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                 fc->num_background--;
 319                 fc->active_background--;
 320                 flush_bg_queue(fc);
 321                 spin_unlock(&fc->bg_lock);
 322         } else {
 323                 /* Wake up waiter sleeping in request_wait_answer() */
 324                 wake_up(&req->waitq);
 325         }
 326
 327         if (test_bit(FR_ASYNC, &req->flags))
 328                 req->args->end(fm, req->args, req->out.h.error);
 329 put_request:
 330         fuse_put_request(req);
 331 }
 332 EXPORT_SYMBOL_GPL(fuse_request_end);
 333
 334 static int queue_interrupt(struct fuse_req *req)
 335 {
 336         struct fuse_iqueue *fiq = &req->fm->fc->iq;
 337
 338         spin_lock(&fiq->lock);
 339         /* Check for we've sent request to interrupt this req */
 340         if (unlikely(!test_bit(FR_INTERRUPTED, &req->flags))) {
 341                 spin_unlock(&fiq->lock);
 342                 return -EINVAL;
 343         }
 344
 345         if (list_empty(&req->intr_entry)) {
 346                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 347                 /*
 348                  * Pairs with smp_mb() implied by test_and_set_bit()
 349                  * from fuse_request_end().
 350                  */
 351                 smp_mb();
 352                 if (test_bit(FR_FINISHED, &req->flags)) {
 353                         list_del_init(&req->intr_entry);
 354                         spin_unlock(&fiq->lock);
 355                         return 0;
 356                 }
 357                 fiq->ops->wake_interrupt_and_unlock(fiq);
 358         } else {
 359                 spin_unlock(&fiq->lock);
 360         }
 361         return 0;
 362 }
 363
 364 static void request_wait_answer(struct fuse_req *req)
 365 {
 366         struct fuse_conn *fc = req->fm->fc;
 367         struct fuse_iqueue *fiq = &fc->iq;
 368         int err;
 369
 370         if (!fc->no_interrupt) {
 371                 /* Any signal may interrupt this */
 372                 err = wait_event_interruptible(req->waitq,
 373                                         test_bit(FR_FINISHED, &req->flags));
 374                 if (!err)
 375                         return;
 376
 377                 set_bit(FR_INTERRUPTED, &req->flags);
 378                 /* matches barrier in fuse_dev_do_read() */
 379                 smp_mb__after_atomic();
 380                 if (test_bit(FR_SENT, &req->flags))
 381                         queue_interrupt(req);
 382         }
 383
 384         if (!test_bit(FR_FORCE, &req->flags)) {
 385                 /* Only fatal signals may interrupt this */
 386                 err = wait_event_killable(req->waitq,
 387                                         test_bit(FR_FINISHED, &req->flags));
 388                 if (!err)
 389                         return;
 390
 391                 spin_lock(&fiq->lock);
 392                 /* Request is not yet in userspace, bail out */
 393                 if (test_bit(FR_PENDING, &req->flags)) {
 394                         list_del(&req->list);
 395                         spin_unlock(&fiq->lock);
 396                         __fuse_put_request(req);
 397                         req->out.h.error = -EINTR;
 398                         return;
 399                 }
 400                 spin_unlock(&fiq->lock);
 401         }
 402
 403         /*
 404          * Either request is already in userspace, or it was forced.
 405          * Wait it out.
 406          */
 407         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 408 }
 409
 410 static void __fuse_request_send(struct fuse_req *req)
 411 {
 412         struct fuse_iqueue *fiq = &req->fm->fc->iq;
 413
 414         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 415         spin_lock(&fiq->lock);
 416         if (!fiq->connected) {
 417                 spin_unlock(&fiq->lock);
 418                 req->out.h.error = -ENOTCONN;
 419         } else {
 420                 req->in.h.unique = fuse_get_unique(fiq);
 421                 /* acquire extra reference, since request is still needed
 422                    after fuse_request_end() */
 423                 __fuse_get_request(req);
 424                 queue_request_and_unlock(fiq, req);
 425
 426                 request_wait_answer(req);
 427                 /* Pairs with smp_wmb() in fuse_request_end() */
 428                 smp_rmb();
 429         }
 430 }
 431
 432 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 433 {
 434         if (fc->minor < 4 && args->opcode == FUSE_STATFS)
 435                 args->out_args[0].size = FUSE_COMPAT_STATFS_SIZE;
 436
 437         if (fc->minor < 9) {
 438                 switch (args->opcode) {
 439                 case FUSE_LOOKUP:
 440                 case FUSE_CREATE:
 441                 case FUSE_MKNOD:
 442                 case FUSE_MKDIR:
 443                 case FUSE_SYMLINK:
 444                 case FUSE_LINK:
 445                         args->out_args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 446                         break;
 447                 case FUSE_GETATTR:
 448                 case FUSE_SETATTR:
 449                         args->out_args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 450                         break;
 451                 }
 452         }
 453         if (fc->minor < 12) {
 454                 switch (args->opcode) {
 455                 case FUSE_CREATE:
 456                         args->in_args[0].size = sizeof(struct fuse_open_in);
 457                         break;
 458                 case FUSE_MKNOD:
 459                         args->in_args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 460                         break;
 461                 }
 462         }
 463 }
 464
 465 static void fuse_force_creds(struct fuse_req *req)
 466 {
 467         struct fuse_conn *fc = req->fm->fc;
 468
 469         req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
 470         req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
 471         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 472 }
 473
 474 static void fuse_args_to_req(struct fuse_req *req, struct fuse_args *args)
 475 {
 476         req->in.h.opcode = args->opcode;
 477         req->in.h.nodeid = args->nodeid;
 478         req->args = args;
 479         if (args->end)
 480                 __set_bit(FR_ASYNC, &req->flags);
 481 }
 482
 483 ssize_t fuse_simple_request(struct fuse_mount *fm, struct fuse_args *args)
 484 {
 485         struct fuse_conn *fc = fm->fc;
 486         struct fuse_req *req;
 487         ssize_t ret;
 488
 489         if (args->force) {
 490                 atomic_inc(&fc->num_waiting);
 491                 req = fuse_request_alloc(fm, GFP_KERNEL | __GFP_NOFAIL);
 492
 493                 if (!args->nocreds)
 494                         fuse_force_creds(req);
 495
 496                 __set_bit(FR_WAITING, &req->flags);
 497                 __set_bit(FR_FORCE, &req->flags);
 498         } else {
 499                 WARN_ON(args->nocreds);
 500                 req = fuse_get_req(fm, false);
 501                 if (IS_ERR(req))
 502                         return PTR_ERR(req);
 503         }
 504
 505         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 506         fuse_adjust_compat(fc, args);
 507         fuse_args_to_req(req, args);
 508
 509         if (!args->noreply)
 510                 __set_bit(FR_ISREPLY, &req->flags);
 511         __fuse_request_send(req);
 512         ret = req->out.h.error;
 513         if (!ret && args->out_argvar) {
 514                 BUG_ON(args->out_numargs == 0);
 515                 ret = args->out_args[args->out_numargs - 1].size;
 516         }
 517         fuse_put_request(req);
 518
 519         return ret;
 520 }
 521
 522 static bool fuse_request_queue_background(struct fuse_req *req)
 523 {
 524         struct fuse_mount *fm = req->fm;
 525         struct fuse_conn *fc = fm->fc;
 526         bool queued = false;
 527
 528         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 529         if (!test_bit(FR_WAITING, &req->flags)) {
 530                 __set_bit(FR_WAITING, &req->flags);
 531                 atomic_inc(&fc->num_waiting);
 532         }
 533         __set_bit(FR_ISREPLY, &req->flags);
 534         spin_lock(&fc->bg_lock);
 535         if (likely(fc->connected)) {
 536                 fc->num_background++;
 537                 if (fc->num_background == fc->max_background)
 538                         fc->blocked = 1;
 539                 list_add_tail(&req->list, &fc->bg_queue);
 540                 flush_bg_queue(fc);
 541                 queued = true;
 542         }
 543         spin_unlock(&fc->bg_lock);
 544
 545         return queued;
 546 }
 547
 548 int fuse_simple_background(struct fuse_mount *fm, struct fuse_args *args,
 549                             gfp_t gfp_flags)
 550 {
 551         struct fuse_req *req;
 552
 553         if (args->force) {
 554                 WARN_ON(!args->nocreds);
 555                 req = fuse_request_alloc(fm, gfp_flags);
 556                 if (!req)
 557                         return -ENOMEM;
 558                 __set_bit(FR_BACKGROUND, &req->flags);
 559         } else {
 560                 WARN_ON(args->nocreds);
 561                 req = fuse_get_req(fm, true);
 562                 if (IS_ERR(req))
 563                         return PTR_ERR(req);
 564         }
 565
 566         fuse_args_to_req(req, args);
 567
 568         if (!fuse_request_queue_background(req)) {
 569                 fuse_put_request(req);
 570                 return -ENOTCONN;
 571         }
 572
 573         return 0;
 574 }
 575 EXPORT_SYMBOL_GPL(fuse_simple_background);
 576
 577 static int fuse_simple_notify_reply(struct fuse_mount *fm,
 578                                     struct fuse_args *args, u64 unique)
 579 {
 580         struct fuse_req *req;
 581         struct fuse_iqueue *fiq = &fm->fc->iq;
 582         int err = 0;
 583
 584         req = fuse_get_req(fm, false);
 585         if (IS_ERR(req))
 586                 return PTR_ERR(req);
 587
 588         __clear_bit(FR_ISREPLY, &req->flags);
 589         req->in.h.unique = unique;
 590
 591         fuse_args_to_req(req, args);
 592
 593         spin_lock(&fiq->lock);
 594         if (fiq->connected) {
 595                 queue_request_and_unlock(fiq, req);
 596         } else {
 597                 err = -ENODEV;
 598                 spin_unlock(&fiq->lock);
 599                 fuse_put_request(req);
 600         }
 601
 602         return err;
 603 }
 604
 605 /*
 606  * Lock the request.  Up to the next unlock_request() there mustn't be
 607  * anything that could cause a page-fault.  If the request was already
 608  * aborted bail out.
