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/sched/signal.h>
  15 #include <linux/uio.h>
  16 #include <linux/miscdevice.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/pipe_fs_i.h>
  21 #include <linux/swap.h>
  22 #include <linux/splice.h>
  23 #include <linux/sched.h>
  24
  25 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  26 MODULE_ALIAS("devname:fuse");
  27
  28 static struct kmem_cache *fuse_req_cachep;
  29
  30 static struct fuse_dev *fuse_get_dev(struct file *file)
  31 {
  32         /*
  33          * Lockless access is OK, because file->private data is set
  34          * once during mount and is valid until the file is released.
  35          */
  36         return ACCESS_ONCE(file->private_data);
  37 }
  38
  39 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  40                               struct fuse_page_desc *page_descs,
  41                               unsigned npages)
  42 {
  43         memset(req, 0, sizeof(*req));
  44         memset(pages, 0, sizeof(*pages) * npages);
  45         memset(page_descs, 0, sizeof(*page_descs) * npages);
  46         INIT_LIST_HEAD(&req->list);
  47         INIT_LIST_HEAD(&req->intr_entry);
  48         init_waitqueue_head(&req->waitq);
  49         refcount_set(&req->count, 1);
  50         req->pages = pages;
  51         req->page_descs = page_descs;
  52         req->max_pages = npages;
  53         __set_bit(FR_PENDING, &req->flags);
  54 }
  55
  56 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  57 {
  58         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, flags);
  59         if (req) {
  60                 struct page **pages;
  61                 struct fuse_page_desc *page_descs;
  62
  63                 if (npages <= FUSE_REQ_INLINE_PAGES) {
  64                         pages = req->inline_pages;
  65                         page_descs = req->inline_page_descs;
  66                 } else {
  67                         pages = kmalloc(sizeof(struct page *) * npages, flags);
  68                         page_descs = kmalloc(sizeof(struct fuse_page_desc) *
  69                                              npages, flags);
  70                 }
  71
  72                 if (!pages || !page_descs) {
  73                         kfree(pages);
  74                         kfree(page_descs);
  75                         kmem_cache_free(fuse_req_cachep, req);
  76                         return NULL;
  77                 }
  78
  79                 fuse_request_init(req, pages, page_descs, npages);
  80         }
  81         return req;
  82 }
  83
  84 struct fuse_req *fuse_request_alloc(unsigned npages)
  85 {
  86         return __fuse_request_alloc(npages, GFP_KERNEL);
  87 }
  88 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  89
  90 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  91 {
  92         return __fuse_request_alloc(npages, GFP_NOFS);
  93 }
  94
  95 void fuse_request_free(struct fuse_req *req)
  96 {
  97         if (req->pages != req->inline_pages) {
  98                 kfree(req->pages);
  99                 kfree(req->page_descs);
 100         }
 101         kmem_cache_free(fuse_req_cachep, req);
 102 }
 103
 104 void __fuse_get_request(struct fuse_req *req)
 105 {
 106         refcount_inc(&req->count);
 107 }
 108
 109 /* Must be called with > 1 refcount */
 110 static void __fuse_put_request(struct fuse_req *req)
 111 {
 112         refcount_dec(&req->count);
 113 }
 114
 115 static void fuse_req_init_context(struct fuse_conn *fc, struct fuse_req *req)
 116 {
 117         req->in.h.uid = from_kuid_munged(&init_user_ns, current_fsuid());
 118         req->in.h.gid = from_kgid_munged(&init_user_ns, current_fsgid());
 119         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 120 }
 121
 122 void fuse_set_initialized(struct fuse_conn *fc)
 123 {
 124         /* Make sure stores before this are seen on another CPU */
 125         smp_wmb();
 126         fc->initialized = 1;
 127 }
 128
 129 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 130 {
 131         return !fc->initialized || (for_background && fc->blocked);
 132 }
 133
 134 static void fuse_drop_waiting(struct fuse_conn *fc)
 135 {
 136         /*
 137          * lockess check of fc->connected is okay, because atomic_dec_and_test()
 138          * provides a memory barrier mached with the one in fuse_wait_aborted()
 139          * to ensure no wake-up is missed.
 140          */
 141         if (atomic_dec_and_test(&fc->num_waiting) &&
 142             !READ_ONCE(fc->connected)) {
 143                 /* wake up aborters */
 144                 wake_up_all(&fc->blocked_waitq);
 145         }
 146 }
 147
 148 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 149                                        bool for_background)
 150 {
 151         struct fuse_req *req;
 152         int err;
 153         atomic_inc(&fc->num_waiting);
 154
 155         if (fuse_block_alloc(fc, for_background)) {
 156                 err = -EINTR;
 157                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 158                                 !fuse_block_alloc(fc, for_background)))
 159                         goto out;
 160         }
 161         /* Matches smp_wmb() in fuse_set_initialized() */
 162         smp_rmb();
 163
 164         err = -ENOTCONN;
 165         if (!fc->connected)
 166                 goto out;
 167
 168         err = -ECONNREFUSED;
 169         if (fc->conn_error)
 170                 goto out;
 171
 172         req = fuse_request_alloc(npages);
 173         err = -ENOMEM;
 174         if (!req) {
 175                 if (for_background)
 176                         wake_up(&fc->blocked_waitq);
 177                 goto out;
 178         }
 179
 180         fuse_req_init_context(fc, req);
 181         __set_bit(FR_WAITING, &req->flags);
 182         if (for_background)
 183                 __set_bit(FR_BACKGROUND, &req->flags);
 184
 185         return req;
 186
 187  out:
 188         fuse_drop_waiting(fc);
 189         return ERR_PTR(err);
 190 }
 191
 192 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 193 {
 194         return __fuse_get_req(fc, npages, false);
 195 }
 196 EXPORT_SYMBOL_GPL(fuse_get_req);
 197
 198 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 199                                              unsigned npages)
 200 {
 201         return __fuse_get_req(fc, npages, true);
 202 }
 203 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 204
 205 /*
 206  * Return request in fuse_file->reserved_req.  However that may
 207  * currently be in use.  If that is the case, wait for it to become
 208  * available.
 209  */
 210 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 211                                          struct file *file)
 212 {
 213         struct fuse_req *req = NULL;
 214         struct fuse_file *ff = file->private_data;
 215
 216         do {
 217                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 218                 spin_lock(&fc->lock);
 219                 if (ff->reserved_req) {
 220                         req = ff->reserved_req;
 221                         ff->reserved_req = NULL;
 222                         req->stolen_file = get_file(file);
 223                 }
 224                 spin_unlock(&fc->lock);
 225         } while (!req);
 226
 227         return req;
 228 }
 229
 230 /*
 231  * Put stolen request back into fuse_file->reserved_req
 232  */
 233 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 234 {
 235         struct file *file = req->stolen_file;
 236         struct fuse_file *ff = file->private_data;
 237
 238         spin_lock(&fc->lock);
 239         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 240         BUG_ON(ff->reserved_req);
 241         ff->reserved_req = req;
 242         wake_up_all(&fc->reserved_req_waitq);
 243         spin_unlock(&fc->lock);
 244         fput(file);
 245 }
 246
 247 /*
 248  * Gets a requests for a file operation, always succeeds
 249  *
 250  * This is used for sending the FLUSH request, which must get to
 251  * userspace, due to POSIX locks which may need to be unlocked.
 252  *
 253  * If allocation fails due to OOM, use the reserved request in
 254  * fuse_file.
 255  *
 256  * This is very unlikely to deadlock accidentally, since the
 257  * filesystem should not have it's own file open.  If deadlock is
 258  * intentional, it can still be broken by "aborting" the filesystem.
