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/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/pipe_fs_i.h>
  20 #include <linux/swap.h>
  21 #include <linux/splice.h>
  22
  23 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  24 MODULE_ALIAS("devname:fuse");
  25
  26 static struct kmem_cache *fuse_req_cachep;
  27
  28 static struct fuse_dev *fuse_get_dev(struct file *file)
  29 {
  30         /*
  31          * Lockless access is OK, because file->private data is set
  32          * once during mount and is valid until the file is released.
  33          */
  34         return ACCESS_ONCE(file->private_data);
  35 }
  36
  37 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  38                               struct fuse_page_desc *page_descs,
  39                               unsigned npages)
  40 {
  41         memset(req, 0, sizeof(*req));
  42         memset(pages, 0, sizeof(*pages) * npages);
  43         memset(page_descs, 0, sizeof(*page_descs) * npages);
  44         INIT_LIST_HEAD(&req->list);
  45         INIT_LIST_HEAD(&req->intr_entry);
  46         init_waitqueue_head(&req->waitq);
  47         atomic_set(&req->count, 1);
  48         req->pages = pages;
  49         req->page_descs = page_descs;
  50         req->max_pages = npages;
  51         __set_bit(FR_PENDING, &req->flags);
  52 }
  53
  54 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  55 {
  56         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, flags);
  57         if (req) {
  58                 struct page **pages;
  59                 struct fuse_page_desc *page_descs;
  60
  61                 if (npages <= FUSE_REQ_INLINE_PAGES) {
  62                         pages = req->inline_pages;
  63                         page_descs = req->inline_page_descs;
  64                 } else {
  65                         pages = kmalloc(sizeof(struct page *) * npages, flags);
  66                         page_descs = kmalloc(sizeof(struct fuse_page_desc) *
  67                                              npages, flags);
  68                 }
  69
  70                 if (!pages || !page_descs) {
  71                         kfree(pages);
  72                         kfree(page_descs);
  73                         kmem_cache_free(fuse_req_cachep, req);
  74                         return NULL;
  75                 }
  76
  77                 fuse_request_init(req, pages, page_descs, npages);
  78         }
  79         return req;
  80 }
  81
  82 struct fuse_req *fuse_request_alloc(unsigned npages)
  83 {
  84         return __fuse_request_alloc(npages, GFP_KERNEL);
  85 }
  86 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  87
  88 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  89 {
  90         return __fuse_request_alloc(npages, GFP_NOFS);
  91 }
  92
  93 void fuse_request_free(struct fuse_req *req)
  94 {
  95         if (req->pages != req->inline_pages) {
  96                 kfree(req->pages);
  97                 kfree(req->page_descs);
  98         }
  99         kmem_cache_free(fuse_req_cachep, req);
 100 }
 101
 102 void __fuse_get_request(struct fuse_req *req)
 103 {
 104         atomic_inc(&req->count);
 105 }
 106
 107 /* Must be called with > 1 refcount */
 108 static void __fuse_put_request(struct fuse_req *req)
 109 {
 110         BUG_ON(atomic_read(&req->count) < 2);
 111         atomic_dec(&req->count);
 112 }
 113
 114 static void fuse_req_init_context(struct fuse_req *req)
 115 {
 116         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 117         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 118         req->in.h.pid = current->pid;
 119 }
 120
 121 void fuse_set_initialized(struct fuse_conn *fc)
 122 {
 123         /* Make sure stores before this are seen on another CPU */
 124         smp_wmb();
 125         fc->initialized = 1;
 126 }
 127
 128 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 129 {
 130         return !fc->initialized || (for_background && fc->blocked);
 131 }
 132
 133 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 134                                        bool for_background)
 135 {
 136         struct fuse_req *req;
 137         int err;
 138         atomic_inc(&fc->num_waiting);
 139
 140         if (fuse_block_alloc(fc, for_background)) {
 141                 err = -EINTR;
 142                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 143                                 !fuse_block_alloc(fc, for_background)))
 144                         goto out;
 145         }
 146         /* Matches smp_wmb() in fuse_set_initialized() */
 147         smp_rmb();
 148
 149         err = -ENOTCONN;
 150         if (!fc->connected)
 151                 goto out;
 152
 153         err = -ECONNREFUSED;
 154         if (fc->conn_error)
 155                 goto out;
 156
 157         req = fuse_request_alloc(npages);
 158         err = -ENOMEM;
 159         if (!req) {
 160                 if (for_background)
 161                         wake_up(&fc->blocked_waitq);
 162                 goto out;
 163         }
 164
 165         fuse_req_init_context(req);
 166         __set_bit(FR_WAITING, &req->flags);
 167         if (for_background)
 168                 __set_bit(FR_BACKGROUND, &req->flags);
 169
 170         return req;
 171
 172  out:
 173         atomic_dec(&fc->num_waiting);
 174         return ERR_PTR(err);
 175 }
 176
 177 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 178 {
 179         return __fuse_get_req(fc, npages, false);
 180 }
 181 EXPORT_SYMBOL_GPL(fuse_get_req);
 182
 183 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 184                                              unsigned npages)
 185 {
 186         return __fuse_get_req(fc, npages, true);
 187 }
 188 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 189
 190 /*
 191  * Return request in fuse_file->reserved_req.  However that may
 192  * currently be in use.  If that is the case, wait for it to become
 193  * available.
 194  */
 195 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 196                                          struct file *file)
 197 {
 198         struct fuse_req *req = NULL;
 199         struct fuse_file *ff = file->private_data;
 200
 201         do {
 202                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 203                 spin_lock(&fc->lock);
 204                 if (ff->reserved_req) {
 205                         req = ff->reserved_req;
 206                         ff->reserved_req = NULL;
 207                         req->stolen_file = get_file(file);
 208                 }
 209                 spin_unlock(&fc->lock);
 210         } while (!req);
 211
 212         return req;
 213 }
 214
 215 /*
 216  * Put stolen request back into fuse_file->reserved_req
 217  */
 218 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 219 {
 220         struct file *file = req->stolen_file;
 221         struct fuse_file *ff = file->private_data;
 222
 223         spin_lock(&fc->lock);
 224         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 225         BUG_ON(ff->reserved_req);
 226         ff->reserved_req = req;
 227         wake_up_all(&fc->reserved_req_waitq);
 228         spin_unlock(&fc->lock);
 229         fput(file);
 230 }
 231
 232 /*
 233  * Gets a requests for a file operation, always succeeds
 234  *
 235  * This is used for sending the FLUSH request, which must get to
 236  * userspace, due to POSIX locks which may need to be unlocked.
 237  *
 238  * If allocation fails due to OOM, use the reserved request in
 239  * fuse_file.
 240  *
 241  * This is very unlikely to deadlock accidentally, since the
 242  * filesystem should not have it's own file open.  If deadlock is
 243  * intentional, it can still be broken by "aborting" the filesystem.
