fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/sched/signal.h>
  15 #include <linux/uio.h>
  16 #include <linux/miscdevice.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/pipe_fs_i.h>
  21 #include <linux/swap.h>
  22 #include <linux/splice.h>
  23 #include <linux/sched.h>
  24
  25 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  26 MODULE_ALIAS("devname:fuse");
  27
  28 /* Ordinary requests have even IDs, while interrupts IDs are odd */
  29 #define FUSE_INT_REQ_BIT (1ULL << 0)
  30 #define FUSE_REQ_ID_STEP (1ULL << 1)
  31
  32 static struct kmem_cache *fuse_req_cachep;
  33
  34 static struct fuse_dev *fuse_get_dev(struct file *file)
  35 {
  36         /*
  37          * Lockless access is OK, because file->private data is set
  38          * once during mount and is valid until the file is released.
  39          */
  40         return READ_ONCE(file->private_data);
  41 }
  42
  43 static void fuse_request_init(struct fuse_req *req)
  44 {
  45         INIT_LIST_HEAD(&req->list);
  46         INIT_LIST_HEAD(&req->intr_entry);
  47         init_waitqueue_head(&req->waitq);
  48         refcount_set(&req->count, 1);
  49         __set_bit(FR_PENDING, &req->flags);
  50 }
  51
  52 static struct fuse_req *fuse_request_alloc(gfp_t flags)
  53 {
  54         struct fuse_req *req = kmem_cache_zalloc(fuse_req_cachep, flags);
  55         if (req)
  56                 fuse_request_init(req);
  57
  58         return req;
  59 }
  60
  61 static void fuse_request_free(struct fuse_req *req)
  62 {
  63         kmem_cache_free(fuse_req_cachep, req);
  64 }
  65
  66 static void __fuse_get_request(struct fuse_req *req)
  67 {
  68         refcount_inc(&req->count);
  69 }
  70
  71 /* Must be called with > 1 refcount */
  72 static void __fuse_put_request(struct fuse_req *req)
  73 {
  74         refcount_dec(&req->count);
  75 }
  76
  77 void fuse_set_initialized(struct fuse_conn *fc)
  78 {
  79         /* Make sure stores before this are seen on another CPU */
  80         smp_wmb();
  81         fc->initialized = 1;
  82 }
  83
  84 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
  85 {
  86         return !fc->initialized || (for_background && fc->blocked);
  87 }
  88
  89 static void fuse_drop_waiting(struct fuse_conn *fc)
  90 {
  91         /*
  92          * lockess check of fc->connected is okay, because atomic_dec_and_test()
  93          * provides a memory barrier mached with the one in fuse_wait_aborted()
  94          * to ensure no wake-up is missed.
  95          */
  96         if (atomic_dec_and_test(&fc->num_waiting) &&
  97             !READ_ONCE(fc->connected)) {
  98                 /* wake up aborters */
  99                 wake_up_all(&fc->blocked_waitq);
 100         }
 101 }
 102
 103 static void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req);
 104
 105 static struct fuse_req *fuse_get_req(struct fuse_conn *fc, bool for_background)
 106 {
 107         struct fuse_req *req;
 108         int err;
 109         atomic_inc(&fc->num_waiting);
 110
 111         if (fuse_block_alloc(fc, for_background)) {
 112                 err = -EINTR;
 113                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 114                                 !fuse_block_alloc(fc, for_background)))
 115                         goto out;
 116         }
 117         /* Matches smp_wmb() in fuse_set_initialized() */
 118         smp_rmb();
 119
 120         err = -ENOTCONN;
 121         if (!fc->connected)
 122                 goto out;
 123
 124         err = -ECONNREFUSED;
 125         if (fc->conn_error)
 126                 goto out;
 127
 128         req = fuse_request_alloc(GFP_KERNEL);
 129         err = -ENOMEM;
 130         if (!req) {
 131                 if (for_background)
 132                         wake_up(&fc->blocked_waitq);
 133                 goto out;
 134         }
 135
 136         req->in.h.uid = from_kuid(fc->user_ns, current_fsuid());
 137         req->in.h.gid = from_kgid(fc->user_ns, current_fsgid());
 138         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 139
 140         __set_bit(FR_WAITING, &req->flags);
 141         if (for_background)
 142                 __set_bit(FR_BACKGROUND, &req->flags);
 143
 144         if (unlikely(req->in.h.uid == ((uid_t)-1) ||
 145                      req->in.h.gid == ((gid_t)-1))) {
 146                 fuse_put_request(fc, req);
 147                 return ERR_PTR(-EOVERFLOW);
 148         }
 149         return req;
 150
 151  out:
 152         fuse_drop_waiting(fc);
 153         return ERR_PTR(err);
 154 }
 155
 156 static void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 157 {
 158         if (refcount_dec_and_test(&req->count)) {
 159                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 160                         /*
 161                          * We get here in the unlikely case that a background
 162                          * request was allocated but not sent
 163                          */
 164                         spin_lock(&fc->bg_lock);
 165                         if (!fc->blocked)
 166                                 wake_up(&fc->blocked_waitq);
 167                         spin_unlock(&fc->bg_lock);
 168                 }
 169
 170                 if (test_bit(FR_WAITING, &req->flags)) {
 171                         __clear_bit(FR_WAITING, &req->flags);
 172                         fuse_drop_waiting(fc);
 173                 }
 174
 175                 fuse_request_free(req);
 176         }
 177 }
 178
 179 unsigned int fuse_len_args(unsigned int numargs, struct fuse_arg *args)
 180 {
 181         unsigned nbytes = 0;
 182         unsigned i;
 183
 184         for (i = 0; i < numargs; i++)
 185                 nbytes += args[i].size;
 186
 187         return nbytes;
 188 }
 189 EXPORT_SYMBOL_GPL(fuse_len_args);
 190
 191 u64 fuse_get_unique(struct fuse_iqueue *fiq)
 192 {
 193         fiq->reqctr += FUSE_REQ_ID_STEP;
 194         return fiq->reqctr;
 195 }
 196 EXPORT_SYMBOL_GPL(fuse_get_unique);
 197
 198 static unsigned int fuse_req_hash(u64 unique)
 199 {
 200         return hash_long(unique & ~FUSE_INT_REQ_BIT, FUSE_PQ_HASH_BITS);
 201 }
 202
 203 /**
 204  * A new request is available, wake fiq->waitq
 205  */
 206 static void fuse_dev_wake_and_unlock(struct fuse_iqueue *fiq)
 207 __releases(fiq->lock)
 208 {
 209         wake_up(&fiq->waitq);
 210         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 211         spin_unlock(&fiq->lock);
 212 }
 213
 214 const struct fuse_iqueue_ops fuse_dev_fiq_ops = {
 215         .wake_forget_and_unlock         = fuse_dev_wake_and_unlock,
 216         .wake_interrupt_and_unlock      = fuse_dev_wake_and_unlock,
 217         .wake_pending_and_unlock        = fuse_dev_wake_and_unlock,
 218 };
 219 EXPORT_SYMBOL_GPL(fuse_dev_fiq_ops);
 220
 221 static void queue_request_and_unlock(struct fuse_iqueue *fiq,
 222                                      struct fuse_req *req)
 223 __releases(fiq->lock)
 224 {
 225         req->in.h.len = sizeof(struct fuse_in_header) +
 226                 fuse_len_args(req->args->in_numargs,
 227                               (struct fuse_arg *) req->args->in_args);
 228         list_add_tail(&req->list, &fiq->pending);
 229         fiq->ops->wake_pending_and_unlock(fiq);
 230 }
 231
 232 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 233                        u64 nodeid, u64 nlookup)
 234 {
 235         struct fuse_iqueue *fiq = &fc->iq;
 236
 237         forget->forget_one.nodeid = nodeid;
 238         forget->forget_one.nlookup = nlookup;
 239
 240         spin_lock(&fiq->lock);
 241         if (fiq->connected) {
 242                 fiq->forget_list_tail->next = forget;
 243                 fiq->forget_list_tail = forget;
 244                 fiq->ops->wake_forget_and_unlock(fiq);
 245         } else {
 246                 kfree(forget);
 247                 spin_unlock(&fiq->lock);
 248         }
 249 }
 250
 251 static void flush_bg_queue(struct fuse_conn *fc)
 252 {
 253         struct fuse_iqueue *fiq = &fc->iq;
 254
 255         while (fc->active_background < fc->max_background &&
 256                !list_empty(&fc->bg_queue)) {
 257                 struct fuse_req *req;
 258
 259                 req = list_first_entry(&fc->bg_queue, struct fuse_req, list);
 260                 list_del(&req->list);
 261                 fc->active_background++;
 262                 spin_lock(&fiq->lock);
 263                 req->in.h.unique = fuse_get_unique(fiq);
 264                 queue_request_and_unlock(fiq, req);
 265         }
 266 }
 267
 268 /*
 269  * This function is called when a request is finished.  Either a reply
 270  * has arrived or it was aborted (and not yet sent) or some error
 271  * occurred during communication with userspace, or the device file
 272  * was closed.  The requester thread is woken up (if still waiting),
 273  * the 'end' callback is called if given, else the reference to the
 274  * request is released
 275  */
 276 void fuse_request_end(struct fuse_conn *fc, struct fuse_req *req)
 277 {
 278         struct fuse_iqueue *fiq = &fc->iq;
 279
 280         if (test_and_set_bit(FR_FINISHED, &req->flags))
 281                 goto put_request;
 282
 283         /*
 284          * test_and_set_bit() implies smp_mb() between bit
 285          * changing and below FR_INTERRUPTED check. Pairs with
 286          * smp_mb() from queue_interrupt().
