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 intr_entry check. Pairs with
 292          * smp_mb() from queue_interrupt().
 293          */
 294         if (!list_empty(&req->intr_entry)) {
 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_reclaim |
 787                1 << PG_waiters))) {
 788                 dump_page(page, "fuse: trying to steal weird page");
 789                 return 1;
 790         }
 791         return 0;
 792 }
 793
 794 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 795 {
 796         int err;
 797         struct page *oldpage = *pagep;
 798         struct page *newpage;
 799         struct pipe_buffer *buf = cs->pipebufs;
 800
 801         get_page(oldpage);
 802         err = unlock_request(cs->req);
 803         if (err)
 804                 goto out_put_old;
 805
 806         fuse_copy_finish(cs);
 807
 808         err = pipe_buf_confirm(cs->pipe, buf);
 809         if (err)
 810                 goto out_put_old;
 811
 812         BUG_ON(!cs->nr_segs);
 813         cs->currbuf = buf;
 814         cs->len = buf->len;
 815         cs->pipebufs++;
 816         cs->nr_segs--;
 817
 818         if (cs->len != PAGE_SIZE)
 819                 goto out_fallback;
 820
 821         if (!pipe_buf_try_steal(cs->pipe, buf))
 822                 goto out_fallback;
 823
 824         newpage = buf->page;
 825
 826         if (!PageUptodate(newpage))
 827                 SetPageUptodate(newpage);
 828
 829         ClearPageMappedToDisk(newpage);
 830
 831         if (fuse_check_page(newpage) != 0)
 832                 goto out_fallback_unlock;
 833
 834         /*
 835          * This is a new and locked page, it shouldn't be mapped or
 836          * have any special flags on it
 837          */
 838         if (WARN_ON(page_mapped(oldpage)))
 839                 goto out_fallback_unlock;
 840         if (WARN_ON(page_has_private(oldpage)))
 841                 goto out_fallback_unlock;
 842         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 843                 goto out_fallback_unlock;
 844         if (WARN_ON(PageMlocked(oldpage)))
 845                 goto out_fallback_unlock;
 846
 847         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 848         if (err) {
 849                 unlock_page(newpage);
 850                 goto out_put_old;
 851         }
 852
 853         get_page(newpage);
 854
 855         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 856                 lru_cache_add(newpage);
 857
 858         err = 0;
 859         spin_lock(&cs->req->waitq.lock);
 860         if (test_bit(FR_ABORTED, &cs->req->flags))
 861                 err = -ENOENT;
 862         else
 863                 *pagep = newpage;
 864         spin_unlock(&cs->req->waitq.lock);
 865
 866         if (err) {
 867                 unlock_page(newpage);
 868                 put_page(newpage);
 869                 goto out_put_old;
 870         }
 871
 872         unlock_page(oldpage);
 873         /* Drop ref for ap->pages[] array */
 874         put_page(oldpage);
 875         cs->len = 0;
 876
 877         err = 0;
 878 out_put_old:
 879         /* Drop ref obtained in this function */
 880         put_page(oldpage);
 881         return err;
 882
 883 out_fallback_unlock:
 884         unlock_page(newpage);
 885 out_fallback:
 886         cs->pg = buf->page;
 887         cs->offset = buf->offset;
 888
 889         err = lock_request(cs->req);
 890         if (!err)
 891                 err = 1;
 892
 893         goto out_put_old;
 894 }
 895
 896 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 897                          unsigned offset, unsigned count)
 898 {
 899         struct pipe_buffer *buf;
 900         int err;
 901
 902         if (cs->nr_segs >= cs->pipe->max_usage)
 903                 return -EIO;
 904
 905         get_page(page);
 906         err = unlock_request(cs->req);
 907         if (err) {
 908                 put_page(page);
 909                 return err;
 910         }
 911
 912         fuse_copy_finish(cs);
 913
 914         buf = cs->pipebufs;
 915         buf->page = page;
 916         buf->offset = offset;
 917         buf->len = count;
 918
 919         cs->pipebufs++;
 920         cs->nr_segs++;
 921         cs->len = 0;
 922
 923         return 0;
 924 }
 925
 926 /*
 927  * Copy a page in the request to/from the userspace buffer.  Must be
 928  * done atomically
 929  */
 930 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 931                           unsigned offset, unsigned count, int zeroing)
 932 {
 933         int err;
 934         struct page *page = *pagep;
 935
 936         if (page && zeroing && count < PAGE_SIZE)
 937                 clear_highpage(page);
 938
 939         while (count) {
 940                 if (cs->write && cs->pipebufs && page) {
 941                         return fuse_ref_page(cs, page, offset, count);
 942                 } else if (!cs->len) {
 943                         if (cs->move_pages && page &&
 944                             offset == 0 && count == PAGE_SIZE) {
 945                                 err = fuse_try_move_page(cs, pagep);
 946                                 if (err <= 0)
 947                                         return err;
 948                         } else {
 949                                 err = fuse_copy_fill(cs);
 950                                 if (err)
 951                                         return err;
 952                         }
 953                 }
 954                 if (page) {
 955                         void *mapaddr = kmap_atomic(page);
 956                         void *buf = mapaddr + offset;
 957                         offset += fuse_copy_do(cs, &buf, &count);
 958                         kunmap_atomic(mapaddr);
 959                 } else
 960                         offset += fuse_copy_do(cs, NULL, &count);
 961         }
 962         if (page && !cs->write)
 963                 flush_dcache_page(page);
 964         return 0;
 965 }
 966
 967 /* Copy pages in the request to/from userspace buffer */
 968 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 969                            int zeroing)
 970 {
 971         unsigned i;
 972         struct fuse_req *req = cs->req;
 973         struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
 974
 975
 976         for (i = 0; i < ap->num_pages && (nbytes || zeroing); i++) {
 977                 int err;
 978                 unsigned int offset = ap->descs[i].offset;
 979                 unsigned int count = min(nbytes, ap->descs[i].length);
 980
 981                 err = fuse_copy_page(cs, &ap->pages[i], offset, count, zeroing);
 982                 if (err)
 983                         return err;
 984
 985                 nbytes -= count;
 986         }
 987         return 0;
 988 }
 989
 990 /* Copy a single argument in the request to/from userspace buffer */
 991 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 992 {
 993         while (size) {
 994                 if (!cs->len) {
 995                         int err = fuse_copy_fill(cs);
 996                         if (err)
 997                                 return err;
 998                 }
 999                 fuse_copy_do(cs, &val, &size);
1000         }
1001         return 0;
1002 }
1003
1004 /* Copy request arguments to/from userspace buffer */
1005 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1006                           unsigned argpages, struct fuse_arg *args,
1007                           int zeroing)
1008 {
1009         int err = 0;
1010         unsigned i;
1011
1012         for (i = 0; !err && i < numargs; i++)  {
1013                 struct fuse_arg *arg = &args[i];
1014                 if (i == numargs - 1 && argpages)
1015                         err = fuse_copy_pages(cs, arg->size, zeroing);
1016                 else
1017                         err = fuse_copy_one(cs, arg->value, arg->size);
1018         }
1019         return err;
1020 }
1021
1022 static int forget_pending(struct fuse_iqueue *fiq)
1023 {
1024         return fiq->forget_list_head.next != NULL;
1025 }
1026
1027 static int request_pending(struct fuse_iqueue *fiq)
1028 {
1029         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1030                 forget_pending(fiq);
1031 }
1032
1033 /*
1034  * Transfer an interrupt request to userspace
1035  *
1036  * Unlike other requests this is assembled on demand, without a need
1037  * to allocate a separate fuse_req structure.
