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 ACCESS_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(sizeof(struct page *) * npages, flags);
  68                         page_descs = kmalloc(sizeof(struct fuse_page_desc) *
  69                                              npages, flags);
  70                 }
  71
  72                 if (!pages || !page_descs) {
  73                         kfree(pages);
  74                         kfree(page_descs);
  75                         kmem_cache_free(fuse_req_cachep, req);
  76                         return NULL;
  77                 }
  78
  79                 fuse_request_init(req, pages, page_descs, npages);
  80         }
  81         return req;
  82 }
  83
  84 struct fuse_req *fuse_request_alloc(unsigned npages)
  85 {
  86         return __fuse_request_alloc(npages, GFP_KERNEL);
  87 }
  88 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  89
  90 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  91 {
  92         return __fuse_request_alloc(npages, GFP_NOFS);
  93 }
  94
  95 void fuse_request_free(struct fuse_req *req)
  96 {
  97         if (req->pages != req->inline_pages) {
  98                 kfree(req->pages);
  99                 kfree(req->page_descs);
 100         }
 101         kmem_cache_free(fuse_req_cachep, req);
 102 }
 103
 104 void __fuse_get_request(struct fuse_req *req)
 105 {
 106         refcount_inc(&req->count);
 107 }
 108
 109 /* Must be called with > 1 refcount */
 110 static void __fuse_put_request(struct fuse_req *req)
 111 {
 112         refcount_dec(&req->count);
 113 }
 114
 115 static void fuse_req_init_context(struct fuse_conn *fc, struct fuse_req *req)
 116 {
 117         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 118         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 119         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 120 }
 121
 122 void fuse_set_initialized(struct fuse_conn *fc)
 123 {
 124         /* Make sure stores before this are seen on another CPU */
 125         smp_wmb();
 126         fc->initialized = 1;
 127 }
 128
 129 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 130 {
 131         return !fc->initialized || (for_background && fc->blocked);
 132 }
 133
 134 static void fuse_drop_waiting(struct fuse_conn *fc)
 135 {
 136         /*
 137          * lockess check of fc->connected is okay, because atomic_dec_and_test()
 138          * provides a memory barrier mached with the one in fuse_wait_aborted()
 139          * to ensure no wake-up is missed.
 140          */
 141         if (atomic_dec_and_test(&fc->num_waiting) &&
 142             !READ_ONCE(fc->connected)) {
 143                 /* wake up aborters */
 144                 wake_up_all(&fc->blocked_waitq);
 145         }
 146 }
 147
 148 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 149                                        bool for_background)
 150 {
 151         struct fuse_req *req;
 152         int err;
 153         atomic_inc(&fc->num_waiting);
 154
 155         if (fuse_block_alloc(fc, for_background)) {
 156                 err = -EINTR;
 157                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 158                                 !fuse_block_alloc(fc, for_background)))
 159                         goto out;
 160         }
 161         /* Matches smp_wmb() in fuse_set_initialized() */
 162         smp_rmb();
 163
 164         err = -ENOTCONN;
 165         if (!fc->connected)
 166                 goto out;
 167
 168         err = -ECONNREFUSED;
 169         if (fc->conn_error)
 170                 goto out;
 171
 172         req = fuse_request_alloc(npages);
 173         err = -ENOMEM;
 174         if (!req) {
 175                 if (for_background)
 176                         wake_up(&fc->blocked_waitq);
 177                 goto out;
 178         }
 179
 180         fuse_req_init_context(fc, req);
 181         __set_bit(FR_WAITING, &req->flags);
 182         if (for_background)
 183                 __set_bit(FR_BACKGROUND, &req->flags);
 184
 185         return req;
 186
 187  out:
 188         fuse_drop_waiting(fc);
 189         return ERR_PTR(err);
 190 }
 191
 192 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 193 {
 194         return __fuse_get_req(fc, npages, false);
 195 }
 196 EXPORT_SYMBOL_GPL(fuse_get_req);
 197
 198 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 199                                              unsigned npages)
 200 {
 201         return __fuse_get_req(fc, npages, true);
 202 }
 203 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 204
 205 /*
 206  * Return request in fuse_file->reserved_req.  However that may
 207  * currently be in use.  If that is the case, wait for it to become
 208  * available.
 209  */
 210 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 211                                          struct file *file)
 212 {
 213         struct fuse_req *req = NULL;
 214         struct fuse_file *ff = file->private_data;
 215
 216         do {
 217                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 218                 spin_lock(&fc->lock);
 219                 if (ff->reserved_req) {
 220                         req = ff->reserved_req;
 221                         ff->reserved_req = NULL;
 222                         req->stolen_file = get_file(file);
 223                 }
 224                 spin_unlock(&fc->lock);
 225         } while (!req);
 226
 227         return req;
 228 }
 229
 230 /*
 231  * Put stolen request back into fuse_file->reserved_req
 232  */
 233 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 234 {
 235         struct file *file = req->stolen_file;
 236         struct fuse_file *ff = file->private_data;
 237
 238         spin_lock(&fc->lock);
 239         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 240         BUG_ON(ff->reserved_req);
 241         ff->reserved_req = req;
 242         wake_up_all(&fc->reserved_req_waitq);
 243         spin_unlock(&fc->lock);
 244         fput(file);
 245 }
 246
 247 /*
 248  * Gets a requests for a file operation, always succeeds
 249  *
 250  * This is used for sending the FLUSH request, which must get to
 251  * userspace, due to POSIX locks which may need to be unlocked.
 252  *
 253  * If allocation fails due to OOM, use the reserved request in
 254  * fuse_file.
 255  *
 256  * This is very unlikely to deadlock accidentally, since the
 257  * filesystem should not have it's own file open.  If deadlock is
 258  * intentional, it can still be broken by "aborting" the filesystem.
 259  */
 260 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 261                                              struct file *file)
 262 {
 263         struct fuse_req *req;
 264
 265         atomic_inc(&fc->num_waiting);
 266         wait_event(fc->blocked_waitq, fc->initialized);
 267         /* Matches smp_wmb() in fuse_set_initialized() */
 268         smp_rmb();
 269         req = fuse_request_alloc(0);
 270         if (!req)
 271                 req = get_reserved_req(fc, file);
 272
 273         fuse_req_init_context(fc, req);
 274         __set_bit(FR_WAITING, &req->flags);
 275         __clear_bit(FR_BACKGROUND, &req->flags);
 276         return req;
 277 }
 278
 279 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 280 {
 281         if (refcount_dec_and_test(&req->count)) {
 282                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 283                         /*
 284                          * We get here in the unlikely case that a background
 285                          * request was allocated but not sent
 286                          */
 287                         spin_lock(&fc->lock);
 288                         if (!fc->blocked)
 289                                 wake_up(&fc->blocked_waitq);
 290                         spin_unlock(&fc->lock);
 291                 }
 292
 293                 if (test_bit(FR_WAITING, &req->flags)) {
 294                         __clear_bit(FR_WAITING, &req->flags);
 295                         fuse_drop_waiting(fc);
 296                 }
 297
 298                 if (req->stolen_file)
 299                         put_reserved_req(fc, req);
 300                 else
 301                         fuse_request_free(req);
 302         }
 303 }
 304 EXPORT_SYMBOL_GPL(fuse_put_request);
 305
 306 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 307 {
 308         unsigned nbytes = 0;
 309         unsigned i;
 310
 311         for (i = 0; i < numargs; i++)
 312                 nbytes += args[i].size;
 313
 314         return nbytes;
 315 }
 316
 317 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 318 {
 319         return ++fiq->reqctr;
 320 }
 321
 322 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 323 {
 324         req->in.h.len = sizeof(struct fuse_in_header) +
 325                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 326         list_add_tail(&req->list, &fiq->pending);
 327         wake_up_locked(&fiq->waitq);
 328         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 329 }
 330
 331 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 332                        u64 nodeid, u64 nlookup)
 333 {
 334         struct fuse_iqueue *fiq = &fc->iq;
 335
 336         forget->forget_one.nodeid = nodeid;
 337         forget->forget_one.nlookup = nlookup;
 338
 339         spin_lock(&fiq->waitq.lock);
 340         if (fiq->connected) {
 341                 fiq->forget_list_tail->next = forget;
 342                 fiq->forget_list_tail = forget;
 343                 wake_up_locked(&fiq->waitq);
 344                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 345         } else {
 346                 kfree(forget);
 347         }
 348         spin_unlock(&fiq->waitq.lock);
 349 }
 350
 351 static void flush_bg_queue(struct fuse_conn *fc)
 352 {
 353         while (fc->active_background < fc->max_background &&
 354                !list_empty(&fc->bg_queue)) {
 355                 struct fuse_req *req;
 356                 struct fuse_iqueue *fiq = &fc->iq;
 357
 358                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 359                 list_del(&req->list);
 360                 fc->active_background++;
 361                 spin_lock(&fiq->waitq.lock);
 362                 req->in.h.unique = fuse_get_unique(fiq);
 363                 queue_request(fiq, req);
 364                 spin_unlock(&fiq->waitq.lock);
 365         }
 366 }
 367
 368 /*
 369  * This function is called when a request is finished.  Either a reply
 370  * has arrived or it was aborted (and not yet sent) or some error
 371  * occurred during communication with userspace, or the device file
 372  * was closed.  The requester thread is woken up (if still waiting),
 373  * the 'end' callback is called if given, else the reference to the
 374  * request is released
 375  */
 376 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 377 {
 378         struct fuse_iqueue *fiq = &fc->iq;
 379
 380         if (test_and_set_bit(FR_FINISHED, &req->flags))
 381                 goto put_request;
 382
 383         spin_lock(&fiq->waitq.lock);
 384         list_del_init(&req->intr_entry);
 385         spin_unlock(&fiq->waitq.lock);
 386         WARN_ON(test_bit(FR_PENDING, &req->flags));
 387         WARN_ON(test_bit(FR_SENT, &req->flags));
 388         if (test_bit(FR_BACKGROUND, &req->flags)) {
 389                 spin_lock(&fc->lock);
 390                 clear_bit(FR_BACKGROUND, &req->flags);
 391                 if (fc->num_background == fc->max_background) {
 392                         fc->blocked = 0;
 393                         wake_up(&fc->blocked_waitq);
 394                 } else if (!fc->blocked) {
 395                         /*
 396                          * Wake up next waiter, if any.  It's okay to use
 397                          * waitqueue_active(), as we've already synced up
 398                          * fc->blocked with waiters with the wake_up() call
 399                          * above.
