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         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 741         if (err) {
 742                 unlock_page(newpage);
 743                 return err;
 744         }
 745
 746         page_cache_get(newpage);
 747
 748         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 749                 lru_cache_add_file(newpage);
 750
 751         err = 0;
 752         spin_lock(&cs->fc->lock);
 753         if (cs->req->aborted)
 754                 err = -ENOENT;
 755         else
 756                 *pagep = newpage;
 757         spin_unlock(&cs->fc->lock);
 758
 759         if (err) {
 760                 unlock_page(newpage);
 761                 page_cache_release(newpage);
 762                 return err;
 763         }
 764
 765         unlock_page(oldpage);
 766         page_cache_release(oldpage);
 767         cs->len = 0;
 768
 769         return 0;
 770
 771 out_fallback_unlock:
 772         unlock_page(newpage);
 773 out_fallback:
 774         cs->mapaddr = buf->ops->map(cs->pipe, buf, 1);
 775         cs->buf = cs->mapaddr + buf->offset;
 776
 777         err = lock_request(cs->fc, cs->req);
 778         if (err)
 779                 return err;
 780
 781         return 1;
 782 }
 783
 784 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 785                          unsigned offset, unsigned count)
 786 {
 787         struct pipe_buffer *buf;
 788
 789         if (cs->nr_segs == cs->pipe->buffers)
 790                 return -EIO;
 791
 792         unlock_request(cs->fc, cs->req);
 793         fuse_copy_finish(cs);
 794
 795         buf = cs->pipebufs;
 796         page_cache_get(page);
 797         buf->page = page;
 798         buf->offset = offset;
 799         buf->len = count;
 800
 801         cs->pipebufs++;
 802         cs->nr_segs++;
 803         cs->len = 0;
 804
 805         return 0;
 806 }
 807
 808 /*
 809  * Copy a page in the request to/from the userspace buffer.  Must be
 810  * done atomically
 811  */
 812 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 813                           unsigned offset, unsigned count, int zeroing)
 814 {
 815         int err;
 816         struct page *page = *pagep;
 817
 818         if (page && zeroing && count < PAGE_SIZE)
 819                 clear_highpage(page);
 820
 821         while (count) {
 822                 if (cs->write && cs->pipebufs && page) {
 823                         return fuse_ref_page(cs, page, offset, count);
 824                 } else if (!cs->len) {
 825                         if (cs->move_pages && page &&
 826                             offset == 0 && count == PAGE_SIZE) {
 827                                 err = fuse_try_move_page(cs, pagep);
 828                                 if (err <= 0)
 829                                         return err;
 830                         } else {
 831                                 err = fuse_copy_fill(cs);
 832                                 if (err)
 833                                         return err;
 834                         }
 835                 }
 836                 if (page) {
 837                         void *mapaddr = kmap_atomic(page, KM_USER0);
 838                         void *buf = mapaddr + offset;
 839                         offset += fuse_copy_do(cs, &buf, &count);
 840                         kunmap_atomic(mapaddr, KM_USER0);
 841                 } else
 842                         offset += fuse_copy_do(cs, NULL, &count);
 843         }
 844         if (page && !cs->write)
 845                 flush_dcache_page(page);
 846         return 0;
 847 }
 848
 849 /* Copy pages in the request to/from userspace buffer */
 850 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 851                            int zeroing)
 852 {
 853         unsigned i;
 854         struct fuse_req *req = cs->req;
 855         unsigned offset = req->page_offset;
 856         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 857
 858         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 859                 int err;
 860
 861                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
 862                                      zeroing);
 863                 if (err)
 864                         return err;
 865
 866                 nbytes -= count;
 867                 count = min(nbytes, (unsigned) PAGE_SIZE);
 868                 offset = 0;
 869         }
 870         return 0;
 871 }
 872
 873 /* Copy a single argument in the request to/from userspace buffer */
 874 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 875 {
 876         while (size) {
 877                 if (!cs->len) {
 878                         int err = fuse_copy_fill(cs);
 879                         if (err)
 880                                 return err;
 881                 }
 882                 fuse_copy_do(cs, &val, &size);
 883         }
 884         return 0;
 885 }
 886
 887 /* Copy request arguments to/from userspace buffer */
 888 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 889                           unsigned argpages, struct fuse_arg *args,
 890                           int zeroing)
 891 {
 892         int err = 0;
 893         unsigned i;
 894
 895         for (i = 0; !err && i < numargs; i++)  {
 896                 struct fuse_arg *arg = &args[i];
 897                 if (i == numargs - 1 && argpages)
 898                         err = fuse_copy_pages(cs, arg->size, zeroing);
 899                 else
 900                         err = fuse_copy_one(cs, arg->value, arg->size);
 901         }
 902         return err;
 903 }
 904
 905 static int forget_pending(struct fuse_conn *fc)
 906 {
 907         return fc->forget_list_head.next != NULL;
 908 }
 909
 910 static int request_pending(struct fuse_conn *fc)
 911 {
 912         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts) ||
 913                 forget_pending(fc);
 914 }
 915
 916 /* Wait until a request is available on the pending list */
 917 static void request_wait(struct fuse_conn *fc)
 918 __releases(fc->lock)
 919 __acquires(fc->lock)
 920 {
 921         DECLARE_WAITQUEUE(wait, current);
 922
 923         add_wait_queue_exclusive(&fc->waitq, &wait);
 924         while (fc->connected && !request_pending(fc)) {
 925                 set_current_state(TASK_INTERRUPTIBLE);
 926                 if (signal_pending(current))
 927                         break;
 928
 929                 spin_unlock(&fc->lock);
 930                 schedule();
 931                 spin_lock(&fc->lock);
 932         }
 933         set_current_state(TASK_RUNNING);
 934         remove_wait_queue(&fc->waitq, &wait);
 935 }
 936
 937 /*
 938  * Transfer an interrupt request to userspace
 939  *
 940  * Unlike other requests this is assembled on demand, without a need
 941  * to allocate a separate fuse_req structure.