1038  *
1039  * Called with fiq->lock held, releases it
1040  */
1041 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1042                                struct fuse_copy_state *cs,
1043                                size_t nbytes, struct fuse_req *req)
1044 __releases(fiq->lock)
1045 {
1046         struct fuse_in_header ih;
1047         struct fuse_interrupt_in arg;
1048         unsigned reqsize = sizeof(ih) + sizeof(arg);
1049         int err;
1050
1051         list_del_init(&req->intr_entry);
1052         memset(&ih, 0, sizeof(ih));
1053         memset(&arg, 0, sizeof(arg));
1054         ih.len = reqsize;
1055         ih.opcode = FUSE_INTERRUPT;
1056         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1057         arg.unique = req->in.h.unique;
1058
1059         spin_unlock(&fiq->lock);
1060         if (nbytes < reqsize)
1061                 return -EINVAL;
1062
1063         err = fuse_copy_one(cs, &ih, sizeof(ih));
1064         if (!err)
1065                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1066         fuse_copy_finish(cs);
1067
1068         return err ? err : reqsize;
1069 }
1070
1071 struct fuse_forget_link *fuse_dequeue_forget(struct fuse_iqueue *fiq,
1072                                              unsigned int max,
1073                                              unsigned int *countp)
1074 {
1075         struct fuse_forget_link *head = fiq->forget_list_head.next;
1076         struct fuse_forget_link **newhead = &head;
1077         unsigned count;
1078
1079         for (count = 0; *newhead != NULL && count < max; count++)
1080                 newhead = &(*newhead)->next;
1081
1082         fiq->forget_list_head.next = *newhead;
1083         *newhead = NULL;
1084         if (fiq->forget_list_head.next == NULL)
1085                 fiq->forget_list_tail = &fiq->forget_list_head;
1086
1087         if (countp != NULL)
1088                 *countp = count;
1089
1090         return head;
1091 }
1092 EXPORT_SYMBOL(fuse_dequeue_forget);
1093
1094 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1095                                    struct fuse_copy_state *cs,
1096                                    size_t nbytes)
1097 __releases(fiq->lock)
1098 {
1099         int err;
1100         struct fuse_forget_link *forget = fuse_dequeue_forget(fiq, 1, NULL);
1101         struct fuse_forget_in arg = {
1102                 .nlookup = forget->forget_one.nlookup,
1103         };
1104         struct fuse_in_header ih = {
1105                 .opcode = FUSE_FORGET,
1106                 .nodeid = forget->forget_one.nodeid,
1107                 .unique = fuse_get_unique(fiq),
1108                 .len = sizeof(ih) + sizeof(arg),
1109         };
1110
1111         spin_unlock(&fiq->lock);
1112         kfree(forget);
1113         if (nbytes < ih.len)
1114                 return -EINVAL;
1115
1116         err = fuse_copy_one(cs, &ih, sizeof(ih));
1117         if (!err)
1118                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1119         fuse_copy_finish(cs);
1120
1121         if (err)
1122                 return err;
1123
1124         return ih.len;
1125 }
1126
1127 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1128                                    struct fuse_copy_state *cs, size_t nbytes)
1129 __releases(fiq->lock)
1130 {
1131         int err;
1132         unsigned max_forgets;
1133         unsigned count;
1134         struct fuse_forget_link *head;
1135         struct fuse_batch_forget_in arg = { .count = 0 };
1136         struct fuse_in_header ih = {
1137                 .opcode = FUSE_BATCH_FORGET,
1138                 .unique = fuse_get_unique(fiq),
1139                 .len = sizeof(ih) + sizeof(arg),
1140         };
1141
1142         if (nbytes < ih.len) {
1143                 spin_unlock(&fiq->lock);
1144                 return -EINVAL;
1145         }
1146
1147         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1148         head = fuse_dequeue_forget(fiq, max_forgets, &count);
1149         spin_unlock(&fiq->lock);
1150
1151         arg.count = count;
1152         ih.len += count * sizeof(struct fuse_forget_one);
1153         err = fuse_copy_one(cs, &ih, sizeof(ih));
1154         if (!err)
1155                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1156
1157         while (head) {
1158                 struct fuse_forget_link *forget = head;
1159
1160                 if (!err) {
1161                         err = fuse_copy_one(cs, &forget->forget_one,
1162                                             sizeof(forget->forget_one));
1163                 }
1164                 head = forget->next;
1165                 kfree(forget);
1166         }
1167
1168         fuse_copy_finish(cs);
1169
1170         if (err)
1171                 return err;
1172
1173         return ih.len;
1174 }
1175
1176 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1177                             struct fuse_copy_state *cs,
1178                             size_t nbytes)
1179 __releases(fiq->lock)
1180 {
1181         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1182                 return fuse_read_single_forget(fiq, cs, nbytes);
1183         else
1184                 return fuse_read_batch_forget(fiq, cs, nbytes);
1185 }
1186
1187 /*
1188  * Read a single request into the userspace filesystem's buffer.  This
1189  * function waits until a request is available, then removes it from
1190  * the pending list and copies request data to userspace buffer.  If
1191  * no reply is needed (FORGET) or request has been aborted or there
1192  * was an error during the copying then it's finished by calling
1193  * fuse_request_end().  Otherwise add it to the processing list, and set
1194  * the 'sent' flag.
