fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/namei.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/pipe_fs_i.h>
  21 #include <linux/swap.h>
  22 #include <linux/splice.h>
  23 #include <linux/freezer.h>
  24
  25 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  26 MODULE_ALIAS("devname:fuse");
  27
  28 static struct kmem_cache *fuse_req_cachep;
  29
  30 static struct fuse_dev *fuse_get_dev(struct file *file)
  31 {
  32         /*
  33          * Lockless access is OK, because file->private data is set
  34          * once during mount and is valid until the file is released.
  35          */
  36         return ACCESS_ONCE(file->private_data);
  37 }
  38
  39 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  40                               struct fuse_page_desc *page_descs,
  41                               unsigned npages)
  42 {
  43         memset(req, 0, sizeof(*req));
  44         memset(pages, 0, sizeof(*pages) * npages);
  45         memset(page_descs, 0, sizeof(*page_descs) * npages);
  46         INIT_LIST_HEAD(&req->list);
  47         INIT_LIST_HEAD(&req->intr_entry);
  48         init_waitqueue_head(&req->waitq);
  49         atomic_set(&req->count, 1);
  50         req->pages = pages;
  51         req->page_descs = page_descs;
  52         req->max_pages = npages;
  53         __set_bit(FR_PENDING, &req->flags);
  54 }
  55
  56 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  57 {
  58         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, flags);
  59         if (req) {
  60                 struct page **pages;
  61                 struct fuse_page_desc *page_descs;
  62
  63                 if (npages <= FUSE_REQ_INLINE_PAGES) {
  64                         pages = req->inline_pages;
  65                         page_descs = req->inline_page_descs;
  66                 } else {
  67                         pages = kmalloc(sizeof(struct page *) * npages, flags);
  68                         page_descs = kmalloc(sizeof(struct fuse_page_desc) *
  69                                              npages, flags);
  70                 }
  71
  72                 if (!pages || !page_descs) {
  73                         kfree(pages);
  74                         kfree(page_descs);
  75                         kmem_cache_free(fuse_req_cachep, req);
  76                         return NULL;
  77                 }
  78
  79                 fuse_request_init(req, pages, page_descs, npages);
  80         }
  81         return req;
  82 }
  83
  84 struct fuse_req *fuse_request_alloc(unsigned npages)
  85 {
  86         return __fuse_request_alloc(npages, GFP_KERNEL);
  87 }
  88 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  89
  90 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  91 {
  92         return __fuse_request_alloc(npages, GFP_NOFS);
  93 }
  94
  95 void fuse_request_free(struct fuse_req *req)
  96 {
  97         if (req->pages != req->inline_pages) {
  98                 kfree(req->pages);
  99                 kfree(req->page_descs);
 100         }
 101         kmem_cache_free(fuse_req_cachep, req);
 102 }
 103
 104 static void block_sigs(sigset_t *oldset)
 105 {
 106         sigset_t mask;
 107
 108         siginitsetinv(&mask, sigmask(SIGKILL));
 109         sigprocmask(SIG_BLOCK, &mask, oldset);
 110 }
 111
 112 static void restore_sigs(sigset_t *oldset)
 113 {
 114         sigprocmask(SIG_SETMASK, oldset, NULL);
 115 }
 116
 117 void __fuse_get_request(struct fuse_req *req)
 118 {
 119         atomic_inc(&req->count);
 120 }
 121
 122 /* Must be called with > 1 refcount */
 123 static void __fuse_put_request(struct fuse_req *req)
 124 {
 125         BUG_ON(atomic_read(&req->count) < 2);
 126         atomic_dec(&req->count);
 127 }
 128
 129 static void fuse_req_init_context(struct fuse_req *req)
 130 {
 131         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 132         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 133         req->in.h.pid = current->pid;
 134 }
 135
 136 void fuse_set_initialized(struct fuse_conn *fc)
 137 {
 138         /* Make sure stores before this are seen on another CPU */
 139         smp_wmb();
 140         fc->initialized = 1;
 141 }
 142
 143 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 144 {
 145         return !fc->initialized || (for_background && fc->blocked);
 146 }
 147
 148 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 149                                        bool for_background)
 150 {
 151         struct fuse_req *req;
 152         int err;
 153         atomic_inc(&fc->num_waiting);
 154
 155         if (fuse_block_alloc(fc, for_background)) {
 156                 sigset_t oldset;
 157                 int intr;
 158
 159                 block_sigs(&oldset);
 160                 intr = wait_event_interruptible_exclusive(fc->blocked_waitq,
 161                                 !fuse_block_alloc(fc, for_background));
 162                 restore_sigs(&oldset);
 163                 err = -EINTR;
 164                 if (intr)
 165                         goto out;
 166         }
 167         /* Matches smp_wmb() in fuse_set_initialized() */
 168         smp_rmb();
 169
 170         err = -ENOTCONN;
 171         if (!fc->connected)
 172                 goto out;
 173
 174         err = -ECONNREFUSED;
 175         if (fc->conn_error)
 176                 goto out;
 177
 178         req = fuse_request_alloc(npages);
 179         err = -ENOMEM;
 180         if (!req) {
 181                 if (for_background)
 182                         wake_up(&fc->blocked_waitq);
 183                 goto out;
 184         }
 185
 186         fuse_req_init_context(req);
 187         __set_bit(FR_WAITING, &req->flags);
 188         if (for_background)
 189                 __set_bit(FR_BACKGROUND, &req->flags);
 190
 191         return req;
 192
 193  out:
 194         atomic_dec(&fc->num_waiting);
 195         return ERR_PTR(err);
 196 }
 197
 198 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 199 {
 200         return __fuse_get_req(fc, npages, false);
 201 }
 202 EXPORT_SYMBOL_GPL(fuse_get_req);
 203
 204 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 205                                              unsigned npages)
 206 {
 207         return __fuse_get_req(fc, npages, true);
 208 }
 209 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 210
 211 /*
 212  * Return request in fuse_file->reserved_req.  However that may
 213  * currently be in use.  If that is the case, wait for it to become
 214  * available.
 215  */
 216 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 217                                          struct file *file)
 218 {
 219         struct fuse_req *req = NULL;
 220         struct fuse_file *ff = file->private_data;
 221
 222         do {
 223                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 224                 spin_lock(&fc->lock);
 225                 if (ff->reserved_req) {
 226                         req = ff->reserved_req;
 227                         ff->reserved_req = NULL;
 228                         req->stolen_file = get_file(file);
 229                 }
 230                 spin_unlock(&fc->lock);
 231         } while (!req);
 232
 233         return req;
 234 }
 235
 236 /*
 237  * Put stolen request back into fuse_file->reserved_req
 238  */
 239 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 240 {
 241         struct file *file = req->stolen_file;
 242         struct fuse_file *ff = file->private_data;
 243
 244         spin_lock(&fc->lock);
 245         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 246         BUG_ON(ff->reserved_req);
 247         ff->reserved_req = req;
 248         wake_up_all(&fc->reserved_req_waitq);
 249         spin_unlock(&fc->lock);
 250         fput(file);
 251 }
 252
 253 /*
 254  * Gets a requests for a file operation, always succeeds
 255  *
 256  * This is used for sending the FLUSH request, which must get to
 257  * userspace, due to POSIX locks which may need to be unlocked.
 258  *
 259  * If allocation fails due to OOM, use the reserved request in
 260  * fuse_file.
 261  *
 262  * This is very unlikely to deadlock accidentally, since the
 263  * filesystem should not have it's own file open.  If deadlock is
 264  * intentional, it can still be broken by "aborting" the filesystem.
