fs/orangefs/devorangefs-req.c

   1 /*
   2  * (C) 2001 Clemson University and The University of Chicago
   3  *
   4  * Changes by Acxiom Corporation to add protocol version to kernel
   5  * communication, Copyright Acxiom Corporation, 2005.
   6  *
   7  * See COPYING in top-level directory.
   8  */
   9
  10 #include "protocol.h"
  11 #include "orangefs-kernel.h"
  12 #include "orangefs-dev-proto.h"
  13 #include "orangefs-bufmap.h"
  14 #include "orangefs-debugfs.h"
  15
  16 #include <linux/debugfs.h>
  17 #include <linux/slab.h>
  18
  19 /* this file implements the /dev/pvfs2-req device node */
  20
  21 uint32_t orangefs_userspace_version;
  22
  23 static int open_access_count;
  24
  25 static DEFINE_MUTEX(devreq_mutex);
  26
  27 #define DUMP_DEVICE_ERROR()                                                   \
  28 do {                                                                          \
  29         gossip_err("*****************************************************\n");\
  30         gossip_err("ORANGEFS Device Error:  You cannot open the device file ");  \
  31         gossip_err("\n/dev/%s more than once.  Please make sure that\nthere " \
  32                    "are no ", ORANGEFS_REQDEVICE_NAME);                          \
  33         gossip_err("instances of a program using this device\ncurrently "     \
  34                    "running. (You must verify this!)\n");                     \
  35         gossip_err("For example, you can use the lsof program as follows:\n");\
  36         gossip_err("'lsof | grep %s' (run this as root)\n",                   \
  37                    ORANGEFS_REQDEVICE_NAME);                                     \
  38         gossip_err("  open_access_count = %d\n", open_access_count);          \
  39         gossip_err("*****************************************************\n");\
  40 } while (0)
  41
  42 static int hash_func(__u64 tag, int table_size)
  43 {
  44         return do_div(tag, (unsigned int)table_size);
  45 }
  46
  47 static void orangefs_devreq_add_op(struct orangefs_kernel_op_s *op)
  48 {
  49         int index = hash_func(op->tag, hash_table_size);
  50
  51         list_add_tail(&op->list, &orangefs_htable_ops_in_progress[index]);
  52 }
  53
  54 /*
  55  * find the op with this tag and remove it from the in progress
  56  * hash table.
  57  */
  58 static struct orangefs_kernel_op_s *orangefs_devreq_remove_op(__u64 tag)
  59 {
  60         struct orangefs_kernel_op_s *op, *next;
  61         int index;
  62
  63         index = hash_func(tag, hash_table_size);
  64
  65         spin_lock(&orangefs_htable_ops_in_progress_lock);
  66         list_for_each_entry_safe(op,
  67                                  next,
  68                                  &orangefs_htable_ops_in_progress[index],
  69                                  list) {
  70                 if (op->tag == tag && !op_state_purged(op) &&
  71                     !op_state_given_up(op)) {
  72                         list_del_init(&op->list);
  73                         spin_unlock(&orangefs_htable_ops_in_progress_lock);
  74                         return op;
  75                 }
  76         }
  77
  78         spin_unlock(&orangefs_htable_ops_in_progress_lock);
  79         return NULL;
  80 }
  81
  82 /* Returns whether any FS are still pending remounted */
  83 static int mark_all_pending_mounts(void)
  84 {
  85         int unmounted = 1;
  86         struct orangefs_sb_info_s *orangefs_sb = NULL;
  87
  88         spin_lock(&orangefs_superblocks_lock);
  89         list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
  90                 /* All of these file system require a remount */
  91                 orangefs_sb->mount_pending = 1;
  92                 unmounted = 0;
  93         }
  94         spin_unlock(&orangefs_superblocks_lock);
  95         return unmounted;
  96 }
  97
  98 /*
  99  * Determine if a given file system needs to be remounted or not
 100  *  Returns -1 on error
 101  *           0 if already mounted
 102  *           1 if needs remount
 103  */
 104 static int fs_mount_pending(__s32 fsid)
 105 {
 106         int mount_pending = -1;
 107         struct orangefs_sb_info_s *orangefs_sb = NULL;
 108
 109         spin_lock(&orangefs_superblocks_lock);
 110         list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
 111                 if (orangefs_sb->fs_id == fsid) {
