drivers/target/target_core_user.c

   1 /*
   2  * Copyright (C) 2013 Shaohua Li <shli@kernel.org>
   3  * Copyright (C) 2014 Red Hat, Inc.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along with
  15  * this program; if not, write to the Free Software Foundation, Inc.,
  16  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  17  */
  18
  19 #include <linux/spinlock.h>
  20 #include <linux/module.h>
  21 #include <linux/idr.h>
  22 #include <linux/timer.h>
  23 #include <linux/parser.h>
  24 #include <scsi/scsi.h>
  25 #include <scsi/scsi_host.h>
  26 #include <linux/uio_driver.h>
  27 #include <net/genetlink.h>
  28 #include <target/target_core_base.h>
  29 #include <target/target_core_fabric.h>
  30 #include <target/target_core_backend.h>
  31 #include <linux/target_core_user.h>
  32
  33 /*
  34  * Define a shared-memory interface for LIO to pass SCSI commands and
  35  * data to userspace for processing. This is to allow backends that
  36  * are too complex for in-kernel support to be possible.
  37  *
  38  * It uses the UIO framework to do a lot of the device-creation and
  39  * introspection work for us.
  40  *
  41  * See the .h file for how the ring is laid out. Note that while the
  42  * command ring is defined, the particulars of the data area are
  43  * not. Offset values in the command entry point to other locations
  44  * internal to the mmap()ed area. There is separate space outside the
  45  * command ring for data buffers. This leaves maximum flexibility for
  46  * moving buffer allocations, or even page flipping or other
  47  * allocation techniques, without altering the command ring layout.
  48  *
  49  * SECURITY:
  50  * The user process must be assumed to be malicious. There's no way to
  51  * prevent it breaking the command ring protocol if it wants, but in
  52  * order to prevent other issues we must only ever read *data* from
  53  * the shared memory area, not offsets or sizes. This applies to
  54  * command ring entries as well as the mailbox. Extra code needed for
  55  * this may have a 'UAM' comment.
  56  */
  57
  58
  59 #define TCMU_TIME_OUT (30 * MSEC_PER_SEC)
  60
  61 #define CMDR_SIZE (16 * 4096)
  62 #define DATA_SIZE (257 * 4096)
  63
  64 #define TCMU_RING_SIZE (CMDR_SIZE + DATA_SIZE)
  65
  66 static struct device *tcmu_root_device;
  67
  68 struct tcmu_hba {
  69         u32 host_id;
  70 };
  71
  72 /* User wants all cmds or just some */
  73 enum passthru_level {
  74         TCMU_PASS_ALL = 0,
  75         TCMU_PASS_IO,
  76         TCMU_PASS_INVALID,
  77 };
  78
  79 #define TCMU_CONFIG_LEN 256
  80
  81 struct tcmu_dev {
  82         struct se_device se_dev;
  83
  84         char *name;
  85         struct se_hba *hba;
  86
  87 #define TCMU_DEV_BIT_OPEN 0
  88 #define TCMU_DEV_BIT_BROKEN 1
  89         unsigned long flags;
  90         enum passthru_level pass_level;
  91
  92         struct uio_info uio_info;
  93
  94         struct tcmu_mailbox *mb_addr;
  95         size_t dev_size;
  96         u32 cmdr_size;
  97         u32 cmdr_last_cleaned;
  98         /* Offset of data ring from start of mb */
  99         size_t data_off;
 100         size_t data_size;
 101         /* Ring head + tail values. */
 102         /* Must add data_off and mb_addr to get the address */
 103         size_t data_head;
 104         size_t data_tail;
 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->data_length when cleaning up expired cmds, because if
 129            cmd has been completed then accessing se_cmd is off limits */
 130         size_t data_length;
 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 };
 159
 160 static struct tcmu_cmd *tcmu_alloc_cmd(struct se_cmd *se_cmd)
 161 {
 162         struct se_device *se_dev = se_cmd->se_dev;
 163         struct tcmu_dev *udev = TCMU_DEV(se_dev);
 164         struct tcmu_cmd *tcmu_cmd;
 165         int cmd_id;
 166
 167         tcmu_cmd = kmem_cache_zalloc(tcmu_cmd_cache, GFP_KERNEL);
 168         if (!tcmu_cmd)
 169                 return NULL;
 170
 171         tcmu_cmd->se_cmd = se_cmd;
 172         tcmu_cmd->tcmu_dev = udev;
 173         tcmu_cmd->data_length = se_cmd->data_length;
 174
 175         tcmu_cmd->deadline = jiffies + msecs_to_jiffies(TCMU_TIME_OUT);
 176
 177         idr_preload(GFP_KERNEL);
 178         spin_lock_irq(&udev->commands_lock);
 179         cmd_id = idr_alloc(&udev->commands, tcmu_cmd, 0,
 180                 USHRT_MAX, GFP_NOWAIT);
 181         spin_unlock_irq(&udev->commands_lock);
 182         idr_preload_end();
 183
 184         if (cmd_id < 0) {
 185                 kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
 186                 return NULL;
 187         }
 188         tcmu_cmd->cmd_id = cmd_id;
 189
 190         return tcmu_cmd;
 191 }
 192
 193 static inline void tcmu_flush_dcache_range(void *vaddr, size_t size)
 194 {
 195         unsigned long offset = (unsigned long) vaddr & ~PAGE_MASK;
 196
 197         size = round_up(size+offset, PAGE_SIZE);
 198         vaddr -= offset;
 199
 200         while (size) {
 201                 flush_dcache_page(virt_to_page(vaddr));
 202                 size -= PAGE_SIZE;
 203         }
 204 }
 205
 206 /*
 207  * Some ring helper functions. We don't assume size is a power of 2 so
 208  * we can't use circ_buf.h.