1195  */
1196 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1197                                 struct fuse_copy_state *cs, size_t nbytes)
1198 {
1199         ssize_t err;
1200         struct fuse_conn *fc = fud->fc;
1201         struct fuse_iqueue *fiq = &fc->iq;
1202         struct fuse_pqueue *fpq = &fud->pq;
1203         struct fuse_req *req;
1204         struct fuse_args *args;
1205         unsigned reqsize;
1206         unsigned int hash;
1207
1208         /*
1209          * Require sane minimum read buffer - that has capacity for fixed part
1210          * of any request header + negotiated max_write room for data.
1211          *
1212          * Historically libfuse reserves 4K for fixed header room, but e.g.
1213          * GlusterFS reserves only 80 bytes
1214          *
1215          *      = `sizeof(fuse_in_header) + sizeof(fuse_write_in)`
1216          *
1217          * which is the absolute minimum any sane filesystem should be using
1218          * for header room.
1219          */
1220         if (nbytes < max_t(size_t, FUSE_MIN_READ_BUFFER,
1221                            sizeof(struct fuse_in_header) +
1222                            sizeof(struct fuse_write_in) +
1223                            fc->max_write))
1224                 return -EINVAL;
1225
1226  restart:
1227         for (;;) {
1228                 spin_lock(&fiq->lock);
1229                 if (!fiq->connected || request_pending(fiq))
1230                         break;
1231                 spin_unlock(&fiq->lock);
1232
1233                 if (file->f_flags & O_NONBLOCK)
1234                         return -EAGAIN;
1235                 err = wait_event_interruptible_exclusive(fiq->waitq,
1236                                 !fiq->connected || request_pending(fiq));
1237                 if (err)
1238                         return err;
1239         }
1240
1241         if (!fiq->connected) {
1242                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1243                 goto err_unlock;
1244         }
1245
1246         if (!list_empty(&fiq->interrupts)) {
1247                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1248                                  intr_entry);
1249                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1250         }
1251
1252         if (forget_pending(fiq)) {
1253                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1254                         return fuse_read_forget(fc, fiq, cs, nbytes);
1255
1256                 if (fiq->forget_batch <= -8)
1257                         fiq->forget_batch = 16;
1258         }
1259
1260         req = list_entry(fiq->pending.next, struct fuse_req, list);
1261         clear_bit(FR_PENDING, &req->flags);
1262         list_del_init(&req->list);
1263         spin_unlock(&fiq->lock);
1264
1265         args = req->args;
1266         reqsize = req->in.h.len;
1267
1268         /* If request is too large, reply with an error and restart the read */
1269         if (nbytes < reqsize) {
1270                 req->out.h.error = -EIO;
1271                 /* SETXATTR is special, since it may contain too large data */
1272                 if (args->opcode == FUSE_SETXATTR)
1273                         req->out.h.error = -E2BIG;
1274                 fuse_request_end(req);
1275                 goto restart;
1276         }
1277         spin_lock(&fpq->lock);
1278         list_add(&req->list, &fpq->io);
1279         spin_unlock(&fpq->lock);
1280         cs->req = req;
1281         err = fuse_copy_one(cs, &req->in.h, sizeof(req->in.h));
1282         if (!err)
1283                 err = fuse_copy_args(cs, args->in_numargs, args->in_pages,
1284                                      (struct fuse_arg *) args->in_args, 0);
1285         fuse_copy_finish(cs);
1286         spin_lock(&fpq->lock);
1287         clear_bit(FR_LOCKED, &req->flags);
1288         if (!fpq->connected) {
1289                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1290                 goto out_end;
1291         }
1292         if (err) {
1293                 req->out.h.error = -EIO;
1294                 goto out_end;
1295         }
1296         if (!test_bit(FR_ISREPLY, &req->flags)) {
1297                 err = reqsize;
1298                 goto out_end;
1299         }
1300         hash = fuse_req_hash(req->in.h.unique);
1301         list_move_tail(&req->list, &fpq->processing[hash]);
1302         __fuse_get_request(req);
1303         set_bit(FR_SENT, &req->flags);
1304         spin_unlock(&fpq->lock);
1305         /* matches barrier in request_wait_answer() */
1306         smp_mb__after_atomic();
1307         if (test_bit(FR_INTERRUPTED, &req->flags))
1308                 queue_interrupt(req);
1309         fuse_put_request(req);
1310
1311         return reqsize;
1312
1313 out_end:
1314         if (!test_bit(FR_PRIVATE, &req->flags))
1315                 list_del_init(&req->list);
1316         spin_unlock(&fpq->lock);
1317         fuse_request_end(req);
1318         return err;
1319
1320  err_unlock:
1321         spin_unlock(&fiq->lock);
1322         return err;
1323 }
1324
1325 static int fuse_dev_open(struct inode *inode, struct file *file)
1326 {
1327         /*
1328          * The fuse device's file's private_data is used to hold
1329          * the fuse_conn(ection) when it is mounted, and is used to
1330          * keep track of whether the file has been mounted already.