 287          */
 288         if (test_bit(FR_INTERRUPTED, &req->flags)) {
 289                 spin_lock(&fiq->lock);
 290                 list_del_init(&req->intr_entry);
 291                 spin_unlock(&fiq->lock);
 292         }
 293         WARN_ON(test_bit(FR_PENDING, &req->flags));
 294         WARN_ON(test_bit(FR_SENT, &req->flags));
 295         if (test_bit(FR_BACKGROUND, &req->flags)) {
 296                 spin_lock(&fc->bg_lock);
 297                 clear_bit(FR_BACKGROUND, &req->flags);
 298                 if (fc->num_background == fc->max_background) {
 299                         fc->blocked = 0;
 300                         wake_up(&fc->blocked_waitq);
 301                 } else if (!fc->blocked) {
 302                         /*
 303                          * Wake up next waiter, if any.  It's okay to use
 304                          * waitqueue_active(), as we've already synced up
 305                          * fc->blocked with waiters with the wake_up() call
 306                          * above.
 307                          */
 308                         if (waitqueue_active(&fc->blocked_waitq))
 309                                 wake_up(&fc->blocked_waitq);
 310                 }
 311
 312                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 313                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 314                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 315                 }
 316                 fc->num_background--;
 317                 fc->active_background--;
 318                 flush_bg_queue(fc);
 319                 spin_unlock(&fc->bg_lock);
 320         } else {
 321                 /* Wake up waiter sleeping in request_wait_answer() */
 322                 wake_up(&req->waitq);
 323         }
 324
 325         if (test_bit(FR_ASYNC, &req->flags))
 326                 req->args->end(fc, req->args, req->out.h.error);
 327 put_request:
 328         fuse_put_request(fc, req);
 329 }
 330 EXPORT_SYMBOL_GPL(fuse_request_end);
 331
 332 static int queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 333 {
 334         spin_lock(&fiq->lock);
 335         /* Check for we've sent request to interrupt this req */
 336         if (unlikely(!test_bit(FR_INTERRUPTED, &req->flags))) {
 337                 spin_unlock(&fiq->lock);
 338                 return -EINVAL;
 339         }
 340
 341         if (list_empty(&req->intr_entry)) {
 342                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 343                 /*
 344                  * Pairs with smp_mb() implied by test_and_set_bit()
 345                  * from request_end().
 346                  */
 347                 smp_mb();
 348                 if (test_bit(FR_FINISHED, &req->flags)) {
 349                         list_del_init(&req->intr_entry);
 350                         spin_unlock(&fiq->lock);
 351                         return 0;
 352                 }
 353                 fiq->ops->wake_interrupt_and_unlock(fiq);
 354         } else {
 355                 spin_unlock(&fiq->lock);
 356         }
 357         return 0;
 358 }
 359
 360 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 361 {
 362         struct fuse_iqueue *fiq = &fc->iq;
 363         int err;
 364
 365         if (!fc->no_interrupt) {
 366                 /* Any signal may interrupt this */
 367                 err = wait_event_interruptible(req->waitq,
 368                                         test_bit(FR_FINISHED, &req->flags));
 369                 if (!err)
 370                         return;
 371
 372                 set_bit(FR_INTERRUPTED, &req->flags);
 373                 /* matches barrier in fuse_dev_do_read() */
 374                 smp_mb__after_atomic();
 375                 if (test_bit(FR_SENT, &req->flags))
 376                         queue_interrupt(fiq, req);
 377         }
 378
 379         if (!test_bit(FR_FORCE, &req->flags)) {
 380                 /* Only fatal signals may interrupt this */
 381                 err = wait_event_killable(req->waitq,
 382                                         test_bit(FR_FINISHED, &req->flags));
 383                 if (!err)
 384                         return;
 385
 386                 spin_lock(&fiq->lock);
 387                 /* Request is not yet in userspace, bail out */
 388                 if (test_bit(FR_PENDING, &req->flags)) {
 389                         list_del(&req->list);
 390                         spin_unlock(&fiq->lock);
 391                         __fuse_put_request(req);
 392                         req->out.h.error = -EINTR;
 393                         return;
 394                 }
 395                 spin_unlock(&fiq->lock);
 396         }
 397
 398         /*
 399          * Either request is already in userspace, or it was forced.
 400          * Wait it out.
 401          */
 402         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 403 }
 404
 405 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 406 {
 407         struct fuse_iqueue *fiq = &fc->iq;
 408
 409         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 410         spin_lock(&fiq->lock);
 411         if (!fiq->connected) {
 412                 spin_unlock(&fiq->lock);
 413                 req->out.h.error = -ENOTCONN;
 414         } else {
 415                 req->in.h.unique = fuse_get_unique(fiq);
 416                 /* acquire extra reference, since request is still needed
 417                    after fuse_request_end() */
 418                 __fuse_get_request(req);
 419                 queue_request_and_unlock(fiq, req);
 420
 421                 request_wait_answer(fc, req);
 422                 /* Pairs with smp_wmb() in fuse_request_end() */
 423                 smp_rmb();
 424         }
 425 }
 426
 427 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 428 {
 429         if (fc->minor < 4 && args->opcode == FUSE_STATFS)
 430                 args->out_args[0].size = FUSE_COMPAT_STATFS_SIZE;
 431
 432         if (fc->minor < 9) {
 433                 switch (args->opcode) {
 434                 case FUSE_LOOKUP:
 435                 case FUSE_CREATE:
 436                 case FUSE_MKNOD:
 437                 case FUSE_MKDIR:
 438                 case FUSE_SYMLINK:
 439                 case FUSE_LINK:
 440                         args->out_args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 441                         break;
 442                 case FUSE_GETATTR:
 443                 case FUSE_SETATTR:
 444                         args->out_args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 445                         break;
 446                 }
 447         }
 448         if (fc->minor < 12) {
 449                 switch (args->opcode) {
 450                 case FUSE_CREATE:
 451                         args->in_args[0].size = sizeof(struct fuse_open_in);
 452                         break;
 453                 case FUSE_MKNOD:
 454                         args->in_args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 455                         break;
 456                 }
 457         }
 458 }
 459
 460 static void fuse_force_creds(struct fuse_conn *fc, struct fuse_req *req)
 461 {
 462         req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
 463         req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
 464         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 465 }
 466
 467 static void fuse_args_to_req(struct fuse_req *req, struct fuse_args *args)
 468 {
 469         req->in.h.opcode = args->opcode;
 470         req->in.h.nodeid = args->nodeid;
 471         req->args = args;
 472         if (args->end)
 473                 __set_bit(FR_ASYNC, &req->flags);
 474 }
 475
 476 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 477 {
 478         struct fuse_req *req;
 479         ssize_t ret;
 480
 481         if (args->force) {
 482                 atomic_inc(&fc->num_waiting);
 483                 req = fuse_request_alloc(GFP_KERNEL | __GFP_NOFAIL);
 484
 485                 if (!args->nocreds)
 486                         fuse_force_creds(fc, req);
 487
 488                 __set_bit(FR_WAITING, &req->flags);
 489                 __set_bit(FR_FORCE, &req->flags);
 490         } else {
 491                 WARN_ON(args->nocreds);
 492                 req = fuse_get_req(fc, false);
 493                 if (IS_ERR(req))
 494                         return PTR_ERR(req);
 495         }
 496
 497         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 498         fuse_adjust_compat(fc, args);
 499         fuse_args_to_req(req, args);
 500
 501         if (!args->noreply)
 502                 __set_bit(FR_ISREPLY, &req->flags);
 503         __fuse_request_send(fc, req);
 504         ret = req->out.h.error;
 505         if (!ret && args->out_argvar) {
 506                 BUG_ON(args->out_numargs == 0);
 507                 ret = args->out_args[args->out_numargs - 1].size;
 508         }
 509         fuse_put_request(fc, req);
 510
 511         return ret;
 512 }
 513
 514 static bool fuse_request_queue_background(struct fuse_conn *fc,
 515                                           struct fuse_req *req)
 516 {
 517         bool queued = false;
 518
 519         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 520         if (!test_bit(FR_WAITING, &req->flags)) {
 521                 __set_bit(FR_WAITING, &req->flags);
 522                 atomic_inc(&fc->num_waiting);
 523         }
 524         __set_bit(FR_ISREPLY, &req->flags);
 525         spin_lock(&fc->bg_lock);
 526         if (likely(fc->connected)) {
 527                 fc->num_background++;
 528                 if (fc->num_background == fc->max_background)
 529                         fc->blocked = 1;
 530                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 531                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 532                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 533                 }
 534                 list_add_tail(&req->list, &fc->bg_queue);
 535                 flush_bg_queue(fc);
 536                 queued = true;
 537         }
 538         spin_unlock(&fc->bg_lock);
 539
 540         return queued;
 541 }
 542
 543 int fuse_simple_background(struct fuse_conn *fc, struct fuse_args *args,
 544                             gfp_t gfp_flags)
 545 {
 546         struct fuse_req *req;
 547
 548         if (args->force) {
 549                 WARN_ON(!args->nocreds);
 550                 req = fuse_request_alloc(gfp_flags);
 551                 if (!req)
 552                         return -ENOMEM;
 553                 __set_bit(FR_BACKGROUND, &req->flags);
 554         } else {
 555                 WARN_ON(args->nocreds);
 556                 req = fuse_get_req(fc, true);
 557                 if (IS_ERR(req))
 558                         return PTR_ERR(req);
 559         }
 560
 561         fuse_args_to_req(req, args);
 562
 563         if (!fuse_request_queue_background(fc, req)) {
 564                 fuse_put_request(fc, req);
 565                 return -ENOTCONN;
 566         }
 567
 568         return 0;
 569 }
 570 EXPORT_SYMBOL_GPL(fuse_simple_background);
 571
 572 static int fuse_simple_notify_reply(struct fuse_conn *fc,
 573                                     struct fuse_args *args, u64 unique)
 574 {
 575         struct fuse_req *req;
 576         struct fuse_iqueue *fiq = &fc->iq;
 577         int err = 0;
 578
 579         req = fuse_get_req(fc, false);
 580         if (IS_ERR(req))
 581                 return PTR_ERR(req);
 582
 583         __clear_bit(FR_ISREPLY, &req->flags);
 584         req->in.h.unique = unique;
 585
 586         fuse_args_to_req(req, args);
 587
 588         spin_lock(&fiq->lock);
 589         if (fiq->connected) {
 590                 queue_request_and_unlock(fiq, req);
 591         } else {
 592                 err = -ENODEV;
 593                 spin_unlock(&fiq->lock);
 594                 fuse_put_request(fc, req);
 595         }
 596
 597         return err;
 598 }
 599
 600 /*
 601  * Lock the request.  Up to the next unlock_request() there mustn't be
 602  * anything that could cause a page-fault.  If the request was already
 603  * aborted bail out.