 244  */
 245 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 246                                              struct file *file)
 247 {
 248         struct fuse_req *req;
 249
 250         atomic_inc(&fc->num_waiting);
 251         wait_event(fc->blocked_waitq, fc->initialized);
 252         /* Matches smp_wmb() in fuse_set_initialized() */
 253         smp_rmb();
 254         req = fuse_request_alloc(0);
 255         if (!req)
 256                 req = get_reserved_req(fc, file);
 257
 258         fuse_req_init_context(req);
 259         __set_bit(FR_WAITING, &req->flags);
 260         __clear_bit(FR_BACKGROUND, &req->flags);
 261         return req;
 262 }
 263
 264 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 265 {
 266         if (atomic_dec_and_test(&req->count)) {
 267                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 268                         /*
 269                          * We get here in the unlikely case that a background
 270                          * request was allocated but not sent
 271                          */
 272                         spin_lock(&fc->lock);
 273                         if (!fc->blocked)
 274                                 wake_up(&fc->blocked_waitq);
 275                         spin_unlock(&fc->lock);
 276                 }
 277
 278                 if (test_bit(FR_WAITING, &req->flags)) {
 279                         __clear_bit(FR_WAITING, &req->flags);
 280                         atomic_dec(&fc->num_waiting);
 281                 }
 282
 283                 if (req->stolen_file)
 284                         put_reserved_req(fc, req);
 285                 else
 286                         fuse_request_free(req);
 287         }
 288 }
 289 EXPORT_SYMBOL_GPL(fuse_put_request);
 290
 291 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 292 {
 293         unsigned nbytes = 0;
 294         unsigned i;
 295
 296         for (i = 0; i < numargs; i++)
 297                 nbytes += args[i].size;
 298
 299         return nbytes;
 300 }
 301
 302 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 303 {
 304         return ++fiq->reqctr;
 305 }
 306
 307 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 308 {
 309         req->in.h.len = sizeof(struct fuse_in_header) +
 310                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 311         list_add_tail(&req->list, &fiq->pending);
 312         wake_up_locked(&fiq->waitq);
 313         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 314 }
 315
 316 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 317                        u64 nodeid, u64 nlookup)
 318 {
 319         struct fuse_iqueue *fiq = &fc->iq;
 320
 321         forget->forget_one.nodeid = nodeid;
 322         forget->forget_one.nlookup = nlookup;
 323
 324         spin_lock(&fiq->waitq.lock);
 325         if (fiq->connected) {
 326                 fiq->forget_list_tail->next = forget;
 327                 fiq->forget_list_tail = forget;
 328                 wake_up_locked(&fiq->waitq);
 329                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 330         } else {
 331                 kfree(forget);
 332         }
 333         spin_unlock(&fiq->waitq.lock);
 334 }
 335
 336 static void flush_bg_queue(struct fuse_conn *fc)
 337 {
 338         while (fc->active_background < fc->max_background &&
 339                !list_empty(&fc->bg_queue)) {
 340                 struct fuse_req *req;
 341                 struct fuse_iqueue *fiq = &fc->iq;
 342
 343                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 344                 list_del(&req->list);
 345                 fc->active_background++;
 346                 spin_lock(&fiq->waitq.lock);
 347                 req->in.h.unique = fuse_get_unique(fiq);
 348                 queue_request(fiq, req);
 349                 spin_unlock(&fiq->waitq.lock);
 350         }
 351 }
 352
 353 /*
 354  * This function is called when a request is finished.  Either a reply
 355  * has arrived or it was aborted (and not yet sent) or some error
 356  * occurred during communication with userspace, or the device file
 357  * was closed.  The requester thread is woken up (if still waiting),
 358  * the 'end' callback is called if given, else the reference to the
 359  * request is released
 360  */
 361 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 362 {
 363         struct fuse_iqueue *fiq = &fc->iq;
 364
 365         if (test_and_set_bit(FR_FINISHED, &req->flags))
 366                 return;
 367
 368         spin_lock(&fiq->waitq.lock);
 369         list_del_init(&req->intr_entry);
 370         spin_unlock(&fiq->waitq.lock);
 371         WARN_ON(test_bit(FR_PENDING, &req->flags));
 372         WARN_ON(test_bit(FR_SENT, &req->flags));
 373         if (test_bit(FR_BACKGROUND, &req->flags)) {
 374                 spin_lock(&fc->lock);
 375                 clear_bit(FR_BACKGROUND, &req->flags);
 376                 if (fc->num_background == fc->max_background)
 377                         fc->blocked = 0;
 378
 379                 /* Wake up next waiter, if any */
 380                 if (!fc->blocked && waitqueue_active(&fc->blocked_waitq))
 381                         wake_up(&fc->blocked_waitq);
 382
 383                 if (fc->num_background == fc->congestion_threshold &&
 384                     fc->connected && fc->bdi_initialized) {
 385                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 386                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 387                 }
 388                 fc->num_background--;
 389                 fc->active_background--;
 390                 flush_bg_queue(fc);
 391                 spin_unlock(&fc->lock);
 392         }
 393         wake_up(&req->waitq);
 394         if (req->end)
 395                 req->end(fc, req);
 396         fuse_put_request(fc, req);
 397 }
 398
 399 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 400 {
 401         spin_lock(&fiq->waitq.lock);
 402         if (test_bit(FR_FINISHED, &req->flags)) {
 403                 spin_unlock(&fiq->waitq.lock);
 404                 return;
 405         }
 406         if (list_empty(&req->intr_entry)) {
 407                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 408                 wake_up_locked(&fiq->waitq);
 409         }
 410         spin_unlock(&fiq->waitq.lock);
 411         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 412 }
 413
 414 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 415 {
 416         struct fuse_iqueue *fiq = &fc->iq;
 417         int err;
 418
 419         if (!fc->no_interrupt) {
 420                 /* Any signal may interrupt this */
 421                 err = wait_event_interruptible(req->waitq,
 422                                         test_bit(FR_FINISHED, &req->flags));
 423                 if (!err)
 424                         return;
 425
 426                 set_bit(FR_INTERRUPTED, &req->flags);
 427                 /* matches barrier in fuse_dev_do_read() */
 428                 smp_mb__after_atomic();
 429                 if (test_bit(FR_SENT, &req->flags))
 430                         queue_interrupt(fiq, req);
 431         }
 432
 433         if (!test_bit(FR_FORCE, &req->flags)) {
 434                 /* Only fatal signals may interrupt this */
 435                 err = wait_event_killable(req->waitq,
 436                                         test_bit(FR_FINISHED, &req->flags));
 437                 if (!err)
 438                         return;
 439
 440                 spin_lock(&fiq->waitq.lock);
 441                 /* Request is not yet in userspace, bail out */
 442                 if (test_bit(FR_PENDING, &req->flags)) {
 443                         list_del(&req->list);
 444                         spin_unlock(&fiq->waitq.lock);
 445                         __fuse_put_request(req);
 446                         req->out.h.error = -EINTR;
 447                         return;
 448                 }
 449                 spin_unlock(&fiq->waitq.lock);
 450         }
 451
 452         /*
 453          * Either request is already in userspace, or it was forced.
 454          * Wait it out.
 455          */
 456         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 457 }
 458
 459 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 460 {
 461         struct fuse_iqueue *fiq = &fc->iq;
 462
 463         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 464         spin_lock(&fiq->waitq.lock);
 465         if (!fiq->connected) {
 466                 spin_unlock(&fiq->waitq.lock);
 467                 req->out.h.error = -ENOTCONN;
 468         } else {
 469                 req->in.h.unique = fuse_get_unique(fiq);
 470                 queue_request(fiq, req);
 471                 /* acquire extra reference, since request is still needed
 472                    after request_end() */
 473                 __fuse_get_request(req);
 474                 spin_unlock(&fiq->waitq.lock);
 475
 476                 request_wait_answer(fc, req);
 477                 /* Pairs with smp_wmb() in request_end() */
 478                 smp_rmb();
 479         }
 480 }
 481
 482 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 483 {
 484         __set_bit(FR_ISREPLY, &req->flags);
 485         if (!test_bit(FR_WAITING, &req->flags)) {
 486                 __set_bit(FR_WAITING, &req->flags);
 487                 atomic_inc(&fc->num_waiting);
 488         }
 489         __fuse_request_send(fc, req);
 490 }
 491 EXPORT_SYMBOL_GPL(fuse_request_send);
 492
 493 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 494 {
