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 struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 147                                        bool for_background)
 148 {
 149         struct fuse_req *req;
 150         int err;
 151         atomic_inc(&fc->num_waiting);
 152
 153         if (fuse_block_alloc(fc, for_background)) {
 154                 sigset_t oldset;
 155                 int intr;
 156
 157                 block_sigs(&oldset);
 158                 intr = wait_event_interruptible_exclusive(fc->blocked_waitq,
 159                                 !fuse_block_alloc(fc, for_background));
 160                 restore_sigs(&oldset);
 161                 err = -EINTR;
 162                 if (intr)
 163                         goto out;
 164         }
 165         /* Matches smp_wmb() in fuse_set_initialized() */
 166         smp_rmb();
 167
 168         err = -ENOTCONN;
 169         if (!fc->connected)
 170                 goto out;
 171
 172         err = -ECONNREFUSED;
 173         if (fc->conn_error)
 174                 goto out;
 175
 176         req = fuse_request_alloc(npages);
 177         err = -ENOMEM;
 178         if (!req) {
 179                 if (for_background)
 180                         wake_up(&fc->blocked_waitq);
 181                 goto out;
 182         }
 183
 184         fuse_req_init_context(req);
 185         __set_bit(FR_WAITING, &req->flags);
 186         if (for_background)
 187                 __set_bit(FR_BACKGROUND, &req->flags);
 188
 189         return req;
 190
 191  out:
 192         atomic_dec(&fc->num_waiting);
 193         return ERR_PTR(err);
 194 }
 195
 196 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 197 {
 198         return __fuse_get_req(fc, npages, false);
 199 }
 200 EXPORT_SYMBOL_GPL(fuse_get_req);
 201
 202 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 203                                              unsigned npages)
 204 {
 205         return __fuse_get_req(fc, npages, true);
 206 }
 207 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 208
 209 /*
 210  * Return request in fuse_file->reserved_req.  However that may
 211  * currently be in use.  If that is the case, wait for it to become
 212  * available.
 213  */
 214 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 215                                          struct file *file)
 216 {
 217         struct fuse_req *req = NULL;
 218         struct fuse_file *ff = file->private_data;
 219
 220         do {
 221                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 222                 spin_lock(&fc->lock);
 223                 if (ff->reserved_req) {
 224                         req = ff->reserved_req;
 225                         ff->reserved_req = NULL;
 226                         req->stolen_file = get_file(file);
 227                 }
 228                 spin_unlock(&fc->lock);
 229         } while (!req);
 230
 231         return req;
 232 }
 233
 234 /*
 235  * Put stolen request back into fuse_file->reserved_req
 236  */
 237 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 238 {
 239         struct file *file = req->stolen_file;
 240         struct fuse_file *ff = file->private_data;
 241
 242         spin_lock(&fc->lock);
 243         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 244         BUG_ON(ff->reserved_req);
 245         ff->reserved_req = req;
 246         wake_up_all(&fc->reserved_req_waitq);
 247         spin_unlock(&fc->lock);
 248         fput(file);
 249 }
 250
 251 /*
 252  * Gets a requests for a file operation, always succeeds
 253  *
 254  * This is used for sending the FLUSH request, which must get to
 255  * userspace, due to POSIX locks which may need to be unlocked.
 256  *
 257  * If allocation fails due to OOM, use the reserved request in
 258  * fuse_file.
 259  *
 260  * This is very unlikely to deadlock accidentally, since the
 261  * filesystem should not have it's own file open.  If deadlock is
 262  * intentional, it can still be broken by "aborting" the filesystem.
 263  */
 264 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 265                                              struct file *file)
 266 {
 267         struct fuse_req *req;
 268
 269         atomic_inc(&fc->num_waiting);
 270         wait_event(fc->blocked_waitq, fc->initialized);
 271         /* Matches smp_wmb() in fuse_set_initialized() */
 272         smp_rmb();
 273         req = fuse_request_alloc(0);
 274         if (!req)
 275                 req = get_reserved_req(fc, file);
 276
 277         fuse_req_init_context(req);
 278         __set_bit(FR_WAITING, &req->flags);
 279         __clear_bit(FR_BACKGROUND, &req->flags);
 280         return req;
 281 }
 282
 283 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 284 {
 285         if (atomic_dec_and_test(&req->count)) {
 286                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 287                         /*
 288                          * We get here in the unlikely case that a background
 289                          * request was allocated but not sent
 290                          */
 291                         spin_lock(&fc->lock);
 292                         if (!fc->blocked)
 293                                 wake_up(&fc->blocked_waitq);
 294                         spin_unlock(&fc->lock);
 295                 }
 296
 297                 if (test_bit(FR_WAITING, &req->flags)) {
 298                         __clear_bit(FR_WAITING, &req->flags);
 299                         atomic_dec(&fc->num_waiting);
 300                 }
 301
 302                 if (req->stolen_file)
 303                         put_reserved_req(fc, req);
 304                 else
 305                         fuse_request_free(req);
 306         }
 307 }
 308 EXPORT_SYMBOL_GPL(fuse_put_request);
 309
 310 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 311 {
 312         unsigned nbytes = 0;
 313         unsigned i;
 314
 315         for (i = 0; i < numargs; i++)
 316                 nbytes += args[i].size;
 317
 318         return nbytes;
 319 }
 320
 321 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 322 {
 323         return ++fiq->reqctr;
 324 }
 325
 326 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 327 {
 328         req->in.h.len = sizeof(struct fuse_in_header) +
 329                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 330         list_add_tail(&req->list, &fiq->pending);
 331         wake_up_locked(&fiq->waitq);
 332         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 333 }
 334
 335 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 336                        u64 nodeid, u64 nlookup)
 337 {
 338         struct fuse_iqueue *fiq = &fc->iq;
 339
 340         forget->forget_one.nodeid = nodeid;
 341         forget->forget_one.nlookup = nlookup;
 342
 343         spin_lock(&fiq->waitq.lock);
 344         if (fiq->connected) {
 345                 fiq->forget_list_tail->next = forget;
 346                 fiq->forget_list_tail = forget;
 347                 wake_up_locked(&fiq->waitq);
 348                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 349         } else {
 350                 kfree(forget);
 351         }
 352         spin_unlock(&fiq->waitq.lock);
 353 }
 354
 355 static void flush_bg_queue(struct fuse_conn *fc)
 356 {
 357         while (fc->active_background < fc->max_background &&
 358                !list_empty(&fc->bg_queue)) {
 359                 struct fuse_req *req;
 360                 struct fuse_iqueue *fiq = &fc->iq;
 361
 362                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 363                 list_del(&req->list);
 364                 fc->active_background++;
 365                 spin_lock(&fiq->waitq.lock);
 366                 req->in.h.unique = fuse_get_unique(fiq);
 367                 queue_request(fiq, req);
 368                 spin_unlock(&fiq->waitq.lock);
 369         }
 370 }
 371
 372 /*
 373  * This function is called when a request is finished.  Either a reply
 374  * has arrived or it was aborted (and not yet sent) or some error
 375  * occurred during communication with userspace, or the device file
 376  * was closed.  The requester thread is woken up (if still waiting),
 377  * the 'end' callback is called if given, else the reference to the
 378  * request is released
 379  */
 380 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 381 {
 382         struct fuse_iqueue *fiq = &fc->iq;
 383
 384         if (test_and_set_bit(FR_FINISHED, &req->flags))
 385                 return;
 386
 387         spin_lock(&fiq->waitq.lock);
 388         list_del_init(&req->intr_entry);
 389         spin_unlock(&fiq->waitq.lock);
 390         WARN_ON(test_bit(FR_PENDING, &req->flags));
 391         WARN_ON(test_bit(FR_SENT, &req->flags));
 392         if (test_bit(FR_BACKGROUND, &req->flags)) {
 393                 spin_lock(&fc->lock);
 394                 clear_bit(FR_BACKGROUND, &req->flags);
 395                 if (fc->num_background == fc->max_background)
 396                         fc->blocked = 0;
 397
 398                 /* Wake up next waiter, if any */
 399                 if (!fc->blocked && waitqueue_active(&fc->blocked_waitq))
 400                         wake_up(&fc->blocked_waitq);
 401
 402                 if (fc->num_background == fc->congestion_threshold &&
 403                     fc->connected && fc->bdi_initialized) {
 404                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 405                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 406                 }
 407                 fc->num_background--;
 408                 fc->active_background--;
 409                 flush_bg_queue(fc);
 410                 spin_unlock(&fc->lock);
 411         }
 412         wake_up(&req->waitq);
 413         if (req->end)
 414                 req->end(fc, req);
 415         fuse_put_request(fc, req);
 416 }
 417
 418 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 419 {
 420         spin_lock(&fiq->waitq.lock);
 421         if (list_empty(&req->intr_entry)) {
 422                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 423                 wake_up_locked(&fiq->waitq);
 424         }
 425         spin_unlock(&fiq->waitq.lock);
 426         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 427 }
 428
 429 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 430 {
 431         struct fuse_iqueue *fiq = &fc->iq;
 432         int err;
 433
 434         if (!fc->no_interrupt) {
 435                 /* Any signal may interrupt this */
 436                 err = wait_event_interruptible(req->waitq,
 437                                         test_bit(FR_FINISHED, &req->flags));
 438                 if (!err)
 439                         return;
 440
 441                 set_bit(FR_INTERRUPTED, &req->flags);
 442                 /* matches barrier in fuse_dev_do_read() */
 443                 smp_mb__after_atomic();
 444                 if (test_bit(FR_SENT, &req->flags))
 445                         queue_interrupt(fiq, req);
 446         }
 447
 448         if (!test_bit(FR_FORCE, &req->flags)) {
 449                 sigset_t oldset;
 450
 451                 /* Only fatal signals may interrupt this */
 452                 block_sigs(&oldset);
 453                 err = wait_event_interruptible(req->waitq,
 454                                         test_bit(FR_FINISHED, &req->flags));
 455                 restore_sigs(&oldset);
 456
 457                 if (!err)
 458                         return;
 459
 460                 spin_lock(&fiq->waitq.lock);
 461                 /* Request is not yet in userspace, bail out */
 462                 if (test_bit(FR_PENDING, &req->flags)) {
 463                         list_del(&req->list);
 464                         spin_unlock(&fiq->waitq.lock);
 465                         __fuse_put_request(req);
 466                         req->out.h.error = -EINTR;
 467                         return;
 468                 }
 469                 spin_unlock(&fiq->waitq.lock);
 470         }
 471
 472         /*
 473          * Either request is already in userspace, or it was forced.
