fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/pipe_fs_i.h>
  20 #include <linux/swap.h>
  21 #include <linux/splice.h>
  22
  23 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  24 MODULE_ALIAS("devname:fuse");
  25
  26 static struct kmem_cache *fuse_req_cachep;
  27
  28 static struct fuse_conn *fuse_get_conn(struct file *file)
  29 {
  30         /*
  31          * Lockless access is OK, because file->private data is set
  32          * once during mount and is valid until the file is released.
  33          */
  34         return file->private_data;
  35 }
  36
  37 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  38                               struct fuse_page_desc *page_descs,
  39                               unsigned npages)
  40 {
  41         memset(req, 0, sizeof(*req));
  42         memset(pages, 0, sizeof(*pages) * npages);
  43         memset(page_descs, 0, sizeof(*page_descs) * npages);
  44         INIT_LIST_HEAD(&req->list);
  45         INIT_LIST_HEAD(&req->intr_entry);
  46         init_waitqueue_head(&req->waitq);
  47         atomic_set(&req->count, 1);
  48         req->pages = pages;
  49         req->page_descs = page_descs;
  50         req->max_pages = npages;
  51         __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  * Called with fc->lock, unlocks it
 381  */
 382 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 383 __releases(fc->lock)
 384 {
 385         struct fuse_iqueue *fiq = &fc->iq;
 386         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 387         req->end = NULL;
 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         smp_wmb();
 394         set_bit(FR_FINISHED, &req->flags);
 395         if (test_bit(FR_BACKGROUND, &req->flags)) {
 396                 clear_bit(FR_BACKGROUND, &req->flags);
 397                 if (fc->num_background == fc->max_background)
 398                         fc->blocked = 0;
 399
 400                 /* Wake up next waiter, if any */
 401                 if (!fc->blocked && waitqueue_active(&fc->blocked_waitq))
 402                         wake_up(&fc->blocked_waitq);
 403
 404                 if (fc->num_background == fc->congestion_threshold &&
 405                     fc->connected && fc->bdi_initialized) {
 406                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 407                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 408                 }
 409                 fc->num_background--;
 410                 fc->active_background--;
 411                 flush_bg_queue(fc);
 412         }
 413         spin_unlock(&fc->lock);
 414         wake_up(&req->waitq);
 415         if (end)
 416                 end(fc, req);
 417         fuse_put_request(fc, req);
 418 }
 419
 420 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 421 {
 422         spin_lock(&fiq->waitq.lock);
 423         if (list_empty(&req->intr_entry)) {
 424                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 425                 wake_up_locked(&fiq->waitq);
 426         }
 427         spin_unlock(&fiq->waitq.lock);
 428         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 429 }
 430
 431 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 432 {
 433         struct fuse_iqueue *fiq = &fc->iq;
 434         int err;
 435
 436         if (!fc->no_interrupt) {
 437                 /* Any signal may interrupt this */
 438                 err = wait_event_interruptible(req->waitq,
 439                                         test_bit(FR_FINISHED, &req->flags));
 440                 if (!err)
 441                         return;
 442
 443                 set_bit(FR_INTERRUPTED, &req->flags);
 444                 /* matches barrier in fuse_dev_do_read() */
 445                 smp_mb__after_atomic();
 446                 if (test_bit(FR_SENT, &req->flags))
 447                         queue_interrupt(fiq, req);
 448         }
 449
 450         if (!test_bit(FR_FORCE, &req->flags)) {
 451                 sigset_t oldset;
 452
 453                 /* Only fatal signals may interrupt this */
 454                 block_sigs(&oldset);
 455                 err = wait_event_interruptible(req->waitq,
 456                                         test_bit(FR_FINISHED, &req->flags));
 457                 restore_sigs(&oldset);
 458
 459                 if (!err)
 460                         return;
 461
 462                 spin_lock(&fiq->waitq.lock);
 463                 /* Request is not yet in userspace, bail out */
 464                 if (test_bit(FR_PENDING, &req->flags)) {
 465                         list_del(&req->list);
 466                         spin_unlock(&fiq->waitq.lock);
 467                         __fuse_put_request(req);
 468                         req->out.h.error = -EINTR;
 469                         return;
 470                 }
 471                 spin_unlock(&fiq->waitq.lock);
 472         }
 473
 474         /*
 475          * Either request is already in userspace, or it was forced.
 476          * Wait it out.
 477          */
 478         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 479 }
 480
 481 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 482 {
 483         struct fuse_iqueue *fiq = &fc->iq;
 484
 485         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 486         spin_lock(&fiq->waitq.lock);
 487         if (!fiq->connected) {
 488                 spin_unlock(&fiq->waitq.lock);
 489                 req->out.h.error = -ENOTCONN;
 490         } else {
 491                 req->in.h.unique = fuse_get_unique(fiq);
 492                 queue_request(fiq, req);
 493                 /* acquire extra reference, since request is still needed
 494                    after request_end() */
 495                 __fuse_get_request(req);
 496                 spin_unlock(&fiq->waitq.lock);
 497
 498                 request_wait_answer(fc, req);
 499                 /* Pairs with smp_wmb() in request_end() */
 500                 smp_rmb();
 501         }
 502 }
 503
 504 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 505 {
 506         __set_bit(FR_ISREPLY, &req->flags);
 507         if (!test_bit(FR_WAITING, &req->flags)) {
 508                 __set_bit(FR_WAITING, &req->flags);
 509                 atomic_inc(&fc->num_waiting);
 510         }
 511         __fuse_request_send(fc, req);
 512 }
 513 EXPORT_SYMBOL_GPL(fuse_request_send);
 514
 515 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 516 {
 517         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 518                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 519
 520         if (fc->minor < 9) {
 521                 switch (args->in.h.opcode) {
 522                 case FUSE_LOOKUP:
 523                 case FUSE_CREATE:
 524                 case FUSE_MKNOD:
 525                 case FUSE_MKDIR:
 526                 case FUSE_SYMLINK:
 527                 case FUSE_LINK:
 528                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 529                         break;
 530                 case FUSE_GETATTR:
 531                 case FUSE_SETATTR:
 532                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 533                         break;
 534                 }
 535         }
 536         if (fc->minor < 12) {
 537                 switch (args->in.h.opcode) {
 538                 case FUSE_CREATE:
 539                         args->in.args[0].size = sizeof(struct fuse_open_in);
 540                         break;
 541                 case FUSE_MKNOD:
 542                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 543                         break;
 544                 }
 545         }
 546 }
 547
 548 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 549 {
 550         struct fuse_req *req;
 551         ssize_t ret;
 552
 553         req = fuse_get_req(fc, 0);
 554         if (IS_ERR(req))
 555                 return PTR_ERR(req);
 556
 557         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 558         fuse_adjust_compat(fc, args);
 559
 560         req->in.h.opcode = args->in.h.opcode;
 561         req->in.h.nodeid = args->in.h.nodeid;
 562         req->in.numargs = args->in.numargs;
 563         memcpy(req->in.args, args->in.args,
 564                args->in.numargs * sizeof(struct fuse_in_arg));
 565         req->out.argvar = args->out.argvar;
 566         req->out.numargs = args->out.numargs;
 567         memcpy(req->out.args, args->out.args,
 568                args->out.numargs * sizeof(struct fuse_arg));
 569         fuse_request_send(fc, req);
 570         ret = req->out.h.error;