 942  *
 943  * Called with fc->lock held, releases it
 944  */
 945 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
 946                                size_t nbytes, struct fuse_req *req)
 947 __releases(fc->lock)
 948 {
 949         struct fuse_in_header ih;
 950         struct fuse_interrupt_in arg;
 951         unsigned reqsize = sizeof(ih) + sizeof(arg);
 952         int err;
 953
 954         list_del_init(&req->intr_entry);
 955         req->intr_unique = fuse_get_unique(fc);
 956         memset(&ih, 0, sizeof(ih));
 957         memset(&arg, 0, sizeof(arg));
 958         ih.len = reqsize;
 959         ih.opcode = FUSE_INTERRUPT;
 960         ih.unique = req->intr_unique;
 961         arg.unique = req->in.h.unique;
 962
 963         spin_unlock(&fc->lock);
 964         if (nbytes < reqsize)
 965                 return -EINVAL;
 966
 967         err = fuse_copy_one(cs, &ih, sizeof(ih));
 968         if (!err)
 969                 err = fuse_copy_one(cs, &arg, sizeof(arg));
 970         fuse_copy_finish(cs);
 971
 972         return err ? err : reqsize;
 973 }
 974
 975 static struct fuse_forget_link *dequeue_forget(struct fuse_conn *fc,
 976                                                unsigned max,
 977                                                unsigned *countp)
 978 {
 979         struct fuse_forget_link *head = fc->forget_list_head.next;
 980         struct fuse_forget_link **newhead = &head;
 981         unsigned count;
 982
 983         for (count = 0; *newhead != NULL && count < max; count++)
 984                 newhead = &(*newhead)->next;
 985
 986         fc->forget_list_head.next = *newhead;
 987         *newhead = NULL;
 988         if (fc->forget_list_head.next == NULL)
 989                 fc->forget_list_tail = &fc->forget_list_head;
 990
 991         if (countp != NULL)
 992                 *countp = count;
 993
 994         return head;
 995 }
 996
 997 static int fuse_read_single_forget(struct fuse_conn *fc,
 998                                    struct fuse_copy_state *cs,
 999                                    size_t nbytes)
1000 __releases(fc->lock)
1001 {
1002         int err;
1003         struct fuse_forget_link *forget = dequeue_forget(fc, 1, NULL);
1004         struct fuse_forget_in arg = {
1005                 .nlookup = forget->forget_one.nlookup,
1006         };
1007         struct fuse_in_header ih = {
1008                 .opcode = FUSE_FORGET,
1009                 .nodeid = forget->forget_one.nodeid,
1010                 .unique = fuse_get_unique(fc),
1011                 .len = sizeof(ih) + sizeof(arg),
1012         };
1013
1014         spin_unlock(&fc->lock);
1015         kfree(forget);
1016         if (nbytes < ih.len)
1017                 return -EINVAL;
1018
1019         err = fuse_copy_one(cs, &ih, sizeof(ih));
1020         if (!err)
1021                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1022         fuse_copy_finish(cs);
1023
1024         if (err)
1025                 return err;
1026
1027         return ih.len;
1028 }
1029
1030 static int fuse_read_batch_forget(struct fuse_conn *fc,
1031                                    struct fuse_copy_state *cs, size_t nbytes)
1032 __releases(fc->lock)
1033 {
1034         int err;
1035         unsigned max_forgets;
1036         unsigned count;
1037         struct fuse_forget_link *head;
1038         struct fuse_batch_forget_in arg = { .count = 0 };
1039         struct fuse_in_header ih = {
1040                 .opcode = FUSE_BATCH_FORGET,
1041                 .unique = fuse_get_unique(fc),
1042                 .len = sizeof(ih) + sizeof(arg),
1043         };
1044
1045         if (nbytes < ih.len) {
1046                 spin_unlock(&fc->lock);
1047                 return -EINVAL;
1048         }
1049
1050         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1051         head = dequeue_forget(fc, max_forgets, &count);
1052         spin_unlock(&fc->lock);
1053
1054         arg.count = count;
1055         ih.len += count * sizeof(struct fuse_forget_one);
1056         err = fuse_copy_one(cs, &ih, sizeof(ih));
1057         if (!err)
1058                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1059
1060         while (head) {
1061                 struct fuse_forget_link *forget = head;
1062
1063                 if (!err) {
1064                         err = fuse_copy_one(cs, &forget->forget_one,
1065                                             sizeof(forget->forget_one));
1066                 }
1067                 head = forget->next;
1068                 kfree(forget);
1069         }
1070
1071         fuse_copy_finish(cs);
1072
1073         if (err)
1074                 return err;
1075
1076         return ih.len;
1077 }
1078
1079 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_copy_state *cs,
1080                             size_t nbytes)
1081 __releases(fc->lock)
1082 {
1083         if (fc->minor < 16 || fc->forget_list_head.next->next == NULL)
1084                 return fuse_read_single_forget(fc, cs, nbytes);
1085         else
1086                 return fuse_read_batch_forget(fc, cs, nbytes);
1087 }
1088
1089 /*
1090  * Read a single request into the userspace filesystem's buffer.  This
1091  * function waits until a request is available, then removes it from
1092  * the pending list and copies request data to userspace buffer.  If
1093  * no reply is needed (FORGET) or request has been aborted or there
1094  * was an error during the copying then it's finished by calling
1095  * request_end().  Otherwise add it to the processing list, and set
1096  * the 'sent' flag.