 609  */
 610 static int lock_request(struct fuse_req *req)
 611 {
 612         int err = 0;
 613         if (req) {
 614                 spin_lock(&req->waitq.lock);
 615                 if (test_bit(FR_ABORTED, &req->flags))
 616                         err = -ENOENT;
 617                 else
 618                         set_bit(FR_LOCKED, &req->flags);
 619                 spin_unlock(&req->waitq.lock);
 620         }
 621         return err;
 622 }
 623
 624 /*
 625  * Unlock request.  If it was aborted while locked, caller is responsible
 626  * for unlocking and ending the request.
 627  */
 628 static int unlock_request(struct fuse_req *req)
 629 {
 630         int err = 0;
 631         if (req) {
 632                 spin_lock(&req->waitq.lock);
 633                 if (test_bit(FR_ABORTED, &req->flags))
 634                         err = -ENOENT;
 635                 else
 636                         clear_bit(FR_LOCKED, &req->flags);
 637                 spin_unlock(&req->waitq.lock);
 638         }
 639         return err;
 640 }
 641
 642 struct fuse_copy_state {
 643         int write;
 644         struct fuse_req *req;
 645         struct iov_iter *iter;
 646         struct pipe_buffer *pipebufs;
 647         struct pipe_buffer *currbuf;
 648         struct pipe_inode_info *pipe;
 649         unsigned long nr_segs;
 650         struct page *pg;
 651         unsigned len;
 652         unsigned offset;
 653         unsigned move_pages:1;
 654 };
 655
 656 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 657                            struct iov_iter *iter)
 658 {
 659         memset(cs, 0, sizeof(*cs));
 660         cs->write = write;
 661         cs->iter = iter;
 662 }
 663
 664 /* Unmap and put previous page of userspace buffer */
 665 static void fuse_copy_finish(struct fuse_copy_state *cs)
 666 {
 667         if (cs->currbuf) {
 668                 struct pipe_buffer *buf = cs->currbuf;
 669
 670                 if (cs->write)
 671                         buf->len = PAGE_SIZE - cs->len;
 672                 cs->currbuf = NULL;
 673         } else if (cs->pg) {
 674                 if (cs->write) {
 675                         flush_dcache_page(cs->pg);
 676                         set_page_dirty_lock(cs->pg);
 677                 }
 678                 put_page(cs->pg);
 679         }
 680         cs->pg = NULL;
 681 }
 682
 683 /*
 684  * Get another pagefull of userspace buffer, and map it to kernel
 685  * address space, and lock request
 686  */
 687 static int fuse_copy_fill(struct fuse_copy_state *cs)
 688 {
 689         struct page *page;
 690         int err;
 691
 692         err = unlock_request(cs->req);
 693         if (err)
 694                 return err;
 695
 696         fuse_copy_finish(cs);
 697         if (cs->pipebufs) {
 698                 struct pipe_buffer *buf = cs->pipebufs;
 699
 700                 if (!cs->write) {
 701                         err = pipe_buf_confirm(cs->pipe, buf);
 702                         if (err)
 703                                 return err;
 704
 705                         BUG_ON(!cs->nr_segs);
 706                         cs->currbuf = buf;
 707                         cs->pg = buf->page;
 708                         cs->offset = buf->offset;
 709                         cs->len = buf->len;
 710                         cs->pipebufs++;
 711                         cs->nr_segs--;
 712                 } else {
 713                         if (cs->nr_segs >= cs->pipe->max_usage)
 714                                 return -EIO;
 715
 716                         page = alloc_page(GFP_HIGHUSER);
 717                         if (!page)
 718                                 return -ENOMEM;
 719
 720                         buf->page = page;
 721                         buf->offset = 0;
 722                         buf->len = 0;
 723
 724                         cs->currbuf = buf;
 725                         cs->pg = page;
 726                         cs->offset = 0;
 727                         cs->len = PAGE_SIZE;
 728                         cs->pipebufs++;
 729                         cs->nr_segs++;
 730                 }
 731         } else {
 732                 size_t off;
 733                 err = iov_iter_get_pages2(cs->iter, &page, PAGE_SIZE, 1, &off);
 734                 if (err < 0)
 735                         return err;
 736                 BUG_ON(!err);
 737                 cs->len = err;
 738                 cs->offset = off;
 739                 cs->pg = page;
 740         }
 741
 742         return lock_request(cs->req);
 743 }
 744
 745 /* Do as much copy to/from userspace buffer as we can */
 746 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 747 {
 748         unsigned ncpy = min(*size, cs->len);
 749         if (val) {
 750                 void *pgaddr = kmap_local_page(cs->pg);
 751                 void *buf = pgaddr + cs->offset;
 752
 753                 if (cs->write)
 754                         memcpy(buf, *val, ncpy);
 755                 else
 756                         memcpy(*val, buf, ncpy);
 757
 758                 kunmap_local(pgaddr);
 759                 *val += ncpy;
 760         }
 761         *size -= ncpy;
 762         cs->len -= ncpy;
 763         cs->offset += ncpy;
 764         return ncpy;
 765 }
 766
 767 static int fuse_check_page(struct page *page)
 768 {
 769         if (page_mapcount(page) ||
 770             page->mapping != NULL ||
 771             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 772              ~(1 << PG_locked |
 773                1 << PG_referenced |
 774                1 << PG_uptodate |
 775                1 << PG_lru |
 776                1 << PG_active |
 777                1 << PG_workingset |
 778                1 << PG_reclaim |
 779                1 << PG_waiters))) {
 780                 dump_page(page, "fuse: trying to steal weird page");
 781                 return 1;
 782         }
 783         return 0;
 784 }
 785
 786 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 787 {
 788         int err;
 789         struct page *oldpage = *pagep;
 790         struct page *newpage;
 791         struct pipe_buffer *buf = cs->pipebufs;
 792
 793         get_page(oldpage);
 794         err = unlock_request(cs->req);
 795         if (err)
 796                 goto out_put_old;
 797
 798         fuse_copy_finish(cs);
 799
 800         err = pipe_buf_confirm(cs->pipe, buf);
 801         if (err)
 802                 goto out_put_old;
 803
 804         BUG_ON(!cs->nr_segs);
 805         cs->currbuf = buf;
 806         cs->len = buf->len;
 807         cs->pipebufs++;
 808         cs->nr_segs--;
 809
 810         if (cs->len != PAGE_SIZE)
 811                 goto out_fallback;
 812
 813         if (!pipe_buf_try_steal(cs->pipe, buf))
 814                 goto out_fallback;
 815
 816         newpage = buf->page;
 817
 818         if (!PageUptodate(newpage))
 819                 SetPageUptodate(newpage);
 820
 821         ClearPageMappedToDisk(newpage);
 822
 823         if (fuse_check_page(newpage) != 0)
 824                 goto out_fallback_unlock;
 825
 826         /*
 827          * This is a new and locked page, it shouldn't be mapped or
 828          * have any special flags on it
 829          */
 830         if (WARN_ON(page_mapped(oldpage)))
 831                 goto out_fallback_unlock;
 832         if (WARN_ON(page_has_private(oldpage)))
 833                 goto out_fallback_unlock;
 834         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 835                 goto out_fallback_unlock;
 836         if (WARN_ON(PageMlocked(oldpage)))
 837                 goto out_fallback_unlock;
 838
 839         replace_page_cache_page(oldpage, newpage);
 840
 841         get_page(newpage);
 842
 843         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 844                 lru_cache_add(newpage);
 845
 846         /*
 847          * Release while we have extra ref on stolen page.  Otherwise
 848          * anon_pipe_buf_release() might think the page can be reused.