 604  */
 605 static int lock_request(struct fuse_req *req)
 606 {
 607         int err = 0;
 608         if (req) {
 609                 spin_lock(&req->waitq.lock);
 610                 if (test_bit(FR_ABORTED, &req->flags))
 611                         err = -ENOENT;
 612                 else
 613                         set_bit(FR_LOCKED, &req->flags);
 614                 spin_unlock(&req->waitq.lock);
 615         }
 616         return err;
 617 }
 618
 619 /*
 620  * Unlock request.  If it was aborted while locked, caller is responsible
 621  * for unlocking and ending the request.
 622  */
 623 static int unlock_request(struct fuse_req *req)
 624 {
 625         int err = 0;
 626         if (req) {
 627                 spin_lock(&req->waitq.lock);
 628                 if (test_bit(FR_ABORTED, &req->flags))
 629                         err = -ENOENT;
 630                 else
 631                         clear_bit(FR_LOCKED, &req->flags);
 632                 spin_unlock(&req->waitq.lock);
 633         }
 634         return err;
 635 }
 636
 637 struct fuse_copy_state {
 638         int write;
 639         struct fuse_req *req;
 640         struct iov_iter *iter;
 641         struct pipe_buffer *pipebufs;
 642         struct pipe_buffer *currbuf;
 643         struct pipe_inode_info *pipe;
 644         unsigned long nr_segs;
 645         struct page *pg;
 646         unsigned len;
 647         unsigned offset;
 648         unsigned move_pages:1;
 649 };
 650
 651 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 652                            struct iov_iter *iter)
 653 {
 654         memset(cs, 0, sizeof(*cs));
 655         cs->write = write;
 656         cs->iter = iter;
 657 }
 658
 659 /* Unmap and put previous page of userspace buffer */
 660 static void fuse_copy_finish(struct fuse_copy_state *cs)
 661 {
 662         if (cs->currbuf) {
 663                 struct pipe_buffer *buf = cs->currbuf;
 664
 665                 if (cs->write)
 666                         buf->len = PAGE_SIZE - cs->len;
 667                 cs->currbuf = NULL;
 668         } else if (cs->pg) {
 669                 if (cs->write) {
 670                         flush_dcache_page(cs->pg);
 671                         set_page_dirty_lock(cs->pg);
 672                 }
 673                 put_page(cs->pg);
 674         }
 675         cs->pg = NULL;
 676 }
 677
 678 /*
 679  * Get another pagefull of userspace buffer, and map it to kernel
 680  * address space, and lock request
 681  */
 682 static int fuse_copy_fill(struct fuse_copy_state *cs)
 683 {
 684         struct page *page;
 685         int err;
 686
 687         err = unlock_request(cs->req);
 688         if (err)
 689                 return err;
 690
 691         fuse_copy_finish(cs);
 692         if (cs->pipebufs) {
 693                 struct pipe_buffer *buf = cs->pipebufs;
 694
 695                 if (!cs->write) {
 696                         err = pipe_buf_confirm(cs->pipe, buf);
 697                         if (err)
 698                                 return err;
 699
 700                         BUG_ON(!cs->nr_segs);
 701                         cs->currbuf = buf;
 702                         cs->pg = buf->page;
 703                         cs->offset = buf->offset;
 704                         cs->len = buf->len;
 705                         cs->pipebufs++;
 706                         cs->nr_segs--;
 707                 } else {
 708                         if (cs->nr_segs == cs->pipe->buffers)
 709                                 return -EIO;
 710
 711                         page = alloc_page(GFP_HIGHUSER);
 712                         if (!page)
 713                                 return -ENOMEM;
 714
 715                         buf->page = page;
 716                         buf->offset = 0;
 717                         buf->len = 0;
 718
 719                         cs->currbuf = buf;
 720                         cs->pg = page;
 721                         cs->offset = 0;
 722                         cs->len = PAGE_SIZE;
 723                         cs->pipebufs++;
 724                         cs->nr_segs++;
 725                 }
 726         } else {
 727                 size_t off;
 728                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 729                 if (err < 0)
 730                         return err;
 731                 BUG_ON(!err);
 732                 cs->len = err;
 733                 cs->offset = off;
 734                 cs->pg = page;
 735                 iov_iter_advance(cs->iter, err);
 736         }
 737
 738         return lock_request(cs->req);
 739 }
 740
 741 /* Do as much copy to/from userspace buffer as we can */
 742 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 743 {
 744         unsigned ncpy = min(*size, cs->len);
 745         if (val) {
 746                 void *pgaddr = kmap_atomic(cs->pg);
 747                 void *buf = pgaddr + cs->offset;
 748
 749                 if (cs->write)
 750                         memcpy(buf, *val, ncpy);
 751                 else
 752                         memcpy(*val, buf, ncpy);
 753
 754                 kunmap_atomic(pgaddr);
 755                 *val += ncpy;
 756         }
 757         *size -= ncpy;
 758         cs->len -= ncpy;
 759         cs->offset += ncpy;
 760         return ncpy;
 761 }
 762
 763 static int fuse_check_page(struct page *page)
 764 {
 765         if (page_mapcount(page) ||
 766             page->mapping != NULL ||
 767             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 768              ~(1 << PG_locked |
 769                1 << PG_referenced |
 770                1 << PG_uptodate |
 771                1 << PG_lru |
 772                1 << PG_active |
 773                1 << PG_workingset |
 774                1 << PG_reclaim |
 775                1 << PG_waiters))) {
 776                 pr_warn("trying to steal weird page\n");
 777                 pr_warn("  page=%p index=%li flags=%08lx, count=%i, mapcount=%i, mapping=%p\n", page, page->index, page->flags, page_count(page), page_mapcount(page), page->mapping);
 778                 return 1;
 779         }
 780         return 0;
 781 }
 782
 783 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 784 {
 785         int err;
 786         struct page *oldpage = *pagep;
 787         struct page *newpage;
 788         struct pipe_buffer *buf = cs->pipebufs;
 789
 790         get_page(oldpage);
 791         err = unlock_request(cs->req);
 792         if (err)
 793                 goto out_put_old;
 794
 795         fuse_copy_finish(cs);
 796
 797         err = pipe_buf_confirm(cs->pipe, buf);
 798         if (err)
 799                 goto out_put_old;
 800
 801         BUG_ON(!cs->nr_segs);
 802         cs->currbuf = buf;
 803         cs->len = buf->len;
 804         cs->pipebufs++;
 805         cs->nr_segs--;
 806
 807         if (cs->len != PAGE_SIZE)
 808                 goto out_fallback;
 809
 810         if (pipe_buf_steal(cs->pipe, buf) != 0)
 811                 goto out_fallback;
 812
 813         newpage = buf->page;
 814
 815         if (!PageUptodate(newpage))
 816                 SetPageUptodate(newpage);
 817
 818         ClearPageMappedToDisk(newpage);
 819
 820         if (fuse_check_page(newpage) != 0)
 821                 goto out_fallback_unlock;
 822
 823         /*
 824          * This is a new and locked page, it shouldn't be mapped or
 825          * have any special flags on it
 826          */
 827         if (WARN_ON(page_mapped(oldpage)))
 828                 goto out_fallback_unlock;
 829         if (WARN_ON(page_has_private(oldpage)))
 830                 goto out_fallback_unlock;
 831         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 832                 goto out_fallback_unlock;
 833         if (WARN_ON(PageMlocked(oldpage)))
 834                 goto out_fallback_unlock;
 835
 836         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 837         if (err) {
 838                 unlock_page(newpage);
 839                 goto out_put_old;
 840         }
 841
 842         get_page(newpage);
 843
 844         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 845                 lru_cache_add_file(newpage);
 846
 847         /*
 848          * Release while we have extra ref on stolen page.  Otherwise
 849          * anon_pipe_buf_release() might think the page can be reused.