 474          * Wait it out.
 475          */
 476         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 477 }
 478
 479 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 480 {
 481         struct fuse_iqueue *fiq = &fc->iq;
 482
 483         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 484         spin_lock(&fiq->waitq.lock);
 485         if (!fiq->connected) {
 486                 spin_unlock(&fiq->waitq.lock);
 487                 req->out.h.error = -ENOTCONN;
 488         } else {
 489                 req->in.h.unique = fuse_get_unique(fiq);
 490                 queue_request(fiq, req);
 491                 /* acquire extra reference, since request is still needed
 492                    after request_end() */
 493                 __fuse_get_request(req);
 494                 spin_unlock(&fiq->waitq.lock);
 495
 496                 request_wait_answer(fc, req);
 497                 /* Pairs with smp_wmb() in request_end() */
 498                 smp_rmb();
 499         }
 500 }
 501
 502 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 503 {
 504         __set_bit(FR_ISREPLY, &req->flags);
 505         if (!test_bit(FR_WAITING, &req->flags)) {
 506                 __set_bit(FR_WAITING, &req->flags);
 507                 atomic_inc(&fc->num_waiting);
 508         }
 509         __fuse_request_send(fc, req);
 510 }
 511 EXPORT_SYMBOL_GPL(fuse_request_send);
 512
 513 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 514 {
 515         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 516                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 517
 518         if (fc->minor < 9) {
 519                 switch (args->in.h.opcode) {
 520                 case FUSE_LOOKUP:
 521                 case FUSE_CREATE:
 522                 case FUSE_MKNOD:
 523                 case FUSE_MKDIR:
 524                 case FUSE_SYMLINK:
 525                 case FUSE_LINK:
 526                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 527                         break;
 528                 case FUSE_GETATTR:
 529                 case FUSE_SETATTR:
 530                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 531                         break;
 532                 }
 533         }
 534         if (fc->minor < 12) {
 535                 switch (args->in.h.opcode) {
 536                 case FUSE_CREATE:
 537                         args->in.args[0].size = sizeof(struct fuse_open_in);
 538                         break;
 539                 case FUSE_MKNOD:
 540                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 541                         break;
 542                 }
 543         }
 544 }
 545
 546 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 547 {
 548         struct fuse_req *req;
 549         ssize_t ret;
 550
 551         req = fuse_get_req(fc, 0);
 552         if (IS_ERR(req))
 553                 return PTR_ERR(req);
 554
 555         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 556         fuse_adjust_compat(fc, args);
 557
 558         req->in.h.opcode = args->in.h.opcode;
 559         req->in.h.nodeid = args->in.h.nodeid;
 560         req->in.numargs = args->in.numargs;
 561         memcpy(req->in.args, args->in.args,
 562                args->in.numargs * sizeof(struct fuse_in_arg));
 563         req->out.argvar = args->out.argvar;
 564         req->out.numargs = args->out.numargs;
 565         memcpy(req->out.args, args->out.args,
 566                args->out.numargs * sizeof(struct fuse_arg));
 567         fuse_request_send(fc, req);
 568         ret = req->out.h.error;
 569         if (!ret && args->out.argvar) {
 570                 BUG_ON(args->out.numargs != 1);
 571                 ret = req->out.args[0].size;
 572         }
 573         fuse_put_request(fc, req);
 574
 575         return ret;
 576 }
 577
 578 /*
 579  * Called under fc->lock
 580  *
 581  * fc->connected must have been checked previously
 582  */
 583 void fuse_request_send_background_locked(struct fuse_conn *fc,
 584                                          struct fuse_req *req)
 585 {
 586         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 587         if (!test_bit(FR_WAITING, &req->flags)) {
 588                 __set_bit(FR_WAITING, &req->flags);
 589                 atomic_inc(&fc->num_waiting);
 590         }
 591         __set_bit(FR_ISREPLY, &req->flags);
 592         fc->num_background++;
 593         if (fc->num_background == fc->max_background)
 594                 fc->blocked = 1;
 595         if (fc->num_background == fc->congestion_threshold &&
 596             fc->bdi_initialized) {
 597                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 598                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 599         }
 600         list_add_tail(&req->list, &fc->bg_queue);
 601         flush_bg_queue(fc);
 602 }
 603
 604 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 605 {
 606         BUG_ON(!req->end);
 607         spin_lock(&fc->lock);
 608         if (fc->connected) {
 609                 fuse_request_send_background_locked(fc, req);
 610                 spin_unlock(&fc->lock);
 611         } else {
 612                 spin_unlock(&fc->lock);
 613                 req->out.h.error = -ENOTCONN;
 614                 req->end(fc, req);
 615                 fuse_put_request(fc, req);
 616         }
 617 }
 618 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 619
 620 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 621                                           struct fuse_req *req, u64 unique)
 622 {
 623         int err = -ENODEV;
 624         struct fuse_iqueue *fiq = &fc->iq;
 625
 626         __clear_bit(FR_ISREPLY, &req->flags);
 627         req->in.h.unique = unique;
 628         spin_lock(&fiq->waitq.lock);
 629         if (fiq->connected) {
 630                 queue_request(fiq, req);
 631                 err = 0;
 632         }
 633         spin_unlock(&fiq->waitq.lock);
 634
 635         return err;
 636 }
 637
 638 void fuse_force_forget(struct file *file, u64 nodeid)
 639 {
 640         struct inode *inode = file_inode(file);
 641         struct fuse_conn *fc = get_fuse_conn(inode);
 642         struct fuse_req *req;
 643         struct fuse_forget_in inarg;
 644
 645         memset(&inarg, 0, sizeof(inarg));
 646         inarg.nlookup = 1;
 647         req = fuse_get_req_nofail_nopages(fc, file);
 648         req->in.h.opcode = FUSE_FORGET;
 649         req->in.h.nodeid = nodeid;
 650         req->in.numargs = 1;
 651         req->in.args[0].size = sizeof(inarg);
 652         req->in.args[0].value = &inarg;
 653         __clear_bit(FR_ISREPLY, &req->flags);
 654         __fuse_request_send(fc, req);
 655         /* ignore errors */
 656         fuse_put_request(fc, req);
 657 }
 658
 659 /*
 660  * Lock the request.  Up to the next unlock_request() there mustn't be
 661  * anything that could cause a page-fault.  If the request was already
 662  * aborted bail out.