 265  */
 266 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 267                                              struct file *file)
 268 {
 269         struct fuse_req *req;
 270
 271         atomic_inc(&fc->num_waiting);
 272         wait_event(fc->blocked_waitq, fc->initialized);
 273         /* Matches smp_wmb() in fuse_set_initialized() */
 274         smp_rmb();
 275         req = fuse_request_alloc(0);
 276         if (!req)
 277                 req = get_reserved_req(fc, file);
 278
 279         fuse_req_init_context(req);
 280         __set_bit(FR_WAITING, &req->flags);
 281         __clear_bit(FR_BACKGROUND, &req->flags);
 282         return req;
 283 }
 284
 285 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 286 {
 287         if (atomic_dec_and_test(&req->count)) {
 288                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 289                         /*
 290                          * We get here in the unlikely case that a background
 291                          * request was allocated but not sent
 292                          */
 293                         spin_lock(&fc->lock);
 294                         if (!fc->blocked)
 295                                 wake_up(&fc->blocked_waitq);
 296                         spin_unlock(&fc->lock);
 297                 }
 298
 299                 if (test_bit(FR_WAITING, &req->flags)) {
 300                         __clear_bit(FR_WAITING, &req->flags);
 301                         atomic_dec(&fc->num_waiting);
 302                 }
 303
 304                 if (req->stolen_file)
 305                         put_reserved_req(fc, req);
 306                 else
 307                         fuse_request_free(req);
 308         }
 309 }
 310 EXPORT_SYMBOL_GPL(fuse_put_request);
 311
 312 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 313 {
 314         unsigned nbytes = 0;
 315         unsigned i;
 316
 317         for (i = 0; i < numargs; i++)
 318                 nbytes += args[i].size;
 319
 320         return nbytes;
 321 }
 322
 323 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 324 {
 325         return ++fiq->reqctr;
 326 }
 327
 328 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 329 {
 330         req->in.h.len = sizeof(struct fuse_in_header) +
 331                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 332         list_add_tail(&req->list, &fiq->pending);
 333         wake_up_locked(&fiq->waitq);
 334         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 335 }
 336
 337 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 338                        u64 nodeid, u64 nlookup)
 339 {
 340         struct fuse_iqueue *fiq = &fc->iq;
 341
 342         forget->forget_one.nodeid = nodeid;
 343         forget->forget_one.nlookup = nlookup;
 344
 345         spin_lock(&fiq->waitq.lock);
 346         if (fiq->connected) {
 347                 fiq->forget_list_tail->next = forget;
 348                 fiq->forget_list_tail = forget;
 349                 wake_up_locked(&fiq->waitq);
 350                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 351         } else {
 352                 kfree(forget);
 353         }
 354         spin_unlock(&fiq->waitq.lock);
 355 }
 356
 357 static void flush_bg_queue(struct fuse_conn *fc)
 358 {
 359         while (fc->active_background < fc->max_background &&
 360                !list_empty(&fc->bg_queue)) {
 361                 struct fuse_req *req;
 362                 struct fuse_iqueue *fiq = &fc->iq;
 363
 364                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 365                 list_del(&req->list);
 366                 fc->active_background++;
 367                 spin_lock(&fiq->waitq.lock);
 368                 req->in.h.unique = fuse_get_unique(fiq);
 369                 queue_request(fiq, req);
 370                 spin_unlock(&fiq->waitq.lock);
 371         }
 372 }
 373
 374 /*
 375  * This function is called when a request is finished.  Either a reply
 376  * has arrived or it was aborted (and not yet sent) or some error
 377  * occurred during communication with userspace, or the device file
 378  * was closed.  The requester thread is woken up (if still waiting),
 379  * the 'end' callback is called if given, else the reference to the
 380  * request is released
 381  */
 382 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 383 {
 384         struct fuse_iqueue *fiq = &fc->iq;
 385
 386         if (test_and_set_bit(FR_FINISHED, &req->flags))
 387                 return;
 388
 389         spin_lock(&fiq->waitq.lock);
 390         list_del_init(&req->intr_entry);
 391         spin_unlock(&fiq->waitq.lock);
 392         WARN_ON(test_bit(FR_PENDING, &req->flags));
 393         WARN_ON(test_bit(FR_SENT, &req->flags));
 394         if (test_bit(FR_BACKGROUND, &req->flags)) {
 395                 spin_lock(&fc->lock);
 396                 clear_bit(FR_BACKGROUND, &req->flags);
 397                 if (fc->num_background == fc->max_background)
 398                         fc->blocked = 0;
 399
 400                 /* Wake up next waiter, if any */
 401                 if (!fc->blocked && waitqueue_active(&fc->blocked_waitq))
 402                         wake_up(&fc->blocked_waitq);
 403
 404                 if (fc->num_background == fc->congestion_threshold &&
 405                     fc->connected && fc->bdi_initialized) {
 406                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 407                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 408                 }
 409                 fc->num_background--;
 410                 fc->active_background--;
 411                 flush_bg_queue(fc);
 412                 spin_unlock(&fc->lock);
 413         }
 414         wake_up(&req->waitq);
 415         if (req->end)
 416                 req->end(fc, req);
 417         fuse_put_request(fc, req);
 418 }
 419
 420 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 421 {
 422         spin_lock(&fiq->waitq.lock);
 423         if (list_empty(&req->intr_entry)) {
 424                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 425                 wake_up_locked(&fiq->waitq);
 426         }
 427         spin_unlock(&fiq->waitq.lock);
 428         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 429 }
 430
 431 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 432 {
 433         struct fuse_iqueue *fiq = &fc->iq;
 434         int err;
 435
 436         if (!fc->no_interrupt) {
 437                 /* Any signal may interrupt this */
 438                 err = wait_event_interruptible(req->waitq,
 439                                         test_bit(FR_FINISHED, &req->flags));
 440                 if (!err)
 441                         return;
 442
 443                 set_bit(FR_INTERRUPTED, &req->flags);
 444                 /* matches barrier in fuse_dev_do_read() */
 445                 smp_mb__after_atomic();
 446                 if (test_bit(FR_SENT, &req->flags))
 447                         queue_interrupt(fiq, req);
 448         }
 449
 450         if (!test_bit(FR_FORCE, &req->flags)) {
 451                 sigset_t oldset;
 452
 453                 /* Only fatal signals may interrupt this */
 454                 block_sigs(&oldset);
 455                 err = wait_event_interruptible(req->waitq,
 456                                         test_bit(FR_FINISHED, &req->flags));
 457                 restore_sigs(&oldset);
 458
 459                 if (!err)
 460                         return;
 461
 462                 spin_lock(&fiq->waitq.lock);
 463                 /* Request is not yet in userspace, bail out */
 464                 if (test_bit(FR_PENDING, &req->flags)) {
 465                         list_del(&req->list);
 466                         spin_unlock(&fiq->waitq.lock);
 467                         __fuse_put_request(req);
 468                         req->out.h.error = -EINTR;
 469                         return;
 470                 }
 471                 spin_unlock(&fiq->waitq.lock);
 472         }
 473
 474         /*
 475          * Either request is already in userspace, or it was forced.
 476          * Wait it out.
 477          */
 478         while (!test_bit(FR_FINISHED, &req->flags))
 479                 wait_event_freezable(req->waitq,
 480                                 test_bit(FR_FINISHED, &req->flags));
 481 }
 482
 483 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 484 {
 485         struct fuse_iqueue *fiq = &fc->iq;
 486
 487         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 488         spin_lock(&fiq->waitq.lock);
 489         if (!fiq->connected) {
 490                 spin_unlock(&fiq->waitq.lock);
 491                 req->out.h.error = -ENOTCONN;
 492         } else {
 493                 req->in.h.unique = fuse_get_unique(fiq);
 494                 queue_request(fiq, req);
 495                 /* acquire extra reference, since request is still needed
 496                    after request_end() */
 497                 __fuse_get_request(req);
 498                 spin_unlock(&fiq->waitq.lock);
 499
 500                 request_wait_answer(fc, req);
 501                 /* Pairs with smp_wmb() in request_end() */
 502                 smp_rmb();
 503         }
 504 }
 505
 506 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 507 {
 508         __set_bit(FR_ISREPLY, &req->flags);
 509         if (!test_bit(FR_WAITING, &req->flags)) {
 510                 __set_bit(FR_WAITING, &req->flags);
 511                 atomic_inc(&fc->num_waiting);
 512         }
 513         __fuse_request_send(fc, req);
 514 }
 515 EXPORT_SYMBOL_GPL(fuse_request_send);
 516
 517 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 518 {
