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 = from_kuid_munged(&init_user_ns, current_fsuid());
  96         req->in.h.gid = from_kgid_munged(&init_user_ns, 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                         req->stolen_file = get_file(file);
 152                 }
 153                 spin_unlock(&fc->lock);
 154         } while (!req);
 155
 156         return req;
 157 }
 158
 159 /*
 160  * Put stolen request back into fuse_file->reserved_req
 161  */
 162 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 163 {
 164         struct file *file = req->stolen_file;
 165         struct fuse_file *ff = file->private_data;
 166
 167         spin_lock(&fc->lock);
 168         fuse_request_init(req);
 169         BUG_ON(ff->reserved_req);
 170         ff->reserved_req = req;
 171         wake_up_all(&fc->reserved_req_waitq);
 172         spin_unlock(&fc->lock);
 173         fput(file);
 174 }
 175
 176 /*
 177  * Gets a requests for a file operation, always succeeds
 178  *
 179  * This is used for sending the FLUSH request, which must get to
 180  * userspace, due to POSIX locks which may need to be unlocked.
 181  *
 182  * If allocation fails due to OOM, use the reserved request in
 183  * fuse_file.
 184  *
 185  * This is very unlikely to deadlock accidentally, since the
 186  * filesystem should not have it's own file open.  If deadlock is
 187  * intentional, it can still be broken by "aborting" the filesystem.
 188  */
 189 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 190 {
 191         struct fuse_req *req;
 192
 193         atomic_inc(&fc->num_waiting);
 194         wait_event(fc->blocked_waitq, !fc->blocked);
 195         req = fuse_request_alloc();
 196         if (!req)
 197                 req = get_reserved_req(fc, file);
 198
 199         fuse_req_init_context(req);
 200         req->waiting = 1;
 201         return req;
 202 }
 203
 204 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 205 {
 206         if (atomic_dec_and_test(&req->count)) {
 207                 if (req->waiting)
 208                         atomic_dec(&fc->num_waiting);
 209
 210                 if (req->stolen_file)
 211                         put_reserved_req(fc, req);
 212                 else
 213                         fuse_request_free(req);
 214         }
 215 }
 216 EXPORT_SYMBOL_GPL(fuse_put_request);
 217
 218 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 219 {
 220         unsigned nbytes = 0;
 221         unsigned i;
 222
 223         for (i = 0; i < numargs; i++)
 224                 nbytes += args[i].size;
 225
 226         return nbytes;
 227 }
 228
 229 static u64 fuse_get_unique(struct fuse_conn *fc)
 230 {
 231         fc->reqctr++;
 232         /* zero is special */
 233         if (fc->reqctr == 0)
 234                 fc->reqctr = 1;
 235
 236         return fc->reqctr;
 237 }
 238
 239 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 240 {
 241         req->in.h.len = sizeof(struct fuse_in_header) +
 242                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 243         list_add_tail(&req->list, &fc->pending);
 244         req->state = FUSE_REQ_PENDING;
 245         if (!req->waiting) {
 246                 req->waiting = 1;
 247                 atomic_inc(&fc->num_waiting);
 248         }
 249         wake_up(&fc->waitq);
 250         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 251 }
 252
 253 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 254                        u64 nodeid, u64 nlookup)
 255 {
 256         forget->forget_one.nodeid = nodeid;
 257         forget->forget_one.nlookup = nlookup;
 258
 259         spin_lock(&fc->lock);
 260         if (fc->connected) {
 261                 fc->forget_list_tail->next = forget;
 262                 fc->forget_list_tail = forget;
 263                 wake_up(&fc->waitq);
 264                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 265         } else {
 266                 kfree(forget);
 267         }
 268         spin_unlock(&fc->lock);
 269 }
 270
 271 static void flush_bg_queue(struct fuse_conn *fc)
 272 {
 273         while (fc->active_background < fc->max_background &&
 274                !list_empty(&fc->bg_queue)) {
 275                 struct fuse_req *req;
 276
 277                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 278                 list_del(&req->list);
 279                 fc->active_background++;
 280                 req->in.h.unique = fuse_get_unique(fc);
 281                 queue_request(fc, req);
 282         }
 283 }
 284
 285 /*
 286  * This function is called when a request is finished.  Either a reply
 287  * has arrived or it was aborted (and not yet sent) or some error
 288  * occurred during communication with userspace, or the device file
 289  * was closed.  The requester thread is woken up (if still waiting),
 290  * the 'end' callback is called if given, else the reference to the
 291  * request is released
 292  *
 293  * Called with fc->lock, unlocks it
 294  */
 295 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 296 __releases(fc->lock)
 297 {
 298         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 299         req->end = NULL;
 300         list_del(&req->list);
 301         list_del(&req->intr_entry);
 302         req->state = FUSE_REQ_FINISHED;
 303         if (req->background) {
 304                 if (fc->num_background == fc->max_background) {
 305                         fc->blocked = 0;
 306                         wake_up_all(&fc->blocked_waitq);
 307                 }
 308                 if (fc->num_background == fc->congestion_threshold &&
 309                     fc->connected && fc->bdi_initialized) {
 310                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 311                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 312                 }
 313                 fc->num_background--;
 314                 fc->active_background--;
 315                 flush_bg_queue(fc);
 316         }
 317         spin_unlock(&fc->lock);
 318         wake_up(&req->waitq);
 319         if (end)
 320                 end(fc, req);
 321         fuse_put_request(fc, req);
 322 }
 323
 324 static void wait_answer_interruptible(struct fuse_conn *fc,
 325                                       struct fuse_req *req)
 326 __releases(fc->lock)
 327 __acquires(fc->lock)
 328 {
 329         if (signal_pending(current))
 330                 return;
 331
 332         spin_unlock(&fc->lock);
 333         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 334         spin_lock(&fc->lock);
 335 }
 336
 337 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 338 {
 339         list_add_tail(&req->intr_entry, &fc->interrupts);
 340         wake_up(&fc->waitq);
 341         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 342 }
 343
 344 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 345 __releases(fc->lock)
 346 __acquires(fc->lock)
 347 {
 348         if (!fc->no_interrupt) {
 349                 /* Any signal may interrupt this */
 350                 wait_answer_interruptible(fc, req);
 351
 352                 if (req->aborted)
 353                         goto aborted;
 354                 if (req->state == FUSE_REQ_FINISHED)
 355                         return;
 356
 357                 req->interrupted = 1;
 358                 if (req->state == FUSE_REQ_SENT)
 359                         queue_interrupt(fc, req);
 360         }
 361
 362         if (!req->force) {
 363                 sigset_t oldset;
 364
 365                 /* Only fatal signals may interrupt this */
 366                 block_sigs(&oldset);
 367                 wait_answer_interruptible(fc, req);
 368                 restore_sigs(&oldset);
 369
 370                 if (req->aborted)
 371                         goto aborted;
 372                 if (req->state == FUSE_REQ_FINISHED)
 373                         return;
 374
 375                 /* Request is not yet in userspace, bail out */
 376                 if (req->state == FUSE_REQ_PENDING) {
 377                         list_del(&req->list);
 378                         __fuse_put_request(req);
 379                         req->out.h.error = -EINTR;
 380                         return;
 381                 }
 382         }
 383
 384         /*
 385          * Either request is already in userspace, or it was forced.
 386          * Wait it out.
