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
  24 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  25 MODULE_ALIAS("devname:fuse");
  26
  27 static struct kmem_cache *fuse_req_cachep;
  28
  29 static struct fuse_dev *fuse_get_dev(struct file *file)
  30 {
  31         /*
  32          * Lockless access is OK, because file->private data is set
  33          * once during mount and is valid until the file is released.
  34          */
  35         return ACCESS_ONCE(file->private_data);
  36 }
  37
  38 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  39                               struct fuse_page_desc *page_descs,
  40                               unsigned npages)
  41 {
  42         memset(req, 0, sizeof(*req));
  43         memset(pages, 0, sizeof(*pages) * npages);
  44         memset(page_descs, 0, sizeof(*page_descs) * npages);
  45         INIT_LIST_HEAD(&req->list);
  46         INIT_LIST_HEAD(&req->intr_entry);
  47         init_waitqueue_head(&req->waitq);
  48         atomic_set(&req->count, 1);
  49         req->pages = pages;
  50         req->page_descs = page_descs;
  51         req->max_pages = npages;
  52         __set_bit(FR_PENDING, &req->flags);
  53 }
  54
  55 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  56 {
  57         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, flags);
  58         if (req) {
  59                 struct page **pages;
  60                 struct fuse_page_desc *page_descs;
  61
  62                 if (npages <= FUSE_REQ_INLINE_PAGES) {
  63                         pages = req->inline_pages;
  64                         page_descs = req->inline_page_descs;
  65                 } else {
  66                         pages = kmalloc(sizeof(struct page *) * npages, flags);
  67                         page_descs = kmalloc(sizeof(struct fuse_page_desc) *
  68                                              npages, flags);
  69                 }
  70
  71                 if (!pages || !page_descs) {
  72                         kfree(pages);
  73                         kfree(page_descs);
  74                         kmem_cache_free(fuse_req_cachep, req);
  75                         return NULL;
  76                 }
  77
  78                 fuse_request_init(req, pages, page_descs, npages);
  79         }
  80         return req;
  81 }
  82
  83 struct fuse_req *fuse_request_alloc(unsigned npages)
  84 {
  85         return __fuse_request_alloc(npages, GFP_KERNEL);
  86 }
  87 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  88
  89 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  90 {
  91         return __fuse_request_alloc(npages, GFP_NOFS);
  92 }
  93
  94 void fuse_request_free(struct fuse_req *req)
  95 {
  96         if (req->pages != req->inline_pages) {
  97                 kfree(req->pages);
  98                 kfree(req->page_descs);
  99         }
 100         kmem_cache_free(fuse_req_cachep, req);
 101 }
 102
 103 void __fuse_get_request(struct fuse_req *req)
 104 {
 105         atomic_inc(&req->count);
 106 }
 107
 108 /* Must be called with > 1 refcount */
 109 static void __fuse_put_request(struct fuse_req *req)
 110 {
 111         BUG_ON(atomic_read(&req->count) < 2);
 112         atomic_dec(&req->count);
 113 }
 114
 115 static void fuse_req_init_context(struct fuse_req *req)
 116 {
 117         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 118         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 119         req->in.h.pid = current->pid;
 120 }
 121
 122 void fuse_set_initialized(struct fuse_conn *fc)
 123 {
 124         /* Make sure stores before this are seen on another CPU */
 125         smp_wmb();
 126         fc->initialized = 1;
 127 }
 128
 129 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 130 {
 131         return !fc->initialized || (for_background && fc->blocked);
 132 }
 133
 134 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 135                                        bool for_background)
 136 {
 137         struct fuse_req *req;
 138         int err;
 139         atomic_inc(&fc->num_waiting);
 140
 141         if (fuse_block_alloc(fc, for_background)) {
 142                 err = -EINTR;
 143                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 144                                 !fuse_block_alloc(fc, for_background)))
 145                         goto out;
 146         }
 147         /* Matches smp_wmb() in fuse_set_initialized() */
 148         smp_rmb();
 149
 150         err = -ENOTCONN;
 151         if (!fc->connected)
 152                 goto out;
 153
 154         err = -ECONNREFUSED;
 155         if (fc->conn_error)
 156                 goto out;
 157
 158         req = fuse_request_alloc(npages);
 159         err = -ENOMEM;
 160         if (!req) {
 161                 if (for_background)
 162                         wake_up(&fc->blocked_waitq);
 163                 goto out;
 164         }
 165
 166         fuse_req_init_context(req);
 167         __set_bit(FR_WAITING, &req->flags);
 168         if (for_background)
 169                 __set_bit(FR_BACKGROUND, &req->flags);
 170
 171         return req;
 172
 173  out:
 174         atomic_dec(&fc->num_waiting);
 175         return ERR_PTR(err);
 176 }
 177
 178 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 179 {
 180         return __fuse_get_req(fc, npages, false);
 181 }
 182 EXPORT_SYMBOL_GPL(fuse_get_req);
 183
 184 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 185                                              unsigned npages)
 186 {
 187         return __fuse_get_req(fc, npages, true);
 188 }
 189 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 190
 191 /*
 192  * Return request in fuse_file->reserved_req.  However that may
 193  * currently be in use.  If that is the case, wait for it to become
 194  * available.
 195  */
 196 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 197                                          struct file *file)
 198 {
 199         struct fuse_req *req = NULL;
 200         struct fuse_file *ff = file->private_data;
 201
 202         do {
 203                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 204                 spin_lock(&fc->lock);
 205                 if (ff->reserved_req) {
 206                         req = ff->reserved_req;
 207                         ff->reserved_req = NULL;
 208                         req->stolen_file = get_file(file);
 209                 }
 210                 spin_unlock(&fc->lock);
 211         } while (!req);
 212
 213         return req;
 214 }
 215
 216 /*
 217  * Put stolen request back into fuse_file->reserved_req
 218  */
 219 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 220 {
 221         struct file *file = req->stolen_file;
 222         struct fuse_file *ff = file->private_data;
 223
 224         spin_lock(&fc->lock);
 225         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 226         BUG_ON(ff->reserved_req);
 227         ff->reserved_req = req;
 228         wake_up_all(&fc->reserved_req_waitq);
 229         spin_unlock(&fc->lock);
 230         fput(file);
 231 }
 232
 233 /*
 234  * Gets a requests for a file operation, always succeeds
 235  *
 236  * This is used for sending the FLUSH request, which must get to
 237  * userspace, due to POSIX locks which may need to be unlocked.
 238  *
 239  * If allocation fails due to OOM, use the reserved request in
 240  * fuse_file.
 241  *
 242  * This is very unlikely to deadlock accidentally, since the
 243  * filesystem should not have it's own file open.  If deadlock is
 244  * intentional, it can still be broken by "aborting" the filesystem.