 663  */
 664 static int lock_request(struct fuse_req *req)
 665 {
 666         int err = 0;
 667         if (req) {
 668                 spin_lock(&req->waitq.lock);
 669                 if (test_bit(FR_ABORTED, &req->flags))
 670                         err = -ENOENT;
 671                 else
 672                         set_bit(FR_LOCKED, &req->flags);
 673                 spin_unlock(&req->waitq.lock);
 674         }
 675         return err;
 676 }
 677
 678 /*
 679  * Unlock request.  If it was aborted while locked, caller is responsible
 680  * for unlocking and ending the request.
 681  */
 682 static int unlock_request(struct fuse_req *req)
 683 {
 684         int err = 0;
 685         if (req) {
 686                 spin_lock(&req->waitq.lock);
 687                 if (test_bit(FR_ABORTED, &req->flags))
 688                         err = -ENOENT;
 689                 else
 690                         clear_bit(FR_LOCKED, &req->flags);
 691                 spin_unlock(&req->waitq.lock);
 692         }
 693         return err;
 694 }
 695
 696 struct fuse_copy_state {
 697         int write;
 698         struct fuse_req *req;
 699         struct iov_iter *iter;
 700         struct pipe_buffer *pipebufs;
 701         struct pipe_buffer *currbuf;
 702         struct pipe_inode_info *pipe;
 703         unsigned long nr_segs;
 704         struct page *pg;
 705         unsigned len;
 706         unsigned offset;
 707         unsigned move_pages:1;
 708 };
 709
 710 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 711                            struct iov_iter *iter)
 712 {
 713         memset(cs, 0, sizeof(*cs));
 714         cs->write = write;
 715         cs->iter = iter;
 716 }
 717
 718 /* Unmap and put previous page of userspace buffer */
 719 static void fuse_copy_finish(struct fuse_copy_state *cs)
 720 {
 721         if (cs->currbuf) {
 722                 struct pipe_buffer *buf = cs->currbuf;
 723
 724                 if (cs->write)
 725                         buf->len = PAGE_SIZE - cs->len;
 726                 cs->currbuf = NULL;
 727         } else if (cs->pg) {
 728                 if (cs->write) {
 729                         flush_dcache_page(cs->pg);
 730                         set_page_dirty_lock(cs->pg);
 731                 }
 732                 put_page(cs->pg);
 733         }
 734         cs->pg = NULL;
 735 }
 736
 737 /*
 738  * Get another pagefull of userspace buffer, and map it to kernel
 739  * address space, and lock request
 740  */
 741 static int fuse_copy_fill(struct fuse_copy_state *cs)
 742 {
 743         struct page *page;
 744         int err;
 745
 746         err = unlock_request(cs->req);
 747         if (err)
 748                 return err;
 749
 750         fuse_copy_finish(cs);
 751         if (cs->pipebufs) {
 752                 struct pipe_buffer *buf = cs->pipebufs;
 753
 754                 if (!cs->write) {
 755                         err = buf->ops->confirm(cs->pipe, buf);
 756                         if (err)
 757                                 return err;
 758
 759                         BUG_ON(!cs->nr_segs);
 760                         cs->currbuf = buf;
 761                         cs->pg = buf->page;
 762                         cs->offset = buf->offset;
 763                         cs->len = buf->len;
 764                         cs->pipebufs++;
 765                         cs->nr_segs--;
 766                 } else {
 767                         if (cs->nr_segs == cs->pipe->buffers)
 768                                 return -EIO;
 769
 770                         page = alloc_page(GFP_HIGHUSER);
 771                         if (!page)
 772                                 return -ENOMEM;
 773
 774                         buf->page = page;
 775                         buf->offset = 0;
 776                         buf->len = 0;
 777
 778                         cs->currbuf = buf;
 779                         cs->pg = page;
 780                         cs->offset = 0;
 781                         cs->len = PAGE_SIZE;
 782                         cs->pipebufs++;
 783                         cs->nr_segs++;
 784                 }
 785         } else {
 786                 size_t off;
 787                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 788                 if (err < 0)
 789                         return err;
 790                 BUG_ON(!err);
 791                 cs->len = err;
 792                 cs->offset = off;
 793                 cs->pg = page;
 794                 cs->offset = off;
 795                 iov_iter_advance(cs->iter, err);
 796         }
 797
 798         return lock_request(cs->req);
 799 }
 800
 801 /* Do as much copy to/from userspace buffer as we can */
 802 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 803 {
 804         unsigned ncpy = min(*size, cs->len);
 805         if (val) {
 806                 void *pgaddr = kmap_atomic(cs->pg);
 807                 void *buf = pgaddr + cs->offset;
 808
 809                 if (cs->write)
 810                         memcpy(buf, *val, ncpy);
 811                 else
 812                         memcpy(*val, buf, ncpy);
 813
 814                 kunmap_atomic(pgaddr);
 815                 *val += ncpy;
 816         }
 817         *size -= ncpy;
 818         cs->len -= ncpy;
 819         cs->offset += ncpy;
 820         return ncpy;
 821 }
 822
 823 static int fuse_check_page(struct page *page)
 824 {
 825         if (page_mapcount(page) ||
 826             page->mapping != NULL ||
 827             page_count(page) != 1 ||
 828             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 829              ~(1 << PG_locked |
 830                1 << PG_referenced |
 831                1 << PG_uptodate |
 832                1 << PG_lru |
 833                1 << PG_active |
 834                1 << PG_reclaim))) {
 835                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 836                 printk(KERN_WARNING "  page=%p index=%li flags=%08lx, count=%i, mapcount=%i, mapping=%p\n", page, page->index, page->flags, page_count(page), page_mapcount(page), page->mapping);
 837                 return 1;
 838         }
 839         return 0;
 840 }
 841
 842 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 843 {
 844         int err;
 845         struct page *oldpage = *pagep;
 846         struct page *newpage;
 847         struct pipe_buffer *buf = cs->pipebufs;
 848
 849         err = unlock_request(cs->req);
 850         if (err)
 851                 return err;
 852
 853         fuse_copy_finish(cs);
 854
 855         err = buf->ops->confirm(cs->pipe, buf);
 856         if (err)
 857                 return err;
 858
 859         BUG_ON(!cs->nr_segs);
 860         cs->currbuf = buf;
 861         cs->len = buf->len;
 862         cs->pipebufs++;
 863         cs->nr_segs--;
 864
 865         if (cs->len != PAGE_SIZE)
 866                 goto out_fallback;
 867
 868         if (buf->ops->steal(cs->pipe, buf) != 0)
 869                 goto out_fallback;
 870
 871         newpage = buf->page;
 872
 873         if (!PageUptodate(newpage))
 874                 SetPageUptodate(newpage);
 875
 876         ClearPageMappedToDisk(newpage);
 877
 878         if (fuse_check_page(newpage) != 0)
 879                 goto out_fallback_unlock;
 880
 881         /*
 882          * This is a new and locked page, it shouldn't be mapped or
 883          * have any special flags on it
 884          */
 885         if (WARN_ON(page_mapped(oldpage)))
 886                 goto out_fallback_unlock;
 887         if (WARN_ON(page_has_private(oldpage)))
 888                 goto out_fallback_unlock;
 889         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 890                 goto out_fallback_unlock;
 891         if (WARN_ON(PageMlocked(oldpage)))
 892                 goto out_fallback_unlock;
 893
 894         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 895         if (err) {
 896                 unlock_page(newpage);
 897                 return err;
 898         }
 899
 900         get_page(newpage);
 901
 902         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 903                 lru_cache_add_file(newpage);
 904
 905         err = 0;
 906         spin_lock(&cs->req->waitq.lock);
 907         if (test_bit(FR_ABORTED, &cs->req->flags))
 908                 err = -ENOENT;
 909         else
 910                 *pagep = newpage;
 911         spin_unlock(&cs->req->waitq.lock);
 912
 913         if (err) {
 914                 unlock_page(newpage);
 915                 put_page(newpage);
 916                 return err;
 917         }
 918
 919         unlock_page(oldpage);
 920         put_page(oldpage);
 921         cs->len = 0;
 922
 923         return 0;
 924
 925 out_fallback_unlock:
 926         unlock_page(newpage);
 927 out_fallback:
 928         cs->pg = buf->page;
 929         cs->offset = buf->offset;
 930
 931         err = lock_request(cs->req);
 932         if (err)
 933                 return err;
 934
 935         return 1;
 936 }
 937
 938 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 939                          unsigned offset, unsigned count)
 940 {
 941         struct pipe_buffer *buf;
 942         int err;
 943
 944         if (cs->nr_segs == cs->pipe->buffers)
 945                 return -EIO;
 946
 947         err = unlock_request(cs->req);
 948         if (err)
 949                 return err;
 950
 951         fuse_copy_finish(cs);
 952
 953         buf = cs->pipebufs;
 954         get_page(page);
 955         buf->page = page;
 956         buf->offset = offset;
 957         buf->len = count;
 958
 959         cs->pipebufs++;
 960         cs->nr_segs++;
 961         cs->len = 0;
 962
 963         return 0;
 964 }
 965
 966 /*
 967  * Copy a page in the request to/from the userspace buffer.  Must be