1097  */
1098 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1099                                 struct fuse_copy_state *cs, size_t nbytes)
1100 {
1101         int err;
1102         struct fuse_req *req;
1103         struct fuse_in *in;
1104         unsigned reqsize;
1105
1106  restart:
1107         spin_lock(&fc->lock);
1108         err = -EAGAIN;
1109         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
1110             !request_pending(fc))
1111                 goto err_unlock;
1112
1113         request_wait(fc);
1114         err = -ENODEV;
1115         if (!fc->connected)
1116                 goto err_unlock;
1117         err = -ERESTARTSYS;
1118         if (!request_pending(fc))
1119                 goto err_unlock;
1120
1121         if (!list_empty(&fc->interrupts)) {
1122                 req = list_entry(fc->interrupts.next, struct fuse_req,
1123                                  intr_entry);
1124                 return fuse_read_interrupt(fc, cs, nbytes, req);
1125         }
1126
1127         if (forget_pending(fc)) {
1128                 if (list_empty(&fc->pending) || fc->forget_batch-- > 0)
1129                         return fuse_read_forget(fc, cs, nbytes);
1130
1131                 if (fc->forget_batch <= -8)
1132                         fc->forget_batch = 16;
1133         }
1134
1135         req = list_entry(fc->pending.next, struct fuse_req, list);
1136         req->state = FUSE_REQ_READING;
1137         list_move(&req->list, &fc->io);
1138
1139         in = &req->in;
1140         reqsize = in->h.len;
1141         /* If request is too large, reply with an error and restart the read */
1142         if (nbytes < reqsize) {
1143                 req->out.h.error = -EIO;
1144                 /* SETXATTR is special, since it may contain too large data */
1145                 if (in->h.opcode == FUSE_SETXATTR)
1146                         req->out.h.error = -E2BIG;
1147                 request_end(fc, req);
1148                 goto restart;
1149         }
1150         spin_unlock(&fc->lock);
1151         cs->req = req;
1152         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1153         if (!err)
1154                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1155                                      (struct fuse_arg *) in->args, 0);
1156         fuse_copy_finish(cs);
1157         spin_lock(&fc->lock);
1158         req->locked = 0;
1159         if (req->aborted) {
1160                 request_end(fc, req);
1161                 return -ENODEV;
1162         }
1163         if (err) {
1164                 req->out.h.error = -EIO;
1165                 request_end(fc, req);
1166                 return err;
1167         }
1168         if (!req->isreply)
1169                 request_end(fc, req);
1170         else {
1171                 req->state = FUSE_REQ_SENT;
1172                 list_move_tail(&req->list, &fc->processing);
1173                 if (req->interrupted)
1174                         queue_interrupt(fc, req);
1175                 spin_unlock(&fc->lock);
1176         }
1177         return reqsize;
1178
1179  err_unlock:
1180         spin_unlock(&fc->lock);
1181         return err;
1182 }
1183
1184 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
1185                               unsigned long nr_segs, loff_t pos)
1186 {
1187         struct fuse_copy_state cs;
1188         struct file *file = iocb->ki_filp;
1189         struct fuse_conn *fc = fuse_get_conn(file);
1190         if (!fc)
1191                 return -EPERM;
1192
1193         fuse_copy_init(&cs, fc, 1, iov, nr_segs);
1194
1195         return fuse_dev_do_read(fc, file, &cs, iov_length(iov, nr_segs));
1196 }
1197
1198 static int fuse_dev_pipe_buf_steal(struct pipe_inode_info *pipe,
1199                                    struct pipe_buffer *buf)
1200 {
1201         return 1;
1202 }
1203
1204 static const struct pipe_buf_operations fuse_dev_pipe_buf_ops = {
1205         .can_merge = 0,
1206         .map = generic_pipe_buf_map,
1207         .unmap = generic_pipe_buf_unmap,
1208         .confirm = generic_pipe_buf_confirm,
1209         .release = generic_pipe_buf_release,
1210         .steal = fuse_dev_pipe_buf_steal,
1211         .get = generic_pipe_buf_get,
1212 };
1213
1214 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1215                                     struct pipe_inode_info *pipe,
1216                                     size_t len, unsigned int flags)
1217 {
1218         int ret;
1219         int page_nr = 0;
1220         int do_wakeup = 0;
1221         struct pipe_buffer *bufs;
1222         struct fuse_copy_state cs;
1223         struct fuse_conn *fc = fuse_get_conn(in);
1224         if (!fc)
1225                 return -EPERM;
1226
1227         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1228         if (!bufs)
1229                 return -ENOMEM;
1230
1231         fuse_copy_init(&cs, fc, 1, NULL, 0);
1232         cs.pipebufs = bufs;
1233         cs.pipe = pipe;
1234         ret = fuse_dev_do_read(fc, in, &cs, len);
1235         if (ret < 0)
1236                 goto out;
1237
1238         ret = 0;
1239         pipe_lock(pipe);
1240
1241         if (!pipe->readers) {
