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/aio.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, struct page **pages,
  39                               struct fuse_page_desc *page_descs,
  40                               unsigned npages)
  41 {
  42         memset(req, 0, sizeof(*req));
  43         memset(pages, 0, sizeof(*pages) * npages);
  44         memset(page_descs, 0, sizeof(*page_descs) * npages);
  45         INIT_LIST_HEAD(&req->list);
  46         INIT_LIST_HEAD(&req->intr_entry);
  47         init_waitqueue_head(&req->waitq);
  48         atomic_set(&req->count, 1);
  49         req->pages = pages;
  50         req->page_descs = page_descs;
  51         req->max_pages = npages;
  52 }
  53
  54 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  55 {
  56         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, flags);
  57         if (req) {
  58                 struct page **pages;
  59                 struct fuse_page_desc *page_descs;
  60
  61                 if (npages <= FUSE_REQ_INLINE_PAGES) {
  62                         pages = req->inline_pages;
  63                         page_descs = req->inline_page_descs;
  64                 } else {
  65                         pages = kmalloc(sizeof(struct page *) * npages, flags);
  66                         page_descs = kmalloc(sizeof(struct fuse_page_desc) *
  67                                              npages, flags);
  68                 }
  69
  70                 if (!pages || !page_descs) {
  71                         kfree(pages);
  72                         kfree(page_descs);
  73                         kmem_cache_free(fuse_req_cachep, req);
  74                         return NULL;
  75                 }
  76
  77                 fuse_request_init(req, pages, page_descs, npages);
  78         }
  79         return req;
  80 }
  81
  82 struct fuse_req *fuse_request_alloc(unsigned npages)
  83 {
  84         return __fuse_request_alloc(npages, GFP_KERNEL);
  85 }
  86 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  87
  88 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  89 {
  90         return __fuse_request_alloc(npages, GFP_NOFS);
  91 }
  92
  93 void fuse_request_free(struct fuse_req *req)
  94 {
  95         if (req->pages != req->inline_pages) {
  96                 kfree(req->pages);
  97                 kfree(req->page_descs);
  98         }
  99         kmem_cache_free(fuse_req_cachep, req);
 100 }
 101
 102 static void block_sigs(sigset_t *oldset)
 103 {
 104         sigset_t mask;
 105
 106         siginitsetinv(&mask, sigmask(SIGKILL));
 107         sigprocmask(SIG_BLOCK, &mask, oldset);
 108 }
 109
 110 static void restore_sigs(sigset_t *oldset)
 111 {
 112         sigprocmask(SIG_SETMASK, oldset, NULL);
 113 }
 114
 115 void __fuse_get_request(struct fuse_req *req)
 116 {
 117         atomic_inc(&req->count);
 118 }
 119
 120 /* Must be called with > 1 refcount */
 121 static void __fuse_put_request(struct fuse_req *req)
 122 {
 123         BUG_ON(atomic_read(&req->count) < 2);
 124         atomic_dec(&req->count);
 125 }
 126
 127 static void fuse_req_init_context(struct fuse_req *req)
 128 {
 129         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 130         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 131         req->in.h.pid = current->pid;
 132 }
 133
 134 void fuse_set_initialized(struct fuse_conn *fc)
 135 {
 136         /* Make sure stores before this are seen on another CPU */
 137         smp_wmb();
 138         fc->initialized = 1;
 139 }
 140
 141 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 142 {
 143         return !fc->initialized || (for_background && fc->blocked);
 144 }
 145
 146 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 147                                        bool for_background)
 148 {
 149         struct fuse_req *req;
 150         int err;
 151         atomic_inc(&fc->num_waiting);
 152
 153         if (fuse_block_alloc(fc, for_background)) {
 154                 sigset_t oldset;
 155                 int intr;
 156
 157                 block_sigs(&oldset);
 158                 intr = wait_event_interruptible_exclusive(fc->blocked_waitq,
 159                                 !fuse_block_alloc(fc, for_background));
 160                 restore_sigs(&oldset);
 161                 err = -EINTR;
 162                 if (intr)
 163                         goto out;
 164         }
 165         /* Matches smp_wmb() in fuse_set_initialized() */
 166         smp_rmb();
 167
 168         err = -ENOTCONN;
 169         if (!fc->connected)
 170                 goto out;
 171
 172         req = fuse_request_alloc(npages);
 173         err = -ENOMEM;
 174         if (!req) {
 175                 if (for_background)
 176                         wake_up(&fc->blocked_waitq);
 177                 goto out;
 178         }
 179
 180         fuse_req_init_context(req);
 181         req->waiting = 1;
 182         req->background = for_background;
 183         return req;
 184
 185  out:
 186         atomic_dec(&fc->num_waiting);
 187         return ERR_PTR(err);
 188 }
 189
 190 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 191 {
 192         return __fuse_get_req(fc, npages, false);
 193 }
 194 EXPORT_SYMBOL_GPL(fuse_get_req);
 195
 196 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 197                                              unsigned npages)
 198 {
 199         return __fuse_get_req(fc, npages, true);
 200 }
 201 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 202
 203 /*
 204  * Return request in fuse_file->reserved_req.  However that may
 205  * currently be in use.  If that is the case, wait for it to become
 206  * available.
 207  */
 208 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 209                                          struct file *file)
 210 {
 211         struct fuse_req *req = NULL;
 212         struct fuse_file *ff = file->private_data;
 213
 214         do {
 215                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 216                 spin_lock(&fc->lock);
 217                 if (ff->reserved_req) {
 218                         req = ff->reserved_req;
 219                         ff->reserved_req = NULL;
 220                         req->stolen_file = get_file(file);
 221                 }
 222                 spin_unlock(&fc->lock);
 223         } while (!req);
 224
 225         return req;
 226 }
 227
 228 /*
 229  * Put stolen request back into fuse_file->reserved_req
 230  */
 231 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 232 {
 233         struct file *file = req->stolen_file;
 234         struct fuse_file *ff = file->private_data;
 235
 236         spin_lock(&fc->lock);
 237         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 238         BUG_ON(ff->reserved_req);
 239         ff->reserved_req = req;
 240         wake_up_all(&fc->reserved_req_waitq);
 241         spin_unlock(&fc->lock);
 242         fput(file);
 243 }
 244
 245 /*
 246  * Gets a requests for a file operation, always succeeds
 247  *
 248  * This is used for sending the FLUSH request, which must get to
 249  * userspace, due to POSIX locks which may need to be unlocked.
 250  *
 251  * If allocation fails due to OOM, use the reserved request in
 252  * fuse_file.
 253  *
 254  * This is very unlikely to deadlock accidentally, since the
 255  * filesystem should not have it's own file open.  If deadlock is
 256  * intentional, it can still be broken by "aborting" the filesystem.
