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         struct iov_iter *iter;
 714         struct pipe_buffer *pipebufs;
 715         struct pipe_buffer *currbuf;
 716         struct pipe_inode_info *pipe;
 717         unsigned long nr_segs;
 718         struct page *pg;
 719         unsigned len;
 720         unsigned offset;
 721         unsigned move_pages:1;
 722 };
 723
 724 static void fuse_copy_init(struct fuse_copy_state *cs,
 725                            struct fuse_conn *fc,
 726                            int write,
 727                            struct iov_iter *iter)
 728 {
 729         memset(cs, 0, sizeof(*cs));
 730         cs->fc = fc;
 731         cs->write = write;
 732         cs->iter = iter;
 733 }
 734
 735 /* Unmap and put previous page of userspace buffer */
 736 static void fuse_copy_finish(struct fuse_copy_state *cs)
 737 {
 738         if (cs->currbuf) {
 739                 struct pipe_buffer *buf = cs->currbuf;
 740
 741                 if (cs->write)
 742                         buf->len = PAGE_SIZE - cs->len;
 743                 cs->currbuf = NULL;
 744         } else if (cs->pg) {
 745                 if (cs->write) {
 746                         flush_dcache_page(cs->pg);
 747                         set_page_dirty_lock(cs->pg);
 748                 }
 749                 put_page(cs->pg);
 750         }
 751         cs->pg = NULL;
 752 }
 753
 754 /*
 755  * Get another pagefull of userspace buffer, and map it to kernel
 756  * address space, and lock request
 757  */
 758 static int fuse_copy_fill(struct fuse_copy_state *cs)
 759 {
 760         struct page *page;
 761         int err;
 762
 763         unlock_request(cs->fc, cs->req);
 764         fuse_copy_finish(cs);
 765         if (cs->pipebufs) {
 766                 struct pipe_buffer *buf = cs->pipebufs;
 767
 768                 if (!cs->write) {
 769                         err = buf->ops->confirm(cs->pipe, buf);
 770                         if (err)
 771                                 return err;
 772
 773                         BUG_ON(!cs->nr_segs);
 774                         cs->currbuf = buf;
 775                         cs->pg = buf->page;
 776                         cs->offset = buf->offset;
 777                         cs->len = buf->len;
 778                         cs->pipebufs++;
 779                         cs->nr_segs--;
 780                 } else {
 781                         if (cs->nr_segs == cs->pipe->buffers)
 782                                 return -EIO;
 783
 784                         page = alloc_page(GFP_HIGHUSER);
 785                         if (!page)
 786                                 return -ENOMEM;
 787
 788                         buf->page = page;
 789                         buf->offset = 0;
 790                         buf->len = 0;
 791
 792                         cs->currbuf = buf;
 793                         cs->pg = page;
 794                         cs->offset = 0;
 795                         cs->len = PAGE_SIZE;
 796                         cs->pipebufs++;
 797                         cs->nr_segs++;
 798                 }
 799         } else {
 800                 size_t off;
 801                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 802                 if (err < 0)
 803                         return err;
 804                 BUG_ON(!err);
 805                 cs->len = err;
 806                 cs->offset = off;
 807                 cs->pg = page;
 808                 cs->offset = off;
 809                 iov_iter_advance(cs->iter, err);
 810         }
 811
 812         return lock_request(cs->fc, cs->req);
 813 }
 814
 815 /* Do as much copy to/from userspace buffer as we can */
 816 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 817 {
 818         unsigned ncpy = min(*size, cs->len);
 819         if (val) {
 820                 void *pgaddr = kmap_atomic(cs->pg);
 821                 void *buf = pgaddr + cs->offset;
 822
 823                 if (cs->write)
 824                         memcpy(buf, *val, ncpy);
 825                 else
 826                         memcpy(*val, buf, ncpy);
 827
 828                 kunmap_atomic(pgaddr);
 829                 *val += ncpy;
 830         }
 831         *size -= ncpy;
 832         cs->len -= ncpy;
 833         cs->offset += ncpy;
 834         return ncpy;
 835 }
 836
 837 static int fuse_check_page(struct page *page)
 838 {
 839         if (page_mapcount(page) ||
 840             page->mapping != NULL ||
 841             page_count(page) != 1 ||
 842             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 843              ~(1 << PG_locked |
 844                1 << PG_referenced |
 845                1 << PG_uptodate |
 846                1 << PG_lru |
 847                1 << PG_active |
 848                1 << PG_reclaim))) {
 849                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 850                 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);
 851                 return 1;
 852         }
 853         return 0;
 854 }
 855
 856 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 857 {
 858         int err;
 859         struct page *oldpage = *pagep;
 860         struct page *newpage;
 861         struct pipe_buffer *buf = cs->pipebufs;
 862
 863         unlock_request(cs->fc, cs->req);
 864         fuse_copy_finish(cs);
 865
 866         err = buf->ops->confirm(cs->pipe, buf);
 867         if (err)
 868                 return err;
 869
 870         BUG_ON(!cs->nr_segs);
 871         cs->currbuf = buf;
 872         cs->len = buf->len;
 873         cs->pipebufs++;
 874         cs->nr_segs--;
 875
 876         if (cs->len != PAGE_SIZE)
 877                 goto out_fallback;
 878
 879         if (buf->ops->steal(cs->pipe, buf) != 0)
 880                 goto out_fallback;
 881
 882         newpage = buf->page;
 883
 884         if (!PageUptodate(newpage))
 885                 SetPageUptodate(newpage);
 886
 887         ClearPageMappedToDisk(newpage);
 888
 889         if (fuse_check_page(newpage) != 0)
 890                 goto out_fallback_unlock;
 891
 892         /*
 893          * This is a new and locked page, it shouldn't be mapped or
 894          * have any special flags on it
 895          */
 896         if (WARN_ON(page_mapped(oldpage)))
 897                 goto out_fallback_unlock;
 898         if (WARN_ON(page_has_private(oldpage)))
 899                 goto out_fallback_unlock;
 900         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 901                 goto out_fallback_unlock;
 902         if (WARN_ON(PageMlocked(oldpage)))
 903                 goto out_fallback_unlock;
 904
