fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   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 matched 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_local_page(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_local(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         replace_page_cache_page(oldpage, newpage);
 849
 850         get_page(newpage);
 851
 852         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 853                 lru_cache_add(newpage);
 854
 855         /*
 856          * Release while we have extra ref on stolen page.  Otherwise
 857          * anon_pipe_buf_release() might think the page can be reused.
 858          */
 859         pipe_buf_release(cs->pipe, buf);
 860
 861         err = 0;
 862         spin_lock(&cs->req->waitq.lock);
 863         if (test_bit(FR_ABORTED, &cs->req->flags))
 864                 err = -ENOENT;
 865         else
 866                 *pagep = newpage;
 867         spin_unlock(&cs->req->waitq.lock);
 868
 869         if (err) {
 870                 unlock_page(newpage);
 871                 put_page(newpage);
 872                 goto out_put_old;
 873         }
 874
 875         unlock_page(oldpage);
 876         /* Drop ref for ap->pages[] array */
 877         put_page(oldpage);
 878         cs->len = 0;
 879
 880         err = 0;
 881 out_put_old:
 882         /* Drop ref obtained in this function */
 883         put_page(oldpage);
 884         return err;
 885
 886 out_fallback_unlock:
 887         unlock_page(newpage);
 888 out_fallback:
 889         cs->pg = buf->page;
 890         cs->offset = buf->offset;
 891
 892         err = lock_request(cs->req);
 893         if (!err)
 894                 err = 1;
 895
 896         goto out_put_old;
 897 }
 898
 899 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 900                          unsigned offset, unsigned count)
 901 {
 902         struct pipe_buffer *buf;
 903         int err;
 904
 905         if (cs->nr_segs >= cs->pipe->max_usage)
 906                 return -EIO;
 907
 908         get_page(page);
 909         err = unlock_request(cs->req);
 910         if (err) {
 911                 put_page(page);
 912                 return err;
 913         }
 914
 915         fuse_copy_finish(cs);
 916
 917         buf = cs->pipebufs;
 918         buf->page = page;
 919         buf->offset = offset;
 920         buf->len = count;
 921
 922         cs->pipebufs++;
 923         cs->nr_segs++;
 924         cs->len = 0;
 925
 926         return 0;
 927 }
 928
 929 /*
 930  * Copy a page in the request to/from the userspace buffer.  Must be
 931  * done atomically
 932  */
 933 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 934                           unsigned offset, unsigned count, int zeroing)
 935 {
 936         int err;
 937         struct page *page = *pagep;
 938
 939         if (page && zeroing && count < PAGE_SIZE)
 940                 clear_highpage(page);
 941
 942         while (count) {
 943                 if (cs->write && cs->pipebufs && page) {
 944                         return fuse_ref_page(cs, page, offset, count);
 945                 } else if (!cs->len) {
 946                         if (cs->move_pages && page &&
 947                             offset == 0 && count == PAGE_SIZE) {
 948                                 err = fuse_try_move_page(cs, pagep);
 949                                 if (err <= 0)
 950                                         return err;
 951                         } else {
 952                                 err = fuse_copy_fill(cs);
 953                                 if (err)
 954                                         return err;
 955                         }
 956                 }
 957                 if (page) {
 958                         void *mapaddr = kmap_local_page(page);
 959                         void *buf = mapaddr + offset;
 960                         offset += fuse_copy_do(cs, &buf, &count);
 961                         kunmap_local(mapaddr);
 962                 } else
 963                         offset += fuse_copy_do(cs, NULL, &count);
 964         }
 965         if (page && !cs->write)
 966                 flush_dcache_page(page);
 967         return 0;
 968 }
 969
 970 /* Copy pages in the request to/from userspace buffer */
 971 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 972                            int zeroing)
 973 {
 974         unsigned i;
 975         struct fuse_req *req = cs->req;
 976         struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
 977
 978
 979         for (i = 0; i < ap->num_pages && (nbytes || zeroing); i++) {
 980                 int err;
 981                 unsigned int offset = ap->descs[i].offset;
 982                 unsigned int count = min(nbytes, ap->descs[i].length);
 983
 984                 err = fuse_copy_page(cs, &ap->pages[i], offset, count, zeroing);
 985                 if (err)
 986                         return err;
 987
 988                 nbytes -= count;
 989         }
 990         return 0;
 991 }
 992
 993 /* Copy a single argument in the request to/from userspace buffer */
 994 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 995 {
 996         while (size) {
 997                 if (!cs->len) {
 998                         int err = fuse_copy_fill(cs);
 999                         if (err)
1000                                 return err;
1001                 }
1002                 fuse_copy_do(cs, &val, &size);
1003         }
1004         return 0;
1005 }
1006
1007 /* Copy request arguments to/from userspace buffer */
1008 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1009                           unsigned argpages, struct fuse_arg *args,
1010                           int zeroing)
1011 {
1012         int err = 0;
1013         unsigned i;
1014
1015         for (i = 0; !err && i < numargs; i++)  {
1016                 struct fuse_arg *arg = &args[i];
1017                 if (i == numargs - 1 && argpages)
1018                         err = fuse_copy_pages(cs, arg->size, zeroing);
1019                 else
1020                         err = fuse_copy_one(cs, arg->value, arg->size);
1021         }
1022         return err;
1023 }
1024
1025 static int forget_pending(struct fuse_iqueue *fiq)
1026 {
1027         return fiq->forget_list_head.next != NULL;
1028 }
1029
1030 static int request_pending(struct fuse_iqueue *fiq)
1031 {
1032         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1033                 forget_pending(fiq);
1034 }
1035
1036 /*
1037  * Transfer an interrupt request to userspace
1038  *
1039  * Unlike other requests this is assembled on demand, without a need
1040  * to allocate a separate fuse_req structure.
