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.
 212                          */
 213                         /* XXX: is there a better way to detect this? */
 214                         } else if (ret == -1 &&
 215                                    !(op->upcall.type ==
 216                                         ORANGEFS_VFS_OP_FS_MOUNT ||
 217                                      op->upcall.type ==
 218                                         ORANGEFS_VFS_OP_GETATTR)) {
 219                                 gossip_debug(GOSSIP_DEV_DEBUG,
 220                                     "orangefs: skipping op tag %llu %s\n",
 221                                     llu(op->tag), get_opname_string(op));
 222                                 gossip_err(
 223                                     "orangefs: ERROR: fs_mount_pending %d\n",
 224                                     fsid);
 225                                 spin_unlock(&op->lock);
 226                                 continue;
 227                         }
 228                 }
 229                 /*
 230                  * Either this op does not pertain to a filesystem, is mounting
 231                  * a filesystem, or pertains to a mounted filesystem. Let it
 232                  * through.
 233                  */
 234                 cur_op = op;
 235                 break;
 236         }
 237
 238         /*
 239          * At this point we either have a valid op and can continue or have not
 240          * found an op and must ask the client to try again later.
 241          */
 242         if (!cur_op) {
 243                 spin_unlock(&orangefs_request_list_lock);
 244                 return -EAGAIN;
 245         }
 246
 247         gossip_debug(GOSSIP_DEV_DEBUG, "%s: reading op tag %llu %s\n",
 248                      __func__,
 249                      llu(cur_op->tag),
 250                      get_opname_string(cur_op));
 251
 252         /*
 253          * Such an op should never be on the list in the first place. If so, we
 254          * will abort.
 255          */
 256         if (op_state_in_progress(cur_op) || op_state_serviced(cur_op)) {
 257                 gossip_err("orangefs: ERROR: Current op already queued.\n");
 258                 list_del_init(&cur_op->list);
 259                 spin_unlock(&cur_op->lock);
 260                 spin_unlock(&orangefs_request_list_lock);
 261                 return -EAGAIN;
 262         }
 263
 264         list_del_init(&cur_op->list);
 265         spin_unlock(&orangefs_request_list_lock);
 266
 267         spin_unlock(&cur_op->lock);
 268
 269         /* Push the upcall out. */
 270         ret = copy_to_user(buf, &proto_ver, sizeof(__s32));
 271         if (ret != 0)
 272                 goto error;
 273         ret = copy_to_user(buf+sizeof(__s32), &magic, sizeof(__s32));
 274         if (ret != 0)
 275                 goto error;
 276         ret = copy_to_user(buf+2 * sizeof(__s32), &cur_op->tag, sizeof(__u64));
 277         if (ret != 0)
 278                 goto error;
 279         ret = copy_to_user(buf+2*sizeof(__s32)+sizeof(__u64), &cur_op->upcall,
 280                            sizeof(struct orangefs_upcall_s));
 281         if (ret != 0)
 282                 goto error;
 283
 284         spin_lock(&orangefs_htable_ops_in_progress_lock);
 285         spin_lock(&cur_op->lock);
 286         if (unlikely(op_state_given_up(cur_op))) {
 287                 spin_unlock(&cur_op->lock);
 288                 spin_unlock(&orangefs_htable_ops_in_progress_lock);
 289                 complete(&cur_op->waitq);
 290                 goto restart;
 291         }
 292
 293         /*
 294          * Set the operation to be in progress and move it between lists since
 295          * it has been sent to the client.
 296          */
 297         set_op_state_inprogress(cur_op);
 298         gossip_debug(GOSSIP_DEV_DEBUG,
 299                      "%s: 1 op:%s: op_state:%d: process:%s:\n",
 300                      __func__,
 301                      get_opname_string(cur_op),
 302                      cur_op->op_state,
 303                      current->comm);
 304         orangefs_devreq_add_op(cur_op);
 305         spin_unlock(&cur_op->lock);
 306         spin_unlock(&orangefs_htable_ops_in_progress_lock);
 307
 308         /* The client only asks to read one size buffer. */
 309         return MAX_DEV_REQ_UPSIZE;
 310 error:
 311         /*
 312          * We were unable to copy the op data to the client. Put the op back in
 313          * list. If client has crashed, the op will be purged later when the
 314          * device is released.