 112                         mount_pending = orangefs_sb->mount_pending;
 113                         break;
 114                 }
 115         }
 116         spin_unlock(&orangefs_superblocks_lock);
 117         return mount_pending;
 118 }
 119
 120 static int orangefs_devreq_open(struct inode *inode, struct file *file)
 121 {
 122         int ret = -EINVAL;
 123
 124         /* in order to ensure that the filesystem driver sees correct UIDs */
 125         if (file->f_cred->user_ns != &init_user_ns) {
 126                 gossip_err("%s: device cannot be opened outside init_user_ns\n",
 127                            __func__);
 128                 goto out;
 129         }
 130
 131         if (!(file->f_flags & O_NONBLOCK)) {
 132                 gossip_err("%s: device cannot be opened in blocking mode\n",
 133                            __func__);
 134                 goto out;
 135         }
 136         ret = -EACCES;
 137         gossip_debug(GOSSIP_DEV_DEBUG, "client-core: opening device\n");
 138         mutex_lock(&devreq_mutex);
 139
 140         if (open_access_count == 0) {
 141                 open_access_count = 1;
 142                 ret = 0;
 143         } else {
 144                 DUMP_DEVICE_ERROR();
 145         }
 146         mutex_unlock(&devreq_mutex);
 147
 148 out:
 149
 150         gossip_debug(GOSSIP_DEV_DEBUG,
 151                      "pvfs2-client-core: open device complete (ret = %d)\n",
 152                      ret);
 153         return ret;
 154 }
 155
 156 /* Function for read() callers into the device */
 157 static ssize_t orangefs_devreq_read(struct file *file,
 158                                  char __user *buf,
 159                                  size_t count, loff_t *offset)
 160 {
 161         struct orangefs_kernel_op_s *op, *temp;
 162         __s32 proto_ver = ORANGEFS_KERNEL_PROTO_VERSION;
 163         static __s32 magic = ORANGEFS_DEVREQ_MAGIC;
 164         struct orangefs_kernel_op_s *cur_op = NULL;
 165         unsigned long ret;
 166
 167         /* We do not support blocking IO. */
 168         if (!(file->f_flags & O_NONBLOCK)) {
 169                 gossip_err("%s: blocking read from client-core.\n",
 170                            __func__);
 171                 return -EINVAL;
 172         }
 173
 174         /*
 175          * The client will do an ioctl to find MAX_DEV_REQ_UPSIZE, then
 176          * always read with that size buffer.
 177          */
 178         if (count != MAX_DEV_REQ_UPSIZE) {
 179                 gossip_err("orangefs: client-core tried to read wrong size\n");
 180                 return -EINVAL;
 181         }
 182
 183 restart:
 184         /* Get next op (if any) from top of list. */
 185         spin_lock(&orangefs_request_list_lock);
 186         list_for_each_entry_safe(op, temp, &orangefs_request_list, list) {
 187                 __s32 fsid;
 188                 /* This lock is held past the end of the loop when we break. */
 189                 spin_lock(&op->lock);
 190                 if (unlikely(op_state_purged(op) || op_state_given_up(op))) {
 191                         spin_unlock(&op->lock);
 192                         continue;
 193                 }
 194
 195                 fsid = fsid_of_op(op);
 196                 if (fsid != ORANGEFS_FS_ID_NULL) {
 197                         int ret;
 198                         /* Skip ops whose filesystem needs to be mounted. */
 199                         ret = fs_mount_pending(fsid);
 200                         if (ret == 1) {
 201                                 gossip_debug(GOSSIP_DEV_DEBUG,
 202                                     "%s: mount pending, skipping op tag "
 203                                     "%llu %s\n",
 204                                     __func__,
 205                                     llu(op->tag),
 206                                     get_opname_string(op));
 207                                 spin_unlock(&op->lock);
 208                                 continue;
 209                         /*
 210                          * Skip ops whose filesystem we don't know about unless
 211                          * it is being mounted or unmounted.  It is possible for
 212                          * a filesystem we don't know about to be unmounted if
 213                          * it fails to mount in the kernel after userspace has
 214                          * been sent the mount request.