 245  */
 246 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 247                                              struct file *file)
 248 {
 249         struct fuse_req *req;
 250
 251         atomic_inc(&fc->num_waiting);
 252         wait_event(fc->blocked_waitq, fc->initialized);
 253         /* Matches smp_wmb() in fuse_set_initialized() */
 254         smp_rmb();
 255         req = fuse_request_alloc(0);
 256         if (!req)
 257                 req = get_reserved_req(fc, file);
 258
 259         fuse_req_init_context(req);
 260         __set_bit(FR_WAITING, &req->flags);
 261         __clear_bit(FR_BACKGROUND, &req->flags);
 262         return req;
 263 }
 264
 265 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 266 {
 267         if (atomic_dec_and_test(&req->count)) {
 268                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 269                         /*
 270                          * We get here in the unlikely case that a background
 271                          * request was allocated but not sent
 272                          */
 273                         spin_lock(&fc->lock);
 274                         if (!fc->blocked)
 275                                 wake_up(&fc->blocked_waitq);
 276                         spin_unlock(&fc->lock);
 277                 }
 278
 279                 if (test_bit(FR_WAITING, &req->flags)) {
 280                         __clear_bit(FR_WAITING, &req->flags);
 281                         atomic_dec(&fc->num_waiting);
 282                 }
 283
 284                 if (req->stolen_file)
 285                         put_reserved_req(fc, req);
 286                 else
 287                         fuse_request_free(req);
 288         }
 289 }
 290 EXPORT_SYMBOL_GPL(fuse_put_request);
 291
 292 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 293 {
 294         unsigned nbytes = 0;
 295         unsigned i;
 296
 297         for (i = 0; i < numargs; i++)
 298                 nbytes += args[i].size;
 299
 300         return nbytes;
 301 }
 302
 303 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 304 {
 305         return ++fiq->reqctr;
 306 }
 307
 308 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 309 {
 310         req->in.h.len = sizeof(struct fuse_in_header) +
 311                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 312         list_add_tail(&req->list, &fiq->pending);
 313         wake_up_locked(&fiq->waitq);
 314         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 315 }
 316
 317 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 318                        u64 nodeid, u64 nlookup)
 319 {
 320         struct fuse_iqueue *fiq = &fc->iq;
 321
 322         forget->forget_one.nodeid = nodeid;
 323         forget->forget_one.nlookup = nlookup;
 324
 325         spin_lock(&fiq->waitq.lock);
 326         if (fiq->connected) {
 327                 fiq->forget_list_tail->next = forget;
 328                 fiq->forget_list_tail = forget;
 329                 wake_up_locked(&fiq->waitq);
 330                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 331         } else {
 332                 kfree(forget);
 333         }
 334         spin_unlock(&fiq->waitq.lock);
 335 }
 336
 337 static void flush_bg_queue(struct fuse_conn *fc)
 338 {
 339         while (fc->active_background < fc->max_background &&
 340                !list_empty(&fc->bg_queue)) {
 341                 struct fuse_req *req;
 342                 struct fuse_iqueue *fiq = &fc->iq;
 343
 344                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 345                 list_del(&req->list);
 346                 fc->active_background++;
 347                 spin_lock(&fiq->waitq.lock);
 348                 req->in.h.unique = fuse_get_unique(fiq);
 349                 queue_request(fiq, req);
 350                 spin_unlock(&fiq->waitq.lock);
 351         }
 352 }
 353
 354 /*
 355  * This function is called when a request is finished.  Either a reply
 356  * has arrived or it was aborted (and not yet sent) or some error
 357  * occurred during communication with userspace, or the device file
 358  * was closed.  The requester thread is woken up (if still waiting),
 359  * the 'end' callback is called if given, else the reference to the
 360  * request is released
 361  */
 362 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 363 {
 364         struct fuse_iqueue *fiq = &fc->iq;
 365
 366         if (test_and_set_bit(FR_FINISHED, &req->flags))
 367                 return;
 368
 369         spin_lock(&fiq->waitq.lock);
 370         list_del_init(&req->intr_entry);
 371         spin_unlock(&fiq->waitq.lock);
 372         WARN_ON(test_bit(FR_PENDING, &req->flags));
 373         WARN_ON(test_bit(FR_SENT, &req->flags));
 374         if (test_bit(FR_BACKGROUND, &req->flags)) {
 375                 spin_lock(&fc->lock);
 376                 clear_bit(FR_BACKGROUND, &req->flags);
 377                 if (fc->num_background == fc->max_background)
 378                         fc->blocked = 0;
 379
 380                 /* Wake up next waiter, if any */
 381                 if (!fc->blocked && waitqueue_active(&fc->blocked_waitq))
 382                         wake_up(&fc->blocked_waitq);
 383
 384                 if (fc->num_background == fc->congestion_threshold &&
 385                     fc->connected && fc->bdi_initialized) {
 386                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 387                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 388                 }
 389                 fc->num_background--;
 390                 fc->active_background--;
 391                 flush_bg_queue(fc);
 392                 spin_unlock(&fc->lock);
 393         }
 394         wake_up(&req->waitq);
 395         if (req->end)
 396                 req->end(fc, req);
 397         fuse_put_request(fc, req);
 398 }
 399
 400 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 401 {
 402         spin_lock(&fiq->waitq.lock);
 403         if (test_bit(FR_FINISHED, &req->flags)) {
 404                 spin_unlock(&fiq->waitq.lock);
 405                 return;
 406         }
 407         if (list_empty(&req->intr_entry)) {
 408                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 409                 wake_up_locked(&fiq->waitq);
 410         }
 411         spin_unlock(&fiq->waitq.lock);
 412         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 413 }
 414
 415 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 416 {
 417         struct fuse_iqueue *fiq = &fc->iq;
 418         int err;
 419
 420         if (!fc->no_interrupt) {
 421                 /* Any signal may interrupt this */
 422                 err = wait_event_interruptible(req->waitq,
 423                                         test_bit(FR_FINISHED, &req->flags));
 424                 if (!err)
 425                         return;
 426
 427                 set_bit(FR_INTERRUPTED, &req->flags);
 428                 /* matches barrier in fuse_dev_do_read() */
 429                 smp_mb__after_atomic();
 430                 if (test_bit(FR_SENT, &req->flags))
 431                         queue_interrupt(fiq, req);
 432         }
 433
 434         if (!test_bit(FR_FORCE, &req->flags)) {
 435                 /* Only fatal signals may interrupt this */
 436                 err = wait_event_killable(req->waitq,
 437                                         test_bit(FR_FINISHED, &req->flags));
 438                 if (!err)
 439                         return;
 440
 441                 spin_lock(&fiq->waitq.lock);
 442                 /* Request is not yet in userspace, bail out */
 443                 if (test_bit(FR_PENDING, &req->flags)) {
 444                         list_del(&req->list);
 445                         spin_unlock(&fiq->waitq.lock);
 446                         __fuse_put_request(req);
 447                         req->out.h.error = -EINTR;
 448                         return;
 449                 }
 450                 spin_unlock(&fiq->waitq.lock);
 451         }
 452
 453         /*
 454          * Either request is already in userspace, or it was forced.
 455          * Wait it out.
 456          */
 457         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 458 }
 459
 460 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 461 {
 462         struct fuse_iqueue *fiq = &fc->iq;
 463
 464         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 465         spin_lock(&fiq->waitq.lock);
 466         if (!fiq->connected) {
 467                 spin_unlock(&fiq->waitq.lock);
 468                 req->out.h.error = -ENOTCONN;
 469         } else {
 470                 req->in.h.unique = fuse_get_unique(fiq);
 471                 queue_request(fiq, req);
 472                 /* acquire extra reference, since request is still needed
 473                    after request_end() */
 474                 __fuse_get_request(req);
 475                 spin_unlock(&fiq->waitq.lock);
 476
 477                 request_wait_answer(fc, req);
 478                 /* Pairs with smp_wmb() in request_end() */
 479                 smp_rmb();
 480         }
 481 }
 482
 483 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 484 {
 485         __set_bit(FR_ISREPLY, &req->flags);
 486         if (!test_bit(FR_WAITING, &req->flags)) {
 487                 __set_bit(FR_WAITING, &req->flags);
 488                 atomic_inc(&fc->num_waiting);
 489         }
 490         __fuse_request_send(fc, req);
 491 }
 492 EXPORT_SYMBOL_GPL(fuse_request_send);
 493
 494 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 495 {
