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 n;
 359         int downcall_size = sizeof(struct orangefs_downcall_s);
 360         int head_size = sizeof(head);
 361
 362         gossip_debug(GOSSIP_DEV_DEBUG, "%s: total:%d: ret:%zd:\n",
 363                      __func__,
 364                      total,
 365                      ret);
 366
 367         if (total < MAX_DEV_REQ_DOWNSIZE) {
 368                 gossip_err("%s: total:%d: must be at least:%u:\n",
 369                            __func__,
 370                            total,
 371                            (unsigned int) MAX_DEV_REQ_DOWNSIZE);
 372                 return -EFAULT;
 373         }
 374
 375         n = copy_from_iter(&head, head_size, iter);
 376         if (n < head_size) {
 377                 gossip_err("%s: failed to copy head.\n", __func__);
 378                 return -EFAULT;
 379         }
 380
 381         if (head.version < ORANGEFS_MINIMUM_USERSPACE_VERSION) {
 382                 gossip_err("%s: userspace claims version"
 383                            "%d, minimum version required: %d.\n",
 384                            __func__,
 385                            head.version,
 386                            ORANGEFS_MINIMUM_USERSPACE_VERSION);
 387                 return -EPROTO;
 388         }
 389
 390         if (head.magic != ORANGEFS_DEVREQ_MAGIC) {
 391                 gossip_err("Error: Device magic number does not match.\n");
 392                 return -EPROTO;
 393         }
 394
 395         if (!orangefs_userspace_version) {
 396                 orangefs_userspace_version = head.version;
 397         } else if (orangefs_userspace_version != head.version) {
 398                 gossip_err("Error: userspace version changes\n");
 399                 return -EPROTO;
 400         }
 401
 402         /* remove the op from the in progress hash table */
 403         op = orangefs_devreq_remove_op(head.tag);
 404         if (!op) {
 405                 gossip_err("WARNING: No one's waiting for tag %llu\n",
 406                            llu(head.tag));
 407                 return ret;
 408         }
 409
 410         n = copy_from_iter(&op->downcall, downcall_size, iter);
 411         if (n != downcall_size) {
 412                 gossip_err("%s: failed to copy downcall.\n", __func__);
 413                 goto Efault;
 414         }
 415
 416         if (op->downcall.status)
 417                 goto wakeup;
 418
 419         /*
 420          * We've successfully peeled off the head and the downcall.
 421          * Something has gone awry if total doesn't equal the
 422          * sum of head_size, downcall_size and trailer_size.
 423          */
 424         if ((head_size + downcall_size + op->downcall.trailer_size) != total) {
 425                 gossip_err("%s: funky write, head_size:%d"
 426                            ": downcall_size:%d: trailer_size:%lld"
 427                            ": total size:%d:\n",
 428                            __func__,
 429                            head_size,
 430                            downcall_size,
 431                            op->downcall.trailer_size,
 432                            total);
 433                 goto Efault;
 434         }
 435
 436         /* Only READDIR operations should have trailers. */
 437         if ((op->downcall.type != ORANGEFS_VFS_OP_READDIR) &&
 438             (op->downcall.trailer_size != 0)) {
 439                 gossip_err("%s: %x operation with trailer.",
 440                            __func__,
 441                            op->downcall.type);
 442                 goto Efault;
 443         }
 444
 445         /* READDIR operations should always have trailers. */
 446         if ((op->downcall.type == ORANGEFS_VFS_OP_READDIR) &&
 447             (op->downcall.trailer_size == 0)) {
 448                 gossip_err("%s: %x operation with no trailer.",
 449                            __func__,
 450                            op->downcall.type);
 451                 goto Efault;
 452         }
 453
 454         if (op->downcall.type != ORANGEFS_VFS_OP_READDIR)
 455                 goto wakeup;
 456
 457         op->downcall.trailer_buf =
 458                 vmalloc(op->downcall.trailer_size);
 459         if (op->downcall.trailer_buf == NULL) {
 460                 gossip_err("%s: failed trailer vmalloc.\n",
 461                            __func__);
 462                 goto Enomem;
 463         }
 464         memset(op->downcall.trailer_buf, 0, op->downcall.trailer_size);
 465         n = copy_from_iter(op->downcall.trailer_buf,
 466                            op->downcall.trailer_size,
 467                            iter);
 468         if (n != op->downcall.trailer_size) {
 469                 gossip_err("%s: failed to copy trailer.\n", __func__);
 470                 vfree(op->downcall.trailer_buf);
 471                 goto Efault;
 472         }
 473
 474 wakeup:
 475         /*
 476          * Return to vfs waitqueue, and back to service_operation
 477          * through wait_for_matching_downcall.