1041  *
1042  * Called with fiq->lock held, releases it
1043  */
1044 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1045                                struct fuse_copy_state *cs,
1046                                size_t nbytes, struct fuse_req *req)
1047 __releases(fiq->lock)
1048 {
1049         struct fuse_in_header ih;
1050         struct fuse_interrupt_in arg;
1051         unsigned reqsize = sizeof(ih) + sizeof(arg);
1052         int err;
1053
1054         list_del_init(&req->intr_entry);
1055         memset(&ih, 0, sizeof(ih));
1056         memset(&arg, 0, sizeof(arg));
1057         ih.len = reqsize;
1058         ih.opcode = FUSE_INTERRUPT;
1059         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1060         arg.unique = req->in.h.unique;
1061
1062         spin_unlock(&fiq->lock);
1063         if (nbytes < reqsize)
1064                 return -EINVAL;
1065
1066         err = fuse_copy_one(cs, &ih, sizeof(ih));
1067         if (!err)
1068                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1069         fuse_copy_finish(cs);
1070
1071         return err ? err : reqsize;
1072 }
1073
1074 struct fuse_forget_link *fuse_dequeue_forget(struct fuse_iqueue *fiq,
1075                                              unsigned int max,
1076                                              unsigned int *countp)
1077 {
1078         struct fuse_forget_link *head = fiq->forget_list_head.next;
1079         struct fuse_forget_link **newhead = &head;
1080         unsigned count;
1081
1082         for (count = 0; *newhead != NULL && count < max; count++)
1083                 newhead = &(*newhead)->next;
1084
1085         fiq->forget_list_head.next = *newhead;
1086         *newhead = NULL;
1087         if (fiq->forget_list_head.next == NULL)
1088                 fiq->forget_list_tail = &fiq->forget_list_head;
1089
1090         if (countp != NULL)
1091                 *countp = count;
1092
1093         return head;
1094 }
1095 EXPORT_SYMBOL(fuse_dequeue_forget);
1096
1097 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1098                                    struct fuse_copy_state *cs,
1099                                    size_t nbytes)
1100 __releases(fiq->lock)
1101 {
1102         int err;
1103         struct fuse_forget_link *forget = fuse_dequeue_forget(fiq, 1, NULL);
1104         struct fuse_forget_in arg = {
1105                 .nlookup = forget->forget_one.nlookup,
1106         };
1107         struct fuse_in_header ih = {
1108                 .opcode = FUSE_FORGET,
1109                 .nodeid = forget->forget_one.nodeid,
1110                 .unique = fuse_get_unique(fiq),
1111                 .len = sizeof(ih) + sizeof(arg),
1112         };
1113
1114         spin_unlock(&fiq->lock);
1115         kfree(forget);
1116         if (nbytes < ih.len)
1117                 return -EINVAL;
1118
1119         err = fuse_copy_one(cs, &ih, sizeof(ih));
1120         if (!err)
1121                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1122         fuse_copy_finish(cs);
1123
1124         if (err)
1125                 return err;
1126
1127         return ih.len;
1128 }
1129
1130 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1131                                    struct fuse_copy_state *cs, size_t nbytes)
1132 __releases(fiq->lock)
1133 {
1134         int err;
1135         unsigned max_forgets;
1136         unsigned count;
1137         struct fuse_forget_link *head;
1138         struct fuse_batch_forget_in arg = { .count = 0 };
1139         struct fuse_in_header ih = {
1140                 .opcode = FUSE_BATCH_FORGET,
1141                 .unique = fuse_get_unique(fiq),
1142                 .len = sizeof(ih) + sizeof(arg),
1143         };
1144
1145         if (nbytes < ih.len) {
1146                 spin_unlock(&fiq->lock);
1147                 return -EINVAL;
1148         }
1149
1150         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1151         head = fuse_dequeue_forget(fiq, max_forgets, &count);
1152         spin_unlock(&fiq->lock);
1153
1154         arg.count = count;
1155         ih.len += count * sizeof(struct fuse_forget_one);
1156         err = fuse_copy_one(cs, &ih, sizeof(ih));
1157         if (!err)
1158                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1159
1160         while (head) {
1161                 struct fuse_forget_link *forget = head;
1162
1163                 if (!err) {
1164                         err = fuse_copy_one(cs, &forget->forget_one,
1165                                             sizeof(forget->forget_one));
1166                 }
1167                 head = forget->next;
1168                 kfree(forget);
1169         }
1170
1171         fuse_copy_finish(cs);
1172
1173         if (err)
1174                 return err;
1175
1176         return ih.len;
1177 }
1178
1179 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1180                             struct fuse_copy_state *cs,
1181                             size_t nbytes)
1182 __releases(fiq->lock)
1183 {
1184         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1185                 return fuse_read_single_forget(fiq, cs, nbytes);
1186         else
1187                 return fuse_read_batch_forget(fiq, cs, nbytes);
1188 }
1189
1190 /*
1191  * Read a single request into the userspace filesystem's buffer.  This
1192  * function waits until a request is available, then removes it from
1193  * the pending list and copies request data to userspace buffer.  If
1194  * no reply is needed (FORGET) or request has been aborted or there
1195  * was an error during the copying then it's finished by calling
1196  * fuse_request_end().  Otherwise add it to the processing list, and set
1197  * the 'sent' flag.
1198  */
1199 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1200                                 struct fuse_copy_state *cs, size_t nbytes)
1201 {
1202         ssize_t err;
1203         struct fuse_conn *fc = fud->fc;
1204         struct fuse_iqueue *fiq = &fc->iq;
1205         struct fuse_pqueue *fpq = &fud->pq;
1206         struct fuse_req *req;
1207         struct fuse_args *args;
1208         unsigned reqsize;
1209         unsigned int hash;
1210
1211         /*
1212          * Require sane minimum read buffer - that has capacity for fixed part
1213          * of any request header + negotiated max_write room for data.
1214          *
1215          * Historically libfuse reserves 4K for fixed header room, but e.g.
1216          * GlusterFS reserves only 80 bytes
1217          *
1218          *      = `sizeof(fuse_in_header) + sizeof(fuse_write_in)`
1219          *
1220          * which is the absolute minimum any sane filesystem should be using
1221          * for header room.
