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         /*
 169          * lockess check of fc->connected is okay, because atomic_dec_and_test()
 170          * provides a memory barrier mached with the one in fuse_wait_aborted()
 171          * to ensure no wake-up is missed.
 172          */
 173         if (atomic_dec_and_test(&fc->num_waiting) &&
 174             !READ_ONCE(fc->connected)) {
 175                 /* wake up aborters */
 176                 wake_up_all(&fc->blocked_waitq);
 177         }
 178 }
 179
 180 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 181                                        bool for_background)
 182 {
 183         struct fuse_req *req;
 184         int err;
 185         atomic_inc(&fc->num_waiting);
 186
 187         if (fuse_block_alloc(fc, for_background)) {
 188                 err = -EINTR;
 189                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 190                                 !fuse_block_alloc(fc, for_background)))
 191                         goto out;
 192         }
 193         /* Matches smp_wmb() in fuse_set_initialized() */
 194         smp_rmb();
 195
 196         err = -ENOTCONN;
 197         if (!fc->connected)
 198                 goto out;
 199
 200         err = -ECONNREFUSED;
 201         if (fc->conn_error)
 202                 goto out;
 203
 204         req = fuse_request_alloc(npages);
 205         err = -ENOMEM;
 206         if (!req) {
 207                 if (for_background)
 208                         wake_up(&fc->blocked_waitq);
 209                 goto out;
 210         }
 211
 212         req->in.h.uid = from_kuid(fc->user_ns, current_fsuid());
 213         req->in.h.gid = from_kgid(fc->user_ns, current_fsgid());
 214         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 215
 216         __set_bit(FR_WAITING, &req->flags);
 217         if (for_background)
 218                 __set_bit(FR_BACKGROUND, &req->flags);
 219
 220         if (unlikely(req->in.h.uid == ((uid_t)-1) ||
 221                      req->in.h.gid == ((gid_t)-1))) {
 222                 fuse_put_request(fc, req);
 223                 return ERR_PTR(-EOVERFLOW);
 224         }
 225         return req;
 226
 227  out:
 228         fuse_drop_waiting(fc);
 229         return ERR_PTR(err);
 230 }
 231
 232 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 233 {
 234         return __fuse_get_req(fc, npages, false);
 235 }
 236 EXPORT_SYMBOL_GPL(fuse_get_req);
 237
 238 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 239                                              unsigned npages)
 240 {
 241         return __fuse_get_req(fc, npages, true);
 242 }
 243 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 244
 245 /*
 246  * Return request in fuse_file->reserved_req.  However that may
 247  * currently be in use.  If that is the case, wait for it to become
 248  * available.
 249  */
 250 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 251                                          struct file *file)
 252 {
 253         struct fuse_req *req = NULL;
 254         struct fuse_file *ff = file->private_data;
 255
 256         do {
 257                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 258                 spin_lock(&fc->lock);
 259                 if (ff->reserved_req) {
 260                         req = ff->reserved_req;
 261                         ff->reserved_req = NULL;
 262                         req->stolen_file = get_file(file);
 263                 }
 264                 spin_unlock(&fc->lock);
 265         } while (!req);
 266
 267         return req;
 268 }
 269
 270 /*
 271  * Put stolen request back into fuse_file->reserved_req
 272  */
 273 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 274 {
 275         struct file *file = req->stolen_file;
 276         struct fuse_file *ff = file->private_data;
 277
 278         WARN_ON(req->max_pages);
 279         spin_lock(&fc->lock);
 280         memset(req, 0, sizeof(*req));
 281         fuse_request_init(req, NULL, NULL, 0);
 282         BUG_ON(ff->reserved_req);
 283         ff->reserved_req = req;
 284         wake_up_all(&fc->reserved_req_waitq);
 285         spin_unlock(&fc->lock);
 286         fput(file);
 287 }
 288
 289 /*
 290  * Gets a requests for a file operation, always succeeds
 291  *
 292  * This is used for sending the FLUSH request, which must get to
 293  * userspace, due to POSIX locks which may need to be unlocked.
 294  *
 295  * If allocation fails due to OOM, use the reserved request in
 296  * fuse_file.
 297  *
 298  * This is very unlikely to deadlock accidentally, since the
 299  * filesystem should not have it's own file open.  If deadlock is
 300  * intentional, it can still be broken by "aborting" the filesystem.
 301  */
 302 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 303                                              struct file *file)
 304 {
 305         struct fuse_req *req;
 306
 307         atomic_inc(&fc->num_waiting);
 308         wait_event(fc->blocked_waitq, fc->initialized);
 309         /* Matches smp_wmb() in fuse_set_initialized() */
 310         smp_rmb();
 311         req = fuse_request_alloc(0);
 312         if (!req)
 313                 req = get_reserved_req(fc, file);
 314
 315         req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
 316         req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
 317         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 318
 319         __set_bit(FR_WAITING, &req->flags);
 320         __clear_bit(FR_BACKGROUND, &req->flags);
 321         return req;
 322 }
 323
 324 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 325 {
 326         if (refcount_dec_and_test(&req->count)) {
 327                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 328                         /*
 329                          * We get here in the unlikely case that a background
 330                          * request was allocated but not sent
 331                          */
 332                         spin_lock(&fc->bg_lock);
 333                         if (!fc->blocked)
 334                                 wake_up(&fc->blocked_waitq);
 335                         spin_unlock(&fc->bg_lock);
 336                 }
 337
 338                 if (test_bit(FR_WAITING, &req->flags)) {
 339                         __clear_bit(FR_WAITING, &req->flags);
 340                         fuse_drop_waiting(fc);
 341                 }
 342
 343                 if (req->stolen_file)
 344                         put_reserved_req(fc, req);
 345                 else
 346                         fuse_request_free(req);
 347         }
 348 }
 349 EXPORT_SYMBOL_GPL(fuse_put_request);
 350
 351 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 352 {
 353         unsigned nbytes = 0;
 354         unsigned i;
 355
 356         for (i = 0; i < numargs; i++)
 357                 nbytes += args[i].size;
 358
 359         return nbytes;
 360 }
 361
 362 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 363 {
 364         fiq->reqctr += FUSE_REQ_ID_STEP;
 365         return fiq->reqctr;
 366 }
 367
 368 static unsigned int fuse_req_hash(u64 unique)
 369 {
 370         return hash_long(unique & ~FUSE_INT_REQ_BIT, FUSE_PQ_HASH_BITS);
 371 }
 372
 373 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 374 {
 375         req->in.h.len = sizeof(struct fuse_in_header) +
 376                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 377         list_add_tail(&req->list, &fiq->pending);
 378         wake_up_locked(&fiq->waitq);
 379         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 380 }
 381
 382 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 383                        u64 nodeid, u64 nlookup)
 384 {
 385         struct fuse_iqueue *fiq = &fc->iq;
 386
 387         forget->forget_one.nodeid = nodeid;
 388         forget->forget_one.nlookup = nlookup;
 389
 390         spin_lock(&fiq->waitq.lock);
 391         if (fiq->connected) {
 392                 fiq->forget_list_tail->next = forget;
 393                 fiq->forget_list_tail = forget;
 394                 wake_up_locked(&fiq->waitq);
 395                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 396         } else {
 397                 kfree(forget);
 398         }
 399         spin_unlock(&fiq->waitq.lock);
 400 }
 401
 402 static void flush_bg_queue(struct fuse_conn *fc)
 403 {
 404         struct fuse_iqueue *fiq = &fc->iq;
 405
 406         while (fc->active_background < fc->max_background &&
 407                !list_empty(&fc->bg_queue)) {
 408                 struct fuse_req *req;
 409
 410                 req = list_first_entry(&fc->bg_queue, struct fuse_req, list);
 411                 list_del(&req->list);
 412                 fc->active_background++;
 413                 spin_lock(&fiq->waitq.lock);
 414                 req->in.h.unique = fuse_get_unique(fiq);
 415                 queue_request(fiq, req);
 416                 spin_unlock(&fiq->waitq.lock);
 417         }
 418 }
 419
 420 /*
 421  * This function is called when a request is finished.  Either a reply
 422  * has arrived or it was aborted (and not yet sent) or some error
 423  * occurred during communication with userspace, or the device file
 424  * was closed.  The requester thread is woken up (if still waiting),
 425  * the 'end' callback is called if given, else the reference to the
 426  * request is released
 427  */
 428 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 429 {
 430         struct fuse_iqueue *fiq = &fc->iq;
 431
 432         if (test_and_set_bit(FR_FINISHED, &req->flags))
 433                 goto put_request;
 434
 435         spin_lock(&fiq->waitq.lock);
 436         list_del_init(&req->intr_entry);
 437         spin_unlock(&fiq->waitq.lock);
 438         WARN_ON(test_bit(FR_PENDING, &req->flags));
 439         WARN_ON(test_bit(FR_SENT, &req->flags));
 440         if (test_bit(FR_BACKGROUND, &req->flags)) {
 441                 spin_lock(&fc->bg_lock);
 442                 clear_bit(FR_BACKGROUND, &req->flags);
 443                 if (fc->num_background == fc->max_background) {
 444                         fc->blocked = 0;
 445                         wake_up(&fc->blocked_waitq);
 446                 } else if (!fc->blocked) {
 447                         /*
 448                          * Wake up next waiter, if any.  It's okay to use
 449                          * waitqueue_active(), as we've already synced up
 450                          * fc->blocked with waiters with the wake_up() call
 451                          * above.
