fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/pipe_fs_i.h>
  20 #include <linux/swap.h>
  21 #include <linux/splice.h>
  22
  23 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  24 MODULE_ALIAS("devname:fuse");
  25
  26 static struct kmem_cache *fuse_req_cachep;
  27
  28 static struct fuse_dev *fuse_get_dev(struct file *file)
  29 {
  30         /*
  31          * Lockless access is OK, because file->private data is set
  32          * once during mount and is valid until the file is released.
  33          */
  34         return ACCESS_ONCE(file->private_data);
  35 }
  36
  37 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  38                               struct fuse_page_desc *page_descs,
  39                               unsigned npages)
  40 {
  41         memset(req, 0, sizeof(*req));
  42         memset(pages, 0, sizeof(*pages) * npages);
  43         memset(page_descs, 0, sizeof(*page_descs) * npages);
  44         INIT_LIST_HEAD(&req->list);
  45         INIT_LIST_HEAD(&req->intr_entry);
  46         init_waitqueue_head(&req->waitq);
  47         atomic_set(&req->count, 1);
  48         req->pages = pages;
  49         req->page_descs = page_descs;
  50         req->max_pages = npages;
  51         __set_bit(FR_PENDING, &req->flags);
  52 }
  53
  54 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  55 {
  56         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, flags);
  57         if (req) {
  58                 struct page **pages;
  59                 struct fuse_page_desc *page_descs;
  60
  61                 if (npages <= FUSE_REQ_INLINE_PAGES) {
  62                         pages = req->inline_pages;
  63                         page_descs = req->inline_page_descs;
  64                 } else {
  65                         pages = kmalloc(sizeof(struct page *) * npages, flags);
  66                         page_descs = kmalloc(sizeof(struct fuse_page_desc) *
  67                                              npages, flags);
  68                 }
  69
  70                 if (!pages || !page_descs) {
  71                         kfree(pages);
  72                         kfree(page_descs);
  73                         kmem_cache_free(fuse_req_cachep, req);
  74                         return NULL;
  75                 }
  76
  77                 fuse_request_init(req, pages, page_descs, npages);
  78         }
  79         return req;
  80 }
  81
  82 struct fuse_req *fuse_request_alloc(unsigned npages)
  83 {
  84         return __fuse_request_alloc(npages, GFP_KERNEL);
  85 }
  86 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  87
  88 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  89 {
  90         return __fuse_request_alloc(npages, GFP_NOFS);
  91 }
  92
  93 void fuse_request_free(struct fuse_req *req)
  94 {
  95         if (req->pages != req->inline_pages) {
  96                 kfree(req->pages);
  97                 kfree(req->page_descs);
  98         }
  99         kmem_cache_free(fuse_req_cachep, req);
 100 }
 101
 102 void __fuse_get_request(struct fuse_req *req)
 103 {
 104         atomic_inc(&req->count);
 105 }
 106
 107 /* Must be called with > 1 refcount */
 108 static void __fuse_put_request(struct fuse_req *req)
 109 {
 110         BUG_ON(atomic_read(&req->count) < 2);
 111         atomic_dec(&req->count);
 112 }
 113
 114 static void fuse_req_init_context(struct fuse_req *req)
 115 {
 116         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 117         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 118         req->in.h.pid = current->pid;
 119 }
 120
 121 void fuse_set_initialized(struct fuse_conn *fc)
 122 {
 123         /* Make sure stores before this are seen on another CPU */
 124         smp_wmb();
 125         fc->initialized = 1;
 126 }
 127
 128 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 129 {
 130         return !fc->initialized || (for_background && fc->blocked);
 131 }
 132
 133 static void fuse_drop_waiting(struct fuse_conn *fc)
 134 {
 135         /*
 136          * lockess check of fc->connected is okay, because atomic_dec_and_test()
 137          * provides a memory barrier mached with the one in fuse_wait_aborted()
 138          * to ensure no wake-up is missed.
 139          */
 140         if (atomic_dec_and_test(&fc->num_waiting) &&
 141             !READ_ONCE(fc->connected)) {
 142                 /* wake up aborters */
 143                 wake_up_all(&fc->blocked_waitq);
 144         }
 145 }
 146
 147 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 148                                        bool for_background)
 149 {
 150         struct fuse_req *req;
 151         int err;
 152         atomic_inc(&fc->num_waiting);
 153
 154         if (fuse_block_alloc(fc, for_background)) {
 155                 err = -EINTR;
 156                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 157                                 !fuse_block_alloc(fc, for_background)))
 158                         goto out;
 159         }
 160         /* Matches smp_wmb() in fuse_set_initialized() */
 161         smp_rmb();
 162
 163         err = -ENOTCONN;
 164         if (!fc->connected)
 165                 goto out;
 166
 167         err = -ECONNREFUSED;
 168         if (fc->conn_error)
 169                 goto out;
 170
 171         req = fuse_request_alloc(npages);
 172         err = -ENOMEM;
 173         if (!req) {
 174                 if (for_background)
 175                         wake_up(&fc->blocked_waitq);
 176                 goto out;
 177         }
 178
 179         fuse_req_init_context(req);
 180         __set_bit(FR_WAITING, &req->flags);
 181         if (for_background)
 182                 __set_bit(FR_BACKGROUND, &req->flags);
 183
 184         return req;
 185
 186  out:
 187         fuse_drop_waiting(fc);
 188         return ERR_PTR(err);
 189 }
 190
 191 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 192 {
 193         return __fuse_get_req(fc, npages, false);
 194 }
 195 EXPORT_SYMBOL_GPL(fuse_get_req);
 196
 197 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 198                                              unsigned npages)
 199 {
 200         return __fuse_get_req(fc, npages, true);
 201 }
 202 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 203
 204 /*
 205  * Return request in fuse_file->reserved_req.  However that may
 206  * currently be in use.  If that is the case, wait for it to become
 207  * available.
 208  */
 209 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 210                                          struct file *file)
 211 {
 212         struct fuse_req *req = NULL;
 213         struct fuse_file *ff = file->private_data;
 214
 215         do {
 216                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 217                 spin_lock(&fc->lock);
 218                 if (ff->reserved_req) {
 219                         req = ff->reserved_req;
 220                         ff->reserved_req = NULL;
 221                         req->stolen_file = get_file(file);
 222                 }
 223                 spin_unlock(&fc->lock);
 224         } while (!req);
 225
 226         return req;
 227 }
 228
 229 /*
 230  * Put stolen request back into fuse_file->reserved_req
 231  */
 232 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 233 {
 234         struct file *file = req->stolen_file;
 235         struct fuse_file *ff = file->private_data;
 236
 237         spin_lock(&fc->lock);
 238         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 239         BUG_ON(ff->reserved_req);
 240         ff->reserved_req = req;
 241         wake_up_all(&fc->reserved_req_waitq);
 242         spin_unlock(&fc->lock);
 243         fput(file);
 244 }
 245
 246 /*
 247  * Gets a requests for a file operation, always succeeds
 248  *
 249  * This is used for sending the FLUSH request, which must get to
 250  * userspace, due to POSIX locks which may need to be unlocked.
 251  *
 252  * If allocation fails due to OOM, use the reserved request in
 253  * fuse_file.
 254  *
 255  * This is very unlikely to deadlock accidentally, since the
 256  * filesystem should not have it's own file open.  If deadlock is
 257  * intentional, it can still be broken by "aborting" the filesystem.