 519         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 520                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 521
 522         if (fc->minor < 9) {
 523                 switch (args->in.h.opcode) {
 524                 case FUSE_LOOKUP:
 525                 case FUSE_CREATE:
 526                 case FUSE_MKNOD:
 527                 case FUSE_MKDIR:
 528                 case FUSE_SYMLINK:
 529                 case FUSE_LINK:
 530                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 531                         break;
 532                 case FUSE_GETATTR:
 533                 case FUSE_SETATTR:
 534                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 535                         break;
 536                 }
 537         }
 538         if (fc->minor < 12) {
 539                 switch (args->in.h.opcode) {
 540                 case FUSE_CREATE:
 541                         args->in.args[0].size = sizeof(struct fuse_open_in);
 542                         break;
 543                 case FUSE_MKNOD:
 544                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 545                         break;
 546                 }
 547         }
 548 }
 549
 550 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 551 {
 552         struct fuse_req *req;
 553         ssize_t ret;
 554
 555         req = fuse_get_req(fc, 0);
 556         if (IS_ERR(req))
 557                 return PTR_ERR(req);
 558
 559         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 560         fuse_adjust_compat(fc, args);
 561
 562         req->in.h.opcode = args->in.h.opcode;
 563         req->in.h.nodeid = args->in.h.nodeid;
 564         req->in.numargs = args->in.numargs;
 565         memcpy(req->in.args, args->in.args,
 566                args->in.numargs * sizeof(struct fuse_in_arg));
 567         req->out.argvar = args->out.argvar;
 568         req->out.numargs = args->out.numargs;
 569         memcpy(req->out.args, args->out.args,
 570                args->out.numargs * sizeof(struct fuse_arg));
 571         fuse_request_send(fc, req);
 572         ret = req->out.h.error;
 573         if (!ret && args->out.argvar) {
 574                 BUG_ON(args->out.numargs != 1);
 575                 ret = req->out.args[0].size;
 576         }
 577         fuse_put_request(fc, req);
 578
 579         return ret;
 580 }
 581
 582 /*
 583  * Called under fc->lock
 584  *
 585  * fc->connected must have been checked previously
 586  */
 587 void fuse_request_send_background_locked(struct fuse_conn *fc,
 588                                          struct fuse_req *req)
 589 {
 590         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 591         if (!test_bit(FR_WAITING, &req->flags)) {
 592                 __set_bit(FR_WAITING, &req->flags);
 593                 atomic_inc(&fc->num_waiting);
 594         }
 595         __set_bit(FR_ISREPLY, &req->flags);
 596         fc->num_background++;
 597         if (fc->num_background == fc->max_background)
 598                 fc->blocked = 1;
 599         if (fc->num_background == fc->congestion_threshold &&
 600             fc->bdi_initialized) {
 601                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 602                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 603         }
 604         list_add_tail(&req->list, &fc->bg_queue);
 605         flush_bg_queue(fc);
 606 }
 607
 608 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 609 {
 610         BUG_ON(!req->end);
 611         spin_lock(&fc->lock);
 612         if (fc->connected) {
 613                 fuse_request_send_background_locked(fc, req);
 614                 spin_unlock(&fc->lock);
 615         } else {
 616                 spin_unlock(&fc->lock);
 617                 req->out.h.error = -ENOTCONN;
 618                 req->end(fc, req);
 619                 fuse_put_request(fc, req);
 620         }
 621 }
 622 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 623
 624 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 625                                           struct fuse_req *req, u64 unique)
 626 {
 627         int err = -ENODEV;
 628         struct fuse_iqueue *fiq = &fc->iq;
 629
 630         __clear_bit(FR_ISREPLY, &req->flags);
 631         req->in.h.unique = unique;
 632         spin_lock(&fiq->waitq.lock);
 633         if (fiq->connected) {
 634                 queue_request(fiq, req);
 635                 err = 0;
 636         }
 637         spin_unlock(&fiq->waitq.lock);
 638
 639         return err;
 640 }
 641
 642 void fuse_force_forget(struct file *file, u64 nodeid)
 643 {
 644         struct inode *inode = file_inode(file);
 645         struct fuse_conn *fc = get_fuse_conn(inode);
 646         struct fuse_req *req;
 647         struct fuse_forget_in inarg;
 648
 649         memset(&inarg, 0, sizeof(inarg));
 650         inarg.nlookup = 1;
 651         req = fuse_get_req_nofail_nopages(fc, file);
 652         req->in.h.opcode = FUSE_FORGET;
 653         req->in.h.nodeid = nodeid;
 654         req->in.numargs = 1;
 655         req->in.args[0].size = sizeof(inarg);
 656         req->in.args[0].value = &inarg;
 657         __clear_bit(FR_ISREPLY, &req->flags);
 658         __fuse_request_send(fc, req);
 659         /* ignore errors */
 660         fuse_put_request(fc, req);
 661 }
 662
 663 /*
 664  * Lock the request.  Up to the next unlock_request() there mustn't be
 665  * anything that could cause a page-fault.  If the request was already
 666  * aborted bail out.
 667  */
 668 static int lock_request(struct fuse_req *req)
 669 {
 670         int err = 0;
 671         if (req) {
 672                 spin_lock(&req->waitq.lock);
 673                 if (test_bit(FR_ABORTED, &req->flags))
 674                         err = -ENOENT;
 675                 else
 676                         set_bit(FR_LOCKED, &req->flags);
 677                 spin_unlock(&req->waitq.lock);
 678         }
 679         return err;
 680 }
 681
 682 /*
 683  * Unlock request.  If it was aborted while locked, caller is responsible
 684  * for unlocking and ending the request.
 685  */
 686 static int unlock_request(struct fuse_req *req)
 687 {
 688         int err = 0;
 689         if (req) {
 690                 spin_lock(&req->waitq.lock);
 691                 if (test_bit(FR_ABORTED, &req->flags))
 692                         err = -ENOENT;
 693                 else
 694                         clear_bit(FR_LOCKED, &req->flags);
 695                 spin_unlock(&req->waitq.lock);
 696         }
 697         return err;
 698 }
 699
 700 struct fuse_copy_state {
 701         int write;
 702         struct fuse_req *req;
 703         struct iov_iter *iter;
 704         struct pipe_buffer *pipebufs;
 705         struct pipe_buffer *currbuf;
 706         struct pipe_inode_info *pipe;
 707         unsigned long nr_segs;
 708         struct page *pg;
 709         unsigned len;
 710         unsigned offset;
 711         unsigned move_pages:1;
 712 };
 713
 714 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 715                            struct iov_iter *iter)
 716 {
 717         memset(cs, 0, sizeof(*cs));
 718         cs->write = write;
 719         cs->iter = iter;
 720 }
 721
 722 /* Unmap and put previous page of userspace buffer */
 723 static void fuse_copy_finish(struct fuse_copy_state *cs)
 724 {
 725         if (cs->currbuf) {
 726                 struct pipe_buffer *buf = cs->currbuf;
 727
 728                 if (cs->write)
 729                         buf->len = PAGE_SIZE - cs->len;
 730                 cs->currbuf = NULL;
 731         } else if (cs->pg) {
 732                 if (cs->write) {
 733                         flush_dcache_page(cs->pg);
 734                         set_page_dirty_lock(cs->pg);
 735                 }
 736                 put_page(cs->pg);
 737         }
 738         cs->pg = NULL;
 739 }
 740
 741 /*
 742  * Get another pagefull of userspace buffer, and map it to kernel
 743  * address space, and lock request
 744  */
 745 static int fuse_copy_fill(struct fuse_copy_state *cs)
 746 {
 747         struct page *page;
 748         int err;
 749
 750         err = unlock_request(cs->req);
 751         if (err)
 752                 return err;
 753
 754         fuse_copy_finish(cs);
 755         if (cs->pipebufs) {
 756                 struct pipe_buffer *buf = cs->pipebufs;
 757
 758                 if (!cs->write) {
 759                         err = buf->ops->confirm(cs->pipe, buf);
 760                         if (err)
 761                                 return err;
 762
 763                         BUG_ON(!cs->nr_segs);
 764                         cs->currbuf = buf;
 765                         cs->pg = buf->page;
 766                         cs->offset = buf->offset;
 767                         cs->len = buf->len;
 768                         cs->pipebufs++;
 769                         cs->nr_segs--;
 770                 } else {
 771                         if (cs->nr_segs == cs->pipe->buffers)
 772                                 return -EIO;
 773
 774                         page = alloc_page(GFP_HIGHUSER);
 775                         if (!page)
 776                                 return -ENOMEM;
 777
 778                         buf->page = page;
 779                         buf->offset = 0;
 780                         buf->len = 0;
 781
 782                         cs->currbuf = buf;
 783                         cs->pg = page;
 784                         cs->offset = 0;
 785                         cs->len = PAGE_SIZE;
 786                         cs->pipebufs++;
 787                         cs->nr_segs++;
 788                 }
 789         } else {
 790                 size_t off;
