fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/pipe_fs_i.h>
  20 #include <linux/swap.h>
  21 #include <linux/splice.h>
  22 #include <linux/freezer.h>
  23
  24 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  25 MODULE_ALIAS("devname:fuse");
  26
  27 static struct kmem_cache *fuse_req_cachep;
  28
  29 static struct fuse_dev *fuse_get_dev(struct file *file)
  30 {
  31         /*
  32          * Lockless access is OK, because file->private data is set
  33          * once during mount and is valid until the file is released.
  34          */
  35         return ACCESS_ONCE(file->private_data);
  36 }
  37
  38 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  39                               struct fuse_page_desc *page_descs,
  40                               unsigned npages)
  41 {
  42         memset(req, 0, sizeof(*req));
  43         memset(pages, 0, sizeof(*pages) * npages);
  44         memset(page_descs, 0, sizeof(*page_descs) * npages);
  45         INIT_LIST_HEAD(&req->list);
  46         INIT_LIST_HEAD(&req->intr_entry);
  47         init_waitqueue_head(&req->waitq);
  48         atomic_set(&req->count, 1);
  49         req->pages = pages;
  50         req->page_descs = page_descs;
  51         req->max_pages = npages;
  52         __set_bit(FR_PENDING, &req->flags);
  53 }
  54
  55 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  56 {
  57         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, flags);
  58         if (req) {
  59                 struct page **pages;
  60                 struct fuse_page_desc *page_descs;
  61
  62                 if (npages <= FUSE_REQ_INLINE_PAGES) {
  63                         pages = req->inline_pages;
  64                         page_descs = req->inline_page_descs;
  65                 } else {
  66                         pages = kmalloc(sizeof(struct page *) * npages, flags);
  67                         page_descs = kmalloc(sizeof(struct fuse_page_desc) *
  68                                              npages, flags);
  69                 }
  70
  71                 if (!pages || !page_descs) {
  72                         kfree(pages);
  73                         kfree(page_descs);
  74                         kmem_cache_free(fuse_req_cachep, req);
  75                         return NULL;
  76                 }
  77
  78                 fuse_request_init(req, pages, page_descs, npages);
  79         }
  80         return req;
  81 }
  82
  83 struct fuse_req *fuse_request_alloc(unsigned npages)
  84 {
  85         return __fuse_request_alloc(npages, GFP_KERNEL);
  86 }
  87 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  88
  89 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  90 {
  91         return __fuse_request_alloc(npages, GFP_NOFS);
  92 }
  93
  94 void fuse_request_free(struct fuse_req *req)
  95 {
  96         if (req->pages != req->inline_pages) {
  97                 kfree(req->pages);
  98                 kfree(req->page_descs);
  99         }
 100         kmem_cache_free(fuse_req_cachep, req);
 101 }
 102
 103 static void block_sigs(sigset_t *oldset)
 104 {
 105         sigset_t mask;
 106
 107         siginitsetinv(&mask, sigmask(SIGKILL));
 108         sigprocmask(SIG_BLOCK, &mask, oldset);
 109 }
 110
 111 static void restore_sigs(sigset_t *oldset)
 112 {
 113         sigprocmask(SIG_SETMASK, oldset, NULL);
 114 }
 115
 116 void __fuse_get_request(struct fuse_req *req)
 117 {
 118         atomic_inc(&req->count);
 119 }
 120
 121 /* Must be called with > 1 refcount */
 122 static void __fuse_put_request(struct fuse_req *req)
 123 {
 124         BUG_ON(atomic_read(&req->count) < 2);
 125         atomic_dec(&req->count);
 126 }
 127
 128 static void fuse_req_init_context(struct fuse_req *req)
 129 {
 130         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 131         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 132         req->in.h.pid = current->pid;
 133 }
 134
 135 void fuse_set_initialized(struct fuse_conn *fc)
 136 {
 137         /* Make sure stores before this are seen on another CPU */
 138         smp_wmb();
 139         fc->initialized = 1;
 140 }
 141
 142 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 143 {
 144         return !fc->initialized || (for_background && fc->blocked);
 145 }
 146
 147 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 148                                        bool for_background)
 149 {
 150         struct fuse_req *req;
 151         int err;
 152         atomic_inc(&fc->num_waiting);
 153
 154         if (fuse_block_alloc(fc, for_background)) {
 155                 sigset_t oldset;
 156                 int intr;
 157
 158                 block_sigs(&oldset);
 159                 intr = wait_event_interruptible_exclusive(fc->blocked_waitq,
 160                                 !fuse_block_alloc(fc, for_background));
 161                 restore_sigs(&oldset);
 162                 err = -EINTR;
 163                 if (intr)
 164                         goto out;
 165         }
 166         /* Matches smp_wmb() in fuse_set_initialized() */
 167         smp_rmb();
 168
 169         err = -ENOTCONN;
 170         if (!fc->connected)
 171                 goto out;
 172
 173         err = -ECONNREFUSED;
 174         if (fc->conn_error)
 175                 goto out;
 176
 177         req = fuse_request_alloc(npages);
 178         err = -ENOMEM;
 179         if (!req) {
 180                 if (for_background)
 181                         wake_up(&fc->blocked_waitq);
 182                 goto out;
 183         }
 184
 185         fuse_req_init_context(req);
 186         __set_bit(FR_WAITING, &req->flags);
 187         if (for_background)
 188                 __set_bit(FR_BACKGROUND, &req->flags);
 189
 190         return req;
 191
 192  out:
 193         atomic_dec(&fc->num_waiting);
 194         return ERR_PTR(err);
 195 }
 196
 197 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 198 {
 199         return __fuse_get_req(fc, npages, false);
 200 }
 201 EXPORT_SYMBOL_GPL(fuse_get_req);
 202
 203 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 204                                              unsigned npages)
 205 {
 206         return __fuse_get_req(fc, npages, true);
 207 }
 208 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 209
 210 /*
 211  * Return request in fuse_file->reserved_req.  However that may
 212  * currently be in use.  If that is the case, wait for it to become
 213  * available.
 214  */
 215 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 216                                          struct file *file)
 217 {
 218         struct fuse_req *req = NULL;
 219         struct fuse_file *ff = file->private_data;
 220
 221         do {
 222                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 223                 spin_lock(&fc->lock);
 224                 if (ff->reserved_req) {
 225                         req = ff->reserved_req;
 226                         ff->reserved_req = NULL;
 227                         req->stolen_file = get_file(file);
 228                 }
 229                 spin_unlock(&fc->lock);
 230         } while (!req);
 231
 232         return req;
 233 }
 234
 235 /*
 236  * Put stolen request back into fuse_file->reserved_req
 237  */
 238 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 239 {
 240         struct file *file = req->stolen_file;
 241         struct fuse_file *ff = file->private_data;
 242
 243         spin_lock(&fc->lock);
 244         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 245         BUG_ON(ff->reserved_req);
 246         ff->reserved_req = req;
 247         wake_up_all(&fc->reserved_req_waitq);
 248         spin_unlock(&fc->lock);
 249         fput(file);
 250 }
 251
 252 /*
 253  * Gets a requests for a file operation, always succeeds
 254  *
 255  * This is used for sending the FLUSH request, which must get to
 256  * userspace, due to POSIX locks which may need to be unlocked.
 257  *
 258  * If allocation fails due to OOM, use the reserved request in
 259  * fuse_file.
 260  *
 261  * This is very unlikely to deadlock accidentally, since the
 262  * filesystem should not have it's own file open.  If deadlock is
 263  * intentional, it can still be broken by "aborting" the filesystem.