 905         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 906         if (err) {
 907                 unlock_page(newpage);
 908                 return err;
 909         }
 910
 911         page_cache_get(newpage);
 912
 913         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 914                 lru_cache_add_file(newpage);
 915
 916         err = 0;
 917         spin_lock(&cs->fc->lock);
 918         if (cs->req->aborted)
 919                 err = -ENOENT;
 920         else
 921                 *pagep = newpage;
 922         spin_unlock(&cs->fc->lock);
 923
 924         if (err) {
 925                 unlock_page(newpage);
 926                 page_cache_release(newpage);
 927                 return err;
 928         }
 929
 930         unlock_page(oldpage);
 931         page_cache_release(oldpage);
 932         cs->len = 0;
 933
 934         return 0;
 935
 936 out_fallback_unlock:
 937         unlock_page(newpage);
 938 out_fallback:
 939         cs->pg = buf->page;
 940         cs->offset = buf->offset;
 941
 942         err = lock_request(cs->fc, cs->req);
 943         if (err)
 944                 return err;
 945
 946         return 1;
 947 }
 948
 949 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 950                          unsigned offset, unsigned count)
 951 {
 952         struct pipe_buffer *buf;
 953
 954         if (cs->nr_segs == cs->pipe->buffers)
 955                 return -EIO;
 956
 957         unlock_request(cs->fc, cs->req);
 958         fuse_copy_finish(cs);
 959
 960         buf = cs->pipebufs;
 961         page_cache_get(page);
 962         buf->page = page;
 963         buf->offset = offset;
 964         buf->len = count;
 965
 966         cs->pipebufs++;
 967         cs->nr_segs++;
 968         cs->len = 0;
 969
 970         return 0;
 971 }
 972
 973 /*
 974  * Copy a page in the request to/from the userspace buffer.  Must be
 975  * done atomically
 976  */
 977 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 978                           unsigned offset, unsigned count, int zeroing)
 979 {
 980         int err;
 981         struct page *page = *pagep;
 982
 983         if (page && zeroing && count < PAGE_SIZE)
 984                 clear_highpage(page);
 985
 986         while (count) {
 987                 if (cs->write && cs->pipebufs && page) {
 988                         return fuse_ref_page(cs, page, offset, count);
 989                 } else if (!cs->len) {
 990                         if (cs->move_pages && page &&
 991                             offset == 0 && count == PAGE_SIZE) {
 992                                 err = fuse_try_move_page(cs, pagep);
 993                                 if (err <= 0)
 994                                         return err;
 995                         } else {
 996                                 err = fuse_copy_fill(cs);
 997                                 if (err)
 998                                         return err;
 999                         }
1000                 }
1001                 if (page) {
1002                         void *mapaddr = kmap_atomic(page);
1003                         void *buf = mapaddr + offset;
1004                         offset += fuse_copy_do(cs, &buf, &count);
1005                         kunmap_atomic(mapaddr);
1006                 } else
1007                         offset += fuse_copy_do(cs, NULL, &count);
1008         }
1009         if (page && !cs->write)
1010                 flush_dcache_page(page);
1011         return 0;
1012 }
1013
1014 /* Copy pages in the request to/from userspace buffer */
1015 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1016                            int zeroing)
1017 {
1018         unsigned i;
1019         struct fuse_req *req = cs->req;
1020
1021         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1022                 int err;
1023                 unsigned offset = req->page_descs[i].offset;
1024                 unsigned count = min(nbytes, req->page_descs[i].length);
1025
1026                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1027                                      zeroing);
1028                 if (err)
1029                         return err;
1030
1031                 nbytes -= count;
1032         }
1033         return 0;
1034 }
1035
1036 /* Copy a single argument in the request to/from userspace buffer */
1037 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1038 {
1039         while (size) {
1040                 if (!cs->len) {
1041                         int err = fuse_copy_fill(cs);
1042                         if (err)
1043                                 return err;
1044                 }
1045                 fuse_copy_do(cs, &val, &size);
1046         }
1047         return 0;
1048 }
1049
1050 /* Copy request arguments to/from userspace buffer */
1051 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1052                           unsigned argpages, struct fuse_arg *args,
1053                           int zeroing)
1054 {
1055         int err = 0;
1056         unsigned i;
1057
1058         for (i = 0; !err && i < numargs; i++)  {
1059                 struct fuse_arg *arg = &args[i];
1060                 if (i == numargs - 1 && argpages)
1061                         err = fuse_copy_pages(cs, arg->size, zeroing);
1062                 else
1063                         err = fuse_copy_one(cs, arg->value, arg->size);
1064         }
1065         return err;
1066 }
1067
1068 static int forget_pending(struct fuse_conn *fc)
1069 {
1070         return fc->forget_list_head.next != NULL;
1071 }
1072
1073 static int request_pending(struct fuse_conn *fc)
1074 {
1075         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts) ||
1076                 forget_pending(fc);
1077 }
1078
1079 /* Wait until a request is available on the pending list */
1080 static void request_wait(struct fuse_conn *fc)
1081 __releases(fc->lock)
1082 __acquires(fc->lock)
1083 {
1084         DECLARE_WAITQUEUE(wait, current);
1085
1086         add_wait_queue_exclusive(&fc->waitq, &wait);
1087         while (fc->connected && !request_pending(fc)) {
1088                 set_current_state(TASK_INTERRUPTIBLE);
1089                 if (signal_pending(current))
1090                         break;
1091
1092                 spin_unlock(&fc->lock);
1093                 schedule();
1094                 spin_lock(&fc->lock);
1095         }
1096         set_current_state(TASK_RUNNING);
1097         remove_wait_queue(&fc->waitq, &wait);
1098 }
1099
1100 /*
1101  * Transfer an interrupt request to userspace
1102  *
1103  * Unlike other requests this is assembled on demand, without a need
1104  * to allocate a separate fuse_req structure.
