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
  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         __set_bit(FR_PENDING, &req->flags);
  52 }
  53
  54 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  55 {
  56         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, flags);
  57         if (req) {
  58                 struct page **pages;
  59                 struct fuse_page_desc *page_descs;
  60
  61                 if (npages <= FUSE_REQ_INLINE_PAGES) {
  62                         pages = req->inline_pages;
  63                         page_descs = req->inline_page_descs;
  64                 } else {
  65                         pages = kmalloc(sizeof(struct page *) * npages, flags);
  66                         page_descs = kmalloc(sizeof(struct fuse_page_desc) *
  67                                              npages, flags);
  68                 }
  69
  70                 if (!pages || !page_descs) {
  71                         kfree(pages);
  72                         kfree(page_descs);
  73                         kmem_cache_free(fuse_req_cachep, req);
  74                         return NULL;
  75                 }
  76
  77                 fuse_request_init(req, pages, page_descs, npages);
  78         }
  79         return req;
  80 }
  81
  82 struct fuse_req *fuse_request_alloc(unsigned npages)
  83 {
  84         return __fuse_request_alloc(npages, GFP_KERNEL);
  85 }
  86 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  87
  88 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  89 {
  90         return __fuse_request_alloc(npages, GFP_NOFS);
  91 }
  92
  93 void fuse_request_free(struct fuse_req *req)
  94 {
  95         if (req->pages != req->inline_pages) {
  96                 kfree(req->pages);
  97                 kfree(req->page_descs);
  98         }
  99         kmem_cache_free(fuse_req_cachep, req);
 100 }
 101
 102 static void block_sigs(sigset_t *oldset)
 103 {
 104         sigset_t mask;
 105
 106         siginitsetinv(&mask, sigmask(SIGKILL));
 107         sigprocmask(SIG_BLOCK, &mask, oldset);
 108 }
 109
 110 static void restore_sigs(sigset_t *oldset)
 111 {
 112         sigprocmask(SIG_SETMASK, oldset, NULL);
 113 }
 114
 115 void __fuse_get_request(struct fuse_req *req)
 116 {
 117         atomic_inc(&req->count);
 118 }
 119
 120 /* Must be called with > 1 refcount */
 121 static void __fuse_put_request(struct fuse_req *req)
 122 {
 123         BUG_ON(atomic_read(&req->count) < 2);
 124         atomic_dec(&req->count);
 125 }
 126
 127 static void fuse_req_init_context(struct fuse_req *req)
 128 {
 129         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 130         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 131         req->in.h.pid = current->pid;
 132 }
 133
 134 void fuse_set_initialized(struct fuse_conn *fc)
 135 {
 136         /* Make sure stores before this are seen on another CPU */
 137         smp_wmb();
 138         fc->initialized = 1;
 139 }
 140
 141 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 142 {
 143         return !fc->initialized || (for_background && fc->blocked);
 144 }
 145
 146 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 147                                        bool for_background)
 148 {
 149         struct fuse_req *req;
 150         int err;
 151         atomic_inc(&fc->num_waiting);
 152
 153         if (fuse_block_alloc(fc, for_background)) {
 154                 sigset_t oldset;
 155                 int intr;
 156
 157                 block_sigs(&oldset);
 158                 intr = wait_event_interruptible_exclusive(fc->blocked_waitq,
 159                                 !fuse_block_alloc(fc, for_background));
 160                 restore_sigs(&oldset);
 161                 err = -EINTR;
 162                 if (intr)
 163                         goto out;
 164         }
 165         /* Matches smp_wmb() in fuse_set_initialized() */
 166         smp_rmb();
 167
 168         err = -ENOTCONN;
 169         if (!fc->connected)
 170                 goto out;
 171
 172         err = -ECONNREFUSED;
 173         if (fc->conn_error)
 174                 goto out;
 175
 176         req = fuse_request_alloc(npages);
 177         err = -ENOMEM;
 178         if (!req) {
 179                 if (for_background)
 180                         wake_up(&fc->blocked_waitq);
 181                 goto out;
 182         }
 183
 184         fuse_req_init_context(req);
 185         __set_bit(FR_WAITING, &req->flags);
 186         if (for_background)
 187                 __set_bit(FR_BACKGROUND, &req->flags);
 188
 189         return req;
 190
 191  out:
 192         atomic_dec(&fc->num_waiting);
 193         return ERR_PTR(err);
 194 }
 195
 196 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 197 {
 198         return __fuse_get_req(fc, npages, false);
 199 }
 200 EXPORT_SYMBOL_GPL(fuse_get_req);
 201
 202 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 203                                              unsigned npages)
 204 {
 205         return __fuse_get_req(fc, npages, true);
 206 }
 207 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 208
 209 /*
 210  * Return request in fuse_file->reserved_req.  However that may
 211  * currently be in use.  If that is the case, wait for it to become
 212  * available.
 213  */
 214 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 215                                          struct file *file)
 216 {
 217         struct fuse_req *req = NULL;
 218         struct fuse_file *ff = file->private_data;
 219
 220         do {
 221                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 222                 spin_lock(&fc->lock);
 223                 if (ff->reserved_req) {
 224                         req = ff->reserved_req;
 225                         ff->reserved_req = NULL;
 226                         req->stolen_file = get_file(file);
 227                 }
 228                 spin_unlock(&fc->lock);
 229         } while (!req);
 230
 231         return req;
 232 }
 233
 234 /*
 235  * Put stolen request back into fuse_file->reserved_req
 236  */
 237 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 238 {
 239         struct file *file = req->stolen_file;
 240         struct fuse_file *ff = file->private_data;
 241
 242         spin_lock(&fc->lock);
 243         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 244         BUG_ON(ff->reserved_req);
 245         ff->reserved_req = req;
 246         wake_up_all(&fc->reserved_req_waitq);
 247         spin_unlock(&fc->lock);
 248         fput(file);
 249 }
 250
 251 /*
 252  * Gets a requests for a file operation, always succeeds
 253  *
 254  * This is used for sending the FLUSH request, which must get to
 255  * userspace, due to POSIX locks which may need to be unlocked.
 256  *
 257  * If allocation fails due to OOM, use the reserved request in
 258  * fuse_file.
 259  *
 260  * This is very unlikely to deadlock accidentally, since the
 261  * filesystem should not have it's own file open.  If deadlock is
 262  * intentional, it can still be broken by "aborting" the filesystem.