 791                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 792                 if (err < 0)
 793                         return err;
 794                 BUG_ON(!err);
 795                 cs->len = err;
 796                 cs->offset = off;
 797                 cs->pg = page;
 798                 cs->offset = off;
 799                 iov_iter_advance(cs->iter, err);
 800         }
 801
 802         return lock_request(cs->req);
 803 }
 804
 805 /* Do as much copy to/from userspace buffer as we can */
 806 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 807 {
 808         unsigned ncpy = min(*size, cs->len);
 809         if (val) {
 810                 void *pgaddr = kmap_atomic(cs->pg);
 811                 void *buf = pgaddr + cs->offset;
 812
 813                 if (cs->write)
 814                         memcpy(buf, *val, ncpy);
 815                 else
 816                         memcpy(*val, buf, ncpy);
 817
 818                 kunmap_atomic(pgaddr);
 819                 *val += ncpy;
 820         }
 821         *size -= ncpy;
 822         cs->len -= ncpy;
 823         cs->offset += ncpy;
 824         return ncpy;
 825 }
 826
 827 static int fuse_check_page(struct page *page)
 828 {
 829         if (page_mapcount(page) ||
 830             page->mapping != NULL ||
 831             page_count(page) != 1 ||
 832             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 833              ~(1 << PG_locked |
 834                1 << PG_referenced |
 835                1 << PG_uptodate |
 836                1 << PG_lru |
 837                1 << PG_active |
 838                1 << PG_reclaim))) {
 839                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 840                 printk(KERN_WARNING "  page=%p index=%li flags=%08lx, count=%i, mapcount=%i, mapping=%p\n", page, page->index, page->flags, page_count(page), page_mapcount(page), page->mapping);
 841                 return 1;
 842         }
 843         return 0;
 844 }
 845
 846 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 847 {
 848         int err;
 849         struct page *oldpage = *pagep;
 850         struct page *newpage;
 851         struct pipe_buffer *buf = cs->pipebufs;
 852
 853         err = unlock_request(cs->req);
 854         if (err)
 855                 return err;
 856
 857         fuse_copy_finish(cs);
 858
 859         err = buf->ops->confirm(cs->pipe, buf);
 860         if (err)
 861                 return err;
 862
 863         BUG_ON(!cs->nr_segs);
 864         cs->currbuf = buf;
 865         cs->len = buf->len;
 866         cs->pipebufs++;
 867         cs->nr_segs--;
 868
 869         if (cs->len != PAGE_SIZE)
 870                 goto out_fallback;
 871
 872         if (buf->ops->steal(cs->pipe, buf) != 0)
 873                 goto out_fallback;
 874
 875         newpage = buf->page;
 876
 877         if (!PageUptodate(newpage))
 878                 SetPageUptodate(newpage);
 879
 880         ClearPageMappedToDisk(newpage);
 881
 882         if (fuse_check_page(newpage) != 0)
 883                 goto out_fallback_unlock;
 884
 885         /*
 886          * This is a new and locked page, it shouldn't be mapped or
 887          * have any special flags on it
 888          */
 889         if (WARN_ON(page_mapped(oldpage)))
 890                 goto out_fallback_unlock;
 891         if (WARN_ON(page_has_private(oldpage)))
 892                 goto out_fallback_unlock;
 893         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 894                 goto out_fallback_unlock;
 895         if (WARN_ON(PageMlocked(oldpage)))
 896                 goto out_fallback_unlock;
 897
 898         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 899         if (err) {
 900                 unlock_page(newpage);
 901                 return err;
 902         }
 903
 904         page_cache_get(newpage);
 905
 906         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 907                 lru_cache_add_file(newpage);
 908
 909         err = 0;
 910         spin_lock(&cs->req->waitq.lock);
 911         if (test_bit(FR_ABORTED, &cs->req->flags))
 912                 err = -ENOENT;
 913         else
 914                 *pagep = newpage;
 915         spin_unlock(&cs->req->waitq.lock);
 916
 917         if (err) {
 918                 unlock_page(newpage);
 919                 page_cache_release(newpage);
 920                 return err;
 921         }
 922
 923         unlock_page(oldpage);
 924         page_cache_release(oldpage);
 925         cs->len = 0;
 926
 927         return 0;
 928
 929 out_fallback_unlock:
 930         unlock_page(newpage);
 931 out_fallback:
 932         cs->pg = buf->page;
 933         cs->offset = buf->offset;
 934
 935         err = lock_request(cs->req);
 936         if (err)
 937                 return err;
 938
 939         return 1;
 940 }
 941
 942 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 943                          unsigned offset, unsigned count)
 944 {
 945         struct pipe_buffer *buf;
 946         int err;
 947
 948         if (cs->nr_segs == cs->pipe->buffers)
 949                 return -EIO;
 950
 951         err = unlock_request(cs->req);
 952         if (err)
 953                 return err;
 954
 955         fuse_copy_finish(cs);
 956
 957         buf = cs->pipebufs;
 958         page_cache_get(page);
 959         buf->page = page;
 960         buf->offset = offset;
 961         buf->len = count;
 962
 963         cs->pipebufs++;
 964         cs->nr_segs++;
 965         cs->len = 0;
 966
 967         return 0;
 968 }
 969
 970 /*
 971  * Copy a page in the request to/from the userspace buffer.  Must be
 972  * done atomically
 973  */
 974 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 975                           unsigned offset, unsigned count, int zeroing)
 976 {
 977         int err;
 978         struct page *page = *pagep;
 979
 980         if (page && zeroing && count < PAGE_SIZE)
 981                 clear_highpage(page);
 982
 983         while (count) {
 984                 if (cs->write && cs->pipebufs && page) {
 985                         return fuse_ref_page(cs, page, offset, count);
 986                 } else if (!cs->len) {
 987                         if (cs->move_pages && page &&
 988                             offset == 0 && count == PAGE_SIZE) {
 989                                 err = fuse_try_move_page(cs, pagep);
 990                                 if (err <= 0)
 991                                         return err;
 992                         } else {
 993                                 err = fuse_copy_fill(cs);
 994                                 if (err)
 995                                         return err;
 996                         }
 997                 }
 998                 if (page) {
 999                         void *mapaddr = kmap_atomic(page);
1000                         void *buf = mapaddr + offset;
1001                         offset += fuse_copy_do(cs, &buf, &count);
1002                         kunmap_atomic(mapaddr);
1003                 } else
1004                         offset += fuse_copy_do(cs, NULL, &count);
1005         }
1006         if (page && !cs->write)
1007                 flush_dcache_page(page);
1008         return 0;
1009 }
1010
1011 /* Copy pages in the request to/from userspace buffer */
1012 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1013                            int zeroing)
1014 {
1015         unsigned i;
1016         struct fuse_req *req = cs->req;
1017
1018         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1019                 int err;
1020                 unsigned offset = req->page_descs[i].offset;
1021                 unsigned count = min(nbytes, req->page_descs[i].length);
1022
1023                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1024                                      zeroing);
1025                 if (err)
1026                         return err;
1027
1028                 nbytes -= count;
1029         }
1030         return 0;
1031 }
1032
1033 /* Copy a single argument in the request to/from userspace buffer */
1034 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1035 {
1036         while (size) {
1037                 if (!cs->len) {
1038                         int err = fuse_copy_fill(cs);
1039                         if (err)
1040                                 return err;
1041                 }
1042                 fuse_copy_do(cs, &val, &size);
1043         }
1044         return 0;
1045 }
1046
1047 /* Copy request arguments to/from userspace buffer */
1048 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1049                           unsigned argpages, struct fuse_arg *args,
1050                           int zeroing)
1051 {
1052         int err = 0;
1053         unsigned i;
1054
1055         for (i = 0; !err && i < numargs; i++)  {
1056                 struct fuse_arg *arg = &args[i];
1057                 if (i == numargs - 1 && argpages)
1058                         err = fuse_copy_pages(cs, arg->size, zeroing);
1059                 else
1060                         err = fuse_copy_one(cs, arg->value, arg->size);
1061         }
1062         return err;
1063 }
1064
1065 static int forget_pending(struct fuse_iqueue *fiq)
1066 {
1067         return fiq->forget_list_head.next != NULL;
1068 }
1069
1070 static int request_pending(struct fuse_iqueue *fiq)
1071 {
1072         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1073                 forget_pending(fiq);
1074 }
1075
1076 /*
1077  * Transfer an interrupt request to userspace
1078  *
1079  * Unlike other requests this is assembled on demand, without a need
1080  * to allocate a separate fuse_req structure.
