fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/sched/signal.h>
  15 #include <linux/uio.h>
  16 #include <linux/miscdevice.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/pipe_fs_i.h>
  21 #include <linux/swap.h>
  22 #include <linux/splice.h>
  23 #include <linux/sched.h>
  24
  25 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  26 MODULE_ALIAS("devname:fuse");
  27
  28 /* Ordinary requests have even IDs, while interrupts IDs are odd */
  29 #define FUSE_INT_REQ_BIT (1ULL << 0)
  30 #define FUSE_REQ_ID_STEP (1ULL << 1)
  31
  32 static struct kmem_cache *fuse_req_cachep;
  33
  34 static struct fuse_dev *fuse_get_dev(struct file *file)
  35 {
  36         /*
  37          * Lockless access is OK, because file->private data is set
  38          * once during mount and is valid until the file is released.
  39          */
  40         return READ_ONCE(file->private_data);
  41 }
  42
  43 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  44                               struct fuse_page_desc *page_descs,
  45                               unsigned npages)
  46 {
  47         INIT_LIST_HEAD(&req->list);
  48         INIT_LIST_HEAD(&req->intr_entry);
  49         init_waitqueue_head(&req->waitq);
  50         refcount_set(&req->count, 1);
  51         req->pages = pages;
  52         req->page_descs = page_descs;
  53         req->max_pages = npages;
  54         __set_bit(FR_PENDING, &req->flags);
  55 }
  56
  57 static struct page **fuse_req_pages_alloc(unsigned int npages, gfp_t flags,
  58                                           struct fuse_page_desc **desc)
  59 {
  60         struct page **pages;
  61
  62         pages = kzalloc(npages * (sizeof(struct page *) +
  63                                   sizeof(struct fuse_page_desc)), flags);
  64         *desc = (void *) pages + npages * sizeof(struct page *);
  65
  66         return pages;
  67 }
  68
  69 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  70 {
  71         struct fuse_req *req = kmem_cache_zalloc(fuse_req_cachep, flags);
  72         if (req) {
  73                 struct page **pages = NULL;
  74                 struct fuse_page_desc *page_descs = NULL;
  75
  76                 WARN_ON(npages > FUSE_MAX_MAX_PAGES);
  77                 if (npages > FUSE_REQ_INLINE_PAGES) {
  78                         pages = fuse_req_pages_alloc(npages, flags,
  79                                                      &page_descs);
  80                         if (!pages) {
  81                                 kmem_cache_free(fuse_req_cachep, req);
  82                                 return NULL;
  83                         }
  84                 } else if (npages) {
  85                         pages = req->inline_pages;
  86                         page_descs = req->inline_page_descs;
  87                 }
  88
  89                 fuse_request_init(req, pages, page_descs, npages);
  90         }
  91         return req;
  92 }
  93
  94 struct fuse_req *fuse_request_alloc(unsigned npages)
  95 {
  96         return __fuse_request_alloc(npages, GFP_KERNEL);
  97 }
  98 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  99
 100 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
 101 {
 102         return __fuse_request_alloc(npages, GFP_NOFS);
 103 }
 104
 105 static void fuse_req_pages_free(struct fuse_req *req)
 106 {
 107         if (req->pages != req->inline_pages)
 108                 kfree(req->pages);
 109 }
 110
 111 bool fuse_req_realloc_pages(struct fuse_conn *fc, struct fuse_req *req,
 112                             gfp_t flags)
 113 {
 114         struct page **pages;
 115         struct fuse_page_desc *page_descs;
 116         unsigned int npages = min_t(unsigned int,
 117                                     max_t(unsigned int, req->max_pages * 2,
 118                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
 119                                     fc->max_pages);
 120         WARN_ON(npages <= req->max_pages);
 121
 122         pages = fuse_req_pages_alloc(npages, flags, &page_descs);
 123         if (!pages)
 124                 return false;
 125
 126         memcpy(pages, req->pages, sizeof(struct page *) * req->max_pages);
 127         memcpy(page_descs, req->page_descs,
 128                sizeof(struct fuse_page_desc) * req->max_pages);
 129         fuse_req_pages_free(req);
 130         req->pages = pages;
 131         req->page_descs = page_descs;
 132         req->max_pages = npages;
 133
 134         return true;
 135 }
 136
 137 void fuse_request_free(struct fuse_req *req)
 138 {
 139         fuse_req_pages_free(req);
 140         kmem_cache_free(fuse_req_cachep, req);
 141 }
 142
 143 void __fuse_get_request(struct fuse_req *req)
 144 {
 145         refcount_inc(&req->count);
 146 }
 147
 148 /* Must be called with > 1 refcount */
 149 static void __fuse_put_request(struct fuse_req *req)
 150 {
 151         refcount_dec(&req->count);
 152 }
 153
 154 void fuse_set_initialized(struct fuse_conn *fc)
 155 {
 156         /* Make sure stores before this are seen on another CPU */
 157         smp_wmb();
 158         fc->initialized = 1;
 159 }
 160
 161 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 162 {
 163         return !fc->initialized || (for_background && fc->blocked);
 164 }
 165
 166 static void fuse_drop_waiting(struct fuse_conn *fc)
 167 {
 168         if (fc->connected) {
 169                 atomic_dec(&fc->num_waiting);
 170         } else if (atomic_dec_and_test(&fc->num_waiting)) {
 171                 /* wake up aborters */
 172                 wake_up_all(&fc->blocked_waitq);
 173         }
 174 }
 175
 176 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 177                                        bool for_background)
 178 {
 179         struct fuse_req *req;
 180         int err;
 181         atomic_inc(&fc->num_waiting);
 182
 183         if (fuse_block_alloc(fc, for_background)) {
 184                 err = -EINTR;
 185                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 186                                 !fuse_block_alloc(fc, for_background)))
 187                         goto out;
 188         }
 189         /* Matches smp_wmb() in fuse_set_initialized() */
 190         smp_rmb();
 191
 192         err = -ENOTCONN;
 193         if (!fc->connected)
 194                 goto out;
 195
 196         err = -ECONNREFUSED;
 197         if (fc->conn_error)
 198                 goto out;
 199
 200         req = fuse_request_alloc(npages);
 201         err = -ENOMEM;
 202         if (!req) {
 203                 if (for_background)
 204                         wake_up(&fc->blocked_waitq);
 205                 goto out;
 206         }
 207
 208         req->in.h.uid = from_kuid(fc->user_ns, current_fsuid());
 209         req->in.h.gid = from_kgid(fc->user_ns, current_fsgid());
 210         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 211
 212         __set_bit(FR_WAITING, &req->flags);
 213         if (for_background)
 214                 __set_bit(FR_BACKGROUND, &req->flags);
 215
 216         if (unlikely(req->in.h.uid == ((uid_t)-1) ||
 217                      req->in.h.gid == ((gid_t)-1))) {
 218                 fuse_put_request(fc, req);
 219                 return ERR_PTR(-EOVERFLOW);
 220         }
 221         return req;
 222
 223  out:
 224         fuse_drop_waiting(fc);
 225         return ERR_PTR(err);
 226 }
 227
 228 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 229 {
 230         return __fuse_get_req(fc, npages, false);
 231 }
 232 EXPORT_SYMBOL_GPL(fuse_get_req);
 233
 234 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 235                                              unsigned npages)
 236 {
 237         return __fuse_get_req(fc, npages, true);
 238 }
 239 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 240
 241 /*
 242  * Return request in fuse_file->reserved_req.  However that may
 243  * currently be in use.  If that is the case, wait for it to become
 244  * available.
 245  */
 246 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 247                                          struct file *file)
 248 {
 249         struct fuse_req *req = NULL;
 250         struct fuse_file *ff = file->private_data;
 251
 252         do {
 253                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 254                 spin_lock(&fc->lock);
 255                 if (ff->reserved_req) {
 256                         req = ff->reserved_req;
 257                         ff->reserved_req = NULL;
 258                         req->stolen_file = get_file(file);
 259                 }
 260                 spin_unlock(&fc->lock);
 261         } while (!req);
 262
 263         return req;
 264 }
 265
 266 /*
 267  * Put stolen request back into fuse_file->reserved_req
 268  */
 269 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 270 {
 271         struct file *file = req->stolen_file;
 272         struct fuse_file *ff = file->private_data;
 273
 274         WARN_ON(req->max_pages);
 275         spin_lock(&fc->lock);
 276         memset(req, 0, sizeof(*req));
 277         fuse_request_init(req, NULL, NULL, 0);
 278         BUG_ON(ff->reserved_req);
 279         ff->reserved_req = req;
 280         wake_up_all(&fc->reserved_req_waitq);
 281         spin_unlock(&fc->lock);
 282         fput(file);
 283 }
 284
 285 /*
 286  * Gets a requests for a file operation, always succeeds
 287  *
 288  * This is used for sending the FLUSH request, which must get to
 289  * userspace, due to POSIX locks which may need to be unlocked.
 290  *
 291  * If allocation fails due to OOM, use the reserved request in
 292  * fuse_file.
 293  *
 294  * This is very unlikely to deadlock accidentally, since the
 295  * filesystem should not have it's own file open.  If deadlock is
 296  * intentional, it can still be broken by "aborting" the filesystem.