 452                          */
 453                         if (waitqueue_active(&fc->blocked_waitq))
 454                                 wake_up(&fc->blocked_waitq);
 455                 }
 456
 457                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 458                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 459                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 460                 }
 461                 fc->num_background--;
 462                 fc->active_background--;
 463                 flush_bg_queue(fc);
 464                 spin_unlock(&fc->bg_lock);
 465         }
 466         wake_up(&req->waitq);
 467         if (req->end)
 468                 req->end(fc, req);
 469 put_request:
 470         fuse_put_request(fc, req);
 471 }
 472
 473 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 474 {
 475         spin_lock(&fiq->waitq.lock);
 476         if (test_bit(FR_FINISHED, &req->flags)) {
 477                 spin_unlock(&fiq->waitq.lock);
 478                 return;
 479         }
 480         if (list_empty(&req->intr_entry)) {
 481                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 482                 wake_up_locked(&fiq->waitq);
 483         }
 484         spin_unlock(&fiq->waitq.lock);
 485         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 486 }
 487
 488 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 489 {
 490         struct fuse_iqueue *fiq = &fc->iq;
 491         int err;
 492
 493         if (!fc->no_interrupt) {
 494                 /* Any signal may interrupt this */
 495                 err = wait_event_interruptible(req->waitq,
 496                                         test_bit(FR_FINISHED, &req->flags));
 497                 if (!err)
 498                         return;
 499
 500                 set_bit(FR_INTERRUPTED, &req->flags);
 501                 /* matches barrier in fuse_dev_do_read() */
 502                 smp_mb__after_atomic();
 503                 if (test_bit(FR_SENT, &req->flags))
 504                         queue_interrupt(fiq, req);
 505         }
 506
 507         if (!test_bit(FR_FORCE, &req->flags)) {
 508                 /* Only fatal signals may interrupt this */
 509                 err = wait_event_killable(req->waitq,
 510                                         test_bit(FR_FINISHED, &req->flags));
 511                 if (!err)
 512                         return;
 513
 514                 spin_lock(&fiq->waitq.lock);
 515                 /* Request is not yet in userspace, bail out */
 516                 if (test_bit(FR_PENDING, &req->flags)) {
 517                         list_del(&req->list);
 518                         spin_unlock(&fiq->waitq.lock);
 519                         __fuse_put_request(req);
 520                         req->out.h.error = -EINTR;
 521                         return;
 522                 }
 523                 spin_unlock(&fiq->waitq.lock);
 524         }
 525
 526         /*
 527          * Either request is already in userspace, or it was forced.
 528          * Wait it out.
 529          */
 530         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 531 }
 532
 533 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 534 {
 535         struct fuse_iqueue *fiq = &fc->iq;
 536
 537         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 538         spin_lock(&fiq->waitq.lock);
 539         if (!fiq->connected) {
 540                 spin_unlock(&fiq->waitq.lock);
 541                 req->out.h.error = -ENOTCONN;
 542         } else {
 543                 req->in.h.unique = fuse_get_unique(fiq);
 544                 queue_request(fiq, req);
 545                 /* acquire extra reference, since request is still needed
 546                    after request_end() */
 547                 __fuse_get_request(req);
 548                 spin_unlock(&fiq->waitq.lock);
 549
 550                 request_wait_answer(fc, req);
 551                 /* Pairs with smp_wmb() in request_end() */
 552                 smp_rmb();
 553         }
 554 }
 555
 556 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 557 {
 558         __set_bit(FR_ISREPLY, &req->flags);
 559         if (!test_bit(FR_WAITING, &req->flags)) {
 560                 __set_bit(FR_WAITING, &req->flags);
 561                 atomic_inc(&fc->num_waiting);
 562         }
 563         __fuse_request_send(fc, req);
 564 }
 565 EXPORT_SYMBOL_GPL(fuse_request_send);
 566
 567 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 568 {
 569         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 570                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 571
 572         if (fc->minor < 9) {
 573                 switch (args->in.h.opcode) {
 574                 case FUSE_LOOKUP:
 575                 case FUSE_CREATE:
 576                 case FUSE_MKNOD:
 577                 case FUSE_MKDIR:
 578                 case FUSE_SYMLINK:
 579                 case FUSE_LINK:
 580                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 581                         break;
 582                 case FUSE_GETATTR:
 583                 case FUSE_SETATTR:
 584                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 585                         break;
 586                 }
 587         }
 588         if (fc->minor < 12) {
 589                 switch (args->in.h.opcode) {
 590                 case FUSE_CREATE:
 591                         args->in.args[0].size = sizeof(struct fuse_open_in);
 592                         break;
 593                 case FUSE_MKNOD:
 594                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 595                         break;
 596                 }
 597         }
 598 }
 599
 600 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 601 {
 602         struct fuse_req *req;
 603         ssize_t ret;
 604
 605         req = fuse_get_req(fc, 0);
 606         if (IS_ERR(req))
 607                 return PTR_ERR(req);
 608
 609         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 610         fuse_adjust_compat(fc, args);
 611
 612         req->in.h.opcode = args->in.h.opcode;
 613         req->in.h.nodeid = args->in.h.nodeid;
 614         req->in.numargs = args->in.numargs;
 615         memcpy(req->in.args, args->in.args,
 616                args->in.numargs * sizeof(struct fuse_in_arg));
 617         req->out.argvar = args->out.argvar;
 618         req->out.numargs = args->out.numargs;
 619         memcpy(req->out.args, args->out.args,
 620                args->out.numargs * sizeof(struct fuse_arg));
 621         fuse_request_send(fc, req);
 622         ret = req->out.h.error;
 623         if (!ret && args->out.argvar) {
 624                 BUG_ON(args->out.numargs != 1);
 625                 ret = req->out.args[0].size;
 626         }
 627         fuse_put_request(fc, req);
 628
 629         return ret;
 630 }
 631
 632 bool fuse_request_queue_background(struct fuse_conn *fc, struct fuse_req *req)
 633 {
 634         bool queued = false;
 635
 636         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 637         if (!test_bit(FR_WAITING, &req->flags)) {
 638                 __set_bit(FR_WAITING, &req->flags);
 639                 atomic_inc(&fc->num_waiting);
 640         }
 641         __set_bit(FR_ISREPLY, &req->flags);
 642         spin_lock(&fc->bg_lock);
 643         if (likely(fc->connected)) {
 644                 fc->num_background++;
 645                 if (fc->num_background == fc->max_background)
 646                         fc->blocked = 1;
 647                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 648                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 649                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 650                 }
 651                 list_add_tail(&req->list, &fc->bg_queue);
 652                 flush_bg_queue(fc);
 653                 queued = true;
 654         }
 655         spin_unlock(&fc->bg_lock);
 656
 657         return queued;
 658 }
 659
 660 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 661 {
 662         WARN_ON(!req->end);
 663         if (!fuse_request_queue_background(fc, req)) {
 664                 req->out.h.error = -ENOTCONN;
 665                 req->end(fc, req);
 666                 fuse_put_request(fc, req);
 667         }
 668 }
 669 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 670
 671 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 672                                           struct fuse_req *req, u64 unique)
 673 {
 674         int err = -ENODEV;
 675         struct fuse_iqueue *fiq = &fc->iq;
 676
 677         __clear_bit(FR_ISREPLY, &req->flags);
 678         req->in.h.unique = unique;
 679         spin_lock(&fiq->waitq.lock);
 680         if (fiq->connected) {
 681                 queue_request(fiq, req);
 682                 err = 0;
 683         }
 684         spin_unlock(&fiq->waitq.lock);
 685
 686         return err;
 687 }
 688
 689 void fuse_force_forget(struct file *file, u64 nodeid)
 690 {
 691         struct inode *inode = file_inode(file);
 692         struct fuse_conn *fc = get_fuse_conn(inode);
 693         struct fuse_req *req;
 694         struct fuse_forget_in inarg;
 695
 696         memset(&inarg, 0, sizeof(inarg));
 697         inarg.nlookup = 1;
 698         req = fuse_get_req_nofail_nopages(fc, file);
 699         req->in.h.opcode = FUSE_FORGET;
 700         req->in.h.nodeid = nodeid;
 701         req->in.numargs = 1;
 702         req->in.args[0].size = sizeof(inarg);
 703         req->in.args[0].value = &inarg;
 704         __clear_bit(FR_ISREPLY, &req->flags);
 705         __fuse_request_send(fc, req);
 706         /* ignore errors */
 707         fuse_put_request(fc, req);
 708 }
 709
 710 /*
 711  * Lock the request.  Up to the next unlock_request() there mustn't be
 712  * anything that could cause a page-fault.  If the request was already
 713  * aborted bail out.