1081  *
1082  * Called with fiq->waitq.lock held, releases it
1083  */
1084 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1085                                struct fuse_copy_state *cs,
1086                                size_t nbytes, struct fuse_req *req)
1087 __releases(fiq->waitq.lock)
1088 {
1089         struct fuse_in_header ih;
1090         struct fuse_interrupt_in arg;
1091         unsigned reqsize = sizeof(ih) + sizeof(arg);
1092         int err;
1093
1094         list_del_init(&req->intr_entry);
1095         req->intr_unique = fuse_get_unique(fiq);
1096         memset(&ih, 0, sizeof(ih));
1097         memset(&arg, 0, sizeof(arg));
1098         ih.len = reqsize;
1099         ih.opcode = FUSE_INTERRUPT;
1100         ih.unique = req->intr_unique;
1101         arg.unique = req->in.h.unique;
1102
1103         spin_unlock(&fiq->waitq.lock);
1104         if (nbytes < reqsize)
1105                 return -EINVAL;
1106
1107         err = fuse_copy_one(cs, &ih, sizeof(ih));
1108         if (!err)
1109                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1110         fuse_copy_finish(cs);
1111
1112         return err ? err : reqsize;
1113 }
1114
1115 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1116                                                unsigned max,
1117                                                unsigned *countp)
1118 {
1119         struct fuse_forget_link *head = fiq->forget_list_head.next;
1120         struct fuse_forget_link **newhead = &head;
1121         unsigned count;
1122
1123         for (count = 0; *newhead != NULL && count < max; count++)
1124                 newhead = &(*newhead)->next;
1125
1126         fiq->forget_list_head.next = *newhead;
1127         *newhead = NULL;
1128         if (fiq->forget_list_head.next == NULL)
1129                 fiq->forget_list_tail = &fiq->forget_list_head;
1130
1131         if (countp != NULL)
1132                 *countp = count;
1133
1134         return head;
1135 }
1136
1137 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1138                                    struct fuse_copy_state *cs,
1139                                    size_t nbytes)
1140 __releases(fiq->waitq.lock)
1141 {
1142         int err;
1143         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1144         struct fuse_forget_in arg = {
1145                 .nlookup = forget->forget_one.nlookup,
1146         };
1147         struct fuse_in_header ih = {
1148                 .opcode = FUSE_FORGET,
1149                 .nodeid = forget->forget_one.nodeid,
1150                 .unique = fuse_get_unique(fiq),
1151                 .len = sizeof(ih) + sizeof(arg),
1152         };
1153
1154         spin_unlock(&fiq->waitq.lock);
1155         kfree(forget);
1156         if (nbytes < ih.len)
1157                 return -EINVAL;
1158
1159         err = fuse_copy_one(cs, &ih, sizeof(ih));
1160         if (!err)
1161                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1162         fuse_copy_finish(cs);
1163
1164         if (err)
1165                 return err;
1166
1167         return ih.len;
1168 }
1169
1170 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1171                                    struct fuse_copy_state *cs, size_t nbytes)
1172 __releases(fiq->waitq.lock)
1173 {
1174         int err;
1175         unsigned max_forgets;
1176         unsigned count;
1177         struct fuse_forget_link *head;
1178         struct fuse_batch_forget_in arg = { .count = 0 };
1179         struct fuse_in_header ih = {
1180                 .opcode = FUSE_BATCH_FORGET,
1181                 .unique = fuse_get_unique(fiq),
1182                 .len = sizeof(ih) + sizeof(arg),
1183         };
1184
1185         if (nbytes < ih.len) {
1186                 spin_unlock(&fiq->waitq.lock);
1187                 return -EINVAL;
1188         }
1189
1190         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1191         head = dequeue_forget(fiq, max_forgets, &count);
1192         spin_unlock(&fiq->waitq.lock);
1193
1194         arg.count = count;
1195         ih.len += count * sizeof(struct fuse_forget_one);
1196         err = fuse_copy_one(cs, &ih, sizeof(ih));
1197         if (!err)
1198                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1199
1200         while (head) {
1201                 struct fuse_forget_link *forget = head;
1202
1203                 if (!err) {
1204                         err = fuse_copy_one(cs, &forget->forget_one,
1205                                             sizeof(forget->forget_one));
1206                 }
1207                 head = forget->next;
1208                 kfree(forget);
1209         }
1210
1211         fuse_copy_finish(cs);
1212
1213         if (err)
1214                 return err;
1215
1216         return ih.len;
1217 }
1218
1219 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1220                             struct fuse_copy_state *cs,
1221                             size_t nbytes)
1222 __releases(fiq->waitq.lock)
1223 {
1224         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1225                 return fuse_read_single_forget(fiq, cs, nbytes);
1226         else
1227                 return fuse_read_batch_forget(fiq, cs, nbytes);
1228 }
1229
1230 /*
1231  * Read a single request into the userspace filesystem's buffer.  This
1232  * function waits until a request is available, then removes it from
1233  * the pending list and copies request data to userspace buffer.  If
1234  * no reply is needed (FORGET) or request has been aborted or there
1235  * was an error during the copying then it's finished by calling
1236  * request_end().  Otherwise add it to the processing list, and set
1237  * the 'sent' flag.
1238  */
1239 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1240                                 struct fuse_copy_state *cs, size_t nbytes)
1241 {
1242         ssize_t err;
1243         struct fuse_conn *fc = fud->fc;
1244         struct fuse_iqueue *fiq = &fc->iq;
1245         struct fuse_pqueue *fpq = &fud->pq;
1246         struct fuse_req *req;
1247         struct fuse_in *in;
1248         unsigned reqsize;
1249
1250  restart:
1251         spin_lock(&fiq->waitq.lock);
1252         err = -EAGAIN;
1253         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1254             !request_pending(fiq))
1255                 goto err_unlock;
1256
1257         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1258                                 !fiq->connected || request_pending(fiq));
1259         if (err)
1260                 goto err_unlock;
1261
1262         err = -ENODEV;
1263         if (!fiq->connected)
1264                 goto err_unlock;
1265
1266         if (!list_empty(&fiq->interrupts)) {
1267                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1268                                  intr_entry);
1269                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1270         }
1271
1272         if (forget_pending(fiq)) {
1273                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1274                         return fuse_read_forget(fc, fiq, cs, nbytes);
1275
1276                 if (fiq->forget_batch <= -8)
1277                         fiq->forget_batch = 16;
1278         }
1279
1280         req = list_entry(fiq->pending.next, struct fuse_req, list);
1281         clear_bit(FR_PENDING, &req->flags);
1282         list_del_init(&req->list);
1283         spin_unlock(&fiq->waitq.lock);
1284
1285         in = &req->in;
1286         reqsize = in->h.len;
1287         /* If request is too large, reply with an error and restart the read */
1288         if (nbytes < reqsize) {
1289                 req->out.h.error = -EIO;
1290                 /* SETXATTR is special, since it may contain too large data */
1291                 if (in->h.opcode == FUSE_SETXATTR)
1292                         req->out.h.error = -E2BIG;
1293                 request_end(fc, req);
1294                 goto restart;
1295         }
1296         spin_lock(&fpq->lock);
1297         list_add(&req->list, &fpq->io);
1298         spin_unlock(&fpq->lock);
1299         cs->req = req;
1300         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1301         if (!err)
1302                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1303                                      (struct fuse_arg *) in->args, 0);
1304         fuse_copy_finish(cs);
1305         spin_lock(&fpq->lock);
1306         clear_bit(FR_LOCKED, &req->flags);
1307         if (!fpq->connected) {
1308                 err = -ENODEV;
1309                 goto out_end;
1310         }
1311         if (err) {
1312                 req->out.h.error = -EIO;
1313                 goto out_end;
1314         }
1315         if (!test_bit(FR_ISREPLY, &req->flags)) {
1316                 err = reqsize;
1317                 goto out_end;
1318         }
1319         list_move_tail(&req->list, &fpq->processing);
1320         spin_unlock(&fpq->lock);
1321         set_bit(FR_SENT, &req->flags);
1322         /* matches barrier in request_wait_answer() */
1323         smp_mb__after_atomic();
1324         if (test_bit(FR_INTERRUPTED, &req->flags))
1325                 queue_interrupt(fiq, req);
1326
1327         return reqsize;
1328
1329 out_end:
1330         if (!test_bit(FR_PRIVATE, &req->flags))
1331                 list_del_init(&req->list);
1332         spin_unlock(&fpq->lock);
1333         request_end(fc, req);
1334         return err;
1335
1336  err_unlock:
1337         spin_unlock(&fiq->waitq.lock);
1338         return err;
1339 }
1340
1341 static int fuse_dev_open(struct inode *inode, struct file *file)
1342 {
1343         /*
1344          * The fuse device's file's private_data is used to hold
1345          * the fuse_conn(ection) when it is mounted, and is used to
1346          * keep track of whether the file has been mounted already.
1347          */
1348         file->private_data = NULL;
1349         return 0;
1350 }
1351
1352 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1353 {
1354         struct fuse_copy_state cs;
1355         struct file *file = iocb->ki_filp;
1356         struct fuse_dev *fud = fuse_get_dev(file);
1357
1358         if (!fud)
1359                 return -EPERM;
1360
1361         if (!iter_is_iovec(to))
1362                 return -EINVAL;
1363
1364         fuse_copy_init(&cs, 1, to);
1365
1366         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1367 }
1368
1369 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1370                                     struct pipe_inode_info *pipe,
1371                                     size_t len, unsigned int flags)
1372 {
1373         int ret;
1374         int page_nr = 0;
1375         int do_wakeup = 0;
1376         struct pipe_buffer *bufs;
1377         struct fuse_copy_state cs;
1378         struct fuse_dev *fud = fuse_get_dev(in);
1379
1380         if (!fud)
1381                 return -EPERM;
1382
1383         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1384         if (!bufs)
1385                 return -ENOMEM;
1386
1387         fuse_copy_init(&cs, 1, NULL);
1388         cs.pipebufs = bufs;
1389         cs.pipe = pipe;
1390         ret = fuse_dev_do_read(fud, in, &cs, len);
1391         if (ret < 0)
1392                 goto out;
1393
1394         ret = 0;
1395         pipe_lock(pipe);
1396
1397         if (!pipe->readers) {
1398                 send_sig(SIGPIPE, current, 0);
1399                 if (!ret)
1400                         ret = -EPIPE;
1401                 goto out_unlock;
1402         }
1403
1404         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1405                 ret = -EIO;
1406                 goto out_unlock;
1407         }
1408
1409         while (page_nr < cs.nr_segs) {
1410                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1411                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1412
1413                 buf->page = bufs[page_nr].page;
1414                 buf->offset = bufs[page_nr].offset;
1415                 buf->len = bufs[page_nr].len;
1416                 /*
1417                  * Need to be careful about this.  Having buf->ops in module
1418                  * code can Oops if the buffer persists after module unload.