 297  */
 298 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 299                                              struct file *file)
 300 {
 301         struct fuse_req *req;
 302
 303         atomic_inc(&fc->num_waiting);
 304         wait_event(fc->blocked_waitq, fc->initialized);
 305         /* Matches smp_wmb() in fuse_set_initialized() */
 306         smp_rmb();
 307         req = fuse_request_alloc(0);
 308         if (!req)
 309                 req = get_reserved_req(fc, file);
 310
 311         req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
 312         req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
 313         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 314
 315         __set_bit(FR_WAITING, &req->flags);
 316         __clear_bit(FR_BACKGROUND, &req->flags);
 317         return req;
 318 }
 319
 320 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 321 {
 322         if (refcount_dec_and_test(&req->count)) {
 323                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 324                         /*
 325                          * We get here in the unlikely case that a background
 326                          * request was allocated but not sent
 327                          */
 328                         spin_lock(&fc->bg_lock);
 329                         if (!fc->blocked)
 330                                 wake_up(&fc->blocked_waitq);
 331                         spin_unlock(&fc->bg_lock);
 332                 }
 333
 334                 if (test_bit(FR_WAITING, &req->flags)) {
 335                         __clear_bit(FR_WAITING, &req->flags);
 336                         fuse_drop_waiting(fc);
 337                 }
 338
 339                 if (req->stolen_file)
 340                         put_reserved_req(fc, req);
 341                 else
 342                         fuse_request_free(req);
 343         }
 344 }
 345 EXPORT_SYMBOL_GPL(fuse_put_request);
 346
 347 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 348 {
 349         unsigned nbytes = 0;
 350         unsigned i;
 351
 352         for (i = 0; i < numargs; i++)
 353                 nbytes += args[i].size;
 354
 355         return nbytes;
 356 }
 357
 358 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 359 {
 360         fiq->reqctr += FUSE_REQ_ID_STEP;
 361         return fiq->reqctr;
 362 }
 363
 364 static unsigned int fuse_req_hash(u64 unique)
 365 {
 366         return hash_long(unique & ~FUSE_INT_REQ_BIT, FUSE_PQ_HASH_BITS);
 367 }
 368
 369 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 370 {
 371         req->in.h.len = sizeof(struct fuse_in_header) +
 372                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 373         list_add_tail(&req->list, &fiq->pending);
 374         wake_up_locked(&fiq->waitq);
 375         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 376 }
 377
 378 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 379                        u64 nodeid, u64 nlookup)
 380 {
 381         struct fuse_iqueue *fiq = &fc->iq;
 382
 383         forget->forget_one.nodeid = nodeid;
 384         forget->forget_one.nlookup = nlookup;
 385
 386         spin_lock(&fiq->waitq.lock);
 387         if (fiq->connected) {
 388                 fiq->forget_list_tail->next = forget;
 389                 fiq->forget_list_tail = forget;
 390                 wake_up_locked(&fiq->waitq);
 391                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 392         } else {
 393                 kfree(forget);
 394         }
 395         spin_unlock(&fiq->waitq.lock);
 396 }
 397
 398 static void flush_bg_queue(struct fuse_conn *fc)
 399 {
 400         struct fuse_iqueue *fiq = &fc->iq;
 401
 402         while (fc->active_background < fc->max_background &&
 403                !list_empty(&fc->bg_queue)) {
 404                 struct fuse_req *req;
 405
 406                 req = list_first_entry(&fc->bg_queue, struct fuse_req, list);
 407                 list_del(&req->list);
 408                 fc->active_background++;
 409                 spin_lock(&fiq->waitq.lock);
 410                 req->in.h.unique = fuse_get_unique(fiq);
 411                 queue_request(fiq, req);
 412                 spin_unlock(&fiq->waitq.lock);
 413         }
 414 }
 415
 416 /*
 417  * This function is called when a request is finished.  Either a reply
 418  * has arrived or it was aborted (and not yet sent) or some error
 419  * occurred during communication with userspace, or the device file
 420  * was closed.  The requester thread is woken up (if still waiting),
 421  * the 'end' callback is called if given, else the reference to the
 422  * request is released
 423  */
 424 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 425 {
 426         struct fuse_iqueue *fiq = &fc->iq;
 427
 428         if (test_and_set_bit(FR_FINISHED, &req->flags))
 429                 goto put_request;
 430
 431         spin_lock(&fiq->waitq.lock);
 432         list_del_init(&req->intr_entry);
 433         spin_unlock(&fiq->waitq.lock);
 434         WARN_ON(test_bit(FR_PENDING, &req->flags));
 435         WARN_ON(test_bit(FR_SENT, &req->flags));
 436         if (test_bit(FR_BACKGROUND, &req->flags)) {
 437                 spin_lock(&fc->bg_lock);
 438                 clear_bit(FR_BACKGROUND, &req->flags);
 439                 if (fc->num_background == fc->max_background) {
 440                         fc->blocked = 0;
 441                         wake_up(&fc->blocked_waitq);
 442                 } else if (!fc->blocked) {
 443                         /*
 444                          * Wake up next waiter, if any.  It's okay to use
 445                          * waitqueue_active(), as we've already synced up
 446                          * fc->blocked with waiters with the wake_up() call
 447                          * above.
 448                          */
 449                         if (waitqueue_active(&fc->blocked_waitq))
 450                                 wake_up(&fc->blocked_waitq);
 451                 }
 452
 453                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 454                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 455                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 456                 }
 457                 fc->num_background--;
 458                 fc->active_background--;
 459                 flush_bg_queue(fc);
 460                 spin_unlock(&fc->bg_lock);
 461         }
 462         wake_up(&req->waitq);
 463         if (req->end)
 464                 req->end(fc, req);
 465 put_request:
 466         fuse_put_request(fc, req);
 467 }
 468
 469 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 470 {
 471         spin_lock(&fiq->waitq.lock);
 472         if (test_bit(FR_FINISHED, &req->flags)) {
 473                 spin_unlock(&fiq->waitq.lock);
 474                 return;
 475         }
 476         if (list_empty(&req->intr_entry)) {
 477                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 478                 wake_up_locked(&fiq->waitq);
 479         }
 480         spin_unlock(&fiq->waitq.lock);
 481         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 482 }
 483
 484 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 485 {
 486         struct fuse_iqueue *fiq = &fc->iq;
 487         int err;
 488
 489         if (!fc->no_interrupt) {
 490                 /* Any signal may interrupt this */
 491                 err = wait_event_interruptible(req->waitq,
 492                                         test_bit(FR_FINISHED, &req->flags));
 493                 if (!err)
 494                         return;
 495
 496                 set_bit(FR_INTERRUPTED, &req->flags);
 497                 /* matches barrier in fuse_dev_do_read() */
 498                 smp_mb__after_atomic();
 499                 if (test_bit(FR_SENT, &req->flags))
 500                         queue_interrupt(fiq, req);
 501         }
 502
 503         if (!test_bit(FR_FORCE, &req->flags)) {
 504                 /* Only fatal signals may interrupt this */
 505                 err = wait_event_killable(req->waitq,
 506                                         test_bit(FR_FINISHED, &req->flags));
 507                 if (!err)
 508                         return;
 509
 510                 spin_lock(&fiq->waitq.lock);
 511                 /* Request is not yet in userspace, bail out */
 512                 if (test_bit(FR_PENDING, &req->flags)) {
 513                         list_del(&req->list);
 514                         spin_unlock(&fiq->waitq.lock);
 515                         __fuse_put_request(req);
 516                         req->out.h.error = -EINTR;
 517                         return;
 518                 }
 519                 spin_unlock(&fiq->waitq.lock);
 520         }
 521
 522         /*
 523          * Either request is already in userspace, or it was forced.
 524          * Wait it out.
