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         if (fc->connected) {
 262                 fc->forget_list_tail->next = forget;
 263                 fc->forget_list_tail = forget;
 264                 wake_up(&fc->waitq);
 265                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 266         } else {
 267                 kfree(forget);
 268         }
 269         spin_unlock(&fc->lock);
 270 }
 271
 272 static void flush_bg_queue(struct fuse_conn *fc)
 273 {
 274         while (fc->active_background < fc->max_background &&
 275                !list_empty(&fc->bg_queue)) {
 276                 struct fuse_req *req;
 277
 278                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 279                 list_del(&req->list);
 280                 fc->active_background++;
 281                 req->in.h.unique = fuse_get_unique(fc);
 282                 queue_request(fc, req);
 283         }
 284 }
 285
 286 /*
 287  * This function is called when a request is finished.  Either a reply
 288  * has arrived or it was aborted (and not yet sent) or some error
 289  * occurred during communication with userspace, or the device file
 290  * was closed.  The requester thread is woken up (if still waiting),
 291  * the 'end' callback is called if given, else the reference to the
 292  * request is released
 293  *
 294  * Called with fc->lock, unlocks it
 295  */
 296 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 297 __releases(fc->lock)
 298 {
 299         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 300         req->end = NULL;
 301         list_del(&req->list);
 302         list_del(&req->intr_entry);
 303         req->state = FUSE_REQ_FINISHED;
 304         if (req->background) {
 305                 if (fc->num_background == fc->max_background) {
 306                         fc->blocked = 0;
 307                         wake_up_all(&fc->blocked_waitq);
 308                 }
 309                 if (fc->num_background == fc->congestion_threshold &&
 310                     fc->connected && fc->bdi_initialized) {
 311                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 312                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 313                 }
 314                 fc->num_background--;
 315                 fc->active_background--;
 316                 flush_bg_queue(fc);
 317         }
 318         spin_unlock(&fc->lock);
 319         wake_up(&req->waitq);
 320         if (end)
 321                 end(fc, req);
 322         fuse_put_request(fc, req);
 323 }
 324
 325 static void wait_answer_interruptible(struct fuse_conn *fc,
 326                                       struct fuse_req *req)
 327 __releases(fc->lock)
 328 __acquires(fc->lock)
 329 {
 330         if (signal_pending(current))
 331                 return;
 332
 333         spin_unlock(&fc->lock);
 334         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 335         spin_lock(&fc->lock);
 336 }
 337
 338 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 339 {
 340         list_add_tail(&req->intr_entry, &fc->interrupts);
 341         wake_up(&fc->waitq);
 342         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 343 }
 344
 345 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 346 __releases(fc->lock)
 347 __acquires(fc->lock)
 348 {
 349         if (!fc->no_interrupt) {
 350                 /* Any signal may interrupt this */
 351                 wait_answer_interruptible(fc, req);
 352
 353                 if (req->aborted)
 354                         goto aborted;
 355                 if (req->state == FUSE_REQ_FINISHED)
 356                         return;
 357
 358                 req->interrupted = 1;
 359                 if (req->state == FUSE_REQ_SENT)
 360                         queue_interrupt(fc, req);
 361         }
 362
 363         if (!req->force) {
 364                 sigset_t oldset;
 365
 366                 /* Only fatal signals may interrupt this */
 367                 block_sigs(&oldset);
 368                 wait_answer_interruptible(fc, req);
 369                 restore_sigs(&oldset);
 370
 371                 if (req->aborted)
 372                         goto aborted;
 373                 if (req->state == FUSE_REQ_FINISHED)
 374                         return;
 375
 376                 /* Request is not yet in userspace, bail out */
 377                 if (req->state == FUSE_REQ_PENDING) {
 378                         list_del(&req->list);
 379                         __fuse_put_request(req);
 380                         req->out.h.error = -EINTR;
 381                         return;
 382                 }
 383         }
 384
 385         /*
 386          * Either request is already in userspace, or it was forced.
 387          * Wait it out.
 388          */
 389         spin_unlock(&fc->lock);
 390         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 391         spin_lock(&fc->lock);
 392
 393         if (!req->aborted)
 394                 return;
 395
 396  aborted:
 397         BUG_ON(req->state != FUSE_REQ_FINISHED);
 398         if (req->locked) {
 399                 /* This is uninterruptible sleep, because data is
 400                    being copied to/from the buffers of req.  During
 401                    locked state, there mustn't be any filesystem
 402                    operation (e.g. page fault), since that could lead
 403                    to deadlock */
 404                 spin_unlock(&fc->lock);
 405                 wait_event(req->waitq, !req->locked);
 406                 spin_lock(&fc->lock);
 407         }
 408 }
 409
 410 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 411 {
 412         req->isreply = 1;
 413         spin_lock(&fc->lock);
 414         if (!fc->connected)
 415                 req->out.h.error = -ENOTCONN;
 416         else if (fc->conn_error)
 417                 req->out.h.error = -ECONNREFUSED;
 418         else {
 419                 req->in.h.unique = fuse_get_unique(fc);
 420                 queue_request(fc, req);
 421                 /* acquire extra reference, since request is still needed
 422                    after request_end() */
 423                 __fuse_get_request(req);
 424
 425                 request_wait_answer(fc, req);
 426         }
 427         spin_unlock(&fc->lock);
 428 }
 429 EXPORT_SYMBOL_GPL(fuse_request_send);
 430
 431 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 432                                             struct fuse_req *req)
 433 {
 434         req->background = 1;
 435         fc->num_background++;
 436         if (fc->num_background == fc->max_background)
 437                 fc->blocked = 1;
 438         if (fc->num_background == fc->congestion_threshold &&
 439             fc->bdi_initialized) {
 440                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 441                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 442         }
 443         list_add_tail(&req->list, &fc->bg_queue);
 444         flush_bg_queue(fc);
 445 }
 446
 447 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 448 {
 449         spin_lock(&fc->lock);
 450         if (fc->connected) {
 451                 fuse_request_send_nowait_locked(fc, req);
 452                 spin_unlock(&fc->lock);
 453         } else {
 454                 req->out.h.error = -ENOTCONN;
 455                 request_end(fc, req);
 456         }
 457 }
 458
 459 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 460 {
 461         req->isreply = 1;
 462         fuse_request_send_nowait(fc, req);
 463 }
 464 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 465
 466 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 467                                           struct fuse_req *req, u64 unique)
 468 {
 469         int err = -ENODEV;
 470
 471         req->isreply = 0;
 472         req->in.h.unique = unique;
 473         spin_lock(&fc->lock);
 474         if (fc->connected) {
 475                 queue_request(fc, req);
 476                 err = 0;
 477         }
 478         spin_unlock(&fc->lock);
 479
 480         return err;
 481 }
 482
 483 /*
 484  * Called under fc->lock
 485  *
 486  * fc->connected must have been checked previously
 487  */
 488 void fuse_request_send_background_locked(struct fuse_conn *fc,
 489                                          struct fuse_req *req)
 490 {
 491         req->isreply = 1;
 492         fuse_request_send_nowait_locked(fc, req);
 493 }
 494
 495 /*
 496  * Lock the request.  Up to the next unlock_request() there mustn't be
 497  * anything that could cause a page-fault.  If the request was already
 498  * aborted bail out.
