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 static void flush_bg_queue(struct fuse_conn *fc)
 255 {
 256         while (fc->active_background < fc->max_background &&
 257                !list_empty(&fc->bg_queue)) {
 258                 struct fuse_req *req;
 259
 260                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 261                 list_del(&req->list);
 262                 fc->active_background++;
 263                 req->in.h.unique = fuse_get_unique(fc);
 264                 queue_request(fc, req);
 265         }
 266 }
 267
 268 /*
 269  * This function is called when a request is finished.  Either a reply
 270  * has arrived or it was aborted (and not yet sent) or some error
 271  * occurred during communication with userspace, or the device file
 272  * was closed.  The requester thread is woken up (if still waiting),
 273  * the 'end' callback is called if given, else the reference to the
 274  * request is released
 275  *
 276  * Called with fc->lock, unlocks it
 277  */
 278 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 279 __releases(fc->lock)
 280 {
 281         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 282         req->end = NULL;
 283         list_del(&req->list);
 284         list_del(&req->intr_entry);
 285         req->state = FUSE_REQ_FINISHED;
 286         if (req->background) {
 287                 if (fc->num_background == fc->max_background) {
 288                         fc->blocked = 0;
 289                         wake_up_all(&fc->blocked_waitq);
 290                 }
 291                 if (fc->num_background == fc->congestion_threshold &&
 292                     fc->connected && fc->bdi_initialized) {
 293                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 294                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 295                 }
 296                 fc->num_background--;
 297                 fc->active_background--;
 298                 flush_bg_queue(fc);
 299         }
 300         spin_unlock(&fc->lock);
 301         wake_up(&req->waitq);
 302         if (end)
 303                 end(fc, req);
 304         fuse_put_request(fc, req);
 305 }
 306
 307 static void wait_answer_interruptible(struct fuse_conn *fc,
 308                                       struct fuse_req *req)
 309 __releases(fc->lock)
 310 __acquires(fc->lock)
 311 {
 312         if (signal_pending(current))
 313                 return;
 314
 315         spin_unlock(&fc->lock);
 316         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 317         spin_lock(&fc->lock);
 318 }
 319
 320 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 321 {
 322         list_add_tail(&req->intr_entry, &fc->interrupts);
 323         wake_up(&fc->waitq);
 324         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 325 }
 326
 327 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 328 __releases(fc->lock)
 329 __acquires(fc->lock)
 330 {
 331         if (!fc->no_interrupt) {
 332                 /* Any signal may interrupt this */
 333                 wait_answer_interruptible(fc, req);
 334
 335                 if (req->aborted)
 336                         goto aborted;
 337                 if (req->state == FUSE_REQ_FINISHED)
 338                         return;
 339
 340                 req->interrupted = 1;
 341                 if (req->state == FUSE_REQ_SENT)
 342                         queue_interrupt(fc, req);
 343         }
 344
 345         if (!req->force) {
 346                 sigset_t oldset;
 347
 348                 /* Only fatal signals may interrupt this */
 349                 block_sigs(&oldset);
 350                 wait_answer_interruptible(fc, req);
 351                 restore_sigs(&oldset);
 352
 353                 if (req->aborted)
 354                         goto aborted;
 355                 if (req->state == FUSE_REQ_FINISHED)
 356                         return;
 357
 358                 /* Request is not yet in userspace, bail out */
 359                 if (req->state == FUSE_REQ_PENDING) {
 360                         list_del(&req->list);
 361                         __fuse_put_request(req);
 362                         req->out.h.error = -EINTR;
 363                         return;
 364                 }
 365         }
 366
 367         /*
 368          * Either request is already in userspace, or it was forced.
 369          * Wait it out.
 370          */
 371         spin_unlock(&fc->lock);
 372         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 373         spin_lock(&fc->lock);
 374
 375         if (!req->aborted)
 376                 return;
 377
 378  aborted:
 379         BUG_ON(req->state != FUSE_REQ_FINISHED);
 380         if (req->locked) {
 381                 /* This is uninterruptible sleep, because data is
 382                    being copied to/from the buffers of req.  During
 383                    locked state, there mustn't be any filesystem
 384                    operation (e.g. page fault), since that could lead
 385                    to deadlock */
 386                 spin_unlock(&fc->lock);
 387                 wait_event(req->waitq, !req->locked);
 388                 spin_lock(&fc->lock);
 389         }
 390 }
 391
 392 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 393 {
 394         req->isreply = 1;
 395         spin_lock(&fc->lock);
 396         if (!fc->connected)
 397                 req->out.h.error = -ENOTCONN;
 398         else if (fc->conn_error)
 399                 req->out.h.error = -ECONNREFUSED;
 400         else {
 401                 req->in.h.unique = fuse_get_unique(fc);
 402                 queue_request(fc, req);
 403                 /* acquire extra reference, since request is still needed
 404                    after request_end() */
 405                 __fuse_get_request(req);
 406
 407                 request_wait_answer(fc, req);
 408         }
 409         spin_unlock(&fc->lock);
 410 }
 411 EXPORT_SYMBOL_GPL(fuse_request_send);
 412
 413 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 414                                             struct fuse_req *req)
 415 {
 416         req->background = 1;
 417         fc->num_background++;
 418         if (fc->num_background == fc->max_background)
 419                 fc->blocked = 1;
 420         if (fc->num_background == fc->congestion_threshold &&
 421             fc->bdi_initialized) {
 422                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 423                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 424         }
 425         list_add_tail(&req->list, &fc->bg_queue);
 426         flush_bg_queue(fc);
 427 }
 428
 429 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 430 {
 431         spin_lock(&fc->lock);
 432         if (fc->connected) {
 433                 fuse_request_send_nowait_locked(fc, req);
 434                 spin_unlock(&fc->lock);
 435         } else {
 436                 req->out.h.error = -ENOTCONN;
 437                 request_end(fc, req);
 438         }
 439 }
 440
 441 void fuse_request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 442 {
 443         req->isreply = 0;
 444         fuse_request_send_nowait(fc, req);
 445 }
 446
 447 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 448 {
 449         req->isreply = 1;
 450         fuse_request_send_nowait(fc, req);
 451 }
 452 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 453
 454 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 455                                           struct fuse_req *req, u64 unique)
 456 {
 457         int err = -ENODEV;
 458
 459         req->isreply = 0;
 460         req->in.h.unique = unique;
 461         spin_lock(&fc->lock);
 462         if (fc->connected) {
 463                 queue_request(fc, req);
 464                 err = 0;
 465         }
 466         spin_unlock(&fc->lock);
 467
 468         return err;
 469 }
 470
 471 /*
 472  * Called under fc->lock
 473  *
 474  * fc->connected must have been checked previously
 475  */
 476 void fuse_request_send_background_locked(struct fuse_conn *fc,
 477                                          struct fuse_req *req)
 478 {
 479         req->isreply = 1;
 480         fuse_request_send_nowait_locked(fc, req);
 481 }
 482
 483 /*
 484  * Lock the request.  Up to the next unlock_request() there mustn't be
 485  * anything that could cause a page-fault.  If the request was already
 486  * aborted bail out.