 315          */
 316         gossip_err("orangefs: Failed to copy data to user space\n");
 317         spin_lock(&orangefs_request_list_lock);
 318         spin_lock(&cur_op->lock);
 319         if (likely(!op_state_given_up(cur_op))) {
 320                 set_op_state_waiting(cur_op);
 321                 gossip_debug(GOSSIP_DEV_DEBUG,
 322                              "%s: 2 op:%s: op_state:%d: process:%s:\n",
 323                              __func__,
 324                              get_opname_string(cur_op),
 325                              cur_op->op_state,
 326                              current->comm);
 327                 list_add(&cur_op->list, &orangefs_request_list);
 328                 spin_unlock(&cur_op->lock);
 329         } else {
 330                 spin_unlock(&cur_op->lock);
 331                 complete(&cur_op->waitq);
 332         }
 333         spin_unlock(&orangefs_request_list_lock);
 334         return -EFAULT;
 335 }
 336
 337 /*
 338  * Function for writev() callers into the device.
 339  *
 340  * Userspace should have written:
 341  *  - __u32 version
 342  *  - __u32 magic
 343  *  - __u64 tag
 344  *  - struct orangefs_downcall_s
 345  *  - trailer buffer (in the case of READDIR operations)
 346  */
 347 static ssize_t orangefs_devreq_write_iter(struct kiocb *iocb,
 348                                       struct iov_iter *iter)
 349 {
 350         ssize_t ret;
 351         struct orangefs_kernel_op_s *op = NULL;
 352         struct {
 353                 __u32 version;
 354                 __u32 magic;
 355                 __u64 tag;
 356         } head;
 357         int total = ret = iov_iter_count(iter);
 358         int downcall_size = sizeof(struct orangefs_downcall_s);
 359         int head_size = sizeof(head);
 360
 361         gossip_debug(GOSSIP_DEV_DEBUG, "%s: total:%d: ret:%zd:\n",
 362                      __func__,
 363                      total,
 364                      ret);
 365
 366         if (total < MAX_DEV_REQ_DOWNSIZE) {
 367                 gossip_err("%s: total:%d: must be at least:%u:\n",
 368                            __func__,
 369                            total,
 370                            (unsigned int) MAX_DEV_REQ_DOWNSIZE);
 371                 return -EFAULT;
 372         }
 373
 374         if (!copy_from_iter_full(&head, head_size, iter)) {
 375                 gossip_err("%s: failed to copy head.\n", __func__);
 376                 return -EFAULT;
 377         }
 378
 379         if (head.version < ORANGEFS_MINIMUM_USERSPACE_VERSION) {
 380                 gossip_err("%s: userspace claims version"
 381                            "%d, minimum version required: %d.\n",
 382                            __func__,
 383                            head.version,
 384                            ORANGEFS_MINIMUM_USERSPACE_VERSION);
 385                 return -EPROTO;
 386         }
 387
 388         if (head.magic != ORANGEFS_DEVREQ_MAGIC) {
 389                 gossip_err("Error: Device magic number does not match.\n");
 390                 return -EPROTO;
 391         }
 392
 393         if (!orangefs_userspace_version) {
 394                 orangefs_userspace_version = head.version;
 395         } else if (orangefs_userspace_version != head.version) {
 396                 gossip_err("Error: userspace version changes\n");
 397                 return -EPROTO;
 398         }
 399
 400         /* remove the op from the in progress hash table */
 401         op = orangefs_devreq_remove_op(head.tag);
 402         if (!op) {
 403                 gossip_err("WARNING: No one's waiting for tag %llu\n",
 404                            llu(head.tag));
 405                 return ret;
 406         }
 407
 408         if (!copy_from_iter_full(&op->downcall, downcall_size, iter)) {
 409                 gossip_err("%s: failed to copy downcall.\n", __func__);
 410                 goto Efault;
 411         }
 412
 413         if (op->downcall.status)
 414                 goto wakeup;
 415
 416         /*
 417          * We've successfully peeled off the head and the downcall.
 418          * Something has gone awry if total doesn't equal the
 419          * sum of head_size, downcall_size and trailer_size.