 209  */
 210 static inline size_t spc_used(size_t head, size_t tail, size_t size)
 211 {
 212         int diff = head - tail;
 213
 214         if (diff >= 0)
 215                 return diff;
 216         else
 217                 return size + diff;
 218 }
 219
 220 static inline size_t spc_free(size_t head, size_t tail, size_t size)
 221 {
 222         /* Keep 1 byte unused or we can't tell full from empty */
 223         return (size - spc_used(head, tail, size) - 1);
 224 }
 225
 226 static inline size_t head_to_end(size_t head, size_t size)
 227 {
 228         return size - head;
 229 }
 230
 231 #define UPDATE_HEAD(head, used, size) smp_store_release(&head, ((head % size) + used) % size)
 232
 233 /*
 234  * We can't queue a command until we have space available on the cmd ring *and* space
 235  * space avail on the data ring.
 236  *
 237  * Called with ring lock held.
 238  */
 239 static bool is_ring_space_avail(struct tcmu_dev *udev, size_t cmd_needed, size_t data_needed)
 240 {
 241         struct tcmu_mailbox *mb = udev->mb_addr;
 242         size_t space;
 243         u32 cmd_head;
 244
 245         tcmu_flush_dcache_range(mb, sizeof(*mb));
 246
 247         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 248
 249         space = spc_free(cmd_head, udev->cmdr_last_cleaned, udev->cmdr_size);
 250         if (space < cmd_needed) {
 251                 pr_debug("no cmd space: %u %u %u\n", cmd_head,
 252                        udev->cmdr_last_cleaned, udev->cmdr_size);
 253                 return false;
 254         }
 255
 256         space = spc_free(udev->data_head, udev->data_tail, udev->data_size);
 257         if (space < data_needed) {
 258                 pr_debug("no data space: %zu %zu %zu\n", udev->data_head,
 259                        udev->data_tail, udev->data_size);
 260                 return false;
 261         }
 262
 263         return true;
 264 }
 265
 266 static int tcmu_queue_cmd_ring(struct tcmu_cmd *tcmu_cmd)
 267 {
 268         struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
 269         struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
 270         size_t base_command_size, command_size;
 271         size_t cmdr_space_needed;
 272         struct tcmu_mailbox *mb;
 273         size_t pad_size;
 274         struct tcmu_cmd_entry *entry;
 275         int i;
 276         struct scatterlist *sg;
 277         struct iovec *iov;
 278         int iov_cnt = 0;
 279         uint32_t cmd_head;
 280         uint64_t cdb_off;
 281
 282         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags))
 283                 return -EINVAL;
 284
 285         /*
 286          * Must be a certain minimum size for response sense info, but
 287          * also may be larger if the iov array is large.
 288          *
 289          * iovs = sgl_nents+1, for end-of-ring case, plus another 1
 290          * b/c size == offsetof one-past-element.
 291         */
 292         base_command_size = max(offsetof(struct tcmu_cmd_entry,
 293                                          req.iov[se_cmd->t_data_nents + 2]),
 294                                 sizeof(struct tcmu_cmd_entry));
 295         command_size = base_command_size
 296                 + round_up(scsi_command_size(se_cmd->t_task_cdb), TCMU_OP_ALIGN_SIZE);
 297
 298         WARN_ON(command_size & (TCMU_OP_ALIGN_SIZE-1));
 299
 300         spin_lock_irq(&udev->cmdr_lock);
 301
 302         mb = udev->mb_addr;
 303         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 304         if ((command_size > (udev->cmdr_size / 2))
 305             || tcmu_cmd->data_length > (udev->data_size - 1))
 306                 pr_warn("TCMU: Request of size %zu/%zu may be too big for %u/%zu "
 307                         "cmd/data ring buffers\n", command_size, tcmu_cmd->data_length,
 308                         udev->cmdr_size, udev->data_size);
 309
 310         /*
 311          * Cmd end-of-ring space is too small so we need space for a NOP plus
 312          * original cmd - cmds are internally contiguous.