1419                  */
1420                 buf->ops = &nosteal_pipe_buf_ops;
1421
1422                 pipe->nrbufs++;
1423                 page_nr++;
1424                 ret += buf->len;
1425
1426                 if (pipe->files)
1427                         do_wakeup = 1;
1428         }
1429
1430 out_unlock:
1431         pipe_unlock(pipe);
1432
1433         if (do_wakeup) {
1434                 smp_mb();
1435                 if (waitqueue_active(&pipe->wait))
1436                         wake_up_interruptible(&pipe->wait);
1437                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1438         }
1439
1440 out:
1441         for (; page_nr < cs.nr_segs; page_nr++)
1442                 page_cache_release(bufs[page_nr].page);
1443
1444         kfree(bufs);
1445         return ret;
1446 }
1447
1448 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1449                             struct fuse_copy_state *cs)
1450 {
1451         struct fuse_notify_poll_wakeup_out outarg;
1452         int err = -EINVAL;
1453
1454         if (size != sizeof(outarg))
1455                 goto err;
1456
1457         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1458         if (err)
1459                 goto err;
1460
1461         fuse_copy_finish(cs);
1462         return fuse_notify_poll_wakeup(fc, &outarg);
1463
1464 err:
1465         fuse_copy_finish(cs);
1466         return err;
1467 }
1468
1469 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1470                                    struct fuse_copy_state *cs)
1471 {
1472         struct fuse_notify_inval_inode_out outarg;
1473         int err = -EINVAL;
1474
1475         if (size != sizeof(outarg))
1476                 goto err;
1477
1478         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1479         if (err)
1480                 goto err;
1481         fuse_copy_finish(cs);
1482
1483         down_read(&fc->killsb);
1484         err = -ENOENT;
1485         if (fc->sb) {
1486                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1487                                                outarg.off, outarg.len);
1488         }
1489         up_read(&fc->killsb);
1490         return err;
1491
1492 err:
1493         fuse_copy_finish(cs);
1494         return err;
1495 }
1496
1497 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1498                                    struct fuse_copy_state *cs)
1499 {
1500         struct fuse_notify_inval_entry_out outarg;
1501         int err = -ENOMEM;
1502         char *buf;
1503         struct qstr name;
1504
1505         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1506         if (!buf)
1507                 goto err;
1508
1509         err = -EINVAL;
1510         if (size < sizeof(outarg))
1511                 goto err;
1512
1513         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1514         if (err)
1515                 goto err;
1516
1517         err = -ENAMETOOLONG;
1518         if (outarg.namelen > FUSE_NAME_MAX)
1519                 goto err;
1520
1521         err = -EINVAL;
1522         if (size != sizeof(outarg) + outarg.namelen + 1)
1523                 goto err;
1524
1525         name.name = buf;
1526         name.len = outarg.namelen;
1527         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1528         if (err)
1529                 goto err;
1530         fuse_copy_finish(cs);
1531         buf[outarg.namelen] = 0;
1532         name.hash = full_name_hash(name.name, name.len);
1533
1534         down_read(&fc->killsb);
1535         err = -ENOENT;
1536         if (fc->sb)
1537                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1538         up_read(&fc->killsb);
1539         kfree(buf);
1540         return err;
1541
1542 err:
1543         kfree(buf);
1544         fuse_copy_finish(cs);
1545         return err;
1546 }
1547
1548 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1549                               struct fuse_copy_state *cs)
1550 {
1551         struct fuse_notify_delete_out outarg;
1552         int err = -ENOMEM;
1553         char *buf;
1554         struct qstr name;
1555
1556         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1557         if (!buf)
1558                 goto err;
1559
1560         err = -EINVAL;
1561         if (size < sizeof(outarg))
1562                 goto err;
1563
1564         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1565         if (err)
1566                 goto err;
1567
1568         err = -ENAMETOOLONG;
1569         if (outarg.namelen > FUSE_NAME_MAX)
1570                 goto err;
1571
1572         err = -EINVAL;
1573         if (size != sizeof(outarg) + outarg.namelen + 1)
1574                 goto err;
1575
1576         name.name = buf;
1577         name.len = outarg.namelen;
1578         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1579         if (err)
1580                 goto err;
1581         fuse_copy_finish(cs);
1582         buf[outarg.namelen] = 0;
1583         name.hash = full_name_hash(name.name, name.len);
1584
1585         down_read(&fc->killsb);
1586         err = -ENOENT;
1587         if (fc->sb)
1588                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1589                                                outarg.child, &name);
1590         up_read(&fc->killsb);
1591         kfree(buf);
1592         return err;
1593
1594 err:
1595         kfree(buf);
1596         fuse_copy_finish(cs);
1597         return err;
1598 }
1599
1600 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1601                              struct fuse_copy_state *cs)
1602 {
1603         struct fuse_notify_store_out outarg;
1604         struct inode *inode;
1605         struct address_space *mapping;
1606         u64 nodeid;
1607         int err;
1608         pgoff_t index;
1609         unsigned int offset;
1610         unsigned int num;
1611         loff_t file_size;
1612         loff_t end;
1613
1614         err = -EINVAL;
1615         if (size < sizeof(outarg))
1616                 goto out_finish;
1617
1618         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1619         if (err)
1620                 goto out_finish;
1621
1622         err = -EINVAL;
1623         if (size - sizeof(outarg) != outarg.size)
1624                 goto out_finish;
1625
1626         nodeid = outarg.nodeid;
1627
1628         down_read(&fc->killsb);
1629
1630         err = -ENOENT;
1631         if (!fc->sb)
1632                 goto out_up_killsb;
1633
1634         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1635         if (!inode)
1636                 goto out_up_killsb;
1637
1638         mapping = inode->i_mapping;
1639         index = outarg.offset >> PAGE_CACHE_SHIFT;
1640         offset = outarg.offset & ~PAGE_CACHE_MASK;
1641         file_size = i_size_read(inode);
1642         end = outarg.offset + outarg.size;
1643         if (end > file_size) {
1644                 file_size = end;
1645                 fuse_write_update_size(inode, file_size);
1646         }
1647
1648         num = outarg.size;
1649         while (num) {
1650                 struct page *page;
1651                 unsigned int this_num;
1652
1653                 err = -ENOMEM;
1654                 page = find_or_create_page(mapping, index,
1655                                            mapping_gfp_mask(mapping));
1656                 if (!page)
1657                         goto out_iput;
1658
1659                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1660                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1661                 if (!err && offset == 0 &&
1662                     (this_num == PAGE_CACHE_SIZE || file_size == end))
1663                         SetPageUptodate(page);
1664                 unlock_page(page);
1665                 page_cache_release(page);
1666
1667                 if (err)
1668                         goto out_iput;
1669
1670                 num -= this_num;
1671                 offset = 0;
1672                 index++;
1673         }
1674
1675         err = 0;
1676
1677 out_iput:
1678         iput(inode);
1679 out_up_killsb:
1680         up_read(&fc->killsb);
1681 out_finish:
1682         fuse_copy_finish(cs);
1683         return err;
1684 }
1685
1686 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1687 {
1688         release_pages(req->pages, req->num_pages, false);
1689 }
1690
1691 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1692                          struct fuse_notify_retrieve_out *outarg)
1693 {
1694         int err;
1695         struct address_space *mapping = inode->i_mapping;
1696         struct fuse_req *req;
1697         pgoff_t index;
1698         loff_t file_size;
1699         unsigned int num;
1700         unsigned int offset;
1701         size_t total_len = 0;
1702         int num_pages;
1703
1704         offset = outarg->offset & ~PAGE_CACHE_MASK;
1705         file_size = i_size_read(inode);
1706
1707         num = outarg->size;
1708         if (outarg->offset > file_size)
1709                 num = 0;
1710         else if (outarg->offset + num > file_size)
1711                 num = file_size - outarg->offset;
1712
1713         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1714         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1715
1716         req = fuse_get_req(fc, num_pages);
1717         if (IS_ERR(req))
1718                 return PTR_ERR(req);
1719
1720         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1721         req->in.h.nodeid = outarg->nodeid;
1722         req->in.numargs = 2;
1723         req->in.argpages = 1;
1724         req->page_descs[0].offset = offset;
1725         req->end = fuse_retrieve_end;
1726
1727         index = outarg->offset >> PAGE_CACHE_SHIFT;
1728
1729         while (num && req->num_pages < num_pages) {
1730                 struct page *page;
1731                 unsigned int this_num;
1732
1733                 page = find_get_page(mapping, index);
1734                 if (!page)
1735                         break;
1736
1737                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1738                 req->pages[req->num_pages] = page;
1739                 req->page_descs[req->num_pages].length = this_num;
1740                 req->num_pages++;
1741
1742                 offset = 0;
1743                 num -= this_num;
1744                 total_len += this_num;
1745                 index++;
1746         }
1747         req->misc.retrieve_in.offset = outarg->offset;
1748         req->misc.retrieve_in.size = total_len;
1749         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1750         req->in.args[0].value = &req->misc.retrieve_in;
1751         req->in.args[1].size = total_len;
1752
1753         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1754         if (err)
1755                 fuse_retrieve_end(fc, req);
1756
1757         return err;
1758 }
1759
1760 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1761                                 struct fuse_copy_state *cs)
1762 {
1763         struct fuse_notify_retrieve_out outarg;
1764         struct inode *inode;
1765         int err;
1766
1767         err = -EINVAL;
1768         if (size != sizeof(outarg))
1769                 goto copy_finish;
1770
1771         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1772         if (err)
1773                 goto copy_finish;
1774
1775         fuse_copy_finish(cs);
