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)
  38 {
  39         memset(req, 0, sizeof(*req));
  40         INIT_LIST_HEAD(&req->list);
  41         INIT_LIST_HEAD(&req->intr_entry);
  42         init_waitqueue_head(&req->waitq);
  43         atomic_set(&req->count, 1);
  44 }
  45
  46 struct fuse_req *fuse_request_alloc(void)
  47 {
  48         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_KERNEL);
  49         if (req)
  50                 fuse_request_init(req);
  51         return req;
  52 }
  53 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  54
  55 struct fuse_req *fuse_request_alloc_nofs(void)
  56 {
  57         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_NOFS);
  58         if (req)
  59                 fuse_request_init(req);
  60         return req;
  61 }
  62
  63 void fuse_request_free(struct fuse_req *req)
  64 {
  65         kmem_cache_free(fuse_req_cachep, req);
  66 }
  67
  68 static void block_sigs(sigset_t *oldset)
  69 {
  70         sigset_t mask;
  71
  72         siginitsetinv(&mask, sigmask(SIGKILL));
  73         sigprocmask(SIG_BLOCK, &mask, oldset);
  74 }
  75
  76 static void restore_sigs(sigset_t *oldset)
  77 {
  78         sigprocmask(SIG_SETMASK, oldset, NULL);
  79 }
  80
  81 static void __fuse_get_request(struct fuse_req *req)
  82 {
  83         atomic_inc(&req->count);
  84 }
  85
  86 /* Must be called with > 1 refcount */
  87 static void __fuse_put_request(struct fuse_req *req)
  88 {
  89         BUG_ON(atomic_read(&req->count) < 2);
  90         atomic_dec(&req->count);
  91 }
  92
  93 static void fuse_req_init_context(struct fuse_req *req)
  94 {
  95         req->in.h.uid = current_fsuid();
  96         req->in.h.gid = current_fsgid();
  97         req->in.h.pid = current->pid;
  98 }
  99
 100 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
 101 {
 102         struct fuse_req *req;
 103         sigset_t oldset;
 104         int intr;
 105         int err;
 106
 107         atomic_inc(&fc->num_waiting);
 108         block_sigs(&oldset);
 109         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
 110         restore_sigs(&oldset);
 111         err = -EINTR;
 112         if (intr)
 113                 goto out;
 114
 115         err = -ENOTCONN;
 116         if (!fc->connected)
 117                 goto out;
 118
 119         req = fuse_request_alloc();
 120         err = -ENOMEM;
 121         if (!req)
 122                 goto out;
 123
 124         fuse_req_init_context(req);
 125         req->waiting = 1;
 126         return req;
 127
 128  out:
 129         atomic_dec(&fc->num_waiting);
 130         return ERR_PTR(err);
 131 }
 132 EXPORT_SYMBOL_GPL(fuse_get_req);
 133
 134 /*
 135  * Return request in fuse_file->reserved_req.  However that may
 136  * currently be in use.  If that is the case, wait for it to become
 137  * available.
 138  */
 139 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 140                                          struct file *file)
 141 {
 142         struct fuse_req *req = NULL;
 143         struct fuse_file *ff = file->private_data;
 144
 145         do {
 146                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 147                 spin_lock(&fc->lock);
 148                 if (ff->reserved_req) {
 149                         req = ff->reserved_req;
 150                         ff->reserved_req = NULL;
 151                         get_file(file);
 152                         req->stolen_file = file;
 153                 }
 154                 spin_unlock(&fc->lock);
 155         } while (!req);
 156
 157         return req;
 158 }
 159
 160 /*
 161  * Put stolen request back into fuse_file->reserved_req
 162  */
 163 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 164 {
 165         struct file *file = req->stolen_file;
 166         struct fuse_file *ff = file->private_data;
 167
 168         spin_lock(&fc->lock);
 169         fuse_request_init(req);
 170         BUG_ON(ff->reserved_req);
 171         ff->reserved_req = req;
 172         wake_up_all(&fc->reserved_req_waitq);
 173         spin_unlock(&fc->lock);
 174         fput(file);
 175 }
 176
 177 /*
 178  * Gets a requests for a file operation, always succeeds
 179  *
 180  * This is used for sending the FLUSH request, which must get to
 181  * userspace, due to POSIX locks which may need to be unlocked.
 182  *
 183  * If allocation fails due to OOM, use the reserved request in
 184  * fuse_file.
 185  *
 186  * This is very unlikely to deadlock accidentally, since the
 187  * filesystem should not have it's own file open.  If deadlock is
 188  * intentional, it can still be broken by "aborting" the filesystem.
 189  */
 190 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 191 {
 192         struct fuse_req *req;
 193
 194         atomic_inc(&fc->num_waiting);
 195         wait_event(fc->blocked_waitq, !fc->blocked);
 196         req = fuse_request_alloc();
 197         if (!req)
 198                 req = get_reserved_req(fc, file);
 199
 200         fuse_req_init_context(req);
 201         req->waiting = 1;
 202         return req;
 203 }
 204
 205 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 206 {
 207         if (atomic_dec_and_test(&req->count)) {
 208                 if (req->waiting)
 209                         atomic_dec(&fc->num_waiting);
 210
 211                 if (req->stolen_file)
 212                         put_reserved_req(fc, req);
 213                 else
 214                         fuse_request_free(req);
 215         }
 216 }
 217 EXPORT_SYMBOL_GPL(fuse_put_request);
 218
 219 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 220 {
 221         unsigned nbytes = 0;
 222         unsigned i;
 223
 224         for (i = 0; i < numargs; i++)
 225                 nbytes += args[i].size;
 226
 227         return nbytes;
 228 }
 229
 230 static u64 fuse_get_unique(struct fuse_conn *fc)
 231 {
 232         fc->reqctr++;
 233         /* zero is special */
 234         if (fc->reqctr == 0)
 235                 fc->reqctr = 1;
 236
 237         return fc->reqctr;
 238 }
 239
 240 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 241 {
 242         req->in.h.len = sizeof(struct fuse_in_header) +
 243                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 244         list_add_tail(&req->list, &fc->pending);
 245         req->state = FUSE_REQ_PENDING;
 246         if (!req->waiting) {
 247                 req->waiting = 1;
 248                 atomic_inc(&fc->num_waiting);
 249         }
 250         wake_up(&fc->waitq);
 251         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 252 }
 253
 254 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 255                        u64 nodeid, u64 nlookup)
 256 {
 257         forget->forget_one.nodeid = nodeid;
 258         forget->forget_one.nlookup = nlookup;
 259
 260         spin_lock(&fc->lock);
 261         fc->forget_list_tail->next = forget;
 262         fc->forget_list_tail = forget;
 263         wake_up(&fc->waitq);
 264         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 265         spin_unlock(&fc->lock);
 266 }
 267
 268 static void flush_bg_queue(struct fuse_conn *fc)
 269 {
 270         while (fc->active_background < fc->max_background &&
 271                !list_empty(&fc->bg_queue)) {
 272                 struct fuse_req *req;
 273
 274                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 275                 list_del(&req->list);
 276                 fc->active_background++;
 277                 req->in.h.unique = fuse_get_unique(fc);
 278                 queue_request(fc, req);
 279         }
 280 }
 281
 282 /*
 283  * This function is called when a request is finished.  Either a reply
 284  * has arrived or it was aborted (and not yet sent) or some error
 285  * occurred during communication with userspace, or the device file
 286  * was closed.  The requester thread is woken up (if still waiting),