 496         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 497                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 498
 499         if (fc->minor < 9) {
 500                 switch (args->in.h.opcode) {
 501                 case FUSE_LOOKUP:
 502                 case FUSE_CREATE:
 503                 case FUSE_MKNOD:
 504                 case FUSE_MKDIR:
 505                 case FUSE_SYMLINK:
 506                 case FUSE_LINK:
 507                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 508                         break;
 509                 case FUSE_GETATTR:
 510                 case FUSE_SETATTR:
 511                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 512                         break;
 513                 }
 514         }
 515         if (fc->minor < 12) {
 516                 switch (args->in.h.opcode) {
 517                 case FUSE_CREATE:
 518                         args->in.args[0].size = sizeof(struct fuse_open_in);
 519                         break;
 520                 case FUSE_MKNOD:
 521                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 522                         break;
 523                 }
 524         }
 525 }
 526
 527 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 528 {
 529         struct fuse_req *req;
 530         ssize_t ret;
 531
 532         req = fuse_get_req(fc, 0);
 533         if (IS_ERR(req))
 534                 return PTR_ERR(req);
 535
 536         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 537         fuse_adjust_compat(fc, args);
 538
 539         req->in.h.opcode = args->in.h.opcode;
 540         req->in.h.nodeid = args->in.h.nodeid;
 541         req->in.numargs = args->in.numargs;
 542         memcpy(req->in.args, args->in.args,
 543                args->in.numargs * sizeof(struct fuse_in_arg));
 544         req->out.argvar = args->out.argvar;
 545         req->out.numargs = args->out.numargs;
 546         memcpy(req->out.args, args->out.args,
 547                args->out.numargs * sizeof(struct fuse_arg));
 548         fuse_request_send(fc, req);
 549         ret = req->out.h.error;
 550         if (!ret && args->out.argvar) {
 551                 BUG_ON(args->out.numargs != 1);
 552                 ret = req->out.args[0].size;
 553         }
 554         fuse_put_request(fc, req);
 555
 556         return ret;
 557 }
 558
 559 /*
 560  * Called under fc->lock
 561  *
 562  * fc->connected must have been checked previously
 563  */
 564 void fuse_request_send_background_locked(struct fuse_conn *fc,
 565                                          struct fuse_req *req)
 566 {
 567         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 568         if (!test_bit(FR_WAITING, &req->flags)) {
 569                 __set_bit(FR_WAITING, &req->flags);
 570                 atomic_inc(&fc->num_waiting);
 571         }
 572         __set_bit(FR_ISREPLY, &req->flags);
 573         fc->num_background++;
 574         if (fc->num_background == fc->max_background)
 575                 fc->blocked = 1;
 576         if (fc->num_background == fc->congestion_threshold &&
 577             fc->bdi_initialized) {
 578                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 579                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 580         }
 581         list_add_tail(&req->list, &fc->bg_queue);
 582         flush_bg_queue(fc);
 583 }
 584
 585 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 586 {
 587         BUG_ON(!req->end);
 588         spin_lock(&fc->lock);
 589         if (fc->connected) {
 590                 fuse_request_send_background_locked(fc, req);
 591                 spin_unlock(&fc->lock);
 592         } else {
 593                 spin_unlock(&fc->lock);
 594                 req->out.h.error = -ENOTCONN;
 595                 req->end(fc, req);
 596                 fuse_put_request(fc, req);
 597         }
 598 }
 599 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 600
 601 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 602                                           struct fuse_req *req, u64 unique)
 603 {
 604         int err = -ENODEV;
 605         struct fuse_iqueue *fiq = &fc->iq;
 606
 607         __clear_bit(FR_ISREPLY, &req->flags);
 608         req->in.h.unique = unique;
 609         spin_lock(&fiq->waitq.lock);
 610         if (fiq->connected) {
 611                 queue_request(fiq, req);
 612                 err = 0;
 613         }
 614         spin_unlock(&fiq->waitq.lock);
 615
 616         return err;
 617 }
 618
 619 void fuse_force_forget(struct file *file, u64 nodeid)
 620 {
 621         struct inode *inode = file_inode(file);
 622         struct fuse_conn *fc = get_fuse_conn(inode);
 623         struct fuse_req *req;
 624         struct fuse_forget_in inarg;
 625
 626         memset(&inarg, 0, sizeof(inarg));
 627         inarg.nlookup = 1;
 628         req = fuse_get_req_nofail_nopages(fc, file);
 629         req->in.h.opcode = FUSE_FORGET;
 630         req->in.h.nodeid = nodeid;
 631         req->in.numargs = 1;
 632         req->in.args[0].size = sizeof(inarg);
 633         req->in.args[0].value = &inarg;
 634         __clear_bit(FR_ISREPLY, &req->flags);
 635         __fuse_request_send(fc, req);
 636         /* ignore errors */
 637         fuse_put_request(fc, req);
 638 }
 639
 640 /*
 641  * Lock the request.  Up to the next unlock_request() there mustn't be
 642  * anything that could cause a page-fault.  If the request was already
 643  * aborted bail out.
 644  */
 645 static int lock_request(struct fuse_req *req)
 646 {
 647         int err = 0;
 648         if (req) {
 649                 spin_lock(&req->waitq.lock);
 650                 if (test_bit(FR_ABORTED, &req->flags))
 651                         err = -ENOENT;
 652                 else
 653                         set_bit(FR_LOCKED, &req->flags);
 654                 spin_unlock(&req->waitq.lock);
 655         }
 656         return err;
 657 }
 658
 659 /*
 660  * Unlock request.  If it was aborted while locked, caller is responsible
 661  * for unlocking and ending the request.
 662  */
 663 static int unlock_request(struct fuse_req *req)
 664 {
 665         int err = 0;
 666         if (req) {
 667                 spin_lock(&req->waitq.lock);
 668                 if (test_bit(FR_ABORTED, &req->flags))
 669                         err = -ENOENT;
 670                 else
 671                         clear_bit(FR_LOCKED, &req->flags);
 672                 spin_unlock(&req->waitq.lock);
 673         }
 674         return err;
 675 }
 676
 677 struct fuse_copy_state {
 678         int write;
 679         struct fuse_req *req;
 680         struct iov_iter *iter;
 681         struct pipe_buffer *pipebufs;
 682         struct pipe_buffer *currbuf;
 683         struct pipe_inode_info *pipe;
 684         unsigned long nr_segs;
 685         struct page *pg;
 686         unsigned len;
 687         unsigned offset;
 688         unsigned move_pages:1;
 689 };
 690
 691 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 692                            struct iov_iter *iter)
 693 {
 694         memset(cs, 0, sizeof(*cs));
 695         cs->write = write;
 696         cs->iter = iter;
 697 }
 698
 699 /* Unmap and put previous page of userspace buffer */
 700 static void fuse_copy_finish(struct fuse_copy_state *cs)
 701 {
 702         if (cs->currbuf) {
 703                 struct pipe_buffer *buf = cs->currbuf;
 704
 705                 if (cs->write)
 706                         buf->len = PAGE_SIZE - cs->len;
 707                 cs->currbuf = NULL;
 708         } else if (cs->pg) {
 709                 if (cs->write) {
 710                         flush_dcache_page(cs->pg);
 711                         set_page_dirty_lock(cs->pg);
 712                 }
 713                 put_page(cs->pg);
 714         }
 715         cs->pg = NULL;
 716 }
 717
 718 /*
 719  * Get another pagefull of userspace buffer, and map it to kernel
 720  * address space, and lock request
 721  */
 722 static int fuse_copy_fill(struct fuse_copy_state *cs)
 723 {
 724         struct page *page;
 725         int err;
 726
 727         err = unlock_request(cs->req);
 728         if (err)
 729                 return err;
 730
 731         fuse_copy_finish(cs);
 732         if (cs->pipebufs) {
 733                 struct pipe_buffer *buf = cs->pipebufs;
 734
 735                 if (!cs->write) {
 736                         err = pipe_buf_confirm(cs->pipe, buf);
 737                         if (err)
 738                                 return err;
 739
 740                         BUG_ON(!cs->nr_segs);
 741                         cs->currbuf = buf;
 742                         cs->pg = buf->page;
 743                         cs->offset = buf->offset;
 744                         cs->len = buf->len;
 745                         cs->pipebufs++;
 746                         cs->nr_segs--;
 747                 } else {
 748                         if (cs->nr_segs == cs->pipe->buffers)
 749                                 return -EIO;
 750
 751                         page = alloc_page(GFP_HIGHUSER);
 752                         if (!page)
 753                                 return -ENOMEM;
 754
 755                         buf->page = page;
 756                         buf->offset = 0;
 757                         buf->len = 0;
 758
 759                         cs->currbuf = buf;
 760                         cs->pg = page;
 761                         cs->offset = 0;
 762                         cs->len = PAGE_SIZE;
 763                         cs->pipebufs++;
 764                         cs->nr_segs++;