 263  */
 264 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 265                                              struct file *file)
 266 {
 267         struct fuse_req *req;
 268
 269         atomic_inc(&fc->num_waiting);
 270         wait_event(fc->blocked_waitq, fc->initialized);
 271         /* Matches smp_wmb() in fuse_set_initialized() */
 272         smp_rmb();
 273         req = fuse_request_alloc(0);
 274         if (!req)
 275                 req = get_reserved_req(fc, file);
 276
 277         fuse_req_init_context(req);
 278         __set_bit(FR_WAITING, &req->flags);
 279         __clear_bit(FR_BACKGROUND, &req->flags);
 280         return req;
 281 }
 282
 283 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 284 {
 285         if (atomic_dec_and_test(&req->count)) {
 286                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 287                         /*
 288                          * We get here in the unlikely case that a background
 289                          * request was allocated but not sent
 290                          */
 291                         spin_lock(&fc->lock);
 292                         if (!fc->blocked)
 293                                 wake_up(&fc->blocked_waitq);
 294                         spin_unlock(&fc->lock);
 295                 }
 296
 297                 if (test_bit(FR_WAITING, &req->flags)) {
 298                         __clear_bit(FR_WAITING, &req->flags);
 299                         atomic_dec(&fc->num_waiting);
 300                 }
 301
 302                 if (req->stolen_file)
 303                         put_reserved_req(fc, req);
 304                 else
 305                         fuse_request_free(req);
 306         }
 307 }
 308 EXPORT_SYMBOL_GPL(fuse_put_request);
 309
 310 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 311 {
 312         unsigned nbytes = 0;
 313         unsigned i;
 314
 315         for (i = 0; i < numargs; i++)
 316                 nbytes += args[i].size;
 317
 318         return nbytes;
 319 }
 320
 321 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 322 {
 323         return ++fiq->reqctr;
 324 }
 325
 326 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 327 {
 328         req->in.h.len = sizeof(struct fuse_in_header) +
 329                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 330         list_add_tail(&req->list, &fiq->pending);
 331         wake_up_locked(&fiq->waitq);
 332         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 333 }
 334
 335 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 336                        u64 nodeid, u64 nlookup)
 337 {
 338         struct fuse_iqueue *fiq = &fc->iq;
 339
 340         forget->forget_one.nodeid = nodeid;
 341         forget->forget_one.nlookup = nlookup;
 342
 343         spin_lock(&fiq->waitq.lock);
 344         if (fiq->connected) {
 345                 fiq->forget_list_tail->next = forget;
 346                 fiq->forget_list_tail = forget;
 347                 wake_up_locked(&fiq->waitq);
 348                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 349         } else {
 350                 kfree(forget);
 351         }
 352         spin_unlock(&fiq->waitq.lock);
 353 }
 354
 355 static void flush_bg_queue(struct fuse_conn *fc)
 356 {
 357         while (fc->active_background < fc->max_background &&
 358                !list_empty(&fc->bg_queue)) {
 359                 struct fuse_req *req;
 360                 struct fuse_iqueue *fiq = &fc->iq;
 361
 362                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 363                 list_del(&req->list);
 364                 fc->active_background++;
 365                 spin_lock(&fiq->waitq.lock);
 366                 req->in.h.unique = fuse_get_unique(fiq);
 367                 queue_request(fiq, req);
 368                 spin_unlock(&fiq->waitq.lock);
 369         }
 370 }
 371
 372 /*
 373  * This function is called when a request is finished.  Either a reply
 374  * has arrived or it was aborted (and not yet sent) or some error
 375  * occurred during communication with userspace, or the device file
 376  * was closed.  The requester thread is woken up (if still waiting),
 377  * the 'end' callback is called if given, else the reference to the
 378  * request is released
 379  */
 380 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 381 {
 382         struct fuse_iqueue *fiq = &fc->iq;
 383
 384         if (test_and_set_bit(FR_FINISHED, &req->flags))
 385                 return;
 386
 387         spin_lock(&fiq->waitq.lock);
 388         list_del_init(&req->intr_entry);
 389         spin_unlock(&fiq->waitq.lock);
 390         WARN_ON(test_bit(FR_PENDING, &req->flags));
 391         WARN_ON(test_bit(FR_SENT, &req->flags));
 392         if (test_bit(FR_BACKGROUND, &req->flags)) {
 393                 spin_lock(&fc->lock);
 394                 clear_bit(FR_BACKGROUND, &req->flags);
 395                 if (fc->num_background == fc->max_background)
 396                         fc->blocked = 0;
 397
 398                 /* Wake up next waiter, if any */
 399                 if (!fc->blocked && waitqueue_active(&fc->blocked_waitq))
 400                         wake_up(&fc->blocked_waitq);
 401
 402                 if (fc->num_background == fc->congestion_threshold &&
 403                     fc->connected && fc->bdi_initialized) {
 404                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 405                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 406                 }
 407                 fc->num_background--;
 408                 fc->active_background--;
 409                 flush_bg_queue(fc);
 410                 spin_unlock(&fc->lock);
 411         }
 412         wake_up(&req->waitq);
 413         if (req->end)
 414                 req->end(fc, req);
 415         fuse_put_request(fc, req);
 416 }
 417
 418 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 419 {
 420         spin_lock(&fiq->waitq.lock);
 421         if (list_empty(&req->intr_entry)) {
 422                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 423                 wake_up_locked(&fiq->waitq);
 424         }
 425         spin_unlock(&fiq->waitq.lock);
 426         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 427 }
 428
 429 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 430 {
 431         struct fuse_iqueue *fiq = &fc->iq;
 432         int err;
 433
 434         if (!fc->no_interrupt) {
 435                 /* Any signal may interrupt this */
 436                 err = wait_event_interruptible(req->waitq,
 437                                         test_bit(FR_FINISHED, &req->flags));
 438                 if (!err)
 439                         return;
 440
 441                 set_bit(FR_INTERRUPTED, &req->flags);
 442                 /* matches barrier in fuse_dev_do_read() */
 443                 smp_mb__after_atomic();
 444                 if (test_bit(FR_SENT, &req->flags))
 445                         queue_interrupt(fiq, req);
 446         }
 447
 448         if (!test_bit(FR_FORCE, &req->flags)) {
 449                 sigset_t oldset;
 450
 451                 /* Only fatal signals may interrupt this */
 452                 block_sigs(&oldset);
 453                 err = wait_event_interruptible(req->waitq,
 454                                         test_bit(FR_FINISHED, &req->flags));
 455                 restore_sigs(&oldset);
 456
 457                 if (!err)
 458                         return;
 459
 460                 spin_lock(&fiq->waitq.lock);
 461                 /* Request is not yet in userspace, bail out */
 462                 if (test_bit(FR_PENDING, &req->flags)) {
 463                         list_del(&req->list);
 464                         spin_unlock(&fiq->waitq.lock);
 465                         __fuse_put_request(req);
 466                         req->out.h.error = -EINTR;
 467                         return;
 468                 }
 469                 spin_unlock(&fiq->waitq.lock);
 470         }
 471
 472         /*
 473          * Either request is already in userspace, or it was forced.
 474          * Wait it out.
 475          */
 476         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 477 }
 478
 479 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 480 {
 481         struct fuse_iqueue *fiq = &fc->iq;
 482
 483         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 484         spin_lock(&fiq->waitq.lock);
 485         if (!fiq->connected) {
 486                 spin_unlock(&fiq->waitq.lock);
 487                 req->out.h.error = -ENOTCONN;
 488         } else {
 489                 req->in.h.unique = fuse_get_unique(fiq);
 490                 queue_request(fiq, req);
 491                 /* acquire extra reference, since request is still needed
 492                    after request_end() */
 493                 __fuse_get_request(req);
 494                 spin_unlock(&fiq->waitq.lock);
 495
 496                 request_wait_answer(fc, req);
 497                 /* Pairs with smp_wmb() in request_end() */
 498                 smp_rmb();
 499         }
 500 }
 501
 502 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 503 {
 504         __set_bit(FR_ISREPLY, &req->flags);
 505         if (!test_bit(FR_WAITING, &req->flags)) {
 506                 __set_bit(FR_WAITING, &req->flags);
 507                 atomic_inc(&fc->num_waiting);
 508         }
 509         __fuse_request_send(fc, req);
 510 }
 511 EXPORT_SYMBOL_GPL(fuse_request_send);
 512
 513 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 514 {