 525          */
 526         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 527 }
 528
 529 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 530 {
 531         struct fuse_iqueue *fiq = &fc->iq;
 532
 533         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 534         spin_lock(&fiq->waitq.lock);
 535         if (!fiq->connected) {
 536                 spin_unlock(&fiq->waitq.lock);
 537                 req->out.h.error = -ENOTCONN;
 538         } else {
 539                 req->in.h.unique = fuse_get_unique(fiq);
 540                 queue_request(fiq, req);
 541                 /* acquire extra reference, since request is still needed
 542                    after request_end() */
 543                 __fuse_get_request(req);
 544                 spin_unlock(&fiq->waitq.lock);
 545
 546                 request_wait_answer(fc, req);
 547                 /* Pairs with smp_wmb() in request_end() */
 548                 smp_rmb();
 549         }
 550 }
 551
 552 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 553 {
 554         __set_bit(FR_ISREPLY, &req->flags);
 555         if (!test_bit(FR_WAITING, &req->flags)) {
 556                 __set_bit(FR_WAITING, &req->flags);
 557                 atomic_inc(&fc->num_waiting);
 558         }
 559         __fuse_request_send(fc, req);
 560 }
 561 EXPORT_SYMBOL_GPL(fuse_request_send);
 562
 563 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 564 {
 565         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 566                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 567
 568         if (fc->minor < 9) {
 569                 switch (args->in.h.opcode) {
 570                 case FUSE_LOOKUP:
 571                 case FUSE_CREATE:
 572                 case FUSE_MKNOD:
 573                 case FUSE_MKDIR:
 574                 case FUSE_SYMLINK:
 575                 case FUSE_LINK:
 576                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 577                         break;
 578                 case FUSE_GETATTR:
 579                 case FUSE_SETATTR:
 580                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 581                         break;
 582                 }
 583         }
 584         if (fc->minor < 12) {
 585                 switch (args->in.h.opcode) {
 586                 case FUSE_CREATE:
 587                         args->in.args[0].size = sizeof(struct fuse_open_in);
 588                         break;
 589                 case FUSE_MKNOD:
 590                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 591                         break;
 592                 }
 593         }
 594 }
 595
 596 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 597 {
 598         struct fuse_req *req;
 599         ssize_t ret;
 600
 601         req = fuse_get_req(fc, 0);
 602         if (IS_ERR(req))
 603                 return PTR_ERR(req);
 604
 605         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 606         fuse_adjust_compat(fc, args);
 607
 608         req->in.h.opcode = args->in.h.opcode;
 609         req->in.h.nodeid = args->in.h.nodeid;
 610         req->in.numargs = args->in.numargs;
 611         memcpy(req->in.args, args->in.args,
 612                args->in.numargs * sizeof(struct fuse_in_arg));
 613         req->out.argvar = args->out.argvar;
 614         req->out.numargs = args->out.numargs;
 615         memcpy(req->out.args, args->out.args,
 616                args->out.numargs * sizeof(struct fuse_arg));
 617         fuse_request_send(fc, req);
 618         ret = req->out.h.error;
 619         if (!ret && args->out.argvar) {
 620                 BUG_ON(args->out.numargs != 1);
 621                 ret = req->out.args[0].size;
 622         }
 623         fuse_put_request(fc, req);
 624
 625         return ret;
 626 }
 627
 628 bool fuse_request_queue_background(struct fuse_conn *fc, struct fuse_req *req)
 629 {
 630         bool queued = false;
 631
 632         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 633         if (!test_bit(FR_WAITING, &req->flags)) {
 634                 __set_bit(FR_WAITING, &req->flags);
 635                 atomic_inc(&fc->num_waiting);
 636         }
 637         __set_bit(FR_ISREPLY, &req->flags);
 638         spin_lock(&fc->bg_lock);
 639         if (likely(fc->connected)) {
 640                 fc->num_background++;
 641                 if (fc->num_background == fc->max_background)
 642                         fc->blocked = 1;
 643                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 644                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 645                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 646                 }
 647                 list_add_tail(&req->list, &fc->bg_queue);
 648                 flush_bg_queue(fc);
 649                 queued = true;
 650         }
 651         spin_unlock(&fc->bg_lock);
 652
 653         return queued;
 654 }
 655
 656 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 657 {
 658         WARN_ON(!req->end);
 659         if (!fuse_request_queue_background(fc, req)) {
 660                 req->out.h.error = -ENOTCONN;
 661                 req->end(fc, req);
 662                 fuse_put_request(fc, req);
 663         }
 664 }
 665 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 666
 667 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 668                                           struct fuse_req *req, u64 unique)
 669 {
 670         int err = -ENODEV;
 671         struct fuse_iqueue *fiq = &fc->iq;
 672
 673         __clear_bit(FR_ISREPLY, &req->flags);
 674         req->in.h.unique = unique;
 675         spin_lock(&fiq->waitq.lock);
 676         if (fiq->connected) {
 677                 queue_request(fiq, req);
 678                 err = 0;
 679         }
 680         spin_unlock(&fiq->waitq.lock);
 681
 682         return err;
 683 }
 684
 685 void fuse_force_forget(struct file *file, u64 nodeid)
 686 {
 687         struct inode *inode = file_inode(file);
 688         struct fuse_conn *fc = get_fuse_conn(inode);
 689         struct fuse_req *req;
 690         struct fuse_forget_in inarg;
 691
 692         memset(&inarg, 0, sizeof(inarg));
 693         inarg.nlookup = 1;
 694         req = fuse_get_req_nofail_nopages(fc, file);
 695         req->in.h.opcode = FUSE_FORGET;
 696         req->in.h.nodeid = nodeid;
 697         req->in.numargs = 1;
 698         req->in.args[0].size = sizeof(inarg);
 699         req->in.args[0].value = &inarg;
 700         __clear_bit(FR_ISREPLY, &req->flags);
 701         __fuse_request_send(fc, req);
 702         /* ignore errors */
 703         fuse_put_request(fc, req);
 704 }
 705
 706 /*
 707  * Lock the request.  Up to the next unlock_request() there mustn't be
 708  * anything that could cause a page-fault.  If the request was already
 709  * aborted bail out.
 710  */
 711 static int lock_request(struct fuse_req *req)
 712 {
 713         int err = 0;
 714         if (req) {
 715                 spin_lock(&req->waitq.lock);
 716                 if (test_bit(FR_ABORTED, &req->flags))
 717                         err = -ENOENT;
 718                 else
 719                         set_bit(FR_LOCKED, &req->flags);
 720                 spin_unlock(&req->waitq.lock);
 721         }
 722         return err;
 723 }
 724
 725 /*
 726  * Unlock request.  If it was aborted while locked, caller is responsible
 727  * for unlocking and ending the request.
 728  */
 729 static int unlock_request(struct fuse_req *req)
 730 {
 731         int err = 0;
 732         if (req) {
 733                 spin_lock(&req->waitq.lock);
 734                 if (test_bit(FR_ABORTED, &req->flags))
 735                         err = -ENOENT;
 736                 else
 737                         clear_bit(FR_LOCKED, &req->flags);
 738                 spin_unlock(&req->waitq.lock);
 739         }
 740         return err;
 741 }
 742
 743 struct fuse_copy_state {
 744         int write;
 745         struct fuse_req *req;
 746         struct iov_iter *iter;
 747         struct pipe_buffer *pipebufs;
 748         struct pipe_buffer *currbuf;
 749         struct pipe_inode_info *pipe;
 750         unsigned long nr_segs;
 751         struct page *pg;
 752         unsigned len;
 753         unsigned offset;
 754         unsigned move_pages:1;
 755 };
 756
 757 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 758                            struct iov_iter *iter)
 759 {
 760         memset(cs, 0, sizeof(*cs));
 761         cs->write = write;
 762         cs->iter = iter;
 763 }
 764
 765 /* Unmap and put previous page of userspace buffer */
 766 static void fuse_copy_finish(struct fuse_copy_state *cs)
 767 {
 768         if (cs->currbuf) {
 769                 struct pipe_buffer *buf = cs->currbuf;
 770
 771                 if (cs->write)
 772                         buf->len = PAGE_SIZE - cs->len;
 773                 cs->currbuf = NULL;
 774         } else if (cs->pg) {
 775                 if (cs->write) {
 776                         flush_dcache_page(cs->pg);
 777                         set_page_dirty_lock(cs->pg);
 778                 }
 779                 put_page(cs->pg);
 780         }
 781         cs->pg = NULL;
 782 }
 783
 784 /*
 785  * Get another pagefull of userspace buffer, and map it to kernel
 786  * address space, and lock request
 787  */
 788 static int fuse_copy_fill(struct fuse_copy_state *cs)
 789 {
 790         struct page *page;
 791         int err;
 792
 793         err = unlock_request(cs->req);
 794         if (err)
 795                 return err;
 796
 797         fuse_copy_finish(cs);
 798         if (cs->pipebufs) {
 799                 struct pipe_buffer *buf = cs->pipebufs;
 800
 801                 if (!cs->write) {
 802                         err = pipe_buf_confirm(cs->pipe, buf);
 803                         if (err)
 804                                 return err;
 805
 806                         BUG_ON(!cs->nr_segs);
 807                         cs->currbuf = buf;
 808                         cs->pg = buf->page;
 809                         cs->offset = buf->offset;
 810                         cs->len = buf->len;
 811                         cs->pipebufs++;
 812                         cs->nr_segs--;
 813                 } else {