 968  * done atomically
 969  */
 970 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 971                           unsigned offset, unsigned count, int zeroing)
 972 {
 973         int err;
 974         struct page *page = *pagep;
 975
 976         if (page && zeroing && count < PAGE_SIZE)
 977                 clear_highpage(page);
 978
 979         while (count) {
 980                 if (cs->write && cs->pipebufs && page) {
 981                         return fuse_ref_page(cs, page, offset, count);
 982                 } else if (!cs->len) {
 983                         if (cs->move_pages && page &&
 984                             offset == 0 && count == PAGE_SIZE) {
 985                                 err = fuse_try_move_page(cs, pagep);
 986                                 if (err <= 0)
 987                                         return err;
 988                         } else {
 989                                 err = fuse_copy_fill(cs);
 990                                 if (err)
 991                                         return err;
 992                         }
 993                 }
 994                 if (page) {
 995                         void *mapaddr = kmap_atomic(page);
 996                         void *buf = mapaddr + offset;
 997                         offset += fuse_copy_do(cs, &buf, &count);
 998                         kunmap_atomic(mapaddr);
 999                 } else
1000                         offset += fuse_copy_do(cs, NULL, &count);
1001         }
1002         if (page && !cs->write)
1003                 flush_dcache_page(page);
1004         return 0;
1005 }
1006
1007 /* Copy pages in the request to/from userspace buffer */
1008 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1009                            int zeroing)
1010 {
1011         unsigned i;
1012         struct fuse_req *req = cs->req;
1013
1014         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1015                 int err;
1016                 unsigned offset = req->page_descs[i].offset;
1017                 unsigned count = min(nbytes, req->page_descs[i].length);
1018
1019                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1020                                      zeroing);
1021                 if (err)
1022                         return err;
1023
1024                 nbytes -= count;
1025         }
1026         return 0;
1027 }
1028
1029 /* Copy a single argument in the request to/from userspace buffer */
1030 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1031 {
1032         while (size) {
1033                 if (!cs->len) {
1034                         int err = fuse_copy_fill(cs);
1035                         if (err)
1036                                 return err;
1037                 }
1038                 fuse_copy_do(cs, &val, &size);
1039         }
1040         return 0;
1041 }
1042
1043 /* Copy request arguments to/from userspace buffer */
1044 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1045                           unsigned argpages, struct fuse_arg *args,
1046                           int zeroing)
1047 {
1048         int err = 0;
1049         unsigned i;
1050
1051         for (i = 0; !err && i < numargs; i++)  {
1052                 struct fuse_arg *arg = &args[i];
1053                 if (i == numargs - 1 && argpages)
1054                         err = fuse_copy_pages(cs, arg->size, zeroing);
1055                 else
1056                         err = fuse_copy_one(cs, arg->value, arg->size);
1057         }
1058         return err;
1059 }
1060
1061 static int forget_pending(struct fuse_iqueue *fiq)
1062 {
1063         return fiq->forget_list_head.next != NULL;
1064 }
1065
1066 static int request_pending(struct fuse_iqueue *fiq)
1067 {
1068         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1069                 forget_pending(fiq);
1070 }
1071
1072 /*
1073  * Transfer an interrupt request to userspace
1074  *
1075  * Unlike other requests this is assembled on demand, without a need
1076  * to allocate a separate fuse_req structure.
1077  *
1078  * Called with fiq->waitq.lock held, releases it
1079  */
1080 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1081                                struct fuse_copy_state *cs,
1082                                size_t nbytes, struct fuse_req *req)
1083 __releases(fiq->waitq.lock)
1084 {
1085         struct fuse_in_header ih;
1086         struct fuse_interrupt_in arg;
1087         unsigned reqsize = sizeof(ih) + sizeof(arg);
1088         int err;
1089
1090         list_del_init(&req->intr_entry);
1091         req->intr_unique = fuse_get_unique(fiq);
1092         memset(&ih, 0, sizeof(ih));
1093         memset(&arg, 0, sizeof(arg));
1094         ih.len = reqsize;
1095         ih.opcode = FUSE_INTERRUPT;
1096         ih.unique = req->intr_unique;
1097         arg.unique = req->in.h.unique;
1098
1099         spin_unlock(&fiq->waitq.lock);
1100         if (nbytes < reqsize)
1101                 return -EINVAL;
1102
1103         err = fuse_copy_one(cs, &ih, sizeof(ih));
1104         if (!err)
1105                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1106         fuse_copy_finish(cs);
1107
1108         return err ? err : reqsize;
1109 }
1110
1111 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1112                                                unsigned max,
1113                                                unsigned *countp)
1114 {
1115         struct fuse_forget_link *head = fiq->forget_list_head.next;
1116         struct fuse_forget_link **newhead = &head;
1117         unsigned count;
1118
1119         for (count = 0; *newhead != NULL && count < max; count++)
1120                 newhead = &(*newhead)->next;
1121
1122         fiq->forget_list_head.next = *newhead;
1123         *newhead = NULL;
1124         if (fiq->forget_list_head.next == NULL)
1125                 fiq->forget_list_tail = &fiq->forget_list_head;
1126
1127         if (countp != NULL)
1128                 *countp = count;
1129
1130         return head;
1131 }
1132
1133 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1134                                    struct fuse_copy_state *cs,
1135                                    size_t nbytes)
1136 __releases(fiq->waitq.lock)
1137 {
1138         int err;
1139         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1140         struct fuse_forget_in arg = {
1141                 .nlookup = forget->forget_one.nlookup,
1142         };
1143         struct fuse_in_header ih = {
1144                 .opcode = FUSE_FORGET,
1145                 .nodeid = forget->forget_one.nodeid,
1146                 .unique = fuse_get_unique(fiq),
1147                 .len = sizeof(ih) + sizeof(arg),
1148         };
1149
1150         spin_unlock(&fiq->waitq.lock);
1151         kfree(forget);
1152         if (nbytes < ih.len)
1153                 return -EINVAL;
1154
1155         err = fuse_copy_one(cs, &ih, sizeof(ih));
1156         if (!err)
1157                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1158         fuse_copy_finish(cs);
1159
1160         if (err)
1161                 return err;
1162
1163         return ih.len;
1164 }
1165
1166 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1167                                    struct fuse_copy_state *cs, size_t nbytes)
1168 __releases(fiq->waitq.lock)
1169 {
1170         int err;
1171         unsigned max_forgets;
1172         unsigned count;
1173         struct fuse_forget_link *head;
1174         struct fuse_batch_forget_in arg = { .count = 0 };
1175         struct fuse_in_header ih = {
1176                 .opcode = FUSE_BATCH_FORGET,
1177                 .unique = fuse_get_unique(fiq),
1178                 .len = sizeof(ih) + sizeof(arg),
1179         };
1180
1181         if (nbytes < ih.len) {
1182                 spin_unlock(&fiq->waitq.lock);
1183                 return -EINVAL;
1184         }
1185
1186         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1187         head = dequeue_forget(fiq, max_forgets, &count);
1188         spin_unlock(&fiq->waitq.lock);
1189
1190         arg.count = count;
1191         ih.len += count * sizeof(struct fuse_forget_one);
1192         err = fuse_copy_one(cs, &ih, sizeof(ih));
1193         if (!err)
1194                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1195
1196         while (head) {
1197                 struct fuse_forget_link *forget = head;
1198
1199                 if (!err) {
1200                         err = fuse_copy_one(cs, &forget->forget_one,
1201                                             sizeof(forget->forget_one));
1202                 }
1203                 head = forget->next;
1204                 kfree(forget);
1205         }
1206
1207         fuse_copy_finish(cs);
1208
1209         if (err)
1210                 return err;
1211
1212         return ih.len;
1213 }
1214
1215 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1216                             struct fuse_copy_state *cs,
1217                             size_t nbytes)
1218 __releases(fiq->waitq.lock)
1219 {
1220         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1221                 return fuse_read_single_forget(fiq, cs, nbytes);
1222         else
1223                 return fuse_read_batch_forget(fiq, cs, nbytes);
1224 }
1225
1226 /*
1227  * Read a single request into the userspace filesystem's buffer.  This
1228  * function waits until a request is available, then removes it from
1229  * the pending list and copies request data to userspace buffer.  If
1230  * no reply is needed (FORGET) or request has been aborted or there
1231  * was an error during the copying then it's finished by calling
1232  * request_end().  Otherwise add it to the processing list, and set
1233  * the 'sent' flag.