 257  */
 258 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 259                                              struct file *file)
 260 {
 261         struct fuse_req *req;
 262
 263         atomic_inc(&fc->num_waiting);
 264         wait_event(fc->blocked_waitq, fc->initialized);
 265         /* Matches smp_wmb() in fuse_set_initialized() */
 266         smp_rmb();
 267         req = fuse_request_alloc(0);
 268         if (!req)
 269                 req = get_reserved_req(fc, file);
 270
 271         fuse_req_init_context(req);
 272         req->waiting = 1;
 273         req->background = 0;
 274         return req;
 275 }
 276
 277 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 278 {
 279         if (atomic_dec_and_test(&req->count)) {
 280                 if (unlikely(req->background)) {
 281                         /*
 282                          * We get here in the unlikely case that a background
 283                          * request was allocated but not sent
 284                          */
 285                         spin_lock(&fc->lock);
 286                         if (!fc->blocked)
 287                                 wake_up(&fc->blocked_waitq);
 288                         spin_unlock(&fc->lock);
 289                 }
 290
 291                 if (req->waiting)
 292                         atomic_dec(&fc->num_waiting);
 293
 294                 if (req->stolen_file)
 295                         put_reserved_req(fc, req);
 296                 else
 297                         fuse_request_free(req);
 298         }
 299 }
 300 EXPORT_SYMBOL_GPL(fuse_put_request);
 301
 302 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 303 {
 304         unsigned nbytes = 0;
 305         unsigned i;
 306
 307         for (i = 0; i < numargs; i++)
 308                 nbytes += args[i].size;
 309
 310         return nbytes;
 311 }
 312
 313 static u64 fuse_get_unique(struct fuse_conn *fc)
 314 {
 315         fc->reqctr++;
 316         /* zero is special */
 317         if (fc->reqctr == 0)
 318                 fc->reqctr = 1;
 319
 320         return fc->reqctr;
 321 }
 322
 323 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 324 {
 325         req->in.h.len = sizeof(struct fuse_in_header) +
 326                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 327         list_add_tail(&req->list, &fc->pending);
 328         req->state = FUSE_REQ_PENDING;
 329         if (!req->waiting) {
 330                 req->waiting = 1;
 331                 atomic_inc(&fc->num_waiting);
 332         }
 333         wake_up(&fc->waitq);
 334         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 335 }
 336
 337 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 338                        u64 nodeid, u64 nlookup)
 339 {
 340         forget->forget_one.nodeid = nodeid;
 341         forget->forget_one.nlookup = nlookup;
 342
 343         spin_lock(&fc->lock);
 344         if (fc->connected) {
 345                 fc->forget_list_tail->next = forget;
 346                 fc->forget_list_tail = forget;
 347                 wake_up(&fc->waitq);
 348                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 349         } else {
 350                 kfree(forget);
 351         }
 352         spin_unlock(&fc->lock);
 353 }
 354
 355 static void flush_bg_queue(struct fuse_conn *fc)
 356 {
 357         while (fc->active_background < fc->max_background &&
 358                !list_empty(&fc->bg_queue)) {
 359                 struct fuse_req *req;
 360
 361                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 362                 list_del(&req->list);
 363                 fc->active_background++;
 364                 req->in.h.unique = fuse_get_unique(fc);
 365                 queue_request(fc, req);
 366         }
 367 }
 368
 369 /*
 370  * This function is called when a request is finished.  Either a reply
 371  * has arrived or it was aborted (and not yet sent) or some error
 372  * occurred during communication with userspace, or the device file
 373  * was closed.  The requester thread is woken up (if still waiting),
 374  * the 'end' callback is called if given, else the reference to the
 375  * request is released
 376  *
 377  * Called with fc->lock, unlocks it
 378  */
 379 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 380 __releases(fc->lock)
 381 {
 382         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 383         req->end = NULL;
 384         list_del(&req->list);
 385         list_del(&req->intr_entry);
 386         req->state = FUSE_REQ_FINISHED;
 387         if (req->background) {
 388                 req->background = 0;
 389
 390                 if (fc->num_background == fc->max_background)
 391                         fc->blocked = 0;
 392
 393                 /* Wake up next waiter, if any */
 394                 if (!fc->blocked && waitqueue_active(&fc->blocked_waitq))
 395                         wake_up(&fc->blocked_waitq);
 396
 397                 if (fc->num_background == fc->congestion_threshold &&
 398                     fc->connected && fc->bdi_initialized) {
 399                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 400                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 401                 }
 402                 fc->num_background--;
 403                 fc->active_background--;
 404                 flush_bg_queue(fc);
 405         }
 406         spin_unlock(&fc->lock);
 407         wake_up(&req->waitq);
 408         if (end)
 409                 end(fc, req);
 410         fuse_put_request(fc, req);
 411 }
 412
 413 static void wait_answer_interruptible(struct fuse_conn *fc,
 414                                       struct fuse_req *req)
 415 __releases(fc->lock)
 416 __acquires(fc->lock)
 417 {
 418         if (signal_pending(current))
 419                 return;
 420
 421         spin_unlock(&fc->lock);
 422         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 423         spin_lock(&fc->lock);
 424 }
 425
 426 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 427 {
 428         list_add_tail(&req->intr_entry, &fc->interrupts);
 429         wake_up(&fc->waitq);
 430         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 431 }
 432
 433 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 434 __releases(fc->lock)
 435 __acquires(fc->lock)
 436 {
 437         if (!fc->no_interrupt) {
 438                 /* Any signal may interrupt this */
 439                 wait_answer_interruptible(fc, req);
 440
 441                 if (req->aborted)
 442                         goto aborted;
 443                 if (req->state == FUSE_REQ_FINISHED)
 444                         return;
 445
 446                 req->interrupted = 1;
 447                 if (req->state == FUSE_REQ_SENT)
 448                         queue_interrupt(fc, req);
 449         }
 450
 451         if (!req->force) {
 452                 sigset_t oldset;
 453
 454                 /* Only fatal signals may interrupt this */
 455                 block_sigs(&oldset);
 456                 wait_answer_interruptible(fc, req);
 457                 restore_sigs(&oldset);
 458
 459                 if (req->aborted)
 460                         goto aborted;
 461                 if (req->state == FUSE_REQ_FINISHED)
 462                         return;
 463
 464                 /* Request is not yet in userspace, bail out */
 465                 if (req->state == FUSE_REQ_PENDING) {
 466                         list_del(&req->list);
 467                         __fuse_put_request(req);
 468                         req->out.h.error = -EINTR;
 469                         return;
 470                 }
 471         }
 472
 473         /*
 474          * Either request is already in userspace, or it was forced.
 475          * Wait it out.
 476          */
 477         spin_unlock(&fc->lock);
 478         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 479         spin_lock(&fc->lock);
 480
 481         if (!req->aborted)
 482                 return;
 483
 484  aborted:
 485         BUG_ON(req->state != FUSE_REQ_FINISHED);
 486         if (req->locked) {
 487                 /* This is uninterruptible sleep, because data is
 488                    being copied to/from the buffers of req.  During
 489                    locked state, there mustn't be any filesystem
 490                    operation (e.g. page fault), since that could lead
 491                    to deadlock */
 492                 spin_unlock(&fc->lock);
 493                 wait_event(req->waitq, !req->locked);
 494                 spin_lock(&fc->lock);
 495         }
 496 }
 497
 498 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 499 {
 500         BUG_ON(req->background);
 501         spin_lock(&fc->lock);
 502         if (!fc->connected)
 503                 req->out.h.error = -ENOTCONN;
 504         else if (fc->conn_error)
 505                 req->out.h.error = -ECONNREFUSED;
 506         else {
 507                 req->in.h.unique = fuse_get_unique(fc);
 508                 queue_request(fc, req);
 509                 /* acquire extra reference, since request is still needed
 510                    after request_end() */
 511                 __fuse_get_request(req);
 512
 513                 request_wait_answer(fc, req);
 514         }
 515         spin_unlock(&fc->lock);
 516 }
 517
 518 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 519 {
 520         req->isreply = 1;
 521         __fuse_request_send(fc, req);
 522 }
 523 EXPORT_SYMBOL_GPL(fuse_request_send);
 524
 525 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 526 {
 527         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 528                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 529
 530         if (fc->minor < 9) {
 531                 switch (args->in.h.opcode) {
 532                 case FUSE_LOOKUP:
 533                 case FUSE_CREATE:
 534                 case FUSE_MKNOD:
 535                 case FUSE_MKDIR:
 536                 case FUSE_SYMLINK:
 537                 case FUSE_LINK:
 538                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 539                         break;
 540                 case FUSE_GETATTR:
 541                 case FUSE_SETATTR:
 542                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 543                         break;
 544                 }
 545         }
 546         if (fc->minor < 12) {
 547                 switch (args->in.h.opcode) {
 548                 case FUSE_CREATE:
 549                         args->in.args[0].size = sizeof(struct fuse_open_in);
 550                         break;
 551                 case FUSE_MKNOD:
 552                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 553                         break;
 554                 }
 555         }
 556 }
 557
 558 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 559 {
 560         struct fuse_req *req;
 561         ssize_t ret;
 562
 563         req = fuse_get_req(fc, 0);
 564         if (IS_ERR(req))
 565                 return PTR_ERR(req);
 566
 567         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 568         fuse_adjust_compat(fc, args);
 569
 570         req->in.h.opcode = args->in.h.opcode;
 571         req->in.h.nodeid = args->in.h.nodeid;
 572         req->in.numargs = args->in.numargs;
 573         memcpy(req->in.args, args->in.args,
 574                args->in.numargs * sizeof(struct fuse_in_arg));
 575         req->out.argvar = args->out.argvar;
 576         req->out.numargs = args->out.numargs;
 577         memcpy(req->out.args, args->out.args,
 578                args->out.numargs * sizeof(struct fuse_arg));
 579         fuse_request_send(fc, req);
 580         ret = req->out.h.error;
 581         if (!ret && args->out.argvar) {
 582                 BUG_ON(args->out.numargs != 1);
 583                 ret = req->out.args[0].size;
 584         }
 585         fuse_put_request(fc, req);
 586
 587         return ret;
 588 }
 589
 590 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 591                                             struct fuse_req *req)
 592 {
 593         BUG_ON(!req->background);
 594         fc->num_background++;
 595         if (fc->num_background == fc->max_background)
 596                 fc->blocked = 1;
 597         if (fc->num_background == fc->congestion_threshold &&
 598             fc->bdi_initialized) {
 599                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 600                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 601         }
 602         list_add_tail(&req->list, &fc->bg_queue);
 603         flush_bg_queue(fc);
 604 }
 605
 606 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 607 {
 608         spin_lock(&fc->lock);
 609         if (fc->connected) {
 610                 fuse_request_send_nowait_locked(fc, req);
 611                 spin_unlock(&fc->lock);
 612         } else {
 613                 req->out.h.error = -ENOTCONN;
 614                 request_end(fc, req);
 615         }
 616 }
 617
 618 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 619 {
 620         req->isreply = 1;
 621         fuse_request_send_nowait(fc, req);
 622 }
 623 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 624
 625 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 626                                           struct fuse_req *req, u64 unique)
 627 {
 628         int err = -ENODEV;
 629
 630         req->isreply = 0;
 631         req->in.h.unique = unique;
 632         spin_lock(&fc->lock);
 633         if (fc->connected) {
 634                 queue_request(fc, req);
 635                 err = 0;
 636         }
 637         spin_unlock(&fc->lock);
 638
 639         return err;
 640 }
 641
 642 /*
 643  * Called under fc->lock
 644  *
 645  * fc->connected must have been checked previously
 646  */
 647 void fuse_request_send_background_locked(struct fuse_conn *fc,
 648                                          struct fuse_req *req)
 649 {
 650         req->isreply = 1;
 651         fuse_request_send_nowait_locked(fc, req);
 652 }
 653
 654 void fuse_force_forget(struct file *file, u64 nodeid)
 655 {
 656         struct inode *inode = file_inode(file);
 657         struct fuse_conn *fc = get_fuse_conn(inode);
 658         struct fuse_req *req;
 659         struct fuse_forget_in inarg;
 660
 661         memset(&inarg, 0, sizeof(inarg));
 662         inarg.nlookup = 1;
 663         req = fuse_get_req_nofail_nopages(fc, file);
 664         req->in.h.opcode = FUSE_FORGET;
 665         req->in.h.nodeid = nodeid;
 666         req->in.numargs = 1;
 667         req->in.args[0].size = sizeof(inarg);
 668         req->in.args[0].value = &inarg;
 669         req->isreply = 0;
 670         __fuse_request_send(fc, req);
 671         /* ignore errors */
 672         fuse_put_request(fc, req);
 673 }
 674
 675 /*
 676  * Lock the request.  Up to the next unlock_request() there mustn't be
 677  * anything that could cause a page-fault.  If the request was already
 678  * aborted bail out.