 499  */
 500 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 501 {
 502         int err = 0;
 503         if (req) {
 504                 spin_lock(&fc->lock);
 505                 if (req->aborted)
 506                         err = -ENOENT;
 507                 else
 508                         req->locked = 1;
 509                 spin_unlock(&fc->lock);
 510         }
 511         return err;
 512 }
 513
 514 /*
 515  * Unlock request.  If it was aborted during being locked, the
 516  * requester thread is currently waiting for it to be unlocked, so
 517  * wake it up.
 518  */
 519 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 520 {
 521         if (req) {
 522                 spin_lock(&fc->lock);
 523                 req->locked = 0;
 524                 if (req->aborted)
 525                         wake_up(&req->waitq);
 526                 spin_unlock(&fc->lock);
 527         }
 528 }
 529
 530 struct fuse_copy_state {
 531         struct fuse_conn *fc;
 532         int write;
 533         struct fuse_req *req;
 534         const struct iovec *iov;
 535         struct pipe_buffer *pipebufs;
 536         struct pipe_buffer *currbuf;
 537         struct pipe_inode_info *pipe;
 538         unsigned long nr_segs;
 539         unsigned long seglen;
 540         unsigned long addr;
 541         struct page *pg;
 542         void *mapaddr;
 543         void *buf;
 544         unsigned len;
 545         unsigned move_pages:1;
 546 };
 547
 548 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 549                            int write,
 550                            const struct iovec *iov, unsigned long nr_segs)
 551 {
 552         memset(cs, 0, sizeof(*cs));
 553         cs->fc = fc;
 554         cs->write = write;
 555         cs->iov = iov;
 556         cs->nr_segs = nr_segs;
 557 }
 558
 559 /* Unmap and put previous page of userspace buffer */
 560 static void fuse_copy_finish(struct fuse_copy_state *cs)
 561 {
 562         if (cs->currbuf) {
 563                 struct pipe_buffer *buf = cs->currbuf;
 564
 565                 if (!cs->write) {
 566                         buf->ops->unmap(cs->pipe, buf, cs->mapaddr);
 567                 } else {
 568                         kunmap(buf->page);
 569                         buf->len = PAGE_SIZE - cs->len;
 570                 }
 571                 cs->currbuf = NULL;
 572                 cs->mapaddr = NULL;
 573         } else if (cs->mapaddr) {
 574                 kunmap(cs->pg);
 575                 if (cs->write) {
 576                         flush_dcache_page(cs->pg);
 577                         set_page_dirty_lock(cs->pg);
 578                 }
 579                 put_page(cs->pg);
 580                 cs->mapaddr = NULL;
 581         }
 582 }
 583
 584 /*
 585  * Get another pagefull of userspace buffer, and map it to kernel
 586  * address space, and lock request
 587  */
 588 static int fuse_copy_fill(struct fuse_copy_state *cs)
 589 {
 590         unsigned long offset;
 591         int err;
 592
 593         unlock_request(cs->fc, cs->req);
 594         fuse_copy_finish(cs);
 595         if (cs->pipebufs) {
 596                 struct pipe_buffer *buf = cs->pipebufs;
 597
 598                 if (!cs->write) {
 599                         err = buf->ops->confirm(cs->pipe, buf);
 600                         if (err)
 601                                 return err;
 602
 603                         BUG_ON(!cs->nr_segs);
 604                         cs->currbuf = buf;
 605                         cs->mapaddr = buf->ops->map(cs->pipe, buf, 0);
 606                         cs->len = buf->len;
 607                         cs->buf = cs->mapaddr + buf->offset;
 608                         cs->pipebufs++;
 609                         cs->nr_segs--;
 610                 } else {
 611                         struct page *page;
 612
 613                         if (cs->nr_segs == cs->pipe->buffers)
 614                                 return -EIO;
 615
 616                         page = alloc_page(GFP_HIGHUSER);
 617                         if (!page)
 618                                 return -ENOMEM;
 619
 620                         buf->page = page;
 621                         buf->offset = 0;
 622                         buf->len = 0;
 623
 624                         cs->currbuf = buf;
 625                         cs->mapaddr = kmap(page);
 626                         cs->buf = cs->mapaddr;
 627                         cs->len = PAGE_SIZE;
 628                         cs->pipebufs++;
 629                         cs->nr_segs++;
 630                 }
 631         } else {
 632                 if (!cs->seglen) {
 633                         BUG_ON(!cs->nr_segs);
 634                         cs->seglen = cs->iov[0].iov_len;
 635                         cs->addr = (unsigned long) cs->iov[0].iov_base;
 636                         cs->iov++;
 637                         cs->nr_segs--;
 638                 }
 639                 err = get_user_pages_fast(cs->addr, 1, cs->write, &cs->pg);
 640                 if (err < 0)
 641                         return err;
 642                 BUG_ON(err != 1);
 643                 offset = cs->addr % PAGE_SIZE;
 644                 cs->mapaddr = kmap(cs->pg);
 645                 cs->buf = cs->mapaddr + offset;
 646                 cs->len = min(PAGE_SIZE - offset, cs->seglen);
 647                 cs->seglen -= cs->len;
 648                 cs->addr += cs->len;
 649         }
 650
 651         return lock_request(cs->fc, cs->req);
 652 }
 653
 654 /* Do as much copy to/from userspace buffer as we can */
 655 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 656 {
 657         unsigned ncpy = min(*size, cs->len);
 658         if (val) {
 659                 if (cs->write)
 660                         memcpy(cs->buf, *val, ncpy);
 661                 else
 662                         memcpy(*val, cs->buf, ncpy);
 663                 *val += ncpy;
 664         }
 665         *size -= ncpy;
 666         cs->len -= ncpy;
 667         cs->buf += ncpy;
 668         return ncpy;
 669 }
 670
 671 static int fuse_check_page(struct page *page)
 672 {
 673         if (page_mapcount(page) ||
 674             page->mapping != NULL ||
 675             page_count(page) != 1 ||
 676             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 677              ~(1 << PG_locked |
 678                1 << PG_referenced |
 679                1 << PG_uptodate |
 680                1 << PG_lru |
 681                1 << PG_active |
 682                1 << PG_reclaim))) {
 683                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 684                 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);
 685                 return 1;
 686         }
 687         return 0;
 688 }
 689
 690 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 691 {
 692         int err;
 693         struct page *oldpage = *pagep;
 694         struct page *newpage;
 695         struct pipe_buffer *buf = cs->pipebufs;
 696         struct address_space *mapping;
 697         pgoff_t index;
 698
 699         unlock_request(cs->fc, cs->req);
 700         fuse_copy_finish(cs);
 701
 702         err = buf->ops->confirm(cs->pipe, buf);
 703         if (err)
 704                 return err;
 705
 706         BUG_ON(!cs->nr_segs);
 707         cs->currbuf = buf;
 708         cs->len = buf->len;
 709         cs->pipebufs++;
 710         cs->nr_segs--;
 711
 712         if (cs->len != PAGE_SIZE)
 713                 goto out_fallback;
 714
 715         if (buf->ops->steal(cs->pipe, buf) != 0)
 716                 goto out_fallback;
 717
 718         newpage = buf->page;
 719
