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                         return fuse_ref_page(cs, page, offset, count);
 995                 } else if (!cs->len) {
 996                         if (cs->move_pages && page &&
 997                             offset == 0 && count == PAGE_SIZE) {
 998                                 err = fuse_try_move_page(cs, pagep);
 999                                 if (err <= 0)
1000                                         return err;
1001                         } else {
1002                                 err = fuse_copy_fill(cs);
1003                                 if (err)
1004                                         return err;
1005                         }
1006                 }
1007                 if (page) {
1008                         void *mapaddr = kmap_atomic(page);
1009                         void *buf = mapaddr + offset;
1010                         offset += fuse_copy_do(cs, &buf, &count);
1011                         kunmap_atomic(mapaddr);
1012                 } else
1013                         offset += fuse_copy_do(cs, NULL, &count);
1014         }
1015         if (page && !cs->write)
1016                 flush_dcache_page(page);
1017         return 0;
1018 }
1019
1020 /* Copy pages in the request to/from userspace buffer */
1021 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1022                            int zeroing)
1023 {
1024         unsigned i;
1025         struct fuse_req *req = cs->req;
1026
1027         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1028                 int err;
1029                 unsigned offset = req->page_descs[i].offset;
1030                 unsigned count = min(nbytes, req->page_descs[i].length);
1031
1032                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1033                                      zeroing);
1034                 if (err)
1035                         return err;
1036
1037                 nbytes -= count;
1038         }
1039         return 0;
1040 }
1041
1042 /* Copy a single argument in the request to/from userspace buffer */
1043 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1044 {
1045         while (size) {
1046                 if (!cs->len) {
1047                         int err = fuse_copy_fill(cs);
1048                         if (err)
1049                                 return err;
1050                 }
1051                 fuse_copy_do(cs, &val, &size);
1052         }
1053         return 0;
1054 }
1055
1056 /* Copy request arguments to/from userspace buffer */
1057 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1058                           unsigned argpages, struct fuse_arg *args,
1059                           int zeroing)
1060 {
1061         int err = 0;
1062         unsigned i;
1063
1064         for (i = 0; !err && i < numargs; i++)  {
1065                 struct fuse_arg *arg = &args[i];
1066                 if (i == numargs - 1 && argpages)
1067                         err = fuse_copy_pages(cs, arg->size, zeroing);
1068                 else
1069                         err = fuse_copy_one(cs, arg->value, arg->size);
1070         }
1071         return err;
1072 }
1073
1074 static int forget_pending(struct fuse_iqueue *fiq)
1075 {
1076         return fiq->forget_list_head.next != NULL;
1077 }
1078
1079 static int request_pending(struct fuse_iqueue *fiq)
1080 {
1081         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1082                 forget_pending(fiq);
1083 }
1084
1085 /*
1086  * Transfer an interrupt request to userspace
1087  *
1088  * Unlike other requests this is assembled on demand, without a need
1089  * to allocate a separate fuse_req structure.
1090  *
1091  * Called with fiq->waitq.lock held, releases it
1092  */
1093 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1094                                struct fuse_copy_state *cs,
1095                                size_t nbytes, struct fuse_req *req)
1096 __releases(fiq->waitq.lock)
1097 {
1098         struct fuse_in_header ih;
1099         struct fuse_interrupt_in arg;
1100         unsigned reqsize = sizeof(ih) + sizeof(arg);
1101         int err;
1102
1103         list_del_init(&req->intr_entry);
1104         req->intr_unique = fuse_get_unique(fiq);
1105         memset(&ih, 0, sizeof(ih));
1106         memset(&arg, 0, sizeof(arg));
1107         ih.len = reqsize;
1108         ih.opcode = FUSE_INTERRUPT;
1109         ih.unique = req->intr_unique;
1110         arg.unique = req->in.h.unique;
1111
1112         spin_unlock(&fiq->waitq.lock);
1113         if (nbytes < reqsize)
1114                 return -EINVAL;
1115
1116         err = fuse_copy_one(cs, &ih, sizeof(ih));
1117         if (!err)
1118                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1119         fuse_copy_finish(cs);
1120
1121         return err ? err : reqsize;
1122 }
1123
1124 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1125                                                unsigned max,
1126                                                unsigned *countp)
1127 {
1128         struct fuse_forget_link *head = fiq->forget_list_head.next;
1129         struct fuse_forget_link **newhead = &head;
1130         unsigned count;
1131
1132         for (count = 0; *newhead != NULL && count < max; count++)
1133                 newhead = &(*newhead)->next;
