd = tcmu_alloc_cmd(se_cmd);
 543         if (!tcmu_cmd)
 544                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
 545
 546         ret = tcmu_queue_cmd_ring(tcmu_cmd);
 547         if (ret != TCM_NO_SENSE) {
 548                 pr_err("TCMU: Could not queue command\n");
 549                 spin_lock_irq(&udev->commands_lock);
 550                 idr_remove(&udev->commands, tcmu_cmd->cmd_id);
 551                 spin_unlock_irq(&udev->commands_lock);
 552
 553                 kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
 554         }
 555
 556         return ret;
 557 }
 558
 559 static void tcmu_handle_completion(struct tcmu_cmd *cmd, struct tcmu_cmd_entry *entry)
 560 {
 561         struct se_cmd *se_cmd = cmd->se_cmd;
 562         struct tcmu_dev *udev = cmd->tcmu_dev;
 563
 564         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
 565                 /*
 566                  * cmd has been completed already from timeout, just reclaim
 567                  * data area space and free cmd
 568                  */
 569                 free_data_area(udev, cmd);
 570
 571                 kmem_cache_free(tcmu_cmd_cache, cmd);
 572                 return;
 573         }
 574
 575         if (entry->hdr.uflags & TCMU_UFLAG_UNKNOWN_OP) {
 576                 free_data_area(udev, cmd);
 577                 pr_warn("TCMU: Userspace set UNKNOWN_OP flag on se_cmd %p\n",
 578                         cmd->se_cmd);
 579                 entry->rsp.scsi_status = SAM_STAT_CHECK_CONDITION;
 580         } else if (entry->rsp.scsi_status == SAM_STAT_CHECK_CONDITION) {
 581                 memcpy(se_cmd->sense_buffer, entry->rsp.sense_buffer,
 582                                se_cmd->scsi_sense_length);
 583                 free_data_area(udev, cmd);
 584         } else if (se_cmd->se_cmd_flags & SCF_BIDI) {
 585                 DECLARE_BITMAP(bitmap, DATA_BLOCK_BITS);
 586
 587                 /* Get Data-In buffer before clean up */
 588                 bitmap_copy(bitmap, cmd->data_bitmap, DATA_BLOCK_BITS);
 589                 gather_data_area(udev, bitmap,
 590                         se_cmd->t_bidi_data_sg, se_cmd->t_bidi_data_nents);
 591                 free_data_area(udev, cmd);
 592         } else if (se_cmd->data_direction == DMA_FROM_DEVICE) {
 593                 DECLARE_BITMAP(bitmap, DATA_BLOCK_BITS);
 594
 595                 bitmap_copy(bitmap, cmd->data_bitmap, DATA_BLOCK_BITS);
 596                 gather_data_area(udev, bitmap,
 597                         se_cmd->t_data_sg, se_cmd->t_data_nents);
 598                 free_data_area(udev, cmd);
 599         } else if (se_cmd->data_direction == DMA_TO_DEVICE) {
 600                 free_data_area(udev, cmd);
 601         } else if (se_cmd->data_direction != DMA_NONE) {
 602                 pr_warn("TCMU: data direction was %d!\n",
 603                         se_cmd->data_direction);
 604         }
 605
 606         target_complete_cmd(cmd->se_cmd, entry->rsp.scsi_status);
 607         cmd->se_cmd = NULL;
 608
 609         kmem_cache_free(tcmu_cmd_cache, cmd);
 610 }
 611
 612 static unsigned int tcmu_handle_completions(struct tcmu_dev *udev)
 613 {
 614         struct tcmu_mailbox *mb;
 615         unsigned long flags;
 616         int handled = 0;
 617
 618         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags)) {
 619                 pr_err("ring broken, not handling completions\n");
 620                 return 0;
 621         }
 622
 623         spin_lock_irqsave(&udev->cmdr_lock, flags);
 624
 625         mb = udev->mb_addr;
 626         tcmu_flush_dcache_range(mb, sizeof(*mb));
 627
 628         while (udev->cmdr_last_cleaned != ACCESS_ONCE(mb->cmd_tail)) {
 629
 630                 struct tcmu_cmd_entry *entry = (void *) mb + CMDR_OFF + udev->cmdr_last_cleaned;
 631                 struct tcmu_cmd *cmd;
 632
 633                 tcmu_flush_dcache_range(entry, sizeof(*entry));
 634
 635                 if (tcmu_hdr_get_op(entry->hdr.len_op) == TCMU_OP_PAD) {
 636                         UPDATE_HEAD(udev->cmdr_last_cleaned,
 637                                     tcmu_hdr_get_len(entry->hdr.len_op),
 638                                     udev->cmdr_size);