 478          */
 479         spin_lock(&op->lock);
 480         if (unlikely(op_is_cancel(op))) {
 481                 spin_unlock(&op->lock);
 482                 put_cancel(op);
 483         } else if (unlikely(op_state_given_up(op))) {
 484                 spin_unlock(&op->lock);
 485                 complete(&op->waitq);
 486         } else {
 487                 set_op_state_serviced(op);
 488                 gossip_debug(GOSSIP_DEV_DEBUG,
 489                              "%s: op:%s: op_state:%d: process:%s:\n",
 490                              __func__,
 491                              get_opname_string(op),
 492                              op->op_state,
 493                              current->comm);
 494                 spin_unlock(&op->lock);
 495         }
 496         return ret;
 497
 498 Efault:
 499         op->downcall.status = -(ORANGEFS_ERROR_BIT | 9);
 500         ret = -EFAULT;
 501         goto wakeup;
 502
 503 Enomem:
 504         op->downcall.status = -(ORANGEFS_ERROR_BIT | 8);
 505         ret = -ENOMEM;
 506         goto wakeup;
 507 }
 508
 509 /*
 510  * NOTE: gets called when the last reference to this device is dropped.
 511  * Using the open_access_count variable, we enforce a reference count
 512  * on this file so that it can be opened by only one process at a time.
 513  * the devreq_mutex is used to make sure all i/o has completed
 514  * before we call orangefs_bufmap_finalize, and similar such tricky
 515  * situations
 516  */
 517 static int orangefs_devreq_release(struct inode *inode, struct file *file)
 518 {
 519         int unmounted = 0;
 520
 521         gossip_debug(GOSSIP_DEV_DEBUG,
 522                      "%s:pvfs2-client-core: exiting, closing device\n",
 523                      __func__);
 524
 525         mutex_lock(&devreq_mutex);
 526         orangefs_bufmap_finalize();
 527
 528         open_access_count = -1;
 529
 530         unmounted = mark_all_pending_mounts();
 531         gossip_debug(GOSSIP_DEV_DEBUG, "ORANGEFS Device Close: Filesystem(s) %s\n",
 532                      (unmounted ? "UNMOUNTED" : "MOUNTED"));
 533
 534         purge_waiting_ops();
 535         purge_inprogress_ops();
 536
 537         orangefs_bufmap_run_down();
 538
 539         gossip_debug(GOSSIP_DEV_DEBUG,
 540                      "pvfs2-client-core: device close complete\n");
 541         open_access_count = 0;
 542         orangefs_userspace_version = 0;
 543         mutex_unlock(&devreq_mutex);
 544         return 0;
 545 }
 546
 547 int is_daemon_in_service(void)
 548 {
 549         int in_service;
 550
 551         /*
 552          * What this function does is checks if client-core is alive
 553          * based on the access count we maintain on the device.
