fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/pipe_fs_i.h>
  20 #include <linux/swap.h>
  21 #include <linux/splice.h>
  22
  23 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  24 MODULE_ALIAS("devname:fuse");
  25
  26 static struct kmem_cache *fuse_req_cachep;
  27
  28 static struct fuse_conn *fuse_get_conn(struct file *file)
  29 {
  30         /*
  31          * Lockless access is OK, because file->private data is set
  32          * once during mount and is valid until the file is released.
  33          */
  34         return file->private_data;
  35 }
  36
  37 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  38                               struct fuse_page_desc *page_descs,
  39                               unsigned npages)
  40 {
  41         memset(req, 0, sizeof(*req));
  42         memset(pages, 0, sizeof(*pages) * npages);
  43         memset(page_descs, 0, sizeof(*page_descs) * npages);
  44         INIT_LIST_HEAD(&req->list);
  45         INIT_LIST_HEAD(&req->intr_entry);
  46         init_waitqueue_head(&req->waitq);
  47         atomic_set(&req->count, 1);
  48         req->pages = pages;
  49         req->page_descs = page_descs;
  50         req->max_pages = npages;
  51 }
  52
  53 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  54 {
  55         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, flags);
  56         if (req) {
  57                 struct page **pages;
  58                 struct fuse_page_desc *page_descs;
  59
  60                 if (npages <= FUSE_REQ_INLINE_PAGES) {
  61                         pages = req->inline_pages;
  62                         page_descs = req->inline_page_descs;
  63                 } else {
  64                         pages = kmalloc(sizeof(struct page *) * npages, flags);
  65                         page_descs = kmalloc(sizeof(struct fuse_page_desc) *
  66                                              npages, flags);
  67                 }
  68
  69                 if (!pages || !page_descs) {
  70                         kfree(pages);
  71                         kfree(page_descs);
  72                         kmem_cache_free(fuse_req_cachep, req);
  73                         return NULL;
  74                 }
  75
  76                 fuse_request_init(req, pages, page_descs, npages);
  77         }
  78         return req;
  79 }
  80
  81 struct fuse_req *fuse_request_alloc(unsigned npages)
  82 {
  83         return __fuse_request_alloc(npages, GFP_KERNEL);
  84 }
  85 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  86
  87 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  88 {
  89         return __fuse_request_alloc(npages, GFP_NOFS);
  90 }
  91
  92 void fuse_request_free(struct fuse_req *req)
  93 {
  94         if (req->pages != req->inline_pages) {
  95                 kfree(req->pages);
  96                 kfree(req->page_descs);
  97         }
  98         kmem_cache_free(fuse_req_cachep, req);
  99 }
 100
 101 static void block_sigs(sigset_t *oldset)
 102 {
 103         sigset_t mask;
 104
 105         siginitsetinv(&mask, sigmask(SIGKILL));
 106         sigprocmask(SIG_BLOCK, &mask, oldset);
 107 }
 108
 109 static void restore_sigs(sigset_t *oldset)
 110 {
 111         sigprocmask(SIG_SETMASK, oldset, NULL);
 112 }
 113
 114 void __fuse_get_request(struct fuse_req *req)
 115 {
 116         atomic_inc(&req->count);
 117 }
 118
 119 /* Must be called with > 1 refcount */
 120 static void __fuse_put_request(struct fuse_req *req)
 121 {
 122         BUG_ON(atomic_read(&req->count) < 2);
 123         atomic_dec(&req->count);
 124 }
 125
 126 static void fuse_req_init_context(struct fuse_req *req)
 127 {
 128         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 129         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 130         req->in.h.pid = current->pid;
 131 }
 132
 133 void fuse_set_initialized(struct fuse_conn *fc)
 134 {
 135         /* Make sure stores before this are seen on another CPU */
 136         smp_wmb();
 137         fc->initialized = 1;
 138 }
 139
 140 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 141 {
 142         return !fc->initialized || (for_background && fc->blocked);
 143 }
 144
 145 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 146                                        bool for_background)
 147 {
 148         struct fuse_req *req;
 149         int err;
 150         atomic_inc(&fc->num_waiting);
 151
 152         if (fuse_block_alloc(fc, for_background)) {
 153                 sigset_t oldset;
 154                 int intr;
 155
 156                 block_sigs(&oldset);
 157                 intr = wait_event_interruptible_exclusive(fc->blocked_waitq,
 158                                 !fuse_block_alloc(fc, for_background));
 159                 restore_sigs(&oldset);
 160                 err = -EINTR;
 161                 if (intr)
 162                         goto out;
 163         }
 164         /* Matches smp_wmb() in fuse_set_initialized() */
 165         smp_rmb();
 166
 167         err = -ENOTCONN;
 168         if (!fc->connected)
 169                 goto out;
 170
 171         req = fuse_request_alloc(npages);
 172         err = -ENOMEM;
 173         if (!req) {
 174                 if (for_background)
 175                         wake_up(&fc->blocked_waitq);
 176                 goto out;
 177         }
 178
 179         fuse_req_init_context(req);
 180         req->waiting = 1;
 181         req->background = for_background;
 182         return req;
 183
 184  out:
 185         atomic_dec(&fc->num_waiting);
 186         return ERR_PTR(err);
 187 }
 188
 189 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 190 {
 191         return __fuse_get_req(fc, npages, false);
 192 }
 193 EXPORT_SYMBOL_GPL(fuse_get_req);
 194
 195 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 196                                              unsigned npages)
 197 {
 198         return __fuse_get_req(fc, npages, true);
 199 }
 200 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 201
 202 /*
 203  * Return request in fuse_file->reserved_req.  However that may
 204  * currently be in use.  If that is the case, wait for it to become
 205  * available.
 206  */
 207 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 208                                          struct file *file)
 209 {
 210         struct fuse_req *req = NULL;
 211         struct fuse_file *ff = file->private_data;
 212
 213         do {
 214                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 215                 spin_lock(&fc->lock);
 216                 if (ff->reserved_req) {
 217                         req = ff->reserved_req;
 218                         ff->reserved_req = NULL;
 219                         req->stolen_file = get_file(file);
 220                 }
 221                 spin_unlock(&fc->lock);
 222         } while (!req);
 223
 224         return req;
 225 }
 226
 227 /*
 228  * Put stolen request back into fuse_file->reserved_req
 229  */
 230 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 231 {
 232         struct file *file = req->stolen_file;
 233         struct fuse_file *ff = file->private_data;
 234
 235         spin_lock(&fc->lock);
 236         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 237         BUG_ON(ff->reserved_req);
 238         ff->reserved_req = req;
 239         wake_up_all(&fc->reserved_req_waitq);
 240         spin_unlock(&fc->lock);
 241         fput(file);
 242 }
 243
 244 /*
 245  * Gets a requests for a file operation, always succeeds
 246  *
 247  * This is used for sending the FLUSH request, which must get to
 248  * userspace, due to POSIX locks which may need to be unlocked.
 249  *
 250  * If allocation fails due to OOM, use the reserved request in
 251  * fuse_file.
 252  *
 253  * This is very unlikely to deadlock accidentally, since the
 254  * filesystem should not have it's own file open.  If deadlock is
 255  * intentional, it can still be broken by "aborting" the filesystem.