 258  */
 259 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 260                                              struct file *file)
 261 {
 262         struct fuse_req *req;
 263
 264         atomic_inc(&fc->num_waiting);
 265         wait_event(fc->blocked_waitq, fc->initialized);
 266         /* Matches smp_wmb() in fuse_set_initialized() */
 267         smp_rmb();
 268         req = fuse_request_alloc(0);
 269         if (!req)
 270                 req = get_reserved_req(fc, file);
 271
 272         fuse_req_init_context(req);
 273         __set_bit(FR_WAITING, &req->flags);
 274         __clear_bit(FR_BACKGROUND, &req->flags);
 275         return req;
 276 }
 277
 278 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 279 {
 280         if (atomic_dec_and_test(&req->count)) {
 281                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 282                         /*
 283                          * We get here in the unlikely case that a background
 284                          * request was allocated but not sent
 285                          */
 286                         spin_lock(&fc->lock);
 287                         if (!fc->blocked)
 288                                 wake_up(&fc->blocked_waitq);
 289                         spin_unlock(&fc->lock);
 290                 }
 291
 292                 if (test_bit(FR_WAITING, &req->flags)) {
 293                         __clear_bit(FR_WAITING, &req->flags);
 294                         fuse_drop_waiting(fc);
 295                 }
 296
 297                 if (req->stolen_file)
 298                         put_reserved_req(fc, req);
 299                 else
 300                         fuse_request_free(req);
 301         }
 302 }
 303 EXPORT_SYMBOL_GPL(fuse_put_request);
 304
 305 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 306 {
 307         unsigned nbytes = 0;
 308         unsigned i;
 309
 310         for (i = 0; i < numargs; i++)
 311                 nbytes += args[i].size;
 312
 313         return nbytes;
 314 }
 315
 316 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 317 {
 318         return ++fiq->reqctr;
 319 }
 320
 321 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 322 {
 323         req->in.h.len = sizeof(struct fuse_in_header) +
 324                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 325         list_add_tail(&req->list, &fiq->pending);
 326         wake_up_locked(&fiq->waitq);
 327         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 328 }
 329
 330 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 331                        u64 nodeid, u64 nlookup)
 332 {
 333         struct fuse_iqueue *fiq = &fc->iq;
 334
 335         forget->forget_one.nodeid = nodeid;
 336         forget->forget_one.nlookup = nlookup;
 337
 338         spin_lock(&fiq->waitq.lock);
 339         if (fiq->connected) {
 340                 fiq->forget_list_tail->next = forget;
 341                 fiq->forget_list_tail = forget;
 342                 wake_up_locked(&fiq->waitq);
 343                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 344         } else {
 345                 kfree(forget);
 346         }
 347         spin_unlock(&fiq->waitq.lock);
 348 }
 349
 350 static void flush_bg_queue(struct fuse_conn *fc)
 351 {
 352         while (fc->active_background < fc->max_background &&
 353                !list_empty(&fc->bg_queue)) {
 354                 struct fuse_req *req;
 355                 struct fuse_iqueue *fiq = &fc->iq;
 356
 357                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 358                 list_del(&req->list);
 359                 fc->active_background++;
 360                 spin_lock(&fiq->waitq.lock);
 361                 req->in.h.unique = fuse_get_unique(fiq);
 362                 queue_request(fiq, req);
 363                 spin_unlock(&fiq->waitq.lock);
 364         }
 365 }
 366
 367 /*
 368  * This function is called when a request is finished.  Either a reply
 369  * has arrived or it was aborted (and not yet sent) or some error
 370  * occurred during communication with userspace, or the device file
 371  * was closed.  The requester thread is woken up (if still waiting),
 372  * the 'end' callback is called if given, else the reference to the
 373  * request is released
 374  */
 375 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 376 {
 377         struct fuse_iqueue *fiq = &fc->iq;
 378
 379         if (test_and_set_bit(FR_FINISHED, &req->flags))
 380                 goto put_request;
 381
 382         spin_lock(&fiq->waitq.lock);
 383         list_del_init(&req->intr_entry);
 384         spin_unlock(&fiq->waitq.lock);
 385         WARN_ON(test_bit(FR_PENDING, &req->flags));
 386         WARN_ON(test_bit(FR_SENT, &req->flags));
 387         if (test_bit(FR_BACKGROUND, &req->flags)) {
 388                 spin_lock(&fc->lock);
 389                 clear_bit(FR_BACKGROUND, &req->flags);
 390                 if (fc->num_background == fc->max_background) {
 391                         fc->blocked = 0;
 392                         wake_up(&fc->blocked_waitq);
 393                 } else if (!fc->blocked) {
 394                         /*
 395                          * Wake up next waiter, if any.  It's okay to use
 396                          * waitqueue_active(), as we've already synced up
 397                          * fc->blocked with waiters with the wake_up() call
 398                          * above.
 399                          */
 400                         if (waitqueue_active(&fc->blocked_waitq))
 401                                 wake_up(&fc->blocked_waitq);
 402                 }
 403
 404                 if (fc->num_background == fc->congestion_threshold &&
 405                     fc->connected && fc->bdi_initialized) {
 406                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 407                         clear_bdi_congested(&fc->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 &&
 598             fc->bdi_initialized) {
 599                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 600                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 601         }
 602         list_add_tail(&req->list, &fc->bg_queue);
 603         flush_bg_queue(fc);
 604 }
 605
 606 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 607 {
 608         BUG_ON(!req->end);
 609         spin_lock(&fc->lock);
 610         if (fc->connected) {
 611                 fuse_request_send_background_locked(fc, req);
 612                 spin_unlock(&fc->lock);
 613         } else {
 614                 spin_unlock(&fc->lock);
 615                 req->out.h.error = -ENOTCONN;
 616                 req->end(fc, req);
 617                 fuse_put_request(fc, req);
 618         }
 619 }
 620 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 621
 622 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 623                                           struct fuse_req *req, u64 unique)
 624 {
 625         int err = -ENODEV;
 626         struct fuse_iqueue *fiq = &fc->iq;
 627
 628         __clear_bit(FR_ISREPLY, &req->flags);
 629         req->in.h.unique = unique;
 630         spin_lock(&fiq->waitq.lock);
 631         if (fiq->connected) {
 632                 queue_request(fiq, req);
 633                 err = 0;
 634         }
 635         spin_unlock(&fiq->waitq.lock);
 636
 637         return err;
 638 }
 639
 640 void fuse_force_forget(struct file *file, u64 nodeid)
 641 {
 642         struct inode *inode = file_inode(file);
 643         struct fuse_conn *fc = get_fuse_conn(inode);
 644         struct fuse_req *req;
 645         struct fuse_forget_in inarg;
 646
 647         memset(&inarg, 0, sizeof(inarg));
 648         inarg.nlookup = 1;
 649         req = fuse_get_req_nofail_nopages(fc, file);
 650         req->in.h.opcode = FUSE_FORGET;
 651         req->in.h.nodeid = nodeid;
 652         req->in.numargs = 1;
 653         req->in.args[0].size = sizeof(inarg);
 654         req->in.args[0].value = &inarg;
 655         __clear_bit(FR_ISREPLY, &req->flags);
 656         __fuse_request_send(fc, req);
 657         /* ignore errors */
 658         fuse_put_request(fc, req);
 659 }
 660
 661 /*
 662  * Lock the request.  Up to the next unlock_request() there mustn't be
 663  * anything that could cause a page-fault.  If the request was already
 664  * aborted bail out.
 665  */
 666 static int lock_request(struct fuse_req *req)
 667 {
 668         int err = 0;
 669         if (req) {
 670                 spin_lock(&req->waitq.lock);
 671                 if (test_bit(FR_ABORTED, &req->flags))
 672                         err = -ENOENT;
 673                 else
 674                         set_bit(FR_LOCKED, &req->flags);
 675                 spin_unlock(&req->waitq.lock);
 676         }
 677         return err;
 678 }
 679
 680 /*
 681  * Unlock request.  If it was aborted while locked, caller is responsible
 682  * for unlocking and ending the request.