 765                 }
 766         } else {
 767                 size_t off;
 768                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 769                 if (err < 0)
 770                         return err;
 771                 BUG_ON(!err);
 772                 cs->len = err;
 773                 cs->offset = off;
 774                 cs->pg = page;
 775                 iov_iter_advance(cs->iter, err);
 776         }
 777
 778         return lock_request(cs->req);
 779 }
 780
 781 /* Do as much copy to/from userspace buffer as we can */
 782 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 783 {
 784         unsigned ncpy = min(*size, cs->len);
 785         if (val) {
 786                 void *pgaddr = kmap_atomic(cs->pg);
 787                 void *buf = pgaddr + cs->offset;
 788
 789                 if (cs->write)
 790                         memcpy(buf, *val, ncpy);
 791                 else
 792                         memcpy(*val, buf, ncpy);
 793
 794                 kunmap_atomic(pgaddr);
 795                 *val += ncpy;
 796         }
 797         *size -= ncpy;
 798         cs->len -= ncpy;
 799         cs->offset += ncpy;
 800         return ncpy;
 801 }
 802
 803 static int fuse_check_page(struct page *page)
 804 {
 805         if (page_mapcount(page) ||
 806             page->mapping != NULL ||
 807             page_count(page) != 1 ||
 808             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 809              ~(1 << PG_locked |
 810                1 << PG_referenced |
 811                1 << PG_uptodate |
 812                1 << PG_lru |
 813                1 << PG_active |
 814                1 << PG_reclaim))) {
 815                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 816                 printk(KERN_WARNING "  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);
 817                 return 1;
 818         }
 819         return 0;
 820 }
 821
 822 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 823 {
 824         int err;
 825         struct page *oldpage = *pagep;
 826         struct page *newpage;
 827         struct pipe_buffer *buf = cs->pipebufs;
 828
 829         err = unlock_request(cs->req);
 830         if (err)
 831                 return err;
 832
 833         fuse_copy_finish(cs);
 834
 835         err = pipe_buf_confirm(cs->pipe, buf);
 836         if (err)
 837                 return err;
 838
 839         BUG_ON(!cs->nr_segs);
 840         cs->currbuf = buf;
 841         cs->len = buf->len;
 842         cs->pipebufs++;
 843         cs->nr_segs--;
 844
 845         if (cs->len != PAGE_SIZE)
 846                 goto out_fallback;
 847
 848         if (pipe_buf_steal(cs->pipe, buf) != 0)
 849                 goto out_fallback;
 850
 851         newpage = buf->page;
 852
 853         if (!PageUptodate(newpage))
 854                 SetPageUptodate(newpage);
 855
 856         ClearPageMappedToDisk(newpage);
 857
 858         if (fuse_check_page(newpage) != 0)
 859                 goto out_fallback_unlock;
 860
 861         /*
 862          * This is a new and locked page, it shouldn't be mapped or
 863          * have any special flags on it
 864          */
 865         if (WARN_ON(page_mapped(oldpage)))
 866                 goto out_fallback_unlock;
 867         if (WARN_ON(page_has_private(oldpage)))
 868                 goto out_fallback_unlock;
 869         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 870                 goto out_fallback_unlock;
 871         if (WARN_ON(PageMlocked(oldpage)))
 872                 goto out_fallback_unlock;
 873
 874         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 875         if (err) {
 876                 unlock_page(newpage);
 877                 return err;
 878         }
 879
 880         get_page(newpage);
 881
 882         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 883                 lru_cache_add_file(newpage);
 884
 885         err = 0;
 886         spin_lock(&cs->req->waitq.lock);
 887         if (test_bit(FR_ABORTED, &cs->req->flags))
 888                 err = -ENOENT;
 889         else
 890                 *pagep = newpage;
 891         spin_unlock(&cs->req->waitq.lock);
 892
 893         if (err) {
 894                 unlock_page(newpage);
 895                 put_page(newpage);
 896                 return err;
 897         }
 898
 899         unlock_page(oldpage);
 900         put_page(oldpage);
 901         cs->len = 0;
 902
 903         return 0;
 904
 905 out_fallback_unlock:
 906         unlock_page(newpage);
 907 out_fallback:
 908         cs->pg = buf->page;
 909         cs->offset = buf->offset;
 910
 911         err = lock_request(cs->req);
 912         if (err)
 913                 return err;
 914
 915         return 1;
 916 }
 917
 918 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 919                          unsigned offset, unsigned count)
 920 {
 921         struct pipe_buffer *buf;
 922         int err;
 923
 924         if (cs->nr_segs == cs->pipe->buffers)
 925                 return -EIO;
 926
 927         err = unlock_request(cs->req);
 928         if (err)
 929                 return err;
 930
 931         fuse_copy_finish(cs);
 932
 933         buf = cs->pipebufs;
 934         get_page(page);
 935         buf->page = page;
 936         buf->offset = offset;
 937         buf->len = count;
 938
 939         cs->pipebufs++;
 940         cs->nr_segs++;
 941         cs->len = 0;
 942
 943         return 0;
 944 }
 945
 946 /*
 947  * Copy a page in the request to/from the userspace buffer.  Must be
 948  * done atomically
 949  */
 950 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 951                           unsigned offset, unsigned count, int zeroing)
 952 {
 953         int err;
 954         struct page *page = *pagep;
 955
 956         if (page && zeroing && count < PAGE_SIZE)
 957                 clear_highpage(page);
 958
 959         while (count) {
 960                 if (cs->write && cs->pipebufs && page) {
 961                         return fuse_ref_page(cs, page, offset, count);
 962                 } else if (!cs->len) {
 963                         if (cs->move_pages && page &&
 964                             offset == 0 && count == PAGE_SIZE) {
 965                                 err = fuse_try_move_page(cs, pagep);
 966                                 if (err <= 0)
 967                                         return err;
 968                         } else {
 969                                 err = fuse_copy_fill(cs);
 970                                 if (err)
 971                                         return err;
 972                         }
 973                 }
 974                 if (page) {
 975                         void *mapaddr = kmap_atomic(page);
 976                         void *buf = mapaddr + offset;
 977                         offset += fuse_copy_do(cs, &buf, &count);
 978                         kunmap_atomic(mapaddr);
 979                 } else
 980                         offset += fuse_copy_do(cs, NULL, &count);
 981         }
 982         if (page && !cs->write)
 983                 flush_dcache_page(page);
 984         return 0;
 985 }
 986
 987 /* Copy pages in the request to/from userspace buffer */
 988 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 989                            int zeroing)
 990 {
 991         unsigned i;
 992         struct fuse_req *req = cs->req;
 993
 994         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 995                 int err;
 996                 unsigned offset = req->page_descs[i].offset;
 997                 unsigned count = min(nbytes, req->page_descs[i].length);
 998
 999                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1000                                      zeroing);
1001                 if (err)
1002                         return err;
1003
1004                 nbytes -= count;
1005         }
1006         return 0;
1007 }
1008
1009 /* Copy a single argument in the request to/from userspace buffer */
1010 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1011 {
1012         while (size) {
1013                 if (!cs->len) {
1014                         int err = fuse_copy_fill(cs);
1015                         if (err)
1016                                 return err;
1017                 }
1018                 fuse_copy_do(cs, &val, &size);
1019         }
1020         return 0;
1021 }
1022
1023 /* Copy request arguments to/from userspace buffer */
1024 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1025                           unsigned argpages, struct fuse_arg *args,
1026                           int zeroing)
1027 {
1028         int err = 0;
1029         unsigned i;
1030
1031         for (i = 0; !err && i < numargs; i++)  {
1032                 struct fuse_arg *arg = &args[i];
1033                 if (i == numargs - 1 && argpages)
1034                         err = fuse_copy_pages(cs, arg->size, zeroing);
1035                 else
1036                         err = fuse_copy_one(cs, arg->value, arg->size);
1037         }
1038         return err;
1039 }
1040
1041 static int forget_pending(struct fuse_iqueue *fiq)
1042 {
1043         return fiq->forget_list_head.next != NULL;
1044 }
1045
1046 static int request_pending(struct fuse_iqueue *fiq)
1047 {
1048         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1049                 forget_pending(fiq);
1050 }
1051
1052 /*
1053  * Transfer an interrupt request to userspace
1054  *
1055  * Unlike other requests this is assembled on demand, without a need
1056  * to allocate a separate fuse_req structure.
