fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/pipe_fs_i.h>
  20 #include <linux/swap.h>
  21 #include <linux/splice.h>
  22
  23 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  24 MODULE_ALIAS("devname:fuse");
  25
  26 static struct kmem_cache *fuse_req_cachep;
  27
  28 static struct fuse_conn *fuse_get_conn(struct file *file)
  29 {
  30         /*
  31          * Lockless access is OK, because file->private data is set
  32          * once during mount and is valid until the file is released.
  33          */
  34         return file->private_data;
  35 }
  36
  37 static void fuse_request_init(struct fuse_req *req)
  38 {
  39         memset(req, 0, sizeof(*req));
  40         INIT_LIST_HEAD(&req->list);
  41         INIT_LIST_HEAD(&req->intr_entry);
  42         init_waitqueue_head(&req->waitq);
  43         atomic_set(&req->count, 1);
  44 }
  45
  46 struct fuse_req *fuse_request_alloc(void)
  47 {
  48         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_KERNEL);
  49         if (req)
  50                 fuse_request_init(req);
  51         return req;
  52 }
  53 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  54
  55 struct fuse_req *fuse_request_alloc_nofs(void)
  56 {
  57         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_NOFS);
  58         if (req)
  59                 fuse_request_init(req);
  60         return req;
  61 }
  62
  63 void fuse_request_free(struct fuse_req *req)
  64 {
  65         kmem_cache_free(fuse_req_cachep, req);
  66 }
  67
  68 static void block_sigs(sigset_t *oldset)
  69 {
  70         sigset_t mask;
  71
  72         siginitsetinv(&mask, sigmask(SIGKILL));
  73         sigprocmask(SIG_BLOCK, &mask, oldset);
  74 }
  75
  76 static void restore_sigs(sigset_t *oldset)
  77 {
  78         sigprocmask(SIG_SETMASK, oldset, NULL);
  79 }
  80
  81 static void __fuse_get_request(struct fuse_req *req)
  82 {
  83         atomic_inc(&req->count);
  84 }
  85
  86 /* Must be called with > 1 refcount */
  87 static void __fuse_put_request(struct fuse_req *req)
  88 {
  89         BUG_ON(atomic_read(&req->count) < 2);
  90         atomic_dec(&req->count);
  91 }
  92
  93 static void fuse_req_init_context(struct fuse_req *req)
  94 {
  95         req->in.h.uid = current_fsuid();
  96         req->in.h.gid = current_fsgid();
  97         req->in.h.pid = current->pid;
  98 }
  99
 100 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
 101 {
 102         struct fuse_req *req;
 103         sigset_t oldset;
 104         int intr;
 105         int err;
 106
 107         atomic_inc(&fc->num_waiting);
 108         block_sigs(&oldset);
 109         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
 110         restore_sigs(&oldset);
 111         err = -EINTR;
 112         if (intr)
 113                 goto out;
 114
 115         err = -ENOTCONN;
 116         if (!fc->connected)
 117                 goto out;
 118
 119         req = fuse_request_alloc();
 120         err = -ENOMEM;
 121         if (!req)
 122                 goto out;
 123
 124         fuse_req_init_context(req);
 125         req->waiting = 1;
 126         return req;
 127
 128  out:
 129         atomic_dec(&fc->num_waiting);
 130         return ERR_PTR(err);
 131 }
 132 EXPORT_SYMBOL_GPL(fuse_get_req);
 133
 134 /*
 135  * Return request in fuse_file->reserved_req.  However that may
 136  * currently be in use.  If that is the case, wait for it to become
 137  * available.
 138  */
 139 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 140                                          struct file *file)
 141 {
 142         struct fuse_req *req = NULL;
 143         struct fuse_file *ff = file->private_data;
 144
 145         do {
 146                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 147                 spin_lock(&fc->lock);
 148                 if (ff->reserved_req) {
 149                         req = ff->reserved_req;
 150                         ff->reserved_req = NULL;
 151                         get_file(file);
 152                         req->stolen_file = file;
 153                 }
 154                 spin_unlock(&fc->lock);
 155         } while (!req);
 156
 157         return req;
 158 }
 159
 160 /*
 161  * Put stolen request back into fuse_file->reserved_req
 162  */
 163 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 164 {
 165         struct file *file = req->stolen_file;
 166         struct fuse_file *ff = file->private_data;
 167
 168         spin_lock(&fc->lock);
 169         fuse_request_init(req);
 170         BUG_ON(ff->reserved_req);
 171         ff->reserved_req = req;
 172         wake_up_all(&fc->reserved_req_waitq);
 173         spin_unlock(&fc->lock);
 174         fput(file);
 175 }
 176
 177 /*
 178  * Gets a requests for a file operation, always succeeds
 179  *
 180  * This is used for sending the FLUSH request, which must get to
 181  * userspace, due to POSIX locks which may need to be unlocked.
 182  *
 183  * If allocation fails due to OOM, use the reserved request in
 184  * fuse_file.
 185  *
 186  * This is very unlikely to deadlock accidentally, since the
 187  * filesystem should not have it's own file open.  If deadlock is
 188  * intentional, it can still be broken by "aborting" the filesystem.
