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         /* Check for an empty list before locking. */
 184         if (list_empty(&orangefs_request_list))
 185                 return -EAGAIN;
 186
 187 restart:
 188         /* Get next op (if any) from top of list. */
 189         spin_lock(&orangefs_request_list_lock);
 190         list_for_each_entry_safe(op, temp, &orangefs_request_list, list) {
 191                 __s32 fsid;
 192                 /* This lock is held past the end of the loop when we break. */
 193                 spin_lock(&op->lock);
 194                 if (unlikely(op_state_purged(op) || op_state_given_up(op))) {
 195                         spin_unlock(&op->lock);
 196                         continue;
 197                 }
 198
 199                 fsid = fsid_of_op(op);
 200                 if (fsid != ORANGEFS_FS_ID_NULL) {
 201                         int ret;
 202                         /* Skip ops whose filesystem needs to be mounted. */
 203                         ret = fs_mount_pending(fsid);
 204                         if (ret == 1) {
 205                                 gossip_debug(GOSSIP_DEV_DEBUG,
 206                                     "%s: mount pending, skipping op tag "
 207                                     "%llu %s\n",
 208                                     __func__,
 209                                     llu(op->tag),
 210                                     get_opname_string(op));
 211                                 spin_unlock(&op->lock);
 212                                 continue;
 213                         /*
 214                          * Skip ops whose filesystem we don't know about unless
 215                          * it is being mounted or unmounted.  It is possible for
 216                          * a filesystem we don't know about to be unmounted if
 217                          * it fails to mount in the kernel after userspace has
 218                          * been sent the mount request.
 219                          */
 220                         /* XXX: is there a better way to detect this? */
 221                         } else if (ret == -1 &&
 222                                    !(op->upcall.type ==
 223                                         ORANGEFS_VFS_OP_FS_MOUNT ||
 224                                      op->upcall.type ==
 225                                         ORANGEFS_VFS_OP_GETATTR ||
 226                                      op->upcall.type ==
 227                                         ORANGEFS_VFS_OP_FS_UMOUNT)) {
 228                                 gossip_debug(GOSSIP_DEV_DEBUG,
 229                                     "orangefs: skipping op tag %llu %s\n",
 230                                     llu(op->tag), get_opname_string(op));
 231                                 gossip_err(
 232                                     "orangefs: ERROR: fs_mount_pending %d\n",
 233                                     fsid);
 234                                 spin_unlock(&op->lock);
 235                                 continue;
 236                         }
 237                 }
 238                 /*
 239                  * Either this op does not pertain to a filesystem, is mounting
 240                  * a filesystem, or pertains to a mounted filesystem. Let it
 241                  * through.
 242                  */
 243                 cur_op = op;
 244                 break;
 245         }
 246
 247         /*
 248          * At this point we either have a valid op and can continue or have not
 249          * found an op and must ask the client to try again later.
 250          */
 251         if (!cur_op) {
 252                 spin_unlock(&orangefs_request_list_lock);
 253                 return -EAGAIN;
 254         }
 255
 256         gossip_debug(GOSSIP_DEV_DEBUG, "%s: reading op tag %llu %s\n",
 257                      __func__,
 258                      llu(cur_op->tag),
 259                      get_opname_string(cur_op));
 260
 261         /*
 262          * Such an op should never be on the list in the first place. If so, we
 263          * will abort.
