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 /* Ordinary requests have even IDs, while interrupts IDs are odd */
  29 #define FUSE_INT_REQ_BIT (1ULL << 0)
  30 #define FUSE_REQ_ID_STEP (1ULL << 1)
  31
  32 static struct kmem_cache *fuse_req_cachep;
  33
  34 static struct fuse_dev *fuse_get_dev(struct file *file)
  35 {
  36         /*
  37          * Lockless access is OK, because file->private data is set
  38          * once during mount and is valid until the file is released.
  39          */
  40         return READ_ONCE(file->private_data);
  41 }
  42
  43 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  44                               struct fuse_page_desc *page_descs,
  45                               unsigned npages)
  46 {
  47         memset(req, 0, sizeof(*req));
  48         memset(pages, 0, sizeof(*pages) * npages);
  49         memset(page_descs, 0, sizeof(*page_descs) * npages);
  50         INIT_LIST_HEAD(&req->list);
  51         INIT_LIST_HEAD(&req->intr_entry);
  52         init_waitqueue_head(&req->waitq);
  53         refcount_set(&req->count, 1);
  54         req->pages = pages;
  55         req->page_descs = page_descs;
  56         req->max_pages = npages;
  57         __set_bit(FR_PENDING, &req->flags);
  58 }
  59
  60 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  61 {
  62         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, flags);
  63         if (req) {
  64                 struct page **pages;
  65                 struct fuse_page_desc *page_descs;
  66
  67                 if (npages <= FUSE_REQ_INLINE_PAGES) {
  68                         pages = req->inline_pages;
  69                         page_descs = req->inline_page_descs;
  70                 } else {
  71                         pages = kmalloc_array(npages, sizeof(struct page *),
  72                                               flags);
  73                         page_descs =
  74                                 kmalloc_array(npages,
  75                                               sizeof(struct fuse_page_desc),
  76                                               flags);
  77                 }
  78
  79                 if (!pages || !page_descs) {
  80                         kfree(pages);
  81                         kfree(page_descs);
  82                         kmem_cache_free(fuse_req_cachep, req);
  83                         return NULL;
  84                 }
  85
  86                 fuse_request_init(req, pages, page_descs, npages);
  87         }
  88         return req;
  89 }
  90
  91 struct fuse_req *fuse_request_alloc(unsigned npages)
  92 {
  93         return __fuse_request_alloc(npages, GFP_KERNEL);
  94 }
  95 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  96
  97 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
  98 {
  99         return __fuse_request_alloc(npages, GFP_NOFS);
 100 }
 101
 102 void fuse_request_free(struct fuse_req *req)
 103 {
 104         if (req->pages != req->inline_pages) {
 105                 kfree(req->pages);
 106                 kfree(req->page_descs);
 107         }
 108         kmem_cache_free(fuse_req_cachep, req);
 109 }
 110
 111 void __fuse_get_request(struct fuse_req *req)
 112 {
 113         refcount_inc(&req->count);
 114 }
 115
 116 /* Must be called with > 1 refcount */
 117 static void __fuse_put_request(struct fuse_req *req)
 118 {
 119         refcount_dec(&req->count);
 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         if (fc->connected) {
 137                 atomic_dec(&fc->num_waiting);
 138         } else if (atomic_dec_and_test(&fc->num_waiting)) {
 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->bg_lock);
 295                         if (!fc->blocked)
 296                                 wake_up(&fc->blocked_waitq);
 297                         spin_unlock(&fc->bg_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         fiq->reqctr += FUSE_REQ_ID_STEP;
 327         return fiq->reqctr;
 328 }
 329
 330 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 331 {
 332         req->in.h.len = sizeof(struct fuse_in_header) +
 333                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 334         list_add_tail(&req->list, &fiq->pending);
 335         wake_up_locked(&fiq->waitq);
 336         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 337 }
 338
 339 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 340                        u64 nodeid, u64 nlookup)
 341 {
 342         struct fuse_iqueue *fiq = &fc->iq;
 343
 344         forget->forget_one.nodeid = nodeid;
 345         forget->forget_one.nlookup = nlookup;
 346
 347         spin_lock(&fiq->waitq.lock);
 348         if (fiq->connected) {
 349                 fiq->forget_list_tail->next = forget;
 350                 fiq->forget_list_tail = forget;
 351                 wake_up_locked(&fiq->waitq);
 352                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 353         } else {
 354                 kfree(forget);
 355         }
 356         spin_unlock(&fiq->waitq.lock);
 357 }
 358
 359 static void flush_bg_queue(struct fuse_conn *fc)
 360 {
 361         struct fuse_iqueue *fiq = &fc->iq;
 362
 363         while (fc->active_background < fc->max_background &&
 364                !list_empty(&fc->bg_queue)) {
 365                 struct fuse_req *req;
 366
 367                 req = list_first_entry(&fc->bg_queue, struct fuse_req, list);
 368                 list_del(&req->list);
 369                 fc->active_background++;
 370                 spin_lock(&fiq->waitq.lock);
 371                 req->in.h.unique = fuse_get_unique(fiq);
 372                 queue_request(fiq, req);
 373                 spin_unlock(&fiq->waitq.lock);
 374         }
 375 }
 376
 377 /*
 378  * This function is called when a request is finished.  Either a reply
 379  * has arrived or it was aborted (and not yet sent) or some error
 380  * occurred during communication with userspace, or the device file
 381  * was closed.  The requester thread is woken up (if still waiting),
 382  * the 'end' callback is called if given, else the reference to the
 383  * request is released
 384  */
 385 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 386 {
 387         struct fuse_iqueue *fiq = &fc->iq;
 388
 389         if (test_and_set_bit(FR_FINISHED, &req->flags))
 390                 goto put_request;
 391
 392         spin_lock(&fiq->waitq.lock);
 393         list_del_init(&req->intr_entry);
 394         spin_unlock(&fiq->waitq.lock);
 395         WARN_ON(test_bit(FR_PENDING, &req->flags));
 396         WARN_ON(test_bit(FR_SENT, &req->flags));
 397         if (test_bit(FR_BACKGROUND, &req->flags)) {
 398                 spin_lock(&fc->bg_lock);
 399                 clear_bit(FR_BACKGROUND, &req->flags);
 400                 if (fc->num_background == fc->max_background) {
 401                         fc->blocked = 0;
 402                         wake_up(&fc->blocked_waitq);
 403                 } else if (!fc->blocked) {
 404                         /*
 405                          * Wake up next waiter, if any.  It's okay to use
 406                          * waitqueue_active(), as we've already synced up
 407                          * fc->blocked with waiters with the wake_up() call
 408                          * above.
