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
 280         if (test_and_set_bit(FR_FINISHED, &req->flags))
 281                 goto put_request;
 282
 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 (test_bit(FR_ASYNC, &req->flags))
 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 fuse_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         if (args->end)
 473                 __set_bit(FR_ASYNC, &req->flags);
 474 }
 475
 476 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 477 {
 478         struct fuse_req *req;
 479         ssize_t ret;
 480
 481         if (args->force) {
 482                 atomic_inc(&fc->num_waiting);
 483                 req = fuse_request_alloc(GFP_KERNEL | __GFP_NOFAIL);
 484
 485                 if (!args->nocreds)
 486                         fuse_force_creds(fc, req);
 487
 488                 __set_bit(FR_WAITING, &req->flags);
 489                 __set_bit(FR_FORCE, &req->flags);
 490         } else {
 491                 WARN_ON(args->nocreds);
 492                 req = fuse_get_req(fc, false);
 493                 if (IS_ERR(req))
 494                         return PTR_ERR(req);
 495         }
 496
 497         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 498         fuse_adjust_compat(fc, args);
 499         fuse_args_to_req(req, args);
 500
 501         if (!args->noreply)
 502                 __set_bit(FR_ISREPLY, &req->flags);
 503         __fuse_request_send(fc, req);
 504         ret = req->out.h.error;
 505         if (!ret && args->out_argvar) {
 506                 BUG_ON(args->out_numargs == 0);
 507                 ret = args->out_args[args->out_numargs - 1].size;
 508         }
 509         fuse_put_request(fc, req);
 510
 511         return ret;
 512 }
 513
 514 static bool fuse_request_queue_background(struct fuse_conn *fc,
 515                                           struct fuse_req *req)
 516 {
 517         bool queued = false;
 518
 519         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 520         if (!test_bit(FR_WAITING, &req->flags)) {
 521                 __set_bit(FR_WAITING, &req->flags);
 522                 atomic_inc(&fc->num_waiting);
 523         }
 524         __set_bit(FR_ISREPLY, &req->flags);
 525         spin_lock(&fc->bg_lock);
 526         if (likely(fc->connected)) {
 527                 fc->num_background++;
 528                 if (fc->num_background == fc->max_background)
 529                         fc->blocked = 1;
 530                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 531                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 532                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 533                 }
 534                 list_add_tail(&req->list, &fc->bg_queue);
 535                 flush_bg_queue(fc);
 536                 queued = true;
 537         }
 538         spin_unlock(&fc->bg_lock);
 539
 540         return queued;
 541 }
 542
 543 int fuse_simple_background(struct fuse_conn *fc, struct fuse_args *args,
 544                             gfp_t gfp_flags)
 545 {
 546         struct fuse_req *req;
 547
 548         if (args->force) {
 549                 WARN_ON(!args->nocreds);
 550                 req = fuse_request_alloc(gfp_flags);
 551                 if (!req)
 552                         return -ENOMEM;
 553                 __set_bit(FR_BACKGROUND, &req->flags);
 554         } else {
 555                 WARN_ON(args->nocreds);
 556                 req = fuse_get_req(fc, true);
 557                 if (IS_ERR(req))
 558                         return PTR_ERR(req);
 559         }
 560
 561         fuse_args_to_req(req, args);
 562
 563         if (!fuse_request_queue_background(fc, req)) {
 564                 fuse_put_request(fc, req);
 565                 return -ENOTCONN;
 566         }
 567
 568         return 0;
 569 }
 570 EXPORT_SYMBOL_GPL(fuse_simple_background);
 571
 572 static int fuse_simple_notify_reply(struct fuse_conn *fc,
 573                                     struct fuse_args *args, u64 unique)
 574 {
 575         struct fuse_req *req;
 576         struct fuse_iqueue *fiq = &fc->iq;
 577         int err = 0;
 578
 579         req = fuse_get_req(fc, false);
 580         if (IS_ERR(req))
 581                 return PTR_ERR(req);
 582
 583         __clear_bit(FR_ISREPLY, &req->flags);
 584         req->in.h.unique = unique;
 585
 586         fuse_args_to_req(req, args);
 587
 588         spin_lock(&fiq->lock);
 589         if (fiq->connected) {
 590                 queue_request_and_unlock(fiq, req);
 591         } else {
 592                 err = -ENODEV;
 593                 spin_unlock(&fiq->lock);
 594                 fuse_put_request(fc, req);
 595         }
 596
 597         return err;
 598 }
 599
 600 /*
 601  * Lock the request.  Up to the next unlock_request() there mustn't be
 602  * anything that could cause a page-fault.  If the request was already
 603  * aborted bail out.
 604  */
 605 static int lock_request(struct fuse_req *req)
 606 {
 607         int err = 0;
 608         if (req) {
 609                 spin_lock(&req->waitq.lock);
 610                 if (test_bit(FR_ABORTED, &req->flags))
 611                         err = -ENOENT;
 612                 else
 613                         set_bit(FR_LOCKED, &req->flags);
 614                 spin_unlock(&req->waitq.lock);
 615         }
 616         return err;
 617 }
 618
 619 /*
 620  * Unlock request.  If it was aborted while locked, caller is responsible
 621  * for unlocking and ending the request.
