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_mount *fm, 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         req->fm = fm;
  51 }
  52
  53 static struct fuse_req *fuse_request_alloc(struct fuse_mount *fm, gfp_t flags)
  54 {
  55         struct fuse_req *req = kmem_cache_zalloc(fuse_req_cachep, flags);
  56         if (req)
  57                 fuse_request_init(fm, req);
  58
  59         return req;
  60 }
  61
  62 static void fuse_request_free(struct fuse_req *req)
  63 {
  64         kmem_cache_free(fuse_req_cachep, req);
  65 }
  66
  67 static void __fuse_get_request(struct fuse_req *req)
  68 {
  69         refcount_inc(&req->count);
  70 }
  71
  72 /* Must be called with > 1 refcount */
  73 static void __fuse_put_request(struct fuse_req *req)
  74 {
  75         refcount_dec(&req->count);
  76 }
  77
  78 void fuse_set_initialized(struct fuse_conn *fc)
  79 {
  80         /* Make sure stores before this are seen on another CPU */
  81         smp_wmb();
  82         fc->initialized = 1;
  83 }
  84
  85 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
  86 {
  87         return !fc->initialized || (for_background && fc->blocked);
  88 }
  89
  90 static void fuse_drop_waiting(struct fuse_conn *fc)
  91 {
  92         /*
  93          * lockess check of fc->connected is okay, because atomic_dec_and_test()
  94          * provides a memory barrier mached with the one in fuse_wait_aborted()
  95          * to ensure no wake-up is missed.
  96          */
  97         if (atomic_dec_and_test(&fc->num_waiting) &&
  98             !READ_ONCE(fc->connected)) {
  99                 /* wake up aborters */
 100                 wake_up_all(&fc->blocked_waitq);
 101         }
 102 }
 103
 104 static void fuse_put_request(struct fuse_req *req);
 105
 106 static struct fuse_req *fuse_get_req(struct fuse_mount *fm, bool for_background)
 107 {
 108         struct fuse_conn *fc = fm->fc;
 109         struct fuse_req *req;
 110         int err;
 111         atomic_inc(&fc->num_waiting);
 112
 113         if (fuse_block_alloc(fc, for_background)) {
 114                 err = -EINTR;
 115                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 116                                 !fuse_block_alloc(fc, for_background)))
 117                         goto out;
 118         }
 119         /* Matches smp_wmb() in fuse_set_initialized() */
 120         smp_rmb();
 121
 122         err = -ENOTCONN;
 123         if (!fc->connected)
 124                 goto out;
 125
 126         err = -ECONNREFUSED;
 127         if (fc->conn_error)
 128                 goto out;
 129
 130         req = fuse_request_alloc(fm, GFP_KERNEL);
 131         err = -ENOMEM;
 132         if (!req) {
 133                 if (for_background)
 134                         wake_up(&fc->blocked_waitq);
 135                 goto out;
 136         }
 137
 138         req->in.h.uid = from_kuid(fc->user_ns, current_fsuid());
 139         req->in.h.gid = from_kgid(fc->user_ns, current_fsgid());
 140         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 141
 142         __set_bit(FR_WAITING, &req->flags);
 143         if (for_background)
 144                 __set_bit(FR_BACKGROUND, &req->flags);
 145
 146         if (unlikely(req->in.h.uid == ((uid_t)-1) ||
 147                      req->in.h.gid == ((gid_t)-1))) {
 148                 fuse_put_request(req);
 149                 return ERR_PTR(-EOVERFLOW);
 150         }
 151         return req;
 152
 153  out:
 154         fuse_drop_waiting(fc);
 155         return ERR_PTR(err);
 156 }
 157
 158 static void fuse_put_request(struct fuse_req *req)
 159 {
 160         struct fuse_conn *fc = req->fm->fc;
 161
 162         if (refcount_dec_and_test(&req->count)) {
 163                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 164                         /*
 165                          * We get here in the unlikely case that a background
 166                          * request was allocated but not sent
 167                          */
 168                         spin_lock(&fc->bg_lock);
 169                         if (!fc->blocked)
 170                                 wake_up(&fc->blocked_waitq);
 171                         spin_unlock(&fc->bg_lock);
 172                 }
 173
 174                 if (test_bit(FR_WAITING, &req->flags)) {
 175                         __clear_bit(FR_WAITING, &req->flags);
 176                         fuse_drop_waiting(fc);
 177                 }
 178
 179                 fuse_request_free(req);
 180         }
 181 }
 182
 183 unsigned int fuse_len_args(unsigned int numargs, struct fuse_arg *args)
 184 {
 185         unsigned nbytes = 0;
 186         unsigned i;
 187
 188         for (i = 0; i < numargs; i++)
 189                 nbytes += args[i].size;
 190
 191         return nbytes;
 192 }
 193 EXPORT_SYMBOL_GPL(fuse_len_args);
 194
 195 u64 fuse_get_unique(struct fuse_iqueue *fiq)
 196 {
 197         fiq->reqctr += FUSE_REQ_ID_STEP;
 198         return fiq->reqctr;
 199 }
 200 EXPORT_SYMBOL_GPL(fuse_get_unique);
 201
 202 static unsigned int fuse_req_hash(u64 unique)
 203 {
 204         return hash_long(unique & ~FUSE_INT_REQ_BIT, FUSE_PQ_HASH_BITS);
 205 }
 206
 207 /**
 208  * A new request is available, wake fiq->waitq
 209  */
 210 static void fuse_dev_wake_and_unlock(struct fuse_iqueue *fiq)
 211 __releases(fiq->lock)
 212 {
 213         wake_up(&fiq->waitq);
 214         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 215         spin_unlock(&fiq->lock);
 216 }
 217
 218 const struct fuse_iqueue_ops fuse_dev_fiq_ops = {
 219         .wake_forget_and_unlock         = fuse_dev_wake_and_unlock,
 220         .wake_interrupt_and_unlock      = fuse_dev_wake_and_unlock,
 221         .wake_pending_and_unlock        = fuse_dev_wake_and_unlock,
 222 };
 223 EXPORT_SYMBOL_GPL(fuse_dev_fiq_ops);
 224
 225 static void queue_request_and_unlock(struct fuse_iqueue *fiq,
 226                                      struct fuse_req *req)
 227 __releases(fiq->lock)
 228 {
 229         req->in.h.len = sizeof(struct fuse_in_header) +
 230                 fuse_len_args(req->args->in_numargs,
 231                               (struct fuse_arg *) req->args->in_args);
 232         list_add_tail(&req->list, &fiq->pending);
 233         fiq->ops->wake_pending_and_unlock(fiq);
 234 }
 235
 236 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 237                        u64 nodeid, u64 nlookup)
 238 {
 239         struct fuse_iqueue *fiq = &fc->iq;
 240
 241         forget->forget_one.nodeid = nodeid;
 242         forget->forget_one.nlookup = nlookup;
 243
 244         spin_lock(&fiq->lock);
 245         if (fiq->connected) {
 246                 fiq->forget_list_tail->next = forget;
 247                 fiq->forget_list_tail = forget;
 248                 fiq->ops->wake_forget_and_unlock(fiq);
 249         } else {
 250                 kfree(forget);
 251                 spin_unlock(&fiq->lock);
 252         }
 253 }
 254
 255 static void flush_bg_queue(struct fuse_conn *fc)
 256 {
 257         struct fuse_iqueue *fiq = &fc->iq;
 258
 259         while (fc->active_background < fc->max_background &&
 260                !list_empty(&fc->bg_queue)) {
 261                 struct fuse_req *req;
 262
 263                 req = list_first_entry(&fc->bg_queue, struct fuse_req, list);
 264                 list_del(&req->list);
 265                 fc->active_background++;
 266                 spin_lock(&fiq->lock);
 267                 req->in.h.unique = fuse_get_unique(fiq);
 268                 queue_request_and_unlock(fiq, req);
 269         }
 270 }
 271
 272 /*
 273  * This function is called when a request is finished.  Either a reply
 274  * has arrived or it was aborted (and not yet sent) or some error
 275  * occurred during communication with userspace, or the device file
 276  * was closed.  The requester thread is woken up (if still waiting),
 277  * the 'end' callback is called if given, else the reference to the
 278  * request is released
 279  */
 280 void fuse_request_end(struct fuse_req *req)
 281 {
 282         struct fuse_mount *fm = req->fm;
 283         struct fuse_conn *fc = fm->fc;
 284         struct fuse_iqueue *fiq = &fc->iq;
 285
 286         if (test_and_set_bit(FR_FINISHED, &req->flags))
 287                 goto put_request;
 288
 289         /*
 290          * test_and_set_bit() implies smp_mb() between bit
 291          * changing and below FR_INTERRUPTED check. Pairs with
 292          * smp_mb() from queue_interrupt().
 293          */
 294         if (test_bit(FR_INTERRUPTED, &req->flags)) {
 295                 spin_lock(&fiq->lock);
 296                 list_del_init(&req->intr_entry);
 297                 spin_unlock(&fiq->lock);
 298         }
 299         WARN_ON(test_bit(FR_PENDING, &req->flags));
 300         WARN_ON(test_bit(FR_SENT, &req->flags));
 301         if (test_bit(FR_BACKGROUND, &req->flags)) {
 302                 spin_lock(&fc->bg_lock);
 303                 clear_bit(FR_BACKGROUND, &req->flags);
 304                 if (fc->num_background == fc->max_background) {
 305                         fc->blocked = 0;
 306                         wake_up(&fc->blocked_waitq);
 307                 } else if (!fc->blocked) {
 308                         /*
 309                          * Wake up next waiter, if any.  It's okay to use
 310                          * waitqueue_active(), as we've already synced up
 311                          * fc->blocked with waiters with the wake_up() call
 312                          * above.