 639                         continue;
 640                 }
 641                 WARN_ON(tcmu_hdr_get_op(entry->hdr.len_op) != TCMU_OP_CMD);
 642
 643                 spin_lock(&udev->commands_lock);
 644                 cmd = idr_find(&udev->commands, entry->hdr.cmd_id);
 645                 if (cmd)
 646                         idr_remove(&udev->commands, cmd->cmd_id);
 647                 spin_unlock(&udev->commands_lock);
 648
 649                 if (!cmd) {
 650                         pr_err("cmd_id not found, ring is broken\n");
 651                         set_bit(TCMU_DEV_BIT_BROKEN, &udev->flags);
 652                         break;
 653                 }
 654
 655                 tcmu_handle_completion(cmd, entry);
 656
 657                 UPDATE_HEAD(udev->cmdr_last_cleaned,
 658                             tcmu_hdr_get_len(entry->hdr.len_op),
 659                             udev->cmdr_size);
 660
 661                 handled++;
 662         }
 663
 664         if (mb->cmd_tail == mb->cmd_head)
 665                 del_timer(&udev->timeout); /* no more pending cmds */
 666
 667         spin_unlock_irqrestore(&udev->cmdr_lock, flags);
 668
 669         wake_up(&udev->wait_cmdr);
 670
 671         return handled;
 672 }
 673
 674 static int tcmu_check_expired_cmd(int id, void *p, void *data)
 675 {
 676         struct tcmu_cmd *cmd = p;
 677
 678         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
 679                 return 0;
 680
 681         if (!time_after(jiffies, cmd->deadline))
 682                 return 0;
 683
 684         set_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags);
 685         target_complete_cmd(cmd->se_cmd, SAM_STAT_CHECK_CONDITION);
 686         cmd->se_cmd = NULL;
 687
 688         kmem_cache_free(tcmu_cmd_cache, cmd);
 689
 690         return 0;
 691 }
 692
 693 static void tcmu_device_timedout(unsigned long data)
 694 {
 695         struct tcmu_dev *udev = (struct tcmu_dev *)data;
 696         unsigned long flags;
 697         int handled;
 698
 699         handled = tcmu_handle_completions(udev);
 700
 701         pr_warn("%d completions handled from timeout\n", handled);
 702
 703         spin_lock_irqsave(&udev->commands_lock, flags);
 704         idr_for_each(&udev->commands, tcmu_check_expired_cmd, NULL);
 705         spin_unlock_irqrestore(&udev->commands_lock, flags);
 706
 707         /*
 708          * We don't need to wakeup threads on wait_cmdr since they have their
 709          * own timeout.
 710          */
 711 }
 712
 713 static int tcmu_attach_hba(struct se_hba *hba, u32 host_id)
 714 {
 715         struct tcmu_hba *tcmu_hba;
 716
 717         tcmu_hba = kzalloc(sizeof(struct tcmu_hba), GFP_KERNEL);
 718         if (!tcmu_hba)
 719                 return -ENOMEM;
 720
 721         tcmu_hba->host_id = host_id;
 722         hba->hba_ptr = tcmu_hba;
 723
 724         return 0;
 725 }
 726
 727 static void tcmu_detach_hba(struct se_hba *hba)
 728 {
 729         kfree(hba->hba_ptr);
 730         hba->hba_ptr = NULL;
 731 }
 732
 733 static struct se_device *tcmu_alloc_device(struct se_hba *hba, const char *name)
 734 {
 735         struct tcmu_dev *udev;
 736
 737         udev = kzalloc(sizeof(struct tcmu_dev), GFP_KERNEL);
 738         if (!udev)
 739                 return NULL;
 740
 741         udev->name = kstrdup(name, GFP_KERNEL);
 742         if (!udev->name) {
 743                 kfree(udev);
 744                 return NULL;
 745         }
 746
 747         udev->hba = hba;
 748
 749         init_waitqueue_head(&udev->wait_cmdr);
 750         spin_lock_init(&udev->cmdr_lock);
 751
 752         idr_init(&udev->commands);
 753         spin_lock_init(&udev->commands_lock);
 754
 755         setup_timer(&udev->timeout, tcmu_device_timedout,
 756                 (unsigned long)udev);
 757
 758         return &udev->se_dev;
 759 }
 760
 761 static int tcmu_irqcontrol(struct uio_info *info, s32 irq_on)
 762 {
 763         struct tcmu_dev *tcmu_dev = container_of(info, struct tcmu_dev, uio_info);
 764
 765         tcmu_handle_completions(tcmu_dev);
 766
 767         return 0;
 768 }
 769
 770 /*
 771  * mmap code from uio.c. Copied here because we want to hook mmap()
 772  * and this stuff must come along.