 189  */
 190 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 191 {
 192         struct fuse_req *req;
 193
 194         atomic_inc(&fc->num_waiting);
 195         wait_event(fc->blocked_waitq, !fc->blocked);
 196         req = fuse_request_alloc();
 197         if (!req)
 198                 req = get_reserved_req(fc, file);
 199
 200         fuse_req_init_context(req);
 201         req->waiting = 1;
 202         return req;
 203 }
 204
 205 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 206 {
 207         if (atomic_dec_and_test(&req->count)) {
 208                 if (req->waiting)
 209                         atomic_dec(&fc->num_waiting);
 210
 211                 if (req->stolen_file)
 212                         put_reserved_req(fc, req);
 213                 else
 214                         fuse_request_free(req);
 215         }
 216 }
 217 EXPORT_SYMBOL_GPL(fuse_put_request);
 218
 219 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 220 {
 221         unsigned nbytes = 0;
 222         unsigned i;
 223
 224         for (i = 0; i < numargs; i++)
 225                 nbytes += args[i].size;
 226
 227         return nbytes;
 228 }
 229
 230 static u64 fuse_get_unique(struct fuse_conn *fc)
 231 {
 232         fc->reqctr++;
 233         /* zero is special */
 234         if (fc->reqctr == 0)
 235                 fc->reqctr = 1;
 236
 237         return fc->reqctr;
 238 }
 239
 240 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 241 {
 242         req->in.h.len = sizeof(struct fuse_in_header) +
 243                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 244         list_add_tail(&req->list, &fc->pending);
 245         req->state = FUSE_REQ_PENDING;
 246         if (!req->waiting) {
 247                 req->waiting = 1;
 248                 atomic_inc(&fc->num_waiting);
 249         }
 250         wake_up(&fc->waitq);
 251         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 252 }
 253
 254 static void flush_bg_queue(struct fuse_conn *fc)
 255 {
 256         while (fc->active_background < fc->max_background &&
 257                !list_empty(&fc->bg_queue)) {
 258                 struct fuse_req *req;
 259
 260                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 261                 list_del(&req->list);
 262                 fc->active_background++;
 263                 req->in.h.unique = fuse_get_unique(fc);
 264                 queue_request(fc, req);
 265         }
 266 }
 267
 268 /*
 269  * This function is called when a request is finished.  Either a reply
 270  * has arrived or it was aborted (and not yet sent) or some error
 271  * occurred during communication with userspace, or the device file
 272  * was closed.  The requester thread is woken up (if still waiting),
 273  * the 'end' callback is called if given, else the reference to the
 274  * request is released
 275  *
 276  * Called with fc->lock, unlocks it
 277  */
 278 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 279 __releases(fc->lock)
 280 {
 281         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 282         req->end = NULL;
 283         list_del(&req->list);
 284         list_del(&req->intr_entry);
 285         req->state = FUSE_REQ_FINISHED;
 286         if (req->background) {
 287                 if (fc->num_background == fc->max_background) {
 288                         fc->blocked = 0;
 289                         wake_up_all(&fc->blocked_waitq);
 290                 }
 291                 if (fc->num_background == fc->congestion_threshold &&
 292                     fc->connected && fc->bdi_initialized) {
 293                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 294                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 295                 }
 296                 fc->num_background--;
 297                 fc->active_background--;
 298                 flush_bg_queue(fc);
 299         }
 300         spin_unlock(&fc->lock);
 301         wake_up(&req->waitq);
 302         if (end)
 303                 end(fc, req);
 304         fuse_put_request(fc, req);
 305 }
 306
 307 static void wait_answer_interruptible(struct fuse_conn *fc,
 308                                       struct fuse_req *req)
 309 __releases(fc->lock)
 310 __acquires(fc->lock)
 311 {
 312         if (signal_pending(current))
 313                 return;
 314
 315         spin_unlock(&fc->lock);
 316         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 317         spin_lock(&fc->lock);
 318 }
 319
 320 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 321 {
 322         list_add_tail(&req->intr_entry, &fc->interrupts);
 323         wake_up(&fc->waitq);
 324         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 325 }
 326
 327 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 328 __releases(fc->lock)
 329 __acquires(fc->lock)
 330 {
 331         if (!fc->no_interrupt) {
 332                 /* Any signal may interrupt this */
 333                 wait_answer_interruptible(fc, req);
 334
 335                 if (req->aborted)
 336                         goto aborted;
 337                 if (req->state == FUSE_REQ_FINISHED)
 338                         return;
 339
 340                 req->interrupted = 1;
 341                 if (req->state == FUSE_REQ_SENT)
 342                         queue_interrupt(fc, req);
 343         }
 344
 345         if (!req->force) {
 346                 sigset_t oldset;
 347
 348                 /* Only fatal signals may interrupt this */
 349                 block_sigs(&oldset);
 350                 wait_answer_interruptible(fc, req);
 351                 restore_sigs(&oldset);
 352
 353                 if (req->aborted)
 354                         goto aborted;
 355                 if (req->state == FUSE_REQ_FINISHED)
 356                         return;
 357
 358                 /* Request is not yet in userspace, bail out */
 359                 if (req->state == FUSE_REQ_PENDING) {
 360                         list_del(&req->list);
 361                         __fuse_put_request(req);
 362                         req->out.h.error = -EINTR;
 363                         return;
 364                 }
 365         }
 366
 367         /*
 368          * Either request is already in userspace, or it was forced.
 369          * Wait it out.
 370          */
 371         spin_unlock(&fc->lock);
 372         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 373         spin_lock(&fc->lock);
 374
 375         if (!req->aborted)
 376                 return;
 377
 378  aborted:
 379         BUG_ON(req->state != FUSE_REQ_FINISHED);
 380         if (req->locked) {
 381                 /* This is uninterruptible sleep, because data is
 382                    being copied to/from the buffers of req.  During
 383                    locked state, there mustn't be any filesystem
 384                    operation (e.g. page fault), since that could lead
 385                    to deadlock */
 386                 spin_unlock(&fc->lock);
 387                 wait_event(req->waitq, !req->locked);
 388                 spin_lock(&fc->lock);
 389         }
 390 }
 391
 392 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 393 {
 394         req->isreply = 1;
 395         spin_lock(&fc->lock);
 396         if (!fc->connected)
 397                 req->out.h.error = -ENOTCONN;
 398         else if (fc->conn_error)
 399                 req->out.h.error = -ECONNREFUSED;
 400         else {
 401                 req->in.h.unique = fuse_get_unique(fc);
 402                 queue_request(fc, req);
 403                 /* acquire extra reference, since request is still needed
 404                    after request_end() */
 405                 __fuse_get_request(req);
 406
 407                 request_wait_answer(fc, req);
 408         }
 409         spin_unlock(&fc->lock);
 410 }
 411 EXPORT_SYMBOL_GPL(fuse_request_send);
 412
 413 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 414                                             struct fuse_req *req)
 415 {
 416         req->background = 1;
 417         fc->num_background++;
 418         if (fc->num_background == fc->max_background)
 419                 fc->blocked = 1;
 420         if (fc->num_background == fc->congestion_threshold &&
 421             fc->bdi_initialized) {
 422                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 423                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 424         }
 425         list_add_tail(&req->list, &fc->bg_queue);
 426         flush_bg_queue(fc);
 427 }
 428
 429 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 430 {
 431         spin_lock(&fc->lock);
 432         if (fc->connected) {
 433                 fuse_request_send_nowait_locked(fc, req);
 434                 spin_unlock(&fc->lock);
 435         } else {
 436                 req->out.h.error = -ENOTCONN;
 437                 request_end(fc, req);
 438         }
 439 }
 440
 441 void fuse_request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 442 {
 443         req->isreply = 0;
 444         fuse_request_send_nowait(fc, req);
 445 }
 446
 447 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 448 {
 449         req->isreply = 1;
 450         fuse_request_send_nowait(fc, req);
 451 }
 452 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 453
 454 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 455                                           struct fuse_req *req, u64 unique)
 456 {
 457         int err = -ENODEV;
 458
 459         req->isreply = 0;
 460         req->in.h.unique = unique;
 461         spin_lock(&fc->lock);
 462         if (fc->connected) {
 463                 queue_request(fc, req);
 464                 err = 0;
 465         }
 466         spin_unlock(&fc->lock);
 467
 468         return err;
 469 }
 470
 471 /*
 472  * Called under fc->lock
 473  *
 474  * fc->connected must have been checked previously
 475  */
 476 void fuse_request_send_background_locked(struct fuse_conn *fc,
 477                                          struct fuse_req *req)
 478 {
 479         req->isreply = 1;
 480         fuse_request_send_nowait_locked(fc, req);
 481 }
 482
 483 /*
 484  * Lock the request.  Up to the next unlock_request() there mustn't be
 485  * anything that could cause a page-fault.  If the request was already
 486  * aborted bail out.
 487  */
 488 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 489 {
 490         int err = 0;
 491         if (req) {
 492                 spin_lock(&fc->lock);
 493                 if (req->aborted)
 494                         err = -ENOENT;
 495                 else
 496                         req->locked = 1;
 497                 spin_unlock(&fc->lock);
 498         }
 499         return err;
 500 }
 501
 502 /*
 503  * Unlock request.  If it was aborted during being locked, the
 504  * requester thread is currently waiting for it to be unlocked, so
 505  * wake it up.