 409                          */
 410                         if (waitqueue_active(&fc->blocked_waitq))
 411                                 wake_up(&fc->blocked_waitq);
 412                 }
 413
 414                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 415                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 416                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 417                 }
 418                 fc->num_background--;
 419                 fc->active_background--;
 420                 flush_bg_queue(fc);
 421                 spin_unlock(&fc->bg_lock);
 422         }
 423         wake_up(&req->waitq);
 424         if (req->end)
 425                 req->end(fc, req);
 426 put_request:
 427         fuse_put_request(fc, req);
 428 }
 429
 430 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 431 {
 432         spin_lock(&fiq->waitq.lock);
 433         if (test_bit(FR_FINISHED, &req->flags)) {
 434                 spin_unlock(&fiq->waitq.lock);
 435                 return;
 436         }
 437         if (list_empty(&req->intr_entry)) {
 438                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 439                 wake_up_locked(&fiq->waitq);
 440         }
 441         spin_unlock(&fiq->waitq.lock);
 442         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 443 }
 444
 445 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 446 {
 447         struct fuse_iqueue *fiq = &fc->iq;
 448         int err;
 449
 450         if (!fc->no_interrupt) {
 451                 /* Any signal may interrupt this */
 452                 err = wait_event_interruptible(req->waitq,
 453                                         test_bit(FR_FINISHED, &req->flags));
 454                 if (!err)
 455                         return;
 456
 457                 set_bit(FR_INTERRUPTED, &req->flags);
 458                 /* matches barrier in fuse_dev_do_read() */
 459                 smp_mb__after_atomic();
 460                 if (test_bit(FR_SENT, &req->flags))
 461                         queue_interrupt(fiq, req);
 462         }
 463
 464         if (!test_bit(FR_FORCE, &req->flags)) {
 465                 /* Only fatal signals may interrupt this */
 466                 err = wait_event_killable(req->waitq,
 467                                         test_bit(FR_FINISHED, &req->flags));
 468                 if (!err)
 469                         return;
 470
 471                 spin_lock(&fiq->waitq.lock);
 472                 /* Request is not yet in userspace, bail out */
 473                 if (test_bit(FR_PENDING, &req->flags)) {
 474                         list_del(&req->list);
 475                         spin_unlock(&fiq->waitq.lock);
 476                         __fuse_put_request(req);
 477                         req->out.h.error = -EINTR;
 478                         return;
 479                 }
 480                 spin_unlock(&fiq->waitq.lock);
 481         }
 482
 483         /*
 484          * Either request is already in userspace, or it was forced.
 485          * Wait it out.