 554          */
 555         mutex_lock(&devreq_mutex);
 556         in_service = open_access_count == 1 ? 0 : -EIO;
 557         mutex_unlock(&devreq_mutex);
 558         return in_service;
 559 }
 560
 561 bool __is_daemon_in_service(void)
 562 {
 563         return open_access_count == 1;
 564 }
 565
 566 static inline long check_ioctl_command(unsigned int command)
 567 {
 568         /* Check for valid ioctl codes */
 569         if (_IOC_TYPE(command) != ORANGEFS_DEV_MAGIC) {
 570                 gossip_err("device ioctl magic numbers don't match! Did you rebuild pvfs2-client-core/libpvfs2? [cmd %x, magic %x != %x]\n",
 571                         command,
 572                         _IOC_TYPE(command),
 573                         ORANGEFS_DEV_MAGIC);
 574                 return -EINVAL;
 575         }
 576         /* and valid ioctl commands */
 577         if (_IOC_NR(command) >= ORANGEFS_DEV_MAXNR || _IOC_NR(command) <= 0) {
 578                 gossip_err("Invalid ioctl command number [%d >= %d]\n",
 579                            _IOC_NR(command), ORANGEFS_DEV_MAXNR);
 580                 return -ENOIOCTLCMD;
 581         }
 582         return 0;
 583 }
 584
 585 static long dispatch_ioctl_command(unsigned int command, unsigned long arg)
 586 {
 587         static __s32 magic = ORANGEFS_DEVREQ_MAGIC;
 588         static __s32 max_up_size = MAX_DEV_REQ_UPSIZE;
 589         static __s32 max_down_size = MAX_DEV_REQ_DOWNSIZE;
 590         struct ORANGEFS_dev_map_desc user_desc;
 591         int ret = 0;
 592         int upstream_kmod = 1;
 593         struct orangefs_sb_info_s *orangefs_sb;
 594
 595         /* mtmoore: add locking here */
 596
 597         switch (command) {
 598         case ORANGEFS_DEV_GET_MAGIC:
 599                 return ((put_user(magic, (__s32 __user *) arg) == -EFAULT) ?
 600                         -EIO :
 601                         0);
 602         case ORANGEFS_DEV_GET_MAX_UPSIZE:
 603                 return ((put_user(max_up_size,
 604                                   (__s32 __user *) arg) == -EFAULT) ?
 605                                         -EIO :
 606                                         0);
 607         case ORANGEFS_DEV_GET_MAX_DOWNSIZE:
 608                 return ((put_user(max_down_size,
 609                                   (__s32 __user *) arg) == -EFAULT) ?
 610                                         -EIO :
 611                                         0);
 612         case ORANGEFS_DEV_MAP:
 613                 ret = copy_from_user(&user_desc,
 614                                      (struct ORANGEFS_dev_map_desc __user *)
 615                                      arg,
 616                                      sizeof(struct ORANGEFS_dev_map_desc));
 617                 /* WTF -EIO and not -EFAULT? */
 618                 return ret ? -EIO : orangefs_bufmap_initialize(&user_desc);
 619         case ORANGEFS_DEV_REMOUNT_ALL:
 620                 gossip_debug(GOSSIP_DEV_DEBUG,
 621                              "%s: got ORANGEFS_DEV_REMOUNT_ALL\n",
 622                              __func__);
 623
 624                 /*
 625                  * remount all mounted orangefs volumes to regain the lost
 626                  * dynamic mount tables (if any) -- NOTE: this is done
 627                  * without keeping the superblock list locked due to the
 628                  * upcall/downcall waiting.  also, the request mutex is
 629                  * used to ensure that no operations will be serviced until
 630                  * all of the remounts are serviced (to avoid ops between
 631                  * mounts to fail)
 632                  */
 633                 ret = mutex_lock_interruptible(&orangefs_request_mutex);
 634                 if (ret < 0)
 635                         return ret;
 636                 gossip_debug(GOSSIP_DEV_DEBUG,
 637                              "%s: priority remount in progress\n",
 638                              __func__);
 639                 spin_lock(&orangefs_superblocks_lock);
 640                 list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
 641                         /*
 642                          * We have to drop the spinlock, so entries can be
 643                          * removed.  They can't be freed, though, so we just
 644                          * keep the forward pointers and zero the back ones -
 645                          * that way we can get to the rest of the list.