 506  */
 507 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 508 {
 509         if (req) {
 510                 spin_lock(&fc->lock);
 511                 req->locked = 0;
 512                 if (req->aborted)
 513                         wake_up(&req->waitq);
 514                 spin_unlock(&fc->lock);
 515         }
 516 }
 517
 518 struct fuse_copy_state {
 519         struct fuse_conn *fc;
 520         int write;
 521         struct fuse_req *req;
 522         const struct iovec *iov;
 523         struct pipe_buffer *pipebufs;
 524         struct pipe_buffer *currbuf;
 525         struct pipe_inode_info *pipe;
 526         unsigned long nr_segs;
 527         unsigned long seglen;
 528         unsigned long addr;
 529         struct page *pg;
 530         void *mapaddr;
 531         void *buf;
 532         unsigned len;
 533         unsigned move_pages:1;
 534 };
 535
 536 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 537                            int write,
 538                            const struct iovec *iov, unsigned long nr_segs)
 539 {
 540         memset(cs, 0, sizeof(*cs));
 541         cs->fc = fc;
 542         cs->write = write;
 543         cs->iov = iov;
 544         cs->nr_segs = nr_segs;
 545 }
 546
 547 /* Unmap and put previous page of userspace buffer */
 548 static void fuse_copy_finish(struct fuse_copy_state *cs)
 549 {
 550         if (cs->currbuf) {
 551                 struct pipe_buffer *buf = cs->currbuf;
 552
 553                 if (!cs->write) {
 554                         buf->ops->unmap(cs->pipe, buf, cs->mapaddr);
 555                 } else {
 556                         kunmap(buf->page);
 557                         buf->len = PAGE_SIZE - cs->len;
 558                 }
 559                 cs->currbuf = NULL;
 560                 cs->mapaddr = NULL;
 561         } else if (cs->mapaddr) {
 562                 kunmap(cs->pg);
 563                 if (cs->write) {
 564                         flush_dcache_page(cs->pg);
 565                         set_page_dirty_lock(cs->pg);
 566                 }
 567                 put_page(cs->pg);
 568                 cs->mapaddr = NULL;
 569         }
 570 }
 571
 572 /*
 573  * Get another pagefull of userspace buffer, and map it to kernel
 574  * address space, and lock request
 575  */
 576 static int fuse_copy_fill(struct fuse_copy_state *cs)
 577 {
 578         unsigned long offset;
 579         int err;
 580
 581         unlock_request(cs->fc, cs->req);
 582         fuse_copy_finish(cs);
 583         if (cs->pipebufs) {
 584                 struct pipe_buffer *buf = cs->pipebufs;
 585
 586                 if (!cs->write) {
 587                         err = buf->ops->confirm(cs->pipe, buf);
 588                         if (err)
 589                                 return err;
 590
 591                         BUG_ON(!cs->nr_segs);
 592                         cs->currbuf = buf;
 593                         cs->mapaddr = buf->ops->map(cs->pipe, buf, 0);
 594                         cs->len = buf->len;
 595                         cs->buf = cs->mapaddr + buf->offset;
 596                         cs->pipebufs++;
 597                         cs->nr_segs--;
 598                 } else {
 599                         struct page *page;
 600
 601                         if (cs->nr_segs == cs->pipe->buffers)
 602                                 return -EIO;
 603
 604                         page = alloc_page(GFP_HIGHUSER);
 605                         if (!page)
 606                                 return -ENOMEM;
 607
 608                         buf->page = page;
 609                         buf->offset = 0;
 610                         buf->len = 0;
 611
 612                         cs->currbuf = buf;
 613                         cs->mapaddr = kmap(page);
 614                         cs->buf = cs->mapaddr;
 615                         cs->len = PAGE_SIZE;
 616                         cs->pipebufs++;
 617                         cs->nr_segs++;
 618                 }
 619         } else {
 620                 if (!cs->seglen) {
 621                         BUG_ON(!cs->nr_segs);
 622                         cs->seglen = cs->iov[0].iov_len;
 623                         cs->addr = (unsigned long) cs->iov[0].iov_base;
 624                         cs->iov++;
 625                         cs->nr_segs--;
 626                 }
 627                 err = get_user_pages_fast(cs->addr, 1, cs->write, &cs->pg);
 628                 if (err < 0)
 629                         return err;
 630                 BUG_ON(err != 1);
 631                 offset = cs->addr % PAGE_SIZE;
 632                 cs->mapaddr = kmap(cs->pg);
 633                 cs->buf = cs->mapaddr + offset;
 634                 cs->len = min(PAGE_SIZE - offset, cs->seglen);
 635                 cs->seglen -= cs->len;
 636                 cs->addr += cs->len;
 637         }
 638
 639         return lock_request(cs->fc, cs->req);
 640 }
 641
 642 /* Do as much copy to/from userspace buffer as we can */
 643 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 644 {
 645         unsigned ncpy = min(*size, cs->len);
 646         if (val) {
 647                 if (cs->write)
 648                         memcpy(cs->buf, *val, ncpy);
 649                 else
 650                         memcpy(*val, cs->buf, ncpy);
 651                 *val += ncpy;
 652         }
 653         *size -= ncpy;
 654         cs->len -= ncpy;
 655         cs->buf += ncpy;
 656         return ncpy;
 657 }
 658
 659 static int fuse_check_page(struct page *page)
 660 {
 661         if (page_mapcount(page) ||
 662             page->mapping != NULL ||
 663             page_count(page) != 1 ||
 664             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 665              ~(1 << PG_locked |
 666                1 << PG_referenced |
 667                1 << PG_uptodate |
 668                1 << PG_lru |
 669                1 << PG_active |
 670                1 << PG_reclaim))) {
 671                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 672                 printk(KERN_WARNING "  page=%p index=%li flags=%08lx, count=%i, mapcount=%i, mapping=%p\n", page, page->index, page->flags, page_count(page), page_mapcount(page), page->mapping);
 673                 return 1;
 674         }
 675         return 0;
 676 }
 677
 678 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 679 {
 680         int err;
 681         struct page *oldpage = *pagep;
 682         struct page *newpage;
 683         struct pipe_buffer *buf = cs->pipebufs;
 684         struct address_space *mapping;
 685         pgoff_t index;
 686
 687         unlock_request(cs->fc, cs->req);
 688         fuse_copy_finish(cs);
 689
 690         err = buf->ops->confirm(cs->pipe, buf);
 691         if (err)
 692                 return err;
 693
 694         BUG_ON(!cs->nr_segs);
 695         cs->currbuf = buf;
 696         cs->len = buf->len;
 697         cs->pipebufs++;
 698         cs->nr_segs--;
 699
 700         if (cs->len != PAGE_SIZE)
 701                 goto out_fallback;
 702
 703         if (buf->ops->steal(cs->pipe, buf) != 0)
 704                 goto out_fallback;
 705
 706         newpage = buf->page;
 707
 708         if (WARN_ON(!PageUptodate(newpage)))
 709                 return -EIO;
 710
 711         ClearPageMappedToDisk(newpage);
 712
 713         if (fuse_check_page(newpage) != 0)
 714                 goto out_fallback_unlock;
 715
 716         mapping = oldpage->mapping;
 717         index = oldpage->index;
 718
 719         /*
 720          * This is a new and locked page, it shouldn't be mapped or
 721          * have any special flags on it
 722          */
 723         if (WARN_ON(page_mapped(oldpage)))
 724                 goto out_fallback_unlock;
 725         if (WARN_ON(page_has_private(oldpage)))
 726                 goto out_fallback_unlock;
 727         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 728                 goto out_fallback_unlock;
 729         if (WARN_ON(PageMlocked(oldpage)))
 730                 goto out_fallback_unlock;
 731
 732         remove_from_page_cache(oldpage);
 733         page_cache_release(oldpage);
 734
 735         err = add_to_page_cache_locked(newpage, mapping, index, GFP_KERNEL);