 387          */
 388         spin_unlock(&fc->lock);
 389         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 390         spin_lock(&fc->lock);
 391
 392         if (!req->aborted)
 393                 return;
 394
 395  aborted:
 396         BUG_ON(req->state != FUSE_REQ_FINISHED);
 397         if (req->locked) {
 398                 /* This is uninterruptible sleep, because data is
 399                    being copied to/from the buffers of req.  During
 400                    locked state, there mustn't be any filesystem
 401                    operation (e.g. page fault), since that could lead
 402                    to deadlock */
 403                 spin_unlock(&fc->lock);
 404                 wait_event(req->waitq, !req->locked);
 405                 spin_lock(&fc->lock);
 406         }
 407 }
 408
 409 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 410 {
 411         req->isreply = 1;
 412         spin_lock(&fc->lock);
 413         if (!fc->connected)
 414                 req->out.h.error = -ENOTCONN;
 415         else if (fc->conn_error)
 416                 req->out.h.error = -ECONNREFUSED;
 417         else {
 418                 req->in.h.unique = fuse_get_unique(fc);
 419                 queue_request(fc, req);
 420                 /* acquire extra reference, since request is still needed
 421                    after request_end() */
 422                 __fuse_get_request(req);
 423
 424                 request_wait_answer(fc, req);
 425         }
 426         spin_unlock(&fc->lock);
 427 }
 428 EXPORT_SYMBOL_GPL(fuse_request_send);
 429
 430 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 431                                             struct fuse_req *req)
 432 {
 433         req->background = 1;
 434         fc->num_background++;
 435         if (fc->num_background == fc->max_background)
 436                 fc->blocked = 1;
 437         if (fc->num_background == fc->congestion_threshold &&
 438             fc->bdi_initialized) {
 439                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 440                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 441         }
 442         list_add_tail(&req->list, &fc->bg_queue);
 443         flush_bg_queue(fc);
 444 }
 445
 446 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 447 {
 448         spin_lock(&fc->lock);
 449         if (fc->connected) {
 450                 fuse_request_send_nowait_locked(fc, req);
 451                 spin_unlock(&fc->lock);
 452         } else {
 453                 req->out.h.error = -ENOTCONN;
 454                 request_end(fc, req);
 455         }
 456 }
 457
 458 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 459 {
 460         req->isreply = 1;
 461         fuse_request_send_nowait(fc, req);
 462 }
 463 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 464
 465 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 466                                           struct fuse_req *req, u64 unique)
 467 {
 468         int err = -ENODEV;
 469
 470         req->isreply = 0;
 471         req->in.h.unique = unique;
 472         spin_lock(&fc->lock);
 473         if (fc->connected) {
 474                 queue_request(fc, req);
 475                 err = 0;
 476         }
 477         spin_unlock(&fc->lock);
 478
 479         return err;
 480 }
 481
 482 /*
 483  * Called under fc->lock
 484  *
 485  * fc->connected must have been checked previously
 486  */
 487 void fuse_request_send_background_locked(struct fuse_conn *fc,
 488                                          struct fuse_req *req)
 489 {
 490         req->isreply = 1;
 491         fuse_request_send_nowait_locked(fc, req);
 492 }
 493
 494 /*
 495  * Lock the request.  Up to the next unlock_request() there mustn't be
 496  * anything that could cause a page-fault.  If the request was already
 497  * aborted bail out.
 498  */
 499 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 500 {
 501         int err = 0;
 502         if (req) {
 503                 spin_lock(&fc->lock);
 504                 if (req->aborted)
 505                         err = -ENOENT;
 506                 else
 507                         req->locked = 1;
 508                 spin_unlock(&fc->lock);
 509         }
 510         return err;
 511 }
 512
 513 /*
 514  * Unlock request.  If it was aborted during being locked, the
 515  * requester thread is currently waiting for it to be unlocked, so
 516  * wake it up.
 517  */
 518 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 519 {
 520         if (req) {
 521                 spin_lock(&fc->lock);
 522                 req->locked = 0;
 523                 if (req->aborted)
 524                         wake_up(&req->waitq);
 525                 spin_unlock(&fc->lock);
 526         }
 527 }
 528
 529 struct fuse_copy_state {
 530         struct fuse_conn *fc;
 531         int write;
 532         struct fuse_req *req;
 533         const struct iovec *iov;
 534         struct pipe_buffer *pipebufs;
 535         struct pipe_buffer *currbuf;
 536         struct pipe_inode_info *pipe;
 537         unsigned long nr_segs;
 538         unsigned long seglen;
 539         unsigned long addr;
 540         struct page *pg;
 541         void *mapaddr;
 542         void *buf;
 543         unsigned len;
 544         unsigned move_pages:1;
 545 };
 546
 547 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 548                            int write,
 549                            const struct iovec *iov, unsigned long nr_segs)
 550 {
 551         memset(cs, 0, sizeof(*cs));
 552         cs->fc = fc;
 553         cs->write = write;
 554         cs->iov = iov;
 555         cs->nr_segs = nr_segs;
 556 }
 557
 558 /* Unmap and put previous page of userspace buffer */
 559 static void fuse_copy_finish(struct fuse_copy_state *cs)
 560 {
 561         if (cs->currbuf) {
 562                 struct pipe_buffer *buf = cs->currbuf;
 563
 564                 if (!cs->write) {
 565                         buf->ops->unmap(cs->pipe, buf, cs->mapaddr);
 566                 } else {
 567                         kunmap(buf->page);
 568                         buf->len = PAGE_SIZE - cs->len;
 569                 }
 570                 cs->currbuf = NULL;
 571                 cs->mapaddr = NULL;
 572         } else if (cs->mapaddr) {
 573                 kunmap(cs->pg);
 574                 if (cs->write) {
 575                         flush_dcache_page(cs->pg);
 576                         set_page_dirty_lock(cs->pg);
 577                 }
 578                 put_page(cs->pg);
 579                 cs->mapaddr = NULL;
 580         }
 581 }
 582
 583 /*
 584  * Get another pagefull of userspace buffer, and map it to kernel
 585  * address space, and lock request
 586  */
 587 static int fuse_copy_fill(struct fuse_copy_state *cs)
 588 {
 589         unsigned long offset;
 590         int err;
 591
 592         unlock_request(cs->fc, cs->req);
 593         fuse_copy_finish(cs);
 594         if (cs->pipebufs) {
 595                 struct pipe_buffer *buf = cs->pipebufs;
 596
 597                 if (!cs->write) {
 598                         err = buf->ops->confirm(cs->pipe, buf);
 599                         if (err)
 600                                 return err;
 601
 602                         BUG_ON(!cs->nr_segs);
 603                         cs->currbuf = buf;
 604                         cs->mapaddr = buf->ops->map(cs->pipe, buf, 0);
 605                         cs->len = buf->len;
 606                         cs->buf = cs->mapaddr + buf->offset;
 607                         cs->pipebufs++;
 608                         cs->nr_segs--;
 609                 } else {
 610                         struct page *page;
 611
 612                         if (cs->nr_segs == cs->pipe->buffers)
 613                                 return -EIO;
 614
 615                         page = alloc_page(GFP_HIGHUSER);
 616                         if (!page)
 617                                 return -ENOMEM;
 618
 619                         buf->page = page;
 620                         buf->offset = 0;
 621                         buf->len = 0;
 622
 623                         cs->currbuf = buf;
 624                         cs->mapaddr = kmap(page);
 625                         cs->buf = cs->mapaddr;
 626                         cs->len = PAGE_SIZE;
 627                         cs->pipebufs++;
 628                         cs->nr_segs++;
 629                 }
 630         } else {
 631                 if (!cs->seglen) {
 632                         BUG_ON(!cs->nr_segs);
 633                         cs->seglen = cs->iov[0].iov_len;
 634                         cs->addr = (unsigned long) cs->iov[0].iov_base;
 635                         cs->iov++;
 636                         cs->nr_segs--;
 637                 }
 638                 err = get_user_pages_fast(cs->addr, 1, cs->write, &cs->pg);
 639                 if (err < 0)
 640                         return err;
 641                 BUG_ON(err != 1);
 642                 offset = cs->addr % PAGE_SIZE;
 643                 cs->mapaddr = kmap(cs->pg);
 644                 cs->buf = cs->mapaddr + offset;
 645                 cs->len = min(PAGE_SIZE - offset, cs->seglen);
 646                 cs->seglen -= cs->len;
 647                 cs->addr += cs->len;
 648         }
 649
 650         return lock_request(cs->fc, cs->req);
 651 }
 652
 653 /* Do as much copy to/from userspace buffer as we can */
 654 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 655 {