1105  *
1106  * Called with fc->lock held, releases it
1107  */
1108 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
1109                                size_t nbytes, struct fuse_req *req)
1110 __releases(fc->lock)
1111 {
1112         struct fuse_in_header ih;
1113         struct fuse_interrupt_in arg;
1114         unsigned reqsize = sizeof(ih) + sizeof(arg);
1115         int err;
1116
1117         list_del_init(&req->intr_entry);
1118         req->intr_unique = fuse_get_unique(fc);
1119         memset(&ih, 0, sizeof(ih));
1120         memset(&arg, 0, sizeof(arg));
1121         ih.len = reqsize;
1122         ih.opcode = FUSE_INTERRUPT;
1123         ih.unique = req->intr_unique;
1124         arg.unique = req->in.h.unique;
1125
1126         spin_unlock(&fc->lock);
1127         if (nbytes < reqsize)
1128                 return -EINVAL;
1129
1130         err = fuse_copy_one(cs, &ih, sizeof(ih));
1131         if (!err)
1132                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1133         fuse_copy_finish(cs);
1134
1135         return err ? err : reqsize;
1136 }
1137
1138 static struct fuse_forget_link *dequeue_forget(struct fuse_conn *fc,
1139                                                unsigned max,
1140                                                unsigned *countp)
1141 {
1142         struct fuse_forget_link *head = fc->forget_list_head.next;
1143         struct fuse_forget_link **newhead = &head;
1144         unsigned count;
1145
1146         for (count = 0; *newhead != NULL && count < max; count++)
1147                 newhead = &(*newhead)->next;
1148
1149         fc->forget_list_head.next = *newhead;
1150         *newhead = NULL;
1151         if (fc->forget_list_head.next == NULL)
1152                 fc->forget_list_tail = &fc->forget_list_head;
1153
1154         if (countp != NULL)
1155                 *countp = count;
1156
1157         return head;
1158 }
1159
1160 static int fuse_read_single_forget(struct fuse_conn *fc,
1161                                    struct fuse_copy_state *cs,
1162                                    size_t nbytes)
1163 __releases(fc->lock)
1164 {
1165         int err;
1166         struct fuse_forget_link *forget = dequeue_forget(fc, 1, NULL);
1167         struct fuse_forget_in arg = {
1168                 .nlookup = forget->forget_one.nlookup,
1169         };
1170         struct fuse_in_header ih = {
1171                 .opcode = FUSE_FORGET,
1172                 .nodeid = forget->forget_one.nodeid,
1173                 .unique = fuse_get_unique(fc),
1174                 .len = sizeof(ih) + sizeof(arg),
1175         };
1176
1177         spin_unlock(&fc->lock);
1178         kfree(forget);
1179         if (nbytes < ih.len)
1180                 return -EINVAL;
1181
1182         err = fuse_copy_one(cs, &ih, sizeof(ih));
1183         if (!err)
1184                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1185         fuse_copy_finish(cs);
1186
1187         if (err)
1188                 return err;
1189
1190         return ih.len;
1191 }
1192
1193 static int fuse_read_batch_forget(struct fuse_conn *fc,
1194                                    struct fuse_copy_state *cs, size_t nbytes)
1195 __releases(fc->lock)
1196 {
1197         int err;
1198         unsigned max_forgets;
1199         unsigned count;
1200         struct fuse_forget_link *head;
1201         struct fuse_batch_forget_in arg = { .count = 0 };
1202         struct fuse_in_header ih = {
1203                 .opcode = FUSE_BATCH_FORGET,
1204                 .unique = fuse_get_unique(fc),
1205                 .len = sizeof(ih) + sizeof(arg),
1206         };
1207
1208         if (nbytes < ih.len) {
1209                 spin_unlock(&fc->lock);
1210                 return -EINVAL;
1211         }
1212
1213         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1214         head = dequeue_forget(fc, max_forgets, &count);
1215         spin_unlock(&fc->lock);
1216
1217         arg.count = count;
1218         ih.len += count * sizeof(struct fuse_forget_one);
1219         err = fuse_copy_one(cs, &ih, sizeof(ih));
1220         if (!err)
1221                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1222
1223         while (head) {
1224                 struct fuse_forget_link *forget = head;
1225
1226                 if (!err) {
1227                         err = fuse_copy_one(cs, &forget->forget_one,
1228                                             sizeof(forget->forget_one));
1229                 }
1230                 head = forget->next;
1231                 kfree(forget);
1232         }
1233
1234         fuse_copy_finish(cs);
1235
1236         if (err)
1237                 return err;
1238
1239         return ih.len;
1240 }
1241
1242 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_copy_state *cs,
1243                             size_t nbytes)
1244 __releases(fc->lock)
1245 {
1246         if (fc->minor < 16 || fc->forget_list_head.next->next == NULL)
1247                 return fuse_read_single_forget(fc, cs, nbytes);
1248         else
1249                 return fuse_read_batch_forget(fc, cs, nbytes);
1250 }
1251
1252 /*
1253  * Read a single request into the userspace filesystem's buffer.  This
1254  * function waits until a request is available, then removes it from
1255  * the pending list and copies request data to userspace buffer.  If
1256  * no reply is needed (FORGET) or request has been aborted or there
1257  * was an error during the copying then it's finished by calling
1258  * request_end().  Otherwise add it to the processing list, and set
1259  * the 'sent' flag.
