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  * Copyright (C) 2017 Chinamobile, Inc.
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms and conditions of the GNU General Public License,
   9  * version 2, as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  *
  16  * You should have received a copy of the GNU General Public License along with
  17  * this program; if not, write to the Free Software Foundation, Inc.,
  18  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  19  */
  20
  21 #include <linux/spinlock.h>
  22 #include <linux/module.h>
  23 #include <linux/idr.h>
  24 #include <linux/kernel.h>
  25 #include <linux/timer.h>
  26 #include <linux/parser.h>
  27 #include <linux/vmalloc.h>
  28 #include <linux/uio_driver.h>
  29 #include <linux/radix-tree.h>
  30 #include <linux/stringify.h>
  31 #include <linux/bitops.h>
  32 #include <linux/highmem.h>
  33 #include <linux/configfs.h>
  34 #include <linux/mutex.h>
  35 #include <linux/kthread.h>
  36 #include <net/genetlink.h>
  37 #include <scsi/scsi_common.h>
  38 #include <scsi/scsi_proto.h>
  39 #include <target/target_core_base.h>
  40 #include <target/target_core_fabric.h>
  41 #include <target/target_core_backend.h>
  42
  43 #include <linux/target_core_user.h>
  44
  45 /*
  46  * Define a shared-memory interface for LIO to pass SCSI commands and
  47  * data to userspace for processing. This is to allow backends that
  48  * are too complex for in-kernel support to be possible.
  49  *
  50  * It uses the UIO framework to do a lot of the device-creation and
  51  * introspection work for us.
  52  *
  53  * See the .h file for how the ring is laid out. Note that while the
  54  * command ring is defined, the particulars of the data area are
  55  * not. Offset values in the command entry point to other locations
  56  * internal to the mmap()ed area. There is separate space outside the
  57  * command ring for data buffers. This leaves maximum flexibility for
  58  * moving buffer allocations, or even page flipping or other
  59  * allocation techniques, without altering the command ring layout.
  60  *
  61  * SECURITY:
  62  * The user process must be assumed to be malicious. There's no way to
  63  * prevent it breaking the command ring protocol if it wants, but in
  64  * order to prevent other issues we must only ever read *data* from
  65  * the shared memory area, not offsets or sizes. This applies to
  66  * command ring entries as well as the mailbox. Extra code needed for
  67  * this may have a 'UAM' comment.
  68  */
  69
  70 #define TCMU_TIME_OUT (30 * MSEC_PER_SEC)
  71
  72 /* For cmd area, the size is fixed 8MB */
  73 #define CMDR_SIZE (8 * 1024 * 1024)
  74
  75 /*
  76  * For data area, the block size is PAGE_SIZE and
  77  * the total size is 256K * PAGE_SIZE.
  78  */
  79 #define DATA_BLOCK_SIZE PAGE_SIZE
  80 #define DATA_BLOCK_BITS (256 * 1024)
  81 #define DATA_SIZE (DATA_BLOCK_BITS * DATA_BLOCK_SIZE)
  82 #define DATA_BLOCK_INIT_BITS 128
  83
  84 /* The total size of the ring is 8M + 256K * PAGE_SIZE */
  85 #define TCMU_RING_SIZE (CMDR_SIZE + DATA_SIZE)
  86
  87 /* Default maximum of the global data blocks(512K * PAGE_SIZE) */
  88 #define TCMU_GLOBAL_MAX_BLOCKS (512 * 1024)
  89
  90 static struct device *tcmu_root_device;
  91
  92 struct tcmu_hba {
  93         u32 host_id;
  94 };
  95
  96 #define TCMU_CONFIG_LEN 256
  97
  98 struct tcmu_dev {
  99         struct list_head node;
 100         struct kref kref;
 101         struct se_device se_dev;
 102
 103         char *name;
 104         struct se_hba *hba;
 105
 106 #define TCMU_DEV_BIT_OPEN 0
 107 #define TCMU_DEV_BIT_BROKEN 1
 108         unsigned long flags;
 109
 110         struct uio_info uio_info;
 111
 112         struct inode *inode;
 113
 114         struct tcmu_mailbox *mb_addr;
 115         size_t dev_size;
 116         u32 cmdr_size;
 117         u32 cmdr_last_cleaned;
 118         /* Offset of data area from start of mb */
 119         /* Must add data_off and mb_addr to get the address */
 120         size_t data_off;
 121         size_t data_size;
 122
 123         wait_queue_head_t wait_cmdr;
 124         struct mutex cmdr_lock;
 125
 126         bool waiting_global;
 127         uint32_t dbi_max;
 128         uint32_t dbi_thresh;
 129         DECLARE_BITMAP(data_bitmap, DATA_BLOCK_BITS);
 130         struct radix_tree_root data_blocks;
 131
 132         struct idr commands;
 133         spinlock_t commands_lock;
 134
 135         struct timer_list timeout;
 136         unsigned int cmd_time_out;
 137
 138         char dev_config[TCMU_CONFIG_LEN];
 139 };
 140
 141 #define TCMU_DEV(_se_dev) container_of(_se_dev, struct tcmu_dev, se_dev)
 142
 143 #define CMDR_OFF sizeof(struct tcmu_mailbox)
 144
 145 struct tcmu_cmd {
 146         struct se_cmd *se_cmd;
 147         struct tcmu_dev *tcmu_dev;
 148
 149         uint16_t cmd_id;
 150
 151         /* Can't use se_cmd when cleaning up expired cmds, because if
 152            cmd has been completed then accessing se_cmd is off limits */
 153         uint32_t dbi_cnt;
 154         uint32_t dbi_cur;
 155         uint32_t *dbi;
 156
 157         unsigned long deadline;
 158
 159 #define TCMU_CMD_BIT_EXPIRED 0
 160         unsigned long flags;
 161 };
 162
 163 static struct task_struct *unmap_thread;
 164 static wait_queue_head_t unmap_wait;
 165 static DEFINE_MUTEX(root_udev_mutex);
 166 static LIST_HEAD(root_udev);
 167
 168 static atomic_t global_db_count = ATOMIC_INIT(0);
 169
 170 static struct kmem_cache *tcmu_cmd_cache;
 171
 172 /* multicast group */
 173 enum tcmu_multicast_groups {
 174         TCMU_MCGRP_CONFIG,
 175 };
 176
 177 static const struct genl_multicast_group tcmu_mcgrps[] = {
 178         [TCMU_MCGRP_CONFIG] = { .name = "config", },
 179 };
 180
 181 /* Our generic netlink family */
 182 static struct genl_family tcmu_genl_family __ro_after_init = {
 183         .module = THIS_MODULE,
 184         .hdrsize = 0,
 185         .name = "TCM-USER",
 186         .version = 1,
 187         .maxattr = TCMU_ATTR_MAX,
 188         .mcgrps = tcmu_mcgrps,
 189         .n_mcgrps = ARRAY_SIZE(tcmu_mcgrps),
 190         .netnsok = true,
 191 };
 192
 193 #define tcmu_cmd_set_dbi_cur(cmd, index) ((cmd)->dbi_cur = (index))
 194 #define tcmu_cmd_reset_dbi_cur(cmd) tcmu_cmd_set_dbi_cur(cmd, 0)
 195 #define tcmu_cmd_set_dbi(cmd, index) ((cmd)->dbi[(cmd)->dbi_cur++] = (index))
 196 #define tcmu_cmd_get_dbi(cmd) ((cmd)->dbi[(cmd)->dbi_cur++])
 197
 198 static void tcmu_cmd_free_data(struct tcmu_cmd *tcmu_cmd, uint32_t len)
 199 {
 200         struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
 201         uint32_t i;
 202
 203         for (i = 0; i < len; i++)
 204                 clear_bit(tcmu_cmd->dbi[i], udev->data_bitmap);
 205 }
 206
 207 static inline bool tcmu_get_empty_block(struct tcmu_dev *udev,
 208                                         struct tcmu_cmd *tcmu_cmd)
 209 {
 210         struct page *page;
 211         int ret, dbi;
 212
 213         dbi = find_first_zero_bit(udev->data_bitmap, udev->dbi_thresh);
 214         if (dbi == udev->dbi_thresh)
 215                 return false;
 216
 217         page = radix_tree_lookup(&udev->data_blocks, dbi);
 218         if (!page) {
 219
 220                 if (atomic_add_return(1, &global_db_count) >
 221                                         TCMU_GLOBAL_MAX_BLOCKS) {
 222                         atomic_dec(&global_db_count);
 223                         return false;
 224                 }
 225
 226                 /* try to get new page from the mm */
 227                 page = alloc_page(GFP_KERNEL);
 228                 if (!page)
 229                         return false;
 230
 231                 ret = radix_tree_insert(&udev->data_blocks, dbi, page);
 232                 if (ret) {
 233                         __free_page(page);
 234                         return false;
 235                 }
 236
 237         }
 238
 239         if (dbi > udev->dbi_max)
 240                 udev->dbi_max = dbi;
 241
 242         set_bit(dbi, udev->data_bitmap);
 243         tcmu_cmd_set_dbi(tcmu_cmd, dbi);
 244
 245         return true;
 246 }
 247
 248 static bool tcmu_get_empty_blocks(struct tcmu_dev *udev,
 249                                   struct tcmu_cmd *tcmu_cmd)
 250 {
 251         int i;
 252
 253         udev->waiting_global = false;
 254
 255         for (i = tcmu_cmd->dbi_cur; i < tcmu_cmd->dbi_cnt; i++) {
 256                 if (!tcmu_get_empty_block(udev, tcmu_cmd))