 849          */
 850         pipe_buf_release(cs->pipe, buf);
 851
 852         err = 0;
 853         spin_lock(&cs->req->waitq.lock);
 854         if (test_bit(FR_ABORTED, &cs->req->flags))
 855                 err = -ENOENT;
 856         else
 857                 *pagep = newpage;
 858         spin_unlock(&cs->req->waitq.lock);
 859
 860         if (err) {
 861                 unlock_page(newpage);
 862                 put_page(newpage);
 863                 goto out_put_old;
 864         }
 865
 866         unlock_page(oldpage);
 867         /* Drop ref for ap->pages[] array */
 868         put_page(oldpage);
 869         cs->len = 0;
 870
 871         err = 0;
 872 out_put_old:
 873         /* Drop ref obtained in this function */
 874         put_page(oldpage);
 875         return err;
 876
 877 out_fallback_unlock:
 878         unlock_page(newpage);
 879 out_fallback:
 880         cs->pg = buf->page;
 881         cs->offset = buf->offset;
 882
 883         err = lock_request(cs->req);
 884         if (!err)
 885                 err = 1;
 886
 887         goto out_put_old;
 888 }
 889
 890 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 891                          unsigned offset, unsigned count)
 892 {
 893         struct pipe_buffer *buf;
 894         int err;
 895
 896         if (cs->nr_segs >= cs->pipe->max_usage)
 897                 return -EIO;
 898
 899         get_page(page);
 900         err = unlock_request(cs->req);
 901         if (err) {
 902                 put_page(page);
 903                 return err;
 904         }
 905
 906         fuse_copy_finish(cs);
 907
 908         buf = cs->pipebufs;
 909         buf->page = page;
 910         buf->offset = offset;
 911         buf->len = count;
 912
 913         cs->pipebufs++;
 914         cs->nr_segs++;
 915         cs->len = 0;
 916
 917         return 0;
 918 }
 919
 920 /*
 921  * Copy a page in the request to/from the userspace buffer.  Must be
 922  * done atomically
 923  */
 924 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 925                           unsigned offset, unsigned count, int zeroing)
 926 {
 927         int err;
 928         struct page *page = *pagep;
 929
 930         if (page && zeroing && count < PAGE_SIZE)
 931                 clear_highpage(page);
 932
 933         while (count) {
 934                 if (cs->write && cs->pipebufs && page) {
 935                         /*
 936                          * Can't control lifetime of pipe buffers, so always
 937                          * copy user pages.
 938                          */
 939                         if (cs->req->args->user_pages) {
 940                                 err = fuse_copy_fill(cs);
 941                                 if (err)
 942                                         return err;
 943                         } else {
 944                                 return fuse_ref_page(cs, page, offset, count);
 945                         }
 946                 } else if (!cs->len) {
 947                         if (cs->move_pages && page &&
 948                             offset == 0 && count == PAGE_SIZE) {
 949                                 err = fuse_try_move_page(cs, pagep);
 950                                 if (err <= 0)
 951                                         return err;
 952                         } else {
 953                                 err = fuse_copy_fill(cs);
 954                                 if (err)
 955                                         return err;
 956                         }
 957                 }
 958                 if (page) {
 959                         void *mapaddr = kmap_local_page(page);
 960                         void *buf = mapaddr + offset;
 961                         offset += fuse_copy_do(cs, &buf, &count);
 962                         kunmap_local(mapaddr);
 963                 } else
 964                         offset += fuse_copy_do(cs, NULL, &count);
 965         }
 966         if (page && !cs->write)
 967                 flush_dcache_page(page);
 968         return 0;
 969 }
 970
 971 /* Copy pages in the request to/from userspace buffer */
 972 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 973                            int zeroing)
 974 {
 975         unsigned i;
 976         struct fuse_req *req = cs->req;
 977         struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
 978
 979
 980         for (i = 0; i < ap->num_pages && (nbytes || zeroing); i++) {
 981                 int err;
 982                 unsigned int offset = ap->descs[i].offset;
 983                 unsigned int count = min(nbytes, ap->descs[i].length);
 984
 985                 err = fuse_copy_page(cs, &ap->pages[i], offset, count, zeroing);
 986                 if (err)
 987                         return err;
 988
 989                 nbytes -= count;
 990         }
 991         return 0;
 992 }
 993
 994 /* Copy a single argument in the request to/from userspace buffer */
 995 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 996 {
 997         while (size) {
 998                 if (!cs->len) {
 999                         int err = fuse_copy_fill(cs);
1000                         if (err)
1001                                 return err;
1002                 }
1003                 fuse_copy_do(cs, &val, &size);
1004         }
1005         return 0;
1006 }
1007
1008 /* Copy request arguments to/from userspace buffer */
1009 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1010                           unsigned argpages, struct fuse_arg *args,
1011                           int zeroing)
1012 {
1013         int err = 0;
1014         unsigned i;
1015
1016         for (i = 0; !err && i < numargs; i++)  {
1017                 struct fuse_arg *arg = &args[i];
1018                 if (i == numargs - 1 && argpages)
1019                         err = fuse_copy_pages(cs, arg->size, zeroing);
1020                 else
1021                         err = fuse_copy_one(cs, arg->value, arg->size);
1022         }
1023         return err;
1024 }
1025
1026 static int forget_pending(struct fuse_iqueue *fiq)
1027 {
1028         return fiq->forget_list_head.next != NULL;
1029 }
1030
1031 static int request_pending(struct fuse_iqueue *fiq)
1032 {
1033         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1034                 forget_pending(fiq);
1035 }
1036
1037 /*
1038  * Transfer an interrupt request to userspace
1039  *
1040  * Unlike other requests this is assembled on demand, without a need
1041  * to allocate a separate fuse_req structure.