1260  */
1261 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1262                                 struct fuse_copy_state *cs, size_t nbytes)
1263 {
1264         int err;
1265         struct fuse_req *req;
1266         struct fuse_in *in;
1267         unsigned reqsize;
1268
1269  restart:
1270         spin_lock(&fc->lock);
1271         err = -EAGAIN;
1272         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
1273             !request_pending(fc))
1274                 goto err_unlock;
1275
1276         request_wait(fc);
1277         err = -ENODEV;
1278         if (!fc->connected)
1279                 goto err_unlock;
1280         err = -ERESTARTSYS;
1281         if (!request_pending(fc))
1282                 goto err_unlock;
1283
1284         if (!list_empty(&fc->interrupts)) {
1285                 req = list_entry(fc->interrupts.next, struct fuse_req,
1286                                  intr_entry);
1287                 return fuse_read_interrupt(fc, cs, nbytes, req);
1288         }
1289
1290         if (forget_pending(fc)) {
1291                 if (list_empty(&fc->pending) || fc->forget_batch-- > 0)
1292                         return fuse_read_forget(fc, cs, nbytes);
1293
1294                 if (fc->forget_batch <= -8)
1295                         fc->forget_batch = 16;
1296         }
1297
1298         req = list_entry(fc->pending.next, struct fuse_req, list);
1299         req->state = FUSE_REQ_READING;
1300         list_move(&req->list, &fc->io);
1301
1302         in = &req->in;
1303         reqsize = in->h.len;
1304         /* If request is too large, reply with an error and restart the read */
1305         if (nbytes < reqsize) {
1306                 req->out.h.error = -EIO;
1307                 /* SETXATTR is special, since it may contain too large data */
1308                 if (in->h.opcode == FUSE_SETXATTR)
1309                         req->out.h.error = -E2BIG;
1310                 request_end(fc, req);
1311                 goto restart;
1312         }
1313         spin_unlock(&fc->lock);
1314         cs->req = req;
1315         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1316         if (!err)
1317                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1318                                      (struct fuse_arg *) in->args, 0);
1319         fuse_copy_finish(cs);
1320         spin_lock(&fc->lock);
1321         req->locked = 0;
1322         if (req->aborted) {
1323                 request_end(fc, req);
1324                 return -ENODEV;
1325         }
1326         if (err) {
1327                 req->out.h.error = -EIO;
1328                 request_end(fc, req);
1329                 return err;
1330         }
1331         if (!req->isreply)
1332                 request_end(fc, req);
1333         else {
1334                 req->state = FUSE_REQ_SENT;
1335                 list_move_tail(&req->list, &fc->processing);
1336                 if (req->interrupted)
1337                         queue_interrupt(fc, req);
1338                 spin_unlock(&fc->lock);
1339         }
1340         return reqsize;
1341
1342  err_unlock:
1343         spin_unlock(&fc->lock);
1344         return err;
1345 }
1346
1347 static int fuse_dev_open(struct inode *inode, struct file *file)
1348 {
1349         /*
1350          * The fuse device's file's private_data is used to hold
1351          * the fuse_conn(ection) when it is mounted, and is used to
1352          * keep track of whether the file has been mounted already.
1353          */
1354         file->private_data = NULL;
1355         return 0;
1356 }
1357
1358 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1359 {
1360         struct fuse_copy_state cs;
1361         struct file *file = iocb->ki_filp;
1362         struct fuse_conn *fc = fuse_get_conn(file);
1363         if (!fc)
1364                 return -EPERM;
1365
1366         if (!iter_is_iovec(to))
1367                 return -EINVAL;
1368
1369         fuse_copy_init(&cs, fc, 1, to);
1370
1371         return fuse_dev_do_read(fc, file, &cs, iov_iter_count(to));
1372 }
1373
1374 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1375                                     struct pipe_inode_info *pipe,
1376                                     size_t len, unsigned int flags)
1377 {
1378         int ret;
1379         int page_nr = 0;
1380         int do_wakeup = 0;
1381         struct pipe_buffer *bufs;
1382         struct fuse_copy_state cs;
1383         struct fuse_conn *fc = fuse_get_conn(in);
1384         if (!fc)
1385                 return -EPERM;
1386
1387         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1388         if (!bufs)
1389                 return -ENOMEM;
1390
1391         fuse_copy_init(&cs, fc, 1, NULL);
1392         cs.pipebufs = bufs;
1393         cs.pipe = pipe;
1394         ret = fuse_dev_do_read(fc, in, &cs, len);
1395         if (ret < 0)
1396                 goto out;
1397
1398         ret = 0;
1399         pipe_lock(pipe);
1400
1401         if (!pipe->readers) {
1402                 send_sig(SIGPIPE, current, 0);
1403                 if (!ret)
1404                         ret = -EPIPE;
1405                 goto out_unlock;
1406         }
1407
1408         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1409                 ret = -EIO;
1410                 goto out_unlock;
1411         }
1412
1413         while (page_nr < cs.nr_segs) {
1414                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1415                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1416
1417                 buf->page = bufs[page_nr].page;
1418                 buf->offset = bufs[page_nr].offset;
1419                 buf->len = bufs[page_nr].len;
1420                 /*
1421                  * Need to be careful about this.  Having buf->ops in module
1422                  * code can Oops if the buffer persists after module unload.