 257                         goto err;
 258         }
 259         return true;
 260
 261 err:
 262         udev->waiting_global = true;
 263         /* Try to wake up the unmap thread */
 264         wake_up(&unmap_wait);
 265         return false;
 266 }
 267
 268 static inline struct page *
 269 tcmu_get_block_page(struct tcmu_dev *udev, uint32_t dbi)
 270 {
 271         return radix_tree_lookup(&udev->data_blocks, dbi);
 272 }
 273
 274 static inline void tcmu_free_cmd(struct tcmu_cmd *tcmu_cmd)
 275 {
 276         kfree(tcmu_cmd->dbi);
 277         kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
 278 }
 279
 280 static inline size_t tcmu_cmd_get_data_length(struct tcmu_cmd *tcmu_cmd)
 281 {
 282         struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
 283         size_t data_length = round_up(se_cmd->data_length, DATA_BLOCK_SIZE);
 284
 285         if (se_cmd->se_cmd_flags & SCF_BIDI) {
 286                 BUG_ON(!(se_cmd->t_bidi_data_sg && se_cmd->t_bidi_data_nents));
 287                 data_length += round_up(se_cmd->t_bidi_data_sg->length,
 288                                 DATA_BLOCK_SIZE);
 289         }
 290
 291         return data_length;
 292 }
 293
 294 static inline uint32_t tcmu_cmd_get_block_cnt(struct tcmu_cmd *tcmu_cmd)
 295 {
 296         size_t data_length = tcmu_cmd_get_data_length(tcmu_cmd);
 297
 298         return data_length / DATA_BLOCK_SIZE;
 299 }
 300
 301 static struct tcmu_cmd *tcmu_alloc_cmd(struct se_cmd *se_cmd)
 302 {
 303         struct se_device *se_dev = se_cmd->se_dev;
 304         struct tcmu_dev *udev = TCMU_DEV(se_dev);
 305         struct tcmu_cmd *tcmu_cmd;
 306         int cmd_id;
 307
 308         tcmu_cmd = kmem_cache_zalloc(tcmu_cmd_cache, GFP_KERNEL);
 309         if (!tcmu_cmd)
 310                 return NULL;
 311
 312         tcmu_cmd->se_cmd = se_cmd;
 313         tcmu_cmd->tcmu_dev = udev;
 314         if (udev->cmd_time_out)
 315                 tcmu_cmd->deadline = jiffies +
 316                                         msecs_to_jiffies(udev->cmd_time_out);
 317
 318         tcmu_cmd_reset_dbi_cur(tcmu_cmd);
 319         tcmu_cmd->dbi_cnt = tcmu_cmd_get_block_cnt(tcmu_cmd);
 320         tcmu_cmd->dbi = kcalloc(tcmu_cmd->dbi_cnt, sizeof(uint32_t),
 321                                 GFP_KERNEL);
 322         if (!tcmu_cmd->dbi) {
 323                 kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
 324                 return NULL;
 325         }
 326
 327         idr_preload(GFP_KERNEL);
 328         spin_lock_irq(&udev->commands_lock);
 329         cmd_id = idr_alloc(&udev->commands, tcmu_cmd, 0,
 330                 USHRT_MAX, GFP_NOWAIT);
 331         spin_unlock_irq(&udev->commands_lock);
 332         idr_preload_end();
 333
 334         if (cmd_id < 0) {
 335                 tcmu_free_cmd(tcmu_cmd);
 336                 return NULL;
 337         }
 338         tcmu_cmd->cmd_id = cmd_id;
 339
 340         return tcmu_cmd;
 341 }
 342
 343 static inline void tcmu_flush_dcache_range(void *vaddr, size_t size)
 344 {
 345         unsigned long offset = offset_in_page(vaddr);
 346
 347         size = round_up(size+offset, PAGE_SIZE);
 348         vaddr -= offset;
 349
 350         while (size) {
 351                 flush_dcache_page(virt_to_page(vaddr));
 352                 size -= PAGE_SIZE;
 353         }
 354 }
 355
 356 /*
 357  * Some ring helper functions. We don't assume size is a power of 2 so
 358  * we can't use circ_buf.h.
 359  */
 360 static inline size_t spc_used(size_t head, size_t tail, size_t size)
 361 {
 362         int diff = head - tail;
 363
 364         if (diff >= 0)
 365                 return diff;
 366         else
 367                 return size + diff;
 368 }
 369
 370 static inline size_t spc_free(size_t head, size_t tail, size_t size)
 371 {
 372         /* Keep 1 byte unused or we can't tell full from empty */
 373         return (size - spc_used(head, tail, size) - 1);
 374 }
 375
 376 static inline size_t head_to_end(size_t head, size_t size)
 377 {
 378         return size - head;
 379 }
 380
 381 static inline void new_iov(struct iovec **iov, int *iov_cnt,
 382                            struct tcmu_dev *udev)
 383 {
 384         struct iovec *iovec;
 385
 386         if (*iov_cnt != 0)
 387                 (*iov)++;
 388         (*iov_cnt)++;
 389
 390         iovec = *iov;
 391         memset(iovec, 0, sizeof(struct iovec));
 392 }
 393
 394 #define UPDATE_HEAD(head, used, size) smp_store_release(&head, ((head % size) + used) % size)
 395
 396 /* offset is relative to mb_addr */
 397 static inline size_t get_block_offset_user(struct tcmu_dev *dev,
 398                 int dbi, int remaining)
 399 {
 400         return dev->data_off + dbi * DATA_BLOCK_SIZE +
 401                 DATA_BLOCK_SIZE - remaining;
 402 }
 403
 404 static inline size_t iov_tail(struct tcmu_dev *udev, struct iovec *iov)
 405 {
 406         return (size_t)iov->iov_base + iov->iov_len;
 407 }
 408
 409 static int scatter_data_area(struct tcmu_dev *udev,
 410         struct tcmu_cmd *tcmu_cmd, struct scatterlist *data_sg,
 411         unsigned int data_nents, struct iovec **iov,
 412         int *iov_cnt, bool copy_data)
 413 {
 414         int i, dbi;
 415         int block_remaining = 0;
 416         void *from, *to = NULL;
 417         size_t copy_bytes, to_offset, offset;
 418         struct scatterlist *sg;
 419         struct page *page;
 420
 421         for_each_sg(data_sg, sg, data_nents, i) {
 422                 int sg_remaining = sg->length;
 423                 from = kmap_atomic(sg_page(sg)) + sg->offset;
 424                 while (sg_remaining > 0) {
 425                         if (block_remaining == 0) {
 426                                 if (to)
 427                                         kunmap_atomic(to);
 428
 429                                 block_remaining = DATA_BLOCK_SIZE;
 430                                 dbi = tcmu_cmd_get_dbi(tcmu_cmd);
 431                                 page = tcmu_get_block_page(udev, dbi);
 432                                 to = kmap_atomic(page);
 433                         }
 434
 435                         copy_bytes = min_t(size_t, sg_remaining,
 436                                         block_remaining);
 437                         to_offset = get_block_offset_user(udev, dbi,
 438                                         block_remaining);
 439                         offset = DATA_BLOCK_SIZE - block_remaining;
 440                         to = (void *)(unsigned long)to + offset;
 441
 442                         if (*iov_cnt != 0 &&
 443                             to_offset == iov_tail(udev, *iov)) {
 444                                 (*iov)->iov_len += copy_bytes;
 445                         } else {
 446                                 new_iov(iov, iov_cnt, udev);
 447                                 (*iov)->iov_base = (void __user *)to_offset;
 448                                 (*iov)->iov_len = copy_bytes;
 449                         }
 450                         if (copy_data) {
 451                                 memcpy(to, from + sg->length - sg_remaining,
 452                                         copy_bytes);
 453                                 tcmu_flush_dcache_range(to, copy_bytes);
 454                         }
 455                         sg_remaining -= copy_bytes;
 456                         block_remaining -= copy_bytes;
 457                 }
 458                 kunmap_atomic(from - sg->offset);
 459         }
 460         if (to)
 461                 kunmap_atomic(to);
 462
 463         return 0;
 464 }
 465
 466 static void gather_data_area(struct tcmu_dev *udev, struct tcmu_cmd *cmd,
 467                              bool bidi)
 468 {
 469         struct se_cmd *se_cmd = cmd->se_cmd;
 470         int i, dbi;
 471         int block_remaining = 0;
 472         void *from = NULL, *to;
 473         size_t copy_bytes, offset;
 474         struct scatterlist *sg, *data_sg;
 475         struct page *page;
 476         unsigned int data_nents;
 477         uint32_t count = 0;
 478
 479         if (!bidi) {
 480                 data_sg = se_cmd->t_data_sg;
 481                 data_nents = se_cmd->t_data_nents;
 482         } else {
 483
 484                 /*
 485                  * For bidi case, the first count blocks are for Data-Out
 486                  * buffer blocks, and before gathering the Data-In buffer
 487                  * the Data-Out buffer blocks should be discarded.