 515         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 516                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 517
 518         if (fc->minor < 9) {
 519                 switch (args->in.h.opcode) {
 520                 case FUSE_LOOKUP:
 521                 case FUSE_CREATE:
 522                 case FUSE_MKNOD:
 523                 case FUSE_MKDIR:
 524                 case FUSE_SYMLINK:
 525                 case FUSE_LINK:
 526                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 527                         break;
 528                 case FUSE_GETATTR:
 529                 case FUSE_SETATTR:
 530                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 531                         break;
 532                 }
 533         }
 534         if (fc->minor < 12) {
 535                 switch (args->in.h.opcode) {
 536                 case FUSE_CREATE:
 537                         args->in.args[0].size = sizeof(struct fuse_open_in);
 538                         break;
 539                 case FUSE_MKNOD:
 540                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 541                         break;
 542                 }
 543         }
 544 }
 545
 546 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 547 {
 548         struct fuse_req *req;
 549         ssize_t ret;
 550
 551         req = fuse_get_req(fc, 0);
 552         if (IS_ERR(req))
 553                 return PTR_ERR(req);
 554
 555         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 556         fuse_adjust_compat(fc, args);
 557
 558         req->in.h.opcode = args->in.h.opcode;
 559         req->in.h.nodeid = args->in.h.nodeid;
 560         req->in.numargs = args->in.numargs;
 561         memcpy(req->in.args, args->in.args,
 562                args->in.numargs * sizeof(struct fuse_in_arg));
 563         req->out.argvar = args->out.argvar;
 564         req->out.numargs = args->out.numargs;
 565         memcpy(req->out.args, args->out.args,
 566                args->out.numargs * sizeof(struct fuse_arg));
 567         fuse_request_send(fc, req);
 568         ret = req->out.h.error;
 569         if (!ret && args->out.argvar) {
 570                 BUG_ON(args->out.numargs != 1);
 571                 ret = req->out.args[0].size;
 572         }
 573         fuse_put_request(fc, req);
 574
 575         return ret;
 576 }
 577
 578 /*
 579  * Called under fc->lock
 580  *
 581  * fc->connected must have been checked previously
 582  */
 583 void fuse_request_send_background_locked(struct fuse_conn *fc,
 584                                          struct fuse_req *req)
 585 {
 586         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 587         if (!test_bit(FR_WAITING, &req->flags)) {
 588                 __set_bit(FR_WAITING, &req->flags);
 589                 atomic_inc(&fc->num_waiting);
 590         }
 591         __set_bit(FR_ISREPLY, &req->flags);
 592         fc->num_background++;
 593         if (fc->num_background == fc->max_background)
 594                 fc->blocked = 1;
 595         if (fc->num_background == fc->congestion_threshold &&
 596             fc->bdi_initialized) {
 597                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 598                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 599         }
 600         list_add_tail(&req->list, &fc->bg_queue);
 601         flush_bg_queue(fc);
 602 }
 603
 604 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 605 {
 606         BUG_ON(!req->end);
 607         spin_lock(&fc->lock);
 608         if (fc->connected) {
 609                 fuse_request_send_background_locked(fc, req);
 610                 spin_unlock(&fc->lock);
 611         } else {
 612                 spin_unlock(&fc->lock);
 613                 req->out.h.error = -ENOTCONN;
 614                 req->end(fc, req);
 615                 fuse_put_request(fc, req);
 616         }
 617 }
 618 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 619
 620 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 621                                           struct fuse_req *req, u64 unique)
 622 {
 623         int err = -ENODEV;
 624         struct fuse_iqueue *fiq = &fc->iq;
 625
 626         __clear_bit(FR_ISREPLY, &req->flags);
 627         req->in.h.unique = unique;
 628         spin_lock(&fiq->waitq.lock);
 629         if (fiq->connected) {
 630                 queue_request(fiq, req);
 631                 err = 0;
 632         }
 633         spin_unlock(&fiq->waitq.lock);
 634
 635         return err;
 636 }
 637
 638 void fuse_force_forget(struct file *file, u64 nodeid)
 639 {
 640         struct inode *inode = file_inode(file);
 641         struct fuse_conn *fc = get_fuse_conn(inode);
 642         struct fuse_req *req;
 643         struct fuse_forget_in inarg;
 644
 645         memset(&inarg, 0, sizeof(inarg));
 646         inarg.nlookup = 1;
 647         req = fuse_get_req_nofail_nopages(fc, file);
 648         req->in.h.opcode = FUSE_FORGET;
 649         req->in.h.nodeid = nodeid;
 650         req->in.numargs = 1;
 651         req->in.args[0].size = sizeof(inarg);
 652         req->in.args[0].value = &inarg;
 653         __clear_bit(FR_ISREPLY, &req->flags);
 654         __fuse_request_send(fc, req);
 655         /* ignore errors */
 656         fuse_put_request(fc, req);
 657 }
 658
 659 /*
 660  * Lock the request.  Up to the next unlock_request() there mustn't be
 661  * anything that could cause a page-fault.  If the request was already
 662  * aborted bail out.
 663  */
 664 static int lock_request(struct fuse_req *req)
 665 {
 666         int err = 0;
 667         if (req) {
 668                 spin_lock(&req->waitq.lock);
 669                 if (test_bit(FR_ABORTED, &req->flags))
 670                         err = -ENOENT;
 671                 else
 672                         set_bit(FR_LOCKED, &req->flags);
 673                 spin_unlock(&req->waitq.lock);
 674         }
 675         return err;
 676 }
 677
 678 /*
 679  * Unlock request.  If it was aborted while locked, caller is responsible
 680  * for unlocking and ending the request.
 681  */
 682 static int unlock_request(struct fuse_req *req)
 683 {
 684         int err = 0;
 685         if (req) {
 686                 spin_lock(&req->waitq.lock);
 687                 if (test_bit(FR_ABORTED, &req->flags))
 688                         err = -ENOENT;
 689                 else
 690                         clear_bit(FR_LOCKED, &req->flags);
 691                 spin_unlock(&req->waitq.lock);
 692         }
 693         return err;
 694 }
 695
 696 struct fuse_copy_state {
 697         int write;
 698         struct fuse_req *req;
 699         struct iov_iter *iter;
 700         struct pipe_buffer *pipebufs;
 701         struct pipe_buffer *currbuf;
 702         struct pipe_inode_info *pipe;
 703         unsigned long nr_segs;
 704         struct page *pg;
 705         unsigned len;
 706         unsigned offset;
 707         unsigned move_pages:1;
 708 };
 709
 710 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 711                            struct iov_iter *iter)
 712 {
 713         memset(cs, 0, sizeof(*cs));
 714         cs->write = write;
 715         cs->iter = iter;
 716 }
 717
 718 /* Unmap and put previous page of userspace buffer */
 719 static void fuse_copy_finish(struct fuse_copy_state *cs)
 720 {
 721         if (cs->currbuf) {
 722                 struct pipe_buffer *buf = cs->currbuf;
 723
 724                 if (cs->write)
 725                         buf->len = PAGE_SIZE - cs->len;
 726                 cs->currbuf = NULL;
 727         } else if (cs->pg) {
 728                 if (cs->write) {
 729                         flush_dcache_page(cs->pg);
 730                         set_page_dirty_lock(cs->pg);
 731                 }
 732                 put_page(cs->pg);
 733         }
 734         cs->pg = NULL;
 735 }
 736
 737 /*
 738  * Get another pagefull of userspace buffer, and map it to kernel
 739  * address space, and lock request
 740  */
 741 static int fuse_copy_fill(struct fuse_copy_state *cs)
 742 {
 743         struct page *page;
 744         int err;
 745
 746         err = unlock_request(cs->req);
 747         if (err)
 748                 return err;
 749
 750         fuse_copy_finish(cs);
 751         if (cs->pipebufs) {
 752                 struct pipe_buffer *buf = cs->pipebufs;
 753
 754                 if (!cs->write) {
 755                         err = buf->ops->confirm(cs->pipe, buf);
 756                         if (err)
 757                                 return err;
 758
 759                         BUG_ON(!cs->nr_segs);
 760                         cs->currbuf = buf;
 761                         cs->pg = buf->page;
 762                         cs->offset = buf->offset;
 763                         cs->len = buf->len;
 764                         cs->pipebufs++;
 765                         cs->nr_segs--;
 766                 } else {
 767                         if (cs->nr_segs == cs->pipe->buffers)
 768                                 return -EIO;
 769
 770                         page = alloc_page(GFP_HIGHUSER);
 771                         if (!page)
 772                                 return -ENOMEM;
 773
 774                         buf->page = page;
 775                         buf->offset = 0;
 776                         buf->len = 0;
 777
 778                         cs->currbuf = buf;
 779                         cs->pg = page;
 780                         cs->offset = 0;
 781                         cs->len = PAGE_SIZE;
 782                         cs->pipebufs++;
 783                         cs->nr_segs++;
 784                 }