 313                          */
 314                         if (waitqueue_active(&fc->blocked_waitq))
 315                                 wake_up(&fc->blocked_waitq);
 316                 }
 317
 318                 if (fc->num_background == fc->congestion_threshold && fm->sb) {
 319                         clear_bdi_congested(fm->sb->s_bdi, BLK_RW_SYNC);
 320                         clear_bdi_congested(fm->sb->s_bdi, BLK_RW_ASYNC);
 321                 }
 322                 fc->num_background--;
 323                 fc->active_background--;
 324                 flush_bg_queue(fc);
 325                 spin_unlock(&fc->bg_lock);
 326         } else {
 327                 /* Wake up waiter sleeping in request_wait_answer() */
 328                 wake_up(&req->waitq);
 329         }
 330
 331         if (test_bit(FR_ASYNC, &req->flags))
 332                 req->args->end(fm, req->args, req->out.h.error);
 333 put_request:
 334         fuse_put_request(req);
 335 }
 336 EXPORT_SYMBOL_GPL(fuse_request_end);
 337
 338 static int queue_interrupt(struct fuse_req *req)
 339 {
 340         struct fuse_iqueue *fiq = &req->fm->fc->iq;
 341
 342         spin_lock(&fiq->lock);
 343         /* Check for we've sent request to interrupt this req */
 344         if (unlikely(!test_bit(FR_INTERRUPTED, &req->flags))) {
 345                 spin_unlock(&fiq->lock);
 346                 return -EINVAL;
 347         }
 348
 349         if (list_empty(&req->intr_entry)) {
 350                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 351                 /*
 352                  * Pairs with smp_mb() implied by test_and_set_bit()
 353                  * from fuse_request_end().
 354                  */
 355                 smp_mb();
 356                 if (test_bit(FR_FINISHED, &req->flags)) {
 357                         list_del_init(&req->intr_entry);
 358                         spin_unlock(&fiq->lock);
 359                         return 0;
 360                 }
 361                 fiq->ops->wake_interrupt_and_unlock(fiq);
 362         } else {
 363                 spin_unlock(&fiq->lock);
 364         }
 365         return 0;
 366 }
 367
 368 static void request_wait_answer(struct fuse_req *req)
 369 {
 370         struct fuse_conn *fc = req->fm->fc;
 371         struct fuse_iqueue *fiq = &fc->iq;
 372         int err;
 373
 374         if (!fc->no_interrupt) {
 375                 /* Any signal may interrupt this */
 376                 err = wait_event_interruptible(req->waitq,
 377                                         test_bit(FR_FINISHED, &req->flags));
 378                 if (!err)
 379                         return;
 380
 381                 set_bit(FR_INTERRUPTED, &req->flags);
 382                 /* matches barrier in fuse_dev_do_read() */
 383                 smp_mb__after_atomic();
 384                 if (test_bit(FR_SENT, &req->flags))
 385                         queue_interrupt(req);
 386         }
 387
 388         if (!test_bit(FR_FORCE, &req->flags)) {
 389                 /* Only fatal signals may interrupt this */
 390                 err = wait_event_killable(req->waitq,
 391                                         test_bit(FR_FINISHED, &req->flags));
 392                 if (!err)
 393                         return;
 394
 395                 spin_lock(&fiq->lock);
 396                 /* Request is not yet in userspace, bail out */
 397                 if (test_bit(FR_PENDING, &req->flags)) {
 398                         list_del(&req->list);
 399                         spin_unlock(&fiq->lock);
 400                         __fuse_put_request(req);
 401                         req->out.h.error = -EINTR;
 402                         return;
 403                 }
 404                 spin_unlock(&fiq->lock);
 405         }
 406
 407         /*
 408          * Either request is already in userspace, or it was forced.
 409          * Wait it out.
 410          */
 411         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 412 }
 413
 414 static void __fuse_request_send(struct fuse_req *req)
 415 {
 416         struct fuse_iqueue *fiq = &req->fm->fc->iq;
 417
 418         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 419         spin_lock(&fiq->lock);
 420         if (!fiq->connected) {
 421                 spin_unlock(&fiq->lock);
 422                 req->out.h.error = -ENOTCONN;
 423         } else {
 424                 req->in.h.unique = fuse_get_unique(fiq);
 425                 /* acquire extra reference, since request is still needed
 426                    after fuse_request_end() */
 427                 __fuse_get_request(req);
 428                 queue_request_and_unlock(fiq, req);
 429
 430                 request_wait_answer(req);
 431                 /* Pairs with smp_wmb() in fuse_request_end() */
 432                 smp_rmb();
 433         }
 434 }
 435
 436 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 437 {
 438         if (fc->minor < 4 && args->opcode == FUSE_STATFS)
 439                 args->out_args[0].size = FUSE_COMPAT_STATFS_SIZE;
 440
 441         if (fc->minor < 9) {
 442                 switch (args->opcode) {
 443                 case FUSE_LOOKUP:
 444                 case FUSE_CREATE:
 445                 case FUSE_MKNOD:
 446                 case FUSE_MKDIR:
 447                 case FUSE_SYMLINK:
 448                 case FUSE_LINK:
 449                         args->out_args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 450                         break;
 451                 case FUSE_GETATTR:
 452                 case FUSE_SETATTR:
 453                         args->out_args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 454                         break;
 455                 }
 456         }
 457         if (fc->minor < 12) {
 458                 switch (args->opcode) {
 459                 case FUSE_CREATE:
 460                         args->in_args[0].size = sizeof(struct fuse_open_in);
 461                         break;
 462                 case FUSE_MKNOD:
 463                         args->in_args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 464                         break;
 465                 }
 466         }
 467 }
 468
 469 static void fuse_force_creds(struct fuse_req *req)
 470 {
 471         struct fuse_conn *fc = req->fm->fc;
 472
 473         req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
 474         req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
 475         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 476 }
 477
 478 static void fuse_args_to_req(struct fuse_req *req, struct fuse_args *args)
 479 {
 480         req->in.h.opcode = args->opcode;
 481         req->in.h.nodeid = args->nodeid;
 482         req->args = args;
 483         if (args->end)
 484                 __set_bit(FR_ASYNC, &req->flags);
 485 }
 486
 487 ssize_t fuse_simple_request(struct fuse_mount *fm, struct fuse_args *args)
 488 {
 489         struct fuse_conn *fc = fm->fc;
 490         struct fuse_req *req;
 491         ssize_t ret;
 492
 493         if (args->force) {
 494                 atomic_inc(&fc->num_waiting);
 495                 req = fuse_request_alloc(fm, GFP_KERNEL | __GFP_NOFAIL);
 496
 497                 if (!args->nocreds)
 498                         fuse_force_creds(req);
 499
 500                 __set_bit(FR_WAITING, &req->flags);
 501                 __set_bit(FR_FORCE, &req->flags);
 502         } else {
 503                 WARN_ON(args->nocreds);
 504                 req = fuse_get_req(fm, false);
 505                 if (IS_ERR(req))
 506                         return PTR_ERR(req);
 507         }
 508
 509         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 510         fuse_adjust_compat(fc, args);
 511         fuse_args_to_req(req, args);
 512
 513         if (!args->noreply)
 514                 __set_bit(FR_ISREPLY, &req->flags);
 515         __fuse_request_send(req);
 516         ret = req->out.h.error;
 517         if (!ret && args->out_argvar) {
 518                 BUG_ON(args->out_numargs == 0);
 519                 ret = args->out_args[args->out_numargs - 1].size;
 520         }
 521         fuse_put_request(req);
 522
 523         return ret;
 524 }
 525
 526 static bool fuse_request_queue_background(struct fuse_req *req)
 527 {
 528         struct fuse_mount *fm = req->fm;
 529         struct fuse_conn *fc = fm->fc;
 530         bool queued = false;
 531
 532         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 533         if (!test_bit(FR_WAITING, &req->flags)) {
 534                 __set_bit(FR_WAITING, &req->flags);
 535                 atomic_inc(&fc->num_waiting);
 536         }
 537         __set_bit(FR_ISREPLY, &req->flags);
 538         spin_lock(&fc->bg_lock);
 539         if (likely(fc->connected)) {
 540                 fc->num_background++;
 541                 if (fc->num_background == fc->max_background)
 542                         fc->blocked = 1;
 543                 if (fc->num_background == fc->congestion_threshold && fm->sb) {
 544                         set_bdi_congested(fm->sb->s_bdi, BLK_RW_SYNC);
 545                         set_bdi_congested(fm->sb->s_bdi, BLK_RW_ASYNC);
 546                 }
 547                 list_add_tail(&req->list, &fc->bg_queue);
 548                 flush_bg_queue(fc);
 549                 queued = true;
 550         }
 551         spin_unlock(&fc->bg_lock);
 552
 553         return queued;
 554 }
 555
 556 int fuse_simple_background(struct fuse_mount *fm, struct fuse_args *args,
 557                             gfp_t gfp_flags)
 558 {
 559         struct fuse_req *req;
 560
 561         if (args->force) {
 562                 WARN_ON(!args->nocreds);
 563                 req = fuse_request_alloc(fm, gfp_flags);
 564                 if (!req)
 565                         return -ENOMEM;
 566                 __set_bit(FR_BACKGROUND, &req->flags);
 567         } else {
 568                 WARN_ON(args->nocreds);
 569                 req = fuse_get_req(fm, true);
 570                 if (IS_ERR(req))
 571                         return PTR_ERR(req);
 572         }
 573
 574         fuse_args_to_req(req, args);
 575
 576         if (!fuse_request_queue_background(req)) {
 577                 fuse_put_request(req);
 578                 return -ENOTCONN;
 579         }
 580
 581         return 0;
 582 }
 583 EXPORT_SYMBOL_GPL(fuse_simple_background);
 584
 585 static int fuse_simple_notify_reply(struct fuse_mount *fm,
 586                                     struct fuse_args *args, u64 unique)
 587 {
 588         struct fuse_req *req;
 589         struct fuse_iqueue *fiq = &fm->fc->iq;
 590         int err = 0;
 591
 592         req = fuse_get_req(fm, false);
 593         if (IS_ERR(req))
 594                 return PTR_ERR(req);
 595
 596         __clear_bit(FR_ISREPLY, &req->flags);
 597         req->in.h.unique = unique;
 598
 599         fuse_args_to_req(req, args);
 600
 601         spin_lock(&fiq->lock);
 602         if (fiq->connected) {
 603                 queue_request_and_unlock(fiq, req);
 604         } else {
 605                 err = -ENODEV;
 606                 spin_unlock(&fiq->lock);
 607                 fuse_put_request(req);
 608         }
 609
 610         return err;
 611 }
 612
 613 /*
 614  * Lock the request.  Up to the next unlock_request() there mustn't be
 615  * anything that could cause a page-fault.  If the request was already
 616  * aborted bail out.