1042  *
1043  * Called with fiq->lock held, releases it
1044  */
1045 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1046                                struct fuse_copy_state *cs,
1047                                size_t nbytes, struct fuse_req *req)
1048 __releases(fiq->lock)
1049 {
1050         struct fuse_in_header ih;
1051         struct fuse_interrupt_in arg;
1052         unsigned reqsize = sizeof(ih) + sizeof(arg);
1053         int err;
1054
1055         list_del_init(&req->intr_entry);
1056         memset(&ih, 0, sizeof(ih));
1057         memset(&arg, 0, sizeof(arg));
1058         ih.len = reqsize;
1059         ih.opcode = FUSE_INTERRUPT;
1060         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1061         arg.unique = req->in.h.unique;
1062
1063         spin_unlock(&fiq->lock);
1064         if (nbytes < reqsize)
1065                 return -EINVAL;
1066
1067         err = fuse_copy_one(cs, &ih, sizeof(ih));
1068         if (!err)
1069                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1070         fuse_copy_finish(cs);
1071
1072         return err ? err : reqsize;
1073 }
1074
1075 struct fuse_forget_link *fuse_dequeue_forget(struct fuse_iqueue *fiq,
1076                                              unsigned int max,
1077                                              unsigned int *countp)
1078 {
1079         struct fuse_forget_link *head = fiq->forget_list_head.next;
1080         struct fuse_forget_link **newhead = &head;
1081         unsigned count;
1082
1083         for (count = 0; *newhead != NULL && count < max; count++)
1084                 newhead = &(*newhead)->next;
1085
1086         fiq->forget_list_head.next = *newhead;
1087         *newhead = NULL;
1088         if (fiq->forget_list_head.next == NULL)
1089                 fiq->forget_list_tail = &fiq->forget_list_head;
1090
1091         if (countp != NULL)
1092                 *countp = count;
1093
1094         return head;
1095 }
1096 EXPORT_SYMBOL(fuse_dequeue_forget);
1097
1098 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1099                                    struct fuse_copy_state *cs,
1100                                    size_t nbytes)
1101 __releases(fiq->lock)
1102 {
1103         int err;
1104         struct fuse_forget_link *forget = fuse_dequeue_forget(fiq, 1, NULL);
1105         struct fuse_forget_in arg = {
1106                 .nlookup = forget->forget_one.nlookup,
1107         };
1108         struct fuse_in_header ih = {
1109                 .opcode = FUSE_FORGET,
1110                 .nodeid = forget->forget_one.nodeid,
1111                 .unique = fuse_get_unique(fiq),
1112                 .len = sizeof(ih) + sizeof(arg),
1113         };
1114
1115         spin_unlock(&fiq->lock);
1116         kfree(forget);
1117         if (nbytes < ih.len)
1118                 return -EINVAL;
1119
1120         err = fuse_copy_one(cs, &ih, sizeof(ih));
1121         if (!err)
1122                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1123         fuse_copy_finish(cs);
1124
1125         if (err)
1126                 return err;
1127
1128         return ih.len;
1129 }
1130
1131 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1132                                    struct fuse_copy_state *cs, size_t nbytes)
1133 __releases(fiq->lock)
1134 {
1135         int err;
1136         unsigned max_forgets;
1137         unsigned count;
1138         struct fuse_forget_link *head;
1139         struct fuse_batch_forget_in arg = { .count = 0 };
1140         struct fuse_in_header ih = {
1141                 .opcode = FUSE_BATCH_FORGET,
1142                 .unique = fuse_get_unique(fiq),
1143                 .len = sizeof(ih) + sizeof(arg),
1144         };
1145
1146         if (nbytes < ih.len) {
1147                 spin_unlock(&fiq->lock);
1148                 return -EINVAL;
1149         }
1150
1151         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1152         head = fuse_dequeue_forget(fiq, max_forgets, &count);
1153         spin_unlock(&fiq->lock);
1154
1155         arg.count = count;
1156         ih.len += count * sizeof(struct fuse_forget_one);
1157         err = fuse_copy_one(cs, &ih, sizeof(ih));
1158         if (!err)
1159                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1160
1161         while (head) {
1162                 struct fuse_forget_link *forget = head;
1163
1164                 if (!err) {
1165                         err = fuse_copy_one(cs, &forget->forget_one,
1166                                             sizeof(forget->forget_one));
1167                 }
1168                 head = forget->next;
1169                 kfree(forget);
1170         }
1171
1172         fuse_copy_finish(cs);
1173
1174         if (err)
1175                 return err;
1176
1177         return ih.len;
1178 }
1179
1180 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1181                             struct fuse_copy_state *cs,
1182                             size_t nbytes)
1183 __releases(fiq->lock)
1184 {
1185         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1186                 return fuse_read_single_forget(fiq, cs, nbytes);
1187         else
1188                 return fuse_read_batch_forget(fiq, cs, nbytes);
1189 }
1190
1191 /*
1192  * Read a single request into the userspace filesystem's buffer.  This
1193  * function waits until a request is available, then removes it from
1194  * the pending list and copies request data to userspace buffer.  If
1195  * no reply is needed (FORGET) or request has been aborted or there
1196  * was an error during the copying then it's finished by calling
1197  * fuse_request_end().  Otherwise add it to the processing list, and set
1198  * the 'sent' flag.
1199  */
1200 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1201                                 struct fuse_copy_state *cs, size_t nbytes)
1202 {
1203         ssize_t err;
1204         struct fuse_conn *fc = fud->fc;
1205         struct fuse_iqueue *fiq = &fc->iq;
1206         struct fuse_pqueue *fpq = &fud->pq;
1207         struct fuse_req *req;
1208         struct fuse_args *args;
1209         unsigned reqsize;
1210         unsigned int hash;
1211
1212         /*
1213          * Require sane minimum read buffer - that has capacity for fixed part
1214          * of any request header + negotiated max_write room for data.
1215          *
1216          * Historically libfuse reserves 4K for fixed header room, but e.g.
1217          * GlusterFS reserves only 80 bytes
1218          *
1219          *      = `sizeof(fuse_in_header) + sizeof(fuse_write_in)`
1220          *
1221          * which is the absolute minimum any sane filesystem should be using
1222          * for header room.
1223          */
1224         if (nbytes < max_t(size_t, FUSE_MIN_READ_BUFFER,
1225                            sizeof(struct fuse_in_header) +
1226                            sizeof(struct fuse_write_in) +
1227                            fc->max_write))
1228                 return -EINVAL;
1229
1230  restart:
1231         for (;;) {
1232                 spin_lock(&fiq->lock);
1233                 if (!fiq->connected || request_pending(fiq))
1234                         break;
1235                 spin_unlock(&fiq->lock);
1236
1237                 if (file->f_flags & O_NONBLOCK)
1238                         return -EAGAIN;
1239                 err = wait_event_interruptible_exclusive(fiq->waitq,
1240                                 !fiq->connected || request_pending(fiq));
1241                 if (err)
1242                         return err;
1243         }
1244
1245         if (!fiq->connected) {
1246                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1247                 goto err_unlock;
1248         }
1249
1250         if (!list_empty(&fiq->interrupts)) {
1251                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1252                                  intr_entry);
1253                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1254         }
1255
1256         if (forget_pending(fiq)) {
1257                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1258                         return fuse_read_forget(fc, fiq, cs, nbytes);
1259
1260                 if (fiq->forget_batch <= -8)
1261                         fiq->forget_batch = 16;
1262         }
1263
1264         req = list_entry(fiq->pending.next, struct fuse_req, list);
1265         clear_bit(FR_PENDING, &req->flags);
1266         list_del_init(&req->list);
1267         spin_unlock(&fiq->lock);
1268
1269         args = req->args;
1270         reqsize = req->in.h.len;
1271
1272         /* If request is too large, reply with an error and restart the read */
1273         if (nbytes < reqsize) {
1274                 req->out.h.error = -EIO;
1275                 /* SETXATTR is special, since it may contain too large data */
1276                 if (args->opcode == FUSE_SETXATTR)
1277                         req->out.h.error = -E2BIG;
1278                 fuse_request_end(req);
1279                 goto restart;
1280         }
1281         spin_lock(&fpq->lock);
1282         /*
1283          *  Must not put request on fpq->io queue after having been shut down by
1284          *  fuse_abort_conn()
1285          */
1286         if (!fpq->connected) {
1287                 req->out.h.error = err = -ECONNABORTED;
1288                 goto out_end;
1289
1290         }
1291         list_add(&req->list, &fpq->io);
1292         spin_unlock(&fpq->lock);
1293         cs->req = req;
1294         err = fuse_copy_one(cs, &req->in.h, sizeof(req->in.h));
1295         if (!err)
1296                 err = fuse_copy_args(cs, args->in_numargs, args->in_pages,
1297                                      (struct fuse_arg *) args->in_args, 0);
1298         fuse_copy_finish(cs);
1299         spin_lock(&fpq->lock);
1300         clear_bit(FR_LOCKED, &req->flags);
1301         if (!fpq->connected) {
1302                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1303                 goto out_end;
1304         }
1305         if (err) {
1306                 req->out.h.error = -EIO;
1307                 goto out_end;
1308         }
1309         if (!test_bit(FR_ISREPLY, &req->flags)) {
1310                 err = reqsize;
1311                 goto out_end;
1312         }
1313         hash = fuse_req_hash(req->in.h.unique);
1314         list_move_tail(&req->list, &fpq->processing[hash]);
1315         __fuse_get_request(req);
1316         set_bit(FR_SENT, &req->flags);
1317         spin_unlock(&fpq->lock);
1318         /* matches barrier in request_wait_answer() */
1319         smp_mb__after_atomic();
1320         if (test_bit(FR_INTERRUPTED, &req->flags))
1321                 queue_interrupt(req);
1322         fuse_put_request(req);
1323
1324         return reqsize;
1325
1326 out_end:
1327         if (!test_bit(FR_PRIVATE, &req->flags))
1328                 list_del_init(&req->list);
1329         spin_unlock(&fpq->lock);
1330         fuse_request_end(req);
1331         return err;
1332
1333  err_unlock:
1334         spin_unlock(&fiq->lock);
1335         return err;
1336 }
1337
1338 static int fuse_dev_open(struct inode *inode, struct file *file)
1339 {
1340         /*
1341          * The fuse device's file's private_data is used to hold
1342          * the fuse_conn(ection) when it is mounted, and is used to
1343          * keep track of whether the file has been mounted already.