 814                         if (cs->nr_segs == cs->pipe->buffers)
 815                                 return -EIO;
 816
 817                         page = alloc_page(GFP_HIGHUSER);
 818                         if (!page)
 819                                 return -ENOMEM;
 820
 821                         buf->page = page;
 822                         buf->offset = 0;
 823                         buf->len = 0;
 824
 825                         cs->currbuf = buf;
 826                         cs->pg = page;
 827                         cs->offset = 0;
 828                         cs->len = PAGE_SIZE;
 829                         cs->pipebufs++;
 830                         cs->nr_segs++;
 831                 }
 832         } else {
 833                 size_t off;
 834                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 835                 if (err < 0)
 836                         return err;
 837                 BUG_ON(!err);
 838                 cs->len = err;
 839                 cs->offset = off;
 840                 cs->pg = page;
 841                 iov_iter_advance(cs->iter, err);
 842         }
 843
 844         return lock_request(cs->req);
 845 }
 846
 847 /* Do as much copy to/from userspace buffer as we can */
 848 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 849 {
 850         unsigned ncpy = min(*size, cs->len);
 851         if (val) {
 852                 void *pgaddr = kmap_atomic(cs->pg);
 853                 void *buf = pgaddr + cs->offset;
 854
 855                 if (cs->write)
 856                         memcpy(buf, *val, ncpy);
 857                 else
 858                         memcpy(*val, buf, ncpy);
 859
 860                 kunmap_atomic(pgaddr);
 861                 *val += ncpy;
 862         }
 863         *size -= ncpy;
 864         cs->len -= ncpy;
 865         cs->offset += ncpy;
 866         return ncpy;
 867 }
 868
 869 static int fuse_check_page(struct page *page)
 870 {
 871         if (page_mapcount(page) ||
 872             page->mapping != NULL ||
 873             page_count(page) != 1 ||
 874             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 875              ~(1 << PG_locked |
 876                1 << PG_referenced |
 877                1 << PG_uptodate |
 878                1 << PG_lru |
 879                1 << PG_active |
 880                1 << PG_reclaim))) {
 881                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 882                 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);
 883                 return 1;
 884         }
 885         return 0;
 886 }
 887
 888 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 889 {
 890         int err;
 891         struct page *oldpage = *pagep;
 892         struct page *newpage;
 893         struct pipe_buffer *buf = cs->pipebufs;
 894
 895         err = unlock_request(cs->req);
 896         if (err)
 897                 return err;
 898
 899         fuse_copy_finish(cs);
 900
 901         err = pipe_buf_confirm(cs->pipe, buf);
 902         if (err)
 903                 return err;
 904
 905         BUG_ON(!cs->nr_segs);
 906         cs->currbuf = buf;
 907         cs->len = buf->len;
 908         cs->pipebufs++;
 909         cs->nr_segs--;
 910
 911         if (cs->len != PAGE_SIZE)
 912                 goto out_fallback;
 913
 914         if (pipe_buf_steal(cs->pipe, buf) != 0)
 915                 goto out_fallback;
 916
 917         newpage = buf->page;
 918
 919         if (!PageUptodate(newpage))
 920                 SetPageUptodate(newpage);
 921
 922         ClearPageMappedToDisk(newpage);
 923
 924         if (fuse_check_page(newpage) != 0)
 925                 goto out_fallback_unlock;
 926
 927         /*
 928          * This is a new and locked page, it shouldn't be mapped or
 929          * have any special flags on it
 930          */
 931         if (WARN_ON(page_mapped(oldpage)))
 932                 goto out_fallback_unlock;
 933         if (WARN_ON(page_has_private(oldpage)))
 934                 goto out_fallback_unlock;
 935         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 936                 goto out_fallback_unlock;
 937         if (WARN_ON(PageMlocked(oldpage)))
 938                 goto out_fallback_unlock;
 939
 940         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 941         if (err) {
 942                 unlock_page(newpage);
 943                 return err;
 944         }
 945
 946         get_page(newpage);
 947
 948         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 949                 lru_cache_add_file(newpage);
 950
 951         err = 0;
 952         spin_lock(&cs->req->waitq.lock);
 953         if (test_bit(FR_ABORTED, &cs->req->flags))
 954                 err = -ENOENT;
 955         else
 956                 *pagep = newpage;
 957         spin_unlock(&cs->req->waitq.lock);
 958
 959         if (err) {
 960                 unlock_page(newpage);
 961                 put_page(newpage);
 962                 return err;
 963         }
 964
 965         unlock_page(oldpage);
 966         put_page(oldpage);
 967         cs->len = 0;
 968
 969         return 0;
 970
 971 out_fallback_unlock:
 972         unlock_page(newpage);
 973 out_fallback:
 974         cs->pg = buf->page;
 975         cs->offset = buf->offset;
 976
 977         err = lock_request(cs->req);
 978         if (err)
 979                 return err;
 980
 981         return 1;
 982 }
 983
 984 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 985                          unsigned offset, unsigned count)
 986 {
 987         struct pipe_buffer *buf;
 988         int err;
 989
 990         if (cs->nr_segs == cs->pipe->buffers)
 991                 return -EIO;
 992
 993         err = unlock_request(cs->req);
 994         if (err)
 995                 return err;
 996
 997         fuse_copy_finish(cs);
 998
 999         buf = cs->pipebufs;
1000         get_page(page);
1001         buf->page = page;
1002         buf->offset = offset;
1003         buf->len = count;
1004
1005         cs->pipebufs++;
1006         cs->nr_segs++;
1007         cs->len = 0;
1008
1009         return 0;
1010 }
1011
1012 /*
1013  * Copy a page in the request to/from the userspace buffer.  Must be
1014  * done atomically
1015  */
1016 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
1017                           unsigned offset, unsigned count, int zeroing)
1018 {
1019         int err;
1020         struct page *page = *pagep;
1021
1022         if (page && zeroing && count < PAGE_SIZE)
1023                 clear_highpage(page);
1024
1025         while (count) {
1026                 if (cs->write && cs->pipebufs && page) {
1027                         return fuse_ref_page(cs, page, offset, count);
1028                 } else if (!cs->len) {
1029                         if (cs->move_pages && page &&
1030                             offset == 0 && count == PAGE_SIZE) {
1031                                 err = fuse_try_move_page(cs, pagep);
1032                                 if (err <= 0)
1033                                         return err;
1034                         } else {
1035                                 err = fuse_copy_fill(cs);
1036                                 if (err)
1037                                         return err;
1038                         }
1039                 }
1040                 if (page) {
1041                         void *mapaddr = kmap_atomic(page);
1042                         void *buf = mapaddr + offset;
1043                         offset += fuse_copy_do(cs, &buf, &count);
1044                         kunmap_atomic(mapaddr);
1045                 } else
1046                         offset += fuse_copy_do(cs, NULL, &count);
1047         }
1048         if (page && !cs->write)
1049                 flush_dcache_page(page);
1050         return 0;
1051 }
1052
1053 /* Copy pages in the request to/from userspace buffer */
1054 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1055                            int zeroing)
1056 {
1057         unsigned i;
1058         struct fuse_req *req = cs->req;
1059
1060         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1061                 int err;
1062                 unsigned offset = req->page_descs[i].offset;
1063                 unsigned count = min(nbytes, req->page_descs[i].length);
1064
1065                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1066                                      zeroing);
1067                 if (err)
1068                         return err;
1069
1070                 nbytes -= count;
1071         }
1072         return 0;
1073 }
1074
1075 /* Copy a single argument in the request to/from userspace buffer */
1076 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1077 {
1078         while (size) {
1079                 if (!cs->len) {
1080                         int err = fuse_copy_fill(cs);
1081                         if (err)
1082                                 return err;
1083                 }
1084                 fuse_copy_do(cs, &val, &size);
1085         }
1086         return 0;
1087 }
1088
1089 /* Copy request arguments to/from userspace buffer */
1090 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1091                           unsigned argpages, struct fuse_arg *args,
1092                           int zeroing)
1093 {
1094         int err = 0;
1095         unsigned i;
1096
1097         for (i = 0; !err && i < numargs; i++)  {
1098                 struct fuse_arg *arg = &args[i];
1099                 if (i == numargs - 1 && argpages)
1100                         err = fuse_copy_pages(cs, arg->size, zeroing);
1101                 else
1102                         err = fuse_copy_one(cs, arg->value, arg->size);
1103         }
1104         return err;
1105 }
1106
1107 static int forget_pending(struct fuse_iqueue *fiq)
1108 {
1109         return fiq->forget_list_head.next != NULL;
1110 }
1111
1112 static int request_pending(struct fuse_iqueue *fiq)
1113 {
1114         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1115                 forget_pending(fiq);
1116 }
1117
1118 /*
1119  * Transfer an interrupt request to userspace
1120  *
1121  * Unlike other requests this is assembled on demand, without a need
1122  * to allocate a separate fuse_req structure.