 313          */
 314         if (head_to_end(cmd_head, udev->cmdr_size) >= command_size)
 315                 pad_size = 0;
 316         else
 317                 pad_size = head_to_end(cmd_head, udev->cmdr_size);
 318         cmdr_space_needed = command_size + pad_size;
 319
 320         while (!is_ring_space_avail(udev, cmdr_space_needed, tcmu_cmd->data_length)) {
 321                 int ret;
 322                 DEFINE_WAIT(__wait);
 323
 324                 prepare_to_wait(&udev->wait_cmdr, &__wait, TASK_INTERRUPTIBLE);
 325
 326                 pr_debug("sleeping for ring space\n");
 327                 spin_unlock_irq(&udev->cmdr_lock);
 328                 ret = schedule_timeout(msecs_to_jiffies(TCMU_TIME_OUT));
 329                 finish_wait(&udev->wait_cmdr, &__wait);
 330                 if (!ret) {
 331                         pr_warn("tcmu: command timed out\n");
 332                         return -ETIMEDOUT;
 333                 }
 334
 335                 spin_lock_irq(&udev->cmdr_lock);
 336
 337                 /* We dropped cmdr_lock, cmd_head is stale */
 338                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 339         }
 340
 341         if (pad_size) {
 342                 entry = (void *) mb + CMDR_OFF + cmd_head;
 343                 tcmu_flush_dcache_range(entry, sizeof(*entry));
 344                 tcmu_hdr_set_op(&entry->hdr, TCMU_OP_PAD);
 345                 tcmu_hdr_set_len(&entry->hdr, pad_size);
 346
 347                 UPDATE_HEAD(mb->cmd_head, pad_size, udev->cmdr_size);
 348
 349                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
 350                 WARN_ON(cmd_head != 0);
 351         }
 352
 353         entry = (void *) mb + CMDR_OFF + cmd_head;
 354         tcmu_flush_dcache_range(entry, sizeof(*entry));
 355         tcmu_hdr_set_op(&entry->hdr, TCMU_OP_CMD);
 356         tcmu_hdr_set_len(&entry->hdr, command_size);
 357         entry->cmd_id = tcmu_cmd->cmd_id;
 358
 359         /*
 360          * Fix up iovecs, and handle if allocation in data ring wrapped.
 361          */
 362         iov = &entry->req.iov[0];
 363         for_each_sg(se_cmd->t_data_sg, sg, se_cmd->t_data_nents, i) {
 364                 size_t copy_bytes = min((size_t)sg->length,
 365                                      head_to_end(udev->data_head, udev->data_size));
 366                 void *from = kmap_atomic(sg_page(sg)) + sg->offset;
 367                 void *to = (void *) mb + udev->data_off + udev->data_head;
 368
 369                 if (tcmu_cmd->se_cmd->data_direction == DMA_TO_DEVICE) {
 370                         memcpy(to, from, copy_bytes);
 371                         tcmu_flush_dcache_range(to, copy_bytes);
 372                 }
 373
 374                 /* Even iov_base is relative to mb_addr */
 375                 iov->iov_len = copy_bytes;
 376                 iov->iov_base = (void *) udev->data_off + udev->data_head;
 377                 iov_cnt++;
 378                 iov++;
 379
 380                 UPDATE_HEAD(udev->data_head, copy_bytes, udev->data_size);
 381
 382                 /* Uh oh, we wrapped the buffer. Must split sg across 2 iovs. */
 383                 if (sg->length != copy_bytes) {
 384                         from += copy_bytes;
 385                         copy_bytes = sg->length - copy_bytes;
 386
 387                         iov->iov_len = copy_bytes;
 388                         iov->iov_base = (void *) udev->data_off + udev->data_head;
 389
 390                         if (se_cmd->data_direction == DMA_TO_DEVICE) {
 391                                 to = (void *) mb + udev->data_off + udev->data_head;
 392                                 memcpy(to, from, copy_bytes);
 393                                 tcmu_flush_dcache_range(to, copy_bytes);
 394                         }
 395
 396                         iov_cnt++;
 397                         iov++;
 398
 399                         UPDATE_HEAD(udev->data_head, copy_bytes, udev->data_size);
 400                 }
 401
 402                 kunmap_atomic(from);
 403         }
 404         entry->req.iov_cnt = iov_cnt;
 405
 406         /* All offsets relative to mb_addr, not start of entry! */
 407         cdb_off = CMDR_OFF + cmd_head + base_command_size;
 408         memcpy((void *) mb + cdb_off, se_cmd->t_task_cdb, scsi_command_size(se_cmd->t_task_cdb));
 409         entry->req.cdb_off = cdb_off;
 410         tcmu_flush_dcache_range(entry, sizeof(*entry));
 411
 412         UPDATE_HEAD(mb->cmd_head, command_size, udev->cmdr_size);
 413         tcmu_flush_dcache_range(mb, sizeof(*mb));
 414
 415         spin_unlock_irq(&udev->cmdr_lock);
 416
 417         /* TODO: only if FLUSH and FUA? */
 418         uio_event_notify(&udev->uio_info);
 419
 420         mod_timer(&udev->timeout,
 421                 round_jiffies_up(jiffies + msecs_to_jiffies(TCMU_TIME_OUT)));
 422
 423         return 0;
 424 }
 425
 426 static int tcmu_queue_cmd(struct se_cmd *se_cmd)
 427 {
 428         struct se_device *se_dev = se_cmd->se_dev;
 429         struct tcmu_dev *udev = TCMU_DEV(se_dev);
 430         struct tcmu_cmd *tcmu_cmd;
 431         int ret;
 432
 433         tcmu_cmd = tcmu_alloc_cmd(se_cmd);
 434         if (!tcmu_cmd)
 435                 return -ENOMEM;
 436
 437         ret = tcmu_queue_cmd_ring(tcmu_cmd);
 438         if (ret < 0) {
 439                 pr_err("TCMU: Could not queue command\n");
 440                 spin_lock_irq(&udev->commands_lock);
 441                 idr_remove(&udev->commands, tcmu_cmd->cmd_id);
 442                 spin_unlock_irq(&udev->commands_lock);
 443
 444                 kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
 445         }
 446
 447         return ret;
 448 }
 449
 450 static void tcmu_handle_completion(struct tcmu_cmd *cmd, struct tcmu_cmd_entry *entry)
 451 {
 452         struct se_cmd *se_cmd = cmd->se_cmd;
 453         struct tcmu_dev *udev = cmd->tcmu_dev;
 454
 455         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
 456                 /* cmd has been completed already from timeout, just reclaim data
 457                    ring space */
 458                 UPDATE_HEAD(udev->data_tail, cmd->data_length, udev->data_size);