1222          */
1223         if (nbytes < max_t(size_t, FUSE_MIN_READ_BUFFER,
1224                            sizeof(struct fuse_in_header) +
1225                            sizeof(struct fuse_write_in) +
1226                            fc->max_write))
1227                 return -EINVAL;
1228
1229  restart:
1230         for (;;) {
1231                 spin_lock(&fiq->lock);
1232                 if (!fiq->connected || request_pending(fiq))
1233                         break;
1234                 spin_unlock(&fiq->lock);
1235
1236                 if (file->f_flags & O_NONBLOCK)
1237                         return -EAGAIN;
1238                 err = wait_event_interruptible_exclusive(fiq->waitq,
1239                                 !fiq->connected || request_pending(fiq));
1240                 if (err)
1241                         return err;
1242         }
1243
1244         if (!fiq->connected) {
1245                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1246                 goto err_unlock;
1247         }
1248
1249         if (!list_empty(&fiq->interrupts)) {
1250                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1251                                  intr_entry);
1252                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1253         }
1254
1255         if (forget_pending(fiq)) {
1256                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1257                         return fuse_read_forget(fc, fiq, cs, nbytes);
1258
1259                 if (fiq->forget_batch <= -8)
1260                         fiq->forget_batch = 16;
1261         }
1262
1263         req = list_entry(fiq->pending.next, struct fuse_req, list);
1264         clear_bit(FR_PENDING, &req->flags);
1265         list_del_init(&req->list);
1266         spin_unlock(&fiq->lock);
1267
1268         args = req->args;
1269         reqsize = req->in.h.len;
1270
1271         /* If request is too large, reply with an error and restart the read */
1272         if (nbytes < reqsize) {
1273                 req->out.h.error = -EIO;
1274                 /* SETXATTR is special, since it may contain too large data */
1275                 if (args->opcode == FUSE_SETXATTR)
1276                         req->out.h.error = -E2BIG;
1277                 fuse_request_end(req);
1278                 goto restart;
1279         }
1280         spin_lock(&fpq->lock);
1281         /*
1282          *  Must not put request on fpq->io queue after having been shut down by
1283          *  fuse_abort_conn()
1284          */
1285         if (!fpq->connected) {
1286                 req->out.h.error = err = -ECONNABORTED;
1287                 goto out_end;
1288
1289         }
1290         list_add(&req->list, &fpq->io);
1291         spin_unlock(&fpq->lock);
1292         cs->req = req;
1293         err = fuse_copy_one(cs, &req->in.h, sizeof(req->in.h));
1294         if (!err)
1295                 err = fuse_copy_args(cs, args->in_numargs, args->in_pages,
1296                                      (struct fuse_arg *) args->in_args, 0);
1297         fuse_copy_finish(cs);
1298         spin_lock(&fpq->lock);
1299         clear_bit(FR_LOCKED, &req->flags);
1300         if (!fpq->connected) {
1301                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1302                 goto out_end;
1303         }
1304         if (err) {
1305                 req->out.h.error = -EIO;
1306                 goto out_end;
1307         }
1308         if (!test_bit(FR_ISREPLY, &req->flags)) {
1309                 err = reqsize;
1310                 goto out_end;
1311         }
1312         hash = fuse_req_hash(req->in.h.unique);
1313         list_move_tail(&req->list, &fpq->processing[hash]);
1314         __fuse_get_request(req);
1315         set_bit(FR_SENT, &req->flags);
1316         spin_unlock(&fpq->lock);
1317         /* matches barrier in request_wait_answer() */
1318         smp_mb__after_atomic();
1319         if (test_bit(FR_INTERRUPTED, &req->flags))
1320                 queue_interrupt(req);
1321         fuse_put_request(req);
1322
1323         return reqsize;
1324
1325 out_end:
1326         if (!test_bit(FR_PRIVATE, &req->flags))
1327                 list_del_init(&req->list);
1328         spin_unlock(&fpq->lock);
1329         fuse_request_end(req);
1330         return err;
1331
1332  err_unlock:
1333         spin_unlock(&fiq->lock);
1334         return err;
1335 }
1336
1337 static int fuse_dev_open(struct inode *inode, struct file *file)
1338 {
1339         /*
1340          * The fuse device's file's private_data is used to hold
1341          * the fuse_conn(ection) when it is mounted, and is used to
1342          * keep track of whether the file has been mounted already.
1343          */
1344         file->private_data = NULL;
1345         return 0;
1346 }
1347
1348 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1349 {
1350         struct fuse_copy_state cs;
1351         struct file *file = iocb->ki_filp;
1352         struct fuse_dev *fud = fuse_get_dev(file);
1353
1354         if (!fud)
1355                 return -EPERM;
1356
1357         if (!iter_is_iovec(to))
1358                 return -EINVAL;
1359
1360         fuse_copy_init(&cs, 1, to);
1361
1362         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1363 }
1364
1365 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1366                                     struct pipe_inode_info *pipe,
1367                                     size_t len, unsigned int flags)
1368 {
1369         int total, ret;
1370         int page_nr = 0;
1371         struct pipe_buffer *bufs;
1372         struct fuse_copy_state cs;
1373         struct fuse_dev *fud = fuse_get_dev(in);
1374
1375         if (!fud)
1376                 return -EPERM;
1377
1378         bufs = kvmalloc_array(pipe->max_usage, sizeof(struct pipe_buffer),
1379                               GFP_KERNEL);
1380         if (!bufs)
1381                 return -ENOMEM;
1382
1383         fuse_copy_init(&cs, 1, NULL);
1384         cs.pipebufs = bufs;
1385         cs.pipe = pipe;
1386         ret = fuse_dev_do_read(fud, in, &cs, len);
1387         if (ret < 0)
1388                 goto out;
1389
1390         if (pipe_occupancy(pipe->head, pipe->tail) + cs.nr_segs > pipe->max_usage) {
1391                 ret = -EIO;
1392                 goto out;
1393         }
1394
1395         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1396                 /*
1397                  * Need to be careful about this.  Having buf->ops in module
1398                  * code can Oops if the buffer persists after module unload.