 622  */
 623 static int unlock_request(struct fuse_req *req)
 624 {
 625         int err = 0;
 626         if (req) {
 627                 spin_lock(&req->waitq.lock);
 628                 if (test_bit(FR_ABORTED, &req->flags))
 629                         err = -ENOENT;
 630                 else
 631                         clear_bit(FR_LOCKED, &req->flags);
 632                 spin_unlock(&req->waitq.lock);
 633         }
 634         return err;
 635 }
 636
 637 struct fuse_copy_state {
 638         int write;
 639         struct fuse_req *req;
 640         struct iov_iter *iter;
 641         struct pipe_buffer *pipebufs;
 642         struct pipe_buffer *currbuf;
 643         struct pipe_inode_info *pipe;
 644         unsigned long nr_segs;
 645         struct page *pg;
 646         unsigned len;
 647         unsigned offset;
 648         unsigned move_pages:1;
 649 };
 650
 651 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 652                            struct iov_iter *iter)
 653 {
 654         memset(cs, 0, sizeof(*cs));
 655         cs->write = write;
 656         cs->iter = iter;
 657 }
 658
 659 /* Unmap and put previous page of userspace buffer */
 660 static void fuse_copy_finish(struct fuse_copy_state *cs)
 661 {
 662         if (cs->currbuf) {
 663                 struct pipe_buffer *buf = cs->currbuf;
 664
 665                 if (cs->write)
 666                         buf->len = PAGE_SIZE - cs->len;
 667                 cs->currbuf = NULL;
 668         } else if (cs->pg) {
 669                 if (cs->write) {
 670                         flush_dcache_page(cs->pg);
 671                         set_page_dirty_lock(cs->pg);
 672                 }
 673                 put_page(cs->pg);
 674         }
 675         cs->pg = NULL;
 676 }
 677
 678 /*
 679  * Get another pagefull of userspace buffer, and map it to kernel
 680  * address space, and lock request
 681  */
 682 static int fuse_copy_fill(struct fuse_copy_state *cs)
 683 {
 684         struct page *page;
 685         int err;
 686
 687         err = unlock_request(cs->req);
 688         if (err)
 689                 return err;
 690
 691         fuse_copy_finish(cs);
 692         if (cs->pipebufs) {
 693                 struct pipe_buffer *buf = cs->pipebufs;
 694
 695                 if (!cs->write) {
 696                         err = pipe_buf_confirm(cs->pipe, buf);
 697                         if (err)
 698                                 return err;
 699
 700                         BUG_ON(!cs->nr_segs);
 701                         cs->currbuf = buf;
 702                         cs->pg = buf->page;
 703                         cs->offset = buf->offset;
 704                         cs->len = buf->len;
 705                         cs->pipebufs++;
 706                         cs->nr_segs--;
 707                 } else {
 708                         if (cs->nr_segs >= cs->pipe->max_usage)
 709                                 return -EIO;
 710
 711                         page = alloc_page(GFP_HIGHUSER);
 712                         if (!page)
 713                                 return -ENOMEM;
 714
 715                         buf->page = page;
 716                         buf->offset = 0;
 717                         buf->len = 0;
 718
 719                         cs->currbuf = buf;
 720                         cs->pg = page;
 721                         cs->offset = 0;
 722                         cs->len = PAGE_SIZE;
 723                         cs->pipebufs++;
 724                         cs->nr_segs++;
 725                 }
 726         } else {
 727                 size_t off;
 728                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 729                 if (err < 0)
 730                         return err;
 731                 BUG_ON(!err);
 732                 cs->len = err;
 733                 cs->offset = off;
 734                 cs->pg = page;
 735                 iov_iter_advance(cs->iter, err);
 736         }
 737
 738         return lock_request(cs->req);
 739 }
 740
 741 /* Do as much copy to/from userspace buffer as we can */
 742 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 743 {
 744         unsigned ncpy = min(*size, cs->len);
 745         if (val) {
 746                 void *pgaddr = kmap_atomic(cs->pg);
 747                 void *buf = pgaddr + cs->offset;
 748
 749                 if (cs->write)
 750                         memcpy(buf, *val, ncpy);
 751                 else
 752                         memcpy(*val, buf, ncpy);
 753
 754                 kunmap_atomic(pgaddr);
 755                 *val += ncpy;
 756         }
 757         *size -= ncpy;
 758         cs->len -= ncpy;
 759         cs->offset += ncpy;
 760         return ncpy;
 761 }
 762
 763 static int fuse_check_page(struct page *page)
 764 {
 765         if (page_mapcount(page) ||
 766             page->mapping != NULL ||
 767             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 768              ~(1 << PG_locked |
 769                1 << PG_referenced |
 770                1 << PG_uptodate |
 771                1 << PG_lru |
 772                1 << PG_active |
 773                1 << PG_reclaim |
 774                1 << PG_waiters))) {
 775                 dump_page(page, "fuse: trying to steal weird page");
 776                 return 1;
 777         }
 778         return 0;
 779 }
 780
 781 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 782 {
 783         int err;
 784         struct page *oldpage = *pagep;
 785         struct page *newpage;
 786         struct pipe_buffer *buf = cs->pipebufs;
 787
 788         err = unlock_request(cs->req);
 789         if (err)
 790                 return err;
 791
 792         fuse_copy_finish(cs);
 793
 794         err = pipe_buf_confirm(cs->pipe, buf);
 795         if (err)
 796                 return err;
 797
 798         BUG_ON(!cs->nr_segs);
 799         cs->currbuf = buf;
 800         cs->len = buf->len;
 801         cs->pipebufs++;
 802         cs->nr_segs--;
 803
 804         if (cs->len != PAGE_SIZE)
 805                 goto out_fallback;
 806
 807         if (!pipe_buf_try_steal(cs->pipe, buf))
 808                 goto out_fallback;
 809
 810         newpage = buf->page;
 811
 812         if (!PageUptodate(newpage))
 813                 SetPageUptodate(newpage);
 814
 815         ClearPageMappedToDisk(newpage);
 816
 817         if (fuse_check_page(newpage) != 0)
 818                 goto out_fallback_unlock;
 819
 820         /*
 821          * This is a new and locked page, it shouldn't be mapped or
 822          * have any special flags on it
 823          */
 824         if (WARN_ON(page_mapped(oldpage)))
 825                 goto out_fallback_unlock;
 826         if (WARN_ON(page_has_private(oldpage)))
 827                 goto out_fallback_unlock;
 828         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 829                 goto out_fallback_unlock;
 830         if (WARN_ON(PageMlocked(oldpage)))
 831                 goto out_fallback_unlock;
 832
 833         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 834         if (err) {
 835                 unlock_page(newpage);
 836                 return err;
 837         }
 838
 839         get_page(newpage);
 840
 841         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 842                 lru_cache_add(newpage);
 843
 844         err = 0;
 845         spin_lock(&cs->req->waitq.lock);
 846         if (test_bit(FR_ABORTED, &cs->req->flags))
 847                 err = -ENOENT;
 848         else
 849                 *pagep = newpage;
 850         spin_unlock(&cs->req->waitq.lock);
 851
 852         if (err) {
 853                 unlock_page(newpage);
 854                 put_page(newpage);
 855                 return err;
 856         }
 857
 858         unlock_page(oldpage);
 859         put_page(oldpage);
 860         cs->len = 0;
 861
 862         return 0;
 863
 864 out_fallback_unlock:
 865         unlock_page(newpage);
 866 out_fallback:
 867         cs->pg = buf->page;
 868         cs->offset = buf->offset;
 869
 870         err = lock_request(cs->req);
 871         if (err)
 872                 return err;
 873
 874         return 1;
 875 }
 876
 877 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 878                          unsigned offset, unsigned count)
 879 {
 880         struct pipe_buffer *buf;
 881         int err;
 882
 883         if (cs->nr_segs >= cs->pipe->max_usage)
 884                 return -EIO;
 885
 886         err = unlock_request(cs->req);
 887         if (err)
 888                 return err;
 889
 890         fuse_copy_finish(cs);
 891
 892         buf = cs->pipebufs;
 893         get_page(page);
 894         buf->page = page;
 895         buf->offset = offset;
 896         buf->len = count;
 897
 898         cs->pipebufs++;
 899         cs->nr_segs++;
 900         cs->len = 0;
 901
 902         return 0;
 903 }
 904
 905 /*
 906  * Copy a page in the request to/from the userspace buffer.  Must be
 907  * done atomically
 908  */
 909 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 910                           unsigned offset, unsigned count, int zeroing)
 911 {
 912         int err;
 913         struct page *page = *pagep;
 914
 915         if (page && zeroing && count < PAGE_SIZE)
 916                 clear_highpage(page);
 917
 918         while (count) {
 919                 if (cs->write && cs->pipebufs && page) {