 488                  */
 489                 count = DIV_ROUND_UP(se_cmd->data_length, DATA_BLOCK_SIZE);
 490
 491                 data_sg = se_cmd->t_bidi_data_sg;
 492                 data_nents = se_cmd->t_bidi_data_nents;
 493         }
 494
 495         tcmu_cmd_set_dbi_cur(cmd, count);
 496
 497         for_each_sg(data_sg, sg, data_nents, i) {
 498                 int sg_remaining = sg->length;
 499                 to = kmap_atomic(sg_page(sg)) + sg->offset;
 500                 while (sg_remaining > 0) {
 501                         if (block_remaining == 0) {
 502                                 if (from)
 503                                         kunmap_atomic(from);
 504
 505                                 block_remaining = DATA_BLOCK_SIZE;
 506                                 dbi = tcmu_cmd_get_dbi(cmd);
 507                                 page = tcmu_get_block_page(udev, dbi);
 508                                 from = kmap_atomic(page);
 509                         }
 510                         copy_bytes = min_t(size_t, sg_remaining,
 511                                         block_remaining);
 512                         offset = DATA_BLOCK_SIZE - block_remaining;
 513                         from = (void *)(unsigned long)from + offset;
 514                         tcmu_flush_dcache_range(from, copy_bytes);
 515                         memcpy(to + sg->length - sg_remaining, from,
 516                                         copy_bytes);
 517
 518                         sg_remaining -= copy_bytes;
 519                         block_remaining -= copy_bytes;
 520                 }
 521                 kunmap_atomic(to - sg->offset);
 522         }
 523         if (from)
 524                 kunmap_atomic(from);
 525 }
 526
 527 static inline size_t spc_bitmap_free(unsigned long *bitmap, uint32_t thresh)
 528 {
 529         return DATA_BLOCK_SIZE * (thresh - bitmap_weight(bitmap, thresh));
 530 }
 531
 532 /*
 533  * We can't queue a command until we have space available on the cmd ring *and*
 534  * space available on the data area.
 535  *
 536  * Called with ring lock held.
 537  */
 538 static bool is_ring_space_avail(struct tcmu_dev *udev, struct tcmu_cmd *cmd,
 539                 size_t cmd_size, size_t data_needed)
 540 {
 541         struct tcmu_mailbox *mb = udev->mb_addr;
 542         uint32_t blocks_needed = (data_needed + DATA_BLOCK_SIZE - 1)
 543                                 / DATA_BLOCK_SIZE;
 544         size_t space, cmd_needed;
 545         u32 cmd_head;
 546
 547         tcmu_flush_dcache_range(mb, sizeof(*mb));
 548
 549         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 550
 551         /*
 552          * If cmd end-of-ring space is too small then we need space for a NOP plus
 553          * original cmd - cmds are internally contiguous.
 554          */
 555         if (head_to_end(cmd_head, udev->cmdr_size) >= cmd_size)
 556                 cmd_needed = cmd_size;
 557         else
 558                 cmd_needed = cmd_size + head_to_end(cmd_head, udev->cmdr_size);
 559
 560         space = spc_free(cmd_head, udev->cmdr_last_cleaned, udev->cmdr_size);
 561         if (space < cmd_needed) {
 562                 pr_debug("no cmd space: %u %u %u\n", cmd_head,
 563                        udev->cmdr_last_cleaned, udev->cmdr_size);
 564                 return false;
 565         }
 566
 567         /* try to check and get the data blocks as needed */
 568         space = spc_bitmap_free(udev->data_bitmap, udev->dbi_thresh);
 569         if (space < data_needed) {
 570                 unsigned long blocks_left = DATA_BLOCK_BITS - udev->dbi_thresh;
 571                 unsigned long grow;
 572
 573                 if (blocks_left < blocks_needed) {
 574                         pr_debug("no data space: only %lu available, but ask for %zu\n",
 575                                         blocks_left * DATA_BLOCK_SIZE,
 576                                         data_needed);
 577                         return false;
 578                 }
 579
 580                 /* Try to expand the thresh */
 581                 if (!udev->dbi_thresh) {
 582                         /* From idle state */
 583                         uint32_t init_thresh = DATA_BLOCK_INIT_BITS;
 584
 585                         udev->dbi_thresh = max(blocks_needed, init_thresh);
 586                 } else {
 587                         /*
 588                          * Grow the data area by max(blocks needed,
 589                          * dbi_thresh / 2), but limited to the max
 590                          * DATA_BLOCK_BITS size.
 591                          */
 592                         grow = max(blocks_needed, udev->dbi_thresh / 2);
 593                         udev->dbi_thresh += grow;
 594                         if (udev->dbi_thresh > DATA_BLOCK_BITS)
 595                                 udev->dbi_thresh = DATA_BLOCK_BITS;
 596                 }
 597         }
 598
 599         if (!tcmu_get_empty_blocks(udev, cmd))
 600                 return false;
 601
 602         return true;
 603 }
 604
 605 static inline size_t tcmu_cmd_get_base_cmd_size(size_t iov_cnt)
 606 {
 607         return max(offsetof(struct tcmu_cmd_entry, req.iov[iov_cnt]),
 608                         sizeof(struct tcmu_cmd_entry));
 609 }
 610
 611 static inline size_t tcmu_cmd_get_cmd_size(struct tcmu_cmd *tcmu_cmd,
 612                                            size_t base_command_size)
 613 {
 614         struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
 615         size_t command_size;
 616
 617         command_size = base_command_size +
 618                 round_up(scsi_command_size(se_cmd->t_task_cdb),
 619                                 TCMU_OP_ALIGN_SIZE);
 620
 621         WARN_ON(command_size & (TCMU_OP_ALIGN_SIZE-1));
 622
 623         return command_size;
 624 }
 625
 626 static sense_reason_t
 627 tcmu_queue_cmd_ring(struct tcmu_cmd *tcmu_cmd)
 628 {
 629         struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
 630         struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
 631         size_t base_command_size, command_size;
 632         struct tcmu_mailbox *mb;
 633         struct tcmu_cmd_entry *entry;
 634         struct iovec *iov;
 635         int iov_cnt, ret;
 636         uint32_t cmd_head;
 637         uint64_t cdb_off;
 638         bool copy_to_data_area;
 639         size_t data_length = tcmu_cmd_get_data_length(tcmu_cmd);
 640
 641         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags))
 642                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
 643
 644         /*
 645          * Must be a certain minimum size for response sense info, but
 646          * also may be larger if the iov array is large.
 647          *
 648          * We prepare as many iovs as possbile for potential uses here,
 649          * because it's expensive to tell how many regions are freed in
 650          * the bitmap & global data pool, as the size calculated here
 651          * will only be used to do the checks.
 652          *
 653          * The size will be recalculated later as actually needed to save
 654          * cmd area memories.