 264          */
 265         if (op_state_in_progress(cur_op) || op_state_serviced(cur_op)) {
 266                 gossip_err("orangefs: ERROR: Current op already queued.\n");
 267                 list_del_init(&cur_op->list);
 268                 spin_unlock(&cur_op->lock);
 269                 spin_unlock(&orangefs_request_list_lock);
 270                 return -EAGAIN;
 271         }
 272
 273         list_del_init(&cur_op->list);
 274         spin_unlock(&orangefs_request_list_lock);
 275
 276         spin_unlock(&cur_op->lock);
 277
 278         /* Push the upcall out. */
 279         ret = copy_to_user(buf, &proto_ver, sizeof(__s32));
 280         if (ret != 0)
 281                 goto error;
 282         ret = copy_to_user(buf+sizeof(__s32), &magic, sizeof(__s32));
 283         if (ret != 0)
 284                 goto error;
 285         ret = copy_to_user(buf+2 * sizeof(__s32), &cur_op->tag, sizeof(__u64));
 286         if (ret != 0)
 287                 goto error;
 288         ret = copy_to_user(buf+2*sizeof(__s32)+sizeof(__u64), &cur_op->upcall,
 289                            sizeof(struct orangefs_upcall_s));
 290         if (ret != 0)
 291                 goto error;
 292
 293         spin_lock(&orangefs_htable_ops_in_progress_lock);
 294         spin_lock(&cur_op->lock);
 295         if (unlikely(op_state_given_up(cur_op))) {
 296                 spin_unlock(&cur_op->lock);
 297                 spin_unlock(&orangefs_htable_ops_in_progress_lock);
 298                 complete(&cur_op->waitq);
 299                 goto restart;
 300         }
 301
 302         /*
 303          * Set the operation to be in progress and move it between lists since
 304          * it has been sent to the client.
 305          */
 306         set_op_state_inprogress(cur_op);
 307         gossip_debug(GOSSIP_DEV_DEBUG,
 308                      "%s: 1 op:%s: op_state:%d: process:%s:\n",
 309                      __func__,
 310                      get_opname_string(cur_op),
 311                      cur_op->op_state,
 312                      current->comm);
 313         orangefs_devreq_add_op(cur_op);
 314         spin_unlock(&cur_op->lock);
 315         spin_unlock(&orangefs_htable_ops_in_progress_lock);
 316
 317         /* The client only asks to read one size buffer. */
 318         return MAX_DEV_REQ_UPSIZE;
 319 error:
 320         /*
 321          * We were unable to copy the op data to the client. Put the op back in
 322          * list. If client has crashed, the op will be purged later when the
 323          * device is released.
 324          */
 325         gossip_err("orangefs: Failed to copy data to user space\n");
 326         spin_lock(&orangefs_request_list_lock);
 327         spin_lock(&cur_op->lock);
 328         if (likely(!op_state_given_up(cur_op))) {
 329                 set_op_state_waiting(cur_op);
 330                 gossip_debug(GOSSIP_DEV_DEBUG,
 331                              "%s: 2 op:%s: op_state:%d: process:%s:\n",
 332                              __func__,
 333                              get_opname_string(cur_op),
 334                              cur_op->op_state,
 335                              current->comm);
 336                 list_add(&cur_op->list, &orangefs_request_list);
 337                 spin_unlock(&cur_op->lock);
 338         } else {
 339                 spin_unlock(&cur_op->lock);
 340                 complete(&cur_op->waitq);
 341         }
 342         spin_unlock(&orangefs_request_list_lock);
 343         return -EFAULT;
 344 }
 345
 346 /*
 347  * Function for writev() callers into the device.