 920                         return fuse_ref_page(cs, page, offset, count);
 921                 } else if (!cs->len) {
 922                         if (cs->move_pages && page &&
 923                             offset == 0 && count == PAGE_SIZE) {
 924                                 err = fuse_try_move_page(cs, pagep);
 925                                 if (err <= 0)
 926                                         return err;
 927                         } else {
 928                                 err = fuse_copy_fill(cs);
 929                                 if (err)
 930                                         return err;
 931                         }
 932                 }
 933                 if (page) {
 934                         void *mapaddr = kmap_atomic(page);
 935                         void *buf = mapaddr + offset;
 936                         offset += fuse_copy_do(cs, &buf, &count);
 937                         kunmap_atomic(mapaddr);
 938                 } else
 939                         offset += fuse_copy_do(cs, NULL, &count);
 940         }
 941         if (page && !cs->write)
 942                 flush_dcache_page(page);
 943         return 0;
 944 }
 945
 946 /* Copy pages in the request to/from userspace buffer */
 947 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 948                            int zeroing)
 949 {
 950         unsigned i;
 951         struct fuse_req *req = cs->req;
 952         struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
 953
 954
 955         for (i = 0; i < ap->num_pages && (nbytes || zeroing); i++) {
 956                 int err;
 957                 unsigned int offset = ap->descs[i].offset;
 958                 unsigned int count = min(nbytes, ap->descs[i].length);
 959
 960                 err = fuse_copy_page(cs, &ap->pages[i], offset, count, zeroing);
 961                 if (err)
 962                         return err;
 963
 964                 nbytes -= count;
 965         }
 966         return 0;
 967 }
 968
 969 /* Copy a single argument in the request to/from userspace buffer */
 970 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 971 {
 972         while (size) {
 973                 if (!cs->len) {
 974                         int err = fuse_copy_fill(cs);
 975                         if (err)
 976                                 return err;
 977                 }
 978                 fuse_copy_do(cs, &val, &size);
 979         }
 980         return 0;
 981 }
 982
 983 /* Copy request arguments to/from userspace buffer */
 984 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 985                           unsigned argpages, struct fuse_arg *args,
 986                           int zeroing)
 987 {
 988         int err = 0;
 989         unsigned i;
 990
 991         for (i = 0; !err && i < numargs; i++)  {
 992                 struct fuse_arg *arg = &args[i];
 993                 if (i == numargs - 1 && argpages)
 994                         err = fuse_copy_pages(cs, arg->size, zeroing);
 995                 else
 996                         err = fuse_copy_one(cs, arg->value, arg->size);
 997         }
 998         return err;
 999 }
1000
1001 static int forget_pending(struct fuse_iqueue *fiq)
1002 {
1003         return fiq->forget_list_head.next != NULL;
1004 }
1005
1006 static int request_pending(struct fuse_iqueue *fiq)
1007 {
1008         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1009                 forget_pending(fiq);
1010 }
1011
1012 /*
1013  * Transfer an interrupt request to userspace
1014  *
1015  * Unlike other requests this is assembled on demand, without a need
1016  * to allocate a separate fuse_req structure.
1017  *
1018  * Called with fiq->lock held, releases it
1019  */
1020 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1021                                struct fuse_copy_state *cs,
1022                                size_t nbytes, struct fuse_req *req)
1023 __releases(fiq->lock)
1024 {
1025         struct fuse_in_header ih;
1026         struct fuse_interrupt_in arg;
1027         unsigned reqsize = sizeof(ih) + sizeof(arg);
1028         int err;
1029
1030         list_del_init(&req->intr_entry);
1031         memset(&ih, 0, sizeof(ih));
1032         memset(&arg, 0, sizeof(arg));
1033         ih.len = reqsize;
1034         ih.opcode = FUSE_INTERRUPT;
1035         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1036         arg.unique = req->in.h.unique;
1037
1038         spin_unlock(&fiq->lock);
1039         if (nbytes < reqsize)
1040                 return -EINVAL;
1041
1042         err = fuse_copy_one(cs, &ih, sizeof(ih));
1043         if (!err)
1044                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1045         fuse_copy_finish(cs);
1046
1047         return err ? err : reqsize;
1048 }
1049
1050 struct fuse_forget_link *fuse_dequeue_forget(struct fuse_iqueue *fiq,
1051                                              unsigned int max,
1052                                              unsigned int *countp)
1053 {
1054         struct fuse_forget_link *head = fiq->forget_list_head.next;
1055         struct fuse_forget_link **newhead = &head;
1056         unsigned count;
1057
1058         for (count = 0; *newhead != NULL && count < max; count++)
1059                 newhead = &(*newhead)->next;
1060
1061         fiq->forget_list_head.next = *newhead;
1062         *newhead = NULL;
1063         if (fiq->forget_list_head.next == NULL)
1064                 fiq->forget_list_tail = &fiq->forget_list_head;
1065
1066         if (countp != NULL)
1067                 *countp = count;
1068
1069         return head;
1070 }
1071 EXPORT_SYMBOL(fuse_dequeue_forget);
1072
1073 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1074                                    struct fuse_copy_state *cs,
1075                                    size_t nbytes)
1076 __releases(fiq->lock)
1077 {
1078         int err;
1079         struct fuse_forget_link *forget = fuse_dequeue_forget(fiq, 1, NULL);
1080         struct fuse_forget_in arg = {
1081                 .nlookup = forget->forget_one.nlookup,
1082         };
1083         struct fuse_in_header ih = {
1084                 .opcode = FUSE_FORGET,
1085                 .nodeid = forget->forget_one.nodeid,
1086                 .unique = fuse_get_unique(fiq),
1087                 .len = sizeof(ih) + sizeof(arg),
1088         };
1089
1090         spin_unlock(&fiq->lock);
1091         kfree(forget);
1092         if (nbytes < ih.len)
1093                 return -EINVAL;
1094
1095         err = fuse_copy_one(cs, &ih, sizeof(ih));
1096         if (!err)
1097                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1098         fuse_copy_finish(cs);
1099
1100         if (err)
1101                 return err;
1102
1103         return ih.len;
1104 }
1105
1106 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1107                                    struct fuse_copy_state *cs, size_t nbytes)
1108 __releases(fiq->lock)
1109 {
1110         int err;
1111         unsigned max_forgets;
1112         unsigned count;
1113         struct fuse_forget_link *head;
1114         struct fuse_batch_forget_in arg = { .count = 0 };
1115         struct fuse_in_header ih = {
1116                 .opcode = FUSE_BATCH_FORGET,
1117                 .unique = fuse_get_unique(fiq),
1118                 .len = sizeof(ih) + sizeof(arg),
1119         };
1120
1121         if (nbytes < ih.len) {
1122                 spin_unlock(&fiq->lock);
1123                 return -EINVAL;
1124         }
1125
1126         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1127         head = fuse_dequeue_forget(fiq, max_forgets, &count);
1128         spin_unlock(&fiq->lock);
1129
1130         arg.count = count;
1131         ih.len += count * sizeof(struct fuse_forget_one);
1132         err = fuse_copy_one(cs, &ih, sizeof(ih));
1133         if (!err)
1134                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1135
1136         while (head) {
1137                 struct fuse_forget_link *forget = head;
1138
1139                 if (!err) {
1140                         err = fuse_copy_one(cs, &forget->forget_one,
1141                                             sizeof(forget->forget_one));
1142                 }
1143                 head = forget->next;
1144                 kfree(forget);
1145         }
1146
1147         fuse_copy_finish(cs);
1148
1149         if (err)
1150                 return err;
1151
1152         return ih.len;
1153 }
1154
1155 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1156                             struct fuse_copy_state *cs,
1157                             size_t nbytes)
1158 __releases(fiq->lock)
1159 {
1160         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1161                 return fuse_read_single_forget(fiq, cs, nbytes);
1162         else
1163                 return fuse_read_batch_forget(fiq, cs, nbytes);
1164 }
1165
1166 /*
1167  * Read a single request into the userspace filesystem's buffer.  This
1168  * function waits until a request is available, then removes it from
1169  * the pending list and copies request data to userspace buffer.  If
1170  * no reply is needed (FORGET) or request has been aborted or there
1171  * was an error during the copying then it's finished by calling
1172  * fuse_request_end().  Otherwise add it to the processing list, and set
1173  * the 'sent' flag.