1234  */
1235 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1236                                 struct fuse_copy_state *cs, size_t nbytes)
1237 {
1238         ssize_t err;
1239         struct fuse_conn *fc = fud->fc;
1240         struct fuse_iqueue *fiq = &fc->iq;
1241         struct fuse_pqueue *fpq = &fud->pq;
1242         struct fuse_req *req;
1243         struct fuse_in *in;
1244         unsigned reqsize;
1245
1246  restart:
1247         spin_lock(&fiq->waitq.lock);
1248         err = -EAGAIN;
1249         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1250             !request_pending(fiq))
1251                 goto err_unlock;
1252
1253         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1254                                 !fiq->connected || request_pending(fiq));
1255         if (err)
1256                 goto err_unlock;
1257
1258         err = -ENODEV;
1259         if (!fiq->connected)
1260                 goto err_unlock;
1261
1262         if (!list_empty(&fiq->interrupts)) {
1263                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1264                                  intr_entry);
1265                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1266         }
1267
1268         if (forget_pending(fiq)) {
1269                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1270                         return fuse_read_forget(fc, fiq, cs, nbytes);
1271
1272                 if (fiq->forget_batch <= -8)
1273                         fiq->forget_batch = 16;
1274         }
1275
1276         req = list_entry(fiq->pending.next, struct fuse_req, list);
1277         clear_bit(FR_PENDING, &req->flags);
1278         list_del_init(&req->list);
1279         spin_unlock(&fiq->waitq.lock);
1280
1281         in = &req->in;
1282         reqsize = in->h.len;
1283         /* If request is too large, reply with an error and restart the read */
1284         if (nbytes < reqsize) {
1285                 req->out.h.error = -EIO;
1286                 /* SETXATTR is special, since it may contain too large data */
1287                 if (in->h.opcode == FUSE_SETXATTR)
1288                         req->out.h.error = -E2BIG;
1289                 request_end(fc, req);
1290                 goto restart;
1291         }
1292         spin_lock(&fpq->lock);
1293         list_add(&req->list, &fpq->io);
1294         spin_unlock(&fpq->lock);
1295         cs->req = req;
1296         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1297         if (!err)
1298                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1299                                      (struct fuse_arg *) in->args, 0);
1300         fuse_copy_finish(cs);
1301         spin_lock(&fpq->lock);
1302         clear_bit(FR_LOCKED, &req->flags);
1303         if (!fpq->connected) {
1304                 err = -ENODEV;
1305                 goto out_end;
1306         }
1307         if (err) {
1308                 req->out.h.error = -EIO;
1309                 goto out_end;
1310         }
1311         if (!test_bit(FR_ISREPLY, &req->flags)) {
1312                 err = reqsize;
1313                 goto out_end;
1314         }
1315         list_move_tail(&req->list, &fpq->processing);
1316         spin_unlock(&fpq->lock);
1317         set_bit(FR_SENT, &req->flags);
1318         /* matches barrier in request_wait_answer() */
1319         smp_mb__after_atomic();
1320         if (test_bit(FR_INTERRUPTED, &req->flags))
1321                 queue_interrupt(fiq, req);
1322
1323         return reqsize;
1324
1325 out_end:
1326         if (!test_bit(FR_PRIVATE, &req->flags))
1327                 list_del_init(&req->list);
1328         spin_unlock(&fpq->lock);
1329         request_end(fc, req);
1330         return err;
1331
1332  err_unlock:
1333         spin_unlock(&fiq->waitq.lock);
1334         return err;
1335 }
1336
1337 static int fuse_dev_open(struct inode *inode, struct file *file)
1338 {
1339         /*
1340          * The fuse device's file's private_data is used to hold
1341          * the fuse_conn(ection) when it is mounted, and is used to
1342          * keep track of whether the file has been mounted already.
1343          */
1344         file->private_data = NULL;
1345         return 0;
1346 }
1347
1348 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1349 {
1350         struct fuse_copy_state cs;
1351         struct file *file = iocb->ki_filp;
1352         struct fuse_dev *fud = fuse_get_dev(file);
1353
1354         if (!fud)
1355                 return -EPERM;
1356
1357         if (!iter_is_iovec(to))
1358                 return -EINVAL;
1359
1360         fuse_copy_init(&cs, 1, to);
1361
1362         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1363 }
1364
1365 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1366                                     struct pipe_inode_info *pipe,
1367                                     size_t len, unsigned int flags)
1368 {
1369         int ret;
1370         int page_nr = 0;
1371         int do_wakeup = 0;
1372         struct pipe_buffer *bufs;
1373         struct fuse_copy_state cs;
1374         struct fuse_dev *fud = fuse_get_dev(in);
1375
1376         if (!fud)
1377                 return -EPERM;
1378
1379         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1380         if (!bufs)
1381                 return -ENOMEM;
1382
1383         fuse_copy_init(&cs, 1, NULL);
1384         cs.pipebufs = bufs;
1385         cs.pipe = pipe;
1386         ret = fuse_dev_do_read(fud, in, &cs, len);
1387         if (ret < 0)
1388                 goto out;
1389
1390         ret = 0;
1391         pipe_lock(pipe);
1392
1393         if (!pipe->readers) {
1394                 send_sig(SIGPIPE, current, 0);
1395                 if (!ret)
1396                         ret = -EPIPE;
1397                 goto out_unlock;
1398         }
1399
1400         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1401                 ret = -EIO;
1402                 goto out_unlock;
1403         }
1404
1405         while (page_nr < cs.nr_segs) {
1406                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1407                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1408
1409                 buf->page = bufs[page_nr].page;
1410                 buf->offset = bufs[page_nr].offset;
1411                 buf->len = bufs[page_nr].len;
1412                 /*
1413                  * Need to be careful about this.  Having buf->ops in module
1414                  * code can Oops if the buffer persists after module unload.