 486          */
 487         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 488 }
 489
 490 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 491 {
 492         struct fuse_iqueue *fiq = &fc->iq;
 493
 494         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 495         spin_lock(&fiq->waitq.lock);
 496         if (!fiq->connected) {
 497                 spin_unlock(&fiq->waitq.lock);
 498                 req->out.h.error = -ENOTCONN;
 499         } else {
 500                 req->in.h.unique = fuse_get_unique(fiq);
 501                 queue_request(fiq, req);
 502                 /* acquire extra reference, since request is still needed
 503                    after request_end() */
 504                 __fuse_get_request(req);
 505                 spin_unlock(&fiq->waitq.lock);
 506
 507                 request_wait_answer(fc, req);
 508                 /* Pairs with smp_wmb() in request_end() */
 509                 smp_rmb();
 510         }
 511 }
 512
 513 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 514 {
 515         __set_bit(FR_ISREPLY, &req->flags);
 516         if (!test_bit(FR_WAITING, &req->flags)) {
 517                 __set_bit(FR_WAITING, &req->flags);
 518                 atomic_inc(&fc->num_waiting);
 519         }
 520         __fuse_request_send(fc, req);
 521 }
 522 EXPORT_SYMBOL_GPL(fuse_request_send);
 523
 524 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 525 {
 526         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 527                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 528
 529         if (fc->minor < 9) {
 530                 switch (args->in.h.opcode) {
 531                 case FUSE_LOOKUP:
 532                 case FUSE_CREATE:
 533                 case FUSE_MKNOD:
 534                 case FUSE_MKDIR:
 535                 case FUSE_SYMLINK:
 536                 case FUSE_LINK:
 537                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 538                         break;
 539                 case FUSE_GETATTR:
 540                 case FUSE_SETATTR:
 541                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 542                         break;
 543                 }
 544         }
 545         if (fc->minor < 12) {
 546                 switch (args->in.h.opcode) {
 547                 case FUSE_CREATE:
 548                         args->in.args[0].size = sizeof(struct fuse_open_in);
 549                         break;
 550                 case FUSE_MKNOD:
 551                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 552                         break;
 553                 }
 554         }
 555 }
 556
 557 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 558 {
 559         struct fuse_req *req;
 560         ssize_t ret;
 561
 562         req = fuse_get_req(fc, 0);
 563         if (IS_ERR(req))
 564                 return PTR_ERR(req);
 565
 566         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 567         fuse_adjust_compat(fc, args);
 568
 569         req->in.h.opcode = args->in.h.opcode;
 570         req->in.h.nodeid = args->in.h.nodeid;
 571         req->in.numargs = args->in.numargs;
 572         memcpy(req->in.args, args->in.args,
 573                args->in.numargs * sizeof(struct fuse_in_arg));
 574         req->out.argvar = args->out.argvar;
 575         req->out.numargs = args->out.numargs;
 576         memcpy(req->out.args, args->out.args,
 577                args->out.numargs * sizeof(struct fuse_arg));
 578         fuse_request_send(fc, req);
 579         ret = req->out.h.error;
 580         if (!ret && args->out.argvar) {
 581                 BUG_ON(args->out.numargs != 1);
 582                 ret = req->out.args[0].size;
 583         }
 584         fuse_put_request(fc, req);
 585
 586         return ret;
 587 }
 588
 589 bool fuse_request_queue_background(struct fuse_conn *fc, struct fuse_req *req)
 590 {
 591         bool queued = false;
 592
 593         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 594         if (!test_bit(FR_WAITING, &req->flags)) {
 595                 __set_bit(FR_WAITING, &req->flags);
 596                 atomic_inc(&fc->num_waiting);
 597         }
 598         __set_bit(FR_ISREPLY, &req->flags);
 599         spin_lock(&fc->bg_lock);
 600         if (likely(fc->connected)) {
 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                 queued = true;
 611         }
 612         spin_unlock(&fc->bg_lock);
 613
 614         return queued;
 615 }
 616
 617 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 618 {
 619         WARN_ON(!req->end);
 620         if (!fuse_request_queue_background(fc, req)) {
 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->waitq.lock);
 637         if (fiq->connected) {
 638                 queue_request(fiq, req);
 639                 err = 0;
 640         }
 641         spin_unlock(&fiq->waitq.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_count(page) != 1 ||
 835             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 836              ~(1 << PG_locked |
 837                1 << PG_referenced |
 838                1 << PG_uptodate |
 839                1 << PG_lru |
 840                1 << PG_active |
 841                1 << PG_reclaim))) {
 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         err = unlock_request(cs->req);
 857         if (err)
 858                 return err;
 859
 860         fuse_copy_finish(cs);
 861
 862         err = pipe_buf_confirm(cs->pipe, buf);
 863         if (err)
 864                 return err;
 865
 866         BUG_ON(!cs->nr_segs);
 867         cs->currbuf = buf;
 868         cs->len = buf->len;
 869         cs->pipebufs++;
 870         cs->nr_segs--;
 871
 872         if (cs->len != PAGE_SIZE)
 873                 goto out_fallback;
 874
 875         if (pipe_buf_steal(cs->pipe, buf) != 0)
 876                 goto out_fallback;
 877
 878         newpage = buf->page;
 879
 880         if (!PageUptodate(newpage))
 881                 SetPageUptodate(newpage);
 882
 883         ClearPageMappedToDisk(newpage);
 884
 885         if (fuse_check_page(newpage) != 0)
 886                 goto out_fallback_unlock;
 887
 888         /*
 889          * This is a new and locked page, it shouldn't be mapped or
 890          * have any special flags on it
 891          */
 892         if (WARN_ON(page_mapped(oldpage)))
