fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   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_conn *fuse_get_conn(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 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         page_cache_get(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                 page_cache_release(newpage);
 916                 return err;
 917         }
 918
 919         unlock_page(oldpage);
 920         page_cache_release(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         page_cache_get(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_conn *fc, struct file *file,
1236                                 struct fuse_copy_state *cs, size_t nbytes)
1237 {
1238         ssize_t err;
1239         struct fuse_iqueue *fiq = &fc->iq;
1240         struct fuse_pqueue *fpq = &fc->pq;
1241         struct fuse_req *req;
1242         struct fuse_in *in;
1243         unsigned reqsize;
1244
1245  restart:
1246         spin_lock(&fiq->waitq.lock);
1247         err = -EAGAIN;
1248         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1249             !request_pending(fiq))
1250                 goto err_unlock;
1251
1252         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1253                                 !fiq->connected || request_pending(fiq));
1254         if (err)
1255                 goto err_unlock;
1256
1257         err = -ENODEV;
1258         if (!fiq->connected)
1259                 goto err_unlock;
1260
1261         if (!list_empty(&fiq->interrupts)) {
1262                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1263                                  intr_entry);
1264                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1265         }
1266
1267         if (forget_pending(fiq)) {
1268                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1269                         return fuse_read_forget(fc, fiq, cs, nbytes);
1270
1271                 if (fiq->forget_batch <= -8)
1272                         fiq->forget_batch = 16;
1273         }
1274
1275         req = list_entry(fiq->pending.next, struct fuse_req, list);
1276         clear_bit(FR_PENDING, &req->flags);
1277         list_del_init(&req->list);
1278         spin_unlock(&fiq->waitq.lock);
1279
1280         spin_lock(&fc->lock);
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                 spin_unlock(&fc->lock);
1290                 request_end(fc, req);
1291                 goto restart;
1292         }
1293         spin_lock(&fpq->lock);
1294         list_add(&req->list, &fpq->io);
1295         spin_unlock(&fpq->lock);
1296         spin_unlock(&fc->lock);
1297         cs->req = req;
1298         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1299         if (!err)
1300                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1301                                      (struct fuse_arg *) in->args, 0);
1302         fuse_copy_finish(cs);
1303         spin_lock(&fc->lock);
1304         spin_lock(&fpq->lock);
1305         clear_bit(FR_LOCKED, &req->flags);
1306         if (!fpq->connected) {
1307                 err = -ENODEV;
1308                 goto out_end;
1309         }
1310         if (err) {
1311                 req->out.h.error = -EIO;
1312                 goto out_end;
1313         }
1314         if (!test_bit(FR_ISREPLY, &req->flags)) {
1315                 err = reqsize;
1316                 goto out_end;
1317         }
1318         list_move_tail(&req->list, &fpq->processing);
1319         spin_unlock(&fpq->lock);
1320         set_bit(FR_SENT, &req->flags);
1321         /* matches barrier in request_wait_answer() */
1322         smp_mb__after_atomic();
1323         if (test_bit(FR_INTERRUPTED, &req->flags))
1324                 queue_interrupt(fiq, req);
1325         spin_unlock(&fc->lock);
1326
1327         return reqsize;
1328
1329 out_end:
1330         if (!test_bit(FR_PRIVATE, &req->flags))
1331                 list_del_init(&req->list);
1332         spin_unlock(&fpq->lock);
1333         spin_unlock(&fc->lock);
1334         request_end(fc, req);
1335         return err;
1336
1337  err_unlock:
1338         spin_unlock(&fiq->waitq.lock);
1339         return err;
1340 }
1341
1342 static int fuse_dev_open(struct inode *inode, struct file *file)
1343 {
1344         /*
1345          * The fuse device's file's private_data is used to hold
1346          * the fuse_conn(ection) when it is mounted, and is used to
1347          * keep track of whether the file has been mounted already.