1415                  */
1416                 buf->ops = &nosteal_pipe_buf_ops;
1417
1418                 pipe->nrbufs++;
1419                 page_nr++;
1420                 ret += buf->len;
1421
1422                 if (pipe->files)
1423                         do_wakeup = 1;
1424         }
1425
1426 out_unlock:
1427         pipe_unlock(pipe);
1428
1429         if (do_wakeup) {
1430                 smp_mb();
1431                 if (waitqueue_active(&pipe->wait))
1432                         wake_up_interruptible(&pipe->wait);
1433                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1434         }
1435
1436 out:
1437         for (; page_nr < cs.nr_segs; page_nr++)
1438                 put_page(bufs[page_nr].page);
1439
1440         kfree(bufs);
1441         return ret;
1442 }
1443
1444 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1445                             struct fuse_copy_state *cs)
1446 {
1447         struct fuse_notify_poll_wakeup_out outarg;
1448         int err = -EINVAL;
1449
1450         if (size != sizeof(outarg))
1451                 goto err;
1452
1453         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1454         if (err)
1455                 goto err;
1456
1457         fuse_copy_finish(cs);
1458         return fuse_notify_poll_wakeup(fc, &outarg);
1459
1460 err:
1461         fuse_copy_finish(cs);
1462         return err;
1463 }
1464
1465 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1466                                    struct fuse_copy_state *cs)
1467 {
1468         struct fuse_notify_inval_inode_out outarg;
1469         int err = -EINVAL;
1470
1471         if (size != sizeof(outarg))
1472                 goto err;
1473
1474         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1475         if (err)
1476                 goto err;
1477         fuse_copy_finish(cs);
1478
1479         down_read(&fc->killsb);
1480         err = -ENOENT;
1481         if (fc->sb) {
1482                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1483                                                outarg.off, outarg.len);
1484         }
1485         up_read(&fc->killsb);
1486         return err;
1487
1488 err:
1489         fuse_copy_finish(cs);
1490         return err;
1491 }
1492
1493 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1494                                    struct fuse_copy_state *cs)
1495 {
1496         struct fuse_notify_inval_entry_out outarg;
1497         int err = -ENOMEM;
1498         char *buf;
1499         struct qstr name;
1500
1501         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1502         if (!buf)
1503                 goto err;
1504
1505         err = -EINVAL;
1506         if (size < sizeof(outarg))
1507                 goto err;
1508
1509         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1510         if (err)
1511                 goto err;
1512
1513         err = -ENAMETOOLONG;
1514         if (outarg.namelen > FUSE_NAME_MAX)
1515                 goto err;
1516
1517         err = -EINVAL;
1518         if (size != sizeof(outarg) + outarg.namelen + 1)
1519                 goto err;
1520
1521         name.name = buf;
1522         name.len = outarg.namelen;
1523         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1524         if (err)
1525                 goto err;
1526         fuse_copy_finish(cs);
1527         buf[outarg.namelen] = 0;
1528         name.hash = full_name_hash(name.name, name.len);
1529
1530         down_read(&fc->killsb);
1531         err = -ENOENT;
1532         if (fc->sb)
1533                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1534         up_read(&fc->killsb);
1535         kfree(buf);
1536         return err;
1537
1538 err:
1539         kfree(buf);
1540         fuse_copy_finish(cs);
1541         return err;
1542 }
1543
1544 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1545                               struct fuse_copy_state *cs)
1546 {
1547         struct fuse_notify_delete_out outarg;
1548         int err = -ENOMEM;
1549         char *buf;
1550         struct qstr name;
1551
1552         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1553         if (!buf)
1554                 goto err;
1555
1556         err = -EINVAL;
1557         if (size < sizeof(outarg))
1558                 goto err;
1559
1560         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1561         if (err)
1562                 goto err;
1563
1564         err = -ENAMETOOLONG;
1565         if (outarg.namelen > FUSE_NAME_MAX)
1566                 goto err;
1567
1568         err = -EINVAL;
1569         if (size != sizeof(outarg) + outarg.namelen + 1)
1570                 goto err;
1571
1572         name.name = buf;
1573         name.len = outarg.namelen;
1574         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1575         if (err)
1576                 goto err;
1577         fuse_copy_finish(cs);
1578         buf[outarg.namelen] = 0;
1579         name.hash = full_name_hash(name.name, name.len);
1580
1581         down_read(&fc->killsb);
1582         err = -ENOENT;
1583         if (fc->sb)
1584                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1585                                                outarg.child, &name);
1586         up_read(&fc->killsb);
1587         kfree(buf);
1588         return err;
1589
1590 err:
1591         kfree(buf);
1592         fuse_copy_finish(cs);
1593         return err;
1594 }
1595
1596 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1597                              struct fuse_copy_state *cs)
1598 {
1599         struct fuse_notify_store_out outarg;
1600         struct inode *inode;
1601         struct address_space *mapping;
1602         u64 nodeid;
1603         int err;
1604         pgoff_t index;
1605         unsigned int offset;
1606         unsigned int num;
1607         loff_t file_size;
1608         loff_t end;
1609
1610         err = -EINVAL;
1611         if (size < sizeof(outarg))
1612                 goto out_finish;
1613
1614         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1615         if (err)
1616                 goto out_finish;
1617
1618         err = -EINVAL;
1619         if (size - sizeof(outarg) != outarg.size)
1620                 goto out_finish;
1621
1622         nodeid = outarg.nodeid;
1623
1624         down_read(&fc->killsb);
1625
1626         err = -ENOENT;
1627         if (!fc->sb)
1628                 goto out_up_killsb;
1629
1630         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1631         if (!inode)
1632                 goto out_up_killsb;
1633
1634         mapping = inode->i_mapping;
1635         index = outarg.offset >> PAGE_SHIFT;
1636         offset = outarg.offset & ~PAGE_MASK;
1637         file_size = i_size_read(inode);
1638         end = outarg.offset + outarg.size;
1639         if (end > file_size) {
1640                 file_size = end;
1641                 fuse_write_update_size(inode, file_size);
1642         }
1643
1644         num = outarg.size;
1645         while (num) {
1646                 struct page *page;
1647                 unsigned int this_num;
1648
1649                 err = -ENOMEM;
1650                 page = find_or_create_page(mapping, index,
1651                                            mapping_gfp_mask(mapping));
1652                 if (!page)
1653                         goto out_iput;
1654
1655                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1656                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1657                 if (!err && offset == 0 &&
1658                     (this_num == PAGE_SIZE || file_size == end))
1659                         SetPageUptodate(page);
1660                 unlock_page(page);
1661                 put_page(page);
1662
1663                 if (err)
1664                         goto out_iput;
1665
1666                 num -= this_num;
1667                 offset = 0;
1668                 index++;
1669         }
1670
1671         err = 0;
1672
1673 out_iput:
1674         iput(inode);
1675 out_up_killsb:
1676         up_read(&fc->killsb);
1677 out_finish:
1678         fuse_copy_finish(cs);
1679         return err;
1680 }
1681
1682 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1683 {
1684         release_pages(req->pages, req->num_pages, false);
1685 }
1686
1687 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1688                          struct fuse_notify_retrieve_out *outarg)
1689 {
1690         int err;
1691         struct address_space *mapping = inode->i_mapping;
1692         struct fuse_req *req;
1693         pgoff_t index;
1694         loff_t file_size;
1695         unsigned int num;
1696         unsigned int offset;
1697         size_t total_len = 0;
1698         int num_pages;
1699
1700         offset = outarg->offset & ~PAGE_MASK;
1701         file_size = i_size_read(inode);
1702
1703         num = outarg->size;
1704         if (outarg->offset > file_size)
1705                 num = 0;
1706         else if (outarg->offset + num > file_size)
1707                 num = file_size - outarg->offset;
1708
1709         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1710         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1711
1712         req = fuse_get_req(fc, num_pages);
1713         if (IS_ERR(req))
1714                 return PTR_ERR(req);
1715
1716         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1717         req->in.h.nodeid = outarg->nodeid;
1718         req->in.numargs = 2;
1719         req->in.argpages = 1;
1720         req->page_descs[0].offset = offset;
1721         req->end = fuse_retrieve_end;
1722
1723         index = outarg->offset >> PAGE_SHIFT;
1724
1725         while (num && req->num_pages < num_pages) {
1726                 struct page *page;
1727                 unsigned int this_num;
1728
1729                 page = find_get_page(mapping, index);
1730                 if (!page)
1731                         break;
1732
1733                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1734                 req->pages[req->num_pages] = page;
1735                 req->page_descs[req->num_pages].length = this_num;
1736                 req->num_pages++;
1737
1738                 offset = 0;
1739                 num -= this_num;
1740                 total_len += this_num;
1741                 index++;
1742         }
1743         req->misc.retrieve_in.offset = outarg->offset;
1744         req->misc.retrieve_in.size = total_len;
1745         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1746         req->in.args[0].value = &req->misc.retrieve_in;
1747         req->in.args[1].size = total_len;
1748
1749         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1750         if (err)
1751                 fuse_retrieve_end(fc, req);
1752
1753         return err;
1754 }
1755
1756 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1757                                 struct fuse_copy_state *cs)
1758 {
1759         struct fuse_notify_retrieve_out outarg;
1760         struct inode *inode;
1761         int err;
1762
1763         err = -EINVAL;
1764         if (size != sizeof(outarg))
1765                 goto copy_finish;
1766
1767         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1768         if (err)
1769                 goto copy_finish;
1770
1771         fuse_copy_finish(cs);
1772
1773         down_read(&fc->killsb);
1774         err = -ENOENT;
1775         if (fc->sb) {
1776                 u64 nodeid = outarg.nodeid;
1777
1778                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1779                 if (inode) {
1780                         err = fuse_retrieve(fc, inode, &outarg);
1781                         iput(inode);
1782                 }
1783         }
1784         up_read(&fc->killsb);
1785
1786         return err;
1787
1788 copy_finish:
1789         fuse_copy_finish(cs);
1790         return err;
1791 }
1792
1793 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1794                        unsigned int size, struct fuse_copy_state *cs)
1795 {
1796         /* Don't try to move pages (yet) */
1797         cs->move_pages = 0;
1798
1799         switch (code) {
1800         case FUSE_NOTIFY_POLL:
1801                 return fuse_notify_poll(fc, size, cs);
1802
1803         case FUSE_NOTIFY_INVAL_INODE:
1804                 return fuse_notify_inval_inode(fc, size, cs);
1805
1806         case FUSE_NOTIFY_INVAL_ENTRY:
1807                 return fuse_notify_inval_entry(fc, size, cs);
1808
1809         case FUSE_NOTIFY_STORE:
1810                 return fuse_notify_store(fc, size, cs);
1811
1812         case FUSE_NOTIFY_RETRIEVE:
1813                 return fuse_notify_retrieve(fc, size, cs);
1814
1815         case FUSE_NOTIFY_DELETE:
1816                 return fuse_notify_delete(fc, size, cs);
1817
1818         default:
1819                 fuse_copy_finish(cs);
1820                 return -EINVAL;
1821         }
1822 }
1823
1824 /* Look up request on processing list by unique ID */
1825 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1826 {
1827         struct fuse_req *req;
1828
1829         list_for_each_entry(req, &fpq->processing, list) {
1830                 if (req->in.h.unique == unique || req->intr_unique == unique)
1831                         return req;
1832         }
1833         return NULL;
1834 }
1835
1836 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1837                          unsigned nbytes)
1838 {
1839         unsigned reqsize = sizeof(struct fuse_out_header);
1840
1841         if (out->h.error)
1842                 return nbytes != reqsize ? -EINVAL : 0;
1843
1844         reqsize += len_args(out->numargs, out->args);
1845
1846         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1847                 return -EINVAL;
1848         else if (reqsize > nbytes) {
1849                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1850                 unsigned diffsize = reqsize - nbytes;
1851                 if (diffsize > lastarg->size)
1852                         return -EINVAL;
1853                 lastarg->size -= diffsize;
1854         }
1855         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1856                               out->page_zeroing);
1857 }
1858
1859 /*
1860  * Write a single reply to a request.  First the header is copied from
1861  * the write buffer.  The request is then searched on the processing
1862  * list by the unique ID found in the header.  If found, then remove
1863  * it from the list and copy the rest of the buffer to the request.