 256  */
 257 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 258                                              struct file *file)
 259 {
 260         struct fuse_req *req;
 261
 262         atomic_inc(&fc->num_waiting);
 263         wait_event(fc->blocked_waitq, fc->initialized);
 264         /* Matches smp_wmb() in fuse_set_initialized() */
 265         smp_rmb();
 266         req = fuse_request_alloc(0);
 267         if (!req)
 268                 req = get_reserved_req(fc, file);
 269
 270         fuse_req_init_context(req);
 271         req->waiting = 1;
 272         req->background = 0;
 273         return req;
 274 }
 275
 276 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 277 {
 278         if (atomic_dec_and_test(&req->count)) {
 279                 if (unlikely(req->background)) {
 280                         /*
 281                          * We get here in the unlikely case that a background
 282                          * request was allocated but not sent
 283                          */
 284                         spin_lock(&fc->lock);
 285                         if (!fc->blocked)
 286                                 wake_up(&fc->blocked_waitq);
 287                         spin_unlock(&fc->lock);
 288                 }
 289
 290                 if (req->waiting)
 291                         atomic_dec(&fc->num_waiting);
 292
 293                 if (req->stolen_file)
 294                         put_reserved_req(fc, req);
 295                 else
 296                         fuse_request_free(req);
 297         }
 298 }
 299 EXPORT_SYMBOL_GPL(fuse_put_request);
 300
 301 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 302 {
 303         unsigned nbytes = 0;
 304         unsigned i;
 305
 306         for (i = 0; i < numargs; i++)
 307                 nbytes += args[i].size;
 308
 309         return nbytes;
 310 }
 311
 312 static u64 fuse_get_unique(struct fuse_conn *fc)
 313 {
 314         fc->reqctr++;
 315         /* zero is special */
 316         if (fc->reqctr == 0)
 317                 fc->reqctr = 1;
 318
 319         return fc->reqctr;
 320 }
 321
 322 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 323 {
 324         req->in.h.len = sizeof(struct fuse_in_header) +
 325                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 326         list_add_tail(&req->list, &fc->pending);
 327         req->state = FUSE_REQ_PENDING;
 328         if (!req->waiting) {
 329                 req->waiting = 1;
 330                 atomic_inc(&fc->num_waiting);
 331         }
 332         wake_up(&fc->waitq);
 333         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 334 }
 335
 336 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 337                        u64 nodeid, u64 nlookup)
 338 {
 339         forget->forget_one.nodeid = nodeid;
 340         forget->forget_one.nlookup = nlookup;
 341
 342         spin_lock(&fc->lock);
 343         if (fc->connected) {
 344                 fc->forget_list_tail->next = forget;
 345                 fc->forget_list_tail = forget;
 346                 wake_up(&fc->waitq);
 347                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 348         } else {
 349                 kfree(forget);
 350         }
 351         spin_unlock(&fc->lock);
 352 }
 353
 354 static void flush_bg_queue(struct fuse_conn *fc)
 355 {
 356         while (fc->active_background < fc->max_background &&
 357                !list_empty(&fc->bg_queue)) {
 358                 struct fuse_req *req;
 359
 360                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 361                 list_del(&req->list);
 362                 fc->active_background++;
 363                 req->in.h.unique = fuse_get_unique(fc);
 364                 queue_request(fc, req);
 365         }
 366 }
 367
 368 /*
 369  * This function is called when a request is finished.  Either a reply
 370  * has arrived or it was aborted (and not yet sent) or some error
 371  * occurred during communication with userspace, or the device file
 372  * was closed.  The requester thread is woken up (if still waiting),
 373  * the 'end' callback is called if given, else the reference to the
 374  * request is released
 375  *
 376  * Called with fc->lock, unlocks it
 377  */
 378 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 379 __releases(fc->lock)
 380 {
 381         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 382         req->end = NULL;
 383         list_del(&req->list);
 384         list_del(&req->intr_entry);
 385         req->state = FUSE_REQ_FINISHED;
 386         if (req->background) {
 387                 req->background = 0;
 388
 389                 if (fc->num_background == fc->max_background)
 390                         fc->blocked = 0;
 391
 392                 /* Wake up next waiter, if any */
 393                 if (!fc->blocked && waitqueue_active(&fc->blocked_waitq))
 394                         wake_up(&fc->blocked_waitq);
 395
 396                 if (fc->num_background == fc->congestion_threshold &&
 397                     fc->connected && fc->bdi_initialized) {
 398                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 399                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 400                 }
 401                 fc->num_background--;
 402                 fc->active_background--;
 403                 flush_bg_queue(fc);
 404         }
 405         spin_unlock(&fc->lock);
 406         wake_up(&req->waitq);
 407         if (end)
 408                 end(fc, req);
 409         fuse_put_request(fc, req);
 410 }
 411
 412 static void wait_answer_interruptible(struct fuse_conn *fc,
 413                                       struct fuse_req *req)
 414 __releases(fc->lock)
 415 __acquires(fc->lock)
 416 {
 417         if (signal_pending(current))
 418                 return;
 419
 420         spin_unlock(&fc->lock);
 421         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 422         spin_lock(&fc->lock);
 423 }
 424
 425 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 426 {
 427         list_add_tail(&req->intr_entry, &fc->interrupts);
 428         wake_up(&fc->waitq);
 429         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 430 }
 431
 432 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 433 __releases(fc->lock)
 434 __acquires(fc->lock)
 435 {
 436         if (!fc->no_interrupt) {
 437                 /* Any signal may interrupt this */
 438                 wait_answer_interruptible(fc, req);
 439
 440                 if (req->aborted)
 441                         goto aborted;
 442                 if (req->state == FUSE_REQ_FINISHED)
 443                         return;
 444
 445                 req->interrupted = 1;
 446                 if (req->state == FUSE_REQ_SENT)
 447                         queue_interrupt(fc, req);
 448         }
 449
 450         if (!req->force) {
 451                 sigset_t oldset;
 452
 453                 /* Only fatal signals may interrupt this */
 454                 block_sigs(&oldset);
 455                 wait_answer_interruptible(fc, req);
 456                 restore_sigs(&oldset);
 457
 458                 if (req->aborted)
 459                         goto aborted;
 460                 if (req->state == FUSE_REQ_FINISHED)
 461                         return;
 462
 463                 /* Request is not yet in userspace, bail out */
 464                 if (req->state == FUSE_REQ_PENDING) {
 465                         list_del(&req->list);
 466                         __fuse_put_request(req);
 467                         req->out.h.error = -EINTR;
 468                         return;
 469                 }
 470         }
 471
 472         /*
 473          * Either request is already in userspace, or it was forced.
 474          * Wait it out.
 475          */
 476         spin_unlock(&fc->lock);
 477         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 478         spin_lock(&fc->lock);
 479
 480         if (!req->aborted)
 481                 return;
 482
 483  aborted:
 484         BUG_ON(req->state != FUSE_REQ_FINISHED);
 485         if (req->locked) {
 486                 /* This is uninterruptible sleep, because data is
 487                    being copied to/from the buffers of req.  During
 488                    locked state, there mustn't be any filesystem
 489                    operation (e.g. page fault), since that could lead
 490                    to deadlock */
 491                 spin_unlock(&fc->lock);
 492                 wait_event(req->waitq, !req->locked);
 493                 spin_lock(&fc->lock);
 494         }
 495 }
 496
 497 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 498 {
 499         BUG_ON(req->background);
 500         spin_lock(&fc->lock);
 501         if (!fc->connected)
 502                 req->out.h.error = -ENOTCONN;
 503         else if (fc->conn_error)
 504                 req->out.h.error = -ECONNREFUSED;
 505         else {
 506                 req->in.h.unique = fuse_get_unique(fc);
 507                 queue_request(fc, req);
 508                 /* acquire extra reference, since request is still needed
 509                    after request_end() */
 510                 __fuse_get_request(req);
 511
 512                 request_wait_answer(fc, req);
 513         }
 514         spin_unlock(&fc->lock);
 515 }
 516
 517 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 518 {
 519         req->isreply = 1;
 520         __fuse_request_send(fc, req);
 521 }
 522 EXPORT_SYMBOL_GPL(fuse_request_send);
 523
 524 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 525 {
 526         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 527                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 528
 529         if (fc->minor < 9) {
 530                 switch (args->in.h.opcode) {
 531                 case FUSE_LOOKUP:
 532                 case FUSE_CREATE:
 533                 case FUSE_MKNOD:
 534                 case FUSE_MKDIR:
 535                 case FUSE_SYMLINK:
 536                 case FUSE_LINK:
 537                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 538                         break;
 539                 case FUSE_GETATTR:
 540                 case FUSE_SETATTR:
 541                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 542                         break;
 543                 }
 544         }
 545         if (fc->minor < 12) {
 546                 switch (args->in.h.opcode) {
 547                 case FUSE_CREATE:
 548                         args->in.args[0].size = sizeof(struct fuse_open_in);
 549                         break;
 550                 case FUSE_MKNOD:
 551                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 552                         break;
 553                 }
 554         }
 555 }
 556
 557 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 558 {
 559         struct fuse_req *req;
 560         ssize_t ret;
 561
 562         req = fuse_get_req(fc, 0);
 563         if (IS_ERR(req))
 564                 return PTR_ERR(req);
 565
 566         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 567         fuse_adjust_compat(fc, args);
 568
 569         req->in.h.opcode = args->in.h.opcode;
 570         req->in.h.nodeid = args->in.h.nodeid;
 571         req->in.numargs = args->in.numargs;
 572         memcpy(req->in.args, args->in.args,
 573                args->in.numargs * sizeof(struct fuse_in_arg));
 574         req->out.argvar = args->out.argvar;
 575         req->out.numargs = args->out.numargs;
 576         memcpy(req->out.args, args->out.args,
 577                args->out.numargs * sizeof(struct fuse_arg));
 578         fuse_request_send(fc, req);
 579         ret = req->out.h.error;
 580         if (!ret && args->out.argvar) {
 581                 BUG_ON(args->out.numargs != 1);
 582                 ret = req->out.args[0].size;
 583         }
 584         fuse_put_request(fc, req);
 585
 586         return ret;
 587 }
 588
 589 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 590                                             struct fuse_req *req)
 591 {
 592         BUG_ON(!req->background);
 593         fc->num_background++;
 594         if (fc->num_background == fc->max_background)
 595                 fc->blocked = 1;
 596         if (fc->num_background == fc->congestion_threshold &&
 597             fc->bdi_initialized) {
 598                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 599                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 600         }
 601         list_add_tail(&req->list, &fc->bg_queue);
 602         flush_bg_queue(fc);
 603 }
 604
 605 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 606 {
 607         spin_lock(&fc->lock);
 608         if (fc->connected) {
 609                 fuse_request_send_nowait_locked(fc, req);
 610                 spin_unlock(&fc->lock);
 611         } else {
 612                 req->out.h.error = -ENOTCONN;
 613                 request_end(fc, req);
 614         }
 615 }
 616
 617 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 618 {
 619         req->isreply = 1;
 620         fuse_request_send_nowait(fc, req);
 621 }
 622 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 623
 624 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 625                                           struct fuse_req *req, u64 unique)
 626 {
 627         int err = -ENODEV;
 628
 629         req->isreply = 0;
 630         req->in.h.unique = unique;
 631         spin_lock(&fc->lock);
 632         if (fc->connected) {
 633                 queue_request(fc, req);
 634                 err = 0;
 635         }
 636         spin_unlock(&fc->lock);
 637
 638         return err;
 639 }
 640
 641 /*
 642  * Called under fc->lock
 643  *
 644  * fc->connected must have been checked previously
 645  */
 646 void fuse_request_send_background_locked(struct fuse_conn *fc,
 647                                          struct fuse_req *req)
 648 {
 649         req->isreply = 1;
 650         fuse_request_send_nowait_locked(fc, req);
 651 }
 652
 653 void fuse_force_forget(struct file *file, u64 nodeid)
 654 {
 655         struct inode *inode = file_inode(file);
 656         struct fuse_conn *fc = get_fuse_conn(inode);
 657         struct fuse_req *req;
 658         struct fuse_forget_in inarg;
 659
 660         memset(&inarg, 0, sizeof(inarg));
 661         inarg.nlookup = 1;
 662         req = fuse_get_req_nofail_nopages(fc, file);
 663         req->in.h.opcode = FUSE_FORGET;
 664         req->in.h.nodeid = nodeid;
 665         req->in.numargs = 1;
 666         req->in.args[0].size = sizeof(inarg);
 667         req->in.args[0].value = &inarg;
 668         req->isreply = 0;
 669         __fuse_request_send(fc, req);
 670         /* ignore errors */
 671         fuse_put_request(fc, req);
 672 }
 673
 674 /*
 675  * Lock the request.  Up to the next unlock_request() there mustn't be
 676  * anything that could cause a page-fault.  If the request was already
 677  * aborted bail out.
 678  */
 679 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 680 {
 681         int err = 0;
 682         if (req) {
 683                 spin_lock(&fc->lock);
 684                 if (req->aborted)
 685                         err = -ENOENT;
 686                 else
 687                         req->locked = 1;
 688                 spin_unlock(&fc->lock);
 689         }
 690         return err;
 691 }
 692
 693 /*
 694  * Unlock request.  If it was aborted during being locked, the
 695  * requester thread is currently waiting for it to be unlocked, so
 696  * wake it up.
 697  */
 698 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 699 {
 700         if (req) {
 701                 spin_lock(&fc->lock);
 702                 req->locked = 0;
 703                 if (req->aborted)
 704                         wake_up(&req->waitq);
 705                 spin_unlock(&fc->lock);
 706         }
 707 }
 708
 709 struct fuse_copy_state {
 710         struct fuse_conn *fc;
 711         int write;
 712         struct fuse_req *req;
 713         const struct iovec *iov;
 714         struct pipe_buffer *pipebufs;
 715         struct pipe_buffer *currbuf;
 716         struct pipe_inode_info *pipe;
 717         unsigned long nr_segs;
 718         unsigned long seglen;
 719         unsigned long addr;
 720         struct page *pg;
 721         unsigned len;
 722         unsigned offset;
 723         unsigned move_pages:1;
 724 };
 725
 726 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 727                            int write,
 728                            const struct iovec *iov, unsigned long nr_segs)
 729 {
 730         memset(cs, 0, sizeof(*cs));
 731         cs->fc = fc;
 732         cs->write = write;
 733         cs->iov = iov;
 734         cs->nr_segs = nr_segs;
 735 }
 736
 737 /* Unmap and put previous page of userspace buffer */
 738 static void fuse_copy_finish(struct fuse_copy_state *cs)
 739 {
 740         if (cs->currbuf) {
 741                 struct pipe_buffer *buf = cs->currbuf;
 742
 743                 if (cs->write)
 744                         buf->len = PAGE_SIZE - cs->len;
 745                 cs->currbuf = NULL;
 746         } else if (cs->pg) {
 747                 if (cs->write) {
 748                         flush_dcache_page(cs->pg);
 749                         set_page_dirty_lock(cs->pg);
 750                 }
 751                 put_page(cs->pg);
 752         }
 753         cs->pg = NULL;
 754 }
 755
 756 /*
 757  * Get another pagefull of userspace buffer, and map it to kernel
 758  * address space, and lock request
 759  */
 760 static int fuse_copy_fill(struct fuse_copy_state *cs)
 761 {
 762         struct page *page;
 763         int err;
 764
 765         unlock_request(cs->fc, cs->req);
 766         fuse_copy_finish(cs);
 767         if (cs->pipebufs) {
 768                 struct pipe_buffer *buf = cs->pipebufs;
 769
 770                 if (!cs->write) {
 771                         err = buf->ops->confirm(cs->pipe, buf);
 772                         if (err)
 773                                 return err;
 774
 775                         BUG_ON(!cs->nr_segs);
 776                         cs->currbuf = buf;
 777                         cs->pg = buf->page;
 778                         cs->offset = buf->offset;
 779                         cs->len = buf->len;
 780                         cs->pipebufs++;
 781                         cs->nr_segs--;
 782                 } else {
 783                         if (cs->nr_segs == cs->pipe->buffers)
 784                                 return -EIO;
 785
 786                         page = alloc_page(GFP_HIGHUSER);
 787                         if (!page)
 788                                 return -ENOMEM;
 789
 790                         buf->page = page;
 791                         buf->offset = 0;
 792                         buf->len = 0;
 793
 794                         cs->currbuf = buf;
 795                         cs->pg = page;
 796                         cs->offset = 0;
 797                         cs->len = PAGE_SIZE;
 798                         cs->pipebufs++;
 799                         cs->nr_segs++;
 800                 }
 801         } else {
 802                 if (!cs->seglen) {
 803                         BUG_ON(!cs->nr_segs);
 804                         cs->seglen = cs->iov[0].iov_len;
 805                         cs->addr = (unsigned long) cs->iov[0].iov_base;
 806                         cs->iov++;
 807                         cs->nr_segs--;
 808                 }
 809                 err = get_user_pages_fast(cs->addr, 1, cs->write, &page);
 810                 if (err < 0)
 811                         return err;
 812                 BUG_ON(err != 1);
 813                 cs->pg = page;
 814                 cs->offset = cs->addr % PAGE_SIZE;
