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 READ_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_array(npages, sizeof(struct page *),
  68                                               flags);
  69                         page_descs =
  70                                 kmalloc_array(npages,
  71                                               sizeof(struct fuse_page_desc),
  72                                               flags);
  73                 }
  74
  75                 if (!pages || !page_descs) {
  76                         kfree(pages);
  77                         kfree(page_descs);
  78                         kmem_cache_free(fuse_req_cachep, req);
  79                         return NULL;
  80                 }
  81
  82                 fuse_request_init(req, pages, page_descs, npages);
  83         }
  84         return req;
  85 }
  86
  87 struct fuse_req *fuse_request_alloc(unsigned npages)
  88 {
  89         return __fuse_request_alloc(npages, GFP_KERNEL);
  90 }
  91 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  92
  93 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  94 {
  95         return __fuse_request_alloc(npages, GFP_NOFS);
  96 }
  97
  98 void fuse_request_free(struct fuse_req *req)
  99 {
 100         if (req->pages != req->inline_pages) {
 101                 kfree(req->pages);
 102                 kfree(req->page_descs);
 103         }
 104         kmem_cache_free(fuse_req_cachep, req);
 105 }
 106
 107 void __fuse_get_request(struct fuse_req *req)
 108 {
 109         refcount_inc(&req->count);
 110 }
 111
 112 /* Must be called with > 1 refcount */
 113 static void __fuse_put_request(struct fuse_req *req)
 114 {
 115         refcount_dec(&req->count);
 116 }
 117
 118 void fuse_set_initialized(struct fuse_conn *fc)
 119 {
 120         /* Make sure stores before this are seen on another CPU */
 121         smp_wmb();
 122         fc->initialized = 1;
 123 }
 124
 125 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 126 {
 127         return !fc->initialized || (for_background && fc->blocked);
 128 }
 129
 130 static void fuse_drop_waiting(struct fuse_conn *fc)
 131 {
 132         /*
 133          * lockess check of fc->connected is okay, because atomic_dec_and_test()
 134          * provides a memory barrier mached with the one in fuse_wait_aborted()
 135          * to ensure no wake-up is missed.
 136          */
 137         if (atomic_dec_and_test(&fc->num_waiting) &&
 138             !READ_ONCE(fc->connected)) {
 139                 /* wake up aborters */
 140                 wake_up_all(&fc->blocked_waitq);
 141         }
 142 }
 143
 144 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 145                                        bool for_background)
 146 {
 147         struct fuse_req *req;
 148         int err;
 149         atomic_inc(&fc->num_waiting);
 150
 151         if (fuse_block_alloc(fc, for_background)) {
 152                 err = -EINTR;
 153                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 154                                 !fuse_block_alloc(fc, for_background)))
 155                         goto out;
 156         }
 157         /* Matches smp_wmb() in fuse_set_initialized() */
 158         smp_rmb();
 159
 160         err = -ENOTCONN;
 161         if (!fc->connected)
 162                 goto out;
 163
 164         err = -ECONNREFUSED;
 165         if (fc->conn_error)
 166                 goto out;
 167
 168         req = fuse_request_alloc(npages);
 169         err = -ENOMEM;
 170         if (!req) {
 171                 if (for_background)
 172                         wake_up(&fc->blocked_waitq);
 173                 goto out;
 174         }
 175
 176         req->in.h.uid = from_kuid(fc->user_ns, current_fsuid());
 177         req->in.h.gid = from_kgid(fc->user_ns, current_fsgid());
 178         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 179
 180         __set_bit(FR_WAITING, &req->flags);
 181         if (for_background)
 182                 __set_bit(FR_BACKGROUND, &req->flags);
 183
 184         if (unlikely(req->in.h.uid == ((uid_t)-1) ||
 185                      req->in.h.gid == ((gid_t)-1))) {
 186                 fuse_put_request(fc, req);
 187                 return ERR_PTR(-EOVERFLOW);
 188         }
 189         return req;
 190
 191  out:
 192         fuse_drop_waiting(fc);
 193         return ERR_PTR(err);
 194 }
 195
 196 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 197 {
 198         return __fuse_get_req(fc, npages, false);
 199 }
 200 EXPORT_SYMBOL_GPL(fuse_get_req);
 201
 202 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 203                                              unsigned npages)
 204 {
 205         return __fuse_get_req(fc, npages, true);
 206 }
 207 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 208
 209 /*
 210  * Return request in fuse_file->reserved_req.  However that may
 211  * currently be in use.  If that is the case, wait for it to become
 212  * available.
 213  */
 214 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 215                                          struct file *file)
 216 {
 217         struct fuse_req *req = NULL;
 218         struct fuse_file *ff = file->private_data;
 219
 220         do {
 221                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 222                 spin_lock(&fc->lock);
 223                 if (ff->reserved_req) {
 224                         req = ff->reserved_req;
 225                         ff->reserved_req = NULL;
 226                         req->stolen_file = get_file(file);
 227                 }
 228                 spin_unlock(&fc->lock);
 229         } while (!req);
 230
 231         return req;
 232 }
 233
 234 /*
 235  * Put stolen request back into fuse_file->reserved_req
 236  */
 237 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 238 {
 239         struct file *file = req->stolen_file;
 240         struct fuse_file *ff = file->private_data;
 241
 242         spin_lock(&fc->lock);
 243         fuse_request_init(req, req->pages, req->page_descs, req->max_pages);
 244         BUG_ON(ff->reserved_req);
 245         ff->reserved_req = req;
 246         wake_up_all(&fc->reserved_req_waitq);
 247         spin_unlock(&fc->lock);
 248         fput(file);
 249 }
 250
 251 /*
 252  * Gets a requests for a file operation, always succeeds
 253  *
 254  * This is used for sending the FLUSH request, which must get to
 255  * userspace, due to POSIX locks which may need to be unlocked.
 256  *
 257  * If allocation fails due to OOM, use the reserved request in
 258  * fuse_file.
 259  *
 260  * This is very unlikely to deadlock accidentally, since the
 261  * filesystem should not have it's own file open.  If deadlock is
 262  * intentional, it can still be broken by "aborting" the filesystem.
 263  */
 264 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 265                                              struct file *file)
 266 {
 267         struct fuse_req *req;
 268
 269         atomic_inc(&fc->num_waiting);
 270         wait_event(fc->blocked_waitq, fc->initialized);
 271         /* Matches smp_wmb() in fuse_set_initialized() */
 272         smp_rmb();
 273         req = fuse_request_alloc(0);
 274         if (!req)
 275                 req = get_reserved_req(fc, file);
 276
 277         req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
 278         req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
 279         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 280
 281         __set_bit(FR_WAITING, &req->flags);
 282         __clear_bit(FR_BACKGROUND, &req->flags);
 283         return req;
 284 }
 285
 286 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 287 {
 288         if (refcount_dec_and_test(&req->count)) {
 289                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 290                         /*
 291                          * We get here in the unlikely case that a background
 292                          * request was allocated but not sent
 293                          */
 294                         spin_lock(&fc->lock);
 295                         if (!fc->blocked)
 296                                 wake_up(&fc->blocked_waitq);
 297                         spin_unlock(&fc->lock);
 298                 }
 299
 300                 if (test_bit(FR_WAITING, &req->flags)) {
 301                         __clear_bit(FR_WAITING, &req->flags);
 302                         fuse_drop_waiting(fc);
 303                 }
 304
 305                 if (req->stolen_file)
 306                         put_reserved_req(fc, req);
 307                 else
 308                         fuse_request_free(req);
 309         }
 310 }
 311 EXPORT_SYMBOL_GPL(fuse_put_request);
 312
 313 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 314 {
 315         unsigned nbytes = 0;
 316         unsigned i;
 317
 318         for (i = 0; i < numargs; i++)
 319                 nbytes += args[i].size;
 320
 321         return nbytes;
 322 }
 323
 324 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 325 {
 326         return ++fiq->reqctr;
 327 }
 328
 329 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 330 {
 331         req->in.h.len = sizeof(struct fuse_in_header) +
 332                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 333         list_add_tail(&req->list, &fiq->pending);
 334         wake_up(&fiq->waitq);
 335         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 336 }
 337
 338 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 339                        u64 nodeid, u64 nlookup)
 340 {
 341         struct fuse_iqueue *fiq = &fc->iq;
 342
 343         forget->forget_one.nodeid = nodeid;
 344         forget->forget_one.nlookup = nlookup;
 345
 346         spin_lock(&fiq->lock);
 347         if (fiq->connected) {
 348                 fiq->forget_list_tail->next = forget;
 349                 fiq->forget_list_tail = forget;
 350                 wake_up(&fiq->waitq);
 351                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 352         } else {
 353                 kfree(forget);
 354         }
 355         spin_unlock(&fiq->lock);
 356 }
 357
 358 static void flush_bg_queue(struct fuse_conn *fc)
 359 {
 360         while (fc->active_background < fc->max_background &&
 361                !list_empty(&fc->bg_queue)) {
 362                 struct fuse_req *req;
 363                 struct fuse_iqueue *fiq = &fc->iq;
 364
 365                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 366                 list_del(&req->list);
 367                 fc->active_background++;
 368                 spin_lock(&fiq->lock);
 369                 req->in.h.unique = fuse_get_unique(fiq);
 370                 queue_request(fiq, req);
 371                 spin_unlock(&fiq->lock);
 372         }
 373 }
 374
 375 /*
 376  * This function is called when a request is finished.  Either a reply
 377  * has arrived or it was aborted (and not yet sent) or some error
 378  * occurred during communication with userspace, or the device file
 379  * was closed.  The requester thread is woken up (if still waiting),
 380  * the 'end' callback is called if given, else the reference to the
 381  * request is released
 382  */
 383 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 384 {
 385         struct fuse_iqueue *fiq = &fc->iq;
 386
 387         if (test_and_set_bit(FR_FINISHED, &req->flags))
 388                 goto put_request;
 389
 390         spin_lock(&fiq->lock);
 391         list_del_init(&req->intr_entry);
 392         spin_unlock(&fiq->lock);
 393         WARN_ON(test_bit(FR_PENDING, &req->flags));
 394         WARN_ON(test_bit(FR_SENT, &req->flags));
 395         if (test_bit(FR_BACKGROUND, &req->flags)) {
 396                 spin_lock(&fc->lock);
 397                 clear_bit(FR_BACKGROUND, &req->flags);
 398                 if (fc->num_background == fc->max_background) {
 399                         fc->blocked = 0;
 400                         wake_up(&fc->blocked_waitq);
 401                 } else if (!fc->blocked) {
 402                         /*
 403                          * Wake up next waiter, if any.  It's okay to use
 404                          * waitqueue_active(), as we've already synced up
 405                          * fc->blocked with waiters with the wake_up() call
 406                          * above.