 495         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 496                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 497
 498         if (fc->minor < 9) {
 499                 switch (args->in.h.opcode) {
 500                 case FUSE_LOOKUP:
 501                 case FUSE_CREATE:
 502                 case FUSE_MKNOD:
 503                 case FUSE_MKDIR:
 504                 case FUSE_SYMLINK:
 505                 case FUSE_LINK:
 506                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 507                         break;
 508                 case FUSE_GETATTR:
 509                 case FUSE_SETATTR:
 510                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 511                         break;
 512                 }
 513         }
 514         if (fc->minor < 12) {
 515                 switch (args->in.h.opcode) {
 516                 case FUSE_CREATE:
 517                         args->in.args[0].size = sizeof(struct fuse_open_in);
 518                         break;
 519                 case FUSE_MKNOD:
 520                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 521                         break;
 522                 }
 523         }
 524 }
 525
 526 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 527 {
 528         struct fuse_req *req;
 529         ssize_t ret;
 530
 531         req = fuse_get_req(fc, 0);
 532         if (IS_ERR(req))
 533                 return PTR_ERR(req);
 534
 535         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 536         fuse_adjust_compat(fc, args);
 537
 538         req->in.h.opcode = args->in.h.opcode;
 539         req->in.h.nodeid = args->in.h.nodeid;
 540         req->in.numargs = args->in.numargs;
 541         memcpy(req->in.args, args->in.args,
 542                args->in.numargs * sizeof(struct fuse_in_arg));
 543         req->out.argvar = args->out.argvar;
 544         req->out.numargs = args->out.numargs;
 545         memcpy(req->out.args, args->out.args,
 546                args->out.numargs * sizeof(struct fuse_arg));
 547         fuse_request_send(fc, req);
 548         ret = req->out.h.error;
 549         if (!ret && args->out.argvar) {
 550                 BUG_ON(args->out.numargs != 1);
 551                 ret = req->out.args[0].size;
 552         }
 553         fuse_put_request(fc, req);
 554
 555         return ret;
 556 }
 557
 558 /*
 559  * Called under fc->lock
 560  *
 561  * fc->connected must have been checked previously
 562  */
 563 void fuse_request_send_background_locked(struct fuse_conn *fc,
 564                                          struct fuse_req *req)
 565 {
 566         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 567         if (!test_bit(FR_WAITING, &req->flags)) {
 568                 __set_bit(FR_WAITING, &req->flags);
 569                 atomic_inc(&fc->num_waiting);
 570         }
 571         __set_bit(FR_ISREPLY, &req->flags);
 572         fc->num_background++;
 573         if (fc->num_background == fc->max_background)
 574                 fc->blocked = 1;
 575         if (fc->num_background == fc->congestion_threshold &&
 576             fc->bdi_initialized) {
 577                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 578                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 579         }
 580         list_add_tail(&req->list, &fc->bg_queue);
 581         flush_bg_queue(fc);
 582 }
 583
 584 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 585 {
 586         BUG_ON(!req->end);
 587         spin_lock(&fc->lock);
 588         if (fc->connected) {
 589                 fuse_request_send_background_locked(fc, req);
 590                 spin_unlock(&fc->lock);
 591         } else {
 592                 spin_unlock(&fc->lock);
 593                 req->out.h.error = -ENOTCONN;
 594                 req->end(fc, req);
 595                 fuse_put_request(fc, req);
 596         }
 597 }
 598 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 599
 600 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 601                                           struct fuse_req *req, u64 unique)
 602 {
 603         int err = -ENODEV;
 604         struct fuse_iqueue *fiq = &fc->iq;
 605
 606         __clear_bit(FR_ISREPLY, &req->flags);
 607         req->in.h.unique = unique;
 608         spin_lock(&fiq->waitq.lock);
 609         if (fiq->connected) {
 610                 queue_request(fiq, req);
 611                 err = 0;
 612         }
 613         spin_unlock(&fiq->waitq.lock);
 614
 615         return err;
 616 }
 617
 618 void fuse_force_forget(struct file *file, u64 nodeid)
 619 {
 620         struct inode *inode = file_inode(file);
 621         struct fuse_conn *fc = get_fuse_conn(inode);
 622         struct fuse_req *req;
 623         struct fuse_forget_in inarg;
 624
 625         memset(&inarg, 0, sizeof(inarg));
 626         inarg.nlookup = 1;
 627         req = fuse_get_req_nofail_nopages(fc, file);
 628         req->in.h.opcode = FUSE_FORGET;
 629         req->in.h.nodeid = nodeid;
 630         req->in.numargs = 1;
 631         req->in.args[0].size = sizeof(inarg);
 632         req->in.args[0].value = &inarg;
 633         __clear_bit(FR_ISREPLY, &req->flags);
 634         __fuse_request_send(fc, req);
 635         /* ignore errors */
 636         fuse_put_request(fc, req);
 637 }
 638
 639 /*
 640  * Lock the request.  Up to the next unlock_request() there mustn't be
 641  * anything that could cause a page-fault.  If the request was already
 642  * aborted bail out.
 643  */
 644 static int lock_request(struct fuse_req *req)
 645 {
 646         int err = 0;
 647         if (req) {
 648                 spin_lock(&req->waitq.lock);
 649                 if (test_bit(FR_ABORTED, &req->flags))
 650                         err = -ENOENT;
 651                 else
 652                         set_bit(FR_LOCKED, &req->flags);
 653                 spin_unlock(&req->waitq.lock);
 654         }
 655         return err;
 656 }
 657
 658 /*
 659  * Unlock request.  If it was aborted while locked, caller is responsible
 660  * for unlocking and ending the request.
 661  */
 662 static int unlock_request(struct fuse_req *req)
 663 {
 664         int err = 0;
 665         if (req) {
 666                 spin_lock(&req->waitq.lock);
 667                 if (test_bit(FR_ABORTED, &req->flags))
 668                         err = -ENOENT;
 669                 else
 670                         clear_bit(FR_LOCKED, &req->flags);
 671                 spin_unlock(&req->waitq.lock);
 672         }
 673         return err;
 674 }
 675
 676 struct fuse_copy_state {
 677         int write;
 678         struct fuse_req *req;
 679         struct iov_iter *iter;
 680         struct pipe_buffer *pipebufs;
 681         struct pipe_buffer *currbuf;
 682         struct pipe_inode_info *pipe;
 683         unsigned long nr_segs;
 684         struct page *pg;
 685         unsigned len;
 686         unsigned offset;
 687         unsigned move_pages:1;
 688 };
 689
 690 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 691                            struct iov_iter *iter)
 692 {
 693         memset(cs, 0, sizeof(*cs));
 694         cs->write = write;
 695         cs->iter = iter;
 696 }
 697
 698 /* Unmap and put previous page of userspace buffer */
 699 static void fuse_copy_finish(struct fuse_copy_state *cs)
 700 {
 701         if (cs->currbuf) {
 702                 struct pipe_buffer *buf = cs->currbuf;
 703
 704                 if (cs->write)
 705                         buf->len = PAGE_SIZE - cs->len;
 706                 cs->currbuf = NULL;
 707         } else if (cs->pg) {
 708                 if (cs->write) {
 709                         flush_dcache_page(cs->pg);
 710                         set_page_dirty_lock(cs->pg);
 711                 }
 712                 put_page(cs->pg);
 713         }
 714         cs->pg = NULL;
 715 }
 716
 717 /*
 718  * Get another pagefull of userspace buffer, and map it to kernel
 719  * address space, and lock request
 720  */
 721 static int fuse_copy_fill(struct fuse_copy_state *cs)
 722 {
 723         struct page *page;
 724         int err;
 725
 726         err = unlock_request(cs->req);
 727         if (err)
 728                 return err;
 729
 730         fuse_copy_finish(cs);
 731         if (cs->pipebufs) {
 732                 struct pipe_buffer *buf = cs->pipebufs;
 733
 734                 if (!cs->write) {
 735                         err = pipe_buf_confirm(cs->pipe, buf);
 736                         if (err)
 737                                 return err;
 738
 739                         BUG_ON(!cs->nr_segs);
 740                         cs->currbuf = buf;
 741                         cs->pg = buf->page;
 742                         cs->offset = buf->offset;
 743                         cs->len = buf->len;
 744                         cs->pipebufs++;
 745                         cs->nr_segs--;
 746                 } else {
 747                         if (cs->nr_segs == cs->pipe->buffers)
 748                                 return -EIO;
 749
 750                         page = alloc_page(GFP_HIGHUSER);
 751                         if (!page)
 752                                 return -ENOMEM;
 753
 754                         buf->page = page;
 755                         buf->offset = 0;
 756                         buf->len = 0;
 757
 758                         cs->currbuf = buf;
 759                         cs->pg = page;
 760                         cs->offset = 0;
 761                         cs->len = PAGE_SIZE;
 762                         cs->pipebufs++;
 763                         cs->nr_segs++;