 785         } else {
 786                 size_t off;
 787                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 788                 if (err < 0)
 789                         return err;
 790                 BUG_ON(!err);
 791                 cs->len = err;
 792                 cs->offset = off;
 793                 cs->pg = page;
 794                 cs->offset = off;
 795                 iov_iter_advance(cs->iter, err);
 796         }
 797
 798         return lock_request(cs->req);
 799 }
 800
 801 /* Do as much copy to/from userspace buffer as we can */
 802 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 803 {
 804         unsigned ncpy = min(*size, cs->len);
 805         if (val) {
 806                 void *pgaddr = kmap_atomic(cs->pg);
 807                 void *buf = pgaddr + cs->offset;
 808
 809                 if (cs->write)
 810                         memcpy(buf, *val, ncpy);
 811                 else
 812                         memcpy(*val, buf, ncpy);
 813
 814                 kunmap_atomic(pgaddr);
 815                 *val += ncpy;
 816         }
 817         *size -= ncpy;
 818         cs->len -= ncpy;
 819         cs->offset += ncpy;
 820         return ncpy;
 821 }
 822
 823 static int fuse_check_page(struct page *page)
 824 {
 825         if (page_mapcount(page) ||
 826             page->mapping != NULL ||
 827             page_count(page) != 1 ||
 828             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 829              ~(1 << PG_locked |
 830                1 << PG_referenced |
 831                1 << PG_uptodate |
 832                1 << PG_lru |
 833                1 << PG_active |
 834                1 << PG_reclaim))) {
 835                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 836                 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);
 837                 return 1;
 838         }
 839         return 0;
 840 }
 841
 842 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 843 {
 844         int err;
 845         struct page *oldpage = *pagep;
 846         struct page *newpage;
 847         struct pipe_buffer *buf = cs->pipebufs;
 848
 849         err = unlock_request(cs->req);
 850         if (err)
 851                 return err;
 852
 853         fuse_copy_finish(cs);
 854
 855         err = buf->ops->confirm(cs->pipe, buf);
 856         if (err)
 857                 return err;
 858
 859         BUG_ON(!cs->nr_segs);
 860         cs->currbuf = buf;
 861         cs->len = buf->len;
 862         cs->pipebufs++;
 863         cs->nr_segs--;
 864
 865         if (cs->len != PAGE_SIZE)
 866                 goto out_fallback;
 867
 868         if (buf->ops->steal(cs->pipe, buf) != 0)
 869                 goto out_fallback;
 870
 871         newpage = buf->page;
 872
 873         if (!PageUptodate(newpage))
 874                 SetPageUptodate(newpage);
 875
 876         ClearPageMappedToDisk(newpage);
 877
 878         if (fuse_check_page(newpage) != 0)
 879                 goto out_fallback_unlock;
 880
 881         /*
 882          * This is a new and locked page, it shouldn't be mapped or
 883          * have any special flags on it
 884          */
 885         if (WARN_ON(page_mapped(oldpage)))
 886                 goto out_fallback_unlock;
 887         if (WARN_ON(page_has_private(oldpage)))
 888                 goto out_fallback_unlock;
 889         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 890                 goto out_fallback_unlock;
 891         if (WARN_ON(PageMlocked(oldpage)))
 892                 goto out_fallback_unlock;
 893
 894         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 895         if (err) {
 896                 unlock_page(newpage);
 897                 return err;
 898         }
 899
 900         page_cache_get(newpage);
 901
 902         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 903                 lru_cache_add_file(newpage);
 904
 905         err = 0;
 906         spin_lock(&cs->req->waitq.lock);
 907         if (test_bit(FR_ABORTED, &cs->req->flags))
 908                 err = -ENOENT;
 909         else
 910                 *pagep = newpage;
 911         spin_unlock(&cs->req->waitq.lock);
 912
 913         if (err) {
 914                 unlock_page(newpage);
 915                 page_cache_release(newpage);
 916                 return err;
 917         }
 918
 919         unlock_page(oldpage);
 920         page_cache_release(oldpage);
 921         cs->len = 0;
 922
 923         return 0;
 924
 925 out_fallback_unlock:
 926         unlock_page(newpage);
 927 out_fallback:
 928         cs->pg = buf->page;
 929         cs->offset = buf->offset;
 930
 931         err = lock_request(cs->req);
 932         if (err)
 933                 return err;
 934
 935         return 1;
 936 }
 937
 938 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 939                          unsigned offset, unsigned count)
 940 {
 941         struct pipe_buffer *buf;
 942         int err;
 943
 944         if (cs->nr_segs == cs->pipe->buffers)
 945                 return -EIO;
 946
 947         err = unlock_request(cs->req);
 948         if (err)
 949                 return err;
 950
 951         fuse_copy_finish(cs);
 952
 953         buf = cs->pipebufs;
 954         page_cache_get(page);
 955         buf->page = page;
 956         buf->offset = offset;
 957         buf->len = count;
 958
 959         cs->pipebufs++;
 960         cs->nr_segs++;
 961         cs->len = 0;
 962
 963         return 0;
 964 }
 965
 966 /*
 967  * Copy a page in the request to/from the userspace buffer.  Must be
 968  * done atomically
 969  */
 970 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 971                           unsigned offset, unsigned count, int zeroing)
 972 {
 973         int err;
 974         struct page *page = *pagep;
 975
 976         if (page && zeroing && count < PAGE_SIZE)
 977                 clear_highpage(page);
 978
 979         while (count) {
 980                 if (cs->write && cs->pipebufs && page) {
 981                         return fuse_ref_page(cs, page, offset, count);
 982                 } else if (!cs->len) {
 983                         if (cs->move_pages && page &&
 984                             offset == 0 && count == PAGE_SIZE) {
 985                                 err = fuse_try_move_page(cs, pagep);
 986                                 if (err <= 0)
 987                                         return err;
 988                         } else {
 989                                 err = fuse_copy_fill(cs);
 990                                 if (err)
 991                                         return err;
 992                         }
 993                 }
 994                 if (page) {
 995                         void *mapaddr = kmap_atomic(page);
 996                         void *buf = mapaddr + offset;
 997                         offset += fuse_copy_do(cs, &buf, &count);
 998                         kunmap_atomic(mapaddr);
 999                 } else
1000                         offset += fuse_copy_do(cs, NULL, &count);
1001         }
1002         if (page && !cs->write)
1003                 flush_dcache_page(page);
1004         return 0;
1005 }
1006
1007 /* Copy pages in the request to/from userspace buffer */
1008 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1009                            int zeroing)
1010 {
1011         unsigned i;
1012         struct fuse_req *req = cs->req;
1013
1014         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1015                 int err;
1016                 unsigned offset = req->page_descs[i].offset;
1017                 unsigned count = min(nbytes, req->page_descs[i].length);
1018
1019                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1020                                      zeroing);
1021                 if (err)
1022                         return err;
1023
1024                 nbytes -= count;
1025         }
1026         return 0;
1027 }
1028
1029 /* Copy a single argument in the request to/from userspace buffer */
1030 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1031 {
1032         while (size) {
1033                 if (!cs->len) {
1034                         int err = fuse_copy_fill(cs);
1035                         if (err)
1036                                 return err;
1037                 }
1038                 fuse_copy_do(cs, &val, &size);
1039         }
1040         return 0;
1041 }
1042
1043 /* Copy request arguments to/from userspace buffer */
1044 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1045                           unsigned argpages, struct fuse_arg *args,
1046                           int zeroing)
1047 {
1048         int err = 0;
1049         unsigned i;
1050
1051         for (i = 0; !err && i < numargs; i++)  {
1052                 struct fuse_arg *arg = &args[i];
1053                 if (i == numargs - 1 && argpages)
1054                         err = fuse_copy_pages(cs, arg->size, zeroing);
1055                 else
1056                         err = fuse_copy_one(cs, arg->value, arg->size);
1057         }
1058         return err;
1059 }
1060
1061 static int forget_pending(struct fuse_iqueue *fiq)
1062 {
1063         return fiq->forget_list_head.next != NULL;
1064 }
1065
1066 static int request_pending(struct fuse_iqueue *fiq)
1067 {
1068         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1069                 forget_pending(fiq);
1070 }
1071
1072 /*
1073  * Transfer an interrupt request to userspace
1074  *
1075  * Unlike other requests this is assembled on demand, without a need
1076  * to allocate a separate fuse_req structure.