 215                          */
 216                         /* XXX: is there a better way to detect this? */
 217                         } else if (ret == -1 &&
 218                                    !(op->upcall.type ==
 219                                         ORANGEFS_VFS_OP_FS_MOUNT ||
 220                                      op->upcall.type ==
 221                                         ORANGEFS_VFS_OP_GETATTR ||
 222                                      op->upcall.type ==
 223                                         ORANGEFS_VFS_OP_FS_UMOUNT)) {
 224                                 gossip_debug(GOSSIP_DEV_DEBUG,
 225                                     "orangefs: skipping op tag %llu %s\n",
 226                                     llu(op->tag), get_opname_string(op));
 227                                 gossip_err(
 228                                     "orangefs: ERROR: fs_mount_pending %d\n",
 229                                     fsid);
 230                                 spin_unlock(&op->lock);
 231                                 continue;
 232                         }
 233                 }
 234                 /*
 235                  * Either this op does not pertain to a filesystem, is mounting
 236                  * a filesystem, or pertains to a mounted filesystem. Let it
 237                  * through.
 238                  */
 239                 cur_op = op;
 240                 break;
 241         }
 242
 243         /*
 244          * At this point we either have a valid op and can continue or have not
 245          * found an op and must ask the client to try again later.
 246          */
 247         if (!cur_op) {
 248                 spin_unlock(&orangefs_request_list_lock);
 249                 return -EAGAIN;
 250         }
 251
 252         gossip_debug(GOSSIP_DEV_DEBUG, "%s: reading op tag %llu %s\n",
 253                      __func__,
 254                      llu(cur_op->tag),
 255                      get_opname_string(cur_op));
 256
 257         /*
 258          * Such an op should never be on the list in the first place. If so, we
 259          * will abort.
 260          */
 261         if (op_state_in_progress(cur_op) || op_state_serviced(cur_op)) {
 262                 gossip_err("orangefs: ERROR: Current op already queued.\n");
 263                 list_del_init(&cur_op->list);
 264                 spin_unlock(&cur_op->lock);
 265                 spin_unlock(&orangefs_request_list_lock);
 266                 return -EAGAIN;
 267         }
 268
 269         list_del_init(&cur_op->list);
 270         spin_unlock(&orangefs_request_list_lock);
 271
 272         spin_unlock(&cur_op->lock);
 273
 274         /* Push the upcall out. */
 275         ret = copy_to_user(buf, &proto_ver, sizeof(__s32));
 276         if (ret != 0)
 277                 goto error;
 278         ret = copy_to_user(buf+sizeof(__s32), &magic, sizeof(__s32));
 279         if (ret != 0)
 280                 goto error;
 281         ret = copy_to_user(buf+2 * sizeof(__s32), &cur_op->tag, sizeof(__u64));
 282         if (ret != 0)
 283                 goto error;
 284         ret = copy_to_user(buf+2*sizeof(__s32)+sizeof(__u64), &cur_op->upcall,
 285                            sizeof(struct orangefs_upcall_s));
 286         if (ret != 0)
 287                 goto error;
 288
 289         spin_lock(&orangefs_htable_ops_in_progress_lock);
 290         spin_lock(&cur_op->lock);
 291         if (unlikely(op_state_given_up(cur_op))) {
 292                 spin_unlock(&cur_op->lock);
 293                 spin_unlock(&orangefs_htable_ops_in_progress_lock);
 294                 complete(&cur_op->waitq);
 295                 goto restart;
 296         }
 297
 298         /*
 299          * Set the operation to be in progress and move it between lists since
 300          * it has been sent to the client.
 301          */
 302         set_op_state_inprogress(cur_op);
 303         gossip_debug(GOSSIP_DEV_DEBUG,
 304                      "%s: 1 op:%s: op_state:%d: process:%s:\n",
 305                      __func__,
 306                      get_opname_string(cur_op),
 307                      cur_op->op_state,
 308                      current->comm);
 309         orangefs_devreq_add_op(cur_op);
 310         spin_unlock(&cur_op->lock);
 311         spin_unlock(&orangefs_htable_ops_in_progress_lock);
 312
 313         /* The client only asks to read one size buffer. */
 314         return MAX_DEV_REQ_UPSIZE;
 315 error:
 316         /*
 317          * We were unable to copy the op data to the client. Put the op back in
 318          * list. If client has crashed, the op will be purged later when the
 319          * device is released.