 736         if (err) {
 737                 printk(KERN_WARNING "fuse_try_move_page: failed to add page");
 738                 goto out_fallback_unlock;
 739         }
 740         page_cache_get(newpage);
 741
 742         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 743                 lru_cache_add_file(newpage);
 744
 745         err = 0;
 746         spin_lock(&cs->fc->lock);
 747         if (cs->req->aborted)
 748                 err = -ENOENT;
 749         else
 750                 *pagep = newpage;
 751         spin_unlock(&cs->fc->lock);
 752
 753         if (err) {
 754                 unlock_page(newpage);
 755                 page_cache_release(newpage);
 756                 return err;
 757         }
 758
 759         unlock_page(oldpage);
 760         page_cache_release(oldpage);
 761         cs->len = 0;
 762
 763         return 0;
 764
 765 out_fallback_unlock:
 766         unlock_page(newpage);
 767 out_fallback:
 768         cs->mapaddr = buf->ops->map(cs->pipe, buf, 1);
 769         cs->buf = cs->mapaddr + buf->offset;
 770
 771         err = lock_request(cs->fc, cs->req);
 772         if (err)
 773                 return err;
 774
 775         return 1;
 776 }
 777
 778 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 779                          unsigned offset, unsigned count)
 780 {
 781         struct pipe_buffer *buf;
 782
 783         if (cs->nr_segs == cs->pipe->buffers)
 784                 return -EIO;
 785
 786         unlock_request(cs->fc, cs->req);
 787         fuse_copy_finish(cs);
 788
 789         buf = cs->pipebufs;
 790         page_cache_get(page);
 791         buf->page = page;
 792         buf->offset = offset;
 793         buf->len = count;
 794
 795         cs->pipebufs++;
 796         cs->nr_segs++;
 797         cs->len = 0;
 798
 799         return 0;
 800 }
 801
 802 /*
 803  * Copy a page in the request to/from the userspace buffer.  Must be
 804  * done atomically
 805  */
 806 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 807                           unsigned offset, unsigned count, int zeroing)
 808 {
 809         int err;
 810         struct page *page = *pagep;
 811
 812         if (page && zeroing && count < PAGE_SIZE)
 813                 clear_highpage(page);
 814
 815         while (count) {
 816                 if (cs->write && cs->pipebufs && page) {
 817                         return fuse_ref_page(cs, page, offset, count);
 818                 } else if (!cs->len) {
 819                         if (cs->move_pages && page &&
 820                             offset == 0 && count == PAGE_SIZE) {
 821                                 err = fuse_try_move_page(cs, pagep);
 822                                 if (err <= 0)
 823                                         return err;
 824                         } else {
 825                                 err = fuse_copy_fill(cs);
 826                                 if (err)
 827                                         return err;
 828                         }
 829                 }
 830                 if (page) {
 831                         void *mapaddr = kmap_atomic(page, KM_USER0);
 832                         void *buf = mapaddr + offset;
 833                         offset += fuse_copy_do(cs, &buf, &count);
 834                         kunmap_atomic(mapaddr, KM_USER0);
 835                 } else
 836                         offset += fuse_copy_do(cs, NULL, &count);
 837         }
 838         if (page && !cs->write)
 839                 flush_dcache_page(page);
 840         return 0;
 841 }
 842
 843 /* Copy pages in the request to/from userspace buffer */
 844 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 845                            int zeroing)
 846 {
 847         unsigned i;
 848         struct fuse_req *req = cs->req;
 849         unsigned offset = req->page_offset;
 850         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 851
 852         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 853                 int err;
 854
 855                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
 856                                      zeroing);
 857                 if (err)
 858                         return err;
 859
 860                 nbytes -= count;
 861                 count = min(nbytes, (unsigned) PAGE_SIZE);
 862                 offset = 0;
 863         }
 864         return 0;
 865 }
 866
 867 /* Copy a single argument in the request to/from userspace buffer */
 868 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 869 {
 870         while (size) {
 871                 if (!cs->len) {
 872                         int err = fuse_copy_fill(cs);
 873                         if (err)
 874                                 return err;
 875                 }
 876                 fuse_copy_do(cs, &val, &size);
 877         }
 878         return 0;
 879 }
 880
 881 /* Copy request arguments to/from userspace buffer */
 882 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 883                           unsigned argpages, struct fuse_arg *args,
 884                           int zeroing)
 885 {
 886         int err = 0;
 887         unsigned i;
 888
 889         for (i = 0; !err && i < numargs; i++)  {
 890                 struct fuse_arg *arg = &args[i];
 891                 if (i == numargs - 1 && argpages)
 892                         err = fuse_copy_pages(cs, arg->size, zeroing);
 893                 else
 894                         err = fuse_copy_one(cs, arg->value, arg->size);
 895         }
 896         return err;
 897 }
 898
 899 static int request_pending(struct fuse_conn *fc)
 900 {
 901         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 902 }
 903
 904 /* Wait until a request is available on the pending list */
 905 static void request_wait(struct fuse_conn *fc)
 906 __releases(fc->lock)
 907 __acquires(fc->lock)
 908 {
 909         DECLARE_WAITQUEUE(wait, current);
 910
 911         add_wait_queue_exclusive(&fc->waitq, &wait);
 912         while (fc->connected && !request_pending(fc)) {
 913                 set_current_state(TASK_INTERRUPTIBLE);
 914                 if (signal_pending(current))
 915                         break;
 916
 917                 spin_unlock(&fc->lock);
 918                 schedule();
 919                 spin_lock(&fc->lock);
 920         }
 921         set_current_state(TASK_RUNNING);
 922         remove_wait_queue(&fc->waitq, &wait);
 923 }
 924
 925 /*
 926  * Transfer an interrupt request to userspace
 927  *
 928  * Unlike other requests this is assembled on demand, without a need
 929  * to allocate a separate fuse_req structure.
 930  *
 931  * Called with fc->lock held, releases it
 932  */
 933 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
 934                                size_t nbytes, struct fuse_req *req)
 935 __releases(fc->lock)
 936 {
 937         struct fuse_in_header ih;
 938         struct fuse_interrupt_in arg;
 939         unsigned reqsize = sizeof(ih) + sizeof(arg);
 940         int err;
 941
 942         list_del_init(&req->intr_entry);
 943         req->intr_unique = fuse_get_unique(fc);
 944         memset(&ih, 0, sizeof(ih));
 945         memset(&arg, 0, sizeof(arg));
 946         ih.len = reqsize;
 947         ih.opcode = FUSE_INTERRUPT;
 948         ih.unique = req->intr_unique;
 949         arg.unique = req->in.h.unique;
 950
 951         spin_unlock(&fc->lock);
 952         if (nbytes < reqsize)
 953                 return -EINVAL;
 954
 955         err = fuse_copy_one(cs, &ih, sizeof(ih));
 956         if (!err)
 957                 err = fuse_copy_one(cs, &arg, sizeof(arg));
 958         fuse_copy_finish(cs);
 959
 960         return err ? err : reqsize;
 961 }
 962
 963 /*
 964  * Read a single request into the userspace filesystem's buffer.  This
 965  * function waits until a request is available, then removes it from
 966  * the pending list and copies request data to userspace buffer.  If
 967  * no reply is needed (FORGET) or request has been aborted or there
 968  * was an error during the copying then it's finished by calling
 969  * request_end().  Otherwise add it to the processing list, and set
 970  * the 'sent' flag.