1423                  */
1424                 buf->ops = &nosteal_pipe_buf_ops;
1425
1426                 pipe->nrbufs++;
1427                 page_nr++;
1428                 ret += buf->len;
1429
1430                 if (pipe->files)
1431                         do_wakeup = 1;
1432         }
1433
1434 out_unlock:
1435         pipe_unlock(pipe);
1436
1437         if (do_wakeup) {
1438                 smp_mb();
1439                 if (waitqueue_active(&pipe->wait))
1440                         wake_up_interruptible(&pipe->wait);
1441                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1442         }
1443
1444 out:
1445         for (; page_nr < cs.nr_segs; page_nr++)
1446                 page_cache_release(bufs[page_nr].page);
1447
1448         kfree(bufs);
1449         return ret;
1450 }
1451
1452 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1453                             struct fuse_copy_state *cs)
1454 {
1455         struct fuse_notify_poll_wakeup_out outarg;
1456         int err = -EINVAL;
1457
1458         if (size != sizeof(outarg))
1459                 goto err;
1460
1461         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1462         if (err)
1463                 goto err;
1464
1465         fuse_copy_finish(cs);
1466         return fuse_notify_poll_wakeup(fc, &outarg);
1467
1468 err:
1469         fuse_copy_finish(cs);
1470         return err;
1471 }
1472
1473 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1474                                    struct fuse_copy_state *cs)
1475 {
1476         struct fuse_notify_inval_inode_out outarg;
1477         int err = -EINVAL;
1478
1479         if (size != sizeof(outarg))
1480                 goto err;
1481
1482         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1483         if (err)
1484                 goto err;
1485         fuse_copy_finish(cs);
1486
1487         down_read(&fc->killsb);
1488         err = -ENOENT;
1489         if (fc->sb) {
1490                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1491                                                outarg.off, outarg.len);
1492         }
1493         up_read(&fc->killsb);
1494         return err;
1495
1496 err:
1497         fuse_copy_finish(cs);
1498         return err;
1499 }
1500
1501 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1502                                    struct fuse_copy_state *cs)
1503 {
1504         struct fuse_notify_inval_entry_out outarg;
1505         int err = -ENOMEM;
1506         char *buf;
1507         struct qstr name;
1508
1509         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1510         if (!buf)
1511                 goto err;
1512
1513         err = -EINVAL;
1514         if (size < sizeof(outarg))
1515                 goto err;
1516
1517         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1518         if (err)
1519                 goto err;
1520
1521         err = -ENAMETOOLONG;
1522         if (outarg.namelen > FUSE_NAME_MAX)
1523                 goto err;
1524
1525         err = -EINVAL;
1526         if (size != sizeof(outarg) + outarg.namelen + 1)
1527                 goto err;
1528
1529         name.name = buf;
1530         name.len = outarg.namelen;
1531         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1532         if (err)
1533                 goto err;
1534         fuse_copy_finish(cs);
1535         buf[outarg.namelen] = 0;
1536         name.hash = full_name_hash(name.name, name.len);
1537
1538         down_read(&fc->killsb);
1539         err = -ENOENT;
1540         if (fc->sb)
1541                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1542         up_read(&fc->killsb);
1543         kfree(buf);
1544         return err;
1545
1546 err:
1547         kfree(buf);
1548         fuse_copy_finish(cs);
1549         return err;
1550 }
1551
1552 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1553                               struct fuse_copy_state *cs)
1554 {
1555         struct fuse_notify_delete_out outarg;
1556         int err = -ENOMEM;
1557         char *buf;
1558         struct qstr name;
1559
1560         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1561         if (!buf)
1562                 goto err;
1563
1564         err = -EINVAL;
1565         if (size < sizeof(outarg))
1566                 goto err;
1567
1568         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1569         if (err)
1570                 goto err;
1571
1572         err = -ENAMETOOLONG;
1573         if (outarg.namelen > FUSE_NAME_MAX)
1574                 goto err;
1575
1576         err = -EINVAL;
1577         if (size != sizeof(outarg) + outarg.namelen + 1)
1578                 goto err;
1579
1580         name.name = buf;
1581         name.len = outarg.namelen;
1582         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1583         if (err)
1584                 goto err;
1585         fuse_copy_finish(cs);
1586         buf[outarg.namelen] = 0;
1587         name.hash = full_name_hash(name.name, name.len);
1588
1589         down_read(&fc->killsb);
1590         err = -ENOENT;
1591         if (fc->sb)
1592                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1593                                                outarg.child, &name);
1594         up_read(&fc->killsb);
1595         kfree(buf);
1596         return err;
1597
1598 err:
1599         kfree(buf);
1600         fuse_copy_finish(cs);
1601         return err;
1602 }
1603
1604 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1605                              struct fuse_copy_state *cs)
1606 {
1607         struct fuse_notify_store_out outarg;
1608         struct inode *inode;
1609         struct address_space *mapping;
1610         u64 nodeid;
1611         int err;
1612         pgoff_t index;
1613         unsigned int offset;
1614         unsigned int num;
1615         loff_t file_size;
1616         loff_t end;
1617
1618         err = -EINVAL;
1619         if (size < sizeof(outarg))
1620                 goto out_finish;
1621
1622         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1623         if (err)
1624                 goto out_finish;
1625
1626         err = -EINVAL;
1627         if (size - sizeof(outarg) != outarg.size)
1628                 goto out_finish;
1629
1630         nodeid = outarg.nodeid;
1631
1632         down_read(&fc->killsb);
1633
1634         err = -ENOENT;
1635         if (!fc->sb)
1636                 goto out_up_killsb;
1637
1638         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1639         if (!inode)
1640                 goto out_up_killsb;
1641
1642         mapping = inode->i_mapping;
1643         index = outarg.offset >> PAGE_CACHE_SHIFT;
1644         offset = outarg.offset & ~PAGE_CACHE_MASK;
1645         file_size = i_size_read(inode);
1646         end = outarg.offset + outarg.size;
1647         if (end > file_size) {
1648                 file_size = end;
1649                 fuse_write_update_size(inode, file_size);
1650         }
1651
1652         num = outarg.size;
1653         while (num) {
1654                 struct page *page;
1655                 unsigned int this_num;
1656
1657                 err = -ENOMEM;
1658                 page = find_or_create_page(mapping, index,
1659                                            mapping_gfp_mask(mapping));
1660                 if (!page)
1661                         goto out_iput;
1662
1663                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1664                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1665                 if (!err && offset == 0 &&
1666                     (this_num == PAGE_CACHE_SIZE || file_size == end))
1667                         SetPageUptodate(page);