 400                          */
 401                         if (waitqueue_active(&fc->blocked_waitq))
 402                                 wake_up(&fc->blocked_waitq);
 403                 }
 404
 405                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 406                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 407                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 408                 }
 409                 fc->num_background--;
 410                 fc->active_background--;
 411                 flush_bg_queue(fc);
 412                 spin_unlock(&fc->lock);
 413         }
 414         wake_up(&req->waitq);
 415         if (req->end)
 416                 req->end(fc, req);
 417 put_request:
 418         fuse_put_request(fc, req);
 419 }
 420
 421 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 422 {
 423         spin_lock(&fiq->waitq.lock);
 424         if (test_bit(FR_FINISHED, &req->flags)) {
 425                 spin_unlock(&fiq->waitq.lock);
 426                 return;
 427         }
 428         if (list_empty(&req->intr_entry)) {
 429                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 430                 wake_up_locked(&fiq->waitq);
 431         }
 432         spin_unlock(&fiq->waitq.lock);
 433         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 434 }
 435
 436 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 437 {
 438         struct fuse_iqueue *fiq = &fc->iq;
 439         int err;
 440
 441         if (!fc->no_interrupt) {
 442                 /* Any signal may interrupt this */
 443                 err = wait_event_interruptible(req->waitq,
 444                                         test_bit(FR_FINISHED, &req->flags));
 445                 if (!err)
 446                         return;
 447
 448                 set_bit(FR_INTERRUPTED, &req->flags);
 449                 /* matches barrier in fuse_dev_do_read() */
 450                 smp_mb__after_atomic();
 451                 if (test_bit(FR_SENT, &req->flags))
 452                         queue_interrupt(fiq, req);
 453         }
 454
 455         if (!test_bit(FR_FORCE, &req->flags)) {
 456                 /* Only fatal signals may interrupt this */
 457                 err = wait_event_killable(req->waitq,
 458                                         test_bit(FR_FINISHED, &req->flags));
 459                 if (!err)
 460                         return;
 461
 462                 spin_lock(&fiq->waitq.lock);
 463                 /* Request is not yet in userspace, bail out */
 464                 if (test_bit(FR_PENDING, &req->flags)) {
 465                         list_del(&req->list);
 466                         spin_unlock(&fiq->waitq.lock);
 467                         __fuse_put_request(req);
 468                         req->out.h.error = -EINTR;
 469                         return;
 470                 }
 471                 spin_unlock(&fiq->waitq.lock);
 472         }
 473
 474         /*
 475          * Either request is already in userspace, or it was forced.
 476          * Wait it out.
 477          */
 478         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 479 }
 480
 481 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 482 {
 483         struct fuse_iqueue *fiq = &fc->iq;
 484
 485         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 486         spin_lock(&fiq->waitq.lock);
 487         if (!fiq->connected) {
 488                 spin_unlock(&fiq->waitq.lock);
 489                 req->out.h.error = -ENOTCONN;
 490         } else {
 491                 req->in.h.unique = fuse_get_unique(fiq);
 492                 queue_request(fiq, req);
 493                 /* acquire extra reference, since request is still needed
 494                    after request_end() */
 495                 __fuse_get_request(req);
 496                 spin_unlock(&fiq->waitq.lock);
 497
 498                 request_wait_answer(fc, req);
 499                 /* Pairs with smp_wmb() in request_end() */
 500                 smp_rmb();
 501         }
 502 }
 503
 504 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 505 {
 506         __set_bit(FR_ISREPLY, &req->flags);
 507         if (!test_bit(FR_WAITING, &req->flags)) {
 508                 __set_bit(FR_WAITING, &req->flags);
 509                 atomic_inc(&fc->num_waiting);
 510         }
 511         __fuse_request_send(fc, req);
 512 }
 513 EXPORT_SYMBOL_GPL(fuse_request_send);
 514
 515 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 516 {
 517         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 518                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 519
 520         if (fc->minor < 9) {
 521                 switch (args->in.h.opcode) {
 522                 case FUSE_LOOKUP:
 523                 case FUSE_CREATE:
 524                 case FUSE_MKNOD:
 525                 case FUSE_MKDIR:
 526                 case FUSE_SYMLINK:
 527                 case FUSE_LINK:
 528                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 529                         break;
 530                 case FUSE_GETATTR:
 531                 case FUSE_SETATTR:
 532                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 533                         break;
 534                 }
 535         }
 536         if (fc->minor < 12) {
 537                 switch (args->in.h.opcode) {
 538                 case FUSE_CREATE:
 539                         args->in.args[0].size = sizeof(struct fuse_open_in);
 540                         break;
 541                 case FUSE_MKNOD:
 542                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 543                         break;
 544                 }
 545         }
 546 }
 547
 548 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 549 {
 550         struct fuse_req *req;
 551         ssize_t ret;
 552
 553         req = fuse_get_req(fc, 0);
 554         if (IS_ERR(req))
 555                 return PTR_ERR(req);
 556
 557         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 558         fuse_adjust_compat(fc, args);
 559
 560         req->in.h.opcode = args->in.h.opcode;
 561         req->in.h.nodeid = args->in.h.nodeid;
 562         req->in.numargs = args->in.numargs;
 563         memcpy(req->in.args, args->in.args,
 564                args->in.numargs * sizeof(struct fuse_in_arg));
 565         req->out.argvar = args->out.argvar;
 566         req->out.numargs = args->out.numargs;
 567         memcpy(req->out.args, args->out.args,
 568                args->out.numargs * sizeof(struct fuse_arg));
 569         fuse_request_send(fc, req);
 570         ret = req->out.h.error;
 571         if (!ret && args->out.argvar) {
 572                 BUG_ON(args->out.numargs != 1);
 573                 ret = req->out.args[0].size;
 574         }
 575         fuse_put_request(fc, req);
 576
 577         return ret;
 578 }
 579
 580 /*
 581  * Called under fc->lock
 582  *
 583  * fc->connected must have been checked previously
 584  */
 585 void fuse_request_send_background_locked(struct fuse_conn *fc,
 586                                          struct fuse_req *req)
 587 {
 588         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 589         if (!test_bit(FR_WAITING, &req->flags)) {
 590                 __set_bit(FR_WAITING, &req->flags);
 591                 atomic_inc(&fc->num_waiting);
 592         }
 593         __set_bit(FR_ISREPLY, &req->flags);
 594         fc->num_background++;
 595         if (fc->num_background == fc->max_background)
 596                 fc->blocked = 1;
 597         if (fc->num_background == fc->congestion_threshold && fc->sb) {
 598                 set_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 599                 set_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 600         }
 601         list_add_tail(&req->list, &fc->bg_queue);
 602         flush_bg_queue(fc);
 603 }
 604
 605 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 606 {
 607         BUG_ON(!req->end);
 608         spin_lock(&fc->lock);
 609         if (fc->connected) {
 610                 fuse_request_send_background_locked(fc, req);
 611                 spin_unlock(&fc->lock);
 612         } else {
 613                 spin_unlock(&fc->lock);
 614                 req->out.h.error = -ENOTCONN;
 615                 req->end(fc, req);
 616                 fuse_put_request(fc, req);
 617         }
 618 }
 619 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 620
 621 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 622                                           struct fuse_req *req, u64 unique)
 623 {
 624         int err = -ENODEV;
 625         struct fuse_iqueue *fiq = &fc->iq;
 626
 627         __clear_bit(FR_ISREPLY, &req->flags);
 628         req->in.h.unique = unique;
 629         spin_lock(&fiq->waitq.lock);
 630         if (fiq->connected) {
 631                 queue_request(fiq, req);
 632                 err = 0;
 633         }
 634         spin_unlock(&fiq->waitq.lock);
 635
 636         return err;
 637 }
 638
 639 void fuse_force_forget(struct file *file, u64 nodeid)
 640 {
 641         struct inode *inode = file_inode(file);
 642         struct fuse_conn *fc = get_fuse_conn(inode);
 643         struct fuse_req *req;
 644         struct fuse_forget_in inarg;
 645
 646         memset(&inarg, 0, sizeof(inarg));
 647         inarg.nlookup = 1;
 648         req = fuse_get_req_nofail_nopages(fc, file);
 649         req->in.h.opcode = FUSE_FORGET;
 650         req->in.h.nodeid = nodeid;
 651         req->in.numargs = 1;
 652         req->in.args[0].size = sizeof(inarg);
 653         req->in.args[0].value = &inarg;
 654         __clear_bit(FR_ISREPLY, &req->flags);
 655         __fuse_request_send(fc, req);
 656         /* ignore errors */
 657         fuse_put_request(fc, req);
 658 }
 659
 660 /*
 661  * Lock the request.  Up to the next unlock_request() there mustn't be
 662  * anything that could cause a page-fault.  If the request was already
 663  * aborted bail out.