1344          */
1345         file->private_data = NULL;
1346         return 0;
1347 }
1348
1349 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1350 {
1351         struct fuse_copy_state cs;
1352         struct file *file = iocb->ki_filp;
1353         struct fuse_dev *fud = fuse_get_dev(file);
1354
1355         if (!fud)
1356                 return -EPERM;
1357
1358         if (!user_backed_iter(to))
1359                 return -EINVAL;
1360
1361         fuse_copy_init(&cs, 1, to);
1362
1363         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1364 }
1365
1366 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1367                                     struct pipe_inode_info *pipe,
1368                                     size_t len, unsigned int flags)
1369 {
1370         int total, ret;
1371         int page_nr = 0;
1372         struct pipe_buffer *bufs;
1373         struct fuse_copy_state cs;
1374         struct fuse_dev *fud = fuse_get_dev(in);
1375
1376         if (!fud)
1377                 return -EPERM;
1378
1379         bufs = kvmalloc_array(pipe->max_usage, sizeof(struct pipe_buffer),
1380                               GFP_KERNEL);
1381         if (!bufs)
1382                 return -ENOMEM;
1383
1384         fuse_copy_init(&cs, 1, NULL);
1385         cs.pipebufs = bufs;
1386         cs.pipe = pipe;
1387         ret = fuse_dev_do_read(fud, in, &cs, len);
1388         if (ret < 0)
1389                 goto out;
1390
1391         if (pipe_occupancy(pipe->head, pipe->tail) + cs.nr_segs > pipe->max_usage) {
1392                 ret = -EIO;
1393                 goto out;
1394         }
1395
1396         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1397                 /*
1398                  * Need to be careful about this.  Having buf->ops in module
1399                  * code can Oops if the buffer persists after module unload.
1400                  */
1401                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1402                 bufs[page_nr].flags = 0;
1403                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1404                 if (unlikely(ret < 0))
1405                         break;
1406         }
1407         if (total)
1408                 ret = total;
1409 out:
1410         for (; page_nr < cs.nr_segs; page_nr++)
1411                 put_page(bufs[page_nr].page);
1412
1413         kvfree(bufs);
1414         return ret;
1415 }
1416
1417 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1418                             struct fuse_copy_state *cs)
1419 {
1420         struct fuse_notify_poll_wakeup_out outarg;
1421         int err = -EINVAL;
1422
1423         if (size != sizeof(outarg))
1424                 goto err;
1425
1426         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1427         if (err)
1428                 goto err;
1429
1430         fuse_copy_finish(cs);
1431         return fuse_notify_poll_wakeup(fc, &outarg);
1432
1433 err:
1434         fuse_copy_finish(cs);
1435         return err;
1436 }
1437
1438 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1439                                    struct fuse_copy_state *cs)
1440 {
1441         struct fuse_notify_inval_inode_out outarg;
1442         int err = -EINVAL;
1443
1444         if (size != sizeof(outarg))
1445                 goto err;
1446
1447         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1448         if (err)
1449                 goto err;
1450         fuse_copy_finish(cs);
1451
1452         down_read(&fc->killsb);
1453         err = fuse_reverse_inval_inode(fc, outarg.ino,
1454                                        outarg.off, outarg.len);
1455         up_read(&fc->killsb);
1456         return err;
1457
1458 err:
1459         fuse_copy_finish(cs);
1460         return err;
1461 }
1462
1463 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1464                                    struct fuse_copy_state *cs)
1465 {
1466         struct fuse_notify_inval_entry_out outarg;
1467         int err = -ENOMEM;
1468         char *buf;
1469         struct qstr name;
1470
1471         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1472         if (!buf)
1473                 goto err;
1474
1475         err = -EINVAL;
1476         if (size < sizeof(outarg))
1477                 goto err;
1478
1479         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1480         if (err)
1481                 goto err;
1482
1483         err = -ENAMETOOLONG;
1484         if (outarg.namelen > FUSE_NAME_MAX)
1485                 goto err;
1486
1487         err = -EINVAL;
1488         if (size != sizeof(outarg) + outarg.namelen + 1)
1489                 goto err;
1490
1491         name.name = buf;
1492         name.len = outarg.namelen;
1493         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1494         if (err)
1495                 goto err;
1496         fuse_copy_finish(cs);
1497         buf[outarg.namelen] = 0;
1498
1499         down_read(&fc->killsb);
1500         err = fuse_reverse_inval_entry(fc, outarg.parent, 0, &name);
1501         up_read(&fc->killsb);
1502         kfree(buf);
1503         return err;
1504
1505 err:
1506         kfree(buf);
1507         fuse_copy_finish(cs);
1508         return err;
1509 }
1510
1511 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1512                               struct fuse_copy_state *cs)
1513 {
1514         struct fuse_notify_delete_out outarg;
1515         int err = -ENOMEM;
1516         char *buf;
1517         struct qstr name;
1518
1519         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1520         if (!buf)
1521                 goto err;
1522
1523         err = -EINVAL;
1524         if (size < sizeof(outarg))
1525                 goto err;
1526
1527         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1528         if (err)
1529                 goto err;
1530
1531         err = -ENAMETOOLONG;
1532         if (outarg.namelen > FUSE_NAME_MAX)
1533                 goto err;
1534
1535         err = -EINVAL;
1536         if (size != sizeof(outarg) + outarg.namelen + 1)
1537                 goto err;
1538
1539         name.name = buf;
1540         name.len = outarg.namelen;
1541         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1542         if (err)
1543                 goto err;
1544         fuse_copy_finish(cs);
1545         buf[outarg.namelen] = 0;
1546
1547         down_read(&fc->killsb);
1548         err = fuse_reverse_inval_entry(fc, outarg.parent, outarg.child, &name);
1549         up_read(&fc->killsb);
1550         kfree(buf);
1551         return err;
1552
1553 err:
1554         kfree(buf);
1555         fuse_copy_finish(cs);
1556         return err;
1557 }
1558
1559 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1560                              struct fuse_copy_state *cs)
1561 {
1562         struct fuse_notify_store_out outarg;
1563         struct inode *inode;
1564         struct address_space *mapping;
1565         u64 nodeid;
1566         int err;
1567         pgoff_t index;
1568         unsigned int offset;
1569         unsigned int num;
1570         loff_t file_size;
1571         loff_t end;
1572
1573         err = -EINVAL;
1574         if (size < sizeof(outarg))
1575                 goto out_finish;
1576
1577         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1578         if (err)
1579                 goto out_finish;
1580
1581         err = -EINVAL;
1582         if (size - sizeof(outarg) != outarg.size)
1583                 goto out_finish;
1584
1585         nodeid = outarg.nodeid;
1586
1587         down_read(&fc->killsb);
1588
1589         err = -ENOENT;
1590         inode = fuse_ilookup(fc, nodeid,  NULL);
1591         if (!inode)
1592                 goto out_up_killsb;
1593
1594         mapping = inode->i_mapping;
1595         index = outarg.offset >> PAGE_SHIFT;
1596         offset = outarg.offset & ~PAGE_MASK;
1597         file_size = i_size_read(inode);
1598         end = outarg.offset + outarg.size;
1599         if (end > file_size) {
1600                 file_size = end;
1601                 fuse_write_update_attr(inode, file_size, outarg.size);
1602         }
1603
1604         num = outarg.size;
1605         while (num) {
1606                 struct page *page;