1134
1135         fiq->forget_list_head.next = *newhead;
1136         *newhead = NULL;
1137         if (fiq->forget_list_head.next == NULL)
1138                 fiq->forget_list_tail = &fiq->forget_list_head;
1139
1140         if (countp != NULL)
1141                 *countp = count;
1142
1143         return head;
1144 }
1145
1146 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1147                                    struct fuse_copy_state *cs,
1148                                    size_t nbytes)
1149 __releases(fiq->waitq.lock)
1150 {
1151         int err;
1152         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1153         struct fuse_forget_in arg = {
1154                 .nlookup = forget->forget_one.nlookup,
1155         };
1156         struct fuse_in_header ih = {
1157                 .opcode = FUSE_FORGET,
1158                 .nodeid = forget->forget_one.nodeid,
1159                 .unique = fuse_get_unique(fiq),
1160                 .len = sizeof(ih) + sizeof(arg),
1161         };
1162
1163         spin_unlock(&fiq->waitq.lock);
1164         kfree(forget);
1165         if (nbytes < ih.len)
1166                 return -EINVAL;
1167
1168         err = fuse_copy_one(cs, &ih, sizeof(ih));
1169         if (!err)
1170                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1171         fuse_copy_finish(cs);
1172
1173         if (err)
1174                 return err;
1175
1176         return ih.len;
1177 }
1178
1179 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1180                                    struct fuse_copy_state *cs, size_t nbytes)
1181 __releases(fiq->waitq.lock)
1182 {
1183         int err;
1184         unsigned max_forgets;
1185         unsigned count;
1186         struct fuse_forget_link *head;
1187         struct fuse_batch_forget_in arg = { .count = 0 };
1188         struct fuse_in_header ih = {
1189                 .opcode = FUSE_BATCH_FORGET,
1190                 .unique = fuse_get_unique(fiq),
1191                 .len = sizeof(ih) + sizeof(arg),
1192         };
1193
1194         if (nbytes < ih.len) {
1195                 spin_unlock(&fiq->waitq.lock);
1196                 return -EINVAL;
1197         }
1198
1199         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1200         head = dequeue_forget(fiq, max_forgets, &count);
1201         spin_unlock(&fiq->waitq.lock);
1202
1203         arg.count = count;
1204         ih.len += count * sizeof(struct fuse_forget_one);
1205         err = fuse_copy_one(cs, &ih, sizeof(ih));
1206         if (!err)
1207                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1208
1209         while (head) {
1210                 struct fuse_forget_link *forget = head;
1211
1212                 if (!err) {
1213                         err = fuse_copy_one(cs, &forget->forget_one,
1214                                             sizeof(forget->forget_one));
1215                 }
1216                 head = forget->next;
1217                 kfree(forget);
1218         }
1219
1220         fuse_copy_finish(cs);
1221
1222         if (err)
1223                 return err;
1224
1225         return ih.len;
1226 }
1227
1228 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1229                             struct fuse_copy_state *cs,
1230                             size_t nbytes)
1231 __releases(fiq->waitq.lock)
1232 {
1233         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1234                 return fuse_read_single_forget(fiq, cs, nbytes);
1235         else
1236                 return fuse_read_batch_forget(fiq, cs, nbytes);
1237 }
1238
1239 /*
1240  * Read a single request into the userspace filesystem's buffer.  This
1241  * function waits until a request is available, then removes it from
1242  * the pending list and copies request data to userspace buffer.  If
1243  * no reply is needed (FORGET) or request has been aborted or there
1244  * was an error during the copying then it's finished by calling
1245  * request_end().  Otherwise add it to the processing list, and set
1246  * the 'sent' flag.
1247  */
1248 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1249                                 struct fuse_copy_state *cs, size_t nbytes)
1250 {
1251         ssize_t err;
1252         struct fuse_conn *fc = fud->fc;
1253         struct fuse_iqueue *fiq = &fc->iq;
1254         struct fuse_pqueue *fpq = &fud->pq;
1255         struct fuse_req *req;
1256         struct fuse_in *in;
1257         unsigned reqsize;
1258
1259  restart:
1260         spin_lock(&fiq->waitq.lock);
1261         err = -EAGAIN;
1262         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1263             !request_pending(fiq))
1264                 goto err_unlock;
1265
1266         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1267                                 !fiq->connected || request_pending(fiq));
1268         if (err)
1269                 goto err_unlock;
1270
1271         err = -ENODEV;
1272         if (!fiq->connected)
1273                 goto err_unlock;
1274
1275         if (!list_empty(&fiq->interrupts)) {
1276                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1277                                  intr_entry);
1278                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1279         }
1280