 617  */
 618 static int lock_request(struct fuse_req *req)
 619 {
 620         int err = 0;
 621         if (req) {
 622                 spin_lock(&req->waitq.lock);
 623                 if (test_bit(FR_ABORTED, &req->flags))
 624                         err = -ENOENT;
 625                 else
 626                         set_bit(FR_LOCKED, &req->flags);
 627                 spin_unlock(&req->waitq.lock);
 628         }
 629         return err;
 630 }
 631
 632 /*
 633  * Unlock request.  If it was aborted while locked, caller is responsible
 634  * for unlocking and ending the request.
 635  */
 636 static int unlock_request(struct fuse_req *req)
 637 {
 638         int err = 0;
 639         if (req) {
 640                 spin_lock(&req->waitq.lock);
 641                 if (test_bit(FR_ABORTED, &req->flags))
 642                         err = -ENOENT;
 643                 else
 644                         clear_bit(FR_LOCKED, &req->flags);
 645                 spin_unlock(&req->waitq.lock);
 646         }
 647         return err;
 648 }
 649
 650 struct fuse_copy_state {
 651         int write;
 652         struct fuse_req *req;
 653         struct iov_iter *iter;
 654         struct pipe_buffer *pipebufs;
 655         struct pipe_buffer *currbuf;
 656         struct pipe_inode_info *pipe;
 657         unsigned long nr_segs;
 658         struct page *pg;
 659         unsigned len;
 660         unsigned offset;
 661         unsigned move_pages:1;
 662 };
 663
 664 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 665                            struct iov_iter *iter)
 666 {
 667         memset(cs, 0, sizeof(*cs));
 668         cs->write = write;
 669         cs->iter = iter;
 670 }
 671
 672 /* Unmap and put previous page of userspace buffer */
 673 static void fuse_copy_finish(struct fuse_copy_state *cs)
 674 {
 675         if (cs->currbuf) {
 676                 struct pipe_buffer *buf = cs->currbuf;
 677
 678                 if (cs->write)
 679                         buf->len = PAGE_SIZE - cs->len;
 680                 cs->currbuf = NULL;
 681         } else if (cs->pg) {
 682                 if (cs->write) {
 683                         flush_dcache_page(cs->pg);
 684                         set_page_dirty_lock(cs->pg);
 685                 }
 686                 put_page(cs->pg);
 687         }
 688         cs->pg = NULL;
 689 }
 690
 691 /*
 692  * Get another pagefull of userspace buffer, and map it to kernel
 693  * address space, and lock request
 694  */
 695 static int fuse_copy_fill(struct fuse_copy_state *cs)
 696 {
 697         struct page *page;
 698         int err;
 699
 700         err = unlock_request(cs->req);
 701         if (err)
 702                 return err;
 703
 704         fuse_copy_finish(cs);
 705         if (cs->pipebufs) {
 706                 struct pipe_buffer *buf = cs->pipebufs;
 707
 708                 if (!cs->write) {
 709                         err = pipe_buf_confirm(cs->pipe, buf);
 710                         if (err)
 711                                 return err;
 712
 713                         BUG_ON(!cs->nr_segs);
 714                         cs->currbuf = buf;
 715                         cs->pg = buf->page;
 716                         cs->offset = buf->offset;
 717                         cs->len = buf->len;
 718                         cs->pipebufs++;
 719                         cs->nr_segs--;
 720                 } else {
 721                         if (cs->nr_segs >= cs->pipe->max_usage)
 722                                 return -EIO;
 723
 724                         page = alloc_page(GFP_HIGHUSER);
 725                         if (!page)
 726                                 return -ENOMEM;
 727
 728                         buf->page = page;
 729                         buf->offset = 0;
 730                         buf->len = 0;
 731
 732                         cs->currbuf = buf;
 733                         cs->pg = page;
 734                         cs->offset = 0;
 735                         cs->len = PAGE_SIZE;
 736                         cs->pipebufs++;
 737                         cs->nr_segs++;
 738                 }
 739         } else {
 740                 size_t off;
 741                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 742                 if (err < 0)
 743                         return err;
 744                 BUG_ON(!err);
 745                 cs->len = err;
 746                 cs->offset = off;
 747                 cs->pg = page;
 748                 iov_iter_advance(cs->iter, err);
 749         }
 750
 751         return lock_request(cs->req);
 752 }
 753
 754 /* Do as much copy to/from userspace buffer as we can */
 755 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 756 {
 757         unsigned ncpy = min(*size, cs->len);
 758         if (val) {
 759                 void *pgaddr = kmap_atomic(cs->pg);
 760                 void *buf = pgaddr + cs->offset;
 761
 762                 if (cs->write)
 763                         memcpy(buf, *val, ncpy);
 764                 else
 765                         memcpy(*val, buf, ncpy);
 766
 767                 kunmap_atomic(pgaddr);
 768                 *val += ncpy;
 769         }
 770         *size -= ncpy;
 771         cs->len -= ncpy;
 772         cs->offset += ncpy;
 773         return ncpy;
 774 }
 775
 776 static int fuse_check_page(struct page *page)
 777 {
 778         if (page_mapcount(page) ||
 779             page->mapping != NULL ||
 780             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 781              ~(1 << PG_locked |
 782                1 << PG_referenced |
 783                1 << PG_uptodate |
 784                1 << PG_lru |
 785                1 << PG_active |
 786                1 << PG_workingset |
 787                1 << PG_reclaim |
 788                1 << PG_waiters))) {
 789                 dump_page(page, "fuse: trying to steal weird page");
 790                 return 1;
 791         }
 792         return 0;
 793 }
 794
 795 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 796 {
 797         int err;
 798         struct page *oldpage = *pagep;
 799         struct page *newpage;
 800         struct pipe_buffer *buf = cs->pipebufs;
 801
 802         get_page(oldpage);
 803         err = unlock_request(cs->req);
 804         if (err)
 805                 goto out_put_old;
 806
 807         fuse_copy_finish(cs);
 808
 809         err = pipe_buf_confirm(cs->pipe, buf);
 810         if (err)
 811                 goto out_put_old;
 812
 813         BUG_ON(!cs->nr_segs);
 814         cs->currbuf = buf;
 815         cs->len = buf->len;
 816         cs->pipebufs++;
 817         cs->nr_segs--;
 818
 819         if (cs->len != PAGE_SIZE)
 820                 goto out_fallback;
 821
 822         if (!pipe_buf_try_steal(cs->pipe, buf))
 823                 goto out_fallback;
 824
 825         newpage = buf->page;
 826
 827         if (!PageUptodate(newpage))
 828                 SetPageUptodate(newpage);
 829
 830         ClearPageMappedToDisk(newpage);
 831
 832         if (fuse_check_page(newpage) != 0)
 833                 goto out_fallback_unlock;
 834
 835         /*
 836          * This is a new and locked page, it shouldn't be mapped or
 837          * have any special flags on it
 838          */
 839         if (WARN_ON(page_mapped(oldpage)))
 840                 goto out_fallback_unlock;
 841         if (WARN_ON(page_has_private(oldpage)))
 842                 goto out_fallback_unlock;
 843         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 844                 goto out_fallback_unlock;
 845         if (WARN_ON(PageMlocked(oldpage)))
 846                 goto out_fallback_unlock;
 847
 848         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 849         if (err) {
 850                 unlock_page(newpage);
 851                 goto out_put_old;
 852         }
 853
 854         get_page(newpage);
 855
 856         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 857                 lru_cache_add(newpage);
 858
 859         /*
 860          * Release while we have extra ref on stolen page.  Otherwise
 861          * anon_pipe_buf_release() might think the page can be reused.