 893                 goto out_fallback_unlock;
 894         if (WARN_ON(page_has_private(oldpage)))
 895                 goto out_fallback_unlock;
 896         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 897                 goto out_fallback_unlock;
 898         if (WARN_ON(PageMlocked(oldpage)))
 899                 goto out_fallback_unlock;
 900
 901         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 902         if (err) {
 903                 unlock_page(newpage);
 904                 return err;
 905         }
 906
 907         get_page(newpage);
 908
 909         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 910                 lru_cache_add_file(newpage);
 911
 912         err = 0;
 913         spin_lock(&cs->req->waitq.lock);
 914         if (test_bit(FR_ABORTED, &cs->req->flags))
 915                 err = -ENOENT;
 916         else
 917                 *pagep = newpage;
 918         spin_unlock(&cs->req->waitq.lock);
 919
 920         if (err) {
 921                 unlock_page(newpage);
 922                 put_page(newpage);
 923                 return err;
 924         }
 925
 926         unlock_page(oldpage);
 927         put_page(oldpage);
 928         cs->len = 0;
 929
 930         return 0;
 931
 932 out_fallback_unlock:
 933         unlock_page(newpage);
 934 out_fallback:
 935         cs->pg = buf->page;
 936         cs->offset = buf->offset;
 937
 938         err = lock_request(cs->req);
 939         if (err)
 940                 return err;
 941
 942         return 1;
 943 }
 944
 945 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 946                          unsigned offset, unsigned count)
 947 {
 948         struct pipe_buffer *buf;
 949         int err;
 950
 951         if (cs->nr_segs == cs->pipe->buffers)
 952                 return -EIO;
 953
 954         err = unlock_request(cs->req);
 955         if (err)
 956                 return err;
 957
 958         fuse_copy_finish(cs);
 959
 960         buf = cs->pipebufs;
 961         get_page(page);
 962         buf->page = page;
 963         buf->offset = offset;
 964         buf->len = count;
 965
 966         cs->pipebufs++;
 967         cs->nr_segs++;
 968         cs->len = 0;
 969
 970         return 0;
 971 }
 972
 973 /*
 974  * Copy a page in the request to/from the userspace buffer.  Must be
 975  * done atomically
 976  */
 977 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 978                           unsigned offset, unsigned count, int zeroing)
 979 {
 980         int err;
 981         struct page *page = *pagep;
 982
 983         if (page && zeroing && count < PAGE_SIZE)
 984                 clear_highpage(page);
 985
 986         while (count) {
 987                 if (cs->write && cs->pipebufs && page) {
 988                         return fuse_ref_page(cs, page, offset, count);
 989                 } else if (!cs->len) {
 990                         if (cs->move_pages && page &&
 991                             offset == 0 && count == PAGE_SIZE) {
 992                                 err = fuse_try_move_page(cs, pagep);
 993                                 if (err <= 0)
 994                                         return err;
 995                         } else {
 996                                 err = fuse_copy_fill(cs);
 997                                 if (err)
 998                                         return err;
 999                         }
1000                 }
1001                 if (page) {
1002                         void *mapaddr = kmap_atomic(page);
1003                         void *buf = mapaddr + offset;
1004                         offset += fuse_copy_do(cs, &buf, &count);
1005                         kunmap_atomic(mapaddr);
1006                 } else
1007                         offset += fuse_copy_do(cs, NULL, &count);
1008         }
1009         if (page && !cs->write)
1010                 flush_dcache_page(page);
1011         return 0;
1012 }
1013
1014 /* Copy pages in the request to/from userspace buffer */
1015 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1016                            int zeroing)
1017 {
1018         unsigned i;
1019         struct fuse_req *req = cs->req;
1020
1021         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1022                 int err;
1023                 unsigned offset = req->page_descs[i].offset;
1024                 unsigned count = min(nbytes, req->page_descs[i].length);
1025
1026                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1027                                      zeroing);
1028                 if (err)
1029                         return err;
1030
1031                 nbytes -= count;
1032         }
1033         return 0;
1034 }
1035
1036 /* Copy a single argument in the request to/from userspace buffer */
1037 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1038 {
1039         while (size) {
1040                 if (!cs->len) {
1041                         int err = fuse_copy_fill(cs);
1042                         if (err)
1043                                 return err;
1044                 }
1045                 fuse_copy_do(cs, &val, &size);
1046         }
1047         return 0;
1048 }
1049
1050 /* Copy request arguments to/from userspace buffer */
1051 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1052                           unsigned argpages, struct fuse_arg *args,
1053                           int zeroing)
1054 {
1055         int err = 0;
1056         unsigned i;
1057
1058         for (i = 0; !err && i < numargs; i++)  {
1059                 struct fuse_arg *arg = &args[i];
1060                 if (i == numargs - 1 && argpages)
1061                         err = fuse_copy_pages(cs, arg->size, zeroing);
1062                 else
1063                         err = fuse_copy_one(cs, arg->value, arg->size);
1064         }
1065         return err;
1066 }
1067
1068 static int forget_pending(struct fuse_iqueue *fiq)
1069 {
1070         return fiq->forget_list_head.next != NULL;
1071 }
1072
1073 static int request_pending(struct fuse_iqueue *fiq)
1074 {
1075         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1076                 forget_pending(fiq);
1077 }
1078
1079 /*
1080  * Transfer an interrupt request to userspace
1081  *
1082  * Unlike other requests this is assembled on demand, without a need
1083  * to allocate a separate fuse_req structure.