 720         if (WARN_ON(!PageUptodate(newpage)))
 721                 return -EIO;
 722
 723         ClearPageMappedToDisk(newpage);
 724
 725         if (fuse_check_page(newpage) != 0)
 726                 goto out_fallback_unlock;
 727
 728         mapping = oldpage->mapping;
 729         index = oldpage->index;
 730
 731         /*
 732          * This is a new and locked page, it shouldn't be mapped or
 733          * have any special flags on it
 734          */
 735         if (WARN_ON(page_mapped(oldpage)))
 736                 goto out_fallback_unlock;
 737         if (WARN_ON(page_has_private(oldpage)))
 738                 goto out_fallback_unlock;
 739         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 740                 goto out_fallback_unlock;
 741         if (WARN_ON(PageMlocked(oldpage)))
 742                 goto out_fallback_unlock;
 743
 744         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 745         if (err) {
 746                 unlock_page(newpage);
 747                 return err;
 748         }
 749
 750         page_cache_get(newpage);
 751
 752         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 753                 lru_cache_add_file(newpage);
 754
 755         err = 0;
 756         spin_lock(&cs->fc->lock);
 757         if (cs->req->aborted)
 758                 err = -ENOENT;
 759         else
 760                 *pagep = newpage;
 761         spin_unlock(&cs->fc->lock);
 762
 763         if (err) {
 764                 unlock_page(newpage);
 765                 page_cache_release(newpage);
 766                 return err;
 767         }
 768
 769         unlock_page(oldpage);
 770         page_cache_release(oldpage);
 771         cs->len = 0;
 772
 773         return 0;
 774
 775 out_fallback_unlock:
 776         unlock_page(newpage);
 777 out_fallback:
 778         cs->mapaddr = buf->ops->map(cs->pipe, buf, 1);
 779         cs->buf = cs->mapaddr + buf->offset;
 780
 781         err = lock_request(cs->fc, cs->req);
 782         if (err)
 783                 return err;
 784
 785         return 1;
 786 }
 787
 788 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 789                          unsigned offset, unsigned count)
 790 {
 791         struct pipe_buffer *buf;
 792
 793         if (cs->nr_segs == cs->pipe->buffers)
 794                 return -EIO;
 795
 796         unlock_request(cs->fc, cs->req);
 797         fuse_copy_finish(cs);
 798
 799         buf = cs->pipebufs;
 800         page_cache_get(page);
 801         buf->page = page;
 802         buf->offset = offset;
 803         buf->len = count;
 804
 805         cs->pipebufs++;
 806         cs->nr_segs++;
 807         cs->len = 0;
 808
 809         return 0;
 810 }
 811
 812 /*
 813  * Copy a page in the request to/from the userspace buffer.  Must be
 814  * done atomically
 815  */
 816 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 817                           unsigned offset, unsigned count, int zeroing)
 818 {
 819         int err;
 820         struct page *page = *pagep;
 821
 822         if (page && zeroing && count < PAGE_SIZE)
 823                 clear_highpage(page);
 824
 825         while (count) {
 826                 if (cs->write && cs->pipebufs && page) {
 827                         return fuse_ref_page(cs, page, offset, count);
 828                 } else if (!cs->len) {
 829                         if (cs->move_pages && page &&
 830                             offset == 0 && count == PAGE_SIZE) {
 831                                 err = fuse_try_move_page(cs, pagep);
 832                                 if (err <= 0)
 833                                         return err;
 834                         } else {
 835                                 err = fuse_copy_fill(cs);
 836                                 if (err)
 837                                         return err;
 838                         }
 839                 }
 840                 if (page) {
 841                         void *mapaddr = kmap_atomic(page, KM_USER0);
 842                         void *buf = mapaddr + offset;
 843                         offset += fuse_copy_do(cs, &buf, &count);
 844                         kunmap_atomic(mapaddr, KM_USER0);
 845                 } else
 846                         offset += fuse_copy_do(cs, NULL, &count);
 847         }
 848         if (page && !cs->write)
 849                 flush_dcache_page(page);
 850         return 0;
 851 }
 852
 853 /* Copy pages in the request to/from userspace buffer */
 854 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 855                            int zeroing)
 856 {
 857         unsigned i;
 858         struct fuse_req *req = cs->req;
 859         unsigned offset = req->page_offset;
 860         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 861
 862         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 863                 int err;
 864
 865                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
 866                                      zeroing);
 867                 if (err)
 868                         return err;
 869
 870                 nbytes -= count;
 871                 count = min(nbytes, (unsigned) PAGE_SIZE);
 872                 offset = 0;
 873         }
 874         return 0;
 875 }
 876
 877 /* Copy a single argument in the request to/from userspace buffer */
 878 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 879 {
 880         while (size) {
 881                 if (!cs->len) {
 882                         int err = fuse_copy_fill(cs);
 883                         if (err)
 884                                 return err;
 885                 }
 886                 fuse_copy_do(cs, &val, &size);
 887         }
 888         return 0;
 889 }
 890
 891 /* Copy request arguments to/from userspace buffer */
 892 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 893                           unsigned argpages, struct fuse_arg *args,
 894                           int zeroing)
 895 {
 896         int err = 0;
 897         unsigned i;
 898
 899         for (i = 0; !err && i < numargs; i++)  {
 900                 struct fuse_arg *arg = &args[i];
 901                 if (i == numargs - 1 && argpages)
 902                         err = fuse_copy_pages(cs, arg->size, zeroing);
 903                 else
 904                         err = fuse_copy_one(cs, arg->value, arg->size);
 905         }
 906         return err;
 907 }
 908
 909 static int forget_pending(struct fuse_conn *fc)
 910 {
 911         return fc->forget_list_head.next != NULL;
 912 }
 913
 914 static int request_pending(struct fuse_conn *fc)
 915 {
 916         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts) ||
 917                 forget_pending(fc);
 918 }
 919
 920 /* Wait until a request is available on the pending list */
 921 static void request_wait(struct fuse_conn *fc)
 922 __releases(fc->lock)
 923 __acquires(fc->lock)
 924 {
 925         DECLARE_WAITQUEUE(wait, current);
 926
 927         add_wait_queue_exclusive(&fc->waitq, &wait);
 928         while (fc->connected && !request_pending(fc)) {
 929                 set_current_state(TASK_INTERRUPTIBLE);
 930                 if (signal_pending(current))
 931                         break;
 932
 933                 spin_unlock(&fc->lock);
 934                 schedule();
 935                 spin_lock(&fc->lock);
 936         }
 937         set_current_state(TASK_RUNNING);
 938         remove_wait_queue(&fc->waitq, &wait);
 939 }
 940
 941 /*
 942  * Transfer an interrupt request to userspace
 943  *
 944  * Unlike other requests this is assembled on demand, without a need
 945  * to allocate a separate fuse_req structure.