 259  */
 260 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 261                                              struct file *file)
 262 {
 263         struct fuse_req *req;
 264
 265         atomic_inc(&fc->num_waiting);
 266         wait_event(fc->blocked_waitq, fc->initialized);
 267         /* Matches smp_wmb() in fuse_set_initialized() */
 268         smp_rmb();
 269         req = fuse_request_alloc(0);
 270         if (!req)
 271                 req = get_reserved_req(fc, file);
 272
 273         fuse_req_init_context(fc, req);
 274         __set_bit(FR_WAITING, &req->flags);
 275         __clear_bit(FR_BACKGROUND, &req->flags);
 276         return req;
 277 }
 278
 279 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 280 {
 281         if (refcount_dec_and_test(&req->count)) {
 282                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 283                         /*
 284                          * We get here in the unlikely case that a background
 285                          * request was allocated but not sent
 286                          */
 287                         spin_lock(&fc->lock);
 288                         if (!fc->blocked)
 289                                 wake_up(&fc->blocked_waitq);
 290                         spin_unlock(&fc->lock);
 291                 }
 292
 293                 if (test_bit(FR_WAITING, &req->flags)) {
 294                         __clear_bit(FR_WAITING, &req->flags);
 295                         fuse_drop_waiting(fc);
 296                 }
 297
 298                 if (req->stolen_file)
 299                         put_reserved_req(fc, req);
 300                 else
 301                         fuse_request_free(req);
 302         }
 303 }
 304 EXPORT_SYMBOL_GPL(fuse_put_request);
 305
 306 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 307 {
 308         unsigned nbytes = 0;
 309         unsigned i;
 310
 311         for (i = 0; i < numargs; i++)
 312                 nbytes += args[i].size;
 313
 314         return nbytes;
 315 }
 316
 317 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 318 {
 319         return ++fiq->reqctr;
 320 }
 321
 322 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 323 {
 324         req->in.h.len = sizeof(struct fuse_in_header) +
 325                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 326         list_add_tail(&req->list, &fiq->pending);
 327         wake_up_locked(&fiq->waitq);
 328         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 329 }
 330
 331 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 332                        u64 nodeid, u64 nlookup)
 333 {
 334         struct fuse_iqueue *fiq = &fc->iq;
 335
 336         forget->forget_one.nodeid = nodeid;
 337         forget->forget_one.nlookup = nlookup;
 338
 339         spin_lock(&fiq->waitq.lock);
 340         if (fiq->connected) {
 341                 fiq->forget_list_tail->next = forget;
 342                 fiq->forget_list_tail = forget;
 343                 wake_up_locked(&fiq->waitq);
 344                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 345         } else {
 346                 kfree(forget);
 347         }
 348         spin_unlock(&fiq->waitq.lock);
 349 }
 350
 351 static void flush_bg_queue(struct fuse_conn *fc)
 352 {
 353         while (fc->active_background < fc->max_background &&
 354                !list_empty(&fc->bg_queue)) {
 355                 struct fuse_req *req;
 356                 struct fuse_iqueue *fiq = &fc->iq;
 357
 358                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 359                 list_del(&req->list);
 360                 fc->active_background++;
 361                 spin_lock(&fiq->waitq.lock);
 362                 req->in.h.unique = fuse_get_unique(fiq);
 363                 queue_request(fiq, req);
 364                 spin_unlock(&fiq->waitq.lock);
 365         }
 366 }
 367
 368 /*
 369  * This function is called when a request is finished.  Either a reply
 370  * has arrived or it was aborted (and not yet sent) or some error
 371  * occurred during communication with userspace, or the device file
 372  * was closed.  The requester thread is woken up (if still waiting),
 373  * the 'end' callback is called if given, else the reference to the
 374  * request is released
 375  */
 376 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 377 {
 378         struct fuse_iqueue *fiq = &fc->iq;
 379
 380         if (test_and_set_bit(FR_FINISHED, &req->flags))
 381                 goto put_request;
 382
 383         spin_lock(&fiq->waitq.lock);
 384         list_del_init(&req->intr_entry);
 385         spin_unlock(&fiq->waitq.lock);
 386         WARN_ON(test_bit(FR_PENDING, &req->flags));
 387         WARN_ON(test_bit(FR_SENT, &req->flags));
 388         if (test_bit(FR_BACKGROUND, &req->flags)) {
 389                 spin_lock(&fc->lock);
 390                 clear_bit(FR_BACKGROUND, &req->flags);
 391                 if (fc->num_background == fc->max_background) {
 392                         fc->blocked = 0;
 393                         wake_up(&fc->blocked_waitq);
 394                 } else if (!fc->blocked) {
 395                         /*
 396                          * Wake up next waiter, if any.  It's okay to use
 397                          * waitqueue_active(), as we've already synced up
 398                          * fc->blocked with waiters with the wake_up() call
 399                          * above.
 400                          */
 401                         if (waitqueue_active(&fc->blocked_waitq))
 402                                 wake_up(&fc->blocked_waitq);
 403                 }
 404
 405                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 406                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 407                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 408                 }
 409                 fc->num_background--;
 410                 fc->active_background--;
 411                 flush_bg_queue(fc);
 412                 spin_unlock(&fc->lock);
 413         }
 414         wake_up(&req->waitq);
 415         if (req->end)
 416                 req->end(fc, req);
 417 put_request:
 418         fuse_put_request(fc, req);
 419 }
 420
 421 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 422 {
 423         spin_lock(&fiq->waitq.lock);
 424         if (test_bit(FR_FINISHED, &req->flags)) {
 425                 spin_unlock(&fiq->waitq.lock);
 426                 return;
 427         }
 428         if (list_empty(&req->intr_entry)) {
 429                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 430                 wake_up_locked(&fiq->waitq);
 431         }
 432         spin_unlock(&fiq->waitq.lock);
 433         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 434 }
 435
 436 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 437 {
 438         struct fuse_iqueue *fiq = &fc->iq;
 439         int err;
 440
 441         if (!fc->no_interrupt) {
 442                 /* Any signal may interrupt this */
 443                 err = wait_event_interruptible(req->waitq,
 444                                         test_bit(FR_FINISHED, &req->flags));
 445                 if (!err)
 446                         return;
 447
 448                 set_bit(FR_INTERRUPTED, &req->flags);
 449                 /* matches barrier in fuse_dev_do_read() */
 450                 smp_mb__after_atomic();
 451                 if (test_bit(FR_SENT, &req->flags))
 452                         queue_interrupt(fiq, req);
 453         }
 454
 455         if (!test_bit(FR_FORCE, &req->flags)) {
 456                 /* Only fatal signals may interrupt this */
 457                 err = wait_event_killable(req->waitq,
 458                                         test_bit(FR_FINISHED, &req->flags));
 459                 if (!err)
 460                         return;
 461
 462                 spin_lock(&fiq->waitq.lock);
 463                 /* Request is not yet in userspace, bail out */
 464                 if (test_bit(FR_PENDING, &req->flags)) {
 465                         list_del(&req->list);
 466                         spin_unlock(&fiq->waitq.lock);
 467                         __fuse_put_request(req);
 468                         req->out.h.error = -EINTR;
 469                         return;
 470                 }
 471                 spin_unlock(&fiq->waitq.lock);
 472         }
 473
 474         /*
 475          * Either request is already in userspace, or it was forced.
 476          * Wait it out.
 477          */
 478         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 479 }
 480
 481 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 482 {
 483         struct fuse_iqueue *fiq = &fc->iq;
 484
 485         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 486         spin_lock(&fiq->waitq.lock);
 487         if (!fiq->connected) {
 488                 spin_unlock(&fiq->waitq.lock);
 489                 req->out.h.error = -ENOTCONN;
 490         } else {
 491                 req->in.h.unique = fuse_get_unique(fiq);
 492                 queue_request(fiq, req);
 493                 /* acquire extra reference, since request is still needed
 494                    after request_end() */
 495                 __fuse_get_request(req);
 496                 spin_unlock(&fiq->waitq.lock);
 497
 498                 request_wait_answer(fc, req);
 499                 /* Pairs with smp_wmb() in request_end() */
 500                 smp_rmb();
 501         }
 502 }
 503
 504 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 505 {
 506         __set_bit(FR_ISREPLY, &req->flags);
 507         if (!test_bit(FR_WAITING, &req->flags)) {
 508                 __set_bit(FR_WAITING, &req->flags);
 509                 atomic_inc(&fc->num_waiting);
 510         }
 511         __fuse_request_send(fc, req);
 512 }
 513 EXPORT_SYMBOL_GPL(fuse_request_send);
 514
 515 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 516 {
 517         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 518                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 519
 520         if (fc->minor < 9) {
 521                 switch (args->in.h.opcode) {
 522                 case FUSE_LOOKUP:
 523                 case FUSE_CREATE:
 524                 case FUSE_MKNOD:
 525                 case FUSE_MKDIR:
 526                 case FUSE_SYMLINK:
 527                 case FUSE_LINK:
 528                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 529                         break;
 530                 case FUSE_GETATTR:
 531                 case FUSE_SETATTR:
 532                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 533                         break;
 534                 }
