fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/pipe_fs_i.h>
  20 #include <linux/swap.h>
  21 #include <linux/splice.h>
  22 #include <linux/freezer.h>
  23
  24 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  25 MODULE_ALIAS("devname:fuse");
  26
  27 static struct kmem_cache *fuse_req_cachep;
  28
  29 static struct fuse_conn *fuse_get_conn(struct file *file)
  30 {
  31         /*
  32          * Lockless access is OK, because file->private data is set
  33          * once during mount and is valid until the file is released.
  34          */
  35         return file->private_data;
  36 }
  37
  38 static void fuse_request_init(struct fuse_req *req)
  39 {
  40         memset(req, 0, sizeof(*req));
  41         INIT_LIST_HEAD(&req->list);
  42         INIT_LIST_HEAD(&req->intr_entry);
  43         init_waitqueue_head(&req->waitq);
  44         atomic_set(&req->count, 1);
  45 }
  46
  47 struct fuse_req *fuse_request_alloc(void)
  48 {
  49         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_KERNEL);
  50         if (req)
  51                 fuse_request_init(req);
  52         return req;
  53 }
  54 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  55
  56 struct fuse_req *fuse_request_alloc_nofs(void)
  57 {
  58         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_NOFS);
  59         if (req)
  60                 fuse_request_init(req);
  61         return req;
  62 }
  63
  64 void fuse_request_free(struct fuse_req *req)
  65 {
  66         kmem_cache_free(fuse_req_cachep, req);
  67 }
  68
  69 static void block_sigs(sigset_t *oldset)
  70 {
  71         sigset_t mask;
  72
  73         siginitsetinv(&mask, sigmask(SIGKILL));
  74         sigprocmask(SIG_BLOCK, &mask, oldset);
  75 }
  76
  77 static void restore_sigs(sigset_t *oldset)
  78 {
  79         sigprocmask(SIG_SETMASK, oldset, NULL);
  80 }
  81
  82 static void __fuse_get_request(struct fuse_req *req)
  83 {
  84         atomic_inc(&req->count);
  85 }
  86
  87 /* Must be called with > 1 refcount */
  88 static void __fuse_put_request(struct fuse_req *req)
  89 {
  90         BUG_ON(atomic_read(&req->count) < 2);
  91         atomic_dec(&req->count);
  92 }
  93
  94 static void fuse_req_init_context(struct fuse_req *req)
  95 {
  96         req->in.h.uid = current_fsuid();
  97         req->in.h.gid = current_fsgid();
  98         req->in.h.pid = current->pid;
  99 }
 100
 101 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
 102 {
 103         struct fuse_req *req;
 104         sigset_t oldset;
 105         int intr;
 106         int err;
 107
 108         atomic_inc(&fc->num_waiting);
 109         block_sigs(&oldset);
 110         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
 111         restore_sigs(&oldset);
 112         err = -EINTR;
 113         if (intr)
 114                 goto out;
 115
 116         err = -ENOTCONN;
 117         if (!fc->connected)
 118                 goto out;
 119
 120         req = fuse_request_alloc();
 121         err = -ENOMEM;
 122         if (!req)
 123                 goto out;
 124
 125         fuse_req_init_context(req);
 126         req->waiting = 1;
 127         return req;
 128
 129  out:
 130         atomic_dec(&fc->num_waiting);
 131         return ERR_PTR(err);
 132 }
 133 EXPORT_SYMBOL_GPL(fuse_get_req);
 134
 135 /*
 136  * Return request in fuse_file->reserved_req.  However that may
 137  * currently be in use.  If that is the case, wait for it to become
 138  * available.
 139  */
 140 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 141                                          struct file *file)
 142 {
 143         struct fuse_req *req = NULL;
 144         struct fuse_file *ff = file->private_data;
 145
 146         do {
 147                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 148                 spin_lock(&fc->lock);
 149                 if (ff->reserved_req) {
 150                         req = ff->reserved_req;
 151                         ff->reserved_req = NULL;
 152                         get_file(file);
 153                         req->stolen_file = file;
 154                 }
 155                 spin_unlock(&fc->lock);
 156         } while (!req);
 157
 158         return req;
 159 }
 160
 161 /*
 162  * Put stolen request back into fuse_file->reserved_req
 163  */
 164 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 165 {
 166         struct file *file = req->stolen_file;
 167         struct fuse_file *ff = file->private_data;
 168
 169         spin_lock(&fc->lock);
 170         fuse_request_init(req);
 171         BUG_ON(ff->reserved_req);
 172         ff->reserved_req = req;
 173         wake_up_all(&fc->reserved_req_waitq);
 174         spin_unlock(&fc->lock);
 175         fput(file);
 176 }
 177
 178 /*
 179  * Gets a requests for a file operation, always succeeds
 180  *
 181  * This is used for sending the FLUSH request, which must get to
 182  * userspace, due to POSIX locks which may need to be unlocked.
 183  *
 184  * If allocation fails due to OOM, use the reserved request in
 185  * fuse_file.
 186  *
 187  * This is very unlikely to deadlock accidentally, since the
 188  * filesystem should not have it's own file open.  If deadlock is
 189  * intentional, it can still be broken by "aborting" the filesystem.
 190  */
 191 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 192 {
 193         struct fuse_req *req;
 194
 195         atomic_inc(&fc->num_waiting);
 196         wait_event(fc->blocked_waitq, !fc->blocked);
 197         req = fuse_request_alloc();
 198         if (!req)
 199                 req = get_reserved_req(fc, file);
 200
 201         fuse_req_init_context(req);
 202         req->waiting = 1;
 203         return req;
 204 }
 205
 206 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 207 {
 208         if (atomic_dec_and_test(&req->count)) {
 209                 if (req->waiting)
 210                         atomic_dec(&fc->num_waiting);
 211
 212                 if (req->stolen_file)
 213                         put_reserved_req(fc, req);
 214                 else
 215                         fuse_request_free(req);
 216         }
 217 }
 218 EXPORT_SYMBOL_GPL(fuse_put_request);
 219
 220 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 221 {
 222         unsigned nbytes = 0;
 223         unsigned i;
 224
 225         for (i = 0; i < numargs; i++)
 226                 nbytes += args[i].size;
 227
 228         return nbytes;
 229 }
 230
 231 static u64 fuse_get_unique(struct fuse_conn *fc)
 232 {
 233         fc->reqctr++;
 234         /* zero is special */
 235         if (fc->reqctr == 0)
 236                 fc->reqctr = 1;
 237
 238         return fc->reqctr;
 239 }
 240
 241 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 242 {
 243         req->in.h.len = sizeof(struct fuse_in_header) +
 244                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 245         list_add_tail(&req->list, &fc->pending);
 246         req->state = FUSE_REQ_PENDING;
 247         if (!req->waiting) {
 248                 req->waiting = 1;
 249                 atomic_inc(&fc->num_waiting);
 250         }
 251         wake_up(&fc->waitq);
 252         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 253 }
 254
 255 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 256                        u64 nodeid, u64 nlookup)
 257 {
 258         forget->forget_one.nodeid = nodeid;
 259         forget->forget_one.nlookup = nlookup;
 260
 261         spin_lock(&fc->lock);
 262         if (fc->connected) {
 263                 fc->forget_list_tail->next = forget;
 264                 fc->forget_list_tail = forget;
 265                 wake_up(&fc->waitq);
 266                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 267         } else {
 268                 kfree(forget);
 269         }
 270         spin_unlock(&fc->lock);
 271 }
 272
 273 static void flush_bg_queue(struct fuse_conn *fc)
 274 {
 275         while (fc->active_background < fc->max_background &&
 276                !list_empty(&fc->bg_queue)) {
 277                 struct fuse_req *req;
 278
 279                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 280                 list_del(&req->list);
 281                 fc->active_background++;
 282                 req->in.h.unique = fuse_get_unique(fc);
 283                 queue_request(fc, req);
 284         }
 285 }
 286
 287 /*