1607                 unsigned int this_num;
1608
1609                 err = -ENOMEM;
1610                 page = find_or_create_page(mapping, index,
1611                                            mapping_gfp_mask(mapping));
1612                 if (!page)
1613                         goto out_iput;
1614
1615                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1616                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1617                 if (!err && offset == 0 &&
1618                     (this_num == PAGE_SIZE || file_size == end))
1619                         SetPageUptodate(page);
1620                 unlock_page(page);
1621                 put_page(page);
1622
1623                 if (err)
1624                         goto out_iput;
1625
1626                 num -= this_num;
1627                 offset = 0;
1628                 index++;
1629         }
1630
1631         err = 0;
1632
1633 out_iput:
1634         iput(inode);
1635 out_up_killsb:
1636         up_read(&fc->killsb);
1637 out_finish:
1638         fuse_copy_finish(cs);
1639         return err;
1640 }
1641
1642 struct fuse_retrieve_args {
1643         struct fuse_args_pages ap;
1644         struct fuse_notify_retrieve_in inarg;
1645 };
1646
1647 static void fuse_retrieve_end(struct fuse_mount *fm, struct fuse_args *args,
1648                               int error)
1649 {
1650         struct fuse_retrieve_args *ra =
1651                 container_of(args, typeof(*ra), ap.args);
1652
1653         release_pages(ra->ap.pages, ra->ap.num_pages);
1654         kfree(ra);
1655 }
1656
1657 static int fuse_retrieve(struct fuse_mount *fm, struct inode *inode,
1658                          struct fuse_notify_retrieve_out *outarg)
1659 {
1660         int err;
1661         struct address_space *mapping = inode->i_mapping;
1662         pgoff_t index;
1663         loff_t file_size;
1664         unsigned int num;
1665         unsigned int offset;
1666         size_t total_len = 0;
1667         unsigned int num_pages;
1668         struct fuse_conn *fc = fm->fc;
1669         struct fuse_retrieve_args *ra;
1670         size_t args_size = sizeof(*ra);
1671         struct fuse_args_pages *ap;
1672         struct fuse_args *args;
1673
1674         offset = outarg->offset & ~PAGE_MASK;
1675         file_size = i_size_read(inode);
1676
1677         num = min(outarg->size, fc->max_write);
1678         if (outarg->offset > file_size)
1679                 num = 0;
1680         else if (outarg->offset + num > file_size)
1681                 num = file_size - outarg->offset;
1682
1683         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1684         num_pages = min(num_pages, fc->max_pages);
1685
1686         args_size += num_pages * (sizeof(ap->pages[0]) + sizeof(ap->descs[0]));
1687
1688         ra = kzalloc(args_size, GFP_KERNEL);
1689         if (!ra)
1690                 return -ENOMEM;
1691
1692         ap = &ra->ap;
1693         ap->pages = (void *) (ra + 1);
1694         ap->descs = (void *) (ap->pages + num_pages);
1695
1696         args = &ap->args;
1697         args->nodeid = outarg->nodeid;
1698         args->opcode = FUSE_NOTIFY_REPLY;
1699         args->in_numargs = 2;
1700         args->in_pages = true;
1701         args->end = fuse_retrieve_end;
1702
1703         index = outarg->offset >> PAGE_SHIFT;
1704
1705         while (num && ap->num_pages < num_pages) {
1706                 struct page *page;
1707                 unsigned int this_num;
1708
1709                 page = find_get_page(mapping, index);
1710                 if (!page)
1711                         break;
1712
1713                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1714                 ap->pages[ap->num_pages] = page;
1715                 ap->descs[ap->num_pages].offset = offset;
1716                 ap->descs[ap->num_pages].length = this_num;
1717                 ap->num_pages++;
1718
1719                 offset = 0;
1720                 num -= this_num;
1721                 total_len += this_num;
1722                 index++;
1723         }
1724         ra->inarg.offset = outarg->offset;
1725         ra->inarg.size = total_len;
1726         args->in_args[0].size = sizeof(ra->inarg);
1727         args->in_args[0].value = &ra->inarg;
1728         args->in_args[1].size = total_len;
1729
1730         err = fuse_simple_notify_reply(fm, args, outarg->notify_unique);
1731         if (err)
1732                 fuse_retrieve_end(fm, args, err);
1733
1734         return err;
1735 }
1736
1737 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1738                                 struct fuse_copy_state *cs)
1739 {
1740         struct fuse_notify_retrieve_out outarg;
1741         struct fuse_mount *fm;
1742         struct inode *inode;
1743         u64 nodeid;
1744         int err;
1745
1746         err = -EINVAL;
1747         if (size != sizeof(outarg))
1748                 goto copy_finish;
1749
1750         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1751         if (err)
1752                 goto copy_finish;
1753
1754         fuse_copy_finish(cs);
1755
1756         down_read(&fc->killsb);
1757         err = -ENOENT;
1758         nodeid = outarg.nodeid;
1759
1760         inode = fuse_ilookup(fc, nodeid, &fm);
1761         if (inode) {
1762                 err = fuse_retrieve(fm, inode, &outarg);
1763                 iput(inode);
1764         }
1765         up_read(&fc->killsb);
1766
1767         return err;
1768
1769 copy_finish:
1770         fuse_copy_finish(cs);
1771         return err;
1772 }
1773
1774 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1775                        unsigned int size, struct fuse_copy_state *cs)
1776 {
1777         /* Don't try to move pages (yet) */
1778         cs->move_pages = 0;
1779
1780         switch (code) {
1781         case FUSE_NOTIFY_POLL:
1782                 return fuse_notify_poll(fc, size, cs);
1783
1784         case FUSE_NOTIFY_INVAL_INODE:
1785                 return fuse_notify_inval_inode(fc, size, cs);
1786
1787         case FUSE_NOTIFY_INVAL_ENTRY:
1788                 return fuse_notify_inval_entry(fc, size, cs);
1789
1790         case FUSE_NOTIFY_STORE:
1791                 return fuse_notify_store(fc, size, cs);
1792
1793         case FUSE_NOTIFY_RETRIEVE:
1794                 return fuse_notify_retrieve(fc, size, cs);
1795
1796         case FUSE_NOTIFY_DELETE:
1797                 return fuse_notify_delete(fc, size, cs);
1798
1799         default:
1800                 fuse_copy_finish(cs);
1801                 return -EINVAL;
1802         }
1803 }
1804
1805 /* Look up request on processing list by unique ID */
1806 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1807 {
1808         unsigned int hash = fuse_req_hash(unique);
1809         struct fuse_req *req;
1810
1811         list_for_each_entry(req, &fpq->processing[hash], list) {
1812                 if (req->in.h.unique == unique)
1813                         return req;
1814         }
1815         return NULL;
1816 }
1817
1818 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_args *args,
1819                          unsigned nbytes)
1820 {
1821         unsigned reqsize = sizeof(struct fuse_out_header);
1822
1823         reqsize += fuse_len_args(args->out_numargs, args->out_args);
1824
1825         if (reqsize < nbytes || (reqsize > nbytes && !args->out_argvar))
1826                 return -EINVAL;
1827         else if (reqsize > nbytes) {
1828                 struct fuse_arg *lastarg = &args->out_args[args->out_numargs-1];
1829                 unsigned diffsize = reqsize - nbytes;
1830
1831                 if (diffsize > lastarg->size)
1832                         return -EINVAL;
1833                 lastarg->size -= diffsize;
1834         }
1835         return fuse_copy_args(cs, args->out_numargs, args->out_pages,
1836                               args->out_args, args->page_zeroing);
1837 }
1838
1839 /*
1840  * Write a single reply to a request.  First the header is copied from
1841  * the write buffer.  The request is then searched on the processing
1842  * list by the unique ID found in the header.  If found, then remove
1843  * it from the list and copy the rest of the buffer to the request.
