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