fs/fuse/dev.c

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