 679  */
 680 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 681 {
 682         int err = 0;
 683         if (req) {
 684                 spin_lock(&fc->lock);
 685                 if (req->aborted)
 686                         err = -ENOENT;
 687                 else
 688                         req->locked = 1;
 689                 spin_unlock(&fc->lock);
 690         }
 691         return err;
 692 }
 693
 694 /*
 695  * Unlock request.  If it was aborted during being locked, the
 696  * requester thread is currently waiting for it to be unlocked, so
 697  * wake it up.
 698  */
 699 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 700 {
 701         if (req) {
 702                 spin_lock(&fc->lock);
 703                 req->locked = 0;
 704                 if (req->aborted)
 705                         wake_up(&req->waitq);
 706                 spin_unlock(&fc->lock);
 707         }
 708 }
 709
 710 struct fuse_copy_state {
 711         struct fuse_conn *fc;
 712         int write;
 713         struct fuse_req *req;
 714         const struct iovec *iov;
 715         struct pipe_buffer *pipebufs;
 716         struct pipe_buffer *currbuf;
 717         struct pipe_inode_info *pipe;
 718         unsigned long nr_segs;
 719         unsigned long seglen;
 720         unsigned long addr;
 721         struct page *pg;
 722         unsigned len;
 723         unsigned offset;
 724         unsigned move_pages:1;
 725 };
 726
 727 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 728                            int write,
 729                            const struct iovec *iov, unsigned long nr_segs)
 730 {
 731         memset(cs, 0, sizeof(*cs));
 732         cs->fc = fc;
 733         cs->write = write;
 734         cs->iov = iov;
 735         cs->nr_segs = nr_segs;
 736 }
 737
 738 /* Unmap and put previous page of userspace buffer */
 739 static void fuse_copy_finish(struct fuse_copy_state *cs)
 740 {
 741         if (cs->currbuf) {
 742                 struct pipe_buffer *buf = cs->currbuf;
 743
 744                 if (cs->write)
 745                         buf->len = PAGE_SIZE - cs->len;
 746                 cs->currbuf = NULL;
 747         } else if (cs->pg) {
 748                 if (cs->write) {
 749                         flush_dcache_page(cs->pg);
 750                         set_page_dirty_lock(cs->pg);
 751                 }
 752                 put_page(cs->pg);
 753         }
 754         cs->pg = NULL;
 755 }
 756
 757 /*
 758  * Get another pagefull of userspace buffer, and map it to kernel
 759  * address space, and lock request
 760  */
 761 static int fuse_copy_fill(struct fuse_copy_state *cs)
 762 {
 763         struct page *page;
 764         int err;
 765
 766         unlock_request(cs->fc, cs->req);
 767         fuse_copy_finish(cs);
 768         if (cs->pipebufs) {
 769                 struct pipe_buffer *buf = cs->pipebufs;
 770
 771                 if (!cs->write) {
 772                         err = buf->ops->confirm(cs->pipe, buf);
 773                         if (err)
 774                                 return err;
 775
 776                         BUG_ON(!cs->nr_segs);
 777                         cs->currbuf = buf;
 778                         cs->pg = buf->page;
 779                         cs->offset = buf->offset;
 780                         cs->len = buf->len;
 781                         cs->pipebufs++;
 782                         cs->nr_segs--;
 783                 } else {
 784                         if (cs->nr_segs == cs->pipe->buffers)
 785                                 return -EIO;
 786
 787                         page = alloc_page(GFP_HIGHUSER);
 788                         if (!page)
 789                                 return -ENOMEM;
 790
 791                         buf->page = page;
 792                         buf->offset = 0;
 793                         buf->len = 0;
 794
 795                         cs->currbuf = buf;
 796                         cs->pg = page;
 797                         cs->offset = 0;
 798                         cs->len = PAGE_SIZE;
 799                         cs->pipebufs++;
 800                         cs->nr_segs++;
 801                 }
 802         } else {
 803                 if (!cs->seglen) {
 804                         BUG_ON(!cs->nr_segs);
 805                         cs->seglen = cs->iov[0].iov_len;
 806                         cs->addr = (unsigned long) cs->iov[0].iov_base;
 807                         cs->iov++;
 808                         cs->nr_segs--;
 809                 }
 810                 err = get_user_pages_fast(cs->addr, 1, cs->write, &page);
 811                 if (err < 0)
 812                         return err;
 813                 BUG_ON(err != 1);
 814                 cs->pg = page;
 815                 cs->offset = cs->addr % PAGE_SIZE;
 816                 cs->len = min(PAGE_SIZE - cs->offset, cs->seglen);
 817                 cs->seglen -= cs->len;
 818                 cs->addr += cs->len;
 819         }
 820
 821         return lock_request(cs->fc, cs->req);
 822 }
 823
 824 /* Do as much copy to/from userspace buffer as we can */
 825 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 826 {
 827         unsigned ncpy = min(*size, cs->len);
 828         if (val) {
 829                 void *pgaddr = kmap_atomic(cs->pg);
 830                 void *buf = pgaddr + cs->offset;
 831
 832                 if (cs->write)
 833                         memcpy(buf, *val, ncpy);
 834                 else
 835                         memcpy(*val, buf, ncpy);
 836
 837                 kunmap_atomic(pgaddr);
 838                 *val += ncpy;
 839         }
 840         *size -= ncpy;
 841         cs->len -= ncpy;
 842         cs->offset += ncpy;
 843         return ncpy;
 844 }
 845
 846 static int fuse_check_page(struct page *page)
 847 {
 848         if (page_mapcount(page) ||
 849             page->mapping != NULL ||
 850             page_count(page) != 1 ||
 851             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 852              ~(1 << PG_locked |
 853                1 << PG_referenced |
 854                1 << PG_uptodate |
 855                1 << PG_lru |
 856                1 << PG_active |
 857                1 << PG_reclaim))) {
 858                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 859                 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);
 860                 return 1;
 861         }
 862         return 0;
 863 }
 864
 865 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 866 {
 867         int err;
 868         struct page *oldpage = *pagep;
 869         struct page *newpage;
 870         struct pipe_buffer *buf = cs->pipebufs;
 871
 872         unlock_request(cs->fc, cs->req);
 873         fuse_copy_finish(cs);
 874
 875         err = buf->ops->confirm(cs->pipe, buf);
 876         if (err)
 877                 return err;
 878
 879         BUG_ON(!cs->nr_segs);
 880         cs->currbuf = buf;
 881         cs->len = buf->len;
 882         cs->pipebufs++;
 883         cs->nr_segs--;
 884
 885         if (cs->len != PAGE_SIZE)
 886                 goto out_fallback;
 887
 888         if (buf->ops->steal(cs->pipe, buf) != 0)
 889                 goto out_fallback;
 890
 891         newpage = buf->page;
 892
 893         if (!PageUptodate(newpage))
 894                 SetPageUptodate(newpage);
 895