 773  */
 774 static int tcmu_find_mem_index(struct vm_area_struct *vma)
 775 {
 776         struct tcmu_dev *udev = vma->vm_private_data;
 777         struct uio_info *info = &udev->uio_info;
 778
 779         if (vma->vm_pgoff < MAX_UIO_MAPS) {
 780                 if (info->mem[vma->vm_pgoff].size == 0)
 781                         return -1;
 782                 return (int)vma->vm_pgoff;
 783         }
 784         return -1;
 785 }
 786
 787 static int tcmu_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 788 {
 789         struct tcmu_dev *udev = vma->vm_private_data;
 790         struct uio_info *info = &udev->uio_info;
 791         struct page *page;
 792         unsigned long offset;
 793         void *addr;
 794
 795         int mi = tcmu_find_mem_index(vma);
 796         if (mi < 0)
 797                 return VM_FAULT_SIGBUS;
 798
 799         /*
 800          * We need to subtract mi because userspace uses offset = N*PAGE_SIZE
 801          * to use mem[N].
 802          */
 803         offset = (vmf->pgoff - mi) << PAGE_SHIFT;
 804
 805         addr = (void *)(unsigned long)info->mem[mi].addr + offset;
 806         if (info->mem[mi].memtype == UIO_MEM_LOGICAL)
 807                 page = virt_to_page(addr);
 808         else
 809                 page = vmalloc_to_page(addr);
 810         get_page(page);
 811         vmf->page = page;
 812         return 0;
 813 }
 814
 815 static const struct vm_operations_struct tcmu_vm_ops = {
 816         .fault = tcmu_vma_fault,
 817 };
 818
 819 static int tcmu_mmap(struct uio_info *info, struct vm_area_struct *vma)
 820 {
 821         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 822
 823         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
 824         vma->vm_ops = &tcmu_vm_ops;
 825
 826         vma->vm_private_data = udev;
 827
 828         /* Ensure the mmap is exactly the right size */
 829         if (vma_pages(vma) != (TCMU_RING_SIZE >> PAGE_SHIFT))
 830                 return -EINVAL;
 831
 832         return 0;
 833 }
 834
 835 static int tcmu_open(struct uio_info *info, struct inode *inode)
 836 {
 837         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 838
 839         /* O_EXCL not supported for char devs, so fake it? */
 840         if (test_and_set_bit(TCMU_DEV_BIT_OPEN, &udev->flags))
 841                 return -EBUSY;
 842
 843         pr_debug("open\n");
 844
 845         return 0;
 846 }
 847
 848 static int tcmu_release(struct uio_info *info, struct inode *inode)
 849 {
 850         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 851
 852         clear_bit(TCMU_DEV_BIT_OPEN, &udev->flags);
 853
 854         pr_debug("close\n");
 855
 856         return 0;
 857 }
 858
 859 static int tcmu_netlink_event(enum tcmu_genl_cmd cmd, const char *name, int minor)
 860 {
 861         struct sk_buff *skb;
 862         void *msg_header;
 863         int ret = -ENOMEM;
 864
 865         skb = genlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
 866         if (!skb)
 867                 return ret;
 868
 869         msg_header = genlmsg_put(skb, 0, 0, &tcmu_genl_family, 0, cmd);
 870         if (!msg_header)
 871                 goto free_skb;
 872
 873         ret = nla_put_string(skb, TCMU_ATTR_DEVICE, name);
 874         if (ret < 0)
 875                 goto free_skb;
 876
 877         ret = nla_put_u32(skb, TCMU_ATTR_MINOR, minor);
 878         if (ret < 0)
 879                 goto free_skb;
 880
 881         genlmsg_end(skb, msg_header);
 882
 883         ret = genlmsg_multicast_allns(&tcmu_genl_family, skb, 0,
 884                                 TCMU_MCGRP_CONFIG, GFP_KERNEL);
 885
 886         /* We don't care if no one is listening */
 887         if (ret == -ESRCH)
 888                 ret = 0;
 889
 890         return ret;
 891 free_skb:
 892         nlmsg_free(skb);
 893         return ret;
 894 }
 895