 946  *
 947  * Called with fc->lock held, releases it
 948  */
 949 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
 950                                size_t nbytes, struct fuse_req *req)
 951 __releases(fc->lock)
 952 {
 953         struct fuse_in_header ih;
 954         struct fuse_interrupt_in arg;
 955         unsigned reqsize = sizeof(ih) + sizeof(arg);
 956         int err;
 957
 958         list_del_init(&req->intr_entry);
 959         req->intr_unique = fuse_get_unique(fc);
 960         memset(&ih, 0, sizeof(ih));
 961         memset(&arg, 0, sizeof(arg));
 962         ih.len = reqsize;
 963         ih.opcode = FUSE_INTERRUPT;
 964         ih.unique = req->intr_unique;
 965         arg.unique = req->in.h.unique;
 966
 967         spin_unlock(&fc->lock);
 968         if (nbytes < reqsize)
 969                 return -EINVAL;
 970
 971         err = fuse_copy_one(cs, &ih, sizeof(ih));
 972         if (!err)
 973                 err = fuse_copy_one(cs, &arg, sizeof(arg));
 974         fuse_copy_finish(cs);
 975
 976         return err ? err : reqsize;
 977 }
 978
 979 static struct fuse_forget_link *dequeue_forget(struct fuse_conn *fc,
 980                                                unsigned max,
 981                                                unsigned *countp)
 982 {
 983         struct fuse_forget_link *head = fc->forget_list_head.next;
 984         struct fuse_forget_link **newhead = &head;
 985         unsigned count;
 986
 987         for (count = 0; *newhead != NULL && count < max; count++)
 988                 newhead = &(*newhead)->next;
 989
 990         fc->forget_list_head.next = *newhead;
 991         *newhead = NULL;
 992         if (fc->forget_list_head.next == NULL)
 993                 fc->forget_list_tail = &fc->forget_list_head;
 994
 995         if (countp != NULL)
 996                 *countp = count;
 997
 998         return head;
 999 }
1000
1001 static int fuse_read_single_forget(struct fuse_conn *fc,
1002                                    struct fuse_copy_state *cs,
1003                                    size_t nbytes)
1004 __releases(fc->lock)
1005 {
1006         int err;
1007         struct fuse_forget_link *forget = dequeue_forget(fc, 1, NULL);
1008         struct fuse_forget_in arg = {
1009                 .nlookup = forget->forget_one.nlookup,
1010         };
1011         struct fuse_in_header ih = {
1012                 .opcode = FUSE_FORGET,
1013                 .nodeid = forget->forget_one.nodeid,
1014                 .unique = fuse_get_unique(fc),
1015                 .len = sizeof(ih) + sizeof(arg),
1016         };
1017
1018         spin_unlock(&fc->lock);
1019         kfree(forget);
1020         if (nbytes < ih.len)
1021                 return -EINVAL;
1022
1023         err = fuse_copy_one(cs, &ih, sizeof(ih));
1024         if (!err)
1025                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1026         fuse_copy_finish(cs);
1027
1028         if (err)
1029                 return err;
1030
1031         return ih.len;
1032 }
1033
1034 static int fuse_read_batch_forget(struct fuse_conn *fc,
1035                                    struct fuse_copy_state *cs, size_t nbytes)
1036 __releases(fc->lock)
1037 {
1038         int err;
1039         unsigned max_forgets;
1040         unsigned count;
1041         struct fuse_forget_link *head;
1042         struct fuse_batch_forget_in arg = { .count = 0 };
1043         struct fuse_in_header ih = {
1044                 .opcode = FUSE_BATCH_FORGET,
1045                 .unique = fuse_get_unique(fc),
1046                 .len = sizeof(ih) + sizeof(arg),
1047         };
1048
1049         if (nbytes < ih.len) {
1050                 spin_unlock(&fc->lock);
1051                 return -EINVAL;
1052         }
1053
1054         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1055         head = dequeue_forget(fc, max_forgets, &count);
1056         spin_unlock(&fc->lock);
1057
1058         arg.count = count;
1059         ih.len += count * sizeof(struct fuse_forget_one);
1060         err = fuse_copy_one(cs, &ih, sizeof(ih));
1061         if (!err)
1062                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1063
1064         while (head) {
1065                 struct fuse_forget_link *forget = head;
1066
1067                 if (!err) {
1068                         err = fuse_copy_one(cs, &forget->forget_one,
1069                                             sizeof(forget->forget_one));
1070                 }
1071                 head = forget->next;
1072                 kfree(forget);
1073         }
1074
1075         fuse_copy_finish(cs);
1076
1077         if (err)
1078                 return err;
1079
1080         return ih.len;
1081 }
1082
1083 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_copy_state *cs,
1084                             size_t nbytes)
1085 __releases(fc->lock)
1086 {
1087         if (fc->minor < 16 || fc->forget_list_head.next->next == NULL)
1088                 return fuse_read_single_forget(fc, cs, nbytes);
1089         else
1090                 return fuse_read_batch_forget(fc, cs, nbytes);
1091 }
1092
1093 /*
1094  * Read a single request into the userspace filesystem's buffer.  This
1095  * function waits until a request is available, then removes it from
1096  * the pending list and copies request data to userspace buffer.  If
1097  * no reply is needed (FORGET) or request has been aborted or there
1098  * was an error during the copying then it's finished by calling
1099  * request_end().  Otherwise add it to the processing list, and set
1100  * the 'sent' flag.