1281         if (forget_pending(fiq)) {
1282                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1283                         return fuse_read_forget(fc, fiq, cs, nbytes);
1284
1285                 if (fiq->forget_batch <= -8)
1286                         fiq->forget_batch = 16;
1287         }
1288
1289         req = list_entry(fiq->pending.next, struct fuse_req, list);
1290         clear_bit(FR_PENDING, &req->flags);
1291         list_del_init(&req->list);
1292         spin_unlock(&fiq->waitq.lock);
1293
1294         in = &req->in;
1295         reqsize = in->h.len;
1296
1297         if (task_active_pid_ns(current) != fc->pid_ns) {
1298                 rcu_read_lock();
1299                 in->h.pid = pid_vnr(find_pid_ns(in->h.pid, fc->pid_ns));
1300                 rcu_read_unlock();
1301         }
1302
1303         /* If request is too large, reply with an error and restart the read */
1304         if (nbytes < reqsize) {
1305                 req->out.h.error = -EIO;
1306                 /* SETXATTR is special, since it may contain too large data */
1307                 if (in->h.opcode == FUSE_SETXATTR)
1308                         req->out.h.error = -E2BIG;
1309                 request_end(fc, req);
1310                 goto restart;
1311         }
1312         spin_lock(&fpq->lock);
1313         /*
1314          *  Must not put request on fpq->io queue after having been shut down by
1315          *  fuse_abort_conn()
1316          */
1317         if (!fpq->connected) {
1318                 req->out.h.error = err = -ECONNABORTED;
1319                 goto out_end;
1320
1321         }
1322         list_add(&req->list, &fpq->io);
1323         spin_unlock(&fpq->lock);
1324         cs->req = req;
1325         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1326         if (!err)
1327                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1328                                      (struct fuse_arg *) in->args, 0);
1329         fuse_copy_finish(cs);
1330         spin_lock(&fpq->lock);
1331         clear_bit(FR_LOCKED, &req->flags);
1332         if (!fpq->connected) {
1333                 err = -ENODEV;
1334                 goto out_end;
1335         }
1336         if (err) {
1337                 req->out.h.error = -EIO;
1338                 goto out_end;
1339         }
1340         if (!test_bit(FR_ISREPLY, &req->flags)) {
1341                 err = reqsize;
1342                 goto out_end;
1343         }
1344         list_move_tail(&req->list, &fpq->processing);
1345         __fuse_get_request(req);
1346         set_bit(FR_SENT, &req->flags);
1347         spin_unlock(&fpq->lock);
1348         /* matches barrier in request_wait_answer() */
1349         smp_mb__after_atomic();
1350         if (test_bit(FR_INTERRUPTED, &req->flags))
1351                 queue_interrupt(fiq, req);
1352         fuse_put_request(fc, req);
1353
1354         return reqsize;
1355
1356 out_end:
1357         if (!test_bit(FR_PRIVATE, &req->flags))
1358                 list_del_init(&req->list);
1359         spin_unlock(&fpq->lock);
1360         request_end(fc, req);
1361         return err;
1362
1363  err_unlock:
1364         spin_unlock(&fiq->waitq.lock);
1365         return err;
1366 }
1367
1368 static int fuse_dev_open(struct inode *inode, struct file *file)
1369 {
1370         /*
1371          * The fuse device's file's private_data is used to hold
1372          * the fuse_conn(ection) when it is mounted, and is used to
1373          * keep track of whether the file has been mounted already.
1374          */
1375         file->private_data = NULL;
1376         return 0;
1377 }
1378
1379 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1380 {
1381         struct fuse_copy_state cs;
1382         struct file *file = iocb->ki_filp;
1383         struct fuse_dev *fud = fuse_get_dev(file);
1384
1385         if (!fud)
1386                 return -EPERM;
1387
1388         if (!iter_is_iovec(to))
1389                 return -EINVAL;
1390
1391         fuse_copy_init(&cs, 1, to);
1392
1393         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1394 }
1395
1396 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1397                                     struct pipe_inode_info *pipe,
1398                                     size_t len, unsigned int flags)
1399 {
1400         int total, ret;
1401         int page_nr = 0;
1402         struct pipe_buffer *bufs;
1403         struct fuse_copy_state cs;
1404         struct fuse_dev *fud = fuse_get_dev(in);
1405
1406         if (!fud)
1407                 return -EPERM;
1408
1409         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1410         if (!bufs)
1411                 return -ENOMEM;
1412
1413         fuse_copy_init(&cs, 1, NULL);
1414         cs.pipebufs = bufs;
1415         cs.pipe = pipe;
1416         ret = fuse_dev_do_read(fud, in, &cs, len);
1417         if (ret < 0)
1418                 goto out;
1419
1420         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1421                 ret = -EIO;
1422                 goto out;
1423         }
1424
1425         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1426                 /*
1427                  * Need to be careful about this.  Having buf->ops in module
1428                  * code can Oops if the buffer persists after module unload.