 655          */
 656         base_command_size = tcmu_cmd_get_base_cmd_size(tcmu_cmd->dbi_cnt);
 657         command_size = tcmu_cmd_get_cmd_size(tcmu_cmd, base_command_size);
 658
 659         mutex_lock(&udev->cmdr_lock);
 660
 661         mb = udev->mb_addr;
 662         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 663         if ((command_size > (udev->cmdr_size / 2)) ||
 664             data_length > udev->data_size) {
 665                 pr_warn("TCMU: Request of size %zu/%zu is too big for %u/%zu "
 666                         "cmd ring/data area\n", command_size, data_length,
 667                         udev->cmdr_size, udev->data_size);
 668                 mutex_unlock(&udev->cmdr_lock);
 669                 return TCM_INVALID_CDB_FIELD;
 670         }
 671
 672         while (!is_ring_space_avail(udev, tcmu_cmd, command_size, data_length)) {
 673                 int ret;
 674                 DEFINE_WAIT(__wait);
 675
 676                 prepare_to_wait(&udev->wait_cmdr, &__wait, TASK_INTERRUPTIBLE);
 677
 678                 pr_debug("sleeping for ring space\n");
 679                 mutex_unlock(&udev->cmdr_lock);
 680                 if (udev->cmd_time_out)
 681                         ret = schedule_timeout(
 682                                         msecs_to_jiffies(udev->cmd_time_out));
 683                 else
 684                         ret = schedule_timeout(msecs_to_jiffies(TCMU_TIME_OUT));
 685                 finish_wait(&udev->wait_cmdr, &__wait);
 686                 if (!ret) {
 687                         pr_warn("tcmu: command timed out\n");
 688                         return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
 689                 }
 690
 691                 mutex_lock(&udev->cmdr_lock);
 692
 693                 /* We dropped cmdr_lock, cmd_head is stale */
 694                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 695         }
 696
 697         /* Insert a PAD if end-of-ring space is too small */
 698         if (head_to_end(cmd_head, udev->cmdr_size) < command_size) {
 699                 size_t pad_size = head_to_end(cmd_head, udev->cmdr_size);
 700
 701                 entry = (void *) mb + CMDR_OFF + cmd_head;
 702                 tcmu_flush_dcache_range(entry, sizeof(*entry));
 703                 tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_PAD);
 704                 tcmu_hdr_set_len(&entry->hdr.len_op, pad_size);
 705                 entry->hdr.cmd_id = 0; /* not used for PAD */
 706                 entry->hdr.kflags = 0;
 707                 entry->hdr.uflags = 0;
 708
 709                 UPDATE_HEAD(mb->cmd_head, pad_size, udev->cmdr_size);
 710
 711                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 712                 WARN_ON(cmd_head != 0);
 713         }
 714
 715         entry = (void *) mb + CMDR_OFF + cmd_head;
 716         tcmu_flush_dcache_range(entry, sizeof(*entry));
 717         tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_CMD);
 718         entry->hdr.cmd_id = tcmu_cmd->cmd_id;
 719         entry->hdr.kflags = 0;
 720         entry->hdr.uflags = 0;
 721
 722         /* Handle allocating space from the data area */
 723         tcmu_cmd_reset_dbi_cur(tcmu_cmd);
 724         iov = &entry->req.iov[0];
 725         iov_cnt = 0;
 726         copy_to_data_area = (se_cmd->data_direction == DMA_TO_DEVICE
 727                 || se_cmd->se_cmd_flags & SCF_BIDI);
 728         ret = scatter_data_area(udev, tcmu_cmd, se_cmd->t_data_sg,
 729                                 se_cmd->t_data_nents, &iov, &iov_cnt,
 730                                 copy_to_data_area);
 731         if (ret) {
 732                 tcmu_cmd_free_data(tcmu_cmd, tcmu_cmd->dbi_cnt);
 733                 mutex_unlock(&udev->cmdr_lock);
 734
 735                 pr_err("tcmu: alloc and scatter data failed\n");
 736                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
 737         }
 738         entry->req.iov_cnt = iov_cnt;
 739         entry->req.iov_dif_cnt = 0;
 740
 741         /* Handle BIDI commands */
 742         if (se_cmd->se_cmd_flags & SCF_BIDI) {
 743                 iov_cnt = 0;
 744                 iov++;
 745                 ret = scatter_data_area(udev, tcmu_cmd,
 746                                         se_cmd->t_bidi_data_sg,
 747                                         se_cmd->t_bidi_data_nents,
 748                                         &iov, &iov_cnt, false);
 749                 if (ret) {
 750                         tcmu_cmd_free_data(tcmu_cmd, tcmu_cmd->dbi_cnt);
 751                         mutex_unlock(&udev->cmdr_lock);
 752
 753                         pr_err("tcmu: alloc and scatter bidi data failed\n");
 754                         return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
 755                 }
 756                 entry->req.iov_bidi_cnt = iov_cnt;
 757         }
 758
 759         /*
 760          * Recalaulate the command's base size and size according
 761          * to the actual needs
 762          */
 763         base_command_size = tcmu_cmd_get_base_cmd_size(entry->req.iov_cnt +
 764                                                        entry->req.iov_bidi_cnt);
 765         command_size = tcmu_cmd_get_cmd_size(tcmu_cmd, base_command_size);
 766
 767         tcmu_hdr_set_len(&entry->hdr.len_op, command_size);
 768
 769         /* All offsets relative to mb_addr, not start of entry! */
 770         cdb_off = CMDR_OFF + cmd_head + base_command_size;
 771         memcpy((void *) mb + cdb_off, se_cmd->t_task_cdb, scsi_command_size(se_cmd->t_task_cdb));
 772         entry->req.cdb_off = cdb_off;
 773         tcmu_flush_dcache_range(entry, sizeof(*entry));
 774
 775         UPDATE_HEAD(mb->cmd_head, command_size, udev->cmdr_size);
 776         tcmu_flush_dcache_range(mb, sizeof(*mb));
 777         mutex_unlock(&udev->cmdr_lock);
 778
 779         /* TODO: only if FLUSH and FUA? */
 780         uio_event_notify(&udev->uio_info);
 781
 782         if (udev->cmd_time_out)
 783                 mod_timer(&udev->timeout, round_jiffies_up(jiffies +
 784                           msecs_to_jiffies(udev->cmd_time_out)));
 785
 786         return TCM_NO_SENSE;
 787 }
 788
 789 static sense_reason_t
 790 tcmu_queue_cmd(struct se_cmd *se_cmd)
 791 {
 792         struct se_device *se_dev = se_cmd->se_dev;
 793         struct tcmu_dev *udev = TCMU_DEV(se_dev);
 794         struct tcmu_cmd *tcmu_cmd;
 795         sense_reason_t ret;
 796
 797         tcmu_cmd = tcmu_alloc_cmd(se_cmd);
 798         if (!tcmu_cmd)
 799                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
 800
 801         ret = tcmu_queue_cmd_ring(tcmu_cmd);
 802         if (ret != TCM_NO_SENSE) {
 803                 pr_err("TCMU: Could not queue command\n");
 804                 spin_lock_irq(&udev->commands_lock);
 805                 idr_remove(&udev->commands, tcmu_cmd->cmd_id);
 806                 spin_unlock_irq(&udev->commands_lock);
 807
 808                 tcmu_free_cmd(tcmu_cmd);
 809         }
 810
 811         return ret;
 812 }
 813
 814 static void tcmu_handle_completion(struct tcmu_cmd *cmd, struct tcmu_cmd_entry *entry)
 815 {
 816         struct se_cmd *se_cmd = cmd->se_cmd;
 817         struct tcmu_dev *udev = cmd->tcmu_dev;
 818
 819         /*
 820          * cmd has been completed already from timeout, just reclaim
 821          * data area space and free cmd
 822          */
 823         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
 824                 goto out;
 825
 826         tcmu_cmd_reset_dbi_cur(cmd);
 827
 828         if (entry->hdr.uflags & TCMU_UFLAG_UNKNOWN_OP) {
 829                 pr_warn("TCMU: Userspace set UNKNOWN_OP flag on se_cmd %p\n",
 830                         cmd->se_cmd);
 831                 entry->rsp.scsi_status = SAM_STAT_CHECK_CONDITION;
 832         } else if (entry->rsp.scsi_status == SAM_STAT_CHECK_CONDITION) {
 833                 memcpy(se_cmd->sense_buffer, entry->rsp.sense_buffer,
 834                                se_cmd->scsi_sense_length);
 835         } else if (se_cmd->se_cmd_flags & SCF_BIDI) {
 836                 /* Get Data-In buffer before clean up */
