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