 487  */
 488 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 489 {
 490         int err = 0;
 491         if (req) {
 492                 spin_lock(&fc->lock);
 493                 if (req->aborted)
 494                         err = -ENOENT;
 495                 else
 496                         req->locked = 1;
 497                 spin_unlock(&fc->lock);
 498         }
 499         return err;
 500 }
 501
 502 /*
 503  * Unlock request.  If it was aborted during being locked, the
 504  * requester thread is currently waiting for it to be unlocked, so
 505  * wake it up.
 506  */
 507 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 508 {
 509         if (req) {
 510                 spin_lock(&fc->lock);
 511                 req->locked = 0;
 512                 if (req->aborted)
 513                         wake_up(&req->waitq);
 514                 spin_unlock(&fc->lock);
 515         }
 516 }
 517
 518 struct fuse_copy_state {
 519         struct fuse_conn *fc;
 520         int write;
 521         struct fuse_req *req;
 522         const struct iovec *iov;
 523         struct pipe_buffer *pipebufs;
 524         struct pipe_buffer *currbuf;
 525         struct pipe_inode_info *pipe;
 526         unsigned long nr_segs;
 527         unsigned long seglen;
 528         unsigned long addr;
 529         struct page *pg;
 530         void *mapaddr;
 531         void *buf;
 532         unsigned len;
 533         unsigned move_pages:1;
 534 };
 535
 536 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 537                            int write,
 538                            const struct iovec *iov, unsigned long nr_segs)
 539 {
 540         memset(cs, 0, sizeof(*cs));
 541         cs->fc = fc;
 542         cs->write = write;
 543         cs->iov = iov;
 544         cs->nr_segs = nr_segs;
 545 }
 546
 547 /* Unmap and put previous page of userspace buffer */
 548 static void fuse_copy_finish(struct fuse_copy_state *cs)
 549 {
 550         if (cs->currbuf) {
 551                 struct pipe_buffer *buf = cs->currbuf;
 552
 553                 if (!cs->write) {
 554                         buf->ops->unmap(cs->pipe, buf, cs->mapaddr);
 555                 } else {
 556                         kunmap(buf->page);
 557                         buf->len = PAGE_SIZE - cs->len;
 558                 }
 559                 cs->currbuf = NULL;
 560                 cs->mapaddr = NULL;
 561         } else if (cs->mapaddr) {
 562                 kunmap(cs->pg);
 563                 if (cs->write) {
 564                         flush_dcache_page(cs->pg);
 565                         set_page_dirty_lock(cs->pg);
 566                 }
 567                 put_page(cs->pg);
 568                 cs->mapaddr = NULL;
 569         }
 570 }
 571
 572 /*
 573  * Get another pagefull of userspace buffer, and map it to kernel
 574  * address space, and lock request
 575  */
 576 static int fuse_copy_fill(struct fuse_copy_state *cs)
 577 {
 578         unsigned long offset;
 579         int err;
 580
 581         unlock_request(cs->fc, cs->req);
 582         fuse_copy_finish(cs);
 583         if (cs->pipebufs) {
 584                 struct pipe_buffer *buf = cs->pipebufs;
 585
 586                 if (!cs->write) {
 587                         err = buf->ops->confirm(cs->pipe, buf);
 588                         if (err)
 589                                 return err;
 590
 591                         BUG_ON(!cs->nr_segs);
 592                         cs->currbuf = buf;
 593                         cs->mapaddr = buf->ops->map(cs->pipe, buf, 0);
 594                         cs->len = buf->len;
 595                         cs->buf = cs->mapaddr + buf->offset;
 596                         cs->pipebufs++;
 597                         cs->nr_segs--;
 598                 } else {
 599                         struct page *page;
 600
 601                         if (cs->nr_segs == cs->pipe->buffers)
 602                                 return -EIO;
 603
 604                         page = alloc_page(GFP_HIGHUSER);
 605                         if (!page)
 606                                 return -ENOMEM;
 607
 608                         buf->page = page;
 609                         buf->offset = 0;
 610                         buf->len = 0;
 611
 612                         cs->currbuf = buf;
 613                         cs->mapaddr = kmap(page);
 614                         cs->buf = cs->mapaddr;
 615                         cs->len = PAGE_SIZE;
 616                         cs->pipebufs++;
 617                         cs->nr_segs++;
 618                 }
 619         } else {
 620                 if (!cs->seglen) {
 621                         BUG_ON(!cs->nr_segs);
 622                         cs->seglen = cs->iov[0].iov_len;
 623                         cs->addr = (unsigned long) cs->iov[0].iov_base;
 624                         cs->iov++;
 625                         cs->nr_segs--;
 626                 }
 627                 err = get_user_pages_fast(cs->addr, 1, cs->write, &cs->pg);
 628                 if (err < 0)
 629                         return err;
 630                 BUG_ON(err != 1);
 631                 offset = cs->addr % PAGE_SIZE;
 632                 cs->mapaddr = kmap(cs->pg);
 633                 cs->buf = cs->mapaddr + offset;
 634                 cs->len = min(PAGE_SIZE - offset, cs->seglen);
 635                 cs->seglen -= cs->len;
 636                 cs->addr += cs->len;
 637         }
 638
 639         return lock_request(cs->fc, cs->req);
 640 }
 641
 642 /* Do as much copy to/from userspace buffer as we can */
 643 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 644 {
 645         unsigned ncpy = min(*size, cs->len);
 646         if (val) {
 647                 if (cs->write)
 648                         memcpy(cs->buf, *val, ncpy);
 649                 else
 650                         memcpy(*val, cs->buf, ncpy);
 651                 *val += ncpy;
 652         }
 653         *size -= ncpy;
 654         cs->len -= ncpy;
 655         cs->buf += ncpy;
 656         return ncpy;
 657 }
 658
 659 static int fuse_check_page(struct page *page)
 660 {
 661         if (page_mapcount(page) ||
 662             page->mapping != NULL ||
 663             page_count(page) != 1 ||
 664             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 665              ~(1 << PG_locked |
 666                1 << PG_referenced |
 667                1 << PG_uptodate |
 668                1 << PG_lru |
 669                1 << PG_active |
 670                1 << PG_reclaim))) {
 671                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 672                 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);
 673                 return 1;
 674         }
 675         return 0;
 676 }
 677
 678 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 679 {
 680         int err;
 681         struct page *oldpage = *pagep;
 682         struct page *newpage;
 683         struct pipe_buffer *buf = cs->pipebufs;