 288  * This function is called when a request is finished.  Either a reply
 289  * has arrived or it was aborted (and not yet sent) or some error
 290  * occurred during communication with userspace, or the device file
 291  * was closed.  The requester thread is woken up (if still waiting),
 292  * the 'end' callback is called if given, else the reference to the
 293  * request is released
 294  *
 295  * Called with fc->lock, unlocks it
 296  */
 297 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 298 __releases(fc->lock)
 299 {
 300         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 301         req->end = NULL;
 302         list_del(&req->list);
 303         list_del(&req->intr_entry);
 304         req->state = FUSE_REQ_FINISHED;
 305         if (req->background) {
 306                 if (fc->num_background == fc->max_background) {
 307                         fc->blocked = 0;
 308                         wake_up_all(&fc->blocked_waitq);
 309                 }
 310                 if (fc->num_background == fc->congestion_threshold &&
 311                     fc->connected && fc->bdi_initialized) {
 312                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 313                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 314                 }
 315                 fc->num_background--;
 316                 fc->active_background--;
 317                 flush_bg_queue(fc);
 318         }
 319         spin_unlock(&fc->lock);
 320         wake_up(&req->waitq);
 321         if (end)
 322                 end(fc, req);
 323         fuse_put_request(fc, req);
 324 }
 325
 326 static void wait_answer_interruptible(struct fuse_conn *fc,
 327                                       struct fuse_req *req)
 328 __releases(fc->lock)
 329 __acquires(fc->lock)
 330 {
 331         if (signal_pending(current))
 332                 return;
 333
 334         spin_unlock(&fc->lock);
 335         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 336         spin_lock(&fc->lock);
 337 }
 338
 339 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 340 {
 341         list_add_tail(&req->intr_entry, &fc->interrupts);
 342         wake_up(&fc->waitq);
 343         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 344 }
 345
 346 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 347 __releases(fc->lock)
 348 __acquires(fc->lock)
 349 {
 350         if (!fc->no_interrupt) {
 351                 /* Any signal may interrupt this */
 352                 wait_answer_interruptible(fc, req);
 353
 354                 if (req->aborted)
 355                         goto aborted;
 356                 if (req->state == FUSE_REQ_FINISHED)
 357                         return;
 358
 359                 req->interrupted = 1;
 360                 if (req->state == FUSE_REQ_SENT)
 361                         queue_interrupt(fc, req);
 362         }
 363
 364         if (!req->force) {
 365                 sigset_t oldset;
 366
 367                 /* Only fatal signals may interrupt this */
 368                 block_sigs(&oldset);
 369                 wait_answer_interruptible(fc, req);
 370                 restore_sigs(&oldset);
 371
 372                 if (req->aborted)
 373                         goto aborted;
 374                 if (req->state == FUSE_REQ_FINISHED)
 375                         return;
 376
 377                 /* Request is not yet in userspace, bail out */
 378                 if (req->state == FUSE_REQ_PENDING) {
 379                         list_del(&req->list);
 380                         __fuse_put_request(req);
 381                         req->out.h.error = -EINTR;
 382                         return;
 383                 }
 384         }
 385
 386         /*
 387          * Either request is already in userspace, or it was forced.
 388          * Wait it out.
 389          */
 390         spin_unlock(&fc->lock);
 391
 392         while (req->state != FUSE_REQ_FINISHED)
 393                 wait_event_freezable(req->waitq,
 394                                      req->state == FUSE_REQ_FINISHED);
 395         spin_lock(&fc->lock);
 396
 397         if (!req->aborted)
 398                 return;
 399
 400  aborted:
 401         BUG_ON(req->state != FUSE_REQ_FINISHED);
 402         if (req->locked) {
 403                 /* This is uninterruptible sleep, because data is
 404                    being copied to/from the buffers of req.  During
 405                    locked state, there mustn't be any filesystem
 406                    operation (e.g. page fault), since that could lead
 407                    to deadlock */
 408                 spin_unlock(&fc->lock);
 409                 wait_event(req->waitq, !req->locked);
 410                 spin_lock(&fc->lock);
 411         }
 412 }
 413
 414 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 415 {
 416         req->isreply = 1;
 417         spin_lock(&fc->lock);
 418         if (!fc->connected)
 419                 req->out.h.error = -ENOTCONN;
 420         else if (fc->conn_error)
 421                 req->out.h.error = -ECONNREFUSED;
 422         else {
 423                 req->in.h.unique = fuse_get_unique(fc);
 424                 queue_request(fc, req);
 425                 /* acquire extra reference, since request is still needed
 426                    after request_end() */
 427                 __fuse_get_request(req);
 428
 429                 request_wait_answer(fc, req);
 430         }
 431         spin_unlock(&fc->lock);
 432 }
 433 EXPORT_SYMBOL_GPL(fuse_request_send);
 434
 435 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 436                                             struct fuse_req *req)
 437 {
 438         req->background = 1;
 439         fc->num_background++;
 440         if (fc->num_background == fc->max_background)
 441                 fc->blocked = 1;
 442         if (fc->num_background == fc->congestion_threshold &&
 443             fc->bdi_initialized) {
 444                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 445                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 446         }
 447         list_add_tail(&req->list, &fc->bg_queue);
 448         flush_bg_queue(fc);
 449 }
 450
 451 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 452 {
 453         spin_lock(&fc->lock);
 454         if (fc->connected) {
 455                 fuse_request_send_nowait_locked(fc, req);
 456                 spin_unlock(&fc->lock);
 457         } else {
 458                 req->out.h.error = -ENOTCONN;
 459                 request_end(fc, req);
 460         }
 461 }
 462
 463 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 464 {
 465         req->isreply = 1;
 466         fuse_request_send_nowait(fc, req);
 467 }
 468 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 469
 470 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 471                                           struct fuse_req *req, u64 unique)
 472 {
 473         int err = -ENODEV;
 474
 475         req->isreply = 0;
 476         req->in.h.unique = unique;
 477         spin_lock(&fc->lock);
 478         if (fc->connected) {
 479                 queue_request(fc, req);
 480                 err = 0;
 481         }
 482         spin_unlock(&fc->lock);
 483
 484         return err;
 485 }
 486
 487 /*
 488  * Called under fc->lock
 489  *
 490  * fc->connected must have been checked previously
 491  */
 492 void fuse_request_send_background_locked(struct fuse_conn *fc,
 493                                          struct fuse_req *req)
 494 {
 495         req->isreply = 1;
 496         fuse_request_send_nowait_locked(fc, req);
 497 }
 498
 499 /*
 500  * Lock the request.  Up to the next unlock_request() there mustn't be
 501  * anything that could cause a page-fault.  If the request was already
 502  * aborted bail out.
 503  */
 504 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 505 {
 506         int err = 0;
 507         if (req) {
 508                 spin_lock(&fc->lock);
 509                 if (req->aborted)
 510                         err = -ENOENT;
 511                 else
 512                         req->locked = 1;
 513                 spin_unlock(&fc->lock);
 514         }
 515         return err;
 516 }
 517
 518 /*
 519  * Unlock request.  If it was aborted during being locked, the
 520  * requester thread is currently waiting for it to be unlocked, so
 521  * wake it up.