 348  *
 349  * Userspace should have written:
 350  *  - __u32 version
 351  *  - __u32 magic
 352  *  - __u64 tag
 353  *  - struct orangefs_downcall_s
 354  *  - trailer buffer (in the case of READDIR operations)
 355  */
 356 static ssize_t orangefs_devreq_write_iter(struct kiocb *iocb,
 357                                       struct iov_iter *iter)
 358 {
 359         ssize_t ret;
 360         struct orangefs_kernel_op_s *op = NULL;
 361         struct {
 362                 __u32 version;
 363                 __u32 magic;
 364                 __u64 tag;
 365         } head;
 366         int total = ret = iov_iter_count(iter);
 367         int downcall_size = sizeof(struct orangefs_downcall_s);
 368         int head_size = sizeof(head);
 369
 370         gossip_debug(GOSSIP_DEV_DEBUG, "%s: total:%d: ret:%zd:\n",
 371                      __func__,
 372                      total,
 373                      ret);
 374
 375         if (total < MAX_DEV_REQ_DOWNSIZE) {
 376                 gossip_err("%s: total:%d: must be at least:%u:\n",
 377                            __func__,
 378                            total,
 379                            (unsigned int) MAX_DEV_REQ_DOWNSIZE);
 380                 return -EFAULT;
 381         }
 382
 383         if (!copy_from_iter_full(&head, head_size, iter)) {
 384                 gossip_err("%s: failed to copy head.\n", __func__);
 385                 return -EFAULT;
 386         }
 387
 388         if (head.version < ORANGEFS_MINIMUM_USERSPACE_VERSION) {
 389                 gossip_err("%s: userspace claims version"
 390                            "%d, minimum version required: %d.\n",
 391                            __func__,
 392                            head.version,
 393                            ORANGEFS_MINIMUM_USERSPACE_VERSION);
 394                 return -EPROTO;
 395         }
 396
 397         if (head.magic != ORANGEFS_DEVREQ_MAGIC) {
 398                 gossip_err("Error: Device magic number does not match.\n");
 399                 return -EPROTO;
 400         }
 401
 402         if (!orangefs_userspace_version) {
 403                 orangefs_userspace_version = head.version;
 404         } else if (orangefs_userspace_version != head.version) {
 405                 gossip_err("Error: userspace version changes\n");
 406                 return -EPROTO;
 407         }
 408
 409         /* remove the op from the in progress hash table */
 410         op = orangefs_devreq_remove_op(head.tag);
 411         if (!op) {
 412                 gossip_debug(GOSSIP_DEV_DEBUG,
 413                              "%s: No one's waiting for tag %llu\n",
 414                              __func__, llu(head.tag));
 415                 return ret;
 416         }
 417
 418         if (!copy_from_iter_full(&op->downcall, downcall_size, iter)) {
 419                 gossip_err("%s: failed to copy downcall.\n", __func__);
 420                 goto Efault;
 421         }
 422
 423         if (op->downcall.status)
 424                 goto wakeup;
 425
 426         /*
 427          * We've successfully peeled off the head and the downcall.
 428          * Something has gone awry if total doesn't equal the
 429          * sum of head_size, downcall_size and trailer_size.
 430          */
 431         if ((head_size + downcall_size + op->downcall.trailer_size) != total) {
 432                 gossip_err("%s: funky write, head_size:%d"
 433                            ": downcall_size:%d: trailer_size:%lld"
 434                            ": total size:%d:\n",
 435                            __func__,
 436                            head_size,
 437                            downcall_size,
 438                            op->downcall.trailer_size,
 439                            total);
 440                 goto Efault;
 441         }
 442
 443         /* Only READDIR operations should have trailers. */
 444         if ((op->downcall.type != ORANGEFS_VFS_OP_READDIR) &&
 445             (op->downcall.trailer_size != 0)) {
 446                 gossip_err("%s: %x operation with trailer.",
 447                            __func__,
 448                            op->downcall.type);
 449                 goto Efault;
 450         }
 451
 452         /* READDIR operations should always have trailers. */
 453         if ((op->downcall.type == ORANGEFS_VFS_OP_READDIR) &&
 454             (op->downcall.trailer_size == 0)) {
 455                 gossip_err("%s: %x operation with no trailer.",
 456                            __func__,
 457                            op->downcall.type);
 458                 goto Efault;
 459         }
 460
 461         if (op->downcall.type != ORANGEFS_VFS_OP_READDIR)
 462                 goto wakeup;
 463
 464         op->downcall.trailer_buf =
 465                 vmalloc(op->downcall.trailer_size);
 466         if (op->downcall.trailer_buf == NULL) {
 467                 gossip_err("%s: failed trailer vmalloc.\n",
 468                            __func__);
 469                 goto Enomem;
 470         }
 471         memset(op->downcall.trailer_buf, 0, op->downcall.trailer_size);
 472         if (!copy_from_iter_full(op->downcall.trailer_buf,
 473                                  op->downcall.trailer_size, iter)) {
 474                 gossip_err("%s: failed to copy trailer.\n", __func__);
 475                 vfree(op->downcall.trailer_buf);
 476                 goto Efault;
 477         }
 478
 479 wakeup:
 480         /*
 481          * Return to vfs waitqueue, and back to service_operation
 482          * through wait_for_matching_downcall.