 850          */
 851         pipe_buf_release(cs->pipe, buf);
 852
 853         err = 0;
 854         spin_lock(&cs->req->waitq.lock);
 855         if (test_bit(FR_ABORTED, &cs->req->flags))
 856                 err = -ENOENT;
 857         else
 858                 *pagep = newpage;
 859         spin_unlock(&cs->req->waitq.lock);
 860
 861         if (err) {
 862                 unlock_page(newpage);
 863                 put_page(newpage);
 864                 goto out_put_old;
 865         }
 866
 867         unlock_page(oldpage);
 868         /* Drop ref for ap->pages[] array */
 869         put_page(oldpage);
 870         cs->len = 0;
 871
 872         err = 0;
 873 out_put_old:
 874         /* Drop ref obtained in this function */
 875         put_page(oldpage);
 876         return err;
 877
 878 out_fallback_unlock:
 879         unlock_page(newpage);
 880 out_fallback:
 881         cs->pg = buf->page;
 882         cs->offset = buf->offset;
 883
 884         err = lock_request(cs->req);
 885         if (!err)
 886                 err = 1;
 887
 888         goto out_put_old;
 889 }
 890
 891 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 892                          unsigned offset, unsigned count)
 893 {
 894         struct pipe_buffer *buf;
 895         int err;
 896
 897         if (cs->nr_segs == cs->pipe->buffers)
 898                 return -EIO;
 899
 900         get_page(page);
 901         err = unlock_request(cs->req);
 902         if (err) {
 903                 put_page(page);
 904                 return err;
 905         }
 906
 907         fuse_copy_finish(cs);
 908
 909         buf = cs->pipebufs;
 910         buf->page = page;
 911         buf->offset = offset;
 912         buf->len = count;
 913
 914         cs->pipebufs++;
 915         cs->nr_segs++;
 916         cs->len = 0;
 917
 918         return 0;
 919 }
 920
 921 /*
 922  * Copy a page in the request to/from the userspace buffer.  Must be
 923  * done atomically
 924  */
 925 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 926                           unsigned offset, unsigned count, int zeroing)
 927 {
 928         int err;
 929         struct page *page = *pagep;
 930
 931         if (page && zeroing && count < PAGE_SIZE)
 932                 clear_highpage(page);
 933
 934         while (count) {
 935                 if (cs->write && cs->pipebufs && page) {
 936                         /*
 937                          * Can't control lifetime of pipe buffers, so always
 938                          * copy user pages.
 939                          */
 940                         if (cs->req->args->user_pages) {
 941                                 err = fuse_copy_fill(cs);
 942                                 if (err)
 943                                         return err;
 944                         } else {
 945                                 return fuse_ref_page(cs, page, offset, count);
 946                         }
 947                 } else if (!cs->len) {
 948                         if (cs->move_pages && page &&
 949                             offset == 0 && count == PAGE_SIZE) {
 950                                 err = fuse_try_move_page(cs, pagep);
 951                                 if (err <= 0)
 952                                         return err;
 953                         } else {
 954                                 err = fuse_copy_fill(cs);
 955                                 if (err)
 956                                         return err;
 957                         }
 958                 }
 959                 if (page) {
 960                         void *mapaddr = kmap_atomic(page);
 961                         void *buf = mapaddr + offset;
 962                         offset += fuse_copy_do(cs, &buf, &count);
 963                         kunmap_atomic(mapaddr);
 964                 } else
 965                         offset += fuse_copy_do(cs, NULL, &count);
 966         }
 967         if (page && !cs->write)
 968                 flush_dcache_page(page);
 969         return 0;
 970 }
 971
 972 /* Copy pages in the request to/from userspace buffer */
 973 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 974                            int zeroing)
 975 {
 976         unsigned i;
 977         struct fuse_req *req = cs->req;
 978         struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
 979
 980
 981         for (i = 0; i < ap->num_pages && (nbytes || zeroing); i++) {
 982                 int err;
 983                 unsigned int offset = ap->descs[i].offset;
 984                 unsigned int count = min(nbytes, ap->descs[i].length);
 985
 986                 err = fuse_copy_page(cs, &ap->pages[i], offset, count, zeroing);
 987                 if (err)
 988                         return err;
 989
 990                 nbytes -= count;
 991         }
 992         return 0;
 993 }
 994
 995 /* Copy a single argument in the request to/from userspace buffer */
 996 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 997 {
 998         while (size) {
 999                 if (!cs->len) {
1000                         int err = fuse_copy_fill(cs);
1001                         if (err)
1002                                 return err;
1003                 }
1004                 fuse_copy_do(cs, &val, &size);
1005         }
1006         return 0;
1007 }
1008
1009 /* Copy request arguments to/from userspace buffer */
1010 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1011                           unsigned argpages, struct fuse_arg *args,
1012                           int zeroing)
1013 {
1014         int err = 0;
1015         unsigned i;
1016
1017         for (i = 0; !err && i < numargs; i++)  {
1018                 struct fuse_arg *arg = &args[i];
1019                 if (i == numargs - 1 && argpages)
1020                         err = fuse_copy_pages(cs, arg->size, zeroing);
1021                 else
1022                         err = fuse_copy_one(cs, arg->value, arg->size);
1023         }
1024         return err;
1025 }
1026
1027 static int forget_pending(struct fuse_iqueue *fiq)
1028 {
1029         return fiq->forget_list_head.next != NULL;
1030 }
1031
1032 static int request_pending(struct fuse_iqueue *fiq)
1033 {
1034         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1035                 forget_pending(fiq);
1036 }
1037
1038 /*
1039  * Transfer an interrupt request to userspace
1040  *
1041  * Unlike other requests this is assembled on demand, without a need
1042  * to allocate a separate fuse_req structure.
1043  *
1044  * Called with fiq->lock held, releases it
1045  */
1046 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1047                                struct fuse_copy_state *cs,
1048                                size_t nbytes, struct fuse_req *req)
1049 __releases(fiq->lock)
1050 {
1051         struct fuse_in_header ih;
1052         struct fuse_interrupt_in arg;
1053         unsigned reqsize = sizeof(ih) + sizeof(arg);
1054         int err;
1055
1056         list_del_init(&req->intr_entry);
1057         memset(&ih, 0, sizeof(ih));
1058         memset(&arg, 0, sizeof(arg));
1059         ih.len = reqsize;
1060         ih.opcode = FUSE_INTERRUPT;
1061         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1062         arg.unique = req->in.h.unique;
1063
1064         spin_unlock(&fiq->lock);
1065         if (nbytes < reqsize)
1066                 return -EINVAL;
1067
1068         err = fuse_copy_one(cs, &ih, sizeof(ih));
1069         if (!err)
1070                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1071         fuse_copy_finish(cs);
1072
1073         return err ? err : reqsize;
1074 }
1075
1076 struct fuse_forget_link *fuse_dequeue_forget(struct fuse_iqueue *fiq,
1077                                              unsigned int max,
1078                                              unsigned int *countp)
1079 {
1080         struct fuse_forget_link *head = fiq->forget_list_head.next;
1081         struct fuse_forget_link **newhead = &head;
1082         unsigned count;
1083
1084         for (count = 0; *newhead != NULL && count < max; count++)
1085                 newhead = &(*newhead)->next;
1086
1087         fiq->forget_list_head.next = *newhead;
1088         *newhead = NULL;
1089         if (fiq->forget_list_head.next == NULL)
1090                 fiq->forget_list_tail = &fiq->forget_list_head;
1091
1092         if (countp != NULL)
1093                 *countp = count;
1094
1095         return head;
1096 }
1097 EXPORT_SYMBOL(fuse_dequeue_forget);
1098
1099 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1100                                    struct fuse_copy_state *cs,
1101                                    size_t nbytes)
1102 __releases(fiq->lock)
1103 {
1104         int err;
1105         struct fuse_forget_link *forget = fuse_dequeue_forget(fiq, 1, NULL);
1106         struct fuse_forget_in arg = {
1107                 .nlookup = forget->forget_one.nlookup,
1108         };
1109         struct fuse_in_header ih = {
1110                 .opcode = FUSE_FORGET,
1111                 .nodeid = forget->forget_one.nodeid,
1112                 .unique = fuse_get_unique(fiq),
1113                 .len = sizeof(ih) + sizeof(arg),
1114         };
1115
1116         spin_unlock(&fiq->lock);
1117         kfree(forget);
1118         if (nbytes < ih.len)
1119                 return -EINVAL;
1120
1121         err = fuse_copy_one(cs, &ih, sizeof(ih));
1122         if (!err)
1123                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1124         fuse_copy_finish(cs);
1125
1126         if (err)
1127                 return err;
1128
1129         return ih.len;
1130 }
1131
1132 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1133                                    struct fuse_copy_state *cs, size_t nbytes)
1134 __releases(fiq->lock)
1135 {
1136         int err;
1137         unsigned max_forgets;
1138         unsigned count;
1139         struct fuse_forget_link *head;
1140         struct fuse_batch_forget_in arg = { .count = 0 };
1141         struct fuse_in_header ih = {
1142                 .opcode = FUSE_BATCH_FORGET,
1143                 .unique = fuse_get_unique(fiq),
1144                 .len = sizeof(ih) + sizeof(arg),
1145         };
1146
1147         if (nbytes < ih.len) {
1148                 spin_unlock(&fiq->lock);
1149                 return -EINVAL;
1150         }
1151
1152         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1153         head = fuse_dequeue_forget(fiq, max_forgets, &count);
1154         spin_unlock(&fiq->lock);
1155
1156         arg.count = count;
1157         ih.len += count * sizeof(struct fuse_forget_one);
1158         err = fuse_copy_one(cs, &ih, sizeof(ih));
1159         if (!err)
1160                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1161
1162         while (head) {
1163                 struct fuse_forget_link *forget = head;
1164
1165                 if (!err) {
1166                         err = fuse_copy_one(cs, &forget->forget_one,
1167                                             sizeof(forget->forget_one));
1168                 }
1169                 head = forget->next;
1170                 kfree(forget);
1171         }
1172
1173         fuse_copy_finish(cs);
1174
1175         if (err)
1176                 return err;
1177
1178         return ih.len;
1179 }
1180
1181 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1182                             struct fuse_copy_state *cs,
1183                             size_t nbytes)
1184 __releases(fiq->lock)
1185 {
1186         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1187                 return fuse_read_single_forget(fiq, cs, nbytes);
1188         else
1189                 return fuse_read_batch_forget(fiq, cs, nbytes);
1190 }
1191
1192 /*
1193  * Read a single request into the userspace filesystem's buffer.  This
1194  * function waits until a request is available, then removes it from
1195  * the pending list and copies request data to userspace buffer.  If
1196  * no reply is needed (FORGET) or request has been aborted or there
1197  * was an error during the copying then it's finished by calling
1198  * fuse_request_end().  Otherwise add it to the processing list, and set
1199  * the 'sent' flag.