1174  */
1175 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1176                                 struct fuse_copy_state *cs, size_t nbytes)
1177 {
1178         ssize_t err;
1179         struct fuse_conn *fc = fud->fc;
1180         struct fuse_iqueue *fiq = &fc->iq;
1181         struct fuse_pqueue *fpq = &fud->pq;
1182         struct fuse_req *req;
1183         struct fuse_args *args;
1184         unsigned reqsize;
1185         unsigned int hash;
1186
1187         /*
1188          * Require sane minimum read buffer - that has capacity for fixed part
1189          * of any request header + negotiated max_write room for data.
1190          *
1191          * Historically libfuse reserves 4K for fixed header room, but e.g.
1192          * GlusterFS reserves only 80 bytes
1193          *
1194          *      = `sizeof(fuse_in_header) + sizeof(fuse_write_in)`
1195          *
1196          * which is the absolute minimum any sane filesystem should be using
1197          * for header room.
1198          */
1199         if (nbytes < max_t(size_t, FUSE_MIN_READ_BUFFER,
1200                            sizeof(struct fuse_in_header) +
1201                            sizeof(struct fuse_write_in) +
1202                            fc->max_write))
1203                 return -EINVAL;
1204
1205  restart:
1206         for (;;) {
1207                 spin_lock(&fiq->lock);
1208                 if (!fiq->connected || request_pending(fiq))
1209                         break;
1210                 spin_unlock(&fiq->lock);
1211
1212                 if (file->f_flags & O_NONBLOCK)
1213                         return -EAGAIN;
1214                 err = wait_event_interruptible_exclusive(fiq->waitq,
1215                                 !fiq->connected || request_pending(fiq));
1216                 if (err)
1217                         return err;
1218         }
1219
1220         if (!fiq->connected) {
1221                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1222                 goto err_unlock;
1223         }
1224
1225         if (!list_empty(&fiq->interrupts)) {
1226                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1227                                  intr_entry);
1228                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1229         }
1230
1231         if (forget_pending(fiq)) {
1232                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1233                         return fuse_read_forget(fc, fiq, cs, nbytes);
1234
1235                 if (fiq->forget_batch <= -8)
1236                         fiq->forget_batch = 16;
1237         }
1238
1239         req = list_entry(fiq->pending.next, struct fuse_req, list);
1240         clear_bit(FR_PENDING, &req->flags);
1241         list_del_init(&req->list);
1242         spin_unlock(&fiq->lock);
1243
1244         args = req->args;
1245         reqsize = req->in.h.len;
1246
1247         /* If request is too large, reply with an error and restart the read */
1248         if (nbytes < reqsize) {
1249                 req->out.h.error = -EIO;
1250                 /* SETXATTR is special, since it may contain too large data */
1251                 if (args->opcode == FUSE_SETXATTR)
1252                         req->out.h.error = -E2BIG;
1253                 fuse_request_end(fc, req);
1254                 goto restart;
1255         }
1256         spin_lock(&fpq->lock);
1257         list_add(&req->list, &fpq->io);
1258         spin_unlock(&fpq->lock);
1259         cs->req = req;
1260         err = fuse_copy_one(cs, &req->in.h, sizeof(req->in.h));
1261         if (!err)
1262                 err = fuse_copy_args(cs, args->in_numargs, args->in_pages,
1263                                      (struct fuse_arg *) args->in_args, 0);
1264         fuse_copy_finish(cs);
1265         spin_lock(&fpq->lock);
1266         clear_bit(FR_LOCKED, &req->flags);
1267         if (!fpq->connected) {
1268                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1269                 goto out_end;
1270         }
1271         if (err) {
1272                 req->out.h.error = -EIO;
1273                 goto out_end;
1274         }
1275         if (!test_bit(FR_ISREPLY, &req->flags)) {
1276                 err = reqsize;
1277                 goto out_end;
1278         }
1279         hash = fuse_req_hash(req->in.h.unique);
1280         list_move_tail(&req->list, &fpq->processing[hash]);
1281         __fuse_get_request(req);
1282         set_bit(FR_SENT, &req->flags);
1283         spin_unlock(&fpq->lock);
1284         /* matches barrier in request_wait_answer() */
1285         smp_mb__after_atomic();
1286         if (test_bit(FR_INTERRUPTED, &req->flags))
1287                 queue_interrupt(fiq, req);
1288         fuse_put_request(fc, req);
1289
1290         return reqsize;
1291
1292 out_end:
1293         if (!test_bit(FR_PRIVATE, &req->flags))
1294                 list_del_init(&req->list);
1295         spin_unlock(&fpq->lock);
1296         fuse_request_end(fc, req);
1297         return err;
1298
1299  err_unlock:
1300         spin_unlock(&fiq->lock);
1301         return err;
1302 }
1303
1304 static int fuse_dev_open(struct inode *inode, struct file *file)
1305 {
1306         /*
1307          * The fuse device's file's private_data is used to hold
1308          * the fuse_conn(ection) when it is mounted, and is used to
1309          * keep track of whether the file has been mounted already.
1310          */
1311         file->private_data = NULL;
1312         return 0;
1313 }
1314
1315 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1316 {
1317         struct fuse_copy_state cs;
1318         struct file *file = iocb->ki_filp;
1319         struct fuse_dev *fud = fuse_get_dev(file);
1320
1321         if (!fud)
1322                 return -EPERM;
1323
1324         if (!iter_is_iovec(to))
1325                 return -EINVAL;
1326
1327         fuse_copy_init(&cs, 1, to);
1328
1329         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1330 }
1331
1332 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1333                                     struct pipe_inode_info *pipe,
1334                                     size_t len, unsigned int flags)
1335 {
1336         int total, ret;
1337         int page_nr = 0;
1338         struct pipe_buffer *bufs;
1339         struct fuse_copy_state cs;
1340         struct fuse_dev *fud = fuse_get_dev(in);
1341
1342         if (!fud)
1343                 return -EPERM;
1344
1345         bufs = kvmalloc_array(pipe->max_usage, sizeof(struct pipe_buffer),
1346                               GFP_KERNEL);
1347         if (!bufs)
1348                 return -ENOMEM;
1349
1350         fuse_copy_init(&cs, 1, NULL);
1351         cs.pipebufs = bufs;
1352         cs.pipe = pipe;
1353         ret = fuse_dev_do_read(fud, in, &cs, len);
1354         if (ret < 0)
1355                 goto out;
1356
1357         if (pipe_occupancy(pipe->head, pipe->tail) + cs.nr_segs > pipe->max_usage) {
1358                 ret = -EIO;
1359                 goto out;
1360         }
1361
1362         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1363                 /*
1364                  * Need to be careful about this.  Having buf->ops in module
1365                  * code can Oops if the buffer persists after module unload.