1077  *
1078  * Called with fiq->waitq.lock held, releases it
1079  */
1080 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1081                                struct fuse_copy_state *cs,
1082                                size_t nbytes, struct fuse_req *req)
1083 __releases(fiq->waitq.lock)
1084 {
1085         struct fuse_in_header ih;
1086         struct fuse_interrupt_in arg;
1087         unsigned reqsize = sizeof(ih) + sizeof(arg);
1088         int err;
1089
1090         list_del_init(&req->intr_entry);
1091         req->intr_unique = fuse_get_unique(fiq);
1092         memset(&ih, 0, sizeof(ih));
1093         memset(&arg, 0, sizeof(arg));
1094         ih.len = reqsize;
1095         ih.opcode = FUSE_INTERRUPT;
1096         ih.unique = req->intr_unique;
1097         arg.unique = req->in.h.unique;
1098
1099         spin_unlock(&fiq->waitq.lock);
1100         if (nbytes < reqsize)
1101                 return -EINVAL;
1102
1103         err = fuse_copy_one(cs, &ih, sizeof(ih));
1104         if (!err)
1105                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1106         fuse_copy_finish(cs);
1107
1108         return err ? err : reqsize;
1109 }
1110
1111 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1112                                                unsigned max,
1113                                                unsigned *countp)
1114 {
1115         struct fuse_forget_link *head = fiq->forget_list_head.next;
1116         struct fuse_forget_link **newhead = &head;
1117         unsigned count;
1118
1119         for (count = 0; *newhead != NULL && count < max; count++)
1120                 newhead = &(*newhead)->next;
1121
1122         fiq->forget_list_head.next = *newhead;
1123         *newhead = NULL;
1124         if (fiq->forget_list_head.next == NULL)
1125                 fiq->forget_list_tail = &fiq->forget_list_head;
1126
1127         if (countp != NULL)
1128                 *countp = count;
1129
1130         return head;
1131 }
1132
1133 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1134                                    struct fuse_copy_state *cs,
1135                                    size_t nbytes)
1136 __releases(fiq->waitq.lock)
1137 {
1138         int err;
1139         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1140         struct fuse_forget_in arg = {
1141                 .nlookup = forget->forget_one.nlookup,
1142         };
1143         struct fuse_in_header ih = {
1144                 .opcode = FUSE_FORGET,
1145                 .nodeid = forget->forget_one.nodeid,
1146                 .unique = fuse_get_unique(fiq),
1147                 .len = sizeof(ih) + sizeof(arg),
1148         };
1149
1150         spin_unlock(&fiq->waitq.lock);
1151         kfree(forget);
1152         if (nbytes < ih.len)
1153                 return -EINVAL;
1154
1155         err = fuse_copy_one(cs, &ih, sizeof(ih));
1156         if (!err)
1157                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1158         fuse_copy_finish(cs);
1159
1160         if (err)
1161                 return err;
1162
1163         return ih.len;
1164 }
1165
1166 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1167                                    struct fuse_copy_state *cs, size_t nbytes)
1168 __releases(fiq->waitq.lock)
1169 {
1170         int err;
1171         unsigned max_forgets;
1172         unsigned count;
1173         struct fuse_forget_link *head;
1174         struct fuse_batch_forget_in arg = { .count = 0 };
1175         struct fuse_in_header ih = {
1176                 .opcode = FUSE_BATCH_FORGET,
1177                 .unique = fuse_get_unique(fiq),
1178                 .len = sizeof(ih) + sizeof(arg),
1179         };
1180
1181         if (nbytes < ih.len) {
1182                 spin_unlock(&fiq->waitq.lock);
1183                 return -EINVAL;
1184         }
1185
1186         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1187         head = dequeue_forget(fiq, max_forgets, &count);
1188         spin_unlock(&fiq->waitq.lock);
1189
1190         arg.count = count;
1191         ih.len += count * sizeof(struct fuse_forget_one);
1192         err = fuse_copy_one(cs, &ih, sizeof(ih));
1193         if (!err)
1194                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1195
1196         while (head) {
1197                 struct fuse_forget_link *forget = head;
1198
1199                 if (!err) {
1200                         err = fuse_copy_one(cs, &forget->forget_one,
1201                                             sizeof(forget->forget_one));
1202                 }
1203                 head = forget->next;
1204                 kfree(forget);
1205         }
1206
1207         fuse_copy_finish(cs);
1208
1209         if (err)
1210                 return err;
1211
1212         return ih.len;
1213 }
1214
1215 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1216                             struct fuse_copy_state *cs,
1217                             size_t nbytes)
1218 __releases(fiq->waitq.lock)
1219 {
1220         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1221                 return fuse_read_single_forget(fiq, cs, nbytes);
1222         else
1223                 return fuse_read_batch_forget(fiq, cs, nbytes);
1224 }
1225
1226 /*
1227  * Read a single request into the userspace filesystem's buffer.  This
1228  * function waits until a request is available, then removes it from
1229  * the pending list and copies request data to userspace buffer.  If
1230  * no reply is needed (FORGET) or request has been aborted or there
1231  * was an error during the copying then it's finished by calling
1232  * request_end().  Otherwise add it to the processing list, and set
1233  * the 'sent' flag.
1234  */
1235 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1236                                 struct fuse_copy_state *cs, size_t nbytes)
1237 {
1238         ssize_t err;
1239         struct fuse_iqueue *fiq = &fc->iq;
1240         struct fuse_pqueue *fpq = &fc->pq;
1241         struct fuse_req *req;
1242         struct fuse_in *in;
1243         unsigned reqsize;
1244
1245  restart:
1246         spin_lock(&fiq->waitq.lock);
1247         err = -EAGAIN;
1248         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1249             !request_pending(fiq))
1250                 goto err_unlock;
1251
1252         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1253                                 !fiq->connected || request_pending(fiq));
1254         if (err)
1255                 goto err_unlock;
1256
1257         err = -ENODEV;
1258         if (!fiq->connected)
1259                 goto err_unlock;
1260
1261         if (!list_empty(&fiq->interrupts)) {
1262                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1263                                  intr_entry);
1264                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1265         }
1266
1267         if (forget_pending(fiq)) {
1268                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1269                         return fuse_read_forget(fc, fiq, cs, nbytes);
1270
1271                 if (fiq->forget_batch <= -8)
1272                         fiq->forget_batch = 16;
1273         }
1274
1275         req = list_entry(fiq->pending.next, struct fuse_req, list);
1276         clear_bit(FR_PENDING, &req->flags);
1277         list_del_init(&req->list);
1278         spin_unlock(&fiq->waitq.lock);
1279
1280         in = &req->in;
1281         reqsize = in->h.len;
1282         /* If request is too large, reply with an error and restart the read */
1283         if (nbytes < reqsize) {
1284                 req->out.h.error = -EIO;
1285                 /* SETXATTR is special, since it may contain too large data */
1286                 if (in->h.opcode == FUSE_SETXATTR)
1287                         req->out.h.error = -E2BIG;
1288                 request_end(fc, req);
1289                 goto restart;
1290         }
1291         spin_lock(&fpq->lock);
1292         list_add(&req->list, &fpq->io);
1293         spin_unlock(&fpq->lock);
1294         cs->req = req;
1295         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1296         if (!err)
1297                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1298                                      (struct fuse_arg *) in->args, 0);
1299         fuse_copy_finish(cs);
1300         spin_lock(&fpq->lock);
1301         clear_bit(FR_LOCKED, &req->flags);
1302         if (!fpq->connected) {
1303                 err = -ENODEV;
1304                 goto out_end;
1305         }
1306         if (err) {
1307                 req->out.h.error = -EIO;
1308                 goto out_end;
1309         }
1310         if (!test_bit(FR_ISREPLY, &req->flags)) {
1311                 err = reqsize;
1312                 goto out_end;
1313         }
1314         list_move_tail(&req->list, &fpq->processing);
1315         spin_unlock(&fpq->lock);
1316         set_bit(FR_SENT, &req->flags);
1317         /* matches barrier in request_wait_answer() */
1318         smp_mb__after_atomic();
1319         if (test_bit(FR_INTERRUPTED, &req->flags))
1320                 queue_interrupt(fiq, req);
1321
1322         return reqsize;
1323
1324 out_end:
1325         if (!test_bit(FR_PRIVATE, &req->flags))
1326                 list_del_init(&req->list);
1327         spin_unlock(&fpq->lock);
1328         request_end(fc, req);
1329         return err;
1330
1331  err_unlock:
1332         spin_unlock(&fiq->waitq.lock);
1333         return err;
1334 }
1335
1336 static int fuse_dev_open(struct inode *inode, struct file *file)
1337 {
1338         /*
1339          * The fuse device's file's private_data is used to hold
1340          * the fuse_conn(ection) when it is mounted, and is used to
1341          * keep track of whether the file has been mounted already.
1342          */
1343         file->private_data = NULL;
1344         return 0;
1345 }
1346
1347 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1348 {
1349         struct fuse_copy_state cs;
1350         struct file *file = iocb->ki_filp;
1351         struct fuse_conn *fc = fuse_get_conn(file);
1352         if (!fc)
1353                 return -EPERM;
1354
1355         if (!iter_is_iovec(to))
1356                 return -EINVAL;
1357
1358         fuse_copy_init(&cs, 1, to);
1359
1360         return fuse_dev_do_read(fc, file, &cs, iov_iter_count(to));
1361 }
1362
1363 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1364                                     struct pipe_inode_info *pipe,
1365                                     size_t len, unsigned int flags)
1366 {
1367         int ret;
1368         int page_nr = 0;
1369         int do_wakeup = 0;
1370         struct pipe_buffer *bufs;
1371         struct fuse_copy_state cs;
1372         struct fuse_conn *fc = fuse_get_conn(in);
1373         if (!fc)
1374                 return -EPERM;
1375
1376         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1377         if (!bufs)
1378                 return -ENOMEM;
1379
1380         fuse_copy_init(&cs, 1, NULL);
1381         cs.pipebufs = bufs;
1382         cs.pipe = pipe;
1383         ret = fuse_dev_do_read(fc, in, &cs, len);
1384         if (ret < 0)
1385                 goto out;
1386
1387         ret = 0;
1388         pipe_lock(pipe);
1389
1390         if (!pipe->readers) {
1391                 send_sig(SIGPIPE, current, 0);
1392                 if (!ret)
1393                         ret = -EPIPE;
1394                 goto out_unlock;
1395         }
1396
1397         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1398                 ret = -EIO;
1399                 goto out_unlock;
1400         }
1401
1402         while (page_nr < cs.nr_segs) {
1403                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1404                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1405
1406                 buf->page = bufs[page_nr].page;
1407                 buf->offset = bufs[page_nr].offset;
1408                 buf->len = bufs[page_nr].len;
1409                 /*
1410                  * Need to be careful about this.  Having buf->ops in module
1411                  * code can Oops if the buffer persists after module unload.