1200  */
1201 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1202                                 struct fuse_copy_state *cs, size_t nbytes)
1203 {
1204         ssize_t err;
1205         struct fuse_conn *fc = fud->fc;
1206         struct fuse_iqueue *fiq = &fc->iq;
1207         struct fuse_pqueue *fpq = &fud->pq;
1208         struct fuse_req *req;
1209         struct fuse_args *args;
1210         unsigned reqsize;
1211         unsigned int hash;
1212
1213         /*
1214          * Require sane minimum read buffer - that has capacity for fixed part
1215          * of any request header + negotiated max_write room for data.
1216          *
1217          * Historically libfuse reserves 4K for fixed header room, but e.g.
1218          * GlusterFS reserves only 80 bytes
1219          *
1220          *      = `sizeof(fuse_in_header) + sizeof(fuse_write_in)`
1221          *
1222          * which is the absolute minimum any sane filesystem should be using
1223          * for header room.
1224          */
1225         if (nbytes < max_t(size_t, FUSE_MIN_READ_BUFFER,
1226                            sizeof(struct fuse_in_header) +
1227                            sizeof(struct fuse_write_in) +
1228                            fc->max_write))
1229                 return -EINVAL;
1230
1231  restart:
1232         for (;;) {
1233                 spin_lock(&fiq->lock);
1234                 if (!fiq->connected || request_pending(fiq))
1235                         break;
1236                 spin_unlock(&fiq->lock);
1237
1238                 if (file->f_flags & O_NONBLOCK)
1239                         return -EAGAIN;
1240                 err = wait_event_interruptible_exclusive(fiq->waitq,
1241                                 !fiq->connected || request_pending(fiq));
1242                 if (err)
1243                         return err;
1244         }
1245
1246         if (!fiq->connected) {
1247                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1248                 goto err_unlock;
1249         }
1250
1251         if (!list_empty(&fiq->interrupts)) {
1252                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1253                                  intr_entry);
1254                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1255         }
1256
1257         if (forget_pending(fiq)) {
1258                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1259                         return fuse_read_forget(fc, fiq, cs, nbytes);
1260
1261                 if (fiq->forget_batch <= -8)
1262                         fiq->forget_batch = 16;
1263         }
1264
1265         req = list_entry(fiq->pending.next, struct fuse_req, list);
1266         clear_bit(FR_PENDING, &req->flags);
1267         list_del_init(&req->list);
1268         spin_unlock(&fiq->lock);
1269
1270         args = req->args;
1271         reqsize = req->in.h.len;
1272
1273         /* If request is too large, reply with an error and restart the read */
1274         if (nbytes < reqsize) {
1275                 req->out.h.error = -EIO;
1276                 /* SETXATTR is special, since it may contain too large data */
1277                 if (args->opcode == FUSE_SETXATTR)
1278                         req->out.h.error = -E2BIG;
1279                 fuse_request_end(fc, req);
1280                 goto restart;
1281         }
1282         spin_lock(&fpq->lock);
1283         /*
1284          *  Must not put request on fpq->io queue after having been shut down by
1285          *  fuse_abort_conn()
1286          */
1287         if (!fpq->connected) {
1288                 req->out.h.error = err = -ECONNABORTED;
1289                 goto out_end;
1290
1291         }
1292         list_add(&req->list, &fpq->io);
1293         spin_unlock(&fpq->lock);
1294         cs->req = req;
1295         err = fuse_copy_one(cs, &req->in.h, sizeof(req->in.h));
1296         if (!err)
1297                 err = fuse_copy_args(cs, args->in_numargs, args->in_pages,
1298                                      (struct fuse_arg *) args->in_args, 0);
1299         fuse_copy_finish(cs);
1300         spin_lock(&fpq->lock);
1301         clear_bit(FR_LOCKED, &req->flags);
1302         if (!fpq->connected) {
1303                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1304                 goto out_end;
1305         }
1306         if (err) {
1307                 req->out.h.error = -EIO;
1308                 goto out_end;
1309         }
1310         if (!test_bit(FR_ISREPLY, &req->flags)) {
1311                 err = reqsize;
1312                 goto out_end;
1313         }
1314         hash = fuse_req_hash(req->in.h.unique);
1315         list_move_tail(&req->list, &fpq->processing[hash]);
1316         __fuse_get_request(req);
1317         set_bit(FR_SENT, &req->flags);
1318         spin_unlock(&fpq->lock);
1319         /* matches barrier in request_wait_answer() */
1320         smp_mb__after_atomic();
1321         if (test_bit(FR_INTERRUPTED, &req->flags))
1322                 queue_interrupt(fiq, req);
1323         fuse_put_request(fc, req);
1324
1325         return reqsize;
1326
1327 out_end:
1328         if (!test_bit(FR_PRIVATE, &req->flags))
1329                 list_del_init(&req->list);
1330         spin_unlock(&fpq->lock);
1331         fuse_request_end(fc, req);
1332         return err;
1333
1334  err_unlock:
1335         spin_unlock(&fiq->lock);
1336         return err;
1337 }
1338
1339 static int fuse_dev_open(struct inode *inode, struct file *file)
1340 {
1341         /*
1342          * The fuse device's file's private_data is used to hold
1343          * the fuse_conn(ection) when it is mounted, and is used to
1344          * keep track of whether the file has been mounted already.
1345          */
1346         file->private_data = NULL;
1347         return 0;
1348 }
1349
1350 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1351 {
1352         struct fuse_copy_state cs;
1353         struct file *file = iocb->ki_filp;
1354         struct fuse_dev *fud = fuse_get_dev(file);
1355
1356         if (!fud)
1357                 return -EPERM;
1358
1359         if (!iter_is_iovec(to))
1360                 return -EINVAL;
1361
1362         fuse_copy_init(&cs, 1, to);
1363
1364         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1365 }
1366
1367 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1368                                     struct pipe_inode_info *pipe,
1369                                     size_t len, unsigned int flags)
1370 {
1371         int total, ret;
1372         int page_nr = 0;
1373         struct pipe_buffer *bufs;
1374         struct fuse_copy_state cs;
1375         struct fuse_dev *fud = fuse_get_dev(in);
1376
1377         if (!fud)
1378                 return -EPERM;
1379
1380         bufs = kvmalloc_array(pipe->buffers, sizeof(struct pipe_buffer),
1381                               GFP_KERNEL);
1382         if (!bufs)
1383                 return -ENOMEM;
1384
1385         fuse_copy_init(&cs, 1, NULL);
1386         cs.pipebufs = bufs;
1387         cs.pipe = pipe;
1388         ret = fuse_dev_do_read(fud, in, &cs, len);
1389         if (ret < 0)
1390                 goto out;
1391
1392         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1393                 ret = -EIO;
1394                 goto out;
1395         }
1396
1397         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1398                 /*
1399                  * Need to be careful about this.  Having buf->ops in module
1400                  * code can Oops if the buffer persists after module unload.
