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 static void block_sigs(sigset_t *oldset)
 103 {
 104         sigset_t mask;
 105
 106         siginitsetinv(&mask, sigmask(SIGKILL));
 107         sigprocmask(SIG_BLOCK, &mask, oldset);
 108 }
 109
 110 static void restore_sigs(sigset_t *oldset)
 111 {
 112         sigprocmask(SIG_SETMASK, oldset, NULL);
 113 }
 114
 115 void __fuse_get_request(struct fuse_req *req)
 116 {
 117         atomic_inc(&req->count);
 118 }
 119
 120 /* Must be called with > 1 refcount */
 121 static void __fuse_put_request(struct fuse_req *req)
 122 {
 123         BUG_ON(atomic_read(&req->count) < 2);
 124         atomic_dec(&req->count);
 125 }
 126
 127 static void fuse_req_init_context(struct fuse_req *req)
 128 {
 129         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 130         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 131         req->in.h.pid = current->pid;
 132 }
 133
 134 void fuse_set_initialized(struct fuse_conn *fc)
 135 {
 136         /* Make sure stores before this are seen on another CPU */
 137         smp_wmb();
 138         fc->initialized = 1;
 139 }
 140
 141 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 142 {
 143         return !fc->initialized || (for_background && fc->blocked);
 144 }
 145
 146 static void fuse_drop_waiting(struct fuse_conn *fc)
 147 {
 148         /*
 149          * lockess check of fc->connected is okay, because atomic_dec_and_test()
 150          * provides a memory barrier mached with the one in fuse_wait_aborted()
 151          * to ensure no wake-up is missed.
 152          */
 153         if (atomic_dec_and_test(&fc->num_waiting) &&
 154             !READ_ONCE(fc->connected)) {
 155                 /* wake up aborters */
 156                 wake_up_all(&fc->blocked_waitq);
 157         }
 158 }
 159
 160 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 161                                        bool for_background)
 162 {
 163         struct fuse_req *req;
 164         int err;
 165         atomic_inc(&fc->num_waiting);
 166
 167         if (fuse_block_alloc(fc, for_background)) {
 168                 sigset_t oldset;
 169                 int intr;
 170
 171                 block_sigs(&oldset);
 172                 intr = wait_event_interruptible_exclusive(fc->blocked_waitq,
 173                                 !fuse_block_alloc(fc, for_background));
 174                 restore_sigs(&oldset);
 175                 err = -EINTR;
 176                 if (intr)
 177                         goto out;
 178         }
 179         /* Matches smp_wmb() in fuse_set_initialized() */
 180         smp_rmb();
 181
 182         err = -ENOTCONN;
 183         if (!fc->connected)
 184                 goto out;
 185
 186         err = -ECONNREFUSED;
 187         if (fc->conn_error)
 188                 goto out;
 189
 190         req = fuse_request_alloc(npages);
 191         err = -ENOMEM;
 192         if (!req) {
 193                 if (for_background)
 194                         wake_up(&fc->blocked_waitq);
 195                 goto out;
 196         }
 197
 198         fuse_req_init_context(req);
 199         __set_bit(FR_WAITING, &req->flags);
 200         if (for_background)
 201                 __set_bit(FR_BACKGROUND, &req->flags);
 202
 203         return req;
 204
 205  out:
 206         fuse_drop_waiting(fc);
 207         return ERR_PTR(err);
 208 }
 209
 210 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 211 {
 212         return __fuse_get_req(fc, npages, false);
 213 }
 214 EXPORT_SYMBOL_GPL(fuse_get_req);
 215
 216 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 217                                              unsigned npages)
 218 {
 219         return __fuse_get_req(fc, npages, true);
 220 }
 221 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 222
 223 /*
 224  * Return request in fuse_file->reserved_req.  However that may
 225  * currently be in use.  If that is the case, wait for it to become
 226  * available.
 227  */
 228 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 229                                          struct file *file)
 230 {
 231         struct fuse_req *req = NULL;
 232         struct fuse_file *ff = file->private_data;
 233
 234         do {
 235                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 236                 spin_lock(&fc->lock);
 237                 if (ff->reserved_req) {
 238                         req = ff->reserved_req;
 239                         ff->reserved_req = NULL;
 240                         req->stolen_file = get_file(file);
 241                 }
 242                 spin_unlock(&fc->lock);
 243         } while (!req);
 244
 245         return req;
 246 }
 247
 248 /*
 249  * Put stolen request back into fuse_file->reserved_req
 250  */
 251 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 252 {
 253         struct file *file = req->stolen_file;
 254         struct fuse_file *ff = file->private_data;
 255
 256         spin_lock(&fc->lock);
 257         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 258         BUG_ON(ff->reserved_req);
 259         ff->reserved_req = req;
 260         wake_up_all(&fc->reserved_req_waitq);
 261         spin_unlock(&fc->lock);
 262         fput(file);
 263 }
 264
 265 /*
 266  * Gets a requests for a file operation, always succeeds
 267  *
 268  * This is used for sending the FLUSH request, which must get to
 269  * userspace, due to POSIX locks which may need to be unlocked.
 270  *
 271  * If allocation fails due to OOM, use the reserved request in
 272  * fuse_file.
 273  *
 274  * This is very unlikely to deadlock accidentally, since the
 275  * filesystem should not have it's own file open.  If deadlock is
 276  * intentional, it can still be broken by "aborting" the filesystem.
 277  */
 278 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 279                                              struct file *file)
 280 {
 281         struct fuse_req *req;
 282
 283         atomic_inc(&fc->num_waiting);
 284         wait_event(fc->blocked_waitq, fc->initialized);
 285         /* Matches smp_wmb() in fuse_set_initialized() */
 286         smp_rmb();
 287         req = fuse_request_alloc(0);
 288         if (!req)
 289                 req = get_reserved_req(fc, file);
 290
 291         fuse_req_init_context(req);
 292         __set_bit(FR_WAITING, &req->flags);
 293         __clear_bit(FR_BACKGROUND, &req->flags);
 294         return req;
 295 }
 296
 297 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 298 {
 299         if (atomic_dec_and_test(&req->count)) {
 300                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 301                         /*
 302                          * We get here in the unlikely case that a background
 303                          * request was allocated but not sent
 304                          */
 305                         spin_lock(&fc->lock);
 306                         if (!fc->blocked)
 307                                 wake_up(&fc->blocked_waitq);
 308                         spin_unlock(&fc->lock);
 309                 }
 310
 311                 if (test_bit(FR_WAITING, &req->flags)) {
 312                         __clear_bit(FR_WAITING, &req->flags);
 313                         fuse_drop_waiting(fc);
 314                 }
 315
 316                 if (req->stolen_file)
 317                         put_reserved_req(fc, req);
 318                 else
 319                         fuse_request_free(req);
 320         }
 321 }
 322 EXPORT_SYMBOL_GPL(fuse_put_request);
 323
 324 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 325 {
 326         unsigned nbytes = 0;
 327         unsigned i;
 328
 329         for (i = 0; i < numargs; i++)
 330                 nbytes += args[i].size;
 331
 332         return nbytes;
 333 }
 334
 335 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 336 {
 337         return ++fiq->reqctr;
 338 }
 339
 340 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 341 {
 342         req->in.h.len = sizeof(struct fuse_in_header) +
 343                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 344         list_add_tail(&req->list, &fiq->pending);
 345         wake_up_locked(&fiq->waitq);
 346         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 347 }
 348
 349 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 350                        u64 nodeid, u64 nlookup)
 351 {
 352         struct fuse_iqueue *fiq = &fc->iq;
 353
 354         forget->forget_one.nodeid = nodeid;
 355         forget->forget_one.nlookup = nlookup;
 356
 357         spin_lock(&fiq->waitq.lock);
 358         if (fiq->connected) {
 359                 fiq->forget_list_tail->next = forget;
 360                 fiq->forget_list_tail = forget;
 361                 wake_up_locked(&fiq->waitq);
 362                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 363         } else {
 364                 kfree(forget);
 365         }
 366         spin_unlock(&fiq->waitq.lock);
 367 }
 368
 369 static void flush_bg_queue(struct fuse_conn *fc)
 370 {
 371         while (fc->active_background < fc->max_background &&
 372                !list_empty(&fc->bg_queue)) {
 373                 struct fuse_req *req;
 374                 struct fuse_iqueue *fiq = &fc->iq;
 375
 376                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 377                 list_del(&req->list);
 378                 fc->active_background++;
 379                 spin_lock(&fiq->waitq.lock);
 380                 req->in.h.unique = fuse_get_unique(fiq);
 381                 queue_request(fiq, req);
 382                 spin_unlock(&fiq->waitq.lock);
 383         }
 384 }
 385
 386 /*
 387  * This function is called when a request is finished.  Either a reply
 388  * has arrived or it was aborted (and not yet sent) or some error
 389  * occurred during communication with userspace, or the device file
 390  * was closed.  The requester thread is woken up (if still waiting),
 391  * the 'end' callback is called if given, else the reference to the
 392  * request is released
 393  */
 394 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 395 {
 396         struct fuse_iqueue *fiq = &fc->iq;
 397
 398         if (test_and_set_bit(FR_FINISHED, &req->flags))
 399                 goto put_request;
 400
 401         spin_lock(&fiq->waitq.lock);
 402         list_del_init(&req->intr_entry);
 403         spin_unlock(&fiq->waitq.lock);
 404         WARN_ON(test_bit(FR_PENDING, &req->flags));
 405         WARN_ON(test_bit(FR_SENT, &req->flags));
 406         if (test_bit(FR_BACKGROUND, &req->flags)) {
 407                 spin_lock(&fc->lock);
 408                 clear_bit(FR_BACKGROUND, &req->flags);
 409                 if (fc->num_background == fc->max_background) {
 410                         fc->blocked = 0;
 411                         wake_up(&fc->blocked_waitq);
 412                 } else if (!fc->blocked) {
 413                         /*
 414                          * Wake up next waiter, if any.  It's okay to use
 415                          * waitqueue_active(), as we've already synced up
 416                          * fc->blocked with waiters with the wake_up() call
 417                          * above.