1429                  */
1430                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1431                 bufs[page_nr].flags = 0;
1432                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1433                 if (unlikely(ret < 0))
1434                         break;
1435         }
1436         if (total)
1437                 ret = total;
1438 out:
1439         for (; page_nr < cs.nr_segs; page_nr++)
1440                 put_page(bufs[page_nr].page);
1441
1442         kfree(bufs);
1443         return ret;
1444 }
1445
1446 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1447                             struct fuse_copy_state *cs)
1448 {
1449         struct fuse_notify_poll_wakeup_out outarg;
1450         int err = -EINVAL;
1451
1452         if (size != sizeof(outarg))
1453                 goto err;
1454
1455         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1456         if (err)
1457                 goto err;
1458
1459         fuse_copy_finish(cs);
1460         return fuse_notify_poll_wakeup(fc, &outarg);
1461
1462 err:
1463         fuse_copy_finish(cs);
1464         return err;
1465 }
1466
1467 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1468                                    struct fuse_copy_state *cs)
1469 {
1470         struct fuse_notify_inval_inode_out outarg;
1471         int err = -EINVAL;
1472
1473         if (size != sizeof(outarg))
1474                 goto err;
1475
1476         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1477         if (err)
1478                 goto err;
1479         fuse_copy_finish(cs);
1480
1481         down_read(&fc->killsb);
1482         err = -ENOENT;
1483         if (fc->sb) {
1484                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1485                                                outarg.off, outarg.len);
1486         }
1487         up_read(&fc->killsb);
1488         return err;
1489
1490 err:
1491         fuse_copy_finish(cs);
1492         return err;
1493 }
1494
1495 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1496                                    struct fuse_copy_state *cs)
1497 {
1498         struct fuse_notify_inval_entry_out outarg;
1499         int err = -ENOMEM;
1500         char *buf;
1501         struct qstr name;
1502
1503         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1504         if (!buf)
1505                 goto err;
1506
1507         err = -EINVAL;
1508         if (size < sizeof(outarg))
1509                 goto err;
1510
1511         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1512         if (err)
1513                 goto err;
1514
1515         err = -ENAMETOOLONG;
1516         if (outarg.namelen > FUSE_NAME_MAX)
1517                 goto err;
1518
1519         err = -EINVAL;
1520         if (size != sizeof(outarg) + outarg.namelen + 1)
1521                 goto err;
1522
1523         name.name = buf;
1524         name.len = outarg.namelen;
1525         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1526         if (err)
1527                 goto err;
1528         fuse_copy_finish(cs);
1529         buf[outarg.namelen] = 0;
1530
1531         down_read(&fc->killsb);
1532         err = -ENOENT;
1533         if (fc->sb)
1534                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1535         up_read(&fc->killsb);
1536         kfree(buf);
1537         return err;
1538
1539 err:
1540         kfree(buf);
1541         fuse_copy_finish(cs);
1542         return err;
1543 }
1544
1545 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1546                               struct fuse_copy_state *cs)
1547 {
1548         struct fuse_notify_delete_out outarg;
1549         int err = -ENOMEM;
1550         char *buf;
1551         struct qstr name;
1552
1553         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1554         if (!buf)
1555                 goto err;
1556
1557         err = -EINVAL;
1558         if (size < sizeof(outarg))
1559                 goto err;
1560
1561         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1562         if (err)
1563                 goto err;
1564
1565         err = -ENAMETOOLONG;
1566         if (outarg.namelen > FUSE_NAME_MAX)
1567                 goto err;
1568
1569         err = -EINVAL;
1570         if (size != sizeof(outarg) + outarg.namelen + 1)
1571                 goto err;
1572
1573         name.name = buf;
1574         name.len = outarg.namelen;
1575         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1576         if (err)
1577                 goto err;
1578         fuse_copy_finish(cs);
1579         buf[outarg.namelen] = 0;
1580
1581         down_read(&fc->killsb);
1582         err = -ENOENT;
1583         if (fc->sb)
1584                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1585                                                outarg.child, &name);
1586         up_read(&fc->killsb);
1587         kfree(buf);
1588         return err;
1589
1590 err:
1591         kfree(buf);
1592         fuse_copy_finish(cs);
1593         return err;
1594 }
1595
1596 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1597                              struct fuse_copy_state *cs)
1598 {
1599         struct fuse_notify_store_out outarg;
1600         struct inode *inode;
1601         struct address_space *mapping;
1602         u64 nodeid;
1603         int err;
1604         pgoff_t index;
1605         unsigned int offset;
1606         unsigned int num;
1607         loff_t file_size;
1608         loff_t end;
1609
1610         err = -EINVAL;
1611         if (size < sizeof(outarg))
1612                 goto out_finish;
1613
1614         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1615         if (err)
1616                 goto out_finish;
1617
1618         err = -EINVAL;
1619         if (size - sizeof(outarg) != outarg.size)
1620                 goto out_finish;
1621
1622         nodeid = outarg.nodeid;
1623
1624         down_read(&fc->killsb);
1625
1626         err = -ENOENT;
1627         if (!fc->sb)
1628                 goto out_up_killsb;
1629
1630         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1631         if (!inode)
1632                 goto out_up_killsb;
1633
1634         mapping = inode->i_mapping;
1635         index = outarg.offset >> PAGE_SHIFT;
1636         offset = outarg.offset & ~PAGE_MASK;
1637         file_size = i_size_read(inode);
1638         end = outarg.offset + outarg.size;
1639         if (end > file_size) {
1640                 file_size = end;
1641                 fuse_write_update_size(inode, file_size);
1642         }
1643
1644         num = outarg.size;
1645         while (num) {
1646                 struct page *page;
1647                 unsigned int this_num;
1648
1649                 err = -ENOMEM;
1650                 page = find_or_create_page(mapping, index,
1651                                            mapping_gfp_mask(mapping));
1652                 if (!page)
1653                         goto out_iput;
1654
1655                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1656                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1657                 if (!err && offset == 0 &&