 522  */
 523 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 524 {
 525         if (req) {
 526                 spin_lock(&fc->lock);
 527                 req->locked = 0;
 528                 if (req->aborted)
 529                         wake_up(&req->waitq);
 530                 spin_unlock(&fc->lock);
 531         }
 532 }
 533
 534 struct fuse_copy_state {
 535         struct fuse_conn *fc;
 536         int write;
 537         struct fuse_req *req;
 538         const struct iovec *iov;
 539         struct pipe_buffer *pipebufs;
 540         struct pipe_buffer *currbuf;
 541         struct pipe_inode_info *pipe;
 542         unsigned long nr_segs;
 543         unsigned long seglen;
 544         unsigned long addr;
 545         struct page *pg;
 546         void *mapaddr;
 547         void *buf;
 548         unsigned len;
 549         unsigned move_pages:1;
 550 };
 551
 552 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 553                            int write,
 554                            const struct iovec *iov, unsigned long nr_segs)
 555 {
 556         memset(cs, 0, sizeof(*cs));
 557         cs->fc = fc;
 558         cs->write = write;
 559         cs->iov = iov;
 560         cs->nr_segs = nr_segs;
 561 }
 562
 563 /* Unmap and put previous page of userspace buffer */
 564 static void fuse_copy_finish(struct fuse_copy_state *cs)
 565 {
 566         if (cs->currbuf) {
 567                 struct pipe_buffer *buf = cs->currbuf;
 568
 569                 if (!cs->write) {
 570                         buf->ops->unmap(cs->pipe, buf, cs->mapaddr);
 571                 } else {
 572                         kunmap(buf->page);
 573                         buf->len = PAGE_SIZE - cs->len;
 574                 }
 575                 cs->currbuf = NULL;
 576                 cs->mapaddr = NULL;
 577         } else if (cs->mapaddr) {
 578                 kunmap(cs->pg);
 579                 if (cs->write) {
 580                         flush_dcache_page(cs->pg);
 581                         set_page_dirty_lock(cs->pg);
 582                 }
 583                 put_page(cs->pg);
 584                 cs->mapaddr = NULL;
 585         }
 586 }
 587
 588 /*
 589  * Get another pagefull of userspace buffer, and map it to kernel
 590  * address space, and lock request
 591  */
 592 static int fuse_copy_fill(struct fuse_copy_state *cs)
 593 {
 594         unsigned long offset;
 595         int err;
 596
 597         unlock_request(cs->fc, cs->req);
 598         fuse_copy_finish(cs);
 599         if (cs->pipebufs) {
 600                 struct pipe_buffer *buf = cs->pipebufs;
 601
 602                 if (!cs->write) {
 603                         err = buf->ops->confirm(cs->pipe, buf);
 604                         if (err)
 605                                 return err;
 606
 607                         BUG_ON(!cs->nr_segs);
 608                         cs->currbuf = buf;
 609                         cs->mapaddr = buf->ops->map(cs->pipe, buf, 0);
 610                         cs->len = buf->len;
 611                         cs->buf = cs->mapaddr + buf->offset;
 612                         cs->pipebufs++;
 613                         cs->nr_segs--;
 614                 } else {
 615                         struct page *page;
 616
 617                         if (cs->nr_segs == cs->pipe->buffers)
 618                                 return -EIO;
 619
 620                         page = alloc_page(GFP_HIGHUSER);
 621                         if (!page)
 622                                 return -ENOMEM;
 623
 624                         buf->page = page;
 625                         buf->offset = 0;
 626                         buf->len = 0;
 627
 628                         cs->currbuf = buf;
 629                         cs->mapaddr = kmap(page);
 630                         cs->buf = cs->mapaddr;
 631                         cs->len = PAGE_SIZE;
 632                         cs->pipebufs++;
 633                         cs->nr_segs++;
 634                 }
 635         } else {
 636                 if (!cs->seglen) {
 637                         BUG_ON(!cs->nr_segs);
 638                         cs->seglen = cs->iov[0].iov_len;
 639                         cs->addr = (unsigned long) cs->iov[0].iov_base;
 640                         cs->iov++;
 641                         cs->nr_segs--;
 642                 }
 643                 err = get_user_pages_fast(cs->addr, 1, cs->write, &cs->pg);
 644                 if (err < 0)
 645                         return err;
 646                 BUG_ON(err != 1);
 647                 offset = cs->addr % PAGE_SIZE;
 648                 cs->mapaddr = kmap(cs->pg);
 649                 cs->buf = cs->mapaddr + offset;
 650                 cs->len = min(PAGE_SIZE - offset, cs->seglen);
 651                 cs->seglen -= cs->len;
 652                 cs->addr += cs->len;
 653         }
 654
 655         return lock_request(cs->fc, cs->req);
 656 }
 657
 658 /* Do as much copy to/from userspace buffer as we can */
 659 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 660 {
 661         unsigned ncpy = min(*size, cs->len);
 662         if (val) {
 663                 if (cs->write)
 664                         memcpy(cs->buf, *val, ncpy);
 665                 else
 666                         memcpy(*val, cs->buf, ncpy);
 667                 *val += ncpy;
 668         }
 669         *size -= ncpy;
 670         cs->len -= ncpy;
 671         cs->buf += ncpy;
 672         return ncpy;
 673 }
 674
 675 static int fuse_check_page(struct page *page)
 676 {
 677         if (page_mapcount(page) ||
 678             page->mapping != NULL ||
 679             page_count(page) != 1 ||
 680             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 681              ~(1 << PG_locked |
 682                1 << PG_referenced |
 683                1 << PG_uptodate |
 684                1 << PG_lru |
 685                1 << PG_active |
 686                1 << PG_reclaim))) {
 687                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 688                 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);
 689                 return 1;
 690         }
 691         return 0;
 692 }
 693
 694 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 695 {
 696         int err;
 697         struct page *oldpage = *pagep;
 698         struct page *newpage;
 699         struct pipe_buffer *buf = cs->pipebufs;
 700         struct address_space *mapping;
 701         pgoff_t index;
 702
 703         unlock_request(cs->fc, cs->req);
 704         fuse_copy_finish(cs);
 705
 706         err = buf->ops->confirm(cs->pipe, buf);
 707         if (err)
 708                 return err;
 709
 710         BUG_ON(!cs->nr_segs);
 711         cs->currbuf = buf;
 712         cs->len = buf->len;
 713         cs->pipebufs++;
 714         cs->nr_segs--;
 715
 716         if (cs->len != PAGE_SIZE)
 717                 goto out_fallback;
 718
 719         if (buf->ops->steal(cs->pipe, buf) != 0)
 720                 goto out_fallback;
 721
 722         newpage = buf->page;
 723
 724         if (WARN_ON(!PageUptodate(newpage)))
 725                 return -EIO;
 726
 727         ClearPageMappedToDisk(newpage);
 728
 729         if (fuse_check_page(newpage) != 0)
 730                 goto out_fallback_unlock;
 731
 732         mapping = oldpage->mapping;
 733         index = oldpage->index;
 734
 735         /*
 736          * This is a new and locked page, it shouldn't be mapped or
 737          * have any special flags on it
 738          */
 739         if (WARN_ON(page_mapped(oldpage)))
 740                 goto out_fallback_unlock;
 741         if (WARN_ON(page_has_private(oldpage)))
 742                 goto out_fallback_unlock;
 743         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 744                 goto out_fallback_unlock;
 745         if (WARN_ON(PageMlocked(oldpage)))
 746                 goto out_fallback_unlock;
 747
 748         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 749         if (err) {
 750                 unlock_page(newpage);
 751                 return err;
 752         }
 753
 754         page_cache_get(newpage);
 755
 756         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 757                 lru_cache_add_file(newpage);
 758
 759         err = 0;
 760         spin_lock(&cs->fc->lock);
 761         if (cs->req->aborted)
 762                 err = -ENOENT;
 763         else
 764                 *pagep = newpage;
 765         spin_unlock(&cs->fc->lock);
 766
 767         if (err) {
 768                 unlock_page(newpage);
 769                 page_cache_release(newpage);
 770                 return err;
 771         }
 772
 773         unlock_page(oldpage);
 774         page_cache_release(oldpage);
 775         cs->len = 0;
 776
 777         return 0;
 778
 779 out_fallback_unlock:
 780         unlock_page(newpage);
 781 out_fallback:
 782         cs->mapaddr = buf->ops->map(cs->pipe, buf, 1);
 783         cs->buf = cs->mapaddr + buf->offset;
 784
 785         err = lock_request(cs->fc, cs->req);
 786         if (err)
 787                 return err;
 788
 789         return 1;
 790 }
 791
 792 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 793                          unsigned offset, unsigned count)
 794 {
 795         struct pipe_buffer *buf;
 796
 797         if (cs->nr_segs == cs->pipe->buffers)
 798                 return -EIO;
 799
 800         unlock_request(cs->fc, cs->req);
 801         fuse_copy_finish(cs);
 802
 803         buf = cs->pipebufs;
 804         page_cache_get(page);
 805         buf->page = page;
 806         buf->offset = offset;
 807         buf->len = count;
 808
 809         cs->pipebufs++;
 810         cs->nr_segs++;
 811         cs->len = 0;
 812
 813         return 0;
 814 }
 815
 816 /*
 817  * Copy a page in the request to/from the userspace buffer.  Must be
 818  * done atomically
 819  */
 820 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 821                           unsigned offset, unsigned count, int zeroing)
 822 {
 823         int err;
 824         struct page *page = *pagep;
 825
 826         if (page && zeroing && count < PAGE_SIZE)