 264  */
 265 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 266                                              struct file *file)
 267 {
 268         struct fuse_req *req;
 269
 270         atomic_inc(&fc->num_waiting);
 271         wait_event(fc->blocked_waitq, fc->initialized);
 272         /* Matches smp_wmb() in fuse_set_initialized() */
 273         smp_rmb();
 274         req = fuse_request_alloc(0);
 275         if (!req)
 276                 req = get_reserved_req(fc, file);
 277
 278         fuse_req_init_context(req);
 279         __set_bit(FR_WAITING, &req->flags);
 280         __clear_bit(FR_BACKGROUND, &req->flags);
 281         return req;
 282 }
 283
 284 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 285 {
 286         if (atomic_dec_and_test(&req->count)) {
 287                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 288                         /*
 289                          * We get here in the unlikely case that a background
 290                          * request was allocated but not sent
 291                          */
 292                         spin_lock(&fc->lock);
 293                         if (!fc->blocked)
 294                                 wake_up(&fc->blocked_waitq);
 295                         spin_unlock(&fc->lock);
 296                 }
 297
 298                 if (test_bit(FR_WAITING, &req->flags)) {
 299                         __clear_bit(FR_WAITING, &req->flags);
 300                         atomic_dec(&fc->num_waiting);
 301                 }
 302
 303                 if (req->stolen_file)
 304                         put_reserved_req(fc, req);
 305                 else
 306                         fuse_request_free(req);
 307         }
 308 }
 309 EXPORT_SYMBOL_GPL(fuse_put_request);
 310
 311 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 312 {
 313         unsigned nbytes = 0;
 314         unsigned i;
 315
 316         for (i = 0; i < numargs; i++)
 317                 nbytes += args[i].size;
 318
 319         return nbytes;
 320 }
 321
 322 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 323 {
 324         return ++fiq->reqctr;
 325 }
 326
 327 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 328 {
 329         req->in.h.len = sizeof(struct fuse_in_header) +
 330                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 331         list_add_tail(&req->list, &fiq->pending);
 332         wake_up_locked(&fiq->waitq);
 333         kill_fasync(&fiq->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         struct fuse_iqueue *fiq = &fc->iq;
 340
 341         forget->forget_one.nodeid = nodeid;
 342         forget->forget_one.nlookup = nlookup;
 343
 344         spin_lock(&fiq->waitq.lock);
 345         if (fiq->connected) {
 346                 fiq->forget_list_tail->next = forget;
 347                 fiq->forget_list_tail = forget;
 348                 wake_up_locked(&fiq->waitq);
 349                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 350         } else {
 351                 kfree(forget);
 352         }
 353         spin_unlock(&fiq->waitq.lock);
 354 }
 355
 356 static void flush_bg_queue(struct fuse_conn *fc)
 357 {
 358         while (fc->active_background < fc->max_background &&
 359                !list_empty(&fc->bg_queue)) {
 360                 struct fuse_req *req;
 361                 struct fuse_iqueue *fiq = &fc->iq;
 362
 363                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 364                 list_del(&req->list);
 365                 fc->active_background++;
 366                 spin_lock(&fiq->waitq.lock);
 367                 req->in.h.unique = fuse_get_unique(fiq);
 368                 queue_request(fiq, req);
 369                 spin_unlock(&fiq->waitq.lock);
 370         }
 371 }
 372
 373 /*
 374  * This function is called when a request is finished.  Either a reply
 375  * has arrived or it was aborted (and not yet sent) or some error
 376  * occurred during communication with userspace, or the device file
 377  * was closed.  The requester thread is woken up (if still waiting),
 378  * the 'end' callback is called if given, else the reference to the
 379  * request is released
 380  */
 381 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 382 {
 383         struct fuse_iqueue *fiq = &fc->iq;
 384
 385         if (test_and_set_bit(FR_FINISHED, &req->flags))
 386                 return;
 387
 388         spin_lock(&fiq->waitq.lock);
 389         list_del_init(&req->intr_entry);
 390         spin_unlock(&fiq->waitq.lock);
 391         WARN_ON(test_bit(FR_PENDING, &req->flags));
 392         WARN_ON(test_bit(FR_SENT, &req->flags));
 393         if (test_bit(FR_BACKGROUND, &req->flags)) {
 394                 spin_lock(&fc->lock);
 395                 clear_bit(FR_BACKGROUND, &req->flags);
 396                 if (fc->num_background == fc->max_background)
 397                         fc->blocked = 0;
 398
 399                 /* Wake up next waiter, if any */
 400                 if (!fc->blocked && waitqueue_active(&fc->blocked_waitq))
 401                         wake_up(&fc->blocked_waitq);
 402
 403                 if (fc->num_background == fc->congestion_threshold &&
 404                     fc->connected && fc->bdi_initialized) {
 405                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 406                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 407                 }
 408                 fc->num_background--;
 409                 fc->active_background--;
 410                 flush_bg_queue(fc);
 411                 spin_unlock(&fc->lock);
 412         }
 413         wake_up(&req->waitq);
 414         if (req->end)
 415                 req->end(fc, req);
 416         fuse_put_request(fc, req);
 417 }
 418
 419 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 420 {
 421         spin_lock(&fiq->waitq.lock);
 422         if (list_empty(&req->intr_entry)) {
 423                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 424                 wake_up_locked(&fiq->waitq);
 425         }
 426         spin_unlock(&fiq->waitq.lock);
 427         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 428 }
 429
 430 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 431 {
 432         struct fuse_iqueue *fiq = &fc->iq;
 433         int err;
 434
 435         if (!fc->no_interrupt) {
 436                 /* Any signal may interrupt this */
 437                 err = wait_event_interruptible(req->waitq,
 438                                         test_bit(FR_FINISHED, &req->flags));
 439                 if (!err)
 440                         return;
 441
 442                 set_bit(FR_INTERRUPTED, &req->flags);
 443                 /* matches barrier in fuse_dev_do_read() */
 444                 smp_mb__after_atomic();
 445                 if (test_bit(FR_SENT, &req->flags))
 446                         queue_interrupt(fiq, req);
 447         }
 448
 449         if (!test_bit(FR_FORCE, &req->flags)) {
 450                 sigset_t oldset;
 451
 452                 /* Only fatal signals may interrupt this */
 453                 block_sigs(&oldset);
 454                 err = wait_event_interruptible(req->waitq,
 455                                         test_bit(FR_FINISHED, &req->flags));
 456                 restore_sigs(&oldset);
 457
 458                 if (!err)
 459                         return;
 460
 461                 spin_lock(&fiq->waitq.lock);
 462                 /* Request is not yet in userspace, bail out */
 463                 if (test_bit(FR_PENDING, &req->flags)) {
 464                         list_del(&req->list);
 465                         spin_unlock(&fiq->waitq.lock);
 466                         __fuse_put_request(req);
 467                         req->out.h.error = -EINTR;
 468                         return;
 469                 }
 470                 spin_unlock(&fiq->waitq.lock);
 471         }
 472
 473         /*
 474          * Either request is already in userspace, or it was forced.
 475          * Wait it out.