 418                          */
 419                         if (waitqueue_active(&fc->blocked_waitq))
 420                                 wake_up(&fc->blocked_waitq);
 421                 }
 422
 423                 if (fc->num_background == fc->congestion_threshold &&
 424                     fc->connected && fc->bdi_initialized) {
 425                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 426                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 427                 }
 428                 fc->num_background--;
 429                 fc->active_background--;
 430                 flush_bg_queue(fc);
 431                 spin_unlock(&fc->lock);
 432         }
 433         wake_up(&req->waitq);
 434         if (req->end)
 435                 req->end(fc, req);
 436 put_request:
 437         fuse_put_request(fc, req);
 438 }
 439
 440 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 441 {
 442         spin_lock(&fiq->waitq.lock);
 443         if (test_bit(FR_FINISHED, &req->flags)) {
 444                 spin_unlock(&fiq->waitq.lock);
 445                 return;
 446         }
 447         if (list_empty(&req->intr_entry)) {
 448                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 449                 wake_up_locked(&fiq->waitq);
 450         }
 451         spin_unlock(&fiq->waitq.lock);
 452         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 453 }
 454
 455 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 456 {
 457         struct fuse_iqueue *fiq = &fc->iq;
 458         int err;
 459
 460         if (!fc->no_interrupt) {
 461                 /* Any signal may interrupt this */
 462                 err = wait_event_interruptible(req->waitq,
 463                                         test_bit(FR_FINISHED, &req->flags));
 464                 if (!err)
 465                         return;
 466
 467                 set_bit(FR_INTERRUPTED, &req->flags);
 468                 /* matches barrier in fuse_dev_do_read() */
 469                 smp_mb__after_atomic();
 470                 if (test_bit(FR_SENT, &req->flags))
 471                         queue_interrupt(fiq, req);
 472         }
 473
 474         if (!test_bit(FR_FORCE, &req->flags)) {
 475                 sigset_t oldset;
 476
 477                 /* Only fatal signals may interrupt this */
 478                 block_sigs(&oldset);
 479                 err = wait_event_interruptible(req->waitq,
 480                                         test_bit(FR_FINISHED, &req->flags));
 481                 restore_sigs(&oldset);
 482
 483                 if (!err)
 484                         return;
 485
 486                 spin_lock(&fiq->waitq.lock);
 487                 /* Request is not yet in userspace, bail out */
 488                 if (test_bit(FR_PENDING, &req->flags)) {
 489                         list_del(&req->list);
 490                         spin_unlock(&fiq->waitq.lock);
 491                         __fuse_put_request(req);
 492                         req->out.h.error = -EINTR;
 493                         return;
 494                 }
 495                 spin_unlock(&fiq->waitq.lock);
 496         }
 497
 498         /*
 499          * Either request is already in userspace, or it was forced.
 500          * Wait it out.
 501          */
 502         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 503 }
 504
 505 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 506 {
 507         struct fuse_iqueue *fiq = &fc->iq;
 508
 509         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 510         spin_lock(&fiq->waitq.lock);
 511         if (!fiq->connected) {
 512                 spin_unlock(&fiq->waitq.lock);
 513                 req->out.h.error = -ENOTCONN;
 514         } else {
 515                 req->in.h.unique = fuse_get_unique(fiq);
 516                 queue_request(fiq, req);
 517                 /* acquire extra reference, since request is still needed
 518                    after request_end() */
 519                 __fuse_get_request(req);
 520                 spin_unlock(&fiq->waitq.lock);
 521
 522                 request_wait_answer(fc, req);
 523                 /* Pairs with smp_wmb() in request_end() */
 524                 smp_rmb();
 525         }
 526 }
 527
 528 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 529 {
 530         __set_bit(FR_ISREPLY, &req->flags);
 531         if (!test_bit(FR_WAITING, &req->flags)) {
 532                 __set_bit(FR_WAITING, &req->flags);
 533                 atomic_inc(&fc->num_waiting);
 534         }
 535         __fuse_request_send(fc, req);
 536 }
 537 EXPORT_SYMBOL_GPL(fuse_request_send);
 538
 539 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 540 {
 541         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 542                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 543
 544         if (fc->minor < 9) {
 545                 switch (args->in.h.opcode) {
 546                 case FUSE_LOOKUP:
 547                 case FUSE_CREATE:
 548                 case FUSE_MKNOD:
 549                 case FUSE_MKDIR:
 550                 case FUSE_SYMLINK:
 551                 case FUSE_LINK:
 552                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 553                         break;
 554                 case FUSE_GETATTR:
 555                 case FUSE_SETATTR:
 556                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 557                         break;
 558                 }
 559         }
 560         if (fc->minor < 12) {
 561                 switch (args->in.h.opcode) {
 562                 case FUSE_CREATE:
 563                         args->in.args[0].size = sizeof(struct fuse_open_in);
 564                         break;
 565                 case FUSE_MKNOD:
 566                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 567                         break;
 568                 }
 569         }
 570 }
 571
 572 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 573 {
 574         struct fuse_req *req;
 575         ssize_t ret;
 576
 577         req = fuse_get_req(fc, 0);
 578         if (IS_ERR(req))
 579                 return PTR_ERR(req);
 580
 581         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 582         fuse_adjust_compat(fc, args);
 583
 584         req->in.h.opcode = args->in.h.opcode;
 585         req->in.h.nodeid = args->in.h.nodeid;
 586         req->in.numargs = args->in.numargs;
 587         memcpy(req->in.args, args->in.args,
 588                args->in.numargs * sizeof(struct fuse_in_arg));
 589         req->out.argvar = args->out.argvar;
 590         req->out.numargs = args->out.numargs;
 591         memcpy(req->out.args, args->out.args,
 592                args->out.numargs * sizeof(struct fuse_arg));
 593         fuse_request_send(fc, req);
 594         ret = req->out.h.error;
 595         if (!ret && args->out.argvar) {
 596                 BUG_ON(args->out.numargs != 1);
 597                 ret = req->out.args[0].size;
 598         }
 599         fuse_put_request(fc, req);
 600
 601         return ret;
 602 }
 603
 604 /*
 605  * Called under fc->lock
 606  *
 607  * fc->connected must have been checked previously
 608  */
 609 void fuse_request_send_background_locked(struct fuse_conn *fc,
 610                                          struct fuse_req *req)
 611 {
 612         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 613         if (!test_bit(FR_WAITING, &req->flags)) {
 614                 __set_bit(FR_WAITING, &req->flags);
 615                 atomic_inc(&fc->num_waiting);
 616         }
 617         __set_bit(FR_ISREPLY, &req->flags);
 618         fc->num_background++;
 619         if (fc->num_background == fc->max_background)
 620                 fc->blocked = 1;
 621         if (fc->num_background == fc->congestion_threshold &&
 622             fc->bdi_initialized) {
 623                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 624                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 625         }
 626         list_add_tail(&req->list, &fc->bg_queue);
 627         flush_bg_queue(fc);
 628 }
 629
 630 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 631 {
 632         BUG_ON(!req->end);
 633         spin_lock(&fc->lock);
 634         if (fc->connected) {
 635                 fuse_request_send_background_locked(fc, req);
 636                 spin_unlock(&fc->lock);
 637         } else {
 638                 spin_unlock(&fc->lock);
 639                 req->out.h.error = -ENOTCONN;
 640                 req->end(fc, req);
 641                 fuse_put_request(fc, req);
 642         }
 643 }
 644 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 645
 646 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 647                                           struct fuse_req *req, u64 unique)
 648 {
 649         int err = -ENODEV;
 650         struct fuse_iqueue *fiq = &fc->iq;
 651
 652         __clear_bit(FR_ISREPLY, &req->flags);
 653         req->in.h.unique = unique;
 654         spin_lock(&fiq->waitq.lock);
 655         if (fiq->connected) {
 656                 queue_request(fiq, req);
 657                 err = 0;
 658         }
 659         spin_unlock(&fiq->waitq.lock);
 660
 661         return err;
 662 }
 663
 664 void fuse_force_forget(struct file *file, u64 nodeid)
 665 {
 666         struct inode *inode = file_inode(file);
 667         struct fuse_conn *fc = get_fuse_conn(inode);
 668         struct fuse_req *req;
 669         struct fuse_forget_in inarg;
 670
 671         memset(&inarg, 0, sizeof(inarg));
 672         inarg.nlookup = 1;
 673         req = fuse_get_req_nofail_nopages(fc, file);
 674         req->in.h.opcode = FUSE_FORGET;
 675         req->in.h.nodeid = nodeid;
 676         req->in.numargs = 1;
 677         req->in.args[0].size = sizeof(inarg);
 678         req->in.args[0].value = &inarg;
 679         __clear_bit(FR_ISREPLY, &req->flags);
 680         __fuse_request_send(fc, req);
 681         /* ignore errors */
 682         fuse_put_request(fc, req);
 683 }
 684
 685 /*
 686  * Lock the request.  Up to the next unlock_request() there mustn't be
 687  * anything that could cause a page-fault.  If the request was already
 688  * aborted bail out.