 287  * the 'end' callback is called if given, else the reference to the
 288  * request is released
 289  *
 290  * Called with fc->lock, unlocks it
 291  */
 292 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 293 __releases(fc->lock)
 294 {
 295         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 296         req->end = NULL;
 297         list_del(&req->list);
 298         list_del(&req->intr_entry);
 299         req->state = FUSE_REQ_FINISHED;
 300         if (req->background) {
 301                 if (fc->num_background == fc->max_background) {
 302                         fc->blocked = 0;
 303                         wake_up_all(&fc->blocked_waitq);
 304                 }
 305                 if (fc->num_background == fc->congestion_threshold &&
 306                     fc->connected && fc->bdi_initialized) {
 307                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 308                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 309                 }
 310                 fc->num_background--;
 311                 fc->active_background--;
 312                 flush_bg_queue(fc);
 313         }
 314         spin_unlock(&fc->lock);
 315         wake_up(&req->waitq);
 316         if (end)
 317                 end(fc, req);
 318         fuse_put_request(fc, req);
 319 }
 320
 321 static void wait_answer_interruptible(struct fuse_conn *fc,
 322                                       struct fuse_req *req)
 323 __releases(fc->lock)
 324 __acquires(fc->lock)
 325 {
 326         if (signal_pending(current))
 327                 return;
 328
 329         spin_unlock(&fc->lock);
 330         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 331         spin_lock(&fc->lock);
 332 }
 333
 334 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 335 {
 336         list_add_tail(&req->intr_entry, &fc->interrupts);
 337         wake_up(&fc->waitq);
 338         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 339 }
 340
 341 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 342 __releases(fc->lock)
 343 __acquires(fc->lock)
 344 {
 345         if (!fc->no_interrupt) {
 346                 /* Any signal may interrupt this */
 347                 wait_answer_interruptible(fc, req);
 348
 349                 if (req->aborted)
 350                         goto aborted;
 351                 if (req->state == FUSE_REQ_FINISHED)
 352                         return;
 353
 354                 req->interrupted = 1;
 355                 if (req->state == FUSE_REQ_SENT)
 356                         queue_interrupt(fc, req);
 357         }
 358
 359         if (!req->force) {
 360                 sigset_t oldset;
 361
 362                 /* Only fatal signals may interrupt this */
 363                 block_sigs(&oldset);
 364                 wait_answer_interruptible(fc, req);
 365                 restore_sigs(&oldset);
 366
 367                 if (req->aborted)
 368                         goto aborted;
 369                 if (req->state == FUSE_REQ_FINISHED)
 370                         return;
 371
 372                 /* Request is not yet in userspace, bail out */
 373                 if (req->state == FUSE_REQ_PENDING) {
 374                         list_del(&req->list);
 375                         __fuse_put_request(req);
 376                         req->out.h.error = -EINTR;
 377                         return;
 378                 }
 379         }
 380
 381         /*
 382          * Either request is already in userspace, or it was forced.
 383          * Wait it out.
 384          */
 385         spin_unlock(&fc->lock);
 386         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 387         spin_lock(&fc->lock);
 388
 389         if (!req->aborted)
 390                 return;
 391
 392  aborted:
 393         BUG_ON(req->state != FUSE_REQ_FINISHED);
 394         if (req->locked) {
 395                 /* This is uninterruptible sleep, because data is
 396                    being copied to/from the buffers of req.  During
 397                    locked state, there mustn't be any filesystem
 398                    operation (e.g. page fault), since that could lead
 399                    to deadlock */
 400                 spin_unlock(&fc->lock);
 401                 wait_event(req->waitq, !req->locked);
 402                 spin_lock(&fc->lock);
 403         }
 404 }
 405
 406 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 407 {
 408         req->isreply = 1;
 409         spin_lock(&fc->lock);
 410         if (!fc->connected)
 411                 req->out.h.error = -ENOTCONN;
 412         else if (fc->conn_error)
 413                 req->out.h.error = -ECONNREFUSED;
 414         else {
 415                 req->in.h.unique = fuse_get_unique(fc);
 416                 queue_request(fc, req);
 417                 /* acquire extra reference, since request is still needed
 418                    after request_end() */
 419                 __fuse_get_request(req);
 420
 421                 request_wait_answer(fc, req);
 422         }
 423         spin_unlock(&fc->lock);
 424 }
 425 EXPORT_SYMBOL_GPL(fuse_request_send);
 426
 427 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 428                                             struct fuse_req *req)
 429 {
 430         req->background = 1;
 431         fc->num_background++;
 432         if (fc->num_background == fc->max_background)
 433                 fc->blocked = 1;
 434         if (fc->num_background == fc->congestion_threshold &&
 435             fc->bdi_initialized) {
 436                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 437                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 438         }
 439         list_add_tail(&req->list, &fc->bg_queue);
 440         flush_bg_queue(fc);
 441 }
 442
 443 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 444 {
 445         spin_lock(&fc->lock);
 446         if (fc->connected) {
 447                 fuse_request_send_nowait_locked(fc, req);
 448                 spin_unlock(&fc->lock);
 449         } else {
 450                 req->out.h.error = -ENOTCONN;
 451                 request_end(fc, req);
 452         }
 453 }
 454
 455 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 456 {
 457         req->isreply = 1;
 458         fuse_request_send_nowait(fc, req);
 459 }
 460 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 461
 462 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 463                                           struct fuse_req *req, u64 unique)
 464 {
 465         int err = -ENODEV;
 466
 467         req->isreply = 0;
 468         req->in.h.unique = unique;
 469         spin_lock(&fc->lock);
 470         if (fc->connected) {
 471                 queue_request(fc, req);
 472                 err = 0;
 473         }
 474         spin_unlock(&fc->lock);
 475
 476         return err;
 477 }
 478
 479 /*
 480  * Called under fc->lock
 481  *
 482  * fc->connected must have been checked previously
 483  */
 484 void fuse_request_send_background_locked(struct fuse_conn *fc,
 485                                          struct fuse_req *req)
 486 {
 487         req->isreply = 1;
 488         fuse_request_send_nowait_locked(fc, req);
 489 }
 490
 491 /*
 492  * Lock the request.  Up to the next unlock_request() there mustn't be
 493  * anything that could cause a page-fault.  If the request was already
 494  * aborted bail out.
 495  */
 496 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 497 {
 498         int err = 0;
 499         if (req) {
 500                 spin_lock(&fc->lock);
 501                 if (req->aborted)
 502                         err = -ENOENT;
 503                 else
 504                         req->locked = 1;
 505                 spin_unlock(&fc->lock);
 506         }
 507         return err;
 508 }
 509
 510 /*
 511  * Unlock request.  If it was aborted during being locked, the
 512  * requester thread is currently waiting for it to be unlocked, so
 513  * wake it up.