 862          */
 863         pipe_buf_release(cs->pipe, buf);
 864
 865         err = 0;
 866         spin_lock(&cs->req->waitq.lock);
 867         if (test_bit(FR_ABORTED, &cs->req->flags))
 868                 err = -ENOENT;
 869         else
 870                 *pagep = newpage;
 871         spin_unlock(&cs->req->waitq.lock);
 872
 873         if (err) {
 874                 unlock_page(newpage);
 875                 put_page(newpage);
 876                 goto out_put_old;
 877         }
 878
 879         unlock_page(oldpage);
 880         /* Drop ref for ap->pages[] array */
 881         put_page(oldpage);
 882         cs->len = 0;
 883
 884         err = 0;
 885 out_put_old:
 886         /* Drop ref obtained in this function */
 887         put_page(oldpage);
 888         return err;
 889
 890 out_fallback_unlock:
 891         unlock_page(newpage);
 892 out_fallback:
 893         cs->pg = buf->page;
 894         cs->offset = buf->offset;
 895
 896         err = lock_request(cs->req);
 897         if (!err)
 898                 err = 1;
 899
 900         goto out_put_old;
 901 }
 902
 903 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 904                          unsigned offset, unsigned count)
 905 {
 906         struct pipe_buffer *buf;
 907         int err;
 908
 909         if (cs->nr_segs >= cs->pipe->max_usage)
 910                 return -EIO;
 911
 912         get_page(page);
 913         err = unlock_request(cs->req);
 914         if (err) {
 915                 put_page(page);
 916                 return err;
 917         }
 918
 919         fuse_copy_finish(cs);
 920
 921         buf = cs->pipebufs;
 922         buf->page = page;
 923         buf->offset = offset;
 924         buf->len = count;
 925
 926         cs->pipebufs++;
 927         cs->nr_segs++;
 928         cs->len = 0;
 929
 930         return 0;
 931 }
 932
 933 /*
 934  * Copy a page in the request to/from the userspace buffer.  Must be
 935  * done atomically
 936  */
 937 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 938                           unsigned offset, unsigned count, int zeroing)
 939 {
 940         int err;
 941         struct page *page = *pagep;
 942
 943         if (page && zeroing && count < PAGE_SIZE)
 944                 clear_highpage(page);
 945
 946         while (count) {
 947                 if (cs->write && cs->pipebufs && page) {
 948                         /*
 949                          * Can't control lifetime of pipe buffers, so always
 950                          * copy user pages.
 951                          */
 952                         if (cs->req->args->user_pages) {
 953                                 err = fuse_copy_fill(cs);
 954                                 if (err)
 955                                         return err;
 956                         } else {
 957                                 return fuse_ref_page(cs, page, offset, count);
 958                         }
 959                 } else if (!cs->len) {
 960                         if (cs->move_pages && page &&
 961                             offset == 0 && count == PAGE_SIZE) {
 962                                 err = fuse_try_move_page(cs, pagep);
 963                                 if (err <= 0)
 964                                         return err;
 965                         } else {
 966                                 err = fuse_copy_fill(cs);
 967                                 if (err)
 968                                         return err;
 969                         }
 970                 }
 971                 if (page) {
 972                         void *mapaddr = kmap_atomic(page);
 973                         void *buf = mapaddr + offset;
 974                         offset += fuse_copy_do(cs, &buf, &count);
 975                         kunmap_atomic(mapaddr);
 976                 } else
 977                         offset += fuse_copy_do(cs, NULL, &count);
 978         }
 979         if (page && !cs->write)
 980                 flush_dcache_page(page);
 981         return 0;
 982 }
 983
 984 /* Copy pages in the request to/from userspace buffer */
 985 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 986                            int zeroing)
 987 {
 988         unsigned i;
 989         struct fuse_req *req = cs->req;
 990         struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
 991
 992
 993         for (i = 0; i < ap->num_pages && (nbytes || zeroing); i++) {
 994                 int err;
 995                 unsigned int offset = ap->descs[i].offset;
 996                 unsigned int count = min(nbytes, ap->descs[i].length);
 997
 998                 err = fuse_copy_page(cs, &ap->pages[i], offset, count, zeroing);
 999                 if (err)
1000                         return err;
1001
1002                 nbytes -= count;
1003         }
1004         return 0;
1005 }
1006
1007 /* Copy a single argument in the request to/from userspace buffer */
1008 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1009 {
1010         while (size) {
1011                 if (!cs->len) {
1012                         int err = fuse_copy_fill(cs);
1013                         if (err)
1014                                 return err;
1015                 }
1016                 fuse_copy_do(cs, &val, &size);
1017         }
1018         return 0;
1019 }
1020
1021 /* Copy request arguments to/from userspace buffer */
1022 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1023                           unsigned argpages, struct fuse_arg *args,
1024                           int zeroing)
1025 {
1026         int err = 0;
1027         unsigned i;
1028
1029         for (i = 0; !err && i < numargs; i++)  {
1030                 struct fuse_arg *arg = &args[i];
1031                 if (i == numargs - 1 && argpages)
1032                         err = fuse_copy_pages(cs, arg->size, zeroing);
1033                 else
1034                         err = fuse_copy_one(cs, arg->value, arg->size);
1035         }
1036         return err;
1037 }
1038
1039 static int forget_pending(struct fuse_iqueue *fiq)
1040 {
1041         return fiq->forget_list_head.next != NULL;
1042 }
1043
1044 static int request_pending(struct fuse_iqueue *fiq)
1045 {
1046         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1047                 forget_pending(fiq);
1048 }
1049
1050 /*
1051  * Transfer an interrupt request to userspace
1052  *
1053  * Unlike other requests this is assembled on demand, without a need
1054  * to allocate a separate fuse_req structure.
1055  *
1056  * Called with fiq->lock held, releases it
1057  */
1058 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1059                                struct fuse_copy_state *cs,
1060                                size_t nbytes, struct fuse_req *req)
1061 __releases(fiq->lock)
1062 {
1063         struct fuse_in_header ih;
1064         struct fuse_interrupt_in arg;
1065         unsigned reqsize = sizeof(ih) + sizeof(arg);
1066         int err;
1067
1068         list_del_init(&req->intr_entry);
1069         memset(&ih, 0, sizeof(ih));
1070         memset(&arg, 0, sizeof(arg));
1071         ih.len = reqsize;
1072         ih.opcode = FUSE_INTERRUPT;
1073         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1074         arg.unique = req->in.h.unique;
1075
1076         spin_unlock(&fiq->lock);
1077         if (nbytes < reqsize)
1078                 return -EINVAL;
1079
1080         err = fuse_copy_one(cs, &ih, sizeof(ih));
1081         if (!err)
1082                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1083         fuse_copy_finish(cs);
1084
1085         return err ? err : reqsize;
1086 }
1087
1088 struct fuse_forget_link *fuse_dequeue_forget(struct fuse_iqueue *fiq,
1089                                              unsigned int max,
1090                                              unsigned int *countp)
1091 {
1092         struct fuse_forget_link *head = fiq->forget_list_head.next;
1093         struct fuse_forget_link **newhead = &head;
1094         unsigned count;
1095
1096         for (count = 0; *newhead != NULL && count < max; count++)
1097                 newhead = &(*newhead)->next;
1098
1099         fiq->forget_list_head.next = *newhead;
1100         *newhead = NULL;
1101         if (fiq->forget_list_head.next == NULL)
1102                 fiq->forget_list_tail = &fiq->forget_list_head;
1103
1104         if (countp != NULL)
1105                 *countp = count;
1106
1107         return head;
1108 }
1109 EXPORT_SYMBOL(fuse_dequeue_forget);
1110
1111 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1112                                    struct fuse_copy_state *cs,
1113                                    size_t nbytes)
1114 __releases(fiq->lock)
1115 {
1116         int err;
1117         struct fuse_forget_link *forget = fuse_dequeue_forget(fiq, 1, NULL);
1118         struct fuse_forget_in arg = {
1119                 .nlookup = forget->forget_one.nlookup,
1120         };
1121         struct fuse_in_header ih = {
1122                 .opcode = FUSE_FORGET,
1123                 .nodeid = forget->forget_one.nodeid,
1124                 .unique = fuse_get_unique(fiq),
1125                 .len = sizeof(ih) + sizeof(arg),
1126         };
1127
1128         spin_unlock(&fiq->lock);
1129         kfree(forget);
1130         if (nbytes < ih.len)
1131                 return -EINVAL;
1132
1133         err = fuse_copy_one(cs, &ih, sizeof(ih));
1134         if (!err)
1135                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1136         fuse_copy_finish(cs);
1137
1138         if (err)
1139                 return err;
1140
1141         return ih.len;
1142 }
1143
1144 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1145                                    struct fuse_copy_state *cs, size_t nbytes)
1146 __releases(fiq->lock)
1147 {
1148         int err;
1149         unsigned max_forgets;
1150         unsigned count;
1151         struct fuse_forget_link *head;
1152         struct fuse_batch_forget_in arg = { .count = 0 };
1153         struct fuse_in_header ih = {
1154                 .opcode = FUSE_BATCH_FORGET,
1155                 .unique = fuse_get_unique(fiq),
1156                 .len = sizeof(ih) + sizeof(arg),
1157         };
1158
1159         if (nbytes < ih.len) {
1160                 spin_unlock(&fiq->lock);
1161                 return -EINVAL;
1162         }
1163
1164         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1165         head = fuse_dequeue_forget(fiq, max_forgets, &count);
1166         spin_unlock(&fiq->lock);
1167
1168         arg.count = count;
1169         ih.len += count * sizeof(struct fuse_forget_one);
1170         err = fuse_copy_one(cs, &ih, sizeof(ih));
1171         if (!err)
1172                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1173
1174         while (head) {
1175                 struct fuse_forget_link *forget = head;
1176
1177                 if (!err) {
1178                         err = fuse_copy_one(cs, &forget->forget_one,
1179                                             sizeof(forget->forget_one));
1180                 }
1181                 head = forget->next;
1182                 kfree(forget);
1183         }
1184
1185         fuse_copy_finish(cs);
1186
1187         if (err)
1188                 return err;
1189
1190         return ih.len;
1191 }
1192
1193 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1194                             struct fuse_copy_state *cs,
1195                             size_t nbytes)
1196 __releases(fiq->lock)
1197 {
1198         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1199                 return fuse_read_single_forget(fiq, cs, nbytes);
1200         else
1201                 return fuse_read_batch_forget(fiq, cs, nbytes);
1202 }
1203
1204 /*
1205  * Read a single request into the userspace filesystem's buffer.  This
1206  * function waits until a request is available, then removes it from
1207  * the pending list and copies request data to userspace buffer.  If
1208  * no reply is needed (FORGET) or request has been aborted or there
1209  * was an error during the copying then it's finished by calling
1210  * fuse_request_end().  Otherwise add it to the processing list, and set
1211  * the 'sent' flag.