 764                 }
 765         } else {
 766                 size_t off;
 767                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 768                 if (err < 0)
 769                         return err;
 770                 BUG_ON(!err);
 771                 cs->len = err;
 772                 cs->offset = off;
 773                 cs->pg = page;
 774                 iov_iter_advance(cs->iter, err);
 775         }
 776
 777         return lock_request(cs->req);
 778 }
 779
 780 /* Do as much copy to/from userspace buffer as we can */
 781 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 782 {
 783         unsigned ncpy = min(*size, cs->len);
 784         if (val) {
 785                 void *pgaddr = kmap_atomic(cs->pg);
 786                 void *buf = pgaddr + cs->offset;
 787
 788                 if (cs->write)
 789                         memcpy(buf, *val, ncpy);
 790                 else
 791                         memcpy(*val, buf, ncpy);
 792
 793                 kunmap_atomic(pgaddr);
 794                 *val += ncpy;
 795         }
 796         *size -= ncpy;
 797         cs->len -= ncpy;
 798         cs->offset += ncpy;
 799         return ncpy;
 800 }
 801
 802 static int fuse_check_page(struct page *page)
 803 {
 804         if (page_mapcount(page) ||
 805             page->mapping != NULL ||
 806             page_count(page) != 1 ||
 807             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 808              ~(1 << PG_locked |
 809                1 << PG_referenced |
 810                1 << PG_uptodate |
 811                1 << PG_lru |
 812                1 << PG_active |
 813                1 << PG_reclaim))) {
 814                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 815                 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);
 816                 return 1;
 817         }
 818         return 0;
 819 }
 820
 821 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 822 {
 823         int err;
 824         struct page *oldpage = *pagep;
 825         struct page *newpage;
 826         struct pipe_buffer *buf = cs->pipebufs;
 827
 828         err = unlock_request(cs->req);
 829         if (err)
 830                 return err;
 831
 832         fuse_copy_finish(cs);
 833
 834         err = pipe_buf_confirm(cs->pipe, buf);
 835         if (err)
 836                 return err;
 837
 838         BUG_ON(!cs->nr_segs);
 839         cs->currbuf = buf;
 840         cs->len = buf->len;
 841         cs->pipebufs++;
 842         cs->nr_segs--;
 843
 844         if (cs->len != PAGE_SIZE)
 845                 goto out_fallback;
 846
 847         if (pipe_buf_steal(cs->pipe, buf) != 0)
 848                 goto out_fallback;
 849
 850         newpage = buf->page;
 851
 852         if (!PageUptodate(newpage))
 853                 SetPageUptodate(newpage);
 854
 855         ClearPageMappedToDisk(newpage);
 856
 857         if (fuse_check_page(newpage) != 0)
 858                 goto out_fallback_unlock;
 859
 860         /*
 861          * This is a new and locked page, it shouldn't be mapped or
 862          * have any special flags on it
 863          */
 864         if (WARN_ON(page_mapped(oldpage)))
 865                 goto out_fallback_unlock;
 866         if (WARN_ON(page_has_private(oldpage)))
 867                 goto out_fallback_unlock;
 868         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 869                 goto out_fallback_unlock;
 870         if (WARN_ON(PageMlocked(oldpage)))
 871                 goto out_fallback_unlock;
 872
 873         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 874         if (err) {
 875                 unlock_page(newpage);
 876                 return err;
 877         }
 878
 879         get_page(newpage);
 880
 881         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 882                 lru_cache_add_file(newpage);
 883
 884         err = 0;
 885         spin_lock(&cs->req->waitq.lock);
 886         if (test_bit(FR_ABORTED, &cs->req->flags))
 887                 err = -ENOENT;
 888         else
 889                 *pagep = newpage;
 890         spin_unlock(&cs->req->waitq.lock);
 891
 892         if (err) {
 893                 unlock_page(newpage);
 894                 put_page(newpage);
 895                 return err;
 896         }
 897
 898         unlock_page(oldpage);
 899         put_page(oldpage);
 900         cs->len = 0;
 901
 902         return 0;
 903
 904 out_fallback_unlock:
 905         unlock_page(newpage);
 906 out_fallback:
 907         cs->pg = buf->page;
 908         cs->offset = buf->offset;
 909
 910         err = lock_request(cs->req);
 911         if (err)
 912                 return err;
 913
 914         return 1;
 915 }
 916
 917 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 918                          unsigned offset, unsigned count)
 919 {
 920         struct pipe_buffer *buf;
 921         int err;
 922
 923         if (cs->nr_segs == cs->pipe->buffers)
 924                 return -EIO;
 925
 926         err = unlock_request(cs->req);
 927         if (err)
 928                 return err;
 929
 930         fuse_copy_finish(cs);
 931
 932         buf = cs->pipebufs;
 933         get_page(page);
 934         buf->page = page;
 935         buf->offset = offset;
 936         buf->len = count;
 937
 938         cs->pipebufs++;
 939         cs->nr_segs++;
 940         cs->len = 0;
 941
 942         return 0;
 943 }
 944
 945 /*
 946  * Copy a page in the request to/from the userspace buffer.  Must be
 947  * done atomically
 948  */
 949 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 950                           unsigned offset, unsigned count, int zeroing)
 951 {
 952         int err;
 953         struct page *page = *pagep;
 954
 955         if (page && zeroing && count < PAGE_SIZE)
 956                 clear_highpage(page);
 957
 958         while (count) {
 959                 if (cs->write && cs->pipebufs && page) {
 960                         return fuse_ref_page(cs, page, offset, count);
 961                 } else if (!cs->len) {
 962                         if (cs->move_pages && page &&
 963                             offset == 0 && count == PAGE_SIZE) {
 964                                 err = fuse_try_move_page(cs, pagep);
 965                                 if (err <= 0)
 966                                         return err;
 967                         } else {
 968                                 err = fuse_copy_fill(cs);
 969                                 if (err)
 970                                         return err;
 971                         }
 972                 }
 973                 if (page) {
 974                         void *mapaddr = kmap_atomic(page);
 975                         void *buf = mapaddr + offset;
 976                         offset += fuse_copy_do(cs, &buf, &count);
 977                         kunmap_atomic(mapaddr);
 978                 } else
 979                         offset += fuse_copy_do(cs, NULL, &count);
 980         }
 981         if (page && !cs->write)
 982                 flush_dcache_page(page);
 983         return 0;
 984 }
 985
 986 /* Copy pages in the request to/from userspace buffer */
 987 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 988                            int zeroing)
 989 {
 990         unsigned i;
 991         struct fuse_req *req = cs->req;
 992
 993         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 994                 int err;
 995                 unsigned offset = req->page_descs[i].offset;
 996                 unsigned count = min(nbytes, req->page_descs[i].length);
 997
 998                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
 999                                      zeroing);
1000                 if (err)
1001                         return err;
1002
1003                 nbytes -= count;
1004         }
1005         return 0;
1006 }
1007
1008 /* Copy a single argument in the request to/from userspace buffer */
1009 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1010 {
1011         while (size) {
1012                 if (!cs->len) {
1013                         int err = fuse_copy_fill(cs);
1014                         if (err)
1015                                 return err;
1016                 }
1017                 fuse_copy_do(cs, &val, &size);
1018         }
1019         return 0;
1020 }
1021
1022 /* Copy request arguments to/from userspace buffer */
1023 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1024                           unsigned argpages, struct fuse_arg *args,
1025                           int zeroing)
1026 {
1027         int err = 0;
1028         unsigned i;
1029
1030         for (i = 0; !err && i < numargs; i++)  {
1031                 struct fuse_arg *arg = &args[i];
1032                 if (i == numargs - 1 && argpages)
1033                         err = fuse_copy_pages(cs, arg->size, zeroing);
1034                 else
1035                         err = fuse_copy_one(cs, arg->value, arg->size);
1036         }
1037         return err;
1038 }
1039
1040 static int forget_pending(struct fuse_iqueue *fiq)
1041 {
1042         return fiq->forget_list_head.next != NULL;
1043 }
1044
1045 static int request_pending(struct fuse_iqueue *fiq)
1046 {
1047         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1048                 forget_pending(fiq);
1049 }
1050
1051 /*
1052  * Transfer an interrupt request to userspace
1053  *
1054  * Unlike other requests this is assembled on demand, without a need
1055  * to allocate a separate fuse_req structure.