 656         unsigned ncpy = min(*size, cs->len);
 657         if (val) {
 658                 if (cs->write)
 659                         memcpy(cs->buf, *val, ncpy);
 660                 else
 661                         memcpy(*val, cs->buf, ncpy);
 662                 *val += ncpy;
 663         }
 664         *size -= ncpy;
 665         cs->len -= ncpy;
 666         cs->buf += ncpy;
 667         return ncpy;
 668 }
 669
 670 static int fuse_check_page(struct page *page)
 671 {
 672         if (page_mapcount(page) ||
 673             page->mapping != NULL ||
 674             page_count(page) != 1 ||
 675             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 676              ~(1 << PG_locked |
 677                1 << PG_referenced |
 678                1 << PG_uptodate |
 679                1 << PG_lru |
 680                1 << PG_active |
 681                1 << PG_reclaim))) {
 682                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 683                 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);
 684                 return 1;
 685         }
 686         return 0;
 687 }
 688
 689 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 690 {
 691         int err;
 692         struct page *oldpage = *pagep;
 693         struct page *newpage;
 694         struct pipe_buffer *buf = cs->pipebufs;
 695
 696         unlock_request(cs->fc, cs->req);
 697         fuse_copy_finish(cs);
 698
 699         err = buf->ops->confirm(cs->pipe, buf);
 700         if (err)
 701                 return err;
 702
 703         BUG_ON(!cs->nr_segs);
 704         cs->currbuf = buf;
 705         cs->len = buf->len;
 706         cs->pipebufs++;
 707         cs->nr_segs--;
 708
 709         if (cs->len != PAGE_SIZE)
 710                 goto out_fallback;
 711
 712         if (buf->ops->steal(cs->pipe, buf) != 0)
 713                 goto out_fallback;
 714
 715         newpage = buf->page;
 716
 717         if (WARN_ON(!PageUptodate(newpage)))
 718                 return -EIO;
 719
 720         ClearPageMappedToDisk(newpage);
 721
 722         if (fuse_check_page(newpage) != 0)
 723                 goto out_fallback_unlock;
 724
 725         /*
 726          * This is a new and locked page, it shouldn't be mapped or
 727          * have any special flags on it
 728          */
 729         if (WARN_ON(page_mapped(oldpage)))
 730                 goto out_fallback_unlock;
 731         if (WARN_ON(page_has_private(oldpage)))
 732                 goto out_fallback_unlock;
 733         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 734                 goto out_fallback_unlock;
 735         if (WARN_ON(PageMlocked(oldpage)))
 736                 goto out_fallback_unlock;
 737
 738         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 739         if (err) {
 740                 unlock_page(newpage);
 741                 return err;
 742         }
 743
 744         page_cache_get(newpage);
 745
 746         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 747                 lru_cache_add_file(newpage);
 748
 749         err = 0;
 750         spin_lock(&cs->fc->lock);
 751         if (cs->req->aborted)
 752                 err = -ENOENT;
 753         else
 754                 *pagep = newpage;
 755         spin_unlock(&cs->fc->lock);
 756
 757         if (err) {
 758                 unlock_page(newpage);
 759                 page_cache_release(newpage);
 760                 return err;
 761         }
 762
 763         unlock_page(oldpage);
 764         page_cache_release(oldpage);
 765         cs->len = 0;
 766
 767         return 0;
 768
 769 out_fallback_unlock:
 770         unlock_page(newpage);
 771 out_fallback:
 772         cs->mapaddr = buf->ops->map(cs->pipe, buf, 1);
 773         cs->buf = cs->mapaddr + buf->offset;
 774
 775         err = lock_request(cs->fc, cs->req);
 776         if (err)
 777                 return err;
 778
 779         return 1;
 780 }
 781
 782 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 783                          unsigned offset, unsigned count)
 784 {
 785         struct pipe_buffer *buf;
 786
 787         if (cs->nr_segs == cs->pipe->buffers)
 788                 return -EIO;
 789
 790         unlock_request(cs->fc, cs->req);
 791         fuse_copy_finish(cs);
 792
 793         buf = cs->pipebufs;
 794         page_cache_get(page);
 795         buf->page = page;
 796         buf->offset = offset;
 797         buf->len = count;
 798
 799         cs->pipebufs++;
 800         cs->nr_segs++;
 801         cs->len = 0;
 802
 803         return 0;
 804 }
 805
 806 /*
 807  * Copy a page in the request to/from the userspace buffer.  Must be
 808  * done atomically
 809  */
 810 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 811                           unsigned offset, unsigned count, int zeroing)
 812 {
 813         int err;
 814         struct page *page = *pagep;
 815
 816         if (page && zeroing && count < PAGE_SIZE)
 817                 clear_highpage(page);
 818
 819         while (count) {
 820                 if (cs->write && cs->pipebufs && page) {
 821                         return fuse_ref_page(cs, page, offset, count);
 822                 } else if (!cs->len) {
 823                         if (cs->move_pages && page &&
 824                             offset == 0 && count == PAGE_SIZE) {
 825                                 err = fuse_try_move_page(cs, pagep);
 826                                 if (err <= 0)
 827                                         return err;
 828                         } else {
 829                                 err = fuse_copy_fill(cs);
 830                                 if (err)
 831                                         return err;
 832                         }
 833                 }
 834                 if (page) {
 835                         void *mapaddr = kmap_atomic(page);
 836                         void *buf = mapaddr + offset;
 837                         offset += fuse_copy_do(cs, &buf, &count);
 838                         kunmap_atomic(mapaddr);
 839                 } else
 840                         offset += fuse_copy_do(cs, NULL, &count);
 841         }
 842         if (page && !cs->write)
 843                 flush_dcache_page(page);
 844         return 0;
 845 }
 846
 847 /* Copy pages in the request to/from userspace buffer */
 848 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 849                            int zeroing)
 850 {
 851         unsigned i;
 852         struct fuse_req *req = cs->req;
 853         unsigned offset = req->page_offset;
 854         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 855
 856         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 857                 int err;
 858
 859                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
 860                                      zeroing);
 861                 if (err)
 862                         return err;
 863
 864                 nbytes -= count;
 865                 count = min(nbytes, (unsigned) PAGE_SIZE);
 866                 offset = 0;
 867         }
 868         return 0;
 869 }
 870
 871 /* Copy a single argument in the request to/from userspace buffer */
 872 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 873 {
 874         while (size) {
 875                 if (!cs->len) {
 876                         int err = fuse_copy_fill(cs);
 877                         if (err)
 878                                 return err;
 879                 }
 880                 fuse_copy_do(cs, &val, &size);
 881         }
 882         return 0;
 883 }
 884
 885 /* Copy request arguments to/from userspace buffer */
 886 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 887                           unsigned argpages, struct fuse_arg *args,
 888                           int zeroing)
 889 {
 890         int err = 0;
 891         unsigned i;
 892
 893         for (i = 0; !err && i < numargs; i++)  {
 894                 struct fuse_arg *arg = &args[i];
 895                 if (i == numargs - 1 && argpages)
 896                         err = fuse_copy_pages(cs, arg->size, zeroing);
 897                 else
 898                         err = fuse_copy_one(cs, arg->value, arg->size);
 899         }
 900         return err;
 901 }
 902
 903 static int forget_pending(struct fuse_conn *fc)
 904 {
 905         return fc->forget_list_head.next != NULL;
 906 }
 907
 908 static int request_pending(struct fuse_conn *fc)
 909 {
 910         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts) ||
 911                 forget_pending(fc);
 912 }
 913
 914 /* Wait until a request is available on the pending list */
 915 static void request_wait(struct fuse_conn *fc)
 916 __releases(fc->lock)
 917 __acquires(fc->lock)
 918 {
 919         DECLARE_WAITQUEUE(wait, current);
 920
 921         add_wait_queue_exclusive(&fc->waitq, &wait);
 922         while (fc->connected && !request_pending(fc)) {
 923                 set_current_state(TASK_INTERRUPTIBLE);
 924                 if (signal_pending(current))
 925                         break;
 926
 927                 spin_unlock(&fc->lock);
 928                 schedule();
 929                 spin_lock(&fc->lock);
 930         }
 931         set_current_state(TASK_RUNNING);
 932         remove_wait_queue(&fc->waitq, &wait);
 933 }
 934
 935 /*
 936  * Transfer an interrupt request to userspace
 937  *
 938  * Unlike other requests this is assembled on demand, without a need
 939  * to allocate a separate fuse_req structure.