 459                 return;
 460         }
 461
 462         if (entry->rsp.scsi_status == SAM_STAT_CHECK_CONDITION) {
 463                 memcpy(se_cmd->sense_buffer, entry->rsp.sense_buffer,
 464                                se_cmd->scsi_sense_length);
 465
 466                 UPDATE_HEAD(udev->data_tail, cmd->data_length, udev->data_size);
 467         }
 468         else if (se_cmd->data_direction == DMA_FROM_DEVICE) {
 469                 struct scatterlist *sg;
 470                 int i;
 471
 472                 /* It'd be easier to look at entry's iovec again, but UAM */
 473                 for_each_sg(se_cmd->t_data_sg, sg, se_cmd->t_data_nents, i) {
 474                         size_t copy_bytes;
 475                         void *to;
 476                         void *from;
 477
 478                         copy_bytes = min((size_t)sg->length,
 479                                          head_to_end(udev->data_tail, udev->data_size));
 480
 481                         to = kmap_atomic(sg_page(sg)) + sg->offset;
 482                         WARN_ON(sg->length + sg->offset > PAGE_SIZE);
 483                         from = (void *) udev->mb_addr + udev->data_off + udev->data_tail;
 484                         tcmu_flush_dcache_range(from, copy_bytes);
 485                         memcpy(to, from, copy_bytes);
 486
 487                         UPDATE_HEAD(udev->data_tail, copy_bytes, udev->data_size);
 488
 489                         /* Uh oh, wrapped the data buffer for this sg's data */
 490                         if (sg->length != copy_bytes) {
 491                                 from = (void *) udev->mb_addr + udev->data_off + udev->data_tail;
 492                                 WARN_ON(udev->data_tail);
 493                                 to += copy_bytes;
 494                                 copy_bytes = sg->length - copy_bytes;
 495                                 tcmu_flush_dcache_range(from, copy_bytes);
 496                                 memcpy(to, from, copy_bytes);
 497
 498                                 UPDATE_HEAD(udev->data_tail, copy_bytes, udev->data_size);
 499                         }
 500
 501                         kunmap_atomic(to);
 502                 }
 503
 504         } else if (se_cmd->data_direction == DMA_TO_DEVICE) {
 505                 UPDATE_HEAD(udev->data_tail, cmd->data_length, udev->data_size);
 506         } else {
 507                 pr_warn("TCMU: data direction was %d!\n", se_cmd->data_direction);
 508         }
 509
 510         target_complete_cmd(cmd->se_cmd, entry->rsp.scsi_status);
 511         cmd->se_cmd = NULL;
 512
 513         kmem_cache_free(tcmu_cmd_cache, cmd);
 514 }
 515
 516 static unsigned int tcmu_handle_completions(struct tcmu_dev *udev)
 517 {
 518         struct tcmu_mailbox *mb;
 519         LIST_HEAD(cpl_cmds);
 520         unsigned long flags;
 521         int handled = 0;
 522
 523         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags)) {
 524                 pr_err("ring broken, not handling completions\n");
 525                 return 0;
 526         }
 527
 528         spin_lock_irqsave(&udev->cmdr_lock, flags);
 529
 530         mb = udev->mb_addr;
 531         tcmu_flush_dcache_range(mb, sizeof(*mb));
 532
 533         while (udev->cmdr_last_cleaned != ACCESS_ONCE(mb->cmd_tail)) {
 534
 535                 struct tcmu_cmd_entry *entry = (void *) mb + CMDR_OFF + udev->cmdr_last_cleaned;
 536                 struct tcmu_cmd *cmd;
 537
 538                 tcmu_flush_dcache_range(entry, sizeof(*entry));
 539
 540                 if (tcmu_hdr_get_op(&entry->hdr) == TCMU_OP_PAD) {
 541                         UPDATE_HEAD(udev->cmdr_last_cleaned, tcmu_hdr_get_len(&entry->hdr), udev->cmdr_size);
 542                         continue;
 543                 }
 544                 WARN_ON(tcmu_hdr_get_op(&entry->hdr) != TCMU_OP_CMD);
 545
 546                 spin_lock(&udev->commands_lock);
 547                 cmd = idr_find(&udev->commands, entry->cmd_id);
 548                 if (cmd)
 549                         idr_remove(&udev->commands, cmd->cmd_id);
 550                 spin_unlock(&udev->commands_lock);
 551
 552                 if (!cmd) {
 553                         pr_err("cmd_id not found, ring is broken\n");
 554                         set_bit(TCMU_DEV_BIT_BROKEN, &udev->flags);
 555                         break;
 556                 }
 557
 558                 tcmu_handle_completion(cmd, entry);
 559
 560                 UPDATE_HEAD(udev->cmdr_last_cleaned, tcmu_hdr_get_len(&entry->hdr), udev->cmdr_size);
 561
 562                 handled++;
 563         }
 564
 565         if (mb->cmd_tail == mb->cmd_head)
 566                 del_timer(&udev->timeout); /* no more pending cmds */
 567
 568         spin_unlock_irqrestore(&udev->cmdr_lock, flags);
 569
 570         wake_up(&udev->wait_cmdr);
 571
 572         return handled;
 573 }
 574
 575 static int tcmu_check_expired_cmd(int id, void *p, void *data)
 576 {
 577         struct tcmu_cmd *cmd = p;
 578
 579         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
 580                 return 0;
 581
 582         if (!time_after(cmd->deadline, jiffies))
 583                 return 0;
 584
 585         set_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags);
 586         target_complete_cmd(cmd->se_cmd, SAM_STAT_CHECK_CONDITION);
 587         cmd->se_cmd = NULL;
 588
 589         kmem_cache_free(tcmu_cmd_cache, cmd);
 590
 591         return 0;
 592 }
 593
 594 static void tcmu_device_timedout(unsigned long data)
 595 {
 596         struct tcmu_dev *udev = (struct tcmu_dev *)data;
 597         unsigned long flags;
 598         int handled;
 599
 600         handled = tcmu_handle_completions(udev);
 601
 602         pr_warn("%d completions handled from timeout\n", handled);
 603
 604         spin_lock_irqsave(&udev->commands_lock, flags);
 605         idr_for_each(&udev->commands, tcmu_check_expired_cmd, NULL);
 606         spin_unlock_irqrestore(&udev->commands_lock, flags);
 607
 608         /*
 609          * We don't need to wakeup threads on wait_cmdr since they have their
 610          * own timeout.