 483          */
 484         spin_lock(&op->lock);
 485         if (unlikely(op_is_cancel(op))) {
 486                 spin_unlock(&op->lock);
 487                 put_cancel(op);
 488         } else if (unlikely(op_state_given_up(op))) {
 489                 spin_unlock(&op->lock);
 490                 complete(&op->waitq);
 491         } else {
 492                 set_op_state_serviced(op);
 493                 gossip_debug(GOSSIP_DEV_DEBUG,
 494                              "%s: op:%s: op_state:%d: process:%s:\n",
 495                              __func__,
 496                              get_opname_string(op),
 497                              op->op_state,
 498                              current->comm);
 499                 spin_unlock(&op->lock);
 500         }
 501         return ret;
 502
 503 Efault:
 504         op->downcall.status = -(ORANGEFS_ERROR_BIT | 9);
 505         ret = -EFAULT;
 506         goto wakeup;
 507
 508 Enomem:
 509         op->downcall.status = -(ORANGEFS_ERROR_BIT | 8);
 510         ret = -ENOMEM;
 511         goto wakeup;
 512 }
 513
 514 /*
 515  * NOTE: gets called when the last reference to this device is dropped.
 516  * Using the open_access_count variable, we enforce a reference count
 517  * on this file so that it can be opened by only one process at a time.
 518  * the devreq_mutex is used to make sure all i/o has completed
 519  * before we call orangefs_bufmap_finalize, and similar such tricky
 520  * situations
 521  */
 522 static int orangefs_devreq_release(struct inode *inode, struct file *file)
 523 {
 524         int unmounted = 0;
 525
 526         gossip_debug(GOSSIP_DEV_DEBUG,
 527                      "%s:pvfs2-client-core: exiting, closing device\n",
 528                      __func__);
 529
 530         mutex_lock(&devreq_mutex);
 531         orangefs_bufmap_finalize();
 532
 533         open_access_count = -1;
 534
 535         unmounted = mark_all_pending_mounts();
 536         gossip_debug(GOSSIP_DEV_DEBUG, "ORANGEFS Device Close: Filesystem(s) %s\n",
 537                      (unmounted ? "UNMOUNTED" : "MOUNTED"));
 538
 539         purge_waiting_ops();
 540         purge_inprogress_ops();
 541
 542         orangefs_bufmap_run_down();
 543
 544         gossip_debug(GOSSIP_DEV_DEBUG,
 545                      "pvfs2-client-core: device close complete\n");
 546         open_access_count = 0;
 547         orangefs_userspace_version = 0;
 548         mutex_unlock(&devreq_mutex);
 549         return 0;
 550 }
 551
 552 int is_daemon_in_service(void)
 553 {
 554         int in_service;
 555
 556         /*
 557          * What this function does is checks if client-core is alive
 558          * based on the access count we maintain on the device.