1242                 send_sig(SIGPIPE, current, 0);
1243                 if (!ret)
1244                         ret = -EPIPE;
1245                 goto out_unlock;
1246         }
1247
1248         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1249                 ret = -EIO;
1250                 goto out_unlock;
1251         }
1252
1253         while (page_nr < cs.nr_segs) {
1254                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1255                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1256
1257                 buf->page = bufs[page_nr].page;
1258                 buf->offset = bufs[page_nr].offset;
1259                 buf->len = bufs[page_nr].len;
1260                 buf->ops = &fuse_dev_pipe_buf_ops;
1261
1262                 pipe->nrbufs++;
1263                 page_nr++;
1264                 ret += buf->len;
1265
1266                 if (pipe->inode)
1267                         do_wakeup = 1;
1268         }
1269
1270 out_unlock:
1271         pipe_unlock(pipe);
1272
1273         if (do_wakeup) {
1274                 smp_mb();
1275                 if (waitqueue_active(&pipe->wait))
1276                         wake_up_interruptible(&pipe->wait);
1277                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1278         }
1279
1280 out:
1281         for (; page_nr < cs.nr_segs; page_nr++)
1282                 page_cache_release(bufs[page_nr].page);
1283
1284         kfree(bufs);
1285         return ret;
1286 }
1287
1288 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1289                             struct fuse_copy_state *cs)
1290 {
1291         struct fuse_notify_poll_wakeup_out outarg;
1292         int err = -EINVAL;
1293
1294         if (size != sizeof(outarg))
1295                 goto err;
1296
1297         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1298         if (err)
1299                 goto err;
1300
1301         fuse_copy_finish(cs);
1302         return fuse_notify_poll_wakeup(fc, &outarg);
1303
1304 err:
1305         fuse_copy_finish(cs);
1306         return err;
1307 }
1308
1309 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1310                                    struct fuse_copy_state *cs)
1311 {
1312         struct fuse_notify_inval_inode_out outarg;
1313         int err = -EINVAL;
1314
1315         if (size != sizeof(outarg))
1316                 goto err;
1317
1318         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1319         if (err)
1320                 goto err;
1321         fuse_copy_finish(cs);
1322
1323         down_read(&fc->killsb);
1324         err = -ENOENT;
1325         if (fc->sb) {
1326                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1327                                                outarg.off, outarg.len);
1328         }
1329         up_read(&fc->killsb);
1330         return err;
1331
1332 err:
1333         fuse_copy_finish(cs);
1334         return err;
1335 }
1336
1337 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1338                                    struct fuse_copy_state *cs)
1339 {
1340         struct fuse_notify_inval_entry_out outarg;
1341         int err = -ENOMEM;
1342         char *buf;
1343         struct qstr name;
1344
1345         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1346         if (!buf)
1347                 goto err;
1348
1349         err = -EINVAL;
1350         if (size < sizeof(outarg))
1351                 goto err;
1352
1353         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1354         if (err)
1355                 goto err;
1356
1357         err = -ENAMETOOLONG;
1358         if (outarg.namelen > FUSE_NAME_MAX)
1359                 goto err;
1360
1361         name.name = buf;
1362         name.len = outarg.namelen;
1363         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1364         if (err)
1365                 goto err;
1366         fuse_copy_finish(cs);
1367         buf[outarg.namelen] = 0;
1368         name.hash = full_name_hash(name.name, name.len);
1369
1370         down_read(&fc->killsb);
1371         err = -ENOENT;
1372         if (fc->sb)
1373                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, &name);
1374         up_read(&fc->killsb);
1375         kfree(buf);
1376         return err;
1377
1378 err:
1379         kfree(buf);
1380         fuse_copy_finish(cs);
1381         return err;
1382 }
1383
1384 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1385                              struct fuse_copy_state *cs)
1386 {
1387         struct fuse_notify_store_out outarg;
1388         struct inode *inode;
1389         struct address_space *mapping;
1390         u64 nodeid;
1391         int err;
1392         pgoff_t index;
1393         unsigned int offset;
1394         unsigned int num;
1395         loff_t file_size;
1396         loff_t end;
1397
1398         err = -EINVAL;
1399         if (size < sizeof(outarg))
1400                 goto out_finish;
1401
1402         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1403         if (err)
1404                 goto out_finish;
1405
1406         err = -EINVAL;
1407         if (size - sizeof(outarg) != outarg.size)
1408                 goto out_finish;
1409
