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/sched/signal.h>
  15 #include <linux/uio.h>
  16 #include <linux/miscdevice.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/pipe_fs_i.h>
  21 #include <linux/swap.h>
  22 #include <linux/splice.h>
  23 #include <linux/sched.h>
  24
  25 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  26 MODULE_ALIAS("devname:fuse");
  27
  28 /* Ordinary requests have even IDs, while interrupts IDs are odd */
  29 #define FUSE_INT_REQ_BIT (1ULL << 0)
  30 #define FUSE_REQ_ID_STEP (1ULL << 1)
  31
  32 static struct kmem_cache *fuse_req_cachep;
  33
  34 static struct fuse_dev *fuse_get_dev(struct file *file)
  35 {
  36         /*
  37          * Lockless access is OK, because file->private data is set
  38          * once during mount and is valid until the file is released.
  39          */
  40         return READ_ONCE(file->private_data);
  41 }
  42
  43 static void fuse_request_init(struct fuse_req *req)
  44 {
  45         INIT_LIST_HEAD(&req->list);
  46         INIT_LIST_HEAD(&req->intr_entry);
  47         init_waitqueue_head(&req->waitq);
  48         refcount_set(&req->count, 1);
  49         __set_bit(FR_PENDING, &req->flags);
  50 }
  51
  52 static struct fuse_req *fuse_request_alloc(gfp_t flags)
  53 {
  54         struct fuse_req *req = kmem_cache_zalloc(fuse_req_cachep, flags);
  55         if (req)
  56                 fuse_request_init(req);
  57
  58         return req;
  59 }
  60
  61 static void fuse_request_free(struct fuse_req *req)
  62 {
  63         kmem_cache_free(fuse_req_cachep, req);
  64 }
  65
  66 static void __fuse_get_request(struct fuse_req *req)
  67 {
  68         refcount_inc(&req->count);
  69 }
  70
  71 /* Must be called with > 1 refcount */
  72 static void __fuse_put_request(struct fuse_req *req)
  73 {
  74         refcount_dec(&req->count);
  75 }
  76
  77 void fuse_set_initialized(struct fuse_conn *fc)
  78 {
  79         /* Make sure stores before this are seen on another CPU */
  80         smp_wmb();
  81         fc->initialized = 1;
  82 }
  83
  84 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
  85 {
  86         return !fc->initialized || (for_background && fc->blocked);
  87 }
  88
  89 static void fuse_drop_waiting(struct fuse_conn *fc)
  90 {
  91         /*
  92          * lockess check of fc->connected is okay, because atomic_dec_and_test()
  93          * provides a memory barrier mached with the one in fuse_wait_aborted()
  94          * to ensure no wake-up is missed.
  95          */
  96         if (atomic_dec_and_test(&fc->num_waiting) &&
  97             !READ_ONCE(fc->connected)) {
  98                 /* wake up aborters */
  99                 wake_up_all(&fc->blocked_waitq);
 100         }
 101 }
 102
 103 static void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req);
 104
 105 static struct fuse_req *fuse_get_req(struct fuse_conn *fc, bool for_background)
 106 {
 107         struct fuse_req *req;
 108         int err;
 109         atomic_inc(&fc->num_waiting);
 110
 111         if (fuse_block_alloc(fc, for_background)) {
 112                 err = -EINTR;
 113                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 114                                 !fuse_block_alloc(fc, for_background)))
 115                         goto out;
 116         }
 117         /* Matches smp_wmb() in fuse_set_initialized() */
 118         smp_rmb();
 119
 120         err = -ENOTCONN;
 121         if (!fc->connected)
 122                 goto out;
 123
 124         err = -ECONNREFUSED;
 125         if (fc->conn_error)
 126                 goto out;
 127
 128         req = fuse_request_alloc(GFP_KERNEL);
 129         err = -ENOMEM;
 130         if (!req) {
 131                 if (for_background)
 132                         wake_up(&fc->blocked_waitq);
 133                 goto out;
 134         }
 135
 136         req->in.h.uid = from_kuid(fc->user_ns, current_fsuid());
 137         req->in.h.gid = from_kgid(fc->user_ns, current_fsgid());
 138         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 139
 140         __set_bit(FR_WAITING, &req->flags);
 141         if (for_background)
 142                 __set_bit(FR_BACKGROUND, &req->flags);
 143
 144         if (unlikely(req->in.h.uid == ((uid_t)-1) ||
 145                      req->in.h.gid == ((gid_t)-1))) {
 146                 fuse_put_request(fc, req);
 147                 return ERR_PTR(-EOVERFLOW);
 148         }
 149         return req;
 150
 151  out:
 152         fuse_drop_waiting(fc);
 153         return ERR_PTR(err);
 154 }
 155
 156 static void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 157 {
 158         if (refcount_dec_and_test(&req->count)) {
 159                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 160                         /*
 161                          * We get here in the unlikely case that a background
 162                          * request was allocated but not sent
 163                          */
 164                         spin_lock(&fc->bg_lock);
 165                         if (!fc->blocked)
 166                                 wake_up(&fc->blocked_waitq);
 167                         spin_unlock(&fc->bg_lock);
 168                 }
 169
 170                 if (test_bit(FR_WAITING, &req->flags)) {
 171                         __clear_bit(FR_WAITING, &req->flags);
 172                         fuse_drop_waiting(fc);
 173                 }
 174
 175                 fuse_request_free(req);
 176         }
 177 }
 178
 179 unsigned int fuse_len_args(unsigned int numargs, struct fuse_arg *args)
 180 {
 181         unsigned nbytes = 0;
 182         unsigned i;
 183
 184         for (i = 0; i < numargs; i++)
 185                 nbytes += args[i].size;
 186
 187         return nbytes;
 188 }
 189 EXPORT_SYMBOL_GPL(fuse_len_args);
 190
 191 u64 fuse_get_unique(struct fuse_iqueue *fiq)
 192 {
 193         fiq->reqctr += FUSE_REQ_ID_STEP;
 194         return fiq->reqctr;
 195 }
 196 EXPORT_SYMBOL_GPL(fuse_get_unique);
 197
 198 static unsigned int fuse_req_hash(u64 unique)
 199 {
 200         return hash_long(unique & ~FUSE_INT_REQ_BIT, FUSE_PQ_HASH_BITS);
 201 }
 202
 203 /**
 204  * A new request is available, wake fiq->waitq
 205  */
 206 static void fuse_dev_wake_and_unlock(struct fuse_iqueue *fiq)
 207 __releases(fiq->lock)
 208 {
 209         wake_up(&fiq->waitq);
 210         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 211         spin_unlock(&fiq->lock);
 212 }
 213
 214 const struct fuse_iqueue_ops fuse_dev_fiq_ops = {
 215         .wake_forget_and_unlock         = fuse_dev_wake_and_unlock,
 216         .wake_interrupt_and_unlock      = fuse_dev_wake_and_unlock,
 217         .wake_pending_and_unlock        = fuse_dev_wake_and_unlock,
 218 };
 219 EXPORT_SYMBOL_GPL(fuse_dev_fiq_ops);
 220
 221 static void queue_request_and_unlock(struct fuse_iqueue *fiq,
 222                                      struct fuse_req *req)
 223 __releases(fiq->lock)
 224 {
 225         req->in.h.len = sizeof(struct fuse_in_header) +
 226                 fuse_len_args(req->args->in_numargs,
 227                               (struct fuse_arg *) req->args->in_args);
 228         list_add_tail(&req->list, &fiq->pending);
 229         fiq->ops->wake_pending_and_unlock(fiq);
 230 }
 231
 232 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 233                        u64 nodeid, u64 nlookup)
 234 {
 235         struct fuse_iqueue *fiq = &fc->iq;
 236
 237         forget->forget_one.nodeid = nodeid;
 238         forget->forget_one.nlookup = nlookup;
 239
 240         spin_lock(&fiq->lock);
 241         if (fiq->connected) {
 242                 fiq->forget_list_tail->next = forget;
 243                 fiq->forget_list_tail = forget;
 244                 fiq->ops->wake_forget_and_unlock(fiq);
 245         } else {
 246                 kfree(forget);
 247                 spin_unlock(&fiq->lock);
 248         }
 249 }
 250
 251 static void flush_bg_queue(struct fuse_conn *fc)
 252 {
 253         struct fuse_iqueue *fiq = &fc->iq;
 254
 255         while (fc->active_background < fc->max_background &&
 256                !list_empty(&fc->bg_queue)) {
 257                 struct fuse_req *req;
 258
 259                 req = list_first_entry(&fc->bg_queue, struct fuse_req, list);
 260                 list_del(&req->list);
 261                 fc->active_background++;
 262                 spin_lock(&fiq->lock);
 263                 req->in.h.unique = fuse_get_unique(fiq);
 264                 queue_request_and_unlock(fiq, req);
 265         }
 266 }
 267
 268 /*
 269  * This function is called when a request is finished.  Either a reply
 270  * has arrived or it was aborted (and not yet sent) or some error
 271  * occurred during communication with userspace, or the device file
 272  * was closed.  The requester thread is woken up (if still waiting),
 273  * the 'end' callback is called if given, else the reference to the
 274  * request is released
 275  */
 276 void fuse_request_end(struct fuse_conn *fc, struct fuse_req *req)
 277 {
 278         struct fuse_iqueue *fiq = &fc->iq;
 279         bool async = req->args->end;
 280
 281         if (test_and_set_bit(FR_FINISHED, &req->flags))
 282                 goto put_request;
 283         /*
 284          * test_and_set_bit() implies smp_mb() between bit
 285          * changing and below intr_entry check. Pairs with
 286          * smp_mb() from queue_interrupt().