1668                 unlock_page(page);
1669                 page_cache_release(page);
1670
1671                 if (err)
1672                         goto out_iput;
1673
1674                 num -= this_num;
1675                 offset = 0;
1676                 index++;
1677         }
1678
1679         err = 0;
1680
1681 out_iput:
1682         iput(inode);
1683 out_up_killsb:
1684         up_read(&fc->killsb);
1685 out_finish:
1686         fuse_copy_finish(cs);
1687         return err;
1688 }
1689
1690 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1691 {
1692         release_pages(req->pages, req->num_pages, false);
1693 }
1694
1695 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1696                          struct fuse_notify_retrieve_out *outarg)
1697 {
1698         int err;
1699         struct address_space *mapping = inode->i_mapping;
1700         struct fuse_req *req;
1701         pgoff_t index;
1702         loff_t file_size;
1703         unsigned int num;
1704         unsigned int offset;
1705         size_t total_len = 0;
1706         int num_pages;
1707
1708         offset = outarg->offset & ~PAGE_CACHE_MASK;
1709         file_size = i_size_read(inode);
1710
1711         num = outarg->size;
1712         if (outarg->offset > file_size)
1713                 num = 0;
1714         else if (outarg->offset + num > file_size)
1715                 num = file_size - outarg->offset;
1716
1717         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1718         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1719
1720         req = fuse_get_req(fc, num_pages);
1721         if (IS_ERR(req))
1722                 return PTR_ERR(req);
1723
1724         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1725         req->in.h.nodeid = outarg->nodeid;
1726         req->in.numargs = 2;
1727         req->in.argpages = 1;
1728         req->page_descs[0].offset = offset;
1729         req->end = fuse_retrieve_end;
1730
1731         index = outarg->offset >> PAGE_CACHE_SHIFT;
1732
1733         while (num && req->num_pages < num_pages) {
1734                 struct page *page;
1735                 unsigned int this_num;
1736
1737                 page = find_get_page(mapping, index);
1738                 if (!page)
1739                         break;
1740
1741                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1742                 req->pages[req->num_pages] = page;
1743                 req->page_descs[req->num_pages].length = this_num;
1744                 req->num_pages++;
1745
1746                 offset = 0;
1747                 num -= this_num;
1748                 total_len += this_num;
1749                 index++;
1750         }
1751         req->misc.retrieve_in.offset = outarg->offset;
1752         req->misc.retrieve_in.size = total_len;
1753         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1754         req->in.args[0].value = &req->misc.retrieve_in;
1755         req->in.args[1].size = total_len;
1756
1757         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1758         if (err)
1759                 fuse_retrieve_end(fc, req);
1760
1761         return err;
1762 }
1763
1764 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1765                                 struct fuse_copy_state *cs)
1766 {
1767         struct fuse_notify_retrieve_out outarg;
1768         struct inode *inode;
1769         int err;
1770
1771         err = -EINVAL;
1772         if (size != sizeof(outarg))
1773                 goto copy_finish;
1774
1775         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1776         if (err)
1777                 goto copy_finish;
1778
1779         fuse_copy_finish(cs);
1780
1781         down_read(&fc->killsb);
1782         err = -ENOENT;
1783         if (fc->sb) {
1784                 u64 nodeid = outarg.nodeid;
1785
1786                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1787                 if (inode) {
1788                         err = fuse_retrieve(fc, inode, &outarg);
1789                         iput(inode);
1790                 }
1791         }
1792         up_read(&fc->killsb);
1793
1794         return err;
1795
1796 copy_finish:
1797         fuse_copy_finish(cs);
1798         return err;
1799 }
1800
1801 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1802                        unsigned int size, struct fuse_copy_state *cs)
1803 {
1804         /* Don't try to move pages (yet) */
1805         cs->move_pages = 0;
1806
1807         switch (code) {
1808         case FUSE_NOTIFY_POLL:
1809                 return fuse_notify_poll(fc, size, cs);
1810
1811         case FUSE_NOTIFY_INVAL_INODE:
1812                 return fuse_notify_inval_inode(fc, size, cs);
1813
1814         case FUSE_NOTIFY_INVAL_ENTRY:
1815                 return fuse_notify_inval_entry(fc, size, cs);
1816
1817         case FUSE_NOTIFY_STORE:
1818                 return fuse_notify_store(fc, size, cs);
1819
1820         case FUSE_NOTIFY_RETRIEVE:
1821                 return fuse_notify_retrieve(fc, size, cs);
1822
1823         case FUSE_NOTIFY_DELETE:
1824                 return fuse_notify_delete(fc, size, cs);
1825
1826         default:
1827                 fuse_copy_finish(cs);
1828                 return -EINVAL;
1829         }
1830 }
1831
1832 /* Look up request on processing list by unique ID */
1833 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1834 {
1835         struct fuse_req *req;
1836
1837         list_for_each_entry(req, &fc->processing, list) {
1838                 if (req->in.h.unique == unique || req->intr_unique == unique)
1839                         return req;
1840         }
1841         return NULL;
1842 }
1843
1844 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1845                          unsigned nbytes)
1846 {
1847         unsigned reqsize = sizeof(struct fuse_out_header);
1848
1849         if (out->h.error)
1850                 return nbytes != reqsize ? -EINVAL : 0;
1851
1852         reqsize += len_args(out->numargs, out->args);
1853
1854         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1855                 return -EINVAL;
1856         else if (reqsize > nbytes) {
1857                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1858                 unsigned diffsize = reqsize - nbytes;
1859                 if (diffsize > lastarg->size)
1860                         return -EINVAL;
1861                 lastarg->size -= diffsize;
1862         }
1863         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1864                               out->page_zeroing);
1865 }
1866
1867 /*
1868  * Write a single reply to a request.  First the header is copied from
1869  * the write buffer.  The request is then searched on the processing
1870  * list by the unique ID found in the header.  If found, then remove
1871  * it from the list and copy the rest of the buffer to the request.
1872  * The request is finished by calling request_end()
1873  */
1874 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1875                                  struct fuse_copy_state *cs, size_t nbytes)
1876 {
1877         int err;
1878         struct fuse_req *req;
1879         struct fuse_out_header oh;
1880
1881         if (nbytes < sizeof(struct fuse_out_header))
1882                 return -EINVAL;
1883
1884         err = fuse_copy_one(cs, &oh, sizeof(oh));
1885         if (err)
1886                 goto err_finish;
1887
1888         err = -EINVAL;
1889         if (oh.len != nbytes)
1890                 goto err_finish;
1891
1892         /*
1893          * Zero oh.unique indicates unsolicited notification message
1894          * and error contains notification code.