 611          */
 612 }
 613
 614 static int tcmu_attach_hba(struct se_hba *hba, u32 host_id)
 615 {
 616         struct tcmu_hba *tcmu_hba;
 617
 618         tcmu_hba = kzalloc(sizeof(struct tcmu_hba), GFP_KERNEL);
 619         if (!tcmu_hba)
 620                 return -ENOMEM;
 621
 622         tcmu_hba->host_id = host_id;
 623         hba->hba_ptr = tcmu_hba;
 624
 625         return 0;
 626 }
 627
 628 static void tcmu_detach_hba(struct se_hba *hba)
 629 {
 630         kfree(hba->hba_ptr);
 631         hba->hba_ptr = NULL;
 632 }
 633
 634 static struct se_device *tcmu_alloc_device(struct se_hba *hba, const char *name)
 635 {
 636         struct tcmu_dev *udev;
 637
 638         udev = kzalloc(sizeof(struct tcmu_dev), GFP_KERNEL);
 639         if (!udev)
 640                 return NULL;
 641
 642         udev->name = kstrdup(name, GFP_KERNEL);
 643         if (!udev->name) {
 644                 kfree(udev);
 645                 return NULL;
 646         }
 647
 648         udev->hba = hba;
 649
 650         init_waitqueue_head(&udev->wait_cmdr);
 651         spin_lock_init(&udev->cmdr_lock);
 652
 653         idr_init(&udev->commands);
 654         spin_lock_init(&udev->commands_lock);
 655
 656         setup_timer(&udev->timeout, tcmu_device_timedout,
 657                 (unsigned long)udev);
 658
 659         udev->pass_level = TCMU_PASS_ALL;
 660
 661         return &udev->se_dev;
 662 }
 663
 664 static int tcmu_irqcontrol(struct uio_info *info, s32 irq_on)
 665 {
 666         struct tcmu_dev *tcmu_dev = container_of(info, struct tcmu_dev, uio_info);
 667
 668         tcmu_handle_completions(tcmu_dev);
 669
 670         return 0;
 671 }
 672
 673 /*
 674  * mmap code from uio.c. Copied here because we want to hook mmap()
 675  * and this stuff must come along.
 676  */
 677 static int tcmu_find_mem_index(struct vm_area_struct *vma)
 678 {
 679         struct tcmu_dev *udev = vma->vm_private_data;
 680         struct uio_info *info = &udev->uio_info;
 681
 682         if (vma->vm_pgoff < MAX_UIO_MAPS) {
 683                 if (info->mem[vma->vm_pgoff].size == 0)
 684                         return -1;
 685                 return (int)vma->vm_pgoff;
 686         }
 687         return -1;
 688 }
 689
 690 static int tcmu_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 691 {
 692         struct tcmu_dev *udev = vma->vm_private_data;
 693         struct uio_info *info = &udev->uio_info;
 694         struct page *page;
 695         unsigned long offset;
 696         void *addr;
 697
 698         int mi = tcmu_find_mem_index(vma);
 699         if (mi < 0)
 700                 return VM_FAULT_SIGBUS;
 701
 702         /*
 703          * We need to subtract mi because userspace uses offset = N*PAGE_SIZE
 704          * to use mem[N].