 896         ClearPageMappedToDisk(newpage);
 897
 898         if (fuse_check_page(newpage) != 0)
 899                 goto out_fallback_unlock;
 900
 901         /*
 902          * This is a new and locked page, it shouldn't be mapped or
 903          * have any special flags on it
 904          */
 905         if (WARN_ON(page_mapped(oldpage)))
 906                 goto out_fallback_unlock;
 907         if (WARN_ON(page_has_private(oldpage)))
 908                 goto out_fallback_unlock;
 909         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 910                 goto out_fallback_unlock;
 911         if (WARN_ON(PageMlocked(oldpage)))
 912                 goto out_fallback_unlock;
 913
 914         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 915         if (err) {
 916                 unlock_page(newpage);
 917                 return err;
 918         }
 919
 920         page_cache_get(newpage);
 921
 922         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 923                 lru_cache_add_file(newpage);
 924
 925         err = 0;
 926         spin_lock(&cs->fc->lock);
 927         if (cs->req->aborted)
 928                 err = -ENOENT;
 929         else
 930                 *pagep = newpage;
 931         spin_unlock(&cs->fc->lock);
 932
 933         if (err) {
 934                 unlock_page(newpage);
 935                 page_cache_release(newpage);
 936                 return err;
 937         }
 938
 939         unlock_page(oldpage);
 940         page_cache_release(oldpage);
 941         cs->len = 0;
 942
 943         return 0;
 944
 945 out_fallback_unlock:
 946         unlock_page(newpage);
 947 out_fallback:
 948         cs->pg = buf->page;
 949         cs->offset = buf->offset;
 950
 951         err = lock_request(cs->fc, cs->req);
 952         if (err)
 953                 return err;
 954
 955         return 1;
 956 }
 957
 958 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 959                          unsigned offset, unsigned count)
 960 {
 961         struct pipe_buffer *buf;
 962
 963         if (cs->nr_segs == cs->pipe->buffers)
 964                 return -EIO;
 965
 966         unlock_request(cs->fc, cs->req);
 967         fuse_copy_finish(cs);
 968
 969         buf = cs->pipebufs;
 970         page_cache_get(page);
 971         buf->page = page;
 972         buf->offset = offset;
 973         buf->len = count;
 974
 975         cs->pipebufs++;
 976         cs->nr_segs++;
 977         cs->len = 0;
 978
 979         return 0;
 980 }
 981
 982 /*
 983  * Copy a page in the request to/from the userspace buffer.  Must be
 984  * done atomically
 985  */
 986 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 987                           unsigned offset, unsigned count, int zeroing)
 988 {
 989         int err;
 990         struct page *page = *pagep;
 991
 992         if (page && zeroing && count < PAGE_SIZE)
 993                 clear_highpage(page);
 994
 995         while (count) {
 996                 if (cs->write && cs->pipebufs && page) {
 997                         return fuse_ref_page(cs, page, offset, count);
 998                 } else if (!cs->len) {
 999                         if (cs->move_pages && page &&
1000                             offset == 0 && count == PAGE_SIZE) {
1001                                 err = fuse_try_move_page(cs, pagep);
1002                                 if (err <= 0)
1003                                         return err;
1004                         } else {
1005                                 err = fuse_copy_fill(cs);
1006                                 if (err)
1007                                         return err;
1008                         }
1009                 }
1010                 if (page) {
1011                         void *mapaddr = kmap_atomic(page);
1012                         void *buf = mapaddr + offset;
1013                         offset += fuse_copy_do(cs, &buf, &count);
1014                         kunmap_atomic(mapaddr);
1015                 } else
1016                         offset += fuse_copy_do(cs, NULL, &count);
1017         }
1018         if (page && !cs->write)
1019                 flush_dcache_page(page);
1020         return 0;
1021 }
1022
1023 /* Copy pages in the request to/from userspace buffer */
1024 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1025                            int zeroing)
1026 {
1027         unsigned i;
1028         struct fuse_req *req = cs->req;
1029
1030         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1031                 int err;
1032                 unsigned offset = req->page_descs[i].offset;
1033                 unsigned count = min(nbytes, req->page_descs[i].length);
1034
1035                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1036                                      zeroing);
1037                 if (err)
1038                         return err;
1039
1040                 nbytes -= count;
1041         }
1042         return 0;
1043 }
1044
1045 /* Copy a single argument in the request to/from userspace buffer */
1046 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1047 {
1048         while (size) {
1049                 if (!cs->len) {
1050                         int err = fuse_copy_fill(cs);
1051                         if (err)
1052                                 return err;
1053                 }
1054                 fuse_copy_do(cs, &val, &size);
1055         }
1056         return 0;
1057 }
1058
1059 /* Copy request arguments to/from userspace buffer */
1060 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1061                           unsigned argpages, struct fuse_arg *args,
1062                           int zeroing)
1063 {
1064         int err = 0;
1065         unsigned i;
1066
1067         for (i = 0; !err && i < numargs; i++)  {
1068                 struct fuse_arg *arg = &args[i];
1069                 if (i == numargs - 1 && argpages)
1070                         err = fuse_copy_pages(cs, arg->size, zeroing);
1071                 else
1072                         err = fuse_copy_one(cs, arg->value, arg->size);
1073         }
1074         return err;
1075 }
1076
1077 static int forget_pending(struct fuse_conn *fc)
1078 {
1079         return fc->forget_list_head.next != NULL;
1080 }
1081
1082 static int request_pending(struct fuse_conn *fc)
1083 {
1084         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts) ||
1085                 forget_pending(fc);
1086 }
1087
1088 /* Wait until a request is available on the pending list */
1089 static void request_wait(struct fuse_conn *fc)
1090 __releases(fc->lock)
1091 __acquires(fc->lock)
1092 {
1093         DECLARE_WAITQUEUE(wait, current);
1094
1095         add_wait_queue_exclusive(&fc->waitq, &wait);
1096         while (fc->connected && !request_pending(fc)) {
1097                 set_current_state(TASK_INTERRUPTIBLE);
1098                 if (signal_pending(current))
1099                         break;
1100
1101                 spin_unlock(&fc->lock);
1102                 schedule();
1103                 spin_lock(&fc->lock);
1104         }
1105         set_current_state(TASK_RUNNING);
1106         remove_wait_queue(&fc->waitq, &wait);
1107 }
1108
1109 /*
1110  * Transfer an interrupt request to userspace
1111  *
1112  * Unlike other requests this is assembled on demand, without a need
1113  * to allocate a separate fuse_req structure.