1399                  */
1400                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1401                 bufs[page_nr].flags = 0;
1402                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1403                 if (unlikely(ret < 0))
1404                         break;
1405         }
1406         if (total)
1407                 ret = total;
1408 out:
1409         for (; page_nr < cs.nr_segs; page_nr++)
1410                 put_page(bufs[page_nr].page);
1411
1412         kvfree(bufs);
1413         return ret;
1414 }
1415
1416 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1417                             struct fuse_copy_state *cs)
1418 {
1419         struct fuse_notify_poll_wakeup_out outarg;
1420         int err = -EINVAL;
1421
1422         if (size != sizeof(outarg))
1423                 goto err;
1424
1425         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1426         if (err)
1427                 goto err;
1428
1429         fuse_copy_finish(cs);
1430         return fuse_notify_poll_wakeup(fc, &outarg);
1431
1432 err:
1433         fuse_copy_finish(cs);
1434         return err;
1435 }
1436
1437 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1438                                    struct fuse_copy_state *cs)
1439 {
1440         struct fuse_notify_inval_inode_out outarg;
1441         int err = -EINVAL;
1442
1443         if (size != sizeof(outarg))
1444                 goto err;
1445
1446         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1447         if (err)
1448                 goto err;
1449         fuse_copy_finish(cs);
1450
1451         down_read(&fc->killsb);
1452         err = fuse_reverse_inval_inode(fc, outarg.ino,
1453                                        outarg.off, outarg.len);
1454         up_read(&fc->killsb);
1455         return err;
1456
1457 err:
1458         fuse_copy_finish(cs);
1459         return err;
1460 }
1461
1462 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1463                                    struct fuse_copy_state *cs)
1464 {
1465         struct fuse_notify_inval_entry_out outarg;
1466         int err = -ENOMEM;
1467         char *buf;
1468         struct qstr name;
1469
1470         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1471         if (!buf)
1472                 goto err;
1473
1474         err = -EINVAL;
1475         if (size < sizeof(outarg))
1476                 goto err;
1477
1478         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1479         if (err)
1480                 goto err;
1481
1482         err = -ENAMETOOLONG;
1483         if (outarg.namelen > FUSE_NAME_MAX)
1484                 goto err;
1485
1486         err = -EINVAL;
1487         if (size != sizeof(outarg) + outarg.namelen + 1)
1488                 goto err;
1489
1490         name.name = buf;
1491         name.len = outarg.namelen;
1492         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1493         if (err)
1494                 goto err;
1495         fuse_copy_finish(cs);
1496         buf[outarg.namelen] = 0;
1497
1498         down_read(&fc->killsb);
1499         err = fuse_reverse_inval_entry(fc, outarg.parent, 0, &name);
1500         up_read(&fc->killsb);
1501         kfree(buf);
1502         return err;
1503
1504 err:
1505         kfree(buf);
1506         fuse_copy_finish(cs);
1507         return err;
1508 }
1509
1510 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1511                               struct fuse_copy_state *cs)
1512 {
1513         struct fuse_notify_delete_out outarg;
1514         int err = -ENOMEM;
1515         char *buf;
1516         struct qstr name;
1517
1518         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1519         if (!buf)
1520                 goto err;
1521
1522         err = -EINVAL;
1523         if (size < sizeof(outarg))
1524                 goto err;
1525
1526         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1527         if (err)
1528                 goto err;
1529
1530         err = -ENAMETOOLONG;
1531         if (outarg.namelen > FUSE_NAME_MAX)
1532                 goto err;
1533
1534         err = -EINVAL;
1535         if (size != sizeof(outarg) + outarg.namelen + 1)
1536                 goto err;
1537
1538         name.name = buf;
1539         name.len = outarg.namelen;
1540         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1541         if (err)
1542                 goto err;
1543         fuse_copy_finish(cs);
1544         buf[outarg.namelen] = 0;
1545
1546         down_read(&fc->killsb);
1547         err = fuse_reverse_inval_entry(fc, outarg.parent, outarg.child, &name);
1548         up_read(&fc->killsb);
1549         kfree(buf);
1550         return err;
1551
1552 err:
1553         kfree(buf);
1554         fuse_copy_finish(cs);
1555         return err;
1556 }
1557
1558 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1559                              struct fuse_copy_state *cs)
1560 {
1561         struct fuse_notify_store_out outarg;
1562         struct inode *inode;
1563         struct address_space *mapping;
1564         u64 nodeid;
1565         int err;
1566         pgoff_t index;
1567         unsigned int offset;
1568         unsigned int num;
1569         loff_t file_size;
1570         loff_t end;
1571
1572         err = -EINVAL;
1573         if (size < sizeof(outarg))
1574                 goto out_finish;
1575
1576         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1577         if (err)
1578                 goto out_finish;
1579
1580         err = -EINVAL;
1581         if (size - sizeof(outarg) != outarg.size)
1582                 goto out_finish;
1583
1584         nodeid = outarg.nodeid;
1585
1586         down_read(&fc->killsb);
1587
1588         err = -ENOENT;
1589         inode = fuse_ilookup(fc, nodeid,  NULL);
1590         if (!inode)
1591                 goto out_up_killsb;
1592
1593         mapping = inode->i_mapping;
1594         index = outarg.offset >> PAGE_SHIFT;
1595         offset = outarg.offset & ~PAGE_MASK;
1596         file_size = i_size_read(inode);
1597         end = outarg.offset + outarg.size;
1598         if (end > file_size) {
1599                 file_size = end;
1600                 fuse_write_update_attr(inode, file_size, outarg.size);
1601         }
1602
1603         num = outarg.size;
1604         while (num) {
1605                 struct page *page;
1606                 unsigned int this_num;
1607
1608                 err = -ENOMEM;
1609                 page = find_or_create_page(mapping, index,
1610                                            mapping_gfp_mask(mapping));
1611                 if (!page)
1612                         goto out_iput;
1613
1614                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1615                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1616                 if (!err && offset == 0 &&
1617                     (this_num == PAGE_SIZE || file_size == end))
1618                         SetPageUptodate(page);
1619                 unlock_page(page);
1620                 put_page(page);
1621
1622                 if (err)
1623                         goto out_iput;
1624
1625                 num -= this_num;
1626                 offset = 0;
1627                 index++;
1628         }
1629
1630         err = 0;