 714  */
 715 static int lock_request(struct fuse_req *req)
 716 {
 717         int err = 0;
 718         if (req) {
 719                 spin_lock(&req->waitq.lock);
 720                 if (test_bit(FR_ABORTED, &req->flags))
 721                         err = -ENOENT;
 722                 else
 723                         set_bit(FR_LOCKED, &req->flags);
 724                 spin_unlock(&req->waitq.lock);
 725         }
 726         return err;
 727 }
 728
 729 /*
 730  * Unlock request.  If it was aborted while locked, caller is responsible
 731  * for unlocking and ending the request.
 732  */
 733 static int unlock_request(struct fuse_req *req)
 734 {
 735         int err = 0;
 736         if (req) {
 737                 spin_lock(&req->waitq.lock);
 738                 if (test_bit(FR_ABORTED, &req->flags))
 739                         err = -ENOENT;
 740                 else
 741                         clear_bit(FR_LOCKED, &req->flags);
 742                 spin_unlock(&req->waitq.lock);
 743         }
 744         return err;
 745 }
 746
 747 struct fuse_copy_state {
 748         int write;
 749         struct fuse_req *req;
 750         struct iov_iter *iter;
 751         struct pipe_buffer *pipebufs;
 752         struct pipe_buffer *currbuf;
 753         struct pipe_inode_info *pipe;
 754         unsigned long nr_segs;
 755         struct page *pg;
 756         unsigned len;
 757         unsigned offset;
 758         unsigned move_pages:1;
 759 };
 760
 761 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 762                            struct iov_iter *iter)
 763 {
 764         memset(cs, 0, sizeof(*cs));
 765         cs->write = write;
 766         cs->iter = iter;
 767 }
 768
 769 /* Unmap and put previous page of userspace buffer */
 770 static void fuse_copy_finish(struct fuse_copy_state *cs)
 771 {
 772         if (cs->currbuf) {
 773                 struct pipe_buffer *buf = cs->currbuf;
 774
 775                 if (cs->write)
 776                         buf->len = PAGE_SIZE - cs->len;
 777                 cs->currbuf = NULL;
 778         } else if (cs->pg) {
 779                 if (cs->write) {
 780                         flush_dcache_page(cs->pg);
 781                         set_page_dirty_lock(cs->pg);
 782                 }
 783                 put_page(cs->pg);
 784         }
 785         cs->pg = NULL;
 786 }
 787
 788 /*
 789  * Get another pagefull of userspace buffer, and map it to kernel
 790  * address space, and lock request
 791  */
 792 static int fuse_copy_fill(struct fuse_copy_state *cs)
 793 {
 794         struct page *page;
 795         int err;
 796
 797         err = unlock_request(cs->req);
 798         if (err)
 799                 return err;
 800
 801         fuse_copy_finish(cs);
 802         if (cs->pipebufs) {
 803                 struct pipe_buffer *buf = cs->pipebufs;
 804
 805                 if (!cs->write) {
 806                         err = pipe_buf_confirm(cs->pipe, buf);
 807                         if (err)
 808                                 return err;
 809
 810                         BUG_ON(!cs->nr_segs);
 811                         cs->currbuf = buf;
 812                         cs->pg = buf->page;
 813                         cs->offset = buf->offset;
 814                         cs->len = buf->len;
 815                         cs->pipebufs++;
 816                         cs->nr_segs--;
 817                 } else {
 818                         if (cs->nr_segs == cs->pipe->buffers)
 819                                 return -EIO;
 820
 821                         page = alloc_page(GFP_HIGHUSER);
 822                         if (!page)
 823                                 return -ENOMEM;
 824
 825                         buf->page = page;
 826                         buf->offset = 0;
 827                         buf->len = 0;
 828
 829                         cs->currbuf = buf;
 830                         cs->pg = page;
 831                         cs->offset = 0;
 832                         cs->len = PAGE_SIZE;
 833                         cs->pipebufs++;
 834                         cs->nr_segs++;
 835                 }
 836         } else {
 837                 size_t off;
 838                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 839                 if (err < 0)
 840                         return err;
 841                 BUG_ON(!err);
 842                 cs->len = err;
 843                 cs->offset = off;
 844                 cs->pg = page;
 845                 iov_iter_advance(cs->iter, err);
 846         }
 847
 848         return lock_request(cs->req);
 849 }
 850
 851 /* Do as much copy to/from userspace buffer as we can */
 852 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 853 {
 854         unsigned ncpy = min(*size, cs->len);
 855         if (val) {
 856                 void *pgaddr = kmap_atomic(cs->pg);
 857                 void *buf = pgaddr + cs->offset;
 858
 859                 if (cs->write)
 860                         memcpy(buf, *val, ncpy);
 861                 else
 862                         memcpy(*val, buf, ncpy);
 863
 864                 kunmap_atomic(pgaddr);
 865                 *val += ncpy;
 866         }
 867         *size -= ncpy;
 868         cs->len -= ncpy;
 869         cs->offset += ncpy;
 870         return ncpy;
 871 }
 872
 873 static int fuse_check_page(struct page *page)
 874 {
 875         if (page_mapcount(page) ||
 876             page->mapping != NULL ||
 877             page_count(page) != 1 ||
 878             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 879              ~(1 << PG_locked |
 880                1 << PG_referenced |
 881                1 << PG_uptodate |
 882                1 << PG_lru |
 883                1 << PG_active |
 884                1 << PG_reclaim))) {
 885                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 886                 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);
 887                 return 1;
 888         }
 889         return 0;
 890 }
 891
 892 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 893 {
 894         int err;
 895         struct page *oldpage = *pagep;
 896         struct page *newpage;
 897         struct pipe_buffer *buf = cs->pipebufs;
 898
 899         err = unlock_request(cs->req);
 900         if (err)
 901                 return err;
 902
 903         fuse_copy_finish(cs);
 904
 905         err = pipe_buf_confirm(cs->pipe, buf);
 906         if (err)
 907                 return err;
 908
 909         BUG_ON(!cs->nr_segs);
 910         cs->currbuf = buf;
 911         cs->len = buf->len;
 912         cs->pipebufs++;
 913         cs->nr_segs--;
 914
 915         if (cs->len != PAGE_SIZE)
 916                 goto out_fallback;
 917
 918         if (pipe_buf_steal(cs->pipe, buf) != 0)
 919                 goto out_fallback;
 920
 921         newpage = buf->page;
 922
 923         if (!PageUptodate(newpage))
 924                 SetPageUptodate(newpage);
 925
 926         ClearPageMappedToDisk(newpage);
 927
 928         if (fuse_check_page(newpage) != 0)
 929                 goto out_fallback_unlock;
 930
 931         /*
 932          * This is a new and locked page, it shouldn't be mapped or
 933          * have any special flags on it
 934          */
 935         if (WARN_ON(page_mapped(oldpage)))
 936                 goto out_fallback_unlock;
 937         if (WARN_ON(page_has_private(oldpage)))
 938                 goto out_fallback_unlock;
 939         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 940                 goto out_fallback_unlock;
 941         if (WARN_ON(PageMlocked(oldpage)))
 942                 goto out_fallback_unlock;
 943
 944         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 945         if (err) {
 946                 unlock_page(newpage);
 947                 return err;
 948         }
 949
 950         get_page(newpage);
 951
 952         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 953                 lru_cache_add_file(newpage);
 954
 955         err = 0;
 956         spin_lock(&cs->req->waitq.lock);
 957         if (test_bit(FR_ABORTED, &cs->req->flags))
 958                 err = -ENOENT;
 959         else
 960                 *pagep = newpage;
 961         spin_unlock(&cs->req->waitq.lock);
 962
 963         if (err) {
 964                 unlock_page(newpage);
 965                 put_page(newpage);
 966                 return err;
 967         }
 968
 969         unlock_page(oldpage);
 970         put_page(oldpage);
 971         cs->len = 0;
 972
 973         return 0;
 974
 975 out_fallback_unlock:
 976         unlock_page(newpage);
 977 out_fallback:
 978         cs->pg = buf->page;
 979         cs->offset = buf->offset;
 980
 981         err = lock_request(cs->req);
 982         if (err)
 983                 return err;
 984
 985         return 1;
 986 }
 987
 988 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 989                          unsigned offset, unsigned count)
 990 {
 991         struct pipe_buffer *buf;
 992         int err;
 993
 994         if (cs->nr_segs == cs->pipe->buffers)
 995                 return -EIO;
 996
 997         err = unlock_request(cs->req);
 998         if (err)
 999                 return err;
1000
1001         fuse_copy_finish(cs);
1002
1003         buf = cs->pipebufs;
1004         get_page(page);
1005         buf->page = page;
1006         buf->offset = offset;
1007         buf->len = count;
1008
1009         cs->pipebufs++;
1010         cs->nr_segs++;
1011         cs->len = 0;
1012
1013         return 0;
1014 }
1015
1016 /*
1017  * Copy a page in the request to/from the userspace buffer.  Must be
1018  * done atomically
1019  */