1366                  */
1367                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1368                 bufs[page_nr].flags = 0;
1369                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1370                 if (unlikely(ret < 0))
1371                         break;
1372         }
1373         if (total)
1374                 ret = total;
1375 out:
1376         for (; page_nr < cs.nr_segs; page_nr++)
1377                 put_page(bufs[page_nr].page);
1378
1379         kvfree(bufs);
1380         return ret;
1381 }
1382
1383 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1384                             struct fuse_copy_state *cs)
1385 {
1386         struct fuse_notify_poll_wakeup_out outarg;
1387         int err = -EINVAL;
1388
1389         if (size != sizeof(outarg))
1390                 goto err;
1391
1392         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1393         if (err)
1394                 goto err;
1395
1396         fuse_copy_finish(cs);
1397         return fuse_notify_poll_wakeup(fc, &outarg);
1398
1399 err:
1400         fuse_copy_finish(cs);
1401         return err;
1402 }
1403
1404 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1405                                    struct fuse_copy_state *cs)
1406 {
1407         struct fuse_notify_inval_inode_out outarg;
1408         int err = -EINVAL;
1409
1410         if (size != sizeof(outarg))
1411                 goto err;
1412
1413         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1414         if (err)
1415                 goto err;
1416         fuse_copy_finish(cs);
1417
1418         down_read(&fc->killsb);
1419         err = -ENOENT;
1420         if (fc->sb) {
1421                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1422                                                outarg.off, outarg.len);
1423         }
1424         up_read(&fc->killsb);
1425         return err;
1426
1427 err:
1428         fuse_copy_finish(cs);
1429         return err;
1430 }
1431
1432 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1433                                    struct fuse_copy_state *cs)
1434 {
1435         struct fuse_notify_inval_entry_out outarg;
1436         int err = -ENOMEM;
1437         char *buf;
1438         struct qstr name;
1439
1440         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1441         if (!buf)
1442                 goto err;
1443
1444         err = -EINVAL;
1445         if (size < sizeof(outarg))
1446                 goto err;
1447
1448         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1449         if (err)
1450                 goto err;
1451
1452         err = -ENAMETOOLONG;
1453         if (outarg.namelen > FUSE_NAME_MAX)
1454                 goto err;
1455
1456         err = -EINVAL;
1457         if (size != sizeof(outarg) + outarg.namelen + 1)
1458                 goto err;
1459
1460         name.name = buf;
1461         name.len = outarg.namelen;
1462         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1463         if (err)
1464                 goto err;
1465         fuse_copy_finish(cs);
1466         buf[outarg.namelen] = 0;
1467
1468         down_read(&fc->killsb);
1469         err = -ENOENT;
1470         if (fc->sb)
1471                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1472         up_read(&fc->killsb);
1473         kfree(buf);
1474         return err;
1475
1476 err:
1477         kfree(buf);
1478         fuse_copy_finish(cs);
1479         return err;
1480 }
1481
1482 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1483                               struct fuse_copy_state *cs)
1484 {
1485         struct fuse_notify_delete_out outarg;
1486         int err = -ENOMEM;
1487         char *buf;
1488         struct qstr name;
1489
1490         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1491         if (!buf)
1492                 goto err;
1493
1494         err = -EINVAL;
1495         if (size < sizeof(outarg))
1496                 goto err;
1497
1498         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1499         if (err)
1500                 goto err;
1501
1502         err = -ENAMETOOLONG;
1503         if (outarg.namelen > FUSE_NAME_MAX)
1504                 goto err;
1505
1506         err = -EINVAL;
1507         if (size != sizeof(outarg) + outarg.namelen + 1)
1508                 goto err;
1509
1510         name.name = buf;
1511         name.len = outarg.namelen;
1512         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1513         if (err)
1514                 goto err;
1515         fuse_copy_finish(cs);
1516         buf[outarg.namelen] = 0;
1517
1518         down_read(&fc->killsb);
1519         err = -ENOENT;
1520         if (fc->sb)
1521                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1522                                                outarg.child, &name);
1523         up_read(&fc->killsb);
1524         kfree(buf);
1525         return err;
1526
1527 err:
1528         kfree(buf);
1529         fuse_copy_finish(cs);
1530         return err;
1531 }
1532
1533 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1534                              struct fuse_copy_state *cs)
1535 {
1536         struct fuse_notify_store_out outarg;
1537         struct inode *inode;
1538         struct address_space *mapping;
1539         u64 nodeid;
1540         int err;
1541         pgoff_t index;
1542         unsigned int offset;
1543         unsigned int num;
1544         loff_t file_size;
1545         loff_t end;
1546
1547         err = -EINVAL;
1548         if (size < sizeof(outarg))
1549                 goto out_finish;
1550
1551         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1552         if (err)
1553                 goto out_finish;
1554
1555         err = -EINVAL;
1556         if (size - sizeof(outarg) != outarg.size)
1557                 goto out_finish;
1558
1559         nodeid = outarg.nodeid;
1560
1561         down_read(&fc->killsb);
1562
1563         err = -ENOENT;
1564         if (!fc->sb)
1565                 goto out_up_killsb;
1566
1567         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1568         if (!inode)
1569                 goto out_up_killsb;
1570
1571         mapping = inode->i_mapping;
1572         index = outarg.offset >> PAGE_SHIFT;
1573         offset = outarg.offset & ~PAGE_MASK;
1574         file_size = i_size_read(inode);
1575         end = outarg.offset + outarg.size;
1576         if (end > file_size) {
1577                 file_size = end;
1578                 fuse_write_update_size(inode, file_size);
1579         }
1580
1581         num = outarg.size;
1582         while (num) {
1583                 struct page *page;
1584                 unsigned int this_num;
1585
1586                 err = -ENOMEM;
1587                 page = find_or_create_page(mapping, index,
1588                                            mapping_gfp_mask(mapping));
1589                 if (!page)
1590                         goto out_iput;
1591
1592                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1593                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1594                 if (!err && offset == 0 &&
1595                     (this_num == PAGE_SIZE || file_size == end))
1596                         SetPageUptodate(page);
1597                 unlock_page(page);
1598                 put_page(page);
1599
1600                 if (err)
1601                         goto out_iput;
1602
1603                 num -= this_num;
1604                 offset = 0;
1605                 index++;
1606         }
1607
1608         err = 0;
1609
1610 out_iput:
1611         iput(inode);
1612 out_up_killsb:
1613         up_read(&fc->killsb);
1614 out_finish:
1615         fuse_copy_finish(cs);
1616         return err;
1617 }
1618
1619 struct fuse_retrieve_args {
1620         struct fuse_args_pages ap;