 571         if (!ret && args->out.argvar) {
 572                 BUG_ON(args->out.numargs != 1);
 573                 ret = req->out.args[0].size;
 574         }
 575         fuse_put_request(fc, req);
 576
 577         return ret;
 578 }
 579
 580 /*
 581  * Called under fc->lock
 582  *
 583  * fc->connected must have been checked previously
 584  */
 585 void fuse_request_send_background_locked(struct fuse_conn *fc,
 586                                          struct fuse_req *req)
 587 {
 588         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 589         if (!test_bit(FR_WAITING, &req->flags)) {
 590                 __set_bit(FR_WAITING, &req->flags);
 591                 atomic_inc(&fc->num_waiting);
 592         }
 593         __set_bit(FR_ISREPLY, &req->flags);
 594         fc->num_background++;
 595         if (fc->num_background == fc->max_background)
 596                 fc->blocked = 1;
 597         if (fc->num_background == fc->congestion_threshold &&
 598             fc->bdi_initialized) {
 599                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 600                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 601         }
 602         list_add_tail(&req->list, &fc->bg_queue);
 603         flush_bg_queue(fc);
 604 }
 605
 606 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 607 {
 608         BUG_ON(!req->end);
 609         spin_lock(&fc->lock);
 610         if (fc->connected) {
 611                 fuse_request_send_background_locked(fc, req);
 612                 spin_unlock(&fc->lock);
 613         } else {
 614                 spin_unlock(&fc->lock);
 615                 req->out.h.error = -ENOTCONN;
 616                 req->end(fc, req);
 617                 fuse_put_request(fc, req);
 618         }
 619 }
 620 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 621
 622 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 623                                           struct fuse_req *req, u64 unique)
 624 {
 625         int err = -ENODEV;
 626         struct fuse_iqueue *fiq = &fc->iq;
 627
 628         __clear_bit(FR_ISREPLY, &req->flags);
 629         req->in.h.unique = unique;
 630         spin_lock(&fiq->waitq.lock);
 631         if (fiq->connected) {
 632                 queue_request(fiq, req);
 633                 err = 0;
 634         }
 635         spin_unlock(&fiq->waitq.lock);
 636
 637         return err;
 638 }
 639
 640 void fuse_force_forget(struct file *file, u64 nodeid)
 641 {
 642         struct inode *inode = file_inode(file);
 643         struct fuse_conn *fc = get_fuse_conn(inode);
 644         struct fuse_req *req;
 645         struct fuse_forget_in inarg;
 646
 647         memset(&inarg, 0, sizeof(inarg));
 648         inarg.nlookup = 1;
 649         req = fuse_get_req_nofail_nopages(fc, file);
 650         req->in.h.opcode = FUSE_FORGET;
 651         req->in.h.nodeid = nodeid;
 652         req->in.numargs = 1;
 653         req->in.args[0].size = sizeof(inarg);
 654         req->in.args[0].value = &inarg;
 655         __clear_bit(FR_ISREPLY, &req->flags);
 656         __fuse_request_send(fc, req);
 657         /* ignore errors */
 658         fuse_put_request(fc, req);
 659 }
 660
 661 /*
 662  * Lock the request.  Up to the next unlock_request() there mustn't be
 663  * anything that could cause a page-fault.  If the request was already
 664  * aborted bail out.
 665  */
 666 static int lock_request(struct fuse_req *req)
 667 {
 668         int err = 0;
 669         if (req) {
 670                 spin_lock(&req->waitq.lock);
 671                 if (test_bit(FR_ABORTED, &req->flags))
 672                         err = -ENOENT;
 673                 else
 674                         set_bit(FR_LOCKED, &req->flags);
 675                 spin_unlock(&req->waitq.lock);
 676         }
 677         return err;
 678 }
 679
 680 /*
 681  * Unlock request.  If it was aborted while locked, caller is responsible
 682  * for unlocking and ending the request.
 683  */
 684 static int unlock_request(struct fuse_req *req)
 685 {
 686         int err = 0;
 687         if (req) {
 688                 spin_lock(&req->waitq.lock);
 689                 if (test_bit(FR_ABORTED, &req->flags))
 690                         err = -ENOENT;
 691                 else
 692                         clear_bit(FR_LOCKED, &req->flags);
 693                 spin_unlock(&req->waitq.lock);
 694         }
 695         return err;
 696 }
 697
 698 struct fuse_copy_state {
 699         int write;
 700         struct fuse_req *req;
 701         struct iov_iter *iter;
 702         struct pipe_buffer *pipebufs;
 703         struct pipe_buffer *currbuf;
 704         struct pipe_inode_info *pipe;
 705         unsigned long nr_segs;
 706         struct page *pg;
 707         unsigned len;
 708         unsigned offset;
 709         unsigned move_pages:1;
 710 };
 711
 712 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 713                            struct iov_iter *iter)
 714 {
 715         memset(cs, 0, sizeof(*cs));
 716         cs->write = write;
 717         cs->iter = iter;
 718 }
 719
 720 /* Unmap and put previous page of userspace buffer */
 721 static void fuse_copy_finish(struct fuse_copy_state *cs)
 722 {
 723         if (cs->currbuf) {
 724                 struct pipe_buffer *buf = cs->currbuf;
 725
 726                 if (cs->write)
 727                         buf->len = PAGE_SIZE - cs->len;
 728                 cs->currbuf = NULL;
 729         } else if (cs->pg) {
 730                 if (cs->write) {
 731                         flush_dcache_page(cs->pg);
 732                         set_page_dirty_lock(cs->pg);
 733                 }
 734                 put_page(cs->pg);
 735         }
 736         cs->pg = NULL;
 737 }
 738
 739 /*
 740  * Get another pagefull of userspace buffer, and map it to kernel
 741  * address space, and lock request
 742  */
 743 static int fuse_copy_fill(struct fuse_copy_state *cs)
 744 {
 745         struct page *page;
 746         int err;
 747
 748         err = unlock_request(cs->req);
 749         if (err)
 750                 return err;
 751
 752         fuse_copy_finish(cs);
 753         if (cs->pipebufs) {
 754                 struct pipe_buffer *buf = cs->pipebufs;
 755
 756                 if (!cs->write) {
 757                         err = buf->ops->confirm(cs->pipe, buf);
 758                         if (err)
 759                                 return err;
 760
 761                         BUG_ON(!cs->nr_segs);
 762                         cs->currbuf = buf;
 763                         cs->pg = buf->page;
 764                         cs->offset = buf->offset;
 765                         cs->len = buf->len;
 766                         cs->pipebufs++;
 767                         cs->nr_segs--;
 768                 } else {
 769                         if (cs->nr_segs == cs->pipe->buffers)
 770                                 return -EIO;
 771
 772                         page = alloc_page(GFP_HIGHUSER);
 773                         if (!page)
 774                                 return -ENOMEM;
 775
 776                         buf->page = page;
 777                         buf->offset = 0;
 778                         buf->len = 0;
 779
 780                         cs->currbuf = buf;
 781                         cs->pg = page;
 782                         cs->offset = 0;
 783                         cs->len = PAGE_SIZE;
 784                         cs->pipebufs++;
 785                         cs->nr_segs++;
 786                 }
 787         } else {
 788                 size_t off;
 789                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 790                 if (err < 0)
 791                         return err;
 792                 BUG_ON(!err);
 793                 cs->len = err;
 794                 cs->offset = off;
 795                 cs->pg = page;
 796                 cs->offset = off;
 797                 iov_iter_advance(cs->iter, err);
 798         }
 799
 800         return lock_request(cs->req);
 801 }
 802
 803 /* Do as much copy to/from userspace buffer as we can */
 804 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 805 {
 806         unsigned ncpy = min(*size, cs->len);
 807         if (val) {
 808                 void *pgaddr = kmap_atomic(cs->pg);
 809                 void *buf = pgaddr + cs->offset;
 810
 811                 if (cs->write)
 812                         memcpy(buf, *val, ncpy);
 813                 else
 814                         memcpy(*val, buf, ncpy);
 815
 816                 kunmap_atomic(pgaddr);