1331          */
1332         file->private_data = NULL;
1333         return 0;
1334 }
1335
1336 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1337 {
1338         struct fuse_copy_state cs;
1339         struct file *file = iocb->ki_filp;
1340         struct fuse_dev *fud = fuse_get_dev(file);
1341
1342         if (!fud)
1343                 return -EPERM;
1344
1345         if (!iter_is_iovec(to))
1346                 return -EINVAL;
1347
1348         fuse_copy_init(&cs, 1, to);
1349
1350         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1351 }
1352
1353 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1354                                     struct pipe_inode_info *pipe,
1355                                     size_t len, unsigned int flags)
1356 {
1357         int total, ret;
1358         int page_nr = 0;
1359         struct pipe_buffer *bufs;
1360         struct fuse_copy_state cs;
1361         struct fuse_dev *fud = fuse_get_dev(in);
1362
1363         if (!fud)
1364                 return -EPERM;
1365
1366         bufs = kvmalloc_array(pipe->max_usage, sizeof(struct pipe_buffer),
1367                               GFP_KERNEL);
1368         if (!bufs)
1369                 return -ENOMEM;
1370
1371         fuse_copy_init(&cs, 1, NULL);
1372         cs.pipebufs = bufs;
1373         cs.pipe = pipe;
1374         ret = fuse_dev_do_read(fud, in, &cs, len);
1375         if (ret < 0)
1376                 goto out;
1377
1378         if (pipe_occupancy(pipe->head, pipe->tail) + cs.nr_segs > pipe->max_usage) {
1379                 ret = -EIO;
1380                 goto out;
1381         }
1382
1383         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1384                 /*
1385                  * Need to be careful about this.  Having buf->ops in module
1386                  * code can Oops if the buffer persists after module unload.
1387                  */
1388                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1389                 bufs[page_nr].flags = 0;
1390                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1391                 if (unlikely(ret < 0))
1392                         break;
1393         }
1394         if (total)
1395                 ret = total;
1396 out:
1397         for (; page_nr < cs.nr_segs; page_nr++)
1398                 put_page(bufs[page_nr].page);
1399
1400         kvfree(bufs);
1401         return ret;
1402 }
1403
1404 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1405                             struct fuse_copy_state *cs)
1406 {
1407         struct fuse_notify_poll_wakeup_out outarg;
1408         int err = -EINVAL;
1409
1410         if (size != sizeof(outarg))
1411                 goto err;
1412
1413         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1414         if (err)
1415                 goto err;
1416
1417         fuse_copy_finish(cs);
1418         return fuse_notify_poll_wakeup(fc, &outarg);
1419
1420 err:
1421         fuse_copy_finish(cs);
1422         return err;
1423 }
1424
1425 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1426                                    struct fuse_copy_state *cs)
1427 {
1428         struct fuse_notify_inval_inode_out outarg;
1429         int err = -EINVAL;
1430
1431         if (size != sizeof(outarg))
1432                 goto err;
1433
1434         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1435         if (err)
1436                 goto err;
1437         fuse_copy_finish(cs);
1438
1439         down_read(&fc->killsb);
1440         err = fuse_reverse_inval_inode(fc, outarg.ino,
1441                                        outarg.off, outarg.len);
1442         up_read(&fc->killsb);
1443         return err;
1444
1445 err:
1446         fuse_copy_finish(cs);
1447         return err;
1448 }
1449
1450 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1451                                    struct fuse_copy_state *cs)
1452 {
1453         struct fuse_notify_inval_entry_out outarg;
1454         int err = -ENOMEM;
1455         char *buf;
1456         struct qstr name;
1457
1458         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1459         if (!buf)
1460                 goto err;
1461
1462         err = -EINVAL;
1463         if (size < sizeof(outarg))
1464                 goto err;
1465
1466         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1467         if (err)
1468                 goto err;
1469
1470         err = -ENAMETOOLONG;
1471         if (outarg.namelen > FUSE_NAME_MAX)
1472                 goto err;
1473
1474         err = -EINVAL;
1475         if (size != sizeof(outarg) + outarg.namelen + 1)
1476                 goto err;
1477
1478         name.name = buf;
1479         name.len = outarg.namelen;
1480         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1481         if (err)
1482                 goto err;
1483         fuse_copy_finish(cs);
1484         buf[outarg.namelen] = 0;
1485
1486         down_read(&fc->killsb);
1487         err = fuse_reverse_inval_entry(fc, outarg.parent, 0, &name);
1488         up_read(&fc->killsb);
1489         kfree(buf);
1490         return err;
1491
1492 err:
1493         kfree(buf);
1494         fuse_copy_finish(cs);
1495         return err;
1496 }
1497
1498 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1499                               struct fuse_copy_state *cs)
1500 {
1501         struct fuse_notify_delete_out outarg;
1502         int err = -ENOMEM;
1503         char *buf;
1504         struct qstr name;
1505
1506         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1507         if (!buf)
1508                 goto err;
1509
1510         err = -EINVAL;
1511         if (size < sizeof(outarg))
1512                 goto err;
1513
1514         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1515         if (err)
1516                 goto err;
1517
1518         err = -ENAMETOOLONG;
1519         if (outarg.namelen > FUSE_NAME_MAX)
1520                 goto err;
1521
1522         err = -EINVAL;
1523         if (size != sizeof(outarg) + outarg.namelen + 1)
1524                 goto err;
1525
1526         name.name = buf;
1527         name.len = outarg.namelen;
1528         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1529         if (err)
1530                 goto err;
1531         fuse_copy_finish(cs);
1532         buf[outarg.namelen] = 0;