 664  */
 665 static int lock_request(struct fuse_req *req)
 666 {
 667         int err = 0;
 668         if (req) {
 669                 spin_lock(&req->waitq.lock);
 670                 if (test_bit(FR_ABORTED, &req->flags))
 671                         err = -ENOENT;
 672                 else
 673                         set_bit(FR_LOCKED, &req->flags);
 674                 spin_unlock(&req->waitq.lock);
 675         }
 676         return err;
 677 }
 678
 679 /*
 680  * Unlock request.  If it was aborted while locked, caller is responsible
 681  * for unlocking and ending the request.
 682  */
 683 static int unlock_request(struct fuse_req *req)
 684 {
 685         int err = 0;
 686         if (req) {
 687                 spin_lock(&req->waitq.lock);
 688                 if (test_bit(FR_ABORTED, &req->flags))
 689                         err = -ENOENT;
 690                 else
 691                         clear_bit(FR_LOCKED, &req->flags);
 692                 spin_unlock(&req->waitq.lock);
 693         }
 694         return err;
 695 }
 696
 697 struct fuse_copy_state {
 698         int write;
 699         struct fuse_req *req;
 700         struct iov_iter *iter;
 701         struct pipe_buffer *pipebufs;
 702         struct pipe_buffer *currbuf;
 703         struct pipe_inode_info *pipe;
 704         unsigned long nr_segs;
 705         struct page *pg;
 706         unsigned len;
 707         unsigned offset;
 708         unsigned move_pages:1;
 709 };
 710
 711 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 712                            struct iov_iter *iter)
 713 {
 714         memset(cs, 0, sizeof(*cs));
 715         cs->write = write;
 716         cs->iter = iter;
 717 }
 718
 719 /* Unmap and put previous page of userspace buffer */
 720 static void fuse_copy_finish(struct fuse_copy_state *cs)
 721 {
 722         if (cs->currbuf) {
 723                 struct pipe_buffer *buf = cs->currbuf;
 724
 725                 if (cs->write)
 726                         buf->len = PAGE_SIZE - cs->len;
 727                 cs->currbuf = NULL;
 728         } else if (cs->pg) {
 729                 if (cs->write) {
 730                         flush_dcache_page(cs->pg);
 731                         set_page_dirty_lock(cs->pg);
 732                 }
 733                 put_page(cs->pg);
 734         }
 735         cs->pg = NULL;
 736 }
 737
 738 /*
 739  * Get another pagefull of userspace buffer, and map it to kernel
 740  * address space, and lock request
 741  */
 742 static int fuse_copy_fill(struct fuse_copy_state *cs)
 743 {
 744         struct page *page;
 745         int err;
 746
 747         err = unlock_request(cs->req);
 748         if (err)
 749                 return err;
 750
 751         fuse_copy_finish(cs);
 752         if (cs->pipebufs) {
 753                 struct pipe_buffer *buf = cs->pipebufs;
 754
 755                 if (!cs->write) {
 756                         err = pipe_buf_confirm(cs->pipe, buf);
 757                         if (err)
 758                                 return err;
 759
 760                         BUG_ON(!cs->nr_segs);
 761                         cs->currbuf = buf;
 762                         cs->pg = buf->page;
 763                         cs->offset = buf->offset;
 764                         cs->len = buf->len;
 765                         cs->pipebufs++;
 766                         cs->nr_segs--;
 767                 } else {
 768                         if (cs->nr_segs == cs->pipe->buffers)
 769                                 return -EIO;
 770
 771                         page = alloc_page(GFP_HIGHUSER);
 772                         if (!page)
 773                                 return -ENOMEM;
 774
 775                         buf->page = page;
 776                         buf->offset = 0;
 777                         buf->len = 0;
 778
 779                         cs->currbuf = buf;
 780                         cs->pg = page;
 781                         cs->offset = 0;
 782                         cs->len = PAGE_SIZE;
 783                         cs->pipebufs++;
 784                         cs->nr_segs++;
 785                 }
 786         } else {
 787                 size_t off;
 788                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 789                 if (err < 0)
 790                         return err;
 791                 BUG_ON(!err);
 792                 cs->len = err;
 793                 cs->offset = off;
 794                 cs->pg = page;
 795                 iov_iter_advance(cs->iter, err);
 796         }
 797
 798         return lock_request(cs->req);
 799 }
 800
 801 /* Do as much copy to/from userspace buffer as we can */
 802 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 803 {
 804         unsigned ncpy = min(*size, cs->len);
 805         if (val) {
 806                 void *pgaddr = kmap_atomic(cs->pg);
 807                 void *buf = pgaddr + cs->offset;
 808
 809                 if (cs->write)
 810                         memcpy(buf, *val, ncpy);
 811                 else
 812                         memcpy(*val, buf, ncpy);
 813
 814                 kunmap_atomic(pgaddr);
 815                 *val += ncpy;
 816         }
 817         *size -= ncpy;
 818         cs->len -= ncpy;
 819         cs->offset += ncpy;
 820         return ncpy;
 821 }
 822
 823 static int fuse_check_page(struct page *page)
 824 {
 825         if (page_mapcount(page) ||
 826             page->mapping != NULL ||
 827             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 828              ~(1 << PG_locked |
 829                1 << PG_referenced |
 830                1 << PG_uptodate |
 831                1 << PG_lru |
 832                1 << PG_active |
 833                1 << PG_reclaim))) {
 834                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 835                 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);
 836                 return 1;
 837         }
 838         return 0;
 839 }
 840
 841 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 842 {
 843         int err;
 844         struct page *oldpage = *pagep;
 845         struct page *newpage;
 846         struct pipe_buffer *buf = cs->pipebufs;
 847
 848         get_page(oldpage);
 849         err = unlock_request(cs->req);
 850         if (err)
 851                 goto out_put_old;
 852
 853         fuse_copy_finish(cs);
 854
 855         err = pipe_buf_confirm(cs->pipe, buf);
 856         if (err)
 857                 goto out_put_old;
 858
 859         BUG_ON(!cs->nr_segs);
 860         cs->currbuf = buf;
 861         cs->len = buf->len;
 862         cs->pipebufs++;
 863         cs->nr_segs--;
 864
 865         if (cs->len != PAGE_SIZE)
 866                 goto out_fallback;
 867
 868         if (pipe_buf_steal(cs->pipe, buf) != 0)
 869                 goto out_fallback;
 870
 871         newpage = buf->page;
 872
 873         if (!PageUptodate(newpage))
 874                 SetPageUptodate(newpage);
 875
 876         ClearPageMappedToDisk(newpage);
 877
 878         if (fuse_check_page(newpage) != 0)
 879                 goto out_fallback_unlock;
 880
 881         /*
 882          * This is a new and locked page, it shouldn't be mapped or
 883          * have any special flags on it
 884          */
 885         if (WARN_ON(page_mapped(oldpage)))
 886                 goto out_fallback_unlock;
 887         if (WARN_ON(page_has_private(oldpage)))
 888                 goto out_fallback_unlock;
 889         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 890                 goto out_fallback_unlock;
 891         if (WARN_ON(PageMlocked(oldpage)))
 892                 goto out_fallback_unlock;
 893
 894         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 895         if (err) {
 896                 unlock_page(newpage);
 897                 goto out_put_old;
 898         }
 899
 900         get_page(newpage);
 901
 902         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 903                 lru_cache_add_file(newpage);
 904
 905         err = 0;
 906         spin_lock(&cs->req->waitq.lock);
 907         if (test_bit(FR_ABORTED, &cs->req->flags))
 908                 err = -ENOENT;
 909         else
 910                 *pagep = newpage;
 911         spin_unlock(&cs->req->waitq.lock);
 912
 913         if (err) {
 914                 unlock_page(newpage);
 915                 put_page(newpage);
 916                 goto out_put_old;
 917         }
 918
 919         unlock_page(oldpage);
 920         /* Drop ref for ap->pages[] array */
 921         put_page(oldpage);
 922         cs->len = 0;
 923
 924         err = 0;
 925 out_put_old:
 926         /* Drop ref obtained in this function */
 927         put_page(oldpage);
 928         return err;
 929
 930 out_fallback_unlock:
 931         unlock_page(newpage);
 932 out_fallback:
 933         cs->pg = buf->page;
 934         cs->offset = buf->offset;
 935
 936         err = lock_request(cs->req);
 937         if (!err)
 938                 err = 1;
 939
 940         goto out_put_old;
 941 }
 942
 943 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 944                          unsigned offset, unsigned count)
 945 {
 946         struct pipe_buffer *buf;
 947         int err;
 948
 949         if (cs->nr_segs == cs->pipe->buffers)
 950                 return -EIO;
 951
 952         get_page(page);
 953         err = unlock_request(cs->req);
 954         if (err) {
 955                 put_page(page);
 956                 return err;
 957         }
 958
 959         fuse_copy_finish(cs);
 960
 961         buf = cs->pipebufs;
 962         buf->page = page;
 963         buf->offset = offset;
 964         buf->len = count;
 965
 966         cs->pipebufs++;
 967         cs->nr_segs++;
 968         cs->len = 0;
 969
 970         return 0;
 971 }
 972
 973 /*