 817                 *val += ncpy;
 818         }
 819         *size -= ncpy;
 820         cs->len -= ncpy;
 821         cs->offset += ncpy;
 822         return ncpy;
 823 }
 824
 825 static int fuse_check_page(struct page *page)
 826 {
 827         if (page_mapcount(page) ||
 828             page->mapping != NULL ||
 829             page_count(page) != 1 ||
 830             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 831              ~(1 << PG_locked |
 832                1 << PG_referenced |
 833                1 << PG_uptodate |
 834                1 << PG_lru |
 835                1 << PG_active |
 836                1 << PG_reclaim))) {
 837                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 838                 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);
 839                 return 1;
 840         }
 841         return 0;
 842 }
 843
 844 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 845 {
 846         int err;
 847         struct page *oldpage = *pagep;
 848         struct page *newpage;
 849         struct pipe_buffer *buf = cs->pipebufs;
 850
 851         err = unlock_request(cs->req);
 852         if (err)
 853                 return err;
 854
 855         fuse_copy_finish(cs);
 856
 857         err = buf->ops->confirm(cs->pipe, buf);
 858         if (err)
 859                 return err;
 860
 861         BUG_ON(!cs->nr_segs);
 862         cs->currbuf = buf;
 863         cs->len = buf->len;
 864         cs->pipebufs++;
 865         cs->nr_segs--;
 866
 867         if (cs->len != PAGE_SIZE)
 868                 goto out_fallback;
 869
 870         if (buf->ops->steal(cs->pipe, buf) != 0)
 871                 goto out_fallback;
 872
 873         newpage = buf->page;
 874
 875         if (!PageUptodate(newpage))
 876                 SetPageUptodate(newpage);
 877
 878         ClearPageMappedToDisk(newpage);
 879
 880         if (fuse_check_page(newpage) != 0)
 881                 goto out_fallback_unlock;
 882
 883         /*
 884          * This is a new and locked page, it shouldn't be mapped or
 885          * have any special flags on it
 886          */
 887         if (WARN_ON(page_mapped(oldpage)))
 888                 goto out_fallback_unlock;
 889         if (WARN_ON(page_has_private(oldpage)))
 890                 goto out_fallback_unlock;
 891         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 892                 goto out_fallback_unlock;
 893         if (WARN_ON(PageMlocked(oldpage)))
 894                 goto out_fallback_unlock;
 895
 896         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 897         if (err) {
 898                 unlock_page(newpage);
 899                 return err;
 900         }
 901
 902         page_cache_get(newpage);
 903
 904         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 905                 lru_cache_add_file(newpage);
 906
 907         err = 0;
 908         spin_lock(&cs->req->waitq.lock);
 909         if (test_bit(FR_ABORTED, &cs->req->flags))
 910                 err = -ENOENT;
 911         else
 912                 *pagep = newpage;
 913         spin_unlock(&cs->req->waitq.lock);
 914
 915         if (err) {
 916                 unlock_page(newpage);
 917                 page_cache_release(newpage);
 918                 return err;
 919         }
 920
 921         unlock_page(oldpage);
 922         page_cache_release(oldpage);
 923         cs->len = 0;
 924
 925         return 0;
 926
 927 out_fallback_unlock:
 928         unlock_page(newpage);
 929 out_fallback:
 930         cs->pg = buf->page;
 931         cs->offset = buf->offset;
 932
 933         err = lock_request(cs->req);
 934         if (err)
 935                 return err;
 936
 937         return 1;
 938 }
 939
 940 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 941                          unsigned offset, unsigned count)
 942 {
 943         struct pipe_buffer *buf;
 944         int err;
 945
 946         if (cs->nr_segs == cs->pipe->buffers)
 947                 return -EIO;
 948
 949         err = unlock_request(cs->req);
 950         if (err)
 951                 return err;
 952
 953         fuse_copy_finish(cs);
 954
 955         buf = cs->pipebufs;
 956         page_cache_get(page);
 957         buf->page = page;
 958         buf->offset = offset;
 959         buf->len = count;
 960
 961         cs->pipebufs++;
 962         cs->nr_segs++;
 963         cs->len = 0;
 964
 965         return 0;
 966 }
 967
 968 /*
 969  * Copy a page in the request to/from the userspace buffer.  Must be
 970  * done atomically
 971  */
 972 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 973                           unsigned offset, unsigned count, int zeroing)
 974 {
 975         int err;
 976         struct page *page = *pagep;
 977
 978         if (page && zeroing && count < PAGE_SIZE)
 979                 clear_highpage(page);
 980
 981         while (count) {
 982                 if (cs->write && cs->pipebufs && page) {
 983                         return fuse_ref_page(cs, page, offset, count);
 984                 } else if (!cs->len) {
 985                         if (cs->move_pages && page &&
 986                             offset == 0 && count == PAGE_SIZE) {
 987                                 err = fuse_try_move_page(cs, pagep);
 988                                 if (err <= 0)
 989                                         return err;
 990                         } else {
 991                                 err = fuse_copy_fill(cs);
 992                                 if (err)
 993                                         return err;
 994                         }
 995                 }
 996                 if (page) {
 997                         void *mapaddr = kmap_atomic(page);
 998                         void *buf = mapaddr + offset;
 999                         offset += fuse_copy_do(cs, &buf, &count);
1000                         kunmap_atomic(mapaddr);
1001                 } else
1002                         offset += fuse_copy_do(cs, NULL, &count);
1003         }
1004         if (page && !cs->write)
1005                 flush_dcache_page(page);
1006         return 0;
1007 }
1008
1009 /* Copy pages in the request to/from userspace buffer */
1010 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1011                            int zeroing)
1012 {
1013         unsigned i;
1014         struct fuse_req *req = cs->req;
1015
1016         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1017                 int err;
1018                 unsigned offset = req->page_descs[i].offset;
1019                 unsigned count = min(nbytes, req->page_descs[i].length);
1020
1021                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1022                                      zeroing);
1023                 if (err)
1024                         return err;
1025
1026                 nbytes -= count;
1027         }
1028         return 0;
1029 }
1030
1031 /* Copy a single argument in the request to/from userspace buffer */
1032 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1033 {
1034         while (size) {
1035                 if (!cs->len) {
1036                         int err = fuse_copy_fill(cs);
1037                         if (err)
1038                                 return err;
1039                 }
1040                 fuse_copy_do(cs, &val, &size);
1041         }
1042         return 0;
1043 }
1044
1045 /* Copy request arguments to/from userspace buffer */
1046 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1047                           unsigned argpages, struct fuse_arg *args,
1048                           int zeroing)
1049 {
1050         int err = 0;
1051         unsigned i;
1052
1053         for (i = 0; !err && i < numargs; i++)  {
1054                 struct fuse_arg *arg = &args[i];
1055                 if (i == numargs - 1 && argpages)
1056                         err = fuse_copy_pages(cs, arg->size, zeroing);
1057                 else
1058                         err = fuse_copy_one(cs, arg->value, arg->size);
1059         }
1060         return err;
1061 }
1062
1063 static int forget_pending(struct fuse_iqueue *fiq)
1064 {
1065         return fiq->forget_list_head.next != NULL;
1066 }
1067
1068 static int request_pending(struct fuse_iqueue *fiq)
1069 {
1070         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1071                 forget_pending(fiq);
1072 }
1073
1074 /*
1075  * Transfer an interrupt request to userspace
1076  *
1077  * Unlike other requests this is assembled on demand, without a need
1078  * to allocate a separate fuse_req structure.