 683  */
 684 static int unlock_request(struct fuse_req *req)
 685 {
 686         int err = 0;
 687         if (req) {
 688                 spin_lock(&req->waitq.lock);
 689                 if (test_bit(FR_ABORTED, &req->flags))
 690                         err = -ENOENT;
 691                 else
 692                         clear_bit(FR_LOCKED, &req->flags);
 693                 spin_unlock(&req->waitq.lock);
 694         }
 695         return err;
 696 }
 697
 698 struct fuse_copy_state {
 699         int write;
 700         struct fuse_req *req;
 701         struct iov_iter *iter;
 702         struct pipe_buffer *pipebufs;
 703         struct pipe_buffer *currbuf;
 704         struct pipe_inode_info *pipe;
 705         unsigned long nr_segs;
 706         struct page *pg;
 707         unsigned len;
 708         unsigned offset;
 709         unsigned move_pages:1;
 710 };
 711
 712 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 713                            struct iov_iter *iter)
 714 {
 715         memset(cs, 0, sizeof(*cs));
 716         cs->write = write;
 717         cs->iter = iter;
 718 }
 719
 720 /* Unmap and put previous page of userspace buffer */
 721 static void fuse_copy_finish(struct fuse_copy_state *cs)
 722 {
 723         if (cs->currbuf) {
 724                 struct pipe_buffer *buf = cs->currbuf;
 725
 726                 if (cs->write)
 727                         buf->len = PAGE_SIZE - cs->len;
 728                 cs->currbuf = NULL;
 729         } else if (cs->pg) {
 730                 if (cs->write) {
 731                         flush_dcache_page(cs->pg);
 732                         set_page_dirty_lock(cs->pg);
 733                 }
 734                 put_page(cs->pg);
 735         }
 736         cs->pg = NULL;
 737 }
 738
 739 /*
 740  * Get another pagefull of userspace buffer, and map it to kernel
 741  * address space, and lock request
 742  */
 743 static int fuse_copy_fill(struct fuse_copy_state *cs)
 744 {
 745         struct page *page;
 746         int err;
 747
 748         err = unlock_request(cs->req);
 749         if (err)
 750                 return err;
 751
 752         fuse_copy_finish(cs);
 753         if (cs->pipebufs) {
 754                 struct pipe_buffer *buf = cs->pipebufs;
 755
 756                 if (!cs->write) {
 757                         err = pipe_buf_confirm(cs->pipe, buf);
 758                         if (err)
 759                                 return err;
 760
 761                         BUG_ON(!cs->nr_segs);
 762                         cs->currbuf = buf;
 763                         cs->pg = buf->page;
 764                         cs->offset = buf->offset;
 765                         cs->len = buf->len;
 766                         cs->pipebufs++;
 767                         cs->nr_segs--;
 768                 } else {
 769                         if (cs->nr_segs == cs->pipe->buffers)
 770                                 return -EIO;
 771
 772                         page = alloc_page(GFP_HIGHUSER);
 773                         if (!page)
 774                                 return -ENOMEM;
 775
 776                         buf->page = page;
 777                         buf->offset = 0;
 778                         buf->len = 0;
 779
 780                         cs->currbuf = buf;
 781                         cs->pg = page;
 782                         cs->offset = 0;
 783                         cs->len = PAGE_SIZE;
 784                         cs->pipebufs++;
 785                         cs->nr_segs++;
 786                 }
 787         } else {
 788                 size_t off;
 789                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 790                 if (err < 0)
 791                         return err;
 792                 BUG_ON(!err);
 793                 cs->len = err;
 794                 cs->offset = off;
 795                 cs->pg = page;
 796                 iov_iter_advance(cs->iter, err);
 797         }
 798
 799         return lock_request(cs->req);
 800 }
 801
 802 /* Do as much copy to/from userspace buffer as we can */
 803 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 804 {
 805         unsigned ncpy = min(*size, cs->len);
 806         if (val) {
 807                 void *pgaddr = kmap_atomic(cs->pg);
 808                 void *buf = pgaddr + cs->offset;
 809
 810                 if (cs->write)
 811                         memcpy(buf, *val, ncpy);
 812                 else
 813                         memcpy(*val, buf, ncpy);
 814
 815                 kunmap_atomic(pgaddr);
 816                 *val += ncpy;
 817         }
 818         *size -= ncpy;
 819         cs->len -= ncpy;
 820         cs->offset += ncpy;
 821         return ncpy;
 822 }
 823
 824 static int fuse_check_page(struct page *page)
 825 {
 826         if (page_mapcount(page) ||
 827             page->mapping != NULL ||
 828             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 829              ~(1 << PG_locked |
 830                1 << PG_referenced |
 831                1 << PG_uptodate |
 832                1 << PG_lru |
 833                1 << PG_active |
 834                1 << PG_reclaim))) {
 835                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 836                 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);
 837                 return 1;
 838         }
 839         return 0;
 840 }
 841
 842 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 843 {
 844         int err;
 845         struct page *oldpage = *pagep;
 846         struct page *newpage;
 847         struct pipe_buffer *buf = cs->pipebufs;
 848
 849         get_page(oldpage);
 850         err = unlock_request(cs->req);
 851         if (err)
 852                 goto out_put_old;
 853
 854         fuse_copy_finish(cs);
 855
 856         err = pipe_buf_confirm(cs->pipe, buf);
 857         if (err)
 858                 goto out_put_old;
 859
 860         BUG_ON(!cs->nr_segs);
 861         cs->currbuf = buf;
 862         cs->len = buf->len;
 863         cs->pipebufs++;
 864         cs->nr_segs--;
 865
 866         if (cs->len != PAGE_SIZE)
 867                 goto out_fallback;
 868
 869         if (pipe_buf_steal(cs->pipe, buf) != 0)
 870                 goto out_fallback;
 871
 872         newpage = buf->page;
 873
 874         if (!PageUptodate(newpage))
 875                 SetPageUptodate(newpage);
 876
 877         ClearPageMappedToDisk(newpage);
 878
 879         if (fuse_check_page(newpage) != 0)
 880                 goto out_fallback_unlock;
 881
 882         /*
 883          * This is a new and locked page, it shouldn't be mapped or
 884          * have any special flags on it
 885          */
 886         if (WARN_ON(page_mapped(oldpage)))
 887                 goto out_fallback_unlock;
 888         if (WARN_ON(page_has_private(oldpage)))
 889                 goto out_fallback_unlock;
 890         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 891                 goto out_fallback_unlock;
 892         if (WARN_ON(PageMlocked(oldpage)))
 893                 goto out_fallback_unlock;
 894
 895         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 896         if (err) {
 897                 unlock_page(newpage);
 898                 goto out_put_old;
 899         }
 900
 901         get_page(newpage);
 902
 903         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 904                 lru_cache_add_file(newpage);
 905
 906         /*
 907          * Release while we have extra ref on stolen page.  Otherwise
 908          * anon_pipe_buf_release() might think the page can be reused.
 909          */
 910         pipe_buf_release(cs->pipe, buf);
 911
 912         err = 0;
 913         spin_lock(&cs->req->waitq.lock);
 914         if (test_bit(FR_ABORTED, &cs->req->flags))
 915                 err = -ENOENT;
 916         else
 917                 *pagep = newpage;
 918         spin_unlock(&cs->req->waitq.lock);
 919
 920         if (err) {
 921                 unlock_page(newpage);
 922                 put_page(newpage);
 923                 goto out_put_old;
 924         }
 925
 926         unlock_page(oldpage);
 927         /* Drop ref for ap->pages[] array */
 928         put_page(oldpage);
 929         cs->len = 0;
 930
 931         err = 0;
 932 out_put_old:
 933         /* Drop ref obtained in this function */
 934         put_page(oldpage);
 935         return err;
 936
 937 out_fallback_unlock:
 938         unlock_page(newpage);
 939 out_fallback:
 940         cs->pg = buf->page;
 941         cs->offset = buf->offset;
 942
 943         err = lock_request(cs->req);
 944         if (!err)
 945                 err = 1;
 946
 947         goto out_put_old;
 948 }
 949
 950 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 951                          unsigned offset, unsigned count)
 952 {
 953         struct pipe_buffer *buf;
 954         int err;
 955
 956         if (cs->nr_segs == cs->pipe->buffers)
 957                 return -EIO;
 958
 959         get_page(page);
 960         err = unlock_request(cs->req);
 961         if (err) {
 962                 put_page(page);
 963                 return err;
 964         }
 965
 966         fuse_copy_finish(cs);
 967
 968         buf = cs->pipebufs;
 969         buf->page = page;
 970         buf->offset = offset;
 971         buf->len = count;
 972
 973         cs->pipebufs++;
 974         cs->nr_segs++;
 975         cs->len = 0;
 976
 977         return 0;
 978 }
 979
 980 /*
 981  * Copy a page in the request to/from the userspace buffer.  Must be
 982  * done atomically
 983  */
 984 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 985                           unsigned offset, unsigned count, int zeroing)
 986 {
 987         int err;
 988         struct page *page = *pagep;
 989
 990         if (page && zeroing && count < PAGE_SIZE)
 991                 clear_highpage(page);
 992
 993         while (count) {
 994                 if (cs->write && cs->pipebufs && page) {
 995                         /*
 996                          * Can't control lifetime of pipe buffers, so always
 997                          * copy user pages.