 974  * Copy a page in the request to/from the userspace buffer.  Must be
 975  * done atomically
 976  */
 977 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 978                           unsigned offset, unsigned count, int zeroing)
 979 {
 980         int err;
 981         struct page *page = *pagep;
 982
 983         if (page && zeroing && count < PAGE_SIZE)
 984                 clear_highpage(page);
 985
 986         while (count) {
 987                 if (cs->write && cs->pipebufs && page) {
 988                         return fuse_ref_page(cs, page, offset, count);
 989                 } else if (!cs->len) {
 990                         if (cs->move_pages && page &&
 991                             offset == 0 && count == PAGE_SIZE) {
 992                                 err = fuse_try_move_page(cs, pagep);
 993                                 if (err <= 0)
 994                                         return err;
 995                         } else {
 996                                 err = fuse_copy_fill(cs);
 997                                 if (err)
 998                                         return err;
 999                         }
1000                 }
1001                 if (page) {
1002                         void *mapaddr = kmap_atomic(page);
1003                         void *buf = mapaddr + offset;
1004                         offset += fuse_copy_do(cs, &buf, &count);
1005                         kunmap_atomic(mapaddr);
1006                 } else
1007                         offset += fuse_copy_do(cs, NULL, &count);
1008         }
1009         if (page && !cs->write)
1010                 flush_dcache_page(page);
1011         return 0;
1012 }
1013
1014 /* Copy pages in the request to/from userspace buffer */
1015 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1016                            int zeroing)
1017 {
1018         unsigned i;
1019         struct fuse_req *req = cs->req;
1020
1021         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1022                 int err;
1023                 unsigned offset = req->page_descs[i].offset;
1024                 unsigned count = min(nbytes, req->page_descs[i].length);
1025
1026                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1027                                      zeroing);
1028                 if (err)
1029                         return err;
1030
1031                 nbytes -= count;
1032         }
1033         return 0;
1034 }
1035
1036 /* Copy a single argument in the request to/from userspace buffer */
1037 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1038 {
1039         while (size) {
1040                 if (!cs->len) {
1041                         int err = fuse_copy_fill(cs);
1042                         if (err)
1043                                 return err;
1044                 }
1045                 fuse_copy_do(cs, &val, &size);
1046         }
1047         return 0;
1048 }
1049
1050 /* Copy request arguments to/from userspace buffer */
1051 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1052                           unsigned argpages, struct fuse_arg *args,
1053                           int zeroing)
1054 {
1055         int err = 0;
1056         unsigned i;
1057
1058         for (i = 0; !err && i < numargs; i++)  {
1059                 struct fuse_arg *arg = &args[i];
1060                 if (i == numargs - 1 && argpages)
1061                         err = fuse_copy_pages(cs, arg->size, zeroing);
1062                 else
1063                         err = fuse_copy_one(cs, arg->value, arg->size);
1064         }
1065         return err;
1066 }
1067
1068 static int forget_pending(struct fuse_iqueue *fiq)
1069 {
1070         return fiq->forget_list_head.next != NULL;
1071 }
1072
1073 static int request_pending(struct fuse_iqueue *fiq)
1074 {
1075         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1076                 forget_pending(fiq);
1077 }
1078
1079 /*
1080  * Transfer an interrupt request to userspace
1081  *
1082  * Unlike other requests this is assembled on demand, without a need
1083  * to allocate a separate fuse_req structure.
1084  *
1085  * Called with fiq->waitq.lock held, releases it
1086  */
1087 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1088                                struct fuse_copy_state *cs,
1089                                size_t nbytes, struct fuse_req *req)
1090 __releases(fiq->waitq.lock)
1091 {
1092         struct fuse_in_header ih;
1093         struct fuse_interrupt_in arg;
1094         unsigned reqsize = sizeof(ih) + sizeof(arg);
1095         int err;
1096
1097         list_del_init(&req->intr_entry);
1098         req->intr_unique = fuse_get_unique(fiq);
1099         memset(&ih, 0, sizeof(ih));
1100         memset(&arg, 0, sizeof(arg));
1101         ih.len = reqsize;
1102         ih.opcode = FUSE_INTERRUPT;
1103         ih.unique = req->intr_unique;
1104         arg.unique = req->in.h.unique;
1105
1106         spin_unlock(&fiq->waitq.lock);
1107         if (nbytes < reqsize)
1108                 return -EINVAL;
1109
1110         err = fuse_copy_one(cs, &ih, sizeof(ih));
1111         if (!err)
1112                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1113         fuse_copy_finish(cs);
1114
1115         return err ? err : reqsize;
1116 }
1117
1118 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1119                                                unsigned max,
1120                                                unsigned *countp)
1121 {
1122         struct fuse_forget_link *head = fiq->forget_list_head.next;
1123         struct fuse_forget_link **newhead = &head;
1124         unsigned count;
1125
1126         for (count = 0; *newhead != NULL && count < max; count++)
1127                 newhead = &(*newhead)->next;
1128
1129         fiq->forget_list_head.next = *newhead;
1130         *newhead = NULL;
1131         if (fiq->forget_list_head.next == NULL)
1132                 fiq->forget_list_tail = &fiq->forget_list_head;
1133
1134         if (countp != NULL)
1135                 *countp = count;
1136
1137         return head;
1138 }
1139
1140 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1141                                    struct fuse_copy_state *cs,
1142                                    size_t nbytes)
1143 __releases(fiq->waitq.lock)
1144 {
1145         int err;
1146         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1147         struct fuse_forget_in arg = {
1148                 .nlookup = forget->forget_one.nlookup,
1149         };
1150         struct fuse_in_header ih = {
1151                 .opcode = FUSE_FORGET,
1152                 .nodeid = forget->forget_one.nodeid,
1153                 .unique = fuse_get_unique(fiq),
1154                 .len = sizeof(ih) + sizeof(arg),
1155         };
1156
1157         spin_unlock(&fiq->waitq.lock);
1158         kfree(forget);
1159         if (nbytes < ih.len)
1160                 return -EINVAL;
1161
1162         err = fuse_copy_one(cs, &ih, sizeof(ih));
1163         if (!err)
1164                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1165         fuse_copy_finish(cs);
1166
1167         if (err)
1168                 return err;
1169
1170         return ih.len;
1171 }
1172
1173 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1174                                    struct fuse_copy_state *cs, size_t nbytes)
1175 __releases(fiq->waitq.lock)
1176 {
1177         int err;
1178         unsigned max_forgets;
1179         unsigned count;
1180         struct fuse_forget_link *head;
1181         struct fuse_batch_forget_in arg = { .count = 0 };
1182         struct fuse_in_header ih = {
1183                 .opcode = FUSE_BATCH_FORGET,
1184                 .unique = fuse_get_unique(fiq),
1185                 .len = sizeof(ih) + sizeof(arg),
1186         };
1187
1188         if (nbytes < ih.len) {
1189                 spin_unlock(&fiq->waitq.lock);
1190                 return -EINVAL;
1191         }
1192
1193         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1194         head = dequeue_forget(fiq, max_forgets, &count);
1195         spin_unlock(&fiq->waitq.lock);
1196
1197         arg.count = count;
1198         ih.len += count * sizeof(struct fuse_forget_one);
1199         err = fuse_copy_one(cs, &ih, sizeof(ih));
1200         if (!err)
1201                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1202
1203         while (head) {
1204                 struct fuse_forget_link *forget = head;
1205
1206                 if (!err) {
1207                         err = fuse_copy_one(cs, &forget->forget_one,
1208                                             sizeof(forget->forget_one));
1209                 }
1210                 head = forget->next;
1211                 kfree(forget);
1212         }
1213
1214         fuse_copy_finish(cs);
1215
1216         if (err)
1217                 return err;
1218
1219         return ih.len;
1220 }
1221
1222 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1223                             struct fuse_copy_state *cs,
1224                             size_t nbytes)
1225 __releases(fiq->waitq.lock)
1226 {
1227         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1228                 return fuse_read_single_forget(fiq, cs, nbytes);
1229         else
1230                 return fuse_read_batch_forget(fiq, cs, nbytes);
1231 }
1232
1233 /*
1234  * Read a single request into the userspace filesystem's buffer.  This
1235  * function waits until a request is available, then removes it from
1236  * the pending list and copies request data to userspace buffer.  If
1237  * no reply is needed (FORGET) or request has been aborted or there
1238  * was an error during the copying then it's finished by calling
1239  * request_end().  Otherwise add it to the processing list, and set
1240  * the 'sent' flag.