1410         nodeid = outarg.nodeid;
1411
1412         down_read(&fc->killsb);
1413
1414         err = -ENOENT;
1415         if (!fc->sb)
1416                 goto out_up_killsb;
1417
1418         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1419         if (!inode)
1420                 goto out_up_killsb;
1421
1422         mapping = inode->i_mapping;
1423         index = outarg.offset >> PAGE_CACHE_SHIFT;
1424         offset = outarg.offset & ~PAGE_CACHE_MASK;
1425         file_size = i_size_read(inode);
1426         end = outarg.offset + outarg.size;
1427         if (end > file_size) {
1428                 file_size = end;
1429                 fuse_write_update_size(inode, file_size);
1430         }
1431
1432         num = outarg.size;
1433         while (num) {
1434                 struct page *page;
1435                 unsigned int this_num;
1436
1437                 err = -ENOMEM;
1438                 page = find_or_create_page(mapping, index,
1439                                            mapping_gfp_mask(mapping));
1440                 if (!page)
1441                         goto out_iput;
1442
1443                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1444                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1445                 if (!err && offset == 0 && (num != 0 || file_size == end))
1446                         SetPageUptodate(page);
1447                 unlock_page(page);
1448                 page_cache_release(page);
1449
1450                 if (err)
1451                         goto out_iput;
1452
1453                 num -= this_num;
1454                 offset = 0;
1455                 index++;
1456         }
1457
1458         err = 0;
1459
1460 out_iput:
1461         iput(inode);
1462 out_up_killsb:
1463         up_read(&fc->killsb);
1464 out_finish:
1465         fuse_copy_finish(cs);
1466         return err;
1467 }
1468
1469 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1470 {
1471         release_pages(req->pages, req->num_pages, 0);
1472 }
1473
1474 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1475                          struct fuse_notify_retrieve_out *outarg)
1476 {
1477         int err;
1478         struct address_space *mapping = inode->i_mapping;
1479         struct fuse_req *req;
1480         pgoff_t index;
1481         loff_t file_size;
1482         unsigned int num;
1483         unsigned int offset;
1484         size_t total_len = 0;
1485
1486         req = fuse_get_req(fc);
1487         if (IS_ERR(req))
1488                 return PTR_ERR(req);
1489
1490         offset = outarg->offset & ~PAGE_CACHE_MASK;
1491
1492         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1493         req->in.h.nodeid = outarg->nodeid;
1494         req->in.numargs = 2;
1495         req->in.argpages = 1;
1496         req->page_offset = offset;
1497         req->end = fuse_retrieve_end;
1498
1499         index = outarg->offset >> PAGE_CACHE_SHIFT;
1500         file_size = i_size_read(inode);
1501         num = outarg->size;
1502         if (outarg->offset > file_size)
1503                 num = 0;
1504         else if (outarg->offset + num > file_size)
1505                 num = file_size - outarg->offset;
1506
1507         while (num) {
1508                 struct page *page;
1509                 unsigned int this_num;
1510
1511                 page = find_get_page(mapping, index);
1512                 if (!page)
1513                         break;
1514
1515                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1516                 req->pages[req->num_pages] = page;
1517                 req->num_pages++;
1518
1519                 num -= this_num;
1520                 total_len += this_num;
1521         }
1522         req->misc.retrieve_in.offset = outarg->offset;
1523         req->misc.retrieve_in.size = total_len;
1524         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1525         req->in.args[0].value = &req->misc.retrieve_in;
1526         req->in.args[1].size = total_len;
1527
1528         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1529         if (err)
1530                 fuse_retrieve_end(fc, req);
1531
1532         return err;
1533 }
1534
1535 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1536                                 struct fuse_copy_state *cs)
1537 {
1538         struct fuse_notify_retrieve_out outarg;
1539         struct inode *inode;
1540         int err;
1541
1542         err = -EINVAL;
1543         if (size != sizeof(outarg))
1544                 goto copy_finish;
1545
1546         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1547         if (err)
1548                 goto copy_finish;
1549
1550         fuse_copy_finish(cs);
1551
1552         down_read(&fc->killsb);
1553         err = -ENOENT;
1554         if (fc->sb) {
1555                 u64 nodeid = outarg.nodeid;
1556
1557                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1558                 if (inode) {
1559                         err = fuse_retrieve(fc, inode, &outarg);
1560                         iput(inode);
1561                 }
1562         }
1563         up_read(&fc->killsb);
1564
1565         return err;