 827                 clear_highpage(page);
 828
 829         while (count) {
 830                 if (cs->write && cs->pipebufs && page) {
 831                         return fuse_ref_page(cs, page, offset, count);
 832                 } else if (!cs->len) {
 833                         if (cs->move_pages && page &&
 834                             offset == 0 && count == PAGE_SIZE) {
 835                                 err = fuse_try_move_page(cs, pagep);
 836                                 if (err <= 0)
 837                                         return err;
 838                         } else {
 839                                 err = fuse_copy_fill(cs);
 840                                 if (err)
 841                                         return err;
 842                         }
 843                 }
 844                 if (page) {
 845                         void *mapaddr = kmap_atomic(page, KM_USER0);
 846                         void *buf = mapaddr + offset;
 847                         offset += fuse_copy_do(cs, &buf, &count);
 848                         kunmap_atomic(mapaddr, KM_USER0);
 849                 } else
 850                         offset += fuse_copy_do(cs, NULL, &count);
 851         }
 852         if (page && !cs->write)
 853                 flush_dcache_page(page);
 854         return 0;
 855 }
 856
 857 /* Copy pages in the request to/from userspace buffer */
 858 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 859                            int zeroing)
 860 {
 861         unsigned i;
 862         struct fuse_req *req = cs->req;
 863         unsigned offset = req->page_offset;
 864         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 865
 866         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 867                 int err;
 868
 869                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
 870                                      zeroing);
 871                 if (err)
 872                         return err;
 873
 874                 nbytes -= count;
 875                 count = min(nbytes, (unsigned) PAGE_SIZE);
 876                 offset = 0;
 877         }
 878         return 0;
 879 }
 880
 881 /* Copy a single argument in the request to/from userspace buffer */
 882 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 883 {
 884         while (size) {
 885                 if (!cs->len) {
 886                         int err = fuse_copy_fill(cs);
 887                         if (err)
 888                                 return err;
 889                 }
 890                 fuse_copy_do(cs, &val, &size);
 891         }
 892         return 0;
 893 }
 894
 895 /* Copy request arguments to/from userspace buffer */
 896 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 897                           unsigned argpages, struct fuse_arg *args,
 898                           int zeroing)
 899 {
 900         int err = 0;
 901         unsigned i;
 902
 903         for (i = 0; !err && i < numargs; i++)  {
 904                 struct fuse_arg *arg = &args[i];
 905                 if (i == numargs - 1 && argpages)
 906                         err = fuse_copy_pages(cs, arg->size, zeroing);
 907                 else
 908                         err = fuse_copy_one(cs, arg->value, arg->size);
 909         }
 910         return err;
 911 }
 912
 913 static int forget_pending(struct fuse_conn *fc)
 914 {
 915         return fc->forget_list_head.next != NULL;
 916 }
 917
 918 static int request_pending(struct fuse_conn *fc)
 919 {
 920         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts) ||
 921                 forget_pending(fc);
 922 }
 923
 924 /* Wait until a request is available on the pending list */
 925 static void request_wait(struct fuse_conn *fc)
 926 __releases(fc->lock)
 927 __acquires(fc->lock)
 928 {
 929         DECLARE_WAITQUEUE(wait, current);
 930
 931         add_wait_queue_exclusive(&fc->waitq, &wait);
 932         while (fc->connected && !request_pending(fc)) {
 933                 set_current_state(TASK_INTERRUPTIBLE);
 934                 if (signal_pending(current))
 935                         break;
 936
 937                 spin_unlock(&fc->lock);
 938                 schedule();
 939                 spin_lock(&fc->lock);
 940         }
 941         set_current_state(TASK_RUNNING);
 942         remove_wait_queue(&fc->waitq, &wait);
 943 }
 944
 945 /*
 946  * Transfer an interrupt request to userspace
 947  *
 948  * Unlike other requests this is assembled on demand, without a need
 949  * to allocate a separate fuse_req structure.
 950  *
 951  * Called with fc->lock held, releases it
 952  */
 953 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
 954                                size_t nbytes, struct fuse_req *req)
 955 __releases(fc->lock)
 956 {
 957         struct fuse_in_header ih;
 958         struct fuse_interrupt_in arg;
 959         unsigned reqsize = sizeof(ih) + sizeof(arg);
 960         int err;
 961
 962         list_del_init(&req->intr_entry);
 963         req->intr_unique = fuse_get_unique(fc);
 964         memset(&ih, 0, sizeof(ih));
 965         memset(&arg, 0, sizeof(arg));
 966         ih.len = reqsize;
 967         ih.opcode = FUSE_INTERRUPT;
 968         ih.unique = req->intr_unique;
 969         arg.unique = req->in.h.unique;
 970
 971         spin_unlock(&fc->lock);
 972         if (nbytes < reqsize)
 973                 return -EINVAL;
 974
 975         err = fuse_copy_one(cs, &ih, sizeof(ih));
 976         if (!err)
 977                 err = fuse_copy_one(cs, &arg, sizeof(arg));
 978         fuse_copy_finish(cs);
 979
 980         return err ? err : reqsize;
 981 }
 982
 983 static struct fuse_forget_link *dequeue_forget(struct fuse_conn *fc,
 984                                                unsigned max,
 985                                                unsigned *countp)
 986 {
 987         struct fuse_forget_link *head = fc->forget_list_head.next;
 988         struct fuse_forget_link **newhead = &head;
 989         unsigned count;
 990
 991         for (count = 0; *newhead != NULL && count < max; count++)
 992                 newhead = &(*newhead)->next;
 993
 994         fc->forget_list_head.next = *newhead;
 995         *newhead = NULL;
 996         if (fc->forget_list_head.next == NULL)
 997                 fc->forget_list_tail = &fc->forget_list_head;
 998
 999         if (countp != NULL)
1000                 *countp = count;
1001
1002         return head;
1003 }
1004
1005 static int fuse_read_single_forget(struct fuse_conn *fc,
1006                                    struct fuse_copy_state *cs,
1007                                    size_t nbytes)
1008 __releases(fc->lock)
1009 {
1010         int err;
1011         struct fuse_forget_link *forget = dequeue_forget(fc, 1, NULL);
1012         struct fuse_forget_in arg = {
1013                 .nlookup = forget->forget_one.nlookup,
1014         };
1015         struct fuse_in_header ih = {
1016                 .opcode = FUSE_FORGET,
1017                 .nodeid = forget->forget_one.nodeid,
1018                 .unique = fuse_get_unique(fc),
1019                 .len = sizeof(ih) + sizeof(arg),
1020         };
1021
1022         spin_unlock(&fc->lock);
1023         kfree(forget);
1024         if (nbytes < ih.len)
1025                 return -EINVAL;
1026
1027         err = fuse_copy_one(cs, &ih, sizeof(ih));
1028         if (!err)
1029                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1030         fuse_copy_finish(cs);
1031
1032         if (err)
1033                 return err;
1034
1035         return ih.len;
1036 }
1037
1038 static int fuse_read_batch_forget(struct fuse_conn *fc,
1039                                    struct fuse_copy_state *cs, size_t nbytes)
1040 __releases(fc->lock)
1041 {
1042         int err;
1043         unsigned max_forgets;
1044         unsigned count;
1045         struct fuse_forget_link *head;
1046         struct fuse_batch_forget_in arg = { .count = 0 };
1047         struct fuse_in_header ih = {
1048                 .opcode = FUSE_BATCH_FORGET,
1049                 .unique = fuse_get_unique(fc),
1050                 .len = sizeof(ih) + sizeof(arg),
1051         };
1052
1053         if (nbytes < ih.len) {
1054                 spin_unlock(&fc->lock);
1055                 return -EINVAL;
1056         }
1057
1058         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1059         head = dequeue_forget(fc, max_forgets, &count);
1060         spin_unlock(&fc->lock);
1061
1062         arg.count = count;
1063         ih.len += count * sizeof(struct fuse_forget_one);
1064         err = fuse_copy_one(cs, &ih, sizeof(ih));
1065         if (!err)
1066                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1067
1068         while (head) {
1069                 struct fuse_forget_link *forget = head;
1070
1071                 if (!err) {
1072                         err = fuse_copy_one(cs, &forget->forget_one,
1073                                             sizeof(forget->forget_one));
1074                 }
1075                 head = forget->next;
1076                 kfree(forget);
1077         }
1078
1079         fuse_copy_finish(cs);
1080
1081         if (err)
1082                 return err;
1083
1084         return ih.len;
1085 }
1086
1087 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_copy_state *cs,
1088                             size_t nbytes)
1089 __releases(fc->lock)
1090 {
1091         if (fc->minor < 16 || fc->forget_list_head.next->next == NULL)
1092                 return fuse_read_single_forget(fc, cs, nbytes);
1093         else
1094                 return fuse_read_batch_forget(fc, cs, nbytes);
1095 }
1096
1097 /*
1098  * Read a single request into the userspace filesystem's buffer.  This
1099  * function waits until a request is available, then removes it from
1100  * the pending list and copies request data to userspace buffer.  If
1101  * no reply is needed (FORGET) or request has been aborted or there
1102  * was an error during the copying then it's finished by calling
1103  * request_end().  Otherwise add it to the processing list, and set
1104  * the 'sent' flag.