 684         struct address_space *mapping;
 685         pgoff_t index;
 686
 687         unlock_request(cs->fc, cs->req);
 688         fuse_copy_finish(cs);
 689
 690         err = buf->ops->confirm(cs->pipe, buf);
 691         if (err)
 692                 return err;
 693
 694         BUG_ON(!cs->nr_segs);
 695         cs->currbuf = buf;
 696         cs->len = buf->len;
 697         cs->pipebufs++;
 698         cs->nr_segs--;
 699
 700         if (cs->len != PAGE_SIZE)
 701                 goto out_fallback;
 702
 703         if (buf->ops->steal(cs->pipe, buf) != 0)
 704                 goto out_fallback;
 705
 706         newpage = buf->page;
 707
 708         if (WARN_ON(!PageUptodate(newpage)))
 709                 return -EIO;
 710
 711         ClearPageMappedToDisk(newpage);
 712
 713         if (fuse_check_page(newpage) != 0)
 714                 goto out_fallback_unlock;
 715
 716         mapping = oldpage->mapping;
 717         index = oldpage->index;
 718
 719         /*
 720          * This is a new and locked page, it shouldn't be mapped or
 721          * have any special flags on it
 722          */
 723         if (WARN_ON(page_mapped(oldpage)))
 724                 goto out_fallback_unlock;
 725         if (WARN_ON(page_has_private(oldpage)))
 726                 goto out_fallback_unlock;
 727         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 728                 goto out_fallback_unlock;
 729         if (WARN_ON(PageMlocked(oldpage)))
 730                 goto out_fallback_unlock;
 731
 732         remove_from_page_cache(oldpage);
 733         page_cache_release(oldpage);
 734
 735         err = add_to_page_cache_locked(newpage, mapping, index, GFP_KERNEL);
 736         if (err) {
 737                 printk(KERN_WARNING "fuse_try_move_page: failed to add page");
 738                 goto out_fallback_unlock;
 739         }
 740         page_cache_get(newpage);
 741
 742         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 743                 lru_cache_add_file(newpage);
 744
 745         err = 0;
 746         spin_lock(&cs->fc->lock);
 747         if (cs->req->aborted)
 748                 err = -ENOENT;
 749         else
 750                 *pagep = newpage;
 751         spin_unlock(&cs->fc->lock);
 752
 753         if (err) {
 754                 unlock_page(newpage);
 755                 page_cache_release(newpage);
 756                 return err;
 757         }
 758
 759         unlock_page(oldpage);
 760         page_cache_release(oldpage);
 761         cs->len = 0;
 762
 763         return 0;
 764
 765 out_fallback_unlock:
 766         unlock_page(newpage);
 767 out_fallback:
 768         cs->mapaddr = buf->ops->map(cs->pipe, buf, 1);
 769         cs->buf = cs->mapaddr + buf->offset;
 770
 771         err = lock_request(cs->fc, cs->req);
 772         if (err)
 773                 return err;
 774
 775         return 1;
 776 }
 777
 778 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 779                          unsigned offset, unsigned count)
 780 {
 781         struct pipe_buffer *buf;
 782
 783         if (cs->nr_segs == cs->pipe->buffers)
 784                 return -EIO;
 785
 786         unlock_request(cs->fc, cs->req);
 787         fuse_copy_finish(cs);
 788
 789         buf = cs->pipebufs;
 790         page_cache_get(page);
 791         buf->page = page;
 792         buf->offset = offset;
 793         buf->len = count;
 794
 795         cs->pipebufs++;
 796         cs->nr_segs++;
 797         cs->len = 0;
 798
 799         return 0;
 800 }
 801
 802 /*
 803  * Copy a page in the request to/from the userspace buffer.  Must be
 804  * done atomically
 805  */
 806 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 807                           unsigned offset, unsigned count, int zeroing)
 808 {
 809         int err;
 810         struct page *page = *pagep;
 811
 812         if (page && zeroing && count < PAGE_SIZE) {
 813                 void *mapaddr = kmap_atomic(page, KM_USER1);
 814                 memset(mapaddr, 0, PAGE_SIZE);
 815                 kunmap_atomic(mapaddr, KM_USER1);
 816         }
 817         while (count) {
 818                 if (cs->write && cs->pipebufs && page) {
 819                         return fuse_ref_page(cs, page, offset, count);
 820                 } else if (!cs->len) {
 821                         if (cs->move_pages && page &&
 822                             offset == 0 && count == PAGE_SIZE) {
 823                                 err = fuse_try_move_page(cs, pagep);
 824                                 if (err <= 0)
 825                                         return err;
 826                         } else {
 827                                 err = fuse_copy_fill(cs);
 828                                 if (err)
 829                                         return err;
 830                         }
 831                 }
 832                 if (page) {
 833                         void *mapaddr = kmap_atomic(page, KM_USER1);
 834                         void *buf = mapaddr + offset;
 835                         offset += fuse_copy_do(cs, &buf, &count);
 836                         kunmap_atomic(mapaddr, KM_USER1);
 837                 } else
 838                         offset += fuse_copy_do(cs, NULL, &count);
 839         }
 840         if (page && !cs->write)
 841                 flush_dcache_page(page);
 842         return 0;
 843 }
 844
 845 /* Copy pages in the request to/from userspace buffer */
 846 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 847                            int zeroing)
 848 {
 849         unsigned i;
 850         struct fuse_req *req = cs->req;
 851         unsigned offset = req->page_offset;
 852         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 853
 854         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 855                 int err;
 856
 857                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
 858                                      zeroing);
 859                 if (err)
 860                         return err;
 861
 862                 nbytes -= count;
 863                 count = min(nbytes, (unsigned) PAGE_SIZE);
 864                 offset = 0;
 865         }
 866         return 0;
 867 }
 868
 869 /* Copy a single argument in the request to/from userspace buffer */
 870 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 871 {
 872         while (size) {
 873                 if (!cs->len) {
 874                         int err = fuse_copy_fill(cs);
 875                         if (err)
 876                                 return err;
 877                 }
 878                 fuse_copy_do(cs, &val, &size);
 879         }
 880         return 0;
 881 }
 882
 883 /* Copy request arguments to/from userspace buffer */
 884 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 885                           unsigned argpages, struct fuse_arg *args,
 886                           int zeroing)
 887 {
 888         int err = 0;
 889         unsigned i;
 890