1621         struct fuse_notify_retrieve_in inarg;
1622 };
1623
1624 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_args *args,
1625                               int error)
1626 {
1627         struct fuse_retrieve_args *ra =
1628                 container_of(args, typeof(*ra), ap.args);
1629
1630         release_pages(ra->ap.pages, ra->ap.num_pages);
1631         kfree(ra);
1632 }
1633
1634 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1635                          struct fuse_notify_retrieve_out *outarg)
1636 {
1637         int err;
1638         struct address_space *mapping = inode->i_mapping;
1639         pgoff_t index;
1640         loff_t file_size;
1641         unsigned int num;
1642         unsigned int offset;
1643         size_t total_len = 0;
1644         unsigned int num_pages;
1645         struct fuse_retrieve_args *ra;
1646         size_t args_size = sizeof(*ra);
1647         struct fuse_args_pages *ap;
1648         struct fuse_args *args;
1649
1650         offset = outarg->offset & ~PAGE_MASK;
1651         file_size = i_size_read(inode);
1652
1653         num = min(outarg->size, fc->max_write);
1654         if (outarg->offset > file_size)
1655                 num = 0;
1656         else if (outarg->offset + num > file_size)
1657                 num = file_size - outarg->offset;
1658
1659         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1660         num_pages = min(num_pages, fc->max_pages);
1661
1662         args_size += num_pages * (sizeof(ap->pages[0]) + sizeof(ap->descs[0]));
1663
1664         ra = kzalloc(args_size, GFP_KERNEL);
1665         if (!ra)
1666                 return -ENOMEM;
1667
1668         ap = &ra->ap;
1669         ap->pages = (void *) (ra + 1);
1670         ap->descs = (void *) (ap->pages + num_pages);
1671
1672         args = &ap->args;
1673         args->nodeid = outarg->nodeid;
1674         args->opcode = FUSE_NOTIFY_REPLY;
1675         args->in_numargs = 2;
1676         args->in_pages = true;
1677         args->end = fuse_retrieve_end;
1678
1679         index = outarg->offset >> PAGE_SHIFT;
1680
1681         while (num && ap->num_pages < num_pages) {
1682                 struct page *page;
1683                 unsigned int this_num;
1684
1685                 page = find_get_page(mapping, index);
1686                 if (!page)
1687                         break;
1688
1689                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1690                 ap->pages[ap->num_pages] = page;
1691                 ap->descs[ap->num_pages].offset = offset;
1692                 ap->descs[ap->num_pages].length = this_num;
1693                 ap->num_pages++;
1694
1695                 offset = 0;
1696                 num -= this_num;
1697                 total_len += this_num;
1698                 index++;
1699         }
1700         ra->inarg.offset = outarg->offset;
1701         ra->inarg.size = total_len;
1702         args->in_args[0].size = sizeof(ra->inarg);
1703         args->in_args[0].value = &ra->inarg;
1704         args->in_args[1].size = total_len;
1705
1706         err = fuse_simple_notify_reply(fc, args, outarg->notify_unique);
1707         if (err)
1708                 fuse_retrieve_end(fc, args, err);
1709
1710         return err;
1711 }
1712
1713 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1714                                 struct fuse_copy_state *cs)
1715 {
1716         struct fuse_notify_retrieve_out outarg;
1717         struct inode *inode;
1718         int err;
1719
1720         err = -EINVAL;
1721         if (size != sizeof(outarg))
1722                 goto copy_finish;
1723
1724         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1725         if (err)
1726                 goto copy_finish;
1727
1728         fuse_copy_finish(cs);
1729
1730         down_read(&fc->killsb);
1731         err = -ENOENT;
1732         if (fc->sb) {
1733                 u64 nodeid = outarg.nodeid;
1734
1735                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1736                 if (inode) {
1737                         err = fuse_retrieve(fc, inode, &outarg);
1738                         iput(inode);
1739                 }
1740         }
1741         up_read(&fc->killsb);
1742
1743         return err;
1744
1745 copy_finish:
1746         fuse_copy_finish(cs);
1747         return err;
1748 }
1749
1750 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1751                        unsigned int size, struct fuse_copy_state *cs)
1752 {
1753         /* Don't try to move pages (yet) */
1754         cs->move_pages = 0;
1755
1756         switch (code) {
1757         case FUSE_NOTIFY_POLL:
1758                 return fuse_notify_poll(fc, size, cs);
1759
1760         case FUSE_NOTIFY_INVAL_INODE:
1761                 return fuse_notify_inval_inode(fc, size, cs);
1762
1763         case FUSE_NOTIFY_INVAL_ENTRY:
1764                 return fuse_notify_inval_entry(fc, size, cs);
1765
1766         case FUSE_NOTIFY_STORE:
1767                 return fuse_notify_store(fc, size, cs);
1768
1769         case FUSE_NOTIFY_RETRIEVE:
1770                 return fuse_notify_retrieve(fc, size, cs);
1771
1772         case FUSE_NOTIFY_DELETE:
1773                 return fuse_notify_delete(fc, size, cs);
1774
1775         default:
1776                 fuse_copy_finish(cs);
1777                 return -EINVAL;
1778         }
1779 }
1780
1781 /* Look up request on processing list by unique ID */
1782 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1783 {
1784         unsigned int hash = fuse_req_hash(unique);
1785         struct fuse_req *req;
1786
1787         list_for_each_entry(req, &fpq->processing[hash], list) {
1788                 if (req->in.h.unique == unique)
1789                         return req;
1790         }
1791         return NULL;
1792 }
1793
1794 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_args *args,
1795                          unsigned nbytes)
1796 {
1797         unsigned reqsize = sizeof(struct fuse_out_header);
1798
1799         reqsize += fuse_len_args(args->out_numargs, args->out_args);
1800
1801         if (reqsize < nbytes || (reqsize > nbytes && !args->out_argvar))
1802                 return -EINVAL;
1803         else if (reqsize > nbytes) {
1804                 struct fuse_arg *lastarg = &args->out_args[args->out_numargs-1];
1805                 unsigned diffsize = reqsize - nbytes;
1806
1807                 if (diffsize > lastarg->size)
1808                         return -EINVAL;
1809                 lastarg->size -= diffsize;
1810         }
1811         return fuse_copy_args(cs, args->out_numargs, args->out_pages,
1812                               args->out_args, args->page_zeroing);
1813 }
1814
1815 /*
1816  * Write a single reply to a request.  First the header is copied from
1817  * the write buffer.  The request is then searched on the processing
1818  * list by the unique ID found in the header.  If found, then remove
1819  * it from the list and copy the rest of the buffer to the request.
1820  * The request is finished by calling fuse_request_end().
1821  */
1822 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1823                                  struct fuse_copy_state *cs, size_t nbytes)
1824 {
1825         int err;
1826         struct fuse_conn *fc = fud->fc;
1827         struct fuse_pqueue *fpq = &fud->pq;
1828         struct fuse_req *req;
1829         struct fuse_out_header oh;
1830
1831         err = -EINVAL;
1832         if (nbytes < sizeof(struct fuse_out_header))
1833                 goto out;
1834
1835         err = fuse_copy_one(cs, &oh, sizeof(oh));
1836         if (err)
1837                 goto copy_finish;
1838
1839         err = -EINVAL;
1840         if (oh.len != nbytes)
1841                 goto copy_finish;
1842
1843         /*
1844          * Zero oh.unique indicates unsolicited notification message
1845          * and error contains notification code.