1084  *
1085  * Called with fiq->waitq.lock held, releases it
1086  */
1087 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1088                                struct fuse_copy_state *cs,
1089                                size_t nbytes, struct fuse_req *req)
1090 __releases(fiq->waitq.lock)
1091 {
1092         struct fuse_in_header ih;
1093         struct fuse_interrupt_in arg;
1094         unsigned reqsize = sizeof(ih) + sizeof(arg);
1095         int err;
1096
1097         list_del_init(&req->intr_entry);
1098         memset(&ih, 0, sizeof(ih));
1099         memset(&arg, 0, sizeof(arg));
1100         ih.len = reqsize;
1101         ih.opcode = FUSE_INTERRUPT;
1102         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1103         arg.unique = req->in.h.unique;
1104
1105         spin_unlock(&fiq->waitq.lock);
1106         if (nbytes < reqsize)
1107                 return -EINVAL;
1108
1109         err = fuse_copy_one(cs, &ih, sizeof(ih));
1110         if (!err)
1111                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1112         fuse_copy_finish(cs);
1113
1114         return err ? err : reqsize;
1115 }
1116
1117 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1118                                                unsigned max,
1119                                                unsigned *countp)
1120 {
1121         struct fuse_forget_link *head = fiq->forget_list_head.next;
1122         struct fuse_forget_link **newhead = &head;
1123         unsigned count;
1124
1125         for (count = 0; *newhead != NULL && count < max; count++)
1126                 newhead = &(*newhead)->next;
1127
1128         fiq->forget_list_head.next = *newhead;
1129         *newhead = NULL;
1130         if (fiq->forget_list_head.next == NULL)
1131                 fiq->forget_list_tail = &fiq->forget_list_head;
1132
1133         if (countp != NULL)
1134                 *countp = count;
1135
1136         return head;
1137 }
1138
1139 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1140                                    struct fuse_copy_state *cs,
1141                                    size_t nbytes)
1142 __releases(fiq->waitq.lock)
1143 {
1144         int err;
1145         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1146         struct fuse_forget_in arg = {
1147                 .nlookup = forget->forget_one.nlookup,
1148         };
1149         struct fuse_in_header ih = {
1150                 .opcode = FUSE_FORGET,
1151                 .nodeid = forget->forget_one.nodeid,
1152                 .unique = fuse_get_unique(fiq),
1153                 .len = sizeof(ih) + sizeof(arg),
1154         };
1155
1156         spin_unlock(&fiq->waitq.lock);
1157         kfree(forget);
1158         if (nbytes < ih.len)
1159                 return -EINVAL;
1160
1161         err = fuse_copy_one(cs, &ih, sizeof(ih));
1162         if (!err)
1163                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1164         fuse_copy_finish(cs);
1165
1166         if (err)
1167                 return err;
1168
1169         return ih.len;
1170 }
1171
1172 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1173                                    struct fuse_copy_state *cs, size_t nbytes)
1174 __releases(fiq->waitq.lock)
1175 {
1176         int err;
1177         unsigned max_forgets;
1178         unsigned count;
1179         struct fuse_forget_link *head;
1180         struct fuse_batch_forget_in arg = { .count = 0 };
1181         struct fuse_in_header ih = {
1182                 .opcode = FUSE_BATCH_FORGET,
1183                 .unique = fuse_get_unique(fiq),
1184                 .len = sizeof(ih) + sizeof(arg),
1185         };
1186
1187         if (nbytes < ih.len) {
1188                 spin_unlock(&fiq->waitq.lock);
1189                 return -EINVAL;
1190         }
1191
1192         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1193         head = dequeue_forget(fiq, max_forgets, &count);
1194         spin_unlock(&fiq->waitq.lock);
1195
1196         arg.count = count;
1197         ih.len += count * sizeof(struct fuse_forget_one);
1198         err = fuse_copy_one(cs, &ih, sizeof(ih));
1199         if (!err)
1200                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1201
1202         while (head) {
1203                 struct fuse_forget_link *forget = head;
1204
1205                 if (!err) {
1206                         err = fuse_copy_one(cs, &forget->forget_one,
1207                                             sizeof(forget->forget_one));
1208                 }
1209                 head = forget->next;
1210                 kfree(forget);
1211         }
1212
1213         fuse_copy_finish(cs);
1214
1215         if (err)
1216                 return err;
1217
1218         return ih.len;
1219 }
1220
1221 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1222                             struct fuse_copy_state *cs,
1223                             size_t nbytes)
1224 __releases(fiq->waitq.lock)
1225 {
1226         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1227                 return fuse_read_single_forget(fiq, cs, nbytes);
1228         else
1229                 return fuse_read_batch_forget(fiq, cs, nbytes);
1230 }
1231
1232 /*
1233  * Read a single request into the userspace filesystem's buffer.  This
1234  * function waits until a request is available, then removes it from
1235  * the pending list and copies request data to userspace buffer.  If
1236  * no reply is needed (FORGET) or request has been aborted or there
1237  * was an error during the copying then it's finished by calling
1238  * request_end().  Otherwise add it to the processing list, and set
1239  * the 'sent' flag.