 407                          */
 408                         if (waitqueue_active(&fc->blocked_waitq))
 409                                 wake_up(&fc->blocked_waitq);
 410                 }
 411
 412                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 413                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 414                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 415                 }
 416                 fc->num_background--;
 417                 fc->active_background--;
 418                 flush_bg_queue(fc);
 419                 spin_unlock(&fc->lock);
 420         }
 421         wake_up(&req->waitq);
 422         if (req->end)
 423                 req->end(fc, req);
 424 put_request:
 425         fuse_put_request(fc, req);
 426 }
 427
 428 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 429 {
 430         spin_lock(&fiq->lock);
 431         if (test_bit(FR_FINISHED, &req->flags)) {
 432                 spin_unlock(&fiq->lock);
 433                 return;
 434         }
 435         if (list_empty(&req->intr_entry)) {
 436                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 437                 wake_up(&fiq->waitq);
 438         }
 439         spin_unlock(&fiq->lock);
 440         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 441 }
 442
 443 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 444 {
 445         struct fuse_iqueue *fiq = &fc->iq;
 446         int err;
 447
 448         if (!fc->no_interrupt) {
 449                 /* Any signal may interrupt this */
 450                 err = wait_event_interruptible(req->waitq,
 451                                         test_bit(FR_FINISHED, &req->flags));
 452                 if (!err)
 453                         return;
 454
 455                 set_bit(FR_INTERRUPTED, &req->flags);
 456                 /* matches barrier in fuse_dev_do_read() */
 457                 smp_mb__after_atomic();
 458                 if (test_bit(FR_SENT, &req->flags))
 459                         queue_interrupt(fiq, req);
 460         }
 461
 462         if (!test_bit(FR_FORCE, &req->flags)) {
 463                 /* Only fatal signals may interrupt this */
 464                 err = wait_event_killable(req->waitq,
 465                                         test_bit(FR_FINISHED, &req->flags));
 466                 if (!err)
 467                         return;
 468
 469                 spin_lock(&fiq->lock);
 470                 /* Request is not yet in userspace, bail out */
 471                 if (test_bit(FR_PENDING, &req->flags)) {
 472                         list_del(&req->list);
 473                         spin_unlock(&fiq->lock);
 474                         __fuse_put_request(req);
 475                         req->out.h.error = -EINTR;
 476                         return;
 477                 }
 478                 spin_unlock(&fiq->lock);
 479         }
 480
 481         /*
 482          * Either request is already in userspace, or it was forced.
 483          * Wait it out.
 484          */
 485         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 486 }
 487
 488 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 489 {
 490         struct fuse_iqueue *fiq = &fc->iq;
 491
 492         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 493         spin_lock(&fiq->lock);
 494         if (!fiq->connected) {
 495                 spin_unlock(&fiq->lock);
 496                 req->out.h.error = -ENOTCONN;
 497         } else {
 498                 req->in.h.unique = fuse_get_unique(fiq);
 499                 queue_request(fiq, req);
 500                 /* acquire extra reference, since request is still needed
 501                    after request_end() */
 502                 __fuse_get_request(req);
 503                 spin_unlock(&fiq->lock);
 504
 505                 request_wait_answer(fc, req);
 506                 /* Pairs with smp_wmb() in request_end() */
 507                 smp_rmb();
 508         }
 509 }
 510
 511 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 512 {
 513         __set_bit(FR_ISREPLY, &req->flags);
 514         if (!test_bit(FR_WAITING, &req->flags)) {
 515                 __set_bit(FR_WAITING, &req->flags);
 516                 atomic_inc(&fc->num_waiting);
 517         }
 518         __fuse_request_send(fc, req);
 519 }
 520 EXPORT_SYMBOL_GPL(fuse_request_send);
 521
 522 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 523 {
 524         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 525                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 526
 527         if (fc->minor < 9) {
 528                 switch (args->in.h.opcode) {
 529                 case FUSE_LOOKUP:
 530                 case FUSE_CREATE:
 531                 case FUSE_MKNOD:
 532                 case FUSE_MKDIR:
 533                 case FUSE_SYMLINK:
 534                 case FUSE_LINK:
 535                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 536                         break;
 537                 case FUSE_GETATTR:
 538                 case FUSE_SETATTR:
 539                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 540                         break;
 541                 }
 542         }
 543         if (fc->minor < 12) {
 544                 switch (args->in.h.opcode) {
 545                 case FUSE_CREATE:
 546                         args->in.args[0].size = sizeof(struct fuse_open_in);
 547                         break;
 548                 case FUSE_MKNOD:
 549                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 550                         break;
 551                 }
 552         }
 553 }
 554
 555 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 556 {
 557         struct fuse_req *req;
 558         ssize_t ret;
 559
 560         req = fuse_get_req(fc, 0);
 561         if (IS_ERR(req))
 562                 return PTR_ERR(req);
 563
 564         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 565         fuse_adjust_compat(fc, args);
 566
 567         req->in.h.opcode = args->in.h.opcode;
 568         req->in.h.nodeid = args->in.h.nodeid;
 569         req->in.numargs = args->in.numargs;
 570         memcpy(req->in.args, args->in.args,
 571                args->in.numargs * sizeof(struct fuse_in_arg));
 572         req->out.argvar = args->out.argvar;
 573         req->out.numargs = args->out.numargs;
 574         memcpy(req->out.args, args->out.args,
 575                args->out.numargs * sizeof(struct fuse_arg));
 576         fuse_request_send(fc, req);
 577         ret = req->out.h.error;
 578         if (!ret && args->out.argvar) {
 579                 BUG_ON(args->out.numargs != 1);
 580                 ret = req->out.args[0].size;
 581         }
 582         fuse_put_request(fc, req);
 583
 584         return ret;
 585 }
 586
 587 /*
 588  * Called under fc->lock
 589  *
 590  * fc->connected must have been checked previously
 591  */
 592 void fuse_request_send_background_locked(struct fuse_conn *fc,
 593                                          struct fuse_req *req)
 594 {
 595         BUG_ON(!test_bit(FR_BACKGROUND, &req->flags));
 596         if (!test_bit(FR_WAITING, &req->flags)) {
 597                 __set_bit(FR_WAITING, &req->flags);
 598                 atomic_inc(&fc->num_waiting);
 599         }
 600         __set_bit(FR_ISREPLY, &req->flags);
 601         fc->num_background++;
 602         if (fc->num_background == fc->max_background)
 603                 fc->blocked = 1;
 604         if (fc->num_background == fc->congestion_threshold && fc->sb) {
 605                 set_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 606                 set_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 607         }
 608         list_add_tail(&req->list, &fc->bg_queue);
 609         flush_bg_queue(fc);
 610 }
 611
 612 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 613 {
 614         BUG_ON(!req->end);
 615         spin_lock(&fc->lock);
 616         if (fc->connected) {
 617                 fuse_request_send_background_locked(fc, req);
 618                 spin_unlock(&fc->lock);
 619         } else {
 620                 spin_unlock(&fc->lock);
 621                 req->out.h.error = -ENOTCONN;
 622                 req->end(fc, req);
 623                 fuse_put_request(fc, req);
 624         }
 625 }
 626 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 627
 628 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 629                                           struct fuse_req *req, u64 unique)
 630 {
 631         int err = -ENODEV;
 632         struct fuse_iqueue *fiq = &fc->iq;
 633
 634         __clear_bit(FR_ISREPLY, &req->flags);
 635         req->in.h.unique = unique;
 636         spin_lock(&fiq->lock);
 637         if (fiq->connected) {
 638                 queue_request(fiq, req);
 639                 err = 0;
 640         }
 641         spin_unlock(&fiq->lock);
 642
 643         return err;
 644 }
 645
 646 void fuse_force_forget(struct file *file, u64 nodeid)
 647 {
 648         struct inode *inode = file_inode(file);
 649         struct fuse_conn *fc = get_fuse_conn(inode);
 650         struct fuse_req *req;
 651         struct fuse_forget_in inarg;
 652
 653         memset(&inarg, 0, sizeof(inarg));
 654         inarg.nlookup = 1;
 655         req = fuse_get_req_nofail_nopages(fc, file);
 656         req->in.h.opcode = FUSE_FORGET;
 657         req->in.h.nodeid = nodeid;
 658         req->in.numargs = 1;
 659         req->in.args[0].size = sizeof(inarg);
 660         req->in.args[0].value = &inarg;
 661         __clear_bit(FR_ISREPLY, &req->flags);
 662         __fuse_request_send(fc, req);
 663         /* ignore errors */
 664         fuse_put_request(fc, req);
 665 }
 666
 667 /*
 668  * Lock the request.  Up to the next unlock_request() there mustn't be
 669  * anything that could cause a page-fault.  If the request was already
 670  * aborted bail out.