 815                 cs->len = min(PAGE_SIZE - cs->offset, cs->seglen);
 816                 cs->seglen -= cs->len;
 817                 cs->addr += cs->len;
 818         }
 819
 820         return lock_request(cs->fc, cs->req);
 821 }
 822
 823 /* Do as much copy to/from userspace buffer as we can */
 824 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 825 {
 826         unsigned ncpy = min(*size, cs->len);
 827         if (val) {
 828                 void *pgaddr = kmap_atomic(cs->pg);
 829                 void *buf = pgaddr + cs->offset;
 830
 831                 if (cs->write)
 832                         memcpy(buf, *val, ncpy);
 833                 else
 834                         memcpy(*val, buf, ncpy);
 835
 836                 kunmap_atomic(pgaddr);
 837                 *val += ncpy;
 838         }
 839         *size -= ncpy;
 840         cs->len -= ncpy;
 841         cs->offset += ncpy;
 842         return ncpy;
 843 }
 844
 845 static int fuse_check_page(struct page *page)
 846 {
 847         if (page_mapcount(page) ||
 848             page->mapping != NULL ||
 849             page_count(page) != 1 ||
 850             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 851              ~(1 << PG_locked |
 852                1 << PG_referenced |
 853                1 << PG_uptodate |
 854                1 << PG_lru |
 855                1 << PG_active |
 856                1 << PG_reclaim))) {
 857                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 858                 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);
 859                 return 1;
 860         }
 861         return 0;
 862 }
 863
 864 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 865 {
 866         int err;
 867         struct page *oldpage = *pagep;
 868         struct page *newpage;
 869         struct pipe_buffer *buf = cs->pipebufs;
 870
 871         unlock_request(cs->fc, cs->req);
 872         fuse_copy_finish(cs);
 873
 874         err = buf->ops->confirm(cs->pipe, buf);
 875         if (err)
 876                 return err;
 877
 878         BUG_ON(!cs->nr_segs);
 879         cs->currbuf = buf;
 880         cs->len = buf->len;
 881         cs->pipebufs++;
 882         cs->nr_segs--;
 883
 884         if (cs->len != PAGE_SIZE)
 885                 goto out_fallback;
 886
 887         if (buf->ops->steal(cs->pipe, buf) != 0)
 888                 goto out_fallback;
 889
 890         newpage = buf->page;
 891
 892         if (WARN_ON(!PageUptodate(newpage)))
 893                 return -EIO;
 894
 895         ClearPageMappedToDisk(newpage);
 896
 897         if (fuse_check_page(newpage) != 0)
 898                 goto out_fallback_unlock;
 899
 900         /*
 901          * This is a new and locked page, it shouldn't be mapped or
 902          * have any special flags on it
 903          */
 904         if (WARN_ON(page_mapped(oldpage)))
 905                 goto out_fallback_unlock;
 906         if (WARN_ON(page_has_private(oldpage)))
 907                 goto out_fallback_unlock;
 908         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 909                 goto out_fallback_unlock;
 910         if (WARN_ON(PageMlocked(oldpage)))
 911                 goto out_fallback_unlock;
 912
 913         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 914         if (err) {
 915                 unlock_page(newpage);
 916                 return err;
 917         }
 918
 919         page_cache_get(newpage);
 920
 921         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 922                 lru_cache_add_file(newpage);
 923
 924         err = 0;
 925         spin_lock(&cs->fc->lock);
 926         if (cs->req->aborted)
 927                 err = -ENOENT;
 928         else
 929                 *pagep = newpage;
 930         spin_unlock(&cs->fc->lock);
 931
 932         if (err) {
 933                 unlock_page(newpage);
 934                 page_cache_release(newpage);
 935                 return err;
 936         }
 937
 938         unlock_page(oldpage);
 939         page_cache_release(oldpage);
 940         cs->len = 0;
 941
 942         return 0;
 943
 944 out_fallback_unlock:
 945         unlock_page(newpage);
 946 out_fallback:
 947         cs->pg = buf->page;
 948         cs->offset = buf->offset;
 949
 950         err = lock_request(cs->fc, cs->req);
 951         if (err)
 952                 return err;
 953
 954         return 1;
 955 }
 956
 957 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 958                          unsigned offset, unsigned count)
 959 {
 960         struct pipe_buffer *buf;
 961
 962         if (cs->nr_segs == cs->pipe->buffers)
 963                 return -EIO;
 964
 965         unlock_request(cs->fc, cs->req);
 966         fuse_copy_finish(cs);
 967
 968         buf = cs->pipebufs;
 969         page_cache_get(page);
 970         buf->page = page;
 971         buf->offset = offset;
 972         buf->len = count;
 973
 974         cs->pipebufs++;
 975         cs->nr_segs++;
 976         cs->len = 0;
 977
 978         return 0;
 979 }
 980
 981 /*
 982  * Copy a page in the request to/from the userspace buffer.  Must be
 983  * done atomically
 984  */
 985 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 986                           unsigned offset, unsigned count, int zeroing)
 987 {
 988         int err;
 989         struct page *page = *pagep;
 990
 991         if (page && zeroing && count < PAGE_SIZE)
 992                 clear_highpage(page);
 993
 994         while (count) {
 995                 if (cs->write && cs->pipebufs && page) {
 996                         return fuse_ref_page(cs, page, offset, count);
 997                 } else if (!cs->len) {
 998                         if (cs->move_pages && page &&
 999                             offset == 0 && count == PAGE_SIZE) {
1000                                 err = fuse_try_move_page(cs, pagep);
1001                                 if (err <= 0)
1002                                         return err;
1003                         } else {
1004                                 err = fuse_copy_fill(cs);
1005                                 if (err)
1006                                         return err;
1007                         }
1008                 }
1009                 if (page) {
1010                         void *mapaddr = kmap_atomic(page);
1011                         void *buf = mapaddr + offset;
1012                         offset += fuse_copy_do(cs, &buf, &count);
1013                         kunmap_atomic(mapaddr);
1014                 } else
1015                         offset += fuse_copy_do(cs, NULL, &count);
1016         }
1017         if (page && !cs->write)
1018                 flush_dcache_page(page);
1019         return 0;
1020 }
1021
1022 /* Copy pages in the request to/from userspace buffer */
1023 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1024                            int zeroing)
1025 {
1026         unsigned i;
1027         struct fuse_req *req = cs->req;
1028
1029         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1030                 int err;
1031                 unsigned offset = req->page_descs[i].offset;
1032                 unsigned count = min(nbytes, req->page_descs[i].length);
1033
1034                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1035                                      zeroing);
1036                 if (err)
1037                         return err;
1038
1039                 nbytes -= count;
1040         }
1041         return 0;
1042 }
1043
1044 /* Copy a single argument in the request to/from userspace buffer */
1045 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1046 {
1047         while (size) {
1048                 if (!cs->len) {
1049                         int err = fuse_copy_fill(cs);
1050                         if (err)
1051                                 return err;
1052                 }
1053                 fuse_copy_do(cs, &val, &size);
1054         }
1055         return 0;
1056 }
1057
1058 /* Copy request arguments to/from userspace buffer */
1059 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1060                           unsigned argpages, struct fuse_arg *args,
1061                           int zeroing)
1062 {
1063         int err = 0;
1064         unsigned i;
1065
1066         for (i = 0; !err && i < numargs; i++)  {
1067                 struct fuse_arg *arg = &args[i];
1068                 if (i == numargs - 1 && argpages)
1069                         err = fuse_copy_pages(cs, arg->size, zeroing);
1070                 else
1071                         err = fuse_copy_one(cs, arg->value, arg->size);
1072         }
1073         return err;
1074 }
1075
1076 static int forget_pending(struct fuse_conn *fc)
1077 {
1078         return fc->forget_list_head.next != NULL;
1079 }
1080
1081 static int request_pending(struct fuse_conn *fc)
1082 {
1083         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts) ||
1084                 forget_pending(fc);
1085 }
1086
1087 /* Wait until a request is available on the pending list */
1088 static void request_wait(struct fuse_conn *fc)
1089 __releases(fc->lock)
1090 __acquires(fc->lock)
1091 {
1092         DECLARE_WAITQUEUE(wait, current);
1093
1094         add_wait_queue_exclusive(&fc->waitq, &wait);
1095         while (fc->connected && !request_pending(fc)) {
1096                 set_current_state(TASK_INTERRUPTIBLE);
1097                 if (signal_pending(current))
1098                         break;
1099
1100                 spin_unlock(&fc->lock);
1101                 schedule();
1102                 spin_lock(&fc->lock);
1103         }
1104         set_current_state(TASK_RUNNING);
1105         remove_wait_queue(&fc->waitq, &wait);
1106 }
1107
1108 /*
1109  * Transfer an interrupt request to userspace
1110  *
1111  * Unlike other requests this is assembled on demand, without a need
1112  * to allocate a separate fuse_req structure.