 476          */
 477         while (!test_bit(FR_FINISHED, &req->flags))
 478                 wait_event_freezable(req->waitq,
 479                                 test_bit(FR_FINISHED, &req->flags));
 480 }
 481
 482 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 483 {
 484         struct fuse_iqueue *fiq = &fc->iq;
 485
 486         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 487         spin_lock(&fiq->waitq.lock);
 488         if (!fiq->connected) {
 489                 spin_unlock(&fiq->waitq.lock);
 490                 req->out.h.error = -ENOTCONN;
 491         } else {
 492                 req->in.h.unique = fuse_get_unique(fiq);
 493                 queue_request(fiq, req);
 494                 /* acquire extra reference, since request is still needed
 495                    after request_end() */
 496                 __fuse_get_request(req);
 497                 spin_unlock(&fiq->waitq.lock);
 498
 499                 request_wait_answer(fc, req);
 500                 /* Pairs with smp_wmb() in request_end() */
 501                 smp_rmb();
 502         }
 503 }
 504
 505 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 506 {
 507         __set_bit(FR_ISREPLY, &req->flags);
 508         if (!test_bit(FR_WAITING, &req->flags)) {
 509                 __set_bit(FR_WAITING, &req->flags);
 510                 atomic_inc(&fc->num_waiting);
 511         }
 512         __fuse_request_send(fc, req);
 513 }
 514 EXPORT_SYMBOL_GPL(fuse_request_send);
 515
 516 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 517 {
 518         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 519                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 520
 521         if (fc->minor < 9) {
 522                 switch (args->in.h.opcode) {
 523                 case FUSE_LOOKUP:
 524                 case FUSE_CREATE:
 525                 case FUSE_MKNOD:
 526                 case FUSE_MKDIR:
 527                 case FUSE_SYMLINK:
 528                 case FUSE_LINK:
 529                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 530                         break;
 531                 case FUSE_GETATTR:
 532                 case FUSE_SETATTR:
 533                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 534                         break;
 535                 }
 536         }
 537         if (fc->minor < 12) {
 538                 switch (args->in.h.opcode) {
 539                 case FUSE_CREATE:
 540                         args->in.args[0].size = sizeof(struct fuse_open_in);
 541                         break;
 542                 case FUSE_MKNOD:
 543                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 544                         break;
 545                 }
 546         }
 547 }
 548
 549 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 550 {
 551         struct fuse_req *req;
 552         ssize_t ret;
 553
 554         req = fuse_get_req(fc, 0);
 555         if (IS_ERR(req))
 556                 return PTR_ERR(req);
 557
 558         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 559         fuse_adjust_compat(fc, args);
 560
 561         req->in.h.opcode = args->in.h.opcode;
 562         req->in.h.nodeid = args->in.h.nodeid;
 563         req->in.numargs = args->in.numargs;
 564         memcpy(req->in.args, args->in.args,
 565                args->in.numargs * sizeof(struct fuse_in_arg));
 566         req->out.argvar = args->out.argvar;
 567         req->out.numargs = args->out.numargs;
 568         memcpy(req->out.args, args->out.args,
 569                args->out.numargs * sizeof(struct fuse_arg));
 570         fuse_request_send(fc, req);
 571         ret = req->out.h.error;
 572         if (!ret && args->out.argvar) {
 573                 BUG_ON(args->out.numargs != 1);
 574                 ret = req->out.args[0].size;
 575         }
 576         fuse_put_request(fc, req);
 577
 578         return ret;
 579 }
 580
 581 /*
 582  * Called under fc->lock
 583  *
 584  * fc->connected must have been checked previously
 585  */
 586 void fuse_request_send_background_locked(struct fuse_conn *fc,
 587                                          struct fuse_req *req)
 588 {
 589         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 590         if (!test_bit(FR_WAITING, &req->flags)) {
 591                 __set_bit(FR_WAITING, &req->flags);
 592                 atomic_inc(&fc->num_waiting);
 593         }
 594         __set_bit(FR_ISREPLY, &req->flags);
 595         fc->num_background++;
 596         if (fc->num_background == fc->max_background)
 597                 fc->blocked = 1;
 598         if (fc->num_background == fc->congestion_threshold &&
 599             fc->bdi_initialized) {
 600                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 601                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 602         }
 603         list_add_tail(&req->list, &fc->bg_queue);
 604         flush_bg_queue(fc);
 605 }
 606
 607 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 608 {
 609         BUG_ON(!req->end);
 610         spin_lock(&fc->lock);
 611         if (fc->connected) {
 612                 fuse_request_send_background_locked(fc, req);
 613                 spin_unlock(&fc->lock);
 614         } else {
 615                 spin_unlock(&fc->lock);
 616                 req->out.h.error = -ENOTCONN;
 617                 req->end(fc, req);
 618                 fuse_put_request(fc, req);
 619         }
 620 }
 621 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 622
 623 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 624                                           struct fuse_req *req, u64 unique)
 625 {
 626         int err = -ENODEV;
 627         struct fuse_iqueue *fiq = &fc->iq;
 628
 629         __clear_bit(FR_ISREPLY, &req->flags);
 630         req->in.h.unique = unique;
 631         spin_lock(&fiq->waitq.lock);
 632         if (fiq->connected) {
 633                 queue_request(fiq, req);
 634                 err = 0;
 635         }
 636         spin_unlock(&fiq->waitq.lock);
 637
 638         return err;
 639 }
 640
 641 void fuse_force_forget(struct file *file, u64 nodeid)
 642 {
 643         struct inode *inode = file_inode(file);
 644         struct fuse_conn *fc = get_fuse_conn(inode);
 645         struct fuse_req *req;
 646         struct fuse_forget_in inarg;
 647
 648         memset(&inarg, 0, sizeof(inarg));
 649         inarg.nlookup = 1;
 650         req = fuse_get_req_nofail_nopages(fc, file);
 651         req->in.h.opcode = FUSE_FORGET;
 652         req->in.h.nodeid = nodeid;
 653         req->in.numargs = 1;
 654         req->in.args[0].size = sizeof(inarg);
 655         req->in.args[0].value = &inarg;
 656         __clear_bit(FR_ISREPLY, &req->flags);
 657         __fuse_request_send(fc, req);
 658         /* ignore errors */
 659         fuse_put_request(fc, req);
 660 }
 661
 662 /*
 663  * Lock the request.  Up to the next unlock_request() there mustn't be
 664  * anything that could cause a page-fault.  If the request was already
 665  * aborted bail out.
 666  */
 667 static int lock_request(struct fuse_req *req)
 668 {
 669         int err = 0;
 670         if (req) {
 671                 spin_lock(&req->waitq.lock);
 672                 if (test_bit(FR_ABORTED, &req->flags))
 673                         err = -ENOENT;
 674                 else
 675                         set_bit(FR_LOCKED, &req->flags);
 676                 spin_unlock(&req->waitq.lock);
 677         }
 678         return err;
 679 }
 680
 681 /*
 682  * Unlock request.  If it was aborted while locked, caller is responsible
 683  * for unlocking and ending the request.