 689  */
 690 static int lock_request(struct fuse_req *req)
 691 {
 692         int err = 0;
 693         if (req) {
 694                 spin_lock(&req->waitq.lock);
 695                 if (test_bit(FR_ABORTED, &req->flags))
 696                         err = -ENOENT;
 697                 else
 698                         set_bit(FR_LOCKED, &req->flags);
 699                 spin_unlock(&req->waitq.lock);
 700         }
 701         return err;
 702 }
 703
 704 /*
 705  * Unlock request.  If it was aborted while locked, caller is responsible
 706  * for unlocking and ending the request.
 707  */
 708 static int unlock_request(struct fuse_req *req)
 709 {
 710         int err = 0;
 711         if (req) {
 712                 spin_lock(&req->waitq.lock);
 713                 if (test_bit(FR_ABORTED, &req->flags))
 714                         err = -ENOENT;
 715                 else
 716                         clear_bit(FR_LOCKED, &req->flags);
 717                 spin_unlock(&req->waitq.lock);
 718         }
 719         return err;
 720 }
 721
 722 struct fuse_copy_state {
 723         int write;
 724         struct fuse_req *req;
 725         struct iov_iter *iter;
 726         struct pipe_buffer *pipebufs;
 727         struct pipe_buffer *currbuf;
 728         struct pipe_inode_info *pipe;
 729         unsigned long nr_segs;
 730         struct page *pg;
 731         unsigned len;
 732         unsigned offset;
 733         unsigned move_pages:1;
 734 };
 735
 736 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 737                            struct iov_iter *iter)
 738 {
 739         memset(cs, 0, sizeof(*cs));
 740         cs->write = write;
 741         cs->iter = iter;
 742 }
 743
 744 /* Unmap and put previous page of userspace buffer */
 745 static void fuse_copy_finish(struct fuse_copy_state *cs)
 746 {
 747         if (cs->currbuf) {
 748                 struct pipe_buffer *buf = cs->currbuf;
 749
 750                 if (cs->write)
 751                         buf->len = PAGE_SIZE - cs->len;
 752                 cs->currbuf = NULL;
 753         } else if (cs->pg) {
 754                 if (cs->write) {
 755                         flush_dcache_page(cs->pg);
 756                         set_page_dirty_lock(cs->pg);
 757                 }
 758                 put_page(cs->pg);
 759         }
 760         cs->pg = NULL;
 761 }
 762
 763 /*
 764  * Get another pagefull of userspace buffer, and map it to kernel
 765  * address space, and lock request
 766  */
 767 static int fuse_copy_fill(struct fuse_copy_state *cs)
 768 {
 769         struct page *page;
 770         int err;
 771
 772         err = unlock_request(cs->req);
 773         if (err)
 774                 return err;
 775
 776         fuse_copy_finish(cs);
 777         if (cs->pipebufs) {
 778                 struct pipe_buffer *buf = cs->pipebufs;
 779
 780                 if (!cs->write) {
 781                         err = buf->ops->confirm(cs->pipe, buf);
 782                         if (err)
 783                                 return err;
 784
 785                         BUG_ON(!cs->nr_segs);
 786                         cs->currbuf = buf;
 787                         cs->pg = buf->page;
 788                         cs->offset = buf->offset;
 789                         cs->len = buf->len;
 790                         cs->pipebufs++;
 791                         cs->nr_segs--;
 792                 } else {
 793                         if (cs->nr_segs == cs->pipe->buffers)
 794                                 return -EIO;
 795
 796                         page = alloc_page(GFP_HIGHUSER);
 797                         if (!page)
 798                                 return -ENOMEM;
 799
 800                         buf->page = page;
 801                         buf->offset = 0;
 802                         buf->len = 0;
 803
 804                         cs->currbuf = buf;
 805                         cs->pg = page;
 806                         cs->offset = 0;
 807                         cs->len = PAGE_SIZE;
 808                         cs->pipebufs++;
 809                         cs->nr_segs++;
 810                 }
 811         } else {
 812                 size_t off;
 813                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 814                 if (err < 0)
 815                         return err;
 816                 BUG_ON(!err);
 817                 cs->len = err;
 818                 cs->offset = off;
 819                 cs->pg = page;
 820                 cs->offset = off;
 821                 iov_iter_advance(cs->iter, err);
 822         }
 823
 824         return lock_request(cs->req);
 825 }
 826
 827 /* Do as much copy to/from userspace buffer as we can */
 828 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 829 {
 830         unsigned ncpy = min(*size, cs->len);
 831         if (val) {
 832                 void *pgaddr = kmap_atomic(cs->pg);
 833                 void *buf = pgaddr + cs->offset;
 834
 835                 if (cs->write)
 836                         memcpy(buf, *val, ncpy);
 837                 else
 838                         memcpy(*val, buf, ncpy);
 839
 840                 kunmap_atomic(pgaddr);
 841                 *val += ncpy;
 842         }
 843         *size -= ncpy;
 844         cs->len -= ncpy;
 845         cs->offset += ncpy;
 846         return ncpy;
 847 }
 848
 849 static int fuse_check_page(struct page *page)
 850 {
 851         if (page_mapcount(page) ||
 852             page->mapping != NULL ||
 853             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 854              ~(1 << PG_locked |
 855                1 << PG_referenced |
 856                1 << PG_uptodate |
 857                1 << PG_lru |
 858                1 << PG_active |
 859                1 << PG_reclaim))) {
 860                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 861                 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);
 862                 return 1;
 863         }
 864         return 0;
 865 }
 866
 867 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 868 {
 869         int err;
 870         struct page *oldpage = *pagep;
 871         struct page *newpage;
 872         struct pipe_buffer *buf = cs->pipebufs;
 873
 874         err = unlock_request(cs->req);
 875         if (err)
 876                 return err;
 877
 878         fuse_copy_finish(cs);
 879
 880         err = buf->ops->confirm(cs->pipe, buf);
 881         if (err)
 882                 return err;
 883
 884         BUG_ON(!cs->nr_segs);
 885         cs->currbuf = buf;
 886         cs->len = buf->len;
 887         cs->pipebufs++;
 888         cs->nr_segs--;
 889
 890         if (cs->len != PAGE_SIZE)
 891                 goto out_fallback;
 892
 893         if (buf->ops->steal(cs->pipe, buf) != 0)
 894                 goto out_fallback;
 895
 896         newpage = buf->page;
 897
 898         if (!PageUptodate(newpage))
 899                 SetPageUptodate(newpage);
 900
 901         ClearPageMappedToDisk(newpage);
 902
 903         if (fuse_check_page(newpage) != 0)
 904                 goto out_fallback_unlock;
 905
 906         /*
 907          * This is a new and locked page, it shouldn't be mapped or
 908          * have any special flags on it
 909          */
 910         if (WARN_ON(page_mapped(oldpage)))
 911                 goto out_fallback_unlock;
 912         if (WARN_ON(page_has_private(oldpage)))
 913                 goto out_fallback_unlock;
 914         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 915                 goto out_fallback_unlock;
 916         if (WARN_ON(PageMlocked(oldpage)))
 917                 goto out_fallback_unlock;
 918
 919         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 920         if (err) {
 921                 unlock_page(newpage);
 922                 return err;
 923         }
 924
 925         page_cache_get(newpage);
 926
 927         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 928                 lru_cache_add_file(newpage);
 929
 930         err = 0;
 931         spin_lock(&cs->req->waitq.lock);
 932         if (test_bit(FR_ABORTED, &cs->req->flags))
 933                 err = -ENOENT;
 934         else
 935                 *pagep = newpage;
 936         spin_unlock(&cs->req->waitq.lock);
 937
 938         if (err) {
 939                 unlock_page(newpage);
 940                 page_cache_release(newpage);
 941                 return err;
 942         }
 943
 944         unlock_page(oldpage);
 945         page_cache_release(oldpage);
 946         cs->len = 0;
 947
 948         return 0;
 949
 950 out_fallback_unlock:
 951         unlock_page(newpage);
 952 out_fallback:
 953         cs->pg = buf->page;
 954         cs->offset = buf->offset;
 955
 956         err = lock_request(cs->req);
 957         if (err)
 958                 return err;
 959
 960         return 1;
 961 }
 962
 963 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 964                          unsigned offset, unsigned count)
 965 {
 966         struct pipe_buffer *buf;
 967         int err;
 968
 969         if (cs->nr_segs == cs->pipe->buffers)
 970                 return -EIO;
 971
 972         err = unlock_request(cs->req);
 973         if (err)
 974                 return err;
 975
 976         fuse_copy_finish(cs);
 977
 978         buf = cs->pipebufs;
 979         page_cache_get(page);
 980         buf->page = page;
 981         buf->offset = offset;
 982         buf->len = count;
 983
 984         cs->pipebufs++;
 985         cs->nr_segs++;
 986         cs->len = 0;
 987
 988         return 0;
 989 }
 990
 991 /*
 992  * Copy a page in the request to/from the userspace buffer.  Must be
 993  * done atomically
 994  */