 320          */
 321         gossip_err("orangefs: Failed to copy data to user space\n");
 322         spin_lock(&orangefs_request_list_lock);
 323         spin_lock(&cur_op->lock);
 324         if (likely(!op_state_given_up(cur_op))) {
 325                 set_op_state_waiting(cur_op);
 326                 gossip_debug(GOSSIP_DEV_DEBUG,
 327                              "%s: 2 op:%s: op_state:%d: process:%s:\n",
 328                              __func__,
 329                              get_opname_string(cur_op),
 330                              cur_op->op_state,
 331                              current->comm);
 332                 list_add(&cur_op->list, &orangefs_request_list);
 333                 spin_unlock(&cur_op->lock);
 334         } else {
 335                 spin_unlock(&cur_op->lock);
 336                 complete(&cur_op->waitq);
 337         }
 338         spin_unlock(&orangefs_request_list_lock);
 339         return -EFAULT;
 340 }
 341
 342 /*
 343  * Function for writev() callers into the device.
 344  *
 345  * Userspace should have written:
 346  *  - __u32 version
 347  *  - __u32 magic
 348  *  - __u64 tag
 349  *  - struct orangefs_downcall_s
 350  *  - trailer buffer (in the case of READDIR operations)
 351  */
 352 static ssize_t orangefs_devreq_write_iter(struct kiocb *iocb,
 353                                       struct iov_iter *iter)
 354 {
 355         ssize_t ret;
 356         struct orangefs_kernel_op_s *op = NULL;
 357         struct {
 358                 __u32 version;
 359                 __u32 magic;
 360                 __u64 tag;
 361         } head;
 362         int total = ret = iov_iter_count(iter);
 363         int downcall_size = sizeof(struct orangefs_downcall_s);
 364         int head_size = sizeof(head);
 365
 366         gossip_debug(GOSSIP_DEV_DEBUG, "%s: total:%d: ret:%zd:\n",
 367                      __func__,
 368                      total,
 369                      ret);
 370
 371         if (total < MAX_DEV_REQ_DOWNSIZE) {
 372                 gossip_err("%s: total:%d: must be at least:%u:\n",
 373                            __func__,
 374                            total,
 375                            (unsigned int) MAX_DEV_REQ_DOWNSIZE);
 376                 return -EFAULT;
 377         }
 378
 379         if (!copy_from_iter_full(&head, head_size, iter)) {
 380                 gossip_err("%s: failed to copy head.\n", __func__);
 381                 return -EFAULT;
 382         }
 383
 384         if (head.version < ORANGEFS_MINIMUM_USERSPACE_VERSION) {
 385                 gossip_err("%s: userspace claims version"
 386                            "%d, minimum version required: %d.\n",
 387                            __func__,
 388                            head.version,
 389                            ORANGEFS_MINIMUM_USERSPACE_VERSION);
 390                 return -EPROTO;
 391         }
 392
 393         if (head.magic != ORANGEFS_DEVREQ_MAGIC) {
 394                 gossip_err("Error: Device magic number does not match.\n");
 395                 return -EPROTO;
 396         }
 397
 398         if (!orangefs_userspace_version) {
 399                 orangefs_userspace_version = head.version;
 400         } else if (orangefs_userspace_version != head.version) {
 401                 gossip_err("Error: userspace version changes\n");
 402                 return -EPROTO;
 403         }
 404
 405         /* remove the op from the in progress hash table */
 406         op = orangefs_devreq_remove_op(head.tag);
 407         if (!op) {
 408                 gossip_debug(GOSSIP_DEV_DEBUG,
 409                              "%s: No one's waiting for tag %llu\n",
 410                              __func__, llu(head.tag));
 411                 return ret;
 412         }
 413
 414         if (!copy_from_iter_full(&op->downcall, downcall_size, iter)) {
 415                 gossip_err("%s: failed to copy downcall.\n", __func__);
 416                 goto Efault;
 417         }
 418
 419         if (op->downcall.status)
 420                 goto wakeup;
 421
 422         /*
 423          * We've successfully peeled off the head and the downcall.