 420          */
 421         if ((head_size + downcall_size + op->downcall.trailer_size) != total) {
 422                 gossip_err("%s: funky write, head_size:%d"
 423                            ": downcall_size:%d: trailer_size:%lld"
 424                            ": total size:%d:\n",
 425                            __func__,
 426                            head_size,
 427                            downcall_size,
 428                            op->downcall.trailer_size,
 429                            total);
 430                 goto Efault;
 431         }
 432
 433         /* Only READDIR operations should have trailers. */
 434         if ((op->downcall.type != ORANGEFS_VFS_OP_READDIR) &&
 435             (op->downcall.trailer_size != 0)) {
 436                 gossip_err("%s: %x operation with trailer.",
 437                            __func__,
 438                            op->downcall.type);
 439                 goto Efault;
 440         }
 441
 442         /* READDIR operations should always have trailers. */
 443         if ((op->downcall.type == ORANGEFS_VFS_OP_READDIR) &&
 444             (op->downcall.trailer_size == 0)) {
 445                 gossip_err("%s: %x operation with no trailer.",
 446                            __func__,
 447                            op->downcall.type);
 448                 goto Efault;
 449         }
 450
 451         if (op->downcall.type != ORANGEFS_VFS_OP_READDIR)
 452                 goto wakeup;
 453
 454         op->downcall.trailer_buf =
 455                 vmalloc(op->downcall.trailer_size);
 456         if (op->downcall.trailer_buf == NULL) {
 457                 gossip_err("%s: failed trailer vmalloc.\n",
 458                            __func__);
 459                 goto Enomem;
 460         }
 461         memset(op->downcall.trailer_buf, 0, op->downcall.trailer_size);
 462         if (!copy_from_iter_full(op->downcall.trailer_buf,
 463                                  op->downcall.trailer_size, iter)) {
 464                 gossip_err("%s: failed to copy trailer.\n", __func__);
 465                 vfree(op->downcall.trailer_buf);
 466                 goto Efault;
 467         }
 468
 469 wakeup:
 470         /*
 471          * Return to vfs waitqueue, and back to service_operation
 472          * through wait_for_matching_downcall.
 473          */
 474         spin_lock(&op->lock);
 475         if (unlikely(op_is_cancel(op))) {
 476                 spin_unlock(&op->lock);
 477                 put_cancel(op);
 478         } else if (unlikely(op_state_given_up(op))) {
 479                 spin_unlock(&op->lock);
 480                 complete(&op->waitq);
 481         } else {
 482                 set_op_state_serviced(op);
 483                 gossip_debug(GOSSIP_DEV_DEBUG,
 484                              "%s: op:%s: op_state:%d: process:%s:\n",
 485                              __func__,
 486                              get_opname_string(op),
 487                              op->op_state,
 488                              current->comm);
 489                 spin_unlock(&op->lock);
 490         }
 491         return ret;
 492
 493 Efault:
 494         op->downcall.status = -(ORANGEFS_ERROR_BIT | 9);
 495         ret = -EFAULT;
 496         goto wakeup;
 497
 498 Enomem:
 499         op->downcall.status = -(ORANGEFS_ERROR_BIT | 8);
 500         ret = -ENOMEM;
 501         goto wakeup;
 502 }
 503
 504 /*
 505  * NOTE: gets called when the last reference to this device is dropped.
 506  * Using the open_access_count variable, we enforce a reference count
 507  * on this file so that it can be opened by only one process at a time.
 508  * the devreq_mutex is used to make sure all i/o has completed
 509  * before we call orangefs_bufmap_finalize, and similar such tricky
 510  * situations
 511  */
 512 static int orangefs_devreq_release(struct inode *inode, struct file *file)
 513 {
 514         int unmounted = 0;
 515
 516         gossip_debug(GOSSIP_DEV_DEBUG,
 517                      "%s:pvfs2-client-core: exiting, closing device\n",
 518                      __func__);
 519
 520         mutex_lock(&devreq_mutex);
 521         orangefs_bufmap_finalize();
 522
 523         open_access_count = -1;
 524
 525         unmounted = mark_all_pending_mounts();
 526         gossip_debug(GOSSIP_DEV_DEBUG, "ORANGEFS Device Close: Filesystem(s) %s\n",
 527                      (unmounted ? "UNMOUNTED" : "MOUNTED"));
 528
 529         purge_waiting_ops();
 530         purge_inprogress_ops();
 531
 532         orangefs_bufmap_run_down();
 533
 534         gossip_debug(GOSSIP_DEV_DEBUG,
 535                      "pvfs2-client-core: device close complete\n");
 536         open_access_count = 0;
 537         orangefs_userspace_version = 0;
 538         mutex_unlock(&devreq_mutex);
 539         return 0;
 540 }
 541
 542 int is_daemon_in_service(void)
 543 {
 544         int in_service;
 545
 546         /*
 547          * What this function does is checks if client-core is alive
 548          * based on the access count we maintain on the device.