1212  */
1213 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1214                                 struct fuse_copy_state *cs, size_t nbytes)
1215 {
1216         ssize_t err;
1217         struct fuse_conn *fc = fud->fc;
1218         struct fuse_iqueue *fiq = &fc->iq;
1219         struct fuse_pqueue *fpq = &fud->pq;
1220         struct fuse_req *req;
1221         struct fuse_args *args;
1222         unsigned reqsize;
1223         unsigned int hash;
1224
1225         /*
1226          * Require sane minimum read buffer - that has capacity for fixed part
1227          * of any request header + negotiated max_write room for data.
1228          *
1229          * Historically libfuse reserves 4K for fixed header room, but e.g.
1230          * GlusterFS reserves only 80 bytes
1231          *
1232          *      = `sizeof(fuse_in_header) + sizeof(fuse_write_in)`
1233          *
1234          * which is the absolute minimum any sane filesystem should be using
1235          * for header room.
1236          */
1237         if (nbytes < max_t(size_t, FUSE_MIN_READ_BUFFER,
1238                            sizeof(struct fuse_in_header) +
1239                            sizeof(struct fuse_write_in) +
1240                            fc->max_write))
1241                 return -EINVAL;
1242
1243  restart:
1244         for (;;) {
1245                 spin_lock(&fiq->lock);
1246                 if (!fiq->connected || request_pending(fiq))
1247                         break;
1248                 spin_unlock(&fiq->lock);
1249
1250                 if (file->f_flags & O_NONBLOCK)
1251                         return -EAGAIN;
1252                 err = wait_event_interruptible_exclusive(fiq->waitq,
1253                                 !fiq->connected || request_pending(fiq));
1254                 if (err)
1255                         return err;
1256         }
1257
1258         if (!fiq->connected) {
1259                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1260                 goto err_unlock;
1261         }
1262
1263         if (!list_empty(&fiq->interrupts)) {
1264                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1265                                  intr_entry);
1266                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1267         }
1268
1269         if (forget_pending(fiq)) {
1270                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1271                         return fuse_read_forget(fc, fiq, cs, nbytes);
1272
1273                 if (fiq->forget_batch <= -8)
1274                         fiq->forget_batch = 16;
1275         }
1276
1277         req = list_entry(fiq->pending.next, struct fuse_req, list);
1278         clear_bit(FR_PENDING, &req->flags);
1279         list_del_init(&req->list);
1280         spin_unlock(&fiq->lock);
1281
1282         args = req->args;
1283         reqsize = req->in.h.len;
1284
1285         /* If request is too large, reply with an error and restart the read */
1286         if (nbytes < reqsize) {
1287                 req->out.h.error = -EIO;
1288                 /* SETXATTR is special, since it may contain too large data */
1289                 if (args->opcode == FUSE_SETXATTR)
1290                         req->out.h.error = -E2BIG;
1291                 fuse_request_end(req);
1292                 goto restart;
1293         }
1294         spin_lock(&fpq->lock);
1295         /*
1296          *  Must not put request on fpq->io queue after having been shut down by
1297          *  fuse_abort_conn()
1298          */
1299         if (!fpq->connected) {
1300                 req->out.h.error = err = -ECONNABORTED;
1301                 goto out_end;
1302
1303         }
1304         list_add(&req->list, &fpq->io);
1305         spin_unlock(&fpq->lock);
1306         cs->req = req;
1307         err = fuse_copy_one(cs, &req->in.h, sizeof(req->in.h));
1308         if (!err)
1309                 err = fuse_copy_args(cs, args->in_numargs, args->in_pages,
1310                                      (struct fuse_arg *) args->in_args, 0);
1311         fuse_copy_finish(cs);
1312         spin_lock(&fpq->lock);
1313         clear_bit(FR_LOCKED, &req->flags);
1314         if (!fpq->connected) {
1315                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1316                 goto out_end;
1317         }
1318         if (err) {
1319                 req->out.h.error = -EIO;
1320                 goto out_end;
1321         }
1322         if (!test_bit(FR_ISREPLY, &req->flags)) {
1323                 err = reqsize;
1324                 goto out_end;
1325         }
1326         hash = fuse_req_hash(req->in.h.unique);
1327         list_move_tail(&req->list, &fpq->processing[hash]);
1328         __fuse_get_request(req);
1329         set_bit(FR_SENT, &req->flags);
1330         spin_unlock(&fpq->lock);
1331         /* matches barrier in request_wait_answer() */
1332         smp_mb__after_atomic();
1333         if (test_bit(FR_INTERRUPTED, &req->flags))
1334                 queue_interrupt(req);
1335         fuse_put_request(req);
1336
1337         return reqsize;
1338
1339 out_end:
1340         if (!test_bit(FR_PRIVATE, &req->flags))
1341                 list_del_init(&req->list);
1342         spin_unlock(&fpq->lock);
1343         fuse_request_end(req);
1344         return err;
1345
1346  err_unlock:
1347         spin_unlock(&fiq->lock);
1348         return err;
1349 }
1350
1351 static int fuse_dev_open(struct inode *inode, struct file *file)
1352 {
1353         /*
1354          * The fuse device's file's private_data is used to hold
1355          * the fuse_conn(ection) when it is mounted, and is used to
1356          * keep track of whether the file has been mounted already.
1357          */
1358         file->private_data = NULL;
1359         return 0;
1360 }
1361
1362 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1363 {
1364         struct fuse_copy_state cs;
1365         struct file *file = iocb->ki_filp;
1366         struct fuse_dev *fud = fuse_get_dev(file);
1367
1368         if (!fud)
1369                 return -EPERM;
1370
1371         if (!iter_is_iovec(to))
1372                 return -EINVAL;
1373
1374         fuse_copy_init(&cs, 1, to);
1375
1376         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1377 }
1378
1379 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1380                                     struct pipe_inode_info *pipe,
1381                                     size_t len, unsigned int flags)
1382 {
1383         int total, ret;
1384         int page_nr = 0;
1385         struct pipe_buffer *bufs;
1386         struct fuse_copy_state cs;
1387         struct fuse_dev *fud = fuse_get_dev(in);
1388
1389         if (!fud)
1390                 return -EPERM;
1391
1392         bufs = kvmalloc_array(pipe->max_usage, sizeof(struct pipe_buffer),
1393                               GFP_KERNEL);
1394         if (!bufs)
1395                 return -ENOMEM;
1396
1397         fuse_copy_init(&cs, 1, NULL);
1398         cs.pipebufs = bufs;
1399         cs.pipe = pipe;
1400         ret = fuse_dev_do_read(fud, in, &cs, len);
1401         if (ret < 0)
1402                 goto out;
1403
1404         if (pipe_occupancy(pipe->head, pipe->tail) + cs.nr_segs > pipe->max_usage) {
1405                 ret = -EIO;
1406                 goto out;
1407         }
1408
1409         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1410                 /*
1411                  * Need to be careful about this.  Having buf->ops in module
1412                  * code can Oops if the buffer persists after module unload.