 896 static int tcmu_configure_device(struct se_device *dev)
 897 {
 898         struct tcmu_dev *udev = TCMU_DEV(dev);
 899         struct tcmu_hba *hba = udev->hba->hba_ptr;
 900         struct uio_info *info;
 901         struct tcmu_mailbox *mb;
 902         size_t size;
 903         size_t used;
 904         int ret = 0;
 905         char *str;
 906
 907         info = &udev->uio_info;
 908
 909         size = snprintf(NULL, 0, "tcm-user/%u/%s/%s", hba->host_id, udev->name,
 910                         udev->dev_config);
 911         size += 1; /* for \0 */
 912         str = kmalloc(size, GFP_KERNEL);
 913         if (!str)
 914                 return -ENOMEM;
 915
 916         used = snprintf(str, size, "tcm-user/%u/%s", hba->host_id, udev->name);
 917
 918         if (udev->dev_config[0])
 919                 snprintf(str + used, size - used, "/%s", udev->dev_config);
 920
 921         info->name = str;
 922
 923         udev->mb_addr = vzalloc(TCMU_RING_SIZE);
 924         if (!udev->mb_addr) {
 925                 ret = -ENOMEM;
 926                 goto err_vzalloc;
 927         }
 928
 929         /* mailbox fits in first part of CMDR space */
 930         udev->cmdr_size = CMDR_SIZE - CMDR_OFF;
 931         udev->data_off = CMDR_SIZE;
 932         udev->data_size = TCMU_RING_SIZE - CMDR_SIZE;
 933
 934         mb = udev->mb_addr;
 935         mb->version = TCMU_MAILBOX_VERSION;
 936         mb->flags = TCMU_MAILBOX_FLAG_CAP_OOOC;
 937         mb->cmdr_off = CMDR_OFF;
 938         mb->cmdr_size = udev->cmdr_size;
 939
 940         WARN_ON(!PAGE_ALIGNED(udev->data_off));
 941         WARN_ON(udev->data_size % PAGE_SIZE);
 942         WARN_ON(udev->data_size % DATA_BLOCK_SIZE);
 943
 944         info->version = __stringify(TCMU_MAILBOX_VERSION);
 945
 946         info->mem[0].name = "tcm-user command & data buffer";
 947         info->mem[0].addr = (phys_addr_t)(uintptr_t)udev->mb_addr;
 948         info->mem[0].size = TCMU_RING_SIZE;
 949         info->mem[0].memtype = UIO_MEM_VIRTUAL;
 950
 951         info->irqcontrol = tcmu_irqcontrol;
 952         info->irq = UIO_IRQ_CUSTOM;
 953
 954         info->mmap = tcmu_mmap;
 955         info->open = tcmu_open;
 956         info->release = tcmu_release;
 957
 958         ret = uio_register_device(tcmu_root_device, info);
 959         if (ret)
 960                 goto err_register;
 961
 962         /* User can set hw_block_size before enable the device */
 963         if (dev->dev_attrib.hw_block_size == 0)
 964                 dev->dev_attrib.hw_block_size = 512;
 965         /* Other attributes can be configured in userspace */
 966         dev->dev_attrib.hw_max_sectors = 128;
 967         dev->dev_attrib.hw_queue_depth = 128;
 968
 969         ret = tcmu_netlink_event(TCMU_CMD_ADDED_DEVICE, udev->uio_info.name,
 970                                  udev->uio_info.uio_dev->minor);
 971         if (ret)
 972                 goto err_netlink;
 973
 974         return 0;
 975
 976 err_netlink:
 977         uio_unregister_device(&udev->uio_info);
 978 err_register:
 979         vfree(udev->mb_addr);
 980 err_vzalloc:
 981         kfree(info->name);
 982
 983         return ret;
 984 }
 985
 986 static int tcmu_check_and_free_pending_cmd(struct tcmu_cmd *cmd)
 987 {
 988         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
 989                 kmem_cache_free(tcmu_cmd_cache, cmd);
 990                 return 0;
 991         }
 992         return -EINVAL;
 993 }
 994
 995 static void tcmu_dev_call_rcu(struct rcu_head *p)
 996 {
 997         struct se_device *dev = container_of(p, struct se_device, rcu_head);
 998         struct tcmu_dev *udev = TCMU_DEV(dev);
 999
1000         kfree(udev);
1001 }
1002
1003 static void tcmu_free_device(struct se_device *dev)
1004 {
1005         struct tcmu_dev *udev = TCMU_DEV(dev);
1006         struct tcmu_cmd *cmd;