1846          */
1847         if (!oh.unique) {
1848                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1849                 goto out;
1850         }
1851
1852         err = -EINVAL;
1853         if (oh.error <= -1000 || oh.error > 0)
1854                 goto copy_finish;
1855
1856         spin_lock(&fpq->lock);
1857         req = NULL;
1858         if (fpq->connected)
1859                 req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1860
1861         err = -ENOENT;
1862         if (!req) {
1863                 spin_unlock(&fpq->lock);
1864                 goto copy_finish;
1865         }
1866
1867         /* Is it an interrupt reply ID? */
1868         if (oh.unique & FUSE_INT_REQ_BIT) {
1869                 __fuse_get_request(req);
1870                 spin_unlock(&fpq->lock);
1871
1872                 err = 0;
1873                 if (nbytes != sizeof(struct fuse_out_header))
1874                         err = -EINVAL;
1875                 else if (oh.error == -ENOSYS)
1876                         fc->no_interrupt = 1;
1877                 else if (oh.error == -EAGAIN)
1878                         err = queue_interrupt(&fc->iq, req);
1879
1880                 fuse_put_request(fc, req);
1881
1882                 goto copy_finish;
1883         }
1884
1885         clear_bit(FR_SENT, &req->flags);
1886         list_move(&req->list, &fpq->io);
1887         req->out.h = oh;
1888         set_bit(FR_LOCKED, &req->flags);
1889         spin_unlock(&fpq->lock);
1890         cs->req = req;
1891         if (!req->args->page_replace)
1892                 cs->move_pages = 0;
1893
1894         if (oh.error)
1895                 err = nbytes != sizeof(oh) ? -EINVAL : 0;
1896         else
1897                 err = copy_out_args(cs, req->args, nbytes);
1898         fuse_copy_finish(cs);
1899
1900         spin_lock(&fpq->lock);
1901         clear_bit(FR_LOCKED, &req->flags);
1902         if (!fpq->connected)
1903                 err = -ENOENT;
1904         else if (err)
1905                 req->out.h.error = -EIO;
1906         if (!test_bit(FR_PRIVATE, &req->flags))
1907                 list_del_init(&req->list);
1908         spin_unlock(&fpq->lock);
1909
1910         fuse_request_end(fc, req);
1911 out:
1912         return err ? err : nbytes;
1913
1914 copy_finish:
1915         fuse_copy_finish(cs);
1916         goto out;
1917 }
1918
1919 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1920 {
1921         struct fuse_copy_state cs;
1922         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1923
1924         if (!fud)
1925                 return -EPERM;
1926
1927         if (!iter_is_iovec(from))
1928                 return -EINVAL;
1929
1930         fuse_copy_init(&cs, 0, from);
1931
1932         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1933 }
1934
1935 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1936                                      struct file *out, loff_t *ppos,
1937                                      size_t len, unsigned int flags)
1938 {
1939         unsigned int head, tail, mask, count;
1940         unsigned nbuf;
1941         unsigned idx;
1942         struct pipe_buffer *bufs;
1943         struct fuse_copy_state cs;
1944         struct fuse_dev *fud;
1945         size_t rem;
1946         ssize_t ret;
1947
1948         fud = fuse_get_dev(out);
1949         if (!fud)
1950                 return -EPERM;
1951
1952         pipe_lock(pipe);
1953
1954         head = pipe->head;
1955         tail = pipe->tail;
1956         mask = pipe->ring_size - 1;
1957         count = head - tail;
1958
1959         bufs = kvmalloc_array(count, sizeof(struct pipe_buffer), GFP_KERNEL);
1960         if (!bufs) {
1961                 pipe_unlock(pipe);
1962                 return -ENOMEM;
1963         }
1964
1965         nbuf = 0;
1966         rem = 0;
1967         for (idx = tail; idx != head && rem < len; idx++)
1968                 rem += pipe->bufs[idx & mask].len;
1969
1970         ret = -EINVAL;
1971         if (rem < len)
1972                 goto out_free;
1973
1974         rem = len;
1975         while (rem) {
1976                 struct pipe_buffer *ibuf;
1977                 struct pipe_buffer *obuf;
1978
1979                 if (WARN_ON(nbuf >= count || tail == head))
1980                         goto out_free;
1981
1982                 ibuf = &pipe->bufs[tail & mask];
1983                 obuf = &bufs[nbuf];
1984
1985                 if (rem >= ibuf->len) {
1986                         *obuf = *ibuf;
1987                         ibuf->ops = NULL;
1988                         tail++;
1989                         pipe->tail = tail;
1990                 } else {
1991                         if (!pipe_buf_get(pipe, ibuf))
1992                                 goto out_free;
1993
1994                         *obuf = *ibuf;
1995                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1996                         obuf->len = rem;
1997                         ibuf->offset += obuf->len;
1998                         ibuf->len -= obuf->len;
1999                 }
2000                 nbuf++;
2001                 rem -= obuf->len;
2002         }
2003         pipe_unlock(pipe);
2004
2005         fuse_copy_init(&cs, 0, NULL);
2006         cs.pipebufs = bufs;
2007         cs.nr_segs = nbuf;
2008         cs.pipe = pipe;
2009
2010         if (flags & SPLICE_F_MOVE)
2011                 cs.move_pages = 1;
2012
2013         ret = fuse_dev_do_write(fud, &cs, len);
2014
2015         pipe_lock(pipe);
2016 out_free:
2017         for (idx = 0; idx < nbuf; idx++)
2018                 pipe_buf_release(pipe, &bufs[idx]);
2019         pipe_unlock(pipe);
2020
2021         kvfree(bufs);
2022         return ret;
2023 }
2024
2025 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2026 {
2027         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2028         struct fuse_iqueue *fiq;
2029         struct fuse_dev *fud = fuse_get_dev(file);
2030
2031         if (!fud)
2032                 return EPOLLERR;
2033
2034         fiq = &fud->fc->iq;
2035         poll_wait(file, &fiq->waitq, wait);
2036
2037         spin_lock(&fiq->lock);
2038         if (!fiq->connected)
2039                 mask = EPOLLERR;
2040         else if (request_pending(fiq))
2041                 mask |= EPOLLIN | EPOLLRDNORM;
2042         spin_unlock(&fiq->lock);
2043
2044         return mask;
2045 }
2046
2047 /* Abort all requests on the given list (pending or processing) */
2048 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2049 {
2050         while (!list_empty(head)) {
2051                 struct fuse_req *req;
2052                 req = list_entry(head->next, struct fuse_req, list);
2053                 req->out.h.error = -ECONNABORTED;
2054                 clear_bit(FR_SENT, &req->flags);
2055                 list_del_init(&req->list);
2056                 fuse_request_end(fc, req);
2057         }
2058 }
2059
2060 static void end_polls(struct fuse_conn *fc)
2061 {
2062         struct rb_node *p;
2063
2064         p = rb_first(&fc->polled_files);
2065
2066         while (p) {
2067                 struct fuse_file *ff;
2068                 ff = rb_entry(p, struct fuse_file, polled_node);
2069                 wake_up_interruptible_all(&ff->poll_wait);
2070
2071                 p = rb_next(p);
2072         }
2073 }
2074
2075 /*
2076  * Abort all requests.
2077  *
2078  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2079  * filesystem.
2080  *
2081  * The same effect is usually achievable through killing the filesystem daemon
2082  * and all users of the filesystem.  The exception is the combination of an
2083  * asynchronous request and the tricky deadlock (see
2084  * Documentation/filesystems/fuse.rst).
2085  *
2086  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2087  * requests, they should be finished off immediately.  Locked requests will be
2088  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2089  * requests.  It is possible that some request will finish before we can.  This
2090  * is OK, the request will in that case be removed from the list before we touch
2091  * it.