1348          */
1349         file->private_data = NULL;
1350         return 0;
1351 }
1352
1353 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1354 {
1355         struct fuse_copy_state cs;
1356         struct file *file = iocb->ki_filp;
1357         struct fuse_conn *fc = fuse_get_conn(file);
1358         if (!fc)
1359                 return -EPERM;
1360
1361         if (!iter_is_iovec(to))
1362                 return -EINVAL;
1363
1364         fuse_copy_init(&cs, 1, to);
1365
1366         return fuse_dev_do_read(fc, file, &cs, iov_iter_count(to));
1367 }
1368
1369 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1370                                     struct pipe_inode_info *pipe,
1371                                     size_t len, unsigned int flags)
1372 {
1373         int ret;
1374         int page_nr = 0;
1375         int do_wakeup = 0;
1376         struct pipe_buffer *bufs;
1377         struct fuse_copy_state cs;
1378         struct fuse_conn *fc = fuse_get_conn(in);
1379         if (!fc)
1380                 return -EPERM;
1381
1382         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1383         if (!bufs)
1384                 return -ENOMEM;
1385
1386         fuse_copy_init(&cs, 1, NULL);
1387         cs.pipebufs = bufs;
1388         cs.pipe = pipe;
1389         ret = fuse_dev_do_read(fc, in, &cs, len);
1390         if (ret < 0)
1391                 goto out;
1392
1393         ret = 0;
1394         pipe_lock(pipe);
1395
1396         if (!pipe->readers) {
1397                 send_sig(SIGPIPE, current, 0);
1398                 if (!ret)
1399                         ret = -EPIPE;
1400                 goto out_unlock;
1401         }
1402
1403         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1404                 ret = -EIO;
1405                 goto out_unlock;
1406         }
1407
1408         while (page_nr < cs.nr_segs) {
1409                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1410                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1411
1412                 buf->page = bufs[page_nr].page;
1413                 buf->offset = bufs[page_nr].offset;
1414                 buf->len = bufs[page_nr].len;
1415                 /*
1416                  * Need to be careful about this.  Having buf->ops in module
1417                  * code can Oops if the buffer persists after module unload.
1418                  */
1419                 buf->ops = &nosteal_pipe_buf_ops;
1420
1421                 pipe->nrbufs++;
1422                 page_nr++;
1423                 ret += buf->len;
1424
1425                 if (pipe->files)
1426                         do_wakeup = 1;
1427         }
1428
1429 out_unlock:
1430         pipe_unlock(pipe);
1431
1432         if (do_wakeup) {
1433                 smp_mb();
1434                 if (waitqueue_active(&pipe->wait))
1435                         wake_up_interruptible(&pipe->wait);
1436                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1437         }
1438
1439 out:
1440         for (; page_nr < cs.nr_segs; page_nr++)
1441                 page_cache_release(bufs[page_nr].page);
1442
1443         kfree(bufs);
1444         return ret;
1445 }
1446
1447 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1448                             struct fuse_copy_state *cs)
1449 {
1450         struct fuse_notify_poll_wakeup_out outarg;
1451         int err = -EINVAL;
1452
1453         if (size != sizeof(outarg))
1454                 goto err;
1455
1456         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1457         if (err)
1458                 goto err;
1459
1460         fuse_copy_finish(cs);
1461         return fuse_notify_poll_wakeup(fc, &outarg);
1462
1463 err:
1464         fuse_copy_finish(cs);
1465         return err;
1466 }
1467
1468 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1469                                    struct fuse_copy_state *cs)
1470 {
1471         struct fuse_notify_inval_inode_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         fuse_copy_finish(cs);
1481
1482         down_read(&fc->killsb);
1483         err = -ENOENT;
1484         if (fc->sb) {
1485                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1486                                                outarg.off, outarg.len);
1487         }
1488         up_read(&fc->killsb);
1489         return err;
1490
1491 err:
1492         fuse_copy_finish(cs);
1493         return err;
1494 }
1495
1496 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1497                                    struct fuse_copy_state *cs)
1498 {
1499         struct fuse_notify_inval_entry_out outarg;
1500         int err = -ENOMEM;
1501         char *buf;
1502         struct qstr name;
1503
1504         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1505         if (!buf)
1506                 goto err;
1507
1508         err = -EINVAL;
1509         if (size < sizeof(outarg))
1510                 goto err;
1511
1512         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1513         if (err)
1514                 goto err;
1515
1516         err = -ENAMETOOLONG;
1517         if (outarg.namelen > FUSE_NAME_MAX)
1518                 goto err;
1519
1520         err = -EINVAL;
1521         if (size != sizeof(outarg) + outarg.namelen + 1)
1522                 goto err;
1523
1524         name.name = buf;
1525         name.len = outarg.namelen;
1526         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1527         if (err)
1528                 goto err;
1529         fuse_copy_finish(cs);
1530         buf[outarg.namelen] = 0;
1531         name.hash = full_name_hash(name.name, name.len);
1532
1533         down_read(&fc->killsb);
1534         err = -ENOENT;
1535         if (fc->sb)
1536                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1537         up_read(&fc->killsb);
1538         kfree(buf);
1539         return err;
1540
1541 err:
1542         kfree(buf);
1543         fuse_copy_finish(cs);
1544         return err;
1545 }
1546
1547 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1548                               struct fuse_copy_state *cs)
1549 {
1550         struct fuse_notify_delete_out outarg;
1551         int err = -ENOMEM;
1552         char *buf;
1553         struct qstr name;
1554
1555         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1556         if (!buf)
1557                 goto err;
1558
1559         err = -EINVAL;
1560         if (size < sizeof(outarg))
1561                 goto err;
1562
1563         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1564         if (err)
1565                 goto err;
1566
1567         err = -ENAMETOOLONG;
1568         if (outarg.namelen > FUSE_NAME_MAX)
1569                 goto err;
1570
1571         err = -EINVAL;
1572         if (size != sizeof(outarg) + outarg.namelen + 1)
1573                 goto err;
1574
1575         name.name = buf;
1576         name.len = outarg.namelen;
1577         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1578         if (err)
1579                 goto err;
1580         fuse_copy_finish(cs);
1581         buf[outarg.namelen] = 0;
1582         name.hash = full_name_hash(name.name, name.len);
1583
1584         down_read(&fc->killsb);
1585         err = -ENOENT;
1586         if (fc->sb)
1587                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1588                                                outarg.child, &name);
1589         up_read(&fc->killsb);
1590         kfree(buf);
1591         return err;
1592
1593 err:
1594         kfree(buf);
1595         fuse_copy_finish(cs);
1596         return err;
1597 }
1598
1599 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1600                              struct fuse_copy_state *cs)
1601 {
1602         struct fuse_notify_store_out outarg;
1603         struct inode *inode;
1604         struct address_space *mapping;
1605         u64 nodeid;
1606         int err;
1607         pgoff_t index;
1608         unsigned int offset;
1609         unsigned int num;
1610         loff_t file_size;
1611         loff_t end;
1612
1613         err = -EINVAL;
1614         if (size < sizeof(outarg))
1615                 goto out_finish;
1616
1617         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1618         if (err)
1619                 goto out_finish;
1620
1621         err = -EINVAL;
1622         if (size - sizeof(outarg) != outarg.size)
1623                 goto out_finish;
1624
1625         nodeid = outarg.nodeid;
1626
1627         down_read(&fc->killsb);
1628
1629         err = -ENOENT;
1630         if (!fc->sb)
1631                 goto out_up_killsb;
1632
1633         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1634         if (!inode)
1635                 goto out_up_killsb;
1636
1637         mapping = inode->i_mapping;