1566
1567 copy_finish:
1568         fuse_copy_finish(cs);
1569         return err;
1570 }
1571
1572 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1573                        unsigned int size, struct fuse_copy_state *cs)
1574 {
1575         switch (code) {
1576         case FUSE_NOTIFY_POLL:
1577                 return fuse_notify_poll(fc, size, cs);
1578
1579         case FUSE_NOTIFY_INVAL_INODE:
1580                 return fuse_notify_inval_inode(fc, size, cs);
1581
1582         case FUSE_NOTIFY_INVAL_ENTRY:
1583                 return fuse_notify_inval_entry(fc, size, cs);
1584
1585         case FUSE_NOTIFY_STORE:
1586                 return fuse_notify_store(fc, size, cs);
1587
1588         case FUSE_NOTIFY_RETRIEVE:
1589                 return fuse_notify_retrieve(fc, size, cs);
1590
1591         default:
1592                 fuse_copy_finish(cs);
1593                 return -EINVAL;
1594         }
1595 }
1596
1597 /* Look up request on processing list by unique ID */
1598 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1599 {
1600         struct list_head *entry;
1601
1602         list_for_each(entry, &fc->processing) {
1603                 struct fuse_req *req;
1604                 req = list_entry(entry, struct fuse_req, list);
1605                 if (req->in.h.unique == unique || req->intr_unique == unique)
1606                         return req;
1607         }
1608         return NULL;
1609 }
1610
1611 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1612                          unsigned nbytes)
1613 {
1614         unsigned reqsize = sizeof(struct fuse_out_header);
1615
1616         if (out->h.error)
1617                 return nbytes != reqsize ? -EINVAL : 0;
1618
1619         reqsize += len_args(out->numargs, out->args);
1620
1621         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1622                 return -EINVAL;
1623         else if (reqsize > nbytes) {
1624                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1625                 unsigned diffsize = reqsize - nbytes;
1626                 if (diffsize > lastarg->size)
1627                         return -EINVAL;
1628                 lastarg->size -= diffsize;
1629         }
1630         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1631                               out->page_zeroing);
1632 }
1633
1634 /*
1635  * Write a single reply to a request.  First the header is copied from
1636  * the write buffer.  The request is then searched on the processing
1637  * list by the unique ID found in the header.  If found, then remove
1638  * it from the list and copy the rest of the buffer to the request.
1639  * The request is finished by calling request_end()
1640  */
1641 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1642                                  struct fuse_copy_state *cs, size_t nbytes)
1643 {
1644         int err;
1645         struct fuse_req *req;
1646         struct fuse_out_header oh;
1647
1648         if (nbytes < sizeof(struct fuse_out_header))
1649                 return -EINVAL;
1650
1651         err = fuse_copy_one(cs, &oh, sizeof(oh));
1652         if (err)
1653                 goto err_finish;
1654
1655         err = -EINVAL;
1656         if (oh.len != nbytes)
1657                 goto err_finish;
1658
1659         /*
1660          * Zero oh.unique indicates unsolicited notification message
1661          * and error contains notification code.
1662          */
1663         if (!oh.unique) {
1664                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1665                 return err ? err : nbytes;
1666         }
1667
1668         err = -EINVAL;
1669         if (oh.error <= -1000 || oh.error > 0)
1670                 goto err_finish;
1671
1672         spin_lock(&fc->lock);
1673         err = -ENOENT;
1674         if (!fc->connected)
1675                 goto err_unlock;
1676
1677         req = request_find(fc, oh.unique);
1678         if (!req)
1679                 goto err_unlock;
1680
1681         if (req->aborted) {
1682                 spin_unlock(&fc->lock);
1683                 fuse_copy_finish(cs);
1684                 spin_lock(&fc->lock);
1685                 request_end(fc, req);
1686                 return -ENOENT;
1687         }
1688         /* Is it an interrupt reply? */
1689         if (req->intr_unique == oh.unique) {
1690                 err = -EINVAL;
1691                 if (nbytes != sizeof(struct fuse_out_header))
1692                         goto err_unlock;
1693
1694                 if (oh.error == -ENOSYS)
1695                         fc->no_interrupt = 1;
1696                 else if (oh.error == -EAGAIN)
1697                         queue_interrupt(fc, req);
1698
1699                 spin_unlock(&fc->lock);
1700                 fuse_copy_finish(cs);
1701                 return nbytes;
1702         }
1703
1704         req->state = FUSE_REQ_WRITING;
1705         list_move(&req->list, &fc->io);
1706         req->out.h = oh;
1707         req->locked = 1;
1708         cs->req = req;
1709         if (!req->out.page_replace)
1710                 cs->move_pages = 0;
1711         spin_unlock(&fc->lock);
1712
1713         err = copy_out_args(cs, &req->out, nbytes);