 535         }
 536         if (fc->minor < 12) {
 537                 switch (args->in.h.opcode) {
 538                 case FUSE_CREATE:
 539                         args->in.args[0].size = sizeof(struct fuse_open_in);
 540                         break;
 541                 case FUSE_MKNOD:
 542                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 543                         break;
 544                 }
 545         }
 546 }
 547
 548 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 549 {
 550         struct fuse_req *req;
 551         ssize_t ret;
 552
 553         req = fuse_get_req(fc, 0);
 554         if (IS_ERR(req))
 555                 return PTR_ERR(req);
 556
 557         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 558         fuse_adjust_compat(fc, args);
 559
 560         req->in.h.opcode = args->in.h.opcode;
 561         req->in.h.nodeid = args->in.h.nodeid;
 562         req->in.numargs = args->in.numargs;
 563         memcpy(req->in.args, args->in.args,
 564                args->in.numargs * sizeof(struct fuse_in_arg));
 565         req->out.argvar = args->out.argvar;
 566         req->out.numargs = args->out.numargs;
 567         memcpy(req->out.args, args->out.args,
 568                args->out.numargs * sizeof(struct fuse_arg));
 569         fuse_request_send(fc, req);
 570         ret = req->out.h.error;
 571         if (!ret && args->out.argvar) {
 572                 BUG_ON(args->out.numargs != 1);
 573                 ret = req->out.args[0].size;
 574         }
 575         fuse_put_request(fc, req);
 576
 577         return ret;
 578 }
 579
 580 /*
 581  * Called under fc->lock
 582  *
 583  * fc->connected must have been checked previously
 584  */
 585 void fuse_request_send_background_locked(struct fuse_conn *fc,
 586                                          struct fuse_req *req)
 587 {
 588         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 589         if (!test_bit(FR_WAITING, &req->flags)) {
 590                 __set_bit(FR_WAITING, &req->flags);
 591                 atomic_inc(&fc->num_waiting);
 592         }
 593         __set_bit(FR_ISREPLY, &req->flags);
 594         fc->num_background++;
 595         if (fc->num_background == fc->max_background)
 596                 fc->blocked = 1;
 597         if (fc->num_background == fc->congestion_threshold && fc->sb) {
 598                 set_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 599                 set_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 600         }
 601         list_add_tail(&req->list, &fc->bg_queue);
 602         flush_bg_queue(fc);
 603 }
 604
 605 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 606 {
 607         BUG_ON(!req->end);
 608         spin_lock(&fc->lock);
 609         if (fc->connected) {
 610                 fuse_request_send_background_locked(fc, req);
 611                 spin_unlock(&fc->lock);
 612         } else {
 613                 spin_unlock(&fc->lock);
 614                 req->out.h.error = -ENOTCONN;
 615                 req->end(fc, req);
 616                 fuse_put_request(fc, req);
 617         }
 618 }
 619 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 620
 621 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 622                                           struct fuse_req *req, u64 unique)
 623 {
 624         int err = -ENODEV;
 625         struct fuse_iqueue *fiq = &fc->iq;
 626
 627         __clear_bit(FR_ISREPLY, &req->flags);
 628         req->in.h.unique = unique;
 629         spin_lock(&fiq->waitq.lock);
 630         if (fiq->connected) {
 631                 queue_request(fiq, req);
 632                 err = 0;
 633         }
 634         spin_unlock(&fiq->waitq.lock);
 635
 636         return err;
 637 }
 638
 639 void fuse_force_forget(struct file *file, u64 nodeid)
 640 {
 641         struct inode *inode = file_inode(file);
 642         struct fuse_conn *fc = get_fuse_conn(inode);
 643         struct fuse_req *req;
 644         struct fuse_forget_in inarg;
 645
 646         memset(&inarg, 0, sizeof(inarg));
 647         inarg.nlookup = 1;
 648         req = fuse_get_req_nofail_nopages(fc, file);
 649         req->in.h.opcode = FUSE_FORGET;
 650         req->in.h.nodeid = nodeid;
 651         req->in.numargs = 1;
 652         req->in.args[0].size = sizeof(inarg);
 653         req->in.args[0].value = &inarg;
 654         __clear_bit(FR_ISREPLY, &req->flags);
 655         __fuse_request_send(fc, req);
 656         /* ignore errors */
 657         fuse_put_request(fc, req);
 658 }
 659
 660 /*
 661  * Lock the request.  Up to the next unlock_request() there mustn't be
 662  * anything that could cause a page-fault.  If the request was already
 663  * aborted bail out.
 664  */
 665 static int lock_request(struct fuse_req *req)
 666 {
 667         int err = 0;
 668         if (req) {
 669                 spin_lock(&req->waitq.lock);
 670                 if (test_bit(FR_ABORTED, &req->flags))
 671                         err = -ENOENT;
 672                 else
 673                         set_bit(FR_LOCKED, &req->flags);
 674                 spin_unlock(&req->waitq.lock);
 675         }
 676         return err;
 677 }
 678
 679 /*
 680  * Unlock request.  If it was aborted while locked, caller is responsible
 681  * for unlocking and ending the request.
 682  */
 683 static int unlock_request(struct fuse_req *req)
 684 {
 685         int err = 0;
 686         if (req) {
 687                 spin_lock(&req->waitq.lock);
 688                 if (test_bit(FR_ABORTED, &req->flags))
 689                         err = -ENOENT;
 690                 else
 691                         clear_bit(FR_LOCKED, &req->flags);
 692                 spin_unlock(&req->waitq.lock);
 693         }
 694         return err;
 695 }
 696
 697 struct fuse_copy_state {
 698         int write;
 699         struct fuse_req *req;
 700         struct iov_iter *iter;
 701         struct pipe_buffer *pipebufs;
 702         struct pipe_buffer *currbuf;
 703         struct pipe_inode_info *pipe;
 704         unsigned long nr_segs;
 705         struct page *pg;
 706         unsigned len;
 707         unsigned offset;
 708         unsigned move_pages:1;
 709 };
 710
 711 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 712                            struct iov_iter *iter)
 713 {
 714         memset(cs, 0, sizeof(*cs));
 715         cs->write = write;
 716         cs->iter = iter;
 717 }
 718
 719 /* Unmap and put previous page of userspace buffer */
 720 static void fuse_copy_finish(struct fuse_copy_state *cs)
 721 {
 722         if (cs->currbuf) {
 723                 struct pipe_buffer *buf = cs->currbuf;
 724
 725                 if (cs->write)
 726                         buf->len = PAGE_SIZE - cs->len;
 727                 cs->currbuf = NULL;
 728         } else if (cs->pg) {
 729                 if (cs->write) {
 730                         flush_dcache_page(cs->pg);
 731                         set_page_dirty_lock(cs->pg);
 732                 }
 733                 put_page(cs->pg);
 734         }
 735         cs->pg = NULL;
 736 }
 737
 738 /*
 739  * Get another pagefull of userspace buffer, and map it to kernel
 740  * address space, and lock request
 741  */
 742 static int fuse_copy_fill(struct fuse_copy_state *cs)
 743 {
 744         struct page *page;
 745         int err;
 746
 747         err = unlock_request(cs->req);
 748         if (err)
 749                 return err;
 750
 751         fuse_copy_finish(cs);
 752         if (cs->pipebufs) {
 753                 struct pipe_buffer *buf = cs->pipebufs;
 754
 755                 if (!cs->write) {
 756                         err = pipe_buf_confirm(cs->pipe, buf);
 757                         if (err)
 758                                 return err;
 759
 760                         BUG_ON(!cs->nr_segs);
 761                         cs->currbuf = buf;
 762                         cs->pg = buf->page;
 763                         cs->offset = buf->offset;
 764                         cs->len = buf->len;
 765                         cs->pipebufs++;
 766                         cs->nr_segs--;
 767                 } else {
 768                         if (cs->nr_segs == cs->pipe->buffers)
 769                                 return -EIO;
 770
 771                         page = alloc_page(GFP_HIGHUSER);
 772                         if (!page)
 773                                 return -ENOMEM;
 774
 775                         buf->page = page;
 776                         buf->offset = 0;
 777                         buf->len = 0;
 778
 779                         cs->currbuf = buf;
 780                         cs->pg = page;
 781                         cs->offset = 0;
 782                         cs->len = PAGE_SIZE;
 783                         cs->pipebufs++;
 784                         cs->nr_segs++;
 785                 }
 786         } else {
 787                 size_t off;
 788                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 789                 if (err < 0)
 790                         return err;
 791                 BUG_ON(!err);
 792                 cs->len = err;
 793                 cs->offset = off;
 794                 cs->pg = page;
 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->flags & PAGE_FLAGS_CHECK_AT_PREP &
 828              ~(1 << PG_locked |
 829                1 << PG_referenced |
 830                1 << PG_uptodate |
 831                1 << PG_lru |
 832                1 << PG_active |
 833                1 << PG_reclaim))) {
 834                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 835                 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);
 836                 return 1;
 837         }
 838         return 0;
 839 }
 840
 841 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 842 {
 843         int err;
 844         struct page *oldpage = *pagep;
 845         struct page *newpage;
 846         struct pipe_buffer *buf = cs->pipebufs;
 847
 848         get_page(oldpage);
 849         err = unlock_request(cs->req);
 850         if (err)
 851                 goto out_put_old;
 852
 853         fuse_copy_finish(cs);
 854
 855         err = pipe_buf_confirm(cs->pipe, buf);
 856         if (err)
 857                 goto out_put_old;
 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 (pipe_buf_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                 goto out_put_old;
 898         }
 899
 900         get_page(newpage);
 901
 902         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 903                 lru_cache_add_file(newpage);
 904
 905         /*
 906          * Release while we have extra ref on stolen page.  Otherwise
 907          * anon_pipe_buf_release() might think the page can be reused.