1056  *
1057  * Called with fiq->waitq.lock held, releases it
1058  */
1059 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1060                                struct fuse_copy_state *cs,
1061                                size_t nbytes, struct fuse_req *req)
1062 __releases(fiq->waitq.lock)
1063 {
1064         struct fuse_in_header ih;
1065         struct fuse_interrupt_in arg;
1066         unsigned reqsize = sizeof(ih) + sizeof(arg);
1067         int err;
1068
1069         list_del_init(&req->intr_entry);
1070         req->intr_unique = fuse_get_unique(fiq);
1071         memset(&ih, 0, sizeof(ih));
1072         memset(&arg, 0, sizeof(arg));
1073         ih.len = reqsize;
1074         ih.opcode = FUSE_INTERRUPT;
1075         ih.unique = req->intr_unique;
1076         arg.unique = req->in.h.unique;
1077
1078         spin_unlock(&fiq->waitq.lock);
1079         if (nbytes < reqsize)
1080                 return -EINVAL;
1081
1082         err = fuse_copy_one(cs, &ih, sizeof(ih));
1083         if (!err)
1084                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1085         fuse_copy_finish(cs);
1086
1087         return err ? err : reqsize;
1088 }
1089
1090 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1091                                                unsigned max,
1092                                                unsigned *countp)
1093 {
1094         struct fuse_forget_link *head = fiq->forget_list_head.next;
1095         struct fuse_forget_link **newhead = &head;
1096         unsigned count;
1097
1098         for (count = 0; *newhead != NULL && count < max; count++)
1099                 newhead = &(*newhead)->next;
1100
1101         fiq->forget_list_head.next = *newhead;
1102         *newhead = NULL;
1103         if (fiq->forget_list_head.next == NULL)
1104                 fiq->forget_list_tail = &fiq->forget_list_head;
1105
1106         if (countp != NULL)
1107                 *countp = count;
1108
1109         return head;
1110 }
1111
1112 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1113                                    struct fuse_copy_state *cs,
1114                                    size_t nbytes)
1115 __releases(fiq->waitq.lock)
1116 {
1117         int err;
1118         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1119         struct fuse_forget_in arg = {
1120                 .nlookup = forget->forget_one.nlookup,
1121         };
1122         struct fuse_in_header ih = {
1123                 .opcode = FUSE_FORGET,
1124                 .nodeid = forget->forget_one.nodeid,
1125                 .unique = fuse_get_unique(fiq),
1126                 .len = sizeof(ih) + sizeof(arg),
1127         };
1128
1129         spin_unlock(&fiq->waitq.lock);
1130         kfree(forget);
1131         if (nbytes < ih.len)
1132                 return -EINVAL;
1133
1134         err = fuse_copy_one(cs, &ih, sizeof(ih));
1135         if (!err)
1136                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1137         fuse_copy_finish(cs);
1138
1139         if (err)
1140                 return err;
1141
1142         return ih.len;
1143 }
1144
1145 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1146                                    struct fuse_copy_state *cs, size_t nbytes)
1147 __releases(fiq->waitq.lock)
1148 {
1149         int err;
1150         unsigned max_forgets;
1151         unsigned count;
1152         struct fuse_forget_link *head;
1153         struct fuse_batch_forget_in arg = { .count = 0 };
1154         struct fuse_in_header ih = {
1155                 .opcode = FUSE_BATCH_FORGET,
1156                 .unique = fuse_get_unique(fiq),
1157                 .len = sizeof(ih) + sizeof(arg),
1158         };
1159
1160         if (nbytes < ih.len) {
1161                 spin_unlock(&fiq->waitq.lock);
1162                 return -EINVAL;
1163         }
1164
1165         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1166         head = dequeue_forget(fiq, max_forgets, &count);
1167         spin_unlock(&fiq->waitq.lock);
1168
1169         arg.count = count;
1170         ih.len += count * sizeof(struct fuse_forget_one);
1171         err = fuse_copy_one(cs, &ih, sizeof(ih));
1172         if (!err)
1173                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1174
1175         while (head) {
1176                 struct fuse_forget_link *forget = head;
1177
1178                 if (!err) {
1179                         err = fuse_copy_one(cs, &forget->forget_one,
1180                                             sizeof(forget->forget_one));
1181                 }
1182                 head = forget->next;
1183                 kfree(forget);
1184         }
1185
1186         fuse_copy_finish(cs);
1187
1188         if (err)
1189                 return err;
1190
1191         return ih.len;
1192 }
1193
1194 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1195                             struct fuse_copy_state *cs,
1196                             size_t nbytes)
1197 __releases(fiq->waitq.lock)
1198 {
1199         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1200                 return fuse_read_single_forget(fiq, cs, nbytes);
1201         else
1202                 return fuse_read_batch_forget(fiq, cs, nbytes);
1203 }
1204
1205 /*
1206  * Read a single request into the userspace filesystem's buffer.  This
1207  * function waits until a request is available, then removes it from
1208  * the pending list and copies request data to userspace buffer.  If
1209  * no reply is needed (FORGET) or request has been aborted or there
1210  * was an error during the copying then it's finished by calling
1211  * request_end().  Otherwise add it to the processing list, and set
1212  * the 'sent' flag.
1213  */
1214 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1215                                 struct fuse_copy_state *cs, size_t nbytes)
1216 {
1217         ssize_t err;
1218         struct fuse_conn *fc = fud->fc;
1219         struct fuse_iqueue *fiq = &fc->iq;
1220         struct fuse_pqueue *fpq = &fud->pq;
1221         struct fuse_req *req;
1222         struct fuse_in *in;
1223         unsigned reqsize;
1224
1225  restart:
1226         spin_lock(&fiq->waitq.lock);
1227         err = -EAGAIN;
1228         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1229             !request_pending(fiq))
1230                 goto err_unlock;
1231
1232         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1233                                 !fiq->connected || request_pending(fiq));
1234         if (err)
1235                 goto err_unlock;
1236
1237         err = -ENODEV;
1238         if (!fiq->connected)
1239                 goto err_unlock;
1240
1241         if (!list_empty(&fiq->interrupts)) {
1242                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1243                                  intr_entry);
1244                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1245         }
1246
1247         if (forget_pending(fiq)) {
1248                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1249                         return fuse_read_forget(fc, fiq, cs, nbytes);
1250
1251                 if (fiq->forget_batch <= -8)
1252                         fiq->forget_batch = 16;
1253         }
1254
1255         req = list_entry(fiq->pending.next, struct fuse_req, list);
1256         clear_bit(FR_PENDING, &req->flags);
1257         list_del_init(&req->list);
1258         spin_unlock(&fiq->waitq.lock);
1259
1260         in = &req->in;
1261         reqsize = in->h.len;
1262         /* If request is too large, reply with an error and restart the read */
1263         if (nbytes < reqsize) {
1264                 req->out.h.error = -EIO;
1265                 /* SETXATTR is special, since it may contain too large data */
1266                 if (in->h.opcode == FUSE_SETXATTR)
1267                         req->out.h.error = -E2BIG;
1268                 request_end(fc, req);
1269                 goto restart;
1270         }
1271         spin_lock(&fpq->lock);
1272         list_add(&req->list, &fpq->io);
1273         spin_unlock(&fpq->lock);
1274         cs->req = req;
1275         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1276         if (!err)
1277                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1278                                      (struct fuse_arg *) in->args, 0);
1279         fuse_copy_finish(cs);
1280         spin_lock(&fpq->lock);
1281         clear_bit(FR_LOCKED, &req->flags);
1282         if (!fpq->connected) {
1283                 err = -ENODEV;
1284                 goto out_end;
1285         }
1286         if (err) {
1287                 req->out.h.error = -EIO;
1288                 goto out_end;
1289         }
1290         if (!test_bit(FR_ISREPLY, &req->flags)) {
1291                 err = reqsize;
1292                 goto out_end;
1293         }
1294         list_move_tail(&req->list, &fpq->processing);
1295         spin_unlock(&fpq->lock);
1296         set_bit(FR_SENT, &req->flags);
1297         /* matches barrier in request_wait_answer() */
1298         smp_mb__after_atomic();
1299         if (test_bit(FR_INTERRUPTED, &req->flags))
1300                 queue_interrupt(fiq, req);
1301
1302         return reqsize;
1303
1304 out_end:
1305         if (!test_bit(FR_PRIVATE, &req->flags))
1306                 list_del_init(&req->list);
1307         spin_unlock(&fpq->lock);
1308         request_end(fc, req);
1309         return err;
1310
1311  err_unlock:
1312         spin_unlock(&fiq->waitq.lock);
1313         return err;
1314 }
1315
1316 static int fuse_dev_open(struct inode *inode, struct file *file)
1317 {
1318         /*
1319          * The fuse device's file's private_data is used to hold
1320          * the fuse_conn(ection) when it is mounted, and is used to
1321          * keep track of whether the file has been mounted already.
1322          */
1323         file->private_data = NULL;
1324         return 0;
1325 }
1326
1327 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1328 {
1329         struct fuse_copy_state cs;
1330         struct file *file = iocb->ki_filp;
1331         struct fuse_dev *fud = fuse_get_dev(file);
1332
1333         if (!fud)
1334                 return -EPERM;
1335
1336         if (!iter_is_iovec(to))
1337                 return -EINVAL;
1338
1339         fuse_copy_init(&cs, 1, to);
1340
1341         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1342 }
1343
1344 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1345                                     struct pipe_inode_info *pipe,
1346                                     size_t len, unsigned int flags)
1347 {
1348         int total, ret;
1349         int page_nr = 0;
1350         struct pipe_buffer *bufs;
1351         struct fuse_copy_state cs;
1352         struct fuse_dev *fud = fuse_get_dev(in);
1353
1354         if (!fud)
1355                 return -EPERM;
1356
1357         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1358         if (!bufs)
1359                 return -ENOMEM;
1360
1361         fuse_copy_init(&cs, 1, NULL);
1362         cs.pipebufs = bufs;
1363         cs.pipe = pipe;
1364         ret = fuse_dev_do_read(fud, in, &cs, len);
1365         if (ret < 0)
1366                 goto out;
1367
1368         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1369                 ret = -EIO;
1370                 goto out;
1371         }
1372
1373         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1374                 /*
1375                  * Need to be careful about this.  Having buf->ops in module
1376                  * code can Oops if the buffer persists after module unload.