 287          */
 288         if (!list_empty(&req->intr_entry)) {
 289                 spin_lock(&fiq->lock);
 290                 list_del_init(&req->intr_entry);
 291                 spin_unlock(&fiq->lock);
 292         }
 293         WARN_ON(test_bit(FR_PENDING, &req->flags));
 294         WARN_ON(test_bit(FR_SENT, &req->flags));
 295         if (test_bit(FR_BACKGROUND, &req->flags)) {
 296                 spin_lock(&fc->bg_lock);
 297                 clear_bit(FR_BACKGROUND, &req->flags);
 298                 if (fc->num_background == fc->max_background) {
 299                         fc->blocked = 0;
 300                         wake_up(&fc->blocked_waitq);
 301                 } else if (!fc->blocked) {
 302                         /*
 303                          * Wake up next waiter, if any.  It's okay to use
 304                          * waitqueue_active(), as we've already synced up
 305                          * fc->blocked with waiters with the wake_up() call
 306                          * above.
 307                          */
 308                         if (waitqueue_active(&fc->blocked_waitq))
 309                                 wake_up(&fc->blocked_waitq);
 310                 }
 311
 312                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 313                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 314                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 315                 }
 316                 fc->num_background--;
 317                 fc->active_background--;
 318                 flush_bg_queue(fc);
 319                 spin_unlock(&fc->bg_lock);
 320         } else {
 321                 /* Wake up waiter sleeping in request_wait_answer() */
 322                 wake_up(&req->waitq);
 323         }
 324
 325         if (async)
 326                 req->args->end(fc, req->args, req->out.h.error);
 327 put_request:
 328         fuse_put_request(fc, req);
 329 }
 330 EXPORT_SYMBOL_GPL(fuse_request_end);
 331
 332 static int queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 333 {
 334         spin_lock(&fiq->lock);
 335         /* Check for we've sent request to interrupt this req */
 336         if (unlikely(!test_bit(FR_INTERRUPTED, &req->flags))) {
 337                 spin_unlock(&fiq->lock);
 338                 return -EINVAL;
 339         }
 340
 341         if (list_empty(&req->intr_entry)) {
 342                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 343                 /*
 344                  * Pairs with smp_mb() implied by test_and_set_bit()
 345                  * from request_end().
 346                  */
 347                 smp_mb();
 348                 if (test_bit(FR_FINISHED, &req->flags)) {
 349                         list_del_init(&req->intr_entry);
 350                         spin_unlock(&fiq->lock);
 351                         return 0;
 352                 }
 353                 fiq->ops->wake_interrupt_and_unlock(fiq);
 354         } else {
 355                 spin_unlock(&fiq->lock);
 356         }
 357         return 0;
 358 }
 359
 360 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 361 {
 362         struct fuse_iqueue *fiq = &fc->iq;
 363         int err;
 364
 365         if (!fc->no_interrupt) {
 366                 /* Any signal may interrupt this */
 367                 err = wait_event_interruptible(req->waitq,
 368                                         test_bit(FR_FINISHED, &req->flags));
 369                 if (!err)
 370                         return;
 371
 372                 set_bit(FR_INTERRUPTED, &req->flags);
 373                 /* matches barrier in fuse_dev_do_read() */
 374                 smp_mb__after_atomic();
 375                 if (test_bit(FR_SENT, &req->flags))
 376                         queue_interrupt(fiq, req);
 377         }
 378
 379         if (!test_bit(FR_FORCE, &req->flags)) {
 380                 /* Only fatal signals may interrupt this */
 381                 err = wait_event_killable(req->waitq,
 382                                         test_bit(FR_FINISHED, &req->flags));
 383                 if (!err)
 384                         return;
 385
 386                 spin_lock(&fiq->lock);
 387                 /* Request is not yet in userspace, bail out */
 388                 if (test_bit(FR_PENDING, &req->flags)) {
 389                         list_del(&req->list);
 390                         spin_unlock(&fiq->lock);
 391                         __fuse_put_request(req);
 392                         req->out.h.error = -EINTR;
 393                         return;
 394                 }
 395                 spin_unlock(&fiq->lock);
 396         }
 397
 398         /*
 399          * Either request is already in userspace, or it was forced.
 400          * Wait it out.
 401          */
 402         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 403 }
 404
 405 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 406 {
 407         struct fuse_iqueue *fiq = &fc->iq;
 408
 409         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 410         spin_lock(&fiq->lock);
 411         if (!fiq->connected) {
 412                 spin_unlock(&fiq->lock);
 413                 req->out.h.error = -ENOTCONN;
 414         } else {
 415                 req->in.h.unique = fuse_get_unique(fiq);
 416                 /* acquire extra reference, since request is still needed
 417                    after fuse_request_end() */
 418                 __fuse_get_request(req);
 419                 queue_request_and_unlock(fiq, req);
 420
 421                 request_wait_answer(fc, req);
 422                 /* Pairs with smp_wmb() in fuse_request_end() */
 423                 smp_rmb();
 424         }
 425 }
 426
 427 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 428 {
 429         if (fc->minor < 4 && args->opcode == FUSE_STATFS)
 430                 args->out_args[0].size = FUSE_COMPAT_STATFS_SIZE;
 431
 432         if (fc->minor < 9) {
 433                 switch (args->opcode) {
 434                 case FUSE_LOOKUP:
 435                 case FUSE_CREATE:
 436                 case FUSE_MKNOD:
 437                 case FUSE_MKDIR:
 438                 case FUSE_SYMLINK:
 439                 case FUSE_LINK:
 440                         args->out_args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 441                         break;
 442                 case FUSE_GETATTR:
 443                 case FUSE_SETATTR:
 444                         args->out_args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 445                         break;
 446                 }
 447         }
 448         if (fc->minor < 12) {
 449                 switch (args->opcode) {
 450                 case FUSE_CREATE:
 451                         args->in_args[0].size = sizeof(struct fuse_open_in);
 452                         break;
 453                 case FUSE_MKNOD:
 454                         args->in_args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 455                         break;
 456                 }
 457         }
 458 }
 459
 460 static void fuse_force_creds(struct fuse_conn *fc, struct fuse_req *req)
 461 {
 462         req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
 463         req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
 464         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 465 }
 466
 467 static void fuse_args_to_req(struct fuse_req *req, struct fuse_args *args)
 468 {
 469         req->in.h.opcode = args->opcode;
 470         req->in.h.nodeid = args->nodeid;
 471         req->args = args;
 472 }
 473
 474 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 475 {
 476         struct fuse_req *req;
 477         ssize_t ret;
 478
 479         if (args->force) {
 480                 atomic_inc(&fc->num_waiting);
 481                 req = fuse_request_alloc(GFP_KERNEL | __GFP_NOFAIL);
 482
 483                 if (!args->nocreds)
 484                         fuse_force_creds(fc, req);
 485
 486                 __set_bit(FR_WAITING, &req->flags);
 487                 __set_bit(FR_FORCE, &req->flags);
 488         } else {
 489                 WARN_ON(args->nocreds);
 490                 req = fuse_get_req(fc, false);
 491                 if (IS_ERR(req))
 492                         return PTR_ERR(req);
 493         }
 494
 495         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 496         fuse_adjust_compat(fc, args);
 497         fuse_args_to_req(req, args);
 498
 499         if (!args->noreply)
 500                 __set_bit(FR_ISREPLY, &req->flags);
 501         __fuse_request_send(fc, req);
 502         ret = req->out.h.error;
 503         if (!ret && args->out_argvar) {
 504                 BUG_ON(args->out_numargs == 0);
 505                 ret = args->out_args[args->out_numargs - 1].size;
 506         }
 507         fuse_put_request(fc, req);
 508
 509         return ret;
 510 }
 511
 512 static bool fuse_request_queue_background(struct fuse_conn *fc,
 513                                           struct fuse_req *req)
 514 {
 515         bool queued = false;
 516
 517         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 518         if (!test_bit(FR_WAITING, &req->flags)) {
 519                 __set_bit(FR_WAITING, &req->flags);
 520                 atomic_inc(&fc->num_waiting);
 521         }
 522         __set_bit(FR_ISREPLY, &req->flags);
 523         spin_lock(&fc->bg_lock);
 524         if (likely(fc->connected)) {
 525                 fc->num_background++;
 526                 if (fc->num_background == fc->max_background)
 527                         fc->blocked = 1;
 528                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 529                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 530                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 531                 }
 532                 list_add_tail(&req->list, &fc->bg_queue);
 533                 flush_bg_queue(fc);
 534                 queued = true;
 535         }
 536         spin_unlock(&fc->bg_lock);
 537
 538         return queued;
 539 }
 540
 541 int fuse_simple_background(struct fuse_conn *fc, struct fuse_args *args,
 542                             gfp_t gfp_flags)
 543 {
 544         struct fuse_req *req;
 545
 546         if (args->force) {
 547                 WARN_ON(!args->nocreds);
 548                 req = fuse_request_alloc(gfp_flags);
 549                 if (!req)
 550                         return -ENOMEM;
 551                 __set_bit(FR_BACKGROUND, &req->flags);
 552         } else {
 553                 WARN_ON(args->nocreds);
 554                 req = fuse_get_req(fc, true);
 555                 if (IS_ERR(req))
 556                         return PTR_ERR(req);
 557         }
 558
 559         fuse_args_to_req(req, args);
 560
 561         if (!fuse_request_queue_background(fc, req)) {
 562                 fuse_put_request(fc, req);
 563                 return -ENOTCONN;
 564         }
 565
 566         return 0;
 567 }
 568 EXPORT_SYMBOL_GPL(fuse_simple_background);
 569
 570 static int fuse_simple_notify_reply(struct fuse_conn *fc,
 571                                     struct fuse_args *args, u64 unique)
 572 {
 573         struct fuse_req *req;
 574         struct fuse_iqueue *fiq = &fc->iq;
 575         int err = 0;
 576
 577         req = fuse_get_req(fc, false);
 578         if (IS_ERR(req))
 579                 return PTR_ERR(req);
 580
 581         __clear_bit(FR_ISREPLY, &req->flags);
 582         req->in.h.unique = unique;
 583
 584         fuse_args_to_req(req, args);
 585
 586         spin_lock(&fiq->lock);
 587         if (fiq->connected) {
 588                 queue_request_and_unlock(fiq, req);
 589         } else {
 590                 err = -ENODEV;
 591                 spin_unlock(&fiq->lock);
 592                 fuse_put_request(fc, req);
 593         }
 594
 595         return err;
 596 }
 597
 598 /*
 599  * Lock the request.  Up to the next unlock_request() there mustn't be
 600  * anything that could cause a page-fault.  If the request was already
 601  * aborted bail out.