 705          */
 706         offset = (vmf->pgoff - mi) << PAGE_SHIFT;
 707
 708         addr = (void *)(unsigned long)info->mem[mi].addr + offset;
 709         if (info->mem[mi].memtype == UIO_MEM_LOGICAL)
 710                 page = virt_to_page(addr);
 711         else
 712                 page = vmalloc_to_page(addr);
 713         get_page(page);
 714         vmf->page = page;
 715         return 0;
 716 }
 717
 718 static const struct vm_operations_struct tcmu_vm_ops = {
 719         .fault = tcmu_vma_fault,
 720 };
 721
 722 static int tcmu_mmap(struct uio_info *info, struct vm_area_struct *vma)
 723 {
 724         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 725
 726         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
 727         vma->vm_ops = &tcmu_vm_ops;
 728
 729         vma->vm_private_data = udev;
 730
 731         /* Ensure the mmap is exactly the right size */
 732         if (vma_pages(vma) != (TCMU_RING_SIZE >> PAGE_SHIFT))
 733                 return -EINVAL;
 734
 735         return 0;
 736 }
 737
 738 static int tcmu_open(struct uio_info *info, struct inode *inode)
 739 {
 740         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 741
 742         /* O_EXCL not supported for char devs, so fake it? */
 743         if (test_and_set_bit(TCMU_DEV_BIT_OPEN, &udev->flags))
 744                 return -EBUSY;
 745
 746         pr_debug("open\n");
 747
 748         return 0;
 749 }
 750
 751 static int tcmu_release(struct uio_info *info, struct inode *inode)
 752 {
 753         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
 754
 755         clear_bit(TCMU_DEV_BIT_OPEN, &udev->flags);
 756
 757         pr_debug("close\n");
 758
 759         return 0;
 760 }
 761
 762 static int tcmu_netlink_event(enum tcmu_genl_cmd cmd, const char *name, int minor)
 763 {
 764         struct sk_buff *skb;
 765         void *msg_header;
 766         int ret;
 767
 768         skb = genlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
 769         if (!skb)
 770                 return -ENOMEM;
 771
 772         msg_header = genlmsg_put(skb, 0, 0, &tcmu_genl_family, 0, cmd);
 773         if (!msg_header) {
 774                 nlmsg_free(skb);
 775                 return -ENOMEM;
 776         }
 777
 778         ret = nla_put_string(skb, TCMU_ATTR_DEVICE, name);
 779
 780         ret = nla_put_u32(skb, TCMU_ATTR_MINOR, minor);
 781
 782         ret = genlmsg_end(skb, msg_header);
 783         if (ret < 0) {
 784                 nlmsg_free(skb);
 785                 return ret;
 786         }
 787
 788         ret = genlmsg_multicast(&tcmu_genl_family, skb, 0,
 789                                 TCMU_MCGRP_CONFIG, GFP_KERNEL);
 790
 791         /* We don't care if no one is listening */
 792         if (ret == -ESRCH)
 793                 ret = 0;
 794
 795         return ret;
 796 }
 797
 798 static int tcmu_configure_device(struct se_device *dev)
 799 {
 800         struct tcmu_dev *udev = TCMU_DEV(dev);
 801         struct tcmu_hba *hba = udev->hba->hba_ptr;
 802         struct uio_info *info;
 803         struct tcmu_mailbox *mb;
 804         size_t size;
 805         size_t used;
 806         int ret = 0;
 807         char *str;
 808
 809         info = &udev->uio_info;
 810
 811         size = snprintf(NULL, 0, "tcm-user/%u/%s/%s", hba->host_id, udev->name,
 812                         udev->dev_config);
 813         size += 1; /* for \0 */
 814         str = kmalloc(size, GFP_KERNEL);
 815         if (!str)
 816                 return -ENOMEM;
 817
 818         used = snprintf(str, size, "tcm-user/%u/%s", hba->host_id, udev->name);
 819
 820         if (udev->dev_config[0])
 821                 snprintf(str + used, size - used, "/%s", udev->dev_config);
 822
 823         info->name = str;
 824
 825         udev->mb_addr = vzalloc(TCMU_RING_SIZE);
 826         if (!udev->mb_addr) {
 827                 ret = -ENOMEM;
 828                 goto err_vzalloc;
 829         }
 830
 831         /* mailbox fits in first part of CMDR space */
 832         udev->cmdr_size = CMDR_SIZE - CMDR_OFF;
 833         udev->data_off = CMDR_SIZE;
 834         udev->data_size = TCMU_RING_SIZE - CMDR_SIZE;
 835
 836         mb = udev->mb_addr;
 837         mb->version = 1;
 838         mb->cmdr_off = CMDR_OFF;
 839         mb->cmdr_size = udev->cmdr_size;
 840
 841         WARN_ON(!PAGE_ALIGNED(udev->data_off));
 842         WARN_ON(udev->data_size % PAGE_SIZE);
 843
 844         info->version = "1";
 845
 846         info->mem[0].name = "tcm-user command & data buffer";
 847         info->mem[0].addr = (phys_addr_t) udev->mb_addr;
 848         info->mem[0].size = TCMU_RING_SIZE;
 849         info->mem[0].memtype = UIO_MEM_VIRTUAL;
 850
 851         info->irqcontrol = tcmu_irqcontrol;
 852         info->irq = UIO_IRQ_CUSTOM;
 853
 854         info->mmap = tcmu_mmap;
 855         info->open = tcmu_open;
 856         info->release = tcmu_release;
 857
 858         ret = uio_register_device(tcmu_root_device, info);
 859         if (ret)
 860                 goto err_register;
 861
 862         /* Other attributes can be configured in userspace */
 863         dev->dev_attrib.hw_block_size = 512;
 864         dev->dev_attrib.hw_max_sectors = 128;
 865         dev->dev_attrib.hw_queue_depth = 128;
 866
 867         ret = tcmu_netlink_event(TCMU_CMD_ADDED_DEVICE, udev->uio_info.name,
 868                                  udev->uio_info.uio_dev->minor);
 869         if (ret)
 870                 goto err_netlink;
 871
 872         return 0;
 873
 874 err_netlink:
 875         uio_unregister_device(&udev->uio_info);
 876 err_register:
 877         vfree(udev->mb_addr);
 878 err_vzalloc:
 879         kfree(info->name);
 880
 881         return ret;
 882 }
 883
 884 static int tcmu_check_pending_cmd(int id, void *p, void *data)
 885 {
 886         struct tcmu_cmd *cmd = p;
 887
 888         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
 889                 return 0;
 890         return -EINVAL;