1020 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
1021                           unsigned offset, unsigned count, int zeroing)
1022 {
1023         int err;
1024         struct page *page = *pagep;
1025
1026         if (page && zeroing && count < PAGE_SIZE)
1027                 clear_highpage(page);
1028
1029         while (count) {
1030                 if (cs->write && cs->pipebufs && page) {
1031                         return fuse_ref_page(cs, page, offset, count);
1032                 } else if (!cs->len) {
1033                         if (cs->move_pages && page &&
1034                             offset == 0 && count == PAGE_SIZE) {
1035                                 err = fuse_try_move_page(cs, pagep);
1036                                 if (err <= 0)
1037                                         return err;
1038                         } else {
1039                                 err = fuse_copy_fill(cs);
1040                                 if (err)
1041                                         return err;
1042                         }
1043                 }
1044                 if (page) {
1045                         void *mapaddr = kmap_atomic(page);
1046                         void *buf = mapaddr + offset;
1047                         offset += fuse_copy_do(cs, &buf, &count);
1048                         kunmap_atomic(mapaddr);
1049                 } else
1050                         offset += fuse_copy_do(cs, NULL, &count);
1051         }
1052         if (page && !cs->write)
1053                 flush_dcache_page(page);
1054         return 0;
1055 }
1056
1057 /* Copy pages in the request to/from userspace buffer */
1058 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1059                            int zeroing)
1060 {
1061         unsigned i;
1062         struct fuse_req *req = cs->req;
1063
1064         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1065                 int err;
1066                 unsigned offset = req->page_descs[i].offset;
1067                 unsigned count = min(nbytes, req->page_descs[i].length);
1068
1069                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1070                                      zeroing);
1071                 if (err)
1072                         return err;
1073
1074                 nbytes -= count;
1075         }
1076         return 0;
1077 }
1078
1079 /* Copy a single argument in the request to/from userspace buffer */
1080 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1081 {
1082         while (size) {
1083                 if (!cs->len) {
1084                         int err = fuse_copy_fill(cs);
1085                         if (err)
1086                                 return err;
1087                 }
1088                 fuse_copy_do(cs, &val, &size);
1089         }
1090         return 0;
1091 }
1092
1093 /* Copy request arguments to/from userspace buffer */
1094 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1095                           unsigned argpages, struct fuse_arg *args,
1096                           int zeroing)
1097 {
1098         int err = 0;
1099         unsigned i;
1100
1101         for (i = 0; !err && i < numargs; i++)  {
1102                 struct fuse_arg *arg = &args[i];
1103                 if (i == numargs - 1 && argpages)
1104                         err = fuse_copy_pages(cs, arg->size, zeroing);
1105                 else
1106                         err = fuse_copy_one(cs, arg->value, arg->size);
1107         }
1108         return err;
1109 }
1110
1111 static int forget_pending(struct fuse_iqueue *fiq)
1112 {
1113         return fiq->forget_list_head.next != NULL;
1114 }
1115
1116 static int request_pending(struct fuse_iqueue *fiq)
1117 {
1118         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1119                 forget_pending(fiq);
1120 }
1121
1122 /*
1123  * Transfer an interrupt request to userspace
1124  *
1125  * Unlike other requests this is assembled on demand, without a need
1126  * to allocate a separate fuse_req structure.
1127  *
1128  * Called with fiq->waitq.lock held, releases it
1129  */
1130 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1131                                struct fuse_copy_state *cs,
1132                                size_t nbytes, struct fuse_req *req)
1133 __releases(fiq->waitq.lock)
1134 {
1135         struct fuse_in_header ih;
1136         struct fuse_interrupt_in arg;
1137         unsigned reqsize = sizeof(ih) + sizeof(arg);
1138         int err;
1139
1140         list_del_init(&req->intr_entry);
1141         memset(&ih, 0, sizeof(ih));
1142         memset(&arg, 0, sizeof(arg));
1143         ih.len = reqsize;
1144         ih.opcode = FUSE_INTERRUPT;
1145         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1146         arg.unique = req->in.h.unique;
1147
1148         spin_unlock(&fiq->waitq.lock);
1149         if (nbytes < reqsize)
1150                 return -EINVAL;
1151
1152         err = fuse_copy_one(cs, &ih, sizeof(ih));
1153         if (!err)
1154                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1155         fuse_copy_finish(cs);
1156
1157         return err ? err : reqsize;
1158 }
1159
1160 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1161                                                unsigned max,
1162                                                unsigned *countp)
1163 {
1164         struct fuse_forget_link *head = fiq->forget_list_head.next;
1165         struct fuse_forget_link **newhead = &head;
1166         unsigned count;
1167
1168         for (count = 0; *newhead != NULL && count < max; count++)
1169                 newhead = &(*newhead)->next;
1170
1171         fiq->forget_list_head.next = *newhead;
1172         *newhead = NULL;
1173         if (fiq->forget_list_head.next == NULL)
1174                 fiq->forget_list_tail = &fiq->forget_list_head;
1175
1176         if (countp != NULL)
1177                 *countp = count;
1178
1179         return head;
1180 }
1181
1182 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1183                                    struct fuse_copy_state *cs,
1184                                    size_t nbytes)
1185 __releases(fiq->waitq.lock)
1186 {
1187         int err;
1188         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1189         struct fuse_forget_in arg = {
1190                 .nlookup = forget->forget_one.nlookup,
1191         };
1192         struct fuse_in_header ih = {
1193                 .opcode = FUSE_FORGET,
1194                 .nodeid = forget->forget_one.nodeid,
1195                 .unique = fuse_get_unique(fiq),
1196                 .len = sizeof(ih) + sizeof(arg),
1197         };
1198
1199         spin_unlock(&fiq->waitq.lock);
1200         kfree(forget);
1201         if (nbytes < ih.len)
1202                 return -EINVAL;
1203
1204         err = fuse_copy_one(cs, &ih, sizeof(ih));
1205         if (!err)
1206                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1207         fuse_copy_finish(cs);
1208
1209         if (err)
1210                 return err;
1211
1212         return ih.len;
1213 }
1214
1215 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1216                                    struct fuse_copy_state *cs, size_t nbytes)
1217 __releases(fiq->waitq.lock)
1218 {
1219         int err;
1220         unsigned max_forgets;
1221         unsigned count;
1222         struct fuse_forget_link *head;
1223         struct fuse_batch_forget_in arg = { .count = 0 };
1224         struct fuse_in_header ih = {
1225                 .opcode = FUSE_BATCH_FORGET,
1226                 .unique = fuse_get_unique(fiq),
1227                 .len = sizeof(ih) + sizeof(arg),
1228         };
1229
1230         if (nbytes < ih.len) {
1231                 spin_unlock(&fiq->waitq.lock);
1232                 return -EINVAL;
1233         }
1234
1235         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1236         head = dequeue_forget(fiq, max_forgets, &count);
1237         spin_unlock(&fiq->waitq.lock);
1238
1239         arg.count = count;
1240         ih.len += count * sizeof(struct fuse_forget_one);
1241         err = fuse_copy_one(cs, &ih, sizeof(ih));
1242         if (!err)
1243                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1244
1245         while (head) {
1246                 struct fuse_forget_link *forget = head;
1247
1248                 if (!err) {
1249                         err = fuse_copy_one(cs, &forget->forget_one,
1250                                             sizeof(forget->forget_one));
1251                 }
1252                 head = forget->next;
1253                 kfree(forget);
1254         }
1255
1256         fuse_copy_finish(cs);
1257
1258         if (err)
1259                 return err;
1260
1261         return ih.len;
1262 }
1263
1264 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1265                             struct fuse_copy_state *cs,
1266                             size_t nbytes)
1267 __releases(fiq->waitq.lock)
1268 {
1269         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1270                 return fuse_read_single_forget(fiq, cs, nbytes);
1271         else
1272                 return fuse_read_batch_forget(fiq, cs, nbytes);
1273 }
1274
1275 /*
1276  * Read a single request into the userspace filesystem's buffer.  This
1277  * function waits until a request is available, then removes it from
1278  * the pending list and copies request data to userspace buffer.  If
1279  * no reply is needed (FORGET) or request has been aborted or there
1280  * was an error during the copying then it's finished by calling
1281  * request_end().  Otherwise add it to the processing list, and set
1282  * the 'sent' flag.