1113  *
1114  * Called with fc->lock held, releases it
1115  */
1116 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
1117                                size_t nbytes, struct fuse_req *req)
1118 __releases(fc->lock)
1119 {
1120         struct fuse_in_header ih;
1121         struct fuse_interrupt_in arg;
1122         unsigned reqsize = sizeof(ih) + sizeof(arg);
1123         int err;
1124
1125         list_del_init(&req->intr_entry);
1126         req->intr_unique = fuse_get_unique(fc);
1127         memset(&ih, 0, sizeof(ih));
1128         memset(&arg, 0, sizeof(arg));
1129         ih.len = reqsize;
1130         ih.opcode = FUSE_INTERRUPT;
1131         ih.unique = req->intr_unique;
1132         arg.unique = req->in.h.unique;
1133
1134         spin_unlock(&fc->lock);
1135         if (nbytes < reqsize)
1136                 return -EINVAL;
1137
1138         err = fuse_copy_one(cs, &ih, sizeof(ih));
1139         if (!err)
1140                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1141         fuse_copy_finish(cs);
1142
1143         return err ? err : reqsize;
1144 }
1145
1146 static struct fuse_forget_link *dequeue_forget(struct fuse_conn *fc,
1147                                                unsigned max,
1148                                                unsigned *countp)
1149 {
1150         struct fuse_forget_link *head = fc->forget_list_head.next;
1151         struct fuse_forget_link **newhead = &head;
1152         unsigned count;
1153
1154         for (count = 0; *newhead != NULL && count < max; count++)
1155                 newhead = &(*newhead)->next;
1156
1157         fc->forget_list_head.next = *newhead;
1158         *newhead = NULL;
1159         if (fc->forget_list_head.next == NULL)
1160                 fc->forget_list_tail = &fc->forget_list_head;
1161
1162         if (countp != NULL)
1163                 *countp = count;
1164
1165         return head;
1166 }
1167
1168 static int fuse_read_single_forget(struct fuse_conn *fc,
1169                                    struct fuse_copy_state *cs,
1170                                    size_t nbytes)
1171 __releases(fc->lock)
1172 {
1173         int err;
1174         struct fuse_forget_link *forget = dequeue_forget(fc, 1, NULL);
1175         struct fuse_forget_in arg = {
1176                 .nlookup = forget->forget_one.nlookup,
1177         };
1178         struct fuse_in_header ih = {
1179                 .opcode = FUSE_FORGET,
1180                 .nodeid = forget->forget_one.nodeid,
1181                 .unique = fuse_get_unique(fc),
1182                 .len = sizeof(ih) + sizeof(arg),
1183         };
1184
1185         spin_unlock(&fc->lock);
1186         kfree(forget);
1187         if (nbytes < ih.len)
1188                 return -EINVAL;
1189
1190         err = fuse_copy_one(cs, &ih, sizeof(ih));
1191         if (!err)
1192                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1193         fuse_copy_finish(cs);
1194
1195         if (err)
1196                 return err;
1197
1198         return ih.len;
1199 }
1200
1201 static int fuse_read_batch_forget(struct fuse_conn *fc,
1202                                    struct fuse_copy_state *cs, size_t nbytes)
1203 __releases(fc->lock)
1204 {
1205         int err;
1206         unsigned max_forgets;
1207         unsigned count;
1208         struct fuse_forget_link *head;
1209         struct fuse_batch_forget_in arg = { .count = 0 };
1210         struct fuse_in_header ih = {
1211                 .opcode = FUSE_BATCH_FORGET,
1212                 .unique = fuse_get_unique(fc),
1213                 .len = sizeof(ih) + sizeof(arg),
1214         };
1215
1216         if (nbytes < ih.len) {
1217                 spin_unlock(&fc->lock);
1218                 return -EINVAL;
1219         }
1220
1221         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1222         head = dequeue_forget(fc, max_forgets, &count);
1223         spin_unlock(&fc->lock);
1224
1225         arg.count = count;
1226         ih.len += count * sizeof(struct fuse_forget_one);
1227         err = fuse_copy_one(cs, &ih, sizeof(ih));
1228         if (!err)
1229                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1230
1231         while (head) {
1232                 struct fuse_forget_link *forget = head;
1233
1234                 if (!err) {
1235                         err = fuse_copy_one(cs, &forget->forget_one,
1236                                             sizeof(forget->forget_one));
1237                 }
1238                 head = forget->next;
1239                 kfree(forget);
1240         }
1241
1242         fuse_copy_finish(cs);
1243
1244         if (err)
1245                 return err;
1246
1247         return ih.len;
1248 }
1249
1250 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_copy_state *cs,
1251                             size_t nbytes)
1252 __releases(fc->lock)
1253 {
1254         if (fc->minor < 16 || fc->forget_list_head.next->next == NULL)
1255                 return fuse_read_single_forget(fc, cs, nbytes);
1256         else
1257                 return fuse_read_batch_forget(fc, cs, nbytes);
1258 }
1259
1260 /*
1261  * Read a single request into the userspace filesystem's buffer.  This
1262  * function waits until a request is available, then removes it from
1263  * the pending list and copies request data to userspace buffer.  If
1264  * no reply is needed (FORGET) or request has been aborted or there
1265  * was an error during the copying then it's finished by calling
1266  * request_end().  Otherwise add it to the processing list, and set
1267  * the 'sent' flag.
1268  */
1269 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1270                                 struct fuse_copy_state *cs, size_t nbytes)
1271 {
1272         int err;
1273         struct fuse_req *req;
1274         struct fuse_in *in;
1275         unsigned reqsize;
1276
1277  restart:
1278         spin_lock(&fc->lock);
1279         err = -EAGAIN;
1280         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
1281             !request_pending(fc))
1282                 goto err_unlock;
1283
1284         request_wait(fc);
1285         err = -ENODEV;
1286         if (!fc->connected)
1287                 goto err_unlock;
1288         err = -ERESTARTSYS;
1289         if (!request_pending(fc))
1290                 goto err_unlock;
1291
1292         if (!list_empty(&fc->interrupts)) {
1293                 req = list_entry(fc->interrupts.next, struct fuse_req,
1294                                  intr_entry);
1295                 return fuse_read_interrupt(fc, cs, nbytes, req);
1296         }
1297
1298         if (forget_pending(fc)) {
1299                 if (list_empty(&fc->pending) || fc->forget_batch-- > 0)
1300                         return fuse_read_forget(fc, cs, nbytes);
1301
1302                 if (fc->forget_batch <= -8)
1303                         fc->forget_batch = 16;
1304         }
1305
1306         req = list_entry(fc->pending.next, struct fuse_req, list);
1307         req->state = FUSE_REQ_READING;
1308         list_move(&req->list, &fc->io);
1309
1310         in = &req->in;
1311         reqsize = in->h.len;
1312         /* If request is too large, reply with an error and restart the read */
1313         if (nbytes < reqsize) {
1314                 req->out.h.error = -EIO;
1315                 /* SETXATTR is special, since it may contain too large data */
1316                 if (in->h.opcode == FUSE_SETXATTR)
1317                         req->out.h.error = -E2BIG;
1318                 request_end(fc, req);
1319                 goto restart;
1320         }
1321         spin_unlock(&fc->lock);
1322         cs->req = req;
1323         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1324         if (!err)
1325                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1326                                      (struct fuse_arg *) in->args, 0);
1327         fuse_copy_finish(cs);
1328         spin_lock(&fc->lock);
1329         req->locked = 0;
1330         if (req->aborted) {
1331                 request_end(fc, req);
1332                 return -ENODEV;
1333         }
1334         if (err) {
1335                 req->out.h.error = -EIO;
1336                 request_end(fc, req);
1337                 return err;
1338         }
1339         if (!req->isreply)
1340                 request_end(fc, req);
1341         else {
1342                 req->state = FUSE_REQ_SENT;
1343                 list_move_tail(&req->list, &fc->processing);
1344                 if (req->interrupted)
1345                         queue_interrupt(fc, req);
1346                 spin_unlock(&fc->lock);
1347         }
1348         return reqsize;
1349
1350  err_unlock:
1351         spin_unlock(&fc->lock);
1352         return err;
1353 }
1354
1355 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
1356                               unsigned long nr_segs, loff_t pos)
1357 {
1358         struct fuse_copy_state cs;
1359         struct file *file = iocb->ki_filp;
1360         struct fuse_conn *fc = fuse_get_conn(file);
1361         if (!fc)
1362                 return -EPERM;
1363
1364         fuse_copy_init(&cs, fc, 1, iov, nr_segs);
1365
1366         return fuse_dev_do_read(fc, file, &cs, iov_length(iov, nr_segs));
1367 }
1368
1369 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1370                                     struct pipe_inode_info *pipe,
1371                                     size_t len, unsigned int flags)
1372 {
1373         int ret;
1374         int page_nr = 0;
1375         int do_wakeup = 0;
1376         struct pipe_buffer *bufs;
1377         struct fuse_copy_state cs;
1378         struct fuse_conn *fc = fuse_get_conn(in);
1379         if (!fc)
1380                 return -EPERM;
1381
1382         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1383         if (!bufs)
1384                 return -ENOMEM;
1385
1386         fuse_copy_init(&cs, fc, 1, NULL, 0);
1387         cs.pipebufs = bufs;
1388         cs.pipe = pipe;
1389         ret = fuse_dev_do_read(fc, in, &cs, len);
1390         if (ret < 0)
1391                 goto out;
1392
1393         ret = 0;
1394         pipe_lock(pipe);
1395
1396         if (!pipe->readers) {
1397                 send_sig(SIGPIPE, current, 0);
1398                 if (!ret)
1399                         ret = -EPIPE;
1400                 goto out_unlock;
1401         }
1402
1403         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1404                 ret = -EIO;
1405                 goto out_unlock;
1406         }
1407
1408         while (page_nr < cs.nr_segs) {
1409                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1410                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1411
1412                 buf->page = bufs[page_nr].page;
1413                 buf->offset = bufs[page_nr].offset;
1414                 buf->len = bufs[page_nr].len;
1415                 /*
1416                  * Need to be careful about this.  Having buf->ops in module
1417                  * code can Oops if the buffer persists after module unload.