2092  */
2093 void fuse_abort_conn(struct fuse_conn *fc)
2094 {
2095         struct fuse_iqueue *fiq = &fc->iq;
2096
2097         spin_lock(&fc->lock);
2098         if (fc->connected) {
2099                 struct fuse_dev *fud;
2100                 struct fuse_req *req, *next;
2101                 LIST_HEAD(to_end);
2102                 unsigned int i;
2103
2104                 /* Background queuing checks fc->connected under bg_lock */
2105                 spin_lock(&fc->bg_lock);
2106                 fc->connected = 0;
2107                 spin_unlock(&fc->bg_lock);
2108
2109                 fuse_set_initialized(fc);
2110                 list_for_each_entry(fud, &fc->devices, entry) {
2111                         struct fuse_pqueue *fpq = &fud->pq;
2112
2113                         spin_lock(&fpq->lock);
2114                         fpq->connected = 0;
2115                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2116                                 req->out.h.error = -ECONNABORTED;
2117                                 spin_lock(&req->waitq.lock);
2118                                 set_bit(FR_ABORTED, &req->flags);
2119                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2120                                         set_bit(FR_PRIVATE, &req->flags);
2121                                         __fuse_get_request(req);
2122                                         list_move(&req->list, &to_end);
2123                                 }
2124                                 spin_unlock(&req->waitq.lock);
2125                         }
2126                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2127                                 list_splice_tail_init(&fpq->processing[i],
2128                                                       &to_end);
2129                         spin_unlock(&fpq->lock);
2130                 }
2131                 spin_lock(&fc->bg_lock);
2132                 fc->blocked = 0;
2133                 fc->max_background = UINT_MAX;
2134                 flush_bg_queue(fc);
2135                 spin_unlock(&fc->bg_lock);
2136
2137                 spin_lock(&fiq->lock);
2138                 fiq->connected = 0;
2139                 list_for_each_entry(req, &fiq->pending, list)
2140                         clear_bit(FR_PENDING, &req->flags);
2141                 list_splice_tail_init(&fiq->pending, &to_end);
2142                 while (forget_pending(fiq))
2143                         kfree(fuse_dequeue_forget(fiq, 1, NULL));
2144                 wake_up_all(&fiq->waitq);
2145                 spin_unlock(&fiq->lock);
2146                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2147                 end_polls(fc);
2148                 wake_up_all(&fc->blocked_waitq);
2149                 spin_unlock(&fc->lock);
2150
2151                 end_requests(fc, &to_end);
2152         } else {
2153                 spin_unlock(&fc->lock);
2154         }
2155 }
2156 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2157
2158 void fuse_wait_aborted(struct fuse_conn *fc)
2159 {
2160         /* matches implicit memory barrier in fuse_drop_waiting() */
2161         smp_mb();
2162         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2163 }
2164
2165 int fuse_dev_release(struct inode *inode, struct file *file)
2166 {
2167         struct fuse_dev *fud = fuse_get_dev(file);
2168
2169         if (fud) {
2170                 struct fuse_conn *fc = fud->fc;
2171                 struct fuse_pqueue *fpq = &fud->pq;
2172                 LIST_HEAD(to_end);
2173                 unsigned int i;
2174
2175                 spin_lock(&fpq->lock);
2176                 WARN_ON(!list_empty(&fpq->io));
2177                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2178                         list_splice_init(&fpq->processing[i], &to_end);
2179                 spin_unlock(&fpq->lock);
2180
2181                 end_requests(fc, &to_end);
2182
2183                 /* Are we the last open device? */
2184                 if (atomic_dec_and_test(&fc->dev_count)) {
2185                         WARN_ON(fc->iq.fasync != NULL);
2186                         fuse_abort_conn(fc);
2187                 }
2188                 fuse_dev_free(fud);
2189         }
2190         return 0;
2191 }
2192 EXPORT_SYMBOL_GPL(fuse_dev_release);
2193
2194 static int fuse_dev_fasync(int fd, struct file *file, int on)
2195 {
2196         struct fuse_dev *fud = fuse_get_dev(file);
2197
2198         if (!fud)
2199                 return -EPERM;
2200
2201         /* No locking - fasync_helper does its own locking */
2202         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2203 }
2204
2205 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2206 {
2207         struct fuse_dev *fud;
2208
2209         if (new->private_data)
2210                 return -EINVAL;
2211
2212         fud = fuse_dev_alloc_install(fc);
2213         if (!fud)
2214                 return -ENOMEM;
2215
2216         new->private_data = fud;
2217         atomic_inc(&fc->dev_count);
2218
2219         return 0;
2220 }
2221
2222 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2223                            unsigned long arg)
2224 {
2225         int err = -ENOTTY;
2226
2227         if (cmd == FUSE_DEV_IOC_CLONE) {
2228                 int oldfd;
2229
2230                 err = -EFAULT;
2231                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2232                         struct file *old = fget(oldfd);
2233
2234                         err = -EINVAL;
2235                         if (old) {
2236                                 struct fuse_dev *fud = NULL;
2237
2238                                 /*
2239                                  * Check against file->f_op because CUSE
2240                                  * uses the same ioctl handler.
2241                                  */
2242                                 if (old->f_op == file->f_op &&
2243                                     old->f_cred->user_ns == file->f_cred->user_ns)
2244                                         fud = fuse_get_dev(old);
2245
2246                                 if (fud) {
2247                                         mutex_lock(&fuse_mutex);
2248                                         err = fuse_device_clone(fud->fc, file);
2249                                         mutex_unlock(&fuse_mutex);
2250                                 }
2251                                 fput(old);
2252                         }
2253                 }
2254         }
2255         return err;
2256 }
2257
2258 const struct file_operations fuse_dev_operations = {
2259         .owner          = THIS_MODULE,
2260         .open           = fuse_dev_open,
2261         .llseek         = no_llseek,
2262         .read_iter      = fuse_dev_read,
2263         .splice_read    = fuse_dev_splice_read,
2264         .write_iter     = fuse_dev_write,
2265         .splice_write   = fuse_dev_splice_write,
2266         .poll           = fuse_dev_poll,
2267         .release        = fuse_dev_release,
2268         .fasync         = fuse_dev_fasync,
2269         .unlocked_ioctl = fuse_dev_ioctl,
2270         .compat_ioctl   = compat_ptr_ioctl,
2271 };
2272 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2273
2274 static struct miscdevice fuse_miscdevice = {
2275         .minor = FUSE_MINOR,
2276         .name  = "fuse",
2277         .fops = &fuse_dev_operations,
2278 };
2279
2280 int __init fuse_dev_init(void)
2281 {
2282         int err = -ENOMEM;
2283         fuse_req_cachep = kmem_cache_create("fuse_request",
2284                                             sizeof(struct fuse_req),
2285                                             0, 0, NULL);
2286         if (!fuse_req_cachep)
2287                 goto out;
2288
2289         err = misc_register(&fuse_miscdevice);
2290         if (err)
2291                 goto out_cache_clean;
2292
2293         return 0;
2294
2295  out_cache_clean:
2296         kmem_cache_destroy(fuse_req_cachep);
2297  out:
2298         return err;
2299 }
2300
2301 void fuse_dev_cleanup(void)
2302 {
2303         misc_deregister(&fuse_miscdevice);
2304         kmem_cache_destroy(fuse_req_cachep);
2305 }