 837                 gather_data_area(udev, cmd, true);
 838         } else if (se_cmd->data_direction == DMA_FROM_DEVICE) {
 839                 gather_data_area(udev, cmd, false);
 840         } else if (se_cmd->data_direction == DMA_TO_DEVICE) {
 841                 /* TODO: */
 842         } else if (se_cmd->data_direction != DMA_NONE) {
 843                 pr_warn("TCMU: data direction was %d!\n",
 844                         se_cmd->data_direction);
 845         }
 846
 847         target_complete_cmd(cmd->se_cmd, entry->rsp.scsi_status);
 848
 849 out:
 850         cmd->se_cmd = NULL;
 851         tcmu_cmd_free_data(cmd, cmd->dbi_cnt);
 852         tcmu_free_cmd(cmd);
 853 }
 854
 855 static unsigned int tcmu_handle_completions(struct tcmu_dev *udev)
 856 {
 857         struct tcmu_mailbox *mb;
 858         int handled = 0;
 859
 860         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags)) {
 861                 pr_err("ring broken, not handling completions\n");
 862                 return 0;
 863         }
 864
 865         mb = udev->mb_addr;
 866         tcmu_flush_dcache_range(mb, sizeof(*mb));
 867
 868         while (udev->cmdr_last_cleaned != ACCESS_ONCE(mb->cmd_tail)) {
 869
 870                 struct tcmu_cmd_entry *entry = (void *) mb + CMDR_OFF + udev->cmdr_last_cleaned;
 871                 struct tcmu_cmd *cmd;
 872
 873                 tcmu_flush_dcache_range(entry, sizeof(*entry));
 874
 875                 if (tcmu_hdr_get_op(entry->hdr.len_op) == TCMU_OP_PAD) {
 876                         UPDATE_HEAD(udev->cmdr_last_cleaned,
 877                                     tcmu_hdr_get_len(entry->hdr.len_op),
 878                                     udev->cmdr_size);
 879                         continue;
 880                 }
 881                 WARN_ON(tcmu_hdr_get_op(entry->hdr.len_op) != TCMU_OP_CMD);
 882
 883                 spin_lock(&udev->commands_lock);
 884                 cmd = idr_remove(&udev->commands, entry->hdr.cmd_id);
 885                 spin_unlock(&udev->commands_lock);
 886
 887                 if (!cmd) {
 888                         pr_err("cmd_id not found, ring is broken\n");
 889                         set_bit(TCMU_DEV_BIT_BROKEN, &udev->flags);
 890                         break;
 891                 }
 892
 893                 tcmu_handle_completion(cmd, entry);
 894
 895                 UPDATE_HEAD(udev->cmdr_last_cleaned,
 896                             tcmu_hdr_get_len(entry->hdr.len_op),
 897                             udev->cmdr_size);
 898
 899                 handled++;
 900         }
 901
 902         if (mb->cmd_tail == mb->cmd_head)
 903                 del_timer(&udev->timeout); /* no more pending cmds */
 904
 905         wake_up(&udev->wait_cmdr);
 906
 907         return handled;
 908 }
 909
 910 static int tcmu_check_expired_cmd(int id, void *p, void *data)
 911 {
 912         struct tcmu_cmd *cmd = p;
 913
 914         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
 915                 return 0;
 916
 917         if (!time_after(jiffies, cmd->deadline))
 918                 return 0;
 919
 920         set_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags);
 921         target_complete_cmd(cmd->se_cmd, SAM_STAT_CHECK_CONDITION);
 922         cmd->se_cmd = NULL;
 923
 924         return 0;
 925 }
 926
 927 static void tcmu_device_timedout(unsigned long data)
 928 {
 929         struct tcmu_dev *udev = (struct tcmu_dev *)data;
 930         unsigned long flags;
 931
 932         spin_lock_irqsave(&udev->commands_lock, flags);
 933         idr_for_each(&udev->commands, tcmu_check_expired_cmd, NULL);
 934         spin_unlock_irqrestore(&udev->commands_lock, flags);
 935
 936         /* Try to wake up the ummap thread */
 937         wake_up(&unmap_wait);
 938
 939         /*
 940          * We don't need to wakeup threads on wait_cmdr since they have their
 941          * own timeout.
 942          */
 943 }
 944
 945 static int tcmu_attach_hba(struct se_hba *hba, u32 host_id)
 946 {
 947         struct tcmu_hba *tcmu_hba;
 948
 949         tcmu_hba = kzalloc(sizeof(struct tcmu_hba), GFP_KERNEL);
 950         if (!tcmu_hba)
 951                 return -ENOMEM;
 952
 953         tcmu_hba->host_id = host_id;
 954         hba->hba_ptr = tcmu_hba;
 955
 956         return 0;
 957 }
 958
 959 static void tcmu_detach_hba(struct se_hba *hba)
 960 {
 961         kfree(hba->hba_ptr);
 962         hba->hba_ptr = NULL;
 963 }
 964
 965 static struct se_device *tcmu_alloc_device(struct se_hba *hba, const char *name)
 966 {
 967         struct tcmu_dev *udev;
 968
 969         udev = kzalloc(sizeof(struct tcmu_dev), GFP_KERNEL);
 970         if (!udev)
 971                 return NULL;
 972         kref_init(&udev->kref);
 973
 974         udev->name = kstrdup(name, GFP_KERNEL);
 975         if (!udev->name) {
 976                 kfree(udev);
 977                 return NULL;
 978         }
 979
 980         udev->hba = hba;
 981         udev->cmd_time_out = TCMU_TIME_OUT;
 982
 983         init_waitqueue_head(&udev->wait_cmdr);
 984         mutex_init(&udev->cmdr_lock);
 985
 986         idr_init(&udev->commands);
 987         spin_lock_init(&udev->commands_lock);
 988
 989         setup_timer(&udev->timeout, tcmu_device_timedout,
 990                 (unsigned long)udev);
 991
 992         return &udev->se_dev;
 993 }
 994
 995 static int tcmu_irqcontrol(struct uio_info *info, s32 irq_on)
 996 {
 997         struct tcmu_dev *tcmu_dev = container_of(info, struct tcmu_dev, uio_info);
 998
 999         mutex_lock(&tcmu_dev->cmdr_lock);
1000         tcmu_handle_completions(tcmu_dev);
1001         mutex_unlock(&tcmu_dev->cmdr_lock);
1002
1003         return 0;
1004 }
1005
1006 /*
1007  * mmap code from uio.c. Copied here because we want to hook mmap()
1008  * and this stuff must come along.
1009  */
1010 static int tcmu_find_mem_index(struct vm_area_struct *vma)
1011 {
1012         struct tcmu_dev *udev = vma->vm_private_data;
1013         struct uio_info *info = &udev->uio_info;
1014
1015         if (vma->vm_pgoff < MAX_UIO_MAPS) {
1016                 if (info->mem[vma->vm_pgoff].size == 0)
1017                         return -1;
1018                 return (int)vma->vm_pgoff;
1019         }
1020         return -1;
1021 }
1022
1023 static struct page *tcmu_try_get_block_page(struct tcmu_dev *udev, uint32_t dbi)
1024 {
1025         struct page *page;
1026         int ret;
1027
1028         mutex_lock(&udev->cmdr_lock);
1029         page = tcmu_get_block_page(udev, dbi);
1030         if (likely(page)) {
1031                 mutex_unlock(&udev->cmdr_lock);
1032                 return page;
1033         }
1034
1035         /*
1036          * Normally it shouldn't be here:
1037          * Only when the userspace has touched the blocks which
1038          * are out of the tcmu_cmd's data iov[], and will return
1039          * one zeroed page.
1040          */
1041         pr_warn("Block(%u) out of cmd's iov[] has been touched!\n", dbi);
1042         pr_warn("Mostly it will be a bug of userspace, please have a check!\n");
1043
1044         if (dbi >= udev->dbi_thresh) {
1045                 /* Extern the udev->dbi_thresh to dbi + 1 */
1046                 udev->dbi_thresh = dbi + 1;
1047                 udev->dbi_max = dbi;
1048         }
1049
1050         page = radix_tree_lookup(&udev->data_blocks, dbi);
1051         if (!page) {
1052                 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1053                 if (!page) {
1054                         mutex_unlock(&udev->cmdr_lock);
1055                         return NULL;
1056                 }
1057
1058                 ret = radix_tree_insert(&udev->data_blocks, dbi, page);
1059                 if (ret) {
1060                         mutex_unlock(&udev->cmdr_lock);
1061                         __free_page(page);
1062                         return NULL;
1063                 }
1064
1065                 /*
1066                  * Since this case is rare in page fault routine, here we
1067                  * will allow the global_db_count >= TCMU_GLOBAL_MAX_BLOCKS
1068                  * to reduce possible page fault call trace.