1283  */
1284 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1285                                 struct fuse_copy_state *cs, size_t nbytes)
1286 {
1287         ssize_t err;
1288         struct fuse_conn *fc = fud->fc;
1289         struct fuse_iqueue *fiq = &fc->iq;
1290         struct fuse_pqueue *fpq = &fud->pq;
1291         struct fuse_req *req;
1292         struct fuse_in *in;
1293         unsigned reqsize;
1294         unsigned int hash;
1295
1296  restart:
1297         spin_lock(&fiq->waitq.lock);
1298         err = -EAGAIN;
1299         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1300             !request_pending(fiq))
1301                 goto err_unlock;
1302
1303         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1304                                 !fiq->connected || request_pending(fiq));
1305         if (err)
1306                 goto err_unlock;
1307
1308         if (!fiq->connected) {
1309                 err = (fc->aborted && fc->abort_err) ? -ECONNABORTED : -ENODEV;
1310                 goto err_unlock;
1311         }
1312
1313         if (!list_empty(&fiq->interrupts)) {
1314                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1315                                  intr_entry);
1316                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1317         }
1318
1319         if (forget_pending(fiq)) {
1320                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1321                         return fuse_read_forget(fc, fiq, cs, nbytes);
1322
1323                 if (fiq->forget_batch <= -8)
1324                         fiq->forget_batch = 16;
1325         }
1326
1327         req = list_entry(fiq->pending.next, struct fuse_req, list);
1328         clear_bit(FR_PENDING, &req->flags);
1329         list_del_init(&req->list);
1330         spin_unlock(&fiq->waitq.lock);
1331
1332         in = &req->in;
1333         reqsize = in->h.len;
1334
1335         /* If request is too large, reply with an error and restart the read */
1336         if (nbytes < reqsize) {
1337                 req->out.h.error = -EIO;
1338                 /* SETXATTR is special, since it may contain too large data */
1339                 if (in->h.opcode == FUSE_SETXATTR)
1340                         req->out.h.error = -E2BIG;
1341                 request_end(fc, req);
1342                 goto restart;
1343         }
1344         spin_lock(&fpq->lock);
1345         list_add(&req->list, &fpq->io);
1346         spin_unlock(&fpq->lock);
1347         cs->req = req;
1348         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1349         if (!err)
1350                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1351                                      (struct fuse_arg *) in->args, 0);
1352         fuse_copy_finish(cs);
1353         spin_lock(&fpq->lock);
1354         clear_bit(FR_LOCKED, &req->flags);
1355         if (!fpq->connected) {
1356                 err = (fc->aborted && fc->abort_err) ? -ECONNABORTED : -ENODEV;
1357                 goto out_end;
1358         }
1359         if (err) {
1360                 req->out.h.error = -EIO;
1361                 goto out_end;
1362         }
1363         if (!test_bit(FR_ISREPLY, &req->flags)) {
1364                 err = reqsize;
1365                 goto out_end;
1366         }
1367         hash = fuse_req_hash(req->in.h.unique);
1368         list_move_tail(&req->list, &fpq->processing[hash]);
1369         __fuse_get_request(req);
1370         set_bit(FR_SENT, &req->flags);
1371         spin_unlock(&fpq->lock);
1372         /* matches barrier in request_wait_answer() */
1373         smp_mb__after_atomic();
1374         if (test_bit(FR_INTERRUPTED, &req->flags))
1375                 queue_interrupt(fiq, req);
1376         fuse_put_request(fc, req);
1377
1378         return reqsize;
1379
1380 out_end:
1381         if (!test_bit(FR_PRIVATE, &req->flags))
1382                 list_del_init(&req->list);
1383         spin_unlock(&fpq->lock);
1384         request_end(fc, req);
1385         return err;
1386
1387  err_unlock:
1388         spin_unlock(&fiq->waitq.lock);
1389         return err;
1390 }
1391
1392 static int fuse_dev_open(struct inode *inode, struct file *file)
1393 {
1394         /*
1395          * The fuse device's file's private_data is used to hold
1396          * the fuse_conn(ection) when it is mounted, and is used to
1397          * keep track of whether the file has been mounted already.
1398          */
1399         file->private_data = NULL;
1400         return 0;
1401 }
1402
1403 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1404 {
1405         struct fuse_copy_state cs;
1406         struct file *file = iocb->ki_filp;
1407         struct fuse_dev *fud = fuse_get_dev(file);
1408
1409         if (!fud)
1410                 return -EPERM;
1411
1412         if (!iter_is_iovec(to))
1413                 return -EINVAL;
1414
1415         fuse_copy_init(&cs, 1, to);
1416
1417         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1418 }
1419
1420 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1421                                     struct pipe_inode_info *pipe,
1422                                     size_t len, unsigned int flags)
1423 {
1424         int total, ret;
1425         int page_nr = 0;
1426         struct pipe_buffer *bufs;
1427         struct fuse_copy_state cs;
1428         struct fuse_dev *fud = fuse_get_dev(in);
1429
1430         if (!fud)
1431                 return -EPERM;
1432
1433         bufs = kvmalloc_array(pipe->buffers, sizeof(struct pipe_buffer),
1434                               GFP_KERNEL);
1435         if (!bufs)
1436                 return -ENOMEM;
1437
1438         fuse_copy_init(&cs, 1, NULL);
1439         cs.pipebufs = bufs;
1440         cs.pipe = pipe;
1441         ret = fuse_dev_do_read(fud, in, &cs, len);
1442         if (ret < 0)
1443                 goto out;
1444
1445         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1446                 ret = -EIO;
1447                 goto out;
1448         }
1449
1450         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1451                 /*
1452                  * Need to be careful about this.  Having buf->ops in module
1453                  * code can Oops if the buffer persists after module unload.