 549          */
 550         mutex_lock(&devreq_mutex);
 551         in_service = open_access_count == 1 ? 0 : -EIO;
 552         mutex_unlock(&devreq_mutex);
 553         return in_service;
 554 }
 555
 556 bool __is_daemon_in_service(void)
 557 {
 558         return open_access_count == 1;
 559 }
 560
 561 static inline long check_ioctl_command(unsigned int command)
 562 {
 563         /* Check for valid ioctl codes */
 564         if (_IOC_TYPE(command) != ORANGEFS_DEV_MAGIC) {
 565                 gossip_err("device ioctl magic numbers don't match! Did you rebuild pvfs2-client-core/libpvfs2? [cmd %x, magic %x != %x]\n",
 566                         command,
 567                         _IOC_TYPE(command),
 568                         ORANGEFS_DEV_MAGIC);
 569                 return -EINVAL;
 570         }
 571         /* and valid ioctl commands */
 572         if (_IOC_NR(command) >= ORANGEFS_DEV_MAXNR || _IOC_NR(command) <= 0) {
 573                 gossip_err("Invalid ioctl command number [%d >= %d]\n",
 574                            _IOC_NR(command), ORANGEFS_DEV_MAXNR);
 575                 return -ENOIOCTLCMD;
 576         }
 577         return 0;
 578 }
 579
 580 static long dispatch_ioctl_command(unsigned int command, unsigned long arg)
 581 {
 582         static __s32 magic = ORANGEFS_DEVREQ_MAGIC;
 583         static __s32 max_up_size = MAX_DEV_REQ_UPSIZE;
 584         static __s32 max_down_size = MAX_DEV_REQ_DOWNSIZE;
 585         struct ORANGEFS_dev_map_desc user_desc;
 586         int ret = 0;
 587         int upstream_kmod = 1;
 588         struct orangefs_sb_info_s *orangefs_sb;
 589
 590         /* mtmoore: add locking here */
 591
 592         switch (command) {
 593         case ORANGEFS_DEV_GET_MAGIC:
 594                 return ((put_user(magic, (__s32 __user *) arg) == -EFAULT) ?
 595                         -EIO :
 596                         0);
 597         case ORANGEFS_DEV_GET_MAX_UPSIZE:
 598                 return ((put_user(max_up_size,
 599                                   (__s32 __user *) arg) == -EFAULT) ?
 600                                         -EIO :
 601                                         0);
 602         case ORANGEFS_DEV_GET_MAX_DOWNSIZE:
 603                 return ((put_user(max_down_size,
 604                                   (__s32 __user *) arg) == -EFAULT) ?
 605                                         -EIO :
 606                                         0);
 607         case ORANGEFS_DEV_MAP:
 608                 ret = copy_from_user(&user_desc,
 609                                      (struct ORANGEFS_dev_map_desc __user *)
 610                                      arg,
 611                                      sizeof(struct ORANGEFS_dev_map_desc));
 612                 /* WTF -EIO and not -EFAULT? */
 613                 return ret ? -EIO : orangefs_bufmap_initialize(&user_desc);
 614         case ORANGEFS_DEV_REMOUNT_ALL:
 615                 gossip_debug(GOSSIP_DEV_DEBUG,
 616                              "%s: got ORANGEFS_DEV_REMOUNT_ALL\n",
 617                              __func__);
 618
 619                 /*
 620                  * remount all mounted orangefs volumes to regain the lost
 621                  * dynamic mount tables (if any) -- NOTE: this is done
 622                  * without keeping the superblock list locked due to the
 623                  * upcall/downcall waiting.  also, the request mutex is
 624                  * used to ensure that no operations will be serviced until
 625                  * all of the remounts are serviced (to avoid ops between
 626                  * mounts to fail)
 627                  */
 628                 ret = mutex_lock_interruptible(&orangefs_request_mutex);
 629                 if (ret < 0)
 630                         return ret;
 631                 gossip_debug(GOSSIP_DEV_DEBUG,
 632                              "%s: priority remount in progress\n",
 633                              __func__);
 634                 spin_lock(&orangefs_superblocks_lock);
 635                 list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
 636                         /*
 637                          * We have to drop the spinlock, so entries can be
 638                          * removed.  They can't be freed, though, so we just
 639                          * keep the forward pointers and zero the back ones -
 640                          * that way we can get to the rest of the list.