 971  */
 972 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
 973                                 struct fuse_copy_state *cs, size_t nbytes)
 974 {
 975         int err;
 976         struct fuse_req *req;
 977         struct fuse_in *in;
 978         unsigned reqsize;
 979
 980  restart:
 981         spin_lock(&fc->lock);
 982         err = -EAGAIN;
 983         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 984             !request_pending(fc))
 985                 goto err_unlock;
 986
 987         request_wait(fc);
 988         err = -ENODEV;
 989         if (!fc->connected)
 990                 goto err_unlock;
 991         err = -ERESTARTSYS;
 992         if (!request_pending(fc))
 993                 goto err_unlock;
 994
 995         if (!list_empty(&fc->interrupts)) {
 996                 req = list_entry(fc->interrupts.next, struct fuse_req,
 997                                  intr_entry);
 998                 return fuse_read_interrupt(fc, cs, nbytes, req);
 999         }
1000
1001         req = list_entry(fc->pending.next, struct fuse_req, list);
1002         req->state = FUSE_REQ_READING;
1003         list_move(&req->list, &fc->io);
1004
1005         in = &req->in;
1006         reqsize = in->h.len;
1007         /* If request is too large, reply with an error and restart the read */
1008         if (nbytes < reqsize) {
1009                 req->out.h.error = -EIO;
1010                 /* SETXATTR is special, since it may contain too large data */
1011                 if (in->h.opcode == FUSE_SETXATTR)
1012                         req->out.h.error = -E2BIG;
1013                 request_end(fc, req);
1014                 goto restart;
1015         }
1016         spin_unlock(&fc->lock);
1017         cs->req = req;
1018         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1019         if (!err)
1020                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1021                                      (struct fuse_arg *) in->args, 0);
1022         fuse_copy_finish(cs);
1023         spin_lock(&fc->lock);
1024         req->locked = 0;
1025         if (req->aborted) {
1026                 request_end(fc, req);
1027                 return -ENODEV;
1028         }
1029         if (err) {
1030                 req->out.h.error = -EIO;
1031                 request_end(fc, req);
1032                 return err;
1033         }
1034         if (!req->isreply)
1035                 request_end(fc, req);
1036         else {
1037                 req->state = FUSE_REQ_SENT;
1038                 list_move_tail(&req->list, &fc->processing);
1039                 if (req->interrupted)
1040                         queue_interrupt(fc, req);
1041                 spin_unlock(&fc->lock);
1042         }
1043         return reqsize;
1044
1045  err_unlock:
1046         spin_unlock(&fc->lock);
1047         return err;
1048 }
1049
1050 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
1051                               unsigned long nr_segs, loff_t pos)
1052 {
1053         struct fuse_copy_state cs;
1054         struct file *file = iocb->ki_filp;
1055         struct fuse_conn *fc = fuse_get_conn(file);
1056         if (!fc)
1057                 return -EPERM;
1058
1059         fuse_copy_init(&cs, fc, 1, iov, nr_segs);
1060
1061         return fuse_dev_do_read(fc, file, &cs, iov_length(iov, nr_segs));
1062 }
1063
1064 static int fuse_dev_pipe_buf_steal(struct pipe_inode_info *pipe,
1065                                    struct pipe_buffer *buf)
1066 {
1067         return 1;
1068 }
1069
1070 static const struct pipe_buf_operations fuse_dev_pipe_buf_ops = {
1071         .can_merge = 0,
1072         .map = generic_pipe_buf_map,
1073         .unmap = generic_pipe_buf_unmap,
1074         .confirm = generic_pipe_buf_confirm,
1075         .release = generic_pipe_buf_release,
1076         .steal = fuse_dev_pipe_buf_steal,
1077         .get = generic_pipe_buf_get,
1078 };
1079
1080 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1081                                     struct pipe_inode_info *pipe,
1082                                     size_t len, unsigned int flags)
1083 {
1084         int ret;
1085         int page_nr = 0;
1086         int do_wakeup = 0;
1087         struct pipe_buffer *bufs;
1088         struct fuse_copy_state cs;
1089         struct fuse_conn *fc = fuse_get_conn(in);
1090         if (!fc)
1091                 return -EPERM;
1092
1093         bufs = kmalloc(pipe->buffers * sizeof (struct pipe_buffer), GFP_KERNEL);
1094         if (!bufs)
1095                 return -ENOMEM;
1096
1097         fuse_copy_init(&cs, fc, 1, NULL, 0);
1098         cs.pipebufs = bufs;
1099         cs.pipe = pipe;
1100         ret = fuse_dev_do_read(fc, in, &cs, len);
1101         if (ret < 0)
1102                 goto out;
1103
1104         ret = 0;
1105         pipe_lock(pipe);
1106
1107         if (!pipe->readers) {
1108                 send_sig(SIGPIPE, current, 0);
1109                 if (!ret)
1110                         ret = -EPIPE;
1111                 goto out_unlock;
1112         }
1113
1114         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1115                 ret = -EIO;
1116                 goto out_unlock;
1117         }
1118
1119         while (page_nr < cs.nr_segs) {
1120                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1121                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1122
1123                 buf->page = bufs[page_nr].page;
1124                 buf->offset = bufs[page_nr].offset;
1125                 buf->len = bufs[page_nr].len;
1126                 buf->ops = &fuse_dev_pipe_buf_ops;
1127
1128                 pipe->nrbufs++;
1129                 page_nr++;
1130                 ret += buf->len;
1131
1132                 if (pipe->inode)
1133                         do_wakeup = 1;
1134         }
1135
1136 out_unlock:
1137         pipe_unlock(pipe);
1138
1139         if (do_wakeup) {
1140                 smp_mb();
1141                 if (waitqueue_active(&pipe->wait))
1142                         wake_up_interruptible(&pipe->wait);
1143                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1144         }
1145
1146 out:
1147         for (; page_nr < cs.nr_segs; page_nr++)
1148                 page_cache_release(bufs[page_nr].page);
1149
1150         kfree(bufs);
1151         return ret;
1152 }
1153
1154 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1155                             struct fuse_copy_state *cs)
1156 {
1157         struct fuse_notify_poll_wakeup_out outarg;
1158         int err = -EINVAL;
1159
1160         if (size != sizeof(outarg))
1161                 goto err;
1162
1163         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1164         if (err)
1165                 goto err;
1166
1167         fuse_copy_finish(cs);
1168         return fuse_notify_poll_wakeup(fc, &outarg);
1169
1170 err:
1171         fuse_copy_finish(cs);
1172         return err;
1173 }
1174
1175 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1176                                    struct fuse_copy_state *cs)
1177 {
1178         struct fuse_notify_inval_inode_out outarg;
1179         int err = -EINVAL;
1180
1181         if (size != sizeof(outarg))
1182                 goto err;
1183
1184         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1185         if (err)
1186                 goto err;
1187         fuse_copy_finish(cs);
1188
1189         down_read(&fc->killsb);
1190         err = -ENOENT;
1191         if (fc->sb) {
1192                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1193                                                outarg.off, outarg.len);
1194         }
1195         up_read(&fc->killsb);
1196         return err;
1197
1198 err:
1199         fuse_copy_finish(cs);
1200         return err;
1201 }
1202
1203 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1204                                    struct fuse_copy_state *cs)
1205 {
1206         struct fuse_notify_inval_entry_out outarg;
1207         int err = -ENOMEM;
1208         char *buf;
1209         struct qstr name;
1210
1211         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1212         if (!buf)
1213                 goto err;
1214
1215         err = -EINVAL;
1216         if (size < sizeof(outarg))
1217                 goto err;
1218
1219         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1220         if (err)
1221                 goto err;
1222
1223         err = -ENAMETOOLONG;
1224         if (outarg.namelen > FUSE_NAME_MAX)
1225                 goto err;
1226
1227         name.name = buf;
1228         name.len = outarg.namelen;
1229         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1230         if (err)
1231                 goto err;
1232         fuse_copy_finish(cs);
1233         buf[outarg.namelen] = 0;
1234         name.hash = full_name_hash(name.name, name.len);
1235
1236         down_read(&fc->killsb);
1237         err = -ENOENT;