 995 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 996                           unsigned offset, unsigned count, int zeroing)
 997 {
 998         int err;
 999         struct page *page = *pagep;
1000
1001         if (page && zeroing && count < PAGE_SIZE)
1002                 clear_highpage(page);
1003
1004         while (count) {
1005                 if (cs->write && cs->pipebufs && page) {
1006                         return fuse_ref_page(cs, page, offset, count);
1007                 } else if (!cs->len) {
1008                         if (cs->move_pages && page &&
1009                             offset == 0 && count == PAGE_SIZE) {
1010                                 err = fuse_try_move_page(cs, pagep);
1011                                 if (err <= 0)
1012                                         return err;
1013                         } else {
1014                                 err = fuse_copy_fill(cs);
1015                                 if (err)
1016                                         return err;
1017                         }
1018                 }
1019                 if (page) {
1020                         void *mapaddr = kmap_atomic(page);
1021                         void *buf = mapaddr + offset;
1022                         offset += fuse_copy_do(cs, &buf, &count);
1023                         kunmap_atomic(mapaddr);
1024                 } else
1025                         offset += fuse_copy_do(cs, NULL, &count);
1026         }
1027         if (page && !cs->write)
1028                 flush_dcache_page(page);
1029         return 0;
1030 }
1031
1032 /* Copy pages in the request to/from userspace buffer */
1033 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1034                            int zeroing)
1035 {
1036         unsigned i;
1037         struct fuse_req *req = cs->req;
1038
1039         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1040                 int err;
1041                 unsigned offset = req->page_descs[i].offset;
1042                 unsigned count = min(nbytes, req->page_descs[i].length);
1043
1044                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1045                                      zeroing);
1046                 if (err)
1047                         return err;
1048
1049                 nbytes -= count;
1050         }
1051         return 0;
1052 }
1053
1054 /* Copy a single argument in the request to/from userspace buffer */
1055 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1056 {
1057         while (size) {
1058                 if (!cs->len) {
1059                         int err = fuse_copy_fill(cs);
1060                         if (err)
1061                                 return err;
1062                 }
1063                 fuse_copy_do(cs, &val, &size);
1064         }
1065         return 0;
1066 }
1067
1068 /* Copy request arguments to/from userspace buffer */
1069 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1070                           unsigned argpages, struct fuse_arg *args,
1071                           int zeroing)
1072 {
1073         int err = 0;
1074         unsigned i;
1075
1076         for (i = 0; !err && i < numargs; i++)  {
1077                 struct fuse_arg *arg = &args[i];
1078                 if (i == numargs - 1 && argpages)
1079                         err = fuse_copy_pages(cs, arg->size, zeroing);
1080                 else
1081                         err = fuse_copy_one(cs, arg->value, arg->size);
1082         }
1083         return err;
1084 }
1085
1086 static int forget_pending(struct fuse_iqueue *fiq)
1087 {
1088         return fiq->forget_list_head.next != NULL;
1089 }
1090
1091 static int request_pending(struct fuse_iqueue *fiq)
1092 {
1093         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1094                 forget_pending(fiq);
1095 }
1096
1097 /*
1098  * Transfer an interrupt request to userspace
1099  *
1100  * Unlike other requests this is assembled on demand, without a need
1101  * to allocate a separate fuse_req structure.
1102  *
1103  * Called with fiq->waitq.lock held, releases it
1104  */
1105 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1106                                struct fuse_copy_state *cs,
1107                                size_t nbytes, struct fuse_req *req)
1108 __releases(fiq->waitq.lock)
1109 {
1110         struct fuse_in_header ih;
1111         struct fuse_interrupt_in arg;
1112         unsigned reqsize = sizeof(ih) + sizeof(arg);
1113         int err;
1114
1115         list_del_init(&req->intr_entry);
1116         req->intr_unique = fuse_get_unique(fiq);
1117         memset(&ih, 0, sizeof(ih));
1118         memset(&arg, 0, sizeof(arg));
1119         ih.len = reqsize;
1120         ih.opcode = FUSE_INTERRUPT;
1121         ih.unique = req->intr_unique;
1122         arg.unique = req->in.h.unique;
1123
1124         spin_unlock(&fiq->waitq.lock);
1125         if (nbytes < reqsize)
1126                 return -EINVAL;
1127
1128         err = fuse_copy_one(cs, &ih, sizeof(ih));
1129         if (!err)
1130                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1131         fuse_copy_finish(cs);
1132
1133         return err ? err : reqsize;
1134 }
1135
1136 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1137                                                unsigned max,
1138                                                unsigned *countp)
1139 {
1140         struct fuse_forget_link *head = fiq->forget_list_head.next;
1141         struct fuse_forget_link **newhead = &head;
1142         unsigned count;
1143
1144         for (count = 0; *newhead != NULL && count < max; count++)
1145                 newhead = &(*newhead)->next;
1146
1147         fiq->forget_list_head.next = *newhead;
1148         *newhead = NULL;
1149         if (fiq->forget_list_head.next == NULL)
1150                 fiq->forget_list_tail = &fiq->forget_list_head;
1151
1152         if (countp != NULL)
1153                 *countp = count;
1154
1155         return head;
1156 }
1157
1158 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1159                                    struct fuse_copy_state *cs,
1160                                    size_t nbytes)
1161 __releases(fiq->waitq.lock)
1162 {
1163         int err;
1164         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1165         struct fuse_forget_in arg = {
1166                 .nlookup = forget->forget_one.nlookup,
1167         };
1168         struct fuse_in_header ih = {
1169                 .opcode = FUSE_FORGET,
1170                 .nodeid = forget->forget_one.nodeid,
1171                 .unique = fuse_get_unique(fiq),
1172                 .len = sizeof(ih) + sizeof(arg),
1173         };
1174
1175         spin_unlock(&fiq->waitq.lock);
1176         kfree(forget);
1177         if (nbytes < ih.len)
1178                 return -EINVAL;
1179
1180         err = fuse_copy_one(cs, &ih, sizeof(ih));
1181         if (!err)
1182                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1183         fuse_copy_finish(cs);
1184
1185         if (err)
1186                 return err;
1187
1188         return ih.len;
1189 }
1190
1191 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1192                                    struct fuse_copy_state *cs, size_t nbytes)
1193 __releases(fiq->waitq.lock)
1194 {
1195         int err;
1196         unsigned max_forgets;
1197         unsigned count;
1198         struct fuse_forget_link *head;
1199         struct fuse_batch_forget_in arg = { .count = 0 };
1200         struct fuse_in_header ih = {
1201                 .opcode = FUSE_BATCH_FORGET,
1202                 .unique = fuse_get_unique(fiq),
1203                 .len = sizeof(ih) + sizeof(arg),
1204         };
1205
1206         if (nbytes < ih.len) {
1207                 spin_unlock(&fiq->waitq.lock);
1208                 return -EINVAL;
1209         }
1210
1211         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1212         head = dequeue_forget(fiq, max_forgets, &count);
1213         spin_unlock(&fiq->waitq.lock);
1214
1215         arg.count = count;
1216         ih.len += count * sizeof(struct fuse_forget_one);
1217         err = fuse_copy_one(cs, &ih, sizeof(ih));
1218         if (!err)
1219                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1220
1221         while (head) {
1222                 struct fuse_forget_link *forget = head;
1223
1224                 if (!err) {
1225                         err = fuse_copy_one(cs, &forget->forget_one,
1226                                             sizeof(forget->forget_one));
1227                 }
1228                 head = forget->next;
1229                 kfree(forget);
1230         }
1231
1232         fuse_copy_finish(cs);
1233
1234         if (err)
1235                 return err;
1236
1237         return ih.len;
1238 }
1239
1240 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1241                             struct fuse_copy_state *cs,
1242                             size_t nbytes)
1243 __releases(fiq->waitq.lock)
1244 {
1245         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1246                 return fuse_read_single_forget(fiq, cs, nbytes);
1247         else
1248                 return fuse_read_batch_forget(fiq, cs, nbytes);
1249 }
1250
1251 /*
1252  * Read a single request into the userspace filesystem's buffer.  This
1253  * function waits until a request is available, then removes it from
1254  * the pending list and copies request data to userspace buffer.  If
1255  * no reply is needed (FORGET) or request has been aborted or there
1256  * was an error during the copying then it's finished by calling
1257  * request_end().  Otherwise add it to the processing list, and set
1258  * the 'sent' flag.