 998                          */
 999                         if (cs->req->user_pages) {
1000                                 err = fuse_copy_fill(cs);
1001                                 if (err)
1002                                         return err;
1003                         } else {
1004                                 return fuse_ref_page(cs, page, offset, count);
1005                         }
1006                 } else if (!cs->len) {
1007                         if (cs->move_pages && page &&
1008                             offset == 0 && count == PAGE_SIZE) {
1009                                 err = fuse_try_move_page(cs, pagep);
1010                                 if (err <= 0)
1011                                         return err;
1012                         } else {
1013                                 err = fuse_copy_fill(cs);
1014                                 if (err)
1015                                         return err;
1016                         }
1017                 }
1018                 if (page) {
1019                         void *mapaddr = kmap_atomic(page);
1020                         void *buf = mapaddr + offset;
1021                         offset += fuse_copy_do(cs, &buf, &count);
1022                         kunmap_atomic(mapaddr);
1023                 } else
1024                         offset += fuse_copy_do(cs, NULL, &count);
1025         }
1026         if (page && !cs->write)
1027                 flush_dcache_page(page);
1028         return 0;
1029 }
1030
1031 /* Copy pages in the request to/from userspace buffer */
1032 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1033                            int zeroing)
1034 {
1035         unsigned i;
1036         struct fuse_req *req = cs->req;
1037
1038         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1039                 int err;
1040                 unsigned offset = req->page_descs[i].offset;
1041                 unsigned count = min(nbytes, req->page_descs[i].length);
1042
1043                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1044                                      zeroing);
1045                 if (err)
1046                         return err;
1047
1048                 nbytes -= count;
1049         }
1050         return 0;
1051 }
1052
1053 /* Copy a single argument in the request to/from userspace buffer */
1054 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1055 {
1056         while (size) {
1057                 if (!cs->len) {
1058                         int err = fuse_copy_fill(cs);
1059                         if (err)
1060                                 return err;
1061                 }
1062                 fuse_copy_do(cs, &val, &size);
1063         }
1064         return 0;
1065 }
1066
1067 /* Copy request arguments to/from userspace buffer */
1068 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1069                           unsigned argpages, struct fuse_arg *args,
1070                           int zeroing)
1071 {
1072         int err = 0;
1073         unsigned i;
1074
1075         for (i = 0; !err && i < numargs; i++)  {
1076                 struct fuse_arg *arg = &args[i];
1077                 if (i == numargs - 1 && argpages)
1078                         err = fuse_copy_pages(cs, arg->size, zeroing);
1079                 else
1080                         err = fuse_copy_one(cs, arg->value, arg->size);
1081         }
1082         return err;
1083 }
1084
1085 static int forget_pending(struct fuse_iqueue *fiq)
1086 {
1087         return fiq->forget_list_head.next != NULL;
1088 }
1089
1090 static int request_pending(struct fuse_iqueue *fiq)
1091 {
1092         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1093                 forget_pending(fiq);
1094 }
1095
1096 /*
1097  * Transfer an interrupt request to userspace
1098  *
1099  * Unlike other requests this is assembled on demand, without a need
1100  * to allocate a separate fuse_req structure.
1101  *
1102  * Called with fiq->waitq.lock held, releases it
1103  */
1104 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1105                                struct fuse_copy_state *cs,
1106                                size_t nbytes, struct fuse_req *req)
1107 __releases(fiq->waitq.lock)
1108 {
1109         struct fuse_in_header ih;
1110         struct fuse_interrupt_in arg;
1111         unsigned reqsize = sizeof(ih) + sizeof(arg);
1112         int err;
1113
1114         list_del_init(&req->intr_entry);
1115         req->intr_unique = fuse_get_unique(fiq);
1116         memset(&ih, 0, sizeof(ih));
1117         memset(&arg, 0, sizeof(arg));
1118         ih.len = reqsize;
1119         ih.opcode = FUSE_INTERRUPT;
1120         ih.unique = req->intr_unique;
1121         arg.unique = req->in.h.unique;
1122
1123         spin_unlock(&fiq->waitq.lock);
1124         if (nbytes < reqsize)
1125                 return -EINVAL;
1126
1127         err = fuse_copy_one(cs, &ih, sizeof(ih));
1128         if (!err)
1129                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1130         fuse_copy_finish(cs);
1131
1132         return err ? err : reqsize;
1133 }
1134
1135 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1136                                                unsigned max,
1137                                                unsigned *countp)
1138 {
1139         struct fuse_forget_link *head = fiq->forget_list_head.next;
1140         struct fuse_forget_link **newhead = &head;
1141         unsigned count;
1142
1143         for (count = 0; *newhead != NULL && count < max; count++)
1144                 newhead = &(*newhead)->next;
1145
1146         fiq->forget_list_head.next = *newhead;
1147         *newhead = NULL;
1148         if (fiq->forget_list_head.next == NULL)
1149                 fiq->forget_list_tail = &fiq->forget_list_head;
1150
1151         if (countp != NULL)
1152                 *countp = count;
1153
1154         return head;
1155 }
1156
1157 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1158                                    struct fuse_copy_state *cs,
1159                                    size_t nbytes)
1160 __releases(fiq->waitq.lock)
1161 {
1162         int err;
1163         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1164         struct fuse_forget_in arg = {
1165                 .nlookup = forget->forget_one.nlookup,
1166         };
1167         struct fuse_in_header ih = {
1168                 .opcode = FUSE_FORGET,
1169                 .nodeid = forget->forget_one.nodeid,
1170                 .unique = fuse_get_unique(fiq),
1171                 .len = sizeof(ih) + sizeof(arg),
1172         };
1173
1174         spin_unlock(&fiq->waitq.lock);
1175         kfree(forget);
1176         if (nbytes < ih.len)
1177                 return -EINVAL;
1178
1179         err = fuse_copy_one(cs, &ih, sizeof(ih));
1180         if (!err)
1181                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1182         fuse_copy_finish(cs);
1183
1184         if (err)
1185                 return err;
1186
1187         return ih.len;
1188 }
1189
1190 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1191                                    struct fuse_copy_state *cs, size_t nbytes)
1192 __releases(fiq->waitq.lock)
1193 {
1194         int err;
1195         unsigned max_forgets;
1196         unsigned count;
1197         struct fuse_forget_link *head;
1198         struct fuse_batch_forget_in arg = { .count = 0 };
1199         struct fuse_in_header ih = {
1200                 .opcode = FUSE_BATCH_FORGET,
1201                 .unique = fuse_get_unique(fiq),
1202                 .len = sizeof(ih) + sizeof(arg),
1203         };
1204
1205         if (nbytes < ih.len) {
1206                 spin_unlock(&fiq->waitq.lock);
1207                 return -EINVAL;
1208         }
1209
1210         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1211         head = dequeue_forget(fiq, max_forgets, &count);
1212         spin_unlock(&fiq->waitq.lock);
1213
1214         arg.count = count;
1215         ih.len += count * sizeof(struct fuse_forget_one);
1216         err = fuse_copy_one(cs, &ih, sizeof(ih));
1217         if (!err)
1218                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1219
1220         while (head) {
1221                 struct fuse_forget_link *forget = head;
1222
1223                 if (!err) {
1224                         err = fuse_copy_one(cs, &forget->forget_one,
1225                                             sizeof(forget->forget_one));
1226                 }
1227                 head = forget->next;
1228                 kfree(forget);
1229         }
1230
1231         fuse_copy_finish(cs);
1232
1233         if (err)
1234                 return err;
1235
1236         return ih.len;
1237 }
1238
1239 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1240                             struct fuse_copy_state *cs,
1241                             size_t nbytes)
1242 __releases(fiq->waitq.lock)
1243 {
1244         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1245                 return fuse_read_single_forget(fiq, cs, nbytes);
1246         else
1247                 return fuse_read_batch_forget(fiq, cs, nbytes);
1248 }
1249
1250 /*
1251  * Read a single request into the userspace filesystem's buffer.  This
1252  * function waits until a request is available, then removes it from
1253  * the pending list and copies request data to userspace buffer.  If
1254  * no reply is needed (FORGET) or request has been aborted or there
1255  * was an error during the copying then it's finished by calling
1256  * request_end().  Otherwise add it to the processing list, and set
1257  * the 'sent' flag.