 891         for (i = 0; !err && i < numargs; i++)  {
 892                 struct fuse_arg *arg = &args[i];
 893                 if (i == numargs - 1 && argpages)
 894                         err = fuse_copy_pages(cs, arg->size, zeroing);
 895                 else
 896                         err = fuse_copy_one(cs, arg->value, arg->size);
 897         }
 898         return err;
 899 }
 900
 901 static int request_pending(struct fuse_conn *fc)
 902 {
 903         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 904 }
 905
 906 /* Wait until a request is available on the pending list */
 907 static void request_wait(struct fuse_conn *fc)
 908 __releases(fc->lock)
 909 __acquires(fc->lock)
 910 {
 911         DECLARE_WAITQUEUE(wait, current);
 912
 913         add_wait_queue_exclusive(&fc->waitq, &wait);
 914         while (fc->connected && !request_pending(fc)) {
 915                 set_current_state(TASK_INTERRUPTIBLE);
 916                 if (signal_pending(current))
 917                         break;
 918
 919                 spin_unlock(&fc->lock);
 920                 schedule();
 921                 spin_lock(&fc->lock);
 922         }
 923         set_current_state(TASK_RUNNING);
 924         remove_wait_queue(&fc->waitq, &wait);
 925 }
 926
 927 /*
 928  * Transfer an interrupt request to userspace
 929  *
 930  * Unlike other requests this is assembled on demand, without a need
 931  * to allocate a separate fuse_req structure.
 932  *
 933  * Called with fc->lock held, releases it
 934  */
 935 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
 936                                size_t nbytes, struct fuse_req *req)
 937 __releases(fc->lock)
 938 {
 939         struct fuse_in_header ih;
 940         struct fuse_interrupt_in arg;
 941         unsigned reqsize = sizeof(ih) + sizeof(arg);
 942         int err;
 943
 944         list_del_init(&req->intr_entry);
 945         req->intr_unique = fuse_get_unique(fc);
 946         memset(&ih, 0, sizeof(ih));
 947         memset(&arg, 0, sizeof(arg));
 948         ih.len = reqsize;
 949         ih.opcode = FUSE_INTERRUPT;
 950         ih.unique = req->intr_unique;
 951         arg.unique = req->in.h.unique;
 952
 953         spin_unlock(&fc->lock);
 954         if (nbytes < reqsize)
 955                 return -EINVAL;
 956
 957         err = fuse_copy_one(cs, &ih, sizeof(ih));
 958         if (!err)
 959                 err = fuse_copy_one(cs, &arg, sizeof(arg));
 960         fuse_copy_finish(cs);
 961
 962         return err ? err : reqsize;
 963 }
 964
 965 /*
 966  * Read a single request into the userspace filesystem's buffer.  This
 967  * function waits until a request is available, then removes it from
 968  * the pending list and copies request data to userspace buffer.  If
 969  * no reply is needed (FORGET) or request has been aborted or there
 970  * was an error during the copying then it's finished by calling
 971  * request_end().  Otherwise add it to the processing list, and set
 972  * the 'sent' flag.
 973  */
 974 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
 975                                 struct fuse_copy_state *cs, size_t nbytes)
 976 {
 977         int err;
 978         struct fuse_req *req;
 979         struct fuse_in *in;
 980         unsigned reqsize;
 981
 982  restart:
 983         spin_lock(&fc->lock);
 984         err = -EAGAIN;
 985         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 986             !request_pending(fc))
 987                 goto err_unlock;
 988
 989         request_wait(fc);
 990         err = -ENODEV;
 991         if (!fc->connected)
 992                 goto err_unlock;
 993         err = -ERESTARTSYS;
 994         if (!request_pending(fc))
 995                 goto err_unlock;
 996
 997         if (!list_empty(&fc->interrupts)) {
 998                 req = list_entry(fc->interrupts.next, struct fuse_req,
 999                                  intr_entry);
1000                 return fuse_read_interrupt(fc, cs, nbytes, req);
1001         }
1002
1003         req = list_entry(fc->pending.next, struct fuse_req, list);
1004         req->state = FUSE_REQ_READING;
1005         list_move(&req->list, &fc->io);
1006
1007         in = &req->in;
1008         reqsize = in->h.len;
1009         /* If request is too large, reply with an error and restart the read */
1010         if (nbytes < reqsize) {
1011                 req->out.h.error = -EIO;
1012                 /* SETXATTR is special, since it may contain too large data */
1013                 if (in->h.opcode == FUSE_SETXATTR)
1014                         req->out.h.error = -E2BIG;
1015                 request_end(fc, req);
1016                 goto restart;
1017         }
1018         spin_unlock(&fc->lock);
1019         cs->req = req;
1020         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1021         if (!err)
1022                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1023                                      (struct fuse_arg *) in->args, 0);
1024         fuse_copy_finish(cs);
1025         spin_lock(&fc->lock);
1026         req->locked = 0;
1027         if (req->aborted) {
1028                 request_end(fc, req);
1029                 return -ENODEV;
1030         }
1031         if (err) {
1032                 req->out.h.error = -EIO;
1033                 request_end(fc, req);
1034                 return err;
1035         }
1036         if (!req->isreply)
1037                 request_end(fc, req);
1038         else {
1039                 req->state = FUSE_REQ_SENT;
1040                 list_move_tail(&req->list, &fc->processing);
1041                 if (req->interrupted)
1042                         queue_interrupt(fc, req);
1043                 spin_unlock(&fc->lock);
1044         }
1045         return reqsize;
1046
1047  err_unlock:
1048         spin_unlock(&fc->lock);
1049         return err;
1050 }
1051
1052 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
1053                               unsigned long nr_segs, loff_t pos)
1054 {
1055         struct fuse_copy_state cs;
1056         struct file *file = iocb->ki_filp;
1057         struct fuse_conn *fc = fuse_get_conn(file);
1058         if (!fc)
1059                 return -EPERM;
1060
1061         fuse_copy_init(&cs, fc, 1, iov, nr_segs);
1062
1063         return fuse_dev_do_read(fc, file, &cs, iov_length(iov, nr_segs));
1064 }
1065
1066 static int fuse_dev_pipe_buf_steal(struct pipe_inode_info *pipe,
1067                                    struct pipe_buffer *buf)
1068 {
1069         return 1;
1070 }
1071
1072 static const struct pipe_buf_operations fuse_dev_pipe_buf_ops = {
1073         .can_merge = 0,
1074         .map = generic_pipe_buf_map,
1075         .unmap = generic_pipe_buf_unmap,
1076         .confirm = generic_pipe_buf_confirm,
1077         .release = generic_pipe_buf_release,
1078         .steal = fuse_dev_pipe_buf_steal,
1079         .get = generic_pipe_buf_get,
1080 };
1081
1082 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1083                                     struct pipe_inode_info *pipe,