1114  *
1115  * Called with fc->lock held, releases it
1116  */
1117 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
1118                                size_t nbytes, struct fuse_req *req)
1119 __releases(fc->lock)
1120 {
1121         struct fuse_in_header ih;
1122         struct fuse_interrupt_in arg;
1123         unsigned reqsize = sizeof(ih) + sizeof(arg);
1124         int err;
1125
1126         list_del_init(&req->intr_entry);
1127         req->intr_unique = fuse_get_unique(fc);
1128         memset(&ih, 0, sizeof(ih));
1129         memset(&arg, 0, sizeof(arg));
1130         ih.len = reqsize;
1131         ih.opcode = FUSE_INTERRUPT;
1132         ih.unique = req->intr_unique;
1133         arg.unique = req->in.h.unique;
1134
1135         spin_unlock(&fc->lock);
1136         if (nbytes < reqsize)
1137                 return -EINVAL;
1138
1139         err = fuse_copy_one(cs, &ih, sizeof(ih));
1140         if (!err)
1141                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1142         fuse_copy_finish(cs);
1143
1144         return err ? err : reqsize;
1145 }
1146
1147 static struct fuse_forget_link *dequeue_forget(struct fuse_conn *fc,
1148                                                unsigned max,
1149                                                unsigned *countp)
1150 {
1151         struct fuse_forget_link *head = fc->forget_list_head.next;
1152         struct fuse_forget_link **newhead = &head;
1153         unsigned count;
1154
1155         for (count = 0; *newhead != NULL && count < max; count++)
1156                 newhead = &(*newhead)->next;
1157
1158         fc->forget_list_head.next = *newhead;
1159         *newhead = NULL;
1160         if (fc->forget_list_head.next == NULL)
1161                 fc->forget_list_tail = &fc->forget_list_head;
1162
1163         if (countp != NULL)
1164                 *countp = count;
1165
1166         return head;
1167 }
1168
1169 static int fuse_read_single_forget(struct fuse_conn *fc,
1170                                    struct fuse_copy_state *cs,
1171                                    size_t nbytes)
1172 __releases(fc->lock)
1173 {
1174         int err;
1175         struct fuse_forget_link *forget = dequeue_forget(fc, 1, NULL);
1176         struct fuse_forget_in arg = {
1177                 .nlookup = forget->forget_one.nlookup,
1178         };
1179         struct fuse_in_header ih = {
1180                 .opcode = FUSE_FORGET,
1181                 .nodeid = forget->forget_one.nodeid,
1182                 .unique = fuse_get_unique(fc),
1183                 .len = sizeof(ih) + sizeof(arg),
1184         };
1185
1186         spin_unlock(&fc->lock);
1187         kfree(forget);
1188         if (nbytes < ih.len)
1189                 return -EINVAL;
1190
1191         err = fuse_copy_one(cs, &ih, sizeof(ih));
1192         if (!err)
1193                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1194         fuse_copy_finish(cs);
1195
1196         if (err)
1197                 return err;
1198
1199         return ih.len;
1200 }
1201
1202 static int fuse_read_batch_forget(struct fuse_conn *fc,
1203                                    struct fuse_copy_state *cs, size_t nbytes)
1204 __releases(fc->lock)
1205 {
1206         int err;
1207         unsigned max_forgets;
1208         unsigned count;
1209         struct fuse_forget_link *head;
1210         struct fuse_batch_forget_in arg = { .count = 0 };
1211         struct fuse_in_header ih = {
1212                 .opcode = FUSE_BATCH_FORGET,
1213                 .unique = fuse_get_unique(fc),
1214                 .len = sizeof(ih) + sizeof(arg),
1215         };
1216
1217         if (nbytes < ih.len) {
1218                 spin_unlock(&fc->lock);
1219                 return -EINVAL;
1220         }
1221
1222         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1223         head = dequeue_forget(fc, max_forgets, &count);
1224         spin_unlock(&fc->lock);
1225
1226         arg.count = count;
1227         ih.len += count * sizeof(struct fuse_forget_one);
1228         err = fuse_copy_one(cs, &ih, sizeof(ih));
1229         if (!err)
1230                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1231
1232         while (head) {
1233                 struct fuse_forget_link *forget = head;
1234
1235                 if (!err) {
1236                         err = fuse_copy_one(cs, &forget->forget_one,
1237                                             sizeof(forget->forget_one));
1238                 }
1239                 head = forget->next;
1240                 kfree(forget);
1241         }
1242
1243         fuse_copy_finish(cs);
1244
1245         if (err)
1246                 return err;
1247
1248         return ih.len;
1249 }
1250
1251 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_copy_state *cs,
1252                             size_t nbytes)
1253 __releases(fc->lock)
1254 {
1255         if (fc->minor < 16 || fc->forget_list_head.next->next == NULL)
1256                 return fuse_read_single_forget(fc, cs, nbytes);
1257         else
1258                 return fuse_read_batch_forget(fc, cs, nbytes);
1259 }
1260
1261 /*
1262  * Read a single request into the userspace filesystem's buffer.  This
1263  * function waits until a request is available, then removes it from
1264  * the pending list and copies request data to userspace buffer.  If
1265  * no reply is needed (FORGET) or request has been aborted or there
1266  * was an error during the copying then it's finished by calling
1267  * request_end().  Otherwise add it to the processing list, and set
1268  * the 'sent' flag.
1269  */
1270 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1271                                 struct fuse_copy_state *cs, size_t nbytes)
1272 {
1273         int err;
1274         struct fuse_req *req;
1275         struct fuse_in *in;
1276         unsigned reqsize;
1277
1278  restart:
1279         spin_lock(&fc->lock);
1280         err = -EAGAIN;
1281         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
1282             !request_pending(fc))
1283                 goto err_unlock;
1284
1285         request_wait(fc);
1286         err = -ENODEV;
1287         if (!fc->connected)
1288                 goto err_unlock;
1289         err = -ERESTARTSYS;
1290         if (!request_pending(fc))
1291                 goto err_unlock;
1292
1293         if (!list_empty(&fc->interrupts)) {
1294                 req = list_entry(fc->interrupts.next, struct fuse_req,
1295                                  intr_entry);
1296                 return fuse_read_interrupt(fc, cs, nbytes, req);
1297         }
1298
1299         if (forget_pending(fc)) {
1300                 if (list_empty(&fc->pending) || fc->forget_batch-- > 0)
1301                         return fuse_read_forget(fc, cs, nbytes);
1302
1303                 if (fc->forget_batch <= -8)
1304                         fc->forget_batch = 16;
1305         }
1306
1307         req = list_entry(fc->pending.next, struct fuse_req, list);
1308         req->state = FUSE_REQ_READING;
1309         list_move(&req->list, &fc->io);
1310
1311         in = &req->in;
1312         reqsize = in->h.len;
1313         /* If request is too large, reply with an error and restart the read */
1314         if (nbytes < reqsize) {
1315                 req->out.h.error = -EIO;
1316                 /* SETXATTR is special, since it may contain too large data */
1317                 if (in->h.opcode == FUSE_SETXATTR)
1318                         req->out.h.error = -E2BIG;
1319                 request_end(fc, req);
1320                 goto restart;
1321         }
1322         spin_unlock(&fc->lock);
1323         cs->req = req;
1324         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1325         if (!err)
1326                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1327                                      (struct fuse_arg *) in->args, 0);
1328         fuse_copy_finish(cs);
1329         spin_lock(&fc->lock);
1330         req->locked = 0;
1331         if (req->aborted) {
1332                 request_end(fc, req);
1333                 return -ENODEV;
1334         }
1335         if (err) {
1336                 req->out.h.error = -EIO;
1337                 request_end(fc, req);
1338                 return err;
1339         }
1340         if (!req->isreply)
1341                 request_end(fc, req);
1342         else {
1343                 req->state = FUSE_REQ_SENT;
1344                 list_move_tail(&req->list, &fc->processing);
1345                 if (req->interrupted)
1346                         queue_interrupt(fc, req);
1347                 spin_unlock(&fc->lock);
1348         }
1349         return reqsize;
1350
1351  err_unlock:
1352         spin_unlock(&fc->lock);
1353         return err;
1354 }
1355
1356 static int fuse_dev_open(struct inode *inode, struct file *file)
1357 {
1358         /*
1359          * The fuse device's file's private_data is used to hold
1360          * the fuse_conn(ection) when it is mounted, and is used to
1361          * keep track of whether the file has been mounted already.
1362          */
1363         file->private_data = NULL;
1364         return 0;
1365 }
1366
1367 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
1368                               unsigned long nr_segs, loff_t pos)
1369 {
1370         struct fuse_copy_state cs;
1371         struct file *file = iocb->ki_filp;
1372         struct fuse_conn *fc = fuse_get_conn(file);
1373         if (!fc)
1374                 return -EPERM;
1375
1376         fuse_copy_init(&cs, fc, 1, iov, nr_segs);
1377
1378         return fuse_dev_do_read(fc, file, &cs, iov_length(iov, nr_segs));
1379 }
1380
1381 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1382                                     struct pipe_inode_info *pipe,
1383                                     size_t len, unsigned int flags)
1384 {
1385         int ret;
1386         int page_nr = 0;
1387         int do_wakeup = 0;
1388         struct pipe_buffer *bufs;
1389         struct fuse_copy_state cs;
1390         struct fuse_conn *fc = fuse_get_conn(in);
1391         if (!fc)
1392                 return -EPERM;
1393
1394         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1395         if (!bufs)
1396                 return -ENOMEM;
1397
1398         fuse_copy_init(&cs, fc, 1, NULL, 0);
1399         cs.pipebufs = bufs;
1400         cs.pipe = pipe;
1401         ret = fuse_dev_do_read(fc, in, &cs, len);
1402         if (ret < 0)
1403                 goto out;
1404
1405         ret = 0;
1406         pipe_lock(pipe);
1407
1408         if (!pipe->readers) {
1409                 send_sig(SIGPIPE, current, 0);
1410                 if (!ret)
1411                         ret = -EPIPE;
1412                 goto out_unlock;
1413         }
1414
1415         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1416                 ret = -EIO;
1417                 goto out_unlock;
1418         }
1419
1420         while (page_nr < cs.nr_segs) {
1421                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1422                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1423
1424                 buf->page = bufs[page_nr].page;
1425                 buf->offset = bufs[page_nr].offset;
1426                 buf->len = bufs[page_nr].len;
1427                 /*
1428                  * Need to be careful about this.  Having buf->ops in module
1429                  * code can Oops if the buffer persists after module unload.