1895          */
1896         if (!oh.unique) {
1897                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1898                 return err ? err : nbytes;
1899         }
1900
1901         err = -EINVAL;
1902         if (oh.error <= -1000 || oh.error > 0)
1903                 goto err_finish;
1904
1905         spin_lock(&fc->lock);
1906         err = -ENOENT;
1907         if (!fc->connected)
1908                 goto err_unlock;
1909
1910         req = request_find(fc, oh.unique);
1911         if (!req)
1912                 goto err_unlock;
1913
1914         if (req->aborted) {
1915                 spin_unlock(&fc->lock);
1916                 fuse_copy_finish(cs);
1917                 spin_lock(&fc->lock);
1918                 request_end(fc, req);
1919                 return -ENOENT;
1920         }
1921         /* Is it an interrupt reply? */
1922         if (req->intr_unique == oh.unique) {
1923                 err = -EINVAL;
1924                 if (nbytes != sizeof(struct fuse_out_header))
1925                         goto err_unlock;
1926
1927                 if (oh.error == -ENOSYS)
1928                         fc->no_interrupt = 1;
1929                 else if (oh.error == -EAGAIN)
1930                         queue_interrupt(fc, req);
1931
1932                 spin_unlock(&fc->lock);
1933                 fuse_copy_finish(cs);
1934                 return nbytes;
1935         }
1936
1937         req->state = FUSE_REQ_WRITING;
1938         list_move(&req->list, &fc->io);
1939         req->out.h = oh;
1940         req->locked = 1;
1941         cs->req = req;
1942         if (!req->out.page_replace)
1943                 cs->move_pages = 0;
1944         spin_unlock(&fc->lock);
1945
1946         err = copy_out_args(cs, &req->out, nbytes);
1947         fuse_copy_finish(cs);
1948
1949         spin_lock(&fc->lock);
1950         req->locked = 0;
1951         if (!err) {
1952                 if (req->aborted)
1953                         err = -ENOENT;
1954         } else if (!req->aborted)
1955                 req->out.h.error = -EIO;
1956         request_end(fc, req);
1957
1958         return err ? err : nbytes;
1959
1960  err_unlock:
1961         spin_unlock(&fc->lock);
1962  err_finish:
1963         fuse_copy_finish(cs);
1964         return err;
1965 }
1966
1967 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1968 {
1969         struct fuse_copy_state cs;
1970         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1971         if (!fc)
1972                 return -EPERM;
1973
1974         if (!iter_is_iovec(from))
1975                 return -EINVAL;
1976
1977         fuse_copy_init(&cs, fc, 0, from);
1978
1979         return fuse_dev_do_write(fc, &cs, iov_iter_count(from));
1980 }
1981
1982 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1983                                      struct file *out, loff_t *ppos,
1984                                      size_t len, unsigned int flags)
1985 {
1986         unsigned nbuf;
1987         unsigned idx;
1988         struct pipe_buffer *bufs;
1989         struct fuse_copy_state cs;
1990         struct fuse_conn *fc;
1991         size_t rem;
1992         ssize_t ret;
1993
1994         fc = fuse_get_conn(out);
1995         if (!fc)
1996                 return -EPERM;
1997
1998         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1999         if (!bufs)
2000                 return -ENOMEM;
2001
2002         pipe_lock(pipe);
2003         nbuf = 0;
2004         rem = 0;
2005         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2006                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2007
2008         ret = -EINVAL;
2009         if (rem < len) {
2010                 pipe_unlock(pipe);
2011                 goto out;
2012         }
2013
2014         rem = len;
2015         while (rem) {
2016                 struct pipe_buffer *ibuf;
2017                 struct pipe_buffer *obuf;
2018
2019                 BUG_ON(nbuf >= pipe->buffers);
2020                 BUG_ON(!pipe->nrbufs);
2021                 ibuf = &pipe->bufs[pipe->curbuf];
2022                 obuf = &bufs[nbuf];
2023
2024                 if (rem >= ibuf->len) {
2025                         *obuf = *ibuf;
2026                         ibuf->ops = NULL;
2027                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2028                         pipe->nrbufs--;
2029                 } else {
2030                         ibuf->ops->get(pipe, ibuf);
2031                         *obuf = *ibuf;
2032                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2033                         obuf->len = rem;
2034                         ibuf->offset += obuf->len;
2035                         ibuf->len -= obuf->len;
2036                 }
2037                 nbuf++;
2038                 rem -= obuf->len;
2039         }
2040         pipe_unlock(pipe);
2041
2042         fuse_copy_init(&cs, fc, 0, NULL);
2043         cs.pipebufs = bufs;
2044         cs.nr_segs = nbuf;
2045         cs.pipe = pipe;
2046
2047         if (flags & SPLICE_F_MOVE)
2048                 cs.move_pages = 1;
2049
2050         ret = fuse_dev_do_write(fc, &cs, len);
2051
2052         for (idx = 0; idx < nbuf; idx++) {
2053                 struct pipe_buffer *buf = &bufs[idx];
2054                 buf->ops->release(pipe, buf);
2055         }
2056 out:
2057         kfree(bufs);
2058         return ret;
2059 }
2060
2061 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2062 {
2063         unsigned mask = POLLOUT | POLLWRNORM;
2064         struct fuse_conn *fc = fuse_get_conn(file);
2065         if (!fc)
2066                 return POLLERR;
2067
2068         poll_wait(file, &fc->waitq, wait);
2069
2070         spin_lock(&fc->lock);
2071         if (!fc->connected)
2072                 mask = POLLERR;
2073         else if (request_pending(fc))
2074                 mask |= POLLIN | POLLRDNORM;
2075         spin_unlock(&fc->lock);
2076
2077         return mask;
2078 }
2079
2080 /*
2081  * Abort all requests on the given list (pending or processing)
2082  *
2083  * This function releases and reacquires fc->lock
2084  */
2085 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2086 __releases(fc->lock)
2087 __acquires(fc->lock)
2088 {
2089         while (!list_empty(head)) {
2090                 struct fuse_req *req;
2091                 req = list_entry(head->next, struct fuse_req, list);
2092                 req->out.h.error = -ECONNABORTED;
2093                 request_end(fc, req);
2094                 spin_lock(&fc->lock);
2095         }
2096 }
2097
2098 /*
2099  * Abort requests under I/O
2100  *
2101  * The requests are set to aborted and finished, and the request
2102  * waiter is woken up.  This will make request_wait_answer() wait
2103  * until the request is unlocked and then return.