1084                                     size_t len, unsigned int flags)
1085 {
1086         int ret;
1087         int page_nr = 0;
1088         int do_wakeup = 0;
1089         struct pipe_buffer *bufs;
1090         struct fuse_copy_state cs;
1091         struct fuse_conn *fc = fuse_get_conn(in);
1092         if (!fc)
1093                 return -EPERM;
1094
1095         bufs = kmalloc(pipe->buffers * sizeof (struct pipe_buffer), GFP_KERNEL);
1096         if (!bufs)
1097                 return -ENOMEM;
1098
1099         fuse_copy_init(&cs, fc, 1, NULL, 0);
1100         cs.pipebufs = bufs;
1101         cs.pipe = pipe;
1102         ret = fuse_dev_do_read(fc, in, &cs, len);
1103         if (ret < 0)
1104                 goto out;
1105
1106         ret = 0;
1107         pipe_lock(pipe);
1108
1109         if (!pipe->readers) {
1110                 send_sig(SIGPIPE, current, 0);
1111                 if (!ret)
1112                         ret = -EPIPE;
1113                 goto out_unlock;
1114         }
1115
1116         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1117                 ret = -EIO;
1118                 goto out_unlock;
1119         }
1120
1121         while (page_nr < cs.nr_segs) {
1122                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1123                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1124
1125                 buf->page = bufs[page_nr].page;
1126                 buf->offset = bufs[page_nr].offset;
1127                 buf->len = bufs[page_nr].len;
1128                 buf->ops = &fuse_dev_pipe_buf_ops;
1129
1130                 pipe->nrbufs++;
1131                 page_nr++;
1132                 ret += buf->len;
1133
1134                 if (pipe->inode)
1135                         do_wakeup = 1;
1136         }
1137
1138 out_unlock:
1139         pipe_unlock(pipe);
1140
1141         if (do_wakeup) {
1142                 smp_mb();
1143                 if (waitqueue_active(&pipe->wait))
1144                         wake_up_interruptible(&pipe->wait);
1145                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1146         }
1147
1148 out:
1149         for (; page_nr < cs.nr_segs; page_nr++)
1150                 page_cache_release(bufs[page_nr].page);
1151
1152         kfree(bufs);
1153         return ret;
1154 }
1155
1156 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1157                             struct fuse_copy_state *cs)
1158 {
1159         struct fuse_notify_poll_wakeup_out outarg;
1160         int err = -EINVAL;
1161
1162         if (size != sizeof(outarg))
1163                 goto err;
1164
1165         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1166         if (err)
1167                 goto err;
1168
1169         fuse_copy_finish(cs);
1170         return fuse_notify_poll_wakeup(fc, &outarg);
1171
1172 err:
1173         fuse_copy_finish(cs);
1174         return err;
1175 }
1176
1177 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1178                                    struct fuse_copy_state *cs)
1179 {
1180         struct fuse_notify_inval_inode_out outarg;
1181         int err = -EINVAL;
1182
1183         if (size != sizeof(outarg))
1184                 goto err;
1185
1186         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1187         if (err)
1188                 goto err;
1189         fuse_copy_finish(cs);
1190
1191         down_read(&fc->killsb);
1192         err = -ENOENT;
1193         if (fc->sb) {
1194                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1195                                                outarg.off, outarg.len);
1196         }
1197         up_read(&fc->killsb);
1198         return err;
1199
1200 err:
1201         fuse_copy_finish(cs);
1202         return err;
1203 }
1204
1205 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1206                                    struct fuse_copy_state *cs)
1207 {
1208         struct fuse_notify_inval_entry_out outarg;
1209         int err = -ENOMEM;
1210         char *buf;
1211         struct qstr name;
1212
1213         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1214         if (!buf)
1215                 goto err;
1216
1217         err = -EINVAL;
1218         if (size < sizeof(outarg))
1219                 goto err;
1220
1221         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1222         if (err)
1223                 goto err;
1224
1225         err = -ENAMETOOLONG;
1226         if (outarg.namelen > FUSE_NAME_MAX)
1227                 goto err;
1228
1229         name.name = buf;
1230         name.len = outarg.namelen;
1231         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1232         if (err)
1233                 goto err;
1234         fuse_copy_finish(cs);
1235         buf[outarg.namelen] = 0;
1236         name.hash = full_name_hash(name.name, name.len);
1237
1238         down_read(&fc->killsb);
1239         err = -ENOENT;
1240         if (fc->sb)
1241                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, &name);
1242         up_read(&fc->killsb);
1243         kfree(buf);
1244         return err;
1245
1246 err:
1247         kfree(buf);
1248         fuse_copy_finish(cs);
1249         return err;
1250 }
1251
1252 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1253                              struct fuse_copy_state *cs)
1254 {
1255         struct fuse_notify_store_out outarg;
1256         struct inode *inode;
1257         struct address_space *mapping;
1258         u64 nodeid;
1259         int err;
1260         pgoff_t index;
1261         unsigned int offset;
1262         unsigned int num;
1263         loff_t file_size;
1264         loff_t end;
1265
1266         err = -EINVAL;
1267         if (size < sizeof(outarg))
1268                 goto out_finish;
1269
1270         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1271         if (err)
1272                 goto out_finish;
1273
1274         err = -EINVAL;
1275         if (size - sizeof(outarg) != outarg.size)
1276                 goto out_finish;
1277
1278         nodeid = outarg.nodeid;
1279
1280         down_read(&fc->killsb);
1281
1282         err = -ENOENT;
1283         if (!fc->sb)
1284                 goto out_up_killsb;
1285
1286         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1287         if (!inode)
1288                 goto out_up_killsb;
1289
1290         mapping = inode->i_mapping;
1291         index = outarg.offset >> PAGE_CACHE_SHIFT;
1292         offset = outarg.offset & ~PAGE_CACHE_MASK;
1293         file_size = i_size_read(inode);
1294         end = outarg.offset + outarg.size;
1295         if (end > file_size) {
1296                 file_size = end;
1297                 fuse_write_update_size(inode, file_size);
1298         }
1299
1300         num = outarg.size;
1301         while (num) {
1302                 struct page *page;
1303                 unsigned int this_num;
1304
1305                 err = -ENOMEM;
1306                 page = find_or_create_page(mapping, index,