1007         bool all_expired = true;
1008         int i;
1009
1010         del_timer_sync(&udev->timeout);
1011
1012         vfree(udev->mb_addr);
1013
1014         /* Upper layer should drain all requests before calling this */
1015         spin_lock_irq(&udev->commands_lock);
1016         idr_for_each_entry(&udev->commands, cmd, i) {
1017                 if (tcmu_check_and_free_pending_cmd(cmd) != 0)
1018                         all_expired = false;
1019         }
1020         idr_destroy(&udev->commands);
1021         spin_unlock_irq(&udev->commands_lock);
1022         WARN_ON(!all_expired);
1023
1024         /* Device was configured */
1025         if (udev->uio_info.uio_dev) {
1026                 tcmu_netlink_event(TCMU_CMD_REMOVED_DEVICE, udev->uio_info.name,
1027                                    udev->uio_info.uio_dev->minor);
1028
1029                 uio_unregister_device(&udev->uio_info);
1030                 kfree(udev->uio_info.name);
1031                 kfree(udev->name);
1032         }
1033         call_rcu(&dev->rcu_head, tcmu_dev_call_rcu);
1034 }
1035
1036 enum {
1037         Opt_dev_config, Opt_dev_size, Opt_hw_block_size, Opt_err,
1038 };
1039
1040 static match_table_t tokens = {
1041         {Opt_dev_config, "dev_config=%s"},
1042         {Opt_dev_size, "dev_size=%u"},
1043         {Opt_hw_block_size, "hw_block_size=%u"},
1044         {Opt_err, NULL}
1045 };
1046
1047 static ssize_t tcmu_set_configfs_dev_params(struct se_device *dev,
1048                 const char *page, ssize_t count)
1049 {
1050         struct tcmu_dev *udev = TCMU_DEV(dev);
1051         char *orig, *ptr, *opts, *arg_p;
1052         substring_t args[MAX_OPT_ARGS];
1053         int ret = 0, token;
1054         unsigned long tmp_ul;
1055
1056         opts = kstrdup(page, GFP_KERNEL);
1057         if (!opts)
1058                 return -ENOMEM;
1059
1060         orig = opts;
1061
1062         while ((ptr = strsep(&opts, ",\n")) != NULL) {
1063                 if (!*ptr)
1064                         continue;
1065
1066                 token = match_token(ptr, tokens, args);
1067                 switch (token) {
1068                 case Opt_dev_config:
1069                         if (match_strlcpy(udev->dev_config, &args[0],
1070                                           TCMU_CONFIG_LEN) == 0) {
1071                                 ret = -EINVAL;
1072                                 break;
1073                         }
1074                         pr_debug("TCMU: Referencing Path: %s\n", udev->dev_config);
1075                         break;
1076                 case Opt_dev_size:
1077                         arg_p = match_strdup(&args[0]);
1078                         if (!arg_p) {
1079                                 ret = -ENOMEM;
1080                                 break;
1081                         }
1082                         ret = kstrtoul(arg_p, 0, (unsigned long *) &udev->dev_size);
1083                         kfree(arg_p);
1084                         if (ret < 0)
1085                                 pr_err("kstrtoul() failed for dev_size=\n");
1086                         break;
1087                 case Opt_hw_block_size:
1088                         arg_p = match_strdup(&args[0]);
1089                         if (!arg_p) {
1090                                 ret = -ENOMEM;
1091                                 break;
1092                         }
1093                         ret = kstrtoul(arg_p, 0, &tmp_ul);
1094                         kfree(arg_p);
1095                         if (ret < 0) {
1096                                 pr_err("kstrtoul() failed for hw_block_size=\n");
1097                                 break;
1098                         }
1099                         if (!tmp_ul) {
1100                                 pr_err("hw_block_size must be nonzero\n");
1101                                 break;