1377                  */
1378                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1379                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1380                 if (unlikely(ret < 0))
1381                         break;
1382         }
1383         if (total)
1384                 ret = total;
1385 out:
1386         for (; page_nr < cs.nr_segs; page_nr++)
1387                 put_page(bufs[page_nr].page);
1388
1389         kfree(bufs);
1390         return ret;
1391 }
1392
1393 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1394                             struct fuse_copy_state *cs)
1395 {
1396         struct fuse_notify_poll_wakeup_out outarg;
1397         int err = -EINVAL;
1398
1399         if (size != sizeof(outarg))
1400                 goto err;
1401
1402         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1403         if (err)
1404                 goto err;
1405
1406         fuse_copy_finish(cs);
1407         return fuse_notify_poll_wakeup(fc, &outarg);
1408
1409 err:
1410         fuse_copy_finish(cs);
1411         return err;
1412 }
1413
1414 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1415                                    struct fuse_copy_state *cs)
1416 {
1417         struct fuse_notify_inval_inode_out outarg;
1418         int err = -EINVAL;
1419
1420         if (size != sizeof(outarg))
1421                 goto err;
1422
1423         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1424         if (err)
1425                 goto err;
1426         fuse_copy_finish(cs);
1427
1428         down_read(&fc->killsb);
1429         err = -ENOENT;
1430         if (fc->sb) {
1431                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1432                                                outarg.off, outarg.len);
1433         }
1434         up_read(&fc->killsb);
1435         return err;
1436
1437 err:
1438         fuse_copy_finish(cs);
1439         return err;
1440 }
1441
1442 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1443                                    struct fuse_copy_state *cs)
1444 {
1445         struct fuse_notify_inval_entry_out outarg;
1446         int err = -ENOMEM;
1447         char *buf;
1448         struct qstr name;
1449
1450         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1451         if (!buf)
1452                 goto err;
1453
1454         err = -EINVAL;
1455         if (size < sizeof(outarg))
1456                 goto err;
1457
1458         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1459         if (err)
1460                 goto err;
1461
1462         err = -ENAMETOOLONG;
1463         if (outarg.namelen > FUSE_NAME_MAX)
1464                 goto err;
1465
1466         err = -EINVAL;
1467         if (size != sizeof(outarg) + outarg.namelen + 1)
1468                 goto err;
1469
1470         name.name = buf;
1471         name.len = outarg.namelen;
1472         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1473         if (err)
1474                 goto err;
1475         fuse_copy_finish(cs);
1476         buf[outarg.namelen] = 0;
1477
1478         down_read(&fc->killsb);
1479         err = -ENOENT;
1480         if (fc->sb)
1481                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1482         up_read(&fc->killsb);
1483         kfree(buf);
1484         return err;
1485
1486 err:
1487         kfree(buf);
1488         fuse_copy_finish(cs);
1489         return err;
1490 }
1491
1492 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1493                               struct fuse_copy_state *cs)
1494 {
1495         struct fuse_notify_delete_out outarg;
1496         int err = -ENOMEM;
1497         char *buf;
1498         struct qstr name;
1499
1500         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1501         if (!buf)
1502                 goto err;
1503
1504         err = -EINVAL;
1505         if (size < sizeof(outarg))
1506                 goto err;
1507
1508         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1509         if (err)
1510                 goto err;
1511
1512         err = -ENAMETOOLONG;
1513         if (outarg.namelen > FUSE_NAME_MAX)
1514                 goto err;
1515
1516         err = -EINVAL;
1517         if (size != sizeof(outarg) + outarg.namelen + 1)
1518                 goto err;
1519
1520         name.name = buf;
1521         name.len = outarg.namelen;
1522         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1523         if (err)
1524                 goto err;
1525         fuse_copy_finish(cs);
1526         buf[outarg.namelen] = 0;
1527
1528         down_read(&fc->killsb);
1529         err = -ENOENT;
1530         if (fc->sb)
1531                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1532                                                outarg.child, &name);
1533         up_read(&fc->killsb);
1534         kfree(buf);
1535         return err;
1536
1537 err:
1538         kfree(buf);
1539         fuse_copy_finish(cs);
1540         return err;
1541 }
1542
1543 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1544                              struct fuse_copy_state *cs)
1545 {
1546         struct fuse_notify_store_out outarg;
1547         struct inode *inode;
1548         struct address_space *mapping;
1549         u64 nodeid;
1550         int err;
1551         pgoff_t index;
1552         unsigned int offset;
1553         unsigned int num;
1554         loff_t file_size;
1555         loff_t end;
1556
1557         err = -EINVAL;
1558         if (size < sizeof(outarg))
1559                 goto out_finish;
1560
1561         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1562         if (err)
1563                 goto out_finish;
1564
1565         err = -EINVAL;
1566         if (size - sizeof(outarg) != outarg.size)
1567                 goto out_finish;
1568
1569         nodeid = outarg.nodeid;
1570
1571         down_read(&fc->killsb);
1572
1573         err = -ENOENT;
1574         if (!fc->sb)
1575                 goto out_up_killsb;
1576
1577         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1578         if (!inode)
1579                 goto out_up_killsb;
1580
1581         mapping = inode->i_mapping;
1582         index = outarg.offset >> PAGE_SHIFT;
1583         offset = outarg.offset & ~PAGE_MASK;
1584         file_size = i_size_read(inode);
1585         end = outarg.offset + outarg.size;
1586         if (end > file_size) {
1587                 file_size = end;
1588                 fuse_write_update_size(inode, file_size);
1589         }
1590
1591         num = outarg.size;
1592         while (num) {
1593                 struct page *page;
1594                 unsigned int this_num;
1595
1596                 err = -ENOMEM;
1597                 page = find_or_create_page(mapping, index,
1598                                            mapping_gfp_mask(mapping));
1599                 if (!page)
1600                         goto out_iput;
1601
1602                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1603                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1604                 if (!err && offset == 0 &&
1605                     (this_num == PAGE_SIZE || file_size == end))
1606                         SetPageUptodate(page);
1607                 unlock_page(page);
1608                 put_page(page);
1609
1610                 if (err)
1611                         goto out_iput;
1612
1613                 num -= this_num;
1614                 offset = 0;
1615                 index++;
1616         }
1617
1618         err = 0;
1619
1620 out_iput:
1621         iput(inode);
1622 out_up_killsb:
1623         up_read(&fc->killsb);
1624 out_finish:
1625         fuse_copy_finish(cs);
1626         return err;
1627 }
1628
1629 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1630 {
1631         release_pages(req->pages, req->num_pages, false);
1632 }
1633
1634 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1635                          struct fuse_notify_retrieve_out *outarg)
1636 {
1637         int err;
1638         struct address_space *mapping = inode->i_mapping;
1639         struct fuse_req *req;
1640         pgoff_t index;
1641         loff_t file_size;
1642         unsigned int num;
1643         unsigned int offset;
1644         size_t total_len = 0;
1645         int num_pages;
1646
1647         offset = outarg->offset & ~PAGE_MASK;
1648         file_size = i_size_read(inode);
1649
1650         num = outarg->size;
1651         if (outarg->offset > file_size)
1652                 num = 0;
1653         else if (outarg->offset + num > file_size)
1654                 num = file_size - outarg->offset;
1655
1656         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1657         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1658
1659         req = fuse_get_req(fc, num_pages);
1660         if (IS_ERR(req))
1661                 return PTR_ERR(req);
1662
1663         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1664         req->in.h.nodeid = outarg->nodeid;
1665         req->in.numargs = 2;
1666         req->in.argpages = 1;
1667         req->page_descs[0].offset = offset;
1668         req->end = fuse_retrieve_end;
1669
1670         index = outarg->offset >> PAGE_SHIFT;
1671
1672         while (num && req->num_pages < num_pages) {
1673                 struct page *page;
1674                 unsigned int this_num;
1675
1676                 page = find_get_page(mapping, index);
1677                 if (!page)
1678                         break;
1679
1680                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1681                 req->pages[req->num_pages] = page;
1682                 req->page_descs[req->num_pages].length = this_num;
1683                 req->num_pages++;
1684
1685                 offset = 0;
1686                 num -= this_num;
1687                 total_len += this_num;
1688                 index++;
1689         }
1690         req->misc.retrieve_in.offset = outarg->offset;
1691         req->misc.retrieve_in.size = total_len;
1692         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1693         req->in.args[0].value = &req->misc.retrieve_in;
1694         req->in.args[1].size = total_len;
1695
1696         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1697         if (err)
1698                 fuse_retrieve_end(fc, req);
1699
1700         return err;
1701 }
1702
1703 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1704                                 struct fuse_copy_state *cs)
1705 {
1706         struct fuse_notify_retrieve_out outarg;
1707         struct inode *inode;
1708         int err;
1709
1710         err = -EINVAL;
1711         if (size != sizeof(outarg))
1712                 goto copy_finish;
1713
1714         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1715         if (err)
1716                 goto copy_finish;
1717
1718         fuse_copy_finish(cs);
1719
1720         down_read(&fc->killsb);
1721         err = -ENOENT;
1722         if (fc->sb) {
1723                 u64 nodeid = outarg.nodeid;
1724
1725                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1726                 if (inode) {
1727                         err = fuse_retrieve(fc, inode, &outarg);
1728                         iput(inode);
1729                 }
1730         }
1731         up_read(&fc->killsb);
1732
1733         return err;
1734
1735 copy_finish:
1736         fuse_copy_finish(cs);
1737         return err;
1738 }
1739
1740 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1741                        unsigned int size, struct fuse_copy_state *cs)
1742 {
1743         /* Don't try to move pages (yet) */
1744         cs->move_pages = 0;
1745
1746         switch (code) {
1747         case FUSE_NOTIFY_POLL:
1748                 return fuse_notify_poll(fc, size, cs);
1749
1750         case FUSE_NOTIFY_INVAL_INODE:
1751                 return fuse_notify_inval_inode(fc, size, cs);
1752
1753         case FUSE_NOTIFY_INVAL_ENTRY:
1754                 return fuse_notify_inval_entry(fc, size, cs);
1755
1756         case FUSE_NOTIFY_STORE:
1757                 return fuse_notify_store(fc, size, cs);
1758
1759         case FUSE_NOTIFY_RETRIEVE:
1760                 return fuse_notify_retrieve(fc, size, cs);
1761
1762         case FUSE_NOTIFY_DELETE:
1763                 return fuse_notify_delete(fc, size, cs);
1764
1765         default:
1766                 fuse_copy_finish(cs);
1767                 return -EINVAL;
1768         }
1769 }
1770
1771 /* Look up request on processing list by unique ID */
1772 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1773 {
1774         struct fuse_req *req;
1775
1776         list_for_each_entry(req, &fpq->processing, list) {
1777                 if (req->in.h.unique == unique || req->intr_unique == unique)
1778                         return req;
1779         }
1780         return NULL;
1781 }
1782
1783 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1784                          unsigned nbytes)
1785 {
1786         unsigned reqsize = sizeof(struct fuse_out_header);
1787
1788         if (out->h.error)
1789                 return nbytes != reqsize ? -EINVAL : 0;
1790
1791         reqsize += len_args(out->numargs, out->args);
1792
1793         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1794                 return -EINVAL;
1795         else if (reqsize > nbytes) {
1796                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1797                 unsigned diffsize = reqsize - nbytes;
1798                 if (diffsize > lastarg->size)
1799                         return -EINVAL;
1800                 lastarg->size -= diffsize;
1801         }
1802         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1803                               out->page_zeroing);
1804 }
1805
1806 /*
1807  * Write a single reply to a request.  First the header is copied from
1808  * the write buffer.  The request is then searched on the processing
1809  * list by the unique ID found in the header.  If found, then remove
1810  * it from the list and copy the rest of the buffer to the request.