1079  *
1080  * Called with fiq->waitq.lock held, releases it
1081  */
1082 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1083                                struct fuse_copy_state *cs,
1084                                size_t nbytes, struct fuse_req *req)
1085 __releases(fiq->waitq.lock)
1086 {
1087         struct fuse_in_header ih;
1088         struct fuse_interrupt_in arg;
1089         unsigned reqsize = sizeof(ih) + sizeof(arg);
1090         int err;
1091
1092         list_del_init(&req->intr_entry);
1093         req->intr_unique = fuse_get_unique(fiq);
1094         memset(&ih, 0, sizeof(ih));
1095         memset(&arg, 0, sizeof(arg));
1096         ih.len = reqsize;
1097         ih.opcode = FUSE_INTERRUPT;
1098         ih.unique = req->intr_unique;
1099         arg.unique = req->in.h.unique;
1100
1101         spin_unlock(&fiq->waitq.lock);
1102         if (nbytes < reqsize)
1103                 return -EINVAL;
1104
1105         err = fuse_copy_one(cs, &ih, sizeof(ih));
1106         if (!err)
1107                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1108         fuse_copy_finish(cs);
1109
1110         return err ? err : reqsize;
1111 }
1112
1113 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1114                                                unsigned max,
1115                                                unsigned *countp)
1116 {
1117         struct fuse_forget_link *head = fiq->forget_list_head.next;
1118         struct fuse_forget_link **newhead = &head;
1119         unsigned count;
1120
1121         for (count = 0; *newhead != NULL && count < max; count++)
1122                 newhead = &(*newhead)->next;
1123
1124         fiq->forget_list_head.next = *newhead;
1125         *newhead = NULL;
1126         if (fiq->forget_list_head.next == NULL)
1127                 fiq->forget_list_tail = &fiq->forget_list_head;
1128
1129         if (countp != NULL)
1130                 *countp = count;
1131
1132         return head;
1133 }
1134
1135 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1136                                    struct fuse_copy_state *cs,
1137                                    size_t nbytes)
1138 __releases(fiq->waitq.lock)
1139 {
1140         int err;
1141         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1142         struct fuse_forget_in arg = {
1143                 .nlookup = forget->forget_one.nlookup,
1144         };
1145         struct fuse_in_header ih = {
1146                 .opcode = FUSE_FORGET,
1147                 .nodeid = forget->forget_one.nodeid,
1148                 .unique = fuse_get_unique(fiq),
1149                 .len = sizeof(ih) + sizeof(arg),
1150         };
1151
1152         spin_unlock(&fiq->waitq.lock);
1153         kfree(forget);
1154         if (nbytes < ih.len)
1155                 return -EINVAL;
1156
1157         err = fuse_copy_one(cs, &ih, sizeof(ih));
1158         if (!err)
1159                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1160         fuse_copy_finish(cs);
1161
1162         if (err)
1163                 return err;
1164
1165         return ih.len;
1166 }
1167
1168 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1169                                    struct fuse_copy_state *cs, size_t nbytes)
1170 __releases(fiq->waitq.lock)
1171 {
1172         int err;
1173         unsigned max_forgets;
1174         unsigned count;
1175         struct fuse_forget_link *head;
1176         struct fuse_batch_forget_in arg = { .count = 0 };
1177         struct fuse_in_header ih = {
1178                 .opcode = FUSE_BATCH_FORGET,
1179                 .unique = fuse_get_unique(fiq),
1180                 .len = sizeof(ih) + sizeof(arg),
1181         };
1182
1183         if (nbytes < ih.len) {
1184                 spin_unlock(&fiq->waitq.lock);
1185                 return -EINVAL;
1186         }
1187
1188         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1189         head = dequeue_forget(fiq, max_forgets, &count);
1190         spin_unlock(&fiq->waitq.lock);
1191
1192         arg.count = count;
1193         ih.len += count * sizeof(struct fuse_forget_one);
1194         err = fuse_copy_one(cs, &ih, sizeof(ih));
1195         if (!err)
1196                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1197
1198         while (head) {
1199                 struct fuse_forget_link *forget = head;
1200
1201                 if (!err) {
1202                         err = fuse_copy_one(cs, &forget->forget_one,
1203                                             sizeof(forget->forget_one));
1204                 }
1205                 head = forget->next;
1206                 kfree(forget);
1207         }
1208
1209         fuse_copy_finish(cs);
1210
1211         if (err)
1212                 return err;
1213
1214         return ih.len;
1215 }
1216
1217 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1218                             struct fuse_copy_state *cs,
1219                             size_t nbytes)
1220 __releases(fiq->waitq.lock)
1221 {
1222         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1223                 return fuse_read_single_forget(fiq, cs, nbytes);
1224         else
1225                 return fuse_read_batch_forget(fiq, cs, nbytes);
1226 }
1227
1228 /*
1229  * Read a single request into the userspace filesystem's buffer.  This
1230  * function waits until a request is available, then removes it from
1231  * the pending list and copies request data to userspace buffer.  If
1232  * no reply is needed (FORGET) or request has been aborted or there
1233  * was an error during the copying then it's finished by calling
1234  * request_end().  Otherwise add it to the processing list, and set
1235  * the 'sent' flag.