1069                  */
1070                 atomic_inc(&global_db_count);
1071         }
1072         mutex_unlock(&udev->cmdr_lock);
1073
1074         return page;
1075 }
1076
1077 static int tcmu_vma_fault(struct vm_fault *vmf)
1078 {
1079         struct tcmu_dev *udev = vmf->vma->vm_private_data;
1080         struct uio_info *info = &udev->uio_info;
1081         struct page *page;
1082         unsigned long offset;
1083         void *addr;
1084
1085         int mi = tcmu_find_mem_index(vmf->vma);
1086         if (mi < 0)
1087                 return VM_FAULT_SIGBUS;
1088
1089         /*
1090          * We need to subtract mi because userspace uses offset = N*PAGE_SIZE
1091          * to use mem[N].
1092          */
1093         offset = (vmf->pgoff - mi) << PAGE_SHIFT;
1094
1095         if (offset < udev->data_off) {
1096                 /* For the vmalloc()ed cmd area pages */
1097                 addr = (void *)(unsigned long)info->mem[mi].addr + offset;
1098                 page = vmalloc_to_page(addr);
1099         } else {
1100                 uint32_t dbi;
1101
1102                 /* For the dynamically growing data area pages */
1103                 dbi = (offset - udev->data_off) / DATA_BLOCK_SIZE;
1104                 page = tcmu_try_get_block_page(udev, dbi);
1105                 if (!page)
1106                         return VM_FAULT_NOPAGE;
1107         }
1108
1109         get_page(page);
1110         vmf->page = page;
1111         return 0;
1112 }
1113
1114 static const struct vm_operations_struct tcmu_vm_ops = {
1115         .fault = tcmu_vma_fault,
1116 };
1117
1118 static int tcmu_mmap(struct uio_info *info, struct vm_area_struct *vma)
1119 {
1120         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
1121
1122         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
1123         vma->vm_ops = &tcmu_vm_ops;
1124
1125         vma->vm_private_data = udev;
1126
1127         /* Ensure the mmap is exactly the right size */
1128         if (vma_pages(vma) != (TCMU_RING_SIZE >> PAGE_SHIFT))
1129                 return -EINVAL;
1130
1131         return 0;
1132 }
1133
1134 static int tcmu_open(struct uio_info *info, struct inode *inode)
1135 {
1136         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
1137
1138         /* O_EXCL not supported for char devs, so fake it? */
1139         if (test_and_set_bit(TCMU_DEV_BIT_OPEN, &udev->flags))
1140                 return -EBUSY;
1141
1142         udev->inode = inode;
1143
1144         pr_debug("open\n");
1145
1146         return 0;
1147 }
1148
1149 static void tcmu_dev_call_rcu(struct rcu_head *p)
1150 {
1151         struct se_device *dev = container_of(p, struct se_device, rcu_head);
1152         struct tcmu_dev *udev = TCMU_DEV(dev);
1153
1154         kfree(udev->uio_info.name);
1155         kfree(udev->name);
1156         kfree(udev);
1157 }
1158
1159 static void tcmu_dev_kref_release(struct kref *kref)
1160 {
1161         struct tcmu_dev *udev = container_of(kref, struct tcmu_dev, kref);
1162         struct se_device *dev = &udev->se_dev;
1163
1164         call_rcu(&dev->rcu_head, tcmu_dev_call_rcu);
1165 }
1166
1167 static int tcmu_release(struct uio_info *info, struct inode *inode)
1168 {
1169         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
1170
1171         clear_bit(TCMU_DEV_BIT_OPEN, &udev->flags);
1172
1173         pr_debug("close\n");
1174         /* release ref from configure */
1175         kref_put(&udev->kref, tcmu_dev_kref_release);
1176         return 0;
1177 }
1178
1179 static int tcmu_netlink_event(enum tcmu_genl_cmd cmd, const char *name, int minor)
1180 {
1181         struct sk_buff *skb;
1182         void *msg_header;
1183         int ret = -ENOMEM;
1184
1185         skb = genlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1186         if (!skb)
1187                 return ret;
1188
1189         msg_header = genlmsg_put(skb, 0, 0, &tcmu_genl_family, 0, cmd);
1190         if (!msg_header)
1191                 goto free_skb;
1192
1193         ret = nla_put_string(skb, TCMU_ATTR_DEVICE, name);
1194         if (ret < 0)
1195                 goto free_skb;
1196
1197         ret = nla_put_u32(skb, TCMU_ATTR_MINOR, minor);
1198         if (ret < 0)
1199                 goto free_skb;
1200
1201         genlmsg_end(skb, msg_header);
1202
1203         ret = genlmsg_multicast_allns(&tcmu_genl_family, skb, 0,
1204                                 TCMU_MCGRP_CONFIG, GFP_KERNEL);
1205
1206         /* We don't care if no one is listening */
1207         if (ret == -ESRCH)
1208                 ret = 0;
1209
1210         return ret;
1211 free_skb:
1212         nlmsg_free(skb);
1213         return ret;
1214 }
1215
1216 static int tcmu_configure_device(struct se_device *dev)
1217 {
1218         struct tcmu_dev *udev = TCMU_DEV(dev);
1219         struct tcmu_hba *hba = udev->hba->hba_ptr;
1220         struct uio_info *info;
1221         struct tcmu_mailbox *mb;
1222         size_t size;
1223         size_t used;
1224         int ret = 0;
1225         char *str;
1226
1227         info = &udev->uio_info;
1228
1229         size = snprintf(NULL, 0, "tcm-user/%u/%s/%s", hba->host_id, udev->name,
1230                         udev->dev_config);
1231         size += 1; /* for \0 */
1232         str = kmalloc(size, GFP_KERNEL);
1233         if (!str)
1234                 return -ENOMEM;
1235
1236         used = snprintf(str, size, "tcm-user/%u/%s", hba->host_id, udev->name);
1237
1238         if (udev->dev_config[0])
1239                 snprintf(str + used, size - used, "/%s", udev->dev_config);
1240
1241         info->name = str;
1242
1243         udev->mb_addr = vzalloc(CMDR_SIZE);
1244         if (!udev->mb_addr) {
1245                 ret = -ENOMEM;
1246                 goto err_vzalloc;
1247         }
1248
1249         /* mailbox fits in first part of CMDR space */
1250         udev->cmdr_size = CMDR_SIZE - CMDR_OFF;
1251         udev->data_off = CMDR_SIZE;
1252         udev->data_size = DATA_SIZE;
1253         udev->dbi_thresh = 0; /* Default in Idle state */
1254         udev->waiting_global = false;
1255
1256         /* Initialise the mailbox of the ring buffer */
1257         mb = udev->mb_addr;
1258         mb->version = TCMU_MAILBOX_VERSION;
1259         mb->flags = TCMU_MAILBOX_FLAG_CAP_OOOC;
1260         mb->cmdr_off = CMDR_OFF;
1261         mb->cmdr_size = udev->cmdr_size;
1262
1263         WARN_ON(!PAGE_ALIGNED(udev->data_off));
1264         WARN_ON(udev->data_size % PAGE_SIZE);
1265         WARN_ON(udev->data_size % DATA_BLOCK_SIZE);
1266
1267         INIT_RADIX_TREE(&udev->data_blocks, GFP_KERNEL);
1268
1269         info->version = __stringify(TCMU_MAILBOX_VERSION);
1270
1271         info->mem[0].name = "tcm-user command & data buffer";
1272         info->mem[0].addr = (phys_addr_t)(uintptr_t)udev->mb_addr;
1273         info->mem[0].size = TCMU_RING_SIZE;
1274         info->mem[0].memtype = UIO_MEM_NONE;
1275
1276         info->irqcontrol = tcmu_irqcontrol;
1277         info->irq = UIO_IRQ_CUSTOM;
1278
1279         info->mmap = tcmu_mmap;
1280         info->open = tcmu_open;
1281         info->release = tcmu_release;
1282
1283         ret = uio_register_device(tcmu_root_device, info);
1284         if (ret)
1285                 goto err_register;
1286
1287         /* User can set hw_block_size before enable the device */
1288         if (dev->dev_attrib.hw_block_size == 0)
1289                 dev->dev_attrib.hw_block_size = 512;
1290         /* Other attributes can be configured in userspace */
1291         if (!dev->dev_attrib.hw_max_sectors)
1292                 dev->dev_attrib.hw_max_sectors = 128;
1293         dev->dev_attrib.hw_queue_depth = 128;
1294
1295         /*
1296          * Get a ref incase userspace does a close on the uio device before
1297          * LIO has initiated tcmu_free_device.