1101  */
1102 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1103                                 struct fuse_copy_state *cs, size_t nbytes)
1104 {
1105         int err;
1106         struct fuse_req *req;
1107         struct fuse_in *in;
1108         unsigned reqsize;
1109
1110  restart:
1111         spin_lock(&fc->lock);
1112         err = -EAGAIN;
1113         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
1114             !request_pending(fc))
1115                 goto err_unlock;
1116
1117         request_wait(fc);
1118         err = -ENODEV;
1119         if (!fc->connected)
1120                 goto err_unlock;
1121         err = -ERESTARTSYS;
1122         if (!request_pending(fc))
1123                 goto err_unlock;
1124
1125         if (!list_empty(&fc->interrupts)) {
1126                 req = list_entry(fc->interrupts.next, struct fuse_req,
1127                                  intr_entry);
1128                 return fuse_read_interrupt(fc, cs, nbytes, req);
1129         }
1130
1131         if (forget_pending(fc)) {
1132                 if (list_empty(&fc->pending) || fc->forget_batch-- > 0)
1133                         return fuse_read_forget(fc, cs, nbytes);
1134
1135                 if (fc->forget_batch <= -8)
1136                         fc->forget_batch = 16;
1137         }
1138
1139         req = list_entry(fc->pending.next, struct fuse_req, list);
1140         req->state = FUSE_REQ_READING;
1141         list_move(&req->list, &fc->io);
1142
1143         in = &req->in;
1144         reqsize = in->h.len;
1145         /* If request is too large, reply with an error and restart the read */
1146         if (nbytes < reqsize) {
1147                 req->out.h.error = -EIO;
1148                 /* SETXATTR is special, since it may contain too large data */
1149                 if (in->h.opcode == FUSE_SETXATTR)
1150                         req->out.h.error = -E2BIG;
1151                 request_end(fc, req);
1152                 goto restart;
1153         }
1154         spin_unlock(&fc->lock);
1155         cs->req = req;
1156         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1157         if (!err)
1158                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1159                                      (struct fuse_arg *) in->args, 0);
1160         fuse_copy_finish(cs);
1161         spin_lock(&fc->lock);
1162         req->locked = 0;
1163         if (req->aborted) {
1164                 request_end(fc, req);
1165                 return -ENODEV;
1166         }
1167         if (err) {
1168                 req->out.h.error = -EIO;
1169                 request_end(fc, req);
1170                 return err;
1171         }
1172         if (!req->isreply)
1173                 request_end(fc, req);
1174         else {
1175                 req->state = FUSE_REQ_SENT;
1176                 list_move_tail(&req->list, &fc->processing);
1177                 if (req->interrupted)
1178                         queue_interrupt(fc, req);
1179                 spin_unlock(&fc->lock);
1180         }
1181         return reqsize;
1182
1183  err_unlock:
1184         spin_unlock(&fc->lock);
1185         return err;
1186 }
1187
1188 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
1189                               unsigned long nr_segs, loff_t pos)
1190 {
1191         struct fuse_copy_state cs;
1192         struct file *file = iocb->ki_filp;
1193         struct fuse_conn *fc = fuse_get_conn(file);
1194         if (!fc)
1195                 return -EPERM;
1196
1197         fuse_copy_init(&cs, fc, 1, iov, nr_segs);
1198
1199         return fuse_dev_do_read(fc, file, &cs, iov_length(iov, nr_segs));
1200 }
1201
1202 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1203                                     struct pipe_inode_info *pipe,
1204                                     size_t len, unsigned int flags)
1205 {
1206         int ret;
1207         int page_nr = 0;
1208         int do_wakeup = 0;
1209         struct pipe_buffer *bufs;
1210         struct fuse_copy_state cs;
1211         struct fuse_conn *fc = fuse_get_conn(in);
1212         if (!fc)
1213                 return -EPERM;
1214
1215         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1216         if (!bufs)
1217                 return -ENOMEM;
1218
1219         fuse_copy_init(&cs, fc, 1, NULL, 0);
1220         cs.pipebufs = bufs;
1221         cs.pipe = pipe;
1222         ret = fuse_dev_do_read(fc, in, &cs, len);
1223         if (ret < 0)
1224                 goto out;
1225
1226         ret = 0;
1227         pipe_lock(pipe);
1228
1229         if (!pipe->readers) {
1230                 send_sig(SIGPIPE, current, 0);
1231                 if (!ret)
1232                         ret = -EPIPE;
1233                 goto out_unlock;
1234         }
1235
1236         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1237                 ret = -EIO;
1238                 goto out_unlock;
1239         }
1240
1241         while (page_nr < cs.nr_segs) {
1242                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1243                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1244
1245                 buf->page = bufs[page_nr].page;
1246                 buf->offset = bufs[page_nr].offset;
1247                 buf->len = bufs[page_nr].len;
1248                 /*
1249                  * Need to be careful about this.  Having buf->ops in module
1250                  * code can Oops if the buffer persists after module unload.
1251                  */
1252                 buf->ops = &nosteal_pipe_buf_ops;
1253
1254                 pipe->nrbufs++;
1255                 page_nr++;
1256                 ret += buf->len;
1257
1258                 if (pipe->inode)
1259                         do_wakeup = 1;
1260         }
1261
1262 out_unlock:
1263         pipe_unlock(pipe);
1264
1265         if (do_wakeup) {
1266                 smp_mb();
1267                 if (waitqueue_active(&pipe->wait))
1268                         wake_up_interruptible(&pipe->wait);
1269                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1270         }
1271
1272 out:
1273         for (; page_nr < cs.nr_segs; page_nr++)
1274                 page_cache_release(bufs[page_nr].page);
1275
1276         kfree(bufs);
1277         return ret;
1278 }
1279
1280 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1281                             struct fuse_copy_state *cs)
1282 {
1283         struct fuse_notify_poll_wakeup_out outarg;
1284         int err = -EINVAL;
1285
1286         if (size != sizeof(outarg))
1287                 goto err;
1288
1289         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1290         if (err)
1291                 goto err;
1292
1293         fuse_copy_finish(cs);
1294         return fuse_notify_poll_wakeup(fc, &outarg);
1295
1296 err:
1297         fuse_copy_finish(cs);
1298         return err;
1299 }
1300
1301 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1302                                    struct fuse_copy_state *cs)
1303 {
1304         struct fuse_notify_inval_inode_out outarg;
1305         int err = -EINVAL;
1306
1307         if (size != sizeof(outarg))
1308                 goto err;
1309
1310         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1311         if (err)
1312                 goto err;
1313         fuse_copy_finish(cs);
1314
1315         down_read(&fc->killsb);
1316         err = -ENOENT;
1317         if (fc->sb) {
1318                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1319                                                outarg.off, outarg.len);
1320         }
1321         up_read(&fc->killsb);
1322         return err;
1323
1324 err:
1325         fuse_copy_finish(cs);
1326         return err;
1327 }
1328
1329 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1330                                    struct fuse_copy_state *cs)
1331 {
1332         struct fuse_notify_inval_entry_out outarg;
1333         int err = -ENOMEM;
1334         char *buf;
1335         struct qstr name;
1336
1337         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1338         if (!buf)
1339                 goto err;
1340
1341         err = -EINVAL;
1342         if (size < sizeof(outarg))
1343                 goto err;
1344
1345         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1346         if (err)
1347                 goto err;
1348
1349         err = -ENAMETOOLONG;
1350         if (outarg.namelen > FUSE_NAME_MAX)
1351                 goto err;
1352
1353         err = -EINVAL;
1354         if (size != sizeof(outarg) + outarg.namelen + 1)
1355                 goto err;
1356
1357         name.name = buf;
1358         name.len = outarg.namelen;
1359         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1360         if (err)
1361                 goto err;
1362         fuse_copy_finish(cs);
1363         buf[outarg.namelen] = 0;
1364         name.hash = full_name_hash(name.name, name.len);
1365
1366         down_read(&fc->killsb);
1367         err = -ENOENT;
1368         if (fc->sb)