1240  */
1241 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1242                                 struct fuse_copy_state *cs, size_t nbytes)
1243 {
1244         ssize_t err;
1245         struct fuse_conn *fc = fud->fc;
1246         struct fuse_iqueue *fiq = &fc->iq;
1247         struct fuse_pqueue *fpq = &fud->pq;
1248         struct fuse_req *req;
1249         struct fuse_in *in;
1250         unsigned reqsize;
1251
1252  restart:
1253         spin_lock(&fiq->waitq.lock);
1254         err = -EAGAIN;
1255         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1256             !request_pending(fiq))
1257                 goto err_unlock;
1258
1259         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1260                                 !fiq->connected || request_pending(fiq));
1261         if (err)
1262                 goto err_unlock;
1263
1264         if (!fiq->connected) {
1265                 err = (fc->aborted && fc->abort_err) ? -ECONNABORTED : -ENODEV;
1266                 goto err_unlock;
1267         }
1268
1269         if (!list_empty(&fiq->interrupts)) {
1270                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1271                                  intr_entry);
1272                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1273         }
1274
1275         if (forget_pending(fiq)) {
1276                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1277                         return fuse_read_forget(fc, fiq, cs, nbytes);
1278
1279                 if (fiq->forget_batch <= -8)
1280                         fiq->forget_batch = 16;
1281         }
1282
1283         req = list_entry(fiq->pending.next, struct fuse_req, list);
1284         clear_bit(FR_PENDING, &req->flags);
1285         list_del_init(&req->list);
1286         spin_unlock(&fiq->waitq.lock);
1287
1288         in = &req->in;
1289         reqsize = in->h.len;
1290
1291         /* If request is too large, reply with an error and restart the read */
1292         if (nbytes < reqsize) {
1293                 req->out.h.error = -EIO;
1294                 /* SETXATTR is special, since it may contain too large data */
1295                 if (in->h.opcode == FUSE_SETXATTR)
1296                         req->out.h.error = -E2BIG;
1297                 request_end(fc, req);
1298                 goto restart;
1299         }
1300         spin_lock(&fpq->lock);
1301         list_add(&req->list, &fpq->io);
1302         spin_unlock(&fpq->lock);
1303         cs->req = req;
1304         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1305         if (!err)
1306                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1307                                      (struct fuse_arg *) in->args, 0);
1308         fuse_copy_finish(cs);
1309         spin_lock(&fpq->lock);
1310         clear_bit(FR_LOCKED, &req->flags);
1311         if (!fpq->connected) {
1312                 err = (fc->aborted && fc->abort_err) ? -ECONNABORTED : -ENODEV;
1313                 goto out_end;
1314         }
1315         if (err) {
1316                 req->out.h.error = -EIO;
1317                 goto out_end;
1318         }
1319         if (!test_bit(FR_ISREPLY, &req->flags)) {
1320                 err = reqsize;
1321                 goto out_end;
1322         }
1323         list_move_tail(&req->list, &fpq->processing);
1324         __fuse_get_request(req);
1325         set_bit(FR_SENT, &req->flags);
1326         spin_unlock(&fpq->lock);
1327         /* matches barrier in request_wait_answer() */
1328         smp_mb__after_atomic();
1329         if (test_bit(FR_INTERRUPTED, &req->flags))
1330                 queue_interrupt(fiq, req);
1331         fuse_put_request(fc, req);
1332
1333         return reqsize;
1334
1335 out_end:
1336         if (!test_bit(FR_PRIVATE, &req->flags))
1337                 list_del_init(&req->list);
1338         spin_unlock(&fpq->lock);
1339         request_end(fc, req);
1340         return err;
1341
1342  err_unlock:
1343         spin_unlock(&fiq->waitq.lock);
1344         return err;
1345 }
1346
1347 static int fuse_dev_open(struct inode *inode, struct file *file)
1348 {
1349         /*
1350          * The fuse device's file's private_data is used to hold
1351          * the fuse_conn(ection) when it is mounted, and is used to
1352          * keep track of whether the file has been mounted already.
1353          */
1354         file->private_data = NULL;
1355         return 0;
1356 }
1357
1358 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1359 {
1360         struct fuse_copy_state cs;
1361         struct file *file = iocb->ki_filp;
1362         struct fuse_dev *fud = fuse_get_dev(file);
1363
1364         if (!fud)
1365                 return -EPERM;
1366
1367         if (!iter_is_iovec(to))
1368                 return -EINVAL;
1369
1370         fuse_copy_init(&cs, 1, to);
1371
1372         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1373 }
1374
1375 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1376                                     struct pipe_inode_info *pipe,
1377                                     size_t len, unsigned int flags)
1378 {
1379         int total, ret;
1380         int page_nr = 0;
1381         struct pipe_buffer *bufs;
1382         struct fuse_copy_state cs;
1383         struct fuse_dev *fud = fuse_get_dev(in);
1384
1385         if (!fud)
1386                 return -EPERM;
1387
1388         bufs = kvmalloc_array(pipe->buffers, sizeof(struct pipe_buffer),
1389                               GFP_KERNEL);
1390         if (!bufs)
1391                 return -ENOMEM;
1392
1393         fuse_copy_init(&cs, 1, NULL);
1394         cs.pipebufs = bufs;
1395         cs.pipe = pipe;
1396         ret = fuse_dev_do_read(fud, in, &cs, len);
1397         if (ret < 0)
1398                 goto out;
1399
1400         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1401                 ret = -EIO;
1402                 goto out;
1403         }
1404
1405         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1406                 /*
1407                  * Need to be careful about this.  Having buf->ops in module
1408                  * code can Oops if the buffer persists after module unload.