1307                                            mapping_gfp_mask(mapping));
1308                 if (!page)
1309                         goto out_iput;
1310
1311                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1312                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1313                 if (!err && offset == 0 && (num != 0 || file_size == end))
1314                         SetPageUptodate(page);
1315                 unlock_page(page);
1316                 page_cache_release(page);
1317
1318                 if (err)
1319                         goto out_iput;
1320
1321                 num -= this_num;
1322                 offset = 0;
1323                 index++;
1324         }
1325
1326         err = 0;
1327
1328 out_iput:
1329         iput(inode);
1330 out_up_killsb:
1331         up_read(&fc->killsb);
1332 out_finish:
1333         fuse_copy_finish(cs);
1334         return err;
1335 }
1336
1337 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1338 {
1339         int i;
1340
1341         for (i = 0; i < req->num_pages; i++) {
1342                 struct page *page = req->pages[i];
1343                 page_cache_release(page);
1344         }
1345 }
1346
1347 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1348                          struct fuse_notify_retrieve_out *outarg)
1349 {
1350         int err;
1351         struct address_space *mapping = inode->i_mapping;
1352         struct fuse_req *req;
1353         pgoff_t index;
1354         loff_t file_size;
1355         unsigned int num;
1356         unsigned int offset;
1357         size_t total_len;
1358
1359         req = fuse_get_req(fc);
1360         if (IS_ERR(req))
1361                 return PTR_ERR(req);
1362
1363         offset = outarg->offset & ~PAGE_CACHE_MASK;
1364
1365         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1366         req->in.h.nodeid = outarg->nodeid;
1367         req->in.numargs = 2;
1368         req->in.argpages = 1;
1369         req->page_offset = offset;
1370         req->end = fuse_retrieve_end;
1371
1372         index = outarg->offset >> PAGE_CACHE_SHIFT;
1373         file_size = i_size_read(inode);
1374         num = outarg->size;
1375         if (outarg->offset > file_size)
1376                 num = 0;
1377         else if (outarg->offset + num > file_size)
1378                 num = file_size - outarg->offset;
1379
1380         while (num) {
1381                 struct page *page;
1382                 unsigned int this_num;
1383
1384                 page = find_get_page(mapping, index);
1385                 if (!page)
1386                         break;
1387
1388                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1389                 req->pages[req->num_pages] = page;
1390                 req->num_pages++;
1391
1392                 num -= this_num;
1393                 total_len += this_num;
1394         }
1395         req->misc.retrieve_in.offset = outarg->offset;
1396         req->misc.retrieve_in.size = total_len;
1397         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1398         req->in.args[0].value = &req->misc.retrieve_in;
1399         req->in.args[1].size = total_len;
1400
1401         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1402         if (err)
1403                 fuse_retrieve_end(fc, req);
1404
1405         return err;
1406 }
1407
1408 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1409                                 struct fuse_copy_state *cs)
1410 {
1411         struct fuse_notify_retrieve_out outarg;
1412         struct inode *inode;
1413         int err;
1414
1415         err = -EINVAL;
1416         if (size != sizeof(outarg))
1417                 goto copy_finish;
1418
1419         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1420         if (err)
1421                 goto copy_finish;
1422
1423         fuse_copy_finish(cs);
1424
1425         down_read(&fc->killsb);
1426         err = -ENOENT;
1427         if (fc->sb) {
1428                 u64 nodeid = outarg.nodeid;
1429
1430                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1431                 if (inode) {
1432                         err = fuse_retrieve(fc, inode, &outarg);
1433                         iput(inode);
1434                 }
1435         }
1436         up_read(&fc->killsb);
1437
1438         return err;
1439
1440 copy_finish:
1441         fuse_copy_finish(cs);
1442         return err;
1443 }
1444
1445 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1446                        unsigned int size, struct fuse_copy_state *cs)
1447 {
1448         switch (code) {
1449         case FUSE_NOTIFY_POLL:
1450                 return fuse_notify_poll(fc, size, cs);
1451
1452         case FUSE_NOTIFY_INVAL_INODE:
1453                 return fuse_notify_inval_inode(fc, size, cs);
1454
1455         case FUSE_NOTIFY_INVAL_ENTRY:
1456                 return fuse_notify_inval_entry(fc, size, cs);
1457
1458         case FUSE_NOTIFY_STORE:
1459                 return fuse_notify_store(fc, size, cs);
1460
1461         case FUSE_NOTIFY_RETRIEVE:
1462                 return fuse_notify_retrieve(fc, size, cs);
1463
1464         default:
1465                 fuse_copy_finish(cs);
1466                 return -EINVAL;
1467         }
1468 }
1469
1470 /* Look up request on processing list by unique ID */
1471 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1472 {
1473         struct list_head *entry;
1474
1475         list_for_each(entry, &fc->processing) {
1476                 struct fuse_req *req;
1477                 req = list_entry(entry, struct fuse_req, list);
1478                 if (req->in.h.unique == unique || req->intr_unique == unique)
1479                         return req;
1480         }
1481         return NULL;
1482 }
1483
1484 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1485                          unsigned nbytes)
1486 {
1487         unsigned reqsize = sizeof(struct fuse_out_header);
1488
1489         if (out->h.error)
1490                 return nbytes != reqsize ? -EINVAL : 0;
1491
1492         reqsize += len_args(out->numargs, out->args);
1493
1494         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1495                 return -EINVAL;
1496         else if (reqsize > nbytes) {
1497                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1498                 unsigned diffsize = reqsize - nbytes;
1499                 if (diffsize > lastarg->size)
1500                         return -EINVAL;
1501                 lastarg->size -= diffsize;
1502         }
1503         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1504                               out->page_zeroing);
1505 }
1506
1507 /*
1508  * Write a single reply to a request.  First the header is copied from
1509  * the write buffer.  The request is then searched on the processing
1510  * list by the unique ID found in the header.  If found, then remove
1511  * it from the list and copy the rest of the buffer to the request.