 908          */
 909         pipe_buf_release(cs->pipe, buf);
 910
 911         err = 0;
 912         spin_lock(&cs->req->waitq.lock);
 913         if (test_bit(FR_ABORTED, &cs->req->flags))
 914                 err = -ENOENT;
 915         else
 916                 *pagep = newpage;
 917         spin_unlock(&cs->req->waitq.lock);
 918
 919         if (err) {
 920                 unlock_page(newpage);
 921                 put_page(newpage);
 922                 goto out_put_old;
 923         }
 924
 925         unlock_page(oldpage);
 926         /* Drop ref for ap->pages[] array */
 927         put_page(oldpage);
 928         cs->len = 0;
 929
 930         err = 0;
 931 out_put_old:
 932         /* Drop ref obtained in this function */
 933         put_page(oldpage);
 934         return err;
 935
 936 out_fallback_unlock:
 937         unlock_page(newpage);
 938 out_fallback:
 939         cs->pg = buf->page;
 940         cs->offset = buf->offset;
 941
 942         err = lock_request(cs->req);
 943         if (!err)
 944                 err = 1;
 945
 946         goto out_put_old;
 947 }
 948
 949 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 950                          unsigned offset, unsigned count)
 951 {
 952         struct pipe_buffer *buf;
 953         int err;
 954
 955         if (cs->nr_segs == cs->pipe->buffers)
 956                 return -EIO;
 957
 958         get_page(page);
 959         err = unlock_request(cs->req);
 960         if (err) {
 961                 put_page(page);
 962                 return err;
 963         }
 964
 965         fuse_copy_finish(cs);
 966
 967         buf = cs->pipebufs;
 968         buf->page = page;
 969         buf->offset = offset;
 970         buf->len = count;
 971
 972         cs->pipebufs++;
 973         cs->nr_segs++;
 974         cs->len = 0;
 975
 976         return 0;
 977 }
 978
 979 /*
 980  * Copy a page in the request to/from the userspace buffer.  Must be
 981  * done atomically
 982  */
 983 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 984                           unsigned offset, unsigned count, int zeroing)
 985 {
 986         int err;
 987         struct page *page = *pagep;
 988
 989         if (page && zeroing && count < PAGE_SIZE)
 990                 clear_highpage(page);
 991
 992         while (count) {
 993                 if (cs->write && cs->pipebufs && page) {
 994                         /*
 995                          * Can't control lifetime of pipe buffers, so always
 996                          * copy user pages.
 997                          */
 998                         if (cs->req->user_pages) {
 999                                 err = fuse_copy_fill(cs);
1000                                 if (err)
1001                                         return err;
1002                         } else {
1003                                 return fuse_ref_page(cs, page, offset, count);
1004                         }
1005                 } else if (!cs->len) {
1006                         if (cs->move_pages && page &&
1007                             offset == 0 && count == PAGE_SIZE) {
1008                                 err = fuse_try_move_page(cs, pagep);
1009                                 if (err <= 0)
1010                                         return err;
1011                         } else {
1012                                 err = fuse_copy_fill(cs);
1013                                 if (err)
1014                                         return err;
1015                         }
1016                 }
1017                 if (page) {
1018                         void *mapaddr = kmap_atomic(page);
1019                         void *buf = mapaddr + offset;
1020                         offset += fuse_copy_do(cs, &buf, &count);
1021                         kunmap_atomic(mapaddr);
1022                 } else
1023                         offset += fuse_copy_do(cs, NULL, &count);
1024         }
1025         if (page && !cs->write)
1026                 flush_dcache_page(page);
1027         return 0;
1028 }
1029
1030 /* Copy pages in the request to/from userspace buffer */
1031 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1032                            int zeroing)
1033 {
1034         unsigned i;
1035         struct fuse_req *req = cs->req;
1036
1037         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1038                 int err;
1039                 unsigned offset = req->page_descs[i].offset;
1040                 unsigned count = min(nbytes, req->page_descs[i].length);
1041
1042                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1043                                      zeroing);
1044                 if (err)
1045                         return err;
1046
1047                 nbytes -= count;
1048         }
1049         return 0;
1050 }
1051
1052 /* Copy a single argument in the request to/from userspace buffer */
1053 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1054 {
1055         while (size) {
1056                 if (!cs->len) {
1057                         int err = fuse_copy_fill(cs);
1058                         if (err)
1059                                 return err;
1060                 }
1061                 fuse_copy_do(cs, &val, &size);
1062         }
1063         return 0;
1064 }
1065
1066 /* Copy request arguments to/from userspace buffer */
1067 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1068                           unsigned argpages, struct fuse_arg *args,
1069                           int zeroing)
1070 {
1071         int err = 0;
1072         unsigned i;
1073
1074         for (i = 0; !err && i < numargs; i++)  {
1075                 struct fuse_arg *arg = &args[i];
1076                 if (i == numargs - 1 && argpages)
1077                         err = fuse_copy_pages(cs, arg->size, zeroing);
1078                 else
1079                         err = fuse_copy_one(cs, arg->value, arg->size);
1080         }
1081         return err;
1082 }
1083
1084 static int forget_pending(struct fuse_iqueue *fiq)
1085 {
1086         return fiq->forget_list_head.next != NULL;
1087 }
1088
1089 static int request_pending(struct fuse_iqueue *fiq)
1090 {
1091         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1092                 forget_pending(fiq);
1093 }
1094
1095 /*
1096  * Transfer an interrupt request to userspace
1097  *
1098  * Unlike other requests this is assembled on demand, without a need
1099  * to allocate a separate fuse_req structure.
1100  *
1101  * Called with fiq->waitq.lock held, releases it
1102  */
1103 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1104                                struct fuse_copy_state *cs,
1105                                size_t nbytes, struct fuse_req *req)
1106 __releases(fiq->waitq.lock)
1107 {
1108         struct fuse_in_header ih;
1109         struct fuse_interrupt_in arg;
1110         unsigned reqsize = sizeof(ih) + sizeof(arg);
1111         int err;
1112
1113         list_del_init(&req->intr_entry);
1114         req->intr_unique = fuse_get_unique(fiq);
1115         memset(&ih, 0, sizeof(ih));
1116         memset(&arg, 0, sizeof(arg));
1117         ih.len = reqsize;
1118         ih.opcode = FUSE_INTERRUPT;
1119         ih.unique = req->intr_unique;
1120         arg.unique = req->in.h.unique;
1121
1122         spin_unlock(&fiq->waitq.lock);
1123         if (nbytes < reqsize)
1124                 return -EINVAL;
1125
1126         err = fuse_copy_one(cs, &ih, sizeof(ih));
1127         if (!err)
1128                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1129         fuse_copy_finish(cs);
1130
1131         return err ? err : reqsize;
1132 }
1133
1134 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1135                                                unsigned max,
1136                                                unsigned *countp)
1137 {
1138         struct fuse_forget_link *head = fiq->forget_list_head.next;
1139         struct fuse_forget_link **newhead = &head;
1140         unsigned count;
1141
1142         for (count = 0; *newhead != NULL && count < max; count++)
1143                 newhead = &(*newhead)->next;
1144
1145         fiq->forget_list_head.next = *newhead;
1146         *newhead = NULL;
1147         if (fiq->forget_list_head.next == NULL)
1148                 fiq->forget_list_tail = &fiq->forget_list_head;
1149
1150         if (countp != NULL)
1151                 *countp = count;
1152
1153         return head;
1154 }
1155
1156 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1157                                    struct fuse_copy_state *cs,
1158                                    size_t nbytes)
1159 __releases(fiq->waitq.lock)
1160 {
1161         int err;
1162         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1163         struct fuse_forget_in arg = {
1164                 .nlookup = forget->forget_one.nlookup,
1165         };
1166         struct fuse_in_header ih = {
1167                 .opcode = FUSE_FORGET,
1168                 .nodeid = forget->forget_one.nodeid,
1169                 .unique = fuse_get_unique(fiq),
1170                 .len = sizeof(ih) + sizeof(arg),
1171         };
1172
1173         spin_unlock(&fiq->waitq.lock);
1174         kfree(forget);
1175         if (nbytes < ih.len)
1176                 return -EINVAL;
1177
1178         err = fuse_copy_one(cs, &ih, sizeof(ih));
1179         if (!err)
1180                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1181         fuse_copy_finish(cs);
1182
1183         if (err)
1184                 return err;
1185
1186         return ih.len;
1187 }
1188
1189 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1190                                    struct fuse_copy_state *cs, size_t nbytes)
1191 __releases(fiq->waitq.lock)
1192 {
1193         int err;
1194         unsigned max_forgets;
1195         unsigned count;
1196         struct fuse_forget_link *head;
1197         struct fuse_batch_forget_in arg = { .count = 0 };
1198         struct fuse_in_header ih = {
1199                 .opcode = FUSE_BATCH_FORGET,
1200                 .unique = fuse_get_unique(fiq),
1201                 .len = sizeof(ih) + sizeof(arg),
1202         };
1203
1204         if (nbytes < ih.len) {
1205                 spin_unlock(&fiq->waitq.lock);
1206                 return -EINVAL;
1207         }
1208
1209         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1210         head = dequeue_forget(fiq, max_forgets, &count);
1211         spin_unlock(&fiq->waitq.lock);
1212
1213         arg.count = count;
1214         ih.len += count * sizeof(struct fuse_forget_one);
1215         err = fuse_copy_one(cs, &ih, sizeof(ih));
1216         if (!err)
1217                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1218
1219         while (head) {
1220                 struct fuse_forget_link *forget = head;
1221
1222                 if (!err) {
1223                         err = fuse_copy_one(cs, &forget->forget_one,
1224                                             sizeof(forget->forget_one));
1225                 }
1226                 head = forget->next;
1227                 kfree(forget);
1228         }
1229
1230         fuse_copy_finish(cs);
1231
1232         if (err)
1233                 return err;
1234
1235         return ih.len;
1236 }
1237
1238 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1239                             struct fuse_copy_state *cs,
1240                             size_t nbytes)
1241 __releases(fiq->waitq.lock)
1242 {
1243         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1244                 return fuse_read_single_forget(fiq, cs, nbytes);
1245         else
1246                 return fuse_read_batch_forget(fiq, cs, nbytes);
1247 }
1248
1249 /*
1250  * Read a single request into the userspace filesystem's buffer.  This
1251  * function waits until a request is available, then removes it from
1252  * the pending list and copies request data to userspace buffer.  If
1253  * no reply is needed (FORGET) or request has been aborted or there
1254  * was an error during the copying then it's finished by calling
1255  * request_end().  Otherwise add it to the processing list, and set
1256  * the 'sent' flag.