1412                  */
1413                 buf->ops = &nosteal_pipe_buf_ops;
1414
1415                 pipe->nrbufs++;
1416                 page_nr++;
1417                 ret += buf->len;
1418
1419                 if (pipe->files)
1420                         do_wakeup = 1;
1421         }
1422
1423 out_unlock:
1424         pipe_unlock(pipe);
1425
1426         if (do_wakeup) {
1427                 smp_mb();
1428                 if (waitqueue_active(&pipe->wait))
1429                         wake_up_interruptible(&pipe->wait);
1430                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1431         }
1432
1433 out:
1434         for (; page_nr < cs.nr_segs; page_nr++)
1435                 page_cache_release(bufs[page_nr].page);
1436
1437         kfree(bufs);
1438         return ret;
1439 }
1440
1441 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1442                             struct fuse_copy_state *cs)
1443 {
1444         struct fuse_notify_poll_wakeup_out outarg;
1445         int err = -EINVAL;
1446
1447         if (size != sizeof(outarg))
1448                 goto err;
1449
1450         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1451         if (err)
1452                 goto err;
1453
1454         fuse_copy_finish(cs);
1455         return fuse_notify_poll_wakeup(fc, &outarg);
1456
1457 err:
1458         fuse_copy_finish(cs);
1459         return err;
1460 }
1461
1462 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1463                                    struct fuse_copy_state *cs)
1464 {
1465         struct fuse_notify_inval_inode_out outarg;
1466         int err = -EINVAL;
1467
1468         if (size != sizeof(outarg))
1469                 goto err;
1470
1471         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1472         if (err)
1473                 goto err;
1474         fuse_copy_finish(cs);
1475
1476         down_read(&fc->killsb);
1477         err = -ENOENT;
1478         if (fc->sb) {
1479                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1480                                                outarg.off, outarg.len);
1481         }
1482         up_read(&fc->killsb);
1483         return err;
1484
1485 err:
1486         fuse_copy_finish(cs);
1487         return err;
1488 }
1489
1490 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1491                                    struct fuse_copy_state *cs)
1492 {
1493         struct fuse_notify_inval_entry_out outarg;
1494         int err = -ENOMEM;
1495         char *buf;
1496         struct qstr name;
1497
1498         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1499         if (!buf)
1500                 goto err;
1501
1502         err = -EINVAL;
1503         if (size < sizeof(outarg))
1504                 goto err;
1505
1506         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1507         if (err)
1508                 goto err;
1509
1510         err = -ENAMETOOLONG;
1511         if (outarg.namelen > FUSE_NAME_MAX)
1512                 goto err;
1513
1514         err = -EINVAL;
1515         if (size != sizeof(outarg) + outarg.namelen + 1)
1516                 goto err;
1517
1518         name.name = buf;
1519         name.len = outarg.namelen;
1520         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1521         if (err)
1522                 goto err;
1523         fuse_copy_finish(cs);
1524         buf[outarg.namelen] = 0;
1525         name.hash = full_name_hash(name.name, name.len);
1526
1527         down_read(&fc->killsb);
1528         err = -ENOENT;
1529         if (fc->sb)
1530                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1531         up_read(&fc->killsb);
1532         kfree(buf);
1533         return err;
1534
1535 err:
1536         kfree(buf);
1537         fuse_copy_finish(cs);
1538         return err;
1539 }
1540
1541 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1542                               struct fuse_copy_state *cs)
1543 {
1544         struct fuse_notify_delete_out outarg;
1545         int err = -ENOMEM;
1546         char *buf;
1547         struct qstr name;
1548
1549         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1550         if (!buf)
1551                 goto err;
1552
1553         err = -EINVAL;
1554         if (size < sizeof(outarg))
1555                 goto err;
1556
1557         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1558         if (err)
1559                 goto err;
1560
1561         err = -ENAMETOOLONG;
1562         if (outarg.namelen > FUSE_NAME_MAX)
1563                 goto err;
1564
1565         err = -EINVAL;
1566         if (size != sizeof(outarg) + outarg.namelen + 1)
1567                 goto err;
1568
1569         name.name = buf;
1570         name.len = outarg.namelen;
1571         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1572         if (err)
1573                 goto err;
1574         fuse_copy_finish(cs);
1575         buf[outarg.namelen] = 0;
1576         name.hash = full_name_hash(name.name, name.len);
1577
1578         down_read(&fc->killsb);
1579         err = -ENOENT;
1580         if (fc->sb)
1581                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1582                                                outarg.child, &name);
1583         up_read(&fc->killsb);
1584         kfree(buf);
1585         return err;
1586
1587 err:
1588         kfree(buf);
1589         fuse_copy_finish(cs);
1590         return err;
1591 }
1592
1593 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1594                              struct fuse_copy_state *cs)
1595 {
1596         struct fuse_notify_store_out outarg;
1597         struct inode *inode;
1598         struct address_space *mapping;
1599         u64 nodeid;
1600         int err;
1601         pgoff_t index;
1602         unsigned int offset;
1603         unsigned int num;
1604         loff_t file_size;
1605         loff_t end;
1606
1607         err = -EINVAL;
1608         if (size < sizeof(outarg))
1609                 goto out_finish;
1610
1611         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1612         if (err)
1613                 goto out_finish;
1614
1615         err = -EINVAL;
1616         if (size - sizeof(outarg) != outarg.size)
1617                 goto out_finish;
1618
1619         nodeid = outarg.nodeid;
1620
1621         down_read(&fc->killsb);
1622
1623         err = -ENOENT;
1624         if (!fc->sb)
1625                 goto out_up_killsb;
1626
1627         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1628         if (!inode)
1629                 goto out_up_killsb;
1630
1631         mapping = inode->i_mapping;
1632         index = outarg.offset >> PAGE_CACHE_SHIFT;
1633         offset = outarg.offset & ~PAGE_CACHE_MASK;
1634         file_size = i_size_read(inode);
1635         end = outarg.offset + outarg.size;
1636         if (end > file_size) {
1637                 file_size = end;
1638                 fuse_write_update_size(inode, file_size);
1639         }
1640
1641         num = outarg.size;
1642         while (num) {
1643                 struct page *page;
1644                 unsigned int this_num;
1645
1646                 err = -ENOMEM;
1647                 page = find_or_create_page(mapping, index,
1648                                            mapping_gfp_mask(mapping));
1649                 if (!page)
1650                         goto out_iput;
1651
1652                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1653                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1654                 if (!err && offset == 0 &&
1655                     (this_num == PAGE_CACHE_SIZE || file_size == end))
1656                         SetPageUptodate(page);
1657                 unlock_page(page);
1658                 page_cache_release(page);
1659
1660                 if (err)
1661                         goto out_iput;
1662
1663                 num -= this_num;
1664                 offset = 0;
1665                 index++;
1666         }
1667
1668         err = 0;
1669
1670 out_iput:
1671         iput(inode);
1672 out_up_killsb:
1673         up_read(&fc->killsb);
1674 out_finish:
1675         fuse_copy_finish(cs);
1676         return err;
1677 }
1678
1679 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1680 {
1681         release_pages(req->pages, req->num_pages, false);
1682 }
1683
1684 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1685                          struct fuse_notify_retrieve_out *outarg)
1686 {
1687         int err;
1688         struct address_space *mapping = inode->i_mapping;
1689         struct fuse_req *req;
1690         pgoff_t index;
1691         loff_t file_size;
1692         unsigned int num;
1693         unsigned int offset;
1694         size_t total_len = 0;
1695         int num_pages;
1696
1697         offset = outarg->offset & ~PAGE_CACHE_MASK;
1698         file_size = i_size_read(inode);
1699
1700         num = outarg->size;
1701         if (outarg->offset > file_size)
1702                 num = 0;
1703         else if (outarg->offset + num > file_size)
1704                 num = file_size - outarg->offset;
1705
1706         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1707         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1708
1709         req = fuse_get_req(fc, num_pages);
1710         if (IS_ERR(req))