1631
1632 out_iput:
1633         iput(inode);
1634 out_up_killsb:
1635         up_read(&fc->killsb);
1636 out_finish:
1637         fuse_copy_finish(cs);
1638         return err;
1639 }
1640
1641 struct fuse_retrieve_args {
1642         struct fuse_args_pages ap;
1643         struct fuse_notify_retrieve_in inarg;
1644 };
1645
1646 static void fuse_retrieve_end(struct fuse_mount *fm, struct fuse_args *args,
1647                               int error)
1648 {
1649         struct fuse_retrieve_args *ra =
1650                 container_of(args, typeof(*ra), ap.args);
1651
1652         release_pages(ra->ap.pages, ra->ap.num_pages);
1653         kfree(ra);
1654 }
1655
1656 static int fuse_retrieve(struct fuse_mount *fm, struct inode *inode,
1657                          struct fuse_notify_retrieve_out *outarg)
1658 {
1659         int err;
1660         struct address_space *mapping = inode->i_mapping;
1661         pgoff_t index;
1662         loff_t file_size;
1663         unsigned int num;
1664         unsigned int offset;
1665         size_t total_len = 0;
1666         unsigned int num_pages;
1667         struct fuse_conn *fc = fm->fc;
1668         struct fuse_retrieve_args *ra;
1669         size_t args_size = sizeof(*ra);
1670         struct fuse_args_pages *ap;
1671         struct fuse_args *args;
1672
1673         offset = outarg->offset & ~PAGE_MASK;
1674         file_size = i_size_read(inode);
1675
1676         num = min(outarg->size, fc->max_write);
1677         if (outarg->offset > file_size)
1678                 num = 0;
1679         else if (outarg->offset + num > file_size)
1680                 num = file_size - outarg->offset;
1681
1682         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1683         num_pages = min(num_pages, fc->max_pages);
1684
1685         args_size += num_pages * (sizeof(ap->pages[0]) + sizeof(ap->descs[0]));
1686
1687         ra = kzalloc(args_size, GFP_KERNEL);
1688         if (!ra)
1689                 return -ENOMEM;
1690
1691         ap = &ra->ap;
1692         ap->pages = (void *) (ra + 1);
1693         ap->descs = (void *) (ap->pages + num_pages);
1694
1695         args = &ap->args;
1696         args->nodeid = outarg->nodeid;
1697         args->opcode = FUSE_NOTIFY_REPLY;
1698         args->in_numargs = 2;
1699         args->in_pages = true;
1700         args->end = fuse_retrieve_end;
1701
1702         index = outarg->offset >> PAGE_SHIFT;
1703
1704         while (num && ap->num_pages < num_pages) {
1705                 struct page *page;
1706                 unsigned int this_num;
1707
1708                 page = find_get_page(mapping, index);
1709                 if (!page)
1710                         break;
1711
1712                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1713                 ap->pages[ap->num_pages] = page;
1714                 ap->descs[ap->num_pages].offset = offset;
1715                 ap->descs[ap->num_pages].length = this_num;
1716                 ap->num_pages++;
1717
1718                 offset = 0;
1719                 num -= this_num;
1720                 total_len += this_num;
1721                 index++;
1722         }
1723         ra->inarg.offset = outarg->offset;
1724         ra->inarg.size = total_len;
1725         args->in_args[0].size = sizeof(ra->inarg);
1726         args->in_args[0].value = &ra->inarg;
1727         args->in_args[1].size = total_len;
1728
1729         err = fuse_simple_notify_reply(fm, args, outarg->notify_unique);
1730         if (err)
1731                 fuse_retrieve_end(fm, args, err);
1732
1733         return err;
1734 }
1735
1736 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1737                                 struct fuse_copy_state *cs)
1738 {
1739         struct fuse_notify_retrieve_out outarg;
1740         struct fuse_mount *fm;
1741         struct inode *inode;
1742         u64 nodeid;
1743         int err;
1744
1745         err = -EINVAL;
1746         if (size != sizeof(outarg))
1747                 goto copy_finish;
1748
1749         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1750         if (err)
1751                 goto copy_finish;
1752
1753         fuse_copy_finish(cs);
1754
1755         down_read(&fc->killsb);
1756         err = -ENOENT;
1757         nodeid = outarg.nodeid;
1758
1759         inode = fuse_ilookup(fc, nodeid, &fm);
1760         if (inode) {
1761                 err = fuse_retrieve(fm, inode, &outarg);
1762                 iput(inode);
1763         }
1764         up_read(&fc->killsb);
1765
1766         return err;
1767
1768 copy_finish:
1769         fuse_copy_finish(cs);
1770         return err;
1771 }
1772
1773 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1774                        unsigned int size, struct fuse_copy_state *cs)
1775 {
1776         /* Don't try to move pages (yet) */
1777         cs->move_pages = 0;
1778
1779         switch (code) {
1780         case FUSE_NOTIFY_POLL:
1781                 return fuse_notify_poll(fc, size, cs);
1782
1783         case FUSE_NOTIFY_INVAL_INODE:
1784                 return fuse_notify_inval_inode(fc, size, cs);
1785
1786         case FUSE_NOTIFY_INVAL_ENTRY:
1787                 return fuse_notify_inval_entry(fc, size, cs);
1788
1789         case FUSE_NOTIFY_STORE:
1790                 return fuse_notify_store(fc, size, cs);
1791
1792         case FUSE_NOTIFY_RETRIEVE:
1793                 return fuse_notify_retrieve(fc, size, cs);
1794
1795         case FUSE_NOTIFY_DELETE:
1796                 return fuse_notify_delete(fc, size, cs);
1797
1798         default:
1799                 fuse_copy_finish(cs);
1800                 return -EINVAL;
1801         }
1802 }
1803
1804 /* Look up request on processing list by unique ID */
1805 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1806 {
1807         unsigned int hash = fuse_req_hash(unique);
1808         struct fuse_req *req;
1809
1810         list_for_each_entry(req, &fpq->processing[hash], list) {
1811                 if (req->in.h.unique == unique)
1812                         return req;
1813         }
1814         return NULL;
1815 }
1816
1817 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_args *args,
1818                          unsigned nbytes)
1819 {
1820         unsigned reqsize = sizeof(struct fuse_out_header);
1821
1822         reqsize += fuse_len_args(args->out_numargs, args->out_args);
1823
1824         if (reqsize < nbytes || (reqsize > nbytes && !args->out_argvar))
1825                 return -EINVAL;
1826         else if (reqsize > nbytes) {
1827                 struct fuse_arg *lastarg = &args->out_args[args->out_numargs-1];
1828                 unsigned diffsize = reqsize - nbytes;
1829
1830                 if (diffsize > lastarg->size)
1831                         return -EINVAL;
1832                 lastarg->size -= diffsize;
1833         }
1834         return fuse_copy_args(cs, args->out_numargs, args->out_pages,
1835                               args->out_args, args->page_zeroing);
1836 }
1837
1838 /*
1839  * Write a single reply to a request.  First the header is copied from
1840  * the write buffer.  The request is then searched on the processing
1841  * list by the unique ID found in the header.  If found, then remove
1842  * it from the list and copy the rest of the buffer to the request.