 684  */
 685 static int unlock_request(struct fuse_req *req)
 686 {
 687         int err = 0;
 688         if (req) {
 689                 spin_lock(&req->waitq.lock);
 690                 if (test_bit(FR_ABORTED, &req->flags))
 691                         err = -ENOENT;
 692                 else
 693                         clear_bit(FR_LOCKED, &req->flags);
 694                 spin_unlock(&req->waitq.lock);
 695         }
 696         return err;
 697 }
 698
 699 struct fuse_copy_state {
 700         int write;
 701         struct fuse_req *req;
 702         struct iov_iter *iter;
 703         struct pipe_buffer *pipebufs;
 704         struct pipe_buffer *currbuf;
 705         struct pipe_inode_info *pipe;
 706         unsigned long nr_segs;
 707         struct page *pg;
 708         unsigned len;
 709         unsigned offset;
 710         unsigned move_pages:1;
 711 };
 712
 713 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 714                            struct iov_iter *iter)
 715 {
 716         memset(cs, 0, sizeof(*cs));
 717         cs->write = write;
 718         cs->iter = iter;
 719 }
 720
 721 /* Unmap and put previous page of userspace buffer */
 722 static void fuse_copy_finish(struct fuse_copy_state *cs)
 723 {
 724         if (cs->currbuf) {
 725                 struct pipe_buffer *buf = cs->currbuf;
 726
 727                 if (cs->write)
 728                         buf->len = PAGE_SIZE - cs->len;
 729                 cs->currbuf = NULL;
 730         } else if (cs->pg) {
 731                 if (cs->write) {
 732                         flush_dcache_page(cs->pg);
 733                         set_page_dirty_lock(cs->pg);
 734                 }
 735                 put_page(cs->pg);
 736         }
 737         cs->pg = NULL;
 738 }
 739
 740 /*
 741  * Get another pagefull of userspace buffer, and map it to kernel
 742  * address space, and lock request
 743  */
 744 static int fuse_copy_fill(struct fuse_copy_state *cs)
 745 {
 746         struct page *page;
 747         int err;
 748
 749         err = unlock_request(cs->req);
 750         if (err)
 751                 return err;
 752
 753         fuse_copy_finish(cs);
 754         if (cs->pipebufs) {
 755                 struct pipe_buffer *buf = cs->pipebufs;
 756
 757                 if (!cs->write) {
 758                         err = buf->ops->confirm(cs->pipe, buf);
 759                         if (err)
 760                                 return err;
 761
 762                         BUG_ON(!cs->nr_segs);
 763                         cs->currbuf = buf;
 764                         cs->pg = buf->page;
 765                         cs->offset = buf->offset;
 766                         cs->len = buf->len;
 767                         cs->pipebufs++;
 768                         cs->nr_segs--;
 769                 } else {
 770                         if (cs->nr_segs == cs->pipe->buffers)
 771                                 return -EIO;
 772
 773                         page = alloc_page(GFP_HIGHUSER);
 774                         if (!page)
 775                                 return -ENOMEM;
 776
 777                         buf->page = page;
 778                         buf->offset = 0;
 779                         buf->len = 0;
 780
 781                         cs->currbuf = buf;
 782                         cs->pg = page;
 783                         cs->offset = 0;
 784                         cs->len = PAGE_SIZE;
 785                         cs->pipebufs++;
 786                         cs->nr_segs++;
 787                 }
 788         } else {
 789                 size_t off;
 790                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 791                 if (err < 0)
 792                         return err;
 793                 BUG_ON(!err);
 794                 cs->len = err;
 795                 cs->offset = off;
 796                 cs->pg = page;
 797                 cs->offset = off;
 798                 iov_iter_advance(cs->iter, err);
 799         }
 800
 801         return lock_request(cs->req);
 802 }
 803
 804 /* Do as much copy to/from userspace buffer as we can */
 805 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 806 {
 807         unsigned ncpy = min(*size, cs->len);
 808         if (val) {
 809                 void *pgaddr = kmap_atomic(cs->pg);
 810                 void *buf = pgaddr + cs->offset;
 811
 812                 if (cs->write)
 813                         memcpy(buf, *val, ncpy);
 814                 else
 815                         memcpy(*val, buf, ncpy);
 816
 817                 kunmap_atomic(pgaddr);
 818                 *val += ncpy;
 819         }
 820         *size -= ncpy;
 821         cs->len -= ncpy;
 822         cs->offset += ncpy;
 823         return ncpy;
 824 }
 825
 826 static int fuse_check_page(struct page *page)
 827 {
 828         if (page_mapcount(page) ||
 829             page->mapping != NULL ||
 830             page_count(page) != 1 ||
 831             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 832              ~(1 << PG_locked |
 833                1 << PG_referenced |
 834                1 << PG_uptodate |
 835                1 << PG_lru |
 836                1 << PG_active |
 837                1 << PG_reclaim))) {
 838                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 839                 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);
 840                 return 1;
 841         }
 842         return 0;
 843 }
 844
 845 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 846 {
 847         int err;
 848         struct page *oldpage = *pagep;
 849         struct page *newpage;
 850         struct pipe_buffer *buf = cs->pipebufs;
 851
 852         err = unlock_request(cs->req);
 853         if (err)
 854                 return err;
 855
 856         fuse_copy_finish(cs);
 857
 858         err = buf->ops->confirm(cs->pipe, buf);
 859         if (err)
 860                 return err;
 861
 862         BUG_ON(!cs->nr_segs);
 863         cs->currbuf = buf;
 864         cs->len = buf->len;
 865         cs->pipebufs++;
 866         cs->nr_segs--;
 867
 868         if (cs->len != PAGE_SIZE)
 869                 goto out_fallback;
 870
 871         if (buf->ops->steal(cs->pipe, buf) != 0)
 872                 goto out_fallback;
 873
 874         newpage = buf->page;
 875
 876         if (!PageUptodate(newpage))
 877                 SetPageUptodate(newpage);
 878
 879         ClearPageMappedToDisk(newpage);
 880
 881         if (fuse_check_page(newpage) != 0)
 882                 goto out_fallback_unlock;
 883
 884         /*
 885          * This is a new and locked page, it shouldn't be mapped or
 886          * have any special flags on it
 887          */
 888         if (WARN_ON(page_mapped(oldpage)))
 889                 goto out_fallback_unlock;
 890         if (WARN_ON(page_has_private(oldpage)))
 891                 goto out_fallback_unlock;
 892         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 893                 goto out_fallback_unlock;
 894         if (WARN_ON(PageMlocked(oldpage)))
 895                 goto out_fallback_unlock;
 896
 897         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 898         if (err) {
 899                 unlock_page(newpage);
 900                 return err;
 901         }
 902
 903         page_cache_get(newpage);
 904
 905         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 906                 lru_cache_add_file(newpage);
 907
 908         err = 0;
 909         spin_lock(&cs->req->waitq.lock);
 910         if (test_bit(FR_ABORTED, &cs->req->flags))
 911                 err = -ENOENT;
 912         else
 913                 *pagep = newpage;
 914         spin_unlock(&cs->req->waitq.lock);
 915
 916         if (err) {
 917                 unlock_page(newpage);
 918                 page_cache_release(newpage);
 919                 return err;
 920         }
 921
 922         unlock_page(oldpage);
 923         page_cache_release(oldpage);
 924         cs->len = 0;
 925
 926         return 0;
 927
 928 out_fallback_unlock:
 929         unlock_page(newpage);
 930 out_fallback:
 931         cs->pg = buf->page;
 932         cs->offset = buf->offset;
 933
 934         err = lock_request(cs->req);
 935         if (err)
 936                 return err;
 937
 938         return 1;
 939 }
 940
 941 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 942                          unsigned offset, unsigned count)
 943 {
 944         struct pipe_buffer *buf;
 945         int err;
 946
 947         if (cs->nr_segs == cs->pipe->buffers)
 948                 return -EIO;
 949
 950         err = unlock_request(cs->req);
 951         if (err)
 952                 return err;
 953
 954         fuse_copy_finish(cs);
 955
 956         buf = cs->pipebufs;
 957         page_cache_get(page);
 958         buf->page = page;
 959         buf->offset = offset;
 960         buf->len = count;
 961
 962         cs->pipebufs++;
 963         cs->nr_segs++;
 964         cs->len = 0;
 965
 966         return 0;
 967 }
 968
 969 /*
 970  * Copy a page in the request to/from the userspace buffer.  Must be
 971  * done atomically
 972  */
 973 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 974                           unsigned offset, unsigned count, int zeroing)
 975 {
 976         int err;
 977         struct page *page = *pagep;
 978
 979         if (page && zeroing && count < PAGE_SIZE)
 980                 clear_highpage(page);