1123  *
1124  * Called with fiq->waitq.lock held, releases it
1125  */
1126 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1127                                struct fuse_copy_state *cs,
1128                                size_t nbytes, struct fuse_req *req)
1129 __releases(fiq->waitq.lock)
1130 {
1131         struct fuse_in_header ih;
1132         struct fuse_interrupt_in arg;
1133         unsigned reqsize = sizeof(ih) + sizeof(arg);
1134         int err;
1135
1136         list_del_init(&req->intr_entry);
1137         memset(&ih, 0, sizeof(ih));
1138         memset(&arg, 0, sizeof(arg));
1139         ih.len = reqsize;
1140         ih.opcode = FUSE_INTERRUPT;
1141         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1142         arg.unique = req->in.h.unique;
1143
1144         spin_unlock(&fiq->waitq.lock);
1145         if (nbytes < reqsize)
1146                 return -EINVAL;
1147
1148         err = fuse_copy_one(cs, &ih, sizeof(ih));
1149         if (!err)
1150                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1151         fuse_copy_finish(cs);
1152
1153         return err ? err : reqsize;
1154 }
1155
1156 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1157                                                unsigned max,
1158                                                unsigned *countp)
1159 {
1160         struct fuse_forget_link *head = fiq->forget_list_head.next;
1161         struct fuse_forget_link **newhead = &head;
1162         unsigned count;
1163
1164         for (count = 0; *newhead != NULL && count < max; count++)
1165                 newhead = &(*newhead)->next;
1166
1167         fiq->forget_list_head.next = *newhead;
1168         *newhead = NULL;
1169         if (fiq->forget_list_head.next == NULL)
1170                 fiq->forget_list_tail = &fiq->forget_list_head;
1171
1172         if (countp != NULL)
1173                 *countp = count;
1174
1175         return head;
1176 }
1177
1178 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1179                                    struct fuse_copy_state *cs,
1180                                    size_t nbytes)
1181 __releases(fiq->waitq.lock)
1182 {
1183         int err;
1184         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1185         struct fuse_forget_in arg = {
1186                 .nlookup = forget->forget_one.nlookup,
1187         };
1188         struct fuse_in_header ih = {
1189                 .opcode = FUSE_FORGET,
1190                 .nodeid = forget->forget_one.nodeid,
1191                 .unique = fuse_get_unique(fiq),
1192                 .len = sizeof(ih) + sizeof(arg),
1193         };
1194
1195         spin_unlock(&fiq->waitq.lock);
1196         kfree(forget);
1197         if (nbytes < ih.len)
1198                 return -EINVAL;
1199
1200         err = fuse_copy_one(cs, &ih, sizeof(ih));
1201         if (!err)
1202                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1203         fuse_copy_finish(cs);
1204
1205         if (err)
1206                 return err;
1207
1208         return ih.len;
1209 }
1210
1211 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1212                                    struct fuse_copy_state *cs, size_t nbytes)
1213 __releases(fiq->waitq.lock)
1214 {
1215         int err;
1216         unsigned max_forgets;
1217         unsigned count;
1218         struct fuse_forget_link *head;
1219         struct fuse_batch_forget_in arg = { .count = 0 };
1220         struct fuse_in_header ih = {
1221                 .opcode = FUSE_BATCH_FORGET,
1222                 .unique = fuse_get_unique(fiq),
1223                 .len = sizeof(ih) + sizeof(arg),
1224         };
1225
1226         if (nbytes < ih.len) {
1227                 spin_unlock(&fiq->waitq.lock);
1228                 return -EINVAL;
1229         }
1230
1231         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1232         head = dequeue_forget(fiq, max_forgets, &count);
1233         spin_unlock(&fiq->waitq.lock);
1234
1235         arg.count = count;
1236         ih.len += count * sizeof(struct fuse_forget_one);
1237         err = fuse_copy_one(cs, &ih, sizeof(ih));
1238         if (!err)
1239                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1240
1241         while (head) {
1242                 struct fuse_forget_link *forget = head;
1243
1244                 if (!err) {
1245                         err = fuse_copy_one(cs, &forget->forget_one,
1246                                             sizeof(forget->forget_one));
1247                 }
1248                 head = forget->next;
1249                 kfree(forget);
1250         }
1251
1252         fuse_copy_finish(cs);
1253
1254         if (err)
1255                 return err;
1256
1257         return ih.len;
1258 }
1259
1260 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1261                             struct fuse_copy_state *cs,
1262                             size_t nbytes)
1263 __releases(fiq->waitq.lock)
1264 {
1265         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1266                 return fuse_read_single_forget(fiq, cs, nbytes);
1267         else
1268                 return fuse_read_batch_forget(fiq, cs, nbytes);
1269 }
1270
1271 /*
1272  * Read a single request into the userspace filesystem's buffer.  This
1273  * function waits until a request is available, then removes it from
1274  * the pending list and copies request data to userspace buffer.  If
1275  * no reply is needed (FORGET) or request has been aborted or there
1276  * was an error during the copying then it's finished by calling
1277  * request_end().  Otherwise add it to the processing list, and set
1278  * the 'sent' flag.
1279  */
1280 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1281                                 struct fuse_copy_state *cs, size_t nbytes)
1282 {
1283         ssize_t err;
1284         struct fuse_conn *fc = fud->fc;
1285         struct fuse_iqueue *fiq = &fc->iq;
1286         struct fuse_pqueue *fpq = &fud->pq;
1287         struct fuse_req *req;
1288         struct fuse_in *in;
1289         unsigned reqsize;
1290         unsigned int hash;
1291
1292  restart:
1293         spin_lock(&fiq->waitq.lock);
1294         err = -EAGAIN;
1295         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1296             !request_pending(fiq))
1297                 goto err_unlock;
1298
1299         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1300                                 !fiq->connected || request_pending(fiq));
1301         if (err)
1302                 goto err_unlock;
1303
1304         if (!fiq->connected) {
1305                 err = (fc->aborted && fc->abort_err) ? -ECONNABORTED : -ENODEV;
1306                 goto err_unlock;
1307         }
1308
1309         if (!list_empty(&fiq->interrupts)) {
1310                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1311                                  intr_entry);
1312                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1313         }
1314
1315         if (forget_pending(fiq)) {
1316                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1317                         return fuse_read_forget(fc, fiq, cs, nbytes);
1318
1319                 if (fiq->forget_batch <= -8)
1320                         fiq->forget_batch = 16;
1321         }
1322
1323         req = list_entry(fiq->pending.next, struct fuse_req, list);
1324         clear_bit(FR_PENDING, &req->flags);
1325         list_del_init(&req->list);
1326         spin_unlock(&fiq->waitq.lock);
1327
1328         in = &req->in;
1329         reqsize = in->h.len;
1330
1331         /* If request is too large, reply with an error and restart the read */
1332         if (nbytes < reqsize) {
1333                 req->out.h.error = -EIO;
1334                 /* SETXATTR is special, since it may contain too large data */
1335                 if (in->h.opcode == FUSE_SETXATTR)
1336                         req->out.h.error = -E2BIG;
1337                 request_end(fc, req);
1338                 goto restart;
1339         }
1340         spin_lock(&fpq->lock);
1341         list_add(&req->list, &fpq->io);
1342         spin_unlock(&fpq->lock);
1343         cs->req = req;
1344         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1345         if (!err)
1346                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1347                                      (struct fuse_arg *) in->args, 0);
1348         fuse_copy_finish(cs);
1349         spin_lock(&fpq->lock);
1350         clear_bit(FR_LOCKED, &req->flags);
1351         if (!fpq->connected) {
1352                 err = (fc->aborted && fc->abort_err) ? -ECONNABORTED : -ENODEV;
1353                 goto out_end;
1354         }
1355         if (err) {
1356                 req->out.h.error = -EIO;
1357                 goto out_end;
1358         }
1359         if (!test_bit(FR_ISREPLY, &req->flags)) {
1360                 err = reqsize;
1361                 goto out_end;
1362         }
1363         hash = fuse_req_hash(req->in.h.unique);
1364         list_move_tail(&req->list, &fpq->processing[hash]);
1365         __fuse_get_request(req);
1366         set_bit(FR_SENT, &req->flags);
1367         spin_unlock(&fpq->lock);
1368         /* matches barrier in request_wait_answer() */
1369         smp_mb__after_atomic();
1370         if (test_bit(FR_INTERRUPTED, &req->flags))
1371                 queue_interrupt(fiq, req);
1372         fuse_put_request(fc, req);
1373
1374         return reqsize;
1375
1376 out_end:
1377         if (!test_bit(FR_PRIVATE, &req->flags))
1378                 list_del_init(&req->list);
1379         spin_unlock(&fpq->lock);
1380         request_end(fc, req);
1381         return err;
1382
1383  err_unlock:
1384         spin_unlock(&fiq->waitq.lock);
1385         return err;
1386 }
1387
1388 static int fuse_dev_open(struct inode *inode, struct file *file)
1389 {
1390         /*
1391          * The fuse device's file's private_data is used to hold
1392          * the fuse_conn(ection) when it is mounted, and is used to
1393          * keep track of whether the file has been mounted already.
1394          */
1395         file->private_data = NULL;
1396         return 0;
1397 }
1398
1399 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1400 {
1401         struct fuse_copy_state cs;
1402         struct file *file = iocb->ki_filp;
1403         struct fuse_dev *fud = fuse_get_dev(file);
1404
1405         if (!fud)
1406                 return -EPERM;
1407
1408         if (!iter_is_iovec(to))
1409                 return -EINVAL;
1410
1411         fuse_copy_init(&cs, 1, to);
1412
1413         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1414 }
1415
1416 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1417                                     struct pipe_inode_info *pipe,
1418                                     size_t len, unsigned int flags)
1419 {
1420         int total, ret;
1421         int page_nr = 0;
1422         struct pipe_buffer *bufs;
1423         struct fuse_copy_state cs;
1424         struct fuse_dev *fud = fuse_get_dev(in);
1425
1426         if (!fud)
1427                 return -EPERM;
1428
1429         bufs = kvmalloc_array(pipe->buffers, sizeof(struct pipe_buffer),
1430                               GFP_KERNEL);
1431         if (!bufs)
1432                 return -ENOMEM;
1433
1434         fuse_copy_init(&cs, 1, NULL);
1435         cs.pipebufs = bufs;
1436         cs.pipe = pipe;
1437         ret = fuse_dev_do_read(fud, in, &cs, len);
1438         if (ret < 0)
1439                 goto out;
1440
1441         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1442                 ret = -EIO;
1443                 goto out;
1444         }
1445
1446         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1447                 /*
1448                  * Need to be careful about this.  Having buf->ops in module
1449                  * code can Oops if the buffer persists after module unload.