1413                  */
1414                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1415                 bufs[page_nr].flags = 0;
1416                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1417                 if (unlikely(ret < 0))
1418                         break;
1419         }
1420         if (total)
1421                 ret = total;
1422 out:
1423         for (; page_nr < cs.nr_segs; page_nr++)
1424                 put_page(bufs[page_nr].page);
1425
1426         kvfree(bufs);
1427         return ret;
1428 }
1429
1430 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1431                             struct fuse_copy_state *cs)
1432 {
1433         struct fuse_notify_poll_wakeup_out outarg;
1434         int err = -EINVAL;
1435
1436         if (size != sizeof(outarg))
1437                 goto err;
1438
1439         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1440         if (err)
1441                 goto err;
1442
1443         fuse_copy_finish(cs);
1444         return fuse_notify_poll_wakeup(fc, &outarg);
1445
1446 err:
1447         fuse_copy_finish(cs);
1448         return err;
1449 }
1450
1451 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1452                                    struct fuse_copy_state *cs)
1453 {
1454         struct fuse_notify_inval_inode_out outarg;
1455         int err = -EINVAL;
1456
1457         if (size != sizeof(outarg))
1458                 goto err;
1459
1460         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1461         if (err)
1462                 goto err;
1463         fuse_copy_finish(cs);
1464
1465         down_read(&fc->killsb);
1466         err = fuse_reverse_inval_inode(fc, outarg.ino,
1467                                        outarg.off, outarg.len);
1468         up_read(&fc->killsb);
1469         return err;
1470
1471 err:
1472         fuse_copy_finish(cs);
1473         return err;
1474 }
1475
1476 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1477                                    struct fuse_copy_state *cs)
1478 {
1479         struct fuse_notify_inval_entry_out outarg;
1480         int err = -ENOMEM;
1481         char *buf;
1482         struct qstr name;
1483
1484         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1485         if (!buf)
1486                 goto err;
1487
1488         err = -EINVAL;
1489         if (size < sizeof(outarg))
1490                 goto err;
1491
1492         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1493         if (err)
1494                 goto err;
1495
1496         err = -ENAMETOOLONG;
1497         if (outarg.namelen > FUSE_NAME_MAX)
1498                 goto err;
1499
1500         err = -EINVAL;
1501         if (size != sizeof(outarg) + outarg.namelen + 1)
1502                 goto err;
1503
1504         name.name = buf;
1505         name.len = outarg.namelen;
1506         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1507         if (err)
1508                 goto err;
1509         fuse_copy_finish(cs);
1510         buf[outarg.namelen] = 0;
1511
1512         down_read(&fc->killsb);
1513         err = fuse_reverse_inval_entry(fc, outarg.parent, 0, &name);
1514         up_read(&fc->killsb);
1515         kfree(buf);
1516         return err;
1517
1518 err:
1519         kfree(buf);
1520         fuse_copy_finish(cs);
1521         return err;
1522 }
1523
1524 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1525                               struct fuse_copy_state *cs)
1526 {
1527         struct fuse_notify_delete_out outarg;
1528         int err = -ENOMEM;
1529         char *buf;
1530         struct qstr name;
1531
1532         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1533         if (!buf)
1534                 goto err;
1535
1536         err = -EINVAL;
1537         if (size < sizeof(outarg))
1538                 goto err;
1539
1540         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1541         if (err)
1542                 goto err;
1543
1544         err = -ENAMETOOLONG;
1545         if (outarg.namelen > FUSE_NAME_MAX)
1546                 goto err;
1547
1548         err = -EINVAL;
1549         if (size != sizeof(outarg) + outarg.namelen + 1)
1550                 goto err;
1551
1552         name.name = buf;
1553         name.len = outarg.namelen;
1554         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1555         if (err)
1556                 goto err;
1557         fuse_copy_finish(cs);
1558         buf[outarg.namelen] = 0;
1559
1560         down_read(&fc->killsb);
1561         err = fuse_reverse_inval_entry(fc, outarg.parent, outarg.child, &name);
1562         up_read(&fc->killsb);
1563         kfree(buf);
1564         return err;
1565
1566 err:
1567         kfree(buf);
1568         fuse_copy_finish(cs);
1569         return err;
1570 }
1571
1572 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1573                              struct fuse_copy_state *cs)
1574 {
1575         struct fuse_notify_store_out outarg;
1576         struct inode *inode;
1577         struct address_space *mapping;
1578         u64 nodeid;
1579         int err;
1580         pgoff_t index;
1581         unsigned int offset;
1582         unsigned int num;
1583         loff_t file_size;
1584         loff_t end;
1585
1586         err = -EINVAL;
1587         if (size < sizeof(outarg))
1588                 goto out_finish;
1589
1590         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1591         if (err)
1592                 goto out_finish;
1593
1594         err = -EINVAL;
1595         if (size - sizeof(outarg) != outarg.size)
1596                 goto out_finish;
1597
1598         nodeid = outarg.nodeid;
1599
1600         down_read(&fc->killsb);
1601
1602         err = -ENOENT;
1603         inode = fuse_ilookup(fc, nodeid,  NULL);
1604         if (!inode)
1605                 goto out_up_killsb;
1606
1607         mapping = inode->i_mapping;
1608         index = outarg.offset >> PAGE_SHIFT;
1609         offset = outarg.offset & ~PAGE_MASK;
1610         file_size = i_size_read(inode);
1611         end = outarg.offset + outarg.size;
1612         if (end > file_size) {
1613                 file_size = end;
1614                 fuse_write_update_size(inode, file_size);
1615         }
1616
1617         num = outarg.size;
1618         while (num) {
1619                 struct page *page;
1620                 unsigned int this_num;
1621
1622                 err = -ENOMEM;
1623                 page = find_or_create_page(mapping, index,
1624                                            mapping_gfp_mask(mapping));
1625                 if (!page)
1626                         goto out_iput;
1627
1628                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1629                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1630                 if (!err && offset == 0 &&
1631                     (this_num == PAGE_SIZE || file_size == end))
1632                         SetPageUptodate(page);
1633                 unlock_page(page);
1634                 put_page(page);
1635
1636                 if (err)
1637                         goto out_iput;
1638
1639                 num -= this_num;
1640                 offset = 0;
1641                 index++;
1642         }
1643
1644         err = 0;
1645
1646 out_iput:
1647         iput(inode);
1648 out_up_killsb:
1649         up_read(&fc->killsb);
1650 out_finish:
1651         fuse_copy_finish(cs);
1652         return err;
1653 }
1654
1655 struct fuse_retrieve_args {
1656         struct fuse_args_pages ap;
1657         struct fuse_notify_retrieve_in inarg;
1658 };
1659
1660 static void fuse_retrieve_end(struct fuse_mount *fm, struct fuse_args *args,
1661                               int error)
1662 {
1663         struct fuse_retrieve_args *ra =
1664                 container_of(args, typeof(*ra), ap.args);
1665
1666         release_pages(ra->ap.pages, ra->ap.num_pages);
1667         kfree(ra);
1668 }
1669
1670 static int fuse_retrieve(struct fuse_mount *fm, struct inode *inode,
1671                          struct fuse_notify_retrieve_out *outarg)
1672 {
1673         int err;
1674         struct address_space *mapping = inode->i_mapping;
1675         pgoff_t index;
1676         loff_t file_size;
1677         unsigned int num;
1678         unsigned int offset;
1679         size_t total_len = 0;
1680         unsigned int num_pages;
1681         struct fuse_conn *fc = fm->fc;
1682         struct fuse_retrieve_args *ra;
1683         size_t args_size = sizeof(*ra);
1684         struct fuse_args_pages *ap;
1685         struct fuse_args *args;
1686
1687         offset = outarg->offset & ~PAGE_MASK;
1688         file_size = i_size_read(inode);
1689
1690         num = min(outarg->size, fc->max_write);
1691         if (outarg->offset > file_size)
1692                 num = 0;
1693         else if (outarg->offset + num > file_size)
1694                 num = file_size - outarg->offset;
1695
1696         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1697         num_pages = min(num_pages, fc->max_pages);
1698
1699         args_size += num_pages * (sizeof(ap->pages[0]) + sizeof(ap->descs[0]));
1700
1701         ra = kzalloc(args_size, GFP_KERNEL);
1702         if (!ra)
1703                 return -ENOMEM;
1704
1705         ap = &ra->ap;
1706         ap->pages = (void *) (ra + 1);
1707         ap->descs = (void *) (ap->pages + num_pages);
1708
1709         args = &ap->args;
1710         args->nodeid = outarg->nodeid;
1711         args->opcode = FUSE_NOTIFY_REPLY;
1712         args->in_numargs = 2;
1713         args->in_pages = true;
1714         args->end = fuse_retrieve_end;
1715
1716         index = outarg->offset >> PAGE_SHIFT;
1717
1718         while (num && ap->num_pages < num_pages) {
1719                 struct page *page;
1720                 unsigned int this_num;
1721
1722                 page = find_get_page(mapping, index);
1723                 if (!page)
1724                         break;
1725
1726                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1727                 ap->pages[ap->num_pages] = page;
1728                 ap->descs[ap->num_pages].offset = offset;
1729                 ap->descs[ap->num_pages].length = this_num;
1730                 ap->num_pages++;
1731
1732                 offset = 0;
1733                 num -= this_num;
1734                 total_len += this_num;
1735                 index++;
1736         }
1737         ra->inarg.offset = outarg->offset;
1738         ra->inarg.size = total_len;
1739         args->in_args[0].size = sizeof(ra->inarg);
1740         args->in_args[0].value = &ra->inarg;
1741         args->in_args[1].size = total_len;
1742
1743         err = fuse_simple_notify_reply(fm, args, outarg->notify_unique);
1744         if (err)
1745                 fuse_retrieve_end(fm, args, err);
1746
1747         return err;
1748 }
1749
1750 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1751                                 struct fuse_copy_state *cs)
1752 {
1753         struct fuse_notify_retrieve_out outarg;
1754         struct fuse_mount *fm;
1755         struct inode *inode;
1756         u64 nodeid;
1757         int err;
1758
1759         err = -EINVAL;
1760         if (size != sizeof(outarg))
1761                 goto copy_finish;
1762
1763         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1764         if (err)
1765                 goto copy_finish;
1766
1767         fuse_copy_finish(cs);
1768