 514  */
 515 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 516 {
 517         if (req) {
 518                 spin_lock(&fc->lock);
 519                 req->locked = 0;
 520                 if (req->aborted)
 521                         wake_up(&req->waitq);
 522                 spin_unlock(&fc->lock);
 523         }
 524 }
 525
 526 struct fuse_copy_state {
 527         struct fuse_conn *fc;
 528         int write;
 529         struct fuse_req *req;
 530         const struct iovec *iov;
 531         struct pipe_buffer *pipebufs;
 532         struct pipe_buffer *currbuf;
 533         struct pipe_inode_info *pipe;
 534         unsigned long nr_segs;
 535         unsigned long seglen;
 536         unsigned long addr;
 537         struct page *pg;
 538         void *mapaddr;
 539         void *buf;
 540         unsigned len;
 541         unsigned move_pages:1;
 542 };
 543
 544 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 545                            int write,
 546                            const struct iovec *iov, unsigned long nr_segs)
 547 {
 548         memset(cs, 0, sizeof(*cs));
 549         cs->fc = fc;
 550         cs->write = write;
 551         cs->iov = iov;
 552         cs->nr_segs = nr_segs;
 553 }
 554
 555 /* Unmap and put previous page of userspace buffer */
 556 static void fuse_copy_finish(struct fuse_copy_state *cs)
 557 {
 558         if (cs->currbuf) {
 559                 struct pipe_buffer *buf = cs->currbuf;
 560
 561                 if (!cs->write) {
 562                         buf->ops->unmap(cs->pipe, buf, cs->mapaddr);
 563                 } else {
 564                         kunmap(buf->page);
 565                         buf->len = PAGE_SIZE - cs->len;
 566                 }
 567                 cs->currbuf = NULL;
 568                 cs->mapaddr = NULL;
 569         } else if (cs->mapaddr) {
 570                 kunmap(cs->pg);
 571                 if (cs->write) {
 572                         flush_dcache_page(cs->pg);
 573                         set_page_dirty_lock(cs->pg);
 574                 }
 575                 put_page(cs->pg);
 576                 cs->mapaddr = NULL;
 577         }
 578 }
 579
 580 /*
 581  * Get another pagefull of userspace buffer, and map it to kernel
 582  * address space, and lock request
 583  */
 584 static int fuse_copy_fill(struct fuse_copy_state *cs)
 585 {
 586         unsigned long offset;
 587         int err;
 588
 589         unlock_request(cs->fc, cs->req);
 590         fuse_copy_finish(cs);
 591         if (cs->pipebufs) {
 592                 struct pipe_buffer *buf = cs->pipebufs;
 593
 594                 if (!cs->write) {
 595                         err = buf->ops->confirm(cs->pipe, buf);
 596                         if (err)
 597                                 return err;
 598
 599                         BUG_ON(!cs->nr_segs);
 600                         cs->currbuf = buf;
 601                         cs->mapaddr = buf->ops->map(cs->pipe, buf, 0);
 602                         cs->len = buf->len;
 603                         cs->buf = cs->mapaddr + buf->offset;
 604                         cs->pipebufs++;
 605                         cs->nr_segs--;
 606                 } else {
 607                         struct page *page;
 608
 609                         if (cs->nr_segs == cs->pipe->buffers)
 610                                 return -EIO;
 611
 612                         page = alloc_page(GFP_HIGHUSER);
 613                         if (!page)
 614                                 return -ENOMEM;
 615
 616                         buf->page = page;
 617                         buf->offset = 0;
 618                         buf->len = 0;
 619
 620                         cs->currbuf = buf;
 621                         cs->mapaddr = kmap(page);
 622                         cs->buf = cs->mapaddr;
 623                         cs->len = PAGE_SIZE;
 624                         cs->pipebufs++;
 625                         cs->nr_segs++;
 626                 }
 627         } else {
 628                 if (!cs->seglen) {
 629                         BUG_ON(!cs->nr_segs);
 630                         cs->seglen = cs->iov[0].iov_len;
 631                         cs->addr = (unsigned long) cs->iov[0].iov_base;
 632                         cs->iov++;
 633                         cs->nr_segs--;
 634                 }
 635                 err = get_user_pages_fast(cs->addr, 1, cs->write, &cs->pg);
 636                 if (err < 0)
 637                         return err;
 638                 BUG_ON(err != 1);
 639                 offset = cs->addr % PAGE_SIZE;
 640                 cs->mapaddr = kmap(cs->pg);
 641                 cs->buf = cs->mapaddr + offset;
 642                 cs->len = min(PAGE_SIZE - offset, cs->seglen);
 643                 cs->seglen -= cs->len;
 644                 cs->addr += cs->len;
 645         }
 646
 647         return lock_request(cs->fc, cs->req);
 648 }
 649
 650 /* Do as much copy to/from userspace buffer as we can */
 651 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 652 {
 653         unsigned ncpy = min(*size, cs->len);
 654         if (val) {
 655                 if (cs->write)
 656                         memcpy(cs->buf, *val, ncpy);
 657                 else
 658                         memcpy(*val, cs->buf, ncpy);
 659                 *val += ncpy;
 660         }
 661         *size -= ncpy;
 662         cs->len -= ncpy;
 663         cs->buf += ncpy;
 664         return ncpy;
 665 }
 666
 667 static int fuse_check_page(struct page *page)
 668 {
 669         if (page_mapcount(page) ||
 670             page->mapping != NULL ||
 671             page_count(page) != 1 ||
 672             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 673              ~(1 << PG_locked |
 674                1 << PG_referenced |
 675                1 << PG_uptodate |
 676                1 << PG_lru |
 677                1 << PG_active |
 678                1 << PG_reclaim))) {
 679                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 680                 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);
 681                 return 1;
 682         }
 683         return 0;
 684 }
 685
 686 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 687 {
 688         int err;
 689         struct page *oldpage = *pagep;
 690         struct page *newpage;
 691         struct pipe_buffer *buf = cs->pipebufs;
 692         struct address_space *mapping;
 693         pgoff_t index;
 694
 695         unlock_request(cs->fc, cs->req);
 696         fuse_copy_finish(cs);
 697
 698         err = buf->ops->confirm(cs->pipe, buf);
 699         if (err)
 700                 return err;
 701
 702         BUG_ON(!cs->nr_segs);
 703         cs->currbuf = buf;
 704         cs->len = buf->len;
 705         cs->pipebufs++;
 706         cs->nr_segs--;
 707
 708         if (cs->len != PAGE_SIZE)
 709                 goto out_fallback;
 710
 711         if (buf->ops->steal(cs->pipe, buf) != 0)
 712                 goto out_fallback;
 713
 714         newpage = buf->page;
 715
 716         if (WARN_ON(!PageUptodate(newpage)))
 717                 return -EIO;
 718
 719         ClearPageMappedToDisk(newpage);
 720
 721         if (fuse_check_page(newpage) != 0)
 722                 goto out_fallback_unlock;
 723
 724         mapping = oldpage->mapping;
 725         index = oldpage->index;
 726
 727         /*
 728          * This is a new and locked page, it shouldn't be mapped or
 729          * have any special flags on it
 730          */
 731         if (WARN_ON(page_mapped(oldpage)))
 732                 goto out_fallback_unlock;
 733         if (WARN_ON(page_has_private(oldpage)))
 734                 goto out_fallback_unlock;
 735         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 736                 goto out_fallback_unlock;
 737         if (WARN_ON(PageMlocked(oldpage)))
 738                 goto out_fallback_unlock;
 739
 740         remove_from_page_cache(oldpage);
 741         page_cache_release(oldpage);
 742
 743         err = add_to_page_cache_locked(newpage, mapping, index, GFP_KERNEL);
 744         if (err) {
 745                 printk(KERN_WARNING "fuse_try_move_page: failed to add page");
 746                 goto out_fallback_unlock;
 747         }
 748         page_cache_get(newpage);
 749
 750         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 751                 lru_cache_add_file(newpage);
 752
 753         err = 0;
 754         spin_lock(&cs->fc->lock);
 755         if (cs->req->aborted)
 756                 err = -ENOENT;
 757         else
 758                 *pagep = newpage;
 759         spin_unlock(&cs->fc->lock);
 760
 761         if (err) {
 762                 unlock_page(newpage);
 763                 page_cache_release(newpage);
 764                 return err;
 765         }
 766
 767         unlock_page(oldpage);
 768         page_cache_release(oldpage);
 769         cs->len = 0;
 770
 771         return 0;
 772
 773 out_fallback_unlock:
 774         unlock_page(newpage);
 775 out_fallback:
 776         cs->mapaddr = buf->ops->map(cs->pipe, buf, 1);
 777         cs->buf = cs->mapaddr + buf->offset;
 778
 779         err = lock_request(cs->fc, cs->req);
 780         if (err)