 646                          */
 647                         if (!orangefs_sb->list.prev)
 648                                 continue;
 649                         gossip_debug(GOSSIP_DEV_DEBUG,
 650                                      "%s: Remounting SB %p\n",
 651                                      __func__,
 652                                      orangefs_sb);
 653
 654                         spin_unlock(&orangefs_superblocks_lock);
 655                         ret = orangefs_remount(orangefs_sb);
 656                         spin_lock(&orangefs_superblocks_lock);
 657                         if (ret) {
 658                                 gossip_debug(GOSSIP_DEV_DEBUG,
 659                                              "SB %p remount failed\n",
 660                                              orangefs_sb);
 661                                 break;
 662                         }
 663                 }
 664                 spin_unlock(&orangefs_superblocks_lock);
 665                 gossip_debug(GOSSIP_DEV_DEBUG,
 666                              "%s: priority remount complete\n",
 667                              __func__);
 668                 mutex_unlock(&orangefs_request_mutex);
 669                 return ret;
 670
 671         case ORANGEFS_DEV_UPSTREAM:
 672                 ret = copy_to_user((void __user *)arg,
 673                                     &upstream_kmod,
 674                                     sizeof(upstream_kmod));
 675
 676                 if (ret != 0)
 677                         return -EIO;
 678                 else
 679                         return ret;
 680
 681         case ORANGEFS_DEV_CLIENT_MASK:
 682                 return orangefs_debugfs_new_client_mask((void __user *)arg);
 683         case ORANGEFS_DEV_CLIENT_STRING:
 684                 return orangefs_debugfs_new_client_string((void __user *)arg);
 685         case ORANGEFS_DEV_DEBUG:
 686                 return orangefs_debugfs_new_debug((void __user *)arg);
 687         default:
 688                 return -ENOIOCTLCMD;
 689         }
 690         return -ENOIOCTLCMD;
 691 }
 692
 693 static long orangefs_devreq_ioctl(struct file *file,
 694                                unsigned int command, unsigned long arg)
 695 {
 696         long ret;
 697
 698         /* Check for properly constructed commands */
 699         ret = check_ioctl_command(command);
 700         if (ret < 0)
 701                 return (int)ret;
 702
 703         return (int)dispatch_ioctl_command(command, arg);
 704 }
 705
 706 #ifdef CONFIG_COMPAT            /* CONFIG_COMPAT is in .config */
 707
 708 /*  Compat structure for the ORANGEFS_DEV_MAP ioctl */
 709 struct ORANGEFS_dev_map_desc32 {
 710         compat_uptr_t ptr;
 711         __s32 total_size;
 712         __s32 size;
 713         __s32 count;
 714 };
 715
 716 static unsigned long translate_dev_map26(unsigned long args, long *error)
 717 {
 718         struct ORANGEFS_dev_map_desc32 __user *p32 = (void __user *)args;
 719         /*
 720          * Depending on the architecture, allocate some space on the
 721          * user-call-stack based on our expected layout.