 891 }
 892
 893 static void tcmu_free_device(struct se_device *dev)
 894 {
 895         struct tcmu_dev *udev = TCMU_DEV(dev);
 896         int i;
 897
 898         del_timer_sync(&udev->timeout);
 899
 900         vfree(udev->mb_addr);
 901
 902         /* Upper layer should drain all requests before calling this */
 903         spin_lock_irq(&udev->commands_lock);
 904         i = idr_for_each(&udev->commands, tcmu_check_pending_cmd, NULL);
 905         idr_destroy(&udev->commands);
 906         spin_unlock_irq(&udev->commands_lock);
 907         WARN_ON(i);
 908
 909         /* Device was configured */
 910         if (udev->uio_info.uio_dev) {
 911                 tcmu_netlink_event(TCMU_CMD_REMOVED_DEVICE, udev->uio_info.name,
 912                                    udev->uio_info.uio_dev->minor);
 913
 914                 uio_unregister_device(&udev->uio_info);
 915                 kfree(udev->uio_info.name);
 916                 kfree(udev->name);
 917         }
 918
 919         kfree(udev);
 920 }
 921
 922 enum {
 923         Opt_dev_config, Opt_dev_size, Opt_err, Opt_pass_level,
 924 };
 925
 926 static match_table_t tokens = {
 927         {Opt_dev_config, "dev_config=%s"},
 928         {Opt_dev_size, "dev_size=%u"},
 929         {Opt_pass_level, "pass_level=%u"},
 930         {Opt_err, NULL}
 931 };
 932
 933 static ssize_t tcmu_set_configfs_dev_params(struct se_device *dev,
 934                 const char *page, ssize_t count)
 935 {
 936         struct tcmu_dev *udev = TCMU_DEV(dev);
 937         char *orig, *ptr, *opts, *arg_p;
 938         substring_t args[MAX_OPT_ARGS];
 939         int ret = 0, token;
 940         int arg;
 941
 942         opts = kstrdup(page, GFP_KERNEL);
 943         if (!opts)
 944                 return -ENOMEM;
 945
 946         orig = opts;
 947
 948         while ((ptr = strsep(&opts, ",\n")) != NULL) {
 949                 if (!*ptr)
 950                         continue;
 951
 952                 token = match_token(ptr, tokens, args);
 953                 switch (token) {
 954                 case Opt_dev_config:
 955                         if (match_strlcpy(udev->dev_config, &args[0],
 956                                           TCMU_CONFIG_LEN) == 0) {
 957                                 ret = -EINVAL;
 958                                 break;
 959                         }
 960                         pr_debug("TCMU: Referencing Path: %s\n", udev->dev_config);
 961                         break;
 962                 case Opt_dev_size:
 963                         arg_p = match_strdup(&args[0]);
 964                         if (!arg_p) {
 965                                 ret = -ENOMEM;
 966                                 break;
 967                         }
 968                         ret = kstrtoul(arg_p, 0, (unsigned long *) &udev->dev_size);
 969                         kfree(arg_p);
 970                         if (ret < 0)
 971                                 pr_err("kstrtoul() failed for dev_size=\n");
 972                         break;
 973                 case Opt_pass_level:
 974                         match_int(args, &arg);
 975                         if (arg >= TCMU_PASS_INVALID) {
 976                                 pr_warn("TCMU: Invalid pass_level: %d\n", arg);
 977                                 break;
 978                         }
 979
 980                         pr_debug("TCMU: Setting pass_level to %d\n", arg);
 981                         udev->pass_level = arg;
 982                         break;
 983                 default:
 984                         break;
 985                 }
 986         }
 987
 988         kfree(orig);
 989         return (!ret) ? count : ret;
 990 }
 991
 992 static ssize_t tcmu_show_configfs_dev_params(struct se_device *dev, char *b)
 993 {
 994         struct tcmu_dev *udev = TCMU_DEV(dev);
 995         ssize_t bl = 0;
 996
 997         bl = sprintf(b + bl, "Config: %s ",
 998                      udev->dev_config[0] ? udev->dev_config : "NULL");
 999         bl += sprintf(b + bl, "Size: %zu PassLevel: %u\n",
1000                       udev->dev_size, udev->pass_level);
1001
1002         return bl;
1003 }
1004
1005 static sector_t tcmu_get_blocks(struct se_device *dev)
1006 {
1007         struct tcmu_dev *udev = TCMU_DEV(dev);
1008
1009         return div_u64(udev->dev_size - dev->dev_attrib.block_size,
1010                        dev->dev_attrib.block_size);
1011 }
1012
1013 static sense_reason_t
1014 tcmu_execute_rw(struct se_cmd *se_cmd, struct scatterlist *sgl, u32 sgl_nents,
1015                 enum dma_data_direction data_direction)
1016 {
1017         int ret;
1018
1019         ret = tcmu_queue_cmd(se_cmd);
1020
1021         if (ret != 0)
1022                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1023         else
1024                 return TCM_NO_SENSE;
1025 }
1026
1027 static sense_reason_t
1028 tcmu_pass_op(struct se_cmd *se_cmd)
1029 {
1030         int ret = tcmu_queue_cmd(se_cmd);
1031
1032         if (ret != 0)
1033                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1034         else
1035                 return TCM_NO_SENSE;
1036 }
1037
1038 static struct sbc_ops tcmu_sbc_ops = {
1039         .execute_rw = tcmu_execute_rw,
1040         .execute_sync_cache     = tcmu_pass_op,
1041         .execute_write_same     = tcmu_pass_op,
1042         .execute_write_same_unmap = tcmu_pass_op,
1043         .execute_unmap          = tcmu_pass_op,
1044 };
1045
1046 static sense_reason_t
1047 tcmu_parse_cdb(struct se_cmd *cmd)
1048 {
1049         unsigned char *cdb = cmd->t_task_cdb;
1050         struct tcmu_dev *udev = TCMU_DEV(cmd->se_dev);
1051         sense_reason_t ret;
1052
1053         switch (udev->pass_level) {
1054         case TCMU_PASS_ALL:
1055                 /* We're just like pscsi, then */
1056                 /*
1057                  * For REPORT LUNS we always need to emulate the response, for everything
1058                  * else, pass it up.