1512  * The request is finished by calling request_end()
1513  */
1514 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1515                                  struct fuse_copy_state *cs, size_t nbytes)
1516 {
1517         int err;
1518         struct fuse_req *req;
1519         struct fuse_out_header oh;
1520
1521         if (nbytes < sizeof(struct fuse_out_header))
1522                 return -EINVAL;
1523
1524         err = fuse_copy_one(cs, &oh, sizeof(oh));
1525         if (err)
1526                 goto err_finish;
1527
1528         err = -EINVAL;
1529         if (oh.len != nbytes)
1530                 goto err_finish;
1531
1532         /*
1533          * Zero oh.unique indicates unsolicited notification message
1534          * and error contains notification code.
1535          */
1536         if (!oh.unique) {
1537                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1538                 return err ? err : nbytes;
1539         }
1540
1541         err = -EINVAL;
1542         if (oh.error <= -1000 || oh.error > 0)
1543                 goto err_finish;
1544
1545         spin_lock(&fc->lock);
1546         err = -ENOENT;
1547         if (!fc->connected)
1548                 goto err_unlock;
1549
1550         req = request_find(fc, oh.unique);
1551         if (!req)
1552                 goto err_unlock;
1553
1554         if (req->aborted) {
1555                 spin_unlock(&fc->lock);
1556                 fuse_copy_finish(cs);
1557                 spin_lock(&fc->lock);
1558                 request_end(fc, req);
1559                 return -ENOENT;
1560         }
1561         /* Is it an interrupt reply? */
1562         if (req->intr_unique == oh.unique) {
1563                 err = -EINVAL;
1564                 if (nbytes != sizeof(struct fuse_out_header))
1565                         goto err_unlock;
1566
1567                 if (oh.error == -ENOSYS)
1568                         fc->no_interrupt = 1;
1569                 else if (oh.error == -EAGAIN)
1570                         queue_interrupt(fc, req);
1571
1572                 spin_unlock(&fc->lock);
1573                 fuse_copy_finish(cs);
1574                 return nbytes;
1575         }
1576
1577         req->state = FUSE_REQ_WRITING;
1578         list_move(&req->list, &fc->io);
1579         req->out.h = oh;
1580         req->locked = 1;
1581         cs->req = req;
1582         if (!req->out.page_replace)
1583                 cs->move_pages = 0;
1584         spin_unlock(&fc->lock);
1585
1586         err = copy_out_args(cs, &req->out, nbytes);
1587         fuse_copy_finish(cs);
1588
1589         spin_lock(&fc->lock);
1590         req->locked = 0;
1591         if (!err) {
1592                 if (req->aborted)
1593                         err = -ENOENT;
1594         } else if (!req->aborted)
1595                 req->out.h.error = -EIO;
1596         request_end(fc, req);
1597
1598         return err ? err : nbytes;
1599
1600  err_unlock:
1601         spin_unlock(&fc->lock);
1602  err_finish:
1603         fuse_copy_finish(cs);
1604         return err;
1605 }
1606
1607 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
1608                               unsigned long nr_segs, loff_t pos)
1609 {
1610         struct fuse_copy_state cs;
1611         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1612         if (!fc)
1613                 return -EPERM;
1614
1615         fuse_copy_init(&cs, fc, 0, iov, nr_segs);
1616
1617         return fuse_dev_do_write(fc, &cs, iov_length(iov, nr_segs));
1618 }
1619
1620 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1621                                      struct file *out, loff_t *ppos,
1622                                      size_t len, unsigned int flags)
1623 {
1624         unsigned nbuf;
1625         unsigned idx;
1626         struct pipe_buffer *bufs;
1627         struct fuse_copy_state cs;
1628         struct fuse_conn *fc;
1629         size_t rem;
1630         ssize_t ret;
1631
1632         fc = fuse_get_conn(out);
1633         if (!fc)
1634                 return -EPERM;
1635
1636         bufs = kmalloc(pipe->buffers * sizeof (struct pipe_buffer), GFP_KERNEL);
1637         if (!bufs)
1638                 return -ENOMEM;
1639
1640         pipe_lock(pipe);
1641         nbuf = 0;
1642         rem = 0;
1643         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1644                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1645
1646         ret = -EINVAL;
1647         if (rem < len) {
1648                 pipe_unlock(pipe);
1649                 goto out;
1650         }
1651
1652         rem = len;
1653         while (rem) {
1654                 struct pipe_buffer *ibuf;
1655                 struct pipe_buffer *obuf;
1656
1657                 BUG_ON(nbuf >= pipe->buffers);
1658                 BUG_ON(!pipe->nrbufs);
1659                 ibuf = &pipe->bufs[pipe->curbuf];
1660                 obuf = &bufs[nbuf];
1661
1662                 if (rem >= ibuf->len) {
1663                         *obuf = *ibuf;
1664                         ibuf->ops = NULL;
1665                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1666                         pipe->nrbufs--;
1667                 } else {
1668                         ibuf->ops->get(pipe, ibuf);
1669                         *obuf = *ibuf;
1670                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1671                         obuf->len = rem;
1672                         ibuf->offset += obuf->len;
1673                         ibuf->len -= obuf->len;
1674                 }
1675                 nbuf++;
1676                 rem -= obuf->len;
1677         }
1678         pipe_unlock(pipe);
1679
1680         fuse_copy_init(&cs, fc, 0, NULL, nbuf);
1681         cs.pipebufs = bufs;
1682         cs.pipe = pipe;
1683
1684         if (flags & SPLICE_F_MOVE)
1685                 cs.move_pages = 1;
1686
1687         ret = fuse_dev_do_write(fc, &cs, len);
1688
1689         for (idx = 0; idx < nbuf; idx++) {
1690                 struct pipe_buffer *buf = &bufs[idx];
1691                 buf->ops->release(pipe, buf);
1692         }
1693 out:
1694         kfree(bufs);
1695         return ret;
1696 }
1697
1698 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1699 {
1700         unsigned mask = POLLOUT | POLLWRNORM;
1701         struct fuse_conn *fc = fuse_get_conn(file);
1702         if (!fc)
1703                 return POLLERR;
1704
1705         poll_wait(file, &fc->waitq, wait);
1706
1707         spin_lock(&fc->lock);
1708         if (!fc->connected)
1709                 mask = POLLERR;
1710         else if (request_pending(fc))
1711                 mask |= POLLIN | POLLRDNORM;
1712         spin_unlock(&fc->lock);
1713
1714         return mask;
1715 }
1716
1717 /*
1718  * Abort all requests on the given list (pending or processing)
1719  *
1720  * This function releases and reacquires fc->lock
1721  */
1722 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1723 __releases(fc->lock)
1724 __acquires(fc->lock)
1725 {
1726         while (!list_empty(head)) {
1727                 struct fuse_req *req;
1728                 req = list_entry(head->next, struct fuse_req, list);
1729                 req->out.h.error = -ECONNABORTED;
1730                 request_end(fc, req);
1731                 spin_lock(&fc->lock);
1732         }
1733 }
1734
1735 /*
1736  * Abort requests under I/O
1737  *
1738  * The requests are set to aborted and finished, and the request
1739  * waiter is woken up.  This will make request_wait_answer() wait
1740  * until the request is unlocked and then return.