1105  */
1106 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1107                                 struct fuse_copy_state *cs, size_t nbytes)
1108 {
1109         int err;
1110         struct fuse_req *req;
1111         struct fuse_in *in;
1112         unsigned reqsize;
1113
1114  restart:
1115         spin_lock(&fc->lock);
1116         err = -EAGAIN;
1117         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
1118             !request_pending(fc))
1119                 goto err_unlock;
1120
1121         request_wait(fc);
1122         err = -ENODEV;
1123         if (!fc->connected)
1124                 goto err_unlock;
1125         err = -ERESTARTSYS;
1126         if (!request_pending(fc))
1127                 goto err_unlock;
1128
1129         if (!list_empty(&fc->interrupts)) {
1130                 req = list_entry(fc->interrupts.next, struct fuse_req,
1131                                  intr_entry);
1132                 return fuse_read_interrupt(fc, cs, nbytes, req);
1133         }
1134
1135         if (forget_pending(fc)) {
1136                 if (list_empty(&fc->pending) || fc->forget_batch-- > 0)
1137                         return fuse_read_forget(fc, cs, nbytes);
1138
1139                 if (fc->forget_batch <= -8)
1140                         fc->forget_batch = 16;
1141         }
1142
1143         req = list_entry(fc->pending.next, struct fuse_req, list);
1144         req->state = FUSE_REQ_READING;
1145         list_move(&req->list, &fc->io);
1146
1147         in = &req->in;
1148         reqsize = in->h.len;
1149         /* If request is too large, reply with an error and restart the read */
1150         if (nbytes < reqsize) {
1151                 req->out.h.error = -EIO;
1152                 /* SETXATTR is special, since it may contain too large data */
1153                 if (in->h.opcode == FUSE_SETXATTR)
1154                         req->out.h.error = -E2BIG;
1155                 request_end(fc, req);
1156                 goto restart;
1157         }
1158         spin_unlock(&fc->lock);
1159         cs->req = req;
1160         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1161         if (!err)
1162                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1163                                      (struct fuse_arg *) in->args, 0);
1164         fuse_copy_finish(cs);
1165         spin_lock(&fc->lock);
1166         req->locked = 0;
1167         if (req->aborted) {
1168                 request_end(fc, req);
1169                 return -ENODEV;
1170         }
1171         if (err) {
1172                 req->out.h.error = -EIO;
1173                 request_end(fc, req);
1174                 return err;
1175         }
1176         if (!req->isreply)
1177                 request_end(fc, req);
1178         else {
1179                 req->state = FUSE_REQ_SENT;
1180                 list_move_tail(&req->list, &fc->processing);
1181                 if (req->interrupted)
1182                         queue_interrupt(fc, req);
1183                 spin_unlock(&fc->lock);
1184         }
1185         return reqsize;
1186
1187  err_unlock:
1188         spin_unlock(&fc->lock);
1189         return err;
1190 }
1191
1192 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
1193                               unsigned long nr_segs, loff_t pos)
1194 {
1195         struct fuse_copy_state cs;
1196         struct file *file = iocb->ki_filp;
1197         struct fuse_conn *fc = fuse_get_conn(file);
1198         if (!fc)
1199                 return -EPERM;
1200
1201         fuse_copy_init(&cs, fc, 1, iov, nr_segs);
1202
1203         return fuse_dev_do_read(fc, file, &cs, iov_length(iov, nr_segs));
1204 }
1205
1206 static int fuse_dev_pipe_buf_steal(struct pipe_inode_info *pipe,
1207                                    struct pipe_buffer *buf)
1208 {
1209         return 1;
1210 }
1211
1212 static const struct pipe_buf_operations fuse_dev_pipe_buf_ops = {
1213         .can_merge = 0,
1214         .map = generic_pipe_buf_map,
1215         .unmap = generic_pipe_buf_unmap,
1216         .confirm = generic_pipe_buf_confirm,
1217         .release = generic_pipe_buf_release,
1218         .steal = fuse_dev_pipe_buf_steal,
1219         .get = generic_pipe_buf_get,
1220 };
1221
1222 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1223                                     struct pipe_inode_info *pipe,
1224                                     size_t len, unsigned int flags)
1225 {
1226         int ret;
1227         int page_nr = 0;
1228         int do_wakeup = 0;
1229         struct pipe_buffer *bufs;
1230         struct fuse_copy_state cs;
1231         struct fuse_conn *fc = fuse_get_conn(in);
1232         if (!fc)
1233                 return -EPERM;
1234
1235         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1236         if (!bufs)
1237                 return -ENOMEM;
1238
1239         fuse_copy_init(&cs, fc, 1, NULL, 0);
1240         cs.pipebufs = bufs;
1241         cs.pipe = pipe;
1242         ret = fuse_dev_do_read(fc, in, &cs, len);
1243         if (ret < 0)
1244                 goto out;
1245
1246         ret = 0;
1247         pipe_lock(pipe);
1248
1249         if (!pipe->readers) {
1250                 send_sig(SIGPIPE, current, 0);
1251                 if (!ret)
1252                         ret = -EPIPE;
1253                 goto out_unlock;
1254         }
1255
1256         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1257                 ret = -EIO;
1258                 goto out_unlock;
1259         }
1260
1261         while (page_nr < cs.nr_segs) {
1262                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1263                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1264
1265                 buf->page = bufs[page_nr].page;
1266                 buf->offset = bufs[page_nr].offset;
1267                 buf->len = bufs[page_nr].len;
1268                 buf->ops = &fuse_dev_pipe_buf_ops;
1269
1270                 pipe->nrbufs++;
1271                 page_nr++;
1272                 ret += buf->len;
1273
1274                 if (pipe->inode)
1275                         do_wakeup = 1;
1276         }
1277
1278 out_unlock:
1279         pipe_unlock(pipe);
1280
1281         if (do_wakeup) {
1282                 smp_mb();
1283                 if (waitqueue_active(&pipe->wait))
1284                         wake_up_interruptible(&pipe->wait);
1285                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1286         }
1287
1288 out:
1289         for (; page_nr < cs.nr_segs; page_nr++)
1290                 page_cache_release(bufs[page_nr].page);
1291
1292         kfree(bufs);
1293         return ret;
1294 }
1295
1296 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1297                             struct fuse_copy_state *cs)
1298 {
1299         struct fuse_notify_poll_wakeup_out outarg;
1300         int err = -EINVAL;
1301
1302         if (size != sizeof(outarg))
1303                 goto err;
1304
1305         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1306         if (err)
1307                 goto err;
1308
1309         fuse_copy_finish(cs);
1310         return fuse_notify_poll_wakeup(fc, &outarg);
1311
1312 err:
1313         fuse_copy_finish(cs);
1314         return err;
1315 }
1316
1317 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1318                                    struct fuse_copy_state *cs)
1319 {
1320         struct fuse_notify_inval_inode_out outarg;
1321         int err = -EINVAL;
1322
1323         if (size != sizeof(outarg))
1324                 goto err;
1325
1326         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1327         if (err)
1328                 goto err;
1329         fuse_copy_finish(cs);
1330
1331         down_read(&fc->killsb);
1332         err = -ENOENT;
1333         if (fc->sb) {
1334                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1335                                                outarg.off, outarg.len);
1336         }
1337         up_read(&fc->killsb);
1338         return err;
1339
1340 err:
1341         fuse_copy_finish(cs);
1342         return err;
1343 }
1344
1345 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1346                                    struct fuse_copy_state *cs)
1347 {
1348         struct fuse_notify_inval_entry_out outarg;
1349         int err = -ENOMEM;
1350         char *buf;
1351         struct qstr name;
1352
1353         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1354         if (!buf)
1355                 goto err;
1356
1357         err = -EINVAL;
1358         if (size < sizeof(outarg))
1359                 goto err;
1360
1361         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1362         if (err)