 940  *
 941  * Called with fc->lock held, releases it
 942  */
 943 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
 944                                size_t nbytes, struct fuse_req *req)
 945 __releases(fc->lock)
 946 {
 947         struct fuse_in_header ih;
 948         struct fuse_interrupt_in arg;
 949         unsigned reqsize = sizeof(ih) + sizeof(arg);
 950         int err;
 951
 952         list_del_init(&req->intr_entry);
 953         req->intr_unique = fuse_get_unique(fc);
 954         memset(&ih, 0, sizeof(ih));
 955         memset(&arg, 0, sizeof(arg));
 956         ih.len = reqsize;
 957         ih.opcode = FUSE_INTERRUPT;
 958         ih.unique = req->intr_unique;
 959         arg.unique = req->in.h.unique;
 960
 961         spin_unlock(&fc->lock);
 962         if (nbytes < reqsize)
 963                 return -EINVAL;
 964
 965         err = fuse_copy_one(cs, &ih, sizeof(ih));
 966         if (!err)
 967                 err = fuse_copy_one(cs, &arg, sizeof(arg));
 968         fuse_copy_finish(cs);
 969
 970         return err ? err : reqsize;
 971 }
 972
 973 static struct fuse_forget_link *dequeue_forget(struct fuse_conn *fc,
 974                                                unsigned max,
 975                                                unsigned *countp)
 976 {
 977         struct fuse_forget_link *head = fc->forget_list_head.next;
 978         struct fuse_forget_link **newhead = &head;
 979         unsigned count;
 980
 981         for (count = 0; *newhead != NULL && count < max; count++)
 982                 newhead = &(*newhead)->next;
 983
 984         fc->forget_list_head.next = *newhead;
 985         *newhead = NULL;
 986         if (fc->forget_list_head.next == NULL)
 987                 fc->forget_list_tail = &fc->forget_list_head;
 988
 989         if (countp != NULL)
 990                 *countp = count;
 991
 992         return head;
 993 }
 994
 995 static int fuse_read_single_forget(struct fuse_conn *fc,
 996                                    struct fuse_copy_state *cs,
 997                                    size_t nbytes)
 998 __releases(fc->lock)
 999 {
1000         int err;
1001         struct fuse_forget_link *forget = dequeue_forget(fc, 1, NULL);
1002         struct fuse_forget_in arg = {
1003                 .nlookup = forget->forget_one.nlookup,
1004         };
1005         struct fuse_in_header ih = {
1006                 .opcode = FUSE_FORGET,
1007                 .nodeid = forget->forget_one.nodeid,
1008                 .unique = fuse_get_unique(fc),
1009                 .len = sizeof(ih) + sizeof(arg),
1010         };
1011
1012         spin_unlock(&fc->lock);
1013         kfree(forget);
1014         if (nbytes < ih.len)
1015                 return -EINVAL;
1016
1017         err = fuse_copy_one(cs, &ih, sizeof(ih));
1018         if (!err)
1019                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1020         fuse_copy_finish(cs);
1021
1022         if (err)
1023                 return err;
1024
1025         return ih.len;
1026 }
1027
1028 static int fuse_read_batch_forget(struct fuse_conn *fc,
1029                                    struct fuse_copy_state *cs, size_t nbytes)
1030 __releases(fc->lock)
1031 {
1032         int err;
1033         unsigned max_forgets;
1034         unsigned count;
1035         struct fuse_forget_link *head;
1036         struct fuse_batch_forget_in arg = { .count = 0 };
1037         struct fuse_in_header ih = {
1038                 .opcode = FUSE_BATCH_FORGET,
1039                 .unique = fuse_get_unique(fc),
1040                 .len = sizeof(ih) + sizeof(arg),
1041         };
1042
1043         if (nbytes < ih.len) {
1044                 spin_unlock(&fc->lock);
1045                 return -EINVAL;
1046         }
1047
1048         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1049         head = dequeue_forget(fc, max_forgets, &count);
1050         spin_unlock(&fc->lock);
1051
1052         arg.count = count;
1053         ih.len += count * sizeof(struct fuse_forget_one);
1054         err = fuse_copy_one(cs, &ih, sizeof(ih));
1055         if (!err)
1056                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1057
1058         while (head) {
1059                 struct fuse_forget_link *forget = head;
1060
1061                 if (!err) {
1062                         err = fuse_copy_one(cs, &forget->forget_one,
1063                                             sizeof(forget->forget_one));
1064                 }
1065                 head = forget->next;
1066                 kfree(forget);
1067         }
1068
1069         fuse_copy_finish(cs);
1070
1071         if (err)
1072                 return err;
1073
1074         return ih.len;
1075 }
1076
1077 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_copy_state *cs,
1078                             size_t nbytes)
1079 __releases(fc->lock)
1080 {
1081         if (fc->minor < 16 || fc->forget_list_head.next->next == NULL)
1082                 return fuse_read_single_forget(fc, cs, nbytes);
1083         else
1084                 return fuse_read_batch_forget(fc, cs, nbytes);
1085 }
1086
1087 /*
1088  * Read a single request into the userspace filesystem's buffer.  This
1089  * function waits until a request is available, then removes it from
1090  * the pending list and copies request data to userspace buffer.  If
1091  * no reply is needed (FORGET) or request has been aborted or there
1092  * was an error during the copying then it's finished by calling
1093  * request_end().  Otherwise add it to the processing list, and set
1094  * the 'sent' flag.