1401                  */
1402                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1403                 bufs[page_nr].flags = 0;
1404                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1405                 if (unlikely(ret < 0))
1406                         break;
1407         }
1408         if (total)
1409                 ret = total;
1410 out:
1411         for (; page_nr < cs.nr_segs; page_nr++)
1412                 put_page(bufs[page_nr].page);
1413
1414         kvfree(bufs);
1415         return ret;
1416 }
1417
1418 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1419                             struct fuse_copy_state *cs)
1420 {
1421         struct fuse_notify_poll_wakeup_out outarg;
1422         int err = -EINVAL;
1423
1424         if (size != sizeof(outarg))
1425                 goto err;
1426
1427         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1428         if (err)
1429                 goto err;
1430
1431         fuse_copy_finish(cs);
1432         return fuse_notify_poll_wakeup(fc, &outarg);
1433
1434 err:
1435         fuse_copy_finish(cs);
1436         return err;
1437 }
1438
1439 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1440                                    struct fuse_copy_state *cs)
1441 {
1442         struct fuse_notify_inval_inode_out outarg;
1443         int err = -EINVAL;
1444
1445         if (size != sizeof(outarg))
1446                 goto err;
1447
1448         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1449         if (err)
1450                 goto err;
1451         fuse_copy_finish(cs);
1452
1453         down_read(&fc->killsb);
1454         err = -ENOENT;
1455         if (fc->sb) {
1456                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1457                                                outarg.off, outarg.len);
1458         }
1459         up_read(&fc->killsb);
1460         return err;
1461
1462 err:
1463         fuse_copy_finish(cs);
1464         return err;
1465 }
1466
1467 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1468                                    struct fuse_copy_state *cs)
1469 {
1470         struct fuse_notify_inval_entry_out outarg;
1471         int err = -ENOMEM;
1472         char *buf;
1473         struct qstr name;
1474
1475         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1476         if (!buf)
1477                 goto err;
1478
1479         err = -EINVAL;
1480         if (size < sizeof(outarg))
1481                 goto err;
1482
1483         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1484         if (err)
1485                 goto err;
1486
1487         err = -ENAMETOOLONG;
1488         if (outarg.namelen > FUSE_NAME_MAX)
1489                 goto err;
1490
1491         err = -EINVAL;
1492         if (size != sizeof(outarg) + outarg.namelen + 1)
1493                 goto err;
1494
1495         name.name = buf;
1496         name.len = outarg.namelen;
1497         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1498         if (err)
1499                 goto err;
1500         fuse_copy_finish(cs);
1501         buf[outarg.namelen] = 0;
1502
1503         down_read(&fc->killsb);
1504         err = -ENOENT;
1505         if (fc->sb)
1506                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1507         up_read(&fc->killsb);
1508         kfree(buf);
1509         return err;
1510
1511 err:
1512         kfree(buf);
1513         fuse_copy_finish(cs);
1514         return err;
1515 }
1516
1517 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1518                               struct fuse_copy_state *cs)
1519 {
1520         struct fuse_notify_delete_out outarg;
1521         int err = -ENOMEM;
1522         char *buf;
1523         struct qstr name;
1524
1525         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1526         if (!buf)
1527                 goto err;
1528
1529         err = -EINVAL;
1530         if (size < sizeof(outarg))
1531                 goto err;
1532
1533         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1534         if (err)
1535                 goto err;
1536
1537         err = -ENAMETOOLONG;
1538         if (outarg.namelen > FUSE_NAME_MAX)
1539                 goto err;
1540
1541         err = -EINVAL;
1542         if (size != sizeof(outarg) + outarg.namelen + 1)
1543                 goto err;
1544
1545         name.name = buf;
1546         name.len = outarg.namelen;
1547         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1548         if (err)
1549                 goto err;
1550         fuse_copy_finish(cs);
1551         buf[outarg.namelen] = 0;
1552
1553         down_read(&fc->killsb);
1554         err = -ENOENT;
1555         if (fc->sb)
1556                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1557                                                outarg.child, &name);
1558         up_read(&fc->killsb);
1559         kfree(buf);
1560         return err;
1561
1562 err:
1563         kfree(buf);
1564         fuse_copy_finish(cs);
1565         return err;
1566 }
1567
1568 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1569                              struct fuse_copy_state *cs)
1570 {
1571         struct fuse_notify_store_out outarg;
1572         struct inode *inode;
1573         struct address_space *mapping;
1574         u64 nodeid;
1575         int err;
1576         pgoff_t index;
1577         unsigned int offset;
1578         unsigned int num;
1579         loff_t file_size;
1580         loff_t end;
1581
1582         err = -EINVAL;
1583         if (size < sizeof(outarg))
1584                 goto out_finish;
1585
1586         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1587         if (err)
1588                 goto out_finish;
1589
1590         err = -EINVAL;
1591         if (size - sizeof(outarg) != outarg.size)
1592                 goto out_finish;
1593
1594         nodeid = outarg.nodeid;
1595
1596         down_read(&fc->killsb);
1597
1598         err = -ENOENT;
1599         if (!fc->sb)
1600                 goto out_up_killsb;
1601
1602         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1603         if (!inode)
1604                 goto out_up_killsb;
1605
1606         mapping = inode->i_mapping;
1607         index = outarg.offset >> PAGE_SHIFT;
1608         offset = outarg.offset & ~PAGE_MASK;
1609         file_size = i_size_read(inode);
1610         end = outarg.offset + outarg.size;
1611         if (end > file_size) {
1612                 file_size = end;
1613                 fuse_write_update_size(inode, file_size);
1614         }
1615
1616         num = outarg.size;
1617         while (num) {
1618                 struct page *page;
1619                 unsigned int this_num;
1620
1621                 err = -ENOMEM;
1622                 page = find_or_create_page(mapping, index,
1623                                            mapping_gfp_mask(mapping));
1624                 if (!page)
1625                         goto out_iput;
1626
1627                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1628                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1629                 if (!err && offset == 0 &&
1630                     (this_num == PAGE_SIZE || file_size == end))
1631                         SetPageUptodate(page);
1632                 unlock_page(page);
1633                 put_page(page);
1634
1635                 if (err)
1636                         goto out_iput;
1637
1638                 num -= this_num;
1639                 offset = 0;
1640                 index++;
1641         }
1642
1643         err = 0;
1644
1645 out_iput:
1646         iput(inode);
1647 out_up_killsb:
1648         up_read(&fc->killsb);
1649 out_finish:
1650         fuse_copy_finish(cs);
1651         return err;
1652 }
1653
1654 struct fuse_retrieve_args {
1655         struct fuse_args_pages ap;
1656         struct fuse_notify_retrieve_in inarg;
1657 };
1658
1659 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_args *args,
1660                               int error)
1661 {
1662         struct fuse_retrieve_args *ra =
1663                 container_of(args, typeof(*ra), ap.args);
1664
1665         release_pages(ra->ap.pages, ra->ap.num_pages);
1666         kfree(ra);
1667 }
1668
1669 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1670                          struct fuse_notify_retrieve_out *outarg)
1671 {
1672         int err;
1673         struct address_space *mapping = inode->i_mapping;
1674         pgoff_t index;
1675         loff_t file_size;
1676         unsigned int num;
1677         unsigned int offset;
1678         size_t total_len = 0;
1679         unsigned int num_pages;
1680         struct fuse_retrieve_args *ra;
1681         size_t args_size = sizeof(*ra);
1682         struct fuse_args_pages *ap;
1683         struct fuse_args *args;
1684
1685         offset = outarg->offset & ~PAGE_MASK;
1686         file_size = i_size_read(inode);
1687
1688         num = min(outarg->size, fc->max_write);
1689         if (outarg->offset > file_size)
1690                 num = 0;
1691         else if (outarg->offset + num > file_size)
1692                 num = file_size - outarg->offset;
1693
1694         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1695         num_pages = min(num_pages, fc->max_pages);
1696
1697         args_size += num_pages * (sizeof(ap->pages[0]) + sizeof(ap->descs[0]));
1698
1699         ra = kzalloc(args_size, GFP_KERNEL);
1700         if (!ra)
1701                 return -ENOMEM;
1702
1703         ap = &ra->ap;
1704         ap->pages = (void *) (ra + 1);
1705         ap->descs = (void *) (ap->pages + num_pages);
1706
1707         args = &ap->args;
1708         args->nodeid = outarg->nodeid;
1709         args->opcode = FUSE_NOTIFY_REPLY;
1710         args->in_numargs = 2;
1711         args->in_pages = true;
1712         args->end = fuse_retrieve_end;
1713
1714         index = outarg->offset >> PAGE_SHIFT;
1715
1716         while (num && ap->num_pages < num_pages) {
1717                 struct page *page;
1718                 unsigned int this_num;
1719
1720                 page = find_get_page(mapping, index);
1721                 if (!page)
1722                         break;
1723
1724                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1725                 ap->pages[ap->num_pages] = page;
1726                 ap->descs[ap->num_pages].offset = offset;
1727                 ap->descs[ap->num_pages].length = this_num;
1728                 ap->num_pages++;
1729
1730                 offset = 0;
1731                 num -= this_num;
1732                 total_len += this_num;
1733                 index++;
1734         }
1735         ra->inarg.offset = outarg->offset;
1736         ra->inarg.size = total_len;
1737         args->in_args[0].size = sizeof(ra->inarg);
1738         args->in_args[0].value = &ra->inarg;
1739         args->in_args[1].size = total_len;
1740
1741         err = fuse_simple_notify_reply(fc, args, outarg->notify_unique);
1742         if (err)
1743                 fuse_retrieve_end(fc, args, err);
1744
1745         return err;
1746 }
1747
1748 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1749                                 struct fuse_copy_state *cs)
1750 {
1751         struct fuse_notify_retrieve_out outarg;
1752         struct inode *inode;
1753         int err;
1754
1755         err = -EINVAL;
1756         if (size != sizeof(outarg))
1757                 goto copy_finish;
1758
1759         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1760         if (err)