1259  */
1260 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1261                                 struct fuse_copy_state *cs, size_t nbytes)
1262 {
1263         ssize_t err;
1264         struct fuse_conn *fc = fud->fc;
1265         struct fuse_iqueue *fiq = &fc->iq;
1266         struct fuse_pqueue *fpq = &fud->pq;
1267         struct fuse_req *req;
1268         struct fuse_in *in;
1269         unsigned reqsize;
1270
1271  restart:
1272         spin_lock(&fiq->waitq.lock);
1273         err = -EAGAIN;
1274         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1275             !request_pending(fiq))
1276                 goto err_unlock;
1277
1278         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1279                                 !fiq->connected || request_pending(fiq));
1280         if (err)
1281                 goto err_unlock;
1282
1283         err = -ENODEV;
1284         if (!fiq->connected)
1285                 goto err_unlock;
1286
1287         if (!list_empty(&fiq->interrupts)) {
1288                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1289                                  intr_entry);
1290                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1291         }
1292
1293         if (forget_pending(fiq)) {
1294                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1295                         return fuse_read_forget(fc, fiq, cs, nbytes);
1296
1297                 if (fiq->forget_batch <= -8)
1298                         fiq->forget_batch = 16;
1299         }
1300
1301         req = list_entry(fiq->pending.next, struct fuse_req, list);
1302         clear_bit(FR_PENDING, &req->flags);
1303         list_del_init(&req->list);
1304         spin_unlock(&fiq->waitq.lock);
1305
1306         in = &req->in;
1307         reqsize = in->h.len;
1308         /* If request is too large, reply with an error and restart the read */
1309         if (nbytes < reqsize) {
1310                 req->out.h.error = -EIO;
1311                 /* SETXATTR is special, since it may contain too large data */
1312                 if (in->h.opcode == FUSE_SETXATTR)
1313                         req->out.h.error = -E2BIG;
1314                 request_end(fc, req);
1315                 goto restart;
1316         }
1317         spin_lock(&fpq->lock);
1318         /*
1319          *  Must not put request on fpq->io queue after having been shut down by
1320          *  fuse_abort_conn()
1321          */
1322         if (!fpq->connected) {
1323                 req->out.h.error = err = -ECONNABORTED;
1324                 goto out_end;
1325
1326         }
1327         list_add(&req->list, &fpq->io);
1328         spin_unlock(&fpq->lock);
1329         cs->req = req;
1330         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1331         if (!err)
1332                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1333                                      (struct fuse_arg *) in->args, 0);
1334         fuse_copy_finish(cs);
1335         spin_lock(&fpq->lock);
1336         clear_bit(FR_LOCKED, &req->flags);
1337         if (!fpq->connected) {
1338                 err = -ENODEV;
1339                 goto out_end;
1340         }
1341         if (err) {
1342                 req->out.h.error = -EIO;
1343                 goto out_end;
1344         }
1345         if (!test_bit(FR_ISREPLY, &req->flags)) {
1346                 err = reqsize;
1347                 goto out_end;
1348         }
1349         list_move_tail(&req->list, &fpq->processing);
1350         __fuse_get_request(req);
1351         set_bit(FR_SENT, &req->flags);
1352         spin_unlock(&fpq->lock);
1353         /* matches barrier in request_wait_answer() */
1354         smp_mb__after_atomic();
1355         if (test_bit(FR_INTERRUPTED, &req->flags))
1356                 queue_interrupt(fiq, req);
1357         fuse_put_request(fc, req);
1358
1359         return reqsize;
1360
1361 out_end:
1362         if (!test_bit(FR_PRIVATE, &req->flags))
1363                 list_del_init(&req->list);
1364         spin_unlock(&fpq->lock);
1365         request_end(fc, req);
1366         return err;
1367
1368  err_unlock:
1369         spin_unlock(&fiq->waitq.lock);
1370         return err;
1371 }
1372
1373 static int fuse_dev_open(struct inode *inode, struct file *file)
1374 {
1375         /*
1376          * The fuse device's file's private_data is used to hold
1377          * the fuse_conn(ection) when it is mounted, and is used to
1378          * keep track of whether the file has been mounted already.
1379          */
1380         file->private_data = NULL;
1381         return 0;
1382 }
1383
1384 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1385 {
1386         struct fuse_copy_state cs;
1387         struct file *file = iocb->ki_filp;
1388         struct fuse_dev *fud = fuse_get_dev(file);
1389
1390         if (!fud)
1391                 return -EPERM;
1392
1393         if (!iter_is_iovec(to))
1394                 return -EINVAL;
1395
1396         fuse_copy_init(&cs, 1, to);
1397
1398         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1399 }
1400
1401 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1402                                     struct pipe_inode_info *pipe,
1403                                     size_t len, unsigned int flags)
1404 {
1405         int ret;
1406         int page_nr = 0;
1407         int do_wakeup = 0;
1408         struct pipe_buffer *bufs;
1409         struct fuse_copy_state cs;
1410         struct fuse_dev *fud = fuse_get_dev(in);
1411
1412         if (!fud)
1413                 return -EPERM;
1414
1415         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1416         if (!bufs)
1417                 return -ENOMEM;
1418
1419         fuse_copy_init(&cs, 1, NULL);
1420         cs.pipebufs = bufs;
1421         cs.pipe = pipe;
1422         ret = fuse_dev_do_read(fud, in, &cs, len);
1423         if (ret < 0)
1424                 goto out;
1425
1426         ret = 0;
1427         pipe_lock(pipe);
1428
1429         if (!pipe->readers) {
1430                 send_sig(SIGPIPE, current, 0);
1431                 if (!ret)
1432                         ret = -EPIPE;
1433                 goto out_unlock;
1434         }
1435
1436         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1437                 ret = -EIO;
1438                 goto out_unlock;
1439         }
1440
1441         while (page_nr < cs.nr_segs) {
1442                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1443                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1444
1445                 buf->page = bufs[page_nr].page;
1446                 buf->offset = bufs[page_nr].offset;
1447                 buf->len = bufs[page_nr].len;
1448                 /*
1449                  * Need to be careful about this.  Having buf->ops in module
1450                  * code can Oops if the buffer persists after module unload.