1418                  */
1419                 buf->ops = &nosteal_pipe_buf_ops;
1420
1421                 pipe->nrbufs++;
1422                 page_nr++;
1423                 ret += buf->len;
1424
1425                 if (pipe->files)
1426                         do_wakeup = 1;
1427         }
1428
1429 out_unlock:
1430         pipe_unlock(pipe);
1431
1432         if (do_wakeup) {
1433                 smp_mb();
1434                 if (waitqueue_active(&pipe->wait))
1435                         wake_up_interruptible(&pipe->wait);
1436                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1437         }
1438
1439 out:
1440         for (; page_nr < cs.nr_segs; page_nr++)
1441                 page_cache_release(bufs[page_nr].page);
1442
1443         kfree(bufs);
1444         return ret;
1445 }
1446
1447 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1448                             struct fuse_copy_state *cs)
1449 {
1450         struct fuse_notify_poll_wakeup_out outarg;
1451         int err = -EINVAL;
1452
1453         if (size != sizeof(outarg))
1454                 goto err;
1455
1456         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1457         if (err)
1458                 goto err;
1459
1460         fuse_copy_finish(cs);
1461         return fuse_notify_poll_wakeup(fc, &outarg);
1462
1463 err:
1464         fuse_copy_finish(cs);
1465         return err;
1466 }
1467
1468 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1469                                    struct fuse_copy_state *cs)
1470 {
1471         struct fuse_notify_inval_inode_out outarg;
1472         int err = -EINVAL;
1473
1474         if (size != sizeof(outarg))
1475                 goto err;
1476
1477         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1478         if (err)
1479                 goto err;
1480         fuse_copy_finish(cs);
1481
1482         down_read(&fc->killsb);
1483         err = -ENOENT;
1484         if (fc->sb) {
1485                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1486                                                outarg.off, outarg.len);
1487         }
1488         up_read(&fc->killsb);
1489         return err;
1490
1491 err:
1492         fuse_copy_finish(cs);
1493         return err;
1494 }
1495
1496 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1497                                    struct fuse_copy_state *cs)
1498 {
1499         struct fuse_notify_inval_entry_out outarg;
1500         int err = -ENOMEM;
1501         char *buf;
1502         struct qstr name;
1503
1504         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1505         if (!buf)
1506                 goto err;
1507
1508         err = -EINVAL;
1509         if (size < sizeof(outarg))
1510                 goto err;
1511
1512         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1513         if (err)
1514                 goto err;
1515
1516         err = -ENAMETOOLONG;
1517         if (outarg.namelen > FUSE_NAME_MAX)
1518                 goto err;
1519
1520         err = -EINVAL;
1521         if (size != sizeof(outarg) + outarg.namelen + 1)
1522                 goto err;
1523
1524         name.name = buf;
1525         name.len = outarg.namelen;
1526         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1527         if (err)
1528                 goto err;
1529         fuse_copy_finish(cs);
1530         buf[outarg.namelen] = 0;
1531         name.hash = full_name_hash(name.name, name.len);
1532
1533         down_read(&fc->killsb);
1534         err = -ENOENT;
1535         if (fc->sb)
1536                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1537         up_read(&fc->killsb);
1538         kfree(buf);
1539         return err;
1540
1541 err:
1542         kfree(buf);
1543         fuse_copy_finish(cs);
1544         return err;
1545 }
1546
1547 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1548                               struct fuse_copy_state *cs)
1549 {
1550         struct fuse_notify_delete_out outarg;
1551         int err = -ENOMEM;
1552         char *buf;
1553         struct qstr name;
1554
1555         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1556         if (!buf)
1557                 goto err;
1558
1559         err = -EINVAL;
1560         if (size < sizeof(outarg))
1561                 goto err;
1562
1563         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1564         if (err)
1565                 goto err;
1566
1567         err = -ENAMETOOLONG;
1568         if (outarg.namelen > FUSE_NAME_MAX)
1569                 goto err;
1570
1571         err = -EINVAL;
1572         if (size != sizeof(outarg) + outarg.namelen + 1)
1573                 goto err;
1574
1575         name.name = buf;
1576         name.len = outarg.namelen;
1577         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1578         if (err)
1579                 goto err;
1580         fuse_copy_finish(cs);
1581         buf[outarg.namelen] = 0;
1582         name.hash = full_name_hash(name.name, name.len);
1583
1584         down_read(&fc->killsb);
1585         err = -ENOENT;
1586         if (fc->sb)
1587                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1588                                                outarg.child, &name);
1589         up_read(&fc->killsb);
1590         kfree(buf);
1591         return err;
1592
1593 err:
1594         kfree(buf);
1595         fuse_copy_finish(cs);
1596         return err;
1597 }
1598
1599 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1600                              struct fuse_copy_state *cs)
1601 {
1602         struct fuse_notify_store_out outarg;
1603         struct inode *inode;
1604         struct address_space *mapping;
1605         u64 nodeid;
1606         int err;
1607         pgoff_t index;
1608         unsigned int offset;
1609         unsigned int num;
1610         loff_t file_size;
1611         loff_t end;
1612
1613         err = -EINVAL;
1614         if (size < sizeof(outarg))
1615                 goto out_finish;
1616
1617         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1618         if (err)
1619                 goto out_finish;
1620
1621         err = -EINVAL;
1622         if (size - sizeof(outarg) != outarg.size)
1623                 goto out_finish;
1624
1625         nodeid = outarg.nodeid;
1626
1627         down_read(&fc->killsb);
1628
1629         err = -ENOENT;
1630         if (!fc->sb)
1631                 goto out_up_killsb;
1632
1633         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1634         if (!inode)
1635                 goto out_up_killsb;
1636
1637         mapping = inode->i_mapping;
1638         index = outarg.offset >> PAGE_CACHE_SHIFT;
1639         offset = outarg.offset & ~PAGE_CACHE_MASK;
1640         file_size = i_size_read(inode);
1641         end = outarg.offset + outarg.size;
1642         if (end > file_size) {
1643                 file_size = end;
1644                 fuse_write_update_size(inode, file_size);
1645         }
1646
1647         num = outarg.size;
1648         while (num) {
1649                 struct page *page;
1650                 unsigned int this_num;
1651
1652                 err = -ENOMEM;
1653                 page = find_or_create_page(mapping, index,
1654                                            mapping_gfp_mask(mapping));
1655                 if (!page)
1656                         goto out_iput;
1657
1658                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1659                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1660                 if (!err && offset == 0 &&
1661                     (this_num == PAGE_CACHE_SIZE || file_size == end))
1662                         SetPageUptodate(page);
1663                 unlock_page(page);
1664                 page_cache_release(page);
1665
1666                 if (err)
1667                         goto out_iput;
1668
1669                 num -= this_num;
1670                 offset = 0;
1671                 index++;
1672         }
1673
1674         err = 0;
1675
1676 out_iput:
1677         iput(inode);
1678 out_up_killsb:
1679         up_read(&fc->killsb);
1680 out_finish:
1681         fuse_copy_finish(cs);
1682         return err;
1683 }
1684
1685 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1686 {
1687         release_pages(req->pages, req->num_pages, false);
1688 }
1689
1690 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1691                          struct fuse_notify_retrieve_out *outarg)
1692 {
1693         int err;
1694         struct address_space *mapping = inode->i_mapping;
1695         struct fuse_req *req;
1696         pgoff_t index;
1697         loff_t file_size;
1698         unsigned int num;
1699         unsigned int offset;
1700         size_t total_len = 0;
1701         int num_pages;
1702
1703         offset = outarg->offset & ~PAGE_CACHE_MASK;
1704         file_size = i_size_read(inode);
1705
1706         num = outarg->size;
1707         if (outarg->offset > file_size)
1708                 num = 0;
1709         else if (outarg->offset + num > file_size)
1710                 num = file_size - outarg->offset;
1711
1712         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1713         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1714
1715         req = fuse_get_req(fc, num_pages);
1716         if (IS_ERR(req))
1717                 return PTR_ERR(req);
1718