1257  */
1258 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1259                                 struct fuse_copy_state *cs, size_t nbytes)
1260 {
1261         ssize_t err;
1262         struct fuse_conn *fc = fud->fc;
1263         struct fuse_iqueue *fiq = &fc->iq;
1264         struct fuse_pqueue *fpq = &fud->pq;
1265         struct fuse_req *req;
1266         struct fuse_in *in;
1267         unsigned reqsize;
1268
1269  restart:
1270         spin_lock(&fiq->waitq.lock);
1271         err = -EAGAIN;
1272         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1273             !request_pending(fiq))
1274                 goto err_unlock;
1275
1276         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1277                                 !fiq->connected || request_pending(fiq));
1278         if (err)
1279                 goto err_unlock;
1280
1281         err = -ENODEV;
1282         if (!fiq->connected)
1283                 goto err_unlock;
1284
1285         if (!list_empty(&fiq->interrupts)) {
1286                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1287                                  intr_entry);
1288                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1289         }
1290
1291         if (forget_pending(fiq)) {
1292                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1293                         return fuse_read_forget(fc, fiq, cs, nbytes);
1294
1295                 if (fiq->forget_batch <= -8)
1296                         fiq->forget_batch = 16;
1297         }
1298
1299         req = list_entry(fiq->pending.next, struct fuse_req, list);
1300         clear_bit(FR_PENDING, &req->flags);
1301         list_del_init(&req->list);
1302         spin_unlock(&fiq->waitq.lock);
1303
1304         in = &req->in;
1305         reqsize = in->h.len;
1306
1307         if (task_active_pid_ns(current) != fc->pid_ns) {
1308                 rcu_read_lock();
1309                 in->h.pid = pid_vnr(find_pid_ns(in->h.pid, fc->pid_ns));
1310                 rcu_read_unlock();
1311         }
1312
1313         /* If request is too large, reply with an error and restart the read */
1314         if (nbytes < reqsize) {
1315                 req->out.h.error = -EIO;
1316                 /* SETXATTR is special, since it may contain too large data */
1317                 if (in->h.opcode == FUSE_SETXATTR)
1318                         req->out.h.error = -E2BIG;
1319                 request_end(fc, req);
1320                 goto restart;
1321         }
1322         spin_lock(&fpq->lock);
1323         /*
1324          *  Must not put request on fpq->io queue after having been shut down by
1325          *  fuse_abort_conn()
1326          */
1327         if (!fpq->connected) {
1328                 req->out.h.error = err = -ECONNABORTED;
1329                 goto out_end;
1330
1331         }
1332         list_add(&req->list, &fpq->io);
1333         spin_unlock(&fpq->lock);
1334         cs->req = req;
1335         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1336         if (!err)
1337                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1338                                      (struct fuse_arg *) in->args, 0);
1339         fuse_copy_finish(cs);
1340         spin_lock(&fpq->lock);
1341         clear_bit(FR_LOCKED, &req->flags);
1342         if (!fpq->connected) {
1343                 err = -ENODEV;
1344                 goto out_end;
1345         }
1346         if (err) {
1347                 req->out.h.error = -EIO;
1348                 goto out_end;
1349         }
1350         if (!test_bit(FR_ISREPLY, &req->flags)) {
1351                 err = reqsize;
1352                 goto out_end;
1353         }
1354         list_move_tail(&req->list, &fpq->processing);
1355         __fuse_get_request(req);
1356         set_bit(FR_SENT, &req->flags);
1357         spin_unlock(&fpq->lock);
1358         /* matches barrier in request_wait_answer() */
1359         smp_mb__after_atomic();
1360         if (test_bit(FR_INTERRUPTED, &req->flags))
1361                 queue_interrupt(fiq, req);
1362         fuse_put_request(fc, req);
1363
1364         return reqsize;
1365
1366 out_end:
1367         if (!test_bit(FR_PRIVATE, &req->flags))
1368                 list_del_init(&req->list);
1369         spin_unlock(&fpq->lock);
1370         request_end(fc, req);
1371         return err;
1372
1373  err_unlock:
1374         spin_unlock(&fiq->waitq.lock);
1375         return err;
1376 }
1377
1378 static int fuse_dev_open(struct inode *inode, struct file *file)
1379 {
1380         /*
1381          * The fuse device's file's private_data is used to hold
1382          * the fuse_conn(ection) when it is mounted, and is used to
1383          * keep track of whether the file has been mounted already.
1384          */
1385         file->private_data = NULL;
1386         return 0;
1387 }
1388
1389 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1390 {
1391         struct fuse_copy_state cs;
1392         struct file *file = iocb->ki_filp;
1393         struct fuse_dev *fud = fuse_get_dev(file);
1394
1395         if (!fud)
1396                 return -EPERM;
1397
1398         if (!iter_is_iovec(to))
1399                 return -EINVAL;
1400
1401         fuse_copy_init(&cs, 1, to);
1402
1403         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1404 }
1405
1406 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1407                                     struct pipe_inode_info *pipe,
1408                                     size_t len, unsigned int flags)
1409 {
1410         int total, ret;
1411         int page_nr = 0;
1412         struct pipe_buffer *bufs;
1413         struct fuse_copy_state cs;
1414         struct fuse_dev *fud = fuse_get_dev(in);
1415
1416         if (!fud)
1417                 return -EPERM;
1418
1419         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1420         if (!bufs)
1421                 return -ENOMEM;
1422
1423         fuse_copy_init(&cs, 1, NULL);
1424         cs.pipebufs = bufs;
1425         cs.pipe = pipe;
1426         ret = fuse_dev_do_read(fud, in, &cs, len);
1427         if (ret < 0)
1428                 goto out;
1429
1430         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1431                 ret = -EIO;
1432                 goto out;
1433         }
1434
1435         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1436                 /*
1437                  * Need to be careful about this.  Having buf->ops in module
1438                  * code can Oops if the buffer persists after module unload.