1658                     (this_num == PAGE_SIZE || file_size == end))
1659                         SetPageUptodate(page);
1660                 unlock_page(page);
1661                 put_page(page);
1662
1663                 if (err)
1664                         goto out_iput;
1665
1666                 num -= this_num;
1667                 offset = 0;
1668                 index++;
1669         }
1670
1671         err = 0;
1672
1673 out_iput:
1674         iput(inode);
1675 out_up_killsb:
1676         up_read(&fc->killsb);
1677 out_finish:
1678         fuse_copy_finish(cs);
1679         return err;
1680 }
1681
1682 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1683 {
1684         release_pages(req->pages, req->num_pages, false);
1685 }
1686
1687 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1688                          struct fuse_notify_retrieve_out *outarg)
1689 {
1690         int err;
1691         struct address_space *mapping = inode->i_mapping;
1692         struct fuse_req *req;
1693         pgoff_t index;
1694         loff_t file_size;
1695         unsigned int num;
1696         unsigned int offset;
1697         size_t total_len = 0;
1698         int num_pages;
1699
1700         offset = outarg->offset & ~PAGE_MASK;
1701         file_size = i_size_read(inode);
1702
1703         num = min(outarg->size, fc->max_write);
1704         if (outarg->offset > file_size)
1705                 num = 0;
1706         else if (outarg->offset + num > file_size)
1707                 num = file_size - outarg->offset;
1708
1709         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1710         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1711
1712         req = fuse_get_req(fc, num_pages);
1713         if (IS_ERR(req))
1714                 return PTR_ERR(req);
1715
1716         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1717         req->in.h.nodeid = outarg->nodeid;
1718         req->in.numargs = 2;
1719         req->in.argpages = 1;
1720         req->end = fuse_retrieve_end;
1721
1722         index = outarg->offset >> PAGE_SHIFT;
1723
1724         while (num && req->num_pages < num_pages) {
1725                 struct page *page;
1726                 unsigned int this_num;
1727
1728                 page = find_get_page(mapping, index);
1729                 if (!page)
1730                         break;
1731
1732                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1733                 req->pages[req->num_pages] = page;
1734                 req->page_descs[req->num_pages].offset = offset;
1735                 req->page_descs[req->num_pages].length = this_num;
1736                 req->num_pages++;
1737
1738                 offset = 0;
1739                 num -= this_num;
1740                 total_len += this_num;
1741                 index++;
1742         }
1743         req->misc.retrieve_in.offset = outarg->offset;
1744         req->misc.retrieve_in.size = total_len;
1745         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1746         req->in.args[0].value = &req->misc.retrieve_in;
1747         req->in.args[1].size = total_len;
1748
1749         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1750         if (err) {
1751                 fuse_retrieve_end(fc, req);
1752                 fuse_put_request(fc, req);
1753         }
1754
1755         return err;
1756 }
1757
1758 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1759                                 struct fuse_copy_state *cs)
1760 {
1761         struct fuse_notify_retrieve_out outarg;
1762         struct inode *inode;
1763         int err;
1764
1765         err = -EINVAL;
1766         if (size != sizeof(outarg))
1767                 goto copy_finish;
1768
1769         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1770         if (err)
1771                 goto copy_finish;
1772
1773         fuse_copy_finish(cs);
1774
1775         down_read(&fc->killsb);
1776         err = -ENOENT;
1777         if (fc->sb) {
1778                 u64 nodeid = outarg.nodeid;
1779
1780                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1781                 if (inode) {
1782                         err = fuse_retrieve(fc, inode, &outarg);
1783                         iput(inode);
1784                 }
1785         }
1786         up_read(&fc->killsb);
1787
1788         return err;
1789
1790 copy_finish:
1791         fuse_copy_finish(cs);
1792         return err;
1793 }
1794
1795 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1796                        unsigned int size, struct fuse_copy_state *cs)
1797 {
1798         /* Don't try to move pages (yet) */
1799         cs->move_pages = 0;
1800
1801         switch (code) {
1802         case FUSE_NOTIFY_POLL:
1803                 return fuse_notify_poll(fc, size, cs);
1804
1805         case FUSE_NOTIFY_INVAL_INODE:
1806                 return fuse_notify_inval_inode(fc, size, cs);
1807
1808         case FUSE_NOTIFY_INVAL_ENTRY:
1809                 return fuse_notify_inval_entry(fc, size, cs);
1810
1811         case FUSE_NOTIFY_STORE:
1812                 return fuse_notify_store(fc, size, cs);
1813
1814         case FUSE_NOTIFY_RETRIEVE:
1815                 return fuse_notify_retrieve(fc, size, cs);
1816
1817         case FUSE_NOTIFY_DELETE:
1818                 return fuse_notify_delete(fc, size, cs);
1819
1820         default:
1821                 fuse_copy_finish(cs);
1822                 return -EINVAL;
1823         }
1824 }
1825
1826 /* Look up request on processing list by unique ID */
1827 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1828 {
1829         struct fuse_req *req;
1830
1831         list_for_each_entry(req, &fpq->processing, list) {
1832                 if (req->in.h.unique == unique || req->intr_unique == unique)
1833                         return req;
1834         }
1835         return NULL;
1836 }
1837
1838 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1839                          unsigned nbytes)
1840 {
1841         unsigned reqsize = sizeof(struct fuse_out_header);
1842
1843         if (out->h.error)
1844                 return nbytes != reqsize ? -EINVAL : 0;
1845
1846         reqsize += len_args(out->numargs, out->args);
1847
1848         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1849                 return -EINVAL;
1850         else if (reqsize > nbytes) {
1851                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1852                 unsigned diffsize = reqsize - nbytes;
1853                 if (diffsize > lastarg->size)
1854                         return -EINVAL;
1855                 lastarg->size -= diffsize;
1856         }
1857         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1858                               out->page_zeroing);
1859 }
1860
1861 /*
1862  * Write a single reply to a request.  First the header is copied from
1863  * the write buffer.  The request is then searched on the processing
1864  * list by the unique ID found in the header.  If found, then remove
1865  * it from the list and copy the rest of the buffer to the request.