1454                  */
1455                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1456                 bufs[page_nr].flags = 0;
1457                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1458                 if (unlikely(ret < 0))
1459                         break;
1460         }
1461         if (total)
1462                 ret = total;
1463 out:
1464         for (; page_nr < cs.nr_segs; page_nr++)
1465                 put_page(bufs[page_nr].page);
1466
1467         kvfree(bufs);
1468         return ret;
1469 }
1470
1471 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1472                             struct fuse_copy_state *cs)
1473 {
1474         struct fuse_notify_poll_wakeup_out outarg;
1475         int err = -EINVAL;
1476
1477         if (size != sizeof(outarg))
1478                 goto err;
1479
1480         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1481         if (err)
1482                 goto err;
1483
1484         fuse_copy_finish(cs);
1485         return fuse_notify_poll_wakeup(fc, &outarg);
1486
1487 err:
1488         fuse_copy_finish(cs);
1489         return err;
1490 }
1491
1492 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1493                                    struct fuse_copy_state *cs)
1494 {
1495         struct fuse_notify_inval_inode_out outarg;
1496         int err = -EINVAL;
1497
1498         if (size != sizeof(outarg))
1499                 goto err;
1500
1501         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1502         if (err)
1503                 goto err;
1504         fuse_copy_finish(cs);
1505
1506         down_read(&fc->killsb);
1507         err = -ENOENT;
1508         if (fc->sb) {
1509                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1510                                                outarg.off, outarg.len);
1511         }
1512         up_read(&fc->killsb);
1513         return err;
1514
1515 err:
1516         fuse_copy_finish(cs);
1517         return err;
1518 }
1519
1520 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1521                                    struct fuse_copy_state *cs)
1522 {
1523         struct fuse_notify_inval_entry_out outarg;
1524         int err = -ENOMEM;
1525         char *buf;
1526         struct qstr name;
1527
1528         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1529         if (!buf)
1530                 goto err;
1531
1532         err = -EINVAL;
1533         if (size < sizeof(outarg))
1534                 goto err;
1535
1536         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1537         if (err)
1538                 goto err;
1539
1540         err = -ENAMETOOLONG;
1541         if (outarg.namelen > FUSE_NAME_MAX)
1542                 goto err;
1543
1544         err = -EINVAL;
1545         if (size != sizeof(outarg) + outarg.namelen + 1)
1546                 goto err;
1547
1548         name.name = buf;
1549         name.len = outarg.namelen;
1550         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1551         if (err)
1552                 goto err;
1553         fuse_copy_finish(cs);
1554         buf[outarg.namelen] = 0;
1555
1556         down_read(&fc->killsb);
1557         err = -ENOENT;
1558         if (fc->sb)
1559                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1560         up_read(&fc->killsb);
1561         kfree(buf);
1562         return err;
1563
1564 err:
1565         kfree(buf);
1566         fuse_copy_finish(cs);
1567         return err;
1568 }
1569
1570 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1571                               struct fuse_copy_state *cs)
1572 {
1573         struct fuse_notify_delete_out outarg;
1574         int err = -ENOMEM;
1575         char *buf;
1576         struct qstr name;
1577
1578         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1579         if (!buf)
1580                 goto err;
1581
1582         err = -EINVAL;
1583         if (size < sizeof(outarg))
1584                 goto err;
1585
1586         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1587         if (err)
1588                 goto err;
1589
1590         err = -ENAMETOOLONG;
1591         if (outarg.namelen > FUSE_NAME_MAX)
1592                 goto err;
1593
1594         err = -EINVAL;
1595         if (size != sizeof(outarg) + outarg.namelen + 1)
1596                 goto err;
1597
1598         name.name = buf;
1599         name.len = outarg.namelen;
1600         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1601         if (err)
1602                 goto err;
1603         fuse_copy_finish(cs);
1604         buf[outarg.namelen] = 0;
1605
1606         down_read(&fc->killsb);
1607         err = -ENOENT;
1608         if (fc->sb)
1609                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1610                                                outarg.child, &name);
1611         up_read(&fc->killsb);
1612         kfree(buf);
1613         return err;
1614
1615 err:
1616         kfree(buf);
1617         fuse_copy_finish(cs);
1618         return err;
1619 }
1620
1621 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1622                              struct fuse_copy_state *cs)
1623 {
1624         struct fuse_notify_store_out outarg;
1625         struct inode *inode;
1626         struct address_space *mapping;
1627         u64 nodeid;
1628         int err;
1629         pgoff_t index;
1630         unsigned int offset;
1631         unsigned int num;
1632         loff_t file_size;
1633         loff_t end;
1634
1635         err = -EINVAL;
1636         if (size < sizeof(outarg))
1637                 goto out_finish;
1638
1639         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1640         if (err)
1641                 goto out_finish;
1642
1643         err = -EINVAL;
1644         if (size - sizeof(outarg) != outarg.size)
1645                 goto out_finish;
1646
1647         nodeid = outarg.nodeid;
1648
1649         down_read(&fc->killsb);
1650
1651         err = -ENOENT;
1652         if (!fc->sb)
1653                 goto out_up_killsb;
1654
1655         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1656         if (!inode)
1657                 goto out_up_killsb;
1658
1659         mapping = inode->i_mapping;
1660         index = outarg.offset >> PAGE_SHIFT;
1661         offset = outarg.offset & ~PAGE_MASK;
1662         file_size = i_size_read(inode);
1663         end = outarg.offset + outarg.size;
1664         if (end > file_size) {
1665                 file_size = end;
1666                 fuse_write_update_size(inode, file_size);
1667         }
1668
1669         num = outarg.size;
1670         while (num) {
1671                 struct page *page;
1672                 unsigned int this_num;
1673
1674                 err = -ENOMEM;
1675                 page = find_or_create_page(mapping, index,
1676                                            mapping_gfp_mask(mapping));
1677                 if (!page)
1678                         goto out_iput;
1679
1680                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1681                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1682                 if (!err && offset == 0 &&
1683                     (this_num == PAGE_SIZE || file_size == end))
1684                         SetPageUptodate(page);
1685                 unlock_page(page);
1686                 put_page(page);
1687
1688                 if (err)
1689                         goto out_iput;
1690
1691                 num -= this_num;
1692                 offset = 0;
1693                 index++;
1694         }
1695
1696         err = 0;
1697
1698 out_iput:
1699         iput(inode);
1700 out_up_killsb:
1701         up_read(&fc->killsb);
1702 out_finish:
1703         fuse_copy_finish(cs);
1704         return err;
1705 }
1706
1707 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1708 {
1709         release_pages(req->pages, req->num_pages);
1710 }
1711
1712 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1713                          struct fuse_notify_retrieve_out *outarg)
1714 {
1715         int err;
1716         struct address_space *mapping = inode->i_mapping;
1717         struct fuse_req *req;
1718         pgoff_t index;
1719         loff_t file_size;
1720         unsigned int num;
1721         unsigned int offset;
1722         size_t total_len = 0;
1723         unsigned int num_pages;
1724
1725         offset = outarg->offset & ~PAGE_MASK;
1726         file_size = i_size_read(inode);
1727
1728         num = outarg->size;
1729         if (outarg->offset > file_size)
1730                 num = 0;
1731         else if (outarg->offset + num > file_size)
1732                 num = file_size - outarg->offset;
1733
1734         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1735         num_pages = min(num_pages, fc->max_pages);
1736
1737         req = fuse_get_req(fc, num_pages);
1738         if (IS_ERR(req))
1739                 return PTR_ERR(req);
1740
1741         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1742         req->in.h.nodeid = outarg->nodeid;
1743         req->in.numargs = 2;
1744         req->in.argpages = 1;
1745         req->page_descs[0].offset = offset;
1746         req->end = fuse_retrieve_end;
1747
1748         index = outarg->offset >> PAGE_SHIFT;
1749
1750         while (num && req->num_pages < num_pages) {
1751                 struct page *page;
1752                 unsigned int this_num;
1753
1754                 page = find_get_page(mapping, index);
1755                 if (!page)
1756                         break;
1757
1758                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1759                 req->pages[req->num_pages] = page;
1760                 req->page_descs[req->num_pages].length = this_num;
1761                 req->num_pages++;
1762
1763                 offset = 0;
1764                 num -= this_num;
1765                 total_len += this_num;
1766                 index++;
1767         }
1768         req->misc.retrieve_in.offset = outarg->offset;
1769         req->misc.retrieve_in.size = total_len;
1770         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1771         req->in.args[0].value = &req->misc.retrieve_in;
1772         req->in.args[1].size = total_len;
1773
1774         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1775         if (err) {
1776                 fuse_retrieve_end(fc, req);
1777                 fuse_put_request(fc, req);
1778         }
1779
1780         return err;
1781 }
1782
1783 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1784                                 struct fuse_copy_state *cs)
1785 {
1786         struct fuse_notify_retrieve_out outarg;
1787         struct inode *inode;
1788         int err;
1789
1790         err = -EINVAL;
1791         if (size != sizeof(outarg))
1792                 goto copy_finish;
1793
1794         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1795         if (err)
1796                 goto copy_finish;
1797
1798         fuse_copy_finish(cs);
1799
1800         down_read(&fc->killsb);
1801         err = -ENOENT;
1802         if (fc->sb) {
1803                 u64 nodeid = outarg.nodeid;
1804
1805                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1806                 if (inode) {
1807                         err = fuse_retrieve(fc, inode, &outarg);
1808                         iput(inode);
1809                 }
1810         }
1811         up_read(&fc->killsb);
1812
1813         return err;
1814
1815 copy_finish:
1816         fuse_copy_finish(cs);
1817         return err;
1818 }
1819
1820 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1821                        unsigned int size, struct fuse_copy_state *cs)
1822 {
1823         /* Don't try to move pages (yet) */
1824         cs->move_pages = 0;
1825
1826         switch (code) {
1827         case FUSE_NOTIFY_POLL:
1828                 return fuse_notify_poll(fc, size, cs);
1829
1830         case FUSE_NOTIFY_INVAL_INODE:
1831                 return fuse_notify_inval_inode(fc, size, cs);
1832
1833         case FUSE_NOTIFY_INVAL_ENTRY:
1834                 return fuse_notify_inval_entry(fc, size, cs);
1835
1836         case FUSE_NOTIFY_STORE:
1837                 return fuse_notify_store(fc, size, cs);
1838
1839         case FUSE_NOTIFY_RETRIEVE:
1840                 return fuse_notify_retrieve(fc, size, cs);
1841
1842         case FUSE_NOTIFY_DELETE:
1843                 return fuse_notify_delete(fc, size, cs);
1844
1845         default:
1846                 fuse_copy_finish(cs);
1847                 return -EINVAL;
1848         }
1849 }
1850
1851 /* Look up request on processing list by unique ID */
1852 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1853 {
1854         unsigned int hash = fuse_req_hash(unique);
1855         struct fuse_req *req;
1856
1857         list_for_each_entry(req, &fpq->processing[hash], list) {
1858                 if (req->in.h.unique == unique)
1859                         return req;
1860         }
1861         return NULL;
1862 }
1863
1864 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1865                          unsigned nbytes)
1866 {
1867         unsigned reqsize = sizeof(struct fuse_out_header);
1868
1869         if (out->h.error)
1870                 return nbytes != reqsize ? -EINVAL : 0;
1871
1872         reqsize += len_args(out->numargs, out->args);
1873
1874         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1875                 return -EINVAL;
1876         else if (reqsize > nbytes) {
1877                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1878                 unsigned diffsize = reqsize - nbytes;
1879                 if (diffsize > lastarg->size)
1880                         return -EINVAL;
1881                 lastarg->size -= diffsize;
1882         }
1883         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1884                               out->page_zeroing);
1885 }
1886
1887 /*
1888  * Write a single reply to a request.  First the header is copied from
1889  * the write buffer.  The request is then searched on the processing
1890  * list by the unique ID found in the header.  If found, then remove
1891  * it from the list and copy the rest of the buffer to the request.