 781                 return err;
 782
 783         return 1;
 784 }
 785
 786 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 787                          unsigned offset, unsigned count)
 788 {
 789         struct pipe_buffer *buf;
 790
 791         if (cs->nr_segs == cs->pipe->buffers)
 792                 return -EIO;
 793
 794         unlock_request(cs->fc, cs->req);
 795         fuse_copy_finish(cs);
 796
 797         buf = cs->pipebufs;
 798         page_cache_get(page);
 799         buf->page = page;
 800         buf->offset = offset;
 801         buf->len = count;
 802
 803         cs->pipebufs++;
 804         cs->nr_segs++;
 805         cs->len = 0;
 806
 807         return 0;
 808 }
 809
 810 /*
 811  * Copy a page in the request to/from the userspace buffer.  Must be
 812  * done atomically
 813  */
 814 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 815                           unsigned offset, unsigned count, int zeroing)
 816 {
 817         int err;
 818         struct page *page = *pagep;
 819
 820         if (page && zeroing && count < PAGE_SIZE)
 821                 clear_highpage(page);
 822
 823         while (count) {
 824                 if (cs->write && cs->pipebufs && page) {
 825                         return fuse_ref_page(cs, page, offset, count);
 826                 } else if (!cs->len) {
 827                         if (cs->move_pages && page &&
 828                             offset == 0 && count == PAGE_SIZE) {
 829                                 err = fuse_try_move_page(cs, pagep);
 830                                 if (err <= 0)
 831                                         return err;
 832                         } else {
 833                                 err = fuse_copy_fill(cs);
 834                                 if (err)
 835                                         return err;
 836                         }
 837                 }
 838                 if (page) {
 839                         void *mapaddr = kmap_atomic(page, KM_USER0);
 840                         void *buf = mapaddr + offset;
 841                         offset += fuse_copy_do(cs, &buf, &count);
 842                         kunmap_atomic(mapaddr, KM_USER0);
 843                 } else
 844                         offset += fuse_copy_do(cs, NULL, &count);
 845         }
 846         if (page && !cs->write)
 847                 flush_dcache_page(page);
 848         return 0;
 849 }
 850
 851 /* Copy pages in the request to/from userspace buffer */
 852 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 853                            int zeroing)
 854 {
 855         unsigned i;
 856         struct fuse_req *req = cs->req;
 857         unsigned offset = req->page_offset;
 858         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 859
 860         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 861                 int err;
 862
 863                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
 864                                      zeroing);
 865                 if (err)
 866                         return err;
 867
 868                 nbytes -= count;
 869                 count = min(nbytes, (unsigned) PAGE_SIZE);
 870                 offset = 0;
 871         }
 872         return 0;
 873 }
 874
 875 /* Copy a single argument in the request to/from userspace buffer */
 876 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 877 {
 878         while (size) {
 879                 if (!cs->len) {
 880                         int err = fuse_copy_fill(cs);
 881                         if (err)
 882                                 return err;
 883                 }
 884                 fuse_copy_do(cs, &val, &size);
 885         }
 886         return 0;
 887 }
 888
 889 /* Copy request arguments to/from userspace buffer */
 890 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 891                           unsigned argpages, struct fuse_arg *args,
 892                           int zeroing)
 893 {
 894         int err = 0;
 895         unsigned i;
 896
 897         for (i = 0; !err && i < numargs; i++)  {
 898                 struct fuse_arg *arg = &args[i];
 899                 if (i == numargs - 1 && argpages)
 900                         err = fuse_copy_pages(cs, arg->size, zeroing);
 901                 else
 902                         err = fuse_copy_one(cs, arg->value, arg->size);
 903         }
 904         return err;
 905 }
 906
 907 static int forget_pending(struct fuse_conn *fc)
 908 {
 909         return fc->forget_list_head.next != NULL;
 910 }
 911
 912 static int request_pending(struct fuse_conn *fc)
 913 {
 914         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts) ||
 915                 forget_pending(fc);
 916 }
 917
 918 /* Wait until a request is available on the pending list */
 919 static void request_wait(struct fuse_conn *fc)
 920 __releases(fc->lock)
 921 __acquires(fc->lock)
 922 {
 923         DECLARE_WAITQUEUE(wait, current);
 924
 925         add_wait_queue_exclusive(&fc->waitq, &wait);
 926         while (fc->connected && !request_pending(fc)) {
 927                 set_current_state(TASK_INTERRUPTIBLE);
 928                 if (signal_pending(current))
 929                         break;
 930
 931                 spin_unlock(&fc->lock);
 932                 schedule();
 933                 spin_lock(&fc->lock);
 934         }
 935         set_current_state(TASK_RUNNING);
 936         remove_wait_queue(&fc->waitq, &wait);
 937 }
 938
 939 /*
 940  * Transfer an interrupt request to userspace
 941  *
 942  * Unlike other requests this is assembled on demand, without a need
 943  * to allocate a separate fuse_req structure.
 944  *
 945  * Called with fc->lock held, releases it
 946  */
 947 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
 948                                size_t nbytes, struct fuse_req *req)
 949 __releases(fc->lock)
 950 {
 951         struct fuse_in_header ih;
 952         struct fuse_interrupt_in arg;
 953         unsigned reqsize = sizeof(ih) + sizeof(arg);
 954         int err;
 955
 956         list_del_init(&req->intr_entry);
 957         req->intr_unique = fuse_get_unique(fc);
 958         memset(&ih, 0, sizeof(ih));
 959         memset(&arg, 0, sizeof(arg));
 960         ih.len = reqsize;
 961         ih.opcode = FUSE_INTERRUPT;
 962         ih.unique = req->intr_unique;
 963         arg.unique = req->in.h.unique;
 964
 965         spin_unlock(&fc->lock);
 966         if (nbytes < reqsize)
 967                 return -EINVAL;
 968
 969         err = fuse_copy_one(cs, &ih, sizeof(ih));
 970         if (!err)
 971                 err = fuse_copy_one(cs, &arg, sizeof(arg));
 972         fuse_copy_finish(cs);
 973
 974         return err ? err : reqsize;
 975 }
 976
 977 static struct fuse_forget_link *dequeue_forget(struct fuse_conn *fc,
 978                                                unsigned max,
 979                                                unsigned *countp)
 980 {
 981         struct fuse_forget_link *head = fc->forget_list_head.next;
 982         struct fuse_forget_link **newhead = &head;
 983         unsigned count;
 984
 985         for (count = 0; *newhead != NULL && count < max; count++)
 986                 newhead = &(*newhead)->next;
 987
 988         fc->forget_list_head.next = *newhead;
 989         *newhead = NULL;
 990         if (fc->forget_list_head.next == NULL)
 991                 fc->forget_list_tail = &fc->forget_list_head;
 992
 993         if (countp != NULL)
 994                 *countp = count;
 995
 996         return head;
 997 }
 998
 999 static int fuse_read_single_forget(struct fuse_conn *fc,
1000                                    struct fuse_copy_state *cs,
1001                                    size_t nbytes)
1002 __releases(fc->lock)
1003 {
1004         int err;
1005         struct fuse_forget_link *forget = dequeue_forget(fc, 1, NULL);
1006         struct fuse_forget_in arg = {
1007                 .nlookup = forget->forget_one.nlookup,
1008         };
1009         struct fuse_in_header ih = {
1010                 .opcode = FUSE_FORGET,
1011                 .nodeid = forget->forget_one.nodeid,
1012                 .unique = fuse_get_unique(fc),
1013                 .len = sizeof(ih) + sizeof(arg),
1014         };
1015
1016         spin_unlock(&fc->lock);
1017         kfree(forget);
1018         if (nbytes < ih.len)
1019                 return -EINVAL;
1020
1021         err = fuse_copy_one(cs, &ih, sizeof(ih));
1022         if (!err)
1023                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1024         fuse_copy_finish(cs);
1025
1026         if (err)
1027                 return err;
1028
1029         return ih.len;
1030 }
1031
1032 static int fuse_read_batch_forget(struct fuse_conn *fc,
1033                                    struct fuse_copy_state *cs, size_t nbytes)
1034 __releases(fc->lock)
1035 {
1036         int err;
1037         unsigned max_forgets;
1038         unsigned count;
1039         struct fuse_forget_link *head;
1040         struct fuse_batch_forget_in arg = { .count = 0 };
1041         struct fuse_in_header ih = {
1042                 .opcode = FUSE_BATCH_FORGET,
1043                 .unique = fuse_get_unique(fc),