1719         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1720         req->in.h.nodeid = outarg->nodeid;
1721         req->in.numargs = 2;
1722         req->in.argpages = 1;
1723         req->page_descs[0].offset = offset;
1724         req->end = fuse_retrieve_end;
1725
1726         index = outarg->offset >> PAGE_CACHE_SHIFT;
1727
1728         while (num && req->num_pages < num_pages) {
1729                 struct page *page;
1730                 unsigned int this_num;
1731
1732                 page = find_get_page(mapping, index);
1733                 if (!page)
1734                         break;
1735
1736                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1737                 req->pages[req->num_pages] = page;
1738                 req->page_descs[req->num_pages].length = this_num;
1739                 req->num_pages++;
1740
1741                 offset = 0;
1742                 num -= this_num;
1743                 total_len += this_num;
1744                 index++;
1745         }
1746         req->misc.retrieve_in.offset = outarg->offset;
1747         req->misc.retrieve_in.size = total_len;
1748         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1749         req->in.args[0].value = &req->misc.retrieve_in;
1750         req->in.args[1].size = total_len;
1751
1752         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1753         if (err)
1754                 fuse_retrieve_end(fc, req);
1755
1756         return err;
1757 }
1758
1759 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1760                                 struct fuse_copy_state *cs)
1761 {
1762         struct fuse_notify_retrieve_out outarg;
1763         struct inode *inode;
1764         int err;
1765
1766         err = -EINVAL;
1767         if (size != sizeof(outarg))
1768                 goto copy_finish;
1769
1770         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1771         if (err)
1772                 goto copy_finish;
1773
1774         fuse_copy_finish(cs);
1775
1776         down_read(&fc->killsb);
1777         err = -ENOENT;
1778         if (fc->sb) {
1779                 u64 nodeid = outarg.nodeid;
1780
1781                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1782                 if (inode) {
1783                         err = fuse_retrieve(fc, inode, &outarg);
1784                         iput(inode);
1785                 }
1786         }
1787         up_read(&fc->killsb);
1788
1789         return err;
1790
1791 copy_finish:
1792         fuse_copy_finish(cs);
1793         return err;
1794 }
1795
1796 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1797                        unsigned int size, struct fuse_copy_state *cs)
1798 {
1799         switch (code) {
1800         case FUSE_NOTIFY_POLL:
1801                 return fuse_notify_poll(fc, size, cs);
1802
1803         case FUSE_NOTIFY_INVAL_INODE:
1804                 return fuse_notify_inval_inode(fc, size, cs);
1805
1806         case FUSE_NOTIFY_INVAL_ENTRY:
1807                 return fuse_notify_inval_entry(fc, size, cs);
1808
1809         case FUSE_NOTIFY_STORE:
1810                 return fuse_notify_store(fc, size, cs);
1811
1812         case FUSE_NOTIFY_RETRIEVE:
1813                 return fuse_notify_retrieve(fc, size, cs);
1814
1815         case FUSE_NOTIFY_DELETE:
1816                 return fuse_notify_delete(fc, size, cs);
1817
1818         default:
1819                 fuse_copy_finish(cs);
1820                 return -EINVAL;
1821         }
1822 }
1823
1824 /* Look up request on processing list by unique ID */
1825 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1826 {
1827         struct fuse_req *req;
1828
1829         list_for_each_entry(req, &fc->processing, list) {
1830                 if (req->in.h.unique == unique || req->intr_unique == unique)
1831                         return req;
1832         }
1833         return NULL;
1834 }
1835
1836 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1837                          unsigned nbytes)
1838 {
1839         unsigned reqsize = sizeof(struct fuse_out_header);
1840
1841         if (out->h.error)
1842                 return nbytes != reqsize ? -EINVAL : 0;
1843
1844         reqsize += len_args(out->numargs, out->args);
1845
1846         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1847                 return -EINVAL;
1848         else if (reqsize > nbytes) {
1849                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1850                 unsigned diffsize = reqsize - nbytes;
1851                 if (diffsize > lastarg->size)
1852                         return -EINVAL;
1853                 lastarg->size -= diffsize;
1854         }
1855         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1856                               out->page_zeroing);
1857 }
1858
1859 /*
1860  * Write a single reply to a request.  First the header is copied from
1861  * the write buffer.  The request is then searched on the processing
1862  * list by the unique ID found in the header.  If found, then remove
1863  * it from the list and copy the rest of the buffer to the request.
1864  * The request is finished by calling request_end()
1865  */
1866 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1867                                  struct fuse_copy_state *cs, size_t nbytes)
1868 {
1869         int err;
1870         struct fuse_req *req;
1871         struct fuse_out_header oh;
1872
1873         if (nbytes < sizeof(struct fuse_out_header))
1874                 return -EINVAL;
1875
1876         err = fuse_copy_one(cs, &oh, sizeof(oh));
1877         if (err)
1878                 goto err_finish;
1879
1880         err = -EINVAL;
1881         if (oh.len != nbytes)
1882                 goto err_finish;
1883
1884         /*
1885          * Zero oh.unique indicates unsolicited notification message
1886          * and error contains notification code.
1887          */
1888         if (!oh.unique) {
1889                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1890                 return err ? err : nbytes;
1891         }
1892
1893         err = -EINVAL;
1894         if (oh.error <= -1000 || oh.error > 0)
1895                 goto err_finish;
1896
1897         spin_lock(&fc->lock);
1898         err = -ENOENT;
1899         if (!fc->connected)
1900                 goto err_unlock;
1901
1902         req = request_find(fc, oh.unique);
1903         if (!req)
1904                 goto err_unlock;
1905
1906         if (req->aborted) {
1907                 spin_unlock(&fc->lock);
1908                 fuse_copy_finish(cs);
1909                 spin_lock(&fc->lock);
1910                 request_end(fc, req);
1911                 return -ENOENT;
1912         }
1913         /* Is it an interrupt reply? */
1914         if (req->intr_unique == oh.unique) {
1915                 err = -EINVAL;
1916                 if (nbytes != sizeof(struct fuse_out_header))
1917                         goto err_unlock;
1918
1919                 if (oh.error == -ENOSYS)
1920                         fc->no_interrupt = 1;
1921                 else if (oh.error == -EAGAIN)
1922                         queue_interrupt(fc, req);
1923
1924                 spin_unlock(&fc->lock);
1925                 fuse_copy_finish(cs);
1926                 return nbytes;
1927         }
1928
1929         req->state = FUSE_REQ_WRITING;
1930         list_move(&req->list, &fc->io);
1931         req->out.h = oh;
1932         req->locked = 1;
1933         cs->req = req;
1934         if (!req->out.page_replace)
1935                 cs->move_pages = 0;
1936         spin_unlock(&fc->lock);
1937
1938         err = copy_out_args(cs, &req->out, nbytes);
1939         fuse_copy_finish(cs);
1940
1941         spin_lock(&fc->lock);
1942         req->locked = 0;
1943         if (!err) {
1944                 if (req->aborted)
1945                         err = -ENOENT;
1946         } else if (!req->aborted)
1947                 req->out.h.error = -EIO;
1948         request_end(fc, req);
1949
1950         return err ? err : nbytes;
1951
1952  err_unlock:
1953         spin_unlock(&fc->lock);
1954  err_finish:
1955         fuse_copy_finish(cs);
1956         return err;
1957 }
1958
1959 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
1960                               unsigned long nr_segs, loff_t pos)
1961 {
1962         struct fuse_copy_state cs;
1963         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1964         if (!fc)
1965                 return -EPERM;
1966
1967         fuse_copy_init(&cs, fc, 0, iov, nr_segs);
1968
1969         return fuse_dev_do_write(fc, &cs, iov_length(iov, nr_segs));
1970 }
1971
1972 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1973                                      struct file *out, loff_t *ppos,
1974                                      size_t len, unsigned int flags)
1975 {
1976         unsigned nbuf;
1977         unsigned idx;
1978         struct pipe_buffer *bufs;
1979         struct fuse_copy_state cs;
1980         struct fuse_conn *fc;
1981         size_t rem;
1982         ssize_t ret;
1983
1984         fc = fuse_get_conn(out);
1985         if (!fc)
1986                 return -EPERM;
1987
1988         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1989         if (!bufs)
1990                 return -ENOMEM;
1991
1992         pipe_lock(pipe);
1993         nbuf = 0;
1994         rem = 0;