 602  */
 603 static int lock_request(struct fuse_req *req)
 604 {
 605         int err = 0;
 606         if (req) {
 607                 spin_lock(&req->waitq.lock);
 608                 if (test_bit(FR_ABORTED, &req->flags))
 609                         err = -ENOENT;
 610                 else
 611                         set_bit(FR_LOCKED, &req->flags);
 612                 spin_unlock(&req->waitq.lock);
 613         }
 614         return err;
 615 }
 616
 617 /*
 618  * Unlock request.  If it was aborted while locked, caller is responsible
 619  * for unlocking and ending the request.
 620  */
 621 static int unlock_request(struct fuse_req *req)
 622 {
 623         int err = 0;
 624         if (req) {
 625                 spin_lock(&req->waitq.lock);
 626                 if (test_bit(FR_ABORTED, &req->flags))
 627                         err = -ENOENT;
 628                 else
 629                         clear_bit(FR_LOCKED, &req->flags);
 630                 spin_unlock(&req->waitq.lock);
 631         }
 632         return err;
 633 }
 634
 635 struct fuse_copy_state {
 636         int write;
 637         struct fuse_req *req;
 638         struct iov_iter *iter;
 639         struct pipe_buffer *pipebufs;
 640         struct pipe_buffer *currbuf;
 641         struct pipe_inode_info *pipe;
 642         unsigned long nr_segs;
 643         struct page *pg;
 644         unsigned len;
 645         unsigned offset;
 646         unsigned move_pages:1;
 647 };
 648
 649 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 650                            struct iov_iter *iter)
 651 {
 652         memset(cs, 0, sizeof(*cs));
 653         cs->write = write;
 654         cs->iter = iter;
 655 }
 656
 657 /* Unmap and put previous page of userspace buffer */
 658 static void fuse_copy_finish(struct fuse_copy_state *cs)
 659 {
 660         if (cs->currbuf) {
 661                 struct pipe_buffer *buf = cs->currbuf;
 662
 663                 if (cs->write)
 664                         buf->len = PAGE_SIZE - cs->len;
 665                 cs->currbuf = NULL;
 666         } else if (cs->pg) {
 667                 if (cs->write) {
 668                         flush_dcache_page(cs->pg);
 669                         set_page_dirty_lock(cs->pg);
 670                 }
 671                 put_page(cs->pg);
 672         }
 673         cs->pg = NULL;
 674 }
 675
 676 /*
 677  * Get another pagefull of userspace buffer, and map it to kernel
 678  * address space, and lock request
 679  */
 680 static int fuse_copy_fill(struct fuse_copy_state *cs)
 681 {
 682         struct page *page;
 683         int err;
 684
 685         err = unlock_request(cs->req);
 686         if (err)
 687                 return err;
 688
 689         fuse_copy_finish(cs);
 690         if (cs->pipebufs) {
 691                 struct pipe_buffer *buf = cs->pipebufs;
 692
 693                 if (!cs->write) {
 694                         err = pipe_buf_confirm(cs->pipe, buf);
 695                         if (err)
 696                                 return err;
 697
 698                         BUG_ON(!cs->nr_segs);
 699                         cs->currbuf = buf;
 700                         cs->pg = buf->page;
 701                         cs->offset = buf->offset;
 702                         cs->len = buf->len;
 703                         cs->pipebufs++;
 704                         cs->nr_segs--;
 705                 } else {
 706                         if (cs->nr_segs == cs->pipe->buffers)
 707                                 return -EIO;
 708
 709                         page = alloc_page(GFP_HIGHUSER);
 710                         if (!page)
 711                                 return -ENOMEM;
 712
 713                         buf->page = page;
 714                         buf->offset = 0;
 715                         buf->len = 0;
 716
 717                         cs->currbuf = buf;
 718                         cs->pg = page;
 719                         cs->offset = 0;
 720                         cs->len = PAGE_SIZE;
 721                         cs->pipebufs++;
 722                         cs->nr_segs++;
 723                 }
 724         } else {
 725                 size_t off;
 726                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 727                 if (err < 0)
 728                         return err;
 729                 BUG_ON(!err);
 730                 cs->len = err;
 731                 cs->offset = off;
 732                 cs->pg = page;
 733                 iov_iter_advance(cs->iter, err);
 734         }
 735
 736         return lock_request(cs->req);
 737 }
 738
 739 /* Do as much copy to/from userspace buffer as we can */
 740 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 741 {
 742         unsigned ncpy = min(*size, cs->len);
 743         if (val) {
 744                 void *pgaddr = kmap_atomic(cs->pg);
 745                 void *buf = pgaddr + cs->offset;
 746
 747                 if (cs->write)
 748                         memcpy(buf, *val, ncpy);
 749                 else
 750                         memcpy(*val, buf, ncpy);
 751
 752                 kunmap_atomic(pgaddr);
 753                 *val += ncpy;
 754         }
 755         *size -= ncpy;
 756         cs->len -= ncpy;
 757         cs->offset += ncpy;
 758         return ncpy;
 759 }
 760
 761 static int fuse_check_page(struct page *page)
 762 {
 763         if (page_mapcount(page) ||
 764             page->mapping != NULL ||
 765             page_count(page) != 1 ||
 766             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 767              ~(1 << PG_locked |
 768                1 << PG_referenced |
 769                1 << PG_uptodate |
 770                1 << PG_lru |
 771                1 << PG_active |
 772                1 << PG_reclaim))) {
 773                 pr_warn("trying to steal weird page\n");
 774                 pr_warn("  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);
 775                 return 1;
 776         }
 777         return 0;
 778 }
 779
 780 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 781 {
 782         int err;
 783         struct page *oldpage = *pagep;
 784         struct page *newpage;
 785         struct pipe_buffer *buf = cs->pipebufs;
 786
 787         err = unlock_request(cs->req);
 788         if (err)
 789                 return err;
 790
 791         fuse_copy_finish(cs);
 792
 793         err = pipe_buf_confirm(cs->pipe, buf);
 794         if (err)
 795                 return err;
 796
 797         BUG_ON(!cs->nr_segs);
 798         cs->currbuf = buf;
 799         cs->len = buf->len;
 800         cs->pipebufs++;
 801         cs->nr_segs--;
 802
 803         if (cs->len != PAGE_SIZE)
 804                 goto out_fallback;
 805
 806         if (pipe_buf_steal(cs->pipe, buf) != 0)
 807                 goto out_fallback;
 808
 809         newpage = buf->page;
 810
 811         if (!PageUptodate(newpage))
 812                 SetPageUptodate(newpage);
 813
 814         ClearPageMappedToDisk(newpage);
 815
 816         if (fuse_check_page(newpage) != 0)
 817                 goto out_fallback_unlock;
 818
 819         /*
 820          * This is a new and locked page, it shouldn't be mapped or
 821          * have any special flags on it
 822          */
 823         if (WARN_ON(page_mapped(oldpage)))
 824                 goto out_fallback_unlock;
 825         if (WARN_ON(page_has_private(oldpage)))
 826                 goto out_fallback_unlock;
 827         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 828                 goto out_fallback_unlock;
 829         if (WARN_ON(PageMlocked(oldpage)))
 830                 goto out_fallback_unlock;
 831
 832         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 833         if (err) {
 834                 unlock_page(newpage);
 835                 return err;
 836         }
 837
 838         get_page(newpage);
 839
 840         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 841                 lru_cache_add_file(newpage);
 842
 843         err = 0;
 844         spin_lock(&cs->req->waitq.lock);
 845         if (test_bit(FR_ABORTED, &cs->req->flags))
 846                 err = -ENOENT;
 847         else
 848                 *pagep = newpage;
 849         spin_unlock(&cs->req->waitq.lock);
 850
 851         if (err) {
 852                 unlock_page(newpage);
 853                 put_page(newpage);
 854                 return err;
 855         }
 856
 857         unlock_page(oldpage);
 858         put_page(oldpage);
 859         cs->len = 0;
 860
 861         return 0;
 862
 863 out_fallback_unlock:
 864         unlock_page(newpage);
 865 out_fallback:
 866         cs->pg = buf->page;
 867         cs->offset = buf->offset;
 868
 869         err = lock_request(cs->req);
 870         if (err)
 871                 return err;
 872
 873         return 1;
 874 }
 875
 876 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 877                          unsigned offset, unsigned count)
 878 {
 879         struct pipe_buffer *buf;
 880         int err;
 881
 882         if (cs->nr_segs == cs->pipe->buffers)
 883                 return -EIO;
 884
 885         err = unlock_request(cs->req);
 886         if (err)
 887                 return err;
 888
 889         fuse_copy_finish(cs);
 890
 891         buf = cs->pipebufs;
 892         get_page(page);
 893         buf->page = page;
 894         buf->offset = offset;
 895         buf->len = count;
 896
 897         cs->pipebufs++;
 898         cs->nr_segs++;
 899         cs->len = 0;
 900
 901         return 0;
 902 }
 903
 904 /*
 905  * Copy a page in the request to/from the userspace buffer.  Must be
 906  * done atomically
 907  */