 559          */
 560         mutex_lock(&devreq_mutex);
 561         in_service = open_access_count == 1 ? 0 : -EIO;
 562         mutex_unlock(&devreq_mutex);
 563         return in_service;
 564 }
 565
 566 bool __is_daemon_in_service(void)
 567 {
 568         return open_access_count == 1;
 569 }
 570
 571 static inline long check_ioctl_command(unsigned int command)
 572 {
 573         /* Check for valid ioctl codes */
 574         if (_IOC_TYPE(command) != ORANGEFS_DEV_MAGIC) {
 575                 gossip_err("device ioctl magic numbers don't match! Did you rebuild pvfs2-client-core/libpvfs2? [cmd %x, magic %x != %x]\n",
 576                         command,
 577                         _IOC_TYPE(command),
 578                         ORANGEFS_DEV_MAGIC);
 579                 return -EINVAL;
 580         }
 581         /* and valid ioctl commands */
 582         if (_IOC_NR(command) >= ORANGEFS_DEV_MAXNR || _IOC_NR(command) <= 0) {
 583                 gossip_err("Invalid ioctl command number [%d >= %d]\n",
 584                            _IOC_NR(command), ORANGEFS_DEV_MAXNR);
 585                 return -ENOIOCTLCMD;
 586         }
 587         return 0;
 588 }
 589
 590 static long dispatch_ioctl_command(unsigned int command, unsigned long arg)
 591 {
 592         static __s32 magic = ORANGEFS_DEVREQ_MAGIC;
 593         static __s32 max_up_size = MAX_DEV_REQ_UPSIZE;
 594         static __s32 max_down_size = MAX_DEV_REQ_DOWNSIZE;
 595         struct ORANGEFS_dev_map_desc user_desc;
 596         int ret = 0;
 597         int upstream_kmod = 1;
 598         struct orangefs_sb_info_s *orangefs_sb;
 599
 600         /* mtmoore: add locking here */
 601
 602         switch (command) {
 603         case ORANGEFS_DEV_GET_MAGIC:
 604                 return ((put_user(magic, (__s32 __user *) arg) == -EFAULT) ?
 605                         -EIO :
 606                         0);
 607         case ORANGEFS_DEV_GET_MAX_UPSIZE:
 608                 return ((put_user(max_up_size,
 609                                   (__s32 __user *) arg) == -EFAULT) ?
 610                                         -EIO :
 611                                         0);
 612         case ORANGEFS_DEV_GET_MAX_DOWNSIZE:
 613                 return ((put_user(max_down_size,
 614                                   (__s32 __user *) arg) == -EFAULT) ?
 615                                         -EIO :
 616                                         0);
 617         case ORANGEFS_DEV_MAP:
 618                 ret = copy_from_user(&user_desc,
 619                                      (struct ORANGEFS_dev_map_desc __user *)
 620                                      arg,
 621                                      sizeof(struct ORANGEFS_dev_map_desc));
 622                 /* WTF -EIO and not -EFAULT? */
 623                 return ret ? -EIO : orangefs_bufmap_initialize(&user_desc);
 624         case ORANGEFS_DEV_REMOUNT_ALL:
 625                 gossip_debug(GOSSIP_DEV_DEBUG,
 626                              "%s: got ORANGEFS_DEV_REMOUNT_ALL\n",
 627                              __func__);
 628
 629                 /*
 630                  * remount all mounted orangefs volumes to regain the lost
 631                  * dynamic mount tables (if any) -- NOTE: this is done
 632                  * without keeping the superblock list locked due to the
 633                  * upcall/downcall waiting.  also, the request mutex is
 634                  * used to ensure that no operations will be serviced until
 635                  * all of the remounts are serviced (to avoid ops between
 636                  * mounts to fail)
 637                  */
 638                 ret = mutex_lock_interruptible(&orangefs_request_mutex);
 639                 if (ret < 0)
 640                         return ret;
 641                 gossip_debug(GOSSIP_DEV_DEBUG,
 642                              "%s: priority remount in progress\n",
 643                              __func__);
 644                 spin_lock(&orangefs_superblocks_lock);
 645                 list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
 646                         /*
 647                          * We have to drop the spinlock, so entries can be
 648                          * removed.  They can't be freed, though, so we just
 649                          * keep the forward pointers and zero the back ones -
 650                          * that way we can get to the rest of the list.