1995         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1996                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1997
1998         ret = -EINVAL;
1999         if (rem < len) {
2000                 pipe_unlock(pipe);
2001                 goto out;
2002         }
2003
2004         rem = len;
2005         while (rem) {
2006                 struct pipe_buffer *ibuf;
2007                 struct pipe_buffer *obuf;
2008
2009                 BUG_ON(nbuf >= pipe->buffers);
2010                 BUG_ON(!pipe->nrbufs);
2011                 ibuf = &pipe->bufs[pipe->curbuf];
2012                 obuf = &bufs[nbuf];
2013
2014                 if (rem >= ibuf->len) {
2015                         *obuf = *ibuf;
2016                         ibuf->ops = NULL;
2017                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2018                         pipe->nrbufs--;
2019                 } else {
2020                         ibuf->ops->get(pipe, ibuf);
2021                         *obuf = *ibuf;
2022                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2023                         obuf->len = rem;
2024                         ibuf->offset += obuf->len;
2025                         ibuf->len -= obuf->len;
2026                 }
2027                 nbuf++;
2028                 rem -= obuf->len;
2029         }
2030         pipe_unlock(pipe);
2031
2032         fuse_copy_init(&cs, fc, 0, NULL, nbuf);
2033         cs.pipebufs = bufs;
2034         cs.pipe = pipe;
2035
2036         if (flags & SPLICE_F_MOVE)
2037                 cs.move_pages = 1;
2038
2039         ret = fuse_dev_do_write(fc, &cs, len);
2040
2041         for (idx = 0; idx < nbuf; idx++) {
2042                 struct pipe_buffer *buf = &bufs[idx];
2043                 buf->ops->release(pipe, buf);
2044         }
2045 out:
2046         kfree(bufs);
2047         return ret;
2048 }
2049
2050 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2051 {
2052         unsigned mask = POLLOUT | POLLWRNORM;
2053         struct fuse_conn *fc = fuse_get_conn(file);
2054         if (!fc)
2055                 return POLLERR;
2056
2057         poll_wait(file, &fc->waitq, wait);
2058
2059         spin_lock(&fc->lock);
2060         if (!fc->connected)
2061                 mask = POLLERR;
2062         else if (request_pending(fc))
2063                 mask |= POLLIN | POLLRDNORM;
2064         spin_unlock(&fc->lock);
2065
2066         return mask;
2067 }
2068
2069 /*
2070  * Abort all requests on the given list (pending or processing)
2071  *
2072  * This function releases and reacquires fc->lock
2073  */
2074 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2075 __releases(fc->lock)
2076 __acquires(fc->lock)
2077 {
2078         while (!list_empty(head)) {
2079                 struct fuse_req *req;
2080                 req = list_entry(head->next, struct fuse_req, list);
2081                 req->out.h.error = -ECONNABORTED;
2082                 request_end(fc, req);
2083                 spin_lock(&fc->lock);
2084         }
2085 }
2086
2087 /*
2088  * Abort requests under I/O
2089  *
2090  * The requests are set to aborted and finished, and the request
2091  * waiter is woken up.  This will make request_wait_answer() wait
2092  * until the request is unlocked and then return.
2093  *
2094  * If the request is asynchronous, then the end function needs to be
2095  * called after waiting for the request to be unlocked (if it was
2096  * locked).
2097  */
2098 static void end_io_requests(struct fuse_conn *fc)
2099 __releases(fc->lock)
2100 __acquires(fc->lock)
2101 {
2102         while (!list_empty(&fc->io)) {
2103                 struct fuse_req *req =
2104                         list_entry(fc->io.next, struct fuse_req, list);
2105                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
2106
2107                 req->aborted = 1;
2108                 req->out.h.error = -ECONNABORTED;
2109                 req->state = FUSE_REQ_FINISHED;
2110                 list_del_init(&req->list);
2111                 wake_up(&req->waitq);
2112                 if (end) {
2113                         req->end = NULL;
2114                         __fuse_get_request(req);
2115                         spin_unlock(&fc->lock);
2116                         wait_event(req->waitq, !req->locked);
2117                         end(fc, req);
2118                         fuse_put_request(fc, req);
2119                         spin_lock(&fc->lock);
2120                 }
2121         }
2122 }
2123
2124 static void end_queued_requests(struct fuse_conn *fc)
2125 __releases(fc->lock)
2126 __acquires(fc->lock)
2127 {
2128         fc->max_background = UINT_MAX;
2129         flush_bg_queue(fc);
2130         end_requests(fc, &fc->pending);
2131         end_requests(fc, &fc->processing);
2132         while (forget_pending(fc))
2133                 kfree(dequeue_forget(fc, 1, NULL));
2134 }
2135
2136 static void end_polls(struct fuse_conn *fc)
2137 {
2138         struct rb_node *p;
2139
2140         p = rb_first(&fc->polled_files);
2141
2142         while (p) {
2143                 struct fuse_file *ff;
2144                 ff = rb_entry(p, struct fuse_file, polled_node);
2145                 wake_up_interruptible_all(&ff->poll_wait);
2146
2147                 p = rb_next(p);
2148         }
2149 }
2150
2151 /*
2152  * Abort all requests.
2153  *
2154  * Emergency exit in case of a malicious or accidental deadlock, or
2155  * just a hung filesystem.
2156  *
2157  * The same effect is usually achievable through killing the
2158  * filesystem daemon and all users of the filesystem.  The exception
2159  * is the combination of an asynchronous request and the tricky
2160  * deadlock (see Documentation/filesystems/fuse.txt).
2161  *
2162  * During the aborting, progression of requests from the pending and
2163  * processing lists onto the io list, and progression of new requests
2164  * onto the pending list is prevented by req->connected being false.
2165  *
2166  * Progression of requests under I/O to the processing list is
2167  * prevented by the req->aborted flag being true for these requests.
2168  * For this reason requests on the io list must be aborted first.
2169  */
2170 void fuse_abort_conn(struct fuse_conn *fc)
2171 {
2172         spin_lock(&fc->lock);
2173         if (fc->connected) {
2174                 fc->connected = 0;
2175                 fc->blocked = 0;
2176                 fuse_set_initialized(fc);
2177                 end_io_requests(fc);
2178                 end_queued_requests(fc);
2179                 end_polls(fc);
2180                 wake_up_all(&fc->waitq);
2181                 wake_up_all(&fc->blocked_waitq);
2182                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
2183         }
2184         spin_unlock(&fc->lock);
2185 }
2186 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2187
2188 int fuse_dev_release(struct inode *inode, struct file *file)
2189 {
2190         struct fuse_conn *fc = fuse_get_conn(file);
2191         if (fc) {
2192                 spin_lock(&fc->lock);
2193                 fc->connected = 0;
2194                 fc->blocked = 0;
2195                 fuse_set_initialized(fc);
2196                 end_queued_requests(fc);
2197                 end_polls(fc);
2198                 wake_up_all(&fc->blocked_waitq);
2199                 spin_unlock(&fc->lock);
2200                 fuse_conn_put(fc);
2201         }
2202
2203         return 0;
2204 }
2205 EXPORT_SYMBOL_GPL(fuse_dev_release);
2206
2207 static int fuse_dev_fasync(int fd, struct file *file, int on)
2208 {
2209         struct fuse_conn *fc = fuse_get_conn(file);
2210         if (!fc)
2211                 return -EPERM;
2212
2213         /* No locking - fasync_helper does its own locking */
2214         return fasync_helper(fd, file, on, &fc->fasync);
2215 }
2216
2217 const struct file_operations fuse_dev_operations = {
2218         .owner          = THIS_MODULE,
2219         .llseek         = no_llseek,
2220         .read           = do_sync_read,
2221         .aio_read       = fuse_dev_read,
2222         .splice_read    = fuse_dev_splice_read,
2223         .write          = do_sync_write,
2224         .aio_write      = fuse_dev_write,
2225         .splice_write   = fuse_dev_splice_write,
2226         .poll           = fuse_dev_poll,
2227         .release        = fuse_dev_release,
2228         .fasync         = fuse_dev_fasync,
2229 };
2230 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2231
2232 static struct miscdevice fuse_miscdevice = {
2233         .minor = FUSE_MINOR,
2234         .name  = "fuse",
2235         .fops = &fuse_dev_operations,
2236 };
2237
2238 int __init fuse_dev_init(void)
2239 {
2240         int err = -ENOMEM;
2241         fuse_req_cachep = kmem_cache_create("fuse_request",
2242                                             sizeof(struct fuse_req),
2243                                             0, 0, NULL);
2244         if (!fuse_req_cachep)
2245                 goto out;
2246
2247         err = misc_register(&fuse_miscdevice);
2248         if (err)
2249                 goto out_cache_clean;
2250
2251         return 0;
2252
2253  out_cache_clean:
2254         kmem_cache_destroy(fuse_req_cachep);
2255  out:
2256         return err;
2257 }
2258
2259 void fuse_dev_cleanup(void)
2260 {
2261         misc_deregister(&fuse_miscdevice);
2262         kmem_cache_destroy(fuse_req_cachep);
2263 }