 671  */
 672 static int lock_request(struct fuse_req *req)
 673 {
 674         int err = 0;
 675         if (req) {
 676                 spin_lock(&req->waitq.lock);
 677                 if (test_bit(FR_ABORTED, &req->flags))
 678                         err = -ENOENT;
 679                 else
 680                         set_bit(FR_LOCKED, &req->flags);
 681                 spin_unlock(&req->waitq.lock);
 682         }
 683         return err;
 684 }
 685
 686 /*
 687  * Unlock request.  If it was aborted while locked, caller is responsible
 688  * for unlocking and ending the request.
 689  */
 690 static int unlock_request(struct fuse_req *req)
 691 {
 692         int err = 0;
 693         if (req) {
 694                 spin_lock(&req->waitq.lock);
 695                 if (test_bit(FR_ABORTED, &req->flags))
 696                         err = -ENOENT;
 697                 else
 698                         clear_bit(FR_LOCKED, &req->flags);
 699                 spin_unlock(&req->waitq.lock);
 700         }
 701         return err;
 702 }
 703
 704 struct fuse_copy_state {
 705         int write;
 706         struct fuse_req *req;
 707         struct iov_iter *iter;
 708         struct pipe_buffer *pipebufs;
 709         struct pipe_buffer *currbuf;
 710         struct pipe_inode_info *pipe;
 711         unsigned long nr_segs;
 712         struct page *pg;
 713         unsigned len;
 714         unsigned offset;
 715         unsigned move_pages:1;
 716 };
 717
 718 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 719                            struct iov_iter *iter)
 720 {
 721         memset(cs, 0, sizeof(*cs));
 722         cs->write = write;
 723         cs->iter = iter;
 724 }
 725
 726 /* Unmap and put previous page of userspace buffer */
 727 static void fuse_copy_finish(struct fuse_copy_state *cs)
 728 {
 729         if (cs->currbuf) {
 730                 struct pipe_buffer *buf = cs->currbuf;
 731
 732                 if (cs->write)
 733                         buf->len = PAGE_SIZE - cs->len;
 734                 cs->currbuf = NULL;
 735         } else if (cs->pg) {
 736                 if (cs->write) {
 737                         flush_dcache_page(cs->pg);
 738                         set_page_dirty_lock(cs->pg);
 739                 }
 740                 put_page(cs->pg);
 741         }
 742         cs->pg = NULL;
 743 }
 744
 745 /*
 746  * Get another pagefull of userspace buffer, and map it to kernel
 747  * address space, and lock request
 748  */
 749 static int fuse_copy_fill(struct fuse_copy_state *cs)
 750 {
 751         struct page *page;
 752         int err;
 753
 754         err = unlock_request(cs->req);
 755         if (err)
 756                 return err;
 757
 758         fuse_copy_finish(cs);
 759         if (cs->pipebufs) {
 760                 struct pipe_buffer *buf = cs->pipebufs;
 761
 762                 if (!cs->write) {
 763                         err = pipe_buf_confirm(cs->pipe, buf);
 764                         if (err)
 765                                 return err;
 766
 767                         BUG_ON(!cs->nr_segs);
 768                         cs->currbuf = buf;
 769                         cs->pg = buf->page;
 770                         cs->offset = buf->offset;
 771                         cs->len = buf->len;
 772                         cs->pipebufs++;
 773                         cs->nr_segs--;
 774                 } else {
 775                         if (cs->nr_segs == cs->pipe->buffers)
 776                                 return -EIO;
 777
 778                         page = alloc_page(GFP_HIGHUSER);
 779                         if (!page)
 780                                 return -ENOMEM;
 781
 782                         buf->page = page;
 783                         buf->offset = 0;
 784                         buf->len = 0;
 785
 786                         cs->currbuf = buf;
 787                         cs->pg = page;
 788                         cs->offset = 0;
 789                         cs->len = PAGE_SIZE;
 790                         cs->pipebufs++;
 791                         cs->nr_segs++;
 792                 }
 793         } else {
 794                 size_t off;
 795                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 796                 if (err < 0)
 797                         return err;
 798                 BUG_ON(!err);
 799                 cs->len = err;
 800                 cs->offset = off;
 801                 cs->pg = page;
 802                 iov_iter_advance(cs->iter, err);
 803         }
 804
 805         return lock_request(cs->req);
 806 }
 807
 808 /* Do as much copy to/from userspace buffer as we can */
 809 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 810 {
 811         unsigned ncpy = min(*size, cs->len);
 812         if (val) {
 813                 void *pgaddr = kmap_atomic(cs->pg);
 814                 void *buf = pgaddr + cs->offset;
 815
 816                 if (cs->write)
 817                         memcpy(buf, *val, ncpy);
 818                 else
 819                         memcpy(*val, buf, ncpy);
 820
 821                 kunmap_atomic(pgaddr);
 822                 *val += ncpy;
 823         }
 824         *size -= ncpy;
 825         cs->len -= ncpy;
 826         cs->offset += ncpy;
 827         return ncpy;
 828 }
 829
 830 static int fuse_check_page(struct page *page)
 831 {
 832         if (page_mapcount(page) ||
 833             page->mapping != NULL ||
 834             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 835              ~(1 << PG_locked |
 836                1 << PG_referenced |
 837                1 << PG_uptodate |
 838                1 << PG_lru |
 839                1 << PG_active |
 840                1 << PG_reclaim |
 841                1 << PG_waiters))) {
 842                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 843                 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);
 844                 return 1;
 845         }
 846         return 0;
 847 }
 848
 849 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 850 {
 851         int err;
 852         struct page *oldpage = *pagep;
 853         struct page *newpage;
 854         struct pipe_buffer *buf = cs->pipebufs;
 855
 856         get_page(oldpage);
 857         err = unlock_request(cs->req);
 858         if (err)
 859                 goto out_put_old;
 860
 861         fuse_copy_finish(cs);
 862
 863         err = pipe_buf_confirm(cs->pipe, buf);
 864         if (err)
 865                 goto out_put_old;
 866
 867         BUG_ON(!cs->nr_segs);
 868         cs->currbuf = buf;
 869         cs->len = buf->len;
 870         cs->pipebufs++;
 871         cs->nr_segs--;
 872
 873         if (cs->len != PAGE_SIZE)
 874                 goto out_fallback;
 875
 876         if (pipe_buf_steal(cs->pipe, buf) != 0)
 877                 goto out_fallback;
 878
 879         newpage = buf->page;
 880
 881         if (!PageUptodate(newpage))
 882                 SetPageUptodate(newpage);
 883
 884         ClearPageMappedToDisk(newpage);
 885
 886         if (fuse_check_page(newpage) != 0)
 887                 goto out_fallback_unlock;
 888
 889         /*
 890          * This is a new and locked page, it shouldn't be mapped or
 891          * have any special flags on it
 892          */
 893         if (WARN_ON(page_mapped(oldpage)))
 894                 goto out_fallback_unlock;
 895         if (WARN_ON(page_has_private(oldpage)))
 896                 goto out_fallback_unlock;
 897         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 898                 goto out_fallback_unlock;
 899         if (WARN_ON(PageMlocked(oldpage)))
 900                 goto out_fallback_unlock;
 901
 902         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 903         if (err) {
 904                 unlock_page(newpage);
 905                 goto out_put_old;
 906         }
 907
 908         get_page(newpage);
 909
 910         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 911                 lru_cache_add_file(newpage);
 912
 913         /*
 914          * Release while we have extra ref on stolen page.  Otherwise
 915          * anon_pipe_buf_release() might think the page can be reused.