1258  */
1259 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1260                                 struct fuse_copy_state *cs, size_t nbytes)
1261 {
1262         ssize_t err;
1263         struct fuse_conn *fc = fud->fc;
1264         struct fuse_iqueue *fiq = &fc->iq;
1265         struct fuse_pqueue *fpq = &fud->pq;
1266         struct fuse_req *req;
1267         struct fuse_in *in;
1268         unsigned reqsize;
1269
1270  restart:
1271         spin_lock(&fiq->waitq.lock);
1272         err = -EAGAIN;
1273         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1274             !request_pending(fiq))
1275                 goto err_unlock;
1276
1277         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1278                                 !fiq->connected || request_pending(fiq));
1279         if (err)
1280                 goto err_unlock;
1281
1282         err = -ENODEV;
1283         if (!fiq->connected)
1284                 goto err_unlock;
1285
1286         if (!list_empty(&fiq->interrupts)) {
1287                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1288                                  intr_entry);
1289                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1290         }
1291
1292         if (forget_pending(fiq)) {
1293                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1294                         return fuse_read_forget(fc, fiq, cs, nbytes);
1295
1296                 if (fiq->forget_batch <= -8)
1297                         fiq->forget_batch = 16;
1298         }
1299
1300         req = list_entry(fiq->pending.next, struct fuse_req, list);
1301         clear_bit(FR_PENDING, &req->flags);
1302         list_del_init(&req->list);
1303         spin_unlock(&fiq->waitq.lock);
1304
1305         in = &req->in;
1306         reqsize = in->h.len;
1307         /* If request is too large, reply with an error and restart the read */
1308         if (nbytes < reqsize) {
1309                 req->out.h.error = -EIO;
1310                 /* SETXATTR is special, since it may contain too large data */
1311                 if (in->h.opcode == FUSE_SETXATTR)
1312                         req->out.h.error = -E2BIG;
1313                 request_end(fc, req);
1314                 goto restart;
1315         }
1316         spin_lock(&fpq->lock);
1317         /*
1318          *  Must not put request on fpq->io queue after having been shut down by
1319          *  fuse_abort_conn()
1320          */
1321         if (!fpq->connected) {
1322                 req->out.h.error = err = -ECONNABORTED;
1323                 goto out_end;
1324
1325         }
1326         list_add(&req->list, &fpq->io);
1327         spin_unlock(&fpq->lock);
1328         cs->req = req;
1329         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1330         if (!err)
1331                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1332                                      (struct fuse_arg *) in->args, 0);
1333         fuse_copy_finish(cs);
1334         spin_lock(&fpq->lock);
1335         clear_bit(FR_LOCKED, &req->flags);
1336         if (!fpq->connected) {
1337                 err = -ENODEV;
1338                 goto out_end;
1339         }
1340         if (err) {
1341                 req->out.h.error = -EIO;
1342                 goto out_end;
1343         }
1344         if (!test_bit(FR_ISREPLY, &req->flags)) {
1345                 err = reqsize;
1346                 goto out_end;
1347         }
1348         list_move_tail(&req->list, &fpq->processing);
1349         __fuse_get_request(req);
1350         set_bit(FR_SENT, &req->flags);
1351         spin_unlock(&fpq->lock);
1352         /* matches barrier in request_wait_answer() */
1353         smp_mb__after_atomic();
1354         if (test_bit(FR_INTERRUPTED, &req->flags))
1355                 queue_interrupt(fiq, req);
1356         fuse_put_request(fc, req);
1357
1358         return reqsize;
1359
1360 out_end:
1361         if (!test_bit(FR_PRIVATE, &req->flags))
1362                 list_del_init(&req->list);
1363         spin_unlock(&fpq->lock);
1364         request_end(fc, req);
1365         return err;
1366
1367  err_unlock:
1368         spin_unlock(&fiq->waitq.lock);
1369         return err;
1370 }
1371
1372 static int fuse_dev_open(struct inode *inode, struct file *file)
1373 {
1374         /*
1375          * The fuse device's file's private_data is used to hold
1376          * the fuse_conn(ection) when it is mounted, and is used to
1377          * keep track of whether the file has been mounted already.
1378          */
1379         file->private_data = NULL;
1380         return 0;
1381 }
1382
1383 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1384 {
1385         struct fuse_copy_state cs;
1386         struct file *file = iocb->ki_filp;
1387         struct fuse_dev *fud = fuse_get_dev(file);
1388
1389         if (!fud)
1390                 return -EPERM;
1391
1392         if (!iter_is_iovec(to))
1393                 return -EINVAL;
1394
1395         fuse_copy_init(&cs, 1, to);
1396
1397         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1398 }
1399
1400 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1401                                     struct pipe_inode_info *pipe,
1402                                     size_t len, unsigned int flags)
1403 {
1404         int total, ret;
1405         int page_nr = 0;
1406         struct pipe_buffer *bufs;
1407         struct fuse_copy_state cs;
1408         struct fuse_dev *fud = fuse_get_dev(in);
1409
1410         if (!fud)
1411                 return -EPERM;
1412
1413         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1414         if (!bufs)
1415                 return -ENOMEM;
1416
1417         fuse_copy_init(&cs, 1, NULL);
1418         cs.pipebufs = bufs;
1419         cs.pipe = pipe;
1420         ret = fuse_dev_do_read(fud, in, &cs, len);
1421         if (ret < 0)
1422                 goto out;
1423
1424         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1425                 ret = -EIO;
1426                 goto out;
1427         }
1428
1429         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1430                 /*
1431                  * Need to be careful about this.  Having buf->ops in module
1432                  * code can Oops if the buffer persists after module unload.
1433                  */
1434                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1435                 bufs[page_nr].flags = 0;
1436                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1437                 if (unlikely(ret < 0))
1438                         break;
1439         }
1440         if (total)
1441                 ret = total;
1442 out:
1443         for (; page_nr < cs.nr_segs; page_nr++)
1444                 put_page(bufs[page_nr].page);
1445
1446         kfree(bufs);
1447         return ret;
1448 }
1449
1450 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1451                             struct fuse_copy_state *cs)
1452 {
1453         struct fuse_notify_poll_wakeup_out outarg;
1454         int err = -EINVAL;
1455
1456         if (size != sizeof(outarg))
1457                 goto err;
1458
1459         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1460         if (err)
1461                 goto err;
1462
1463         fuse_copy_finish(cs);
1464         return fuse_notify_poll_wakeup(fc, &outarg);
1465
1466 err:
1467         fuse_copy_finish(cs);
1468         return err;
1469 }
1470
1471 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1472                                    struct fuse_copy_state *cs)
1473 {
1474         struct fuse_notify_inval_inode_out outarg;
1475         int err = -EINVAL;
1476
1477         if (size != sizeof(outarg))
1478                 goto err;
1479
1480         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1481         if (err)
1482                 goto err;
1483         fuse_copy_finish(cs);
1484
1485         down_read(&fc->killsb);
1486         err = -ENOENT;
1487         if (fc->sb) {
1488                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1489                                                outarg.off, outarg.len);
1490         }
1491         up_read(&fc->killsb);
1492         return err;
1493
1494 err:
1495         fuse_copy_finish(cs);
1496         return err;
1497 }
1498
1499 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1500                                    struct fuse_copy_state *cs)
1501 {
1502         struct fuse_notify_inval_entry_out outarg;
1503         int err = -ENOMEM;
1504         char *buf;
1505         struct qstr name;
1506
1507         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1508         if (!buf)
1509                 goto err;
1510
1511         err = -EINVAL;
1512         if (size < sizeof(outarg))
1513                 goto err;
1514
1515         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1516         if (err)
1517                 goto err;
1518
1519         err = -ENAMETOOLONG;
1520         if (outarg.namelen > FUSE_NAME_MAX)
1521                 goto err;
1522
1523         err = -EINVAL;
1524         if (size != sizeof(outarg) + outarg.namelen + 1)
1525                 goto err;
1526
1527         name.name = buf;
1528         name.len = outarg.namelen;
1529         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1530         if (err)
1531                 goto err;
1532         fuse_copy_finish(cs);
1533         buf[outarg.namelen] = 0;
1534
1535         down_read(&fc->killsb);
1536         err = -ENOENT;
1537         if (fc->sb)
1538                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1539         up_read(&fc->killsb);
1540         kfree(buf);
1541         return err;
1542
1543 err:
1544         kfree(buf);
1545         fuse_copy_finish(cs);
1546         return err;
1547 }
1548
1549 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1550                               struct fuse_copy_state *cs)
1551 {
1552         struct fuse_notify_delete_out outarg;
1553         int err = -ENOMEM;
1554         char *buf;
1555         struct qstr name;
1556
1557         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1558         if (!buf)
1559                 goto err;
1560
1561         err = -EINVAL;
1562         if (size < sizeof(outarg))