1430                  */
1431                 buf->ops = &nosteal_pipe_buf_ops;
1432
1433                 pipe->nrbufs++;
1434                 page_nr++;
1435                 ret += buf->len;
1436
1437                 if (pipe->files)
1438                         do_wakeup = 1;
1439         }
1440
1441 out_unlock:
1442         pipe_unlock(pipe);
1443
1444         if (do_wakeup) {
1445                 smp_mb();
1446                 if (waitqueue_active(&pipe->wait))
1447                         wake_up_interruptible(&pipe->wait);
1448                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1449         }
1450
1451 out:
1452         for (; page_nr < cs.nr_segs; page_nr++)
1453                 page_cache_release(bufs[page_nr].page);
1454
1455         kfree(bufs);
1456         return ret;
1457 }
1458
1459 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1460                             struct fuse_copy_state *cs)
1461 {
1462         struct fuse_notify_poll_wakeup_out outarg;
1463         int err = -EINVAL;
1464
1465         if (size != sizeof(outarg))
1466                 goto err;
1467
1468         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1469         if (err)
1470                 goto err;
1471
1472         fuse_copy_finish(cs);
1473         return fuse_notify_poll_wakeup(fc, &outarg);
1474
1475 err:
1476         fuse_copy_finish(cs);
1477         return err;
1478 }
1479
1480 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1481                                    struct fuse_copy_state *cs)
1482 {
1483         struct fuse_notify_inval_inode_out outarg;
1484         int err = -EINVAL;
1485
1486         if (size != sizeof(outarg))
1487                 goto err;
1488
1489         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1490         if (err)
1491                 goto err;
1492         fuse_copy_finish(cs);
1493
1494         down_read(&fc->killsb);
1495         err = -ENOENT;
1496         if (fc->sb) {
1497                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1498                                                outarg.off, outarg.len);
1499         }
1500         up_read(&fc->killsb);
1501         return err;
1502
1503 err:
1504         fuse_copy_finish(cs);
1505         return err;
1506 }
1507
1508 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1509                                    struct fuse_copy_state *cs)
1510 {
1511         struct fuse_notify_inval_entry_out outarg;
1512         int err = -ENOMEM;
1513         char *buf;
1514         struct qstr name;
1515
1516         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1517         if (!buf)
1518                 goto err;
1519
1520         err = -EINVAL;
1521         if (size < sizeof(outarg))
1522                 goto err;
1523
1524         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1525         if (err)
1526                 goto err;
1527
1528         err = -ENAMETOOLONG;
1529         if (outarg.namelen > FUSE_NAME_MAX)
1530                 goto err;
1531
1532         err = -EINVAL;
1533         if (size != sizeof(outarg) + outarg.namelen + 1)
1534                 goto err;
1535
1536         name.name = buf;
1537         name.len = outarg.namelen;
1538         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1539         if (err)
1540                 goto err;
1541         fuse_copy_finish(cs);
1542         buf[outarg.namelen] = 0;
1543         name.hash = full_name_hash(name.name, name.len);
1544
1545         down_read(&fc->killsb);
1546         err = -ENOENT;
1547         if (fc->sb)
1548                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1549         up_read(&fc->killsb);
1550         kfree(buf);
1551         return err;
1552
1553 err:
1554         kfree(buf);
1555         fuse_copy_finish(cs);
1556         return err;
1557 }
1558
1559 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1560                               struct fuse_copy_state *cs)
1561 {
1562         struct fuse_notify_delete_out outarg;
1563         int err = -ENOMEM;
1564         char *buf;
1565         struct qstr name;
1566
1567         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1568         if (!buf)
1569                 goto err;
1570
1571         err = -EINVAL;
1572         if (size < sizeof(outarg))
1573                 goto err;
1574
1575         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1576         if (err)
1577                 goto err;
1578
1579         err = -ENAMETOOLONG;
1580         if (outarg.namelen > FUSE_NAME_MAX)
1581                 goto err;
1582
1583         err = -EINVAL;
1584         if (size != sizeof(outarg) + outarg.namelen + 1)
1585                 goto err;
1586
1587         name.name = buf;
1588         name.len = outarg.namelen;
1589         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1590         if (err)
1591                 goto err;
1592         fuse_copy_finish(cs);
1593         buf[outarg.namelen] = 0;
1594         name.hash = full_name_hash(name.name, name.len);
1595
1596         down_read(&fc->killsb);
1597         err = -ENOENT;
1598         if (fc->sb)
1599                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1600                                                outarg.child, &name);
1601         up_read(&fc->killsb);
1602         kfree(buf);
1603         return err;
1604
1605 err:
1606         kfree(buf);
1607         fuse_copy_finish(cs);
1608         return err;
1609 }
1610
1611 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1612                              struct fuse_copy_state *cs)
1613 {
1614         struct fuse_notify_store_out outarg;
1615         struct inode *inode;
1616         struct address_space *mapping;
1617         u64 nodeid;
1618         int err;
1619         pgoff_t index;
1620         unsigned int offset;
1621         unsigned int num;
1622         loff_t file_size;
1623         loff_t end;
1624
1625         err = -EINVAL;
1626         if (size < sizeof(outarg))
1627                 goto out_finish;
1628
1629         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1630         if (err)
1631                 goto out_finish;
1632
1633         err = -EINVAL;
1634         if (size - sizeof(outarg) != outarg.size)
1635                 goto out_finish;
1636
1637         nodeid = outarg.nodeid;
1638
1639         down_read(&fc->killsb);
1640
1641         err = -ENOENT;
1642         if (!fc->sb)
1643                 goto out_up_killsb;
1644
1645         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1646         if (!inode)
1647                 goto out_up_killsb;
1648
1649         mapping = inode->i_mapping;
1650         index = outarg.offset >> PAGE_CACHE_SHIFT;
1651         offset = outarg.offset & ~PAGE_CACHE_MASK;
1652         file_size = i_size_read(inode);
1653         end = outarg.offset + outarg.size;
1654         if (end > file_size) {
1655                 file_size = end;
1656                 fuse_write_update_size(inode, file_size);
1657         }
1658
1659         num = outarg.size;
1660         while (num) {
1661                 struct page *page;
1662                 unsigned int this_num;
1663
1664                 err = -ENOMEM;
1665                 page = find_or_create_page(mapping, index,
1666                                            mapping_gfp_mask(mapping));
1667                 if (!page)
1668                         goto out_iput;
1669
1670                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1671                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1672                 if (!err && offset == 0 &&
1673                     (this_num == PAGE_CACHE_SIZE || file_size == end))
1674                         SetPageUptodate(page);
1675                 unlock_page(page);
1676                 page_cache_release(page);
1677
1678                 if (err)
1679                         goto out_iput;
1680
1681                 num -= this_num;
1682                 offset = 0;
1683                 index++;
1684         }
1685
1686         err = 0;
1687
1688 out_iput:
1689         iput(inode);
1690 out_up_killsb:
1691         up_read(&fc->killsb);
1692 out_finish:
1693         fuse_copy_finish(cs);
1694         return err;
1695 }
1696
1697 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1698 {
1699         release_pages(req->pages, req->num_pages, false);
1700 }
1701
1702 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1703                          struct fuse_notify_retrieve_out *outarg)
1704 {
1705         int err;
1706         struct address_space *mapping = inode->i_mapping;
1707         struct fuse_req *req;
1708         pgoff_t index;
1709         loff_t file_size;
1710         unsigned int num;
1711         unsigned int offset;
1712         size_t total_len = 0;
1713         int num_pages;
1714
1715         offset = outarg->offset & ~PAGE_CACHE_MASK;
1716         file_size = i_size_read(inode);
1717
1718         num = outarg->size;
1719         if (outarg->offset > file_size)
1720                 num = 0;
1721         else if (outarg->offset + num > file_size)
1722                 num = file_size - outarg->offset;
1723
1724         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1725         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1726
1727         req = fuse_get_req(fc, num_pages);
1728         if (IS_ERR(req))
1729                 return PTR_ERR(req);
1730