 916          */
 917         pipe_buf_release(cs->pipe, buf);
 918
 919         err = 0;
 920         spin_lock(&cs->req->waitq.lock);
 921         if (test_bit(FR_ABORTED, &cs->req->flags))
 922                 err = -ENOENT;
 923         else
 924                 *pagep = newpage;
 925         spin_unlock(&cs->req->waitq.lock);
 926
 927         if (err) {
 928                 unlock_page(newpage);
 929                 put_page(newpage);
 930                 goto out_put_old;
 931         }
 932
 933         unlock_page(oldpage);
 934         /* Drop ref for ap->pages[] array */
 935         put_page(oldpage);
 936         cs->len = 0;
 937
 938         err = 0;
 939 out_put_old:
 940         /* Drop ref obtained in this function */
 941         put_page(oldpage);
 942         return err;
 943
 944 out_fallback_unlock:
 945         unlock_page(newpage);
 946 out_fallback:
 947         cs->pg = buf->page;
 948         cs->offset = buf->offset;
 949
 950         err = lock_request(cs->req);
 951         if (!err)
 952                 err = 1;
 953
 954         goto out_put_old;
 955 }
 956
 957 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 958                          unsigned offset, unsigned count)
 959 {
 960         struct pipe_buffer *buf;
 961         int err;
 962
 963         if (cs->nr_segs == cs->pipe->buffers)
 964                 return -EIO;
 965
 966         get_page(page);
 967         err = unlock_request(cs->req);
 968         if (err) {
 969                 put_page(page);
 970                 return err;
 971         }
 972
 973         fuse_copy_finish(cs);
 974
 975         buf = cs->pipebufs;
 976         buf->page = page;
 977         buf->offset = offset;
 978         buf->len = count;
 979
 980         cs->pipebufs++;
 981         cs->nr_segs++;
 982         cs->len = 0;
 983
 984         return 0;
 985 }
 986
 987 /*
 988  * Copy a page in the request to/from the userspace buffer.  Must be
 989  * done atomically
 990  */
 991 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 992                           unsigned offset, unsigned count, int zeroing)
 993 {
 994         int err;
 995         struct page *page = *pagep;
 996
 997         if (page && zeroing && count < PAGE_SIZE)
 998                 clear_highpage(page);
 999
1000         while (count) {
1001                 if (cs->write && cs->pipebufs && page) {
1002                         /*
1003                          * Can't control lifetime of pipe buffers, so always
1004                          * copy user pages.
1005                          */
1006                         if (cs->req->user_pages) {
1007                                 err = fuse_copy_fill(cs);
1008                                 if (err)
1009                                         return err;
1010                         } else {
1011                                 return fuse_ref_page(cs, page, offset, count);
1012                         }
1013                 } else if (!cs->len) {
1014                         if (cs->move_pages && page &&
1015                             offset == 0 && count == PAGE_SIZE) {
1016                                 err = fuse_try_move_page(cs, pagep);
1017                                 if (err <= 0)
1018                                         return err;
1019                         } else {
1020                                 err = fuse_copy_fill(cs);
1021                                 if (err)
1022                                         return err;
1023                         }
1024                 }
1025                 if (page) {
1026                         void *mapaddr = kmap_atomic(page);
1027                         void *buf = mapaddr + offset;
1028                         offset += fuse_copy_do(cs, &buf, &count);
1029                         kunmap_atomic(mapaddr);
1030                 } else
1031                         offset += fuse_copy_do(cs, NULL, &count);
1032         }
1033         if (page && !cs->write)
1034                 flush_dcache_page(page);
1035         return 0;
1036 }
1037
1038 /* Copy pages in the request to/from userspace buffer */
1039 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1040                            int zeroing)
1041 {
1042         unsigned i;
1043         struct fuse_req *req = cs->req;
1044
1045         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1046                 int err;
1047                 unsigned offset = req->page_descs[i].offset;
1048                 unsigned count = min(nbytes, req->page_descs[i].length);
1049
1050                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1051                                      zeroing);
1052                 if (err)
1053                         return err;
1054
1055                 nbytes -= count;
1056         }
1057         return 0;
1058 }
1059
1060 /* Copy a single argument in the request to/from userspace buffer */
1061 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1062 {
1063         while (size) {
1064                 if (!cs->len) {
1065                         int err = fuse_copy_fill(cs);
1066                         if (err)
1067                                 return err;
1068                 }
1069                 fuse_copy_do(cs, &val, &size);
1070         }
1071         return 0;
1072 }
1073
1074 /* Copy request arguments to/from userspace buffer */
1075 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1076                           unsigned argpages, struct fuse_arg *args,
1077                           int zeroing)
1078 {
1079         int err = 0;
1080         unsigned i;
1081
1082         for (i = 0; !err && i < numargs; i++)  {
1083                 struct fuse_arg *arg = &args[i];
1084                 if (i == numargs - 1 && argpages)
1085                         err = fuse_copy_pages(cs, arg->size, zeroing);
1086                 else
1087                         err = fuse_copy_one(cs, arg->value, arg->size);
1088         }
1089         return err;
1090 }
1091
1092 static int forget_pending(struct fuse_iqueue *fiq)
1093 {
1094         return fiq->forget_list_head.next != NULL;
1095 }
1096
1097 static int request_pending(struct fuse_iqueue *fiq)
1098 {
1099         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1100                 forget_pending(fiq);
1101 }
1102
1103 /*
1104  * Transfer an interrupt request to userspace
1105  *
1106  * Unlike other requests this is assembled on demand, without a need
1107  * to allocate a separate fuse_req structure.
1108  *
1109  * Called with fiq->lock held, releases it
1110  */
1111 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1112                                struct fuse_copy_state *cs,
1113                                size_t nbytes, struct fuse_req *req)
1114 __releases(fiq->lock)
1115 {
1116         struct fuse_in_header ih;
1117         struct fuse_interrupt_in arg;
1118         unsigned reqsize = sizeof(ih) + sizeof(arg);
1119         int err;
1120
1121         list_del_init(&req->intr_entry);
1122         req->intr_unique = fuse_get_unique(fiq);
1123         memset(&ih, 0, sizeof(ih));
1124         memset(&arg, 0, sizeof(arg));
1125         ih.len = reqsize;
1126         ih.opcode = FUSE_INTERRUPT;
1127         ih.unique = req->intr_unique;
1128         arg.unique = req->in.h.unique;
1129
1130         spin_unlock(&fiq->lock);
1131         if (nbytes < reqsize)
1132                 return -EINVAL;
1133
1134         err = fuse_copy_one(cs, &ih, sizeof(ih));
1135         if (!err)
1136                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1137         fuse_copy_finish(cs);
1138
1139         return err ? err : reqsize;
1140 }
1141
1142 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1143                                                unsigned max,
1144                                                unsigned *countp)
1145 {
1146         struct fuse_forget_link *head = fiq->forget_list_head.next;
1147         struct fuse_forget_link **newhead = &head;
1148         unsigned count;
1149
1150         for (count = 0; *newhead != NULL && count < max; count++)
1151                 newhead = &(*newhead)->next;
1152
1153         fiq->forget_list_head.next = *newhead;
1154         *newhead = NULL;
1155         if (fiq->forget_list_head.next == NULL)
1156                 fiq->forget_list_tail = &fiq->forget_list_head;
1157
1158         if (countp != NULL)
1159                 *countp = count;
1160
1161         return head;
1162 }
1163
1164 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1165                                    struct fuse_copy_state *cs,
1166                                    size_t nbytes)
1167 __releases(fiq->lock)
1168 {
1169         int err;
1170         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1171         struct fuse_forget_in arg = {
1172                 .nlookup = forget->forget_one.nlookup,
1173         };
1174         struct fuse_in_header ih = {
1175                 .opcode = FUSE_FORGET,
1176                 .nodeid = forget->forget_one.nodeid,
1177                 .unique = fuse_get_unique(fiq),
1178                 .len = sizeof(ih) + sizeof(arg),
1179         };
1180
1181         spin_unlock(&fiq->lock);
1182         kfree(forget);
1183         if (nbytes < ih.len)
1184                 return -EINVAL;
1185
1186         err = fuse_copy_one(cs, &ih, sizeof(ih));
1187         if (!err)
1188                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1189         fuse_copy_finish(cs);
1190
1191         if (err)
1192                 return err;
1193
1194         return ih.len;
1195 }
1196
1197 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1198                                    struct fuse_copy_state *cs, size_t nbytes)
1199 __releases(fiq->lock)
1200 {
1201         int err;
1202         unsigned max_forgets;
1203         unsigned count;
1204         struct fuse_forget_link *head;
1205         struct fuse_batch_forget_in arg = { .count = 0 };
1206         struct fuse_in_header ih = {
1207                 .opcode = FUSE_BATCH_FORGET,
1208                 .unique = fuse_get_unique(fiq),
1209                 .len = sizeof(ih) + sizeof(arg),
1210         };
1211
1212         if (nbytes < ih.len) {
1213                 spin_unlock(&fiq->lock);
1214                 return -EINVAL;
1215         }
1216
1217         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1218         head = dequeue_forget(fiq, max_forgets, &count);
1219         spin_unlock(&fiq->lock);
1220
1221         arg.count = count;
1222         ih.len += count * sizeof(struct fuse_forget_one);
1223         err = fuse_copy_one(cs, &ih, sizeof(ih));
1224         if (!err)
1225                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1226
1227         while (head) {
1228                 struct fuse_forget_link *forget = head;
1229
1230                 if (!err) {
1231                         err = fuse_copy_one(cs, &forget->forget_one,
1232                                             sizeof(forget->forget_one));
1233                 }
1234                 head = forget->next;
1235                 kfree(forget);
1236         }
1237
1238         fuse_copy_finish(cs);
1239
1240         if (err)
1241                 return err;
1242
1243         return ih.len;
1244 }
1245
1246 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1247                             struct fuse_copy_state *cs,
1248                             size_t nbytes)
1249 __releases(fiq->lock)
1250 {
1251         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1252                 return fuse_read_single_forget(fiq, cs, nbytes);
1253         else
1254                 return fuse_read_batch_forget(fiq, cs, nbytes);
1255 }
1256
1257 /*
1258  * Read a single request into the userspace filesystem's buffer.  This
1259  * function waits until a request is available, then removes it from
1260  * the pending list and copies request data to userspace buffer.  If
1261  * no reply is needed (FORGET) or request has been aborted or there
1262  * was an error during the copying then it's finished by calling
1263  * request_end().  Otherwise add it to the processing list, and set
1264  * the 'sent' flag.