1563                 goto err;
1564
1565         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1566         if (err)
1567                 goto err;
1568
1569         err = -ENAMETOOLONG;
1570         if (outarg.namelen > FUSE_NAME_MAX)
1571                 goto err;
1572
1573         err = -EINVAL;
1574         if (size != sizeof(outarg) + outarg.namelen + 1)
1575                 goto err;
1576
1577         name.name = buf;
1578         name.len = outarg.namelen;
1579         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1580         if (err)
1581                 goto err;
1582         fuse_copy_finish(cs);
1583         buf[outarg.namelen] = 0;
1584
1585         down_read(&fc->killsb);
1586         err = -ENOENT;
1587         if (fc->sb)
1588                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1589                                                outarg.child, &name);
1590         up_read(&fc->killsb);
1591         kfree(buf);
1592         return err;
1593
1594 err:
1595         kfree(buf);
1596         fuse_copy_finish(cs);
1597         return err;
1598 }
1599
1600 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1601                              struct fuse_copy_state *cs)
1602 {
1603         struct fuse_notify_store_out outarg;
1604         struct inode *inode;
1605         struct address_space *mapping;
1606         u64 nodeid;
1607         int err;
1608         pgoff_t index;
1609         unsigned int offset;
1610         unsigned int num;
1611         loff_t file_size;
1612         loff_t end;
1613
1614         err = -EINVAL;
1615         if (size < sizeof(outarg))
1616                 goto out_finish;
1617
1618         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1619         if (err)
1620                 goto out_finish;
1621
1622         err = -EINVAL;
1623         if (size - sizeof(outarg) != outarg.size)
1624                 goto out_finish;
1625
1626         nodeid = outarg.nodeid;
1627
1628         down_read(&fc->killsb);
1629
1630         err = -ENOENT;
1631         if (!fc->sb)
1632                 goto out_up_killsb;
1633
1634         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1635         if (!inode)
1636                 goto out_up_killsb;
1637
1638         mapping = inode->i_mapping;
1639         index = outarg.offset >> PAGE_SHIFT;
1640         offset = outarg.offset & ~PAGE_MASK;
1641         file_size = i_size_read(inode);
1642         end = outarg.offset + outarg.size;
1643         if (end > file_size) {
1644                 file_size = end;
1645                 fuse_write_update_size(inode, file_size);
1646         }
1647
1648         num = outarg.size;
1649         while (num) {
1650                 struct page *page;
1651                 unsigned int this_num;
1652
1653                 err = -ENOMEM;
1654                 page = find_or_create_page(mapping, index,
1655                                            mapping_gfp_mask(mapping));
1656                 if (!page)
1657                         goto out_iput;
1658
1659                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1660                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1661                 if (!err && offset == 0 &&
1662                     (this_num == PAGE_SIZE || file_size == end))
1663                         SetPageUptodate(page);
1664                 unlock_page(page);
1665                 put_page(page);
1666
1667                 if (err)
1668                         goto out_iput;
1669
1670                 num -= this_num;
1671                 offset = 0;
1672                 index++;
1673         }
1674
1675         err = 0;
1676
1677 out_iput:
1678         iput(inode);
1679 out_up_killsb:
1680         up_read(&fc->killsb);
1681 out_finish:
1682         fuse_copy_finish(cs);
1683         return err;
1684 }
1685
1686 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1687 {
1688         release_pages(req->pages, req->num_pages, false);
1689 }
1690
1691 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1692                          struct fuse_notify_retrieve_out *outarg)
1693 {
1694         int err;
1695         struct address_space *mapping = inode->i_mapping;
1696         struct fuse_req *req;
1697         pgoff_t index;
1698         loff_t file_size;
1699         unsigned int num;
1700         unsigned int offset;
1701         size_t total_len = 0;
1702         int num_pages;
1703
1704         offset = outarg->offset & ~PAGE_MASK;
1705         file_size = i_size_read(inode);
1706
1707         num = min(outarg->size, fc->max_write);
1708         if (outarg->offset > file_size)
1709                 num = 0;
1710         else if (outarg->offset + num > file_size)
1711                 num = file_size - outarg->offset;
1712
1713         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1714         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1715
1716         req = fuse_get_req(fc, num_pages);
1717         if (IS_ERR(req))
1718                 return PTR_ERR(req);
1719
1720         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1721         req->in.h.nodeid = outarg->nodeid;
1722         req->in.numargs = 2;
1723         req->in.argpages = 1;
1724         req->end = fuse_retrieve_end;
1725
1726         index = outarg->offset >> PAGE_SHIFT;
1727
1728         while (num && req->num_pages < num_pages) {
1729                 struct page *page;
1730                 unsigned int this_num;
1731
1732                 page = find_get_page(mapping, index);
1733                 if (!page)
1734                         break;
1735
1736                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1737                 req->pages[req->num_pages] = page;
1738                 req->page_descs[req->num_pages].offset = offset;
1739                 req->page_descs[req->num_pages].length = this_num;
1740                 req->num_pages++;
1741
1742                 offset = 0;
1743                 num -= this_num;
1744                 total_len += this_num;
1745                 index++;
1746         }
1747         req->misc.retrieve_in.offset = outarg->offset;
1748         req->misc.retrieve_in.size = total_len;
1749         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1750         req->in.args[0].value = &req->misc.retrieve_in;
1751         req->in.args[1].size = total_len;
1752
1753         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1754         if (err) {
1755                 fuse_retrieve_end(fc, req);
1756                 fuse_put_request(fc, req);
1757         }
1758
1759         return err;
1760 }
1761
1762 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1763                                 struct fuse_copy_state *cs)
1764 {
1765         struct fuse_notify_retrieve_out outarg;
1766         struct inode *inode;
1767         int err;
1768
1769         err = -EINVAL;
1770         if (size != sizeof(outarg))
1771                 goto copy_finish;
1772
1773         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1774         if (err)
1775                 goto copy_finish;
1776
1777         fuse_copy_finish(cs);
1778
1779         down_read(&fc->killsb);
1780         err = -ENOENT;
1781         if (fc->sb) {
1782                 u64 nodeid = outarg.nodeid;
1783
1784                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1785                 if (inode) {
1786                         err = fuse_retrieve(fc, inode, &outarg);
1787                         iput(inode);
1788                 }
1789         }
1790         up_read(&fc->killsb);
1791
1792         return err;
1793
1794 copy_finish:
1795         fuse_copy_finish(cs);
1796         return err;
1797 }
1798
1799 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1800                        unsigned int size, struct fuse_copy_state *cs)
1801 {
1802         /* Don't try to move pages (yet) */
1803         cs->move_pages = 0;
1804
1805         switch (code) {
1806         case FUSE_NOTIFY_POLL:
1807                 return fuse_notify_poll(fc, size, cs);
1808
1809         case FUSE_NOTIFY_INVAL_INODE:
1810                 return fuse_notify_inval_inode(fc, size, cs);
1811
1812         case FUSE_NOTIFY_INVAL_ENTRY:
1813                 return fuse_notify_inval_entry(fc, size, cs);
1814
1815         case FUSE_NOTIFY_STORE:
1816                 return fuse_notify_store(fc, size, cs);
1817
1818         case FUSE_NOTIFY_RETRIEVE:
1819                 return fuse_notify_retrieve(fc, size, cs);
1820
1821         case FUSE_NOTIFY_DELETE:
1822                 return fuse_notify_delete(fc, size, cs);
1823
1824         default:
1825                 fuse_copy_finish(cs);
1826                 return -EINVAL;
1827         }
1828 }
1829
1830 /* Look up request on processing list by unique ID */
1831 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1832 {
1833         struct fuse_req *req;
1834
1835         list_for_each_entry(req, &fpq->processing, list) {
1836                 if (req->in.h.unique == unique || req->intr_unique == unique)
1837                         return req;
1838         }
1839         return NULL;
1840 }
1841
1842 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1843                          unsigned nbytes)
1844 {
1845         unsigned reqsize = sizeof(struct fuse_out_header);
1846
1847         if (out->h.error)
1848                 return nbytes != reqsize ? -EINVAL : 0;
1849
1850         reqsize += len_args(out->numargs, out->args);
1851
1852         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1853                 return -EINVAL;
1854         else if (reqsize > nbytes) {
1855                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1856                 unsigned diffsize = reqsize - nbytes;
1857                 if (diffsize > lastarg->size)
1858                         return -EINVAL;
1859                 lastarg->size -= diffsize;
1860         }
1861         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1862                               out->page_zeroing);
1863 }
1864
1865 /*
1866  * Write a single reply to a request.  First the header is copied from
1867  * the write buffer.  The request is then searched on the processing
1868  * list by the unique ID found in the header.  If found, then remove
1869  * it from the list and copy the rest of the buffer to the request.