1741  *
1742  * If the request is asynchronous, then the end function needs to be
1743  * called after waiting for the request to be unlocked (if it was
1744  * locked).
1745  */
1746 static void end_io_requests(struct fuse_conn *fc)
1747 __releases(fc->lock)
1748 __acquires(fc->lock)
1749 {
1750         while (!list_empty(&fc->io)) {
1751                 struct fuse_req *req =
1752                         list_entry(fc->io.next, struct fuse_req, list);
1753                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1754
1755                 req->aborted = 1;
1756                 req->out.h.error = -ECONNABORTED;
1757                 req->state = FUSE_REQ_FINISHED;
1758                 list_del_init(&req->list);
1759                 wake_up(&req->waitq);
1760                 if (end) {
1761                         req->end = NULL;
1762                         __fuse_get_request(req);
1763                         spin_unlock(&fc->lock);
1764                         wait_event(req->waitq, !req->locked);
1765                         end(fc, req);
1766                         fuse_put_request(fc, req);
1767                         spin_lock(&fc->lock);
1768                 }
1769         }
1770 }
1771
1772 static void end_queued_requests(struct fuse_conn *fc)
1773 __releases(fc->lock)
1774 __acquires(fc->lock)
1775 {
1776         fc->max_background = UINT_MAX;
1777         flush_bg_queue(fc);
1778         end_requests(fc, &fc->pending);
1779         end_requests(fc, &fc->processing);
1780 }
1781
1782 /*
1783  * Abort all requests.
1784  *
1785  * Emergency exit in case of a malicious or accidental deadlock, or
1786  * just a hung filesystem.
1787  *
1788  * The same effect is usually achievable through killing the
1789  * filesystem daemon and all users of the filesystem.  The exception
1790  * is the combination of an asynchronous request and the tricky
1791  * deadlock (see Documentation/filesystems/fuse.txt).
1792  *
1793  * During the aborting, progression of requests from the pending and
1794  * processing lists onto the io list, and progression of new requests
1795  * onto the pending list is prevented by req->connected being false.
1796  *
1797  * Progression of requests under I/O to the processing list is
1798  * prevented by the req->aborted flag being true for these requests.
1799  * For this reason requests on the io list must be aborted first.
1800  */
1801 void fuse_abort_conn(struct fuse_conn *fc)
1802 {
1803         spin_lock(&fc->lock);
1804         if (fc->connected) {
1805                 fc->connected = 0;
1806                 fc->blocked = 0;
1807                 end_io_requests(fc);
1808                 end_queued_requests(fc);
1809                 wake_up_all(&fc->waitq);
1810                 wake_up_all(&fc->blocked_waitq);
1811                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1812         }
1813         spin_unlock(&fc->lock);
1814 }
1815 EXPORT_SYMBOL_GPL(fuse_abort_conn);
1816
1817 int fuse_dev_release(struct inode *inode, struct file *file)
1818 {
1819         struct fuse_conn *fc = fuse_get_conn(file);
1820         if (fc) {
1821                 spin_lock(&fc->lock);
1822                 fc->connected = 0;
1823                 fc->blocked = 0;
1824                 end_queued_requests(fc);
1825                 wake_up_all(&fc->blocked_waitq);
1826                 spin_unlock(&fc->lock);
1827                 fuse_conn_put(fc);
1828         }
1829
1830         return 0;
1831 }
1832 EXPORT_SYMBOL_GPL(fuse_dev_release);
1833
1834 static int fuse_dev_fasync(int fd, struct file *file, int on)
1835 {
1836         struct fuse_conn *fc = fuse_get_conn(file);
1837         if (!fc)
1838                 return -EPERM;
1839
1840         /* No locking - fasync_helper does its own locking */
1841         return fasync_helper(fd, file, on, &fc->fasync);
1842 }
1843
1844 const struct file_operations fuse_dev_operations = {
1845         .owner          = THIS_MODULE,
1846         .llseek         = no_llseek,
1847         .read           = do_sync_read,
1848         .aio_read       = fuse_dev_read,
1849         .splice_read    = fuse_dev_splice_read,
1850         .write          = do_sync_write,
1851         .aio_write      = fuse_dev_write,
1852         .splice_write   = fuse_dev_splice_write,
1853         .poll           = fuse_dev_poll,
1854         .release        = fuse_dev_release,
1855         .fasync         = fuse_dev_fasync,
1856 };
1857 EXPORT_SYMBOL_GPL(fuse_dev_operations);
1858
1859 static struct miscdevice fuse_miscdevice = {
1860         .minor = FUSE_MINOR,
1861         .name  = "fuse",
1862         .fops = &fuse_dev_operations,
1863 };
1864
1865 int __init fuse_dev_init(void)
1866 {
1867         int err = -ENOMEM;
1868         fuse_req_cachep = kmem_cache_create("fuse_request",
1869                                             sizeof(struct fuse_req),
1870                                             0, 0, NULL);
1871         if (!fuse_req_cachep)
1872                 goto out;
1873
1874         err = misc_register(&fuse_miscdevice);
1875         if (err)
1876                 goto out_cache_clean;
1877
1878         return 0;
1879
1880  out_cache_clean:
1881         kmem_cache_destroy(fuse_req_cachep);
1882  out:
1883         return err;
1884 }
1885
1886 void fuse_dev_cleanup(void)
1887 {
1888         misc_deregister(&fuse_miscdevice);
1889         kmem_cache_destroy(fuse_req_cachep);
1890 }