1238         if (fc->sb)
1239                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, &name);
1240         up_read(&fc->killsb);
1241         kfree(buf);
1242         return err;
1243
1244 err:
1245         kfree(buf);
1246         fuse_copy_finish(cs);
1247         return err;
1248 }
1249
1250 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1251                              struct fuse_copy_state *cs)
1252 {
1253         struct fuse_notify_store_out outarg;
1254         struct inode *inode;
1255         struct address_space *mapping;
1256         u64 nodeid;
1257         int err;
1258         pgoff_t index;
1259         unsigned int offset;
1260         unsigned int num;
1261         loff_t file_size;
1262         loff_t end;
1263
1264         err = -EINVAL;
1265         if (size < sizeof(outarg))
1266                 goto out_finish;
1267
1268         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1269         if (err)
1270                 goto out_finish;
1271
1272         err = -EINVAL;
1273         if (size - sizeof(outarg) != outarg.size)
1274                 goto out_finish;
1275
1276         nodeid = outarg.nodeid;
1277
1278         down_read(&fc->killsb);
1279
1280         err = -ENOENT;
1281         if (!fc->sb)
1282                 goto out_up_killsb;
1283
1284         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1285         if (!inode)
1286                 goto out_up_killsb;
1287
1288         mapping = inode->i_mapping;
1289         index = outarg.offset >> PAGE_CACHE_SHIFT;
1290         offset = outarg.offset & ~PAGE_CACHE_MASK;
1291         file_size = i_size_read(inode);
1292         end = outarg.offset + outarg.size;
1293         if (end > file_size) {
1294                 file_size = end;
1295                 fuse_write_update_size(inode, file_size);
1296         }
1297
1298         num = outarg.size;
1299         while (num) {
1300                 struct page *page;
1301                 unsigned int this_num;
1302
1303                 err = -ENOMEM;
1304                 page = find_or_create_page(mapping, index,
1305                                            mapping_gfp_mask(mapping));
1306                 if (!page)
1307                         goto out_iput;
1308
1309                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1310                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1311                 if (!err && offset == 0 && (num != 0 || file_size == end))
1312                         SetPageUptodate(page);
1313                 unlock_page(page);
1314                 page_cache_release(page);
1315
1316                 if (err)
1317                         goto out_iput;
1318
1319                 num -= this_num;
1320                 offset = 0;
1321                 index++;
1322         }
1323
1324         err = 0;
1325
1326 out_iput:
1327         iput(inode);
1328 out_up_killsb:
1329         up_read(&fc->killsb);
1330 out_finish:
1331         fuse_copy_finish(cs);
1332         return err;
1333 }
1334
1335 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1336 {
1337         release_pages(req->pages, req->num_pages, 0);
1338 }
1339
1340 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1341                          struct fuse_notify_retrieve_out *outarg)
1342 {
1343         int err;
1344         struct address_space *mapping = inode->i_mapping;
1345         struct fuse_req *req;
1346         pgoff_t index;
1347         loff_t file_size;
1348         unsigned int num;
1349         unsigned int offset;
1350         size_t total_len = 0;
1351
1352         req = fuse_get_req(fc);
1353         if (IS_ERR(req))
1354                 return PTR_ERR(req);
1355
1356         offset = outarg->offset & ~PAGE_CACHE_MASK;
1357
1358         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1359         req->in.h.nodeid = outarg->nodeid;
1360         req->in.numargs = 2;
1361         req->in.argpages = 1;
1362         req->page_offset = offset;
1363         req->end = fuse_retrieve_end;
1364
1365         index = outarg->offset >> PAGE_CACHE_SHIFT;
1366         file_size = i_size_read(inode);
1367         num = outarg->size;
1368         if (outarg->offset > file_size)
1369                 num = 0;
1370         else if (outarg->offset + num > file_size)
1371                 num = file_size - outarg->offset;
1372
1373         while (num) {
1374                 struct page *page;
1375                 unsigned int this_num;
1376
1377                 page = find_get_page(mapping, index);
1378                 if (!page)
1379                         break;
1380
1381                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1382                 req->pages[req->num_pages] = page;
1383                 req->num_pages++;
1384
1385                 num -= this_num;
1386                 total_len += this_num;
1387         }
1388         req->misc.retrieve_in.offset = outarg->offset;
1389         req->misc.retrieve_in.size = total_len;
1390         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1391         req->in.args[0].value = &req->misc.retrieve_in;
1392         req->in.args[1].size = total_len;
1393
1394         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1395         if (err)
1396                 fuse_retrieve_end(fc, req);
1397
1398         return err;
1399 }
1400
1401 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1402                                 struct fuse_copy_state *cs)
1403 {
1404         struct fuse_notify_retrieve_out outarg;
1405         struct inode *inode;
1406         int err;
1407
1408         err = -EINVAL;
1409         if (size != sizeof(outarg))
1410                 goto copy_finish;
1411
1412         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1413         if (err)
1414                 goto copy_finish;
1415
1416         fuse_copy_finish(cs);
1417
1418         down_read(&fc->killsb);
1419         err = -ENOENT;
1420         if (fc->sb) {
1421                 u64 nodeid = outarg.nodeid;
1422
1423                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1424                 if (inode) {
1425                         err = fuse_retrieve(fc, inode, &outarg);
1426                         iput(inode);
1427                 }
1428         }
1429         up_read(&fc->killsb);
1430
1431         return err;
1432
1433 copy_finish:
1434         fuse_copy_finish(cs);
1435         return err;
1436 }
1437
1438 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1439                        unsigned int size, struct fuse_copy_state *cs)
1440 {
1441         switch (code) {
1442         case FUSE_NOTIFY_POLL:
1443                 return fuse_notify_poll(fc, size, cs);
1444
1445         case FUSE_NOTIFY_INVAL_INODE:
1446                 return fuse_notify_inval_inode(fc, size, cs);
1447
1448         case FUSE_NOTIFY_INVAL_ENTRY:
1449                 return fuse_notify_inval_entry(fc, size, cs);
1450
1451         case FUSE_NOTIFY_STORE:
1452                 return fuse_notify_store(fc, size, cs);
1453
1454         case FUSE_NOTIFY_RETRIEVE:
1455                 return fuse_notify_retrieve(fc, size, cs);
1456
1457         default:
1458                 fuse_copy_finish(cs);
1459                 return -EINVAL;
1460         }
1461 }
1462
1463 /* Look up request on processing list by unique ID */
1464 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1465 {
1466         struct list_head *entry;
1467
1468         list_for_each(entry, &fc->processing) {
1469                 struct fuse_req *req;
1470                 req = list_entry(entry, struct fuse_req, list);
1471                 if (req->in.h.unique == unique || req->intr_unique == unique)
1472                         return req;
1473         }
1474         return NULL;
1475 }
1476
1477 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1478                          unsigned nbytes)
1479 {
1480         unsigned reqsize = sizeof(struct fuse_out_header);
1481
1482         if (out->h.error)
1483                 return nbytes != reqsize ? -EINVAL : 0;
1484
1485         reqsize += len_args(out->numargs, out->args);
1486
1487         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1488                 return -EINVAL;
1489         else if (reqsize > nbytes) {
1490                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1491                 unsigned diffsize = reqsize - nbytes;
1492                 if (diffsize > lastarg->size)
1493                         return -EINVAL;
1494                 lastarg->size -= diffsize;
1495         }
1496         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1497                               out->page_zeroing);
1498 }
1499
1500 /*
1501  * Write a single reply to a request.  First the header is copied from
1502  * the write buffer.  The request is then searched on the processing
1503  * list by the unique ID found in the header.  If found, then remove
1504  * it from the list and copy the rest of the buffer to the request.