1095  */
1096 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1097                                 struct fuse_copy_state *cs, size_t nbytes)
1098 {
1099         int err;
1100         struct fuse_req *req;
1101         struct fuse_in *in;
1102         unsigned reqsize;
1103
1104  restart:
1105         spin_lock(&fc->lock);
1106         err = -EAGAIN;
1107         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
1108             !request_pending(fc))
1109                 goto err_unlock;
1110
1111         request_wait(fc);
1112         err = -ENODEV;
1113         if (!fc->connected)
1114                 goto err_unlock;
1115         err = -ERESTARTSYS;
1116         if (!request_pending(fc))
1117                 goto err_unlock;
1118
1119         if (!list_empty(&fc->interrupts)) {
1120                 req = list_entry(fc->interrupts.next, struct fuse_req,
1121                                  intr_entry);
1122                 return fuse_read_interrupt(fc, cs, nbytes, req);
1123         }
1124
1125         if (forget_pending(fc)) {
1126                 if (list_empty(&fc->pending) || fc->forget_batch-- > 0)
1127                         return fuse_read_forget(fc, cs, nbytes);
1128
1129                 if (fc->forget_batch <= -8)
1130                         fc->forget_batch = 16;
1131         }
1132
1133         req = list_entry(fc->pending.next, struct fuse_req, list);
1134         req->state = FUSE_REQ_READING;
1135         list_move(&req->list, &fc->io);
1136
1137         in = &req->in;
1138         reqsize = in->h.len;
1139         /* If request is too large, reply with an error and restart the read */
1140         if (nbytes < reqsize) {
1141                 req->out.h.error = -EIO;
1142                 /* SETXATTR is special, since it may contain too large data */
1143                 if (in->h.opcode == FUSE_SETXATTR)
1144                         req->out.h.error = -E2BIG;
1145                 request_end(fc, req);
1146                 goto restart;
1147         }
1148         spin_unlock(&fc->lock);
1149         cs->req = req;
1150         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1151         if (!err)
1152                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1153                                      (struct fuse_arg *) in->args, 0);
1154         fuse_copy_finish(cs);
1155         spin_lock(&fc->lock);
1156         req->locked = 0;
1157         if (req->aborted) {
1158                 request_end(fc, req);
1159                 return -ENODEV;
1160         }
1161         if (err) {
1162                 req->out.h.error = -EIO;
1163                 request_end(fc, req);
1164                 return err;
1165         }
1166         if (!req->isreply)
1167                 request_end(fc, req);
1168         else {
1169                 req->state = FUSE_REQ_SENT;
1170                 list_move_tail(&req->list, &fc->processing);
1171                 if (req->interrupted)
1172                         queue_interrupt(fc, req);
1173                 spin_unlock(&fc->lock);
1174         }
1175         return reqsize;
1176
1177  err_unlock:
1178         spin_unlock(&fc->lock);
1179         return err;
1180 }
1181
1182 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
1183                               unsigned long nr_segs, loff_t pos)
1184 {
1185         struct fuse_copy_state cs;
1186         struct file *file = iocb->ki_filp;
1187         struct fuse_conn *fc = fuse_get_conn(file);
1188         if (!fc)
1189                 return -EPERM;
1190
1191         fuse_copy_init(&cs, fc, 1, iov, nr_segs);
1192
1193         return fuse_dev_do_read(fc, file, &cs, iov_length(iov, nr_segs));
1194 }
1195
1196 static int fuse_dev_pipe_buf_steal(struct pipe_inode_info *pipe,
1197                                    struct pipe_buffer *buf)
1198 {
1199         return 1;
1200 }
1201
1202 static const struct pipe_buf_operations fuse_dev_pipe_buf_ops = {
1203         .can_merge = 0,
1204         .map = generic_pipe_buf_map,
1205         .unmap = generic_pipe_buf_unmap,
1206         .confirm = generic_pipe_buf_confirm,
1207         .release = generic_pipe_buf_release,
1208         .steal = fuse_dev_pipe_buf_steal,
1209         .get = generic_pipe_buf_get,
1210 };
1211
1212 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1213                                     struct pipe_inode_info *pipe,
1214                                     size_t len, unsigned int flags)
1215 {
1216         int ret;
1217         int page_nr = 0;
1218         int do_wakeup = 0;
1219         struct pipe_buffer *bufs;
1220         struct fuse_copy_state cs;
1221         struct fuse_conn *fc = fuse_get_conn(in);
1222         if (!fc)
1223                 return -EPERM;
1224
1225         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1226         if (!bufs)
1227                 return -ENOMEM;
1228
1229         fuse_copy_init(&cs, fc, 1, NULL, 0);
1230         cs.pipebufs = bufs;
1231         cs.pipe = pipe;
1232         ret = fuse_dev_do_read(fc, in, &cs, len);
1233         if (ret < 0)
1234                 goto out;
1235
1236         ret = 0;
1237         pipe_lock(pipe);
1238
1239         if (!pipe->readers) {
1240                 send_sig(SIGPIPE, current, 0);
1241                 if (!ret)
1242                         ret = -EPIPE;
1243                 goto out_unlock;
1244         }
1245
1246         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1247                 ret = -EIO;
1248                 goto out_unlock;
1249         }
1250
1251         while (page_nr < cs.nr_segs) {
1252                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1253                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1254
1255                 buf->page = bufs[page_nr].page;
1256                 buf->offset = bufs[page_nr].offset;
1257                 buf->len = bufs[page_nr].len;
1258                 buf->ops = &fuse_dev_pipe_buf_ops;
1259
1260                 pipe->nrbufs++;
1261                 page_nr++;
1262                 ret += buf->len;
1263
1264                 if (pipe->inode)
1265                         do_wakeup = 1;
1266         }
1267
1268 out_unlock:
1269         pipe_unlock(pipe);
1270
1271         if (do_wakeup) {
1272                 smp_mb();
1273                 if (waitqueue_active(&pipe->wait))
1274                         wake_up_interruptible(&pipe->wait);
1275                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1276         }
1277
1278 out:
1279         for (; page_nr < cs.nr_segs; page_nr++)
1280                 page_cache_release(bufs[page_nr].page);
1281
1282         kfree(bufs);
1283         return ret;
1284 }
1285
1286 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1287                             struct fuse_copy_state *cs)
1288 {
1289         struct fuse_notify_poll_wakeup_out outarg;
1290         int err = -EINVAL;
1291
1292         if (size != sizeof(outarg))
1293                 goto err;
1294
1295         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1296         if (err)
1297                 goto err;
1298
1299         fuse_copy_finish(cs);
1300         return fuse_notify_poll_wakeup(fc, &outarg);
1301
1302 err:
1303         fuse_copy_finish(cs);
1304         return err;
1305 }
1306
1307 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1308                                    struct fuse_copy_state *cs)
1309 {
1310         struct fuse_notify_inval_inode_out outarg;
1311         int err = -EINVAL;
1312
1313         if (size != sizeof(outarg))
1314                 goto err;
1315
1316         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1317         if (err)
1318                 goto err;
1319         fuse_copy_finish(cs);
1320
1321         down_read(&fc->killsb);
1322         err = -ENOENT;
1323         if (fc->sb) {
1324                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1325                                                outarg.off, outarg.len);
1326         }
1327         up_read(&fc->killsb);
1328         return err;
1329
1330 err:
1331         fuse_copy_finish(cs);
1332         return err;
1333 }
1334
1335 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1336                                    struct fuse_copy_state *cs)
1337 {
1338         struct fuse_notify_inval_entry_out outarg;
1339         int err = -ENOMEM;
1340         char *buf;
1341         struct qstr name;
1342
1343         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1344         if (!buf)
1345                 goto err;
1346
1347         err = -EINVAL;
1348         if (size < sizeof(outarg))
1349                 goto err;
1350
1351         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1352         if (err)
1353                 goto err;
1354
1355         err = -ENAMETOOLONG;
1356         if (outarg.namelen > FUSE_NAME_MAX)
1357                 goto err;
1358
1359         err = -EINVAL;
1360         if (size != sizeof(outarg) + outarg.namelen + 1)
1361                 goto err;
1362
1363         name.name = buf;
1364         name.len = outarg.namelen;
1365         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1366         if (err)
1367                 goto err;
1368         fuse_copy_finish(cs);
1369         buf[outarg.namelen] = 0;
1370         name.hash = full_name_hash(name.name, name.len);
1371
1372         down_read(&fc->killsb);
1373         err = -ENOENT;
1374         if (fc->sb)
1375                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1376         up_read(&fc->killsb);
1377         kfree(buf);
1378         return err;
1379
1380 err:
1381         kfree(buf);
1382         fuse_copy_finish(cs);
1383         return err;
1384 }
1385
1386 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1387                               struct fuse_copy_state *cs)
1388 {
1389         struct fuse_notify_delete_out outarg;
1390         int err = -ENOMEM;
1391         char *buf;
1392         struct qstr name;
1393
1394         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1395         if (!buf)
1396                 goto err;
1397
1398         err = -EINVAL;