1265  */
1266 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1267                                 struct fuse_copy_state *cs, size_t nbytes)
1268 {
1269         ssize_t err;
1270         struct fuse_conn *fc = fud->fc;
1271         struct fuse_iqueue *fiq = &fc->iq;
1272         struct fuse_pqueue *fpq = &fud->pq;
1273         struct fuse_req *req;
1274         struct fuse_in *in;
1275         unsigned reqsize;
1276
1277  restart:
1278         for (;;) {
1279                 spin_lock(&fiq->lock);
1280                 if (!fiq->connected || request_pending(fiq))
1281                         break;
1282                 spin_unlock(&fiq->lock);
1283
1284                 if (file->f_flags & O_NONBLOCK)
1285                         return -EAGAIN;
1286                 err = wait_event_interruptible_exclusive(fiq->waitq,
1287                                 !fiq->connected || request_pending(fiq));
1288                 if (err)
1289                         return err;
1290         }
1291
1292         if (!fiq->connected) {
1293                 err = (fc->aborted && fc->abort_err) ? -ECONNABORTED : -ENODEV;
1294                 goto err_unlock;
1295         }
1296
1297         if (!list_empty(&fiq->interrupts)) {
1298                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1299                                  intr_entry);
1300                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1301         }
1302
1303         if (forget_pending(fiq)) {
1304                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1305                         return fuse_read_forget(fc, fiq, cs, nbytes);
1306
1307                 if (fiq->forget_batch <= -8)
1308                         fiq->forget_batch = 16;
1309         }
1310
1311         req = list_entry(fiq->pending.next, struct fuse_req, list);
1312         clear_bit(FR_PENDING, &req->flags);
1313         list_del_init(&req->list);
1314         spin_unlock(&fiq->lock);
1315
1316         in = &req->in;
1317         reqsize = in->h.len;
1318
1319         /* If request is too large, reply with an error and restart the read */
1320         if (nbytes < reqsize) {
1321                 req->out.h.error = -EIO;
1322                 /* SETXATTR is special, since it may contain too large data */
1323                 if (in->h.opcode == FUSE_SETXATTR)
1324                         req->out.h.error = -E2BIG;
1325                 request_end(fc, req);
1326                 goto restart;
1327         }
1328         spin_lock(&fpq->lock);
1329         /*
1330          *  Must not put request on fpq->io queue after having been shut down by
1331          *  fuse_abort_conn()
1332          */
1333         if (!fpq->connected) {
1334                 req->out.h.error = err = -ECONNABORTED;
1335                 goto out_end;
1336
1337         }
1338         list_add(&req->list, &fpq->io);
1339         spin_unlock(&fpq->lock);
1340         cs->req = req;
1341         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1342         if (!err)
1343                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1344                                      (struct fuse_arg *) in->args, 0);
1345         fuse_copy_finish(cs);
1346         spin_lock(&fpq->lock);
1347         clear_bit(FR_LOCKED, &req->flags);
1348         if (!fpq->connected) {
1349                 err = (fc->aborted && fc->abort_err) ? -ECONNABORTED : -ENODEV;
1350                 goto out_end;
1351         }
1352         if (err) {
1353                 req->out.h.error = -EIO;
1354                 goto out_end;
1355         }
1356         if (!test_bit(FR_ISREPLY, &req->flags)) {
1357                 err = reqsize;
1358                 goto out_end;
1359         }
1360         list_move_tail(&req->list, &fpq->processing);
1361         __fuse_get_request(req);
1362         set_bit(FR_SENT, &req->flags);
1363         spin_unlock(&fpq->lock);
1364         /* matches barrier in request_wait_answer() */
1365         smp_mb__after_atomic();
1366         if (test_bit(FR_INTERRUPTED, &req->flags))
1367                 queue_interrupt(fiq, req);
1368         fuse_put_request(fc, req);
1369
1370         return reqsize;
1371
1372 out_end:
1373         if (!test_bit(FR_PRIVATE, &req->flags))
1374                 list_del_init(&req->list);
1375         spin_unlock(&fpq->lock);
1376         request_end(fc, req);
1377         return err;
1378
1379  err_unlock:
1380         spin_unlock(&fiq->lock);
1381         return err;
1382 }
1383
1384 static int fuse_dev_open(struct inode *inode, struct file *file)
1385 {
1386         /*
1387          * The fuse device's file's private_data is used to hold
1388          * the fuse_conn(ection) when it is mounted, and is used to
1389          * keep track of whether the file has been mounted already.
1390          */
1391         file->private_data = NULL;
1392         return 0;
1393 }
1394
1395 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1396 {
1397         struct fuse_copy_state cs;
1398         struct file *file = iocb->ki_filp;
1399         struct fuse_dev *fud = fuse_get_dev(file);
1400
1401         if (!fud)
1402                 return -EPERM;
1403
1404         if (!iter_is_iovec(to))
1405                 return -EINVAL;
1406
1407         fuse_copy_init(&cs, 1, to);
1408
1409         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1410 }
1411
1412 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1413                                     struct pipe_inode_info *pipe,
1414                                     size_t len, unsigned int flags)
1415 {
1416         int total, ret;
1417         int page_nr = 0;
1418         struct pipe_buffer *bufs;
1419         struct fuse_copy_state cs;
1420         struct fuse_dev *fud = fuse_get_dev(in);
1421
1422         if (!fud)
1423                 return -EPERM;
1424
1425         bufs = kvmalloc_array(pipe->buffers, sizeof(struct pipe_buffer),
1426                               GFP_KERNEL);
1427         if (!bufs)
1428                 return -ENOMEM;
1429
1430         fuse_copy_init(&cs, 1, NULL);
1431         cs.pipebufs = bufs;
1432         cs.pipe = pipe;
1433         ret = fuse_dev_do_read(fud, in, &cs, len);
1434         if (ret < 0)
1435                 goto out;
1436
1437         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1438                 ret = -EIO;
1439                 goto out;
1440         }
1441
1442         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1443                 /*
1444                  * Need to be careful about this.  Having buf->ops in module
1445                  * code can Oops if the buffer persists after module unload.