 981
 982         while (count) {
 983                 if (cs->write && cs->pipebufs && page) {
 984                         return fuse_ref_page(cs, page, offset, count);
 985                 } else if (!cs->len) {
 986                         if (cs->move_pages && page &&
 987                             offset == 0 && count == PAGE_SIZE) {
 988                                 err = fuse_try_move_page(cs, pagep);
 989                                 if (err <= 0)
 990                                         return err;
 991                         } else {
 992                                 err = fuse_copy_fill(cs);
 993                                 if (err)
 994                                         return err;
 995                         }
 996                 }
 997                 if (page) {
 998                         void *mapaddr = kmap_atomic(page);
 999                         void *buf = mapaddr + offset;
1000                         offset += fuse_copy_do(cs, &buf, &count);
1001                         kunmap_atomic(mapaddr);
1002                 } else
1003                         offset += fuse_copy_do(cs, NULL, &count);
1004         }
1005         if (page && !cs->write)
1006                 flush_dcache_page(page);
1007         return 0;
1008 }
1009
1010 /* Copy pages in the request to/from userspace buffer */
1011 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1012                            int zeroing)
1013 {
1014         unsigned i;
1015         struct fuse_req *req = cs->req;
1016
1017         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1018                 int err;
1019                 unsigned offset = req->page_descs[i].offset;
1020                 unsigned count = min(nbytes, req->page_descs[i].length);
1021
1022                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1023                                      zeroing);
1024                 if (err)
1025                         return err;
1026
1027                 nbytes -= count;
1028         }
1029         return 0;
1030 }
1031
1032 /* Copy a single argument in the request to/from userspace buffer */
1033 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1034 {
1035         while (size) {
1036                 if (!cs->len) {
1037                         int err = fuse_copy_fill(cs);
1038                         if (err)
1039                                 return err;
1040                 }
1041                 fuse_copy_do(cs, &val, &size);
1042         }
1043         return 0;
1044 }
1045
1046 /* Copy request arguments to/from userspace buffer */
1047 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1048                           unsigned argpages, struct fuse_arg *args,
1049                           int zeroing)
1050 {
1051         int err = 0;
1052         unsigned i;
1053
1054         for (i = 0; !err && i < numargs; i++)  {
1055                 struct fuse_arg *arg = &args[i];
1056                 if (i == numargs - 1 && argpages)
1057                         err = fuse_copy_pages(cs, arg->size, zeroing);
1058                 else
1059                         err = fuse_copy_one(cs, arg->value, arg->size);
1060         }
1061         return err;
1062 }
1063
1064 static int forget_pending(struct fuse_iqueue *fiq)
1065 {
1066         return fiq->forget_list_head.next != NULL;
1067 }
1068
1069 static int request_pending(struct fuse_iqueue *fiq)
1070 {
1071         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1072                 forget_pending(fiq);
1073 }
1074
1075 /*
1076  * Transfer an interrupt request to userspace
1077  *
1078  * Unlike other requests this is assembled on demand, without a need
1079  * to allocate a separate fuse_req structure.
1080  *
1081  * Called with fiq->waitq.lock held, releases it
1082  */
1083 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1084                                struct fuse_copy_state *cs,
1085                                size_t nbytes, struct fuse_req *req)
1086 __releases(fiq->waitq.lock)
1087 {
1088         struct fuse_in_header ih;
1089         struct fuse_interrupt_in arg;
1090         unsigned reqsize = sizeof(ih) + sizeof(arg);
1091         int err;
1092
1093         list_del_init(&req->intr_entry);
1094         req->intr_unique = fuse_get_unique(fiq);
1095         memset(&ih, 0, sizeof(ih));
1096         memset(&arg, 0, sizeof(arg));
1097         ih.len = reqsize;
1098         ih.opcode = FUSE_INTERRUPT;
1099         ih.unique = req->intr_unique;
1100         arg.unique = req->in.h.unique;
1101
1102         spin_unlock(&fiq->waitq.lock);
1103         if (nbytes < reqsize)
1104                 return -EINVAL;
1105
1106         err = fuse_copy_one(cs, &ih, sizeof(ih));
1107         if (!err)
1108                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1109         fuse_copy_finish(cs);
1110
1111         return err ? err : reqsize;
1112 }
1113
1114 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1115                                                unsigned max,
1116                                                unsigned *countp)
1117 {
1118         struct fuse_forget_link *head = fiq->forget_list_head.next;
1119         struct fuse_forget_link **newhead = &head;
1120         unsigned count;
1121
1122         for (count = 0; *newhead != NULL && count < max; count++)
1123                 newhead = &(*newhead)->next;
1124
1125         fiq->forget_list_head.next = *newhead;
1126         *newhead = NULL;
1127         if (fiq->forget_list_head.next == NULL)
1128                 fiq->forget_list_tail = &fiq->forget_list_head;
1129
1130         if (countp != NULL)
1131                 *countp = count;
1132
1133         return head;
1134 }
1135
1136 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1137                                    struct fuse_copy_state *cs,
1138                                    size_t nbytes)
1139 __releases(fiq->waitq.lock)
1140 {
1141         int err;
1142         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1143         struct fuse_forget_in arg = {
1144                 .nlookup = forget->forget_one.nlookup,
1145         };
1146         struct fuse_in_header ih = {
1147                 .opcode = FUSE_FORGET,
1148                 .nodeid = forget->forget_one.nodeid,
1149                 .unique = fuse_get_unique(fiq),
1150                 .len = sizeof(ih) + sizeof(arg),
1151         };
1152
1153         spin_unlock(&fiq->waitq.lock);
1154         kfree(forget);
1155         if (nbytes < ih.len)
1156                 return -EINVAL;
1157
1158         err = fuse_copy_one(cs, &ih, sizeof(ih));
1159         if (!err)
1160                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1161         fuse_copy_finish(cs);
1162
1163         if (err)
1164                 return err;
1165
1166         return ih.len;
1167 }
1168
1169 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1170                                    struct fuse_copy_state *cs, size_t nbytes)
1171 __releases(fiq->waitq.lock)
1172 {
1173         int err;
1174         unsigned max_forgets;
1175         unsigned count;
1176         struct fuse_forget_link *head;
1177         struct fuse_batch_forget_in arg = { .count = 0 };
1178         struct fuse_in_header ih = {
1179                 .opcode = FUSE_BATCH_FORGET,
1180                 .unique = fuse_get_unique(fiq),
1181                 .len = sizeof(ih) + sizeof(arg),
1182         };
1183
1184         if (nbytes < ih.len) {
1185                 spin_unlock(&fiq->waitq.lock);
1186                 return -EINVAL;
1187         }
1188
1189         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1190         head = dequeue_forget(fiq, max_forgets, &count);
1191         spin_unlock(&fiq->waitq.lock);
1192
1193         arg.count = count;
1194         ih.len += count * sizeof(struct fuse_forget_one);
1195         err = fuse_copy_one(cs, &ih, sizeof(ih));
1196         if (!err)
1197                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1198
1199         while (head) {
1200                 struct fuse_forget_link *forget = head;
1201
1202                 if (!err) {
1203                         err = fuse_copy_one(cs, &forget->forget_one,
1204                                             sizeof(forget->forget_one));
1205                 }
1206                 head = forget->next;
1207                 kfree(forget);
1208         }
1209
1210         fuse_copy_finish(cs);
1211
1212         if (err)
1213                 return err;
1214
1215         return ih.len;
1216 }
1217
1218 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1219                             struct fuse_copy_state *cs,
1220                             size_t nbytes)
1221 __releases(fiq->waitq.lock)
1222 {
1223         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1224                 return fuse_read_single_forget(fiq, cs, nbytes);
1225         else
1226                 return fuse_read_batch_forget(fiq, cs, nbytes);
1227 }
1228
1229 /*
1230  * Read a single request into the userspace filesystem's buffer.  This
1231  * function waits until a request is available, then removes it from
1232  * the pending list and copies request data to userspace buffer.  If
1233  * no reply is needed (FORGET) or request has been aborted or there
1234  * was an error during the copying then it's finished by calling
1235  * request_end().  Otherwise add it to the processing list, and set
1236  * the 'sent' flag.