1399         if (size < sizeof(outarg))
1400                 goto err;
1401
1402         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1403         if (err)
1404                 goto err;
1405
1406         err = -ENAMETOOLONG;
1407         if (outarg.namelen > FUSE_NAME_MAX)
1408                 goto err;
1409
1410         err = -EINVAL;
1411         if (size != sizeof(outarg) + outarg.namelen + 1)
1412                 goto err;
1413
1414         name.name = buf;
1415         name.len = outarg.namelen;
1416         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1417         if (err)
1418                 goto err;
1419         fuse_copy_finish(cs);
1420         buf[outarg.namelen] = 0;
1421         name.hash = full_name_hash(name.name, name.len);
1422
1423         down_read(&fc->killsb);
1424         err = -ENOENT;
1425         if (fc->sb)
1426                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1427                                                outarg.child, &name);
1428         up_read(&fc->killsb);
1429         kfree(buf);
1430         return err;
1431
1432 err:
1433         kfree(buf);
1434         fuse_copy_finish(cs);
1435         return err;
1436 }
1437
1438 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1439                              struct fuse_copy_state *cs)
1440 {
1441         struct fuse_notify_store_out outarg;
1442         struct inode *inode;
1443         struct address_space *mapping;
1444         u64 nodeid;
1445         int err;
1446         pgoff_t index;
1447         unsigned int offset;
1448         unsigned int num;
1449         loff_t file_size;
1450         loff_t end;
1451
1452         err = -EINVAL;
1453         if (size < sizeof(outarg))
1454                 goto out_finish;
1455
1456         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1457         if (err)
1458                 goto out_finish;
1459
1460         err = -EINVAL;
1461         if (size - sizeof(outarg) != outarg.size)
1462                 goto out_finish;
1463
1464         nodeid = outarg.nodeid;
1465
1466         down_read(&fc->killsb);
1467
1468         err = -ENOENT;
1469         if (!fc->sb)
1470                 goto out_up_killsb;
1471
1472         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1473         if (!inode)
1474                 goto out_up_killsb;
1475
1476         mapping = inode->i_mapping;
1477         index = outarg.offset >> PAGE_CACHE_SHIFT;
1478         offset = outarg.offset & ~PAGE_CACHE_MASK;
1479         file_size = i_size_read(inode);
1480         end = outarg.offset + outarg.size;
1481         if (end > file_size) {
1482                 file_size = end;
1483                 fuse_write_update_size(inode, file_size);
1484         }
1485
1486         num = outarg.size;
1487         while (num) {
1488                 struct page *page;
1489                 unsigned int this_num;
1490
1491                 err = -ENOMEM;
1492                 page = find_or_create_page(mapping, index,
1493                                            mapping_gfp_mask(mapping));
1494                 if (!page)
1495                         goto out_iput;
1496
1497                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1498                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1499                 if (!err && offset == 0 && (num != 0 || file_size == end))
1500                         SetPageUptodate(page);
1501                 unlock_page(page);
1502                 page_cache_release(page);
1503
1504                 if (err)
1505                         goto out_iput;
1506
1507                 num -= this_num;
1508                 offset = 0;
1509                 index++;
1510         }
1511
1512         err = 0;
1513
1514 out_iput:
1515         iput(inode);
1516 out_up_killsb:
1517         up_read(&fc->killsb);
1518 out_finish:
1519         fuse_copy_finish(cs);
1520         return err;
1521 }
1522
1523 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1524 {
1525         release_pages(req->pages, req->num_pages, 0);
1526 }
1527
1528 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1529                          struct fuse_notify_retrieve_out *outarg)
1530 {
1531         int err;
1532         struct address_space *mapping = inode->i_mapping;
1533         struct fuse_req *req;
1534         pgoff_t index;
1535         loff_t file_size;
1536         unsigned int num;
1537         unsigned int offset;
1538         size_t total_len = 0;
1539
1540         req = fuse_get_req(fc);
1541         if (IS_ERR(req))
1542                 return PTR_ERR(req);
1543
1544         offset = outarg->offset & ~PAGE_CACHE_MASK;
1545
1546         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1547         req->in.h.nodeid = outarg->nodeid;
1548         req->in.numargs = 2;
1549         req->in.argpages = 1;
1550         req->page_offset = offset;
1551         req->end = fuse_retrieve_end;
1552
1553         index = outarg->offset >> PAGE_CACHE_SHIFT;
1554         file_size = i_size_read(inode);
1555         num = outarg->size;
1556         if (outarg->offset > file_size)
1557                 num = 0;
1558         else if (outarg->offset + num > file_size)
1559                 num = file_size - outarg->offset;
1560
1561         while (num && req->num_pages < FUSE_MAX_PAGES_PER_REQ) {
1562                 struct page *page;
1563                 unsigned int this_num;
1564
1565                 page = find_get_page(mapping, index);
1566                 if (!page)
1567                         break;
1568
1569                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1570                 req->pages[req->num_pages] = page;
1571                 req->num_pages++;
1572
1573                 offset = 0;
1574                 num -= this_num;
1575                 total_len += this_num;
1576                 index++;
1577         }
1578         req->misc.retrieve_in.offset = outarg->offset;
1579         req->misc.retrieve_in.size = total_len;
1580         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1581         req->in.args[0].value = &req->misc.retrieve_in;
1582         req->in.args[1].size = total_len;
1583
1584         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1585         if (err)
1586                 fuse_retrieve_end(fc, req);
1587
1588         return err;
1589 }
1590
1591 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1592                                 struct fuse_copy_state *cs)
1593 {
1594         struct fuse_notify_retrieve_out outarg;
1595         struct inode *inode;
1596         int err;
1597
1598         err = -EINVAL;
1599         if (size != sizeof(outarg))
1600                 goto copy_finish;
1601
1602         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1603         if (err)
1604                 goto copy_finish;
1605
1606         fuse_copy_finish(cs);
1607
1608         down_read(&fc->killsb);
1609         err = -ENOENT;
1610         if (fc->sb) {
1611                 u64 nodeid = outarg.nodeid;
1612
1613                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1614                 if (inode) {
1615                         err = fuse_retrieve(fc, inode, &outarg);
1616                         iput(inode);
1617                 }
1618         }
1619         up_read(&fc->killsb);
1620
1621         return err;
1622
1623 copy_finish:
1624         fuse_copy_finish(cs);
1625         return err;
1626 }
1627
1628 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1629                        unsigned int size, struct fuse_copy_state *cs)
1630 {
1631         switch (code) {
1632         case FUSE_NOTIFY_POLL:
1633                 return fuse_notify_poll(fc, size, cs);
1634
1635         case FUSE_NOTIFY_INVAL_INODE:
1636                 return fuse_notify_inval_inode(fc, size, cs);
1637
1638         case FUSE_NOTIFY_INVAL_ENTRY:
1639                 return fuse_notify_inval_entry(fc, size, cs);
1640
1641         case FUSE_NOTIFY_STORE:
1642                 return fuse_notify_store(fc, size, cs);
1643
1644         case FUSE_NOTIFY_RETRIEVE:
1645                 return fuse_notify_retrieve(fc, size, cs);
1646
1647         case FUSE_NOTIFY_DELETE:
1648                 return fuse_notify_delete(fc, size, cs);
1649
1650         default:
1651                 fuse_copy_finish(cs);
1652                 return -EINVAL;
1653         }
1654 }
1655
1656 /* Look up request on processing list by unique ID */
1657 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1658 {
1659         struct list_head *entry;
1660
1661         list_for_each(entry, &fc->processing) {
1662                 struct fuse_req *req;
1663                 req = list_entry(entry, struct fuse_req, list);
1664                 if (req->in.h.unique == unique || req->intr_unique == unique)
1665                         return req;
1666         }
1667         return NULL;
1668 }
1669
1670 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1671                          unsigned nbytes)
1672 {
1673         unsigned reqsize = sizeof(struct fuse_out_header);
1674
1675         if (out->h.error)
1676                 return nbytes != reqsize ? -EINVAL : 0;
1677
1678         reqsize += len_args(out->numargs, out->args);
1679
1680         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1681                 return -EINVAL;
1682         else if (reqsize > nbytes) {
1683                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1684                 unsigned diffsize = reqsize - nbytes;
1685                 if (diffsize > lastarg->size)
1686                         return -EINVAL;
1687                 lastarg->size -= diffsize;
1688         }
1689         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1690                               out->page_zeroing);
1691 }
1692
1693 /*
1694  * Write a single reply to a request.  First the header is copied from
1695  * the write buffer.  The request is then searched on the processing
1696  * list by the unique ID found in the header.  If found, then remove
1697  * it from the list and copy the rest of the buffer to the request.