1044                 .len = sizeof(ih) + sizeof(arg),
1045         };
1046
1047         if (nbytes < ih.len) {
1048                 spin_unlock(&fc->lock);
1049                 return -EINVAL;
1050         }
1051
1052         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1053         head = dequeue_forget(fc, max_forgets, &count);
1054         spin_unlock(&fc->lock);
1055
1056         arg.count = count;
1057         ih.len += count * sizeof(struct fuse_forget_one);
1058         err = fuse_copy_one(cs, &ih, sizeof(ih));
1059         if (!err)
1060                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1061
1062         while (head) {
1063                 struct fuse_forget_link *forget = head;
1064
1065                 if (!err) {
1066                         err = fuse_copy_one(cs, &forget->forget_one,
1067                                             sizeof(forget->forget_one));
1068                 }
1069                 head = forget->next;
1070                 kfree(forget);
1071         }
1072
1073         fuse_copy_finish(cs);
1074
1075         if (err)
1076                 return err;
1077
1078         return ih.len;
1079 }
1080
1081 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_copy_state *cs,
1082                             size_t nbytes)
1083 __releases(fc->lock)
1084 {
1085         if (fc->minor < 16 || fc->forget_list_head.next->next == NULL)
1086                 return fuse_read_single_forget(fc, cs, nbytes);
1087         else
1088                 return fuse_read_batch_forget(fc, cs, nbytes);
1089 }
1090
1091 /*
1092  * Read a single request into the userspace filesystem's buffer.  This
1093  * function waits until a request is available, then removes it from
1094  * the pending list and copies request data to userspace buffer.  If
1095  * no reply is needed (FORGET) or request has been aborted or there
1096  * was an error during the copying then it's finished by calling
1097  * request_end().  Otherwise add it to the processing list, and set
1098  * the 'sent' flag.
1099  */
1100 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1101                                 struct fuse_copy_state *cs, size_t nbytes)
1102 {
1103         int err;
1104         struct fuse_req *req;
1105         struct fuse_in *in;
1106         unsigned reqsize;
1107
1108  restart:
1109         spin_lock(&fc->lock);
1110         err = -EAGAIN;
1111         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
1112             !request_pending(fc))
1113                 goto err_unlock;
1114
1115         request_wait(fc);
1116         err = -ENODEV;
1117         if (!fc->connected)
1118                 goto err_unlock;
1119         err = -ERESTARTSYS;
1120         if (!request_pending(fc))
1121                 goto err_unlock;
1122
1123         if (!list_empty(&fc->interrupts)) {
1124                 req = list_entry(fc->interrupts.next, struct fuse_req,
1125                                  intr_entry);
1126                 return fuse_read_interrupt(fc, cs, nbytes, req);
1127         }
1128
1129         if (forget_pending(fc)) {
1130                 if (list_empty(&fc->pending) || fc->forget_batch-- > 0)
1131                         return fuse_read_forget(fc, cs, nbytes);
1132
1133                 if (fc->forget_batch <= -8)
1134                         fc->forget_batch = 16;
1135         }
1136
1137         req = list_entry(fc->pending.next, struct fuse_req, list);
1138         req->state = FUSE_REQ_READING;
1139         list_move(&req->list, &fc->io);
1140
1141         in = &req->in;
1142         reqsize = in->h.len;
1143         /* If request is too large, reply with an error and restart the read */
1144         if (nbytes < reqsize) {
1145                 req->out.h.error = -EIO;
1146                 /* SETXATTR is special, since it may contain too large data */
1147                 if (in->h.opcode == FUSE_SETXATTR)
1148                         req->out.h.error = -E2BIG;
1149                 request_end(fc, req);
1150                 goto restart;
1151         }
1152         spin_unlock(&fc->lock);
1153         cs->req = req;
1154         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1155         if (!err)
1156                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1157                                      (struct fuse_arg *) in->args, 0);
1158         fuse_copy_finish(cs);
1159         spin_lock(&fc->lock);
1160         req->locked = 0;
1161         if (req->aborted) {
1162                 request_end(fc, req);
1163                 return -ENODEV;
1164         }
1165         if (err) {
1166                 req->out.h.error = -EIO;
1167                 request_end(fc, req);
1168                 return err;
1169         }
1170         if (!req->isreply)
1171                 request_end(fc, req);
1172         else {
1173                 req->state = FUSE_REQ_SENT;
1174                 list_move_tail(&req->list, &fc->processing);
1175                 if (req->interrupted)
1176                         queue_interrupt(fc, req);
1177                 spin_unlock(&fc->lock);
1178         }
1179         return reqsize;
1180
1181  err_unlock:
1182         spin_unlock(&fc->lock);
1183         return err;
1184 }
1185
1186 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
1187                               unsigned long nr_segs, loff_t pos)
1188 {
1189         struct fuse_copy_state cs;
1190         struct file *file = iocb->ki_filp;
1191         struct fuse_conn *fc = fuse_get_conn(file);
1192         if (!fc)
1193                 return -EPERM;
1194
1195         fuse_copy_init(&cs, fc, 1, iov, nr_segs);
1196
1197         return fuse_dev_do_read(fc, file, &cs, iov_length(iov, nr_segs));
1198 }
1199
1200 static int fuse_dev_pipe_buf_steal(struct pipe_inode_info *pipe,
1201                                    struct pipe_buffer *buf)
1202 {
1203         return 1;
1204 }
1205
1206 static const struct pipe_buf_operations fuse_dev_pipe_buf_ops = {
1207         .can_merge = 0,
1208         .map = generic_pipe_buf_map,
1209         .unmap = generic_pipe_buf_unmap,
1210         .confirm = generic_pipe_buf_confirm,
1211         .release = generic_pipe_buf_release,
1212         .steal = fuse_dev_pipe_buf_steal,
1213         .get = generic_pipe_buf_get,
1214 };
1215
1216 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1217                                     struct pipe_inode_info *pipe,
1218                                     size_t len, unsigned int flags)
1219 {
1220         int ret;
1221         int page_nr = 0;
1222         int do_wakeup = 0;
1223         struct pipe_buffer *bufs;
1224         struct fuse_copy_state cs;
1225         struct fuse_conn *fc = fuse_get_conn(in);
1226         if (!fc)
1227                 return -EPERM;
1228
1229         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1230         if (!bufs)
1231                 return -ENOMEM;
1232
1233         fuse_copy_init(&cs, fc, 1, NULL, 0);
1234         cs.pipebufs = bufs;
1235         cs.pipe = pipe;
1236         ret = fuse_dev_do_read(fc, in, &cs, len);
1237         if (ret < 0)
1238                 goto out;
1239
1240         ret = 0;
1241         pipe_lock(pipe);
1242
1243         if (!pipe->readers) {
1244                 send_sig(SIGPIPE, current, 0);
1245                 if (!ret)
1246                         ret = -EPIPE;
1247                 goto out_unlock;
1248         }
1249
1250         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1251                 ret = -EIO;
1252                 goto out_unlock;
1253         }
1254
1255         while (page_nr < cs.nr_segs) {
1256                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1257                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1258
1259                 buf->page = bufs[page_nr].page;
1260                 buf->offset = bufs[page_nr].offset;
1261                 buf->len = bufs[page_nr].len;
1262                 buf->ops = &fuse_dev_pipe_buf_ops;
1263
1264                 pipe->nrbufs++;
1265                 page_nr++;
1266                 ret += buf->len;
1267
1268                 if (pipe->inode)
1269                         do_wakeup = 1;
1270         }
1271
1272 out_unlock:
1273         pipe_unlock(pipe);
1274
1275         if (do_wakeup) {
1276                 smp_mb();
1277                 if (waitqueue_active(&pipe->wait))
1278                         wake_up_interruptible(&pipe->wait);
1279                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1280         }
1281
1282 out:
1283         for (; page_nr < cs.nr_segs; page_nr++)
1284                 page_cache_release(bufs[page_nr].page);
1285
1286         kfree(bufs);
1287         return ret;
1288 }
1289
1290 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1291                             struct fuse_copy_state *cs)
1292 {
1293         struct fuse_notify_poll_wakeup_out outarg;
1294         int err = -EINVAL;
1295
1296         if (size != sizeof(outarg))
1297                 goto err;
1298
1299         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1300         if (err)
1301                 goto err;
1302
1303         fuse_copy_finish(cs);
1304         return fuse_notify_poll_wakeup(fc, &outarg);
1305
1306 err:
1307         fuse_copy_finish(cs);
1308         return err;
1309 }