1236  */
1237 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1238                                 struct fuse_copy_state *cs, size_t nbytes)
1239 {
1240         ssize_t err;
1241         struct fuse_iqueue *fiq = &fc->iq;
1242         struct fuse_pqueue *fpq = &fc->pq;
1243         struct fuse_req *req;
1244         struct fuse_in *in;
1245         unsigned reqsize;
1246
1247  restart:
1248         spin_lock(&fiq->waitq.lock);
1249         err = -EAGAIN;
1250         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1251             !request_pending(fiq))
1252                 goto err_unlock;
1253
1254         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1255                                 !fiq->connected || request_pending(fiq));
1256         if (err)
1257                 goto err_unlock;
1258
1259         err = -ENODEV;
1260         if (!fiq->connected)
1261                 goto err_unlock;
1262
1263         if (!list_empty(&fiq->interrupts)) {
1264                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1265                                  intr_entry);
1266                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1267         }
1268
1269         if (forget_pending(fiq)) {
1270                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1271                         return fuse_read_forget(fc, fiq, cs, nbytes);
1272
1273                 if (fiq->forget_batch <= -8)
1274                         fiq->forget_batch = 16;
1275         }
1276
1277         req = list_entry(fiq->pending.next, struct fuse_req, list);
1278         clear_bit(FR_PENDING, &req->flags);
1279         list_del_init(&req->list);
1280         spin_unlock(&fiq->waitq.lock);
1281
1282         spin_lock(&fc->lock);
1283         in = &req->in;
1284         reqsize = in->h.len;
1285         /* If request is too large, reply with an error and restart the read */
1286         if (nbytes < reqsize) {
1287                 req->out.h.error = -EIO;
1288                 /* SETXATTR is special, since it may contain too large data */
1289                 if (in->h.opcode == FUSE_SETXATTR)
1290                         req->out.h.error = -E2BIG;
1291                 request_end(fc, req);
1292                 goto restart;
1293         }
1294         spin_lock(&fpq->lock);
1295         list_add(&req->list, &fpq->io);
1296         spin_unlock(&fpq->lock);
1297         spin_unlock(&fc->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(&fc->lock);
1305         spin_lock(&fpq->lock);
1306         clear_bit(FR_LOCKED, &req->flags);
1307         if (!fpq->connected) {
1308                 err = -ENODEV;
1309                 goto out_end;
1310         }
1311         if (err) {
1312                 req->out.h.error = -EIO;
1313                 goto out_end;
1314         }
1315         if (!test_bit(FR_ISREPLY, &req->flags)) {
1316                 err = reqsize;
1317                 goto out_end;
1318         }
1319         list_move_tail(&req->list, &fpq->processing);
1320         spin_unlock(&fpq->lock);
1321         set_bit(FR_SENT, &req->flags);
1322         /* matches barrier in request_wait_answer() */
1323         smp_mb__after_atomic();
1324         if (test_bit(FR_INTERRUPTED, &req->flags))
1325                 queue_interrupt(fiq, req);
1326         spin_unlock(&fc->lock);
1327
1328         return reqsize;
1329
1330 out_end:
1331         list_del_init(&req->list);
1332         spin_unlock(&fpq->lock);
1333         request_end(fc, req);
1334         return err;
1335
1336  err_unlock:
1337         spin_unlock(&fiq->waitq.lock);
1338         return err;
1339 }
1340
1341 static int fuse_dev_open(struct inode *inode, struct file *file)
1342 {
1343         /*
1344          * The fuse device's file's private_data is used to hold
1345          * the fuse_conn(ection) when it is mounted, and is used to
1346          * keep track of whether the file has been mounted already.
1347          */
1348         file->private_data = NULL;
1349         return 0;
1350 }
1351
1352 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1353 {
1354         struct fuse_copy_state cs;
1355         struct file *file = iocb->ki_filp;
1356         struct fuse_conn *fc = fuse_get_conn(file);
1357         if (!fc)
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(fc, 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_conn *fc = fuse_get_conn(in);
1378         if (!fc)
1379                 return -EPERM;
1380
1381         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1382         if (!bufs)
1383                 return -ENOMEM;
1384
1385         fuse_copy_init(&cs, 1, NULL);
1386         cs.pipebufs = bufs;
1387         cs.pipe = pipe;
1388         ret = fuse_dev_do_read(fc, in, &cs, len);
1389         if (ret < 0)
1390                 goto out;
1391
1392         ret = 0;
1393         pipe_lock(pipe);
1394
1395         if (!pipe->readers) {
1396                 send_sig(SIGPIPE, current, 0);
1397                 if (!ret)
1398                         ret = -EPIPE;
1399                 goto out_unlock;
1400         }
1401
1402         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1403                 ret = -EIO;
1404                 goto out_unlock;
1405         }
1406
1407         while (page_nr < cs.nr_segs) {
1408                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1409                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1410
1411                 buf->page = bufs[page_nr].page;
1412                 buf->offset = bufs[page_nr].offset;
1413                 buf->len = bufs[page_nr].len;
1414                 /*
1415                  * Need to be careful about this.  Having buf->ops in module
1416                  * code can Oops if the buffer persists after module unload.
1417                  */
1418                 buf->ops = &nosteal_pipe_buf_ops;
1419
1420                 pipe->nrbufs++;
1421                 page_nr++;
1422                 ret += buf->len;
1423
1424                 if (pipe->files)
1425                         do_wakeup = 1;
1426         }
1427
1428 out_unlock:
1429         pipe_unlock(pipe);
1430
1431         if (do_wakeup) {
1432                 smp_mb();
1433                 if (waitqueue_active(&pipe->wait))
1434                         wake_up_interruptible(&pipe->wait);
1435                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1436         }
1437
1438 out:
1439         for (; page_nr < cs.nr_segs; page_nr++)
1440                 page_cache_release(bufs[page_nr].page);
1441
1442         kfree(bufs);
1443         return ret;
1444 }
1445
1446 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1447                             struct fuse_copy_state *cs)
1448 {
1449         struct fuse_notify_poll_wakeup_out outarg;
1450         int err = -EINVAL;
1451
1452         if (size != sizeof(outarg))
1453                 goto err;
1454
1455         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1456         if (err)
1457                 goto err;
1458
1459         fuse_copy_finish(cs);
1460         return fuse_notify_poll_wakeup(fc, &outarg);
1461
1462 err:
1463         fuse_copy_finish(cs);
1464         return err;
1465 }
1466
1467 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1468                                    struct fuse_copy_state *cs)
1469 {
1470         struct fuse_notify_inval_inode_out outarg;
1471         int err = -EINVAL;
1472
1473         if (size != sizeof(outarg))
1474                 goto err;
1475
1476         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1477         if (err)
1478                 goto err;
1479         fuse_copy_finish(cs);
1480
1481         down_read(&fc->killsb);
1482         err = -ENOENT;
1483         if (fc->sb) {
1484                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1485                                                outarg.off, outarg.len);
1486         }
1487         up_read(&fc->killsb);
1488         return err;
1489
1490 err:
1491         fuse_copy_finish(cs);
1492         return err;
1493 }
1494
1495 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1496                                    struct fuse_copy_state *cs)