1241  */
1242 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1243                                 struct fuse_copy_state *cs, size_t nbytes)
1244 {
1245         ssize_t err;
1246         struct fuse_conn *fc = fud->fc;
1247         struct fuse_iqueue *fiq = &fc->iq;
1248         struct fuse_pqueue *fpq = &fud->pq;
1249         struct fuse_req *req;
1250         struct fuse_in *in;
1251         unsigned reqsize;
1252
1253  restart:
1254         spin_lock(&fiq->waitq.lock);
1255         err = -EAGAIN;
1256         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1257             !request_pending(fiq))
1258                 goto err_unlock;
1259
1260         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1261                                 !fiq->connected || request_pending(fiq));
1262         if (err)
1263                 goto err_unlock;
1264
1265         err = -ENODEV;
1266         if (!fiq->connected)
1267                 goto err_unlock;
1268
1269         if (!list_empty(&fiq->interrupts)) {
1270                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1271                                  intr_entry);
1272                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1273         }
1274
1275         if (forget_pending(fiq)) {
1276                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1277                         return fuse_read_forget(fc, fiq, cs, nbytes);
1278
1279                 if (fiq->forget_batch <= -8)
1280                         fiq->forget_batch = 16;
1281         }
1282
1283         req = list_entry(fiq->pending.next, struct fuse_req, list);
1284         clear_bit(FR_PENDING, &req->flags);
1285         list_del_init(&req->list);
1286         spin_unlock(&fiq->waitq.lock);
1287
1288         in = &req->in;
1289         reqsize = in->h.len;
1290
1291         if (task_active_pid_ns(current) != fc->pid_ns) {
1292                 rcu_read_lock();
1293                 in->h.pid = pid_vnr(find_pid_ns(in->h.pid, fc->pid_ns));
1294                 rcu_read_unlock();
1295         }
1296
1297         /* If request is too large, reply with an error and restart the read */
1298         if (nbytes < reqsize) {
1299                 req->out.h.error = -EIO;
1300                 /* SETXATTR is special, since it may contain too large data */
1301                 if (in->h.opcode == FUSE_SETXATTR)
1302                         req->out.h.error = -E2BIG;
1303                 request_end(fc, req);
1304                 goto restart;
1305         }
1306         spin_lock(&fpq->lock);
1307         /*
1308          *  Must not put request on fpq->io queue after having been shut down by
1309          *  fuse_abort_conn()
1310          */
1311         if (!fpq->connected) {
1312                 req->out.h.error = err = -ECONNABORTED;
1313                 goto out_end;
1314
1315         }
1316         list_add(&req->list, &fpq->io);
1317         spin_unlock(&fpq->lock);
1318         cs->req = req;
1319         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1320         if (!err)
1321                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1322                                      (struct fuse_arg *) in->args, 0);
1323         fuse_copy_finish(cs);
1324         spin_lock(&fpq->lock);
1325         clear_bit(FR_LOCKED, &req->flags);
1326         if (!fpq->connected) {
1327                 err = -ENODEV;
1328                 goto out_end;
1329         }
1330         if (err) {
1331                 req->out.h.error = -EIO;
1332                 goto out_end;
1333         }
1334         if (!test_bit(FR_ISREPLY, &req->flags)) {
1335                 err = reqsize;
1336                 goto out_end;
1337         }
1338         list_move_tail(&req->list, &fpq->processing);
1339         __fuse_get_request(req);
1340         set_bit(FR_SENT, &req->flags);
1341         spin_unlock(&fpq->lock);
1342         /* matches barrier in request_wait_answer() */
1343         smp_mb__after_atomic();
1344         if (test_bit(FR_INTERRUPTED, &req->flags))
1345                 queue_interrupt(fiq, req);
1346         fuse_put_request(fc, req);
1347
1348         return reqsize;
1349
1350 out_end:
1351         if (!test_bit(FR_PRIVATE, &req->flags))
1352                 list_del_init(&req->list);
1353         spin_unlock(&fpq->lock);
1354         request_end(fc, req);
1355         return err;
1356
1357  err_unlock:
1358         spin_unlock(&fiq->waitq.lock);
1359         return err;
1360 }
1361
1362 static int fuse_dev_open(struct inode *inode, struct file *file)
1363 {
1364         /*
1365          * The fuse device's file's private_data is used to hold
1366          * the fuse_conn(ection) when it is mounted, and is used to
1367          * keep track of whether the file has been mounted already.
1368          */
1369         file->private_data = NULL;
1370         return 0;
1371 }
1372
1373 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1374 {
1375         struct fuse_copy_state cs;
1376         struct file *file = iocb->ki_filp;
1377         struct fuse_dev *fud = fuse_get_dev(file);
1378
1379         if (!fud)
1380                 return -EPERM;
1381
1382         if (!iter_is_iovec(to))
1383                 return -EINVAL;
1384
1385         fuse_copy_init(&cs, 1, to);
1386
1387         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1388 }
1389
1390 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1391                                     struct pipe_inode_info *pipe,
1392                                     size_t len, unsigned int flags)
1393 {
1394         int total, ret;
1395         int page_nr = 0;
1396         struct pipe_buffer *bufs;
1397         struct fuse_copy_state cs;
1398         struct fuse_dev *fud = fuse_get_dev(in);
1399
1400         if (!fud)
1401                 return -EPERM;
1402
1403         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1404         if (!bufs)
1405                 return -ENOMEM;
1406
1407         fuse_copy_init(&cs, 1, NULL);
1408         cs.pipebufs = bufs;
1409         cs.pipe = pipe;
1410         ret = fuse_dev_do_read(fud, in, &cs, len);
1411         if (ret < 0)
1412                 goto out;
1413
1414         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1415                 ret = -EIO;
1416                 goto out;
1417         }
1418
1419         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1420                 /*
1421                  * Need to be careful about this.  Having buf->ops in module
1422                  * code can Oops if the buffer persists after module unload.