1310
1311 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1312                                    struct fuse_copy_state *cs)
1313 {
1314         struct fuse_notify_inval_inode_out outarg;
1315         int err = -EINVAL;
1316
1317         if (size != sizeof(outarg))
1318                 goto err;
1319
1320         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1321         if (err)
1322                 goto err;
1323         fuse_copy_finish(cs);
1324
1325         down_read(&fc->killsb);
1326         err = -ENOENT;
1327         if (fc->sb) {
1328                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1329                                                outarg.off, outarg.len);
1330         }
1331         up_read(&fc->killsb);
1332         return err;
1333
1334 err:
1335         fuse_copy_finish(cs);
1336         return err;
1337 }
1338
1339 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1340                                    struct fuse_copy_state *cs)
1341 {
1342         struct fuse_notify_inval_entry_out outarg;
1343         int err = -ENOMEM;
1344         char *buf;
1345         struct qstr name;
1346
1347         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1348         if (!buf)
1349                 goto err;
1350
1351         err = -EINVAL;
1352         if (size < sizeof(outarg))
1353                 goto err;
1354
1355         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1356         if (err)
1357                 goto err;
1358
1359         err = -ENAMETOOLONG;
1360         if (outarg.namelen > FUSE_NAME_MAX)
1361                 goto err;
1362
1363         name.name = buf;
1364         name.len = outarg.namelen;
1365         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1366         if (err)
1367                 goto err;
1368         fuse_copy_finish(cs);
1369         buf[outarg.namelen] = 0;
1370         name.hash = full_name_hash(name.name, name.len);
1371
1372         down_read(&fc->killsb);
1373         err = -ENOENT;
1374         if (fc->sb)
1375                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, &name);
1376         up_read(&fc->killsb);
1377         kfree(buf);
1378         return err;
1379
1380 err:
1381         kfree(buf);
1382         fuse_copy_finish(cs);
1383         return err;
1384 }
1385
1386 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1387                              struct fuse_copy_state *cs)
1388 {
1389         struct fuse_notify_store_out outarg;
1390         struct inode *inode;
1391         struct address_space *mapping;
1392         u64 nodeid;
1393         int err;
1394         pgoff_t index;
1395         unsigned int offset;
1396         unsigned int num;
1397         loff_t file_size;
1398         loff_t end;
1399
1400         err = -EINVAL;
1401         if (size < sizeof(outarg))
1402                 goto out_finish;
1403
1404         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1405         if (err)
1406                 goto out_finish;
1407
1408         err = -EINVAL;
1409         if (size - sizeof(outarg) != outarg.size)
1410                 goto out_finish;
1411
1412         nodeid = outarg.nodeid;
1413
1414         down_read(&fc->killsb);
1415
1416         err = -ENOENT;
1417         if (!fc->sb)
1418                 goto out_up_killsb;
1419
1420         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1421         if (!inode)
1422                 goto out_up_killsb;
1423
1424         mapping = inode->i_mapping;
1425         index = outarg.offset >> PAGE_CACHE_SHIFT;
1426         offset = outarg.offset & ~PAGE_CACHE_MASK;
1427         file_size = i_size_read(inode);
1428         end = outarg.offset + outarg.size;
1429         if (end > file_size) {
1430                 file_size = end;
1431                 fuse_write_update_size(inode, file_size);
1432         }
1433
1434         num = outarg.size;
1435         while (num) {
1436                 struct page *page;
1437                 unsigned int this_num;
1438
1439                 err = -ENOMEM;
1440                 page = find_or_create_page(mapping, index,
1441                                            mapping_gfp_mask(mapping));
1442                 if (!page)
1443                         goto out_iput;
1444
1445                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1446                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1447                 if (!err && offset == 0 && (num != 0 || file_size == end))
1448                         SetPageUptodate(page);
1449                 unlock_page(page);
1450                 page_cache_release(page);
1451
1452                 if (err)
1453                         goto out_iput;
1454
1455                 num -= this_num;
1456                 offset = 0;
1457                 index++;
1458         }
1459
1460         err = 0;
1461
1462 out_iput:
1463         iput(inode);
1464 out_up_killsb:
1465         up_read(&fc->killsb);
1466 out_finish:
1467         fuse_copy_finish(cs);
1468         return err;
1469 }
1470
1471 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1472 {
1473         release_pages(req->pages, req->num_pages, 0);
1474 }
1475
1476 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1477                          struct fuse_notify_retrieve_out *outarg)
1478 {
1479         int err;
1480         struct address_space *mapping = inode->i_mapping;
1481         struct fuse_req *req;
1482         pgoff_t index;
1483         loff_t file_size;
1484         unsigned int num;
1485         unsigned int offset;
1486         size_t total_len = 0;
1487
1488         req = fuse_get_req(fc);
1489         if (IS_ERR(req))
1490                 return PTR_ERR(req);
1491
1492         offset = outarg->offset & ~PAGE_CACHE_MASK;
1493
1494         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1495         req->in.h.nodeid = outarg->nodeid;
1496         req->in.numargs = 2;
1497         req->in.argpages = 1;
1498         req->page_offset = offset;
1499         req->end = fuse_retrieve_end;
1500
1501         index = outarg->offset >> PAGE_CACHE_SHIFT;
1502         file_size = i_size_read(inode);
1503         num = outarg->size;
1504         if (outarg->offset > file_size)
1505                 num = 0;
1506         else if (outarg->offset + num > file_size)
1507                 num = file_size - outarg->offset;
1508
1509         while (num) {
1510                 struct page *page;
1511                 unsigned int this_num;
1512
1513                 page = find_get_page(mapping, index);
1514                 if (!page)
1515                         break;
1516
1517                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1518                 req->pages[req->num_pages] = page;
1519                 req->num_pages++;
1520
1521                 num -= this_num;
1522                 total_len += this_num;
1523         }
1524         req->misc.retrieve_in.offset = outarg->offset;
1525         req->misc.retrieve_in.size = total_len;
1526         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1527         req->in.args[0].value = &req->misc.retrieve_in;
1528         req->in.args[1].size = total_len;
1529
1530         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1531         if (err)
1532                 fuse_retrieve_end(fc, req);
1533
1534         return err;
1535 }
1536
1537 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1538                                 struct fuse_copy_state *cs)
1539 {
1540         struct fuse_notify_retrieve_out outarg;
1541         struct inode *inode;
1542         int err;
1543
1544         err = -EINVAL;
1545         if (size != sizeof(outarg))
1546                 goto copy_finish;
1547
1548         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1549         if (err)
1550                 goto copy_finish;
1551
1552         fuse_copy_finish(cs);
1553
1554         down_read(&fc->killsb);
1555         err = -ENOENT;
1556         if (fc->sb) {
1557                 u64 nodeid = outarg.nodeid;
1558
1559                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1560                 if (inode) {
1561                         err = fuse_retrieve(fc, inode, &outarg);
1562                         iput(inode);
1563                 }
1564         }
1565         up_read(&fc->killsb);
1566
1567         return err;
1568
1569 copy_finish:
1570         fuse_copy_finish(cs);
1571         return err;
1572 }
1573
1574 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1575                        unsigned int size, struct fuse_copy_state *cs)
1576 {
1577         switch (code) {
1578         case FUSE_NOTIFY_POLL:
1579                 return fuse_notify_poll(fc, size, cs);
1580
1581         case FUSE_NOTIFY_INVAL_INODE:
1582                 return fuse_notify_inval_inode(fc, size, cs);
1583
1584         case FUSE_NOTIFY_INVAL_ENTRY:
1585                 return fuse_notify_inval_entry(fc, size, cs);
1586
1587         case FUSE_NOTIFY_STORE:
1588                 return fuse_notify_store(fc, size, cs);
1589
1590         case FUSE_NOTIFY_RETRIEVE:
1591                 return fuse_notify_retrieve(fc, size, cs);
1592
1593         default:
1594                 fuse_copy_finish(cs);
1595                 return -EINVAL;
1596         }
1597 }
1598