1363                 goto err;
1364
1365         err = -ENAMETOOLONG;
1366         if (outarg.namelen > FUSE_NAME_MAX)
1367                 goto err;
1368
1369         err = -EINVAL;
1370         if (size != sizeof(outarg) + outarg.namelen + 1)
1371                 goto err;
1372
1373         name.name = buf;
1374         name.len = outarg.namelen;
1375         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1376         if (err)
1377                 goto err;
1378         fuse_copy_finish(cs);
1379         buf[outarg.namelen] = 0;
1380         name.hash = full_name_hash(name.name, name.len);
1381
1382         down_read(&fc->killsb);
1383         err = -ENOENT;
1384         if (fc->sb)
1385                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, &name);
1386         up_read(&fc->killsb);
1387         kfree(buf);
1388         return err;
1389
1390 err:
1391         kfree(buf);
1392         fuse_copy_finish(cs);
1393         return err;
1394 }
1395
1396 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1397                              struct fuse_copy_state *cs)
1398 {
1399         struct fuse_notify_store_out outarg;
1400         struct inode *inode;
1401         struct address_space *mapping;
1402         u64 nodeid;
1403         int err;
1404         pgoff_t index;
1405         unsigned int offset;
1406         unsigned int num;
1407         loff_t file_size;
1408         loff_t end;
1409
1410         err = -EINVAL;
1411         if (size < sizeof(outarg))
1412                 goto out_finish;
1413
1414         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1415         if (err)
1416                 goto out_finish;
1417
1418         err = -EINVAL;
1419         if (size - sizeof(outarg) != outarg.size)
1420                 goto out_finish;
1421
1422         nodeid = outarg.nodeid;
1423
1424         down_read(&fc->killsb);
1425
1426         err = -ENOENT;
1427         if (!fc->sb)
1428                 goto out_up_killsb;
1429
1430         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1431         if (!inode)
1432                 goto out_up_killsb;
1433
1434         mapping = inode->i_mapping;
1435         index = outarg.offset >> PAGE_CACHE_SHIFT;
1436         offset = outarg.offset & ~PAGE_CACHE_MASK;
1437         file_size = i_size_read(inode);
1438         end = outarg.offset + outarg.size;
1439         if (end > file_size) {
1440                 file_size = end;
1441                 fuse_write_update_size(inode, file_size);
1442         }
1443
1444         num = outarg.size;
1445         while (num) {
1446                 struct page *page;
1447                 unsigned int this_num;
1448
1449                 err = -ENOMEM;
1450                 page = find_or_create_page(mapping, index,
1451                                            mapping_gfp_mask(mapping));
1452                 if (!page)
1453                         goto out_iput;
1454
1455                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1456                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1457                 if (!err && offset == 0 && (num != 0 || file_size == end))
1458                         SetPageUptodate(page);
1459                 unlock_page(page);
1460                 page_cache_release(page);
1461
1462                 if (err)
1463                         goto out_iput;
1464
1465                 num -= this_num;
1466                 offset = 0;
1467                 index++;
1468         }
1469
1470         err = 0;
1471
1472 out_iput:
1473         iput(inode);
1474 out_up_killsb:
1475         up_read(&fc->killsb);
1476 out_finish:
1477         fuse_copy_finish(cs);
1478         return err;
1479 }
1480
1481 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1482 {
1483         release_pages(req->pages, req->num_pages, 0);
1484 }
1485
1486 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1487                          struct fuse_notify_retrieve_out *outarg)
1488 {
1489         int err;
1490         struct address_space *mapping = inode->i_mapping;
1491         struct fuse_req *req;
1492         pgoff_t index;
1493         loff_t file_size;
1494         unsigned int num;
1495         unsigned int offset;
1496         size_t total_len = 0;
1497
1498         req = fuse_get_req(fc);
1499         if (IS_ERR(req))
1500                 return PTR_ERR(req);
1501
1502         offset = outarg->offset & ~PAGE_CACHE_MASK;
1503
1504         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1505         req->in.h.nodeid = outarg->nodeid;
1506         req->in.numargs = 2;
1507         req->in.argpages = 1;
1508         req->page_offset = offset;
1509         req->end = fuse_retrieve_end;
1510
1511         index = outarg->offset >> PAGE_CACHE_SHIFT;
1512         file_size = i_size_read(inode);
1513         num = outarg->size;
1514         if (outarg->offset > file_size)
1515                 num = 0;
1516         else if (outarg->offset + num > file_size)
1517                 num = file_size - outarg->offset;
1518
1519         while (num && req->num_pages < FUSE_MAX_PAGES_PER_REQ) {
1520                 struct page *page;
1521                 unsigned int this_num;
1522
1523                 page = find_get_page(mapping, index);
1524                 if (!page)
1525                         break;
1526
1527                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1528                 req->pages[req->num_pages] = page;
1529                 req->num_pages++;
1530
1531                 offset = 0;
1532                 num -= this_num;
1533                 total_len += this_num;
1534                 index++;
1535         }
1536         req->misc.retrieve_in.offset = outarg->offset;
1537         req->misc.retrieve_in.size = total_len;
1538         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1539         req->in.args[0].value = &req->misc.retrieve_in;
1540         req->in.args[1].size = total_len;
1541
1542         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1543         if (err)
1544                 fuse_retrieve_end(fc, req);
1545
1546         return err;
1547 }
1548
1549 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1550                                 struct fuse_copy_state *cs)
1551 {
1552         struct fuse_notify_retrieve_out outarg;
1553         struct inode *inode;
1554         int err;
1555
1556         err = -EINVAL;
1557         if (size != sizeof(outarg))
1558                 goto copy_finish;
1559
1560         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1561         if (err)
1562                 goto copy_finish;
1563
1564         fuse_copy_finish(cs);
1565
1566         down_read(&fc->killsb);
1567         err = -ENOENT;
1568         if (fc->sb) {
1569                 u64 nodeid = outarg.nodeid;
1570
1571                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1572                 if (inode) {
1573                         err = fuse_retrieve(fc, inode, &outarg);
1574                         iput(inode);
1575                 }
1576         }
1577         up_read(&fc->killsb);
1578
1579         return err;
1580
1581 copy_finish:
1582         fuse_copy_finish(cs);
1583         return err;
1584 }
1585
1586 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1587                        unsigned int size, struct fuse_copy_state *cs)
1588 {
1589         switch (code) {
1590         case FUSE_NOTIFY_POLL:
1591                 return fuse_notify_poll(fc, size, cs);
1592
1593         case FUSE_NOTIFY_INVAL_INODE:
1594                 return fuse_notify_inval_inode(fc, size, cs);
1595
1596         case FUSE_NOTIFY_INVAL_ENTRY:
1597                 return fuse_notify_inval_entry(fc, size, cs);
1598
1599         case FUSE_NOTIFY_STORE:
1600                 return fuse_notify_store(fc, size, cs);
1601
1602         case FUSE_NOTIFY_RETRIEVE:
1603                 return fuse_notify_retrieve(fc, size, cs);
1604
1605         default:
1606                 fuse_copy_finish(cs);
1607                 return -EINVAL;
1608         }
1609 }
1610
1611 /* Look up request on processing list by unique ID */
1612 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1613 {
1614         struct list_head *entry;
1615
1616         list_for_each(entry, &fc->processing) {
1617                 struct fuse_req *req;
1618                 req = list_entry(entry, struct fuse_req, list);
1619                 if (req->in.h.unique == unique || req->intr_unique == unique)
1620                         return req;
1621         }
1622         return NULL;
1623 }
1624
1625 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1626                          unsigned nbytes)