1450                  */
1451                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1452                 bufs[page_nr].flags = 0;
1453                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1454                 if (unlikely(ret < 0))
1455                         break;
1456         }
1457         if (total)
1458                 ret = total;
1459 out:
1460         for (; page_nr < cs.nr_segs; page_nr++)
1461                 put_page(bufs[page_nr].page);
1462
1463         kvfree(bufs);
1464         return ret;
1465 }
1466
1467 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1468                             struct fuse_copy_state *cs)
1469 {
1470         struct fuse_notify_poll_wakeup_out outarg;
1471         int err = -EINVAL;
1472
1473         if (size != sizeof(outarg))
1474                 goto err;
1475
1476         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1477         if (err)
1478                 goto err;
1479
1480         fuse_copy_finish(cs);
1481         return fuse_notify_poll_wakeup(fc, &outarg);
1482
1483 err:
1484         fuse_copy_finish(cs);
1485         return err;
1486 }
1487
1488 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1489                                    struct fuse_copy_state *cs)
1490 {
1491         struct fuse_notify_inval_inode_out outarg;
1492         int err = -EINVAL;
1493
1494         if (size != sizeof(outarg))
1495                 goto err;
1496
1497         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1498         if (err)
1499                 goto err;
1500         fuse_copy_finish(cs);
1501
1502         down_read(&fc->killsb);
1503         err = -ENOENT;
1504         if (fc->sb) {
1505                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1506                                                outarg.off, outarg.len);
1507         }
1508         up_read(&fc->killsb);
1509         return err;
1510
1511 err:
1512         fuse_copy_finish(cs);
1513         return err;
1514 }
1515
1516 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1517                                    struct fuse_copy_state *cs)
1518 {
1519         struct fuse_notify_inval_entry_out outarg;
1520         int err = -ENOMEM;
1521         char *buf;
1522         struct qstr name;
1523
1524         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1525         if (!buf)
1526                 goto err;
1527
1528         err = -EINVAL;
1529         if (size < sizeof(outarg))
1530                 goto err;
1531
1532         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1533         if (err)
1534                 goto err;
1535
1536         err = -ENAMETOOLONG;
1537         if (outarg.namelen > FUSE_NAME_MAX)
1538                 goto err;
1539
1540         err = -EINVAL;
1541         if (size != sizeof(outarg) + outarg.namelen + 1)
1542                 goto err;
1543
1544         name.name = buf;
1545         name.len = outarg.namelen;
1546         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1547         if (err)
1548                 goto err;
1549         fuse_copy_finish(cs);
1550         buf[outarg.namelen] = 0;
1551
1552         down_read(&fc->killsb);
1553         err = -ENOENT;
1554         if (fc->sb)
1555                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1556         up_read(&fc->killsb);
1557         kfree(buf);
1558         return err;
1559
1560 err:
1561         kfree(buf);
1562         fuse_copy_finish(cs);
1563         return err;
1564 }
1565
1566 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1567                               struct fuse_copy_state *cs)
1568 {
1569         struct fuse_notify_delete_out outarg;
1570         int err = -ENOMEM;
1571         char *buf;
1572         struct qstr name;
1573
1574         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1575         if (!buf)
1576                 goto err;
1577
1578         err = -EINVAL;
1579         if (size < sizeof(outarg))
1580                 goto err;
1581
1582         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1583         if (err)
1584                 goto err;
1585
1586         err = -ENAMETOOLONG;
1587         if (outarg.namelen > FUSE_NAME_MAX)
1588                 goto err;
1589
1590         err = -EINVAL;
1591         if (size != sizeof(outarg) + outarg.namelen + 1)
1592                 goto err;
1593
1594         name.name = buf;
1595         name.len = outarg.namelen;
1596         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1597         if (err)
1598                 goto err;
1599         fuse_copy_finish(cs);
1600         buf[outarg.namelen] = 0;
1601
1602         down_read(&fc->killsb);
1603         err = -ENOENT;
1604         if (fc->sb)
1605                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1606                                                outarg.child, &name);
1607         up_read(&fc->killsb);
1608         kfree(buf);
1609         return err;
1610
1611 err:
1612         kfree(buf);
1613         fuse_copy_finish(cs);
1614         return err;
1615 }
1616
1617 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1618                              struct fuse_copy_state *cs)
1619 {
1620         struct fuse_notify_store_out outarg;
1621         struct inode *inode;
1622         struct address_space *mapping;
1623         u64 nodeid;
1624         int err;
1625         pgoff_t index;
1626         unsigned int offset;
1627         unsigned int num;
1628         loff_t file_size;
1629         loff_t end;
1630
1631         err = -EINVAL;
1632         if (size < sizeof(outarg))
1633                 goto out_finish;
1634
1635         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1636         if (err)
1637                 goto out_finish;
1638
1639         err = -EINVAL;
1640         if (size - sizeof(outarg) != outarg.size)
1641                 goto out_finish;
1642
1643         nodeid = outarg.nodeid;
1644
1645         down_read(&fc->killsb);
1646
1647         err = -ENOENT;
1648         if (!fc->sb)
1649                 goto out_up_killsb;
1650
1651         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1652         if (!inode)
1653                 goto out_up_killsb;
1654
1655         mapping = inode->i_mapping;
1656         index = outarg.offset >> PAGE_SHIFT;
1657         offset = outarg.offset & ~PAGE_MASK;
1658         file_size = i_size_read(inode);
1659         end = outarg.offset + outarg.size;
1660         if (end > file_size) {
1661                 file_size = end;
1662                 fuse_write_update_size(inode, file_size);
1663         }
1664
1665         num = outarg.size;
1666         while (num) {
1667                 struct page *page;
1668                 unsigned int this_num;
1669
1670                 err = -ENOMEM;
1671                 page = find_or_create_page(mapping, index,
1672                                            mapping_gfp_mask(mapping));
1673                 if (!page)
1674                         goto out_iput;
1675
1676                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1677                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1678                 if (!err && offset == 0 &&
1679                     (this_num == PAGE_SIZE || file_size == end))
1680                         SetPageUptodate(page);
1681                 unlock_page(page);
1682                 put_page(page);
1683
1684                 if (err)
1685                         goto out_iput;
1686
1687                 num -= this_num;
1688                 offset = 0;
1689                 index++;
1690         }
1691
1692         err = 0;
1693
1694 out_iput:
1695         iput(inode);
1696 out_up_killsb:
1697         up_read(&fc->killsb);
1698 out_finish:
1699         fuse_copy_finish(cs);
1700         return err;
1701 }
1702
1703 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1704 {
1705         release_pages(req->pages, req->num_pages);
1706 }
1707
1708 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1709                          struct fuse_notify_retrieve_out *outarg)
1710 {
1711         int err;
1712         struct address_space *mapping = inode->i_mapping;
1713         struct fuse_req *req;
1714         pgoff_t index;
1715         loff_t file_size;
1716         unsigned int num;
1717         unsigned int offset;
1718         size_t total_len = 0;
1719         unsigned int num_pages;
1720
1721         offset = outarg->offset & ~PAGE_MASK;
1722         file_size = i_size_read(inode);
1723
1724         num = outarg->size;
1725         if (outarg->offset > file_size)
1726                 num = 0;
1727         else if (outarg->offset + num > file_size)
1728                 num = file_size - outarg->offset;
1729
1730         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1731         num_pages = min(num_pages, fc->max_pages);
1732
1733         req = fuse_get_req(fc, num_pages);
1734         if (IS_ERR(req))
1735                 return PTR_ERR(req);
1736
1737         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1738         req->in.h.nodeid = outarg->nodeid;
1739         req->in.numargs = 2;
1740         req->in.argpages = 1;
1741         req->page_descs[0].offset = offset;
1742         req->end = fuse_retrieve_end;
1743
1744         index = outarg->offset >> PAGE_SHIFT;
1745
1746         while (num && req->num_pages < num_pages) {
1747                 struct page *page;
1748                 unsigned int this_num;
1749
1750                 page = find_get_page(mapping, index);
1751                 if (!page)
1752                         break;
1753
1754                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1755                 req->pages[req->num_pages] = page;
1756                 req->page_descs[req->num_pages].length = this_num;
1757                 req->num_pages++;
1758
1759                 offset = 0;
1760                 num -= this_num;
1761                 total_len += this_num;
1762                 index++;
1763         }
1764         req->misc.retrieve_in.offset = outarg->offset;
1765         req->misc.retrieve_in.size = total_len;
1766         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1767         req->in.args[0].value = &req->misc.retrieve_in;
1768         req->in.args[1].size = total_len;
1769
1770         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1771         if (err)
1772                 fuse_retrieve_end(fc, req);
1773
1774         return err;
1775 }
1776
1777 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1778                                 struct fuse_copy_state *cs)
1779 {
1780         struct fuse_notify_retrieve_out outarg;
1781         struct inode *inode;
1782         int err;
1783
1784         err = -EINVAL;
1785         if (size != sizeof(outarg))
1786                 goto copy_finish;
1787
1788         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1789         if (err)
1790                 goto copy_finish;
1791
1792         fuse_copy_finish(cs);
1793
1794         down_read(&fc->killsb);
1795         err = -ENOENT;
1796         if (fc->sb) {
1797                 u64 nodeid = outarg.nodeid;
1798
1799                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1800                 if (inode) {