1714         fuse_copy_finish(cs);
1715
1716         spin_lock(&fc->lock);
1717         req->locked = 0;
1718         if (!err) {
1719                 if (req->aborted)
1720                         err = -ENOENT;
1721         } else if (!req->aborted)
1722                 req->out.h.error = -EIO;
1723         request_end(fc, req);
1724
1725         return err ? err : nbytes;
1726
1727  err_unlock:
1728         spin_unlock(&fc->lock);
1729  err_finish:
1730         fuse_copy_finish(cs);
1731         return err;
1732 }
1733
1734 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
1735                               unsigned long nr_segs, loff_t pos)
1736 {
1737         struct fuse_copy_state cs;
1738         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1739         if (!fc)
1740                 return -EPERM;
1741
1742         fuse_copy_init(&cs, fc, 0, iov, nr_segs);
1743
1744         return fuse_dev_do_write(fc, &cs, iov_length(iov, nr_segs));
1745 }
1746
1747 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1748                                      struct file *out, loff_t *ppos,
1749                                      size_t len, unsigned int flags)
1750 {
1751         unsigned nbuf;
1752         unsigned idx;
1753         struct pipe_buffer *bufs;
1754         struct fuse_copy_state cs;
1755         struct fuse_conn *fc;
1756         size_t rem;
1757         ssize_t ret;
1758
1759         fc = fuse_get_conn(out);
1760         if (!fc)
1761                 return -EPERM;
1762
1763         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1764         if (!bufs)
1765                 return -ENOMEM;
1766
1767         pipe_lock(pipe);
1768         nbuf = 0;
1769         rem = 0;
1770         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1771                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1772
1773         ret = -EINVAL;
1774         if (rem < len) {
1775                 pipe_unlock(pipe);
1776                 goto out;
1777         }
1778
1779         rem = len;
1780         while (rem) {
1781                 struct pipe_buffer *ibuf;
1782                 struct pipe_buffer *obuf;
1783
1784                 BUG_ON(nbuf >= pipe->buffers);
1785                 BUG_ON(!pipe->nrbufs);
1786                 ibuf = &pipe->bufs[pipe->curbuf];
1787                 obuf = &bufs[nbuf];
1788
1789                 if (rem >= ibuf->len) {
1790                         *obuf = *ibuf;
1791                         ibuf->ops = NULL;
1792                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1793                         pipe->nrbufs--;
1794                 } else {
1795                         ibuf->ops->get(pipe, ibuf);
1796                         *obuf = *ibuf;
1797                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1798                         obuf->len = rem;
1799                         ibuf->offset += obuf->len;
1800                         ibuf->len -= obuf->len;
1801                 }
1802                 nbuf++;
1803                 rem -= obuf->len;
1804         }
1805         pipe_unlock(pipe);
1806
1807         fuse_copy_init(&cs, fc, 0, NULL, nbuf);
1808         cs.pipebufs = bufs;
1809         cs.pipe = pipe;
1810
1811         if (flags & SPLICE_F_MOVE)
1812                 cs.move_pages = 1;
1813
1814         ret = fuse_dev_do_write(fc, &cs, len);
1815
1816         for (idx = 0; idx < nbuf; idx++) {
1817                 struct pipe_buffer *buf = &bufs[idx];
1818                 buf->ops->release(pipe, buf);
1819         }
1820 out:
1821         kfree(bufs);
1822         return ret;
1823 }
1824
1825 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1826 {
1827         unsigned mask = POLLOUT | POLLWRNORM;
1828         struct fuse_conn *fc = fuse_get_conn(file);
1829         if (!fc)
1830                 return POLLERR;
1831
1832         poll_wait(file, &fc->waitq, wait);
1833
1834         spin_lock(&fc->lock);
1835         if (!fc->connected)
1836                 mask = POLLERR;
1837         else if (request_pending(fc))
1838                 mask |= POLLIN | POLLRDNORM;
1839         spin_unlock(&fc->lock);
1840
1841         return mask;
1842 }
1843
1844 /*
1845  * Abort all requests on the given list (pending or processing)
1846  *
1847  * This function releases and reacquires fc->lock
1848  */
1849 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1850 __releases(fc->lock)
1851 __acquires(fc->lock)
1852 {
1853         while (!list_empty(head)) {
1854                 struct fuse_req *req;
1855                 req = list_entry(head->next, struct fuse_req, list);
1856                 req->out.h.error = -ECONNABORTED;
1857                 request_end(fc, req);
1858                 spin_lock(&fc->lock);
1859         }
1860 }
1861
1862 /*
1863  * Abort requests under I/O
1864  *
1865  * The requests are set to aborted and finished, and the request
1866  * waiter is woken up.  This will make request_wait_answer() wait
1867  * until the request is unlocked and then return.
1868  *
1869  * If the request is asynchronous, then the end function needs to be
1870  * called after waiting for the request to be unlocked (if it was
1871  * locked).