1599 /* Look up request on processing list by unique ID */
1600 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1601 {
1602         struct list_head *entry;
1603
1604         list_for_each(entry, &fc->processing) {
1605                 struct fuse_req *req;
1606                 req = list_entry(entry, struct fuse_req, list);
1607                 if (req->in.h.unique == unique || req->intr_unique == unique)
1608                         return req;
1609         }
1610         return NULL;
1611 }
1612
1613 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1614                          unsigned nbytes)
1615 {
1616         unsigned reqsize = sizeof(struct fuse_out_header);
1617
1618         if (out->h.error)
1619                 return nbytes != reqsize ? -EINVAL : 0;
1620
1621         reqsize += len_args(out->numargs, out->args);
1622
1623         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1624                 return -EINVAL;
1625         else if (reqsize > nbytes) {
1626                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1627                 unsigned diffsize = reqsize - nbytes;
1628                 if (diffsize > lastarg->size)
1629                         return -EINVAL;
1630                 lastarg->size -= diffsize;
1631         }
1632         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1633                               out->page_zeroing);
1634 }
1635
1636 /*
1637  * Write a single reply to a request.  First the header is copied from
1638  * the write buffer.  The request is then searched on the processing
1639  * list by the unique ID found in the header.  If found, then remove
1640  * it from the list and copy the rest of the buffer to the request.
1641  * The request is finished by calling request_end()
1642  */
1643 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1644                                  struct fuse_copy_state *cs, size_t nbytes)
1645 {
1646         int err;
1647         struct fuse_req *req;
1648         struct fuse_out_header oh;
1649
1650         if (nbytes < sizeof(struct fuse_out_header))
1651                 return -EINVAL;
1652
1653         err = fuse_copy_one(cs, &oh, sizeof(oh));
1654         if (err)
1655                 goto err_finish;
1656
1657         err = -EINVAL;
1658         if (oh.len != nbytes)
1659                 goto err_finish;
1660
1661         /*
1662          * Zero oh.unique indicates unsolicited notification message
1663          * and error contains notification code.
1664          */
1665         if (!oh.unique) {
1666                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1667                 return err ? err : nbytes;
1668         }
1669
1670         err = -EINVAL;
1671         if (oh.error <= -1000 || oh.error > 0)
1672                 goto err_finish;
1673
1674         spin_lock(&fc->lock);
1675         err = -ENOENT;
1676         if (!fc->connected)
1677                 goto err_unlock;
1678
1679         req = request_find(fc, oh.unique);
1680         if (!req)
1681                 goto err_unlock;
1682
1683         if (req->aborted) {
1684                 spin_unlock(&fc->lock);
1685                 fuse_copy_finish(cs);
1686                 spin_lock(&fc->lock);
1687                 request_end(fc, req);
1688                 return -ENOENT;
1689         }
1690         /* Is it an interrupt reply? */
1691         if (req->intr_unique == oh.unique) {
1692                 err = -EINVAL;
1693                 if (nbytes != sizeof(struct fuse_out_header))
1694                         goto err_unlock;
1695
1696                 if (oh.error == -ENOSYS)
1697                         fc->no_interrupt = 1;
1698                 else if (oh.error == -EAGAIN)
1699                         queue_interrupt(fc, req);
1700
1701                 spin_unlock(&fc->lock);
1702                 fuse_copy_finish(cs);
1703                 return nbytes;
1704         }
1705
1706         req->state = FUSE_REQ_WRITING;
1707         list_move(&req->list, &fc->io);
1708         req->out.h = oh;
1709         req->locked = 1;
1710         cs->req = req;
1711         if (!req->out.page_replace)
1712                 cs->move_pages = 0;
1713         spin_unlock(&fc->lock);
1714
1715         err = copy_out_args(cs, &req->out, nbytes);
1716         fuse_copy_finish(cs);
1717
1718         spin_lock(&fc->lock);
1719         req->locked = 0;
1720         if (!err) {
1721                 if (req->aborted)
1722                         err = -ENOENT;
1723         } else if (!req->aborted)
1724                 req->out.h.error = -EIO;
1725         request_end(fc, req);
1726
1727         return err ? err : nbytes;
1728
1729  err_unlock:
1730         spin_unlock(&fc->lock);
1731  err_finish:
1732         fuse_copy_finish(cs);
1733         return err;
1734 }
1735
1736 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
1737                               unsigned long nr_segs, loff_t pos)
1738 {
1739         struct fuse_copy_state cs;
1740         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1741         if (!fc)
1742                 return -EPERM;
1743
1744         fuse_copy_init(&cs, fc, 0, iov, nr_segs);
1745
1746         return fuse_dev_do_write(fc, &cs, iov_length(iov, nr_segs));
1747 }
1748
1749 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1750                                      struct file *out, loff_t *ppos,
1751                                      size_t len, unsigned int flags)
1752 {
1753         unsigned nbuf;
1754         unsigned idx;
1755         struct pipe_buffer *bufs;
1756         struct fuse_copy_state cs;
1757         struct fuse_conn *fc;
1758         size_t rem;
1759         ssize_t ret;
1760
1761         fc = fuse_get_conn(out);
1762         if (!fc)
1763                 return -EPERM;
1764
1765         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1766         if (!bufs)
1767                 return -ENOMEM;
1768
1769         pipe_lock(pipe);
1770         nbuf = 0;
1771         rem = 0;
1772         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1773                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1774
1775         ret = -EINVAL;
1776         if (rem < len) {
1777                 pipe_unlock(pipe);
1778                 goto out;
1779         }
1780
1781         rem = len;
1782         while (rem) {
1783                 struct pipe_buffer *ibuf;
1784                 struct pipe_buffer *obuf;
1785
1786                 BUG_ON(nbuf >= pipe->buffers);
1787                 BUG_ON(!pipe->nrbufs);
1788                 ibuf = &pipe->bufs[pipe->curbuf];
1789                 obuf = &bufs[nbuf];
1790
1791                 if (rem >= ibuf->len) {
1792                         *obuf = *ibuf;
1793                         ibuf->ops = NULL;
1794                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1795                         pipe->nrbufs--;
1796                 } else {
1797                         ibuf->ops->get(pipe, ibuf);
1798                         *obuf = *ibuf;
1799                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1800                         obuf->len = rem;
1801                         ibuf->offset += obuf->len;
1802                         ibuf->len -= obuf->len;
1803                 }
1804                 nbuf++;
1805                 rem -= obuf->len;
1806         }
1807         pipe_unlock(pipe);
1808
1809         fuse_copy_init(&cs, fc, 0, NULL, nbuf);
1810         cs.pipebufs = bufs;
1811         cs.pipe = pipe;
1812
1813         if (flags & SPLICE_F_MOVE)
1814                 cs.move_pages = 1;
1815
1816         ret = fuse_dev_do_write(fc, &cs, len);
1817
1818         for (idx = 0; idx < nbuf; idx++) {
1819                 struct pipe_buffer *buf = &bufs[idx];
1820                 buf->ops->release(pipe, buf);
1821         }
1822 out:
1823         kfree(bufs);
1824         return ret;
1825 }
1826
1827 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1828 {
1829         unsigned mask = POLLOUT | POLLWRNORM;
1830         struct fuse_conn *fc = fuse_get_conn(file);
1831         if (!fc)
1832                 return POLLERR;
1833
1834         poll_wait(file, &fc->waitq, wait);
1835
1836         spin_lock(&fc->lock);
1837         if (!fc->connected)
1838                 mask = POLLERR;
1839         else if (request_pending(fc))
1840                 mask |= POLLIN | POLLRDNORM;
1841         spin_unlock(&fc->lock);
1842
1843         return mask;
1844 }
1845
1846 /*
1847  * Abort all requests on the given list (pending or processing)
1848  *
1849  * This function releases and reacquires fc->lock
1850  */
1851 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1852 __releases(fc->lock)
1853 __acquires(fc->lock)
1854 {
1855         while (!list_empty(head)) {
1856                 struct fuse_req *req;
1857                 req = list_entry(head->next, struct fuse_req, list);
1858                 req->out.h.error = -ECONNABORTED;
1859                 request_end(fc, req);
1860                 spin_lock(&fc->lock);
1861         }
1862 }
1863
1864 /*
1865  * Abort requests under I/O
1866  *
1867  * The requests are set to aborted and finished, and the request
1868  * waiter is woken up.  This will make request_wait_answer() wait
1869  * until the request is unlocked and then return.