1731         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1732         req->in.h.nodeid = outarg->nodeid;
1733         req->in.numargs = 2;
1734         req->in.argpages = 1;
1735         req->page_descs[0].offset = offset;
1736         req->end = fuse_retrieve_end;
1737
1738         index = outarg->offset >> PAGE_CACHE_SHIFT;
1739
1740         while (num && req->num_pages < num_pages) {
1741                 struct page *page;
1742                 unsigned int this_num;
1743
1744                 page = find_get_page(mapping, index);
1745                 if (!page)
1746                         break;
1747
1748                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1749                 req->pages[req->num_pages] = page;
1750                 req->page_descs[req->num_pages].length = this_num;
1751                 req->num_pages++;
1752
1753                 offset = 0;
1754                 num -= this_num;
1755                 total_len += this_num;
1756                 index++;
1757         }
1758         req->misc.retrieve_in.offset = outarg->offset;
1759         req->misc.retrieve_in.size = total_len;
1760         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1761         req->in.args[0].value = &req->misc.retrieve_in;
1762         req->in.args[1].size = total_len;
1763
1764         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1765         if (err)
1766                 fuse_retrieve_end(fc, req);
1767
1768         return err;
1769 }
1770
1771 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1772                                 struct fuse_copy_state *cs)
1773 {
1774         struct fuse_notify_retrieve_out outarg;
1775         struct inode *inode;
1776         int err;
1777
1778         err = -EINVAL;
1779         if (size != sizeof(outarg))
1780                 goto copy_finish;
1781
1782         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1783         if (err)
1784                 goto copy_finish;
1785
1786         fuse_copy_finish(cs);
1787
1788         down_read(&fc->killsb);
1789         err = -ENOENT;
1790         if (fc->sb) {
1791                 u64 nodeid = outarg.nodeid;
1792
1793                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1794                 if (inode) {
1795                         err = fuse_retrieve(fc, inode, &outarg);
1796                         iput(inode);
1797                 }
1798         }
1799         up_read(&fc->killsb);
1800
1801         return err;
1802
1803 copy_finish:
1804         fuse_copy_finish(cs);
1805         return err;
1806 }
1807
1808 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1809                        unsigned int size, struct fuse_copy_state *cs)
1810 {
1811         /* Don't try to move pages (yet) */
1812         cs->move_pages = 0;
1813
1814         switch (code) {
1815         case FUSE_NOTIFY_POLL:
1816                 return fuse_notify_poll(fc, size, cs);
1817
1818         case FUSE_NOTIFY_INVAL_INODE:
1819                 return fuse_notify_inval_inode(fc, size, cs);
1820
1821         case FUSE_NOTIFY_INVAL_ENTRY:
1822                 return fuse_notify_inval_entry(fc, size, cs);
1823
1824         case FUSE_NOTIFY_STORE:
1825                 return fuse_notify_store(fc, size, cs);
1826
1827         case FUSE_NOTIFY_RETRIEVE:
1828                 return fuse_notify_retrieve(fc, size, cs);
1829
1830         case FUSE_NOTIFY_DELETE:
1831                 return fuse_notify_delete(fc, size, cs);
1832
1833         default:
1834                 fuse_copy_finish(cs);
1835                 return -EINVAL;
1836         }
1837 }
1838
1839 /* Look up request on processing list by unique ID */
1840 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1841 {
1842         struct fuse_req *req;
1843
1844         list_for_each_entry(req, &fc->processing, list) {
1845                 if (req->in.h.unique == unique || req->intr_unique == unique)
1846                         return req;
1847         }
1848         return NULL;
1849 }
1850
1851 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1852                          unsigned nbytes)
1853 {
1854         unsigned reqsize = sizeof(struct fuse_out_header);
1855
1856         if (out->h.error)
1857                 return nbytes != reqsize ? -EINVAL : 0;
1858
1859         reqsize += len_args(out->numargs, out->args);
1860
1861         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1862                 return -EINVAL;
1863         else if (reqsize > nbytes) {
1864                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1865                 unsigned diffsize = reqsize - nbytes;
1866                 if (diffsize > lastarg->size)
1867                         return -EINVAL;
1868                 lastarg->size -= diffsize;
1869         }
1870         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1871                               out->page_zeroing);
1872 }
1873
1874 /*
1875  * Write a single reply to a request.  First the header is copied from
1876  * the write buffer.  The request is then searched on the processing
1877  * list by the unique ID found in the header.  If found, then remove
1878  * it from the list and copy the rest of the buffer to the request.
1879  * The request is finished by calling request_end()
1880  */
1881 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1882                                  struct fuse_copy_state *cs, size_t nbytes)
1883 {
1884         int err;
1885         struct fuse_req *req;
1886         struct fuse_out_header oh;
1887
1888         if (nbytes < sizeof(struct fuse_out_header))
1889                 return -EINVAL;
1890
1891         err = fuse_copy_one(cs, &oh, sizeof(oh));
1892         if (err)
1893                 goto err_finish;
1894
1895         err = -EINVAL;
1896         if (oh.len != nbytes)
1897                 goto err_finish;
1898
1899         /*
1900          * Zero oh.unique indicates unsolicited notification message
1901          * and error contains notification code.
1902          */
1903         if (!oh.unique) {
1904                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1905                 return err ? err : nbytes;
1906         }
1907
1908         err = -EINVAL;
1909         if (oh.error <= -1000 || oh.error > 0)
1910                 goto err_finish;
1911
1912         spin_lock(&fc->lock);
1913         err = -ENOENT;
1914         if (!fc->connected)
1915                 goto err_unlock;
1916
1917         req = request_find(fc, oh.unique);
1918         if (!req)
1919                 goto err_unlock;
1920
1921         if (req->aborted) {
1922                 spin_unlock(&fc->lock);
1923                 fuse_copy_finish(cs);
1924                 spin_lock(&fc->lock);
1925                 request_end(fc, req);
1926                 return -ENOENT;
1927         }
1928         /* Is it an interrupt reply? */
1929         if (req->intr_unique == oh.unique) {
1930                 err = -EINVAL;
1931                 if (nbytes != sizeof(struct fuse_out_header))
1932                         goto err_unlock;
1933
1934                 if (oh.error == -ENOSYS)
1935                         fc->no_interrupt = 1;
1936                 else if (oh.error == -EAGAIN)
1937                         queue_interrupt(fc, req);
1938
1939                 spin_unlock(&fc->lock);
1940                 fuse_copy_finish(cs);
1941                 return nbytes;
1942         }
1943
1944         req->state = FUSE_REQ_WRITING;
1945         list_move(&req->list, &fc->io);
1946         req->out.h = oh;
1947         req->locked = 1;
1948         cs->req = req;
1949         if (!req->out.page_replace)
1950                 cs->move_pages = 0;
1951         spin_unlock(&fc->lock);
1952
1953         err = copy_out_args(cs, &req->out, nbytes);
1954         fuse_copy_finish(cs);
1955
1956         spin_lock(&fc->lock);
1957         req->locked = 0;
1958         if (!err) {
1959                 if (req->aborted)
1960                         err = -ENOENT;
1961         } else if (!req->aborted)
1962                 req->out.h.error = -EIO;
1963         request_end(fc, req);
1964
1965         return err ? err : nbytes;
1966
1967  err_unlock:
1968         spin_unlock(&fc->lock);
1969  err_finish:
1970         fuse_copy_finish(cs);
1971         return err;
1972 }
1973
1974 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
1975                               unsigned long nr_segs, loff_t pos)
1976 {
1977         struct fuse_copy_state cs;
1978         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1979         if (!fc)
1980                 return -EPERM;
1981
1982         fuse_copy_init(&cs, fc, 0, iov, nr_segs);
1983
1984         return fuse_dev_do_write(fc, &cs, iov_length(iov, nr_segs));
1985 }
1986
1987 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1988                                      struct file *out, loff_t *ppos,
1989                                      size_t len, unsigned int flags)
1990 {
1991         unsigned nbuf;
1992         unsigned idx;
1993         struct pipe_buffer *bufs;
1994         struct fuse_copy_state cs;
1995         struct fuse_conn *fc;
1996         size_t rem;
1997         ssize_t ret;
1998
1999         fc = fuse_get_conn(out);
2000         if (!fc)
2001                 return -EPERM;
2002
2003         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
2004         if (!bufs)
2005                 return -ENOMEM;
2006
2007         pipe_lock(pipe);
2008         nbuf = 0;
2009         rem = 0;