 908 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 909                           unsigned offset, unsigned count, int zeroing)
 910 {
 911         int err;
 912         struct page *page = *pagep;
 913
 914         if (page && zeroing && count < PAGE_SIZE)
 915                 clear_highpage(page);
 916
 917         while (count) {
 918                 if (cs->write && cs->pipebufs && page) {
 919                         return fuse_ref_page(cs, page, offset, count);
 920                 } else if (!cs->len) {
 921                         if (cs->move_pages && page &&
 922                             offset == 0 && count == PAGE_SIZE) {
 923                                 err = fuse_try_move_page(cs, pagep);
 924                                 if (err <= 0)
 925                                         return err;
 926                         } else {
 927                                 err = fuse_copy_fill(cs);
 928                                 if (err)
 929                                         return err;
 930                         }
 931                 }
 932                 if (page) {
 933                         void *mapaddr = kmap_atomic(page);
 934                         void *buf = mapaddr + offset;
 935                         offset += fuse_copy_do(cs, &buf, &count);
 936                         kunmap_atomic(mapaddr);
 937                 } else
 938                         offset += fuse_copy_do(cs, NULL, &count);
 939         }
 940         if (page && !cs->write)
 941                 flush_dcache_page(page);
 942         return 0;
 943 }
 944
 945 /* Copy pages in the request to/from userspace buffer */
 946 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 947                            int zeroing)
 948 {
 949         unsigned i;
 950         struct fuse_req *req = cs->req;
 951         struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
 952
 953
 954         for (i = 0; i < ap->num_pages && (nbytes || zeroing); i++) {
 955                 int err;
 956                 unsigned int offset = ap->descs[i].offset;
 957                 unsigned int count = min(nbytes, ap->descs[i].length);
 958
 959                 err = fuse_copy_page(cs, &ap->pages[i], offset, count, zeroing);
 960                 if (err)
 961                         return err;
 962
 963                 nbytes -= count;
 964         }
 965         return 0;
 966 }
 967
 968 /* Copy a single argument in the request to/from userspace buffer */
 969 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 970 {
 971         while (size) {
 972                 if (!cs->len) {
 973                         int err = fuse_copy_fill(cs);
 974                         if (err)
 975                                 return err;
 976                 }
 977                 fuse_copy_do(cs, &val, &size);
 978         }
 979         return 0;
 980 }
 981
 982 /* Copy request arguments to/from userspace buffer */
 983 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 984                           unsigned argpages, struct fuse_arg *args,
 985                           int zeroing)
 986 {
 987         int err = 0;
 988         unsigned i;
 989
 990         for (i = 0; !err && i < numargs; i++)  {
 991                 struct fuse_arg *arg = &args[i];
 992                 if (i == numargs - 1 && argpages)
 993                         err = fuse_copy_pages(cs, arg->size, zeroing);
 994                 else
 995                         err = fuse_copy_one(cs, arg->value, arg->size);
 996         }
 997         return err;
 998 }
 999
1000 static int forget_pending(struct fuse_iqueue *fiq)
1001 {
1002         return fiq->forget_list_head.next != NULL;
1003 }
1004
1005 static int request_pending(struct fuse_iqueue *fiq)
1006 {
1007         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1008                 forget_pending(fiq);
1009 }
1010
1011 /*
1012  * Transfer an interrupt request to userspace
1013  *
1014  * Unlike other requests this is assembled on demand, without a need
1015  * to allocate a separate fuse_req structure.
1016  *
1017  * Called with fiq->lock held, releases it
1018  */
1019 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1020                                struct fuse_copy_state *cs,
1021                                size_t nbytes, struct fuse_req *req)
1022 __releases(fiq->lock)
1023 {
1024         struct fuse_in_header ih;
1025         struct fuse_interrupt_in arg;
1026         unsigned reqsize = sizeof(ih) + sizeof(arg);
1027         int err;
1028
1029         list_del_init(&req->intr_entry);
1030         memset(&ih, 0, sizeof(ih));
1031         memset(&arg, 0, sizeof(arg));
1032         ih.len = reqsize;
1033         ih.opcode = FUSE_INTERRUPT;
1034         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1035         arg.unique = req->in.h.unique;
1036
1037         spin_unlock(&fiq->lock);
1038         if (nbytes < reqsize)
1039                 return -EINVAL;
1040
1041         err = fuse_copy_one(cs, &ih, sizeof(ih));
1042         if (!err)
1043                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1044         fuse_copy_finish(cs);
1045
1046         return err ? err : reqsize;
1047 }
1048
1049 struct fuse_forget_link *fuse_dequeue_forget(struct fuse_iqueue *fiq,
1050                                              unsigned int max,
1051                                              unsigned int *countp)
1052 {
1053         struct fuse_forget_link *head = fiq->forget_list_head.next;
1054         struct fuse_forget_link **newhead = &head;
1055         unsigned count;
1056
1057         for (count = 0; *newhead != NULL && count < max; count++)
1058                 newhead = &(*newhead)->next;
1059
1060         fiq->forget_list_head.next = *newhead;
1061         *newhead = NULL;
1062         if (fiq->forget_list_head.next == NULL)
1063                 fiq->forget_list_tail = &fiq->forget_list_head;
1064
1065         if (countp != NULL)
1066                 *countp = count;
1067
1068         return head;
1069 }
1070 EXPORT_SYMBOL(fuse_dequeue_forget);
1071
1072 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1073                                    struct fuse_copy_state *cs,
1074                                    size_t nbytes)
1075 __releases(fiq->lock)
1076 {
1077         int err;
1078         struct fuse_forget_link *forget = fuse_dequeue_forget(fiq, 1, NULL);
1079         struct fuse_forget_in arg = {
1080                 .nlookup = forget->forget_one.nlookup,
1081         };
1082         struct fuse_in_header ih = {
1083                 .opcode = FUSE_FORGET,
1084                 .nodeid = forget->forget_one.nodeid,
1085                 .unique = fuse_get_unique(fiq),
1086                 .len = sizeof(ih) + sizeof(arg),
1087         };
1088
1089         spin_unlock(&fiq->lock);
1090         kfree(forget);
1091         if (nbytes < ih.len)
1092                 return -EINVAL;
1093
1094         err = fuse_copy_one(cs, &ih, sizeof(ih));
1095         if (!err)
1096                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1097         fuse_copy_finish(cs);
1098
1099         if (err)
1100                 return err;
1101
1102         return ih.len;
1103 }
1104
1105 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1106                                    struct fuse_copy_state *cs, size_t nbytes)
1107 __releases(fiq->lock)
1108 {
1109         int err;
1110         unsigned max_forgets;
1111         unsigned count;
1112         struct fuse_forget_link *head;
1113         struct fuse_batch_forget_in arg = { .count = 0 };
1114         struct fuse_in_header ih = {
1115                 .opcode = FUSE_BATCH_FORGET,
1116                 .unique = fuse_get_unique(fiq),
1117                 .len = sizeof(ih) + sizeof(arg),
1118         };
1119
1120         if (nbytes < ih.len) {
1121                 spin_unlock(&fiq->lock);
1122                 return -EINVAL;
1123         }
1124
1125         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1126         head = fuse_dequeue_forget(fiq, max_forgets, &count);
1127         spin_unlock(&fiq->lock);
1128
1129         arg.count = count;
1130         ih.len += count * sizeof(struct fuse_forget_one);
1131         err = fuse_copy_one(cs, &ih, sizeof(ih));
1132         if (!err)
1133                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1134
1135         while (head) {
1136                 struct fuse_forget_link *forget = head;
1137
1138                 if (!err) {
1139                         err = fuse_copy_one(cs, &forget->forget_one,
1140                                             sizeof(forget->forget_one));
1141                 }
1142                 head = forget->next;
1143                 kfree(forget);
1144         }
1145
1146         fuse_copy_finish(cs);
1147
1148         if (err)
1149                 return err;
1150
1151         return ih.len;
1152 }
1153
1154 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1155                             struct fuse_copy_state *cs,
1156                             size_t nbytes)
1157 __releases(fiq->lock)
1158 {
1159         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1160                 return fuse_read_single_forget(fiq, cs, nbytes);
1161         else
1162                 return fuse_read_batch_forget(fiq, cs, nbytes);
1163 }
1164
1165 /*
1166  * Read a single request into the userspace filesystem's buffer.  This
1167  * function waits until a request is available, then removes it from
1168  * the pending list and copies request data to userspace buffer.  If
1169  * no reply is needed (FORGET) or request has been aborted or there
1170  * was an error during the copying then it's finished by calling
1171  * fuse_request_end().  Otherwise add it to the processing list, and set
1172  * the 'sent' flag.