 641                          */
 642                         if (!orangefs_sb->list.prev)
 643                                 continue;
 644                         gossip_debug(GOSSIP_DEV_DEBUG,
 645                                      "%s: Remounting SB %p\n",
 646                                      __func__,
 647                                      orangefs_sb);
 648
 649                         spin_unlock(&orangefs_superblocks_lock);
 650                         ret = orangefs_remount(orangefs_sb);
 651                         spin_lock(&orangefs_superblocks_lock);
 652                         if (ret) {
 653                                 gossip_debug(GOSSIP_DEV_DEBUG,
 654                                              "SB %p remount failed\n",
 655                                              orangefs_sb);
 656                                 break;
 657                         }
 658                 }
 659                 spin_unlock(&orangefs_superblocks_lock);
 660                 gossip_debug(GOSSIP_DEV_DEBUG,
 661                              "%s: priority remount complete\n",
 662                              __func__);
 663                 mutex_unlock(&orangefs_request_mutex);
 664                 return ret;
 665
 666         case ORANGEFS_DEV_UPSTREAM:
 667                 ret = copy_to_user((void __user *)arg,
 668                                     &upstream_kmod,
 669                                     sizeof(upstream_kmod));
 670
 671                 if (ret != 0)
 672                         return -EIO;
 673                 else
 674                         return ret;
 675
 676         case ORANGEFS_DEV_CLIENT_MASK:
 677                 return orangefs_debugfs_new_client_mask((void __user *)arg);
 678         case ORANGEFS_DEV_CLIENT_STRING:
 679                 return orangefs_debugfs_new_client_string((void __user *)arg);
 680         case ORANGEFS_DEV_DEBUG:
 681                 return orangefs_debugfs_new_debug((void __user *)arg);
 682         default:
 683                 return -ENOIOCTLCMD;
 684         }
 685         return -ENOIOCTLCMD;
 686 }
 687
 688 static long orangefs_devreq_ioctl(struct file *file,
 689                                unsigned int command, unsigned long arg)
 690 {
 691         long ret;
 692
 693         /* Check for properly constructed commands */
 694         ret = check_ioctl_command(command);
 695         if (ret < 0)
 696                 return (int)ret;
 697
 698         return (int)dispatch_ioctl_command(command, arg);
 699 }
 700
 701 #ifdef CONFIG_COMPAT            /* CONFIG_COMPAT is in .config */
 702
 703 /*  Compat structure for the ORANGEFS_DEV_MAP ioctl */
 704 struct ORANGEFS_dev_map_desc32 {
 705         compat_uptr_t ptr;
 706         __s32 total_size;
 707         __s32 size;
 708         __s32 count;
 709 };
 710
 711 static unsigned long translate_dev_map26(unsigned long args, long *error)
 712 {
 713         struct ORANGEFS_dev_map_desc32 __user *p32 = (void __user *)args;
 714         /*
 715          * Depending on the architecture, allocate some space on the
 716          * user-call-stack based on our expected layout.