1843  * The request is finished by calling fuse_request_end().
1844  */
1845 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1846                                  struct fuse_copy_state *cs, size_t nbytes)
1847 {
1848         int err;
1849         struct fuse_conn *fc = fud->fc;
1850         struct fuse_pqueue *fpq = &fud->pq;
1851         struct fuse_req *req;
1852         struct fuse_out_header oh;
1853
1854         err = -EINVAL;
1855         if (nbytes < sizeof(struct fuse_out_header))
1856                 goto out;
1857
1858         err = fuse_copy_one(cs, &oh, sizeof(oh));
1859         if (err)
1860                 goto copy_finish;
1861
1862         err = -EINVAL;
1863         if (oh.len != nbytes)
1864                 goto copy_finish;
1865
1866         /*
1867          * Zero oh.unique indicates unsolicited notification message
1868          * and error contains notification code.
1869          */
1870         if (!oh.unique) {
1871                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1872                 goto out;
1873         }
1874
1875         err = -EINVAL;
1876         if (oh.error <= -512 || oh.error > 0)
1877                 goto copy_finish;
1878
1879         spin_lock(&fpq->lock);
1880         req = NULL;
1881         if (fpq->connected)
1882                 req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1883
1884         err = -ENOENT;
1885         if (!req) {
1886                 spin_unlock(&fpq->lock);
1887                 goto copy_finish;
1888         }
1889
1890         /* Is it an interrupt reply ID? */
1891         if (oh.unique & FUSE_INT_REQ_BIT) {
1892                 __fuse_get_request(req);
1893                 spin_unlock(&fpq->lock);
1894
1895                 err = 0;
1896                 if (nbytes != sizeof(struct fuse_out_header))
1897                         err = -EINVAL;
1898                 else if (oh.error == -ENOSYS)
1899                         fc->no_interrupt = 1;
1900                 else if (oh.error == -EAGAIN)
1901                         err = queue_interrupt(req);
1902
1903                 fuse_put_request(req);
1904
1905                 goto copy_finish;
1906         }
1907
1908         clear_bit(FR_SENT, &req->flags);
1909         list_move(&req->list, &fpq->io);
1910         req->out.h = oh;
1911         set_bit(FR_LOCKED, &req->flags);
1912         spin_unlock(&fpq->lock);
1913         cs->req = req;
1914         if (!req->args->page_replace)
1915                 cs->move_pages = 0;
1916
1917         if (oh.error)
1918                 err = nbytes != sizeof(oh) ? -EINVAL : 0;
1919         else
1920                 err = copy_out_args(cs, req->args, nbytes);
1921         fuse_copy_finish(cs);
1922
1923         spin_lock(&fpq->lock);
1924         clear_bit(FR_LOCKED, &req->flags);
1925         if (!fpq->connected)
1926                 err = -ENOENT;
1927         else if (err)
1928                 req->out.h.error = -EIO;
1929         if (!test_bit(FR_PRIVATE, &req->flags))
1930                 list_del_init(&req->list);
1931         spin_unlock(&fpq->lock);
1932
1933         fuse_request_end(req);
1934 out:
1935         return err ? err : nbytes;
1936
1937 copy_finish:
1938         fuse_copy_finish(cs);
1939         goto out;
1940 }
1941
1942 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1943 {
1944         struct fuse_copy_state cs;
1945         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1946
1947         if (!fud)
1948                 return -EPERM;
1949
1950         if (!iter_is_iovec(from))
1951                 return -EINVAL;
1952
1953         fuse_copy_init(&cs, 0, from);
1954
1955         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1956 }
1957
1958 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1959                                      struct file *out, loff_t *ppos,
1960                                      size_t len, unsigned int flags)
1961 {
1962         unsigned int head, tail, mask, count;
1963         unsigned nbuf;
1964         unsigned idx;
1965         struct pipe_buffer *bufs;
1966         struct fuse_copy_state cs;
1967         struct fuse_dev *fud;
1968         size_t rem;
1969         ssize_t ret;
1970
1971         fud = fuse_get_dev(out);
1972         if (!fud)
1973                 return -EPERM;
1974
1975         pipe_lock(pipe);
1976
1977         head = pipe->head;
1978         tail = pipe->tail;
1979         mask = pipe->ring_size - 1;
1980         count = head - tail;
1981
1982         bufs = kvmalloc_array(count, sizeof(struct pipe_buffer), GFP_KERNEL);