2104  *
2105  * If the request is asynchronous, then the end function needs to be
2106  * called after waiting for the request to be unlocked (if it was
2107  * locked).
2108  */
2109 static void end_io_requests(struct fuse_conn *fc)
2110 __releases(fc->lock)
2111 __acquires(fc->lock)
2112 {
2113         while (!list_empty(&fc->io)) {
2114                 struct fuse_req *req =
2115                         list_entry(fc->io.next, struct fuse_req, list);
2116                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
2117
2118                 req->aborted = 1;
2119                 req->out.h.error = -ECONNABORTED;
2120                 req->state = FUSE_REQ_FINISHED;
2121                 list_del_init(&req->list);
2122                 wake_up(&req->waitq);
2123                 if (end) {
2124                         req->end = NULL;
2125                         __fuse_get_request(req);
2126                         spin_unlock(&fc->lock);
2127                         wait_event(req->waitq, !req->locked);
2128                         end(fc, req);
2129                         fuse_put_request(fc, req);
2130                         spin_lock(&fc->lock);
2131                 }
2132         }
2133 }
2134
2135 static void end_queued_requests(struct fuse_conn *fc)
2136 __releases(fc->lock)
2137 __acquires(fc->lock)
2138 {
2139         fc->max_background = UINT_MAX;
2140         flush_bg_queue(fc);
2141         end_requests(fc, &fc->pending);
2142         end_requests(fc, &fc->processing);
2143         while (forget_pending(fc))
2144                 kfree(dequeue_forget(fc, 1, NULL));
2145 }
2146
2147 static void end_polls(struct fuse_conn *fc)
2148 {
2149         struct rb_node *p;
2150
2151         p = rb_first(&fc->polled_files);
2152
2153         while (p) {
2154                 struct fuse_file *ff;
2155                 ff = rb_entry(p, struct fuse_file, polled_node);
2156                 wake_up_interruptible_all(&ff->poll_wait);
2157
2158                 p = rb_next(p);
2159         }
2160 }
2161
2162 /*
2163  * Abort all requests.
2164  *
2165  * Emergency exit in case of a malicious or accidental deadlock, or
2166  * just a hung filesystem.
2167  *
2168  * The same effect is usually achievable through killing the
2169  * filesystem daemon and all users of the filesystem.  The exception
2170  * is the combination of an asynchronous request and the tricky
2171  * deadlock (see Documentation/filesystems/fuse.txt).
2172  *
2173  * During the aborting, progression of requests from the pending and
2174  * processing lists onto the io list, and progression of new requests
2175  * onto the pending list is prevented by req->connected being false.
2176  *
2177  * Progression of requests under I/O to the processing list is
2178  * prevented by the req->aborted flag being true for these requests.
2179  * For this reason requests on the io list must be aborted first.
2180  */
2181 void fuse_abort_conn(struct fuse_conn *fc)
2182 {
2183         spin_lock(&fc->lock);
2184         if (fc->connected) {
2185                 fc->connected = 0;
2186                 fc->blocked = 0;
2187                 fuse_set_initialized(fc);
2188                 end_io_requests(fc);
2189                 end_queued_requests(fc);
2190                 end_polls(fc);
2191                 wake_up_all(&fc->waitq);
2192                 wake_up_all(&fc->blocked_waitq);
2193                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
2194         }
2195         spin_unlock(&fc->lock);
2196 }
2197 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2198
2199 int fuse_dev_release(struct inode *inode, struct file *file)
2200 {
2201         struct fuse_conn *fc = fuse_get_conn(file);
2202         if (fc) {
2203                 spin_lock(&fc->lock);
2204                 fc->connected = 0;
2205                 fc->blocked = 0;
2206                 fuse_set_initialized(fc);
2207                 end_queued_requests(fc);
2208                 end_polls(fc);
2209                 wake_up_all(&fc->blocked_waitq);
2210                 spin_unlock(&fc->lock);
2211                 fuse_conn_put(fc);
2212         }
2213
2214         return 0;
2215 }
2216 EXPORT_SYMBOL_GPL(fuse_dev_release);
2217
2218 static int fuse_dev_fasync(int fd, struct file *file, int on)
2219 {
2220         struct fuse_conn *fc = fuse_get_conn(file);
2221         if (!fc)
2222                 return -EPERM;
2223
2224         /* No locking - fasync_helper does its own locking */
2225         return fasync_helper(fd, file, on, &fc->fasync);
2226 }
2227
2228 const struct file_operations fuse_dev_operations = {
2229         .owner          = THIS_MODULE,
2230         .open           = fuse_dev_open,
2231         .llseek         = no_llseek,
2232         .read_iter      = fuse_dev_read,
2233         .splice_read    = fuse_dev_splice_read,
2234         .write_iter     = fuse_dev_write,
2235         .splice_write   = fuse_dev_splice_write,
2236         .poll           = fuse_dev_poll,
2237         .release        = fuse_dev_release,
2238         .fasync         = fuse_dev_fasync,
2239 };
2240 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2241
2242 static struct miscdevice fuse_miscdevice = {
2243         .minor = FUSE_MINOR,
2244         .name  = "fuse",
2245         .fops = &fuse_dev_operations,
2246 };
2247
2248 int __init fuse_dev_init(void)
2249 {
2250         int err = -ENOMEM;
2251         fuse_req_cachep = kmem_cache_create("fuse_request",
2252                                             sizeof(struct fuse_req),
2253                                             0, 0, NULL);
2254         if (!fuse_req_cachep)
2255                 goto out;
2256
2257         err = misc_register(&fuse_miscdevice);
2258         if (err)
2259                 goto out_cache_clean;
2260
2261         return 0;
2262
2263  out_cache_clean:
2264         kmem_cache_destroy(fuse_req_cachep);
2265  out:
2266         return err;
2267 }
2268
2269 void fuse_dev_cleanup(void)
2270 {
2271         misc_deregister(&fuse_miscdevice);
2272         kmem_cache_destroy(fuse_req_cachep);
2273 }