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