1173  */
1174 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1175                                 struct fuse_copy_state *cs, size_t nbytes)
1176 {
1177         ssize_t err;
1178         struct fuse_conn *fc = fud->fc;
1179         struct fuse_iqueue *fiq = &fc->iq;
1180         struct fuse_pqueue *fpq = &fud->pq;
1181         struct fuse_req *req;
1182         struct fuse_args *args;
1183         unsigned reqsize;
1184         unsigned int hash;
1185
1186         /*
1187          * Require sane minimum read buffer - that has capacity for fixed part
1188          * of any request header + negotiated max_write room for data.
1189          *
1190          * Historically libfuse reserves 4K for fixed header room, but e.g.
1191          * GlusterFS reserves only 80 bytes
1192          *
1193          *      = `sizeof(fuse_in_header) + sizeof(fuse_write_in)`
1194          *
1195          * which is the absolute minimum any sane filesystem should be using
1196          * for header room.
1197          */
1198         if (nbytes < max_t(size_t, FUSE_MIN_READ_BUFFER,
1199                            sizeof(struct fuse_in_header) +
1200                            sizeof(struct fuse_write_in) +
1201                            fc->max_write))
1202                 return -EINVAL;
1203
1204  restart:
1205         for (;;) {
1206                 spin_lock(&fiq->lock);
1207                 if (!fiq->connected || request_pending(fiq))
1208                         break;
1209                 spin_unlock(&fiq->lock);
1210
1211                 if (file->f_flags & O_NONBLOCK)
1212                         return -EAGAIN;
1213                 err = wait_event_interruptible_exclusive(fiq->waitq,
1214                                 !fiq->connected || request_pending(fiq));
1215                 if (err)
1216                         return err;
1217         }
1218
1219         if (!fiq->connected) {
1220                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1221                 goto err_unlock;
1222         }
1223
1224         if (!list_empty(&fiq->interrupts)) {
1225                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1226                                  intr_entry);
1227                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1228         }
1229
1230         if (forget_pending(fiq)) {
1231                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1232                         return fuse_read_forget(fc, fiq, cs, nbytes);
1233
1234                 if (fiq->forget_batch <= -8)
1235                         fiq->forget_batch = 16;
1236         }
1237
1238         req = list_entry(fiq->pending.next, struct fuse_req, list);
1239         clear_bit(FR_PENDING, &req->flags);
1240         list_del_init(&req->list);
1241         spin_unlock(&fiq->lock);
1242
1243         args = req->args;
1244         reqsize = req->in.h.len;
1245
1246         /* If request is too large, reply with an error and restart the read */
1247         if (nbytes < reqsize) {
1248                 req->out.h.error = -EIO;
1249                 /* SETXATTR is special, since it may contain too large data */
1250                 if (args->opcode == FUSE_SETXATTR)
1251                         req->out.h.error = -E2BIG;
1252                 fuse_request_end(fc, req);
1253                 goto restart;
1254         }
1255         spin_lock(&fpq->lock);
1256         list_add(&req->list, &fpq->io);
1257         spin_unlock(&fpq->lock);
1258         cs->req = req;
1259         err = fuse_copy_one(cs, &req->in.h, sizeof(req->in.h));
1260         if (!err)
1261                 err = fuse_copy_args(cs, args->in_numargs, args->in_pages,
1262                                      (struct fuse_arg *) args->in_args, 0);
1263         fuse_copy_finish(cs);
1264         spin_lock(&fpq->lock);
1265         clear_bit(FR_LOCKED, &req->flags);
1266         if (!fpq->connected) {
1267                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1268                 goto out_end;
1269         }
1270         if (err) {
1271                 req->out.h.error = -EIO;
1272                 goto out_end;
1273         }
1274         if (!test_bit(FR_ISREPLY, &req->flags)) {
1275                 err = reqsize;
1276                 goto out_end;
1277         }
1278         hash = fuse_req_hash(req->in.h.unique);
1279         list_move_tail(&req->list, &fpq->processing[hash]);
1280         __fuse_get_request(req);
1281         set_bit(FR_SENT, &req->flags);
1282         spin_unlock(&fpq->lock);
1283         /* matches barrier in request_wait_answer() */
1284         smp_mb__after_atomic();
1285         if (test_bit(FR_INTERRUPTED, &req->flags))
1286                 queue_interrupt(fiq, req);
1287         fuse_put_request(fc, req);
1288
1289         return reqsize;
1290
1291 out_end:
1292         if (!test_bit(FR_PRIVATE, &req->flags))
1293                 list_del_init(&req->list);
1294         spin_unlock(&fpq->lock);
1295         fuse_request_end(fc, req);
1296         return err;
1297
1298  err_unlock:
1299         spin_unlock(&fiq->lock);
1300         return err;
1301 }
1302
1303 static int fuse_dev_open(struct inode *inode, struct file *file)
1304 {
1305         /*
1306          * The fuse device's file's private_data is used to hold
1307          * the fuse_conn(ection) when it is mounted, and is used to
1308          * keep track of whether the file has been mounted already.
1309          */
1310         file->private_data = NULL;
1311         return 0;
1312 }
1313
1314 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1315 {
1316         struct fuse_copy_state cs;
1317         struct file *file = iocb->ki_filp;
1318         struct fuse_dev *fud = fuse_get_dev(file);
1319
1320         if (!fud)
1321                 return -EPERM;
1322
1323         if (!iter_is_iovec(to))
1324                 return -EINVAL;
1325
1326         fuse_copy_init(&cs, 1, to);
1327
1328         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1329 }
1330
1331 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1332                                     struct pipe_inode_info *pipe,
1333                                     size_t len, unsigned int flags)
1334 {
1335         int total, ret;
1336         int page_nr = 0;
1337         struct pipe_buffer *bufs;
1338         struct fuse_copy_state cs;
1339         struct fuse_dev *fud = fuse_get_dev(in);
1340
1341         if (!fud)
1342                 return -EPERM;
1343
1344         bufs = kvmalloc_array(pipe->buffers, sizeof(struct pipe_buffer),
1345                               GFP_KERNEL);
1346         if (!bufs)
1347                 return -ENOMEM;
1348
1349         fuse_copy_init(&cs, 1, NULL);
1350         cs.pipebufs = bufs;
1351         cs.pipe = pipe;
1352         ret = fuse_dev_do_read(fud, in, &cs, len);
1353         if (ret < 0)
1354                 goto out;
1355
1356         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1357                 ret = -EIO;
1358                 goto out;
1359         }
1360
1361         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1362                 /*
1363                  * Need to be careful about this.  Having buf->ops in module
1364                  * code can Oops if the buffer persists after module unload.
