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