1983         if (!bufs) {
1984                 pipe_unlock(pipe);
1985                 return -ENOMEM;
1986         }
1987
1988         nbuf = 0;
1989         rem = 0;
1990         for (idx = tail; idx != head && rem < len; idx++)
1991                 rem += pipe->bufs[idx & mask].len;
1992
1993         ret = -EINVAL;
1994         if (rem < len)
1995                 goto out_free;
1996
1997         rem = len;
1998         while (rem) {
1999                 struct pipe_buffer *ibuf;
2000                 struct pipe_buffer *obuf;
2001
2002                 if (WARN_ON(nbuf >= count || tail == head))
2003                         goto out_free;
2004
2005                 ibuf = &pipe->bufs[tail & mask];
2006                 obuf = &bufs[nbuf];
2007
2008                 if (rem >= ibuf->len) {
2009                         *obuf = *ibuf;
2010                         ibuf->ops = NULL;
2011                         tail++;
2012                         pipe->tail = tail;
2013                 } else {
2014                         if (!pipe_buf_get(pipe, ibuf))
2015                                 goto out_free;
2016
2017                         *obuf = *ibuf;
2018                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2019                         obuf->len = rem;
2020                         ibuf->offset += obuf->len;
2021                         ibuf->len -= obuf->len;
2022                 }
2023                 nbuf++;
2024                 rem -= obuf->len;
2025         }
2026         pipe_unlock(pipe);
2027
2028         fuse_copy_init(&cs, 0, NULL);
2029         cs.pipebufs = bufs;
2030         cs.nr_segs = nbuf;
2031         cs.pipe = pipe;
2032
2033         if (flags & SPLICE_F_MOVE)
2034                 cs.move_pages = 1;
2035
2036         ret = fuse_dev_do_write(fud, &cs, len);
2037
2038         pipe_lock(pipe);
2039 out_free:
2040         for (idx = 0; idx < nbuf; idx++) {
2041                 struct pipe_buffer *buf = &bufs[idx];
2042
2043                 if (buf->ops)
2044                         pipe_buf_release(pipe, buf);
2045         }
2046         pipe_unlock(pipe);
2047
2048         kvfree(bufs);
2049         return ret;
2050 }
2051
2052 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2053 {
2054         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2055         struct fuse_iqueue *fiq;
2056         struct fuse_dev *fud = fuse_get_dev(file);
2057
2058         if (!fud)
2059                 return EPOLLERR;
2060
2061         fiq = &fud->fc->iq;
2062         poll_wait(file, &fiq->waitq, wait);
2063
2064         spin_lock(&fiq->lock);
2065         if (!fiq->connected)
2066                 mask = EPOLLERR;
2067         else if (request_pending(fiq))
2068                 mask |= EPOLLIN | EPOLLRDNORM;
2069         spin_unlock(&fiq->lock);
2070
2071         return mask;
2072 }
2073
2074 /* Abort all requests on the given list (pending or processing) */
2075 static void end_requests(struct list_head *head)
2076 {
2077         while (!list_empty(head)) {
2078                 struct fuse_req *req;
2079                 req = list_entry(head->next, struct fuse_req, list);
2080                 req->out.h.error = -ECONNABORTED;
2081                 clear_bit(FR_SENT, &req->flags);
2082                 list_del_init(&req->list);
2083                 fuse_request_end(req);
2084         }
2085 }
2086
2087 static void end_polls(struct fuse_conn *fc)
2088 {
2089         struct rb_node *p;
2090
2091         p = rb_first(&fc->polled_files);
2092
2093         while (p) {
2094                 struct fuse_file *ff;
2095                 ff = rb_entry(p, struct fuse_file, polled_node);
2096                 wake_up_interruptible_all(&ff->poll_wait);
2097
2098                 p = rb_next(p);
2099         }
2100 }
2101
2102 /*
2103  * Abort all requests.
2104  *
2105  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2106  * filesystem.
2107  *
2108  * The same effect is usually achievable through killing the filesystem daemon
2109  * and all users of the filesystem.  The exception is the combination of an
2110  * asynchronous request and the tricky deadlock (see
2111  * Documentation/filesystems/fuse.rst).
2112  *
2113  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2114  * requests, they should be finished off immediately.  Locked requests will be
2115  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2116  * requests.  It is possible that some request will finish before we can.  This
2117  * is OK, the request will in that case be removed from the list before we touch
2118  * it.