1439                  */
1440                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1441                 bufs[page_nr].flags = 0;
1442                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1443                 if (unlikely(ret < 0))
1444                         break;
1445         }
1446         if (total)
1447                 ret = total;
1448 out:
1449         for (; page_nr < cs.nr_segs; page_nr++)
1450                 put_page(bufs[page_nr].page);
1451
1452         kfree(bufs);
1453         return ret;
1454 }
1455
1456 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1457                             struct fuse_copy_state *cs)
1458 {
1459         struct fuse_notify_poll_wakeup_out outarg;
1460         int err = -EINVAL;
1461
1462         if (size != sizeof(outarg))
1463                 goto err;
1464
1465         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1466         if (err)
1467                 goto err;
1468
1469         fuse_copy_finish(cs);
1470         return fuse_notify_poll_wakeup(fc, &outarg);
1471
1472 err:
1473         fuse_copy_finish(cs);
1474         return err;
1475 }
1476
1477 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1478                                    struct fuse_copy_state *cs)
1479 {
1480         struct fuse_notify_inval_inode_out outarg;
1481         int err = -EINVAL;
1482
1483         if (size != sizeof(outarg))
1484                 goto err;
1485
1486         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1487         if (err)
1488                 goto err;
1489         fuse_copy_finish(cs);
1490
1491         down_read(&fc->killsb);
1492         err = -ENOENT;
1493         if (fc->sb) {
1494                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1495                                                outarg.off, outarg.len);
1496         }
1497         up_read(&fc->killsb);
1498         return err;
1499
1500 err:
1501         fuse_copy_finish(cs);
1502         return err;
1503 }
1504
1505 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1506                                    struct fuse_copy_state *cs)
1507 {
1508         struct fuse_notify_inval_entry_out outarg;
1509         int err = -ENOMEM;
1510         char *buf;
1511         struct qstr name;
1512
1513         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1514         if (!buf)
1515                 goto err;
1516
1517         err = -EINVAL;
1518         if (size < sizeof(outarg))
1519                 goto err;
1520
1521         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1522         if (err)
1523                 goto err;
1524
1525         err = -ENAMETOOLONG;
1526         if (outarg.namelen > FUSE_NAME_MAX)
1527                 goto err;
1528
1529         err = -EINVAL;
1530         if (size != sizeof(outarg) + outarg.namelen + 1)
1531                 goto err;
1532
1533         name.name = buf;
1534         name.len = outarg.namelen;
1535         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1536         if (err)
1537                 goto err;
1538         fuse_copy_finish(cs);
1539         buf[outarg.namelen] = 0;
1540
1541         down_read(&fc->killsb);
1542         err = -ENOENT;
1543         if (fc->sb)
1544                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1545         up_read(&fc->killsb);
1546         kfree(buf);
1547         return err;
1548
1549 err:
1550         kfree(buf);
1551         fuse_copy_finish(cs);
1552         return err;
1553 }
1554
1555 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1556                               struct fuse_copy_state *cs)
1557 {
1558         struct fuse_notify_delete_out outarg;
1559         int err = -ENOMEM;
1560         char *buf;
1561         struct qstr name;
1562
1563         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1564         if (!buf)
1565                 goto err;
1566
1567         err = -EINVAL;
1568         if (size < sizeof(outarg))
1569                 goto err;
1570
1571         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1572         if (err)
1573                 goto err;
1574
1575         err = -ENAMETOOLONG;
1576         if (outarg.namelen > FUSE_NAME_MAX)
1577                 goto err;
1578
1579         err = -EINVAL;
1580         if (size != sizeof(outarg) + outarg.namelen + 1)
1581                 goto err;
1582
1583         name.name = buf;
1584         name.len = outarg.namelen;
1585         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1586         if (err)
1587                 goto err;
1588         fuse_copy_finish(cs);
1589         buf[outarg.namelen] = 0;
1590
1591         down_read(&fc->killsb);
1592         err = -ENOENT;
1593         if (fc->sb)
1594                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1595                                                outarg.child, &name);
1596         up_read(&fc->killsb);
1597         kfree(buf);
1598         return err;
1599
1600 err:
1601         kfree(buf);
1602         fuse_copy_finish(cs);
1603         return err;
1604 }
1605
1606 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1607                              struct fuse_copy_state *cs)
1608 {
1609         struct fuse_notify_store_out outarg;
1610         struct inode *inode;
1611         struct address_space *mapping;
1612         u64 nodeid;
1613         int err;
1614         pgoff_t index;
1615         unsigned int offset;
1616         unsigned int num;
1617         loff_t file_size;
1618         loff_t end;
1619
1620         err = -EINVAL;
1621         if (size < sizeof(outarg))
1622                 goto out_finish;
1623
1624         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1625         if (err)
1626                 goto out_finish;
1627
1628         err = -EINVAL;
1629         if (size - sizeof(outarg) != outarg.size)
1630                 goto out_finish;
1631
1632         nodeid = outarg.nodeid;
1633
1634         down_read(&fc->killsb);
1635
1636         err = -ENOENT;
1637         if (!fc->sb)
1638                 goto out_up_killsb;
1639
1640         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1641         if (!inode)
1642                 goto out_up_killsb;
1643
1644         mapping = inode->i_mapping;
1645         index = outarg.offset >> PAGE_SHIFT;
1646         offset = outarg.offset & ~PAGE_MASK;
1647         file_size = i_size_read(inode);
1648         end = outarg.offset + outarg.size;
1649         if (end > file_size) {
1650                 file_size = end;
1651                 fuse_write_update_size(inode, file_size);
1652         }
1653
1654         num = outarg.size;
1655         while (num) {
1656                 struct page *page;
1657                 unsigned int this_num;
1658
1659                 err = -ENOMEM;
1660                 page = find_or_create_page(mapping, index,
1661                                            mapping_gfp_mask(mapping));
1662                 if (!page)
1663                         goto out_iput;
1664
1665                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1666                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1667                 if (!err && offset == 0 &&
1668                     (this_num == PAGE_SIZE || file_size == end))
1669                         SetPageUptodate(page);
1670                 unlock_page(page);
1671                 put_page(page);
1672
1673                 if (err)
1674                         goto out_iput;
1675
1676                 num -= this_num;
1677                 offset = 0;
1678                 index++;
1679         }
1680
1681         err = 0;
1682
1683 out_iput:
1684         iput(inode);
1685 out_up_killsb:
1686         up_read(&fc->killsb);
1687 out_finish:
1688         fuse_copy_finish(cs);
1689         return err;
1690 }
1691
1692 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1693 {
1694         release_pages(req->pages, req->num_pages, false);
1695 }
1696
1697 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1698                          struct fuse_notify_retrieve_out *outarg)
1699 {
1700         int err;
1701         struct address_space *mapping = inode->i_mapping;
1702         struct fuse_req *req;
1703         pgoff_t index;
1704         loff_t file_size;
1705         unsigned int num;
1706         unsigned int offset;
1707         size_t total_len = 0;
1708         int num_pages;
1709
1710         offset = outarg->offset & ~PAGE_MASK;
1711         file_size = i_size_read(inode);
1712
1713         num = min(outarg->size, fc->max_write);
1714         if (outarg->offset > file_size)
1715                 num = 0;
1716         else if (outarg->offset + num > file_size)
1717                 num = file_size - outarg->offset;
1718
1719         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1720         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1721
1722         req = fuse_get_req(fc, num_pages);
1723         if (IS_ERR(req))
1724                 return PTR_ERR(req);
1725
1726         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1727         req->in.h.nodeid = outarg->nodeid;
1728         req->in.numargs = 2;
1729         req->in.argpages = 1;
1730         req->end = fuse_retrieve_end;
1731
1732         index = outarg->offset >> PAGE_SHIFT;
1733
1734         while (num && req->num_pages < num_pages) {
1735                 struct page *page;
1736                 unsigned int this_num;
1737
1738                 page = find_get_page(mapping, index);
1739                 if (!page)
1740                         break;
1741
1742                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1743                 req->pages[req->num_pages] = page;
1744                 req->page_descs[req->num_pages].offset = offset;
1745                 req->page_descs[req->num_pages].length = this_num;
1746                 req->num_pages++;
1747
1748                 offset = 0;
1749                 num -= this_num;
1750                 total_len += this_num;
1751                 index++;
1752         }
1753         req->misc.retrieve_in.offset = outarg->offset;
1754         req->misc.retrieve_in.size = total_len;
1755         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1756         req->in.args[0].value = &req->misc.retrieve_in;
1757         req->in.args[1].size = total_len;
1758
1759         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1760         if (err) {
1761                 fuse_retrieve_end(fc, req);
1762                 fuse_put_request(fc, req);
1763         }
1764
1765         return err;
1766 }
1767
1768 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1769                                 struct fuse_copy_state *cs)
1770 {
1771         struct fuse_notify_retrieve_out outarg;
1772         struct inode *inode;
1773         int err;
1774
1775         err = -EINVAL;
1776         if (size != sizeof(outarg))
1777                 goto copy_finish;
1778
1779         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1780         if (err)
1781                 goto copy_finish;
1782
1783         fuse_copy_finish(cs);
1784
1785         down_read(&fc->killsb);
1786         err = -ENOENT;
1787         if (fc->sb) {
1788                 u64 nodeid = outarg.nodeid;
1789
1790                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1791                 if (inode) {
1792                         err = fuse_retrieve(fc, inode, &outarg);
1793                         iput(inode);
1794                 }
1795         }
1796         up_read(&fc->killsb);
1797
1798         return err;
1799
1800 copy_finish:
1801         fuse_copy_finish(cs);
1802         return err;
1803 }
1804
1805 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1806                        unsigned int size, struct fuse_copy_state *cs)
1807 {
1808         /* Don't try to move pages (yet) */
1809         cs->move_pages = 0;
1810
1811         switch (code) {
1812         case FUSE_NOTIFY_POLL:
1813                 return fuse_notify_poll(fc, size, cs);
1814
1815         case FUSE_NOTIFY_INVAL_INODE:
1816                 return fuse_notify_inval_inode(fc, size, cs);
1817
1818         case FUSE_NOTIFY_INVAL_ENTRY:
1819                 return fuse_notify_inval_entry(fc, size, cs);
1820
1821         case FUSE_NOTIFY_STORE:
1822                 return fuse_notify_store(fc, size, cs);
1823
1824         case FUSE_NOTIFY_RETRIEVE:
1825                 return fuse_notify_retrieve(fc, size, cs);
1826
1827         case FUSE_NOTIFY_DELETE:
1828                 return fuse_notify_delete(fc, size, cs);
1829
1830         default:
1831                 fuse_copy_finish(cs);
1832                 return -EINVAL;
1833         }
1834 }
1835
1836 /* Look up request on processing list by unique ID */
1837 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1838 {
1839         struct fuse_req *req;
1840
1841         list_for_each_entry(req, &fpq->processing, list) {
1842                 if (req->in.h.unique == unique || req->intr_unique == unique)
1843                         return req;
1844         }
1845         return NULL;
1846 }
1847
1848 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1849                          unsigned nbytes)
1850 {
1851         unsigned reqsize = sizeof(struct fuse_out_header);
1852
1853         if (out->h.error)
1854                 return nbytes != reqsize ? -EINVAL : 0;
1855
1856         reqsize += len_args(out->numargs, out->args);
1857
1858         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1859                 return -EINVAL;
1860         else if (reqsize > nbytes) {
1861                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1862                 unsigned diffsize = reqsize - nbytes;
1863                 if (diffsize > lastarg->size)
1864                         return -EINVAL;
1865                 lastarg->size -= diffsize;
1866         }
1867         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1868                               out->page_zeroing);
1869 }
1870
1871 /*
1872  * Write a single reply to a request.  First the header is copied from
1873  * the write buffer.  The request is then searched on the processing
1874  * list by the unique ID found in the header.  If found, then remove
1875  * it from the list and copy the rest of the buffer to the request.