1533
1534         down_read(&fc->killsb);
1535         err = fuse_reverse_inval_entry(fc, outarg.parent, outarg.child, &name);
1536         up_read(&fc->killsb);
1537         kfree(buf);
1538         return err;
1539
1540 err:
1541         kfree(buf);
1542         fuse_copy_finish(cs);
1543         return err;
1544 }
1545
1546 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1547                              struct fuse_copy_state *cs)
1548 {
1549         struct fuse_notify_store_out outarg;
1550         struct inode *inode;
1551         struct address_space *mapping;
1552         u64 nodeid;
1553         int err;
1554         pgoff_t index;
1555         unsigned int offset;
1556         unsigned int num;
1557         loff_t file_size;
1558         loff_t end;
1559
1560         err = -EINVAL;
1561         if (size < sizeof(outarg))
1562                 goto out_finish;
1563
1564         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1565         if (err)
1566                 goto out_finish;
1567
1568         err = -EINVAL;
1569         if (size - sizeof(outarg) != outarg.size)
1570                 goto out_finish;
1571
1572         nodeid = outarg.nodeid;
1573
1574         down_read(&fc->killsb);
1575
1576         err = -ENOENT;
1577         inode = fuse_ilookup(fc, nodeid,  NULL);
1578         if (!inode)
1579                 goto out_up_killsb;
1580
1581         mapping = inode->i_mapping;
1582         index = outarg.offset >> PAGE_SHIFT;
1583         offset = outarg.offset & ~PAGE_MASK;
1584         file_size = i_size_read(inode);
1585         end = outarg.offset + outarg.size;
1586         if (end > file_size) {
1587                 file_size = end;
1588                 fuse_write_update_size(inode, file_size);
1589         }
1590
1591         num = outarg.size;
1592         while (num) {
1593                 struct page *page;
1594                 unsigned int this_num;
1595
1596                 err = -ENOMEM;
1597                 page = find_or_create_page(mapping, index,
1598                                            mapping_gfp_mask(mapping));
1599                 if (!page)
1600                         goto out_iput;
1601
1602                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1603                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1604                 if (!err && offset == 0 &&
1605                     (this_num == PAGE_SIZE || file_size == end))
1606                         SetPageUptodate(page);
1607                 unlock_page(page);
1608                 put_page(page);
1609
1610                 if (err)
1611                         goto out_iput;
1612
1613                 num -= this_num;
1614                 offset = 0;
1615                 index++;
1616         }
1617
1618         err = 0;
1619
1620 out_iput:
1621         iput(inode);
1622 out_up_killsb:
1623         up_read(&fc->killsb);
1624 out_finish:
1625         fuse_copy_finish(cs);
1626         return err;
1627 }
1628
1629 struct fuse_retrieve_args {
1630         struct fuse_args_pages ap;
1631         struct fuse_notify_retrieve_in inarg;
1632 };
1633
1634 static void fuse_retrieve_end(struct fuse_mount *fm, struct fuse_args *args,
1635                               int error)
1636 {
1637         struct fuse_retrieve_args *ra =
1638                 container_of(args, typeof(*ra), ap.args);
1639
1640         release_pages(ra->ap.pages, ra->ap.num_pages);
1641         kfree(ra);
1642 }
1643
1644 static int fuse_retrieve(struct fuse_mount *fm, struct inode *inode,
1645                          struct fuse_notify_retrieve_out *outarg)
1646 {
1647         int err;
1648         struct address_space *mapping = inode->i_mapping;
1649         pgoff_t index;
1650         loff_t file_size;
1651         unsigned int num;
1652         unsigned int offset;
1653         size_t total_len = 0;
1654         unsigned int num_pages;
1655         struct fuse_conn *fc = fm->fc;
1656         struct fuse_retrieve_args *ra;
1657         size_t args_size = sizeof(*ra);
1658         struct fuse_args_pages *ap;
1659         struct fuse_args *args;
1660
1661         offset = outarg->offset & ~PAGE_MASK;
1662         file_size = i_size_read(inode);
1663
1664         num = min(outarg->size, fc->max_write);
1665         if (outarg->offset > file_size)
1666                 num = 0;
1667         else if (outarg->offset + num > file_size)
1668                 num = file_size - outarg->offset;
1669
1670         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1671         num_pages = min(num_pages, fc->max_pages);
1672
1673         args_size += num_pages * (sizeof(ap->pages[0]) + sizeof(ap->descs[0]));
1674
1675         ra = kzalloc(args_size, GFP_KERNEL);
1676         if (!ra)
1677                 return -ENOMEM;
1678
1679         ap = &ra->ap;
1680         ap->pages = (void *) (ra + 1);
1681         ap->descs = (void *) (ap->pages + num_pages);
1682
1683         args = &ap->args;
1684         args->nodeid = outarg->nodeid;
1685         args->opcode = FUSE_NOTIFY_REPLY;
1686         args->in_numargs = 2;
1687         args->in_pages = true;
1688         args->end = fuse_retrieve_end;
1689
1690         index = outarg->offset >> PAGE_SHIFT;
1691
1692         while (num && ap->num_pages < num_pages) {
1693                 struct page *page;
1694                 unsigned int this_num;
1695
1696                 page = find_get_page(mapping, index);
1697                 if (!page)
1698                         break;
1699
1700                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1701                 ap->pages[ap->num_pages] = page;
1702                 ap->descs[ap->num_pages].offset = offset;
1703                 ap->descs[ap->num_pages].length = this_num;
1704                 ap->num_pages++;
1705
1706                 offset = 0;
1707                 num -= this_num;
1708                 total_len += this_num;
1709                 index++;
1710         }
1711         ra->inarg.offset = outarg->offset;
1712         ra->inarg.size = total_len;
1713         args->in_args[0].size = sizeof(ra->inarg);
1714         args->in_args[0].value = &ra->inarg;
1715         args->in_args[1].size = total_len;
1716
1717         err = fuse_simple_notify_reply(fm, args, outarg->notify_unique);
1718         if (err)
1719                 fuse_retrieve_end(fm, args, err);
1720
1721         return err;