 424          * Something has gone awry if total doesn't equal the
 425          * sum of head_size, downcall_size and trailer_size.
 426          */
 427         if ((head_size + downcall_size + op->downcall.trailer_size) != total) {
 428                 gossip_err("%s: funky write, head_size:%d"
 429                            ": downcall_size:%d: trailer_size:%lld"
 430                            ": total size:%d:\n",
 431                            __func__,
 432                            head_size,
 433                            downcall_size,
 434                            op->downcall.trailer_size,
 435                            total);
 436                 goto Efault;
 437         }
 438
 439         /* Only READDIR operations should have trailers. */
 440         if ((op->downcall.type != ORANGEFS_VFS_OP_READDIR) &&
 441             (op->downcall.trailer_size != 0)) {
 442                 gossip_err("%s: %x operation with trailer.",
 443                            __func__,
 444                            op->downcall.type);
 445                 goto Efault;
 446         }
 447
 448         /* READDIR operations should always have trailers. */
 449         if ((op->downcall.type == ORANGEFS_VFS_OP_READDIR) &&
 450             (op->downcall.trailer_size == 0)) {
 451                 gossip_err("%s: %x operation with no trailer.",
 452                            __func__,
 453                            op->downcall.type);
 454                 goto Efault;
 455         }
 456
 457         if (op->downcall.type != ORANGEFS_VFS_OP_READDIR)
 458                 goto wakeup;
 459
 460         op->downcall.trailer_buf =
 461                 vmalloc(op->downcall.trailer_size);
 462         if (op->downcall.trailer_buf == NULL) {
 463                 gossip_err("%s: failed trailer vmalloc.\n",
 464                            __func__);
 465                 goto Enomem;
 466         }
 467         memset(op->downcall.trailer_buf, 0, op->downcall.trailer_size);
 468         if (!copy_from_iter_full(op->downcall.trailer_buf,
 469                                  op->downcall.trailer_size, iter)) {
 470                 gossip_err("%s: failed to copy trailer.\n", __func__);
 471                 vfree(op->downcall.trailer_buf);
 472                 goto Efault;
 473         }
 474
 475 wakeup:
 476         /*
 477          * Return to vfs waitqueue, and back to service_operation
 478          * through wait_for_matching_downcall.
 479          */
 480         spin_lock(&op->lock);
 481         if (unlikely(op_is_cancel(op))) {
 482                 spin_unlock(&op->lock);
 483                 put_cancel(op);
 484         } else if (unlikely(op_state_given_up(op))) {
 485                 spin_unlock(&op->lock);
 486                 complete(&op->waitq);
 487         } else {
 488                 set_op_state_serviced(op);
 489                 gossip_debug(GOSSIP_DEV_DEBUG,
 490                              "%s: op:%s: op_state:%d: process:%s:\n",
 491                              __func__,
 492                              get_opname_string(op),
 493                              op->op_state,
 494                              current->comm);
 495                 spin_unlock(&op->lock);
 496         }
 497         return ret;
 498
 499 Efault:
 500         op->downcall.status = -(ORANGEFS_ERROR_BIT | 9);
 501         ret = -EFAULT;
 502         goto wakeup;
 503
 504 Enomem:
 505         op->downcall.status = -(ORANGEFS_ERROR_BIT | 8);
 506         ret = -ENOMEM;
 507         goto wakeup;
 508 }
 509
 510 /*
 511  * NOTE: gets called when the last reference to this device is dropped.
 512  * Using the open_access_count variable, we enforce a reference count
 513  * on this file so that it can be opened by only one process at a time.