1057  *
1058  * Called with fiq->waitq.lock held, releases it
1059  */
1060 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1061                                struct fuse_copy_state *cs,
1062                                size_t nbytes, struct fuse_req *req)
1063 __releases(fiq->waitq.lock)
1064 {
1065         struct fuse_in_header ih;
1066         struct fuse_interrupt_in arg;
1067         unsigned reqsize = sizeof(ih) + sizeof(arg);
1068         int err;
1069
1070         list_del_init(&req->intr_entry);
1071         req->intr_unique = fuse_get_unique(fiq);
1072         memset(&ih, 0, sizeof(ih));
1073         memset(&arg, 0, sizeof(arg));
1074         ih.len = reqsize;
1075         ih.opcode = FUSE_INTERRUPT;
1076         ih.unique = req->intr_unique;
1077         arg.unique = req->in.h.unique;
1078
1079         spin_unlock(&fiq->waitq.lock);
1080         if (nbytes < reqsize)
1081                 return -EINVAL;
1082
1083         err = fuse_copy_one(cs, &ih, sizeof(ih));
1084         if (!err)
1085                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1086         fuse_copy_finish(cs);
1087
1088         return err ? err : reqsize;
1089 }
1090
1091 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1092                                                unsigned max,
1093                                                unsigned *countp)
1094 {
1095         struct fuse_forget_link *head = fiq->forget_list_head.next;
1096         struct fuse_forget_link **newhead = &head;
1097         unsigned count;
1098
1099         for (count = 0; *newhead != NULL && count < max; count++)
1100                 newhead = &(*newhead)->next;
1101
1102         fiq->forget_list_head.next = *newhead;
1103         *newhead = NULL;
1104         if (fiq->forget_list_head.next == NULL)
1105                 fiq->forget_list_tail = &fiq->forget_list_head;
1106
1107         if (countp != NULL)
1108                 *countp = count;
1109
1110         return head;
1111 }
1112
1113 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1114                                    struct fuse_copy_state *cs,
1115                                    size_t nbytes)
1116 __releases(fiq->waitq.lock)
1117 {
1118         int err;
1119         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1120         struct fuse_forget_in arg = {
1121                 .nlookup = forget->forget_one.nlookup,
1122         };
1123         struct fuse_in_header ih = {
1124                 .opcode = FUSE_FORGET,
1125                 .nodeid = forget->forget_one.nodeid,
1126                 .unique = fuse_get_unique(fiq),
1127                 .len = sizeof(ih) + sizeof(arg),
1128         };
1129
1130         spin_unlock(&fiq->waitq.lock);
1131         kfree(forget);
1132         if (nbytes < ih.len)
1133                 return -EINVAL;
1134
1135         err = fuse_copy_one(cs, &ih, sizeof(ih));
1136         if (!err)
1137                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1138         fuse_copy_finish(cs);
1139
1140         if (err)
1141                 return err;
1142
1143         return ih.len;
1144 }
1145
1146 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1147                                    struct fuse_copy_state *cs, size_t nbytes)
1148 __releases(fiq->waitq.lock)
1149 {
1150         int err;
1151         unsigned max_forgets;
1152         unsigned count;
1153         struct fuse_forget_link *head;
1154         struct fuse_batch_forget_in arg = { .count = 0 };
1155         struct fuse_in_header ih = {
1156                 .opcode = FUSE_BATCH_FORGET,
1157                 .unique = fuse_get_unique(fiq),
1158                 .len = sizeof(ih) + sizeof(arg),
1159         };
1160
1161         if (nbytes < ih.len) {
1162                 spin_unlock(&fiq->waitq.lock);
1163                 return -EINVAL;
1164         }
1165
1166         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1167         head = dequeue_forget(fiq, max_forgets, &count);
1168         spin_unlock(&fiq->waitq.lock);
1169
1170         arg.count = count;
1171         ih.len += count * sizeof(struct fuse_forget_one);
1172         err = fuse_copy_one(cs, &ih, sizeof(ih));
1173         if (!err)
1174                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1175
1176         while (head) {
1177                 struct fuse_forget_link *forget = head;
1178
1179                 if (!err) {
1180                         err = fuse_copy_one(cs, &forget->forget_one,
1181                                             sizeof(forget->forget_one));
1182                 }
1183                 head = forget->next;
1184                 kfree(forget);
1185         }
1186
1187         fuse_copy_finish(cs);
1188
1189         if (err)
1190                 return err;
1191
1192         return ih.len;
1193 }
1194
1195 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1196                             struct fuse_copy_state *cs,
1197                             size_t nbytes)
1198 __releases(fiq->waitq.lock)
1199 {
1200         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1201                 return fuse_read_single_forget(fiq, cs, nbytes);
1202         else
1203                 return fuse_read_batch_forget(fiq, cs, nbytes);
1204 }
1205
1206 /*
1207  * Read a single request into the userspace filesystem's buffer.  This
1208  * function waits until a request is available, then removes it from
1209  * the pending list and copies request data to userspace buffer.  If
1210  * no reply is needed (FORGET) or request has been aborted or there
1211  * was an error during the copying then it's finished by calling
1212  * request_end().  Otherwise add it to the processing list, and set
1213  * the 'sent' flag.
1214  */
1215 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1216                                 struct fuse_copy_state *cs, size_t nbytes)
1217 {
1218         ssize_t err;
1219         struct fuse_conn *fc = fud->fc;
1220         struct fuse_iqueue *fiq = &fc->iq;
1221         struct fuse_pqueue *fpq = &fud->pq;
1222         struct fuse_req *req;
1223         struct fuse_in *in;
1224         unsigned reqsize;
1225
1226  restart:
1227         spin_lock(&fiq->waitq.lock);
1228         err = -EAGAIN;
1229         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1230             !request_pending(fiq))
1231                 goto err_unlock;
1232
1233         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1234                                 !fiq->connected || request_pending(fiq));
1235         if (err)
1236                 goto err_unlock;
1237
1238         err = -ENODEV;
1239         if (!fiq->connected)
1240                 goto err_unlock;
1241
1242         if (!list_empty(&fiq->interrupts)) {
1243                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1244                                  intr_entry);
1245                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1246         }
1247
1248         if (forget_pending(fiq)) {
1249                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1250                         return fuse_read_forget(fc, fiq, cs, nbytes);
1251
1252                 if (fiq->forget_batch <= -8)
1253                         fiq->forget_batch = 16;
1254         }
1255
1256         req = list_entry(fiq->pending.next, struct fuse_req, list);
1257         clear_bit(FR_PENDING, &req->flags);
1258         list_del_init(&req->list);
1259         spin_unlock(&fiq->waitq.lock);
1260
1261         in = &req->in;
1262         reqsize = in->h.len;
1263         /* If request is too large, reply with an error and restart the read */
1264         if (nbytes < reqsize) {
1265                 req->out.h.error = -EIO;
1266                 /* SETXATTR is special, since it may contain too large data */
1267                 if (in->h.opcode == FUSE_SETXATTR)
1268                         req->out.h.error = -E2BIG;
1269                 request_end(fc, req);
1270                 goto restart;
1271         }
1272         spin_lock(&fpq->lock);
1273         list_add(&req->list, &fpq->io);
1274         spin_unlock(&fpq->lock);
1275         cs->req = req;
1276         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1277         if (!err)
1278                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1279                                      (struct fuse_arg *) in->args, 0);
1280         fuse_copy_finish(cs);
1281         spin_lock(&fpq->lock);
1282         clear_bit(FR_LOCKED, &req->flags);
1283         if (!fpq->connected) {
1284                 err = -ENODEV;
1285                 goto out_end;
1286         }
1287         if (err) {
1288                 req->out.h.error = -EIO;
1289                 goto out_end;
1290         }
1291         if (!test_bit(FR_ISREPLY, &req->flags)) {
1292                 err = reqsize;
1293                 goto out_end;
1294         }
1295         list_move_tail(&req->list, &fpq->processing);
1296         spin_unlock(&fpq->lock);
1297         set_bit(FR_SENT, &req->flags);
1298         /* matches barrier in request_wait_answer() */
1299         smp_mb__after_atomic();
1300         if (test_bit(FR_INTERRUPTED, &req->flags))
1301                 queue_interrupt(fiq, req);
1302
1303         return reqsize;
1304
1305 out_end:
1306         if (!test_bit(FR_PRIVATE, &req->flags))
1307                 list_del_init(&req->list);
1308         spin_unlock(&fpq->lock);
1309         request_end(fc, req);
1310         return err;
1311
1312  err_unlock:
1313         spin_unlock(&fiq->waitq.lock);
1314         return err;
1315 }
1316
1317 static int fuse_dev_open(struct inode *inode, struct file *file)
1318 {
1319         /*
1320          * The fuse device's file's private_data is used to hold
1321          * the fuse_conn(ection) when it is mounted, and is used to
1322          * keep track of whether the file has been mounted already.
1323          */
1324         file->private_data = NULL;
1325         return 0;
1326 }
1327
1328 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1329 {
1330         struct fuse_copy_state cs;
1331         struct file *file = iocb->ki_filp;
1332         struct fuse_dev *fud = fuse_get_dev(file);
1333
1334         if (!fud)
1335                 return -EPERM;
1336
1337         if (!iter_is_iovec(to))
1338                 return -EINVAL;
1339
1340         fuse_copy_init(&cs, 1, to);
1341
1342         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1343 }
1344
1345 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1346                                     struct pipe_inode_info *pipe,
1347                                     size_t len, unsigned int flags)
1348 {
1349         int total, ret;
1350         int page_nr = 0;
1351         struct pipe_buffer *bufs;
1352         struct fuse_copy_state cs;
1353         struct fuse_dev *fud = fuse_get_dev(in);
1354
1355         if (!fud)
1356                 return -EPERM;
1357
1358         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1359         if (!bufs)
1360                 return -ENOMEM;
1361
1362         fuse_copy_init(&cs, 1, NULL);
1363         cs.pipebufs = bufs;
1364         cs.pipe = pipe;
1365         ret = fuse_dev_do_read(fud, in, &cs, len);
1366         if (ret < 0)
1367                 goto out;
1368
1369         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1370                 ret = -EIO;
1371                 goto out;
1372         }
1373
1374         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1375                 /*
1376                  * Need to be careful about this.  Having buf->ops in module
1377                  * code can Oops if the buffer persists after module unload.