1844  * The request is finished by calling fuse_request_end().
1845  */
1846 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1847                                  struct fuse_copy_state *cs, size_t nbytes)
1848 {
1849         int err;
1850         struct fuse_conn *fc = fud->fc;
1851         struct fuse_pqueue *fpq = &fud->pq;
1852         struct fuse_req *req;
1853         struct fuse_out_header oh;
1854
1855         err = -EINVAL;
1856         if (nbytes < sizeof(struct fuse_out_header))
1857                 goto out;
1858
1859         err = fuse_copy_one(cs, &oh, sizeof(oh));
1860         if (err)
1861                 goto copy_finish;
1862
1863         err = -EINVAL;
1864         if (oh.len != nbytes)
1865                 goto copy_finish;
1866
1867         /*
1868          * Zero oh.unique indicates unsolicited notification message
1869          * and error contains notification code.
1870          */
1871         if (!oh.unique) {
1872                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1873                 goto out;
1874         }
1875
1876         err = -EINVAL;
1877         if (oh.error <= -512 || oh.error > 0)
1878                 goto copy_finish;
1879
1880         spin_lock(&fpq->lock);
1881         req = NULL;
1882         if (fpq->connected)
1883                 req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1884
1885         err = -ENOENT;
1886         if (!req) {
1887                 spin_unlock(&fpq->lock);
1888                 goto copy_finish;
1889         }
1890
1891         /* Is it an interrupt reply ID? */
1892         if (oh.unique & FUSE_INT_REQ_BIT) {
1893                 __fuse_get_request(req);
1894                 spin_unlock(&fpq->lock);
1895
1896                 err = 0;
1897                 if (nbytes != sizeof(struct fuse_out_header))
1898                         err = -EINVAL;
1899                 else if (oh.error == -ENOSYS)
1900                         fc->no_interrupt = 1;
1901                 else if (oh.error == -EAGAIN)
1902                         err = queue_interrupt(req);
1903
1904                 fuse_put_request(req);
1905
1906                 goto copy_finish;
1907         }
1908
1909         clear_bit(FR_SENT, &req->flags);
1910         list_move(&req->list, &fpq->io);
1911         req->out.h = oh;
1912         set_bit(FR_LOCKED, &req->flags);
1913         spin_unlock(&fpq->lock);
1914         cs->req = req;
1915         if (!req->args->page_replace)
1916                 cs->move_pages = 0;
1917
1918         if (oh.error)
1919                 err = nbytes != sizeof(oh) ? -EINVAL : 0;
1920         else
1921                 err = copy_out_args(cs, req->args, nbytes);
1922         fuse_copy_finish(cs);
1923
1924         spin_lock(&fpq->lock);
1925         clear_bit(FR_LOCKED, &req->flags);
1926         if (!fpq->connected)
1927                 err = -ENOENT;
1928         else if (err)
1929                 req->out.h.error = -EIO;
1930         if (!test_bit(FR_PRIVATE, &req->flags))
1931                 list_del_init(&req->list);
1932         spin_unlock(&fpq->lock);
1933
1934         fuse_request_end(req);
1935 out:
1936         return err ? err : nbytes;
1937
1938 copy_finish:
1939         fuse_copy_finish(cs);
1940         goto out;
1941 }
1942
1943 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1944 {
1945         struct fuse_copy_state cs;
1946         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1947
1948         if (!fud)
1949                 return -EPERM;
1950
1951         if (!user_backed_iter(from))
1952                 return -EINVAL;
1953
1954         fuse_copy_init(&cs, 0, from);
1955
1956         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1957 }
1958
1959 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1960                                      struct file *out, loff_t *ppos,
1961                                      size_t len, unsigned int flags)
1962 {
1963         unsigned int head, tail, mask, count;
1964         unsigned nbuf;
1965         unsigned idx;
1966         struct pipe_buffer *bufs;
1967         struct fuse_copy_state cs;
1968         struct fuse_dev *fud;
1969         size_t rem;
1970         ssize_t ret;
1971
1972         fud = fuse_get_dev(out);
1973         if (!fud)
1974                 return -EPERM;
1975
1976         pipe_lock(pipe);
1977
1978         head = pipe->head;
1979         tail = pipe->tail;
1980         mask = pipe->ring_size - 1;
1981         count = head - tail;
1982
1983         bufs = kvmalloc_array(count, sizeof(struct pipe_buffer), GFP_KERNEL);
1984         if (!bufs) {
1985                 pipe_unlock(pipe);
1986                 return -ENOMEM;
1987         }
1988
1989         nbuf = 0;
1990         rem = 0;
1991         for (idx = tail; idx != head && rem < len; idx++)
1992                 rem += pipe->bufs[idx & mask].len;
1993
1994         ret = -EINVAL;
1995         if (rem < len)
1996                 goto out_free;
1997
1998         rem = len;
1999         while (rem) {
2000                 struct pipe_buffer *ibuf;
2001                 struct pipe_buffer *obuf;
2002
2003                 if (WARN_ON(nbuf >= count || tail == head))
2004                         goto out_free;
2005
2006                 ibuf = &pipe->bufs[tail & mask];
2007                 obuf = &bufs[nbuf];
2008
2009                 if (rem >= ibuf->len) {
2010                         *obuf = *ibuf;
2011                         ibuf->ops = NULL;
2012                         tail++;
2013                         pipe->tail = tail;
2014                 } else {
2015                         if (!pipe_buf_get(pipe, ibuf))
2016                                 goto out_free;
2017
2018                         *obuf = *ibuf;
2019                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2020                         obuf->len = rem;
2021                         ibuf->offset += obuf->len;
2022                         ibuf->len -= obuf->len;
2023                 }
2024                 nbuf++;
2025                 rem -= obuf->len;
2026         }
2027         pipe_unlock(pipe);
2028
2029         fuse_copy_init(&cs, 0, NULL);
2030         cs.pipebufs = bufs;
2031         cs.nr_segs = nbuf;
2032         cs.pipe = pipe;
2033
2034         if (flags & SPLICE_F_MOVE)
2035                 cs.move_pages = 1;
2036
2037         ret = fuse_dev_do_write(fud, &cs, len);
2038
2039         pipe_lock(pipe);
2040 out_free:
2041         for (idx = 0; idx < nbuf; idx++) {
2042                 struct pipe_buffer *buf = &bufs[idx];
2043
2044                 if (buf->ops)
2045                         pipe_buf_release(pipe, buf);
2046         }
2047         pipe_unlock(pipe);
2048
2049         kvfree(bufs);
2050         return ret;
2051 }
2052
2053 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2054 {
2055         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2056         struct fuse_iqueue *fiq;
2057         struct fuse_dev *fud = fuse_get_dev(file);
2058
2059         if (!fud)
2060                 return EPOLLERR;
2061
2062         fiq = &fud->fc->iq;
2063         poll_wait(file, &fiq->waitq, wait);
2064
2065         spin_lock(&fiq->lock);
2066         if (!fiq->connected)
2067                 mask = EPOLLERR;
2068         else if (request_pending(fiq))
2069                 mask |= EPOLLIN | EPOLLRDNORM;
2070         spin_unlock(&fiq->lock);
2071
2072         return mask;
2073 }
2074
2075 /* Abort all requests on the given list (pending or processing) */
2076 static void end_requests(struct list_head *head)
2077 {
2078         while (!list_empty(head)) {
2079                 struct fuse_req *req;
2080                 req = list_entry(head->next, struct fuse_req, list);
2081                 req->out.h.error = -ECONNABORTED;
2082                 clear_bit(FR_SENT, &req->flags);
2083                 list_del_init(&req->list);
2084                 fuse_request_end(req);
2085         }
2086 }
2087
2088 static void end_polls(struct fuse_conn *fc)
2089 {
2090         struct rb_node *p;
2091
2092         p = rb_first(&fc->polled_files);
2093
2094         while (p) {
2095                 struct fuse_file *ff;
2096                 ff = rb_entry(p, struct fuse_file, polled_node);
2097                 wake_up_interruptible_all(&ff->poll_wait);
2098
2099                 p = rb_next(p);
2100         }
2101 }
2102
2103 /*
2104  * Abort all requests.