1505  * The request is finished by calling request_end()
1506  */
1507 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1508                                  struct fuse_copy_state *cs, size_t nbytes)
1509 {
1510         int err;
1511         struct fuse_req *req;
1512         struct fuse_out_header oh;
1513
1514         if (nbytes < sizeof(struct fuse_out_header))
1515                 return -EINVAL;
1516
1517         err = fuse_copy_one(cs, &oh, sizeof(oh));
1518         if (err)
1519                 goto err_finish;
1520
1521         err = -EINVAL;
1522         if (oh.len != nbytes)
1523                 goto err_finish;
1524
1525         /*
1526          * Zero oh.unique indicates unsolicited notification message
1527          * and error contains notification code.
1528          */
1529         if (!oh.unique) {
1530                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1531                 return err ? err : nbytes;
1532         }
1533
1534         err = -EINVAL;
1535         if (oh.error <= -1000 || oh.error > 0)
1536                 goto err_finish;
1537
1538         spin_lock(&fc->lock);
1539         err = -ENOENT;
1540         if (!fc->connected)
1541                 goto err_unlock;
1542
1543         req = request_find(fc, oh.unique);
1544         if (!req)
1545                 goto err_unlock;
1546
1547         if (req->aborted) {
1548                 spin_unlock(&fc->lock);
1549                 fuse_copy_finish(cs);
1550                 spin_lock(&fc->lock);
1551                 request_end(fc, req);
1552                 return -ENOENT;
1553         }
1554         /* Is it an interrupt reply? */
1555         if (req->intr_unique == oh.unique) {
1556                 err = -EINVAL;
1557                 if (nbytes != sizeof(struct fuse_out_header))
1558                         goto err_unlock;
1559
1560                 if (oh.error == -ENOSYS)
1561                         fc->no_interrupt = 1;
1562                 else if (oh.error == -EAGAIN)
1563                         queue_interrupt(fc, req);
1564
1565                 spin_unlock(&fc->lock);
1566                 fuse_copy_finish(cs);
1567                 return nbytes;
1568         }
1569
1570         req->state = FUSE_REQ_WRITING;
1571         list_move(&req->list, &fc->io);
1572         req->out.h = oh;
1573         req->locked = 1;
1574         cs->req = req;
1575         if (!req->out.page_replace)
1576                 cs->move_pages = 0;
1577         spin_unlock(&fc->lock);
1578
1579         err = copy_out_args(cs, &req->out, nbytes);
1580         fuse_copy_finish(cs);
1581
1582         spin_lock(&fc->lock);
1583         req->locked = 0;
1584         if (!err) {
1585                 if (req->aborted)
1586                         err = -ENOENT;
1587         } else if (!req->aborted)
1588                 req->out.h.error = -EIO;
1589         request_end(fc, req);
1590
1591         return err ? err : nbytes;
1592
1593  err_unlock:
1594         spin_unlock(&fc->lock);
1595  err_finish:
1596         fuse_copy_finish(cs);
1597         return err;
1598 }
1599
1600 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
1601                               unsigned long nr_segs, loff_t pos)
1602 {
1603         struct fuse_copy_state cs;
1604         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1605         if (!fc)
1606                 return -EPERM;
1607
1608         fuse_copy_init(&cs, fc, 0, iov, nr_segs);
1609
1610         return fuse_dev_do_write(fc, &cs, iov_length(iov, nr_segs));
1611 }
1612
1613 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1614                                      struct file *out, loff_t *ppos,
1615                                      size_t len, unsigned int flags)
1616 {
1617         unsigned nbuf;
1618         unsigned idx;
1619         struct pipe_buffer *bufs;
1620         struct fuse_copy_state cs;
1621         struct fuse_conn *fc;
1622         size_t rem;
1623         ssize_t ret;
1624
1625         fc = fuse_get_conn(out);
1626         if (!fc)
1627                 return -EPERM;
1628
1629         bufs = kmalloc(pipe->buffers * sizeof (struct pipe_buffer), GFP_KERNEL);
1630         if (!bufs)
1631                 return -ENOMEM;
1632
1633         pipe_lock(pipe);
1634         nbuf = 0;
1635         rem = 0;
1636         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1637                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1638
1639         ret = -EINVAL;
1640         if (rem < len) {
1641                 pipe_unlock(pipe);
1642                 goto out;
1643         }
1644
1645         rem = len;
1646         while (rem) {
1647                 struct pipe_buffer *ibuf;
1648                 struct pipe_buffer *obuf;
1649
1650                 BUG_ON(nbuf >= pipe->buffers);
1651                 BUG_ON(!pipe->nrbufs);
1652                 ibuf = &pipe->bufs[pipe->curbuf];
1653                 obuf = &bufs[nbuf];
1654
1655                 if (rem >= ibuf->len) {
1656                         *obuf = *ibuf;
1657                         ibuf->ops = NULL;
1658                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1659                         pipe->nrbufs--;
1660                 } else {
1661                         ibuf->ops->get(pipe, ibuf);
1662                         *obuf = *ibuf;
1663                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1664                         obuf->len = rem;
1665                         ibuf->offset += obuf->len;
1666                         ibuf->len -= obuf->len;
1667                 }
1668                 nbuf++;
1669                 rem -= obuf->len;
1670         }
1671         pipe_unlock(pipe);
1672
1673         fuse_copy_init(&cs, fc, 0, NULL, nbuf);
1674         cs.pipebufs = bufs;
1675         cs.pipe = pipe;
1676
1677         if (flags & SPLICE_F_MOVE)
1678                 cs.move_pages = 1;
1679
1680         ret = fuse_dev_do_write(fc, &cs, len);
1681
1682         for (idx = 0; idx < nbuf; idx++) {
1683                 struct pipe_buffer *buf = &bufs[idx];
1684                 buf->ops->release(pipe, buf);
1685         }
1686 out:
1687         kfree(bufs);
1688         return ret;
1689 }
1690
1691 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1692 {
1693         unsigned mask = POLLOUT | POLLWRNORM;
1694         struct fuse_conn *fc = fuse_get_conn(file);
1695         if (!fc)
1696                 return POLLERR;
1697
1698         poll_wait(file, &fc->waitq, wait);
1699
1700         spin_lock(&fc->lock);
1701         if (!fc->connected)
1702                 mask = POLLERR;
1703         else if (request_pending(fc))
1704                 mask |= POLLIN | POLLRDNORM;
1705         spin_unlock(&fc->lock);
1706
1707         return mask;
1708 }
1709
1710 /*
1711  * Abort all requests on the given list (pending or processing)
1712  *
1713  * This function releases and reacquires fc->lock
1714  */
1715 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1716 __releases(fc->lock)
1717 __acquires(fc->lock)
1718 {
1719         while (!list_empty(head)) {
1720                 struct fuse_req *req;
1721                 req = list_entry(head->next, struct fuse_req, list);
1722                 req->out.h.error = -ECONNABORTED;
1723                 request_end(fc, req);
1724                 spin_lock(&fc->lock);
1725         }
1726 }
1727
1728 /*
1729  * Abort requests under I/O
1730  *
1731  * The requests are set to aborted and finished, and the request
1732  * waiter is woken up.  This will make request_wait_answer() wait
1733  * until the request is unlocked and then return.