1801                         err = fuse_retrieve(fc, inode, &outarg);
1802                         iput(inode);
1803                 }
1804         }
1805         up_read(&fc->killsb);
1806
1807         return err;
1808
1809 copy_finish:
1810         fuse_copy_finish(cs);
1811         return err;
1812 }
1813
1814 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1815                        unsigned int size, struct fuse_copy_state *cs)
1816 {
1817         /* Don't try to move pages (yet) */
1818         cs->move_pages = 0;
1819
1820         switch (code) {
1821         case FUSE_NOTIFY_POLL:
1822                 return fuse_notify_poll(fc, size, cs);
1823
1824         case FUSE_NOTIFY_INVAL_INODE:
1825                 return fuse_notify_inval_inode(fc, size, cs);
1826
1827         case FUSE_NOTIFY_INVAL_ENTRY:
1828                 return fuse_notify_inval_entry(fc, size, cs);
1829
1830         case FUSE_NOTIFY_STORE:
1831                 return fuse_notify_store(fc, size, cs);
1832
1833         case FUSE_NOTIFY_RETRIEVE:
1834                 return fuse_notify_retrieve(fc, size, cs);
1835
1836         case FUSE_NOTIFY_DELETE:
1837                 return fuse_notify_delete(fc, size, cs);
1838
1839         default:
1840                 fuse_copy_finish(cs);
1841                 return -EINVAL;
1842         }
1843 }
1844
1845 /* Look up request on processing list by unique ID */
1846 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1847 {
1848         unsigned int hash = fuse_req_hash(unique);
1849         struct fuse_req *req;
1850
1851         list_for_each_entry(req, &fpq->processing[hash], list) {
1852                 if (req->in.h.unique == unique)
1853                         return req;
1854         }
1855         return NULL;
1856 }
1857
1858 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1859                          unsigned nbytes)
1860 {
1861         unsigned reqsize = sizeof(struct fuse_out_header);
1862
1863         if (out->h.error)
1864                 return nbytes != reqsize ? -EINVAL : 0;
1865
1866         reqsize += len_args(out->numargs, out->args);
1867
1868         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1869                 return -EINVAL;
1870         else if (reqsize > nbytes) {
1871                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1872                 unsigned diffsize = reqsize - nbytes;
1873                 if (diffsize > lastarg->size)
1874                         return -EINVAL;
1875                 lastarg->size -= diffsize;
1876         }
1877         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1878                               out->page_zeroing);
1879 }
1880
1881 /*
1882  * Write a single reply to a request.  First the header is copied from
1883  * the write buffer.  The request is then searched on the processing
1884  * list by the unique ID found in the header.  If found, then remove
1885  * it from the list and copy the rest of the buffer to the request.
1886  * The request is finished by calling request_end()
1887  */
1888 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1889                                  struct fuse_copy_state *cs, size_t nbytes)
1890 {
1891         int err;
1892         struct fuse_conn *fc = fud->fc;
1893         struct fuse_pqueue *fpq = &fud->pq;
1894         struct fuse_req *req;
1895         struct fuse_out_header oh;
1896
1897         if (nbytes < sizeof(struct fuse_out_header))
1898                 return -EINVAL;
1899
1900         err = fuse_copy_one(cs, &oh, sizeof(oh));
1901         if (err)
1902                 goto err_finish;
1903
1904         err = -EINVAL;
1905         if (oh.len != nbytes)
1906                 goto err_finish;
1907
1908         /*
1909          * Zero oh.unique indicates unsolicited notification message
1910          * and error contains notification code.
1911          */
1912         if (!oh.unique) {
1913                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1914                 return err ? err : nbytes;
1915         }
1916
1917         err = -EINVAL;
1918         if (oh.error <= -1000 || oh.error > 0)
1919                 goto err_finish;
1920
1921         spin_lock(&fpq->lock);
1922         err = -ENOENT;
1923         if (!fpq->connected)
1924                 goto err_unlock_pq;
1925
1926         req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1927         if (!req)
1928                 goto err_unlock_pq;
1929
1930         /* Is it an interrupt reply ID? */
1931         if (oh.unique & FUSE_INT_REQ_BIT) {
1932                 __fuse_get_request(req);
1933                 spin_unlock(&fpq->lock);
1934
1935                 err = -EINVAL;
1936                 if (nbytes != sizeof(struct fuse_out_header)) {
1937                         fuse_put_request(fc, req);
1938                         goto err_finish;
1939                 }
1940
1941                 if (oh.error == -ENOSYS)
1942                         fc->no_interrupt = 1;
1943                 else if (oh.error == -EAGAIN)
1944                         queue_interrupt(&fc->iq, req);
1945                 fuse_put_request(fc, req);
1946
1947                 fuse_copy_finish(cs);
1948                 return nbytes;
1949         }
1950
1951         clear_bit(FR_SENT, &req->flags);
1952         list_move(&req->list, &fpq->io);
1953         req->out.h = oh;
1954         set_bit(FR_LOCKED, &req->flags);
1955         spin_unlock(&fpq->lock);
1956         cs->req = req;
1957         if (!req->out.page_replace)
1958                 cs->move_pages = 0;
1959
1960         err = copy_out_args(cs, &req->out, nbytes);
1961         fuse_copy_finish(cs);
1962
1963         spin_lock(&fpq->lock);
1964         clear_bit(FR_LOCKED, &req->flags);
1965         if (!fpq->connected)
1966                 err = -ENOENT;
1967         else if (err)
1968                 req->out.h.error = -EIO;
1969         if (!test_bit(FR_PRIVATE, &req->flags))
1970                 list_del_init(&req->list);
1971         spin_unlock(&fpq->lock);
1972
1973         request_end(fc, req);
1974
1975         return err ? err : nbytes;
1976
1977  err_unlock_pq:
1978         spin_unlock(&fpq->lock);
1979  err_finish:
1980         fuse_copy_finish(cs);
1981         return err;
1982 }
1983
1984 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1985 {
1986         struct fuse_copy_state cs;
1987         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1988
1989         if (!fud)
1990                 return -EPERM;
1991
1992         if (!iter_is_iovec(from))
1993                 return -EINVAL;
1994
1995         fuse_copy_init(&cs, 0, from);
1996
1997         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1998 }
1999
2000 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
2001                                      struct file *out, loff_t *ppos,
2002                                      size_t len, unsigned int flags)
2003 {
2004         unsigned nbuf;
2005         unsigned idx;
2006         struct pipe_buffer *bufs;
2007         struct fuse_copy_state cs;
2008         struct fuse_dev *fud;
2009         size_t rem;
2010         ssize_t ret;
2011
2012         fud = fuse_get_dev(out);
2013         if (!fud)
2014                 return -EPERM;
2015
2016         pipe_lock(pipe);
2017
2018         bufs = kvmalloc_array(pipe->nrbufs, sizeof(struct pipe_buffer),
2019                               GFP_KERNEL);
2020         if (!bufs) {
2021                 pipe_unlock(pipe);
2022                 return -ENOMEM;
2023         }
2024
2025         nbuf = 0;
2026         rem = 0;
2027         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2028                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2029
2030         ret = -EINVAL;
2031         if (rem < len) {
2032                 pipe_unlock(pipe);
2033                 goto out;
2034         }
2035
2036         rem = len;
2037         while (rem) {
2038                 struct pipe_buffer *ibuf;
2039                 struct pipe_buffer *obuf;
2040
2041                 BUG_ON(nbuf >= pipe->buffers);
2042                 BUG_ON(!pipe->nrbufs);
2043                 ibuf = &pipe->bufs[pipe->curbuf];
2044                 obuf = &bufs[nbuf];
2045
2046                 if (rem >= ibuf->len) {
2047                         *obuf = *ibuf;
2048                         ibuf->ops = NULL;
2049                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2050                         pipe->nrbufs--;
2051                 } else {
2052                         pipe_buf_get(pipe, ibuf);
2053                         *obuf = *ibuf;
2054                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2055                         obuf->len = rem;
2056                         ibuf->offset += obuf->len;
2057                         ibuf->len -= obuf->len;
2058                 }
2059                 nbuf++;
2060                 rem -= obuf->len;
2061         }
2062         pipe_unlock(pipe);
2063
2064         fuse_copy_init(&cs, 0, NULL);
2065         cs.pipebufs = bufs;
2066         cs.nr_segs = nbuf;
2067         cs.pipe = pipe;
2068
2069         if (flags & SPLICE_F_MOVE)
2070                 cs.move_pages = 1;
2071
2072         ret = fuse_dev_do_write(fud, &cs, len);
2073
2074         for (idx = 0; idx < nbuf; idx++)
2075                 pipe_buf_release(pipe, &bufs[idx]);
2076
2077 out:
2078         kvfree(bufs);
2079         return ret;
2080 }
2081
2082 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2083 {
2084         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2085         struct fuse_iqueue *fiq;
2086         struct fuse_dev *fud = fuse_get_dev(file);
2087
2088         if (!fud)
2089                 return EPOLLERR;
2090
2091         fiq = &fud->fc->iq;
2092         poll_wait(file, &fiq->waitq, wait);
2093
2094         spin_lock(&fiq->waitq.lock);
2095         if (!fiq->connected)
2096                 mask = EPOLLERR;
2097         else if (request_pending(fiq))
2098                 mask |= EPOLLIN | EPOLLRDNORM;
2099         spin_unlock(&fiq->waitq.lock);
2100
2101         return mask;
2102 }
2103
2104 /*
2105  * Abort all requests on the given list (pending or processing)
2106  *
2107  * This function releases and reacquires fc->lock
2108  */
2109 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2110 {
2111         while (!list_empty(head)) {
2112                 struct fuse_req *req;
2113                 req = list_entry(head->next, struct fuse_req, list);
2114                 req->out.h.error = -ECONNABORTED;
2115                 clear_bit(FR_SENT, &req->flags);
2116                 list_del_init(&req->list);
2117                 request_end(fc, req);
2118         }
2119 }
2120
2121 static void end_polls(struct fuse_conn *fc)
2122 {
2123         struct rb_node *p;
2124
2125         p = rb_first(&fc->polled_files);
2126
2127         while (p) {
2128                 struct fuse_file *ff;
2129                 ff = rb_entry(p, struct fuse_file, polled_node);
2130                 wake_up_interruptible_all(&ff->poll_wait);
2131
2132                 p = rb_next(p);
2133         }
2134 }
2135
2136 /*
2137  * Abort all requests.