1638         index = outarg.offset >> PAGE_CACHE_SHIFT;
1639         offset = outarg.offset & ~PAGE_CACHE_MASK;
1640         file_size = i_size_read(inode);
1641         end = outarg.offset + outarg.size;
1642         if (end > file_size) {
1643                 file_size = end;
1644                 fuse_write_update_size(inode, file_size);
1645         }
1646
1647         num = outarg.size;
1648         while (num) {
1649                 struct page *page;
1650                 unsigned int this_num;
1651
1652                 err = -ENOMEM;
1653                 page = find_or_create_page(mapping, index,
1654                                            mapping_gfp_mask(mapping));
1655                 if (!page)
1656                         goto out_iput;
1657
1658                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1659                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1660                 if (!err && offset == 0 &&
1661                     (this_num == PAGE_CACHE_SIZE || file_size == end))
1662                         SetPageUptodate(page);
1663                 unlock_page(page);
1664                 page_cache_release(page);
1665
1666                 if (err)
1667                         goto out_iput;
1668
1669                 num -= this_num;
1670                 offset = 0;
1671                 index++;
1672         }
1673
1674         err = 0;
1675
1676 out_iput:
1677         iput(inode);
1678 out_up_killsb:
1679         up_read(&fc->killsb);
1680 out_finish:
1681         fuse_copy_finish(cs);
1682         return err;
1683 }
1684
1685 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1686 {
1687         release_pages(req->pages, req->num_pages, false);
1688 }
1689
1690 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1691                          struct fuse_notify_retrieve_out *outarg)
1692 {
1693         int err;
1694         struct address_space *mapping = inode->i_mapping;
1695         struct fuse_req *req;
1696         pgoff_t index;
1697         loff_t file_size;
1698         unsigned int num;
1699         unsigned int offset;
1700         size_t total_len = 0;
1701         int num_pages;
1702
1703         offset = outarg->offset & ~PAGE_CACHE_MASK;
1704         file_size = i_size_read(inode);
1705
1706         num = outarg->size;
1707         if (outarg->offset > file_size)
1708                 num = 0;
1709         else if (outarg->offset + num > file_size)
1710                 num = file_size - outarg->offset;
1711
1712         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1713         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1714
1715         req = fuse_get_req(fc, num_pages);
1716         if (IS_ERR(req))
1717                 return PTR_ERR(req);
1718
1719         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1720         req->in.h.nodeid = outarg->nodeid;
1721         req->in.numargs = 2;
1722         req->in.argpages = 1;
1723         req->page_descs[0].offset = offset;
1724         req->end = fuse_retrieve_end;
1725
1726         index = outarg->offset >> PAGE_CACHE_SHIFT;
1727
1728         while (num && req->num_pages < num_pages) {
1729                 struct page *page;
1730                 unsigned int this_num;
1731
1732                 page = find_get_page(mapping, index);
1733                 if (!page)
1734                         break;
1735
1736                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1737                 req->pages[req->num_pages] = page;
1738                 req->page_descs[req->num_pages].length = this_num;
1739                 req->num_pages++;
1740
1741                 offset = 0;
1742                 num -= this_num;
1743                 total_len += this_num;
1744                 index++;
1745         }
1746         req->misc.retrieve_in.offset = outarg->offset;
1747         req->misc.retrieve_in.size = total_len;
1748         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1749         req->in.args[0].value = &req->misc.retrieve_in;
1750         req->in.args[1].size = total_len;
1751
1752         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1753         if (err)
1754                 fuse_retrieve_end(fc, req);
1755
1756         return err;
1757 }
1758
1759 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1760                                 struct fuse_copy_state *cs)
1761 {