1627 {
1628         unsigned reqsize = sizeof(struct fuse_out_header);
1629
1630         if (out->h.error)
1631                 return nbytes != reqsize ? -EINVAL : 0;
1632
1633         reqsize += len_args(out->numargs, out->args);
1634
1635         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1636                 return -EINVAL;
1637         else if (reqsize > nbytes) {
1638                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1639                 unsigned diffsize = reqsize - nbytes;
1640                 if (diffsize > lastarg->size)
1641                         return -EINVAL;
1642                 lastarg->size -= diffsize;
1643         }
1644         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1645                               out->page_zeroing);
1646 }
1647
1648 /*
1649  * Write a single reply to a request.  First the header is copied from
1650  * the write buffer.  The request is then searched on the processing
1651  * list by the unique ID found in the header.  If found, then remove
1652  * it from the list and copy the rest of the buffer to the request.
1653  * The request is finished by calling request_end()
1654  */
1655 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1656                                  struct fuse_copy_state *cs, size_t nbytes)
1657 {
1658         int err;
1659         struct fuse_req *req;
1660         struct fuse_out_header oh;
1661
1662         if (nbytes < sizeof(struct fuse_out_header))
1663                 return -EINVAL;
1664
1665         err = fuse_copy_one(cs, &oh, sizeof(oh));
1666         if (err)
1667                 goto err_finish;
1668
1669         err = -EINVAL;
1670         if (oh.len != nbytes)
1671                 goto err_finish;
1672
1673         /*
1674          * Zero oh.unique indicates unsolicited notification message
1675          * and error contains notification code.
1676          */
1677         if (!oh.unique) {
1678                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1679                 return err ? err : nbytes;
1680         }
1681
1682         err = -EINVAL;
1683         if (oh.error <= -1000 || oh.error > 0)
1684                 goto err_finish;
1685
1686         spin_lock(&fc->lock);
1687         err = -ENOENT;
1688         if (!fc->connected)
1689                 goto err_unlock;
1690
1691         req = request_find(fc, oh.unique);
1692         if (!req)
1693                 goto err_unlock;
1694
1695         if (req->aborted) {
1696                 spin_unlock(&fc->lock);
1697                 fuse_copy_finish(cs);
1698                 spin_lock(&fc->lock);
1699                 request_end(fc, req);
1700                 return -ENOENT;
1701         }
1702         /* Is it an interrupt reply? */
1703         if (req->intr_unique == oh.unique) {
1704                 err = -EINVAL;
1705                 if (nbytes != sizeof(struct fuse_out_header))
1706                         goto err_unlock;
1707
1708                 if (oh.error == -ENOSYS)
1709                         fc->no_interrupt = 1;
1710                 else if (oh.error == -EAGAIN)
1711                         queue_interrupt(fc, req);
1712
1713                 spin_unlock(&fc->lock);
1714                 fuse_copy_finish(cs);
1715                 return nbytes;
1716         }
1717
1718         req->state = FUSE_REQ_WRITING;
1719         list_move(&req->list, &fc->io);
1720         req->out.h = oh;
1721         req->locked = 1;
1722         cs->req = req;
1723         if (!req->out.page_replace)
1724                 cs->move_pages = 0;
1725         spin_unlock(&fc->lock);
1726
1727         err = copy_out_args(cs, &req->out, nbytes);
1728         fuse_copy_finish(cs);
1729
1730         spin_lock(&fc->lock);
1731         req->locked = 0;
1732         if (!err) {
1733                 if (req->aborted)
1734                         err = -ENOENT;
1735         } else if (!req->aborted)
1736                 req->out.h.error = -EIO;
1737         request_end(fc, req);
1738
1739         return err ? err : nbytes;
1740
1741  err_unlock:
1742         spin_unlock(&fc->lock);
1743  err_finish:
1744         fuse_copy_finish(cs);
1745         return err;
1746 }
1747
1748 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
1749                               unsigned long nr_segs, loff_t pos)
1750 {
1751         struct fuse_copy_state cs;
1752         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1753         if (!fc)
1754                 return -EPERM;
1755
1756         fuse_copy_init(&cs, fc, 0, iov, nr_segs);
1757
1758         return fuse_dev_do_write(fc, &cs, iov_length(iov, nr_segs));
1759 }
1760
1761 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1762                                      struct file *out, loff_t *ppos,
1763                                      size_t len, unsigned int flags)
1764 {
1765         unsigned nbuf;
1766         unsigned idx;
1767         struct pipe_buffer *bufs;
1768         struct fuse_copy_state cs;
1769         struct fuse_conn *fc;
1770         size_t rem;
1771         ssize_t ret;
1772
1773         fc = fuse_get_conn(out);
1774         if (!fc)
1775                 return -EPERM;
1776
1777         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1778         if (!bufs)
1779                 return -ENOMEM;
1780
1781         pipe_lock(pipe);
1782         nbuf = 0;
1783         rem = 0;
1784         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1785                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1786
1787         ret = -EINVAL;
1788         if (rem < len) {
1789                 pipe_unlock(pipe);
1790                 goto out;
1791         }
1792
1793         rem = len;
1794         while (rem) {
1795                 struct pipe_buffer *ibuf;
1796                 struct pipe_buffer *obuf;
1797
1798                 BUG_ON(nbuf >= pipe->buffers);
1799                 BUG_ON(!pipe->nrbufs);
1800                 ibuf = &pipe->bufs[pipe->curbuf];
1801                 obuf = &bufs[nbuf];
1802
1803                 if (rem >= ibuf->len) {
1804                         *obuf = *ibuf;
1805                         ibuf->ops = NULL;
1806                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1807                         pipe->nrbufs--;
1808                 } else {
1809                         ibuf->ops->get(pipe, ibuf);
1810                         *obuf = *ibuf;
1811                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1812                         obuf->len = rem;
1813                         ibuf->offset += obuf->len;
1814                         ibuf->len -= obuf->len;
1815                 }
1816                 nbuf++;
1817                 rem -= obuf->len;
1818         }
1819         pipe_unlock(pipe);
1820
1821         fuse_copy_init(&cs, fc, 0, NULL, nbuf);
1822         cs.pipebufs = bufs;
1823         cs.pipe = pipe;
1824
1825         if (flags & SPLICE_F_MOVE)
1826                 cs.move_pages = 1;
1827
1828         ret = fuse_dev_do_write(fc, &cs, len);
1829
1830         for (idx = 0; idx < nbuf; idx++) {
1831                 struct pipe_buffer *buf = &bufs[idx];
1832                 buf->ops->release(pipe, buf);
1833         }
1834 out:
1835         kfree(bufs);
1836         return ret;
1837 }
1838
1839 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1840 {
1841         unsigned mask = POLLOUT | POLLWRNORM;
1842         struct fuse_conn *fc = fuse_get_conn(file);
1843         if (!fc)
1844                 return POLLERR;
1845
1846         poll_wait(file, &fc->waitq, wait);
1847
1848         spin_lock(&fc->lock);
1849         if (!fc->connected)
1850                 mask = POLLERR;
1851         else if (request_pending(fc))
1852                 mask |= POLLIN | POLLRDNORM;
1853         spin_unlock(&fc->lock);
1854
1855         return mask;
1856 }
1857
1858 /*
1859  * Abort all requests on the given list (pending or processing)
1860  *
1861  * This function releases and reacquires fc->lock
1862  */
1863 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1864 __releases(fc->lock)
1865 __acquires(fc->lock)
1866 {
1867         while (!list_empty(head)) {
1868                 struct fuse_req *req;
1869                 req = list_entry(head->next, struct fuse_req, list);
1870                 req->out.h.error = -ECONNABORTED;
1871                 request_end(fc, req);
1872                 spin_lock(&fc->lock);
1873         }
1874 }
1875
1876 /*
1877  * Abort requests under I/O
1878  *
1879  * The requests are set to aborted and finished, and the request
1880  * waiter is woken up.  This will make request_wait_answer() wait
1881  * until the request is unlocked and then return.