 514  * the devreq_mutex is used to make sure all i/o has completed
 515  * before we call orangefs_bufmap_finalize, and similar such tricky
 516  * situations
 517  */
 518 static int orangefs_devreq_release(struct inode *inode, struct file *file)
 519 {
 520         int unmounted = 0;
 521
 522         gossip_debug(GOSSIP_DEV_DEBUG,
 523                      "%s:pvfs2-client-core: exiting, closing device\n",
 524                      __func__);
 525
 526         mutex_lock(&devreq_mutex);
 527         orangefs_bufmap_finalize();
 528
 529         open_access_count = -1;
 530
 531         unmounted = mark_all_pending_mounts();
 532         gossip_debug(GOSSIP_DEV_DEBUG, "ORANGEFS Device Close: Filesystem(s) %s\n",
 533                      (unmounted ? "UNMOUNTED" : "MOUNTED"));
 534
 535         purge_waiting_ops();
 536         purge_inprogress_ops();
 537
 538         orangefs_bufmap_run_down();
 539
 540         gossip_debug(GOSSIP_DEV_DEBUG,
 541                      "pvfs2-client-core: device close complete\n");
 542         open_access_count = 0;
 543         orangefs_userspace_version = 0;
 544         mutex_unlock(&devreq_mutex);
 545         return 0;
 546 }
 547
 548 int is_daemon_in_service(void)
 549 {
 550         int in_service;
 551
 552         /*
 553          * What this function does is checks if client-core is alive
 554          * based on the access count we maintain on the device.
 555          */
 556         mutex_lock(&devreq_mutex);
 557         in_service = open_access_count == 1 ? 0 : -EIO;
 558         mutex_unlock(&devreq_mutex);
 559         return in_service;
 560 }
 561
 562 bool __is_daemon_in_service(void)
 563 {
 564         return open_access_count == 1;
 565 }
 566
 567 static inline long check_ioctl_command(unsigned int command)
 568 {
 569         /* Check for valid ioctl codes */
 570         if (_IOC_TYPE(command) != ORANGEFS_DEV_MAGIC) {
 571                 gossip_err("device ioctl magic numbers don't match! Did you rebuild pvfs2-client-core/libpvfs2? [cmd %x, magic %x != %x]\n",
 572                         command,
 573                         _IOC_TYPE(command),
 574                         ORANGEFS_DEV_MAGIC);
 575                 return -EINVAL;
 576         }
 577         /* and valid ioctl commands */
 578         if (_IOC_NR(command) >= ORANGEFS_DEV_MAXNR || _IOC_NR(command) <= 0) {
 579                 gossip_err("Invalid ioctl command number [%d >= %d]\n",
 580                            _IOC_NR(command), ORANGEFS_DEV_MAXNR);
 581                 return -ENOIOCTLCMD;
 582         }
 583         return 0;
 584 }
 585
 586 static long dispatch_ioctl_command(unsigned int command, unsigned long arg)
 587 {
 588         static __s32 magic = ORANGEFS_DEVREQ_MAGIC;
 589         static __s32 max_up_size = MAX_DEV_REQ_UPSIZE;
 590         static __s32 max_down_size = MAX_DEV_REQ_DOWNSIZE;
 591         struct ORANGEFS_dev_map_desc user_desc;
 592         int ret = 0;
 593         int upstream_kmod = 1;
 594         struct orangefs_sb_info_s *orangefs_sb;
 595
 596         /* mtmoore: add locking here */
 597
 598         switch (command) {
 599         case ORANGEFS_DEV_GET_MAGIC:
 600                 return ((put_user(magic, (__s32 __user *) arg) == -EFAULT) ?
 601                         -EIO :
 602                         0);
 603         case ORANGEFS_DEV_GET_MAX_UPSIZE:
 604                 return ((put_user(max_up_size,
 605                                   (__s32 __user *) arg) == -EFAULT) ?
 606                                         -EIO :
 607                                         0);
 608         case ORANGEFS_DEV_GET_MAX_DOWNSIZE:
 609                 return ((put_user(max_down_size,
 610                                   (__s32 __user *) arg) == -EFAULT) ?