1870  * The request is finished by calling request_end()
1871  */
1872 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1873                                  struct fuse_copy_state *cs, size_t nbytes)
1874 {
1875         int err;
1876         struct fuse_conn *fc = fud->fc;
1877         struct fuse_pqueue *fpq = &fud->pq;
1878         struct fuse_req *req;
1879         struct fuse_out_header oh;
1880
1881         if (nbytes < sizeof(struct fuse_out_header))
1882                 return -EINVAL;
1883
1884         err = fuse_copy_one(cs, &oh, sizeof(oh));
1885         if (err)
1886                 goto err_finish;
1887
1888         err = -EINVAL;
1889         if (oh.len != nbytes)
1890                 goto err_finish;
1891
1892         /*
1893          * Zero oh.unique indicates unsolicited notification message
1894          * and error contains notification code.
1895          */
1896         if (!oh.unique) {
1897                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1898                 return err ? err : nbytes;
1899         }
1900
1901         err = -EINVAL;
1902         if (oh.error <= -512 || oh.error > 0)
1903                 goto err_finish;
1904
1905         spin_lock(&fpq->lock);
1906         err = -ENOENT;
1907         if (!fpq->connected)
1908                 goto err_unlock_pq;
1909
1910         req = request_find(fpq, oh.unique);
1911         if (!req)
1912                 goto err_unlock_pq;
1913
1914         /* Is it an interrupt reply? */
1915         if (req->intr_unique == oh.unique) {
1916                 __fuse_get_request(req);
1917                 spin_unlock(&fpq->lock);
1918
1919                 err = -EINVAL;
1920                 if (nbytes != sizeof(struct fuse_out_header)) {
1921                         fuse_put_request(fc, req);
1922                         goto err_finish;
1923                 }
1924
1925                 if (oh.error == -ENOSYS)
1926                         fc->no_interrupt = 1;
1927                 else if (oh.error == -EAGAIN)
1928                         queue_interrupt(&fc->iq, req);
1929                 fuse_put_request(fc, req);
1930
1931                 fuse_copy_finish(cs);
1932                 return nbytes;
1933         }
1934
1935         clear_bit(FR_SENT, &req->flags);
1936         list_move(&req->list, &fpq->io);
1937         req->out.h = oh;
1938         set_bit(FR_LOCKED, &req->flags);
1939         spin_unlock(&fpq->lock);
1940         cs->req = req;
1941         if (!req->out.page_replace)
1942                 cs->move_pages = 0;
1943
1944         err = copy_out_args(cs, &req->out, nbytes);
1945         fuse_copy_finish(cs);
1946
1947         spin_lock(&fpq->lock);
1948         clear_bit(FR_LOCKED, &req->flags);
1949         if (!fpq->connected)
1950                 err = -ENOENT;
1951         else if (err)
1952                 req->out.h.error = -EIO;
1953         if (!test_bit(FR_PRIVATE, &req->flags))
1954                 list_del_init(&req->list);
1955         spin_unlock(&fpq->lock);
1956
1957         request_end(fc, req);
1958
1959         return err ? err : nbytes;
1960
1961  err_unlock_pq:
1962         spin_unlock(&fpq->lock);
1963  err_finish:
1964         fuse_copy_finish(cs);
1965         return err;
1966 }
1967
1968 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1969 {
1970         struct fuse_copy_state cs;
1971         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1972
1973         if (!fud)
1974                 return -EPERM;
1975
1976         if (!iter_is_iovec(from))
1977                 return -EINVAL;
1978
1979         fuse_copy_init(&cs, 0, from);
1980
1981         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1982 }
1983
1984 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1985                                      struct file *out, loff_t *ppos,
1986                                      size_t len, unsigned int flags)
1987 {
1988         unsigned nbuf;
1989         unsigned idx;
1990         struct pipe_buffer *bufs;
1991         struct fuse_copy_state cs;
1992         struct fuse_dev *fud;
1993         size_t rem;
1994         ssize_t ret;
1995
1996         fud = fuse_get_dev(out);
1997         if (!fud)
1998                 return -EPERM;
1999
2000         pipe_lock(pipe);
2001
2002         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
2003         if (!bufs) {
2004                 pipe_unlock(pipe);
2005                 return -ENOMEM;
2006         }
2007
2008         nbuf = 0;
2009         rem = 0;
2010         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2011                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2012
2013         ret = -EINVAL;
2014         if (rem < len)
2015                 goto out_free;
2016
2017         rem = len;
2018         while (rem) {
2019                 struct pipe_buffer *ibuf;
2020                 struct pipe_buffer *obuf;
2021
2022                 BUG_ON(nbuf >= pipe->buffers);
2023                 BUG_ON(!pipe->nrbufs);
2024                 ibuf = &pipe->bufs[pipe->curbuf];
2025                 obuf = &bufs[nbuf];
2026
2027                 if (rem >= ibuf->len) {
2028                         *obuf = *ibuf;
2029                         ibuf->ops = NULL;
2030                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2031                         pipe->nrbufs--;
2032                 } else {
2033                         if (!pipe_buf_get(pipe, ibuf))
2034                                 goto out_free;
2035
2036                         *obuf = *ibuf;
2037                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2038                         obuf->len = rem;
2039                         ibuf->offset += obuf->len;
2040                         ibuf->len -= obuf->len;
2041                 }
2042                 nbuf++;
2043                 rem -= obuf->len;
2044         }
2045         pipe_unlock(pipe);
2046
2047         fuse_copy_init(&cs, 0, NULL);
2048         cs.pipebufs = bufs;
2049         cs.nr_segs = nbuf;
2050         cs.pipe = pipe;
2051
2052         if (flags & SPLICE_F_MOVE)
2053                 cs.move_pages = 1;
2054
2055         ret = fuse_dev_do_write(fud, &cs, len);
2056
2057         pipe_lock(pipe);
2058 out_free:
2059         for (idx = 0; idx < nbuf; idx++) {
2060                 struct pipe_buffer *buf = &bufs[idx];
2061
2062                 if (buf->ops)
2063                         pipe_buf_release(pipe, buf);
2064         }
2065         pipe_unlock(pipe);
2066
2067         kfree(bufs);
2068         return ret;
2069 }
2070
2071 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2072 {
2073         unsigned mask = POLLOUT | POLLWRNORM;
2074         struct fuse_iqueue *fiq;
2075         struct fuse_dev *fud = fuse_get_dev(file);
2076
2077         if (!fud)
2078                 return POLLERR;
2079
2080         fiq = &fud->fc->iq;
2081         poll_wait(file, &fiq->waitq, wait);
2082
2083         spin_lock(&fiq->waitq.lock);
2084         if (!fiq->connected)
2085                 mask = POLLERR;
2086         else if (request_pending(fiq))
2087                 mask |= POLLIN | POLLRDNORM;
2088         spin_unlock(&fiq->waitq.lock);
2089
2090         return mask;
2091 }
2092
2093 /*
2094  * Abort all requests on the given list (pending or processing)
2095  *
2096  * This function releases and reacquires fc->lock
2097  */
2098 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2099 {
2100         while (!list_empty(head)) {
2101                 struct fuse_req *req;
2102                 req = list_entry(head->next, struct fuse_req, list);
2103                 req->out.h.error = -ECONNABORTED;
2104                 clear_bit(FR_SENT, &req->flags);
2105                 list_del_init(&req->list);
2106                 request_end(fc, req);
2107         }
2108 }
2109
2110 static void end_polls(struct fuse_conn *fc)
2111 {
2112         struct rb_node *p;
2113
2114         p = rb_first(&fc->polled_files);
2115
2116         while (p) {
2117                 struct fuse_file *ff;
2118                 ff = rb_entry(p, struct fuse_file, polled_node);
2119                 wake_up_interruptible_all(&ff->poll_wait);
2120
2121                 p = rb_next(p);
2122         }
2123 }
2124
2125 /*
2126  * Abort all requests.