1451                  */
1452                 buf->ops = &nosteal_pipe_buf_ops;
1453
1454                 pipe->nrbufs++;
1455                 page_nr++;
1456                 ret += buf->len;
1457
1458                 if (pipe->files)
1459                         do_wakeup = 1;
1460         }
1461
1462 out_unlock:
1463         pipe_unlock(pipe);
1464
1465         if (do_wakeup) {
1466                 smp_mb();
1467                 if (waitqueue_active(&pipe->wait))
1468                         wake_up_interruptible(&pipe->wait);
1469                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1470         }
1471
1472 out:
1473         for (; page_nr < cs.nr_segs; page_nr++)
1474                 page_cache_release(bufs[page_nr].page);
1475
1476         kfree(bufs);
1477         return ret;
1478 }
1479
1480 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1481                             struct fuse_copy_state *cs)
1482 {
1483         struct fuse_notify_poll_wakeup_out outarg;
1484         int err = -EINVAL;
1485
1486         if (size != sizeof(outarg))
1487                 goto err;
1488
1489         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1490         if (err)
1491                 goto err;
1492
1493         fuse_copy_finish(cs);
1494         return fuse_notify_poll_wakeup(fc, &outarg);
1495
1496 err:
1497         fuse_copy_finish(cs);
1498         return err;
1499 }
1500
1501 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1502                                    struct fuse_copy_state *cs)
1503 {
1504         struct fuse_notify_inval_inode_out outarg;
1505         int err = -EINVAL;
1506
1507         if (size != sizeof(outarg))
1508                 goto err;
1509
1510         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1511         if (err)
1512                 goto err;
1513         fuse_copy_finish(cs);
1514
1515         down_read(&fc->killsb);
1516         err = -ENOENT;
1517         if (fc->sb) {
1518                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1519                                                outarg.off, outarg.len);
1520         }
1521         up_read(&fc->killsb);
1522         return err;
1523
1524 err:
1525         fuse_copy_finish(cs);
1526         return err;
1527 }
1528
1529 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1530                                    struct fuse_copy_state *cs)
1531 {
1532         struct fuse_notify_inval_entry_out outarg;
1533         int err = -ENOMEM;
1534         char *buf;
1535         struct qstr name;
1536
1537         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1538         if (!buf)
1539                 goto err;
1540
1541         err = -EINVAL;
1542         if (size < sizeof(outarg))
1543                 goto err;
1544
1545         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1546         if (err)
1547                 goto err;
1548
1549         err = -ENAMETOOLONG;
1550         if (outarg.namelen > FUSE_NAME_MAX)
1551                 goto err;
1552
1553         err = -EINVAL;
1554         if (size != sizeof(outarg) + outarg.namelen + 1)
1555                 goto err;
1556
1557         name.name = buf;
1558         name.len = outarg.namelen;
1559         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1560         if (err)
1561                 goto err;
1562         fuse_copy_finish(cs);
1563         buf[outarg.namelen] = 0;
1564         name.hash = full_name_hash(name.name, name.len);
1565
1566         down_read(&fc->killsb);
1567         err = -ENOENT;
1568         if (fc->sb)
1569                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1570         up_read(&fc->killsb);
1571         kfree(buf);
1572         return err;
1573
1574 err:
1575         kfree(buf);
1576         fuse_copy_finish(cs);
1577         return err;
1578 }
1579
1580 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1581                               struct fuse_copy_state *cs)
1582 {
1583         struct fuse_notify_delete_out outarg;
1584         int err = -ENOMEM;
1585         char *buf;
1586         struct qstr name;
1587
1588         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1589         if (!buf)
1590                 goto err;
1591
1592         err = -EINVAL;
1593         if (size < sizeof(outarg))
1594                 goto err;
1595
1596         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1597         if (err)
1598                 goto err;
1599
1600         err = -ENAMETOOLONG;
1601         if (outarg.namelen > FUSE_NAME_MAX)
1602                 goto err;
1603
1604         err = -EINVAL;
1605         if (size != sizeof(outarg) + outarg.namelen + 1)
1606                 goto err;
1607
1608         name.name = buf;
1609         name.len = outarg.namelen;
1610         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1611         if (err)
1612                 goto err;
1613         fuse_copy_finish(cs);
1614         buf[outarg.namelen] = 0;
1615         name.hash = full_name_hash(name.name, name.len);
1616
1617         down_read(&fc->killsb);
1618         err = -ENOENT;
1619         if (fc->sb)
1620                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1621                                                outarg.child, &name);
1622         up_read(&fc->killsb);
1623         kfree(buf);
1624         return err;
1625
1626 err:
1627         kfree(buf);
1628         fuse_copy_finish(cs);
1629         return err;
1630 }
1631
1632 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1633                              struct fuse_copy_state *cs)
1634 {
1635         struct fuse_notify_store_out outarg;
1636         struct inode *inode;
1637         struct address_space *mapping;
1638         u64 nodeid;
1639         int err;
1640         pgoff_t index;
1641         unsigned int offset;
1642         unsigned int num;
1643         loff_t file_size;
1644         loff_t end;
1645
1646         err = -EINVAL;
1647         if (size < sizeof(outarg))
1648                 goto out_finish;
1649
1650         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1651         if (err)
1652                 goto out_finish;
1653
1654         err = -EINVAL;
1655         if (size - sizeof(outarg) != outarg.size)
1656                 goto out_finish;
1657
1658         nodeid = outarg.nodeid;
1659
1660         down_read(&fc->killsb);
1661
1662         err = -ENOENT;
1663         if (!fc->sb)
1664                 goto out_up_killsb;
1665
1666         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1667         if (!inode)
1668                 goto out_up_killsb;
1669
1670         mapping = inode->i_mapping;
1671         index = outarg.offset >> PAGE_CACHE_SHIFT;
1672         offset = outarg.offset & ~PAGE_CACHE_MASK;
1673         file_size = i_size_read(inode);
1674         end = outarg.offset + outarg.size;
1675         if (end > file_size) {
1676                 file_size = end;
1677                 fuse_write_update_size(inode, file_size);
1678         }
1679
1680         num = outarg.size;
1681         while (num) {
1682                 struct page *page;
1683                 unsigned int this_num;
1684
1685                 err = -ENOMEM;
1686                 page = find_or_create_page(mapping, index,
1687                                            mapping_gfp_mask(mapping));
1688                 if (!page)
1689                         goto out_iput;
1690
1691                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1692                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1693                 if (!err && offset == 0 &&
1694                     (this_num == PAGE_CACHE_SIZE || file_size == end))
1695                         SetPageUptodate(page);
1696                 unlock_page(page);
1697                 page_cache_release(page);
1698
1699                 if (err)
1700                         goto out_iput;
1701
1702                 num -= this_num;
1703                 offset = 0;
1704                 index++;
1705         }
1706
1707         err = 0;
1708
1709 out_iput:
1710         iput(inode);
1711 out_up_killsb:
1712         up_read(&fc->killsb);
1713 out_finish:
1714         fuse_copy_finish(cs);
1715         return err;
1716 }
1717
1718 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1719 {
1720         release_pages(req->pages, req->num_pages, false);
1721 }
1722
1723 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1724                          struct fuse_notify_retrieve_out *outarg)
1725 {
1726         int err;
1727         struct address_space *mapping = inode->i_mapping;
1728         struct fuse_req *req;
1729         pgoff_t index;
1730         loff_t file_size;
1731         unsigned int num;
1732         unsigned int offset;
1733         size_t total_len = 0;
1734         int num_pages;
1735
1736         offset = outarg->offset & ~PAGE_CACHE_MASK;
1737         file_size = i_size_read(inode);
1738
1739         num = min(outarg->size, fc->max_write);
1740         if (outarg->offset > file_size)
1741                 num = 0;
1742         else if (outarg->offset + num > file_size)
1743                 num = file_size - outarg->offset;
1744
1745         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1746         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1747
1748         req = fuse_get_req(fc, num_pages);
1749         if (IS_ERR(req))
1750                 return PTR_ERR(req);
1751
1752         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1753         req->in.h.nodeid = outarg->nodeid;
1754         req->in.numargs = 2;
1755         req->in.argpages = 1;
1756         req->end = fuse_retrieve_end;
1757
1758         index = outarg->offset >> PAGE_CACHE_SHIFT;
1759
1760         while (num && req->num_pages < num_pages) {
1761                 struct page *page;
1762                 unsigned int this_num;
1763
1764                 page = find_get_page(mapping, index);
1765                 if (!page)
1766                         break;
1767
1768                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1769                 req->pages[req->num_pages] = page;
1770                 req->page_descs[req->num_pages].offset = offset;
1771                 req->page_descs[req->num_pages].length = this_num;
1772                 req->num_pages++;
1773
1774                 offset = 0;
1775                 num -= this_num;
1776                 total_len += this_num;
1777                 index++;
1778         }
1779         req->misc.retrieve_in.offset = outarg->offset;
1780         req->misc.retrieve_in.size = total_len;
1781         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1782         req->in.args[0].value = &req->misc.retrieve_in;
1783         req->in.args[1].size = total_len;
1784
1785         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1786         if (err) {
1787                 fuse_retrieve_end(fc, req);
1788                 fuse_put_request(fc, req);
1789         }
1790
1791         return err;
1792 }
1793
1794 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1795                                 struct fuse_copy_state *cs)
1796 {
1797         struct fuse_notify_retrieve_out outarg;
1798         struct inode *inode;
1799         int err;
1800
1801         err = -EINVAL;
1802         if (size != sizeof(outarg))
1803                 goto copy_finish;
1804
1805         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1806         if (err)
1807                 goto copy_finish;
1808
1809         fuse_copy_finish(cs);
1810
1811         down_read(&fc->killsb);
1812         err = -ENOENT;
1813         if (fc->sb) {
1814                 u64 nodeid = outarg.nodeid;
1815
1816                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1817                 if (inode) {
1818                         err = fuse_retrieve(fc, inode, &outarg);
1819                         iput(inode);
1820                 }
1821         }
1822         up_read(&fc->killsb);
1823
1824         return err;
1825
1826 copy_finish:
1827         fuse_copy_finish(cs);
1828         return err;
1829 }
1830
1831 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1832                        unsigned int size, struct fuse_copy_state *cs)
1833 {
1834         /* Don't try to move pages (yet) */
1835         cs->move_pages = 0;
1836
1837         switch (code) {
1838         case FUSE_NOTIFY_POLL:
1839                 return fuse_notify_poll(fc, size, cs);
1840
1841         case FUSE_NOTIFY_INVAL_INODE:
1842                 return fuse_notify_inval_inode(fc, size, cs);
1843
1844         case FUSE_NOTIFY_INVAL_ENTRY:
1845                 return fuse_notify_inval_entry(fc, size, cs);
1846
1847         case FUSE_NOTIFY_STORE:
1848                 return fuse_notify_store(fc, size, cs);
1849
1850         case FUSE_NOTIFY_RETRIEVE:
1851                 return fuse_notify_retrieve(fc, size, cs);
1852
1853         case FUSE_NOTIFY_DELETE:
1854                 return fuse_notify_delete(fc, size, cs);
1855
1856         default:
1857                 fuse_copy_finish(cs);
1858                 return -EINVAL;
1859         }
1860 }
1861
1862 /* Look up request on processing list by unique ID */
1863 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1864 {
1865         struct fuse_req *req;
1866
1867         list_for_each_entry(req, &fpq->processing, list) {
1868                 if (req->in.h.unique == unique || req->intr_unique == unique)
1869                         return req;
1870         }
1871         return NULL;
1872 }
1873
1874 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1875                          unsigned nbytes)
1876 {
1877         unsigned reqsize = sizeof(struct fuse_out_header);
1878
1879         if (out->h.error)
1880                 return nbytes != reqsize ? -EINVAL : 0;
1881
1882         reqsize += len_args(out->numargs, out->args);
1883
1884         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1885                 return -EINVAL;
1886         else if (reqsize > nbytes) {
1887                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1888                 unsigned diffsize = reqsize - nbytes;
1889                 if (diffsize > lastarg->size)
1890                         return -EINVAL;
1891                 lastarg->size -= diffsize;
1892         }
1893         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1894                               out->page_zeroing);
1895 }
1896
1897 /*
1898  * Write a single reply to a request.  First the header is copied from
1899  * the write buffer.  The request is then searched on the processing
1900  * list by the unique ID found in the header.  If found, then remove
1901  * it from the list and copy the rest of the buffer to the request.