1298          */
1299         kref_get(&udev->kref);
1300
1301         ret = tcmu_netlink_event(TCMU_CMD_ADDED_DEVICE, udev->uio_info.name,
1302                                  udev->uio_info.uio_dev->minor);
1303         if (ret)
1304                 goto err_netlink;
1305
1306         mutex_lock(&root_udev_mutex);
1307         list_add(&udev->node, &root_udev);
1308         mutex_unlock(&root_udev_mutex);
1309
1310         return 0;
1311
1312 err_netlink:
1313         kref_put(&udev->kref, tcmu_dev_kref_release);
1314         uio_unregister_device(&udev->uio_info);
1315 err_register:
1316         vfree(udev->mb_addr);
1317 err_vzalloc:
1318         kfree(info->name);
1319         info->name = NULL;
1320
1321         return ret;
1322 }
1323
1324 static int tcmu_check_and_free_pending_cmd(struct tcmu_cmd *cmd)
1325 {
1326         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
1327                 kmem_cache_free(tcmu_cmd_cache, cmd);
1328                 return 0;
1329         }
1330         return -EINVAL;
1331 }
1332
1333 static bool tcmu_dev_configured(struct tcmu_dev *udev)
1334 {
1335         return udev->uio_info.uio_dev ? true : false;
1336 }
1337
1338 static void tcmu_blocks_release(struct tcmu_dev *udev)
1339 {
1340         int i;
1341         struct page *page;
1342
1343         /* Try to release all block pages */
1344         mutex_lock(&udev->cmdr_lock);
1345         for (i = 0; i <= udev->dbi_max; i++) {
1346                 page = radix_tree_delete(&udev->data_blocks, i);
1347                 if (page) {
1348                         __free_page(page);
1349                         atomic_dec(&global_db_count);
1350                 }
1351         }
1352         mutex_unlock(&udev->cmdr_lock);
1353 }
1354
1355 static void tcmu_free_device(struct se_device *dev)
1356 {
1357         struct tcmu_dev *udev = TCMU_DEV(dev);
1358         struct tcmu_cmd *cmd;
1359         bool all_expired = true;
1360         int i;
1361
1362         del_timer_sync(&udev->timeout);
1363
1364         mutex_lock(&root_udev_mutex);
1365         list_del(&udev->node);
1366         mutex_unlock(&root_udev_mutex);
1367
1368         vfree(udev->mb_addr);
1369
1370         /* Upper layer should drain all requests before calling this */
1371         spin_lock_irq(&udev->commands_lock);
1372         idr_for_each_entry(&udev->commands, cmd, i) {
1373                 if (tcmu_check_and_free_pending_cmd(cmd) != 0)
1374                         all_expired = false;
1375         }
1376         idr_destroy(&udev->commands);
1377         spin_unlock_irq(&udev->commands_lock);
1378         WARN_ON(!all_expired);
1379
1380         tcmu_blocks_release(udev);
1381
1382         if (tcmu_dev_configured(udev)) {
1383                 tcmu_netlink_event(TCMU_CMD_REMOVED_DEVICE, udev->uio_info.name,
1384                                    udev->uio_info.uio_dev->minor);
1385
1386                 uio_unregister_device(&udev->uio_info);
1387         }
1388
1389         /* release ref from init */
1390         kref_put(&udev->kref, tcmu_dev_kref_release);
1391 }
1392
1393 enum {
1394         Opt_dev_config, Opt_dev_size, Opt_hw_block_size, Opt_hw_max_sectors,
1395         Opt_err,
1396 };
1397
1398 static match_table_t tokens = {
1399         {Opt_dev_config, "dev_config=%s"},
1400         {Opt_dev_size, "dev_size=%u"},
1401         {Opt_hw_block_size, "hw_block_size=%u"},
1402         {Opt_hw_max_sectors, "hw_max_sectors=%u"},
1403         {Opt_err, NULL}
1404 };
1405
1406 static int tcmu_set_dev_attrib(substring_t *arg, u32 *dev_attrib)
1407 {
1408         unsigned long tmp_ul;
1409         char *arg_p;
1410         int ret;
1411
1412         arg_p = match_strdup(arg);
1413         if (!arg_p)
1414                 return -ENOMEM;
1415
1416         ret = kstrtoul(arg_p, 0, &tmp_ul);
1417         kfree(arg_p);
1418         if (ret < 0) {
1419                 pr_err("kstrtoul() failed for dev attrib\n");
1420                 return ret;
1421         }
1422         if (!tmp_ul) {
1423                 pr_err("dev attrib must be nonzero\n");
1424                 return -EINVAL;
1425         }
1426         *dev_attrib = tmp_ul;
1427         return 0;
1428 }
1429
1430 static ssize_t tcmu_set_configfs_dev_params(struct se_device *dev,
1431                 const char *page, ssize_t count)
1432 {
1433         struct tcmu_dev *udev = TCMU_DEV(dev);
1434         char *orig, *ptr, *opts, *arg_p;
1435         substring_t args[MAX_OPT_ARGS];
1436         int ret = 0, token;
1437
1438         opts = kstrdup(page, GFP_KERNEL);
1439         if (!opts)
1440                 return -ENOMEM;
1441
1442         orig = opts;
1443
1444         while ((ptr = strsep(&opts, ",\n")) != NULL) {
1445                 if (!*ptr)
1446                         continue;
1447
1448                 token = match_token(ptr, tokens, args);
1449                 switch (token) {
1450                 case Opt_dev_config:
1451                         if (match_strlcpy(udev->dev_config, &args[0],
1452                                           TCMU_CONFIG_LEN) == 0) {
1453                                 ret = -EINVAL;
1454                                 break;
1455                         }
1456                         pr_debug("TCMU: Referencing Path: %s\n", udev->dev_config);
1457                         break;
1458                 case Opt_dev_size:
1459                         arg_p = match_strdup(&args[0]);
1460                         if (!arg_p) {
1461                                 ret = -ENOMEM;
1462                                 break;
1463                         }
1464                         ret = kstrtoul(arg_p, 0, (unsigned long *) &udev->dev_size);
1465                         kfree(arg_p);
1466                         if (ret < 0)
1467                                 pr_err("kstrtoul() failed for dev_size=\n");
1468                         break;
1469                 case Opt_hw_block_size:
1470                         ret = tcmu_set_dev_attrib(&args[0],
1471                                         &(dev->dev_attrib.hw_block_size));
1472                         break;
1473                 case Opt_hw_max_sectors:
1474                         ret = tcmu_set_dev_attrib(&args[0],
1475                                         &(dev->dev_attrib.hw_max_sectors));
1476                         break;
1477                 default:
1478                         break;
1479                 }
1480
1481                 if (ret)
1482                         break;
1483         }
1484
1485         kfree(orig);
1486         return (!ret) ? count : ret;
1487 }
1488
1489 static ssize_t tcmu_show_configfs_dev_params(struct se_device *dev, char *b)
1490 {
1491         struct tcmu_dev *udev = TCMU_DEV(dev);
1492         ssize_t bl = 0;
1493
1494         bl = sprintf(b + bl, "Config: %s ",
1495                      udev->dev_config[0] ? udev->dev_config : "NULL");
1496         bl += sprintf(b + bl, "Size: %zu\n", udev->dev_size);
1497
1498         return bl;
1499 }
1500
1501 static sector_t tcmu_get_blocks(struct se_device *dev)
1502 {
1503         struct tcmu_dev *udev = TCMU_DEV(dev);
1504
1505         return div_u64(udev->dev_size - dev->dev_attrib.block_size,
1506                        dev->dev_attrib.block_size);
1507 }
1508
1509 static sense_reason_t
1510 tcmu_parse_cdb(struct se_cmd *cmd)
1511 {
1512         return passthrough_parse_cdb(cmd, tcmu_queue_cmd);
1513 }
1514
1515 static ssize_t tcmu_cmd_time_out_show(struct config_item *item, char *page)
1516 {
1517         struct se_dev_attrib *da = container_of(to_config_group(item),
1518                                         struct se_dev_attrib, da_group);
1519         struct tcmu_dev *udev = container_of(da->da_dev,
1520                                         struct tcmu_dev, se_dev);
1521
1522         return snprintf(page, PAGE_SIZE, "%lu\n", udev->cmd_time_out / MSEC_PER_SEC);
1523 }
1524
1525 static ssize_t tcmu_cmd_time_out_store(struct config_item *item, const char *page,
1526                                        size_t count)
1527 {