2127  *
2128  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2129  * filesystem.
2130  *
2131  * The same effect is usually achievable through killing the filesystem daemon
2132  * and all users of the filesystem.  The exception is the combination of an
2133  * asynchronous request and the tricky deadlock (see
2134  * Documentation/filesystems/fuse.txt).
2135  *
2136  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2137  * requests, they should be finished off immediately.  Locked requests will be
2138  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2139  * requests.  It is possible that some request will finish before we can.  This
2140  * is OK, the request will in that case be removed from the list before we touch
2141  * it.
2142  */
2143 void fuse_abort_conn(struct fuse_conn *fc)
2144 {
2145         struct fuse_iqueue *fiq = &fc->iq;
2146
2147         spin_lock(&fc->lock);
2148         if (fc->connected) {
2149                 struct fuse_dev *fud;
2150                 struct fuse_req *req, *next;
2151                 LIST_HEAD(to_end1);
2152                 LIST_HEAD(to_end2);
2153
2154                 fc->connected = 0;
2155                 fc->blocked = 0;
2156                 fuse_set_initialized(fc);
2157                 list_for_each_entry(fud, &fc->devices, entry) {
2158                         struct fuse_pqueue *fpq = &fud->pq;
2159
2160                         spin_lock(&fpq->lock);
2161                         fpq->connected = 0;
2162                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2163                                 req->out.h.error = -ECONNABORTED;
2164                                 spin_lock(&req->waitq.lock);
2165                                 set_bit(FR_ABORTED, &req->flags);
2166                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2167                                         set_bit(FR_PRIVATE, &req->flags);
2168                                         __fuse_get_request(req);
2169                                         list_move(&req->list, &to_end1);
2170                                 }
2171                                 spin_unlock(&req->waitq.lock);
2172                         }
2173                         list_splice_init(&fpq->processing, &to_end2);
2174                         spin_unlock(&fpq->lock);
2175                 }
2176                 fc->max_background = UINT_MAX;
2177                 flush_bg_queue(fc);
2178
2179                 spin_lock(&fiq->waitq.lock);
2180                 fiq->connected = 0;
2181                 list_splice_init(&fiq->pending, &to_end2);
2182                 list_for_each_entry(req, &to_end2, list)
2183                         clear_bit(FR_PENDING, &req->flags);
2184                 while (forget_pending(fiq))
2185                         kfree(dequeue_forget(fiq, 1, NULL));
2186                 wake_up_all_locked(&fiq->waitq);
2187                 spin_unlock(&fiq->waitq.lock);
2188                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2189                 end_polls(fc);
2190                 wake_up_all(&fc->blocked_waitq);
2191                 spin_unlock(&fc->lock);
2192
2193                 while (!list_empty(&to_end1)) {
2194                         req = list_first_entry(&to_end1, struct fuse_req, list);
2195                         list_del_init(&req->list);
2196                         request_end(fc, req);
2197                 }
2198                 end_requests(fc, &to_end2);
2199         } else {
2200                 spin_unlock(&fc->lock);
2201         }
2202 }
2203 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2204
2205 void fuse_wait_aborted(struct fuse_conn *fc)
2206 {
2207         /* matches implicit memory barrier in fuse_drop_waiting() */
2208         smp_mb();
2209         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2210 }
2211
2212 int fuse_dev_release(struct inode *inode, struct file *file)
2213 {
2214         struct fuse_dev *fud = fuse_get_dev(file);
2215
2216         if (fud) {
2217                 struct fuse_conn *fc = fud->fc;
2218                 struct fuse_pqueue *fpq = &fud->pq;
2219                 LIST_HEAD(to_end);
2220
2221                 spin_lock(&fpq->lock);
2222                 WARN_ON(!list_empty(&fpq->io));
2223                 list_splice_init(&fpq->processing, &to_end);
2224                 spin_unlock(&fpq->lock);
2225
2226                 end_requests(fc, &to_end);
2227
2228                 /* Are we the last open device? */
2229                 if (atomic_dec_and_test(&fc->dev_count)) {
2230                         WARN_ON(fc->iq.fasync != NULL);
2231                         fuse_abort_conn(fc);
2232                 }
2233                 fuse_dev_free(fud);
2234         }
2235         return 0;
2236 }
2237 EXPORT_SYMBOL_GPL(fuse_dev_release);
2238
2239 static int fuse_dev_fasync(int fd, struct file *file, int on)
2240 {
2241         struct fuse_dev *fud = fuse_get_dev(file);
2242
2243         if (!fud)
2244                 return -EPERM;
2245
2246         /* No locking - fasync_helper does its own locking */
2247         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2248 }
2249
2250 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2251 {
2252         struct fuse_dev *fud;
2253
2254         if (new->private_data)
2255                 return -EINVAL;
2256
2257         fud = fuse_dev_alloc(fc);
2258         if (!fud)
2259                 return -ENOMEM;
2260
2261         new->private_data = fud;
2262         atomic_inc(&fc->dev_count);
2263
2264         return 0;
2265 }
2266
2267 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2268                            unsigned long arg)
2269 {
2270         int err = -ENOTTY;
2271
2272         if (cmd == FUSE_DEV_IOC_CLONE) {
2273                 int oldfd;
2274
2275                 err = -EFAULT;
2276                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2277                         struct file *old = fget(oldfd);
2278
2279                         err = -EINVAL;
2280                         if (old) {
2281                                 struct fuse_dev *fud = NULL;
2282
2283                                 /*
2284                                  * Check against file->f_op because CUSE
2285                                  * uses the same ioctl handler.
2286                                  */
2287                                 if (old->f_op == file->f_op &&
2288                                     old->f_cred->user_ns == file->f_cred->user_ns)
2289                                         fud = fuse_get_dev(old);
2290
2291                                 if (fud) {
2292                                         mutex_lock(&fuse_mutex);
2293                                         err = fuse_device_clone(fud->fc, file);
2294                                         mutex_unlock(&fuse_mutex);
2295                                 }
2296                                 fput(old);
2297                         }
2298                 }
2299         }
2300         return err;
2301 }
2302
2303 const struct file_operations fuse_dev_operations = {
2304         .owner          = THIS_MODULE,
2305         .open           = fuse_dev_open,
2306         .llseek         = no_llseek,
2307         .read_iter      = fuse_dev_read,
2308         .splice_read    = fuse_dev_splice_read,
2309         .write_iter     = fuse_dev_write,
2310         .splice_write   = fuse_dev_splice_write,
2311         .poll           = fuse_dev_poll,
2312         .release        = fuse_dev_release,
2313         .fasync         = fuse_dev_fasync,
2314         .unlocked_ioctl = fuse_dev_ioctl,
2315         .compat_ioctl   = fuse_dev_ioctl,
2316 };
2317 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2318
2319 static struct miscdevice fuse_miscdevice = {
2320         .minor = FUSE_MINOR,
2321         .name  = "fuse",
2322         .fops = &fuse_dev_operations,
2323 };
2324
2325 int __init fuse_dev_init(void)
2326 {
2327         int err = -ENOMEM;
2328         fuse_req_cachep = kmem_cache_create("fuse_request",
2329                                             sizeof(struct fuse_req),
2330                                             0, 0, NULL);
2331         if (!fuse_req_cachep)
2332                 goto out;
2333
2334         err = misc_register(&fuse_miscdevice);
2335         if (err)
2336                 goto out_cache_clean;
2337
2338         return 0;
2339
2340  out_cache_clean:
2341         kmem_cache_destroy(fuse_req_cachep);
2342  out:
2343         return err;
2344 }
2345
2346 void fuse_dev_cleanup(void)
2347 {
2348         misc_deregister(&fuse_miscdevice);
2349         kmem_cache_destroy(fuse_req_cachep);
2350 }