1711                 return PTR_ERR(req);
1712
1713         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1714         req->in.h.nodeid = outarg->nodeid;
1715         req->in.numargs = 2;
1716         req->in.argpages = 1;
1717         req->page_descs[0].offset = offset;
1718         req->end = fuse_retrieve_end;
1719
1720         index = outarg->offset >> PAGE_CACHE_SHIFT;
1721
1722         while (num && req->num_pages < num_pages) {
1723                 struct page *page;
1724                 unsigned int this_num;
1725
1726                 page = find_get_page(mapping, index);
1727                 if (!page)
1728                         break;
1729
1730                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1731                 req->pages[req->num_pages] = page;
1732                 req->page_descs[req->num_pages].length = this_num;
1733                 req->num_pages++;
1734
1735                 offset = 0;
1736                 num -= this_num;
1737                 total_len += this_num;
1738                 index++;
1739         }
1740         req->misc.retrieve_in.offset = outarg->offset;
1741         req->misc.retrieve_in.size = total_len;
1742         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1743         req->in.args[0].value = &req->misc.retrieve_in;
1744         req->in.args[1].size = total_len;
1745
1746         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1747         if (err)
1748                 fuse_retrieve_end(fc, req);
1749
1750         return err;
1751 }
1752
1753 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1754                                 struct fuse_copy_state *cs)
1755 {
1756         struct fuse_notify_retrieve_out outarg;
1757         struct inode *inode;
1758         int err;
1759
1760         err = -EINVAL;
1761         if (size != sizeof(outarg))
1762                 goto copy_finish;
1763
1764         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1765         if (err)
1766                 goto copy_finish;
1767
1768         fuse_copy_finish(cs);
1769
1770         down_read(&fc->killsb);
1771         err = -ENOENT;
1772         if (fc->sb) {
1773                 u64 nodeid = outarg.nodeid;
1774
1775                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1776                 if (inode) {
1777                         err = fuse_retrieve(fc, inode, &outarg);
1778                         iput(inode);
1779                 }
1780         }
1781         up_read(&fc->killsb);
1782
1783         return err;
1784
1785 copy_finish:
1786         fuse_copy_finish(cs);
1787         return err;
1788 }
1789
1790 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1791                        unsigned int size, struct fuse_copy_state *cs)
1792 {
1793         /* Don't try to move pages (yet) */
1794         cs->move_pages = 0;
1795
1796         switch (code) {
1797         case FUSE_NOTIFY_POLL:
1798                 return fuse_notify_poll(fc, size, cs);
1799
1800         case FUSE_NOTIFY_INVAL_INODE:
1801                 return fuse_notify_inval_inode(fc, size, cs);
1802
1803         case FUSE_NOTIFY_INVAL_ENTRY:
1804                 return fuse_notify_inval_entry(fc, size, cs);
1805
1806         case FUSE_NOTIFY_STORE:
1807                 return fuse_notify_store(fc, size, cs);
1808
1809         case FUSE_NOTIFY_RETRIEVE:
1810                 return fuse_notify_retrieve(fc, size, cs);
1811
1812         case FUSE_NOTIFY_DELETE:
1813                 return fuse_notify_delete(fc, size, cs);
1814
1815         default:
1816                 fuse_copy_finish(cs);
1817                 return -EINVAL;
1818         }
1819 }
1820
1821 /* Look up request on processing list by unique ID */
1822 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1823 {
1824         struct fuse_req *req;
1825
1826         list_for_each_entry(req, &fpq->processing, list) {
1827                 if (req->in.h.unique == unique || req->intr_unique == unique)
1828                         return req;
1829         }
1830         return NULL;
1831 }
1832
1833 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1834                          unsigned nbytes)
1835 {
1836         unsigned reqsize = sizeof(struct fuse_out_header);
1837
1838         if (out->h.error)
1839                 return nbytes != reqsize ? -EINVAL : 0;
1840
1841         reqsize += len_args(out->numargs, out->args);
1842
1843         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1844                 return -EINVAL;
1845         else if (reqsize > nbytes) {
1846                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1847                 unsigned diffsize = reqsize - nbytes;
1848                 if (diffsize > lastarg->size)
1849                         return -EINVAL;
1850                 lastarg->size -= diffsize;
1851         }
1852         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1853                               out->page_zeroing);
1854 }
1855
1856 /*
1857  * Write a single reply to a request.  First the header is copied from
1858  * the write buffer.  The request is then searched on the processing
1859  * list by the unique ID found in the header.  If found, then remove
1860  * it from the list and copy the rest of the buffer to the request.
1861  * The request is finished by calling request_end()
1862  */
1863 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1864                                  struct fuse_copy_state *cs, size_t nbytes)
1865 {
1866         int err;
1867         struct fuse_pqueue *fpq = &fc->pq;
1868         struct fuse_req *req;
1869         struct fuse_out_header oh;
1870
1871         if (nbytes < sizeof(struct fuse_out_header))
1872                 return -EINVAL;
1873
1874         err = fuse_copy_one(cs, &oh, sizeof(oh));
1875         if (err)
1876                 goto err_finish;
1877
1878         err = -EINVAL;
1879         if (oh.len != nbytes)
1880                 goto err_finish;
1881
1882         /*
1883          * Zero oh.unique indicates unsolicited notification message
1884          * and error contains notification code.
1885          */
1886         if (!oh.unique) {
1887                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1888                 return err ? err : nbytes;
1889         }
1890
1891         err = -EINVAL;
1892         if (oh.error <= -1000 || oh.error > 0)
1893                 goto err_finish;
1894
1895         spin_lock(&fpq->lock);
1896         err = -ENOENT;
1897         if (!fpq->connected)
1898                 goto err_unlock_pq;
1899
1900         req = request_find(fpq, oh.unique);
1901         if (!req)
1902                 goto err_unlock_pq;
1903
1904         /* Is it an interrupt reply? */
1905         if (req->intr_unique == oh.unique) {
1906                 spin_unlock(&fpq->lock);
1907
1908                 err = -EINVAL;
1909                 if (nbytes != sizeof(struct fuse_out_header))
1910                         goto err_finish;
1911
1912                 if (oh.error == -ENOSYS)
1913                         fc->no_interrupt = 1;
1914                 else if (oh.error == -EAGAIN)
1915                         queue_interrupt(&fc->iq, req);
1916
1917                 fuse_copy_finish(cs);
1918                 return nbytes;
1919         }
1920
1921         clear_bit(FR_SENT, &req->flags);
1922         list_move(&req->list, &fpq->io);
1923         req->out.h = oh;
1924         set_bit(FR_LOCKED, &req->flags);
1925         spin_unlock(&fpq->lock);
1926         cs->req = req;
1927         if (!req->out.page_replace)
1928                 cs->move_pages = 0;
1929
1930         err = copy_out_args(cs, &req->out, nbytes);
1931         fuse_copy_finish(cs);
1932
1933         spin_lock(&fpq->lock);
1934         clear_bit(FR_LOCKED, &req->flags);
1935         if (!fpq->connected)
1936                 err = -ENOENT;
1937         else if (err)
1938                 req->out.h.error = -EIO;
1939         if (!test_bit(FR_PRIVATE, &req->flags))
1940                 list_del_init(&req->list);
1941         spin_unlock(&fpq->lock);
1942
1943         request_end(fc, req);
1944
1945         return err ? err : nbytes;
1946
1947  err_unlock_pq:
1948         spin_unlock(&fpq->lock);
1949  err_finish:
1950         fuse_copy_finish(cs);
1951         return err;
1952 }
1953
1954 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1955 {
1956         struct fuse_copy_state cs;
1957         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1958         if (!fc)
1959                 return -EPERM;
1960
1961         if (!iter_is_iovec(from))
1962                 return -EINVAL;
1963
1964         fuse_copy_init(&cs, 0, from);
1965
1966         return fuse_dev_do_write(fc, &cs, iov_iter_count(from));
1967 }
1968
1969 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1970                                      struct file *out, loff_t *ppos,
1971                                      size_t len, unsigned int flags)
1972 {
1973         unsigned nbuf;
1974         unsigned idx;
1975         struct pipe_buffer *bufs;
1976         struct fuse_copy_state cs;
1977         struct fuse_conn *fc;
1978         size_t rem;
1979         ssize_t ret;
1980
1981         fc = fuse_get_conn(out);
1982         if (!fc)
1983                 return -EPERM;
1984
1985         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1986         if (!bufs)