1102                         }
1103                         dev->dev_attrib.hw_block_size = tmp_ul;
1104                         break;
1105                 default:
1106                         break;
1107                 }
1108         }
1109
1110         kfree(orig);
1111         return (!ret) ? count : ret;
1112 }
1113
1114 static ssize_t tcmu_show_configfs_dev_params(struct se_device *dev, char *b)
1115 {
1116         struct tcmu_dev *udev = TCMU_DEV(dev);
1117         ssize_t bl = 0;
1118
1119         bl = sprintf(b + bl, "Config: %s ",
1120                      udev->dev_config[0] ? udev->dev_config : "NULL");
1121         bl += sprintf(b + bl, "Size: %zu\n", udev->dev_size);
1122
1123         return bl;
1124 }
1125
1126 static sector_t tcmu_get_blocks(struct se_device *dev)
1127 {
1128         struct tcmu_dev *udev = TCMU_DEV(dev);
1129
1130         return div_u64(udev->dev_size - dev->dev_attrib.block_size,
1131                        dev->dev_attrib.block_size);
1132 }
1133
1134 static sense_reason_t
1135 tcmu_parse_cdb(struct se_cmd *cmd)
1136 {
1137         return passthrough_parse_cdb(cmd, tcmu_queue_cmd);
1138 }
1139
1140 static const struct target_backend_ops tcmu_ops = {
1141         .name                   = "user",
1142         .owner                  = THIS_MODULE,
1143         .transport_flags        = TRANSPORT_FLAG_PASSTHROUGH,
1144         .attach_hba             = tcmu_attach_hba,
1145         .detach_hba             = tcmu_detach_hba,
1146         .alloc_device           = tcmu_alloc_device,
1147         .configure_device       = tcmu_configure_device,
1148         .free_device            = tcmu_free_device,
1149         .parse_cdb              = tcmu_parse_cdb,
1150         .set_configfs_dev_params = tcmu_set_configfs_dev_params,
1151         .show_configfs_dev_params = tcmu_show_configfs_dev_params,
1152         .get_device_type        = sbc_get_device_type,
1153         .get_blocks             = tcmu_get_blocks,
1154         .tb_dev_attrib_attrs    = passthrough_attrib_attrs,
1155 };
1156
1157 static int __init tcmu_module_init(void)
1158 {
1159         int ret;
1160
1161         BUILD_BUG_ON((sizeof(struct tcmu_cmd_entry) % TCMU_OP_ALIGN_SIZE) != 0);
1162
1163         tcmu_cmd_cache = kmem_cache_create("tcmu_cmd_cache",
1164                                 sizeof(struct tcmu_cmd),
1165                                 __alignof__(struct tcmu_cmd),
1166                                 0, NULL);
1167         if (!tcmu_cmd_cache)
1168                 return -ENOMEM;
1169
1170         tcmu_root_device = root_device_register("tcm_user");
1171         if (IS_ERR(tcmu_root_device)) {
1172                 ret = PTR_ERR(tcmu_root_device);
1173                 goto out_free_cache;
1174         }
1175
1176         ret = genl_register_family(&tcmu_genl_family);
1177         if (ret < 0) {
1178                 goto out_unreg_device;
1179         }
1180
1181         ret = transport_backend_register(&tcmu_ops);
1182         if (ret)
1183                 goto out_unreg_genl;
1184
1185         return 0;
1186
1187 out_unreg_genl:
1188         genl_unregister_family(&tcmu_genl_family);
1189 out_unreg_device:
1190         root_device_unregister(tcmu_root_device);
1191 out_free_cache:
1192         kmem_cache_destroy(tcmu_cmd_cache);
1193
1194         return ret;
1195 }
1196
1197 static void __exit tcmu_module_exit(void)
1198 {
1199         target_backend_unregister(&tcmu_ops);
1200         genl_unregister_family(&tcmu_genl_family);
1201         root_device_unregister(tcmu_root_device);
1202         kmem_cache_destroy(tcmu_cmd_cache);
1203 }
1204
1205 MODULE_DESCRIPTION("TCM USER subsystem plugin");
1206 MODULE_AUTHOR("Shaohua Li <shli@kernel.org>");
1207 MODULE_AUTHOR("Andy Grover <agrover@redhat.com>");
1208 MODULE_LICENSE("GPL");
1209
1210 module_init(tcmu_module_init);
1211 module_exit(tcmu_module_exit);