1870  *
1871  * If the request is asynchronous, then the end function needs to be
1872  * called after waiting for the request to be unlocked (if it was
1873  * locked).
1874  */
1875 static void end_io_requests(struct fuse_conn *fc)
1876 __releases(fc->lock)
1877 __acquires(fc->lock)
1878 {
1879         while (!list_empty(&fc->io)) {
1880                 struct fuse_req *req =
1881                         list_entry(fc->io.next, struct fuse_req, list);
1882                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1883
1884                 req->aborted = 1;
1885                 req->out.h.error = -ECONNABORTED;
1886                 req->state = FUSE_REQ_FINISHED;
1887                 list_del_init(&req->list);
1888                 wake_up(&req->waitq);
1889                 if (end) {
1890                         req->end = NULL;
1891                         __fuse_get_request(req);
1892                         spin_unlock(&fc->lock);
1893                         wait_event(req->waitq, !req->locked);
1894                         end(fc, req);
1895                         fuse_put_request(fc, req);
1896                         spin_lock(&fc->lock);
1897                 }
1898         }
1899 }
1900
1901 static void end_queued_requests(struct fuse_conn *fc)
1902 __releases(fc->lock)
1903 __acquires(fc->lock)
1904 {
1905         fc->max_background = UINT_MAX;
1906         flush_bg_queue(fc);
1907         end_requests(fc, &fc->pending);
1908         end_requests(fc, &fc->processing);
1909         while (forget_pending(fc))
1910                 kfree(dequeue_forget(fc, 1, NULL));
1911 }
1912
1913 /*
1914  * Abort all requests.
1915  *
1916  * Emergency exit in case of a malicious or accidental deadlock, or
1917  * just a hung filesystem.
1918  *
1919  * The same effect is usually achievable through killing the
1920  * filesystem daemon and all users of the filesystem.  The exception
1921  * is the combination of an asynchronous request and the tricky
1922  * deadlock (see Documentation/filesystems/fuse.txt).
1923  *
1924  * During the aborting, progression of requests from the pending and
1925  * processing lists onto the io list, and progression of new requests
1926  * onto the pending list is prevented by req->connected being false.
1927  *
1928  * Progression of requests under I/O to the processing list is
1929  * prevented by the req->aborted flag being true for these requests.
1930  * For this reason requests on the io list must be aborted first.
1931  */
1932 void fuse_abort_conn(struct fuse_conn *fc)
1933 {
1934         spin_lock(&fc->lock);
1935         if (fc->connected) {
1936                 fc->connected = 0;
1937                 fc->blocked = 0;
1938                 end_io_requests(fc);
1939                 end_queued_requests(fc);
1940                 wake_up_all(&fc->waitq);
1941                 wake_up_all(&fc->blocked_waitq);
1942                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1943         }
1944         spin_unlock(&fc->lock);
1945 }
1946 EXPORT_SYMBOL_GPL(fuse_abort_conn);
1947
1948 int fuse_dev_release(struct inode *inode, struct file *file)
1949 {
1950         struct fuse_conn *fc = fuse_get_conn(file);
1951         if (fc) {
1952                 spin_lock(&fc->lock);
1953                 fc->connected = 0;
1954                 fc->blocked = 0;
1955                 end_queued_requests(fc);
1956                 wake_up_all(&fc->blocked_waitq);
1957                 spin_unlock(&fc->lock);
1958                 fuse_conn_put(fc);
1959         }
1960
1961         return 0;
1962 }
1963 EXPORT_SYMBOL_GPL(fuse_dev_release);
1964
1965 static int fuse_dev_fasync(int fd, struct file *file, int on)
1966 {
1967         struct fuse_conn *fc = fuse_get_conn(file);
1968         if (!fc)
1969                 return -EPERM;
1970
1971         /* No locking - fasync_helper does its own locking */
1972         return fasync_helper(fd, file, on, &fc->fasync);
1973 }
1974
1975 const struct file_operations fuse_dev_operations = {
1976         .owner          = THIS_MODULE,
1977         .llseek         = no_llseek,
1978         .read           = do_sync_read,
1979         .aio_read       = fuse_dev_read,
1980         .splice_read    = fuse_dev_splice_read,
1981         .write          = do_sync_write,
1982         .aio_write      = fuse_dev_write,
1983         .splice_write   = fuse_dev_splice_write,
1984         .poll           = fuse_dev_poll,
1985         .release        = fuse_dev_release,
1986         .fasync         = fuse_dev_fasync,
1987 };
1988 EXPORT_SYMBOL_GPL(fuse_dev_operations);
1989
1990 static struct miscdevice fuse_miscdevice = {
1991         .minor = FUSE_MINOR,
1992         .name  = "fuse",
1993         .fops = &fuse_dev_operations,
1994 };
1995
1996 int __init fuse_dev_init(void)
1997 {
1998         int err = -ENOMEM;
1999         fuse_req_cachep = kmem_cache_create("fuse_request",
2000                                             sizeof(struct fuse_req),
2001                                             0, 0, NULL);
2002         if (!fuse_req_cachep)
2003                 goto out;
2004
2005         err = misc_register(&fuse_miscdevice);
2006         if (err)
2007                 goto out_cache_clean;
2008
2009         return 0;
2010
2011  out_cache_clean:
2012         kmem_cache_destroy(fuse_req_cachep);
2013  out:
2014         return err;
2015 }
2016
2017 void fuse_dev_cleanup(void)
2018 {
2019         misc_deregister(&fuse_miscdevice);
2020         kmem_cache_destroy(fuse_req_cachep);
2021 }