1987                 return -ENOMEM;
1988
1989         pipe_lock(pipe);
1990         nbuf = 0;
1991         rem = 0;
1992         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1993                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1994
1995         ret = -EINVAL;
1996         if (rem < len) {
1997                 pipe_unlock(pipe);
1998                 goto out;
1999         }
2000
2001         rem = len;
2002         while (rem) {
2003                 struct pipe_buffer *ibuf;
2004                 struct pipe_buffer *obuf;
2005
2006                 BUG_ON(nbuf >= pipe->buffers);
2007                 BUG_ON(!pipe->nrbufs);
2008                 ibuf = &pipe->bufs[pipe->curbuf];
2009                 obuf = &bufs[nbuf];
2010
2011                 if (rem >= ibuf->len) {
2012                         *obuf = *ibuf;
2013                         ibuf->ops = NULL;
2014                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2015                         pipe->nrbufs--;
2016                 } else {
2017                         ibuf->ops->get(pipe, ibuf);
2018                         *obuf = *ibuf;
2019                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2020                         obuf->len = rem;
2021                         ibuf->offset += obuf->len;
2022                         ibuf->len -= obuf->len;
2023                 }
2024                 nbuf++;
2025                 rem -= obuf->len;
2026         }
2027         pipe_unlock(pipe);
2028
2029         fuse_copy_init(&cs, 0, NULL);
2030         cs.pipebufs = bufs;
2031         cs.nr_segs = nbuf;
2032         cs.pipe = pipe;
2033
2034         if (flags & SPLICE_F_MOVE)
2035                 cs.move_pages = 1;
2036
2037         ret = fuse_dev_do_write(fc, &cs, len);
2038
2039         for (idx = 0; idx < nbuf; idx++) {
2040                 struct pipe_buffer *buf = &bufs[idx];
2041                 buf->ops->release(pipe, buf);
2042         }
2043 out:
2044         kfree(bufs);
2045         return ret;
2046 }
2047
2048 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2049 {
2050         unsigned mask = POLLOUT | POLLWRNORM;
2051         struct fuse_iqueue *fiq;
2052         struct fuse_conn *fc = fuse_get_conn(file);
2053         if (!fc)
2054                 return POLLERR;
2055
2056         fiq = &fc->iq;
2057         poll_wait(file, &fiq->waitq, wait);
2058
2059         spin_lock(&fiq->waitq.lock);
2060         if (!fiq->connected)
2061                 mask = POLLERR;
2062         else if (request_pending(fiq))
2063                 mask |= POLLIN | POLLRDNORM;
2064         spin_unlock(&fiq->waitq.lock);
2065
2066         return mask;
2067 }
2068
2069 /*
2070  * Abort all requests on the given list (pending or processing)
2071  *
2072  * This function releases and reacquires fc->lock
2073  */
2074 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2075 {
2076         while (!list_empty(head)) {
2077                 struct fuse_req *req;
2078                 req = list_entry(head->next, struct fuse_req, list);
2079                 req->out.h.error = -ECONNABORTED;
2080                 clear_bit(FR_PENDING, &req->flags);
2081                 clear_bit(FR_SENT, &req->flags);
2082                 list_del_init(&req->list);
2083                 request_end(fc, req);
2084         }
2085 }
2086
2087 static void end_polls(struct fuse_conn *fc)
2088 {
2089         struct rb_node *p;
2090
2091         p = rb_first(&fc->polled_files);
2092
2093         while (p) {
2094                 struct fuse_file *ff;
2095                 ff = rb_entry(p, struct fuse_file, polled_node);
2096                 wake_up_interruptible_all(&ff->poll_wait);
2097
2098                 p = rb_next(p);
2099         }
2100 }
2101
2102 /*
2103  * Abort all requests.
2104  *
2105  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2106  * filesystem.
2107  *
2108  * The same effect is usually achievable through killing the filesystem daemon
2109  * and all users of the filesystem.  The exception is the combination of an
2110  * asynchronous request and the tricky deadlock (see
2111  * Documentation/filesystems/fuse.txt).
2112  *
2113  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2114  * requests, they should be finished off immediately.  Locked requests will be
2115  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2116  * requests.  It is possible that some request will finish before we can.  This
2117  * is OK, the request will in that case be removed from the list before we touch
2118  * it.
2119  */
2120 void fuse_abort_conn(struct fuse_conn *fc)
2121 {
2122         struct fuse_iqueue *fiq = &fc->iq;
2123         struct fuse_pqueue *fpq = &fc->pq;
2124
2125         spin_lock(&fc->lock);
2126         if (fc->connected) {
2127                 struct fuse_req *req, *next;
2128                 LIST_HEAD(to_end1);
2129                 LIST_HEAD(to_end2);
2130
2131                 fc->connected = 0;
2132                 fc->blocked = 0;
2133                 fuse_set_initialized(fc);
2134                 spin_lock(&fpq->lock);
2135                 fpq->connected = 0;
2136                 list_for_each_entry_safe(req, next, &fpq->io, list) {
2137                         req->out.h.error = -ECONNABORTED;
2138                         spin_lock(&req->waitq.lock);
2139                         set_bit(FR_ABORTED, &req->flags);
2140                         if (!test_bit(FR_LOCKED, &req->flags)) {
2141                                 set_bit(FR_PRIVATE, &req->flags);
2142                                 list_move(&req->list, &to_end1);
2143                         }
2144                         spin_unlock(&req->waitq.lock);
2145                 }
2146                 list_splice_init(&fpq->processing, &to_end2);
2147                 spin_unlock(&fpq->lock);
2148                 fc->max_background = UINT_MAX;
2149                 flush_bg_queue(fc);
2150
2151                 spin_lock(&fiq->waitq.lock);
2152                 fiq->connected = 0;
2153                 list_splice_init(&fiq->pending, &to_end2);
2154                 while (forget_pending(fiq))
2155                         kfree(dequeue_forget(fiq, 1, NULL));
2156                 wake_up_all_locked(&fiq->waitq);
2157                 spin_unlock(&fiq->waitq.lock);
2158                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2159                 end_polls(fc);
2160                 wake_up_all(&fc->blocked_waitq);
2161                 spin_unlock(&fc->lock);
2162
2163                 while (!list_empty(&to_end1)) {
2164                         req = list_first_entry(&to_end1, struct fuse_req, list);
2165                         __fuse_get_request(req);
2166                         list_del_init(&req->list);
2167                         request_end(fc, req);
2168                 }
2169                 end_requests(fc, &to_end2);
2170         } else {
2171                 spin_unlock(&fc->lock);
2172         }
2173 }
2174 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2175
2176 int fuse_dev_release(struct inode *inode, struct file *file)
2177 {
2178         struct fuse_conn *fc = fuse_get_conn(file);
2179         if (fc) {
2180                 WARN_ON(!list_empty(&fc->pq.io));
2181                 WARN_ON(fc->iq.fasync != NULL);
2182                 fuse_abort_conn(fc);
2183                 fuse_conn_put(fc);
2184         }
2185
2186         return 0;
2187 }
2188 EXPORT_SYMBOL_GPL(fuse_dev_release);
2189
2190 static int fuse_dev_fasync(int fd, struct file *file, int on)
2191 {
2192         struct fuse_conn *fc = fuse_get_conn(file);
2193         if (!fc)
2194                 return -EPERM;
2195
2196         /* No locking - fasync_helper does its own locking */
2197         return fasync_helper(fd, file, on, &fc->iq.fasync);
2198 }
2199
2200 const struct file_operations fuse_dev_operations = {
2201         .owner          = THIS_MODULE,
2202         .open           = fuse_dev_open,
2203         .llseek         = no_llseek,
2204         .read_iter      = fuse_dev_read,
2205         .splice_read    = fuse_dev_splice_read,
2206         .write_iter     = fuse_dev_write,
2207         .splice_write   = fuse_dev_splice_write,
2208         .poll           = fuse_dev_poll,
2209         .release        = fuse_dev_release,
2210         .fasync         = fuse_dev_fasync,
2211 };
2212 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2213
2214 static struct miscdevice fuse_miscdevice = {
2215         .minor = FUSE_MINOR,
2216         .name  = "fuse",
2217         .fops = &fuse_dev_operations,
2218 };
2219
2220 int __init fuse_dev_init(void)
2221 {
2222         int err = -ENOMEM;
2223         fuse_req_cachep = kmem_cache_create("fuse_request",
2224                                             sizeof(struct fuse_req),
2225                                             0, 0, NULL);
2226         if (!fuse_req_cachep)
2227                 goto out;
2228
2229         err = misc_register(&fuse_miscdevice);
2230         if (err)
2231                 goto out_cache_clean;
2232
2233         return 0;
2234
2235  out_cache_clean:
2236         kmem_cache_destroy(fuse_req_cachep);
2237  out:
2238         return err;
2239 }
2240
2241 void fuse_dev_cleanup(void)
2242 {
2243         misc_deregister(&fuse_miscdevice);
2244         kmem_cache_destroy(fuse_req_cachep);
2245 }