1497 {
1498         struct fuse_notify_inval_entry_out outarg;
1499         int err = -ENOMEM;
1500         char *buf;
1501         struct qstr name;
1502
1503         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1504         if (!buf)
1505                 goto err;
1506
1507         err = -EINVAL;
1508         if (size < sizeof(outarg))
1509                 goto err;
1510
1511         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1512         if (err)
1513                 goto err;
1514
1515         err = -ENAMETOOLONG;
1516         if (outarg.namelen > FUSE_NAME_MAX)
1517                 goto err;
1518
1519         err = -EINVAL;
1520         if (size != sizeof(outarg) + outarg.namelen + 1)
1521                 goto err;
1522
1523         name.name = buf;
1524         name.len = outarg.namelen;
1525         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1526         if (err)
1527                 goto err;
1528         fuse_copy_finish(cs);
1529         buf[outarg.namelen] = 0;
1530         name.hash = full_name_hash(name.name, name.len);
1531
1532         down_read(&fc->killsb);
1533         err = -ENOENT;
1534         if (fc->sb)
1535                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1536         up_read(&fc->killsb);
1537         kfree(buf);
1538         return err;
1539
1540 err:
1541         kfree(buf);
1542         fuse_copy_finish(cs);
1543         return err;
1544 }
1545
1546 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1547                               struct fuse_copy_state *cs)
1548 {
1549         struct fuse_notify_delete_out outarg;
1550         int err = -ENOMEM;
1551         char *buf;
1552         struct qstr name;
1553
1554         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1555         if (!buf)
1556                 goto err;
1557
1558         err = -EINVAL;
1559         if (size < sizeof(outarg))
1560                 goto err;
1561
1562         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1563         if (err)
1564                 goto err;
1565
1566         err = -ENAMETOOLONG;
1567         if (outarg.namelen > FUSE_NAME_MAX)
1568                 goto err;
1569
1570         err = -EINVAL;
1571         if (size != sizeof(outarg) + outarg.namelen + 1)
1572                 goto err;
1573
1574         name.name = buf;
1575         name.len = outarg.namelen;
1576         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1577         if (err)
1578                 goto err;
1579         fuse_copy_finish(cs);
1580         buf[outarg.namelen] = 0;
1581         name.hash = full_name_hash(name.name, name.len);
1582
1583         down_read(&fc->killsb);
1584         err = -ENOENT;
1585         if (fc->sb)
1586                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1587                                                outarg.child, &name);
1588         up_read(&fc->killsb);
1589         kfree(buf);
1590         return err;
1591
1592 err:
1593         kfree(buf);
1594         fuse_copy_finish(cs);
1595         return err;
1596 }
1597
1598 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1599                              struct fuse_copy_state *cs)
1600 {
1601         struct fuse_notify_store_out outarg;
1602         struct inode *inode;
1603         struct address_space *mapping;
1604         u64 nodeid;
1605         int err;
1606         pgoff_t index;
1607         unsigned int offset;
1608         unsigned int num;
1609         loff_t file_size;
1610         loff_t end;
1611
1612         err = -EINVAL;
1613         if (size < sizeof(outarg))
1614                 goto out_finish;
1615
1616         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1617         if (err)
1618                 goto out_finish;
1619
1620         err = -EINVAL;
1621         if (size - sizeof(outarg) != outarg.size)
1622                 goto out_finish;
1623
1624         nodeid = outarg.nodeid;
1625
1626         down_read(&fc->killsb);
1627
1628         err = -ENOENT;
1629         if (!fc->sb)
1630                 goto out_up_killsb;
1631
1632         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1633         if (!inode)
1634                 goto out_up_killsb;
1635
1636         mapping = inode->i_mapping;
1637         index = outarg.offset >> PAGE_CACHE_SHIFT;
1638         offset = outarg.offset & ~PAGE_CACHE_MASK;
1639         file_size = i_size_read(inode);
1640         end = outarg.offset + outarg.size;
1641         if (end > file_size) {
1642                 file_size = end;
1643                 fuse_write_update_size(inode, file_size);
1644         }
1645
1646         num = outarg.size;
1647         while (num) {
1648                 struct page *page;
1649                 unsigned int this_num;
1650
1651                 err = -ENOMEM;
1652                 page = find_or_create_page(mapping, index,
1653                                            mapping_gfp_mask(mapping));
1654                 if (!page)
1655                         goto out_iput;
1656
1657                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1658                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1659                 if (!err && offset == 0 &&
1660                     (this_num == PAGE_CACHE_SIZE || file_size == end))
1661                         SetPageUptodate(page);
1662                 unlock_page(page);
1663                 page_cache_release(page);
1664
1665                 if (err)
1666                         goto out_iput;
1667
1668                 num -= this_num;
1669                 offset = 0;
1670                 index++;
1671         }
1672
1673         err = 0;
1674
1675 out_iput:
1676         iput(inode);
1677 out_up_killsb:
1678         up_read(&fc->killsb);
1679 out_finish:
1680         fuse_copy_finish(cs);
1681         return err;
1682 }
1683
1684 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1685 {
1686         release_pages(req->pages, req->num_pages, false);
1687 }
1688
1689 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1690                          struct fuse_notify_retrieve_out *outarg)
1691 {
1692         int err;
1693         struct address_space *mapping = inode->i_mapping;
1694         struct fuse_req *req;
1695         pgoff_t index;
1696         loff_t file_size;
1697         unsigned int num;
1698         unsigned int offset;
1699         size_t total_len = 0;
1700         int num_pages;
1701
1702         offset = outarg->offset & ~PAGE_CACHE_MASK;
1703         file_size = i_size_read(inode);
1704
1705         num = outarg->size;
1706         if (outarg->offset > file_size)
1707                 num = 0;
1708         else if (outarg->offset + num > file_size)
1709                 num = file_size - outarg->offset;
1710
1711         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1712         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1713
1714         req = fuse_get_req(fc, num_pages);
1715         if (IS_ERR(req))
1716                 return PTR_ERR(req);
1717
1718         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1719         req->in.h.nodeid = outarg->nodeid;
1720         req->in.numargs = 2;
1721         req->in.argpages = 1;
1722         req->page_descs[0].offset = offset;
1723         req->end = fuse_retrieve_end;
1724
1725         index = outarg->offset >> PAGE_CACHE_SHIFT;
1726
1727         while (num && req->num_pages < num_pages) {
1728                 struct page *page;
1729                 unsigned int this_num;
1730
1731                 page = find_get_page(mapping, index);
1732                 if (!page)
1733                         break;
1734
1735                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1736                 req->pages[req->num_pages] = page;
1737                 req->page_descs[req->num_pages].length = this_num;
1738                 req->num_pages++;
1739
1740                 offset = 0;
1741                 num -= this_num;
1742                 total_len += this_num;
1743                 index++;
1744         }
1745         req->misc.retrieve_in.offset = outarg->offset;
1746         req->misc.retrieve_in.size = total_len;