1866  * The request is finished by calling request_end()
1867  */
1868 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1869                                  struct fuse_copy_state *cs, size_t nbytes)
1870 {
1871         int err;
1872         struct fuse_conn *fc = fud->fc;
1873         struct fuse_pqueue *fpq = &fud->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 <= -512 || oh.error > 0)
1899                 goto err_finish;
1900
1901         spin_lock(&fpq->lock);
1902         err = -ENOENT;
1903         if (!fpq->connected)
1904                 goto err_unlock_pq;
1905
1906         req = request_find(fpq, oh.unique);
1907         if (!req)
1908                 goto err_unlock_pq;
1909
1910         /* Is it an interrupt reply? */
1911         if (req->intr_unique == oh.unique) {
1912                 __fuse_get_request(req);
1913                 spin_unlock(&fpq->lock);
1914
1915                 err = -EINVAL;
1916                 if (nbytes != sizeof(struct fuse_out_header)) {
1917                         fuse_put_request(fc, req);
1918                         goto err_finish;
1919                 }
1920
1921                 if (oh.error == -ENOSYS)
1922                         fc->no_interrupt = 1;
1923                 else if (oh.error == -EAGAIN)
1924                         queue_interrupt(&fc->iq, req);
1925                 fuse_put_request(fc, req);
1926
1927                 fuse_copy_finish(cs);
1928                 return nbytes;
1929         }
1930
1931         clear_bit(FR_SENT, &req->flags);
1932         list_move(&req->list, &fpq->io);
1933         req->out.h = oh;
1934         set_bit(FR_LOCKED, &req->flags);
1935         spin_unlock(&fpq->lock);
1936         cs->req = req;
1937         if (!req->out.page_replace)
1938                 cs->move_pages = 0;
1939
1940         err = copy_out_args(cs, &req->out, nbytes);
1941         fuse_copy_finish(cs);
1942
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
1953         request_end(fc, req);
1954
1955         return err ? err : nbytes;
1956
1957  err_unlock_pq:
1958         spin_unlock(&fpq->lock);
1959  err_finish:
1960         fuse_copy_finish(cs);
1961         return err;
1962 }
1963
1964 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1965 {
1966         struct fuse_copy_state cs;
1967         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1968
1969         if (!fud)
1970                 return -EPERM;
1971
1972         if (!iter_is_iovec(from))
1973                 return -EINVAL;
1974
1975         fuse_copy_init(&cs, 0, from);
1976
1977         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1978 }
1979
1980 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1981                                      struct file *out, loff_t *ppos,
1982                                      size_t len, unsigned int flags)
1983 {
1984         unsigned nbuf;
1985         unsigned idx;
1986         struct pipe_buffer *bufs;
1987         struct fuse_copy_state cs;
1988         struct fuse_dev *fud;
1989         size_t rem;
1990         ssize_t ret;
1991
1992         fud = fuse_get_dev(out);
1993         if (!fud)
1994                 return -EPERM;
1995
1996         pipe_lock(pipe);
1997
1998         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1999         if (!bufs) {
2000                 pipe_unlock(pipe);
2001                 return -ENOMEM;
2002         }
2003
2004         nbuf = 0;
2005         rem = 0;
2006         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2007                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2008
2009         ret = -EINVAL;
2010         if (rem < len)
2011                 goto out_free;
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                         if (!pipe_buf_get(pipe, ibuf))
2030                                 goto out_free;
2031
2032                         *obuf = *ibuf;
2033                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2034                         obuf->len = rem;
2035                         ibuf->offset += obuf->len;
2036                         ibuf->len -= obuf->len;
2037                 }
2038                 nbuf++;
2039                 rem -= obuf->len;
2040         }
2041         pipe_unlock(pipe);
2042
2043         fuse_copy_init(&cs, 0, NULL);
2044         cs.pipebufs = bufs;
2045         cs.nr_segs = nbuf;
2046         cs.pipe = pipe;
2047
2048         if (flags & SPLICE_F_MOVE)
2049                 cs.move_pages = 1;
2050
2051         ret = fuse_dev_do_write(fud, &cs, len);
2052
2053         pipe_lock(pipe);
2054 out_free:
2055         for (idx = 0; idx < nbuf; idx++) {
2056                 struct pipe_buffer *buf = &bufs[idx];
2057
2058                 if (buf->ops)
2059                         pipe_buf_release(pipe, buf);
2060         }
2061         pipe_unlock(pipe);
2062
2063         kfree(bufs);
2064         return ret;
2065 }
2066
2067 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
2068 {
2069         unsigned mask = POLLOUT | POLLWRNORM;
2070         struct fuse_iqueue *fiq;
2071         struct fuse_dev *fud = fuse_get_dev(file);