1892  * The request is finished by calling request_end()
1893  */
1894 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1895                                  struct fuse_copy_state *cs, size_t nbytes)
1896 {
1897         int err;
1898         struct fuse_conn *fc = fud->fc;
1899         struct fuse_pqueue *fpq = &fud->pq;
1900         struct fuse_req *req;
1901         struct fuse_out_header oh;
1902
1903         if (nbytes < sizeof(struct fuse_out_header))
1904                 return -EINVAL;
1905
1906         err = fuse_copy_one(cs, &oh, sizeof(oh));
1907         if (err)
1908                 goto err_finish;
1909
1910         err = -EINVAL;
1911         if (oh.len != nbytes)
1912                 goto err_finish;
1913
1914         /*
1915          * Zero oh.unique indicates unsolicited notification message
1916          * and error contains notification code.
1917          */
1918         if (!oh.unique) {
1919                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1920                 return err ? err : nbytes;
1921         }
1922
1923         err = -EINVAL;
1924         if (oh.error <= -1000 || oh.error > 0)
1925                 goto err_finish;
1926
1927         spin_lock(&fpq->lock);
1928         err = -ENOENT;
1929         if (!fpq->connected)
1930                 goto err_unlock_pq;
1931
1932         req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1933         if (!req)
1934                 goto err_unlock_pq;
1935
1936         /* Is it an interrupt reply ID? */
1937         if (oh.unique & FUSE_INT_REQ_BIT) {
1938                 __fuse_get_request(req);
1939                 spin_unlock(&fpq->lock);
1940
1941                 err = -EINVAL;
1942                 if (nbytes != sizeof(struct fuse_out_header)) {
1943                         fuse_put_request(fc, req);
1944                         goto err_finish;
1945                 }
1946
1947                 if (oh.error == -ENOSYS)
1948                         fc->no_interrupt = 1;
1949                 else if (oh.error == -EAGAIN)
1950                         queue_interrupt(&fc->iq, req);
1951                 fuse_put_request(fc, req);
1952
1953                 fuse_copy_finish(cs);
1954                 return nbytes;
1955         }
1956
1957         clear_bit(FR_SENT, &req->flags);
1958         list_move(&req->list, &fpq->io);
1959         req->out.h = oh;
1960         set_bit(FR_LOCKED, &req->flags);
1961         spin_unlock(&fpq->lock);
1962         cs->req = req;
1963         if (!req->out.page_replace)
1964                 cs->move_pages = 0;
1965
1966         err = copy_out_args(cs, &req->out, nbytes);
1967         fuse_copy_finish(cs);
1968
1969         spin_lock(&fpq->lock);
1970         clear_bit(FR_LOCKED, &req->flags);
1971         if (!fpq->connected)
1972                 err = -ENOENT;
1973         else if (err)
1974                 req->out.h.error = -EIO;
1975         if (!test_bit(FR_PRIVATE, &req->flags))
1976                 list_del_init(&req->list);
1977         spin_unlock(&fpq->lock);
1978
1979         request_end(fc, req);
1980
1981         return err ? err : nbytes;
1982
1983  err_unlock_pq:
1984         spin_unlock(&fpq->lock);
1985  err_finish:
1986         fuse_copy_finish(cs);
1987         return err;
1988 }
1989
1990 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1991 {
1992         struct fuse_copy_state cs;
1993         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1994
1995         if (!fud)
1996                 return -EPERM;
1997
1998         if (!iter_is_iovec(from))
1999                 return -EINVAL;
2000
2001         fuse_copy_init(&cs, 0, from);
2002
2003         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
2004 }
2005
2006 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
2007                                      struct file *out, loff_t *ppos,
2008                                      size_t len, unsigned int flags)
2009 {
2010         unsigned nbuf;
2011         unsigned idx;
2012         struct pipe_buffer *bufs;
2013         struct fuse_copy_state cs;
2014         struct fuse_dev *fud;
2015         size_t rem;
2016         ssize_t ret;
2017
2018         fud = fuse_get_dev(out);
2019         if (!fud)
2020                 return -EPERM;
2021
2022         pipe_lock(pipe);
2023
2024         bufs = kvmalloc_array(pipe->nrbufs, sizeof(struct pipe_buffer),
2025                               GFP_KERNEL);
2026         if (!bufs) {
2027                 pipe_unlock(pipe);
2028                 return -ENOMEM;
2029         }
2030
2031         nbuf = 0;
2032         rem = 0;
2033         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2034                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2035
2036         ret = -EINVAL;
2037         if (rem < len) {
2038                 pipe_unlock(pipe);
2039                 goto out;
2040         }
2041
2042         rem = len;
2043         while (rem) {
2044                 struct pipe_buffer *ibuf;
2045                 struct pipe_buffer *obuf;
2046
2047                 BUG_ON(nbuf >= pipe->buffers);
2048                 BUG_ON(!pipe->nrbufs);
2049                 ibuf = &pipe->bufs[pipe->curbuf];
2050                 obuf = &bufs[nbuf];
2051
2052                 if (rem >= ibuf->len) {
2053                         *obuf = *ibuf;
2054                         ibuf->ops = NULL;
2055                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2056                         pipe->nrbufs--;
2057                 } else {
2058                         pipe_buf_get(pipe, ibuf);
2059                         *obuf = *ibuf;
2060                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2061                         obuf->len = rem;
2062                         ibuf->offset += obuf->len;
2063                         ibuf->len -= obuf->len;
2064                 }
2065                 nbuf++;
2066                 rem -= obuf->len;
2067         }
2068         pipe_unlock(pipe);
2069
2070         fuse_copy_init(&cs, 0, NULL);
2071         cs.pipebufs = bufs;
2072         cs.nr_segs = nbuf;
2073         cs.pipe = pipe;
2074
2075         if (flags & SPLICE_F_MOVE)
2076                 cs.move_pages = 1;
2077
2078         ret = fuse_dev_do_write(fud, &cs, len);
2079
2080         for (idx = 0; idx < nbuf; idx++)
2081                 pipe_buf_release(pipe, &bufs[idx]);
2082
2083 out:
2084         kvfree(bufs);
2085         return ret;
2086 }
2087
2088 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2089 {
2090         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2091         struct fuse_iqueue *fiq;
2092         struct fuse_dev *fud = fuse_get_dev(file);
2093
2094         if (!fud)
2095                 return EPOLLERR;
2096
2097         fiq = &fud->fc->iq;
2098         poll_wait(file, &fiq->waitq, wait);
2099
2100         spin_lock(&fiq->waitq.lock);
2101         if (!fiq->connected)
2102                 mask = EPOLLERR;
2103         else if (request_pending(fiq))
2104                 mask |= EPOLLIN | EPOLLRDNORM;
2105         spin_unlock(&fiq->waitq.lock);
2106
2107         return mask;
2108 }
2109
2110 /*
2111  * Abort all requests on the given list (pending or processing)
2112  *
2113  * This function releases and reacquires fc->lock
2114  */
2115 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2116 {
2117         while (!list_empty(head)) {
2118                 struct fuse_req *req;
2119                 req = list_entry(head->next, struct fuse_req, list);
2120                 req->out.h.error = -ECONNABORTED;
2121                 clear_bit(FR_SENT, &req->flags);
2122                 list_del_init(&req->list);
2123                 request_end(fc, req);
2124         }
2125 }
2126
2127 static void end_polls(struct fuse_conn *fc)
2128 {
2129         struct rb_node *p;
2130
2131         p = rb_first(&fc->polled_files);
2132
2133         while (p) {
2134                 struct fuse_file *ff;
2135                 ff = rb_entry(p, struct fuse_file, polled_node);
2136                 wake_up_interruptible_all(&ff->poll_wait);
2137
2138                 p = rb_next(p);
2139         }
2140 }
2141
2142 /*
2143  * Abort all requests.