1761                 goto copy_finish;
1762
1763         fuse_copy_finish(cs);
1764
1765         down_read(&fc->killsb);
1766         err = -ENOENT;
1767         if (fc->sb) {
1768                 u64 nodeid = outarg.nodeid;
1769
1770                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1771                 if (inode) {
1772                         err = fuse_retrieve(fc, inode, &outarg);
1773                         iput(inode);
1774                 }
1775         }
1776         up_read(&fc->killsb);
1777
1778         return err;
1779
1780 copy_finish:
1781         fuse_copy_finish(cs);
1782         return err;
1783 }
1784
1785 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1786                        unsigned int size, struct fuse_copy_state *cs)
1787 {
1788         /* Don't try to move pages (yet) */
1789         cs->move_pages = 0;
1790
1791         switch (code) {
1792         case FUSE_NOTIFY_POLL:
1793                 return fuse_notify_poll(fc, size, cs);
1794
1795         case FUSE_NOTIFY_INVAL_INODE:
1796                 return fuse_notify_inval_inode(fc, size, cs);
1797
1798         case FUSE_NOTIFY_INVAL_ENTRY:
1799                 return fuse_notify_inval_entry(fc, size, cs);
1800
1801         case FUSE_NOTIFY_STORE:
1802                 return fuse_notify_store(fc, size, cs);
1803
1804         case FUSE_NOTIFY_RETRIEVE:
1805                 return fuse_notify_retrieve(fc, size, cs);
1806
1807         case FUSE_NOTIFY_DELETE:
1808                 return fuse_notify_delete(fc, size, cs);
1809
1810         default:
1811                 fuse_copy_finish(cs);
1812                 return -EINVAL;
1813         }
1814 }
1815
1816 /* Look up request on processing list by unique ID */
1817 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1818 {
1819         unsigned int hash = fuse_req_hash(unique);
1820         struct fuse_req *req;
1821
1822         list_for_each_entry(req, &fpq->processing[hash], list) {
1823                 if (req->in.h.unique == unique)
1824                         return req;
1825         }
1826         return NULL;
1827 }
1828
1829 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_args *args,
1830                          unsigned nbytes)
1831 {
1832         unsigned reqsize = sizeof(struct fuse_out_header);
1833
1834         reqsize += fuse_len_args(args->out_numargs, args->out_args);
1835
1836         if (reqsize < nbytes || (reqsize > nbytes && !args->out_argvar))
1837                 return -EINVAL;
1838         else if (reqsize > nbytes) {
1839                 struct fuse_arg *lastarg = &args->out_args[args->out_numargs-1];
1840                 unsigned diffsize = reqsize - nbytes;
1841
1842                 if (diffsize > lastarg->size)
1843                         return -EINVAL;
1844                 lastarg->size -= diffsize;
1845         }
1846         return fuse_copy_args(cs, args->out_numargs, args->out_pages,
1847                               args->out_args, args->page_zeroing);
1848 }
1849
1850 /*
1851  * Write a single reply to a request.  First the header is copied from
1852  * the write buffer.  The request is then searched on the processing
1853  * list by the unique ID found in the header.  If found, then remove
1854  * it from the list and copy the rest of the buffer to the request.
1855  * The request is finished by calling fuse_request_end().
1856  */
1857 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1858                                  struct fuse_copy_state *cs, size_t nbytes)
1859 {
1860         int err;
1861         struct fuse_conn *fc = fud->fc;
1862         struct fuse_pqueue *fpq = &fud->pq;
1863         struct fuse_req *req;
1864         struct fuse_out_header oh;
1865
1866         err = -EINVAL;
1867         if (nbytes < sizeof(struct fuse_out_header))
1868                 goto out;
1869
1870         err = fuse_copy_one(cs, &oh, sizeof(oh));
1871         if (err)
1872                 goto copy_finish;
1873
1874         err = -EINVAL;
1875         if (oh.len != nbytes)
1876                 goto copy_finish;
1877
1878         /*
1879          * Zero oh.unique indicates unsolicited notification message
1880          * and error contains notification code.
1881          */
1882         if (!oh.unique) {
1883                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1884                 goto out;
1885         }
1886
1887         err = -EINVAL;
1888         if (oh.error <= -512 || oh.error > 0)
1889                 goto copy_finish;
1890
1891         spin_lock(&fpq->lock);
1892         req = NULL;
1893         if (fpq->connected)
1894                 req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1895
1896         err = -ENOENT;
1897         if (!req) {
1898                 spin_unlock(&fpq->lock);
1899                 goto copy_finish;
1900         }
1901
1902         /* Is it an interrupt reply ID? */
1903         if (oh.unique & FUSE_INT_REQ_BIT) {
1904                 __fuse_get_request(req);
1905                 spin_unlock(&fpq->lock);
1906
1907                 err = 0;
1908                 if (nbytes != sizeof(struct fuse_out_header))
1909                         err = -EINVAL;
1910                 else if (oh.error == -ENOSYS)
1911                         fc->no_interrupt = 1;
1912                 else if (oh.error == -EAGAIN)
1913                         err = queue_interrupt(&fc->iq, req);
1914
1915                 fuse_put_request(fc, req);
1916
1917                 goto copy_finish;
1918         }
1919
1920         clear_bit(FR_SENT, &req->flags);
1921         list_move(&req->list, &fpq->io);
1922         req->out.h = oh;
1923         set_bit(FR_LOCKED, &req->flags);
1924         spin_unlock(&fpq->lock);
1925         cs->req = req;
1926         if (!req->args->page_replace)
1927                 cs->move_pages = 0;
1928
1929         if (oh.error)
1930                 err = nbytes != sizeof(oh) ? -EINVAL : 0;
1931         else
1932                 err = copy_out_args(cs, req->args, nbytes);
1933         fuse_copy_finish(cs);
1934
1935         spin_lock(&fpq->lock);
1936         clear_bit(FR_LOCKED, &req->flags);
1937         if (!fpq->connected)
1938                 err = -ENOENT;
1939         else if (err)
1940                 req->out.h.error = -EIO;
1941         if (!test_bit(FR_PRIVATE, &req->flags))
1942                 list_del_init(&req->list);
1943         spin_unlock(&fpq->lock);
1944
1945         fuse_request_end(fc, req);
1946 out:
1947         return err ? err : nbytes;
1948
1949 copy_finish:
1950         fuse_copy_finish(cs);
1951         goto out;
1952 }
1953
1954 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1955 {
1956         struct fuse_copy_state cs;
1957         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1958
1959         if (!fud)
1960                 return -EPERM;
1961
1962         if (!iter_is_iovec(from))
1963                 return -EINVAL;
1964
1965         fuse_copy_init(&cs, 0, from);
1966
1967         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1968 }
1969
1970 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1971                                      struct file *out, loff_t *ppos,
1972                                      size_t len, unsigned int flags)
1973 {
1974         unsigned nbuf;
1975         unsigned idx;
1976         struct pipe_buffer *bufs;
1977         struct fuse_copy_state cs;
1978         struct fuse_dev *fud;
1979         size_t rem;
1980         ssize_t ret;
1981
1982         fud = fuse_get_dev(out);
1983         if (!fud)
1984                 return -EPERM;
1985
1986         pipe_lock(pipe);
1987
1988         bufs = kvmalloc_array(pipe->nrbufs, sizeof(struct pipe_buffer),
1989                               GFP_KERNEL);
1990         if (!bufs) {
1991                 pipe_unlock(pipe);
1992                 return -ENOMEM;
1993         }
1994
1995         nbuf = 0;
1996         rem = 0;
1997         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1998                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1999
2000         ret = -EINVAL;
2001         if (rem < len)
2002                 goto out_free;
2003
2004         rem = len;
2005         while (rem) {
2006                 struct pipe_buffer *ibuf;
2007                 struct pipe_buffer *obuf;
2008
2009                 BUG_ON(nbuf >= pipe->buffers);
2010                 BUG_ON(!pipe->nrbufs);
2011                 ibuf = &pipe->bufs[pipe->curbuf];
2012                 obuf = &bufs[nbuf];
2013
2014                 if (rem >= ibuf->len) {
2015                         *obuf = *ibuf;
2016                         ibuf->ops = NULL;
2017                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2018                         pipe->nrbufs--;
2019                 } else {
2020                         if (!pipe_buf_get(pipe, ibuf))
2021                                 goto out_free;
2022
2023                         *obuf = *ibuf;
2024                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2025                         obuf->len = rem;
2026                         ibuf->offset += obuf->len;
2027                         ibuf->len -= obuf->len;
2028                 }
2029                 nbuf++;
2030                 rem -= obuf->len;
2031         }
2032         pipe_unlock(pipe);
2033
2034         fuse_copy_init(&cs, 0, NULL);
2035         cs.pipebufs = bufs;
2036         cs.nr_segs = nbuf;
2037         cs.pipe = pipe;
2038
2039         if (flags & SPLICE_F_MOVE)
2040                 cs.move_pages = 1;
2041
2042         ret = fuse_dev_do_write(fud, &cs, len);
2043
2044         pipe_lock(pipe);
2045 out_free:
2046         for (idx = 0; idx < nbuf; idx++) {
2047                 struct pipe_buffer *buf = &bufs[idx];
2048
2049                 if (buf->ops)
2050                         pipe_buf_release(pipe, buf);
2051         }
2052         pipe_unlock(pipe);
2053
2054         kvfree(bufs);
2055         return ret;
2056 }
2057
2058 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2059 {
2060         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2061         struct fuse_iqueue *fiq;
2062         struct fuse_dev *fud = fuse_get_dev(file);
2063
2064         if (!fud)
2065                 return EPOLLERR;
2066
2067         fiq = &fud->fc->iq;
2068         poll_wait(file, &fiq->waitq, wait);
2069
2070         spin_lock(&fiq->lock);
2071         if (!fiq->connected)
2072                 mask = EPOLLERR;
2073         else if (request_pending(fiq))
2074                 mask |= EPOLLIN | EPOLLRDNORM;
2075         spin_unlock(&fiq->lock);
2076
2077         return mask;
2078 }
2079
2080 /* Abort all requests on the given list (pending or processing) */
2081 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2082 {
2083         while (!list_empty(head)) {
2084                 struct fuse_req *req;
2085                 req = list_entry(head->next, struct fuse_req, list);
2086                 req->out.h.error = -ECONNABORTED;
2087                 clear_bit(FR_SENT, &req->flags);
2088                 list_del_init(&req->list);
2089                 fuse_request_end(fc, req);
2090         }
2091 }
2092
2093 static void end_polls(struct fuse_conn *fc)
2094 {
2095         struct rb_node *p;
2096
2097         p = rb_first(&fc->polled_files);
2098
2099         while (p) {
2100                 struct fuse_file *ff;
2101                 ff = rb_entry(p, struct fuse_file, polled_node);
2102                 wake_up_interruptible_all(&ff->poll_wait);
2103
2104                 p = rb_next(p);
2105         }
2106 }
2107
2108 /*
2109  * Abort all requests.