1722 }
1723
1724 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1725                                 struct fuse_copy_state *cs)
1726 {
1727         struct fuse_notify_retrieve_out outarg;
1728         struct fuse_mount *fm;
1729         struct inode *inode;
1730         u64 nodeid;
1731         int err;
1732
1733         err = -EINVAL;
1734         if (size != sizeof(outarg))
1735                 goto copy_finish;
1736
1737         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1738         if (err)
1739                 goto copy_finish;
1740
1741         fuse_copy_finish(cs);
1742
1743         down_read(&fc->killsb);
1744         err = -ENOENT;
1745         nodeid = outarg.nodeid;
1746
1747         inode = fuse_ilookup(fc, nodeid, &fm);
1748         if (inode) {
1749                 err = fuse_retrieve(fm, inode, &outarg);
1750                 iput(inode);
1751         }
1752         up_read(&fc->killsb);
1753
1754         return err;
1755
1756 copy_finish:
1757         fuse_copy_finish(cs);
1758         return err;
1759 }
1760
1761 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1762                        unsigned int size, struct fuse_copy_state *cs)
1763 {
1764         /* Don't try to move pages (yet) */
1765         cs->move_pages = 0;
1766
1767         switch (code) {
1768         case FUSE_NOTIFY_POLL:
1769                 return fuse_notify_poll(fc, size, cs);
1770
1771         case FUSE_NOTIFY_INVAL_INODE:
1772                 return fuse_notify_inval_inode(fc, size, cs);
1773
1774         case FUSE_NOTIFY_INVAL_ENTRY:
1775                 return fuse_notify_inval_entry(fc, size, cs);
1776
1777         case FUSE_NOTIFY_STORE:
1778                 return fuse_notify_store(fc, size, cs);
1779
1780         case FUSE_NOTIFY_RETRIEVE:
1781                 return fuse_notify_retrieve(fc, size, cs);
1782
1783         case FUSE_NOTIFY_DELETE:
1784                 return fuse_notify_delete(fc, size, cs);
1785
1786         default:
1787                 fuse_copy_finish(cs);
1788                 return -EINVAL;
1789         }
1790 }
1791
1792 /* Look up request on processing list by unique ID */
1793 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1794 {
1795         unsigned int hash = fuse_req_hash(unique);
1796         struct fuse_req *req;
1797
1798         list_for_each_entry(req, &fpq->processing[hash], list) {
1799                 if (req->in.h.unique == unique)
1800                         return req;
1801         }
1802         return NULL;
1803 }
1804
1805 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_args *args,
1806                          unsigned nbytes)
1807 {
1808         unsigned reqsize = sizeof(struct fuse_out_header);
1809
1810         reqsize += fuse_len_args(args->out_numargs, args->out_args);
1811
1812         if (reqsize < nbytes || (reqsize > nbytes && !args->out_argvar))
1813                 return -EINVAL;
1814         else if (reqsize > nbytes) {
1815                 struct fuse_arg *lastarg = &args->out_args[args->out_numargs-1];
1816                 unsigned diffsize = reqsize - nbytes;
1817
1818                 if (diffsize > lastarg->size)
1819                         return -EINVAL;
1820                 lastarg->size -= diffsize;
1821         }
1822         return fuse_copy_args(cs, args->out_numargs, args->out_pages,
1823                               args->out_args, args->page_zeroing);
1824 }
1825
1826 /*
1827  * Write a single reply to a request.  First the header is copied from
1828  * the write buffer.  The request is then searched on the processing
1829  * list by the unique ID found in the header.  If found, then remove
1830  * it from the list and copy the rest of the buffer to the request.
1831  * The request is finished by calling fuse_request_end().
1832  */
1833 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1834                                  struct fuse_copy_state *cs, size_t nbytes)
1835 {
1836         int err;
1837         struct fuse_conn *fc = fud->fc;
1838         struct fuse_pqueue *fpq = &fud->pq;
1839         struct fuse_req *req;
1840         struct fuse_out_header oh;
1841
1842         err = -EINVAL;
1843         if (nbytes < sizeof(struct fuse_out_header))
1844                 goto out;
1845
1846         err = fuse_copy_one(cs, &oh, sizeof(oh));
1847         if (err)
1848                 goto copy_finish;
1849
1850         err = -EINVAL;
1851         if (oh.len != nbytes)
1852                 goto copy_finish;
1853
1854         /*
1855          * Zero oh.unique indicates unsolicited notification message
1856          * and error contains notification code.
1857          */
1858         if (!oh.unique) {
1859                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1860                 goto out;
1861         }
1862
1863         err = -EINVAL;
1864         if (oh.error <= -1000 || oh.error > 0)
1865                 goto copy_finish;
1866
1867         spin_lock(&fpq->lock);
1868         req = NULL;
1869         if (fpq->connected)
1870                 req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1871
1872         err = -ENOENT;
1873         if (!req) {
1874                 spin_unlock(&fpq->lock);
1875                 goto copy_finish;
1876         }
1877
1878         /* Is it an interrupt reply ID? */
1879         if (oh.unique & FUSE_INT_REQ_BIT) {
1880                 __fuse_get_request(req);
1881                 spin_unlock(&fpq->lock);
1882
1883                 err = 0;
1884                 if (nbytes != sizeof(struct fuse_out_header))
1885                         err = -EINVAL;
1886                 else if (oh.error == -ENOSYS)
1887                         fc->no_interrupt = 1;
1888                 else if (oh.error == -EAGAIN)
1889                         err = queue_interrupt(req);
1890
1891                 fuse_put_request(req);
1892
1893                 goto copy_finish;
1894         }
1895
1896         clear_bit(FR_SENT, &req->flags);
1897         list_move(&req->list, &fpq->io);
1898         req->out.h = oh;
1899         set_bit(FR_LOCKED, &req->flags);
1900         spin_unlock(&fpq->lock);
1901         cs->req = req;
1902         if (!req->args->page_replace)
1903                 cs->move_pages = 0;
1904
1905         if (oh.error)
1906                 err = nbytes != sizeof(oh) ? -EINVAL : 0;
1907         else