1769         down_read(&fc->killsb);
1770         err = -ENOENT;
1771         nodeid = outarg.nodeid;
1772
1773         inode = fuse_ilookup(fc, nodeid, &fm);
1774         if (inode) {
1775                 err = fuse_retrieve(fm, inode, &outarg);
1776                 iput(inode);
1777         }
1778         up_read(&fc->killsb);
1779
1780         return err;
1781
1782 copy_finish:
1783         fuse_copy_finish(cs);
1784         return err;
1785 }
1786
1787 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1788                        unsigned int size, struct fuse_copy_state *cs)
1789 {
1790         /* Don't try to move pages (yet) */
1791         cs->move_pages = 0;
1792
1793         switch (code) {
1794         case FUSE_NOTIFY_POLL:
1795                 return fuse_notify_poll(fc, size, cs);
1796
1797         case FUSE_NOTIFY_INVAL_INODE:
1798                 return fuse_notify_inval_inode(fc, size, cs);
1799
1800         case FUSE_NOTIFY_INVAL_ENTRY:
1801                 return fuse_notify_inval_entry(fc, size, cs);
1802
1803         case FUSE_NOTIFY_STORE:
1804                 return fuse_notify_store(fc, size, cs);
1805
1806         case FUSE_NOTIFY_RETRIEVE:
1807                 return fuse_notify_retrieve(fc, size, cs);
1808
1809         case FUSE_NOTIFY_DELETE:
1810                 return fuse_notify_delete(fc, size, cs);
1811
1812         default:
1813                 fuse_copy_finish(cs);
1814                 return -EINVAL;
1815         }
1816 }
1817
1818 /* Look up request on processing list by unique ID */
1819 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1820 {
1821         unsigned int hash = fuse_req_hash(unique);
1822         struct fuse_req *req;
1823
1824         list_for_each_entry(req, &fpq->processing[hash], list) {
1825                 if (req->in.h.unique == unique)
1826                         return req;
1827         }
1828         return NULL;
1829 }
1830
1831 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_args *args,
1832                          unsigned nbytes)
1833 {
1834         unsigned reqsize = sizeof(struct fuse_out_header);
1835
1836         reqsize += fuse_len_args(args->out_numargs, args->out_args);
1837
1838         if (reqsize < nbytes || (reqsize > nbytes && !args->out_argvar))
1839                 return -EINVAL;
1840         else if (reqsize > nbytes) {
1841                 struct fuse_arg *lastarg = &args->out_args[args->out_numargs-1];
1842                 unsigned diffsize = reqsize - nbytes;
1843
1844                 if (diffsize > lastarg->size)
1845                         return -EINVAL;
1846                 lastarg->size -= diffsize;
1847         }
1848         return fuse_copy_args(cs, args->out_numargs, args->out_pages,
1849                               args->out_args, args->page_zeroing);
1850 }
1851
1852 /*
1853  * Write a single reply to a request.  First the header is copied from
1854  * the write buffer.  The request is then searched on the processing
1855  * list by the unique ID found in the header.  If found, then remove
1856  * it from the list and copy the rest of the buffer to the request.
1857  * The request is finished by calling fuse_request_end().
1858  */
1859 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1860                                  struct fuse_copy_state *cs, size_t nbytes)
1861 {
1862         int err;
1863         struct fuse_conn *fc = fud->fc;
1864         struct fuse_pqueue *fpq = &fud->pq;
1865         struct fuse_req *req;
1866         struct fuse_out_header oh;
1867
1868         err = -EINVAL;
1869         if (nbytes < sizeof(struct fuse_out_header))
1870                 goto out;
1871
1872         err = fuse_copy_one(cs, &oh, sizeof(oh));
1873         if (err)
1874                 goto copy_finish;
1875
1876         err = -EINVAL;
1877         if (oh.len != nbytes)
1878                 goto copy_finish;
1879
1880         /*
1881          * Zero oh.unique indicates unsolicited notification message
1882          * and error contains notification code.
1883          */
1884         if (!oh.unique) {
1885                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1886                 goto out;
1887         }
1888
1889         err = -EINVAL;
1890         if (oh.error <= -512 || oh.error > 0)
1891                 goto copy_finish;
1892
1893         spin_lock(&fpq->lock);
1894         req = NULL;
1895         if (fpq->connected)
1896                 req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1897
1898         err = -ENOENT;
1899         if (!req) {
1900                 spin_unlock(&fpq->lock);
1901                 goto copy_finish;
1902         }
1903
1904         /* Is it an interrupt reply ID? */
1905         if (oh.unique & FUSE_INT_REQ_BIT) {
1906                 __fuse_get_request(req);
1907                 spin_unlock(&fpq->lock);
1908
1909                 err = 0;
1910                 if (nbytes != sizeof(struct fuse_out_header))
1911                         err = -EINVAL;
1912                 else if (oh.error == -ENOSYS)
1913                         fc->no_interrupt = 1;
1914                 else if (oh.error == -EAGAIN)
1915                         err = queue_interrupt(req);
1916
1917                 fuse_put_request(req);
1918
1919                 goto copy_finish;
1920         }
1921
1922         clear_bit(FR_SENT, &req->flags);
1923         list_move(&req->list, &fpq->io);
1924         req->out.h = oh;
1925         set_bit(FR_LOCKED, &req->flags);
1926         spin_unlock(&fpq->lock);
1927         cs->req = req;
1928         if (!req->args->page_replace)
1929                 cs->move_pages = 0;
1930
1931         if (oh.error)
1932                 err = nbytes != sizeof(oh) ? -EINVAL : 0;
1933         else
1934                 err = copy_out_args(cs, req->args, nbytes);
1935         fuse_copy_finish(cs);
1936
1937         spin_lock(&fpq->lock);
1938         clear_bit(FR_LOCKED, &req->flags);
1939         if (!fpq->connected)
1940                 err = -ENOENT;
1941         else if (err)
1942                 req->out.h.error = -EIO;
1943         if (!test_bit(FR_PRIVATE, &req->flags))
1944                 list_del_init(&req->list);
1945         spin_unlock(&fpq->lock);
1946
1947         fuse_request_end(req);
1948 out:
1949         return err ? err : nbytes;
1950
1951 copy_finish:
1952         fuse_copy_finish(cs);
1953         goto out;
1954 }
1955
1956 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1957 {
1958         struct fuse_copy_state cs;
1959         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1960
1961         if (!fud)
1962                 return -EPERM;
1963
1964         if (!iter_is_iovec(from))
1965                 return -EINVAL;
1966
1967         fuse_copy_init(&cs, 0, from);
1968
1969         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1970 }
1971
1972 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1973                                      struct file *out, loff_t *ppos,
1974                                      size_t len, unsigned int flags)
1975 {
1976         unsigned int head, tail, mask, count;
1977         unsigned nbuf;
1978         unsigned idx;
1979         struct pipe_buffer *bufs;
1980         struct fuse_copy_state cs;
1981         struct fuse_dev *fud;
1982         size_t rem;
1983         ssize_t ret;
1984
1985         fud = fuse_get_dev(out);
1986         if (!fud)
1987                 return -EPERM;
1988
1989         pipe_lock(pipe);
1990
1991         head = pipe->head;
1992         tail = pipe->tail;
1993         mask = pipe->ring_size - 1;
1994         count = head - tail;
1995
1996         bufs = kvmalloc_array(count, sizeof(struct pipe_buffer), GFP_KERNEL);
1997         if (!bufs) {
1998                 pipe_unlock(pipe);
1999                 return -ENOMEM;
2000         }
2001
2002         nbuf = 0;
2003         rem = 0;
2004         for (idx = tail; idx != head && rem < len; idx++)
2005                 rem += pipe->bufs[idx & mask].len;
2006
2007         ret = -EINVAL;
2008         if (rem < len)
2009                 goto out_free;
2010
2011         rem = len;
2012         while (rem) {
2013                 struct pipe_buffer *ibuf;
2014                 struct pipe_buffer *obuf;
2015
2016                 if (WARN_ON(nbuf >= count || tail == head))
2017                         goto out_free;
2018
2019                 ibuf = &pipe->bufs[tail & mask];
2020                 obuf = &bufs[nbuf];
2021
2022                 if (rem >= ibuf->len) {
2023                         *obuf = *ibuf;
2024                         ibuf->ops = NULL;
2025                         tail++;
2026                         pipe->tail = tail;
2027                 } else {
2028                         if (!pipe_buf_get(pipe, ibuf))
2029                                 goto out_free;
2030
2031                         *obuf = *ibuf;
2032                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2033                         obuf->len = rem;
2034                         ibuf->offset += obuf->len;
2035                         ibuf->len -= obuf->len;
2036                 }
2037                 nbuf++;
2038                 rem -= obuf->len;
2039         }
2040         pipe_unlock(pipe);
2041
2042         fuse_copy_init(&cs, 0, NULL);
2043         cs.pipebufs = bufs;
2044         cs.nr_segs = nbuf;
2045         cs.pipe = pipe;
2046
2047         if (flags & SPLICE_F_MOVE)
2048                 cs.move_pages = 1;
2049
2050         ret = fuse_dev_do_write(fud, &cs, len);
2051
2052         pipe_lock(pipe);
2053 out_free:
2054         for (idx = 0; idx < nbuf; idx++) {
2055                 struct pipe_buffer *buf = &bufs[idx];
2056
2057                 if (buf->ops)
2058                         pipe_buf_release(pipe, buf);
2059         }
2060         pipe_unlock(pipe);
2061
2062         kvfree(bufs);
2063         return ret;
2064 }
2065
2066 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2067 {
2068         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2069         struct fuse_iqueue *fiq;
2070         struct fuse_dev *fud = fuse_get_dev(file);
2071
2072         if (!fud)
2073                 return EPOLLERR;
2074
2075         fiq = &fud->fc->iq;
2076         poll_wait(file, &fiq->waitq, wait);
2077
2078         spin_lock(&fiq->lock);
2079         if (!fiq->connected)
2080                 mask = EPOLLERR;
2081         else if (request_pending(fiq))
2082                 mask |= EPOLLIN | EPOLLRDNORM;
2083         spin_unlock(&fiq->lock);
2084
2085         return mask;
2086 }
2087
2088 /* Abort all requests on the given list (pending or processing) */
2089 static void end_requests(struct list_head *head)
2090 {
2091         while (!list_empty(head)) {
2092                 struct fuse_req *req;
2093                 req = list_entry(head->next, struct fuse_req, list);
2094                 req->out.h.error = -ECONNABORTED;
2095                 clear_bit(FR_SENT, &req->flags);
2096                 list_del_init(&req->list);
2097                 fuse_request_end(req);
2098         }
2099 }
2100
2101 static void end_polls(struct fuse_conn *fc)
2102 {
2103         struct rb_node *p;
2104
2105         p = rb_first(&fc->polled_files);
2106
2107         while (p) {
2108                 struct fuse_file *ff;
2109                 ff = rb_entry(p, struct fuse_file, polled_node);
2110                 wake_up_interruptible_all(&ff->poll_wait);
2111
2112                 p = rb_next(p);
2113         }
2114 }
2115
2116 /*
2117  * Abort all requests.