1882  *
1883  * If the request is asynchronous, then the end function needs to be
1884  * called after waiting for the request to be unlocked (if it was
1885  * locked).
1886  */
1887 static void end_io_requests(struct fuse_conn *fc)
1888 __releases(fc->lock)
1889 __acquires(fc->lock)
1890 {
1891         while (!list_empty(&fc->io)) {
1892                 struct fuse_req *req =
1893                         list_entry(fc->io.next, struct fuse_req, list);
1894                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1895
1896                 req->aborted = 1;
1897                 req->out.h.error = -ECONNABORTED;
1898                 req->state = FUSE_REQ_FINISHED;
1899                 list_del_init(&req->list);
1900                 wake_up(&req->waitq);
1901                 if (end) {
1902                         req->end = NULL;
1903                         __fuse_get_request(req);
1904                         spin_unlock(&fc->lock);
1905                         wait_event(req->waitq, !req->locked);
1906                         end(fc, req);
1907                         fuse_put_request(fc, req);
1908                         spin_lock(&fc->lock);
1909                 }
1910         }
1911 }
1912
1913 static void end_queued_requests(struct fuse_conn *fc)
1914 __releases(fc->lock)
1915 __acquires(fc->lock)
1916 {
1917         fc->max_background = UINT_MAX;
1918         flush_bg_queue(fc);
1919         end_requests(fc, &fc->pending);
1920         end_requests(fc, &fc->processing);
1921         while (forget_pending(fc))
1922                 kfree(dequeue_forget(fc, 1, NULL));
1923 }
1924
1925 static void end_polls(struct fuse_conn *fc)
1926 {
1927         struct rb_node *p;
1928
1929         p = rb_first(&fc->polled_files);
1930
1931         while (p) {
1932                 struct fuse_file *ff;
1933                 ff = rb_entry(p, struct fuse_file, polled_node);
1934                 wake_up_interruptible_all(&ff->poll_wait);
1935
1936                 p = rb_next(p);
1937         }
1938 }
1939
1940 /*
1941  * Abort all requests.
1942  *
1943  * Emergency exit in case of a malicious or accidental deadlock, or
1944  * just a hung filesystem.
1945  *
1946  * The same effect is usually achievable through killing the
1947  * filesystem daemon and all users of the filesystem.  The exception
1948  * is the combination of an asynchronous request and the tricky
1949  * deadlock (see Documentation/filesystems/fuse.txt).
1950  *
1951  * During the aborting, progression of requests from the pending and
1952  * processing lists onto the io list, and progression of new requests
1953  * onto the pending list is prevented by req->connected being false.
1954  *
1955  * Progression of requests under I/O to the processing list is
1956  * prevented by the req->aborted flag being true for these requests.
1957  * For this reason requests on the io list must be aborted first.
1958  */
1959 void fuse_abort_conn(struct fuse_conn *fc)
1960 {
1961         spin_lock(&fc->lock);
1962         if (fc->connected) {
1963                 fc->connected = 0;
1964                 fc->blocked = 0;
1965                 end_io_requests(fc);
1966                 end_queued_requests(fc);
1967                 end_polls(fc);
1968                 wake_up_all(&fc->waitq);
1969                 wake_up_all(&fc->blocked_waitq);
1970                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1971         }
1972         spin_unlock(&fc->lock);
1973 }
1974 EXPORT_SYMBOL_GPL(fuse_abort_conn);
1975
1976 int fuse_dev_release(struct inode *inode, struct file *file)
1977 {
1978         struct fuse_conn *fc = fuse_get_conn(file);
1979         if (fc) {
1980                 spin_lock(&fc->lock);
1981                 fc->connected = 0;
1982                 fc->blocked = 0;
1983                 end_queued_requests(fc);
1984                 end_polls(fc);
1985                 wake_up_all(&fc->blocked_waitq);
1986                 spin_unlock(&fc->lock);
1987                 fuse_conn_put(fc);
1988         }
1989
1990         return 0;
1991 }
1992 EXPORT_SYMBOL_GPL(fuse_dev_release);
1993
1994 static int fuse_dev_fasync(int fd, struct file *file, int on)
1995 {
1996         struct fuse_conn *fc = fuse_get_conn(file);
1997         if (!fc)
1998                 return -EPERM;
1999
2000         /* No locking - fasync_helper does its own locking */
2001         return fasync_helper(fd, file, on, &fc->fasync);
2002 }
2003
2004 const struct file_operations fuse_dev_operations = {
2005         .owner          = THIS_MODULE,
2006         .llseek         = no_llseek,
2007         .read           = do_sync_read,
2008         .aio_read       = fuse_dev_read,
2009         .splice_read    = fuse_dev_splice_read,
2010         .write          = do_sync_write,
2011         .aio_write      = fuse_dev_write,
2012         .splice_write   = fuse_dev_splice_write,
2013         .poll           = fuse_dev_poll,
2014         .release        = fuse_dev_release,
2015         .fasync         = fuse_dev_fasync,
2016 };
2017 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2018
2019 static struct miscdevice fuse_miscdevice = {
2020         .minor = FUSE_MINOR,
2021         .name  = "fuse",
2022         .fops = &fuse_dev_operations,
2023 };
2024
2025 int __init fuse_dev_init(void)
2026 {
2027         int err = -ENOMEM;
2028         fuse_req_cachep = kmem_cache_create("fuse_request",
2029                                             sizeof(struct fuse_req),
2030                                             0, 0, NULL);
2031         if (!fuse_req_cachep)
2032                 goto out;
2033
2034         err = misc_register(&fuse_miscdevice);
2035         if (err)
2036                 goto out_cache_clean;
2037
2038         return 0;
2039
2040  out_cache_clean:
2041         kmem_cache_destroy(fuse_req_cachep);
2042  out:
2043         return err;
2044 }
2045
2046 void fuse_dev_cleanup(void)
2047 {
2048         misc_deregister(&fuse_miscdevice);
2049         kmem_cache_destroy(fuse_req_cachep);
2050 }