2072
2073         if (!fud)
2074                 return POLLERR;
2075
2076         fiq = &fud->fc->iq;
2077         poll_wait(file, &fiq->waitq, wait);
2078
2079         spin_lock(&fiq->waitq.lock);
2080         if (!fiq->connected)
2081                 mask = POLLERR;
2082         else if (request_pending(fiq))
2083                 mask |= POLLIN | POLLRDNORM;
2084         spin_unlock(&fiq->waitq.lock);
2085
2086         return mask;
2087 }
2088
2089 /*
2090  * Abort all requests on the given list (pending or processing)
2091  *
2092  * This function releases and reacquires fc->lock
2093  */
2094 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2095 {
2096         while (!list_empty(head)) {
2097                 struct fuse_req *req;
2098                 req = list_entry(head->next, struct fuse_req, list);
2099                 req->out.h.error = -ECONNABORTED;
2100                 clear_bit(FR_SENT, &req->flags);
2101                 list_del_init(&req->list);
2102                 request_end(fc, req);
2103         }
2104 }
2105
2106 static void end_polls(struct fuse_conn *fc)
2107 {
2108         struct rb_node *p;
2109
2110         p = rb_first(&fc->polled_files);
2111
2112         while (p) {
2113                 struct fuse_file *ff;
2114                 ff = rb_entry(p, struct fuse_file, polled_node);
2115                 wake_up_interruptible_all(&ff->poll_wait);
2116
2117                 p = rb_next(p);
2118         }
2119 }
2120
2121 /*
2122  * Abort all requests.
2123  *
2124  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2125  * filesystem.
2126  *
2127  * The same effect is usually achievable through killing the filesystem daemon
2128  * and all users of the filesystem.  The exception is the combination of an
2129  * asynchronous request and the tricky deadlock (see
2130  * Documentation/filesystems/fuse.txt).
2131  *
2132  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2133  * requests, they should be finished off immediately.  Locked requests will be
2134  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2135  * requests.  It is possible that some request will finish before we can.  This
2136  * is OK, the request will in that case be removed from the list before we touch
2137  * it.
2138  */
2139 void fuse_abort_conn(struct fuse_conn *fc)
2140 {
2141         struct fuse_iqueue *fiq = &fc->iq;
2142
2143         spin_lock(&fc->lock);
2144         if (fc->connected) {
2145                 struct fuse_dev *fud;
2146                 struct fuse_req *req, *next;
2147                 LIST_HEAD(to_end1);
2148                 LIST_HEAD(to_end2);
2149
2150                 fc->connected = 0;
2151                 fc->blocked = 0;
2152                 fuse_set_initialized(fc);
2153                 list_for_each_entry(fud, &fc->devices, entry) {
2154                         struct fuse_pqueue *fpq = &fud->pq;
2155
2156                         spin_lock(&fpq->lock);
2157                         fpq->connected = 0;
2158                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2159                                 req->out.h.error = -ECONNABORTED;
2160                                 spin_lock(&req->waitq.lock);
2161                                 set_bit(FR_ABORTED, &req->flags);
2162                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2163                                         set_bit(FR_PRIVATE, &req->flags);
2164                                         __fuse_get_request(req);
2165                                         list_move(&req->list, &to_end1);
2166                                 }
2167                                 spin_unlock(&req->waitq.lock);
2168                         }
2169                         list_splice_init(&fpq->processing, &to_end2);
2170                         spin_unlock(&fpq->lock);
2171                 }
2172                 fc->max_background = UINT_MAX;
2173                 flush_bg_queue(fc);
2174
2175                 spin_lock(&fiq->waitq.lock);
2176                 fiq->connected = 0;
2177                 list_splice_init(&fiq->pending, &to_end2);
2178                 list_for_each_entry(req, &to_end2, list)
2179                         clear_bit(FR_PENDING, &req->flags);
2180                 while (forget_pending(fiq))
2181                         kfree(dequeue_forget(fiq, 1, NULL));
2182                 wake_up_all_locked(&fiq->waitq);
2183                 spin_unlock(&fiq->waitq.lock);
2184                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2185                 end_polls(fc);
2186                 wake_up_all(&fc->blocked_waitq);
2187                 spin_unlock(&fc->lock);
2188
2189                 while (!list_empty(&to_end1)) {
2190                         req = list_first_entry(&to_end1, struct fuse_req, list);
2191                         list_del_init(&req->list);
2192                         request_end(fc, req);
2193                 }
2194                 end_requests(fc, &to_end2);
2195         } else {
2196                 spin_unlock(&fc->lock);
2197         }
2198 }