1409                  */
1410                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1411                 bufs[page_nr].flags = 0;
1412                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1413                 if (unlikely(ret < 0))
1414                         break;
1415         }
1416         if (total)
1417                 ret = total;
1418 out:
1419         for (; page_nr < cs.nr_segs; page_nr++)
1420                 put_page(bufs[page_nr].page);
1421
1422         kvfree(bufs);
1423         return ret;
1424 }
1425
1426 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1427                             struct fuse_copy_state *cs)
1428 {
1429         struct fuse_notify_poll_wakeup_out outarg;
1430         int err = -EINVAL;
1431
1432         if (size != sizeof(outarg))
1433                 goto err;
1434
1435         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1436         if (err)
1437                 goto err;
1438
1439         fuse_copy_finish(cs);
1440         return fuse_notify_poll_wakeup(fc, &outarg);
1441
1442 err:
1443         fuse_copy_finish(cs);
1444         return err;
1445 }
1446
1447 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1448                                    struct fuse_copy_state *cs)
1449 {
1450         struct fuse_notify_inval_inode_out outarg;
1451         int err = -EINVAL;
1452
1453         if (size != sizeof(outarg))
1454                 goto err;
1455
1456         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1457         if (err)
1458                 goto err;
1459         fuse_copy_finish(cs);
1460
1461         down_read(&fc->killsb);
1462         err = -ENOENT;
1463         if (fc->sb) {
1464                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1465                                                outarg.off, outarg.len);
1466         }
1467         up_read(&fc->killsb);
1468         return err;
1469
1470 err:
1471         fuse_copy_finish(cs);
1472         return err;
1473 }
1474
1475 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1476                                    struct fuse_copy_state *cs)
1477 {
1478         struct fuse_notify_inval_entry_out outarg;
1479         int err = -ENOMEM;
1480         char *buf;
1481         struct qstr name;
1482
1483         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1484         if (!buf)
1485                 goto err;
1486
1487         err = -EINVAL;
1488         if (size < sizeof(outarg))
1489                 goto err;
1490
1491         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1492         if (err)
1493                 goto err;
1494
1495         err = -ENAMETOOLONG;
1496         if (outarg.namelen > FUSE_NAME_MAX)
1497                 goto err;
1498
1499         err = -EINVAL;
1500         if (size != sizeof(outarg) + outarg.namelen + 1)
1501                 goto err;
1502
1503         name.name = buf;
1504         name.len = outarg.namelen;
1505         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1506         if (err)
1507                 goto err;
1508         fuse_copy_finish(cs);
1509         buf[outarg.namelen] = 0;
1510
1511         down_read(&fc->killsb);
1512         err = -ENOENT;
1513         if (fc->sb)
1514                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1515         up_read(&fc->killsb);
1516         kfree(buf);
1517         return err;
1518
1519 err:
1520         kfree(buf);
1521         fuse_copy_finish(cs);
1522         return err;
1523 }
1524
1525 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1526                               struct fuse_copy_state *cs)
1527 {
1528         struct fuse_notify_delete_out outarg;
1529         int err = -ENOMEM;
1530         char *buf;
1531         struct qstr name;
1532
1533         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1534         if (!buf)
1535                 goto err;
1536
1537         err = -EINVAL;
1538         if (size < sizeof(outarg))
1539                 goto err;
1540
1541         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1542         if (err)
1543                 goto err;
1544
1545         err = -ENAMETOOLONG;
1546         if (outarg.namelen > FUSE_NAME_MAX)
1547                 goto err;
1548
1549         err = -EINVAL;
1550         if (size != sizeof(outarg) + outarg.namelen + 1)
1551                 goto err;
1552
1553         name.name = buf;
1554         name.len = outarg.namelen;
1555         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1556         if (err)
1557                 goto err;
1558         fuse_copy_finish(cs);
1559         buf[outarg.namelen] = 0;
1560
1561         down_read(&fc->killsb);
1562         err = -ENOENT;
1563         if (fc->sb)
1564                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1565                                                outarg.child, &name);
1566         up_read(&fc->killsb);
1567         kfree(buf);
1568         return err;
1569
1570 err:
1571         kfree(buf);
1572         fuse_copy_finish(cs);
1573         return err;
1574 }
1575
1576 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1577                              struct fuse_copy_state *cs)
1578 {
1579         struct fuse_notify_store_out outarg;
1580         struct inode *inode;
1581         struct address_space *mapping;
1582         u64 nodeid;
1583         int err;
1584         pgoff_t index;
1585         unsigned int offset;
1586         unsigned int num;
1587         loff_t file_size;
1588         loff_t end;
1589
1590         err = -EINVAL;
1591         if (size < sizeof(outarg))
1592                 goto out_finish;
1593
1594         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1595         if (err)
1596                 goto out_finish;
1597
1598         err = -EINVAL;
1599         if (size - sizeof(outarg) != outarg.size)
1600                 goto out_finish;
1601
1602         nodeid = outarg.nodeid;
1603
1604         down_read(&fc->killsb);
1605
1606         err = -ENOENT;
1607         if (!fc->sb)
1608                 goto out_up_killsb;
1609
1610         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1611         if (!inode)
1612                 goto out_up_killsb;
1613
1614         mapping = inode->i_mapping;
1615         index = outarg.offset >> PAGE_SHIFT;
1616         offset = outarg.offset & ~PAGE_MASK;
1617         file_size = i_size_read(inode);
1618         end = outarg.offset + outarg.size;
1619         if (end > file_size) {
1620                 file_size = end;
1621                 fuse_write_update_size(inode, file_size);
1622         }
1623
1624         num = outarg.size;
1625         while (num) {
1626                 struct page *page;
1627                 unsigned int this_num;
1628
1629                 err = -ENOMEM;
1630                 page = find_or_create_page(mapping, index,
1631                                            mapping_gfp_mask(mapping));
1632                 if (!page)
1633                         goto out_iput;
1634