1762         struct fuse_notify_retrieve_out outarg;
1763         struct inode *inode;
1764         int err;
1765
1766         err = -EINVAL;
1767         if (size != sizeof(outarg))
1768                 goto copy_finish;
1769
1770         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1771         if (err)
1772                 goto copy_finish;
1773
1774         fuse_copy_finish(cs);
1775
1776         down_read(&fc->killsb);
1777         err = -ENOENT;
1778         if (fc->sb) {
1779                 u64 nodeid = outarg.nodeid;
1780
1781                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1782                 if (inode) {
1783                         err = fuse_retrieve(fc, inode, &outarg);
1784                         iput(inode);
1785                 }
1786         }
1787         up_read(&fc->killsb);
1788
1789         return err;
1790
1791 copy_finish:
1792         fuse_copy_finish(cs);
1793         return err;
1794 }
1795
1796 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1797                        unsigned int size, struct fuse_copy_state *cs)
1798 {
1799         /* Don't try to move pages (yet) */
1800         cs->move_pages = 0;
1801
1802         switch (code) {
1803         case FUSE_NOTIFY_POLL:
1804                 return fuse_notify_poll(fc, size, cs);
1805
1806         case FUSE_NOTIFY_INVAL_INODE:
1807                 return fuse_notify_inval_inode(fc, size, cs);
1808
1809         case FUSE_NOTIFY_INVAL_ENTRY:
1810                 return fuse_notify_inval_entry(fc, size, cs);
1811
1812         case FUSE_NOTIFY_STORE:
1813                 return fuse_notify_store(fc, size, cs);
1814
1815         case FUSE_NOTIFY_RETRIEVE:
1816                 return fuse_notify_retrieve(fc, size, cs);
1817
1818         case FUSE_NOTIFY_DELETE:
1819                 return fuse_notify_delete(fc, size, cs);
1820
1821         default:
1822                 fuse_copy_finish(cs);
1823                 return -EINVAL;
1824         }
1825 }
1826
1827 /* Look up request on processing list by unique ID */
1828 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1829 {
1830         struct fuse_req *req;
1831
1832         list_for_each_entry(req, &fpq->processing, list) {
1833                 if (req->in.h.unique == unique || req->intr_unique == unique)
1834                         return req;
1835         }
1836         return NULL;
1837 }
1838
1839 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1840                          unsigned nbytes)
1841 {
1842         unsigned reqsize = sizeof(struct fuse_out_header);
1843
1844         if (out->h.error)
1845                 return nbytes != reqsize ? -EINVAL : 0;
1846
1847         reqsize += len_args(out->numargs, out->args);
1848
1849         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1850                 return -EINVAL;
1851         else if (reqsize > nbytes) {
1852                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1853                 unsigned diffsize = reqsize - nbytes;
1854                 if (diffsize > lastarg->size)
1855                         return -EINVAL;
1856                 lastarg->size -= diffsize;
1857         }
1858         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1859                               out->page_zeroing);
1860 }
1861
1862 /*
1863  * Write a single reply to a request.  First the header is copied from
1864  * the write buffer.  The request is then searched on the processing
1865  * list by the unique ID found in the header.  If found, then remove
1866  * it from the list and copy the rest of the buffer to the request.
1867  * The request is finished by calling request_end()
1868  */
1869 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1870                                  struct fuse_copy_state *cs, size_t nbytes)
1871 {
1872         int err;
1873         struct fuse_pqueue *fpq = &fc->pq;
1874         struct fuse_req *req;
1875         struct fuse_out_header oh;
1876
1877         if (nbytes < sizeof(struct fuse_out_header))
1878                 return -EINVAL;
1879
1880         err = fuse_copy_one(cs, &oh, sizeof(oh));
1881         if (err)
1882                 goto err_finish;
1883
1884         err = -EINVAL;
1885         if (oh.len != nbytes)
1886                 goto err_finish;
1887
1888         /*
1889          * Zero oh.unique indicates unsolicited notification message
1890          * and error contains notification code.