1876  * The request is finished by calling request_end()
1877  */
1878 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1879                                  struct fuse_copy_state *cs, size_t nbytes)
1880 {
1881         int err;
1882         struct fuse_conn *fc = fud->fc;
1883         struct fuse_pqueue *fpq = &fud->pq;
1884         struct fuse_req *req;
1885         struct fuse_out_header oh;
1886
1887         if (nbytes < sizeof(struct fuse_out_header))
1888                 return -EINVAL;
1889
1890         err = fuse_copy_one(cs, &oh, sizeof(oh));
1891         if (err)
1892                 goto err_finish;
1893
1894         err = -EINVAL;
1895         if (oh.len != nbytes)
1896                 goto err_finish;
1897
1898         /*
1899          * Zero oh.unique indicates unsolicited notification message
1900          * and error contains notification code.
1901          */
1902         if (!oh.unique) {
1903                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1904                 return err ? err : nbytes;
1905         }
1906
1907         err = -EINVAL;
1908         if (oh.error <= -512 || oh.error > 0)
1909                 goto err_finish;
1910
1911         spin_lock(&fpq->lock);
1912         err = -ENOENT;
1913         if (!fpq->connected)
1914                 goto err_unlock_pq;
1915
1916         req = request_find(fpq, oh.unique);
1917         if (!req)
1918                 goto err_unlock_pq;
1919
1920         /* Is it an interrupt reply? */
1921         if (req->intr_unique == oh.unique) {
1922                 __fuse_get_request(req);
1923                 spin_unlock(&fpq->lock);
1924
1925                 err = -EINVAL;
1926                 if (nbytes != sizeof(struct fuse_out_header)) {
1927                         fuse_put_request(fc, req);
1928                         goto err_finish;
1929                 }
1930
1931                 if (oh.error == -ENOSYS)
1932                         fc->no_interrupt = 1;
1933                 else if (oh.error == -EAGAIN)
1934                         queue_interrupt(&fc->iq, req);
1935                 fuse_put_request(fc, req);
1936
1937                 fuse_copy_finish(cs);
1938                 return nbytes;
1939         }
1940
1941         clear_bit(FR_SENT, &req->flags);
1942         list_move(&req->list, &fpq->io);
1943         req->out.h = oh;
1944         set_bit(FR_LOCKED, &req->flags);
1945         spin_unlock(&fpq->lock);
1946         cs->req = req;
1947         if (!req->out.page_replace)
1948                 cs->move_pages = 0;
1949
1950         err = copy_out_args(cs, &req->out, nbytes);
1951         fuse_copy_finish(cs);
1952
1953         spin_lock(&fpq->lock);
1954         clear_bit(FR_LOCKED, &req->flags);
1955         if (!fpq->connected)
1956                 err = -ENOENT;
1957         else if (err)
1958                 req->out.h.error = -EIO;
1959         if (!test_bit(FR_PRIVATE, &req->flags))
1960                 list_del_init(&req->list);
1961         spin_unlock(&fpq->lock);
1962
1963         request_end(fc, req);
1964
1965         return err ? err : nbytes;
1966
1967  err_unlock_pq:
1968         spin_unlock(&fpq->lock);
1969  err_finish:
1970         fuse_copy_finish(cs);
1971         return err;
1972 }
1973
1974 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1975 {
1976         struct fuse_copy_state cs;
1977         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1978
1979         if (!fud)
1980                 return -EPERM;
1981
1982         if (!iter_is_iovec(from))
1983                 return -EINVAL;
1984
1985         fuse_copy_init(&cs, 0, from);
1986
1987         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1988 }
1989
1990 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1991                                      struct file *out, loff_t *ppos,
1992                                      size_t len, unsigned int flags)
1993 {
1994         unsigned nbuf;
1995         unsigned idx;
1996         struct pipe_buffer *bufs;
1997         struct fuse_copy_state cs;
1998         struct fuse_dev *fud;
1999         size_t rem;
2000         ssize_t ret;
2001
2002         fud = fuse_get_dev(out);
2003         if (!fud)
2004                 return -EPERM;
2005
2006         pipe_lock(pipe);
2007
2008         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
2009         if (!bufs) {
2010                 pipe_unlock(pipe);
2011                 return -ENOMEM;
2012         }
2013
2014         nbuf = 0;
2015         rem = 0;
2016         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2017                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2018
2019         ret = -EINVAL;
2020         if (rem < len)
2021                 goto out_free;
2022
2023         rem = len;
2024         while (rem) {
2025                 struct pipe_buffer *ibuf;
2026                 struct pipe_buffer *obuf;
2027
2028                 BUG_ON(nbuf >= pipe->buffers);
2029                 BUG_ON(!pipe->nrbufs);
2030                 ibuf = &pipe->bufs[pipe->curbuf];
2031                 obuf = &bufs[nbuf];
2032
2033                 if (rem >= ibuf->len) {
2034                         *obuf = *ibuf;
2035                         ibuf->ops = NULL;
2036                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2037                         pipe->nrbufs--;
2038                 } else {
2039                         if (!pipe_buf_get(pipe, ibuf))
2040                                 goto out_free;
2041
2042                         *obuf = *ibuf;
2043                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2044                         obuf->len = rem;
2045                         ibuf->offset += obuf->len;
2046                         ibuf->len -= obuf->len;
2047                 }
2048                 nbuf++;
2049                 rem -= obuf->len;
2050         }
2051         pipe_unlock(pipe);
2052
2053         fuse_copy_init(&cs, 0, NULL);
2054         cs.pipebufs = bufs;
2055         cs.nr_segs = nbuf;
2056         cs.pipe = pipe;
2057
2058         if (flags & SPLICE_F_MOVE)
2059                 cs.move_pages = 1;
2060
2061         ret = fuse_dev_do_write(fud, &cs, len);
2062
2063         pipe_lock(pipe);
2064 out_free:
2065         for (idx = 0; idx < nbuf; idx++) {
2066                 struct pipe_buffer *buf = &bufs[idx];
2067
2068                 if (buf->ops)
2069                         pipe_buf_release(pipe, buf);
2070         }
2071         pipe_unlock(pipe);
2072
2073         kfree(bufs);
2074         return ret;
2075 }
2076
2077 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2078 {
2079         unsigned mask = POLLOUT | POLLWRNORM;
2080         struct fuse_iqueue *fiq;
2081         struct fuse_dev *fud = fuse_get_dev(file);
2082
2083         if (!fud)
2084                 return POLLERR;
2085
2086         fiq = &fud->fc->iq;
2087         poll_wait(file, &fiq->waitq, wait);
2088
2089         spin_lock(&fiq->waitq.lock);
2090         if (!fiq->connected)
2091                 mask = POLLERR;
2092         else if (request_pending(fiq))
2093                 mask |= POLLIN | POLLRDNORM;
2094         spin_unlock(&fiq->waitq.lock);
2095
2096         return mask;
2097 }
2098
2099 /*
2100  * Abort all requests on the given list (pending or processing)
2101  *
2102  * This function releases and reacquires fc->lock
2103  */
2104 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2105 {
2106         while (!list_empty(head)) {
2107                 struct fuse_req *req;
2108                 req = list_entry(head->next, struct fuse_req, list);
2109                 req->out.h.error = -ECONNABORTED;
2110                 clear_bit(FR_SENT, &req->flags);
2111                 list_del_init(&req->list);
2112                 request_end(fc, req);
2113         }
2114 }
2115
2116 static void end_polls(struct fuse_conn *fc)
2117 {
2118         struct rb_node *p;
2119
2120         p = rb_first(&fc->polled_files);
2121
2122         while (p) {
2123                 struct fuse_file *ff;
2124                 ff = rb_entry(p, struct fuse_file, polled_node);
2125                 wake_up_interruptible_all(&ff->poll_wait);
2126
2127                 p = rb_next(p);
2128         }
2129 }
2130
2131 /*
2132  * Abort all requests.