1446                  */
1447                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1448                 bufs[page_nr].flags = 0;
1449                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1450                 if (unlikely(ret < 0))
1451                         break;
1452         }
1453         if (total)
1454                 ret = total;
1455 out:
1456         for (; page_nr < cs.nr_segs; page_nr++)
1457                 put_page(bufs[page_nr].page);
1458
1459         kvfree(bufs);
1460         return ret;
1461 }
1462
1463 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1464                             struct fuse_copy_state *cs)
1465 {
1466         struct fuse_notify_poll_wakeup_out outarg;
1467         int err = -EINVAL;
1468
1469         if (size != sizeof(outarg))
1470                 goto err;
1471
1472         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1473         if (err)
1474                 goto err;
1475
1476         fuse_copy_finish(cs);
1477         return fuse_notify_poll_wakeup(fc, &outarg);
1478
1479 err:
1480         fuse_copy_finish(cs);
1481         return err;
1482 }
1483
1484 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1485                                    struct fuse_copy_state *cs)
1486 {
1487         struct fuse_notify_inval_inode_out outarg;
1488         int err = -EINVAL;
1489
1490         if (size != sizeof(outarg))
1491                 goto err;
1492
1493         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1494         if (err)
1495                 goto err;
1496         fuse_copy_finish(cs);
1497
1498         down_read(&fc->killsb);
1499         err = -ENOENT;
1500         if (fc->sb) {
1501                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1502                                                outarg.off, outarg.len);
1503         }
1504         up_read(&fc->killsb);
1505         return err;
1506
1507 err:
1508         fuse_copy_finish(cs);
1509         return err;
1510 }
1511
1512 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1513                                    struct fuse_copy_state *cs)
1514 {
1515         struct fuse_notify_inval_entry_out outarg;
1516         int err = -ENOMEM;
1517         char *buf;
1518         struct qstr name;
1519
1520         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1521         if (!buf)
1522                 goto err;
1523
1524         err = -EINVAL;
1525         if (size < sizeof(outarg))
1526                 goto err;
1527
1528         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1529         if (err)
1530                 goto err;
1531
1532         err = -ENAMETOOLONG;
1533         if (outarg.namelen > FUSE_NAME_MAX)
1534                 goto err;
1535
1536         err = -EINVAL;
1537         if (size != sizeof(outarg) + outarg.namelen + 1)
1538                 goto err;
1539
1540         name.name = buf;
1541         name.len = outarg.namelen;
1542         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1543         if (err)
1544                 goto err;
1545         fuse_copy_finish(cs);
1546         buf[outarg.namelen] = 0;
1547
1548         down_read(&fc->killsb);
1549         err = -ENOENT;
1550         if (fc->sb)
1551                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1552         up_read(&fc->killsb);
1553         kfree(buf);
1554         return err;
1555
1556 err:
1557         kfree(buf);
1558         fuse_copy_finish(cs);
1559         return err;
1560 }
1561
1562 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1563                               struct fuse_copy_state *cs)
1564 {
1565         struct fuse_notify_delete_out outarg;
1566         int err = -ENOMEM;
1567         char *buf;
1568         struct qstr name;
1569
1570         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1571         if (!buf)
1572                 goto err;
1573
1574         err = -EINVAL;
1575         if (size < sizeof(outarg))
1576                 goto err;
1577
1578         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1579         if (err)
1580                 goto err;
1581
1582         err = -ENAMETOOLONG;
1583         if (outarg.namelen > FUSE_NAME_MAX)
1584                 goto err;
1585
1586         err = -EINVAL;
1587         if (size != sizeof(outarg) + outarg.namelen + 1)
1588                 goto err;
1589
1590         name.name = buf;
1591         name.len = outarg.namelen;
1592         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1593         if (err)
1594                 goto err;
1595         fuse_copy_finish(cs);
1596         buf[outarg.namelen] = 0;
1597
1598         down_read(&fc->killsb);
1599         err = -ENOENT;
1600         if (fc->sb)
1601                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1602                                                outarg.child, &name);
1603         up_read(&fc->killsb);
1604         kfree(buf);
1605         return err;
1606
1607 err:
1608         kfree(buf);
1609         fuse_copy_finish(cs);
1610         return err;
1611 }
1612
1613 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1614                              struct fuse_copy_state *cs)
1615 {
1616         struct fuse_notify_store_out outarg;
1617         struct inode *inode;
1618         struct address_space *mapping;
1619         u64 nodeid;
1620         int err;
1621         pgoff_t index;
1622         unsigned int offset;
1623         unsigned int num;
1624         loff_t file_size;
1625         loff_t end;
1626
1627         err = -EINVAL;
1628         if (size < sizeof(outarg))
1629                 goto out_finish;
1630
1631         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1632         if (err)
1633                 goto out_finish;
1634
1635         err = -EINVAL;
1636         if (size - sizeof(outarg) != outarg.size)
1637                 goto out_finish;
1638
1639         nodeid = outarg.nodeid;
1640
1641         down_read(&fc->killsb);
1642
1643         err = -ENOENT;
1644         if (!fc->sb)
1645                 goto out_up_killsb;
1646
1647         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1648         if (!inode)
1649                 goto out_up_killsb;
1650
1651         mapping = inode->i_mapping;
1652         index = outarg.offset >> PAGE_SHIFT;
1653         offset = outarg.offset & ~PAGE_MASK;
1654         file_size = i_size_read(inode);
1655         end = outarg.offset + outarg.size;
1656         if (end > file_size) {
1657                 file_size = end;
1658                 fuse_write_update_size(inode, file_size);
1659         }
1660
1661         num = outarg.size;
1662         while (num) {
1663                 struct page *page;
1664                 unsigned int this_num;
1665
1666                 err = -ENOMEM;
1667                 page = find_or_create_page(mapping, index,
1668                                            mapping_gfp_mask(mapping));
1669                 if (!page)
1670                         goto out_iput;
1671
1672                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1673                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1674                 if (!err && offset == 0 &&
1675                     (this_num == PAGE_SIZE || file_size == end))
1676                         SetPageUptodate(page);
1677                 unlock_page(page);
1678                 put_page(page);
1679
1680                 if (err)
1681                         goto out_iput;
1682
1683                 num -= this_num;
1684                 offset = 0;
1685                 index++;
1686         }
1687
1688         err = 0;
1689
1690 out_iput:
1691         iput(inode);
1692 out_up_killsb:
1693         up_read(&fc->killsb);
1694 out_finish:
1695         fuse_copy_finish(cs);
1696         return err;
1697 }
1698
1699 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1700 {
1701         release_pages(req->pages, req->num_pages);
1702 }
1703
1704 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1705                          struct fuse_notify_retrieve_out *outarg)
1706 {
1707         int err;
1708         struct address_space *mapping = inode->i_mapping;
1709         struct fuse_req *req;
1710         pgoff_t index;
1711         loff_t file_size;
1712         unsigned int num;
1713         unsigned int offset;
1714         size_t total_len = 0;
1715         int num_pages;
1716
1717         offset = outarg->offset & ~PAGE_MASK;
1718         file_size = i_size_read(inode);
1719
1720         num = min(outarg->size, fc->max_write);
1721         if (outarg->offset > file_size)
1722                 num = 0;
1723         else if (outarg->offset + num > file_size)
1724                 num = file_size - outarg->offset;
1725
1726         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1727         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1728
1729         req = fuse_get_req(fc, num_pages);
1730         if (IS_ERR(req))
1731                 return PTR_ERR(req);
1732
1733         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1734         req->in.h.nodeid = outarg->nodeid;
1735         req->in.numargs = 2;
1736         req->in.argpages = 1;
1737         req->end = fuse_retrieve_end;
1738
1739         index = outarg->offset >> PAGE_SHIFT;
1740
1741         while (num && req->num_pages < num_pages) {
1742                 struct page *page;
1743                 unsigned int this_num;
1744
1745                 page = find_get_page(mapping, index);
1746                 if (!page)
1747                         break;
1748
1749                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1750                 req->pages[req->num_pages] = page;
1751                 req->page_descs[req->num_pages].offset = offset;
1752                 req->page_descs[req->num_pages].length = this_num;
1753                 req->num_pages++;
1754
1755                 offset = 0;
1756                 num -= this_num;
1757                 total_len += this_num;
1758                 index++;
1759         }
1760         req->misc.retrieve_in.offset = outarg->offset;
1761         req->misc.retrieve_in.size = total_len;
1762         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1763         req->in.args[0].value = &req->misc.retrieve_in;
1764         req->in.args[1].size = total_len;
1765
1766         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1767         if (err) {
1768                 fuse_retrieve_end(fc, req);
1769                 fuse_put_request(fc, req);
1770         }
1771
1772         return err;
1773 }
1774
1775 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1776                                 struct fuse_copy_state *cs)
1777 {
1778         struct fuse_notify_retrieve_out outarg;
1779         struct inode *inode;
1780         int err;
1781
1782         err = -EINVAL;
1783         if (size != sizeof(outarg))
1784                 goto copy_finish;
1785
1786         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1787         if (err)
1788                 goto copy_finish;
1789
1790         fuse_copy_finish(cs);
1791
1792         down_read(&fc->killsb);
1793         err = -ENOENT;
1794         if (fc->sb) {
1795                 u64 nodeid = outarg.nodeid;
1796
1797                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1798                 if (inode) {
1799                         err = fuse_retrieve(fc, inode, &outarg);
1800                         iput(inode);
1801                 }
1802         }
1803         up_read(&fc->killsb);
1804
1805         return err;
1806
1807 copy_finish:
1808         fuse_copy_finish(cs);
1809         return err;
1810 }
1811
1812 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1813                        unsigned int size, struct fuse_copy_state *cs)
1814 {
1815         /* Don't try to move pages (yet) */
1816         cs->move_pages = 0;
1817
1818         switch (code) {
1819         case FUSE_NOTIFY_POLL:
1820                 return fuse_notify_poll(fc, size, cs);
1821
1822         case FUSE_NOTIFY_INVAL_INODE:
1823                 return fuse_notify_inval_inode(fc, size, cs);
1824
1825         case FUSE_NOTIFY_INVAL_ENTRY:
1826                 return fuse_notify_inval_entry(fc, size, cs);
1827
1828         case FUSE_NOTIFY_STORE:
1829                 return fuse_notify_store(fc, size, cs);
1830
1831         case FUSE_NOTIFY_RETRIEVE:
1832                 return fuse_notify_retrieve(fc, size, cs);
1833
1834         case FUSE_NOTIFY_DELETE:
1835                 return fuse_notify_delete(fc, size, cs);
1836
1837         default:
1838                 fuse_copy_finish(cs);
1839                 return -EINVAL;
1840         }
1841 }
1842
1843 /* Look up request on processing list by unique ID */
1844 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1845 {
1846         struct fuse_req *req;
1847
1848         list_for_each_entry(req, &fpq->processing, list) {
1849                 if (req->in.h.unique == unique || req->intr_unique == unique)
1850                         return req;
1851         }
1852         return NULL;
1853 }
1854
1855 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1856                          unsigned nbytes)
1857 {
1858         unsigned reqsize = sizeof(struct fuse_out_header);
1859
1860         if (out->h.error)
1861                 return nbytes != reqsize ? -EINVAL : 0;
1862
1863         reqsize += len_args(out->numargs, out->args);
1864
1865         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1866                 return -EINVAL;
1867         else if (reqsize > nbytes) {
1868                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1869                 unsigned diffsize = reqsize - nbytes;
1870                 if (diffsize > lastarg->size)
1871                         return -EINVAL;
1872                 lastarg->size -= diffsize;
1873         }
1874         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1875                               out->page_zeroing);
1876 }
1877
1878 /*
1879  * Write a single reply to a request.  First the header is copied from
1880  * the write buffer.  The request is then searched on the processing
1881  * list by the unique ID found in the header.  If found, then remove
1882  * it from the list and copy the rest of the buffer to the request.