1378                  */
1379                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1380                 bufs[page_nr].flags = 0;
1381                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1382                 if (unlikely(ret < 0))
1383                         break;
1384         }
1385         if (total)
1386                 ret = total;
1387 out:
1388         for (; page_nr < cs.nr_segs; page_nr++)
1389                 put_page(bufs[page_nr].page);
1390
1391         kfree(bufs);
1392         return ret;
1393 }
1394
1395 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1396                             struct fuse_copy_state *cs)
1397 {
1398         struct fuse_notify_poll_wakeup_out outarg;
1399         int err = -EINVAL;
1400
1401         if (size != sizeof(outarg))
1402                 goto err;
1403
1404         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1405         if (err)
1406                 goto err;
1407
1408         fuse_copy_finish(cs);
1409         return fuse_notify_poll_wakeup(fc, &outarg);
1410
1411 err:
1412         fuse_copy_finish(cs);
1413         return err;
1414 }
1415
1416 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1417                                    struct fuse_copy_state *cs)
1418 {
1419         struct fuse_notify_inval_inode_out outarg;
1420         int err = -EINVAL;
1421
1422         if (size != sizeof(outarg))
1423                 goto err;
1424
1425         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1426         if (err)
1427                 goto err;
1428         fuse_copy_finish(cs);
1429
1430         down_read(&fc->killsb);
1431         err = -ENOENT;
1432         if (fc->sb) {
1433                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1434                                                outarg.off, outarg.len);
1435         }
1436         up_read(&fc->killsb);
1437         return err;
1438
1439 err:
1440         fuse_copy_finish(cs);
1441         return err;
1442 }
1443
1444 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1445                                    struct fuse_copy_state *cs)
1446 {
1447         struct fuse_notify_inval_entry_out outarg;
1448         int err = -ENOMEM;
1449         char *buf;
1450         struct qstr name;
1451
1452         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1453         if (!buf)
1454                 goto err;
1455
1456         err = -EINVAL;
1457         if (size < sizeof(outarg))
1458                 goto err;
1459
1460         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1461         if (err)
1462                 goto err;
1463
1464         err = -ENAMETOOLONG;
1465         if (outarg.namelen > FUSE_NAME_MAX)
1466                 goto err;
1467
1468         err = -EINVAL;
1469         if (size != sizeof(outarg) + outarg.namelen + 1)
1470                 goto err;
1471
1472         name.name = buf;
1473         name.len = outarg.namelen;
1474         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1475         if (err)
1476                 goto err;
1477         fuse_copy_finish(cs);
1478         buf[outarg.namelen] = 0;
1479
1480         down_read(&fc->killsb);
1481         err = -ENOENT;
1482         if (fc->sb)
1483                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1484         up_read(&fc->killsb);
1485         kfree(buf);
1486         return err;
1487
1488 err:
1489         kfree(buf);
1490         fuse_copy_finish(cs);
1491         return err;
1492 }
1493
1494 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1495                               struct fuse_copy_state *cs)
1496 {
1497         struct fuse_notify_delete_out outarg;
1498         int err = -ENOMEM;
1499         char *buf;
1500         struct qstr name;
1501
1502         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1503         if (!buf)
1504                 goto err;
1505
1506         err = -EINVAL;
1507         if (size < sizeof(outarg))
1508                 goto err;
1509
1510         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1511         if (err)
1512                 goto err;
1513
1514         err = -ENAMETOOLONG;
1515         if (outarg.namelen > FUSE_NAME_MAX)
1516                 goto err;
1517
1518         err = -EINVAL;
1519         if (size != sizeof(outarg) + outarg.namelen + 1)
1520                 goto err;
1521
1522         name.name = buf;
1523         name.len = outarg.namelen;
1524         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1525         if (err)
1526                 goto err;
1527         fuse_copy_finish(cs);
1528         buf[outarg.namelen] = 0;
1529
1530         down_read(&fc->killsb);
1531         err = -ENOENT;
1532         if (fc->sb)
1533                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1534                                                outarg.child, &name);
1535         up_read(&fc->killsb);
1536         kfree(buf);
1537         return err;
1538
1539 err:
1540         kfree(buf);
1541         fuse_copy_finish(cs);
1542         return err;
1543 }
1544
1545 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1546                              struct fuse_copy_state *cs)
1547 {
1548         struct fuse_notify_store_out outarg;
1549         struct inode *inode;
1550         struct address_space *mapping;
1551         u64 nodeid;
1552         int err;
1553         pgoff_t index;
1554         unsigned int offset;
1555         unsigned int num;
1556         loff_t file_size;
1557         loff_t end;
1558
1559         err = -EINVAL;
1560         if (size < sizeof(outarg))
1561                 goto out_finish;
1562
1563         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1564         if (err)
1565                 goto out_finish;
1566
1567         err = -EINVAL;
1568         if (size - sizeof(outarg) != outarg.size)
1569                 goto out_finish;
1570
1571         nodeid = outarg.nodeid;
1572
1573         down_read(&fc->killsb);
1574
1575         err = -ENOENT;
1576         if (!fc->sb)
1577                 goto out_up_killsb;
1578
1579         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1580         if (!inode)
1581                 goto out_up_killsb;
1582
1583         mapping = inode->i_mapping;
1584         index = outarg.offset >> PAGE_SHIFT;
1585         offset = outarg.offset & ~PAGE_MASK;
1586         file_size = i_size_read(inode);
1587         end = outarg.offset + outarg.size;
1588         if (end > file_size) {
1589                 file_size = end;
1590                 fuse_write_update_size(inode, file_size);
1591         }
1592
1593         num = outarg.size;
1594         while (num) {
1595                 struct page *page;
1596                 unsigned int this_num;
1597
1598                 err = -ENOMEM;
1599                 page = find_or_create_page(mapping, index,
1600                                            mapping_gfp_mask(mapping));
1601                 if (!page)
1602                         goto out_iput;
1603
1604                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1605                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1606                 if (!err && offset == 0 &&
1607                     (this_num == PAGE_SIZE || file_size == end))
1608                         SetPageUptodate(page);
1609                 unlock_page(page);
1610                 put_page(page);
1611
1612                 if (err)
1613                         goto out_iput;
1614
1615                 num -= this_num;
1616                 offset = 0;
1617                 index++;
1618         }
1619
1620         err = 0;
1621
1622 out_iput:
1623         iput(inode);
1624 out_up_killsb:
1625         up_read(&fc->killsb);
1626 out_finish:
1627         fuse_copy_finish(cs);
1628         return err;
1629 }
1630
1631 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1632 {
1633         release_pages(req->pages, req->num_pages, false);
1634 }
1635
1636 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1637                          struct fuse_notify_retrieve_out *outarg)
1638 {
1639         int err;
1640         struct address_space *mapping = inode->i_mapping;
1641         struct fuse_req *req;
1642         pgoff_t index;
1643         loff_t file_size;
1644         unsigned int num;
1645         unsigned int offset;
1646         size_t total_len = 0;
1647         int num_pages;
1648
1649         offset = outarg->offset & ~PAGE_MASK;
1650         file_size = i_size_read(inode);
1651
1652         num = outarg->size;
1653         if (outarg->offset > file_size)
1654                 num = 0;
1655         else if (outarg->offset + num > file_size)
1656                 num = file_size - outarg->offset;
1657
1658         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1659         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1660
1661         req = fuse_get_req(fc, num_pages);
1662         if (IS_ERR(req))
1663                 return PTR_ERR(req);
1664
1665         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1666         req->in.h.nodeid = outarg->nodeid;
1667         req->in.numargs = 2;
1668         req->in.argpages = 1;
1669         req->page_descs[0].offset = offset;
1670         req->end = fuse_retrieve_end;
1671
1672         index = outarg->offset >> PAGE_SHIFT;
1673
1674         while (num && req->num_pages < num_pages) {
1675                 struct page *page;
1676                 unsigned int this_num;
1677
1678                 page = find_get_page(mapping, index);
1679                 if (!page)
1680                         break;
1681
1682                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1683                 req->pages[req->num_pages] = page;
1684                 req->page_descs[req->num_pages].length = this_num;
1685                 req->num_pages++;
1686
1687                 offset = 0;
1688                 num -= this_num;
1689                 total_len += this_num;
1690                 index++;
1691         }
1692         req->misc.retrieve_in.offset = outarg->offset;
1693         req->misc.retrieve_in.size = total_len;
1694         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1695         req->in.args[0].value = &req->misc.retrieve_in;
1696         req->in.args[1].size = total_len;
1697
1698         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1699         if (err)
1700                 fuse_retrieve_end(fc, req);
1701
1702         return err;
1703 }
1704
1705 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1706                                 struct fuse_copy_state *cs)
1707 {
1708         struct fuse_notify_retrieve_out outarg;
1709         struct inode *inode;
1710         int err;
1711
1712         err = -EINVAL;
1713         if (size != sizeof(outarg))
1714                 goto copy_finish;
1715
1716         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1717         if (err)
1718                 goto copy_finish;
1719
1720         fuse_copy_finish(cs);
1721
1722         down_read(&fc->killsb);
1723         err = -ENOENT;
1724         if (fc->sb) {
1725                 u64 nodeid = outarg.nodeid;
1726
1727                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1728                 if (inode) {
1729                         err = fuse_retrieve(fc, inode, &outarg);
1730                         iput(inode);
1731                 }
1732         }
1733         up_read(&fc->killsb);
1734
1735         return err;
1736
1737 copy_finish:
1738         fuse_copy_finish(cs);
1739         return err;
1740 }
1741
1742 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1743                        unsigned int size, struct fuse_copy_state *cs)
1744 {
1745         /* Don't try to move pages (yet) */
1746         cs->move_pages = 0;
1747
1748         switch (code) {
1749         case FUSE_NOTIFY_POLL:
1750                 return fuse_notify_poll(fc, size, cs);
1751
1752         case FUSE_NOTIFY_INVAL_INODE:
1753                 return fuse_notify_inval_inode(fc, size, cs);
1754
1755         case FUSE_NOTIFY_INVAL_ENTRY:
1756                 return fuse_notify_inval_entry(fc, size, cs);
1757
1758         case FUSE_NOTIFY_STORE:
1759                 return fuse_notify_store(fc, size, cs);
1760
1761         case FUSE_NOTIFY_RETRIEVE:
1762                 return fuse_notify_retrieve(fc, size, cs);
1763
1764         case FUSE_NOTIFY_DELETE:
1765                 return fuse_notify_delete(fc, size, cs);
1766
1767         default:
1768                 fuse_copy_finish(cs);
1769                 return -EINVAL;
1770         }
1771 }
1772
1773 /* Look up request on processing list by unique ID */
1774 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1775 {
1776         struct fuse_req *req;
1777
1778         list_for_each_entry(req, &fpq->processing, list) {
1779                 if (req->in.h.unique == unique || req->intr_unique == unique)
1780                         return req;
1781         }
1782         return NULL;
1783 }
1784
1785 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1786                          unsigned nbytes)
1787 {
1788         unsigned reqsize = sizeof(struct fuse_out_header);
1789
1790         if (out->h.error)
1791                 return nbytes != reqsize ? -EINVAL : 0;
1792
1793         reqsize += len_args(out->numargs, out->args);
1794
1795         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1796                 return -EINVAL;
1797         else if (reqsize > nbytes) {
1798                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1799                 unsigned diffsize = reqsize - nbytes;
1800                 if (diffsize > lastarg->size)
1801                         return -EINVAL;
1802                 lastarg->size -= diffsize;
1803         }
1804         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1805                               out->page_zeroing);
1806 }
1807
1808 /*
1809  * Write a single reply to a request.  First the header is copied from
1810  * the write buffer.  The request is then searched on the processing
1811  * list by the unique ID found in the header.  If found, then remove
1812  * it from the list and copy the rest of the buffer to the request.