1864  * The request is finished by calling request_end()
1865  */
1866 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1867                                  struct fuse_copy_state *cs, size_t nbytes)
1868 {
1869         int err;
1870         struct fuse_conn *fc = fud->fc;
1871         struct fuse_pqueue *fpq = &fud->pq;
1872         struct fuse_req *req;
1873         struct fuse_out_header oh;
1874
1875         if (nbytes < sizeof(struct fuse_out_header))
1876                 return -EINVAL;
1877
1878         err = fuse_copy_one(cs, &oh, sizeof(oh));
1879         if (err)
1880                 goto err_finish;
1881
1882         err = -EINVAL;
1883         if (oh.len != nbytes)
1884                 goto err_finish;
1885
1886         /*
1887          * Zero oh.unique indicates unsolicited notification message
1888          * and error contains notification code.
1889          */
1890         if (!oh.unique) {
1891                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1892                 return err ? err : nbytes;
1893         }
1894
1895         err = -EINVAL;
1896         if (oh.error <= -1000 || oh.error > 0)
1897                 goto err_finish;
1898
1899         spin_lock(&fpq->lock);
1900         err = -ENOENT;
1901         if (!fpq->connected)
1902                 goto err_unlock_pq;
1903
1904         req = request_find(fpq, oh.unique);
1905         if (!req)
1906                 goto err_unlock_pq;
1907
1908         /* Is it an interrupt reply? */
1909         if (req->intr_unique == oh.unique) {
1910                 spin_unlock(&fpq->lock);
1911
1912                 err = -EINVAL;
1913                 if (nbytes != sizeof(struct fuse_out_header))
1914                         goto err_finish;
1915
1916                 if (oh.error == -ENOSYS)
1917                         fc->no_interrupt = 1;
1918                 else if (oh.error == -EAGAIN)
1919                         queue_interrupt(&fc->iq, req);
1920
1921                 fuse_copy_finish(cs);
1922                 return nbytes;
1923         }
1924
1925         clear_bit(FR_SENT, &req->flags);
1926         list_move(&req->list, &fpq->io);
1927         req->out.h = oh;
1928         set_bit(FR_LOCKED, &req->flags);
1929         spin_unlock(&fpq->lock);
1930         cs->req = req;
1931         if (!req->out.page_replace)
1932                 cs->move_pages = 0;
1933
1934         err = copy_out_args(cs, &req->out, nbytes);
1935         fuse_copy_finish(cs);
1936
1937         spin_lock(&fpq->lock);
1938         clear_bit(FR_LOCKED, &req->flags);
1939         if (!fpq->connected)
1940                 err = -ENOENT;
1941         else if (err)
1942                 req->out.h.error = -EIO;
1943         if (!test_bit(FR_PRIVATE, &req->flags))
1944                 list_del_init(&req->list);
1945         spin_unlock(&fpq->lock);
1946
1947         request_end(fc, req);
1948
1949         return err ? err : nbytes;
1950
1951  err_unlock_pq:
1952         spin_unlock(&fpq->lock);
1953  err_finish:
1954         fuse_copy_finish(cs);
1955         return err;
1956 }
1957
1958 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1959 {
1960         struct fuse_copy_state cs;
1961         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1962
1963         if (!fud)
1964                 return -EPERM;
1965
1966         if (!iter_is_iovec(from))
1967                 return -EINVAL;
1968
1969         fuse_copy_init(&cs, 0, from);
1970
1971         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1972 }
1973
1974 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1975                                      struct file *out, loff_t *ppos,
1976                                      size_t len, unsigned int flags)
1977 {
1978         unsigned nbuf;
1979         unsigned idx;
1980         struct pipe_buffer *bufs;
1981         struct fuse_copy_state cs;
1982         struct fuse_dev *fud;
1983         size_t rem;
1984         ssize_t ret;
1985
1986         fud = fuse_get_dev(out);
1987         if (!fud)
1988                 return -EPERM;
1989
1990         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1991         if (!bufs)
1992                 return -ENOMEM;
1993
1994         pipe_lock(pipe);
1995         nbuf = 0;
1996         rem = 0;
1997         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1998                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1999
2000         ret = -EINVAL;
2001         if (rem < len) {
2002                 pipe_unlock(pipe);
2003                 goto out;
2004         }
2005
2006         rem = len;
2007         while (rem) {
2008                 struct pipe_buffer *ibuf;
2009                 struct pipe_buffer *obuf;
2010
2011                 BUG_ON(nbuf >= pipe->buffers);
2012                 BUG_ON(!pipe->nrbufs);
2013                 ibuf = &pipe->bufs[pipe->curbuf];
2014                 obuf = &bufs[nbuf];
2015
2016                 if (rem >= ibuf->len) {
2017                         *obuf = *ibuf;
2018                         ibuf->ops = NULL;
2019                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2020                         pipe->nrbufs--;
2021                 } else {
2022                         ibuf->ops->get(pipe, ibuf);
2023                         *obuf = *ibuf;
2024                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2025                         obuf->len = rem;
2026                         ibuf->offset += obuf->len;
2027                         ibuf->len -= obuf->len;
2028                 }
2029                 nbuf++;
2030                 rem -= obuf->len;
2031         }
2032         pipe_unlock(pipe);
2033
2034         fuse_copy_init(&cs, 0, NULL);
2035         cs.pipebufs = bufs;
2036         cs.nr_segs = nbuf;
2037         cs.pipe = pipe;
2038
2039         if (flags & SPLICE_F_MOVE)
2040                 cs.move_pages = 1;
2041
2042         ret = fuse_dev_do_write(fud, &cs, len);
2043
2044         for (idx = 0; idx < nbuf; idx++) {
2045                 struct pipe_buffer *buf = &bufs[idx];
2046                 buf->ops->release(pipe, buf);
2047         }
2048 out:
2049         kfree(bufs);
2050         return ret;
2051 }
2052
2053 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2054 {
2055         unsigned mask = POLLOUT | POLLWRNORM;
2056         struct fuse_iqueue *fiq;
2057         struct fuse_dev *fud = fuse_get_dev(file);
2058
2059         if (!fud)
2060                 return POLLERR;
2061
2062         fiq = &fud->fc->iq;
2063         poll_wait(file, &fiq->waitq, wait);
2064
2065         spin_lock(&fiq->waitq.lock);
2066         if (!fiq->connected)
2067                 mask = POLLERR;
2068         else if (request_pending(fiq))
2069                 mask |= POLLIN | POLLRDNORM;
2070         spin_unlock(&fiq->waitq.lock);
2071
2072         return mask;
2073 }
2074
2075 /*
2076  * Abort all requests on the given list (pending or processing)
2077  *
2078  * This function releases and reacquires fc->lock
2079  */
2080 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2081 {
2082         while (!list_empty(head)) {
2083                 struct fuse_req *req;
2084                 req = list_entry(head->next, struct fuse_req, list);
2085                 req->out.h.error = -ECONNABORTED;
2086                 clear_bit(FR_PENDING, &req->flags);
2087                 clear_bit(FR_SENT, &req->flags);
2088                 list_del_init(&req->list);
2089                 request_end(fc, req);
2090         }
2091 }
2092
2093 static void end_polls(struct fuse_conn *fc)
2094 {
2095         struct rb_node *p;
2096
2097         p = rb_first(&fc->polled_files);
2098
2099         while (p) {
2100                 struct fuse_file *ff;
2101                 ff = rb_entry(p, struct fuse_file, polled_node);
2102                 wake_up_interruptible_all(&ff->poll_wait);
2103
2104                 p = rb_next(p);
2105         }
2106 }
2107
2108 /*
2109  * Abort all requests.