1883  * The request is finished by calling request_end()
1884  */
1885 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1886                                  struct fuse_copy_state *cs, size_t nbytes)
1887 {
1888         int err;
1889         struct fuse_conn *fc = fud->fc;
1890         struct fuse_pqueue *fpq = &fud->pq;
1891         struct fuse_req *req;
1892         struct fuse_out_header oh;
1893
1894         if (nbytes < sizeof(struct fuse_out_header))
1895                 return -EINVAL;
1896
1897         err = fuse_copy_one(cs, &oh, sizeof(oh));
1898         if (err)
1899                 goto err_finish;
1900
1901         err = -EINVAL;
1902         if (oh.len != nbytes)
1903                 goto err_finish;
1904
1905         /*
1906          * Zero oh.unique indicates unsolicited notification message
1907          * and error contains notification code.
1908          */
1909         if (!oh.unique) {
1910                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1911                 return err ? err : nbytes;
1912         }
1913
1914         err = -EINVAL;
1915         if (oh.error <= -512 || oh.error > 0)
1916                 goto err_finish;
1917
1918         spin_lock(&fpq->lock);
1919         err = -ENOENT;
1920         if (!fpq->connected)
1921                 goto err_unlock_pq;
1922
1923         req = request_find(fpq, oh.unique);
1924         if (!req)
1925                 goto err_unlock_pq;
1926
1927         /* Is it an interrupt reply? */
1928         if (req->intr_unique == oh.unique) {
1929                 __fuse_get_request(req);
1930                 spin_unlock(&fpq->lock);
1931
1932                 err = -EINVAL;
1933                 if (nbytes != sizeof(struct fuse_out_header)) {
1934                         fuse_put_request(fc, req);
1935                         goto err_finish;
1936                 }
1937
1938                 if (oh.error == -ENOSYS)
1939                         fc->no_interrupt = 1;
1940                 else if (oh.error == -EAGAIN)
1941                         queue_interrupt(&fc->iq, req);
1942                 fuse_put_request(fc, req);
1943
1944                 fuse_copy_finish(cs);
1945                 return nbytes;
1946         }
1947
1948         clear_bit(FR_SENT, &req->flags);
1949         list_move(&req->list, &fpq->io);
1950         req->out.h = oh;
1951         set_bit(FR_LOCKED, &req->flags);
1952         spin_unlock(&fpq->lock);
1953         cs->req = req;
1954         if (!req->out.page_replace)
1955                 cs->move_pages = 0;
1956
1957         err = copy_out_args(cs, &req->out, nbytes);
1958         fuse_copy_finish(cs);
1959
1960         spin_lock(&fpq->lock);
1961         clear_bit(FR_LOCKED, &req->flags);
1962         if (!fpq->connected)
1963                 err = -ENOENT;
1964         else if (err)
1965                 req->out.h.error = -EIO;
1966         if (!test_bit(FR_PRIVATE, &req->flags))
1967                 list_del_init(&req->list);
1968         spin_unlock(&fpq->lock);
1969
1970         request_end(fc, req);
1971
1972         return err ? err : nbytes;
1973
1974  err_unlock_pq:
1975         spin_unlock(&fpq->lock);
1976  err_finish:
1977         fuse_copy_finish(cs);
1978         return err;
1979 }
1980
1981 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1982 {
1983         struct fuse_copy_state cs;
1984         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1985
1986         if (!fud)
1987                 return -EPERM;
1988
1989         if (!iter_is_iovec(from))
1990                 return -EINVAL;
1991
1992         fuse_copy_init(&cs, 0, from);
1993
1994         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1995 }
1996
1997 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1998                                      struct file *out, loff_t *ppos,
1999                                      size_t len, unsigned int flags)
2000 {
2001         unsigned nbuf;
2002         unsigned idx;
2003         struct pipe_buffer *bufs;
2004         struct fuse_copy_state cs;
2005         struct fuse_dev *fud;
2006         size_t rem;
2007         ssize_t ret;
2008
2009         fud = fuse_get_dev(out);
2010         if (!fud)
2011                 return -EPERM;
2012
2013         pipe_lock(pipe);
2014
2015         bufs = kvmalloc_array(pipe->nrbufs, sizeof(struct pipe_buffer),
2016                               GFP_KERNEL);
2017         if (!bufs) {
2018                 pipe_unlock(pipe);
2019                 return -ENOMEM;
2020         }
2021
2022         nbuf = 0;
2023         rem = 0;
2024         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2025                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2026
2027         ret = -EINVAL;
2028         if (rem < len)
2029                 goto out_free;
2030
2031         rem = len;
2032         while (rem) {
2033                 struct pipe_buffer *ibuf;
2034                 struct pipe_buffer *obuf;
2035
2036                 BUG_ON(nbuf >= pipe->buffers);
2037                 BUG_ON(!pipe->nrbufs);
2038                 ibuf = &pipe->bufs[pipe->curbuf];
2039                 obuf = &bufs[nbuf];
2040
2041                 if (rem >= ibuf->len) {
2042                         *obuf = *ibuf;
2043                         ibuf->ops = NULL;
2044                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2045                         pipe->nrbufs--;
2046                 } else {
2047                         if (!pipe_buf_get(pipe, ibuf))
2048                                 goto out_free;
2049
2050                         *obuf = *ibuf;
2051                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2052                         obuf->len = rem;
2053                         ibuf->offset += obuf->len;
2054                         ibuf->len -= obuf->len;
2055                 }
2056                 nbuf++;
2057                 rem -= obuf->len;
2058         }
2059         pipe_unlock(pipe);
2060
2061         fuse_copy_init(&cs, 0, NULL);
2062         cs.pipebufs = bufs;
2063         cs.nr_segs = nbuf;
2064         cs.pipe = pipe;
2065
2066         if (flags & SPLICE_F_MOVE)
2067                 cs.move_pages = 1;
2068
2069         ret = fuse_dev_do_write(fud, &cs, len);
2070
2071         pipe_lock(pipe);
2072 out_free:
2073         for (idx = 0; idx < nbuf; idx++) {
2074                 struct pipe_buffer *buf = &bufs[idx];
2075
2076                 if (buf->ops)
2077                         pipe_buf_release(pipe, buf);
2078         }
2079         pipe_unlock(pipe);
2080
2081         kvfree(bufs);
2082         return ret;
2083 }
2084
2085 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2086 {
2087         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2088         struct fuse_iqueue *fiq;
2089         struct fuse_dev *fud = fuse_get_dev(file);
2090
2091         if (!fud)
2092                 return EPOLLERR;
2093
2094         fiq = &fud->fc->iq;
2095         poll_wait(file, &fiq->waitq, wait);
2096
2097         spin_lock(&fiq->lock);
2098         if (!fiq->connected)
2099                 mask = EPOLLERR;
2100         else if (request_pending(fiq))
2101                 mask |= EPOLLIN | EPOLLRDNORM;
2102         spin_unlock(&fiq->lock);
2103
2104         return mask;
2105 }
2106
2107 /*
2108  * Abort all requests on the given list (pending or processing)
2109  *
2110  * This function releases and reacquires fc->lock
2111  */
2112 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2113 {
2114         while (!list_empty(head)) {
2115                 struct fuse_req *req;
2116                 req = list_entry(head->next, struct fuse_req, list);
2117                 req->out.h.error = -ECONNABORTED;
2118                 clear_bit(FR_SENT, &req->flags);
2119                 list_del_init(&req->list);
2120                 request_end(fc, req);
2121         }
2122 }
2123
2124 static void end_polls(struct fuse_conn *fc)
2125 {
2126         struct rb_node *p;
2127
2128         p = rb_first(&fc->polled_files);
2129
2130         while (p) {
2131                 struct fuse_file *ff;
2132                 ff = rb_entry(p, struct fuse_file, polled_node);
2133                 wake_up_interruptible_all(&ff->poll_wait);
2134
2135                 p = rb_next(p);
2136         }
2137 }
2138
2139 /*
2140  * Abort all requests.