2199 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2200
2201 void fuse_wait_aborted(struct fuse_conn *fc)
2202 {
2203         /* matches implicit memory barrier in fuse_drop_waiting() */
2204         smp_mb();
2205         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2206 }
2207
2208 int fuse_dev_release(struct inode *inode, struct file *file)
2209 {
2210         struct fuse_dev *fud = fuse_get_dev(file);
2211
2212         if (fud) {
2213                 struct fuse_conn *fc = fud->fc;
2214                 struct fuse_pqueue *fpq = &fud->pq;
2215                 LIST_HEAD(to_end);
2216
2217                 spin_lock(&fpq->lock);
2218                 WARN_ON(!list_empty(&fpq->io));
2219                 list_splice_init(&fpq->processing, &to_end);
2220                 spin_unlock(&fpq->lock);
2221
2222                 end_requests(fc, &to_end);
2223
2224                 /* Are we the last open device? */
2225                 if (atomic_dec_and_test(&fc->dev_count)) {
2226                         WARN_ON(fc->iq.fasync != NULL);
2227                         fuse_abort_conn(fc);
2228                 }
2229                 fuse_dev_free(fud);
2230         }
2231         return 0;
2232 }
2233 EXPORT_SYMBOL_GPL(fuse_dev_release);
2234
2235 static int fuse_dev_fasync(int fd, struct file *file, int on)
2236 {
2237         struct fuse_dev *fud = fuse_get_dev(file);
2238
2239         if (!fud)
2240                 return -EPERM;
2241
2242         /* No locking - fasync_helper does its own locking */
2243         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2244 }
2245
2246 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2247 {
2248         struct fuse_dev *fud;
2249
2250         if (new->private_data)
2251                 return -EINVAL;
2252
2253         fud = fuse_dev_alloc(fc);
2254         if (!fud)
2255                 return -ENOMEM;
2256
2257         new->private_data = fud;
2258         atomic_inc(&fc->dev_count);
2259
2260         return 0;
2261 }
2262
2263 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2264                            unsigned long arg)
2265 {
2266         int err = -ENOTTY;
2267
2268         if (cmd == FUSE_DEV_IOC_CLONE) {
2269                 int oldfd;
2270
2271                 err = -EFAULT;
2272                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2273                         struct file *old = fget(oldfd);
2274
2275                         err = -EINVAL;
2276                         if (old) {
2277                                 struct fuse_dev *fud = NULL;
2278
2279                                 /*
2280                                  * Check against file->f_op because CUSE
2281                                  * uses the same ioctl handler.
2282                                  */
2283                                 if (old->f_op == file->f_op &&
2284                                     old->f_cred->user_ns == file->f_cred->user_ns)
2285                                         fud = fuse_get_dev(old);
2286
2287                                 if (fud) {
2288                                         mutex_lock(&fuse_mutex);
2289                                         err = fuse_device_clone(fud->fc, file);
2290                                         mutex_unlock(&fuse_mutex);
2291                                 }
2292                                 fput(old);
2293                         }
2294                 }
2295         }
2296         return err;
2297 }
2298
2299 const struct file_operations fuse_dev_operations = {
2300         .owner          = THIS_MODULE,
2301         .open           = fuse_dev_open,
2302         .llseek         = no_llseek,
2303         .read_iter      = fuse_dev_read,
2304         .splice_read    = fuse_dev_splice_read,
2305         .write_iter     = fuse_dev_write,
2306         .splice_write   = fuse_dev_splice_write,
2307         .poll           = fuse_dev_poll,
2308         .release        = fuse_dev_release,
2309         .fasync         = fuse_dev_fasync,
2310         .unlocked_ioctl = fuse_dev_ioctl,
2311         .compat_ioctl   = fuse_dev_ioctl,
2312 };
2313 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2314
2315 static struct miscdevice fuse_miscdevice = {
2316         .minor = FUSE_MINOR,
2317         .name  = "fuse",
2318         .fops = &fuse_dev_operations,
2319 };
2320
2321 int __init fuse_dev_init(void)
2322 {
2323         int err = -ENOMEM;
2324         fuse_req_cachep = kmem_cache_create("fuse_request",
2325                                             sizeof(struct fuse_req),
2326                                             0, 0, NULL);
2327         if (!fuse_req_cachep)
2328                 goto out;
2329
2330         err = misc_register(&fuse_miscdevice);
2331         if (err)
2332                 goto out_cache_clean;
2333
2334         return 0;
2335
2336  out_cache_clean:
2337         kmem_cache_destroy(fuse_req_cachep);
2338  out:
2339         return err;
2340 }
2341
2342 void fuse_dev_cleanup(void)
2343 {
2344         misc_deregister(&fuse_miscdevice);
2345         kmem_cache_destroy(fuse_req_cachep);
2346 }