1698  * The request is finished by calling request_end()
1699  */
1700 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1701                                  struct fuse_copy_state *cs, size_t nbytes)
1702 {
1703         int err;
1704         struct fuse_req *req;
1705         struct fuse_out_header oh;
1706
1707         if (nbytes < sizeof(struct fuse_out_header))
1708                 return -EINVAL;
1709
1710         err = fuse_copy_one(cs, &oh, sizeof(oh));
1711         if (err)
1712                 goto err_finish;
1713
1714         err = -EINVAL;
1715         if (oh.len != nbytes)
1716                 goto err_finish;
1717
1718         /*
1719          * Zero oh.unique indicates unsolicited notification message
1720          * and error contains notification code.
1721          */
1722         if (!oh.unique) {
1723                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1724                 return err ? err : nbytes;
1725         }
1726
1727         err = -EINVAL;
1728         if (oh.error <= -1000 || oh.error > 0)
1729                 goto err_finish;
1730
1731         spin_lock(&fc->lock);
1732         err = -ENOENT;
1733         if (!fc->connected)
1734                 goto err_unlock;
1735
1736         req = request_find(fc, oh.unique);
1737         if (!req)
1738                 goto err_unlock;
1739
1740         if (req->aborted) {
1741                 spin_unlock(&fc->lock);
1742                 fuse_copy_finish(cs);
1743                 spin_lock(&fc->lock);
1744                 request_end(fc, req);
1745                 return -ENOENT;
1746         }
1747         /* Is it an interrupt reply? */
1748         if (req->intr_unique == oh.unique) {
1749                 err = -EINVAL;
1750                 if (nbytes != sizeof(struct fuse_out_header))
1751                         goto err_unlock;
1752
1753                 if (oh.error == -ENOSYS)
1754                         fc->no_interrupt = 1;
1755                 else if (oh.error == -EAGAIN)
1756                         queue_interrupt(fc, req);
1757
1758                 spin_unlock(&fc->lock);
1759                 fuse_copy_finish(cs);
1760                 return nbytes;
1761         }
1762
1763         req->state = FUSE_REQ_WRITING;
1764         list_move(&req->list, &fc->io);
1765         req->out.h = oh;
1766         req->locked = 1;
1767         cs->req = req;
1768         if (!req->out.page_replace)
1769                 cs->move_pages = 0;
1770         spin_unlock(&fc->lock);
1771
1772         err = copy_out_args(cs, &req->out, nbytes);
1773         fuse_copy_finish(cs);
1774
1775         spin_lock(&fc->lock);
1776         req->locked = 0;
1777         if (!err) {
1778                 if (req->aborted)
1779                         err = -ENOENT;
1780         } else if (!req->aborted)
1781                 req->out.h.error = -EIO;
1782         request_end(fc, req);
1783
1784         return err ? err : nbytes;
1785
1786  err_unlock:
1787         spin_unlock(&fc->lock);
1788  err_finish:
1789         fuse_copy_finish(cs);
1790         return err;
1791 }
1792
1793 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
1794                               unsigned long nr_segs, loff_t pos)
1795 {
1796         struct fuse_copy_state cs;
1797         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1798         if (!fc)
1799                 return -EPERM;
1800
1801         fuse_copy_init(&cs, fc, 0, iov, nr_segs);
1802
1803         return fuse_dev_do_write(fc, &cs, iov_length(iov, nr_segs));
1804 }
1805
1806 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1807                                      struct file *out, loff_t *ppos,
1808                                      size_t len, unsigned int flags)
1809 {
1810         unsigned nbuf;
1811         unsigned idx;
1812         struct pipe_buffer *bufs;
1813         struct fuse_copy_state cs;
1814         struct fuse_conn *fc;
1815         size_t rem;
1816         ssize_t ret;
1817
1818         fc = fuse_get_conn(out);
1819         if (!fc)
1820                 return -EPERM;
1821
1822         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1823         if (!bufs)
1824                 return -ENOMEM;
1825
1826         pipe_lock(pipe);
1827         nbuf = 0;
1828         rem = 0;
1829         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1830                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1831
1832         ret = -EINVAL;
1833         if (rem < len) {
1834                 pipe_unlock(pipe);
1835                 goto out;
1836         }
1837
1838         rem = len;
1839         while (rem) {
1840                 struct pipe_buffer *ibuf;
1841                 struct pipe_buffer *obuf;
1842
1843                 BUG_ON(nbuf >= pipe->buffers);
1844                 BUG_ON(!pipe->nrbufs);
1845                 ibuf = &pipe->bufs[pipe->curbuf];
1846                 obuf = &bufs[nbuf];
1847
1848                 if (rem >= ibuf->len) {
1849                         *obuf = *ibuf;
1850                         ibuf->ops = NULL;
1851                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1852                         pipe->nrbufs--;
1853                 } else {
1854                         ibuf->ops->get(pipe, ibuf);
1855                         *obuf = *ibuf;
1856                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1857                         obuf->len = rem;
1858                         ibuf->offset += obuf->len;
1859                         ibuf->len -= obuf->len;
1860                 }
1861                 nbuf++;
1862                 rem -= obuf->len;
1863         }
1864         pipe_unlock(pipe);
1865
1866         fuse_copy_init(&cs, fc, 0, NULL, nbuf);
1867         cs.pipebufs = bufs;
1868         cs.pipe = pipe;
1869
1870         if (flags & SPLICE_F_MOVE)
1871                 cs.move_pages = 1;
1872
1873         ret = fuse_dev_do_write(fc, &cs, len);
1874
1875         for (idx = 0; idx < nbuf; idx++) {
1876                 struct pipe_buffer *buf = &bufs[idx];
1877                 buf->ops->release(pipe, buf);
1878         }
1879 out:
1880         kfree(bufs);
1881         return ret;
1882 }
1883
1884 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1885 {
1886         unsigned mask = POLLOUT | POLLWRNORM;
1887         struct fuse_conn *fc = fuse_get_conn(file);
1888         if (!fc)
1889                 return POLLERR;
1890
1891         poll_wait(file, &fc->waitq, wait);
1892
1893         spin_lock(&fc->lock);
1894         if (!fc->connected)
1895                 mask = POLLERR;
1896         else if (request_pending(fc))
1897                 mask |= POLLIN | POLLRDNORM;
1898         spin_unlock(&fc->lock);
1899
1900         return mask;
1901 }
1902
1903 /*
1904  * Abort all requests on the given list (pending or processing)
1905  *
1906  * This function releases and reacquires fc->lock
1907  */
1908 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1909 __releases(fc->lock)
1910 __acquires(fc->lock)
1911 {
1912         while (!list_empty(head)) {
1913                 struct fuse_req *req;
1914                 req = list_entry(head->next, struct fuse_req, list);
1915                 req->out.h.error = -ECONNABORTED;
1916                 request_end(fc, req);
1917                 spin_lock(&fc->lock);
1918         }
1919 }
1920
1921 /*
1922  * Abort requests under I/O
1923  *
1924  * The requests are set to aborted and finished, and the request
1925  * waiter is woken up.  This will make request_wait_answer() wait
1926  * until the request is unlocked and then return.