1813  * The request is finished by calling request_end()
1814  */
1815 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1816                                  struct fuse_copy_state *cs, size_t nbytes)
1817 {
1818         int err;
1819         struct fuse_conn *fc = fud->fc;
1820         struct fuse_pqueue *fpq = &fud->pq;
1821         struct fuse_req *req;
1822         struct fuse_out_header oh;
1823
1824         if (nbytes < sizeof(struct fuse_out_header))
1825                 return -EINVAL;
1826
1827         err = fuse_copy_one(cs, &oh, sizeof(oh));
1828         if (err)
1829                 goto err_finish;
1830
1831         err = -EINVAL;
1832         if (oh.len != nbytes)
1833                 goto err_finish;
1834
1835         /*
1836          * Zero oh.unique indicates unsolicited notification message
1837          * and error contains notification code.
1838          */
1839         if (!oh.unique) {
1840                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1841                 return err ? err : nbytes;
1842         }
1843
1844         err = -EINVAL;
1845         if (oh.error <= -1000 || oh.error > 0)
1846                 goto err_finish;
1847
1848         spin_lock(&fpq->lock);
1849         err = -ENOENT;
1850         if (!fpq->connected)
1851                 goto err_unlock_pq;
1852
1853         req = request_find(fpq, oh.unique);
1854         if (!req)
1855                 goto err_unlock_pq;
1856
1857         /* Is it an interrupt reply? */
1858         if (req->intr_unique == oh.unique) {
1859                 spin_unlock(&fpq->lock);
1860
1861                 err = -EINVAL;
1862                 if (nbytes != sizeof(struct fuse_out_header))
1863                         goto err_finish;
1864
1865                 if (oh.error == -ENOSYS)
1866                         fc->no_interrupt = 1;
1867                 else if (oh.error == -EAGAIN)
1868                         queue_interrupt(&fc->iq, req);
1869
1870                 fuse_copy_finish(cs);
1871                 return nbytes;
1872         }
1873
1874         clear_bit(FR_SENT, &req->flags);
1875         list_move(&req->list, &fpq->io);
1876         req->out.h = oh;
1877         set_bit(FR_LOCKED, &req->flags);
1878         spin_unlock(&fpq->lock);
1879         cs->req = req;
1880         if (!req->out.page_replace)
1881                 cs->move_pages = 0;
1882
1883         err = copy_out_args(cs, &req->out, nbytes);
1884         fuse_copy_finish(cs);
1885
1886         spin_lock(&fpq->lock);
1887         clear_bit(FR_LOCKED, &req->flags);
1888         if (!fpq->connected)
1889                 err = -ENOENT;
1890         else if (err)
1891                 req->out.h.error = -EIO;
1892         if (!test_bit(FR_PRIVATE, &req->flags))
1893                 list_del_init(&req->list);
1894         spin_unlock(&fpq->lock);
1895
1896         request_end(fc, req);
1897
1898         return err ? err : nbytes;
1899
1900  err_unlock_pq:
1901         spin_unlock(&fpq->lock);
1902  err_finish:
1903         fuse_copy_finish(cs);
1904         return err;
1905 }
1906
1907 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1908 {
1909         struct fuse_copy_state cs;
1910         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1911
1912         if (!fud)
1913                 return -EPERM;
1914
1915         if (!iter_is_iovec(from))
1916                 return -EINVAL;
1917
1918         fuse_copy_init(&cs, 0, from);
1919
1920         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1921 }
1922
1923 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1924                                      struct file *out, loff_t *ppos,
1925                                      size_t len, unsigned int flags)
1926 {
1927         unsigned nbuf;
1928         unsigned idx;
1929         struct pipe_buffer *bufs;
1930         struct fuse_copy_state cs;
1931         struct fuse_dev *fud;
1932         size_t rem;
1933         ssize_t ret;
1934
1935         fud = fuse_get_dev(out);
1936         if (!fud)
1937                 return -EPERM;
1938
1939         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1940         if (!bufs)
1941                 return -ENOMEM;
1942
1943         pipe_lock(pipe);
1944         nbuf = 0;
1945         rem = 0;
1946         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1947                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1948
1949         ret = -EINVAL;
1950         if (rem < len) {
1951                 pipe_unlock(pipe);
1952                 goto out;
1953         }
1954
1955         rem = len;
1956         while (rem) {
1957                 struct pipe_buffer *ibuf;
1958                 struct pipe_buffer *obuf;
1959
1960                 BUG_ON(nbuf >= pipe->buffers);
1961                 BUG_ON(!pipe->nrbufs);
1962                 ibuf = &pipe->bufs[pipe->curbuf];
1963                 obuf = &bufs[nbuf];
1964
1965                 if (rem >= ibuf->len) {
1966                         *obuf = *ibuf;
1967                         ibuf->ops = NULL;
1968                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1969                         pipe->nrbufs--;
1970                 } else {
1971                         pipe_buf_get(pipe, ibuf);
1972                         *obuf = *ibuf;
1973                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1974                         obuf->len = rem;
1975                         ibuf->offset += obuf->len;
1976                         ibuf->len -= obuf->len;
1977                 }
1978                 nbuf++;
1979                 rem -= obuf->len;
1980         }
1981         pipe_unlock(pipe);
1982
1983         fuse_copy_init(&cs, 0, NULL);
1984         cs.pipebufs = bufs;
1985         cs.nr_segs = nbuf;
1986         cs.pipe = pipe;
1987
1988         if (flags & SPLICE_F_MOVE)
1989                 cs.move_pages = 1;
1990
1991         ret = fuse_dev_do_write(fud, &cs, len);
1992
1993         for (idx = 0; idx < nbuf; idx++)
1994                 pipe_buf_release(pipe, &bufs[idx]);
1995
1996 out:
1997         kfree(bufs);
1998         return ret;
1999 }
2000
2001 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2002 {
2003         unsigned mask = POLLOUT | POLLWRNORM;
2004         struct fuse_iqueue *fiq;
2005         struct fuse_dev *fud = fuse_get_dev(file);
2006
2007         if (!fud)
2008                 return POLLERR;
2009
2010         fiq = &fud->fc->iq;
2011         poll_wait(file, &fiq->waitq, wait);
2012
2013         spin_lock(&fiq->waitq.lock);
2014         if (!fiq->connected)
2015                 mask = POLLERR;
2016         else if (request_pending(fiq))
2017                 mask |= POLLIN | POLLRDNORM;
2018         spin_unlock(&fiq->waitq.lock);
2019
2020         return mask;
2021 }
2022
2023 /*
2024  * Abort all requests on the given list (pending or processing)
2025  *
2026  * This function releases and reacquires fc->lock
2027  */
2028 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2029 {
2030         while (!list_empty(head)) {
2031                 struct fuse_req *req;
2032                 req = list_entry(head->next, struct fuse_req, list);
2033                 req->out.h.error = -ECONNABORTED;
2034                 clear_bit(FR_SENT, &req->flags);
2035                 list_del_init(&req->list);
2036                 request_end(fc, req);
2037         }
2038 }
2039
2040 static void end_polls(struct fuse_conn *fc)
2041 {
2042         struct rb_node *p;
2043
2044         p = rb_first(&fc->polled_files);
2045
2046         while (p) {
2047                 struct fuse_file *ff;
2048                 ff = rb_entry(p, struct fuse_file, polled_node);
2049                 wake_up_interruptible_all(&ff->poll_wait);
2050
2051                 p = rb_next(p);
2052         }
2053 }
2054
2055 /*
2056  * Abort all requests.