 717          */
 718         struct ORANGEFS_dev_map_desc __user *p =
 719             compat_alloc_user_space(sizeof(*p));
 720         compat_uptr_t addr;
 721
 722         *error = 0;
 723         /* get the ptr from the 32 bit user-space */
 724         if (get_user(addr, &p32->ptr))
 725                 goto err;
 726         /* try to put that into a 64-bit layout */
 727         if (put_user(compat_ptr(addr), &p->ptr))
 728                 goto err;
 729         /* copy the remaining fields */
 730         if (copy_in_user(&p->total_size, &p32->total_size, sizeof(__s32)))
 731                 goto err;
 732         if (copy_in_user(&p->size, &p32->size, sizeof(__s32)))
 733                 goto err;
 734         if (copy_in_user(&p->count, &p32->count, sizeof(__s32)))
 735                 goto err;
 736         return (unsigned long)p;
 737 err:
 738         *error = -EFAULT;
 739         return 0;
 740 }
 741
 742 /*
 743  * 32 bit user-space apps' ioctl handlers when kernel modules
 744  * is compiled as a 64 bit one
 745  */
 746 static long orangefs_devreq_compat_ioctl(struct file *filp, unsigned int cmd,
 747                                       unsigned long args)
 748 {
 749         long ret;
 750         unsigned long arg = args;
 751
 752         /* Check for properly constructed commands */
 753         ret = check_ioctl_command(cmd);
 754         if (ret < 0)
 755                 return ret;
 756         if (cmd == ORANGEFS_DEV_MAP) {
 757                 /*
 758                  * convert the arguments to what we expect internally
 759                  * in kernel space
 760                  */
 761                 arg = translate_dev_map26(args, &ret);
 762                 if (ret < 0) {
 763                         gossip_err("Could not translate dev map\n");
 764                         return ret;
 765                 }
 766         }
 767         /* no other ioctl requires translation */
 768         return dispatch_ioctl_command(cmd, arg);
 769 }
 770
 771 #endif /* CONFIG_COMPAT is in .config */
 772
 773 /* the assigned character device major number */
 774 static int orangefs_dev_major;
 775
 776 /*
 777  * Initialize orangefs device specific state:
 778  * Must be called at module load time only
 779  */
 780 int orangefs_dev_init(void)
 781 {
 782         /* register orangefs-req device  */
 783         orangefs_dev_major = register_chrdev(0,
 784                                           ORANGEFS_REQDEVICE_NAME,
 785                                           &orangefs_devreq_file_operations);
 786         if (orangefs_dev_major < 0) {
 787                 gossip_debug(GOSSIP_DEV_DEBUG,
 788                              "Failed to register /dev/%s (error %d)\n",
 789                              ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
 790                 return orangefs_dev_major;
 791         }
 792
 793         gossip_debug(GOSSIP_DEV_DEBUG,
 794                      "*** /dev/%s character device registered ***\n",
 795                      ORANGEFS_REQDEVICE_NAME);
 796         gossip_debug(GOSSIP_DEV_DEBUG, "'mknod /dev/%s c %d 0'.\n",
 797                      ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
 798         return 0;
 799 }
 800
 801 void orangefs_dev_cleanup(void)
 802 {
 803         unregister_chrdev(orangefs_dev_major, ORANGEFS_REQDEVICE_NAME);
 804         gossip_debug(GOSSIP_DEV_DEBUG,
 805                      "*** /dev/%s character device unregistered ***\n",
 806                      ORANGEFS_REQDEVICE_NAME);
 807 }
 808
 809 static unsigned int orangefs_devreq_poll(struct file *file,
 810                                       struct poll_table_struct *poll_table)
 811 {
 812         int poll_revent_mask = 0;
 813
 814         poll_wait(file, &orangefs_request_list_waitq, poll_table);
 815
 816         if (!list_empty(&orangefs_request_list))
 817                 poll_revent_mask |= POLL_IN;
 818         return poll_revent_mask;
 819 }
 820
 821 const struct file_operations orangefs_devreq_file_operations = {
 822         .owner = THIS_MODULE,
 823         .read = orangefs_devreq_read,
 824         .write_iter = orangefs_devreq_write_iter,
 825         .open = orangefs_devreq_open,
 826         .release = orangefs_devreq_release,
 827         .unlocked_ioctl = orangefs_devreq_ioctl,
 828
 829 #ifdef CONFIG_COMPAT            /* CONFIG_COMPAT is in .config */
 830         .compat_ioctl = orangefs_devreq_compat_ioctl,
 831 #endif
 832         .poll = orangefs_devreq_poll
 833 };