1734  *
1735  * If the request is asynchronous, then the end function needs to be
1736  * called after waiting for the request to be unlocked (if it was
1737  * locked).
1738  */
1739 static void end_io_requests(struct fuse_conn *fc)
1740 __releases(fc->lock)
1741 __acquires(fc->lock)
1742 {
1743         while (!list_empty(&fc->io)) {
1744                 struct fuse_req *req =
1745                         list_entry(fc->io.next, struct fuse_req, list);
1746                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1747
1748                 req->aborted = 1;
1749                 req->out.h.error = -ECONNABORTED;
1750                 req->state = FUSE_REQ_FINISHED;
1751                 list_del_init(&req->list);
1752                 wake_up(&req->waitq);
1753                 if (end) {
1754                         req->end = NULL;
1755                         __fuse_get_request(req);
1756                         spin_unlock(&fc->lock);
1757                         wait_event(req->waitq, !req->locked);
1758                         end(fc, req);
1759                         fuse_put_request(fc, req);
1760                         spin_lock(&fc->lock);
1761                 }
1762         }
1763 }
1764
1765 static void end_queued_requests(struct fuse_conn *fc)
1766 __releases(fc->lock)
1767 __acquires(fc->lock)
1768 {
1769         fc->max_background = UINT_MAX;
1770         flush_bg_queue(fc);
1771         end_requests(fc, &fc->pending);
1772         end_requests(fc, &fc->processing);
1773 }
1774
1775 /*
1776  * Abort all requests.
1777  *
1778  * Emergency exit in case of a malicious or accidental deadlock, or
1779  * just a hung filesystem.
1780  *
1781  * The same effect is usually achievable through killing the
1782  * filesystem daemon and all users of the filesystem.  The exception
1783  * is the combination of an asynchronous request and the tricky
1784  * deadlock (see Documentation/filesystems/fuse.txt).
1785  *
1786  * During the aborting, progression of requests from the pending and
1787  * processing lists onto the io list, and progression of new requests
1788  * onto the pending list is prevented by req->connected being false.
1789  *
1790  * Progression of requests under I/O to the processing list is
1791  * prevented by the req->aborted flag being true for these requests.
1792  * For this reason requests on the io list must be aborted first.
1793  */
1794 void fuse_abort_conn(struct fuse_conn *fc)
1795 {
1796         spin_lock(&fc->lock);
1797         if (fc->connected) {
1798                 fc->connected = 0;
1799                 fc->blocked = 0;
1800                 end_io_requests(fc);
1801                 end_queued_requests(fc);
1802                 wake_up_all(&fc->waitq);
1803                 wake_up_all(&fc->blocked_waitq);
1804                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1805         }
1806         spin_unlock(&fc->lock);
1807 }
1808 EXPORT_SYMBOL_GPL(fuse_abort_conn);
1809
1810 int fuse_dev_release(struct inode *inode, struct file *file)
1811 {
1812         struct fuse_conn *fc = fuse_get_conn(file);
1813         if (fc) {
1814                 spin_lock(&fc->lock);
1815                 fc->connected = 0;
1816                 fc->blocked = 0;
1817                 end_queued_requests(fc);
1818                 wake_up_all(&fc->blocked_waitq);
1819                 spin_unlock(&fc->lock);
1820                 fuse_conn_put(fc);
1821         }
1822
1823         return 0;
1824 }
1825 EXPORT_SYMBOL_GPL(fuse_dev_release);
1826
1827 static int fuse_dev_fasync(int fd, struct file *file, int on)
1828 {
1829         struct fuse_conn *fc = fuse_get_conn(file);
1830         if (!fc)
1831                 return -EPERM;
1832
1833         /* No locking - fasync_helper does its own locking */
1834         return fasync_helper(fd, file, on, &fc->fasync);
1835 }
1836
1837 const struct file_operations fuse_dev_operations = {
1838         .owner          = THIS_MODULE,
1839         .llseek         = no_llseek,
1840         .read           = do_sync_read,
1841         .aio_read       = fuse_dev_read,
1842         .splice_read    = fuse_dev_splice_read,
1843         .write          = do_sync_write,
1844         .aio_write      = fuse_dev_write,
1845         .splice_write   = fuse_dev_splice_write,
1846         .poll           = fuse_dev_poll,
1847         .release        = fuse_dev_release,
1848         .fasync         = fuse_dev_fasync,
1849 };
1850 EXPORT_SYMBOL_GPL(fuse_dev_operations);
1851
1852 static struct miscdevice fuse_miscdevice = {
1853         .minor = FUSE_MINOR,
1854         .name  = "fuse",
1855         .fops = &fuse_dev_operations,
1856 };
1857
1858 int __init fuse_dev_init(void)
1859 {
1860         int err = -ENOMEM;
1861         fuse_req_cachep = kmem_cache_create("fuse_request",
1862                                             sizeof(struct fuse_req),
1863                                             0, 0, NULL);
1864         if (!fuse_req_cachep)
1865                 goto out;
1866
1867         err = misc_register(&fuse_miscdevice);
1868         if (err)
1869                 goto out_cache_clean;
1870
1871         return 0;
1872
1873  out_cache_clean:
1874         kmem_cache_destroy(fuse_req_cachep);
1875  out:
1876         return err;
1877 }
1878
1879 void fuse_dev_cleanup(void)
1880 {
1881         misc_deregister(&fuse_miscdevice);
1882         kmem_cache_destroy(fuse_req_cachep);
1883 }