2118  *
2119  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2120  * filesystem.
2121  *
2122  * The same effect is usually achievable through killing the filesystem daemon
2123  * and all users of the filesystem.  The exception is the combination of an
2124  * asynchronous request and the tricky deadlock (see
2125  * Documentation/filesystems/fuse.rst).
2126  *
2127  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2128  * requests, they should be finished off immediately.  Locked requests will be
2129  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2130  * requests.  It is possible that some request will finish before we can.  This
2131  * is OK, the request will in that case be removed from the list before we touch
2132  * it.
2133  */
2134 void fuse_abort_conn(struct fuse_conn *fc)
2135 {
2136         struct fuse_iqueue *fiq = &fc->iq;
2137
2138         spin_lock(&fc->lock);
2139         if (fc->connected) {
2140                 struct fuse_dev *fud;
2141                 struct fuse_req *req, *next;
2142                 LIST_HEAD(to_end);
2143                 unsigned int i;
2144
2145                 /* Background queuing checks fc->connected under bg_lock */
2146                 spin_lock(&fc->bg_lock);
2147                 fc->connected = 0;
2148                 spin_unlock(&fc->bg_lock);
2149
2150                 fuse_set_initialized(fc);
2151                 list_for_each_entry(fud, &fc->devices, entry) {
2152                         struct fuse_pqueue *fpq = &fud->pq;
2153
2154                         spin_lock(&fpq->lock);
2155                         fpq->connected = 0;
2156                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2157                                 req->out.h.error = -ECONNABORTED;
2158                                 spin_lock(&req->waitq.lock);
2159                                 set_bit(FR_ABORTED, &req->flags);
2160                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2161                                         set_bit(FR_PRIVATE, &req->flags);
2162                                         __fuse_get_request(req);
2163                                         list_move(&req->list, &to_end);
2164                                 }
2165                                 spin_unlock(&req->waitq.lock);
2166                         }
2167                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2168                                 list_splice_tail_init(&fpq->processing[i],
2169                                                       &to_end);
2170                         spin_unlock(&fpq->lock);
2171                 }
2172                 spin_lock(&fc->bg_lock);
2173                 fc->blocked = 0;
2174                 fc->max_background = UINT_MAX;
2175                 flush_bg_queue(fc);
2176                 spin_unlock(&fc->bg_lock);
2177
2178                 spin_lock(&fiq->lock);
2179                 fiq->connected = 0;
2180                 list_for_each_entry(req, &fiq->pending, list)
2181                         clear_bit(FR_PENDING, &req->flags);
2182                 list_splice_tail_init(&fiq->pending, &to_end);
2183                 while (forget_pending(fiq))
2184                         kfree(fuse_dequeue_forget(fiq, 1, NULL));
2185                 wake_up_all(&fiq->waitq);
2186                 spin_unlock(&fiq->lock);
2187                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2188                 end_polls(fc);
2189                 wake_up_all(&fc->blocked_waitq);
2190                 spin_unlock(&fc->lock);
2191
2192                 end_requests(&to_end);
2193         } else {
2194                 spin_unlock(&fc->lock);
2195         }
2196 }
2197 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2198
2199 void fuse_wait_aborted(struct fuse_conn *fc)
2200 {
2201         /* matches implicit memory barrier in fuse_drop_waiting() */
2202         smp_mb();
2203         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2204 }
2205
2206 int fuse_dev_release(struct inode *inode, struct file *file)
2207 {
2208         struct fuse_dev *fud = fuse_get_dev(file);
2209
2210         if (fud) {
2211                 struct fuse_conn *fc = fud->fc;
2212                 struct fuse_pqueue *fpq = &fud->pq;
2213                 LIST_HEAD(to_end);
2214                 unsigned int i;
2215
2216                 spin_lock(&fpq->lock);
2217                 WARN_ON(!list_empty(&fpq->io));
2218                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2219                         list_splice_init(&fpq->processing[i], &to_end);
2220                 spin_unlock(&fpq->lock);
2221
2222                 end_requests(&to_end);
2223
2224                 /* Are we the last open device? */
2225                 if (atomic_dec_and_test(&fc->dev_count)) {
2226                         WARN_ON(fc->iq.fasync != NULL);
2227                         fuse_abort_conn(fc);
2228                 }
2229                 fuse_dev_free(fud);
2230         }
2231         return 0;
2232 }
2233 EXPORT_SYMBOL_GPL(fuse_dev_release);
2234
2235 static int fuse_dev_fasync(int fd, struct file *file, int on)
2236 {
2237         struct fuse_dev *fud = fuse_get_dev(file);
2238
2239         if (!fud)
2240                 return -EPERM;
2241
2242         /* No locking - fasync_helper does its own locking */
2243         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2244 }
2245
2246 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2247 {
2248         struct fuse_dev *fud;
2249
2250         if (new->private_data)
2251                 return -EINVAL;
2252
2253         fud = fuse_dev_alloc_install(fc);
2254         if (!fud)
2255                 return -ENOMEM;
2256
2257         new->private_data = fud;
2258         atomic_inc(&fc->dev_count);
2259
2260         return 0;
2261 }
2262
2263 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2264                            unsigned long arg)
2265 {
2266         int err = -ENOTTY;
2267
2268         if (cmd == FUSE_DEV_IOC_CLONE) {
2269                 int oldfd;
2270
2271                 err = -EFAULT;
2272                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2273                         struct file *old = fget(oldfd);
2274
2275                         err = -EINVAL;
2276                         if (old) {
2277                                 struct fuse_dev *fud = NULL;
2278
2279                                 /*
2280                                  * Check against file->f_op because CUSE
2281                                  * uses the same ioctl handler.
2282                                  */
2283                                 if (old->f_op == file->f_op &&
2284                                     old->f_cred->user_ns == file->f_cred->user_ns)
2285                                         fud = fuse_get_dev(old);
2286
2287                                 if (fud) {
2288                                         mutex_lock(&fuse_mutex);
2289                                         err = fuse_device_clone(fud->fc, file);
2290                                         mutex_unlock(&fuse_mutex);
2291                                 }
2292                                 fput(old);
2293                         }
2294                 }
2295         }
2296         return err;
2297 }
2298
2299 const struct file_operations fuse_dev_operations = {
2300         .owner          = THIS_MODULE,
2301         .open           = fuse_dev_open,
2302         .llseek         = no_llseek,
2303         .read_iter      = fuse_dev_read,
2304         .splice_read    = fuse_dev_splice_read,
2305         .write_iter     = fuse_dev_write,
2306         .splice_write   = fuse_dev_splice_write,
2307         .poll           = fuse_dev_poll,
2308         .release        = fuse_dev_release,
2309         .fasync         = fuse_dev_fasync,
2310         .unlocked_ioctl = fuse_dev_ioctl,
2311         .compat_ioctl   = compat_ptr_ioctl,
2312 };
2313 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2314
2315 static struct miscdevice fuse_miscdevice = {
2316         .minor = FUSE_MINOR,
2317         .name  = "fuse",
2318         .fops = &fuse_dev_operations,
2319 };
2320
2321 int __init fuse_dev_init(void)
2322 {
2323         int err = -ENOMEM;
2324         fuse_req_cachep = kmem_cache_create("fuse_request",
2325                                             sizeof(struct fuse_req),
2326                                             0, 0, NULL);
2327         if (!fuse_req_cachep)
2328                 goto out;
2329
2330         err = misc_register(&fuse_miscdevice);
2331         if (err)
2332                 goto out_cache_clean;
2333
2334         return 0;
2335
2336  out_cache_clean:
2337         kmem_cache_destroy(fuse_req_cachep);
2338  out:
2339         return err;
2340 }
2341
2342 void fuse_dev_cleanup(void)
2343 {
2344         misc_deregister(&fuse_miscdevice);
2345         kmem_cache_destroy(fuse_req_cachep);
2346 }