1423                  */
1424                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1425                 bufs[page_nr].flags = 0;
1426                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1427                 if (unlikely(ret < 0))
1428                         break;
1429         }
1430         if (total)
1431                 ret = total;
1432 out:
1433         for (; page_nr < cs.nr_segs; page_nr++)
1434                 put_page(bufs[page_nr].page);
1435
1436         kfree(bufs);
1437         return ret;
1438 }
1439
1440 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1441                             struct fuse_copy_state *cs)
1442 {
1443         struct fuse_notify_poll_wakeup_out outarg;
1444         int err = -EINVAL;
1445
1446         if (size != sizeof(outarg))
1447                 goto err;
1448
1449         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1450         if (err)
1451                 goto err;
1452
1453         fuse_copy_finish(cs);
1454         return fuse_notify_poll_wakeup(fc, &outarg);
1455
1456 err:
1457         fuse_copy_finish(cs);
1458         return err;
1459 }
1460
1461 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1462                                    struct fuse_copy_state *cs)
1463 {
1464         struct fuse_notify_inval_inode_out outarg;
1465         int err = -EINVAL;
1466
1467         if (size != sizeof(outarg))
1468                 goto err;
1469
1470         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1471         if (err)
1472                 goto err;
1473         fuse_copy_finish(cs);
1474
1475         down_read(&fc->killsb);
1476         err = -ENOENT;
1477         if (fc->sb) {
1478                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1479                                                outarg.off, outarg.len);
1480         }
1481         up_read(&fc->killsb);
1482         return err;
1483
1484 err:
1485         fuse_copy_finish(cs);
1486         return err;
1487 }
1488
1489 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1490                                    struct fuse_copy_state *cs)
1491 {
1492         struct fuse_notify_inval_entry_out outarg;
1493         int err = -ENOMEM;
1494         char *buf;
1495         struct qstr name;
1496
1497         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1498         if (!buf)
1499                 goto err;
1500
1501         err = -EINVAL;
1502         if (size < sizeof(outarg))
1503                 goto err;
1504
1505         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1506         if (err)
1507                 goto err;
1508
1509         err = -ENAMETOOLONG;
1510         if (outarg.namelen > FUSE_NAME_MAX)
1511                 goto err;
1512
1513         err = -EINVAL;
1514         if (size != sizeof(outarg) + outarg.namelen + 1)
1515                 goto err;
1516
1517         name.name = buf;
1518         name.len = outarg.namelen;
1519         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1520         if (err)
1521                 goto err;
1522         fuse_copy_finish(cs);
1523         buf[outarg.namelen] = 0;
1524
1525         down_read(&fc->killsb);
1526         err = -ENOENT;
1527         if (fc->sb)
1528                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1529         up_read(&fc->killsb);
1530         kfree(buf);
1531         return err;
1532
1533 err:
1534         kfree(buf);
1535         fuse_copy_finish(cs);
1536         return err;
1537 }
1538
1539 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1540                               struct fuse_copy_state *cs)
1541 {
1542         struct fuse_notify_delete_out outarg;
1543         int err = -ENOMEM;
1544         char *buf;
1545         struct qstr name;
1546
1547         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1548         if (!buf)
1549                 goto err;
1550
1551         err = -EINVAL;
1552         if (size < sizeof(outarg))
1553                 goto err;
1554
1555         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1556         if (err)
1557                 goto err;
1558
1559         err = -ENAMETOOLONG;
1560         if (outarg.namelen > FUSE_NAME_MAX)
1561                 goto err;
1562
1563         err = -EINVAL;
1564         if (size != sizeof(outarg) + outarg.namelen + 1)
1565                 goto err;
1566
1567         name.name = buf;
1568         name.len = outarg.namelen;
1569         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1570         if (err)
1571                 goto err;
1572         fuse_copy_finish(cs);
1573         buf[outarg.namelen] = 0;
1574
1575         down_read(&fc->killsb);
1576         err = -ENOENT;
1577         if (fc->sb)
1578                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1579                                                outarg.child, &name);
1580         up_read(&fc->killsb);
1581         kfree(buf);
1582         return err;
1583
1584 err:
1585         kfree(buf);
1586         fuse_copy_finish(cs);
1587         return err;
1588 }
1589
1590 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1591                              struct fuse_copy_state *cs)
1592 {
1593         struct fuse_notify_store_out outarg;
1594         struct inode *inode;
1595         struct address_space *mapping;
1596         u64 nodeid;
1597         int err;
1598         pgoff_t index;
1599         unsigned int offset;
1600         unsigned int num;
1601         loff_t file_size;
1602         loff_t end;
1603
1604         err = -EINVAL;
1605         if (size < sizeof(outarg))
1606                 goto out_finish;
1607
1608         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1609         if (err)
1610                 goto out_finish;
1611
1612         err = -EINVAL;
1613         if (size - sizeof(outarg) != outarg.size)
1614                 goto out_finish;
1615
1616         nodeid = outarg.nodeid;
1617
1618         down_read(&fc->killsb);
1619
1620         err = -ENOENT;
1621         if (!fc->sb)
1622                 goto out_up_killsb;
1623
1624         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1625         if (!inode)
1626                 goto out_up_killsb;
1627
1628         mapping = inode->i_mapping;
1629         index = outarg.offset >> PAGE_SHIFT;
1630         offset = outarg.offset & ~PAGE_MASK;
1631         file_size = i_size_read(inode);
1632         end = outarg.offset + outarg.size;
1633         if (end > file_size) {
1634                 file_size = end;
1635                 fuse_write_update_size(inode, file_size);
1636         }
1637
1638         num = outarg.size;
1639         while (num) {
1640                 struct page *page;
1641                 unsigned int this_num;
1642
1643                 err = -ENOMEM;
1644                 page = find_or_create_page(mapping, index,
1645                                            mapping_gfp_mask(mapping));
1646                 if (!page)
1647                         goto out_iput;
1648
1649                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1650                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1651                 if (!err && offset == 0 &&
1652                     (this_num == PAGE_SIZE || file_size == end))
1653                         SetPageUptodate(page);
1654                 unlock_page(page);
1655                 put_page(page);
1656
1657                 if (err)
1658                         goto out_iput;
1659
1660                 num -= this_num;
1661                 offset = 0;
1662                 index++;
1663         }
1664
1665         err = 0;
1666
1667 out_iput:
1668         iput(inode);
1669 out_up_killsb:
1670         up_read(&fc->killsb);
1671 out_finish:
1672         fuse_copy_finish(cs);
1673         return err;
1674 }
1675
1676 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1677 {
1678         release_pages(req->pages, req->num_pages, false);
1679 }
1680
1681 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1682                          struct fuse_notify_retrieve_out *outarg)
1683 {
1684         int err;
1685         struct address_space *mapping = inode->i_mapping;
1686         struct fuse_req *req;
1687         pgoff_t index;
1688         loff_t file_size;
1689         unsigned int num;
1690         unsigned int offset;
1691         size_t total_len = 0;
1692         int num_pages;
1693
1694         offset = outarg->offset & ~PAGE_MASK;
1695         file_size = i_size_read(inode);
1696
1697         num = min(outarg->size, fc->max_write);
1698         if (outarg->offset > file_size)
1699                 num = 0;
1700         else if (outarg->offset + num > file_size)
1701                 num = file_size - outarg->offset;
1702
1703         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1704         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1705
1706         req = fuse_get_req(fc, num_pages);
1707         if (IS_ERR(req))
1708                 return PTR_ERR(req);
1709
1710         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1711         req->in.h.nodeid = outarg->nodeid;
1712         req->in.numargs = 2;
1713         req->in.argpages = 1;
1714         req->end = fuse_retrieve_end;
1715
1716         index = outarg->offset >> PAGE_SHIFT;
1717
1718         while (num && req->num_pages < num_pages) {
1719                 struct page *page;
1720                 unsigned int this_num;
1721
1722                 page = find_get_page(mapping, index);
1723                 if (!page)
1724                         break;
1725
1726                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1727                 req->pages[req->num_pages] = page;
1728                 req->page_descs[req->num_pages].offset = offset;
1729                 req->page_descs[req->num_pages].length = this_num;
1730                 req->num_pages++;
1731
1732                 offset = 0;
1733                 num -= this_num;
1734                 total_len += this_num;
1735                 index++;
1736         }
1737         req->misc.retrieve_in.offset = outarg->offset;
1738         req->misc.retrieve_in.size = total_len;
1739         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1740         req->in.args[0].value = &req->misc.retrieve_in;
1741         req->in.args[1].size = total_len;
1742
1743         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1744         if (err) {
1745                 fuse_retrieve_end(fc, req);
1746                 fuse_put_request(fc, req);
1747         }
1748
1749         return err;
1750 }
1751
1752 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1753                                 struct fuse_copy_state *cs)
1754 {
1755         struct fuse_notify_retrieve_out outarg;
1756         struct inode *inode;
1757         int err;
1758
1759         err = -EINVAL;
1760         if (size != sizeof(outarg))
1761                 goto copy_finish;
1762
1763         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1764         if (err)
1765                 goto copy_finish;
1766
1767         fuse_copy_finish(cs);
1768
1769         down_read(&fc->killsb);
1770         err = -ENOENT;
1771         if (fc->sb) {
1772                 u64 nodeid = outarg.nodeid;
1773
1774                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1775                 if (inode) {
1776                         err = fuse_retrieve(fc, inode, &outarg);
1777                         iput(inode);
1778                 }
1779         }
1780         up_read(&fc->killsb);
1781
1782         return err;
1783
1784 copy_finish:
1785         fuse_copy_finish(cs);
1786         return err;
1787 }
1788
1789 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1790                        unsigned int size, struct fuse_copy_state *cs)
1791 {
1792         /* Don't try to move pages (yet) */
1793         cs->move_pages = 0;
1794
1795         switch (code) {
1796         case FUSE_NOTIFY_POLL:
1797                 return fuse_notify_poll(fc, size, cs);
1798
1799         case FUSE_NOTIFY_INVAL_INODE:
1800                 return fuse_notify_inval_inode(fc, size, cs);
1801
1802         case FUSE_NOTIFY_INVAL_ENTRY:
1803                 return fuse_notify_inval_entry(fc, size, cs);
1804
1805         case FUSE_NOTIFY_STORE:
1806                 return fuse_notify_store(fc, size, cs);
1807
1808         case FUSE_NOTIFY_RETRIEVE:
1809                 return fuse_notify_retrieve(fc, size, cs);
1810
1811         case FUSE_NOTIFY_DELETE:
1812                 return fuse_notify_delete(fc, size, cs);
1813
1814         default:
1815                 fuse_copy_finish(cs);
1816                 return -EINVAL;
1817         }
1818 }
1819
1820 /* Look up request on processing list by unique ID */
1821 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1822 {
1823         struct fuse_req *req;
1824
1825         list_for_each_entry(req, &fpq->processing, list) {
1826                 if (req->in.h.unique == unique || req->intr_unique == unique)
1827                         return req;
1828         }
1829         return NULL;
1830 }
1831
1832 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1833                          unsigned nbytes)
1834 {
1835         unsigned reqsize = sizeof(struct fuse_out_header);
1836
1837         if (out->h.error)
1838                 return nbytes != reqsize ? -EINVAL : 0;
1839
1840         reqsize += len_args(out->numargs, out->args);
1841
1842         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1843                 return -EINVAL;
1844         else if (reqsize > nbytes) {
1845                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1846                 unsigned diffsize = reqsize - nbytes;
1847                 if (diffsize > lastarg->size)
1848                         return -EINVAL;
1849                 lastarg->size -= diffsize;
1850         }
1851         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1852                               out->page_zeroing);
1853 }
1854
1855 /*
1856  * Write a single reply to a request.  First the header is copied from
1857  * the write buffer.  The request is then searched on the processing
1858  * list by the unique ID found in the header.  If found, then remove
1859  * it from the list and copy the rest of the buffer to the request.