1927  *
1928  * If the request is asynchronous, then the end function needs to be
1929  * called after waiting for the request to be unlocked (if it was
1930  * locked).
1931  */
1932 static void end_io_requests(struct fuse_conn *fc)
1933 __releases(fc->lock)
1934 __acquires(fc->lock)
1935 {
1936         while (!list_empty(&fc->io)) {
1937                 struct fuse_req *req =
1938                         list_entry(fc->io.next, struct fuse_req, list);
1939                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1940
1941                 req->aborted = 1;
1942                 req->out.h.error = -ECONNABORTED;
1943                 req->state = FUSE_REQ_FINISHED;
1944                 list_del_init(&req->list);
1945                 wake_up(&req->waitq);
1946                 if (end) {
1947                         req->end = NULL;
1948                         __fuse_get_request(req);
1949                         spin_unlock(&fc->lock);
1950                         wait_event(req->waitq, !req->locked);
1951                         end(fc, req);
1952                         fuse_put_request(fc, req);
1953                         spin_lock(&fc->lock);
1954                 }
1955         }
1956 }
1957
1958 static void end_queued_requests(struct fuse_conn *fc)
1959 __releases(fc->lock)
1960 __acquires(fc->lock)
1961 {
1962         fc->max_background = UINT_MAX;
1963         flush_bg_queue(fc);
1964         end_requests(fc, &fc->pending);
1965         end_requests(fc, &fc->processing);
1966         while (forget_pending(fc))
1967                 kfree(dequeue_forget(fc, 1, NULL));
1968 }
1969
1970 static void end_polls(struct fuse_conn *fc)
1971 {
1972         struct rb_node *p;
1973
1974         p = rb_first(&fc->polled_files);
1975
1976         while (p) {
1977                 struct fuse_file *ff;
1978                 ff = rb_entry(p, struct fuse_file, polled_node);
1979                 wake_up_interruptible_all(&ff->poll_wait);
1980
1981                 p = rb_next(p);
1982         }
1983 }
1984
1985 /*
1986  * Abort all requests.
1987  *
1988  * Emergency exit in case of a malicious or accidental deadlock, or
1989  * just a hung filesystem.
1990  *
1991  * The same effect is usually achievable through killing the
1992  * filesystem daemon and all users of the filesystem.  The exception
1993  * is the combination of an asynchronous request and the tricky
1994  * deadlock (see Documentation/filesystems/fuse.txt).
1995  *
1996  * During the aborting, progression of requests from the pending and
1997  * processing lists onto the io list, and progression of new requests
1998  * onto the pending list is prevented by req->connected being false.
1999  *
2000  * Progression of requests under I/O to the processing list is
2001  * prevented by the req->aborted flag being true for these requests.
2002  * For this reason requests on the io list must be aborted first.
2003  */
2004 void fuse_abort_conn(struct fuse_conn *fc)
2005 {
2006         spin_lock(&fc->lock);
2007         if (fc->connected) {
2008                 fc->connected = 0;
2009                 fc->blocked = 0;
2010                 end_io_requests(fc);
2011                 end_queued_requests(fc);
2012                 end_polls(fc);
2013                 wake_up_all(&fc->waitq);
2014                 wake_up_all(&fc->blocked_waitq);
2015                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
2016         }
2017         spin_unlock(&fc->lock);
2018 }
2019 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2020
2021 int fuse_dev_release(struct inode *inode, struct file *file)
2022 {
2023         struct fuse_conn *fc = fuse_get_conn(file);
2024         if (fc) {
2025                 spin_lock(&fc->lock);
2026                 fc->connected = 0;
2027                 fc->blocked = 0;
2028                 end_queued_requests(fc);
2029                 end_polls(fc);
2030                 wake_up_all(&fc->blocked_waitq);
2031                 spin_unlock(&fc->lock);
2032                 fuse_conn_put(fc);
2033         }
2034
2035         return 0;
2036 }
2037 EXPORT_SYMBOL_GPL(fuse_dev_release);
2038
2039 static int fuse_dev_fasync(int fd, struct file *file, int on)
2040 {
2041         struct fuse_conn *fc = fuse_get_conn(file);
2042         if (!fc)
2043                 return -EPERM;
2044
2045         /* No locking - fasync_helper does its own locking */
2046         return fasync_helper(fd, file, on, &fc->fasync);
2047 }
2048
2049 const struct file_operations fuse_dev_operations = {
2050         .owner          = THIS_MODULE,
2051         .llseek         = no_llseek,
2052         .read           = do_sync_read,
2053         .aio_read       = fuse_dev_read,
2054         .splice_read    = fuse_dev_splice_read,
2055         .write          = do_sync_write,
2056         .aio_write      = fuse_dev_write,
2057         .splice_write   = fuse_dev_splice_write,
2058         .poll           = fuse_dev_poll,
2059         .release        = fuse_dev_release,
2060         .fasync         = fuse_dev_fasync,
2061 };
2062 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2063
2064 static struct miscdevice fuse_miscdevice = {
2065         .minor = FUSE_MINOR,
2066         .name  = "fuse",
2067         .fops = &fuse_dev_operations,
2068 };
2069
2070 int __init fuse_dev_init(void)
2071 {
2072         int err = -ENOMEM;
2073         fuse_req_cachep = kmem_cache_create("fuse_request",
2074                                             sizeof(struct fuse_req),
2075                                             0, 0, NULL);
2076         if (!fuse_req_cachep)
2077                 goto out;
2078
2079         err = misc_register(&fuse_miscdevice);
2080         if (err)
2081                 goto out_cache_clean;
2082
2083         return 0;
2084
2085  out_cache_clean:
2086         kmem_cache_destroy(fuse_req_cachep);
2087  out:
2088         return err;
2089 }
2090
2091 void fuse_dev_cleanup(void)
2092 {
2093         misc_deregister(&fuse_miscdevice);
2094         kmem_cache_destroy(fuse_req_cachep);
2095 }