2110  *
2111  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2112  * filesystem.
2113  *
2114  * The same effect is usually achievable through killing the filesystem daemon
2115  * and all users of the filesystem.  The exception is the combination of an
2116  * asynchronous request and the tricky deadlock (see
2117  * Documentation/filesystems/fuse.txt).
2118  *
2119  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2120  * requests, they should be finished off immediately.  Locked requests will be
2121  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2122  * requests.  It is possible that some request will finish before we can.  This
2123  * is OK, the request will in that case be removed from the list before we touch
2124  * it.
2125  */
2126 void fuse_abort_conn(struct fuse_conn *fc)
2127 {
2128         struct fuse_iqueue *fiq = &fc->iq;
2129
2130         spin_lock(&fc->lock);
2131         if (fc->connected) {
2132                 struct fuse_dev *fud;
2133                 struct fuse_req *req, *next;
2134                 LIST_HEAD(to_end1);
2135                 LIST_HEAD(to_end2);
2136
2137                 fc->connected = 0;
2138                 fc->blocked = 0;
2139                 fuse_set_initialized(fc);
2140                 list_for_each_entry(fud, &fc->devices, entry) {
2141                         struct fuse_pqueue *fpq = &fud->pq;
2142
2143                         spin_lock(&fpq->lock);
2144                         fpq->connected = 0;
2145                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2146                                 req->out.h.error = -ECONNABORTED;
2147                                 spin_lock(&req->waitq.lock);
2148                                 set_bit(FR_ABORTED, &req->flags);
2149                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2150                                         set_bit(FR_PRIVATE, &req->flags);
2151                                         list_move(&req->list, &to_end1);
2152                                 }
2153                                 spin_unlock(&req->waitq.lock);
2154                         }
2155                         list_splice_init(&fpq->processing, &to_end2);
2156                         spin_unlock(&fpq->lock);
2157                 }
2158                 fc->max_background = UINT_MAX;
2159                 flush_bg_queue(fc);
2160
2161                 spin_lock(&fiq->waitq.lock);
2162                 fiq->connected = 0;
2163                 list_splice_init(&fiq->pending, &to_end2);
2164                 while (forget_pending(fiq))
2165                         kfree(dequeue_forget(fiq, 1, NULL));
2166                 wake_up_all_locked(&fiq->waitq);
2167                 spin_unlock(&fiq->waitq.lock);
2168                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2169                 end_polls(fc);
2170                 wake_up_all(&fc->blocked_waitq);
2171                 spin_unlock(&fc->lock);
2172
2173                 while (!list_empty(&to_end1)) {
2174                         req = list_first_entry(&to_end1, struct fuse_req, list);
2175                         __fuse_get_request(req);
2176                         list_del_init(&req->list);
2177                         request_end(fc, req);
2178                 }
2179                 end_requests(fc, &to_end2);
2180         } else {
2181                 spin_unlock(&fc->lock);
2182         }
2183 }
2184 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2185
2186 int fuse_dev_release(struct inode *inode, struct file *file)
2187 {
2188         struct fuse_dev *fud = fuse_get_dev(file);
2189
2190         if (fud) {
2191                 struct fuse_conn *fc = fud->fc;
2192                 struct fuse_pqueue *fpq = &fud->pq;
2193
2194                 WARN_ON(!list_empty(&fpq->io));
2195                 end_requests(fc, &fpq->processing);
2196                 /* Are we the last open device? */
2197                 if (atomic_dec_and_test(&fc->dev_count)) {
2198                         WARN_ON(fc->iq.fasync != NULL);
2199                         fuse_abort_conn(fc);
2200                 }
2201                 fuse_dev_free(fud);
2202         }
2203         return 0;
2204 }
2205 EXPORT_SYMBOL_GPL(fuse_dev_release);
2206
2207 static int fuse_dev_fasync(int fd, struct file *file, int on)
2208 {
2209         struct fuse_dev *fud = fuse_get_dev(file);
2210
2211         if (!fud)
2212                 return -EPERM;
2213
2214         /* No locking - fasync_helper does its own locking */
2215         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2216 }
2217
2218 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2219 {
2220         struct fuse_dev *fud;
2221
2222         if (new->private_data)
2223                 return -EINVAL;
2224
2225         fud = fuse_dev_alloc(fc);
2226         if (!fud)
2227                 return -ENOMEM;
2228
2229         new->private_data = fud;
2230         atomic_inc(&fc->dev_count);
2231
2232         return 0;
2233 }
2234
2235 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2236                            unsigned long arg)
2237 {
2238         int err = -ENOTTY;
2239
2240         if (cmd == FUSE_DEV_IOC_CLONE) {
2241                 int oldfd;
2242
2243                 err = -EFAULT;
2244                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2245                         struct file *old = fget(oldfd);
2246
2247                         err = -EINVAL;
2248                         if (old) {
2249                                 struct fuse_dev *fud = NULL;
2250
2251                                 /*
2252                                  * Check against file->f_op because CUSE
2253                                  * uses the same ioctl handler.
2254                                  */
2255                                 if (old->f_op == file->f_op &&
2256                                     old->f_cred->user_ns == file->f_cred->user_ns)
2257                                         fud = fuse_get_dev(old);
2258
2259                                 if (fud) {
2260                                         mutex_lock(&fuse_mutex);
2261                                         err = fuse_device_clone(fud->fc, file);
2262                                         mutex_unlock(&fuse_mutex);
2263                                 }
2264                                 fput(old);
2265                         }
2266                 }
2267         }
2268         return err;
2269 }
2270
2271 const struct file_operations fuse_dev_operations = {
2272         .owner          = THIS_MODULE,
2273         .open           = fuse_dev_open,
2274         .llseek         = no_llseek,
2275         .read_iter      = fuse_dev_read,
2276         .splice_read    = fuse_dev_splice_read,
2277         .write_iter     = fuse_dev_write,
2278         .splice_write   = fuse_dev_splice_write,
2279         .poll           = fuse_dev_poll,
2280         .release        = fuse_dev_release,
2281         .fasync         = fuse_dev_fasync,
2282         .unlocked_ioctl = fuse_dev_ioctl,
2283         .compat_ioctl   = fuse_dev_ioctl,
2284 };
2285 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2286
2287 static struct miscdevice fuse_miscdevice = {
2288         .minor = FUSE_MINOR,
2289         .name  = "fuse",
2290         .fops = &fuse_dev_operations,
2291 };
2292
2293 int __init fuse_dev_init(void)
2294 {
2295         int err = -ENOMEM;
2296         fuse_req_cachep = kmem_cache_create("fuse_request",
2297                                             sizeof(struct fuse_req),
2298                                             0, 0, NULL);
2299         if (!fuse_req_cachep)
2300                 goto out;
2301
2302         err = misc_register(&fuse_miscdevice);
2303         if (err)
2304                 goto out_cache_clean;
2305
2306         return 0;
2307
2308  out_cache_clean:
2309         kmem_cache_destroy(fuse_req_cachep);
2310  out:
2311         return err;
2312 }
2313
2314 void fuse_dev_cleanup(void)
2315 {
2316         misc_deregister(&fuse_miscdevice);
2317         kmem_cache_destroy(fuse_req_cachep);
2318 }