1860  * The request is finished by calling request_end()
1861  */
1862 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1863                                  struct fuse_copy_state *cs, size_t nbytes)
1864 {
1865         int err;
1866         struct fuse_conn *fc = fud->fc;
1867         struct fuse_pqueue *fpq = &fud->pq;
1868         struct fuse_req *req;
1869         struct fuse_out_header oh;
1870
1871         if (nbytes < sizeof(struct fuse_out_header))
1872                 return -EINVAL;
1873
1874         err = fuse_copy_one(cs, &oh, sizeof(oh));
1875         if (err)
1876                 goto err_finish;
1877
1878         err = -EINVAL;
1879         if (oh.len != nbytes)
1880                 goto err_finish;
1881
1882         /*
1883          * Zero oh.unique indicates unsolicited notification message
1884          * and error contains notification code.
1885          */
1886         if (!oh.unique) {
1887                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1888                 return err ? err : nbytes;
1889         }
1890
1891         err = -EINVAL;
1892         if (oh.error <= -512 || oh.error > 0)
1893                 goto err_finish;
1894
1895         spin_lock(&fpq->lock);
1896         err = -ENOENT;
1897         if (!fpq->connected)
1898                 goto err_unlock_pq;
1899
1900         req = request_find(fpq, oh.unique);
1901         if (!req)
1902                 goto err_unlock_pq;
1903
1904         /* Is it an interrupt reply? */
1905         if (req->intr_unique == oh.unique) {
1906                 __fuse_get_request(req);
1907                 spin_unlock(&fpq->lock);
1908
1909                 err = -EINVAL;
1910                 if (nbytes != sizeof(struct fuse_out_header)) {
1911                         fuse_put_request(fc, req);
1912                         goto err_finish;
1913                 }
1914
1915                 if (oh.error == -ENOSYS)
1916                         fc->no_interrupt = 1;
1917                 else if (oh.error == -EAGAIN)
1918                         queue_interrupt(&fc->iq, req);
1919                 fuse_put_request(fc, req);
1920
1921                 fuse_copy_finish(cs);
1922                 return nbytes;
1923         }
1924
1925         clear_bit(FR_SENT, &req->flags);
1926         list_move(&req->list, &fpq->io);
1927         req->out.h = oh;
1928         set_bit(FR_LOCKED, &req->flags);
1929         spin_unlock(&fpq->lock);
1930         cs->req = req;
1931         if (!req->out.page_replace)
1932                 cs->move_pages = 0;
1933
1934         err = copy_out_args(cs, &req->out, nbytes);
1935         fuse_copy_finish(cs);
1936
1937         spin_lock(&fpq->lock);
1938         clear_bit(FR_LOCKED, &req->flags);
1939         if (!fpq->connected)
1940                 err = -ENOENT;
1941         else if (err)
1942                 req->out.h.error = -EIO;
1943         if (!test_bit(FR_PRIVATE, &req->flags))
1944                 list_del_init(&req->list);
1945         spin_unlock(&fpq->lock);
1946
1947         request_end(fc, req);
1948
1949         return err ? err : nbytes;
1950
1951  err_unlock_pq:
1952         spin_unlock(&fpq->lock);
1953  err_finish:
1954         fuse_copy_finish(cs);
1955         return err;
1956 }
1957
1958 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1959 {
1960         struct fuse_copy_state cs;
1961         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1962
1963         if (!fud)
1964                 return -EPERM;
1965
1966         if (!iter_is_iovec(from))
1967                 return -EINVAL;
1968
1969         fuse_copy_init(&cs, 0, from);
1970
1971         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1972 }
1973
1974 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1975                                      struct file *out, loff_t *ppos,
1976                                      size_t len, unsigned int flags)
1977 {
1978         unsigned nbuf;
1979         unsigned idx;
1980         struct pipe_buffer *bufs;
1981         struct fuse_copy_state cs;
1982         struct fuse_dev *fud;
1983         size_t rem;
1984         ssize_t ret;
1985
1986         fud = fuse_get_dev(out);
1987         if (!fud)
1988                 return -EPERM;
1989
1990         pipe_lock(pipe);
1991
1992         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1993         if (!bufs) {
1994                 pipe_unlock(pipe);
1995                 return -ENOMEM;
1996         }
1997
1998         nbuf = 0;
1999         rem = 0;
2000         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2001                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2002
2003         ret = -EINVAL;
2004         if (rem < len)
2005                 goto out_free;
2006
2007         rem = len;
2008         while (rem) {
2009                 struct pipe_buffer *ibuf;
2010                 struct pipe_buffer *obuf;
2011
2012                 BUG_ON(nbuf >= pipe->buffers);
2013                 BUG_ON(!pipe->nrbufs);
2014                 ibuf = &pipe->bufs[pipe->curbuf];
2015                 obuf = &bufs[nbuf];
2016
2017                 if (rem >= ibuf->len) {
2018                         *obuf = *ibuf;
2019                         ibuf->ops = NULL;
2020                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2021                         pipe->nrbufs--;
2022                 } else {
2023                         if (!pipe_buf_get(pipe, ibuf))
2024                                 goto out_free;
2025
2026                         *obuf = *ibuf;
2027                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2028                         obuf->len = rem;
2029                         ibuf->offset += obuf->len;
2030                         ibuf->len -= obuf->len;
2031                 }
2032                 nbuf++;
2033                 rem -= obuf->len;
2034         }
2035         pipe_unlock(pipe);
2036
2037         fuse_copy_init(&cs, 0, NULL);
2038         cs.pipebufs = bufs;
2039         cs.nr_segs = nbuf;
2040         cs.pipe = pipe;
2041
2042         if (flags & SPLICE_F_MOVE)
2043                 cs.move_pages = 1;
2044
2045         ret = fuse_dev_do_write(fud, &cs, len);
2046
2047         pipe_lock(pipe);
2048 out_free:
2049         for (idx = 0; idx < nbuf; idx++)
2050                 pipe_buf_release(pipe, &bufs[idx]);
2051         pipe_unlock(pipe);
2052
2053         kfree(bufs);
2054         return ret;
2055 }
2056
2057 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2058 {
2059         unsigned mask = POLLOUT | POLLWRNORM;
2060         struct fuse_iqueue *fiq;
2061         struct fuse_dev *fud = fuse_get_dev(file);
2062
2063         if (!fud)
2064                 return POLLERR;
2065
2066         fiq = &fud->fc->iq;
2067         poll_wait(file, &fiq->waitq, wait);
2068
2069         spin_lock(&fiq->waitq.lock);
2070         if (!fiq->connected)
2071                 mask = POLLERR;
2072         else if (request_pending(fiq))
2073                 mask |= POLLIN | POLLRDNORM;
2074         spin_unlock(&fiq->waitq.lock);
2075
2076         return mask;
2077 }
2078
2079 /*
2080  * Abort all requests on the given list (pending or processing)
2081  *
2082  * This function releases and reacquires fc->lock
2083  */
2084 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2085 {
2086         while (!list_empty(head)) {
2087                 struct fuse_req *req;
2088                 req = list_entry(head->next, struct fuse_req, list);
2089                 req->out.h.error = -ECONNABORTED;
2090                 clear_bit(FR_SENT, &req->flags);
2091                 list_del_init(&req->list);
2092                 request_end(fc, req);
2093         }
2094 }
2095
2096 static void end_polls(struct fuse_conn *fc)
2097 {
2098         struct rb_node *p;
2099
2100         p = rb_first(&fc->polled_files);
2101
2102         while (p) {
2103                 struct fuse_file *ff;
2104                 ff = rb_entry(p, struct fuse_file, polled_node);
2105                 wake_up_interruptible_all(&ff->poll_wait);
2106
2107                 p = rb_next(p);
2108         }
2109 }
2110
2111 /*
2112  * Abort all requests.