1237  */
1238 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1239                                 struct fuse_copy_state *cs, size_t nbytes)
1240 {
1241         ssize_t err;
1242         struct fuse_conn *fc = fud->fc;
1243         struct fuse_iqueue *fiq = &fc->iq;
1244         struct fuse_pqueue *fpq = &fud->pq;
1245         struct fuse_req *req;
1246         struct fuse_in *in;
1247         unsigned reqsize;
1248
1249  restart:
1250         spin_lock(&fiq->waitq.lock);
1251         err = -EAGAIN;
1252         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1253             !request_pending(fiq))
1254                 goto err_unlock;
1255
1256         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1257                                 !fiq->connected || request_pending(fiq));
1258         if (err)
1259                 goto err_unlock;
1260
1261         err = -ENODEV;
1262         if (!fiq->connected)
1263                 goto err_unlock;
1264
1265         if (!list_empty(&fiq->interrupts)) {
1266                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1267                                  intr_entry);
1268                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1269         }
1270
1271         if (forget_pending(fiq)) {
1272                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1273                         return fuse_read_forget(fc, fiq, cs, nbytes);
1274
1275                 if (fiq->forget_batch <= -8)
1276                         fiq->forget_batch = 16;
1277         }
1278
1279         req = list_entry(fiq->pending.next, struct fuse_req, list);
1280         clear_bit(FR_PENDING, &req->flags);
1281         list_del_init(&req->list);
1282         spin_unlock(&fiq->waitq.lock);
1283
1284         in = &req->in;
1285         reqsize = in->h.len;
1286         /* If request is too large, reply with an error and restart the read */
1287         if (nbytes < reqsize) {
1288                 req->out.h.error = -EIO;
1289                 /* SETXATTR is special, since it may contain too large data */
1290                 if (in->h.opcode == FUSE_SETXATTR)
1291                         req->out.h.error = -E2BIG;
1292                 request_end(fc, req);
1293                 goto restart;
1294         }
1295         spin_lock(&fpq->lock);
1296         list_add(&req->list, &fpq->io);
1297         spin_unlock(&fpq->lock);
1298         cs->req = req;
1299         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1300         if (!err)
1301                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1302                                      (struct fuse_arg *) in->args, 0);
1303         fuse_copy_finish(cs);
1304         spin_lock(&fpq->lock);
1305         clear_bit(FR_LOCKED, &req->flags);
1306         if (!fpq->connected) {
1307                 err = -ENODEV;
1308                 goto out_end;
1309         }
1310         if (err) {
1311                 req->out.h.error = -EIO;
1312                 goto out_end;
1313         }
1314         if (!test_bit(FR_ISREPLY, &req->flags)) {
1315                 err = reqsize;
1316                 goto out_end;
1317         }
1318         list_move_tail(&req->list, &fpq->processing);
1319         spin_unlock(&fpq->lock);
1320         set_bit(FR_SENT, &req->flags);
1321         /* matches barrier in request_wait_answer() */
1322         smp_mb__after_atomic();
1323         if (test_bit(FR_INTERRUPTED, &req->flags))
1324                 queue_interrupt(fiq, req);
1325
1326         return reqsize;
1327
1328 out_end:
1329         if (!test_bit(FR_PRIVATE, &req->flags))
1330                 list_del_init(&req->list);
1331         spin_unlock(&fpq->lock);
1332         request_end(fc, req);
1333         return err;
1334
1335  err_unlock:
1336         spin_unlock(&fiq->waitq.lock);
1337         return err;
1338 }
1339
1340 static int fuse_dev_open(struct inode *inode, struct file *file)
1341 {
1342         /*
1343          * The fuse device's file's private_data is used to hold
1344          * the fuse_conn(ection) when it is mounted, and is used to
1345          * keep track of whether the file has been mounted already.
1346          */
1347         file->private_data = NULL;
1348         return 0;
1349 }
1350
1351 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1352 {
1353         struct fuse_copy_state cs;
1354         struct file *file = iocb->ki_filp;
1355         struct fuse_dev *fud = fuse_get_dev(file);
1356
1357         if (!fud)
1358                 return -EPERM;
1359
1360         if (!iter_is_iovec(to))
1361                 return -EINVAL;
1362
1363         fuse_copy_init(&cs, 1, to);
1364
1365         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1366 }
1367
1368 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1369                                     struct pipe_inode_info *pipe,
1370                                     size_t len, unsigned int flags)
1371 {
1372         int ret;
1373         int page_nr = 0;
1374         int do_wakeup = 0;
1375         struct pipe_buffer *bufs;
1376         struct fuse_copy_state cs;
1377         struct fuse_dev *fud = fuse_get_dev(in);
1378
1379         if (!fud)
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, 1, NULL);
1387         cs.pipebufs = bufs;
1388         cs.pipe = pipe;
1389         ret = fuse_dev_do_read(fud, 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         /* Don't try to move pages (yet) */
1800         cs->move_pages = 0;
1801
1802         switch (code) {
1803         case FUSE_NOTIFY_POLL:
1804                 return fuse_notify_poll(fc, size, cs);
1805
1806         case FUSE_NOTIFY_INVAL_INODE:
1807                 return fuse_notify_inval_inode(fc, size, cs);
1808
1809         case FUSE_NOTIFY_INVAL_ENTRY:
1810                 return fuse_notify_inval_entry(fc, size, cs);
1811
1812         case FUSE_NOTIFY_STORE:
1813                 return fuse_notify_store(fc, size, cs);
1814
1815         case FUSE_NOTIFY_RETRIEVE:
1816                 return fuse_notify_retrieve(fc, size, cs);
1817
1818         case FUSE_NOTIFY_DELETE:
1819                 return fuse_notify_delete(fc, size, cs);
1820
1821         default:
1822                 fuse_copy_finish(cs);
1823                 return -EINVAL;
1824         }
1825 }
1826
1827 /* Look up request on processing list by unique ID */
1828 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1829 {
1830         struct fuse_req *req;
1831
1832         list_for_each_entry(req, &fpq->processing, list) {
1833                 if (req->in.h.unique == unique || req->intr_unique == unique)
1834                         return req;
1835         }
1836         return NULL;
1837 }
1838
1839 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1840                          unsigned nbytes)
1841 {
1842         unsigned reqsize = sizeof(struct fuse_out_header);
1843
1844         if (out->h.error)
1845                 return nbytes != reqsize ? -EINVAL : 0;
1846
1847         reqsize += len_args(out->numargs, out->args);
1848
1849         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1850                 return -EINVAL;
1851         else if (reqsize > nbytes) {
1852                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1853                 unsigned diffsize = reqsize - nbytes;
1854                 if (diffsize > lastarg->size)
1855                         return -EINVAL;
1856                 lastarg->size -= diffsize;
1857         }
1858         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1859                               out->page_zeroing);
1860 }
1861
1862 /*
1863  * Write a single reply to a request.  First the header is copied from
1864  * the write buffer.  The request is then searched on the processing
1865  * list by the unique ID found in the header.  If found, then remove
1866  * it from the list and copy the rest of the buffer to the request.