2141  *
2142  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2143  * filesystem.
2144  *
2145  * The same effect is usually achievable through killing the filesystem daemon
2146  * and all users of the filesystem.  The exception is the combination of an
2147  * asynchronous request and the tricky deadlock (see
2148  * Documentation/filesystems/fuse.txt).
2149  *
2150  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2151  * requests, they should be finished off immediately.  Locked requests will be
2152  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2153  * requests.  It is possible that some request will finish before we can.  This
2154  * is OK, the request will in that case be removed from the list before we touch
2155  * it.
2156  */
2157 void fuse_abort_conn(struct fuse_conn *fc, bool is_abort)
2158 {
2159         struct fuse_iqueue *fiq = &fc->iq;
2160
2161         spin_lock(&fc->lock);
2162         if (fc->connected) {
2163                 struct fuse_dev *fud;
2164                 struct fuse_req *req, *next;
2165                 LIST_HEAD(to_end);
2166
2167                 fc->connected = 0;
2168                 fc->blocked = 0;
2169                 fc->aborted = is_abort;
2170                 fuse_set_initialized(fc);
2171                 list_for_each_entry(fud, &fc->devices, entry) {
2172                         struct fuse_pqueue *fpq = &fud->pq;
2173
2174                         spin_lock(&fpq->lock);
2175                         fpq->connected = 0;
2176                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2177                                 req->out.h.error = -ECONNABORTED;
2178                                 spin_lock(&req->waitq.lock);
2179                                 set_bit(FR_ABORTED, &req->flags);
2180                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2181                                         set_bit(FR_PRIVATE, &req->flags);
2182                                         __fuse_get_request(req);
2183                                         list_move(&req->list, &to_end);
2184                                 }
2185                                 spin_unlock(&req->waitq.lock);
2186                         }
2187                         list_splice_tail_init(&fpq->processing, &to_end);
2188                         spin_unlock(&fpq->lock);
2189                 }
2190                 fc->max_background = UINT_MAX;
2191                 flush_bg_queue(fc);
2192
2193                 spin_lock(&fiq->lock);
2194                 fiq->connected = 0;
2195                 list_for_each_entry(req, &fiq->pending, list)
2196                         clear_bit(FR_PENDING, &req->flags);
2197                 list_splice_tail_init(&fiq->pending, &to_end);
2198                 while (forget_pending(fiq))
2199                         kfree(dequeue_forget(fiq, 1, NULL));
2200                 wake_up_all(&fiq->waitq);
2201                 spin_unlock(&fiq->lock);
2202                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2203                 end_polls(fc);
2204                 wake_up_all(&fc->blocked_waitq);
2205                 spin_unlock(&fc->lock);
2206
2207                 end_requests(fc, &to_end);
2208         } else {
2209                 spin_unlock(&fc->lock);
2210         }
2211 }
2212 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2213
2214 void fuse_wait_aborted(struct fuse_conn *fc)
2215 {
2216         /* matches implicit memory barrier in fuse_drop_waiting() */
2217         smp_mb();
2218         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2219 }
2220
2221 int fuse_dev_release(struct inode *inode, struct file *file)
2222 {
2223         struct fuse_dev *fud = fuse_get_dev(file);
2224
2225         if (fud) {
2226                 struct fuse_conn *fc = fud->fc;
2227                 struct fuse_pqueue *fpq = &fud->pq;
2228                 LIST_HEAD(to_end);
2229
2230                 spin_lock(&fpq->lock);
2231                 WARN_ON(!list_empty(&fpq->io));
2232                 list_splice_init(&fpq->processing, &to_end);
2233                 spin_unlock(&fpq->lock);
2234
2235                 end_requests(fc, &to_end);
2236
2237                 /* Are we the last open device? */
2238                 if (atomic_dec_and_test(&fc->dev_count)) {
2239                         WARN_ON(fc->iq.fasync != NULL);
2240                         fuse_abort_conn(fc, false);
2241                 }
2242                 fuse_dev_free(fud);
2243         }
2244         return 0;
2245 }
2246 EXPORT_SYMBOL_GPL(fuse_dev_release);
2247
2248 static int fuse_dev_fasync(int fd, struct file *file, int on)
2249 {
2250         struct fuse_dev *fud = fuse_get_dev(file);
2251
2252         if (!fud)
2253                 return -EPERM;
2254
2255         /* No locking - fasync_helper does its own locking */
2256         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2257 }
2258
2259 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2260 {
2261         struct fuse_dev *fud;
2262
2263         if (new->private_data)
2264                 return -EINVAL;
2265
2266         fud = fuse_dev_alloc(fc);
2267         if (!fud)
2268                 return -ENOMEM;
2269
2270         new->private_data = fud;
2271         atomic_inc(&fc->dev_count);
2272
2273         return 0;
2274 }
2275
2276 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2277                            unsigned long arg)
2278 {
2279         int err = -ENOTTY;
2280
2281         if (cmd == FUSE_DEV_IOC_CLONE) {
2282                 int oldfd;
2283
2284                 err = -EFAULT;
2285                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2286                         struct file *old = fget(oldfd);
2287
2288                         err = -EINVAL;
2289                         if (old) {
2290                                 struct fuse_dev *fud = NULL;
2291
2292                                 /*
2293                                  * Check against file->f_op because CUSE
2294                                  * uses the same ioctl handler.
2295                                  */
2296                                 if (old->f_op == file->f_op &&
2297                                     old->f_cred->user_ns == file->f_cred->user_ns)
2298                                         fud = fuse_get_dev(old);
2299
2300                                 if (fud) {
2301                                         mutex_lock(&fuse_mutex);
2302                                         err = fuse_device_clone(fud->fc, file);
2303                                         mutex_unlock(&fuse_mutex);
2304                                 }
2305                                 fput(old);
2306                         }
2307                 }
2308         }
2309         return err;
2310 }
2311
2312 const struct file_operations fuse_dev_operations = {
2313         .owner          = THIS_MODULE,
2314         .open           = fuse_dev_open,
2315         .llseek         = no_llseek,
2316         .read_iter      = fuse_dev_read,
2317         .splice_read    = fuse_dev_splice_read,
2318         .write_iter     = fuse_dev_write,
2319         .splice_write   = fuse_dev_splice_write,
2320         .poll           = fuse_dev_poll,
2321         .release        = fuse_dev_release,
2322         .fasync         = fuse_dev_fasync,
2323         .unlocked_ioctl = fuse_dev_ioctl,
2324         .compat_ioctl   = fuse_dev_ioctl,
2325 };
2326 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2327
2328 static struct miscdevice fuse_miscdevice = {
2329         .minor = FUSE_MINOR,
2330         .name  = "fuse",
2331         .fops = &fuse_dev_operations,
2332 };
2333
2334 int __init fuse_dev_init(void)
2335 {
2336         int err = -ENOMEM;
2337         fuse_req_cachep = kmem_cache_create("fuse_request",
2338                                             sizeof(struct fuse_req),
2339                                             0, 0, NULL);
2340         if (!fuse_req_cachep)
2341                 goto out;
2342
2343         err = misc_register(&fuse_miscdevice);
2344         if (err)
2345                 goto out_cache_clean;
2346
2347         return 0;
2348
2349  out_cache_clean:
2350         kmem_cache_destroy(fuse_req_cachep);
2351  out:
2352         return err;
2353 }
2354
2355 void fuse_dev_cleanup(void)
2356 {
2357         misc_deregister(&fuse_miscdevice);
2358         kmem_cache_destroy(fuse_req_cachep);
2359 }