1369                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, &name);
1370         up_read(&fc->killsb);
1371         kfree(buf);
1372         return err;
1373
1374 err:
1375         kfree(buf);
1376         fuse_copy_finish(cs);
1377         return err;
1378 }
1379
1380 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1381                              struct fuse_copy_state *cs)
1382 {
1383         struct fuse_notify_store_out outarg;
1384         struct inode *inode;
1385         struct address_space *mapping;
1386         u64 nodeid;
1387         int err;
1388         pgoff_t index;
1389         unsigned int offset;
1390         unsigned int num;
1391         loff_t file_size;
1392         loff_t end;
1393
1394         err = -EINVAL;
1395         if (size < sizeof(outarg))
1396                 goto out_finish;
1397
1398         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1399         if (err)
1400                 goto out_finish;
1401
1402         err = -EINVAL;
1403         if (size - sizeof(outarg) != outarg.size)
1404                 goto out_finish;
1405
1406         nodeid = outarg.nodeid;
1407
1408         down_read(&fc->killsb);
1409
1410         err = -ENOENT;
1411         if (!fc->sb)
1412                 goto out_up_killsb;
1413
1414         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1415         if (!inode)
1416                 goto out_up_killsb;
1417
1418         mapping = inode->i_mapping;
1419         index = outarg.offset >> PAGE_CACHE_SHIFT;
1420         offset = outarg.offset & ~PAGE_CACHE_MASK;
1421         file_size = i_size_read(inode);
1422         end = outarg.offset + outarg.size;
1423         if (end > file_size) {
1424                 file_size = end;
1425                 fuse_write_update_size(inode, file_size);
1426         }
1427
1428         num = outarg.size;
1429         while (num) {
1430                 struct page *page;
1431                 unsigned int this_num;
1432
1433                 err = -ENOMEM;
1434                 page = find_or_create_page(mapping, index,
1435                                            mapping_gfp_mask(mapping));
1436                 if (!page)
1437                         goto out_iput;
1438
1439                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1440                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1441                 if (!err && offset == 0 && (num != 0 || file_size == end))
1442                         SetPageUptodate(page);
1443                 unlock_page(page);
1444                 page_cache_release(page);
1445
1446                 if (err)
1447                         goto out_iput;
1448
1449                 num -= this_num;
1450                 offset = 0;
1451                 index++;
1452         }
1453
1454         err = 0;
1455
1456 out_iput:
1457         iput(inode);
1458 out_up_killsb:
1459         up_read(&fc->killsb);
1460 out_finish:
1461         fuse_copy_finish(cs);
1462         return err;
1463 }
1464
1465 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1466 {
1467         release_pages(req->pages, req->num_pages, 0);
1468 }
1469
1470 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1471                          struct fuse_notify_retrieve_out *outarg)
1472 {
1473         int err;
1474         struct address_space *mapping = inode->i_mapping;
1475         struct fuse_req *req;
1476         pgoff_t index;
1477         loff_t file_size;
1478         unsigned int num;
1479         unsigned int offset;
1480         size_t total_len = 0;
1481
1482         req = fuse_get_req(fc);
1483         if (IS_ERR(req))
1484                 return PTR_ERR(req);
1485
1486         offset = outarg->offset & ~PAGE_CACHE_MASK;
1487
1488         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1489         req->in.h.nodeid = outarg->nodeid;
1490         req->in.numargs = 2;
1491         req->in.argpages = 1;
1492         req->page_offset = offset;
1493         req->end = fuse_retrieve_end;
1494
1495         index = outarg->offset >> PAGE_CACHE_SHIFT;
1496         file_size = i_size_read(inode);
1497         num = outarg->size;
1498         if (outarg->offset > file_size)
1499                 num = 0;
1500         else if (outarg->offset + num > file_size)
1501                 num = file_size - outarg->offset;
1502
1503         while (num && req->num_pages < FUSE_MAX_PAGES_PER_REQ) {
1504                 struct page *page;
1505                 unsigned int this_num;
1506
1507                 page = find_get_page(mapping, index);
1508                 if (!page)
1509                         break;
1510
1511                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1512                 req->pages[req->num_pages] = page;
1513                 req->num_pages++;
1514
1515                 offset = 0;
1516                 num -= this_num;
1517                 total_len += this_num;
1518                 index++;
1519         }
1520         req->misc.retrieve_in.offset = outarg->offset;
1521         req->misc.retrieve_in.size = total_len;
1522         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1523         req->in.args[0].value = &req->misc.retrieve_in;
1524         req->in.args[1].size = total_len;
1525
1526         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1527         if (err)
1528                 fuse_retrieve_end(fc, req);
1529
1530         return err;
1531 }
1532
1533 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1534                                 struct fuse_copy_state *cs)
1535 {
1536         struct fuse_notify_retrieve_out outarg;
1537         struct inode *inode;
1538         int err;
1539
1540         err = -EINVAL;
1541         if (size != sizeof(outarg))
1542                 goto copy_finish;
1543
1544         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1545         if (err)
1546                 goto copy_finish;
1547
1548         fuse_copy_finish(cs);
1549
1550         down_read(&fc->killsb);
1551         err = -ENOENT;
1552         if (fc->sb) {
1553                 u64 nodeid = outarg.nodeid;
1554
1555                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1556                 if (inode) {
1557                         err = fuse_retrieve(fc, inode, &outarg);
1558                         iput(inode);
1559                 }
1560         }
1561         up_read(&fc->killsb);
1562
1563         return err;
1564
1565 copy_finish:
1566         fuse_copy_finish(cs);
1567         return err;
1568 }
1569
1570 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1571                        unsigned int size, struct fuse_copy_state *cs)
1572 {
1573         switch (code) {
1574         case FUSE_NOTIFY_POLL:
1575                 return fuse_notify_poll(fc, size, cs);
1576
1577         case FUSE_NOTIFY_INVAL_INODE:
1578                 return fuse_notify_inval_inode(fc, size, cs);
1579
1580         case FUSE_NOTIFY_INVAL_ENTRY:
1581                 return fuse_notify_inval_entry(fc, size, cs);
1582
1583         case FUSE_NOTIFY_STORE:
1584                 return fuse_notify_store(fc, size, cs);
1585
1586         case FUSE_NOTIFY_RETRIEVE:
1587                 return fuse_notify_retrieve(fc, size, cs);
1588
1589         default:
1590                 fuse_copy_finish(cs);
1591                 return -EINVAL;
1592         }
1593 }
1594
1595 /* Look up request on processing list by unique ID */
1596 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1597 {
1598         struct list_head *entry;
1599
1600         list_for_each(entry, &fc->processing) {
1601                 struct fuse_req *req;
1602                 req = list_entry(entry, struct fuse_req, list);
1603                 if (req->in.h.unique == unique || req->intr_unique == unique)
1604                         return req;
1605         }
1606         return NULL;
1607 }
1608
1609 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1610                          unsigned nbytes)
1611 {
1612         unsigned reqsize = sizeof(struct fuse_out_header);
1613
1614         if (out->h.error)
1615                 return nbytes != reqsize ? -EINVAL : 0;
1616
1617         reqsize += len_args(out->numargs, out->args);
1618
1619         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1620                 return -EINVAL;
1621         else if (reqsize > nbytes) {
1622                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1623                 unsigned diffsize = reqsize - nbytes;
1624                 if (diffsize > lastarg->size)
1625                         return -EINVAL;
1626                 lastarg->size -= diffsize;
1627         }
1628         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1629                               out->page_zeroing);
1630 }
1631
1632 /*
1633  * Write a single reply to a request.  First the header is copied from
1634  * the write buffer.  The request is then searched on the processing
1635  * list by the unique ID found in the header.  If found, then remove
1636  * it from the list and copy the rest of the buffer to the request.