 611                                         -EIO :
 612                                         0);
 613         case ORANGEFS_DEV_MAP:
 614                 ret = copy_from_user(&user_desc,
 615                                      (struct ORANGEFS_dev_map_desc __user *)
 616                                      arg,
 617                                      sizeof(struct ORANGEFS_dev_map_desc));
 618                 /* WTF -EIO and not -EFAULT? */
 619                 return ret ? -EIO : orangefs_bufmap_initialize(&user_desc);
 620         case ORANGEFS_DEV_REMOUNT_ALL:
 621                 gossip_debug(GOSSIP_DEV_DEBUG,
 622                              "%s: got ORANGEFS_DEV_REMOUNT_ALL\n",
 623                              __func__);
 624
 625                 /*
 626                  * remount all mounted orangefs volumes to regain the lost
 627                  * dynamic mount tables (if any) -- NOTE: this is done
 628                  * without keeping the superblock list locked due to the
 629                  * upcall/downcall waiting.  also, the request mutex is
 630                  * used to ensure that no operations will be serviced until
 631                  * all of the remounts are serviced (to avoid ops between
 632                  * mounts to fail)
 633                  */
 634                 ret = mutex_lock_interruptible(&orangefs_request_mutex);
 635                 if (ret < 0)
 636                         return ret;
 637                 gossip_debug(GOSSIP_DEV_DEBUG,
 638                              "%s: priority remount in progress\n",
 639                              __func__);
 640                 spin_lock(&orangefs_superblocks_lock);
 641                 list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
 642                         /*
 643                          * We have to drop the spinlock, so entries can be
 644                          * removed.  They can't be freed, though, so we just
 645                          * keep the forward pointers and zero the back ones -
 646                          * that way we can get to the rest of the list.
 647                          */
 648                         if (!orangefs_sb->list.prev)
 649                                 continue;
 650                         gossip_debug(GOSSIP_DEV_DEBUG,
 651                                      "%s: Remounting SB %p\n",
 652                                      __func__,
 653                                      orangefs_sb);
 654
 655                         spin_unlock(&orangefs_superblocks_lock);
 656                         ret = orangefs_remount(orangefs_sb);
 657                         spin_lock(&orangefs_superblocks_lock);
 658                         if (ret) {
 659                                 gossip_debug(GOSSIP_DEV_DEBUG,
 660                                              "SB %p remount failed\n",
 661                                              orangefs_sb);
 662                                 break;
 663                         }
 664                 }
 665                 spin_unlock(&orangefs_superblocks_lock);
 666                 gossip_debug(GOSSIP_DEV_DEBUG,
 667                              "%s: priority remount complete\n",
 668                              __func__);
 669                 mutex_unlock(&orangefs_request_mutex);
 670                 return ret;
 671
 672         case ORANGEFS_DEV_UPSTREAM:
 673                 ret = copy_to_user((void __user *)arg,
 674                                     &upstream_kmod,
 675                                     sizeof(upstream_kmod));
 676
 677                 if (ret != 0)
 678                         return -EIO;
 679                 else
 680                         return ret;
 681
 682         case ORANGEFS_DEV_CLIENT_MASK:
 683                 return orangefs_debugfs_new_client_mask((void __user *)arg);
 684         case ORANGEFS_DEV_CLIENT_STRING:
 685                 return orangefs_debugfs_new_client_string((void __user *)arg);
 686         case ORANGEFS_DEV_DEBUG:
 687                 return orangefs_debugfs_new_debug((void __user *)arg);
 688         default:
 689                 return -ENOIOCTLCMD;
 690         }
 691         return -ENOIOCTLCMD;
 692 }
 693
 694 static long orangefs_devreq_ioctl(struct file *file,
 695                                unsigned int command, unsigned long arg)
 696 {
 697         long ret;
 698
 699         /* Check for properly constructed commands */
 700         ret = check_ioctl_command(command);
 701         if (ret < 0)
 702                 return (int)ret;
 703
 704         return (int)dispatch_ioctl_command(command, arg);
 705 }
 706
 707 #ifdef CONFIG_COMPAT            /* CONFIG_COMPAT is in .config */
 708
 709 /*  Compat structure for the ORANGEFS_DEV_MAP ioctl */
 710 struct ORANGEFS_dev_map_desc32 {
 711         compat_uptr_t ptr;
 712         __s32 total_size;
 713         __s32 size;
 714         __s32 count;
 715 };
 716
 717 static unsigned long translate_dev_map26(unsigned long args, long *error)
 718 {
 719         struct ORANGEFS_dev_map_desc32 __user *p32 = (void __user *)args;
 720         /*
 721          * Depending on the architecture, allocate some space on the
 722          * user-call-stack based on our expected layout.