 651                          */
 652                         if (!orangefs_sb->list.prev)
 653                                 continue;
 654                         gossip_debug(GOSSIP_DEV_DEBUG,
 655                                      "%s: Remounting SB %p\n",
 656                                      __func__,
 657                                      orangefs_sb);
 658
 659                         spin_unlock(&orangefs_superblocks_lock);
 660                         ret = orangefs_remount(orangefs_sb);
 661                         spin_lock(&orangefs_superblocks_lock);
 662                         if (ret) {
 663                                 gossip_debug(GOSSIP_DEV_DEBUG,
 664                                              "SB %p remount failed\n",
 665                                              orangefs_sb);
 666                                 break;
 667                         }
 668                 }
 669                 spin_unlock(&orangefs_superblocks_lock);
 670                 gossip_debug(GOSSIP_DEV_DEBUG,
 671                              "%s: priority remount complete\n",
 672                              __func__);
 673                 mutex_unlock(&orangefs_request_mutex);
 674                 return ret;
 675
 676         case ORANGEFS_DEV_UPSTREAM:
 677                 ret = copy_to_user((void __user *)arg,
 678                                     &upstream_kmod,
 679                                     sizeof(upstream_kmod));
 680
 681                 if (ret != 0)
 682                         return -EIO;
 683                 else
 684                         return ret;
 685
 686         case ORANGEFS_DEV_CLIENT_MASK:
 687                 return orangefs_debugfs_new_client_mask((void __user *)arg);
 688         case ORANGEFS_DEV_CLIENT_STRING:
 689                 return orangefs_debugfs_new_client_string((void __user *)arg);
 690         case ORANGEFS_DEV_DEBUG:
 691                 return orangefs_debugfs_new_debug((void __user *)arg);
 692         default:
 693                 return -ENOIOCTLCMD;
 694         }
 695         return -ENOIOCTLCMD;
 696 }
 697
 698 static long orangefs_devreq_ioctl(struct file *file,
 699                                unsigned int command, unsigned long arg)
 700 {
 701         long ret;
 702
 703         /* Check for properly constructed commands */
 704         ret = check_ioctl_command(command);
 705         if (ret < 0)
 706                 return (int)ret;
 707
 708         return (int)dispatch_ioctl_command(command, arg);
 709 }
 710
 711 #ifdef CONFIG_COMPAT            /* CONFIG_COMPAT is in .config */
 712
 713 /*  Compat structure for the ORANGEFS_DEV_MAP ioctl */
 714 struct ORANGEFS_dev_map_desc32 {
 715         compat_uptr_t ptr;
 716         __s32 total_size;
 717         __s32 size;
 718         __s32 count;
 719 };
 720
 721 static unsigned long translate_dev_map26(unsigned long args, long *error)
 722 {
 723         struct ORANGEFS_dev_map_desc32 __user *p32 = (void __user *)args;
 724         /*
 725          * Depending on the architecture, allocate some space on the
 726          * user-call-stack based on our expected layout.