1637  * The request is finished by calling request_end()
1638  */
1639 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1640                                  struct fuse_copy_state *cs, size_t nbytes)
1641 {
1642         int err;
1643         struct fuse_req *req;
1644         struct fuse_out_header oh;
1645
1646         if (nbytes < sizeof(struct fuse_out_header))
1647                 return -EINVAL;
1648
1649         err = fuse_copy_one(cs, &oh, sizeof(oh));
1650         if (err)
1651                 goto err_finish;
1652
1653         err = -EINVAL;
1654         if (oh.len != nbytes)
1655                 goto err_finish;
1656
1657         /*
1658          * Zero oh.unique indicates unsolicited notification message
1659          * and error contains notification code.
1660          */
1661         if (!oh.unique) {
1662                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1663                 return err ? err : nbytes;
1664         }
1665
1666         err = -EINVAL;
1667         if (oh.error <= -1000 || oh.error > 0)
1668                 goto err_finish;
1669
1670         spin_lock(&fc->lock);
1671         err = -ENOENT;
1672         if (!fc->connected)
1673                 goto err_unlock;
1674
1675         req = request_find(fc, oh.unique);
1676         if (!req)
1677                 goto err_unlock;
1678
1679         if (req->aborted) {
1680                 spin_unlock(&fc->lock);
1681                 fuse_copy_finish(cs);
1682                 spin_lock(&fc->lock);
1683                 request_end(fc, req);
1684                 return -ENOENT;
1685         }
1686         /* Is it an interrupt reply? */
1687         if (req->intr_unique == oh.unique) {
1688                 err = -EINVAL;
1689                 if (nbytes != sizeof(struct fuse_out_header))
1690                         goto err_unlock;
1691
1692                 if (oh.error == -ENOSYS)
1693                         fc->no_interrupt = 1;
1694                 else if (oh.error == -EAGAIN)
1695                         queue_interrupt(fc, req);
1696
1697                 spin_unlock(&fc->lock);
1698                 fuse_copy_finish(cs);
1699                 return nbytes;
1700         }
1701
1702         req->state = FUSE_REQ_WRITING;
1703         list_move(&req->list, &fc->io);
1704         req->out.h = oh;
1705         req->locked = 1;
1706         cs->req = req;
1707         if (!req->out.page_replace)
1708                 cs->move_pages = 0;
1709         spin_unlock(&fc->lock);
1710
1711         err = copy_out_args(cs, &req->out, nbytes);
1712         fuse_copy_finish(cs);
1713
1714         spin_lock(&fc->lock);
1715         req->locked = 0;
1716         if (!err) {
1717                 if (req->aborted)
1718                         err = -ENOENT;
1719         } else if (!req->aborted)
1720                 req->out.h.error = -EIO;
1721         request_end(fc, req);
1722
1723         return err ? err : nbytes;
1724
1725  err_unlock:
1726         spin_unlock(&fc->lock);
1727  err_finish:
1728         fuse_copy_finish(cs);
1729         return err;
1730 }
1731
1732 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
1733                               unsigned long nr_segs, loff_t pos)
1734 {
1735         struct fuse_copy_state cs;
1736         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1737         if (!fc)
1738                 return -EPERM;
1739
1740         fuse_copy_init(&cs, fc, 0, iov, nr_segs);
1741
1742         return fuse_dev_do_write(fc, &cs, iov_length(iov, nr_segs));
1743 }
1744
1745 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1746                                      struct file *out, loff_t *ppos,
1747                                      size_t len, unsigned int flags)
1748 {
1749         unsigned nbuf;
1750         unsigned idx;
1751         struct pipe_buffer *bufs;
1752         struct fuse_copy_state cs;
1753         struct fuse_conn *fc;
1754         size_t rem;
1755         ssize_t ret;
1756
1757         fc = fuse_get_conn(out);
1758         if (!fc)
1759                 return -EPERM;
1760
1761         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1762         if (!bufs)
1763                 return -ENOMEM;
1764
1765         pipe_lock(pipe);
1766         nbuf = 0;
1767         rem = 0;
1768         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1769                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1770
1771         ret = -EINVAL;
1772         if (rem < len) {
1773                 pipe_unlock(pipe);
1774                 goto out;
1775         }
1776
1777         rem = len;
1778         while (rem) {
1779                 struct pipe_buffer *ibuf;
1780                 struct pipe_buffer *obuf;
1781
1782                 BUG_ON(nbuf >= pipe->buffers);
1783                 BUG_ON(!pipe->nrbufs);
1784                 ibuf = &pipe->bufs[pipe->curbuf];
1785                 obuf = &bufs[nbuf];
1786
1787                 if (rem >= ibuf->len) {
1788                         *obuf = *ibuf;
1789                         ibuf->ops = NULL;
1790                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1791                         pipe->nrbufs--;
1792                 } else {
1793                         ibuf->ops->get(pipe, ibuf);
1794                         *obuf = *ibuf;
1795                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1796                         obuf->len = rem;
1797                         ibuf->offset += obuf->len;
1798                         ibuf->len -= obuf->len;
1799                 }
1800                 nbuf++;
1801                 rem -= obuf->len;
1802         }
1803         pipe_unlock(pipe);
1804
1805         fuse_copy_init(&cs, fc, 0, NULL, nbuf);
1806         cs.pipebufs = bufs;
1807         cs.pipe = pipe;
1808
1809         if (flags & SPLICE_F_MOVE)
1810                 cs.move_pages = 1;
1811
1812         ret = fuse_dev_do_write(fc, &cs, len);
1813
1814         for (idx = 0; idx < nbuf; idx++) {
1815                 struct pipe_buffer *buf = &bufs[idx];
1816                 buf->ops->release(pipe, buf);
1817         }
1818 out:
1819         kfree(bufs);
1820         return ret;
1821 }
1822
1823 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1824 {
1825         unsigned mask = POLLOUT | POLLWRNORM;
1826         struct fuse_conn *fc = fuse_get_conn(file);
1827         if (!fc)
1828                 return POLLERR;
1829
1830         poll_wait(file, &fc->waitq, wait);
1831
1832         spin_lock(&fc->lock);
1833         if (!fc->connected)
1834                 mask = POLLERR;
1835         else if (request_pending(fc))
1836                 mask |= POLLIN | POLLRDNORM;
1837         spin_unlock(&fc->lock);
1838
1839         return mask;
1840 }
1841
1842 /*
1843  * Abort all requests on the given list (pending or processing)
1844  *
1845  * This function releases and reacquires fc->lock
1846  */
1847 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1848 __releases(fc->lock)
1849 __acquires(fc->lock)
1850 {
1851         while (!list_empty(head)) {
1852                 struct fuse_req *req;
1853                 req = list_entry(head->next, struct fuse_req, list);
1854                 req->out.h.error = -ECONNABORTED;
1855                 request_end(fc, req);
1856                 spin_lock(&fc->lock);
1857         }
1858 }
1859
1860 /*
1861  * Abort requests under I/O
1862  *
1863  * The requests are set to aborted and finished, and the request
1864  * waiter is woken up.  This will make request_wait_answer() wait
1865  * until the request is unlocked and then return.