1867  * The request is finished by calling request_end()
1868  */
1869 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1870                                  struct fuse_copy_state *cs, size_t nbytes)
1871 {
1872         int err;
1873         struct fuse_conn *fc = fud->fc;
1874         struct fuse_pqueue *fpq = &fud->pq;
1875         struct fuse_req *req;
1876         struct fuse_out_header oh;
1877
1878         if (nbytes < sizeof(struct fuse_out_header))
1879                 return -EINVAL;
1880
1881         err = fuse_copy_one(cs, &oh, sizeof(oh));
1882         if (err)
1883                 goto err_finish;
1884
1885         err = -EINVAL;
1886         if (oh.len != nbytes)
1887                 goto err_finish;
1888
1889         /*
1890          * Zero oh.unique indicates unsolicited notification message
1891          * and error contains notification code.
1892          */
1893         if (!oh.unique) {
1894                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1895                 return err ? err : nbytes;
1896         }
1897
1898         err = -EINVAL;
1899         if (oh.error <= -1000 || oh.error > 0)
1900                 goto err_finish;
1901
1902         spin_lock(&fpq->lock);
1903         err = -ENOENT;
1904         if (!fpq->connected)
1905                 goto err_unlock_pq;
1906
1907         req = request_find(fpq, oh.unique);
1908         if (!req)
1909                 goto err_unlock_pq;
1910
1911         /* Is it an interrupt reply? */
1912         if (req->intr_unique == oh.unique) {
1913                 spin_unlock(&fpq->lock);
1914
1915                 err = -EINVAL;
1916                 if (nbytes != sizeof(struct fuse_out_header))
1917                         goto err_finish;
1918
1919                 if (oh.error == -ENOSYS)
1920                         fc->no_interrupt = 1;
1921                 else if (oh.error == -EAGAIN)
1922                         queue_interrupt(&fc->iq, req);
1923
1924                 fuse_copy_finish(cs);
1925                 return nbytes;
1926         }
1927
1928         clear_bit(FR_SENT, &req->flags);
1929         list_move(&req->list, &fpq->io);
1930         req->out.h = oh;
1931         set_bit(FR_LOCKED, &req->flags);
1932         spin_unlock(&fpq->lock);
1933         cs->req = req;
1934         if (!req->out.page_replace)
1935                 cs->move_pages = 0;
1936
1937         err = copy_out_args(cs, &req->out, nbytes);
1938         fuse_copy_finish(cs);
1939
1940         spin_lock(&fpq->lock);
1941         clear_bit(FR_LOCKED, &req->flags);
1942         if (!fpq->connected)
1943                 err = -ENOENT;
1944         else if (err)
1945                 req->out.h.error = -EIO;
1946         if (!test_bit(FR_PRIVATE, &req->flags))
1947                 list_del_init(&req->list);
1948         spin_unlock(&fpq->lock);
1949
1950         request_end(fc, req);
1951
1952         return err ? err : nbytes;
1953
1954  err_unlock_pq:
1955         spin_unlock(&fpq->lock);
1956  err_finish:
1957         fuse_copy_finish(cs);
1958         return err;
1959 }
1960
1961 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1962 {
1963         struct fuse_copy_state cs;
1964         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1965
1966         if (!fud)
1967                 return -EPERM;
1968
1969         if (!iter_is_iovec(from))
1970                 return -EINVAL;
1971
1972         fuse_copy_init(&cs, 0, from);
1973
1974         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1975 }
1976
1977 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1978                                      struct file *out, loff_t *ppos,
1979                                      size_t len, unsigned int flags)
1980 {
1981         unsigned nbuf;
1982         unsigned idx;
1983         struct pipe_buffer *bufs;
1984         struct fuse_copy_state cs;
1985         struct fuse_dev *fud;
1986         size_t rem;
1987         ssize_t ret;
1988
1989         fud = fuse_get_dev(out);
1990         if (!fud)
1991                 return -EPERM;
1992
1993         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1994         if (!bufs)
1995                 return -ENOMEM;
1996
1997         pipe_lock(pipe);
1998         nbuf = 0;
1999         rem = 0;
2000         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2001                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2002
2003         ret = -EINVAL;
2004         if (rem < len) {
2005                 pipe_unlock(pipe);
2006                 goto out;
2007         }
2008
2009         rem = len;
2010         while (rem) {
2011                 struct pipe_buffer *ibuf;
2012                 struct pipe_buffer *obuf;
2013
2014                 BUG_ON(nbuf >= pipe->buffers);
2015                 BUG_ON(!pipe->nrbufs);
2016                 ibuf = &pipe->bufs[pipe->curbuf];
2017                 obuf = &bufs[nbuf];
2018
2019                 if (rem >= ibuf->len) {
2020                         *obuf = *ibuf;
2021                         ibuf->ops = NULL;
2022                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2023                         pipe->nrbufs--;
2024                 } else {
2025                         ibuf->ops->get(pipe, ibuf);
2026                         *obuf = *ibuf;
2027                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2028                         obuf->len = rem;
2029                         ibuf->offset += obuf->len;
2030                         ibuf->len -= obuf->len;
2031                 }
2032                 nbuf++;
2033                 rem -= obuf->len;
2034         }
2035         pipe_unlock(pipe);
2036
2037         fuse_copy_init(&cs, 0, NULL);
2038         cs.pipebufs = bufs;
2039         cs.nr_segs = nbuf;
2040         cs.pipe = pipe;
2041
2042         if (flags & SPLICE_F_MOVE)
2043                 cs.move_pages = 1;
2044
2045         ret = fuse_dev_do_write(fud, &cs, len);
2046
2047         for (idx = 0; idx < nbuf; idx++) {
2048                 struct pipe_buffer *buf = &bufs[idx];
2049                 buf->ops->release(pipe, buf);
2050         }
2051 out:
2052         kfree(bufs);
2053         return ret;
2054 }
2055
2056 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2057 {
2058         unsigned mask = POLLOUT | POLLWRNORM;
2059         struct fuse_iqueue *fiq;
2060         struct fuse_dev *fud = fuse_get_dev(file);
2061
2062         if (!fud)
2063                 return POLLERR;
2064
2065         fiq = &fud->fc->iq;
2066         poll_wait(file, &fiq->waitq, wait);
2067
2068         spin_lock(&fiq->waitq.lock);
2069         if (!fiq->connected)
2070                 mask = POLLERR;
2071         else if (request_pending(fiq))
2072                 mask |= POLLIN | POLLRDNORM;
2073         spin_unlock(&fiq->waitq.lock);
2074
2075         return mask;
2076 }
2077
2078 /*
2079  * Abort all requests on the given list (pending or processing)
2080  *
2081  * This function releases and reacquires fc->lock
2082  */
2083 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2084 {
2085         while (!list_empty(head)) {
2086                 struct fuse_req *req;
2087                 req = list_entry(head->next, struct fuse_req, list);
2088                 req->out.h.error = -ECONNABORTED;
2089                 clear_bit(FR_PENDING, &req->flags);
2090                 clear_bit(FR_SENT, &req->flags);
2091                 list_del_init(&req->list);
2092                 request_end(fc, req);
2093         }
2094 }
2095
2096 static void end_polls(struct fuse_conn *fc)
2097 {
2098         struct rb_node *p;
2099
2100         p = rb_first(&fc->polled_files);
2101
2102         while (p) {
2103                 struct fuse_file *ff;
2104                 ff = rb_entry(p, struct fuse_file, polled_node);
2105                 wake_up_interruptible_all(&ff->poll_wait);
2106
2107                 p = rb_next(p);
2108         }
2109 }
2110
2111 /*
2112  * Abort all requests.