2110  *
2111  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2112  * filesystem.
2113  *
2114  * The same effect is usually achievable through killing the filesystem daemon
2115  * and all users of the filesystem.  The exception is the combination of an
2116  * asynchronous request and the tricky deadlock (see
2117  * Documentation/filesystems/fuse.txt).
2118  *
2119  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2120  * requests, they should be finished off immediately.  Locked requests will be
2121  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2122  * requests.  It is possible that some request will finish before we can.  This
2123  * is OK, the request will in that case be removed from the list before we touch
2124  * it.
2125  */
2126 void fuse_abort_conn(struct fuse_conn *fc)
2127 {
2128         struct fuse_iqueue *fiq = &fc->iq;
2129
2130         spin_lock(&fc->lock);
2131         if (fc->connected) {
2132                 struct fuse_dev *fud;
2133                 struct fuse_req *req, *next;
2134                 LIST_HEAD(to_end);
2135                 unsigned int i;
2136
2137                 /* Background queuing checks fc->connected under bg_lock */
2138                 spin_lock(&fc->bg_lock);
2139                 fc->connected = 0;
2140                 spin_unlock(&fc->bg_lock);
2141
2142                 fuse_set_initialized(fc);
2143                 list_for_each_entry(fud, &fc->devices, entry) {
2144                         struct fuse_pqueue *fpq = &fud->pq;
2145
2146                         spin_lock(&fpq->lock);
2147                         fpq->connected = 0;
2148                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2149                                 req->out.h.error = -ECONNABORTED;
2150                                 spin_lock(&req->waitq.lock);
2151                                 set_bit(FR_ABORTED, &req->flags);
2152                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2153                                         set_bit(FR_PRIVATE, &req->flags);
2154                                         __fuse_get_request(req);
2155                                         list_move(&req->list, &to_end);
2156                                 }
2157                                 spin_unlock(&req->waitq.lock);
2158                         }
2159                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2160                                 list_splice_tail_init(&fpq->processing[i],
2161                                                       &to_end);
2162                         spin_unlock(&fpq->lock);
2163                 }
2164                 spin_lock(&fc->bg_lock);
2165                 fc->blocked = 0;
2166                 fc->max_background = UINT_MAX;
2167                 flush_bg_queue(fc);
2168                 spin_unlock(&fc->bg_lock);
2169
2170                 spin_lock(&fiq->lock);
2171                 fiq->connected = 0;
2172                 list_for_each_entry(req, &fiq->pending, list)
2173                         clear_bit(FR_PENDING, &req->flags);
2174                 list_splice_tail_init(&fiq->pending, &to_end);
2175                 while (forget_pending(fiq))
2176                         kfree(fuse_dequeue_forget(fiq, 1, NULL));
2177                 wake_up_all(&fiq->waitq);
2178                 spin_unlock(&fiq->lock);
2179                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2180                 end_polls(fc);
2181                 wake_up_all(&fc->blocked_waitq);
2182                 spin_unlock(&fc->lock);
2183
2184                 end_requests(fc, &to_end);
2185         } else {
2186                 spin_unlock(&fc->lock);
2187         }
2188 }
2189 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2190
2191 void fuse_wait_aborted(struct fuse_conn *fc)
2192 {
2193         /* matches implicit memory barrier in fuse_drop_waiting() */
2194         smp_mb();
2195         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2196 }
2197
2198 int fuse_dev_release(struct inode *inode, struct file *file)
2199 {
2200         struct fuse_dev *fud = fuse_get_dev(file);
2201
2202         if (fud) {
2203                 struct fuse_conn *fc = fud->fc;
2204                 struct fuse_pqueue *fpq = &fud->pq;
2205                 LIST_HEAD(to_end);
2206                 unsigned int i;
2207
2208                 spin_lock(&fpq->lock);
2209                 WARN_ON(!list_empty(&fpq->io));
2210                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2211                         list_splice_init(&fpq->processing[i], &to_end);
2212                 spin_unlock(&fpq->lock);
2213
2214                 end_requests(fc, &to_end);
2215
2216                 /* Are we the last open device? */
2217                 if (atomic_dec_and_test(&fc->dev_count)) {
2218                         WARN_ON(fc->iq.fasync != NULL);
2219                         fuse_abort_conn(fc);
2220                 }
2221                 fuse_dev_free(fud);
2222         }
2223         return 0;
2224 }
2225 EXPORT_SYMBOL_GPL(fuse_dev_release);
2226
2227 static int fuse_dev_fasync(int fd, struct file *file, int on)
2228 {
2229         struct fuse_dev *fud = fuse_get_dev(file);
2230
2231         if (!fud)
2232                 return -EPERM;
2233
2234         /* No locking - fasync_helper does its own locking */
2235         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2236 }
2237
2238 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2239 {
2240         struct fuse_dev *fud;
2241
2242         if (new->private_data)
2243                 return -EINVAL;
2244
2245         fud = fuse_dev_alloc_install(fc);
2246         if (!fud)
2247                 return -ENOMEM;
2248
2249         new->private_data = fud;
2250         atomic_inc(&fc->dev_count);
2251
2252         return 0;
2253 }
2254
2255 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2256                            unsigned long arg)
2257 {
2258         int err = -ENOTTY;
2259
2260         if (cmd == FUSE_DEV_IOC_CLONE) {
2261                 int oldfd;
2262
2263                 err = -EFAULT;
2264                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2265                         struct file *old = fget(oldfd);
2266
2267                         err = -EINVAL;
2268                         if (old) {
2269                                 struct fuse_dev *fud = NULL;
2270
2271                                 /*
2272                                  * Check against file->f_op because CUSE
2273                                  * uses the same ioctl handler.
2274                                  */
2275                                 if (old->f_op == file->f_op &&
2276                                     old->f_cred->user_ns == file->f_cred->user_ns)
2277                                         fud = fuse_get_dev(old);
2278
2279                                 if (fud) {
2280                                         mutex_lock(&fuse_mutex);
2281                                         err = fuse_device_clone(fud->fc, file);
2282                                         mutex_unlock(&fuse_mutex);
2283                                 }
2284                                 fput(old);
2285                         }
2286                 }
2287         }
2288         return err;
2289 }
2290
2291 const struct file_operations fuse_dev_operations = {
2292         .owner          = THIS_MODULE,
2293         .open           = fuse_dev_open,
2294         .llseek         = no_llseek,
2295         .read_iter      = fuse_dev_read,
2296         .splice_read    = fuse_dev_splice_read,
2297         .write_iter     = fuse_dev_write,
2298         .splice_write   = fuse_dev_splice_write,
2299         .poll           = fuse_dev_poll,
2300         .release        = fuse_dev_release,
2301         .fasync         = fuse_dev_fasync,
2302         .unlocked_ioctl = fuse_dev_ioctl,
2303         .compat_ioctl   = fuse_dev_ioctl,
2304 };
2305 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2306
2307 static struct miscdevice fuse_miscdevice = {
2308         .minor = FUSE_MINOR,
2309         .name  = "fuse",
2310         .fops = &fuse_dev_operations,
2311 };
2312
2313 int __init fuse_dev_init(void)
2314 {
2315         int err = -ENOMEM;
2316         fuse_req_cachep = kmem_cache_create("fuse_request",
2317                                             sizeof(struct fuse_req),
2318                                             0, 0, NULL);
2319         if (!fuse_req_cachep)
2320                 goto out;
2321
2322         err = misc_register(&fuse_miscdevice);
2323         if (err)
2324                 goto out_cache_clean;
2325
2326         return 0;
2327
2328  out_cache_clean:
2329         kmem_cache_destroy(fuse_req_cachep);
2330  out:
2331         return err;
2332 }
2333
2334 void fuse_dev_cleanup(void)
2335 {
2336         misc_deregister(&fuse_miscdevice);
2337         kmem_cache_destroy(fuse_req_cachep);
2338 }