2057  *
2058  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2059  * filesystem.
2060  *
2061  * The same effect is usually achievable through killing the filesystem daemon
2062  * and all users of the filesystem.  The exception is the combination of an
2063  * asynchronous request and the tricky deadlock (see
2064  * Documentation/filesystems/fuse.txt).
2065  *
2066  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2067  * requests, they should be finished off immediately.  Locked requests will be
2068  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2069  * requests.  It is possible that some request will finish before we can.  This
2070  * is OK, the request will in that case be removed from the list before we touch
2071  * it.
2072  */
2073 void fuse_abort_conn(struct fuse_conn *fc)
2074 {
2075         struct fuse_iqueue *fiq = &fc->iq;
2076
2077         spin_lock(&fc->lock);
2078         if (fc->connected) {
2079                 struct fuse_dev *fud;
2080                 struct fuse_req *req, *next;
2081                 LIST_HEAD(to_end1);
2082                 LIST_HEAD(to_end2);
2083
2084                 fc->connected = 0;
2085                 fc->blocked = 0;
2086                 fuse_set_initialized(fc);
2087                 list_for_each_entry(fud, &fc->devices, entry) {
2088                         struct fuse_pqueue *fpq = &fud->pq;
2089
2090                         spin_lock(&fpq->lock);
2091                         fpq->connected = 0;
2092                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2093                                 req->out.h.error = -ECONNABORTED;
2094                                 spin_lock(&req->waitq.lock);
2095                                 set_bit(FR_ABORTED, &req->flags);
2096                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2097                                         set_bit(FR_PRIVATE, &req->flags);
2098                                         list_move(&req->list, &to_end1);
2099                                 }
2100                                 spin_unlock(&req->waitq.lock);
2101                         }
2102                         list_splice_init(&fpq->processing, &to_end2);
2103                         spin_unlock(&fpq->lock);
2104                 }
2105                 fc->max_background = UINT_MAX;
2106                 flush_bg_queue(fc);
2107
2108                 spin_lock(&fiq->waitq.lock);
2109                 fiq->connected = 0;
2110                 list_splice_init(&fiq->pending, &to_end2);
2111                 list_for_each_entry(req, &to_end2, list)
2112                         clear_bit(FR_PENDING, &req->flags);
2113                 while (forget_pending(fiq))
2114                         kfree(dequeue_forget(fiq, 1, NULL));
2115                 wake_up_all_locked(&fiq->waitq);
2116                 spin_unlock(&fiq->waitq.lock);
2117                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2118                 end_polls(fc);
2119                 wake_up_all(&fc->blocked_waitq);
2120                 spin_unlock(&fc->lock);
2121
2122                 while (!list_empty(&to_end1)) {
2123                         req = list_first_entry(&to_end1, struct fuse_req, list);
2124                         __fuse_get_request(req);
2125                         list_del_init(&req->list);
2126                         request_end(fc, req);
2127                 }
2128                 end_requests(fc, &to_end2);
2129         } else {
2130                 spin_unlock(&fc->lock);
2131         }
2132 }
2133 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2134
2135 int fuse_dev_release(struct inode *inode, struct file *file)
2136 {
2137         struct fuse_dev *fud = fuse_get_dev(file);
2138
2139         if (fud) {
2140                 struct fuse_conn *fc = fud->fc;
2141                 struct fuse_pqueue *fpq = &fud->pq;
2142
2143                 WARN_ON(!list_empty(&fpq->io));
2144                 end_requests(fc, &fpq->processing);
2145                 /* Are we the last open device? */
2146                 if (atomic_dec_and_test(&fc->dev_count)) {
2147                         WARN_ON(fc->iq.fasync != NULL);
2148                         fuse_abort_conn(fc);
2149                 }
2150                 fuse_dev_free(fud);
2151         }
2152         return 0;
2153 }
2154 EXPORT_SYMBOL_GPL(fuse_dev_release);
2155
2156 static int fuse_dev_fasync(int fd, struct file *file, int on)
2157 {
2158         struct fuse_dev *fud = fuse_get_dev(file);
2159
2160         if (!fud)
2161                 return -EPERM;
2162
2163         /* No locking - fasync_helper does its own locking */
2164         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2165 }
2166
2167 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2168 {
2169         struct fuse_dev *fud;
2170
2171         if (new->private_data)
2172                 return -EINVAL;
2173
2174         fud = fuse_dev_alloc(fc);
2175         if (!fud)
2176                 return -ENOMEM;
2177
2178         new->private_data = fud;
2179         atomic_inc(&fc->dev_count);
2180
2181         return 0;
2182 }
2183
2184 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2185                            unsigned long arg)
2186 {
2187         int err = -ENOTTY;
2188
2189         if (cmd == FUSE_DEV_IOC_CLONE) {
2190                 int oldfd;
2191
2192                 err = -EFAULT;
2193                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2194                         struct file *old = fget(oldfd);
2195
2196                         err = -EINVAL;
2197                         if (old) {
2198                                 struct fuse_dev *fud = NULL;
2199
2200                                 /*
2201                                  * Check against file->f_op because CUSE
2202                                  * uses the same ioctl handler.
2203                                  */
2204                                 if (old->f_op == file->f_op &&
2205                                     old->f_cred->user_ns == file->f_cred->user_ns)
2206                                         fud = fuse_get_dev(old);
2207
2208                                 if (fud) {
2209                                         mutex_lock(&fuse_mutex);
2210                                         err = fuse_device_clone(fud->fc, file);
2211                                         mutex_unlock(&fuse_mutex);
2212                                 }
2213                                 fput(old);
2214                         }
2215                 }
2216         }
2217         return err;
2218 }
2219
2220 const struct file_operations fuse_dev_operations = {
2221         .owner          = THIS_MODULE,
2222         .open           = fuse_dev_open,
2223         .llseek         = no_llseek,
2224         .read_iter      = fuse_dev_read,
2225         .splice_read    = fuse_dev_splice_read,
2226         .write_iter     = fuse_dev_write,
2227         .splice_write   = fuse_dev_splice_write,
2228         .poll           = fuse_dev_poll,
2229         .release        = fuse_dev_release,
2230         .fasync         = fuse_dev_fasync,
2231         .unlocked_ioctl = fuse_dev_ioctl,
2232         .compat_ioctl   = fuse_dev_ioctl,
2233 };
2234 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2235
2236 static struct miscdevice fuse_miscdevice = {
2237         .minor = FUSE_MINOR,
2238         .name  = "fuse",
2239         .fops = &fuse_dev_operations,
2240 };
2241
2242 int __init fuse_dev_init(void)
2243 {
2244         int err = -ENOMEM;
2245         fuse_req_cachep = kmem_cache_create("fuse_request",
2246                                             sizeof(struct fuse_req),
2247                                             0, 0, NULL);
2248         if (!fuse_req_cachep)
2249                 goto out;
2250
2251         err = misc_register(&fuse_miscdevice);
2252         if (err)
2253                 goto out_cache_clean;
2254
2255         return 0;
2256
2257  out_cache_clean:
2258         kmem_cache_destroy(fuse_req_cachep);
2259  out:
2260         return err;
2261 }
2262
2263 void fuse_dev_cleanup(void)
2264 {
2265         misc_deregister(&fuse_miscdevice);
2266         kmem_cache_destroy(fuse_req_cachep);
2267 }