1528         struct se_dev_attrib *da = container_of(to_config_group(item),
1529                                         struct se_dev_attrib, da_group);
1530         struct tcmu_dev *udev = container_of(da->da_dev,
1531                                         struct tcmu_dev, se_dev);
1532         u32 val;
1533         int ret;
1534
1535         if (da->da_dev->export_count) {
1536                 pr_err("Unable to set tcmu cmd_time_out while exports exist\n");
1537                 return -EINVAL;
1538         }
1539
1540         ret = kstrtou32(page, 0, &val);
1541         if (ret < 0)
1542                 return ret;
1543
1544         udev->cmd_time_out = val * MSEC_PER_SEC;
1545         return count;
1546 }
1547 CONFIGFS_ATTR(tcmu_, cmd_time_out);
1548
1549 static struct configfs_attribute **tcmu_attrs;
1550
1551 static struct target_backend_ops tcmu_ops = {
1552         .name                   = "user",
1553         .owner                  = THIS_MODULE,
1554         .transport_flags        = TRANSPORT_FLAG_PASSTHROUGH,
1555         .attach_hba             = tcmu_attach_hba,
1556         .detach_hba             = tcmu_detach_hba,
1557         .alloc_device           = tcmu_alloc_device,
1558         .configure_device       = tcmu_configure_device,
1559         .free_device            = tcmu_free_device,
1560         .parse_cdb              = tcmu_parse_cdb,
1561         .set_configfs_dev_params = tcmu_set_configfs_dev_params,
1562         .show_configfs_dev_params = tcmu_show_configfs_dev_params,
1563         .get_device_type        = sbc_get_device_type,
1564         .get_blocks             = tcmu_get_blocks,
1565         .tb_dev_attrib_attrs    = NULL,
1566 };
1567
1568 static int unmap_thread_fn(void *data)
1569 {
1570         struct tcmu_dev *udev;
1571         loff_t off;
1572         uint32_t start, end, block;
1573         struct page *page;
1574         int i;
1575
1576         while (1) {
1577                 DEFINE_WAIT(__wait);
1578
1579                 prepare_to_wait(&unmap_wait, &__wait, TASK_INTERRUPTIBLE);
1580                 schedule();
1581                 finish_wait(&unmap_wait, &__wait);
1582
1583                 if (kthread_should_stop())
1584                         break;
1585
1586                 mutex_lock(&root_udev_mutex);
1587                 list_for_each_entry(udev, &root_udev, node) {
1588                         mutex_lock(&udev->cmdr_lock);
1589
1590                         /* Try to complete the finished commands first */
1591                         tcmu_handle_completions(udev);
1592
1593                         /* Skip the udevs waiting the global pool or in idle */
1594                         if (udev->waiting_global || !udev->dbi_thresh) {
1595                                 mutex_unlock(&udev->cmdr_lock);
1596                                 continue;
1597                         }
1598
1599                         end = udev->dbi_max + 1;
1600                         block = find_last_bit(udev->data_bitmap, end);
1601                         if (block == udev->dbi_max) {
1602                                 /*
1603                                  * The last bit is dbi_max, so there is
1604                                  * no need to shrink any blocks.
1605                                  */
1606                                 mutex_unlock(&udev->cmdr_lock);
1607                                 continue;
1608                         } else if (block == end) {
1609                                 /* The current udev will goto idle state */
1610                                 udev->dbi_thresh = start = 0;
1611                                 udev->dbi_max = 0;
1612                         } else {
1613                                 udev->dbi_thresh = start = block + 1;
1614                                 udev->dbi_max = block;
1615                         }
1616
1617                         /* Here will truncate the data area from off */
1618                         off = udev->data_off + start * DATA_BLOCK_SIZE;
1619                         unmap_mapping_range(udev->inode->i_mapping, off, 0, 1);
1620
1621                         /* Release the block pages */
1622                         for (i = start; i < end; i++) {
1623                                 page = radix_tree_delete(&udev->data_blocks, i);
1624                                 if (page) {
1625                                         __free_page(page);
1626                                         atomic_dec(&global_db_count);
1627                                 }
1628                         }
1629                         mutex_unlock(&udev->cmdr_lock);
1630                 }
1631
1632                 /*
1633                  * Try to wake up the udevs who are waiting
1634                  * for the global data pool.
1635                  */
1636                 list_for_each_entry(udev, &root_udev, node) {
1637                         if (udev->waiting_global)
1638                                 wake_up(&udev->wait_cmdr);
1639                 }
1640                 mutex_unlock(&root_udev_mutex);
1641         }
1642
1643         return 0;
1644 }
1645
1646 static int __init tcmu_module_init(void)
1647 {
1648         int ret, i, len = 0;
1649
1650         BUILD_BUG_ON((sizeof(struct tcmu_cmd_entry) % TCMU_OP_ALIGN_SIZE) != 0);
1651
1652         tcmu_cmd_cache = kmem_cache_create("tcmu_cmd_cache",
1653                                 sizeof(struct tcmu_cmd),
1654                                 __alignof__(struct tcmu_cmd),
1655                                 0, NULL);
1656         if (!tcmu_cmd_cache)
1657                 return -ENOMEM;
1658
1659         tcmu_root_device = root_device_register("tcm_user");
1660         if (IS_ERR(tcmu_root_device)) {
1661                 ret = PTR_ERR(tcmu_root_device);
1662                 goto out_free_cache;
1663         }
1664
1665         ret = genl_register_family(&tcmu_genl_family);
1666         if (ret < 0) {
1667                 goto out_unreg_device;
1668         }
1669
1670         for (i = 0; passthrough_attrib_attrs[i] != NULL; i++) {
1671                 len += sizeof(struct configfs_attribute *);
1672         }
1673         len += sizeof(struct configfs_attribute *) * 2;
1674
1675         tcmu_attrs = kzalloc(len, GFP_KERNEL);
1676         if (!tcmu_attrs) {
1677                 ret = -ENOMEM;
1678                 goto out_unreg_genl;
1679         }
1680
1681         for (i = 0; passthrough_attrib_attrs[i] != NULL; i++) {
1682                 tcmu_attrs[i] = passthrough_attrib_attrs[i];
1683         }
1684         tcmu_attrs[i] = &tcmu_attr_cmd_time_out;
1685         tcmu_ops.tb_dev_attrib_attrs = tcmu_attrs;
1686
1687         ret = transport_backend_register(&tcmu_ops);
1688         if (ret)
1689                 goto out_attrs;
1690
1691         init_waitqueue_head(&unmap_wait);
1692         unmap_thread = kthread_run(unmap_thread_fn, NULL, "tcmu_unmap");
1693         if (IS_ERR(unmap_thread)) {
1694                 ret = PTR_ERR(unmap_thread);
1695                 goto out_unreg_transport;
1696         }
1697
1698         return 0;
1699
1700 out_unreg_transport:
1701         target_backend_unregister(&tcmu_ops);
1702 out_attrs:
1703         kfree(tcmu_attrs);
1704 out_unreg_genl:
1705         genl_unregister_family(&tcmu_genl_family);
1706 out_unreg_device:
1707         root_device_unregister(tcmu_root_device);
1708 out_free_cache:
1709         kmem_cache_destroy(tcmu_cmd_cache);
1710
1711         return ret;
1712 }
1713
1714 static void __exit tcmu_module_exit(void)
1715 {
1716         kthread_stop(unmap_thread);
1717         target_backend_unregister(&tcmu_ops);
1718         kfree(tcmu_attrs);
1719         genl_unregister_family(&tcmu_genl_family);
1720         root_device_unregister(tcmu_root_device);
1721         kmem_cache_destroy(tcmu_cmd_cache);
1722 }
1723
1724 MODULE_DESCRIPTION("TCM USER subsystem plugin");
1725 MODULE_AUTHOR("Shaohua Li <shli@kernel.org>");
1726 MODULE_AUTHOR("Andy Grover <agrover@redhat.com>");
1727 MODULE_LICENSE("GPL");
1728
1729 module_init(tcmu_module_init);
1730 module_exit(tcmu_module_exit);