1902  * The request is finished by calling request_end()
1903  */
1904 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1905                                  struct fuse_copy_state *cs, size_t nbytes)
1906 {
1907         int err;
1908         struct fuse_conn *fc = fud->fc;
1909         struct fuse_pqueue *fpq = &fud->pq;
1910         struct fuse_req *req;
1911         struct fuse_out_header oh;
1912
1913         if (nbytes < sizeof(struct fuse_out_header))
1914                 return -EINVAL;
1915
1916         err = fuse_copy_one(cs, &oh, sizeof(oh));
1917         if (err)
1918                 goto err_finish;
1919
1920         err = -EINVAL;
1921         if (oh.len != nbytes)
1922                 goto err_finish;
1923
1924         /*
1925          * Zero oh.unique indicates unsolicited notification message
1926          * and error contains notification code.
1927          */
1928         if (!oh.unique) {
1929                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1930                 return err ? err : nbytes;
1931         }
1932
1933         err = -EINVAL;
1934         if (oh.error <= -512 || oh.error > 0)
1935                 goto err_finish;
1936
1937         spin_lock(&fpq->lock);
1938         err = -ENOENT;
1939         if (!fpq->connected)
1940                 goto err_unlock_pq;
1941
1942         req = request_find(fpq, oh.unique);
1943         if (!req)
1944                 goto err_unlock_pq;
1945
1946         /* Is it an interrupt reply? */
1947         if (req->intr_unique == oh.unique) {
1948                 __fuse_get_request(req);
1949                 spin_unlock(&fpq->lock);
1950
1951                 err = -EINVAL;
1952                 if (nbytes != sizeof(struct fuse_out_header)) {
1953                         fuse_put_request(fc, req);
1954                         goto err_finish;
1955                 }
1956
1957                 if (oh.error == -ENOSYS)
1958                         fc->no_interrupt = 1;
1959                 else if (oh.error == -EAGAIN)
1960                         queue_interrupt(&fc->iq, req);
1961                 fuse_put_request(fc, req);
1962
1963                 fuse_copy_finish(cs);
1964                 return nbytes;
1965         }
1966
1967         clear_bit(FR_SENT, &req->flags);
1968         list_move(&req->list, &fpq->io);
1969         req->out.h = oh;
1970         set_bit(FR_LOCKED, &req->flags);
1971         spin_unlock(&fpq->lock);
1972         cs->req = req;
1973         if (!req->out.page_replace)
1974                 cs->move_pages = 0;
1975
1976         err = copy_out_args(cs, &req->out, nbytes);
1977         fuse_copy_finish(cs);
1978
1979         spin_lock(&fpq->lock);
1980         clear_bit(FR_LOCKED, &req->flags);
1981         if (!fpq->connected)
1982                 err = -ENOENT;
1983         else if (err)
1984                 req->out.h.error = -EIO;
1985         if (!test_bit(FR_PRIVATE, &req->flags))
1986                 list_del_init(&req->list);
1987         spin_unlock(&fpq->lock);
1988
1989         request_end(fc, req);
1990
1991         return err ? err : nbytes;
1992
1993  err_unlock_pq:
1994         spin_unlock(&fpq->lock);
1995  err_finish:
1996         fuse_copy_finish(cs);
1997         return err;
1998 }
1999
2000 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
2001 {
2002         struct fuse_copy_state cs;
2003         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
2004
2005         if (!fud)
2006                 return -EPERM;
2007
2008         if (!iter_is_iovec(from))
2009                 return -EINVAL;
2010
2011         fuse_copy_init(&cs, 0, from);
2012
2013         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
2014 }
2015
2016 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
2017                                      struct file *out, loff_t *ppos,
2018                                      size_t len, unsigned int flags)
2019 {
2020         unsigned nbuf;
2021         unsigned idx;
2022         struct pipe_buffer *bufs;
2023         struct fuse_copy_state cs;
2024         struct fuse_dev *fud;
2025         size_t rem;
2026         ssize_t ret;
2027
2028         fud = fuse_get_dev(out);
2029         if (!fud)
2030                 return -EPERM;
2031
2032         pipe_lock(pipe);
2033
2034         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
2035         if (!bufs) {
2036                 pipe_unlock(pipe);
2037                 return -ENOMEM;
2038         }
2039
2040         nbuf = 0;
2041         rem = 0;
2042         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2043                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2044
2045         ret = -EINVAL;
2046         if (rem < len)
2047                 goto out_free;
2048
2049         rem = len;
2050         while (rem) {
2051                 struct pipe_buffer *ibuf;
2052                 struct pipe_buffer *obuf;
2053
2054                 BUG_ON(nbuf >= pipe->buffers);
2055                 BUG_ON(!pipe->nrbufs);
2056                 ibuf = &pipe->bufs[pipe->curbuf];
2057                 obuf = &bufs[nbuf];
2058
2059                 if (rem >= ibuf->len) {
2060                         *obuf = *ibuf;
2061                         ibuf->ops = NULL;
2062                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2063                         pipe->nrbufs--;
2064                 } else {
2065                         if (!pipe_buf_get(pipe, ibuf))
2066                                 goto out_free;
2067
2068                         *obuf = *ibuf;
2069                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2070                         obuf->len = rem;
2071                         ibuf->offset += obuf->len;
2072                         ibuf->len -= obuf->len;
2073                 }
2074                 nbuf++;
2075                 rem -= obuf->len;
2076         }
2077         pipe_unlock(pipe);
2078
2079         fuse_copy_init(&cs, 0, NULL);
2080         cs.pipebufs = bufs;
2081         cs.nr_segs = nbuf;
2082         cs.pipe = pipe;
2083
2084         if (flags & SPLICE_F_MOVE)
2085                 cs.move_pages = 1;
2086
2087         ret = fuse_dev_do_write(fud, &cs, len);
2088
2089         pipe_lock(pipe);
2090 out_free:
2091         for (idx = 0; idx < nbuf; idx++) {
2092                 struct pipe_buffer *buf = &bufs[idx];
2093                 buf->ops->release(pipe, buf);
2094         }
2095         pipe_unlock(pipe);
2096
2097         kfree(bufs);
2098         return ret;
2099 }
2100
2101 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2102 {
2103         unsigned mask = POLLOUT | POLLWRNORM;
2104         struct fuse_iqueue *fiq;
2105         struct fuse_dev *fud = fuse_get_dev(file);
2106
2107         if (!fud)
2108                 return POLLERR;
2109
2110         fiq = &fud->fc->iq;
2111         poll_wait(file, &fiq->waitq, wait);
2112
2113         spin_lock(&fiq->waitq.lock);
2114         if (!fiq->connected)
2115                 mask = POLLERR;
2116         else if (request_pending(fiq))
2117                 mask |= POLLIN | POLLRDNORM;
2118         spin_unlock(&fiq->waitq.lock);
2119
2120         return mask;
2121 }
2122
2123 /*
2124  * Abort all requests on the given list (pending or processing)
2125  *
2126  * This function releases and reacquires fc->lock
2127  */
2128 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2129 {
2130         while (!list_empty(head)) {
2131                 struct fuse_req *req;
2132                 req = list_entry(head->next, struct fuse_req, list);
2133                 req->out.h.error = -ECONNABORTED;
2134                 clear_bit(FR_SENT, &req->flags);
2135                 list_del_init(&req->list);
2136                 request_end(fc, req);
2137         }
2138 }
2139
2140 static void end_polls(struct fuse_conn *fc)
2141 {
2142         struct rb_node *p;
2143
2144         p = rb_first(&fc->polled_files);
2145
2146         while (p) {
2147                 struct fuse_file *ff;
2148                 ff = rb_entry(p, struct fuse_file, polled_node);
2149                 wake_up_interruptible_all(&ff->poll_wait);
2150
2151                 p = rb_next(p);
2152         }
2153 }
2154
2155 /*
2156  * Abort all requests.
