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