 722          */
 723         struct ORANGEFS_dev_map_desc __user *p =
 724             compat_alloc_user_space(sizeof(*p));
 725         compat_uptr_t addr;
 726
 727         *error = 0;
 728         /* get the ptr from the 32 bit user-space */
 729         if (get_user(addr, &p32->ptr))
 730                 goto err;
 731         /* try to put that into a 64-bit layout */
 732         if (put_user(compat_ptr(addr), &p->ptr))
 733                 goto err;
 734         /* copy the remaining fields */
 735         if (copy_in_user(&p->total_size, &p32->total_size, sizeof(__s32)))
 736                 goto err;
 737         if (copy_in_user(&p->size, &p32->size, sizeof(__s32)))
 738                 goto err;
 739         if (copy_in_user(&p->count, &p32->count, sizeof(__s32)))
 740                 goto err;
 741         return (unsigned long)p;
 742 err:
 743         *error = -EFAULT;
 744         return 0;
 745 }
 746
 747 /*
 748  * 32 bit user-space apps' ioctl handlers when kernel modules
 749  * is compiled as a 64 bit one
 750  */
 751 static long orangefs_devreq_compat_ioctl(struct file *filp, unsigned int cmd,
 752                                       unsigned long args)
 753 {
 754         long ret;
 755         unsigned long arg = args;
 756
 757         /* Check for properly constructed commands */
 758         ret = check_ioctl_command(cmd);
 759         if (ret < 0)
 760                 return ret;
 761         if (cmd == ORANGEFS_DEV_MAP) {
 762                 /*
 763                  * convert the arguments to what we expect internally
 764                  * in kernel space
 765                  */
 766                 arg = translate_dev_map26(args, &ret);
 767                 if (ret < 0) {
 768                         gossip_err("Could not translate dev map\n");
 769                         return ret;
 770                 }
 771         }
 772         /* no other ioctl requires translation */
 773         return dispatch_ioctl_command(cmd, arg);
 774 }
 775
 776 #endif /* CONFIG_COMPAT is in .config */
 777
 778 /* the assigned character device major number */
 779 static int orangefs_dev_major;
 780
 781 /*
 782  * Initialize orangefs device specific state:
 783  * Must be called at module load time only
 784  */
 785 int orangefs_dev_init(void)
 786 {
 787         /* register orangefs-req device  */
 788         orangefs_dev_major = register_chrdev(0,
 789                                           ORANGEFS_REQDEVICE_NAME,
 790                                           &orangefs_devreq_file_operations);
 791         if (orangefs_dev_major < 0) {
 792                 gossip_debug(GOSSIP_DEV_DEBUG,
 793                              "Failed to register /dev/%s (error %d)\n",
 794                              ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
 795                 return orangefs_dev_major;
 796         }
 797
 798         gossip_debug(GOSSIP_DEV_DEBUG,
 799                      "*** /dev/%s character device registered ***\n",
 800                      ORANGEFS_REQDEVICE_NAME);
 801         gossip_debug(GOSSIP_DEV_DEBUG, "'mknod /dev/%s c %d 0'.\n",
 802                      ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
 803         return 0;
 804 }
 805
 806 void orangefs_dev_cleanup(void)
 807 {
 808         unregister_chrdev(orangefs_dev_major, ORANGEFS_REQDEVICE_NAME);
 809         gossip_debug(GOSSIP_DEV_DEBUG,
 810                      "*** /dev/%s character device unregistered ***\n",
 811                      ORANGEFS_REQDEVICE_NAME);
 812 }
 813
 814 static unsigned int orangefs_devreq_poll(struct file *file,
 815                                       struct poll_table_struct *poll_table)
 816 {
 817         int poll_revent_mask = 0;
 818
 819         poll_wait(file, &orangefs_request_list_waitq, poll_table);
 820
 821         if (!list_empty(&orangefs_request_list))
 822                 poll_revent_mask |= POLL_IN;
 823         return poll_revent_mask;
 824 }
 825
 826 const struct file_operations orangefs_devreq_file_operations = {
 827         .owner = THIS_MODULE,
 828         .read = orangefs_devreq_read,
 829         .write_iter = orangefs_devreq_write_iter,
 830         .open = orangefs_devreq_open,
 831         .release = orangefs_devreq_release,
 832         .unlocked_ioctl = orangefs_devreq_ioctl,
 833
 834 #ifdef CONFIG_COMPAT            /* CONFIG_COMPAT is in .config */
 835         .compat_ioctl = orangefs_devreq_compat_ioctl,
 836 #endif
 837         .poll = orangefs_devreq_poll
 838 };