1059                  */
1060                 switch (cdb[0]) {
1061                 case REPORT_LUNS:
1062                         cmd->execute_cmd = spc_emulate_report_luns;
1063                         break;
1064                 case READ_6:
1065                 case READ_10:
1066                 case READ_12:
1067                 case READ_16:
1068                 case WRITE_6:
1069                 case WRITE_10:
1070                 case WRITE_12:
1071                 case WRITE_16:
1072                 case WRITE_VERIFY:
1073                         cmd->se_cmd_flags |= SCF_SCSI_DATA_CDB;
1074                         /* FALLTHROUGH */
1075                 default:
1076                         cmd->execute_cmd = tcmu_pass_op;
1077                 }
1078                 ret = TCM_NO_SENSE;
1079                 break;
1080         case TCMU_PASS_IO:
1081                 ret = sbc_parse_cdb(cmd, &tcmu_sbc_ops);
1082                 break;
1083         default:
1084                 pr_err("Unknown tcm-user pass level %d\n", udev->pass_level);
1085                 ret = TCM_CHECK_CONDITION_ABORT_CMD;
1086         }
1087
1088         return ret;
1089 }
1090
1091 static struct se_subsystem_api tcmu_template = {
1092         .name                   = "user",
1093         .inquiry_prod           = "USER",
1094         .inquiry_rev            = TCMU_VERSION,
1095         .owner                  = THIS_MODULE,
1096         .transport_type         = TRANSPORT_PLUGIN_VHBA_PDEV,
1097         .attach_hba             = tcmu_attach_hba,
1098         .detach_hba             = tcmu_detach_hba,
1099         .alloc_device           = tcmu_alloc_device,
1100         .configure_device       = tcmu_configure_device,
1101         .free_device            = tcmu_free_device,
1102         .parse_cdb              = tcmu_parse_cdb,
1103         .set_configfs_dev_params = tcmu_set_configfs_dev_params,
1104         .show_configfs_dev_params = tcmu_show_configfs_dev_params,
1105         .get_device_type        = sbc_get_device_type,
1106         .get_blocks             = tcmu_get_blocks,
1107 };
1108
1109 static int __init tcmu_module_init(void)
1110 {
1111         int ret;
1112
1113         BUILD_BUG_ON((sizeof(struct tcmu_cmd_entry) % TCMU_OP_ALIGN_SIZE) != 0);
1114
1115         tcmu_cmd_cache = kmem_cache_create("tcmu_cmd_cache",
1116                                 sizeof(struct tcmu_cmd),
1117                                 __alignof__(struct tcmu_cmd),
1118                                 0, NULL);
1119         if (!tcmu_cmd_cache)
1120                 return -ENOMEM;
1121
1122         tcmu_root_device = root_device_register("tcm_user");
1123         if (IS_ERR(tcmu_root_device)) {
1124                 ret = PTR_ERR(tcmu_root_device);
1125                 goto out_free_cache;
1126         }
1127
1128         ret = genl_register_family(&tcmu_genl_family);
1129         if (ret < 0) {
1130                 goto out_unreg_device;
1131         }
1132
1133         ret = transport_subsystem_register(&tcmu_template);
1134         if (ret)
1135                 goto out_unreg_genl;
1136
1137         return 0;
1138
1139 out_unreg_genl:
1140         genl_unregister_family(&tcmu_genl_family);
1141 out_unreg_device:
1142         root_device_unregister(tcmu_root_device);
1143 out_free_cache:
1144         kmem_cache_destroy(tcmu_cmd_cache);
1145
1146         return ret;
1147 }
1148
1149 static void __exit tcmu_module_exit(void)
1150 {
1151         transport_subsystem_release(&tcmu_template);
1152         genl_unregister_family(&tcmu_genl_family);
1153         root_device_unregister(tcmu_root_device);
1154         kmem_cache_destroy(tcmu_cmd_cache);
1155 }
1156
1157 MODULE_DESCRIPTION("TCM USER subsystem plugin");
1158 MODULE_AUTHOR("Shaohua Li <shli@kernel.org>");
1159 MODULE_AUTHOR("Andy Grover <agrover@redhat.com>");
1160 MODULE_LICENSE("GPL");
1161
1162 module_init(tcmu_module_init);
1163 module_exit(tcmu_module_exit);