 727          */
 728         struct ORANGEFS_dev_map_desc __user *p =
 729             compat_alloc_user_space(sizeof(*p));
 730         compat_uptr_t addr;
 731
 732         *error = 0;
 733         /* get the ptr from the 32 bit user-space */
 734         if (get_user(addr, &p32->ptr))
 735                 goto err;
 736         /* try to put that into a 64-bit layout */
 737         if (put_user(compat_ptr(addr), &p->ptr))
 738                 goto err;
 739         /* copy the remaining fields */
 740         if (copy_in_user(&p->total_size, &p32->total_size, sizeof(__s32)))
 741                 goto err;
 742         if (copy_in_user(&p->size, &p32->size, sizeof(__s32)))
 743                 goto err;
 744         if (copy_in_user(&p->count, &p32->count, sizeof(__s32)))
 745                 goto err;
 746         return (unsigned long)p;
 747 err:
 748         *error = -EFAULT;
 749         return 0;
 750 }
 751
 752 /*
 753  * 32 bit user-space apps' ioctl handlers when kernel modules
 754  * is compiled as a 64 bit one
 755  */
 756 static long orangefs_devreq_compat_ioctl(struct file *filp, unsigned int cmd,
 757                                       unsigned long args)
 758 {
 759         long ret;
 760         unsigned long arg = args;
 761
 762         /* Check for properly constructed commands */
 763         ret = check_ioctl_command(cmd);
 764         if (ret < 0)
 765                 return ret;
 766         if (cmd == ORANGEFS_DEV_MAP) {
 767                 /*
 768                  * convert the arguments to what we expect internally
 769                  * in kernel space
 770                  */
 771                 arg = translate_dev_map26(args, &ret);
 772                 if (ret < 0) {
 773                         gossip_err("Could not translate dev map\n");
 774                         return ret;
 775                 }
 776         }
 777         /* no other ioctl requires translation */
 778         return dispatch_ioctl_command(cmd, arg);
 779 }
 780
 781 #endif /* CONFIG_COMPAT is in .config */
 782
 783 /* the assigned character device major number */
 784 static int orangefs_dev_major;
 785
 786 /*
 787  * Initialize orangefs device specific state:
 788  * Must be called at module load time only
 789  */
 790 int orangefs_dev_init(void)
 791 {
 792         /* register orangefs-req device  */
 793         orangefs_dev_major = register_chrdev(0,
 794                                           ORANGEFS_REQDEVICE_NAME,
 795                                           &orangefs_devreq_file_operations);
 796         if (orangefs_dev_major < 0) {
 797                 gossip_debug(GOSSIP_DEV_DEBUG,
 798                              "Failed to register /dev/%s (error %d)\n",
 799                              ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
 800                 return orangefs_dev_major;
 801         }
 802
 803         gossip_debug(GOSSIP_DEV_DEBUG,
 804                      "*** /dev/%s character device registered ***\n",
 805                      ORANGEFS_REQDEVICE_NAME);
 806         gossip_debug(GOSSIP_DEV_DEBUG, "'mknod /dev/%s c %d 0'.\n",
 807                      ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
 808         return 0;
 809 }
 810
 811 void orangefs_dev_cleanup(void)
 812 {
 813         unregister_chrdev(orangefs_dev_major, ORANGEFS_REQDEVICE_NAME);
 814         gossip_debug(GOSSIP_DEV_DEBUG,
 815                      "*** /dev/%s character device unregistered ***\n",
 816                      ORANGEFS_REQDEVICE_NAME);
 817 }
 818
 819 static unsigned int orangefs_devreq_poll(struct file *file,
 820                                       struct poll_table_struct *poll_table)
 821 {
 822         int poll_revent_mask = 0;
 823
 824         poll_wait(file, &orangefs_request_list_waitq, poll_table);
 825
 826         if (!list_empty(&orangefs_request_list))
 827                 poll_revent_mask |= POLL_IN;
 828         return poll_revent_mask;
 829 }
 830
 831 const struct file_operations orangefs_devreq_file_operations = {
 832         .owner = THIS_MODULE,
 833         .read = orangefs_devreq_read,
 834         .write_iter = orangefs_devreq_write_iter,
 835         .open = orangefs_devreq_open,
 836         .release = orangefs_devreq_release,
 837         .unlocked_ioctl = orangefs_devreq_ioctl,
 838
 839 #ifdef CONFIG_COMPAT            /* CONFIG_COMPAT is in .config */
 840         .compat_ioctl = orangefs_devreq_compat_ioctl,
 841 #endif
 842         .poll = orangefs_devreq_poll
 843 };