2144  *
2145  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2146  * filesystem.
2147  *
2148  * The same effect is usually achievable through killing the filesystem daemon
2149  * and all users of the filesystem.  The exception is the combination of an
2150  * asynchronous request and the tricky deadlock (see
2151  * Documentation/filesystems/fuse.txt).
2152  *
2153  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2154  * requests, they should be finished off immediately.  Locked requests will be
2155  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2156  * requests.  It is possible that some request will finish before we can.  This
2157  * is OK, the request will in that case be removed from the list before we touch
2158  * it.
2159  */
2160 void fuse_abort_conn(struct fuse_conn *fc, bool is_abort)
2161 {
2162         struct fuse_iqueue *fiq = &fc->iq;
2163
2164         spin_lock(&fc->lock);
2165         if (fc->connected) {
2166                 struct fuse_dev *fud;
2167                 struct fuse_req *req, *next;
2168                 LIST_HEAD(to_end);
2169                 unsigned int i;
2170
2171                 /* Background queuing checks fc->connected under bg_lock */
2172                 spin_lock(&fc->bg_lock);
2173                 fc->connected = 0;
2174                 spin_unlock(&fc->bg_lock);
2175
2176                 fc->aborted = is_abort;
2177                 fuse_set_initialized(fc);
2178                 list_for_each_entry(fud, &fc->devices, entry) {
2179                         struct fuse_pqueue *fpq = &fud->pq;
2180
2181                         spin_lock(&fpq->lock);
2182                         fpq->connected = 0;
2183                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2184                                 req->out.h.error = -ECONNABORTED;
2185                                 spin_lock(&req->waitq.lock);
2186                                 set_bit(FR_ABORTED, &req->flags);
2187                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2188                                         set_bit(FR_PRIVATE, &req->flags);
2189                                         __fuse_get_request(req);
2190                                         list_move(&req->list, &to_end);
2191                                 }
2192                                 spin_unlock(&req->waitq.lock);
2193                         }
2194                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2195                                 list_splice_tail_init(&fpq->processing[i],
2196                                                       &to_end);
2197                         spin_unlock(&fpq->lock);
2198                 }
2199                 spin_lock(&fc->bg_lock);
2200                 fc->blocked = 0;
2201                 fc->max_background = UINT_MAX;
2202                 flush_bg_queue(fc);
2203                 spin_unlock(&fc->bg_lock);
2204
2205                 spin_lock(&fiq->waitq.lock);
2206                 fiq->connected = 0;
2207                 list_for_each_entry(req, &fiq->pending, list)
2208                         clear_bit(FR_PENDING, &req->flags);
2209                 list_splice_tail_init(&fiq->pending, &to_end);
2210                 while (forget_pending(fiq))
2211                         kfree(dequeue_forget(fiq, 1, NULL));
2212                 wake_up_all_locked(&fiq->waitq);
2213                 spin_unlock(&fiq->waitq.lock);
2214                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2215                 end_polls(fc);
2216                 wake_up_all(&fc->blocked_waitq);
2217                 spin_unlock(&fc->lock);
2218
2219                 end_requests(fc, &to_end);
2220         } else {
2221                 spin_unlock(&fc->lock);
2222         }
2223 }
2224 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2225
2226 void fuse_wait_aborted(struct fuse_conn *fc)
2227 {
2228         /* matches implicit memory barrier in fuse_drop_waiting() */
2229         smp_mb();
2230         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2231 }
2232
2233 int fuse_dev_release(struct inode *inode, struct file *file)
2234 {
2235         struct fuse_dev *fud = fuse_get_dev(file);
2236
2237         if (fud) {
2238                 struct fuse_conn *fc = fud->fc;
2239                 struct fuse_pqueue *fpq = &fud->pq;
2240                 LIST_HEAD(to_end);
2241                 unsigned int i;
2242
2243                 spin_lock(&fpq->lock);
2244                 WARN_ON(!list_empty(&fpq->io));
2245                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2246                         list_splice_init(&fpq->processing[i], &to_end);
2247                 spin_unlock(&fpq->lock);
2248
2249                 end_requests(fc, &to_end);
2250
2251                 /* Are we the last open device? */
2252                 if (atomic_dec_and_test(&fc->dev_count)) {
2253                         WARN_ON(fc->iq.fasync != NULL);
2254                         fuse_abort_conn(fc, false);
2255                 }
2256                 fuse_dev_free(fud);
2257         }
2258         return 0;
2259 }
2260 EXPORT_SYMBOL_GPL(fuse_dev_release);
2261
2262 static int fuse_dev_fasync(int fd, struct file *file, int on)
2263 {
2264         struct fuse_dev *fud = fuse_get_dev(file);
2265
2266         if (!fud)
2267                 return -EPERM;
2268
2269         /* No locking - fasync_helper does its own locking */
2270         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2271 }
2272
2273 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2274 {
2275         struct fuse_dev *fud;
2276
2277         if (new->private_data)
2278                 return -EINVAL;
2279
2280         fud = fuse_dev_alloc(fc);
2281         if (!fud)
2282                 return -ENOMEM;
2283
2284         new->private_data = fud;
2285         atomic_inc(&fc->dev_count);
2286
2287         return 0;
2288 }
2289
2290 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2291                            unsigned long arg)
2292 {
2293         int err = -ENOTTY;
2294
2295         if (cmd == FUSE_DEV_IOC_CLONE) {
2296                 int oldfd;
2297
2298                 err = -EFAULT;
2299                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2300                         struct file *old = fget(oldfd);
2301
2302                         err = -EINVAL;
2303                         if (old) {
2304                                 struct fuse_dev *fud = NULL;
2305
2306                                 /*
2307                                  * Check against file->f_op because CUSE
2308                                  * uses the same ioctl handler.
2309                                  */
2310                                 if (old->f_op == file->f_op &&
2311                                     old->f_cred->user_ns == file->f_cred->user_ns)
2312                                         fud = fuse_get_dev(old);
2313
2314                                 if (fud) {
2315                                         mutex_lock(&fuse_mutex);
2316                                         err = fuse_device_clone(fud->fc, file);
2317                                         mutex_unlock(&fuse_mutex);
2318                                 }
2319                                 fput(old);
2320                         }
2321                 }
2322         }
2323         return err;
2324 }
2325
2326 const struct file_operations fuse_dev_operations = {
2327         .owner          = THIS_MODULE,
2328         .open           = fuse_dev_open,
2329         .llseek         = no_llseek,
2330         .read_iter      = fuse_dev_read,
2331         .splice_read    = fuse_dev_splice_read,
2332         .write_iter     = fuse_dev_write,
2333         .splice_write   = fuse_dev_splice_write,
2334         .poll           = fuse_dev_poll,
2335         .release        = fuse_dev_release,
2336         .fasync         = fuse_dev_fasync,
2337         .unlocked_ioctl = fuse_dev_ioctl,
2338         .compat_ioctl   = fuse_dev_ioctl,
2339 };
2340 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2341
2342 static struct miscdevice fuse_miscdevice = {
2343         .minor = FUSE_MINOR,
2344         .name  = "fuse",
2345         .fops = &fuse_dev_operations,
2346 };
2347
2348 int __init fuse_dev_init(void)
2349 {
2350         int err = -ENOMEM;
2351         fuse_req_cachep = kmem_cache_create("fuse_request",
2352                                             sizeof(struct fuse_req),
2353                                             0, 0, NULL);
2354         if (!fuse_req_cachep)
2355                 goto out;
2356
2357         err = misc_register(&fuse_miscdevice);
2358         if (err)
2359                 goto out_cache_clean;
2360
2361         return 0;
2362
2363  out_cache_clean:
2364         kmem_cache_destroy(fuse_req_cachep);
2365  out:
2366         return err;
2367 }
2368
2369 void fuse_dev_cleanup(void)
2370 {
2371         misc_deregister(&fuse_miscdevice);
2372         kmem_cache_destroy(fuse_req_cachep);
2373 }