2138  *
2139  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2140  * filesystem.
2141  *
2142  * The same effect is usually achievable through killing the filesystem daemon
2143  * and all users of the filesystem.  The exception is the combination of an
2144  * asynchronous request and the tricky deadlock (see
2145  * Documentation/filesystems/fuse.txt).
2146  *
2147  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2148  * requests, they should be finished off immediately.  Locked requests will be
2149  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2150  * requests.  It is possible that some request will finish before we can.  This
2151  * is OK, the request will in that case be removed from the list before we touch
2152  * it.
2153  */
2154 void fuse_abort_conn(struct fuse_conn *fc, bool is_abort)
2155 {
2156         struct fuse_iqueue *fiq = &fc->iq;
2157
2158         spin_lock(&fc->lock);
2159         if (fc->connected) {
2160                 struct fuse_dev *fud;
2161                 struct fuse_req *req, *next;
2162                 LIST_HEAD(to_end);
2163                 unsigned int i;
2164
2165                 /* Background queuing checks fc->connected under bg_lock */
2166                 spin_lock(&fc->bg_lock);
2167                 fc->connected = 0;
2168                 spin_unlock(&fc->bg_lock);
2169
2170                 fc->aborted = is_abort;
2171                 fuse_set_initialized(fc);
2172                 list_for_each_entry(fud, &fc->devices, entry) {
2173                         struct fuse_pqueue *fpq = &fud->pq;
2174
2175                         spin_lock(&fpq->lock);
2176                         fpq->connected = 0;
2177                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2178                                 req->out.h.error = -ECONNABORTED;
2179                                 spin_lock(&req->waitq.lock);
2180                                 set_bit(FR_ABORTED, &req->flags);
2181                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2182                                         set_bit(FR_PRIVATE, &req->flags);
2183                                         __fuse_get_request(req);
2184                                         list_move(&req->list, &to_end);
2185                                 }
2186                                 spin_unlock(&req->waitq.lock);
2187                         }
2188                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2189                                 list_splice_tail_init(&fpq->processing[i],
2190                                                       &to_end);
2191                         spin_unlock(&fpq->lock);
2192                 }
2193                 spin_lock(&fc->bg_lock);
2194                 fc->blocked = 0;
2195                 fc->max_background = UINT_MAX;
2196                 flush_bg_queue(fc);
2197                 spin_unlock(&fc->bg_lock);
2198
2199                 spin_lock(&fiq->waitq.lock);
2200                 fiq->connected = 0;
2201                 list_for_each_entry(req, &fiq->pending, list)
2202                         clear_bit(FR_PENDING, &req->flags);
2203                 list_splice_tail_init(&fiq->pending, &to_end);
2204                 while (forget_pending(fiq))
2205                         kfree(dequeue_forget(fiq, 1, NULL));
2206                 wake_up_all_locked(&fiq->waitq);
2207                 spin_unlock(&fiq->waitq.lock);
2208                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2209                 end_polls(fc);
2210                 wake_up_all(&fc->blocked_waitq);
2211                 spin_unlock(&fc->lock);
2212
2213                 end_requests(fc, &to_end);
2214         } else {
2215                 spin_unlock(&fc->lock);
2216         }
2217 }
2218 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2219
2220 void fuse_wait_aborted(struct fuse_conn *fc)
2221 {
2222         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2223 }
2224
2225 int fuse_dev_release(struct inode *inode, struct file *file)
2226 {
2227         struct fuse_dev *fud = fuse_get_dev(file);
2228
2229         if (fud) {
2230                 struct fuse_conn *fc = fud->fc;
2231                 struct fuse_pqueue *fpq = &fud->pq;
2232                 LIST_HEAD(to_end);
2233                 unsigned int i;
2234
2235                 spin_lock(&fpq->lock);
2236                 WARN_ON(!list_empty(&fpq->io));
2237                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2238                         list_splice_init(&fpq->processing[i], &to_end);
2239                 spin_unlock(&fpq->lock);
2240
2241                 end_requests(fc, &to_end);
2242
2243                 /* Are we the last open device? */
2244                 if (atomic_dec_and_test(&fc->dev_count)) {
2245                         WARN_ON(fc->iq.fasync != NULL);
2246                         fuse_abort_conn(fc, false);
2247                 }
2248                 fuse_dev_free(fud);
2249         }
2250         return 0;
2251 }
2252 EXPORT_SYMBOL_GPL(fuse_dev_release);
2253
2254 static int fuse_dev_fasync(int fd, struct file *file, int on)
2255 {
2256         struct fuse_dev *fud = fuse_get_dev(file);
2257
2258         if (!fud)
2259                 return -EPERM;
2260
2261         /* No locking - fasync_helper does its own locking */
2262         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2263 }
2264
2265 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2266 {
2267         struct fuse_dev *fud;
2268
2269         if (new->private_data)
2270                 return -EINVAL;
2271
2272         fud = fuse_dev_alloc(fc);
2273         if (!fud)
2274                 return -ENOMEM;
2275
2276         new->private_data = fud;
2277         atomic_inc(&fc->dev_count);
2278
2279         return 0;
2280 }
2281
2282 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2283                            unsigned long arg)
2284 {
2285         int err = -ENOTTY;
2286
2287         if (cmd == FUSE_DEV_IOC_CLONE) {
2288                 int oldfd;
2289
2290                 err = -EFAULT;
2291                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2292                         struct file *old = fget(oldfd);
2293
2294                         err = -EINVAL;
2295                         if (old) {
2296                                 struct fuse_dev *fud = NULL;
2297
2298                                 /*
2299                                  * Check against file->f_op because CUSE
2300                                  * uses the same ioctl handler.
2301                                  */
2302                                 if (old->f_op == file->f_op &&
2303                                     old->f_cred->user_ns == file->f_cred->user_ns)
2304                                         fud = fuse_get_dev(old);
2305
2306                                 if (fud) {
2307                                         mutex_lock(&fuse_mutex);
2308                                         err = fuse_device_clone(fud->fc, file);
2309                                         mutex_unlock(&fuse_mutex);
2310                                 }
2311                                 fput(old);
2312                         }
2313                 }
2314         }
2315         return err;
2316 }
2317
2318 const struct file_operations fuse_dev_operations = {
2319         .owner          = THIS_MODULE,
2320         .open           = fuse_dev_open,
2321         .llseek         = no_llseek,
2322         .read_iter      = fuse_dev_read,
2323         .splice_read    = fuse_dev_splice_read,
2324         .write_iter     = fuse_dev_write,
2325         .splice_write   = fuse_dev_splice_write,
2326         .poll           = fuse_dev_poll,
2327         .release        = fuse_dev_release,
2328         .fasync         = fuse_dev_fasync,
2329         .unlocked_ioctl = fuse_dev_ioctl,
2330         .compat_ioctl   = fuse_dev_ioctl,
2331 };
2332 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2333
2334 static struct miscdevice fuse_miscdevice = {
2335         .minor = FUSE_MINOR,
2336         .name  = "fuse",
2337         .fops = &fuse_dev_operations,
2338 };
2339
2340 int __init fuse_dev_init(void)
2341 {
2342         int err = -ENOMEM;
2343         fuse_req_cachep = kmem_cache_create("fuse_request",
2344                                             sizeof(struct fuse_req),
2345                                             0, 0, NULL);
2346         if (!fuse_req_cachep)
2347                 goto out;
2348
2349         err = misc_register(&fuse_miscdevice);
2350         if (err)
2351                 goto out_cache_clean;
2352
2353         return 0;
2354
2355  out_cache_clean:
2356         kmem_cache_destroy(fuse_req_cachep);
2357  out:
2358         return err;
2359 }
2360
2361 void fuse_dev_cleanup(void)
2362 {
2363         misc_deregister(&fuse_miscdevice);
2364         kmem_cache_destroy(fuse_req_cachep);
2365 }