2133  *
2134  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2135  * filesystem.
2136  *
2137  * The same effect is usually achievable through killing the filesystem daemon
2138  * and all users of the filesystem.  The exception is the combination of an
2139  * asynchronous request and the tricky deadlock (see
2140  * Documentation/filesystems/fuse.txt).
2141  *
2142  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2143  * requests, they should be finished off immediately.  Locked requests will be
2144  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2145  * requests.  It is possible that some request will finish before we can.  This
2146  * is OK, the request will in that case be removed from the list before we touch
2147  * it.
2148  */
2149 void fuse_abort_conn(struct fuse_conn *fc)
2150 {
2151         struct fuse_iqueue *fiq = &fc->iq;
2152
2153         spin_lock(&fc->lock);
2154         if (fc->connected) {
2155                 struct fuse_dev *fud;
2156                 struct fuse_req *req, *next;
2157                 LIST_HEAD(to_end1);
2158                 LIST_HEAD(to_end2);
2159
2160                 fc->connected = 0;
2161                 fc->blocked = 0;
2162                 fuse_set_initialized(fc);
2163                 list_for_each_entry(fud, &fc->devices, entry) {
2164                         struct fuse_pqueue *fpq = &fud->pq;
2165
2166                         spin_lock(&fpq->lock);
2167                         fpq->connected = 0;
2168                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2169                                 req->out.h.error = -ECONNABORTED;
2170                                 spin_lock(&req->waitq.lock);
2171                                 set_bit(FR_ABORTED, &req->flags);
2172                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2173                                         set_bit(FR_PRIVATE, &req->flags);
2174                                         __fuse_get_request(req);
2175                                         list_move(&req->list, &to_end1);
2176                                 }
2177                                 spin_unlock(&req->waitq.lock);
2178                         }
2179                         list_splice_init(&fpq->processing, &to_end2);
2180                         spin_unlock(&fpq->lock);
2181                 }
2182                 fc->max_background = UINT_MAX;
2183                 flush_bg_queue(fc);
2184
2185                 spin_lock(&fiq->waitq.lock);
2186                 fiq->connected = 0;
2187                 list_splice_init(&fiq->pending, &to_end2);
2188                 list_for_each_entry(req, &to_end2, list)
2189                         clear_bit(FR_PENDING, &req->flags);
2190                 while (forget_pending(fiq))
2191                         kfree(dequeue_forget(fiq, 1, NULL));
2192                 wake_up_all_locked(&fiq->waitq);
2193                 spin_unlock(&fiq->waitq.lock);
2194                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2195                 end_polls(fc);
2196                 wake_up_all(&fc->blocked_waitq);
2197                 spin_unlock(&fc->lock);
2198
2199                 while (!list_empty(&to_end1)) {
2200                         req = list_first_entry(&to_end1, struct fuse_req, list);
2201                         list_del_init(&req->list);
2202                         request_end(fc, req);
2203                 }
2204                 end_requests(fc, &to_end2);
2205         } else {
2206                 spin_unlock(&fc->lock);
2207         }
2208 }
2209 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2210
2211 void fuse_wait_aborted(struct fuse_conn *fc)
2212 {
2213         /* matches implicit memory barrier in fuse_drop_waiting() */
2214         smp_mb();
2215         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2216 }
2217
2218 int fuse_dev_release(struct inode *inode, struct file *file)
2219 {
2220         struct fuse_dev *fud = fuse_get_dev(file);
2221
2222         if (fud) {
2223                 struct fuse_conn *fc = fud->fc;
2224                 struct fuse_pqueue *fpq = &fud->pq;
2225                 LIST_HEAD(to_end);
2226
2227                 spin_lock(&fpq->lock);
2228                 WARN_ON(!list_empty(&fpq->io));
2229                 list_splice_init(&fpq->processing, &to_end);
2230                 spin_unlock(&fpq->lock);
2231
2232                 end_requests(fc, &to_end);
2233
2234                 /* Are we the last open device? */
2235                 if (atomic_dec_and_test(&fc->dev_count)) {
2236                         WARN_ON(fc->iq.fasync != NULL);
2237                         fuse_abort_conn(fc);
2238                 }
2239                 fuse_dev_free(fud);
2240         }
2241         return 0;
2242 }
2243 EXPORT_SYMBOL_GPL(fuse_dev_release);
2244
2245 static int fuse_dev_fasync(int fd, struct file *file, int on)
2246 {
2247         struct fuse_dev *fud = fuse_get_dev(file);
2248
2249         if (!fud)
2250                 return -EPERM;
2251
2252         /* No locking - fasync_helper does its own locking */
2253         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2254 }
2255
2256 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2257 {
2258         struct fuse_dev *fud;
2259
2260         if (new->private_data)
2261                 return -EINVAL;
2262
2263         fud = fuse_dev_alloc(fc);
2264         if (!fud)
2265                 return -ENOMEM;
2266
2267         new->private_data = fud;
2268         atomic_inc(&fc->dev_count);
2269
2270         return 0;
2271 }
2272
2273 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2274                            unsigned long arg)
2275 {
2276         int err = -ENOTTY;
2277
2278         if (cmd == FUSE_DEV_IOC_CLONE) {
2279                 int oldfd;
2280
2281                 err = -EFAULT;
2282                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2283                         struct file *old = fget(oldfd);
2284
2285                         err = -EINVAL;
2286                         if (old) {
2287                                 struct fuse_dev *fud = NULL;
2288
2289                                 /*
2290                                  * Check against file->f_op because CUSE
2291                                  * uses the same ioctl handler.
2292                                  */
2293                                 if (old->f_op == file->f_op &&
2294                                     old->f_cred->user_ns == file->f_cred->user_ns)
2295                                         fud = fuse_get_dev(old);
2296
2297                                 if (fud) {
2298                                         mutex_lock(&fuse_mutex);
2299                                         err = fuse_device_clone(fud->fc, file);
2300                                         mutex_unlock(&fuse_mutex);
2301                                 }
2302                                 fput(old);
2303                         }
2304                 }
2305         }
2306         return err;
2307 }
2308
2309 const struct file_operations fuse_dev_operations = {
2310         .owner          = THIS_MODULE,
2311         .open           = fuse_dev_open,
2312         .llseek         = no_llseek,
2313         .read_iter      = fuse_dev_read,
2314         .splice_read    = fuse_dev_splice_read,
2315         .write_iter     = fuse_dev_write,
2316         .splice_write   = fuse_dev_splice_write,
2317         .poll           = fuse_dev_poll,
2318         .release        = fuse_dev_release,
2319         .fasync         = fuse_dev_fasync,
2320         .unlocked_ioctl = fuse_dev_ioctl,
2321         .compat_ioctl   = fuse_dev_ioctl,
2322 };
2323 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2324
2325 static struct miscdevice fuse_miscdevice = {
2326         .minor = FUSE_MINOR,
2327         .name  = "fuse",
2328         .fops = &fuse_dev_operations,
2329 };
2330
2331 int __init fuse_dev_init(void)
2332 {
2333         int err = -ENOMEM;
2334         fuse_req_cachep = kmem_cache_create("fuse_request",
2335                                             sizeof(struct fuse_req),
2336                                             0, 0, NULL);
2337         if (!fuse_req_cachep)
2338                 goto out;
2339
2340         err = misc_register(&fuse_miscdevice);
2341         if (err)
2342                 goto out_cache_clean;
2343
2344         return 0;
2345
2346  out_cache_clean:
2347         kmem_cache_destroy(fuse_req_cachep);
2348  out:
2349         return err;
2350 }
2351
2352 void fuse_dev_cleanup(void)
2353 {
2354         misc_deregister(&fuse_miscdevice);
2355         kmem_cache_destroy(fuse_req_cachep);
2356 }