1776
1777         down_read(&fc->killsb);
1778         err = -ENOENT;
1779         if (fc->sb) {
1780                 u64 nodeid = outarg.nodeid;
1781
1782                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1783                 if (inode) {
1784                         err = fuse_retrieve(fc, inode, &outarg);
1785                         iput(inode);
1786                 }
1787         }
1788         up_read(&fc->killsb);
1789
1790         return err;
1791
1792 copy_finish:
1793         fuse_copy_finish(cs);
1794         return err;
1795 }
1796
1797 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1798                        unsigned int size, struct fuse_copy_state *cs)
1799 {
1800         /* Don't try to move pages (yet) */
1801         cs->move_pages = 0;
1802
1803         switch (code) {
1804         case FUSE_NOTIFY_POLL:
1805                 return fuse_notify_poll(fc, size, cs);
1806
1807         case FUSE_NOTIFY_INVAL_INODE:
1808                 return fuse_notify_inval_inode(fc, size, cs);
1809
1810         case FUSE_NOTIFY_INVAL_ENTRY:
1811                 return fuse_notify_inval_entry(fc, size, cs);
1812
1813         case FUSE_NOTIFY_STORE:
1814                 return fuse_notify_store(fc, size, cs);
1815
1816         case FUSE_NOTIFY_RETRIEVE:
1817                 return fuse_notify_retrieve(fc, size, cs);
1818
1819         case FUSE_NOTIFY_DELETE:
1820                 return fuse_notify_delete(fc, size, cs);
1821
1822         default:
1823                 fuse_copy_finish(cs);
1824                 return -EINVAL;
1825         }
1826 }
1827
1828 /* Look up request on processing list by unique ID */
1829 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1830 {
1831         struct fuse_req *req;
1832
1833         list_for_each_entry(req, &fpq->processing, list) {
1834                 if (req->in.h.unique == unique || req->intr_unique == unique)
1835                         return req;
1836         }
1837         return NULL;
1838 }
1839
1840 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1841                          unsigned nbytes)
1842 {
1843         unsigned reqsize = sizeof(struct fuse_out_header);
1844
1845         if (out->h.error)
1846                 return nbytes != reqsize ? -EINVAL : 0;
1847
1848         reqsize += len_args(out->numargs, out->args);
1849
1850         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1851                 return -EINVAL;
1852         else if (reqsize > nbytes) {
1853                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1854                 unsigned diffsize = reqsize - nbytes;
1855                 if (diffsize > lastarg->size)
1856                         return -EINVAL;
1857                 lastarg->size -= diffsize;
1858         }
1859         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1860                               out->page_zeroing);
1861 }
1862
1863 /*
1864  * Write a single reply to a request.  First the header is copied from
1865  * the write buffer.  The request is then searched on the processing
1866  * list by the unique ID found in the header.  If found, then remove
1867  * it from the list and copy the rest of the buffer to the request.
1868  * The request is finished by calling request_end()
1869  */
1870 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1871                                  struct fuse_copy_state *cs, size_t nbytes)
1872 {
1873         int err;
1874         struct fuse_conn *fc = fud->fc;
1875         struct fuse_pqueue *fpq = &fud->pq;
1876         struct fuse_req *req;
1877         struct fuse_out_header oh;
1878
1879         if (nbytes < sizeof(struct fuse_out_header))
1880                 return -EINVAL;
1881
1882         err = fuse_copy_one(cs, &oh, sizeof(oh));
1883         if (err)
1884                 goto err_finish;
1885
1886         err = -EINVAL;
1887         if (oh.len != nbytes)
1888                 goto err_finish;
1889
1890         /*
1891          * Zero oh.unique indicates unsolicited notification message
1892          * and error contains notification code.
1893          */
1894         if (!oh.unique) {
1895                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1896                 return err ? err : nbytes;
1897         }
1898
1899         err = -EINVAL;
1900         if (oh.error <= -1000 || oh.error > 0)
1901                 goto err_finish;
1902
1903         spin_lock(&fpq->lock);
1904         err = -ENOENT;
1905         if (!fpq->connected)
1906                 goto err_unlock_pq;
1907
1908         req = request_find(fpq, oh.unique);
1909         if (!req)
1910                 goto err_unlock_pq;
1911
1912         /* Is it an interrupt reply? */
1913         if (req->intr_unique == oh.unique) {
1914                 spin_unlock(&fpq->lock);
1915
1916                 err = -EINVAL;
1917                 if (nbytes != sizeof(struct fuse_out_header))
1918                         goto err_finish;
1919
1920                 if (oh.error == -ENOSYS)
1921                         fc->no_interrupt = 1;
1922                 else if (oh.error == -EAGAIN)
1923                         queue_interrupt(&fc->iq, req);
1924
1925                 fuse_copy_finish(cs);
1926                 return nbytes;
1927         }
1928
1929         clear_bit(FR_SENT, &req->flags);
1930         list_move(&req->list, &fpq->io);
1931         req->out.h = oh;
1932         set_bit(FR_LOCKED, &req->flags);
1933         spin_unlock(&fpq->lock);
1934         cs->req = req;
1935         if (!req->out.page_replace)
1936                 cs->move_pages = 0;
1937
1938         err = copy_out_args(cs, &req->out, nbytes);
1939         if (req->in.h.opcode == FUSE_CANONICAL_PATH) {
1940                 req->out.h.error = kern_path((char *)req->out.args[0].value, 0,
1941                                                         req->canonical_path);
1942         }
1943         fuse_copy_finish(cs);
1944
1945         spin_lock(&fpq->lock);
1946         clear_bit(FR_LOCKED, &req->flags);
1947         if (!fpq->connected)
1948                 err = -ENOENT;
1949         else if (err)
1950                 req->out.h.error = -EIO;
1951         if (!test_bit(FR_PRIVATE, &req->flags))
1952                 list_del_init(&req->list);
1953         spin_unlock(&fpq->lock);
1954
1955         request_end(fc, req);
1956
1957         return err ? err : nbytes;
1958
1959  err_unlock_pq:
1960         spin_unlock(&fpq->lock);
1961  err_finish:
1962         fuse_copy_finish(cs);
1963         return err;
1964 }
1965
1966 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1967 {
1968         struct fuse_copy_state cs;
1969         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1970
1971         if (!fud)
1972                 return -EPERM;
1973
1974         if (!iter_is_iovec(from))
1975                 return -EINVAL;
1976
1977         fuse_copy_init(&cs, 0, from);
1978
1979         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1980 }
1981
1982 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1983                                      struct file *out, loff_t *ppos,
1984                                      size_t len, unsigned int flags)
1985 {
1986         unsigned nbuf;
1987         unsigned idx;
1988         struct pipe_buffer *bufs;
1989         struct fuse_copy_state cs;
1990         struct fuse_dev *fud;
1991         size_t rem;
1992         ssize_t ret;
1993
1994         fud = fuse_get_dev(out);
1995         if (!fud)
1996                 return -EPERM;
1997
1998         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1999         if (!bufs)
2000                 return -ENOMEM;
2001
2002         pipe_lock(pipe);
2003         nbuf = 0;
2004         rem = 0;
2005         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2006                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2007
2008         ret = -EINVAL;
2009         if (rem < len) {
2010                 pipe_unlock(pipe);
2011                 goto out;
2012         }
2013
2014         rem = len;
2015         while (rem) {
2016                 struct pipe_buffer *ibuf;
2017                 struct pipe_buffer *obuf;
2018
2019                 BUG_ON(nbuf >= pipe->buffers);
2020                 BUG_ON(!pipe->nrbufs);
2021                 ibuf = &pipe->bufs[pipe->curbuf];
2022                 obuf = &bufs[nbuf];
2023
2024                 if (rem >= ibuf->len) {
2025                         *obuf = *ibuf;
2026                         ibuf->ops = NULL;
2027                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2028                         pipe->nrbufs--;
2029                 } else {
2030                         ibuf->ops->get(pipe, ibuf);
2031                         *obuf = *ibuf;
2032                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2033                         obuf->len = rem;
2034                         ibuf->offset += obuf->len;
2035                         ibuf->len -= obuf->len;
2036                 }
2037                 nbuf++;
2038                 rem -= obuf->len;
2039         }
2040         pipe_unlock(pipe);
2041
2042         fuse_copy_init(&cs, 0, NULL);
2043         cs.pipebufs = bufs;
2044         cs.nr_segs = nbuf;
2045         cs.pipe = pipe;
2046
2047         if (flags & SPLICE_F_MOVE)
2048                 cs.move_pages = 1;
2049
2050         ret = fuse_dev_do_write(fud, &cs, len);
2051
2052         for (idx = 0; idx < nbuf; idx++) {
2053                 struct pipe_buffer *buf = &bufs[idx];
2054                 buf->ops->release(pipe, buf);
2055         }
2056 out:
2057         kfree(bufs);
2058         return ret;
2059 }
2060
2061 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2062 {
2063         unsigned mask = POLLOUT | POLLWRNORM;
2064         struct fuse_iqueue *fiq;
2065         struct fuse_dev *fud = fuse_get_dev(file);
2066
2067         if (!fud)
2068                 return POLLERR;
2069
2070         fiq = &fud->fc->iq;
2071         poll_wait(file, &fiq->waitq, wait);
2072
2073         spin_lock(&fiq->waitq.lock);
2074         if (!fiq->connected)
2075                 mask = POLLERR;
2076         else if (request_pending(fiq))
2077                 mask |= POLLIN | POLLRDNORM;
2078         spin_unlock(&fiq->waitq.lock);
2079
2080         return mask;
2081 }
2082
2083 /*
2084  * Abort all requests on the given list (pending or processing)
2085  *
2086  * This function releases and reacquires fc->lock
2087  */
2088 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2089 {
2090         while (!list_empty(head)) {
2091                 struct fuse_req *req;
2092                 req = list_entry(head->next, struct fuse_req, list);
2093                 req->out.h.error = -ECONNABORTED;
2094                 clear_bit(FR_PENDING, &req->flags);
2095                 clear_bit(FR_SENT, &req->flags);
2096                 list_del_init(&req->list);
2097                 request_end(fc, req);
2098         }
2099 }
2100
2101 static void end_polls(struct fuse_conn *fc)
2102 {
2103         struct rb_node *p;
2104
2105         p = rb_first(&fc->polled_files);
2106
2107         while (p) {
2108                 struct fuse_file *ff;
2109                 ff = rb_entry(p, struct fuse_file, polled_node);
2110                 wake_up_interruptible_all(&ff->poll_wait);
2111
2112                 p = rb_next(p);
2113         }
2114 }
2115
2116 /*
2117  * Abort all requests.