1872  */
1873 static void end_io_requests(struct fuse_conn *fc)
1874 __releases(fc->lock)
1875 __acquires(fc->lock)
1876 {
1877         while (!list_empty(&fc->io)) {
1878                 struct fuse_req *req =
1879                         list_entry(fc->io.next, struct fuse_req, list);
1880                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1881
1882                 req->aborted = 1;
1883                 req->out.h.error = -ECONNABORTED;
1884                 req->state = FUSE_REQ_FINISHED;
1885                 list_del_init(&req->list);
1886                 wake_up(&req->waitq);
1887                 if (end) {
1888                         req->end = NULL;
1889                         __fuse_get_request(req);
1890                         spin_unlock(&fc->lock);
1891                         wait_event(req->waitq, !req->locked);
1892                         end(fc, req);
1893                         fuse_put_request(fc, req);
1894                         spin_lock(&fc->lock);
1895                 }
1896         }
1897 }
1898
1899 static void end_queued_requests(struct fuse_conn *fc)
1900 __releases(fc->lock)
1901 __acquires(fc->lock)
1902 {
1903         fc->max_background = UINT_MAX;
1904         flush_bg_queue(fc);
1905         end_requests(fc, &fc->pending);
1906         end_requests(fc, &fc->processing);
1907         while (forget_pending(fc))
1908                 kfree(dequeue_forget(fc, 1, NULL));
1909 }
1910
1911 static void end_polls(struct fuse_conn *fc)
1912 {
1913         struct rb_node *p;
1914
1915         p = rb_first(&fc->polled_files);
1916
1917         while (p) {
1918                 struct fuse_file *ff;
1919                 ff = rb_entry(p, struct fuse_file, polled_node);
1920                 wake_up_interruptible_all(&ff->poll_wait);
1921
1922                 p = rb_next(p);
1923         }
1924 }
1925
1926 /*
1927  * Abort all requests.
1928  *
1929  * Emergency exit in case of a malicious or accidental deadlock, or
1930  * just a hung filesystem.
1931  *
1932  * The same effect is usually achievable through killing the
1933  * filesystem daemon and all users of the filesystem.  The exception
1934  * is the combination of an asynchronous request and the tricky
1935  * deadlock (see Documentation/filesystems/fuse.txt).
1936  *
1937  * During the aborting, progression of requests from the pending and
1938  * processing lists onto the io list, and progression of new requests
1939  * onto the pending list is prevented by req->connected being false.
1940  *
1941  * Progression of requests under I/O to the processing list is
1942  * prevented by the req->aborted flag being true for these requests.
1943  * For this reason requests on the io list must be aborted first.
1944  */
1945 void fuse_abort_conn(struct fuse_conn *fc)
1946 {
1947         spin_lock(&fc->lock);
1948         if (fc->connected) {
1949                 fc->connected = 0;
1950                 fc->blocked = 0;
1951                 end_io_requests(fc);
1952                 end_queued_requests(fc);
1953                 end_polls(fc);
1954                 wake_up_all(&fc->waitq);
1955                 wake_up_all(&fc->blocked_waitq);
1956                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1957         }
1958         spin_unlock(&fc->lock);
1959 }
1960 EXPORT_SYMBOL_GPL(fuse_abort_conn);
1961
1962 int fuse_dev_release(struct inode *inode, struct file *file)
1963 {
1964         struct fuse_conn *fc = fuse_get_conn(file);
1965         if (fc) {
1966                 spin_lock(&fc->lock);
1967                 fc->connected = 0;
1968                 fc->blocked = 0;
1969                 end_queued_requests(fc);
1970                 end_polls(fc);
1971                 wake_up_all(&fc->blocked_waitq);
1972                 spin_unlock(&fc->lock);
1973                 fuse_conn_put(fc);
1974         }
1975
1976         return 0;
1977 }
1978 EXPORT_SYMBOL_GPL(fuse_dev_release);
1979
1980 static int fuse_dev_fasync(int fd, struct file *file, int on)
1981 {
1982         struct fuse_conn *fc = fuse_get_conn(file);
1983         if (!fc)
1984                 return -EPERM;
1985
1986         /* No locking - fasync_helper does its own locking */
1987         return fasync_helper(fd, file, on, &fc->fasync);
1988 }
1989
1990 const struct file_operations fuse_dev_operations = {
1991         .owner          = THIS_MODULE,
1992         .llseek         = no_llseek,
1993         .read           = do_sync_read,
1994         .aio_read       = fuse_dev_read,
1995         .splice_read    = fuse_dev_splice_read,
1996         .write          = do_sync_write,
1997         .aio_write      = fuse_dev_write,
1998         .splice_write   = fuse_dev_splice_write,
1999         .poll           = fuse_dev_poll,
2000         .release        = fuse_dev_release,
2001         .fasync         = fuse_dev_fasync,
2002 };
2003 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2004
2005 static struct miscdevice fuse_miscdevice = {
2006         .minor = FUSE_MINOR,
2007         .name  = "fuse",
2008         .fops = &fuse_dev_operations,
2009 };
2010
2011 int __init fuse_dev_init(void)
2012 {
2013         int err = -ENOMEM;
2014         fuse_req_cachep = kmem_cache_create("fuse_request",
2015                                             sizeof(struct fuse_req),
2016                                             0, 0, NULL);
2017         if (!fuse_req_cachep)
2018                 goto out;
2019
2020         err = misc_register(&fuse_miscdevice);
2021         if (err)
2022                 goto out_cache_clean;
2023
2024         return 0;
2025
2026  out_cache_clean:
2027         kmem_cache_destroy(fuse_req_cachep);
2028  out:
2029         return err;
2030 }
2031
2032 void fuse_dev_cleanup(void)
2033 {
2034         misc_deregister(&fuse_miscdevice);
2035         kmem_cache_destroy(fuse_req_cachep);
2036 }