1365                  */
1366                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1367                 bufs[page_nr].flags = 0;
1368                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1369                 if (unlikely(ret < 0))
1370                         break;
1371         }
1372         if (total)
1373                 ret = total;
1374 out:
1375         for (; page_nr < cs.nr_segs; page_nr++)
1376                 put_page(bufs[page_nr].page);
1377
1378         kvfree(bufs);
1379         return ret;
1380 }
1381
1382 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1383                             struct fuse_copy_state *cs)
1384 {
1385         struct fuse_notify_poll_wakeup_out outarg;
1386         int err = -EINVAL;
1387
1388         if (size != sizeof(outarg))
1389                 goto err;
1390
1391         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1392         if (err)
1393                 goto err;
1394
1395         fuse_copy_finish(cs);
1396         return fuse_notify_poll_wakeup(fc, &outarg);
1397
1398 err:
1399         fuse_copy_finish(cs);
1400         return err;
1401 }
1402
1403 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1404                                    struct fuse_copy_state *cs)
1405 {
1406         struct fuse_notify_inval_inode_out outarg;
1407         int err = -EINVAL;
1408
1409         if (size != sizeof(outarg))
1410                 goto err;
1411
1412         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1413         if (err)
1414                 goto err;
1415         fuse_copy_finish(cs);
1416
1417         down_read(&fc->killsb);
1418         err = -ENOENT;
1419         if (fc->sb) {
1420                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1421                                                outarg.off, outarg.len);
1422         }
1423         up_read(&fc->killsb);
1424         return err;
1425
1426 err:
1427         fuse_copy_finish(cs);
1428         return err;
1429 }
1430
1431 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1432                                    struct fuse_copy_state *cs)
1433 {
1434         struct fuse_notify_inval_entry_out outarg;
1435         int err = -ENOMEM;
1436         char *buf;
1437         struct qstr name;
1438
1439         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1440         if (!buf)
1441                 goto err;
1442
1443         err = -EINVAL;
1444         if (size < sizeof(outarg))
1445                 goto err;
1446
1447         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1448         if (err)
1449                 goto err;
1450
1451         err = -ENAMETOOLONG;
1452         if (outarg.namelen > FUSE_NAME_MAX)
1453                 goto err;
1454
1455         err = -EINVAL;
1456         if (size != sizeof(outarg) + outarg.namelen + 1)
1457                 goto err;
1458
1459         name.name = buf;
1460         name.len = outarg.namelen;
1461         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1462         if (err)
1463                 goto err;
1464         fuse_copy_finish(cs);
1465         buf[outarg.namelen] = 0;
1466
1467         down_read(&fc->killsb);
1468         err = -ENOENT;
1469         if (fc->sb)
1470                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1471         up_read(&fc->killsb);
1472         kfree(buf);
1473         return err;
1474
1475 err:
1476         kfree(buf);
1477         fuse_copy_finish(cs);
1478         return err;
1479 }
1480
1481 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1482                               struct fuse_copy_state *cs)
1483 {
1484         struct fuse_notify_delete_out outarg;
1485         int err = -ENOMEM;
1486         char *buf;
1487         struct qstr name;
1488
1489         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1490         if (!buf)
1491                 goto err;
1492
1493         err = -EINVAL;
1494         if (size < sizeof(outarg))
1495                 goto err;
1496
1497         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1498         if (err)
1499                 goto err;
1500
1501         err = -ENAMETOOLONG;
1502         if (outarg.namelen > FUSE_NAME_MAX)
1503                 goto err;
1504
1505         err = -EINVAL;
1506         if (size != sizeof(outarg) + outarg.namelen + 1)
1507                 goto err;
1508
1509         name.name = buf;
1510         name.len = outarg.namelen;
1511         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1512         if (err)
1513                 goto err;
1514         fuse_copy_finish(cs);
1515         buf[outarg.namelen] = 0;
1516
1517         down_read(&fc->killsb);
1518         err = -ENOENT;
1519         if (fc->sb)
1520                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1521                                                outarg.child, &name);
1522         up_read(&fc->killsb);
1523         kfree(buf);
1524         return err;
1525
1526 err:
1527         kfree(buf);
1528         fuse_copy_finish(cs);
1529         return err;
1530 }
1531
1532 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1533                              struct fuse_copy_state *cs)
1534 {
1535         struct fuse_notify_store_out outarg;
1536         struct inode *inode;
1537         struct address_space *mapping;
1538         u64 nodeid;
1539         int err;
1540         pgoff_t index;
1541         unsigned int offset;
1542         unsigned int num;
1543         loff_t file_size;
1544         loff_t end;
1545
1546         err = -EINVAL;
1547         if (size < sizeof(outarg))
1548                 goto out_finish;
1549
1550         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1551         if (err)
1552                 goto out_finish;
1553
1554         err = -EINVAL;
1555         if (size - sizeof(outarg) != outarg.size)
1556                 goto out_finish;
1557
1558         nodeid = outarg.nodeid;
1559
1560         down_read(&fc->killsb);
1561
1562         err = -ENOENT;
1563         if (!fc->sb)
1564                 goto out_up_killsb;
1565
1566         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1567         if (!inode)
1568                 goto out_up_killsb;
1569
1570         mapping = inode->i_mapping;
1571         index = outarg.offset >> PAGE_SHIFT;
1572         offset = outarg.offset & ~PAGE_MASK;
1573         file_size = i_size_read(inode);
1574         end = outarg.offset + outarg.size;
1575         if (end > file_size) {
1576                 file_size = end;
1577                 fuse_write_update_size(inode, file_size);
1578         }
1579
1580         num = outarg.size;
1581         while (num) {
1582                 struct page *page;
1583                 unsigned int this_num;
1584
1585                 err = -ENOMEM;
1586                 page = find_or_create_page(mapping, index,
1587                                            mapping_gfp_mask(mapping));
1588                 if (!page)
1589                         goto out_iput;
1590
1591                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1592                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1593                 if (!err && offset == 0 &&
1594                     (this_num == PAGE_SIZE || file_size == end))
1595                         SetPageUptodate(page);
1596                 unlock_page(page);
1597                 put_page(page);
1598
1599                 if (err)
1600                         goto out_iput;
1601
1602                 num -= this_num;
1603                 offset = 0;
1604                 index++;
1605         }
1606
1607         err = 0;
1608
1609 out_iput:
1610         iput(inode);
1611 out_up_killsb:
1612         up_read(&fc->killsb);
1613 out_finish:
1614         fuse_copy_finish(cs);
1615         return err;
1616 }
1617
1618 struct fuse_retrieve_args {
1619         struct fuse_args_pages ap;
1620         struct fuse_notify_retrieve_in inarg;
1621 };
1622
1623 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_args *args,
1624                               int error)
1625 {
1626         struct fuse_retrieve_args *ra =
1627                 container_of(args, typeof(*ra), ap.args);
1628
1629         release_pages(ra->ap.pages, ra->ap.num_pages);
1630         kfree(ra);
1631 }
1632
1633 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1634                          struct fuse_notify_retrieve_out *outarg)
1635 {
1636         int err;
1637         struct address_space *mapping = inode->i_mapping;
1638         pgoff_t index;
1639         loff_t file_size;
1640         unsigned int num;
1641         unsigned int offset;
1642         size_t total_len = 0;
1643         unsigned int num_pages;
1644         struct fuse_retrieve_args *ra;
1645         size_t args_size = sizeof(*ra);
1646         struct fuse_args_pages *ap;
1647         struct fuse_args *args;
1648
1649         offset = outarg->offset & ~PAGE_MASK;
1650         file_size = i_size_read(inode);
1651
1652         num = min(outarg->size, fc->max_write);
1653         if (outarg->offset > file_size)
1654                 num = 0;
1655         else if (outarg->offset + num > file_size)
1656                 num = file_size - outarg->offset;
1657
1658         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1659         num_pages = min(num_pages, fc->max_pages);
1660
1661         args_size += num_pages * (sizeof(ap->pages[0]) + sizeof(ap->descs[0]));
1662
1663         ra = kzalloc(args_size, GFP_KERNEL);
1664         if (!ra)
1665                 return -ENOMEM;
1666
1667         ap = &ra->ap;
1668         ap->pages = (void *) (ra + 1);
1669         ap->descs = (void *) (ap->pages + num_pages);
1670
1671         args = &ap->args;
1672         args->nodeid = outarg->nodeid;
1673         args->opcode = FUSE_NOTIFY_REPLY;
1674         args->in_numargs = 2;
1675         args->in_pages = true;
1676         args->end = fuse_retrieve_end;
1677
1678         index = outarg->offset >> PAGE_SHIFT;
1679
1680         while (num && ap->num_pages < num_pages) {
1681                 struct page *page;
1682                 unsigned int this_num;
1683
1684                 page = find_get_page(mapping, index);
1685                 if (!page)
1686                         break;
1687
1688                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1689                 ap->pages[ap->num_pages] = page;
1690                 ap->descs[ap->num_pages].offset = offset;
1691                 ap->descs[ap->num_pages].length = this_num;
1692                 ap->num_pages++;
1693
1694                 offset = 0;
1695                 num -= this_num;
1696                 total_len += this_num;
1697                 index++;
1698         }
1699         ra->inarg.offset = outarg->offset;
1700         ra->inarg.size = total_len;
1701         args->in_args[0].size = sizeof(ra->inarg);
1702         args->in_args[0].value = &ra->inarg;
1703         args->in_args[1].size = total_len;
1704
1705         err = fuse_simple_notify_reply(fc, args, outarg->notify_unique);
1706         if (err)
1707                 fuse_retrieve_end(fc, args, err);
1708
1709         return err;
1710 }
1711
1712 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1713                                 struct fuse_copy_state *cs)
1714 {
1715         struct fuse_notify_retrieve_out outarg;
1716         struct inode *inode;
1717         int err;
1718
1719         err = -EINVAL;
1720         if (size != sizeof(outarg))
1721                 goto copy_finish;
1722
1723         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1724         if (err)
1725                 goto copy_finish;
1726
1727         fuse_copy_finish(cs);
1728
1729         down_read(&fc->killsb);
1730         err = -ENOENT;
1731         if (fc->sb) {
1732                 u64 nodeid = outarg.nodeid;
1733
1734                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1735                 if (inode) {
1736                         err = fuse_retrieve(fc, inode, &outarg);
1737                         iput(inode);
1738                 }
1739         }
1740         up_read(&fc->killsb);
1741
1742         return err;
1743
1744 copy_finish:
1745         fuse_copy_finish(cs);
1746         return err;
1747 }
1748
1749 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1750                        unsigned int size, struct fuse_copy_state *cs)
1751 {
1752         /* Don't try to move pages (yet) */
1753         cs->move_pages = 0;
1754
1755         switch (code) {
1756         case FUSE_NOTIFY_POLL:
1757                 return fuse_notify_poll(fc, size, cs);
1758
1759         case FUSE_NOTIFY_INVAL_INODE:
1760                 return fuse_notify_inval_inode(fc, size, cs);
1761
1762         case FUSE_NOTIFY_INVAL_ENTRY:
1763                 return fuse_notify_inval_entry(fc, size, cs);
1764
1765         case FUSE_NOTIFY_STORE:
1766                 return fuse_notify_store(fc, size, cs);
1767
1768         case FUSE_NOTIFY_RETRIEVE:
1769                 return fuse_notify_retrieve(fc, size, cs);
1770
1771         case FUSE_NOTIFY_DELETE:
1772                 return fuse_notify_delete(fc, size, cs);
1773
1774         default:
1775                 fuse_copy_finish(cs);
1776                 return -EINVAL;
1777         }
1778 }
1779
1780 /* Look up request on processing list by unique ID */
1781 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1782 {
1783         unsigned int hash = fuse_req_hash(unique);
1784         struct fuse_req *req;
1785
1786         list_for_each_entry(req, &fpq->processing[hash], list) {
1787                 if (req->in.h.unique == unique)
1788                         return req;
1789         }
1790         return NULL;
1791 }
1792
1793 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_args *args,
1794                          unsigned nbytes)
1795 {
1796         unsigned reqsize = sizeof(struct fuse_out_header);
1797
1798         reqsize += fuse_len_args(args->out_numargs, args->out_args);
1799
1800         if (reqsize < nbytes || (reqsize > nbytes && !args->out_argvar))
1801                 return -EINVAL;
1802         else if (reqsize > nbytes) {
1803                 struct fuse_arg *lastarg = &args->out_args[args->out_numargs-1];
1804                 unsigned diffsize = reqsize - nbytes;
1805
1806                 if (diffsize > lastarg->size)
1807                         return -EINVAL;
1808                 lastarg->size -= diffsize;
1809         }
1810         return fuse_copy_args(cs, args->out_numargs, args->out_pages,
1811                               args->out_args, args->page_zeroing);
1812 }
1813
1814 /*
1815  * Write a single reply to a request.  First the header is copied from
1816  * the write buffer.  The request is then searched on the processing
1817  * list by the unique ID found in the header.  If found, then remove
1818  * it from the list and copy the rest of the buffer to the request.