2118  *
2119  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2120  * filesystem.
2121  *
2122  * The same effect is usually achievable through killing the filesystem daemon
2123  * and all users of the filesystem.  The exception is the combination of an
2124  * asynchronous request and the tricky deadlock (see
2125  * Documentation/filesystems/fuse.txt).
2126  *
2127  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2128  * requests, they should be finished off immediately.  Locked requests will be
2129  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2130  * requests.  It is possible that some request will finish before we can.  This
2131  * is OK, the request will in that case be removed from the list before we touch
2132  * it.
2133  */
2134 void fuse_abort_conn(struct fuse_conn *fc)
2135 {
2136         struct fuse_iqueue *fiq = &fc->iq;
2137
2138         spin_lock(&fc->lock);
2139         if (fc->connected) {
2140                 struct fuse_dev *fud;
2141                 struct fuse_req *req, *next;
2142                 LIST_HEAD(to_end1);
2143                 LIST_HEAD(to_end2);
2144
2145                 fc->connected = 0;
2146                 fc->blocked = 0;
2147                 fuse_set_initialized(fc);
2148                 list_for_each_entry(fud, &fc->devices, entry) {
2149                         struct fuse_pqueue *fpq = &fud->pq;
2150
2151                         spin_lock(&fpq->lock);
2152                         fpq->connected = 0;
2153                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2154                                 req->out.h.error = -ECONNABORTED;
2155                                 spin_lock(&req->waitq.lock);
2156                                 set_bit(FR_ABORTED, &req->flags);
2157                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2158                                         set_bit(FR_PRIVATE, &req->flags);
2159                                         list_move(&req->list, &to_end1);
2160                                 }
2161                                 spin_unlock(&req->waitq.lock);
2162                         }
2163                         list_splice_init(&fpq->processing, &to_end2);
2164                         spin_unlock(&fpq->lock);
2165                 }
2166                 fc->max_background = UINT_MAX;
2167                 flush_bg_queue(fc);
2168
2169                 spin_lock(&fiq->waitq.lock);
2170                 fiq->connected = 0;
2171                 list_splice_init(&fiq->pending, &to_end2);
2172                 while (forget_pending(fiq))
2173                         kfree(dequeue_forget(fiq, 1, NULL));
2174                 wake_up_all_locked(&fiq->waitq);
2175                 spin_unlock(&fiq->waitq.lock);
2176                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2177                 end_polls(fc);
2178                 wake_up_all(&fc->blocked_waitq);
2179                 spin_unlock(&fc->lock);
2180
2181                 while (!list_empty(&to_end1)) {
2182                         req = list_first_entry(&to_end1, struct fuse_req, list);
2183                         __fuse_get_request(req);
2184                         list_del_init(&req->list);
2185                         request_end(fc, req);
2186                 }
2187                 end_requests(fc, &to_end2);
2188         } else {
2189                 spin_unlock(&fc->lock);
2190         }
2191 }
2192 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2193
2194 int fuse_dev_release(struct inode *inode, struct file *file)
2195 {
2196         struct fuse_dev *fud = fuse_get_dev(file);
2197
2198         if (fud) {
2199                 struct fuse_conn *fc = fud->fc;
2200                 struct fuse_pqueue *fpq = &fud->pq;
2201
2202                 WARN_ON(!list_empty(&fpq->io));
2203                 end_requests(fc, &fpq->processing);
2204                 /* Are we the last open device? */
2205                 if (atomic_dec_and_test(&fc->dev_count)) {
2206                         WARN_ON(fc->iq.fasync != NULL);
2207                         fuse_abort_conn(fc);
2208                 }
2209                 fuse_dev_free(fud);
2210         }
2211         return 0;
2212 }
2213 EXPORT_SYMBOL_GPL(fuse_dev_release);
2214
2215 static int fuse_dev_fasync(int fd, struct file *file, int on)
2216 {
2217         struct fuse_dev *fud = fuse_get_dev(file);
2218
2219         if (!fud)
2220                 return -EPERM;
2221
2222         /* No locking - fasync_helper does its own locking */
2223         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2224 }
2225
2226 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2227 {
2228         struct fuse_dev *fud;
2229
2230         if (new->private_data)
2231                 return -EINVAL;
2232
2233         fud = fuse_dev_alloc(fc);
2234         if (!fud)
2235                 return -ENOMEM;
2236
2237         new->private_data = fud;
2238         atomic_inc(&fc->dev_count);
2239
2240         return 0;
2241 }
2242
2243 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2244                            unsigned long arg)
2245 {
2246         int err = -ENOTTY;
2247
2248         if (cmd == FUSE_DEV_IOC_CLONE) {
2249                 int oldfd;
2250
2251                 err = -EFAULT;
2252                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2253                         struct file *old = fget(oldfd);
2254
2255                         err = -EINVAL;
2256                         if (old) {
2257                                 struct fuse_dev *fud = NULL;
2258
2259                                 /*
2260                                  * Check against file->f_op because CUSE
2261                                  * uses the same ioctl handler.
2262                                  */
2263                                 if (old->f_op == file->f_op &&
2264                                     old->f_cred->user_ns == file->f_cred->user_ns)
2265                                         fud = fuse_get_dev(old);
2266
2267                                 if (fud) {
2268                                         mutex_lock(&fuse_mutex);
2269                                         err = fuse_device_clone(fud->fc, file);
2270                                         mutex_unlock(&fuse_mutex);
2271                                 }
2272                                 fput(old);
2273                         }
2274                 }
2275         }
2276         return err;
2277 }
2278
2279 const struct file_operations fuse_dev_operations = {
2280         .owner          = THIS_MODULE,
2281         .open           = fuse_dev_open,
2282         .llseek         = no_llseek,
2283         .read_iter      = fuse_dev_read,
2284         .splice_read    = fuse_dev_splice_read,
2285         .write_iter     = fuse_dev_write,
2286         .splice_write   = fuse_dev_splice_write,
2287         .poll           = fuse_dev_poll,
2288         .release        = fuse_dev_release,
2289         .fasync         = fuse_dev_fasync,
2290         .unlocked_ioctl = fuse_dev_ioctl,
2291         .compat_ioctl   = fuse_dev_ioctl,
2292 };
2293 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2294
2295 static struct miscdevice fuse_miscdevice = {
2296         .minor = FUSE_MINOR,
2297         .name  = "fuse",
2298         .fops = &fuse_dev_operations,
2299 };
2300
2301 int __init fuse_dev_init(void)
2302 {
2303         int err = -ENOMEM;
2304         fuse_req_cachep = kmem_cache_create("fuse_request",
2305                                             sizeof(struct fuse_req),
2306                                             0, 0, NULL);
2307         if (!fuse_req_cachep)
2308                 goto out;
2309
2310         err = misc_register(&fuse_miscdevice);
2311         if (err)
2312                 goto out_cache_clean;
2313
2314         return 0;
2315
2316  out_cache_clean:
2317         kmem_cache_destroy(fuse_req_cachep);
2318  out:
2319         return err;
2320 }
2321
2322 void fuse_dev_cleanup(void)
2323 {
2324         misc_deregister(&fuse_miscdevice);
2325         kmem_cache_destroy(fuse_req_cachep);
2326 }