 723          */
 724         struct ORANGEFS_dev_map_desc __user *p =
 725             compat_alloc_user_space(sizeof(*p));
 726         compat_uptr_t addr;
 727
 728         *error = 0;
 729         /* get the ptr from the 32 bit user-space */
 730         if (get_user(addr, &p32->ptr))
 731                 goto err;
 732         /* try to put that into a 64-bit layout */
 733         if (put_user(compat_ptr(addr), &p->ptr))
 734                 goto err;
 735         /* copy the remaining fields */
 736         if (copy_in_user(&p->total_size, &p32->total_size, sizeof(__s32)))
 737                 goto err;
 738         if (copy_in_user(&p->size, &p32->size, sizeof(__s32)))
 739                 goto err;
 740         if (copy_in_user(&p->count, &p32->count, sizeof(__s32)))
 741                 goto err;
 742         return (unsigned long)p;
 743 err:
 744         *error = -EFAULT;
 745         return 0;
 746 }
 747
 748 /*
 749  * 32 bit user-space apps' ioctl handlers when kernel modules
 750  * is compiled as a 64 bit one
 751  */
 752 static long orangefs_devreq_compat_ioctl(struct file *filp, unsigned int cmd,
 753                                       unsigned long args)
 754 {
 755         long ret;
 756         unsigned long arg = args;
 757
 758         /* Check for properly constructed commands */
 759         ret = check_ioctl_command(cmd);
 760         if (ret < 0)
 761                 return ret;
 762         if (cmd == ORANGEFS_DEV_MAP) {
 763                 /*
 764                  * convert the arguments to what we expect internally
 765                  * in kernel space
 766                  */
 767                 arg = translate_dev_map26(args, &ret);
 768                 if (ret < 0) {
 769                         gossip_err("Could not translate dev map\n");
 770                         return ret;
 771                 }
 772         }
 773         /* no other ioctl requires translation */
 774         return dispatch_ioctl_command(cmd, arg);
 775 }
 776
 777 #endif /* CONFIG_COMPAT is in .config */
 778
 779 /* the assigned character device major number */
 780 static int orangefs_dev_major;
 781
 782 /*
 783  * Initialize orangefs device specific state:
 784  * Must be called at module load time only
 785  */
 786 int orangefs_dev_init(void)
 787 {
 788         /* register orangefs-req device  */
 789         orangefs_dev_major = register_chrdev(0,
 790                                           ORANGEFS_REQDEVICE_NAME,
 791                                           &orangefs_devreq_file_operations);
 792         if (orangefs_dev_major < 0) {
 793                 gossip_debug(GOSSIP_DEV_DEBUG,
 794                              "Failed to register /dev/%s (error %d)\n",
 795                              ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
 796                 return orangefs_dev_major;
 797         }
 798
 799         gossip_debug(GOSSIP_DEV_DEBUG,
 800                      "*** /dev/%s character device registered ***\n",
 801                      ORANGEFS_REQDEVICE_NAME);
 802         gossip_debug(GOSSIP_DEV_DEBUG, "'mknod /dev/%s c %d 0'.\n",
 803                      ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
 804         return 0;
 805 }
 806
 807 void orangefs_dev_cleanup(void)
 808 {
 809         unregister_chrdev(orangefs_dev_major, ORANGEFS_REQDEVICE_NAME);
 810         gossip_debug(GOSSIP_DEV_DEBUG,
 811                      "*** /dev/%s character device unregistered ***\n",
 812                      ORANGEFS_REQDEVICE_NAME);
 813 }
 814
 815 static unsigned int orangefs_devreq_poll(struct file *file,
 816                                       struct poll_table_struct *poll_table)
 817 {
 818         int poll_revent_mask = 0;
 819
 820         poll_wait(file, &orangefs_request_list_waitq, poll_table);
 821
 822         if (!list_empty(&orangefs_request_list))
 823                 poll_revent_mask |= POLL_IN;
 824         return poll_revent_mask;
 825 }
 826
 827 const struct file_operations orangefs_devreq_file_operations = {
 828         .owner = THIS_MODULE,
 829         .read = orangefs_devreq_read,
 830         .write_iter = orangefs_devreq_write_iter,
 831         .open = orangefs_devreq_open,
 832         .release = orangefs_devreq_release,
 833         .unlocked_ioctl = orangefs_devreq_ioctl,
 834
 835 #ifdef CONFIG_COMPAT            /* CONFIG_COMPAT is in .config */
 836         .compat_ioctl = orangefs_devreq_compat_ioctl,
 837 #endif
 838         .poll = orangefs_devreq_poll
 839 };