1891          */
1892         if (!oh.unique) {
1893                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1894                 return err ? err : nbytes;
1895         }
1896
1897         err = -EINVAL;
1898         if (oh.error <= -1000 || oh.error > 0)
1899                 goto err_finish;
1900
1901         spin_lock(&fc->lock);
1902         spin_lock(&fpq->lock);
1903         err = -ENOENT;
1904         if (!fpq->connected)
1905                 goto err_unlock_pq;
1906
1907         req = request_find(fpq, oh.unique);
1908         if (!req)
1909                 goto err_unlock_pq;
1910
1911         /* Is it an interrupt reply? */
1912         if (req->intr_unique == oh.unique) {
1913                 spin_unlock(&fpq->lock);
1914
1915                 err = -EINVAL;
1916                 if (nbytes != sizeof(struct fuse_out_header))
1917                         goto err_unlock;
1918
1919                 if (oh.error == -ENOSYS)
1920                         fc->no_interrupt = 1;
1921                 else if (oh.error == -EAGAIN)
1922                         queue_interrupt(&fc->iq, req);
1923
1924                 spin_unlock(&fc->lock);
1925                 fuse_copy_finish(cs);
1926                 return nbytes;
1927         }
1928
1929         clear_bit(FR_SENT, &req->flags);
1930         list_move(&req->list, &fpq->io);
1931         req->out.h = oh;
1932         set_bit(FR_LOCKED, &req->flags);
1933         spin_unlock(&fpq->lock);
1934         cs->req = req;
1935         if (!req->out.page_replace)
1936                 cs->move_pages = 0;
1937         spin_unlock(&fc->lock);
1938
1939         err = copy_out_args(cs, &req->out, nbytes);
1940         fuse_copy_finish(cs);
1941
1942         spin_lock(&fc->lock);
1943         spin_lock(&fpq->lock);
1944         clear_bit(FR_LOCKED, &req->flags);
1945         if (!fpq->connected)
1946                 err = -ENOENT;
1947         else if (err)
1948                 req->out.h.error = -EIO;
1949         if (!test_bit(FR_PRIVATE, &req->flags))
1950                 list_del_init(&req->list);
1951         spin_unlock(&fpq->lock);
1952         spin_unlock(&fc->lock);
1953         request_end(fc, req);
1954
1955         return err ? err : nbytes;
1956
1957  err_unlock_pq:
1958         spin_unlock(&fpq->lock);
1959  err_unlock:
1960         spin_unlock(&fc->lock);
1961  err_finish:
1962         fuse_copy_finish(cs);
1963         return err;
1964 }
1965
1966 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1967 {
1968         struct fuse_copy_state cs;
1969         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1970         if (!fc)
1971                 return -EPERM;
1972
1973         if (!iter_is_iovec(from))
1974                 return -EINVAL;
1975
1976         fuse_copy_init(&cs, 0, from);
1977
1978         return fuse_dev_do_write(fc, &cs, iov_iter_count(from));
1979 }
1980
1981 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1982                                      struct file *out, loff_t *ppos,
1983                                      size_t len, unsigned int flags)
1984 {
1985         unsigned nbuf;
1986         unsigned idx;
1987         struct pipe_buffer *bufs;
1988         struct fuse_copy_state cs;
1989         struct fuse_conn *fc;
1990         size_t rem;
1991         ssize_t ret;
1992
1993         fc = fuse_get_conn(out);
1994         if (!fc)
1995                 return -EPERM;
1996
1997         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1998         if (!bufs)
1999                 return -ENOMEM;
2000
2001         pipe_lock(pipe);
2002         nbuf = 0;
2003         rem = 0;
2004         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2005                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2006
2007         ret = -EINVAL;
2008         if (rem < len) {
2009                 pipe_unlock(pipe);
2010                 goto out;
2011         }
2012
2013         rem = len;
2014         while (rem) {
2015                 struct pipe_buffer *ibuf;
2016                 struct pipe_buffer *obuf;
2017
2018                 BUG_ON(nbuf >= pipe->buffers);