2113  *
2114  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2115  * filesystem.
2116  *
2117  * The same effect is usually achievable through killing the filesystem daemon
2118  * and all users of the filesystem.  The exception is the combination of an
2119  * asynchronous request and the tricky deadlock (see
2120  * Documentation/filesystems/fuse.txt).
2121  *
2122  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2123  * requests, they should be finished off immediately.  Locked requests will be
2124  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2125  * requests.  It is possible that some request will finish before we can.  This
2126  * is OK, the request will in that case be removed from the list before we touch
2127  * it.
2128  */
2129 void fuse_abort_conn(struct fuse_conn *fc)
2130 {
2131         struct fuse_iqueue *fiq = &fc->iq;
2132
2133         spin_lock(&fc->lock);
2134         if (fc->connected) {
2135                 struct fuse_dev *fud;
2136                 struct fuse_req *req, *next;
2137                 LIST_HEAD(to_end1);
2138                 LIST_HEAD(to_end2);
2139
2140                 fc->connected = 0;
2141                 fc->blocked = 0;
2142                 fuse_set_initialized(fc);
2143                 list_for_each_entry(fud, &fc->devices, entry) {
2144                         struct fuse_pqueue *fpq = &fud->pq;
2145
2146                         spin_lock(&fpq->lock);
2147                         fpq->connected = 0;
2148                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2149                                 req->out.h.error = -ECONNABORTED;
2150                                 spin_lock(&req->waitq.lock);
2151                                 set_bit(FR_ABORTED, &req->flags);
2152                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2153                                         set_bit(FR_PRIVATE, &req->flags);
2154                                         __fuse_get_request(req);
2155                                         list_move(&req->list, &to_end1);
2156                                 }
2157                                 spin_unlock(&req->waitq.lock);
2158                         }
2159                         list_splice_init(&fpq->processing, &to_end2);
2160                         spin_unlock(&fpq->lock);
2161                 }
2162                 fc->max_background = UINT_MAX;
2163                 flush_bg_queue(fc);
2164
2165                 spin_lock(&fiq->waitq.lock);
2166                 fiq->connected = 0;
2167                 list_splice_init(&fiq->pending, &to_end2);
2168                 list_for_each_entry(req, &to_end2, list)
2169                         clear_bit(FR_PENDING, &req->flags);
2170                 while (forget_pending(fiq))
2171                         kfree(dequeue_forget(fiq, 1, NULL));
2172                 wake_up_all_locked(&fiq->waitq);
2173                 spin_unlock(&fiq->waitq.lock);
2174                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2175                 end_polls(fc);
2176                 wake_up_all(&fc->blocked_waitq);
2177                 spin_unlock(&fc->lock);
2178
2179                 while (!list_empty(&to_end1)) {
2180                         req = list_first_entry(&to_end1, struct fuse_req, list);
2181                         list_del_init(&req->list);
2182                         request_end(fc, req);
2183                 }
2184                 end_requests(fc, &to_end2);
2185         } else {
2186                 spin_unlock(&fc->lock);
2187         }
2188 }
2189 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2190
2191 void fuse_wait_aborted(struct fuse_conn *fc)
2192 {
2193         /* matches implicit memory barrier in fuse_drop_waiting() */
2194         smp_mb();
2195         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2196 }
2197
2198 int fuse_dev_release(struct inode *inode, struct file *file)
2199 {
2200         struct fuse_dev *fud = fuse_get_dev(file);
2201
2202         if (fud) {
2203                 struct fuse_conn *fc = fud->fc;
2204                 struct fuse_pqueue *fpq = &fud->pq;
2205                 LIST_HEAD(to_end);
2206
2207                 spin_lock(&fpq->lock);
2208                 WARN_ON(!list_empty(&fpq->io));
2209                 list_splice_init(&fpq->processing, &to_end);
2210                 spin_unlock(&fpq->lock);
2211
2212                 end_requests(fc, &to_end);
2213
2214                 /* Are we the last open device? */
2215                 if (atomic_dec_and_test(&fc->dev_count)) {
2216                         WARN_ON(fc->iq.fasync != NULL);
2217                         fuse_abort_conn(fc);
2218                 }
2219                 fuse_dev_free(fud);
2220         }
2221         return 0;
2222 }
2223 EXPORT_SYMBOL_GPL(fuse_dev_release);
2224
2225 static int fuse_dev_fasync(int fd, struct file *file, int on)
2226 {
2227         struct fuse_dev *fud = fuse_get_dev(file);
2228
2229         if (!fud)
2230                 return -EPERM;
2231
2232         /* No locking - fasync_helper does its own locking */
2233         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2234 }
2235
2236 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2237 {
2238         struct fuse_dev *fud;
2239
2240         if (new->private_data)
2241                 return -EINVAL;
2242
2243         fud = fuse_dev_alloc(fc);
2244         if (!fud)
2245                 return -ENOMEM;
2246
2247         new->private_data = fud;
2248         atomic_inc(&fc->dev_count);
2249
2250         return 0;
2251 }
2252
2253 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2254                            unsigned long arg)
2255 {
2256         int err = -ENOTTY;
2257
2258         if (cmd == FUSE_DEV_IOC_CLONE) {
2259                 int oldfd;
2260
2261                 err = -EFAULT;
2262                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2263                         struct file *old = fget(oldfd);
2264
2265                         err = -EINVAL;
2266                         if (old) {
2267                                 struct fuse_dev *fud = NULL;
2268
2269                                 /*
2270                                  * Check against file->f_op because CUSE
2271                                  * uses the same ioctl handler.
2272                                  */
2273                                 if (old->f_op == file->f_op &&
2274                                     old->f_cred->user_ns == file->f_cred->user_ns)
2275                                         fud = fuse_get_dev(old);
2276
2277                                 if (fud) {
2278                                         mutex_lock(&fuse_mutex);
2279                                         err = fuse_device_clone(fud->fc, file);
2280                                         mutex_unlock(&fuse_mutex);
2281                                 }
2282                                 fput(old);
2283                         }
2284                 }
2285         }
2286         return err;
2287 }
2288
2289 const struct file_operations fuse_dev_operations = {
2290         .owner          = THIS_MODULE,
2291         .open           = fuse_dev_open,
2292         .llseek         = no_llseek,
2293         .read_iter      = fuse_dev_read,
2294         .splice_read    = fuse_dev_splice_read,
2295         .write_iter     = fuse_dev_write,
2296         .splice_write   = fuse_dev_splice_write,
2297         .poll           = fuse_dev_poll,
2298         .release        = fuse_dev_release,
2299         .fasync         = fuse_dev_fasync,
2300         .unlocked_ioctl = fuse_dev_ioctl,
2301         .compat_ioctl   = fuse_dev_ioctl,
2302 };
2303 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2304
2305 static struct miscdevice fuse_miscdevice = {
2306         .minor = FUSE_MINOR,
2307         .name  = "fuse",
2308         .fops = &fuse_dev_operations,
2309 };
2310
2311 int __init fuse_dev_init(void)
2312 {
2313         int err = -ENOMEM;
2314         fuse_req_cachep = kmem_cache_create("fuse_request",
2315                                             sizeof(struct fuse_req),
2316                                             0, 0, NULL);
2317         if (!fuse_req_cachep)
2318                 goto out;
2319
2320         err = misc_register(&fuse_miscdevice);
2321         if (err)
2322                 goto out_cache_clean;
2323
2324         return 0;
2325
2326  out_cache_clean:
2327         kmem_cache_destroy(fuse_req_cachep);
2328  out:
2329         return err;
2330 }
2331
2332 void fuse_dev_cleanup(void)
2333 {
2334         misc_deregister(&fuse_miscdevice);
2335         kmem_cache_destroy(fuse_req_cachep);
2336 }