2113  *
2114  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2115  * filesystem.
2116  *
2117  * The same effect is usually achievable through killing the filesystem daemon
2118  * and all users of the filesystem.  The exception is the combination of an
2119  * asynchronous request and the tricky deadlock (see
2120  * Documentation/filesystems/fuse.txt).
2121  *
2122  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2123  * requests, they should be finished off immediately.  Locked requests will be
2124  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2125  * requests.  It is possible that some request will finish before we can.  This
2126  * is OK, the request will in that case be removed from the list before we touch
2127  * it.
2128  */
2129 void fuse_abort_conn(struct fuse_conn *fc)
2130 {
2131         struct fuse_iqueue *fiq = &fc->iq;
2132
2133         spin_lock(&fc->lock);
2134         if (fc->connected) {
2135                 struct fuse_dev *fud;
2136                 struct fuse_req *req, *next;
2137                 LIST_HEAD(to_end1);
2138                 LIST_HEAD(to_end2);
2139
2140                 fc->connected = 0;
2141                 fc->blocked = 0;
2142                 fuse_set_initialized(fc);
2143                 list_for_each_entry(fud, &fc->devices, entry) {
2144                         struct fuse_pqueue *fpq = &fud->pq;
2145
2146                         spin_lock(&fpq->lock);
2147                         fpq->connected = 0;
2148                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2149                                 req->out.h.error = -ECONNABORTED;
2150                                 spin_lock(&req->waitq.lock);
2151                                 set_bit(FR_ABORTED, &req->flags);
2152                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2153                                         set_bit(FR_PRIVATE, &req->flags);
2154                                         list_move(&req->list, &to_end1);
2155                                 }
2156                                 spin_unlock(&req->waitq.lock);
2157                         }
2158                         list_splice_init(&fpq->processing, &to_end2);
2159                         spin_unlock(&fpq->lock);
2160                 }
2161                 fc->max_background = UINT_MAX;
2162                 flush_bg_queue(fc);
2163
2164                 spin_lock(&fiq->waitq.lock);
2165                 fiq->connected = 0;
2166                 list_splice_init(&fiq->pending, &to_end2);
2167                 while (forget_pending(fiq))
2168                         kfree(dequeue_forget(fiq, 1, NULL));
2169                 wake_up_all_locked(&fiq->waitq);
2170                 spin_unlock(&fiq->waitq.lock);
2171                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2172                 end_polls(fc);
2173                 wake_up_all(&fc->blocked_waitq);
2174                 spin_unlock(&fc->lock);
2175
2176                 while (!list_empty(&to_end1)) {
2177                         req = list_first_entry(&to_end1, struct fuse_req, list);
2178                         __fuse_get_request(req);
2179                         list_del_init(&req->list);
2180                         request_end(fc, req);
2181                 }
2182                 end_requests(fc, &to_end2);
2183         } else {
2184                 spin_unlock(&fc->lock);
2185         }
2186 }
2187 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2188
2189 int fuse_dev_release(struct inode *inode, struct file *file)
2190 {
2191         struct fuse_dev *fud = fuse_get_dev(file);
2192
2193         if (fud) {
2194                 struct fuse_conn *fc = fud->fc;
2195                 struct fuse_pqueue *fpq = &fud->pq;
2196
2197                 WARN_ON(!list_empty(&fpq->io));
2198                 end_requests(fc, &fpq->processing);
2199                 /* Are we the last open device? */
2200                 if (atomic_dec_and_test(&fc->dev_count)) {
2201                         WARN_ON(fc->iq.fasync != NULL);
2202                         fuse_abort_conn(fc);
2203                 }
2204                 fuse_dev_free(fud);
2205         }
2206         return 0;
2207 }
2208 EXPORT_SYMBOL_GPL(fuse_dev_release);
2209
2210 static int fuse_dev_fasync(int fd, struct file *file, int on)
2211 {
2212         struct fuse_dev *fud = fuse_get_dev(file);
2213
2214         if (!fud)
2215                 return -EPERM;
2216
2217         /* No locking - fasync_helper does its own locking */
2218         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2219 }
2220
2221 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2222 {
2223         struct fuse_dev *fud;
2224
2225         if (new->private_data)
2226                 return -EINVAL;
2227
2228         fud = fuse_dev_alloc(fc);
2229         if (!fud)
2230                 return -ENOMEM;
2231
2232         new->private_data = fud;
2233         atomic_inc(&fc->dev_count);
2234
2235         return 0;
2236 }
2237
2238 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2239                            unsigned long arg)
2240 {
2241         int err = -ENOTTY;
2242
2243         if (cmd == FUSE_DEV_IOC_CLONE) {
2244                 int oldfd;
2245
2246                 err = -EFAULT;
2247                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2248                         struct file *old = fget(oldfd);
2249
2250                         err = -EINVAL;
2251                         if (old) {
2252                                 struct fuse_dev *fud = NULL;
2253
2254                                 /*
2255                                  * Check against file->f_op because CUSE
2256                                  * uses the same ioctl handler.
2257                                  */
2258                                 if (old->f_op == file->f_op &&
2259                                     old->f_cred->user_ns == file->f_cred->user_ns)
2260                                         fud = fuse_get_dev(old);
2261
2262                                 if (fud) {
2263                                         mutex_lock(&fuse_mutex);
2264                                         err = fuse_device_clone(fud->fc, file);
2265                                         mutex_unlock(&fuse_mutex);
2266                                 }
2267                                 fput(old);
2268                         }
2269                 }
2270         }
2271         return err;
2272 }
2273
2274 const struct file_operations fuse_dev_operations = {
2275         .owner          = THIS_MODULE,
2276         .open           = fuse_dev_open,
2277         .llseek         = no_llseek,
2278         .read_iter      = fuse_dev_read,
2279         .splice_read    = fuse_dev_splice_read,
2280         .write_iter     = fuse_dev_write,
2281         .splice_write   = fuse_dev_splice_write,
2282         .poll           = fuse_dev_poll,
2283         .release        = fuse_dev_release,
2284         .fasync         = fuse_dev_fasync,
2285         .unlocked_ioctl = fuse_dev_ioctl,
2286         .compat_ioctl   = fuse_dev_ioctl,
2287 };
2288 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2289
2290 static struct miscdevice fuse_miscdevice = {
2291         .minor = FUSE_MINOR,
2292         .name  = "fuse",
2293         .fops = &fuse_dev_operations,
2294 };
2295
2296 int __init fuse_dev_init(void)
2297 {
2298         int err = -ENOMEM;
2299         fuse_req_cachep = kmem_cache_create("fuse_request",
2300                                             sizeof(struct fuse_req),
2301                                             0, 0, NULL);
2302         if (!fuse_req_cachep)
2303                 goto out;
2304
2305         err = misc_register(&fuse_miscdevice);
2306         if (err)
2307                 goto out_cache_clean;
2308
2309         return 0;
2310
2311  out_cache_clean:
2312         kmem_cache_destroy(fuse_req_cachep);
2313  out:
2314         return err;
2315 }
2316
2317 void fuse_dev_cleanup(void)
2318 {
2319         misc_deregister(&fuse_miscdevice);
2320         kmem_cache_destroy(fuse_req_cachep);
2321 }