2019                 BUG_ON(!pipe->nrbufs);
2020                 ibuf = &pipe->bufs[pipe->curbuf];
2021                 obuf = &bufs[nbuf];
2022
2023                 if (rem >= ibuf->len) {
2024                         *obuf = *ibuf;
2025                         ibuf->ops = NULL;
2026                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2027                         pipe->nrbufs--;
2028                 } else {
2029                         ibuf->ops->get(pipe, ibuf);
2030                         *obuf = *ibuf;
2031                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2032                         obuf->len = rem;
2033                         ibuf->offset += obuf->len;
2034                         ibuf->len -= obuf->len;
2035                 }
2036                 nbuf++;
2037                 rem -= obuf->len;
2038         }
2039         pipe_unlock(pipe);
2040
2041         fuse_copy_init(&cs, 0, NULL);
2042         cs.pipebufs = bufs;
2043         cs.nr_segs = nbuf;
2044         cs.pipe = pipe;
2045
2046         if (flags & SPLICE_F_MOVE)
2047                 cs.move_pages = 1;
2048
2049         ret = fuse_dev_do_write(fc, &cs, len);
2050
2051         for (idx = 0; idx < nbuf; idx++) {
2052                 struct pipe_buffer *buf = &bufs[idx];
2053                 buf->ops->release(pipe, buf);
2054         }
2055 out:
2056         kfree(bufs);
2057         return ret;
2058 }
2059
2060 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2061 {
2062         unsigned mask = POLLOUT | POLLWRNORM;
2063         struct fuse_iqueue *fiq;
2064         struct fuse_conn *fc = fuse_get_conn(file);
2065         if (!fc)
2066                 return POLLERR;
2067
2068         fiq = &fc->iq;
2069         poll_wait(file, &fiq->waitq, wait);
2070
2071         spin_lock(&fiq->waitq.lock);
2072         if (!fiq->connected)
2073                 mask = POLLERR;
2074         else if (request_pending(fiq))
2075                 mask |= POLLIN | POLLRDNORM;
2076         spin_unlock(&fiq->waitq.lock);
2077
2078         return mask;
2079 }
2080
2081 /*
2082  * Abort all requests on the given list (pending or processing)
2083  *
2084  * This function releases and reacquires fc->lock
2085  */
2086 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2087 __releases(fc->lock)
2088 __acquires(fc->lock)
2089 {
2090         while (!list_empty(head)) {
2091                 struct fuse_req *req;
2092                 req = list_entry(head->next, struct fuse_req, list);
2093                 req->out.h.error = -ECONNABORTED;
2094                 clear_bit(FR_PENDING, &req->flags);
2095                 clear_bit(FR_SENT, &req->flags);
2096                 list_del_init(&req->list);
2097                 spin_unlock(&fc->lock);
2098                 request_end(fc, req);
2099                 spin_lock(&fc->lock);
2100         }
2101 }
2102
2103 static void end_polls(struct fuse_conn *fc)
2104 {
2105         struct rb_node *p;
2106
2107         p = rb_first(&fc->polled_files);
2108
2109         while (p) {
2110                 struct fuse_file *ff;
2111                 ff = rb_entry(p, struct fuse_file, polled_node);
2112                 wake_up_interruptible_all(&ff->poll_wait);
2113
2114                 p = rb_next(p);
2115         }
2116 }
2117
2118 /*
2119  * Abort all requests.
2120  *
2121  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2122  * filesystem.
2123  *
2124  * The same effect is usually achievable through killing the filesystem daemon
2125  * and all users of the filesystem.  The exception is the combination of an
2126  * asynchronous request and the tricky deadlock (see
2127  * Documentation/filesystems/fuse.txt).
2128  *
2129  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2130  * requests, they should be finished off immediately.  Locked requests will be
2131  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2132  * requests.  It is possible that some request will finish before we can.  This
2133  * is OK, the request will in that case be removed from the list before we touch
2134  * it.