2105  *
2106  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2107  * filesystem.
2108  *
2109  * The same effect is usually achievable through killing the filesystem daemon
2110  * and all users of the filesystem.  The exception is the combination of an
2111  * asynchronous request and the tricky deadlock (see
2112  * Documentation/filesystems/fuse.rst).
2113  *
2114  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2115  * requests, they should be finished off immediately.  Locked requests will be
2116  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2117  * requests.  It is possible that some request will finish before we can.  This
2118  * is OK, the request will in that case be removed from the list before we touch
2119  * it.
2120  */
2121 void fuse_abort_conn(struct fuse_conn *fc)
2122 {
2123         struct fuse_iqueue *fiq = &fc->iq;
2124
2125         spin_lock(&fc->lock);
2126         if (fc->connected) {
2127                 struct fuse_dev *fud;
2128                 struct fuse_req *req, *next;
2129                 LIST_HEAD(to_end);
2130                 unsigned int i;
2131
2132                 /* Background queuing checks fc->connected under bg_lock */
2133                 spin_lock(&fc->bg_lock);
2134                 fc->connected = 0;
2135                 spin_unlock(&fc->bg_lock);
2136
2137                 fuse_set_initialized(fc);
2138                 list_for_each_entry(fud, &fc->devices, entry) {
2139                         struct fuse_pqueue *fpq = &fud->pq;
2140
2141                         spin_lock(&fpq->lock);
2142                         fpq->connected = 0;
2143                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2144                                 req->out.h.error = -ECONNABORTED;
2145                                 spin_lock(&req->waitq.lock);
2146                                 set_bit(FR_ABORTED, &req->flags);
2147                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2148                                         set_bit(FR_PRIVATE, &req->flags);
2149                                         __fuse_get_request(req);
2150                                         list_move(&req->list, &to_end);
2151                                 }
2152                                 spin_unlock(&req->waitq.lock);
2153                         }
2154                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2155                                 list_splice_tail_init(&fpq->processing[i],
2156                                                       &to_end);
2157                         spin_unlock(&fpq->lock);
2158                 }
2159                 spin_lock(&fc->bg_lock);
2160                 fc->blocked = 0;
2161                 fc->max_background = UINT_MAX;
2162                 flush_bg_queue(fc);
2163                 spin_unlock(&fc->bg_lock);
2164
2165                 spin_lock(&fiq->lock);
2166                 fiq->connected = 0;
2167                 list_for_each_entry(req, &fiq->pending, list)
2168                         clear_bit(FR_PENDING, &req->flags);
2169                 list_splice_tail_init(&fiq->pending, &to_end);
2170                 while (forget_pending(fiq))
2171                         kfree(fuse_dequeue_forget(fiq, 1, NULL));
2172                 wake_up_all(&fiq->waitq);
2173                 spin_unlock(&fiq->lock);
2174                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2175                 end_polls(fc);
2176                 wake_up_all(&fc->blocked_waitq);
2177                 spin_unlock(&fc->lock);
2178
2179                 end_requests(&to_end);
2180         } else {
2181                 spin_unlock(&fc->lock);
2182         }
2183 }
2184 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2185
2186 void fuse_wait_aborted(struct fuse_conn *fc)
2187 {
2188         /* matches implicit memory barrier in fuse_drop_waiting() */
2189         smp_mb();
2190         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2191 }
2192
2193 int fuse_dev_release(struct inode *inode, struct file *file)
2194 {
2195         struct fuse_dev *fud = fuse_get_dev(file);
2196
2197         if (fud) {
2198                 struct fuse_conn *fc = fud->fc;
2199                 struct fuse_pqueue *fpq = &fud->pq;
2200                 LIST_HEAD(to_end);
2201                 unsigned int i;
2202
2203                 spin_lock(&fpq->lock);
2204                 WARN_ON(!list_empty(&fpq->io));
2205                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2206                         list_splice_init(&fpq->processing[i], &to_end);
2207                 spin_unlock(&fpq->lock);
2208
2209                 end_requests(&to_end);
2210
2211                 /* Are we the last open device? */
2212                 if (atomic_dec_and_test(&fc->dev_count)) {
2213                         WARN_ON(fc->iq.fasync != NULL);
2214                         fuse_abort_conn(fc);
2215                 }
2216                 fuse_dev_free(fud);
2217         }
2218         return 0;
2219 }
2220 EXPORT_SYMBOL_GPL(fuse_dev_release);
2221
2222 static int fuse_dev_fasync(int fd, struct file *file, int on)
2223 {
2224         struct fuse_dev *fud = fuse_get_dev(file);
2225
2226         if (!fud)
2227                 return -EPERM;
2228
2229         /* No locking - fasync_helper does its own locking */
2230         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2231 }
2232
2233 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2234 {
2235         struct fuse_dev *fud;
2236
2237         if (new->private_data)
2238                 return -EINVAL;
2239
2240         fud = fuse_dev_alloc_install(fc);
2241         if (!fud)
2242                 return -ENOMEM;
2243
2244         new->private_data = fud;
2245         atomic_inc(&fc->dev_count);
2246
2247         return 0;
2248 }
2249
2250 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2251                            unsigned long arg)
2252 {
2253         int res;
2254         int oldfd;
2255         struct fuse_dev *fud = NULL;
2256
2257         switch (cmd) {
2258         case FUSE_DEV_IOC_CLONE:
2259                 res = -EFAULT;
2260                 if (!get_user(oldfd, (__u32 __user *)arg)) {
2261                         struct file *old = fget(oldfd);
2262
2263                         res = -EINVAL;
2264                         if (old) {
2265                                 /*
2266                                  * Check against file->f_op because CUSE
2267                                  * uses the same ioctl handler.
2268                                  */
2269                                 if (old->f_op == file->f_op &&
2270                                     old->f_cred->user_ns == file->f_cred->user_ns)
2271                                         fud = fuse_get_dev(old);
2272
2273                                 if (fud) {
2274                                         mutex_lock(&fuse_mutex);
2275                                         res = fuse_device_clone(fud->fc, file);
2276                                         mutex_unlock(&fuse_mutex);
2277                                 }
2278                                 fput(old);
2279                         }
2280                 }
2281                 break;
2282         default:
2283                 res = -ENOTTY;
2284                 break;
2285         }
2286         return res;
2287 }
2288
2289 const struct file_operations fuse_dev_operations = {
2290         .owner          = THIS_MODULE,
2291         .open           = fuse_dev_open,
2292         .llseek         = no_llseek,
2293         .read_iter      = fuse_dev_read,
2294         .splice_read    = fuse_dev_splice_read,
2295         .write_iter     = fuse_dev_write,
2296         .splice_write   = fuse_dev_splice_write,
2297         .poll           = fuse_dev_poll,
2298         .release        = fuse_dev_release,
2299         .fasync         = fuse_dev_fasync,
2300         .unlocked_ioctl = fuse_dev_ioctl,
2301         .compat_ioctl   = compat_ptr_ioctl,
2302 };
2303 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2304
2305 static struct miscdevice fuse_miscdevice = {
2306         .minor = FUSE_MINOR,
2307         .name  = "fuse",
2308         .fops = &fuse_dev_operations,
2309 };
2310
2311 int __init fuse_dev_init(void)
2312 {
2313         int err = -ENOMEM;
2314         fuse_req_cachep = kmem_cache_create("fuse_request",
2315                                             sizeof(struct fuse_req),
2316                                             0, 0, NULL);
2317         if (!fuse_req_cachep)
2318                 goto out;
2319
2320         err = misc_register(&fuse_miscdevice);
2321         if (err)
2322                 goto out_cache_clean;
2323
2324         return 0;
2325
2326  out_cache_clean:
2327         kmem_cache_destroy(fuse_req_cachep);
2328  out:
2329         return err;
2330 }
2331
2332 void fuse_dev_cleanup(void)
2333 {
2334         misc_deregister(&fuse_miscdevice);
2335         kmem_cache_destroy(fuse_req_cachep);
2336 }