1819  * The request is finished by calling fuse_request_end().
1820  */
1821 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1822                                  struct fuse_copy_state *cs, size_t nbytes)
1823 {
1824         int err;
1825         struct fuse_conn *fc = fud->fc;
1826         struct fuse_pqueue *fpq = &fud->pq;
1827         struct fuse_req *req;
1828         struct fuse_out_header oh;
1829
1830         err = -EINVAL;
1831         if (nbytes < sizeof(struct fuse_out_header))
1832                 goto out;
1833
1834         err = fuse_copy_one(cs, &oh, sizeof(oh));
1835         if (err)
1836                 goto copy_finish;
1837
1838         err = -EINVAL;
1839         if (oh.len != nbytes)
1840                 goto copy_finish;
1841
1842         /*
1843          * Zero oh.unique indicates unsolicited notification message
1844          * and error contains notification code.
1845          */
1846         if (!oh.unique) {
1847                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1848                 goto out;
1849         }
1850
1851         err = -EINVAL;
1852         if (oh.error <= -1000 || oh.error > 0)
1853                 goto copy_finish;
1854
1855         spin_lock(&fpq->lock);
1856         req = NULL;
1857         if (fpq->connected)
1858                 req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1859
1860         err = -ENOENT;
1861         if (!req) {
1862                 spin_unlock(&fpq->lock);
1863                 goto copy_finish;
1864         }
1865
1866         /* Is it an interrupt reply ID? */
1867         if (oh.unique & FUSE_INT_REQ_BIT) {
1868                 __fuse_get_request(req);
1869                 spin_unlock(&fpq->lock);
1870
1871                 err = 0;
1872                 if (nbytes != sizeof(struct fuse_out_header))
1873                         err = -EINVAL;
1874                 else if (oh.error == -ENOSYS)
1875                         fc->no_interrupt = 1;
1876                 else if (oh.error == -EAGAIN)
1877                         err = queue_interrupt(&fc->iq, req);
1878
1879                 fuse_put_request(fc, req);
1880
1881                 goto copy_finish;
1882         }
1883
1884         clear_bit(FR_SENT, &req->flags);
1885         list_move(&req->list, &fpq->io);
1886         req->out.h = oh;
1887         set_bit(FR_LOCKED, &req->flags);
1888         spin_unlock(&fpq->lock);
1889         cs->req = req;
1890         if (!req->args->page_replace)
1891                 cs->move_pages = 0;
1892
1893         if (oh.error)
1894                 err = nbytes != sizeof(oh) ? -EINVAL : 0;
1895         else
1896                 err = copy_out_args(cs, req->args, nbytes);
1897         fuse_copy_finish(cs);
1898
1899         spin_lock(&fpq->lock);
1900         clear_bit(FR_LOCKED, &req->flags);
1901         if (!fpq->connected)
1902                 err = -ENOENT;
1903         else if (err)
1904                 req->out.h.error = -EIO;
1905         if (!test_bit(FR_PRIVATE, &req->flags))
1906                 list_del_init(&req->list);
1907         spin_unlock(&fpq->lock);
1908
1909         fuse_request_end(fc, req);
1910 out:
1911         return err ? err : nbytes;
1912
1913 copy_finish:
1914         fuse_copy_finish(cs);
1915         goto out;
1916 }
1917
1918 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1919 {
1920         struct fuse_copy_state cs;
1921         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1922
1923         if (!fud)
1924                 return -EPERM;
1925
1926         if (!iter_is_iovec(from))
1927                 return -EINVAL;
1928
1929         fuse_copy_init(&cs, 0, from);
1930
1931         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1932 }
1933
1934 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1935                                      struct file *out, loff_t *ppos,
1936                                      size_t len, unsigned int flags)
1937 {
1938         unsigned nbuf;
1939         unsigned idx;
1940         struct pipe_buffer *bufs;
1941         struct fuse_copy_state cs;
1942         struct fuse_dev *fud;
1943         size_t rem;
1944         ssize_t ret;
1945
1946         fud = fuse_get_dev(out);
1947         if (!fud)
1948                 return -EPERM;
1949
1950         pipe_lock(pipe);
1951
1952         bufs = kvmalloc_array(pipe->nrbufs, sizeof(struct pipe_buffer),
1953                               GFP_KERNEL);
1954         if (!bufs) {
1955                 pipe_unlock(pipe);
1956                 return -ENOMEM;
1957         }
1958
1959         nbuf = 0;
1960         rem = 0;
1961         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1962                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1963
1964         ret = -EINVAL;
1965         if (rem < len)
1966                 goto out_free;
1967
1968         rem = len;
1969         while (rem) {
1970                 struct pipe_buffer *ibuf;
1971                 struct pipe_buffer *obuf;
1972
1973                 BUG_ON(nbuf >= pipe->buffers);
1974                 BUG_ON(!pipe->nrbufs);
1975                 ibuf = &pipe->bufs[pipe->curbuf];
1976                 obuf = &bufs[nbuf];
1977
1978                 if (rem >= ibuf->len) {
1979                         *obuf = *ibuf;
1980                         ibuf->ops = NULL;
1981                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1982                         pipe->nrbufs--;
1983                 } else {
1984                         if (!pipe_buf_get(pipe, ibuf))
1985                                 goto out_free;
1986
1987                         *obuf = *ibuf;
1988                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1989                         obuf->len = rem;
1990                         ibuf->offset += obuf->len;
1991                         ibuf->len -= obuf->len;
1992                 }
1993                 nbuf++;
1994                 rem -= obuf->len;
1995         }
1996         pipe_unlock(pipe);
1997
1998         fuse_copy_init(&cs, 0, NULL);
1999         cs.pipebufs = bufs;
2000         cs.nr_segs = nbuf;
2001         cs.pipe = pipe;
2002
2003         if (flags & SPLICE_F_MOVE)
2004                 cs.move_pages = 1;
2005
2006         ret = fuse_dev_do_write(fud, &cs, len);
2007
2008         pipe_lock(pipe);
2009 out_free:
2010         for (idx = 0; idx < nbuf; idx++)
2011                 pipe_buf_release(pipe, &bufs[idx]);
2012         pipe_unlock(pipe);
2013
2014         kvfree(bufs);
2015         return ret;
2016 }
2017
2018 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2019 {
2020         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2021         struct fuse_iqueue *fiq;
2022         struct fuse_dev *fud = fuse_get_dev(file);
2023
2024         if (!fud)
2025                 return EPOLLERR;
2026
2027         fiq = &fud->fc->iq;
2028         poll_wait(file, &fiq->waitq, wait);
2029
2030         spin_lock(&fiq->lock);
2031         if (!fiq->connected)
2032                 mask = EPOLLERR;
2033         else if (request_pending(fiq))
2034                 mask |= EPOLLIN | EPOLLRDNORM;
2035         spin_unlock(&fiq->lock);
2036
2037         return mask;
2038 }
2039
2040 /* Abort all requests on the given list (pending or processing) */
2041 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2042 {
2043         while (!list_empty(head)) {
2044                 struct fuse_req *req;
2045                 req = list_entry(head->next, struct fuse_req, list);
2046                 req->out.h.error = -ECONNABORTED;
2047                 clear_bit(FR_SENT, &req->flags);
2048                 list_del_init(&req->list);
2049                 fuse_request_end(fc, req);
2050         }
2051 }
2052
2053 static void end_polls(struct fuse_conn *fc)
2054 {
2055         struct rb_node *p;
2056
2057         p = rb_first(&fc->polled_files);
2058
2059         while (p) {
2060                 struct fuse_file *ff;
2061                 ff = rb_entry(p, struct fuse_file, polled_node);
2062                 wake_up_interruptible_all(&ff->poll_wait);
2063
2064                 p = rb_next(p);
2065         }
2066 }
2067
2068 /*
2069  * Abort all requests.