1747         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1748         req->in.args[0].value = &req->misc.retrieve_in;
1749         req->in.args[1].size = total_len;
1750
1751         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1752         if (err)
1753                 fuse_retrieve_end(fc, req);
1754
1755         return err;
1756 }
1757
1758 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1759                                 struct fuse_copy_state *cs)
1760 {
1761         struct fuse_notify_retrieve_out outarg;
1762         struct inode *inode;
1763         int err;
1764
1765         err = -EINVAL;
1766         if (size != sizeof(outarg))
1767                 goto copy_finish;
1768
1769         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1770         if (err)
1771                 goto copy_finish;
1772
1773         fuse_copy_finish(cs);
1774
1775         down_read(&fc->killsb);
1776         err = -ENOENT;
1777         if (fc->sb) {
1778                 u64 nodeid = outarg.nodeid;
1779
1780                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1781                 if (inode) {
1782                         err = fuse_retrieve(fc, inode, &outarg);
1783                         iput(inode);
1784                 }
1785         }
1786         up_read(&fc->killsb);
1787
1788         return err;
1789
1790 copy_finish:
1791         fuse_copy_finish(cs);
1792         return err;
1793 }
1794
1795 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1796                        unsigned int size, struct fuse_copy_state *cs)
1797 {
1798         /* Don't try to move pages (yet) */
1799         cs->move_pages = 0;
1800
1801         switch (code) {
1802         case FUSE_NOTIFY_POLL:
1803                 return fuse_notify_poll(fc, size, cs);
1804
1805         case FUSE_NOTIFY_INVAL_INODE:
1806                 return fuse_notify_inval_inode(fc, size, cs);
1807
1808         case FUSE_NOTIFY_INVAL_ENTRY:
1809                 return fuse_notify_inval_entry(fc, size, cs);
1810
1811         case FUSE_NOTIFY_STORE:
1812                 return fuse_notify_store(fc, size, cs);
1813
1814         case FUSE_NOTIFY_RETRIEVE:
1815                 return fuse_notify_retrieve(fc, size, cs);
1816
1817         case FUSE_NOTIFY_DELETE:
1818                 return fuse_notify_delete(fc, size, cs);
1819
1820         default:
1821                 fuse_copy_finish(cs);
1822                 return -EINVAL;
1823         }
1824 }
1825
1826 /* Look up request on processing list by unique ID */
1827 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1828 {
1829         struct fuse_req *req;
1830
1831         list_for_each_entry(req, &fpq->processing, list) {
1832                 if (req->in.h.unique == unique || req->intr_unique == unique)
1833                         return req;
1834         }
1835         return NULL;
1836 }
1837
1838 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1839                          unsigned nbytes)
1840 {
1841         unsigned reqsize = sizeof(struct fuse_out_header);
1842
1843         if (out->h.error)
1844                 return nbytes != reqsize ? -EINVAL : 0;
1845
1846         reqsize += len_args(out->numargs, out->args);
1847
1848         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1849                 return -EINVAL;
1850         else if (reqsize > nbytes) {
1851                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1852                 unsigned diffsize = reqsize - nbytes;
1853                 if (diffsize > lastarg->size)
1854                         return -EINVAL;
1855                 lastarg->size -= diffsize;
1856         }
1857         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1858                               out->page_zeroing);
1859 }
1860
1861 /*
1862  * Write a single reply to a request.  First the header is copied from
1863  * the write buffer.  The request is then searched on the processing
1864  * list by the unique ID found in the header.  If found, then remove
1865  * it from the list and copy the rest of the buffer to the request.
1866  * The request is finished by calling request_end()
1867  */
1868 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1869                                  struct fuse_copy_state *cs, size_t nbytes)
1870 {
1871         int err;
1872         struct fuse_pqueue *fpq = &fc->pq;
1873         struct fuse_req *req;
1874         struct fuse_out_header oh;
1875
1876         if (nbytes < sizeof(struct fuse_out_header))
1877                 return -EINVAL;
1878
1879         err = fuse_copy_one(cs, &oh, sizeof(oh));
1880         if (err)
1881                 goto err_finish;
1882
1883         err = -EINVAL;
1884         if (oh.len != nbytes)
1885                 goto err_finish;
1886
1887         /*
1888          * Zero oh.unique indicates unsolicited notification message
1889          * and error contains notification code.
1890          */
1891         if (!oh.unique) {
1892                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1893                 return err ? err : nbytes;
1894         }
1895
1896         err = -EINVAL;
1897         if (oh.error <= -1000 || oh.error > 0)
1898                 goto err_finish;
1899
1900         spin_lock(&fc->lock);
1901         spin_lock(&fpq->lock);
1902         err = -ENOENT;
1903         if (!fpq->connected)
1904                 goto err_unlock_pq;
1905
1906         req = request_find(fpq, oh.unique);
1907         if (!req)
1908                 goto err_unlock_pq;
1909
1910         /* Is it an interrupt reply? */
1911         if (req->intr_unique == oh.unique) {
1912                 spin_unlock(&fpq->lock);
1913
1914                 err = -EINVAL;
1915                 if (nbytes != sizeof(struct fuse_out_header))
1916                         goto err_unlock;
1917
1918                 if (oh.error == -ENOSYS)
1919                         fc->no_interrupt = 1;
1920                 else if (oh.error == -EAGAIN)
1921                         queue_interrupt(&fc->iq, req);
1922
1923                 spin_unlock(&fc->lock);
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         spin_unlock(&fc->lock);
1937
1938         err = copy_out_args(cs, &req->out, nbytes);
1939         fuse_copy_finish(cs);
1940
1941         spin_lock(&fc->lock);
1942         spin_lock(&fpq->lock);
1943         clear_bit(FR_LOCKED, &req->flags);
1944         if (!fpq->connected)
1945                 err = -ENOENT;
1946         else if (err)
1947                 req->out.h.error = -EIO;
1948         list_del_init(&req->list);
1949         spin_unlock(&fpq->lock);
1950         request_end(fc, req);
1951
1952         return err ? err : nbytes;
1953
1954  err_unlock_pq:
1955         spin_unlock(&fpq->lock);
1956  err_unlock:
1957         spin_unlock(&fc->lock);
1958  err_finish:
1959         fuse_copy_finish(cs);
1960         return err;
1961 }
1962
1963 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1964 {
1965         struct fuse_copy_state cs;
1966         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1967         if (!fc)
1968                 return -EPERM;
1969
1970         if (!iter_is_iovec(from))
1971                 return -EINVAL;
1972
1973         fuse_copy_init(&cs, 0, from);
1974
1975         return fuse_dev_do_write(fc, &cs, iov_iter_count(from));
1976 }
1977
1978 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1979                                      struct file *out, loff_t *ppos,
1980                                      size_t len, unsigned int flags)
1981 {
1982         unsigned nbuf;
1983         unsigned idx;
1984         struct pipe_buffer *bufs;
1985         struct fuse_copy_state cs;
1986         struct fuse_conn *fc;
1987         size_t rem;
1988         ssize_t ret;
1989
1990         fc = fuse_get_conn(out);
1991         if (!fc)
1992                 return -EPERM;
1993
1994         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1995         if (!bufs)
1996                 return -ENOMEM;
1997
1998         pipe_lock(pipe);
1999         nbuf = 0;
2000         rem = 0;
2001         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2002                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2003