1866  *
1867  * If the request is asynchronous, then the end function needs to be
1868  * called after waiting for the request to be unlocked (if it was
1869  * locked).
1870  */
1871 static void end_io_requests(struct fuse_conn *fc)
1872 __releases(fc->lock)
1873 __acquires(fc->lock)
1874 {
1875         while (!list_empty(&fc->io)) {
1876                 struct fuse_req *req =
1877                         list_entry(fc->io.next, struct fuse_req, list);
1878                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1879
1880                 req->aborted = 1;
1881                 req->out.h.error = -ECONNABORTED;
1882                 req->state = FUSE_REQ_FINISHED;
1883                 list_del_init(&req->list);
1884                 wake_up(&req->waitq);
1885                 if (end) {
1886                         req->end = NULL;
1887                         __fuse_get_request(req);
1888                         spin_unlock(&fc->lock);
1889                         wait_event(req->waitq, !req->locked);
1890                         end(fc, req);
1891                         fuse_put_request(fc, req);
1892                         spin_lock(&fc->lock);
1893                 }
1894         }
1895 }
1896
1897 static void end_queued_requests(struct fuse_conn *fc)
1898 __releases(fc->lock)
1899 __acquires(fc->lock)
1900 {
1901         fc->max_background = UINT_MAX;
1902         flush_bg_queue(fc);
1903         end_requests(fc, &fc->pending);
1904         end_requests(fc, &fc->processing);
1905         while (forget_pending(fc))
1906                 kfree(dequeue_forget(fc, 1, NULL));
1907 }
1908
1909 static void end_polls(struct fuse_conn *fc)
1910 {
1911         struct rb_node *p;
1912
1913         p = rb_first(&fc->polled_files);
1914
1915         while (p) {
1916                 struct fuse_file *ff;
1917                 ff = rb_entry(p, struct fuse_file, polled_node);
1918                 wake_up_interruptible_all(&ff->poll_wait);
1919
1920                 p = rb_next(p);
1921         }
1922 }
1923
1924 /*
1925  * Abort all requests.
1926  *
1927  * Emergency exit in case of a malicious or accidental deadlock, or
1928  * just a hung filesystem.
1929  *
1930  * The same effect is usually achievable through killing the
1931  * filesystem daemon and all users of the filesystem.  The exception
1932  * is the combination of an asynchronous request and the tricky
1933  * deadlock (see Documentation/filesystems/fuse.txt).
1934  *
1935  * During the aborting, progression of requests from the pending and
1936  * processing lists onto the io list, and progression of new requests
1937  * onto the pending list is prevented by req->connected being false.
1938  *
1939  * Progression of requests under I/O to the processing list is
1940  * prevented by the req->aborted flag being true for these requests.
1941  * For this reason requests on the io list must be aborted first.
1942  */
1943 void fuse_abort_conn(struct fuse_conn *fc)
1944 {
1945         spin_lock(&fc->lock);
1946         if (fc->connected) {
1947                 fc->connected = 0;
1948                 fc->blocked = 0;
1949                 end_io_requests(fc);
1950                 end_queued_requests(fc);
1951                 end_polls(fc);
1952                 wake_up_all(&fc->waitq);
1953                 wake_up_all(&fc->blocked_waitq);
1954                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1955         }
1956         spin_unlock(&fc->lock);
1957 }
1958 EXPORT_SYMBOL_GPL(fuse_abort_conn);
1959
1960 int fuse_dev_release(struct inode *inode, struct file *file)
1961 {
1962         struct fuse_conn *fc = fuse_get_conn(file);
1963         if (fc) {
1964                 spin_lock(&fc->lock);
1965                 fc->connected = 0;
1966                 fc->blocked = 0;
1967                 end_queued_requests(fc);
1968                 end_polls(fc);
1969                 wake_up_all(&fc->blocked_waitq);
1970                 spin_unlock(&fc->lock);
1971                 fuse_conn_put(fc);
1972         }
1973
1974         return 0;
1975 }
1976 EXPORT_SYMBOL_GPL(fuse_dev_release);
1977
1978 static int fuse_dev_fasync(int fd, struct file *file, int on)
1979 {
1980         struct fuse_conn *fc = fuse_get_conn(file);
1981         if (!fc)
1982                 return -EPERM;
1983
1984         /* No locking - fasync_helper does its own locking */
1985         return fasync_helper(fd, file, on, &fc->fasync);
1986 }
1987
1988 const struct file_operations fuse_dev_operations = {
1989         .owner          = THIS_MODULE,
1990         .llseek         = no_llseek,
1991         .read           = do_sync_read,
1992         .aio_read       = fuse_dev_read,
1993         .splice_read    = fuse_dev_splice_read,
1994         .write          = do_sync_write,
1995         .aio_write      = fuse_dev_write,
1996         .splice_write   = fuse_dev_splice_write,
1997         .poll           = fuse_dev_poll,
1998         .release        = fuse_dev_release,
1999         .fasync         = fuse_dev_fasync,
2000 };
2001 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2002
2003 static struct miscdevice fuse_miscdevice = {
2004         .minor = FUSE_MINOR,
2005         .name  = "fuse",
2006         .fops = &fuse_dev_operations,
2007 };
2008
2009 int __init fuse_dev_init(void)
2010 {
2011         int err = -ENOMEM;
2012         fuse_req_cachep = kmem_cache_create("fuse_request",
2013                                             sizeof(struct fuse_req),
2014                                             0, 0, NULL);
2015         if (!fuse_req_cachep)
2016                 goto out;
2017
2018         err = misc_register(&fuse_miscdevice);
2019         if (err)
2020                 goto out_cache_clean;
2021
2022         return 0;
2023
2024  out_cache_clean:
2025         kmem_cache_destroy(fuse_req_cachep);
2026  out:
2027         return err;
2028 }
2029
2030 void fuse_dev_cleanup(void)
2031 {
2032         misc_deregister(&fuse_miscdevice);
2033         kmem_cache_destroy(fuse_req_cachep);
2034 }