2135  */
2136 void fuse_abort_conn(struct fuse_conn *fc)
2137 {
2138         struct fuse_iqueue *fiq = &fc->iq;
2139         struct fuse_pqueue *fpq = &fc->pq;
2140
2141         spin_lock(&fc->lock);
2142         if (fc->connected) {
2143                 struct fuse_req *req, *next;
2144                 LIST_HEAD(to_end1);
2145                 LIST_HEAD(to_end2);
2146
2147                 fc->connected = 0;
2148                 fc->blocked = 0;
2149                 fuse_set_initialized(fc);
2150                 spin_lock(&fpq->lock);
2151                 fpq->connected = 0;
2152                 list_for_each_entry_safe(req, next, &fpq->io, list) {
2153                         req->out.h.error = -ECONNABORTED;
2154                         spin_lock(&req->waitq.lock);
2155                         set_bit(FR_ABORTED, &req->flags);
2156                         if (!test_bit(FR_LOCKED, &req->flags)) {
2157                                 set_bit(FR_PRIVATE, &req->flags);
2158                                 list_move(&req->list, &to_end1);
2159                         }
2160                         spin_unlock(&req->waitq.lock);
2161                 }
2162                 list_splice_init(&fpq->processing, &to_end2);
2163                 spin_unlock(&fpq->lock);
2164                 fc->max_background = UINT_MAX;
2165                 flush_bg_queue(fc);
2166
2167                 spin_lock(&fiq->waitq.lock);
2168                 fiq->connected = 0;
2169                 list_splice_init(&fiq->pending, &to_end2);
2170                 while (forget_pending(fiq))
2171                         kfree(dequeue_forget(fiq, 1, NULL));
2172                 wake_up_all_locked(&fiq->waitq);
2173                 spin_unlock(&fiq->waitq.lock);
2174                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2175
2176                 while (!list_empty(&to_end1)) {
2177                         req = list_first_entry(&to_end1, struct fuse_req, list);
2178                         __fuse_get_request(req);
2179                         list_del_init(&req->list);
2180                         spin_unlock(&fc->lock);
2181                         request_end(fc, req);
2182                         spin_lock(&fc->lock);
2183                 }
2184                 end_requests(fc, &to_end2);
2185                 end_polls(fc);
2186                 wake_up_all(&fc->blocked_waitq);
2187         }
2188         spin_unlock(&fc->lock);
2189 }
2190 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2191
2192 int fuse_dev_release(struct inode *inode, struct file *file)
2193 {
2194         struct fuse_conn *fc = fuse_get_conn(file);
2195         if (fc) {
2196                 WARN_ON(!list_empty(&fc->pq.io));
2197                 WARN_ON(fc->iq.fasync != NULL);
2198                 fuse_abort_conn(fc);
2199                 fuse_conn_put(fc);
2200         }
2201
2202         return 0;
2203 }
2204 EXPORT_SYMBOL_GPL(fuse_dev_release);
2205
2206 static int fuse_dev_fasync(int fd, struct file *file, int on)
2207 {
2208         struct fuse_conn *fc = fuse_get_conn(file);
2209         if (!fc)
2210                 return -EPERM;
2211
2212         /* No locking - fasync_helper does its own locking */
2213         return fasync_helper(fd, file, on, &fc->iq.fasync);
2214 }
2215
2216 const struct file_operations fuse_dev_operations = {
2217         .owner          = THIS_MODULE,
2218         .open           = fuse_dev_open,
2219         .llseek         = no_llseek,
2220         .read_iter      = fuse_dev_read,
2221         .splice_read    = fuse_dev_splice_read,
2222         .write_iter     = fuse_dev_write,
2223         .splice_write   = fuse_dev_splice_write,
2224         .poll           = fuse_dev_poll,
2225         .release        = fuse_dev_release,
2226         .fasync         = fuse_dev_fasync,
2227 };
2228 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2229
2230 static struct miscdevice fuse_miscdevice = {
2231         .minor = FUSE_MINOR,
2232         .name  = "fuse",
2233         .fops = &fuse_dev_operations,
2234 };
2235
2236 int __init fuse_dev_init(void)
2237 {
2238         int err = -ENOMEM;
2239         fuse_req_cachep = kmem_cache_create("fuse_request",
2240                                             sizeof(struct fuse_req),
2241                                             0, 0, NULL);
2242         if (!fuse_req_cachep)
2243                 goto out;
2244
2245         err = misc_register(&fuse_miscdevice);
2246         if (err)
2247                 goto out_cache_clean;
2248
2249         return 0;
2250
2251  out_cache_clean:
2252         kmem_cache_destroy(fuse_req_cachep);
2253  out:
2254         return err;
2255 }
2256
2257 void fuse_dev_cleanup(void)
2258 {
2259         misc_deregister(&fuse_miscdevice);
2260         kmem_cache_destroy(fuse_req_cachep);
2261 }