1908                 err = copy_out_args(cs, req->args, nbytes);
1909         fuse_copy_finish(cs);
1910
1911         spin_lock(&fpq->lock);
1912         clear_bit(FR_LOCKED, &req->flags);
1913         if (!fpq->connected)
1914                 err = -ENOENT;
1915         else if (err)
1916                 req->out.h.error = -EIO;
1917         if (!test_bit(FR_PRIVATE, &req->flags))
1918                 list_del_init(&req->list);
1919         spin_unlock(&fpq->lock);
1920
1921         fuse_request_end(req);
1922 out:
1923         return err ? err : nbytes;
1924
1925 copy_finish:
1926         fuse_copy_finish(cs);
1927         goto out;
1928 }
1929
1930 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1931 {
1932         struct fuse_copy_state cs;
1933         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1934
1935         if (!fud)
1936                 return -EPERM;
1937
1938         if (!iter_is_iovec(from))
1939                 return -EINVAL;
1940
1941         fuse_copy_init(&cs, 0, from);
1942
1943         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1944 }
1945
1946 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1947                                      struct file *out, loff_t *ppos,
1948                                      size_t len, unsigned int flags)
1949 {
1950         unsigned int head, tail, mask, count;
1951         unsigned nbuf;
1952         unsigned idx;
1953         struct pipe_buffer *bufs;
1954         struct fuse_copy_state cs;
1955         struct fuse_dev *fud;
1956         size_t rem;
1957         ssize_t ret;
1958
1959         fud = fuse_get_dev(out);
1960         if (!fud)
1961                 return -EPERM;
1962
1963         pipe_lock(pipe);
1964
1965         head = pipe->head;
1966         tail = pipe->tail;
1967         mask = pipe->ring_size - 1;
1968         count = head - tail;
1969
1970         bufs = kvmalloc_array(count, sizeof(struct pipe_buffer), GFP_KERNEL);
1971         if (!bufs) {
1972                 pipe_unlock(pipe);
1973                 return -ENOMEM;
1974         }
1975
1976         nbuf = 0;
1977         rem = 0;
1978         for (idx = tail; idx != head && rem < len; idx++)
1979                 rem += pipe->bufs[idx & mask].len;
1980
1981         ret = -EINVAL;
1982         if (rem < len)
1983                 goto out_free;
1984
1985         rem = len;
1986         while (rem) {
1987                 struct pipe_buffer *ibuf;
1988                 struct pipe_buffer *obuf;
1989
1990                 if (WARN_ON(nbuf >= count || tail == head))
1991                         goto out_free;
1992
1993                 ibuf = &pipe->bufs[tail & mask];
1994                 obuf = &bufs[nbuf];
1995
1996                 if (rem >= ibuf->len) {
1997                         *obuf = *ibuf;
1998                         ibuf->ops = NULL;
1999                         tail++;
2000                         pipe->tail = tail;
2001                 } else {
2002                         if (!pipe_buf_get(pipe, ibuf))
2003                                 goto out_free;
2004
2005                         *obuf = *ibuf;
2006                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2007                         obuf->len = rem;
2008                         ibuf->offset += obuf->len;
2009                         ibuf->len -= obuf->len;
2010                 }
2011                 nbuf++;
2012                 rem -= obuf->len;
2013         }
2014         pipe_unlock(pipe);
2015
2016         fuse_copy_init(&cs, 0, NULL);
2017         cs.pipebufs = bufs;
2018         cs.nr_segs = nbuf;
2019         cs.pipe = pipe;
2020
2021         if (flags & SPLICE_F_MOVE)
2022                 cs.move_pages = 1;
2023
2024         ret = fuse_dev_do_write(fud, &cs, len);
2025
2026         pipe_lock(pipe);
2027 out_free:
2028         for (idx = 0; idx < nbuf; idx++)
2029                 pipe_buf_release(pipe, &bufs[idx]);
2030         pipe_unlock(pipe);
2031
2032         kvfree(bufs);
2033         return ret;
2034 }
2035
2036 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2037 {
2038         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2039         struct fuse_iqueue *fiq;
2040         struct fuse_dev *fud = fuse_get_dev(file);
2041
2042         if (!fud)
2043                 return EPOLLERR;
2044
2045         fiq = &fud->fc->iq;
2046         poll_wait(file, &fiq->waitq, wait);
2047
2048         spin_lock(&fiq->lock);
2049         if (!fiq->connected)
2050                 mask = EPOLLERR;
2051         else if (request_pending(fiq))
2052                 mask |= EPOLLIN | EPOLLRDNORM;
2053         spin_unlock(&fiq->lock);
2054
2055         return mask;
2056 }
2057
2058 /* Abort all requests on the given list (pending or processing) */
2059 static void end_requests(struct list_head *head)
2060 {
2061         while (!list_empty(head)) {
2062                 struct fuse_req *req;
2063                 req = list_entry(head->next, struct fuse_req, list);
2064                 req->out.h.error = -ECONNABORTED;
2065                 clear_bit(FR_SENT, &req->flags);
2066                 list_del_init(&req->list);
2067                 fuse_request_end(req);
2068         }
2069 }
2070
2071 static void end_polls(struct fuse_conn *fc)
2072 {
2073         struct rb_node *p;
2074
2075         p = rb_first(&fc->polled_files);
2076
2077         while (p) {
2078                 struct fuse_file *ff;
2079                 ff = rb_entry(p, struct fuse_file, polled_node);
2080                 wake_up_interruptible_all(&ff->poll_wait);
2081
2082                 p = rb_next(p);
2083         }
2084 }
2085
2086 /*
2087  * Abort all requests.
2088  *
2089  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2090  * filesystem.
2091  *
2092  * The same effect is usually achievable through killing the filesystem daemon
2093  * and all users of the filesystem.  The exception is the combination of an
2094  * asynchronous request and the tricky deadlock (see
2095  * Documentation/filesystems/fuse.rst).
2096  *
2097  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2098  * requests, they should be finished off immediately.  Locked requests will be
2099  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2100  * requests.  It is possible that some request will finish before we can.  This
2101  * is OK, the request will in that case be removed from the list before we touch
2102  * it.