2157  *
2158  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2159  * filesystem.
2160  *
2161  * The same effect is usually achievable through killing the filesystem daemon
2162  * and all users of the filesystem.  The exception is the combination of an
2163  * asynchronous request and the tricky deadlock (see
2164  * Documentation/filesystems/fuse.txt).
2165  *
2166  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2167  * requests, they should be finished off immediately.  Locked requests will be
2168  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2169  * requests.  It is possible that some request will finish before we can.  This
2170  * is OK, the request will in that case be removed from the list before we touch
2171  * it.
2172  */
2173 void fuse_abort_conn(struct fuse_conn *fc)
2174 {
2175         struct fuse_iqueue *fiq = &fc->iq;
2176
2177         spin_lock(&fc->lock);
2178         if (fc->connected) {
2179                 struct fuse_dev *fud;
2180                 struct fuse_req *req, *next;
2181                 LIST_HEAD(to_end1);
2182                 LIST_HEAD(to_end2);
2183
2184                 fc->connected = 0;
2185                 fc->blocked = 0;
2186                 fuse_set_initialized(fc);
2187                 list_for_each_entry(fud, &fc->devices, entry) {
2188                         struct fuse_pqueue *fpq = &fud->pq;
2189
2190                         spin_lock(&fpq->lock);
2191                         fpq->connected = 0;
2192                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2193                                 req->out.h.error = -ECONNABORTED;
2194                                 spin_lock(&req->waitq.lock);
2195                                 set_bit(FR_ABORTED, &req->flags);
2196                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2197                                         set_bit(FR_PRIVATE, &req->flags);
2198                                         __fuse_get_request(req);
2199                                         list_move(&req->list, &to_end1);
2200                                 }
2201                                 spin_unlock(&req->waitq.lock);
2202                         }
2203                         list_splice_init(&fpq->processing, &to_end2);
2204                         spin_unlock(&fpq->lock);
2205                 }
2206                 fc->max_background = UINT_MAX;
2207                 flush_bg_queue(fc);
2208
2209                 spin_lock(&fiq->waitq.lock);
2210                 fiq->connected = 0;
2211                 list_splice_init(&fiq->pending, &to_end2);
2212                 list_for_each_entry(req, &to_end2, list)
2213                         clear_bit(FR_PENDING, &req->flags);
2214                 while (forget_pending(fiq))
2215                         kfree(dequeue_forget(fiq, 1, NULL));
2216                 wake_up_all_locked(&fiq->waitq);
2217                 spin_unlock(&fiq->waitq.lock);
2218                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2219                 end_polls(fc);
2220                 wake_up_all(&fc->blocked_waitq);
2221                 spin_unlock(&fc->lock);
2222
2223                 while (!list_empty(&to_end1)) {
2224                         req = list_first_entry(&to_end1, struct fuse_req, list);
2225                         list_del_init(&req->list);
2226                         request_end(fc, req);
2227                 }
2228                 end_requests(fc, &to_end2);
2229         } else {
2230                 spin_unlock(&fc->lock);
2231         }
2232 }
2233 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2234
2235 void fuse_wait_aborted(struct fuse_conn *fc)
2236 {
2237         /* matches implicit memory barrier in fuse_drop_waiting() */
2238         smp_mb();
2239         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2240 }
2241
2242 int fuse_dev_release(struct inode *inode, struct file *file)
2243 {
2244         struct fuse_dev *fud = fuse_get_dev(file);
2245
2246         if (fud) {
2247                 struct fuse_conn *fc = fud->fc;
2248                 struct fuse_pqueue *fpq = &fud->pq;
2249                 LIST_HEAD(to_end);
2250
2251                 spin_lock(&fpq->lock);
2252                 WARN_ON(!list_empty(&fpq->io));
2253                 list_splice_init(&fpq->processing, &to_end);
2254                 spin_unlock(&fpq->lock);
2255
2256                 end_requests(fc, &to_end);
2257
2258                 /* Are we the last open device? */
2259                 if (atomic_dec_and_test(&fc->dev_count)) {
2260                         WARN_ON(fc->iq.fasync != NULL);
2261                         fuse_abort_conn(fc);
2262                 }
2263                 fuse_dev_free(fud);
2264         }
2265         return 0;
2266 }
2267 EXPORT_SYMBOL_GPL(fuse_dev_release);
2268
2269 static int fuse_dev_fasync(int fd, struct file *file, int on)
2270 {
2271         struct fuse_dev *fud = fuse_get_dev(file);
2272
2273         if (!fud)
2274                 return -EPERM;
2275
2276         /* No locking - fasync_helper does its own locking */
2277         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2278 }
2279
2280 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2281 {
2282         struct fuse_dev *fud;
2283
2284         if (new->private_data)
2285                 return -EINVAL;
2286
2287         fud = fuse_dev_alloc(fc);
2288         if (!fud)
2289                 return -ENOMEM;
2290
2291         new->private_data = fud;
2292         atomic_inc(&fc->dev_count);
2293
2294         return 0;
2295 }
2296
2297 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2298                            unsigned long arg)
2299 {
2300         int err = -ENOTTY;
2301
2302         if (cmd == FUSE_DEV_IOC_CLONE) {
2303                 int oldfd;
2304
2305                 err = -EFAULT;
2306                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2307                         struct file *old = fget(oldfd);
2308
2309                         err = -EINVAL;
2310                         if (old) {
2311                                 struct fuse_dev *fud = NULL;
2312
2313                                 /*
2314                                  * Check against file->f_op because CUSE
2315                                  * uses the same ioctl handler.
2316                                  */
2317                                 if (old->f_op == file->f_op &&
2318                                     old->f_cred->user_ns == file->f_cred->user_ns)
2319                                         fud = fuse_get_dev(old);
2320
2321                                 if (fud) {
2322                                         mutex_lock(&fuse_mutex);
2323                                         err = fuse_device_clone(fud->fc, file);
2324                                         mutex_unlock(&fuse_mutex);
2325                                 }
2326                                 fput(old);
2327                         }
2328                 }
2329         }
2330         return err;
2331 }
2332
2333 const struct file_operations fuse_dev_operations = {
2334         .owner          = THIS_MODULE,
2335         .open           = fuse_dev_open,
2336         .llseek         = no_llseek,
2337         .read_iter      = fuse_dev_read,
2338         .splice_read    = fuse_dev_splice_read,
2339         .write_iter     = fuse_dev_write,
2340         .splice_write   = fuse_dev_splice_write,
2341         .poll           = fuse_dev_poll,
2342         .release        = fuse_dev_release,
2343         .fasync         = fuse_dev_fasync,
2344         .unlocked_ioctl = fuse_dev_ioctl,
2345         .compat_ioctl   = fuse_dev_ioctl,
2346 };
2347 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2348
2349 static struct miscdevice fuse_miscdevice = {
2350         .minor = FUSE_MINOR,
2351         .name  = "fuse",
2352         .fops = &fuse_dev_operations,
2353 };
2354
2355 int __init fuse_dev_init(void)
2356 {
2357         int err = -ENOMEM;
2358         fuse_req_cachep = kmem_cache_create("fuse_request",
2359                                             sizeof(struct fuse_req),
2360                                             0, 0, NULL);
2361         if (!fuse_req_cachep)
2362                 goto out;
2363
2364         err = misc_register(&fuse_miscdevice);
2365         if (err)
2366                 goto out_cache_clean;
2367
2368         return 0;
2369
2370  out_cache_clean:
2371         kmem_cache_destroy(fuse_req_cachep);
2372  out:
2373         return err;
2374 }
2375
2376 void fuse_dev_cleanup(void)
2377 {
2378         misc_deregister(&fuse_miscdevice);
2379         kmem_cache_destroy(fuse_req_cachep);
2380 }