1811  * The request is finished by calling request_end()
1812  */
1813 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1814                                  struct fuse_copy_state *cs, size_t nbytes)
1815 {
1816         int err;
1817         struct fuse_conn *fc = fud->fc;
1818         struct fuse_pqueue *fpq = &fud->pq;
1819         struct fuse_req *req;
1820         struct fuse_out_header oh;
1821
1822         if (nbytes < sizeof(struct fuse_out_header))
1823                 return -EINVAL;
1824
1825         err = fuse_copy_one(cs, &oh, sizeof(oh));
1826         if (err)
1827                 goto err_finish;
1828
1829         err = -EINVAL;
1830         if (oh.len != nbytes)
1831                 goto err_finish;
1832
1833         /*
1834          * Zero oh.unique indicates unsolicited notification message
1835          * and error contains notification code.
1836          */
1837         if (!oh.unique) {
1838                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1839                 return err ? err : nbytes;
1840         }
1841
1842         err = -EINVAL;
1843         if (oh.error <= -1000 || oh.error > 0)
1844                 goto err_finish;
1845
1846         spin_lock(&fpq->lock);
1847         err = -ENOENT;
1848         if (!fpq->connected)
1849                 goto err_unlock_pq;
1850
1851         req = request_find(fpq, oh.unique);
1852         if (!req)
1853                 goto err_unlock_pq;
1854
1855         /* Is it an interrupt reply? */
1856         if (req->intr_unique == oh.unique) {
1857                 spin_unlock(&fpq->lock);
1858
1859                 err = -EINVAL;
1860                 if (nbytes != sizeof(struct fuse_out_header))
1861                         goto err_finish;
1862
1863                 if (oh.error == -ENOSYS)
1864                         fc->no_interrupt = 1;
1865                 else if (oh.error == -EAGAIN)
1866                         queue_interrupt(&fc->iq, req);
1867
1868                 fuse_copy_finish(cs);
1869                 return nbytes;
1870         }
1871
1872         clear_bit(FR_SENT, &req->flags);
1873         list_move(&req->list, &fpq->io);
1874         req->out.h = oh;
1875         set_bit(FR_LOCKED, &req->flags);
1876         spin_unlock(&fpq->lock);
1877         cs->req = req;
1878         if (!req->out.page_replace)
1879                 cs->move_pages = 0;
1880
1881         err = copy_out_args(cs, &req->out, nbytes);
1882         fuse_copy_finish(cs);
1883
1884         spin_lock(&fpq->lock);
1885         clear_bit(FR_LOCKED, &req->flags);
1886         if (!fpq->connected)
1887                 err = -ENOENT;
1888         else if (err)
1889                 req->out.h.error = -EIO;
1890         if (!test_bit(FR_PRIVATE, &req->flags))
1891                 list_del_init(&req->list);
1892         spin_unlock(&fpq->lock);
1893
1894         request_end(fc, req);
1895
1896         return err ? err : nbytes;
1897
1898  err_unlock_pq:
1899         spin_unlock(&fpq->lock);
1900  err_finish:
1901         fuse_copy_finish(cs);
1902         return err;
1903 }
1904
1905 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1906 {
1907         struct fuse_copy_state cs;
1908         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1909
1910         if (!fud)
1911                 return -EPERM;
1912
1913         if (!iter_is_iovec(from))
1914                 return -EINVAL;
1915
1916         fuse_copy_init(&cs, 0, from);
1917
1918         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1919 }
1920
1921 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1922                                      struct file *out, loff_t *ppos,
1923                                      size_t len, unsigned int flags)
1924 {
1925         unsigned nbuf;
1926         unsigned idx;
1927         struct pipe_buffer *bufs;
1928         struct fuse_copy_state cs;
1929         struct fuse_dev *fud;
1930         size_t rem;
1931         ssize_t ret;
1932
1933         fud = fuse_get_dev(out);
1934         if (!fud)
1935                 return -EPERM;
1936
1937         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1938         if (!bufs)
1939                 return -ENOMEM;
1940
1941         pipe_lock(pipe);
1942         nbuf = 0;
1943         rem = 0;
1944         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1945                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1946
1947         ret = -EINVAL;
1948         if (rem < len) {
1949                 pipe_unlock(pipe);
1950                 goto out;
1951         }
1952
1953         rem = len;
1954         while (rem) {
1955                 struct pipe_buffer *ibuf;
1956                 struct pipe_buffer *obuf;
1957
1958                 BUG_ON(nbuf >= pipe->buffers);
1959                 BUG_ON(!pipe->nrbufs);
1960                 ibuf = &pipe->bufs[pipe->curbuf];
1961                 obuf = &bufs[nbuf];
1962
1963                 if (rem >= ibuf->len) {
1964                         *obuf = *ibuf;
1965                         ibuf->ops = NULL;
1966                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1967                         pipe->nrbufs--;
1968                 } else {
1969                         pipe_buf_get(pipe, ibuf);
1970                         *obuf = *ibuf;
1971                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1972                         obuf->len = rem;
1973                         ibuf->offset += obuf->len;
1974                         ibuf->len -= obuf->len;
1975                 }
1976                 nbuf++;
1977                 rem -= obuf->len;
1978         }
1979         pipe_unlock(pipe);
1980
1981         fuse_copy_init(&cs, 0, NULL);
1982         cs.pipebufs = bufs;
1983         cs.nr_segs = nbuf;
1984         cs.pipe = pipe;
1985
1986         if (flags & SPLICE_F_MOVE)
1987                 cs.move_pages = 1;
1988
1989         ret = fuse_dev_do_write(fud, &cs, len);
1990
1991         for (idx = 0; idx < nbuf; idx++)
1992                 pipe_buf_release(pipe, &bufs[idx]);
1993
1994 out:
1995         kfree(bufs);
1996         return ret;
1997 }
1998
1999 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2000 {
2001         unsigned mask = POLLOUT | POLLWRNORM;
2002         struct fuse_iqueue *fiq;
2003         struct fuse_dev *fud = fuse_get_dev(file);
2004
2005         if (!fud)
2006                 return POLLERR;
2007
2008         fiq = &fud->fc->iq;
2009         poll_wait(file, &fiq->waitq, wait);
2010
2011         spin_lock(&fiq->waitq.lock);
2012         if (!fiq->connected)
2013                 mask = POLLERR;
2014         else if (request_pending(fiq))
2015                 mask |= POLLIN | POLLRDNORM;
2016         spin_unlock(&fiq->waitq.lock);
2017
2018         return mask;
2019 }
2020
2021 /*
2022  * Abort all requests on the given list (pending or processing)
2023  *
2024  * This function releases and reacquires fc->lock
2025  */
2026 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2027 {
2028         while (!list_empty(head)) {
2029                 struct fuse_req *req;
2030                 req = list_entry(head->next, struct fuse_req, list);
2031                 req->out.h.error = -ECONNABORTED;
2032                 clear_bit(FR_SENT, &req->flags);
2033                 list_del_init(&req->list);
2034                 request_end(fc, req);
2035         }
2036 }
2037
2038 static void end_polls(struct fuse_conn *fc)
2039 {
2040         struct rb_node *p;
2041
2042         p = rb_first(&fc->polled_files);
2043
2044         while (p) {
2045                 struct fuse_file *ff;
2046                 ff = rb_entry(p, struct fuse_file, polled_node);
2047                 wake_up_interruptible_all(&ff->poll_wait);
2048
2049                 p = rb_next(p);
2050         }
2051 }
2052
2053 /*
2054  * Abort all requests.