2010         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2011                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2012
2013         ret = -EINVAL;
2014         if (rem < len) {
2015                 pipe_unlock(pipe);
2016                 goto out;
2017         }
2018
2019         rem = len;
2020         while (rem) {
2021                 struct pipe_buffer *ibuf;
2022                 struct pipe_buffer *obuf;
2023
2024                 BUG_ON(nbuf >= pipe->buffers);
2025                 BUG_ON(!pipe->nrbufs);
2026                 ibuf = &pipe->bufs[pipe->curbuf];
2027                 obuf = &bufs[nbuf];
2028
2029                 if (rem >= ibuf->len) {
2030                         *obuf = *ibuf;
2031                         ibuf->ops = NULL;
2032                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2033                         pipe->nrbufs--;
2034                 } else {
2035                         ibuf->ops->get(pipe, ibuf);
2036                         *obuf = *ibuf;
2037                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2038                         obuf->len = rem;
2039                         ibuf->offset += obuf->len;
2040                         ibuf->len -= obuf->len;
2041                 }
2042                 nbuf++;
2043                 rem -= obuf->len;
2044         }
2045         pipe_unlock(pipe);
2046
2047         fuse_copy_init(&cs, fc, 0, NULL, nbuf);
2048         cs.pipebufs = bufs;
2049         cs.pipe = pipe;
2050
2051         if (flags & SPLICE_F_MOVE)
2052                 cs.move_pages = 1;
2053
2054         ret = fuse_dev_do_write(fc, &cs, len);
2055
2056         for (idx = 0; idx < nbuf; idx++) {
2057                 struct pipe_buffer *buf = &bufs[idx];
2058                 buf->ops->release(pipe, buf);
2059         }
2060 out:
2061         kfree(bufs);
2062         return ret;
2063 }
2064
2065 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2066 {
2067         unsigned mask = POLLOUT | POLLWRNORM;
2068         struct fuse_conn *fc = fuse_get_conn(file);
2069         if (!fc)
2070                 return POLLERR;
2071
2072         poll_wait(file, &fc->waitq, wait);
2073
2074         spin_lock(&fc->lock);
2075         if (!fc->connected)
2076                 mask = POLLERR;
2077         else if (request_pending(fc))
2078                 mask |= POLLIN | POLLRDNORM;
2079         spin_unlock(&fc->lock);
2080
2081         return mask;
2082 }
2083
2084 /*
2085  * Abort all requests on the given list (pending or processing)
2086  *
2087  * This function releases and reacquires fc->lock
2088  */
2089 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2090 __releases(fc->lock)
2091 __acquires(fc->lock)
2092 {
2093         while (!list_empty(head)) {
2094                 struct fuse_req *req;
2095                 req = list_entry(head->next, struct fuse_req, list);
2096                 req->out.h.error = -ECONNABORTED;
2097                 request_end(fc, req);
2098                 spin_lock(&fc->lock);
2099         }
2100 }
2101
2102 /*
2103  * Abort requests under I/O
2104  *
2105  * The requests are set to aborted and finished, and the request
2106  * waiter is woken up.  This will make request_wait_answer() wait
2107  * until the request is unlocked and then return.
2108  *
2109  * If the request is asynchronous, then the end function needs to be
2110  * called after waiting for the request to be unlocked (if it was
2111  * locked).
2112  */
2113 static void end_io_requests(struct fuse_conn *fc)
2114 __releases(fc->lock)
2115 __acquires(fc->lock)
2116 {
2117         while (!list_empty(&fc->io)) {
2118                 struct fuse_req *req =
2119                         list_entry(fc->io.next, struct fuse_req, list);
2120                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
2121
2122                 req->aborted = 1;
2123                 req->out.h.error = -ECONNABORTED;
2124                 req->state = FUSE_REQ_FINISHED;
2125                 list_del_init(&req->list);
2126                 wake_up(&req->waitq);
2127                 if (end) {
2128                         req->end = NULL;
2129                         __fuse_get_request(req);
2130                         spin_unlock(&fc->lock);
2131                         wait_event(req->waitq, !req->locked);
2132                         end(fc, req);
2133                         fuse_put_request(fc, req);
2134                         spin_lock(&fc->lock);
2135                 }
2136         }
2137 }
2138
2139 static void end_queued_requests(struct fuse_conn *fc)
2140 __releases(fc->lock)
2141 __acquires(fc->lock)
2142 {
2143         fc->max_background = UINT_MAX;
2144         flush_bg_queue(fc);
2145         end_requests(fc, &fc->pending);
2146         end_requests(fc, &fc->processing);
2147         while (forget_pending(fc))
2148                 kfree(dequeue_forget(fc, 1, NULL));
2149 }
2150
2151 static void end_polls(struct fuse_conn *fc)
2152 {
2153         struct rb_node *p;
2154
2155         p = rb_first(&fc->polled_files);
2156
2157         while (p) {
2158                 struct fuse_file *ff;
2159                 ff = rb_entry(p, struct fuse_file, polled_node);
2160                 wake_up_interruptible_all(&ff->poll_wait);
2161
2162                 p = rb_next(p);
2163         }
2164 }
2165
2166 /*
2167  * Abort all requests.
2168  *
2169  * Emergency exit in case of a malicious or accidental deadlock, or
2170  * just a hung filesystem.
2171  *
2172  * The same effect is usually achievable through killing the
2173  * filesystem daemon and all users of the filesystem.  The exception
2174  * is the combination of an asynchronous request and the tricky
2175  * deadlock (see Documentation/filesystems/fuse.txt).
2176  *
2177  * During the aborting, progression of requests from the pending and
2178  * processing lists onto the io list, and progression of new requests
2179  * onto the pending list is prevented by req->connected being false.
2180  *
2181  * Progression of requests under I/O to the processing list is
2182  * prevented by the req->aborted flag being true for these requests.
2183  * For this reason requests on the io list must be aborted first.
2184  */
2185 void fuse_abort_conn(struct fuse_conn *fc)
2186 {
2187         spin_lock(&fc->lock);
2188         if (fc->connected) {
2189                 fc->connected = 0;
2190                 fc->blocked = 0;
2191                 fuse_set_initialized(fc);
2192                 end_io_requests(fc);
2193                 end_queued_requests(fc);
2194                 end_polls(fc);
2195                 wake_up_all(&fc->waitq);
2196                 wake_up_all(&fc->blocked_waitq);
2197                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
2198         }
2199         spin_unlock(&fc->lock);
2200 }
2201 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2202
2203 int fuse_dev_release(struct inode *inode, struct file *file)
2204 {
2205         struct fuse_conn *fc = fuse_get_conn(file);
2206         if (fc) {
2207                 spin_lock(&fc->lock);
2208                 fc->connected = 0;
2209                 fc->blocked = 0;
2210                 fuse_set_initialized(fc);
2211                 end_queued_requests(fc);
2212                 end_polls(fc);
2213                 wake_up_all(&fc->blocked_waitq);
2214                 spin_unlock(&fc->lock);
2215                 fuse_conn_put(fc);
2216         }
2217
2218         return 0;
2219 }
2220 EXPORT_SYMBOL_GPL(fuse_dev_release);
2221
2222 static int fuse_dev_fasync(int fd, struct file *file, int on)
2223 {
2224         struct fuse_conn *fc = fuse_get_conn(file);
2225         if (!fc)
2226                 return -EPERM;
2227
2228         /* No locking - fasync_helper does its own locking */
2229         return fasync_helper(fd, file, on, &fc->fasync);
2230 }
2231
2232 const struct file_operations fuse_dev_operations = {
2233         .owner          = THIS_MODULE,
2234         .open           = fuse_dev_open,
2235         .llseek         = no_llseek,
2236         .read           = do_sync_read,
2237         .aio_read       = fuse_dev_read,
2238         .splice_read    = fuse_dev_splice_read,
2239         .write          = do_sync_write,
2240         .aio_write      = fuse_dev_write,
2241         .splice_write   = fuse_dev_splice_write,
2242         .poll           = fuse_dev_poll,
2243         .release        = fuse_dev_release,
2244         .fasync         = fuse_dev_fasync,
2245 };
2246 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2247
2248 static struct miscdevice fuse_miscdevice = {
2249         .minor = FUSE_MINOR,
2250         .name  = "fuse",
2251         .fops = &fuse_dev_operations,
2252 };
2253
2254 int __init fuse_dev_init(void)
2255 {
2256         int err = -ENOMEM;
2257         fuse_req_cachep = kmem_cache_create("fuse_request",
2258                                             sizeof(struct fuse_req),
2259                                             0, 0, NULL);
2260         if (!fuse_req_cachep)
2261                 goto out;
2262
2263         err = misc_register(&fuse_miscdevice);
2264         if (err)
2265                 goto out_cache_clean;
2266
2267         return 0;
2268
2269  out_cache_clean:
2270         kmem_cache_destroy(fuse_req_cachep);
2271  out:
2272         return err;
2273 }
2274
2275 void fuse_dev_cleanup(void)
2276 {
2277         misc_deregister(&fuse_miscdevice);
2278         kmem_cache_destroy(fuse_req_cachep);
2279 }