2004         ret = -EINVAL;
2005         if (rem < len) {
2006                 pipe_unlock(pipe);
2007                 goto out;
2008         }
2009
2010         rem = len;
2011         while (rem) {
2012                 struct pipe_buffer *ibuf;
2013                 struct pipe_buffer *obuf;
2014
2015                 BUG_ON(nbuf >= pipe->buffers);
2016                 BUG_ON(!pipe->nrbufs);
2017                 ibuf = &pipe->bufs[pipe->curbuf];
2018                 obuf = &bufs[nbuf];
2019
2020                 if (rem >= ibuf->len) {
2021                         *obuf = *ibuf;
2022                         ibuf->ops = NULL;
2023                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2024                         pipe->nrbufs--;
2025                 } else {
2026                         ibuf->ops->get(pipe, ibuf);
2027                         *obuf = *ibuf;
2028                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2029                         obuf->len = rem;
2030                         ibuf->offset += obuf->len;
2031                         ibuf->len -= obuf->len;
2032                 }
2033                 nbuf++;
2034                 rem -= obuf->len;
2035         }
2036         pipe_unlock(pipe);
2037
2038         fuse_copy_init(&cs, 0, NULL);
2039         cs.pipebufs = bufs;
2040         cs.nr_segs = nbuf;
2041         cs.pipe = pipe;
2042
2043         if (flags & SPLICE_F_MOVE)
2044                 cs.move_pages = 1;
2045
2046         ret = fuse_dev_do_write(fc, &cs, len);
2047
2048         for (idx = 0; idx < nbuf; idx++) {
2049                 struct pipe_buffer *buf = &bufs[idx];
2050                 buf->ops->release(pipe, buf);
2051         }
2052 out:
2053         kfree(bufs);
2054         return ret;
2055 }
2056
2057 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2058 {
2059         unsigned mask = POLLOUT | POLLWRNORM;
2060         struct fuse_iqueue *fiq;
2061         struct fuse_conn *fc = fuse_get_conn(file);
2062         if (!fc)
2063                 return POLLERR;
2064
2065         fiq = &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 __releases(fc->lock)
2085 __acquires(fc->lock)
2086 {
2087         while (!list_empty(head)) {
2088                 struct fuse_req *req;
2089                 req = list_entry(head->next, struct fuse_req, list);
2090                 req->out.h.error = -ECONNABORTED;
2091                 clear_bit(FR_PENDING, &req->flags);
2092                 clear_bit(FR_SENT, &req->flags);
2093                 list_del_init(&req->list);
2094                 request_end(fc, req);
2095                 spin_lock(&fc->lock);
2096         }
2097 }
2098
2099 static void end_polls(struct fuse_conn *fc)
2100 {
2101         struct rb_node *p;
2102
2103         p = rb_first(&fc->polled_files);
2104
2105         while (p) {
2106                 struct fuse_file *ff;
2107                 ff = rb_entry(p, struct fuse_file, polled_node);
2108                 wake_up_interruptible_all(&ff->poll_wait);
2109
2110                 p = rb_next(p);
2111         }
2112 }
2113
2114 /*
2115  * Abort all requests.
2116  *
2117  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2118  * filesystem.
2119  *
2120  * The same effect is usually achievable through killing the filesystem daemon
2121  * and all users of the filesystem.  The exception is the combination of an
2122  * asynchronous request and the tricky deadlock (see
2123  * Documentation/filesystems/fuse.txt).
2124  *
2125  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2126  * requests, they should be finished off immediately.  Locked requests will be
2127  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2128  * requests.  It is possible that some request will finish before we can.  This
2129  * is OK, the request will in that case be removed from the list before we touch
2130  * it.
2131  */
2132 void fuse_abort_conn(struct fuse_conn *fc)
2133 {
2134         struct fuse_iqueue *fiq = &fc->iq;
2135         struct fuse_pqueue *fpq = &fc->pq;
2136
2137         spin_lock(&fc->lock);
2138         if (fc->connected) {
2139                 struct fuse_req *req, *next;
2140                 LIST_HEAD(to_end1);
2141                 LIST_HEAD(to_end2);
2142
2143                 fc->connected = 0;
2144                 fc->blocked = 0;
2145                 fuse_set_initialized(fc);
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                                 list_move(&req->list, &to_end1);
2154                         spin_unlock(&req->waitq.lock);
2155                 }
2156                 list_splice_init(&fpq->processing, &to_end2);
2157                 spin_unlock(&fpq->lock);
2158                 fc->max_background = UINT_MAX;
2159                 flush_bg_queue(fc);
2160
2161                 spin_lock(&fiq->waitq.lock);
2162                 fiq->connected = 0;
2163                 list_splice_init(&fiq->pending, &to_end2);
2164                 while (forget_pending(fiq))
2165                         kfree(dequeue_forget(fiq, 1, NULL));
2166                 wake_up_all_locked(&fiq->waitq);
2167                 spin_unlock(&fiq->waitq.lock);
2168                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2169
2170                 while (!list_empty(&to_end1)) {
2171                         req = list_first_entry(&to_end1, struct fuse_req, list);
2172                         __fuse_get_request(req);
2173                         list_del_init(&req->list);
2174                         request_end(fc, req);
2175                         spin_lock(&fc->lock);
2176                 }
2177                 end_requests(fc, &to_end2);
2178                 end_polls(fc);
2179                 wake_up_all(&fc->blocked_waitq);
2180         }
2181         spin_unlock(&fc->lock);
2182 }
2183 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2184
2185 int fuse_dev_release(struct inode *inode, struct file *file)
2186 {
2187         struct fuse_conn *fc = fuse_get_conn(file);
2188         if (fc) {
2189                 WARN_ON(!list_empty(&fc->pq.io));
2190                 WARN_ON(fc->iq.fasync != NULL);
2191                 fuse_abort_conn(fc);
2192                 fuse_conn_put(fc);
2193         }
2194
2195         return 0;
2196 }
2197 EXPORT_SYMBOL_GPL(fuse_dev_release);
2198
2199 static int fuse_dev_fasync(int fd, struct file *file, int on)
2200 {
2201         struct fuse_conn *fc = fuse_get_conn(file);
2202         if (!fc)
2203                 return -EPERM;
2204
2205         /* No locking - fasync_helper does its own locking */
2206         return fasync_helper(fd, file, on, &fc->iq.fasync);
2207 }
2208
2209 const struct file_operations fuse_dev_operations = {
2210         .owner          = THIS_MODULE,
2211         .open           = fuse_dev_open,
2212         .llseek         = no_llseek,
2213         .read_iter      = fuse_dev_read,
2214         .splice_read    = fuse_dev_splice_read,
2215         .write_iter     = fuse_dev_write,
2216         .splice_write   = fuse_dev_splice_write,
2217         .poll           = fuse_dev_poll,
2218         .release        = fuse_dev_release,
2219         .fasync         = fuse_dev_fasync,
2220 };
2221 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2222
2223 static struct miscdevice fuse_miscdevice = {
2224         .minor = FUSE_MINOR,
2225         .name  = "fuse",
2226         .fops = &fuse_dev_operations,
2227 };
2228
2229 int __init fuse_dev_init(void)
2230 {
2231         int err = -ENOMEM;
2232         fuse_req_cachep = kmem_cache_create("fuse_request",
2233                                             sizeof(struct fuse_req),
2234                                             0, 0, NULL);
2235         if (!fuse_req_cachep)
2236                 goto out;
2237
2238         err = misc_register(&fuse_miscdevice);
2239         if (err)
2240                 goto out_cache_clean;
2241
2242         return 0;
2243
2244  out_cache_clean:
2245         kmem_cache_destroy(fuse_req_cachep);
2246  out:
2247         return err;
2248 }
2249
2250 void fuse_dev_cleanup(void)
2251 {
2252         misc_deregister(&fuse_miscdevice);
2253         kmem_cache_destroy(fuse_req_cachep);
2254 }