2055  *
2056  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2057  * filesystem.
2058  *
2059  * The same effect is usually achievable through killing the filesystem daemon
2060  * and all users of the filesystem.  The exception is the combination of an
2061  * asynchronous request and the tricky deadlock (see
2062  * Documentation/filesystems/fuse.txt).
2063  *
2064  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2065  * requests, they should be finished off immediately.  Locked requests will be
2066  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2067  * requests.  It is possible that some request will finish before we can.  This
2068  * is OK, the request will in that case be removed from the list before we touch
2069  * it.
2070  */
2071 void fuse_abort_conn(struct fuse_conn *fc)
2072 {
2073         struct fuse_iqueue *fiq = &fc->iq;
2074
2075         spin_lock(&fc->lock);
2076         if (fc->connected) {
2077                 struct fuse_dev *fud;
2078                 struct fuse_req *req, *next;
2079                 LIST_HEAD(to_end1);
2080                 LIST_HEAD(to_end2);
2081
2082                 fc->connected = 0;
2083                 fc->blocked = 0;
2084                 fuse_set_initialized(fc);
2085                 list_for_each_entry(fud, &fc->devices, entry) {
2086                         struct fuse_pqueue *fpq = &fud->pq;
2087
2088                         spin_lock(&fpq->lock);
2089                         fpq->connected = 0;
2090                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2091                                 req->out.h.error = -ECONNABORTED;
2092                                 spin_lock(&req->waitq.lock);
2093                                 set_bit(FR_ABORTED, &req->flags);
2094                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2095                                         set_bit(FR_PRIVATE, &req->flags);
2096                                         list_move(&req->list, &to_end1);
2097                                 }
2098                                 spin_unlock(&req->waitq.lock);
2099                         }
2100                         list_splice_init(&fpq->processing, &to_end2);
2101                         spin_unlock(&fpq->lock);
2102                 }
2103                 fc->max_background = UINT_MAX;
2104                 flush_bg_queue(fc);
2105
2106                 spin_lock(&fiq->waitq.lock);
2107                 fiq->connected = 0;
2108                 list_splice_init(&fiq->pending, &to_end2);
2109                 list_for_each_entry(req, &to_end2, list)
2110                         clear_bit(FR_PENDING, &req->flags);
2111                 while (forget_pending(fiq))
2112                         kfree(dequeue_forget(fiq, 1, NULL));
2113                 wake_up_all_locked(&fiq->waitq);
2114                 spin_unlock(&fiq->waitq.lock);
2115                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2116                 end_polls(fc);
2117                 wake_up_all(&fc->blocked_waitq);
2118                 spin_unlock(&fc->lock);
2119
2120                 while (!list_empty(&to_end1)) {
2121                         req = list_first_entry(&to_end1, struct fuse_req, list);
2122                         __fuse_get_request(req);
2123                         list_del_init(&req->list);
2124                         request_end(fc, req);
2125                 }
2126                 end_requests(fc, &to_end2);
2127         } else {
2128                 spin_unlock(&fc->lock);
2129         }
2130 }
2131 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2132
2133 int fuse_dev_release(struct inode *inode, struct file *file)
2134 {
2135         struct fuse_dev *fud = fuse_get_dev(file);
2136
2137         if (fud) {
2138                 struct fuse_conn *fc = fud->fc;
2139                 struct fuse_pqueue *fpq = &fud->pq;
2140
2141                 WARN_ON(!list_empty(&fpq->io));
2142                 end_requests(fc, &fpq->processing);
2143                 /* Are we the last open device? */
2144                 if (atomic_dec_and_test(&fc->dev_count)) {
2145                         WARN_ON(fc->iq.fasync != NULL);
2146                         fuse_abort_conn(fc);
2147                 }
2148                 fuse_dev_free(fud);
2149         }
2150         return 0;
2151 }
2152 EXPORT_SYMBOL_GPL(fuse_dev_release);
2153
2154 static int fuse_dev_fasync(int fd, struct file *file, int on)
2155 {
2156         struct fuse_dev *fud = fuse_get_dev(file);
2157
2158         if (!fud)
2159                 return -EPERM;
2160
2161         /* No locking - fasync_helper does its own locking */
2162         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2163 }
2164
2165 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2166 {
2167         struct fuse_dev *fud;
2168
2169         if (new->private_data)
2170                 return -EINVAL;
2171
2172         fud = fuse_dev_alloc(fc);
2173         if (!fud)
2174                 return -ENOMEM;
2175
2176         new->private_data = fud;
2177         atomic_inc(&fc->dev_count);
2178
2179         return 0;
2180 }
2181
2182 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2183                            unsigned long arg)
2184 {
2185         int err = -ENOTTY;
2186
2187         if (cmd == FUSE_DEV_IOC_CLONE) {
2188                 int oldfd;
2189
2190                 err = -EFAULT;
2191                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2192                         struct file *old = fget(oldfd);
2193
2194                         err = -EINVAL;
2195                         if (old) {
2196                                 struct fuse_dev *fud = NULL;
2197
2198                                 /*
2199                                  * Check against file->f_op because CUSE
2200                                  * uses the same ioctl handler.
2201                                  */
2202                                 if (old->f_op == file->f_op &&
2203                                     old->f_cred->user_ns == file->f_cred->user_ns)
2204                                         fud = fuse_get_dev(old);
2205
2206                                 if (fud) {
2207                                         mutex_lock(&fuse_mutex);
2208                                         err = fuse_device_clone(fud->fc, file);
2209                                         mutex_unlock(&fuse_mutex);
2210                                 }
2211                                 fput(old);
2212                         }
2213                 }
2214         }
2215         return err;
2216 }
2217
2218 const struct file_operations fuse_dev_operations = {
2219         .owner          = THIS_MODULE,
2220         .open           = fuse_dev_open,
2221         .llseek         = no_llseek,
2222         .read_iter      = fuse_dev_read,
2223         .splice_read    = fuse_dev_splice_read,
2224         .write_iter     = fuse_dev_write,
2225         .splice_write   = fuse_dev_splice_write,
2226         .poll           = fuse_dev_poll,
2227         .release        = fuse_dev_release,
2228         .fasync         = fuse_dev_fasync,
2229         .unlocked_ioctl = fuse_dev_ioctl,
2230         .compat_ioctl   = fuse_dev_ioctl,
2231 };
2232 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2233
2234 static struct miscdevice fuse_miscdevice = {
2235         .minor = FUSE_MINOR,
2236         .name  = "fuse",
2237         .fops = &fuse_dev_operations,
2238 };
2239
2240 int __init fuse_dev_init(void)
2241 {
2242         int err = -ENOMEM;
2243         fuse_req_cachep = kmem_cache_create("fuse_request",
2244                                             sizeof(struct fuse_req),
2245                                             0, 0, NULL);
2246         if (!fuse_req_cachep)
2247                 goto out;
2248
2249         err = misc_register(&fuse_miscdevice);
2250         if (err)
2251                 goto out_cache_clean;
2252
2253         return 0;
2254
2255  out_cache_clean:
2256         kmem_cache_destroy(fuse_req_cachep);
2257  out:
2258         return err;
2259 }
2260
2261 void fuse_dev_cleanup(void)
2262 {
2263         misc_deregister(&fuse_miscdevice);
2264         kmem_cache_destroy(fuse_req_cachep);
2265 }