2103  */
2104 void fuse_abort_conn(struct fuse_conn *fc)
2105 {
2106         struct fuse_iqueue *fiq = &fc->iq;
2107
2108         spin_lock(&fc->lock);
2109         if (fc->connected) {
2110                 struct fuse_dev *fud;
2111                 struct fuse_req *req, *next;
2112                 LIST_HEAD(to_end);
2113                 unsigned int i;
2114
2115                 /* Background queuing checks fc->connected under bg_lock */
2116                 spin_lock(&fc->bg_lock);
2117                 fc->connected = 0;
2118                 spin_unlock(&fc->bg_lock);
2119
2120                 fuse_set_initialized(fc);
2121                 list_for_each_entry(fud, &fc->devices, entry) {
2122                         struct fuse_pqueue *fpq = &fud->pq;
2123
2124                         spin_lock(&fpq->lock);
2125                         fpq->connected = 0;
2126                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2127                                 req->out.h.error = -ECONNABORTED;
2128                                 spin_lock(&req->waitq.lock);
2129                                 set_bit(FR_ABORTED, &req->flags);
2130                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2131                                         set_bit(FR_PRIVATE, &req->flags);
2132                                         __fuse_get_request(req);
2133                                         list_move(&req->list, &to_end);
2134                                 }
2135                                 spin_unlock(&req->waitq.lock);
2136                         }
2137                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2138                                 list_splice_tail_init(&fpq->processing[i],
2139                                                       &to_end);
2140                         spin_unlock(&fpq->lock);
2141                 }
2142                 spin_lock(&fc->bg_lock);
2143                 fc->blocked = 0;
2144                 fc->max_background = UINT_MAX;
2145                 flush_bg_queue(fc);
2146                 spin_unlock(&fc->bg_lock);
2147
2148                 spin_lock(&fiq->lock);
2149                 fiq->connected = 0;
2150                 list_for_each_entry(req, &fiq->pending, list)
2151                         clear_bit(FR_PENDING, &req->flags);
2152                 list_splice_tail_init(&fiq->pending, &to_end);
2153                 while (forget_pending(fiq))
2154                         kfree(fuse_dequeue_forget(fiq, 1, NULL));
2155                 wake_up_all(&fiq->waitq);
2156                 spin_unlock(&fiq->lock);
2157                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2158                 end_polls(fc);
2159                 wake_up_all(&fc->blocked_waitq);
2160                 spin_unlock(&fc->lock);
2161
2162                 end_requests(&to_end);
2163         } else {
2164                 spin_unlock(&fc->lock);
2165         }
2166 }
2167 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2168
2169 void fuse_wait_aborted(struct fuse_conn *fc)
2170 {
2171         /* matches implicit memory barrier in fuse_drop_waiting() */
2172         smp_mb();
2173         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2174 }
2175
2176 int fuse_dev_release(struct inode *inode, struct file *file)
2177 {
2178         struct fuse_dev *fud = fuse_get_dev(file);
2179
2180         if (fud) {
2181                 struct fuse_conn *fc = fud->fc;
2182                 struct fuse_pqueue *fpq = &fud->pq;
2183                 LIST_HEAD(to_end);
2184                 unsigned int i;
2185
2186                 spin_lock(&fpq->lock);
2187                 WARN_ON(!list_empty(&fpq->io));
2188                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2189                         list_splice_init(&fpq->processing[i], &to_end);
2190                 spin_unlock(&fpq->lock);
2191
2192                 end_requests(&to_end);
2193
2194                 /* Are we the last open device? */
2195                 if (atomic_dec_and_test(&fc->dev_count)) {
2196                         WARN_ON(fc->iq.fasync != NULL);
2197                         fuse_abort_conn(fc);
2198                 }
2199                 fuse_dev_free(fud);
2200         }
2201         return 0;
2202 }
2203 EXPORT_SYMBOL_GPL(fuse_dev_release);
2204
2205 static int fuse_dev_fasync(int fd, struct file *file, int on)
2206 {
2207         struct fuse_dev *fud = fuse_get_dev(file);
2208
2209         if (!fud)
2210                 return -EPERM;
2211
2212         /* No locking - fasync_helper does its own locking */
2213         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2214 }
2215
2216 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2217 {
2218         struct fuse_dev *fud;
2219
2220         if (new->private_data)
2221                 return -EINVAL;
2222
2223         fud = fuse_dev_alloc_install(fc);
2224         if (!fud)
2225                 return -ENOMEM;
2226
2227         new->private_data = fud;
2228         atomic_inc(&fc->dev_count);
2229
2230         return 0;
2231 }
2232
2233 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2234                            unsigned long arg)
2235 {
2236         int err = -ENOTTY;
2237
2238         if (cmd == FUSE_DEV_IOC_CLONE) {
2239                 int oldfd;
2240
2241                 err = -EFAULT;
2242                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2243                         struct file *old = fget(oldfd);
2244
2245                         err = -EINVAL;
2246                         if (old) {
2247                                 struct fuse_dev *fud = NULL;
2248
2249                                 /*
2250                                  * Check against file->f_op because CUSE
2251                                  * uses the same ioctl handler.
2252                                  */
2253                                 if (old->f_op == file->f_op &&
2254                                     old->f_cred->user_ns == file->f_cred->user_ns)
2255                                         fud = fuse_get_dev(old);
2256
2257                                 if (fud) {
2258                                         mutex_lock(&fuse_mutex);
2259                                         err = fuse_device_clone(fud->fc, file);
2260                                         mutex_unlock(&fuse_mutex);
2261                                 }
2262                                 fput(old);
2263                         }
2264                 }
2265         }
2266         return err;
2267 }
2268
2269 const struct file_operations fuse_dev_operations = {
2270         .owner          = THIS_MODULE,
2271         .open           = fuse_dev_open,
2272         .llseek         = no_llseek,
2273         .read_iter      = fuse_dev_read,
2274         .splice_read    = fuse_dev_splice_read,
2275         .write_iter     = fuse_dev_write,
2276         .splice_write   = fuse_dev_splice_write,
2277         .poll           = fuse_dev_poll,
2278         .release        = fuse_dev_release,
2279         .fasync         = fuse_dev_fasync,
2280         .unlocked_ioctl = fuse_dev_ioctl,
2281         .compat_ioctl   = compat_ptr_ioctl,
2282 };
2283 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2284
2285 static struct miscdevice fuse_miscdevice = {
2286         .minor = FUSE_MINOR,
2287         .name  = "fuse",
2288         .fops = &fuse_dev_operations,
2289 };
2290
2291 int __init fuse_dev_init(void)
2292 {
2293         int err = -ENOMEM;
2294         fuse_req_cachep = kmem_cache_create("fuse_request",
2295                                             sizeof(struct fuse_req),
2296                                             0, 0, NULL);
2297         if (!fuse_req_cachep)
2298                 goto out;
2299
2300         err = misc_register(&fuse_miscdevice);
2301         if (err)
2302                 goto out_cache_clean;
2303
2304         return 0;
2305
2306  out_cache_clean:
2307         kmem_cache_destroy(fuse_req_cachep);
2308  out:
2309         return err;
2310 }
2311
2312 void fuse_dev_cleanup(void)
2313 {
2314         misc_deregister(&fuse_miscdevice);
2315         kmem_cache_destroy(fuse_req_cachep);
2316 }