2070  *
2071  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2072  * filesystem.
2073  *
2074  * The same effect is usually achievable through killing the filesystem daemon
2075  * and all users of the filesystem.  The exception is the combination of an
2076  * asynchronous request and the tricky deadlock (see
2077  * Documentation/filesystems/fuse.txt).
2078  *
2079  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2080  * requests, they should be finished off immediately.  Locked requests will be
2081  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2082  * requests.  It is possible that some request will finish before we can.  This
2083  * is OK, the request will in that case be removed from the list before we touch
2084  * it.
2085  */
2086 void fuse_abort_conn(struct fuse_conn *fc)
2087 {
2088         struct fuse_iqueue *fiq = &fc->iq;
2089
2090         spin_lock(&fc->lock);
2091         if (fc->connected) {
2092                 struct fuse_dev *fud;
2093                 struct fuse_req *req, *next;
2094                 LIST_HEAD(to_end);
2095                 unsigned int i;
2096
2097                 /* Background queuing checks fc->connected under bg_lock */
2098                 spin_lock(&fc->bg_lock);
2099                 fc->connected = 0;
2100                 spin_unlock(&fc->bg_lock);
2101
2102                 fuse_set_initialized(fc);
2103                 list_for_each_entry(fud, &fc->devices, entry) {
2104                         struct fuse_pqueue *fpq = &fud->pq;
2105
2106                         spin_lock(&fpq->lock);
2107                         fpq->connected = 0;
2108                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2109                                 req->out.h.error = -ECONNABORTED;
2110                                 spin_lock(&req->waitq.lock);
2111                                 set_bit(FR_ABORTED, &req->flags);
2112                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2113                                         set_bit(FR_PRIVATE, &req->flags);
2114                                         __fuse_get_request(req);
2115                                         list_move(&req->list, &to_end);
2116                                 }
2117                                 spin_unlock(&req->waitq.lock);
2118                         }
2119                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2120                                 list_splice_tail_init(&fpq->processing[i],
2121                                                       &to_end);
2122                         spin_unlock(&fpq->lock);
2123                 }
2124                 spin_lock(&fc->bg_lock);
2125                 fc->blocked = 0;
2126                 fc->max_background = UINT_MAX;
2127                 flush_bg_queue(fc);
2128                 spin_unlock(&fc->bg_lock);
2129
2130                 spin_lock(&fiq->lock);
2131                 fiq->connected = 0;
2132                 list_for_each_entry(req, &fiq->pending, list)
2133                         clear_bit(FR_PENDING, &req->flags);
2134                 list_splice_tail_init(&fiq->pending, &to_end);
2135                 while (forget_pending(fiq))
2136                         kfree(fuse_dequeue_forget(fiq, 1, NULL));
2137                 wake_up_all(&fiq->waitq);
2138                 spin_unlock(&fiq->lock);
2139                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2140                 end_polls(fc);
2141                 wake_up_all(&fc->blocked_waitq);
2142                 spin_unlock(&fc->lock);
2143
2144                 end_requests(fc, &to_end);
2145         } else {
2146                 spin_unlock(&fc->lock);
2147         }
2148 }
2149 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2150
2151 void fuse_wait_aborted(struct fuse_conn *fc)
2152 {
2153         /* matches implicit memory barrier in fuse_drop_waiting() */
2154         smp_mb();
2155         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2156 }
2157
2158 int fuse_dev_release(struct inode *inode, struct file *file)
2159 {
2160         struct fuse_dev *fud = fuse_get_dev(file);
2161
2162         if (fud) {
2163                 struct fuse_conn *fc = fud->fc;
2164                 struct fuse_pqueue *fpq = &fud->pq;
2165                 LIST_HEAD(to_end);
2166                 unsigned int i;
2167
2168                 spin_lock(&fpq->lock);
2169                 WARN_ON(!list_empty(&fpq->io));
2170                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2171                         list_splice_init(&fpq->processing[i], &to_end);
2172                 spin_unlock(&fpq->lock);
2173
2174                 end_requests(fc, &to_end);
2175
2176                 /* Are we the last open device? */
2177                 if (atomic_dec_and_test(&fc->dev_count)) {
2178                         WARN_ON(fc->iq.fasync != NULL);
2179                         fuse_abort_conn(fc);
2180                 }
2181                 fuse_dev_free(fud);
2182         }
2183         return 0;
2184 }
2185 EXPORT_SYMBOL_GPL(fuse_dev_release);
2186
2187 static int fuse_dev_fasync(int fd, struct file *file, int on)
2188 {
2189         struct fuse_dev *fud = fuse_get_dev(file);
2190
2191         if (!fud)
2192                 return -EPERM;
2193
2194         /* No locking - fasync_helper does its own locking */
2195         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2196 }
2197
2198 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2199 {
2200         struct fuse_dev *fud;
2201
2202         if (new->private_data)
2203                 return -EINVAL;
2204
2205         fud = fuse_dev_alloc_install(fc);
2206         if (!fud)
2207                 return -ENOMEM;
2208
2209         new->private_data = fud;
2210         atomic_inc(&fc->dev_count);
2211
2212         return 0;
2213 }
2214
2215 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2216                            unsigned long arg)
2217 {
2218         int err = -ENOTTY;
2219
2220         if (cmd == FUSE_DEV_IOC_CLONE) {
2221                 int oldfd;
2222
2223                 err = -EFAULT;
2224                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2225                         struct file *old = fget(oldfd);
2226
2227                         err = -EINVAL;
2228                         if (old) {
2229                                 struct fuse_dev *fud = NULL;
2230
2231                                 /*
2232                                  * Check against file->f_op because CUSE
2233                                  * uses the same ioctl handler.
2234                                  */
2235                                 if (old->f_op == file->f_op &&
2236                                     old->f_cred->user_ns == file->f_cred->user_ns)
2237                                         fud = fuse_get_dev(old);
2238
2239                                 if (fud) {
2240                                         mutex_lock(&fuse_mutex);
2241                                         err = fuse_device_clone(fud->fc, file);
2242                                         mutex_unlock(&fuse_mutex);
2243                                 }
2244                                 fput(old);
2245                         }
2246                 }
2247         }
2248         return err;
2249 }
2250
2251 const struct file_operations fuse_dev_operations = {
2252         .owner          = THIS_MODULE,
2253         .open           = fuse_dev_open,
2254         .llseek         = no_llseek,
2255         .read_iter      = fuse_dev_read,
2256         .splice_read    = fuse_dev_splice_read,
2257         .write_iter     = fuse_dev_write,
2258         .splice_write   = fuse_dev_splice_write,
2259         .poll           = fuse_dev_poll,
2260         .release        = fuse_dev_release,
2261         .fasync         = fuse_dev_fasync,
2262         .unlocked_ioctl = fuse_dev_ioctl,
2263         .compat_ioctl   = fuse_dev_ioctl,
2264 };
2265 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2266
2267 static struct miscdevice fuse_miscdevice = {
2268         .minor = FUSE_MINOR,
2269         .name  = "fuse",
2270         .fops = &fuse_dev_operations,
2271 };
2272
2273 int __init fuse_dev_init(void)
2274 {
2275         int err = -ENOMEM;
2276         fuse_req_cachep = kmem_cache_create("fuse_request",
2277                                             sizeof(struct fuse_req),
2278                                             0, 0, NULL);
2279         if (!fuse_req_cachep)
2280                 goto out;
2281
2282         err = misc_register(&fuse_miscdevice);
2283         if (err)
2284                 goto out_cache_clean;
2285
2286         return 0;
2287
2288  out_cache_clean:
2289         kmem_cache_destroy(fuse_req_cachep);
2290  out:
2291         return err;
2292 }
2293
2294 void fuse_dev_cleanup(void)
2295 {
2296         misc_deregister(&fuse_miscdevice);
2297         kmem_cache_destroy(fuse_req_cachep);
2298 }