2119  */
2120 void fuse_abort_conn(struct fuse_conn *fc)
2121 {
2122         struct fuse_iqueue *fiq = &fc->iq;
2123
2124         spin_lock(&fc->lock);
2125         if (fc->connected) {
2126                 struct fuse_dev *fud;
2127                 struct fuse_req *req, *next;
2128                 LIST_HEAD(to_end);
2129                 unsigned int i;
2130
2131                 /* Background queuing checks fc->connected under bg_lock */
2132                 spin_lock(&fc->bg_lock);
2133                 fc->connected = 0;
2134                 spin_unlock(&fc->bg_lock);
2135
2136                 fuse_set_initialized(fc);
2137                 list_for_each_entry(fud, &fc->devices, entry) {
2138                         struct fuse_pqueue *fpq = &fud->pq;
2139
2140                         spin_lock(&fpq->lock);
2141                         fpq->connected = 0;
2142                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2143                                 req->out.h.error = -ECONNABORTED;
2144                                 spin_lock(&req->waitq.lock);
2145                                 set_bit(FR_ABORTED, &req->flags);
2146                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2147                                         set_bit(FR_PRIVATE, &req->flags);
2148                                         __fuse_get_request(req);
2149                                         list_move(&req->list, &to_end);
2150                                 }
2151                                 spin_unlock(&req->waitq.lock);
2152                         }
2153                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2154                                 list_splice_tail_init(&fpq->processing[i],
2155                                                       &to_end);
2156                         spin_unlock(&fpq->lock);
2157                 }
2158                 spin_lock(&fc->bg_lock);
2159                 fc->blocked = 0;
2160                 fc->max_background = UINT_MAX;
2161                 flush_bg_queue(fc);
2162                 spin_unlock(&fc->bg_lock);
2163
2164                 spin_lock(&fiq->lock);
2165                 fiq->connected = 0;
2166                 list_for_each_entry(req, &fiq->pending, list)
2167                         clear_bit(FR_PENDING, &req->flags);
2168                 list_splice_tail_init(&fiq->pending, &to_end);
2169                 while (forget_pending(fiq))
2170                         kfree(fuse_dequeue_forget(fiq, 1, NULL));
2171                 wake_up_all(&fiq->waitq);
2172                 spin_unlock(&fiq->lock);
2173                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2174                 end_polls(fc);
2175                 wake_up_all(&fc->blocked_waitq);
2176                 spin_unlock(&fc->lock);
2177
2178                 end_requests(&to_end);
2179         } else {
2180                 spin_unlock(&fc->lock);
2181         }
2182 }
2183 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2184
2185 void fuse_wait_aborted(struct fuse_conn *fc)
2186 {
2187         /* matches implicit memory barrier in fuse_drop_waiting() */
2188         smp_mb();
2189         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2190 }
2191
2192 int fuse_dev_release(struct inode *inode, struct file *file)
2193 {
2194         struct fuse_dev *fud = fuse_get_dev(file);
2195
2196         if (fud) {
2197                 struct fuse_conn *fc = fud->fc;
2198                 struct fuse_pqueue *fpq = &fud->pq;
2199                 LIST_HEAD(to_end);
2200                 unsigned int i;
2201
2202                 spin_lock(&fpq->lock);
2203                 WARN_ON(!list_empty(&fpq->io));
2204                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2205                         list_splice_init(&fpq->processing[i], &to_end);
2206                 spin_unlock(&fpq->lock);
2207
2208                 end_requests(&to_end);
2209
2210                 /* Are we the last open device? */
2211                 if (atomic_dec_and_test(&fc->dev_count)) {
2212                         WARN_ON(fc->iq.fasync != NULL);
2213                         fuse_abort_conn(fc);
2214                 }
2215                 fuse_dev_free(fud);
2216         }
2217         return 0;
2218 }
2219 EXPORT_SYMBOL_GPL(fuse_dev_release);
2220
2221 static int fuse_dev_fasync(int fd, struct file *file, int on)
2222 {
2223         struct fuse_dev *fud = fuse_get_dev(file);
2224
2225         if (!fud)
2226                 return -EPERM;
2227
2228         /* No locking - fasync_helper does its own locking */
2229         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2230 }
2231
2232 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2233 {
2234         struct fuse_dev *fud;
2235
2236         if (new->private_data)
2237                 return -EINVAL;
2238
2239         fud = fuse_dev_alloc_install(fc);
2240         if (!fud)
2241                 return -ENOMEM;
2242
2243         new->private_data = fud;
2244         atomic_inc(&fc->dev_count);
2245
2246         return 0;
2247 }
2248
2249 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2250                            unsigned long arg)
2251 {
2252         int res;
2253         int oldfd;
2254         struct fuse_dev *fud = NULL;
2255
2256         switch (cmd) {
2257         case FUSE_DEV_IOC_CLONE:
2258                 res = -EFAULT;
2259                 if (!get_user(oldfd, (__u32 __user *)arg)) {
2260                         struct file *old = fget(oldfd);
2261
2262                         res = -EINVAL;
2263                         if (old) {
2264                                 /*
2265                                  * Check against file->f_op because CUSE
2266                                  * uses the same ioctl handler.
2267                                  */
2268                                 if (old->f_op == file->f_op &&
2269                                     old->f_cred->user_ns == file->f_cred->user_ns)
2270                                         fud = fuse_get_dev(old);
2271
2272                                 if (fud) {
2273                                         mutex_lock(&fuse_mutex);
2274                                         res = fuse_device_clone(fud->fc, file);
2275                                         mutex_unlock(&fuse_mutex);
2276                                 }
2277                                 fput(old);
2278                         }
2279                 }
2280                 break;
2281         default:
2282                 res = -ENOTTY;
2283                 break;
2284         }
2285         return res;
2286 }
2287
2288 const struct file_operations fuse_dev_operations = {
2289         .owner          = THIS_MODULE,
2290         .open           = fuse_dev_open,
2291         .llseek         = no_llseek,
2292         .read_iter      = fuse_dev_read,
2293         .splice_read    = fuse_dev_splice_read,
2294         .write_iter     = fuse_dev_write,
2295         .splice_write   = fuse_dev_splice_write,
2296         .poll           = fuse_dev_poll,
2297         .release        = fuse_dev_release,
2298         .fasync         = fuse_dev_fasync,
2299         .unlocked_ioctl = fuse_dev_ioctl,
2300         .compat_ioctl   = compat_ptr_ioctl,
2301 };
2302 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2303
2304 static struct miscdevice fuse_miscdevice = {
2305         .minor = FUSE_MINOR,
2306         .name  = "fuse",
2307         .fops = &fuse_dev_operations,
2308 };
2309
2310 int __init fuse_dev_init(void)
2311 {
2312         int err = -ENOMEM;
2313         fuse_req_cachep = kmem_cache_create("fuse_request",
2314                                             sizeof(struct fuse_req),
2315                                             0, 0, NULL);
2316         if (!fuse_req_cachep)
2317                 goto out;
2318
2319         err = misc_register(&fuse_miscdevice);
2320         if (err)
2321                 goto out_cache_clean;
2322
2323         return 0;
2324
2325  out_cache_clean:
2326         kmem_cache_destroy(fuse_req_cachep);
2327  out:
2328         return err;
2329 }
2330
2331 void fuse_dev_cleanup(void)
2332 {
2333         misc_deregister(&fuse_miscdevice);
2334         kmem_cache_destroy(fuse_req_cachep);
2335 }