1635                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1636                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1637                 if (!err && offset == 0 &&
1638                     (this_num == PAGE_SIZE || file_size == end))
1639                         SetPageUptodate(page);
1640                 unlock_page(page);
1641                 put_page(page);
1642
1643                 if (err)
1644                         goto out_iput;
1645
1646                 num -= this_num;
1647                 offset = 0;
1648                 index++;
1649         }
1650
1651         err = 0;
1652
1653 out_iput:
1654         iput(inode);
1655 out_up_killsb:
1656         up_read(&fc->killsb);
1657 out_finish:
1658         fuse_copy_finish(cs);
1659         return err;
1660 }
1661
1662 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1663 {
1664         release_pages(req->pages, req->num_pages);
1665 }
1666
1667 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1668                          struct fuse_notify_retrieve_out *outarg)
1669 {
1670         int err;
1671         struct address_space *mapping = inode->i_mapping;
1672         struct fuse_req *req;
1673         pgoff_t index;
1674         loff_t file_size;
1675         unsigned int num;
1676         unsigned int offset;
1677         size_t total_len = 0;
1678         int num_pages;
1679
1680         offset = outarg->offset & ~PAGE_MASK;
1681         file_size = i_size_read(inode);
1682
1683         num = outarg->size;
1684         if (outarg->offset > file_size)
1685                 num = 0;
1686         else if (outarg->offset + num > file_size)
1687                 num = file_size - outarg->offset;
1688
1689         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1690         num_pages = min(num_pages, FUSE_MAX_PAGES_PER_REQ);
1691
1692         req = fuse_get_req(fc, num_pages);
1693         if (IS_ERR(req))
1694                 return PTR_ERR(req);
1695
1696         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1697         req->in.h.nodeid = outarg->nodeid;
1698         req->in.numargs = 2;
1699         req->in.argpages = 1;
1700         req->page_descs[0].offset = offset;
1701         req->end = fuse_retrieve_end;
1702
1703         index = outarg->offset >> PAGE_SHIFT;
1704
1705         while (num && req->num_pages < num_pages) {
1706                 struct page *page;
1707                 unsigned int this_num;
1708
1709                 page = find_get_page(mapping, index);
1710                 if (!page)
1711                         break;
1712
1713                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1714                 req->pages[req->num_pages] = page;
1715                 req->page_descs[req->num_pages].length = this_num;
1716                 req->num_pages++;
1717
1718                 offset = 0;
1719                 num -= this_num;
1720                 total_len += this_num;
1721                 index++;
1722         }
1723         req->misc.retrieve_in.offset = outarg->offset;
1724         req->misc.retrieve_in.size = total_len;
1725         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1726         req->in.args[0].value = &req->misc.retrieve_in;
1727         req->in.args[1].size = total_len;
1728
1729         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1730         if (err)
1731                 fuse_retrieve_end(fc, req);
1732
1733         return err;
1734 }
1735
1736 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1737                                 struct fuse_copy_state *cs)
1738 {
1739         struct fuse_notify_retrieve_out outarg;
1740         struct inode *inode;
1741         int err;
1742
1743         err = -EINVAL;
1744         if (size != sizeof(outarg))
1745                 goto copy_finish;
1746
1747         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1748         if (err)
1749                 goto copy_finish;
1750
1751         fuse_copy_finish(cs);
1752
1753         down_read(&fc->killsb);
1754         err = -ENOENT;
1755         if (fc->sb) {
1756                 u64 nodeid = outarg.nodeid;
1757
1758                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1759                 if (inode) {
1760                         err = fuse_retrieve(fc, inode, &outarg);
1761                         iput(inode);
1762                 }
1763         }
1764         up_read(&fc->killsb);
1765
1766         return err;
1767
1768 copy_finish:
1769         fuse_copy_finish(cs);
1770         return err;
1771 }
1772
1773 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1774                        unsigned int size, struct fuse_copy_state *cs)
1775 {
1776         /* Don't try to move pages (yet) */
1777         cs->move_pages = 0;
1778
1779         switch (code) {
1780         case FUSE_NOTIFY_POLL:
1781                 return fuse_notify_poll(fc, size, cs);
1782
1783         case FUSE_NOTIFY_INVAL_INODE:
1784                 return fuse_notify_inval_inode(fc, size, cs);
1785
1786         case FUSE_NOTIFY_INVAL_ENTRY:
1787                 return fuse_notify_inval_entry(fc, size, cs);
1788
1789         case FUSE_NOTIFY_STORE:
1790                 return fuse_notify_store(fc, size, cs);
1791
1792         case FUSE_NOTIFY_RETRIEVE:
1793                 return fuse_notify_retrieve(fc, size, cs);
1794
1795         case FUSE_NOTIFY_DELETE:
1796                 return fuse_notify_delete(fc, size, cs);
1797
1798         default:
1799                 fuse_copy_finish(cs);
1800                 return -EINVAL;
1801         }
1802 }
1803
1804 /* Look up request on processing list by unique ID */
1805 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1806 {
1807         struct fuse_req *req;
1808
1809         list_for_each_entry(req, &fpq->processing, list) {
1810                 if (req->in.h.unique == unique)
1811                         return req;
1812         }
1813         return NULL;
1814 }
1815
1816 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1817                          unsigned nbytes)
1818 {
1819         unsigned reqsize = sizeof(struct fuse_out_header);
1820
1821         if (out->h.error)
1822                 return nbytes != reqsize ? -EINVAL : 0;
1823
1824         reqsize += len_args(out->numargs, out->args);
1825
1826         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1827                 return -EINVAL;
1828         else if (reqsize > nbytes) {
1829                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1830                 unsigned diffsize = reqsize - nbytes;
1831                 if (diffsize > lastarg->size)
1832                         return -EINVAL;
1833                 lastarg->size -= diffsize;
1834         }
1835         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1836                               out->page_zeroing);
1837 }
1838
1839 /*
1840  * Write a single reply to a request.  First the header is copied from
1841  * the write buffer.  The request is then searched on the processing
1842  * list by the unique ID found in the header.  If found, then remove
1843  * it from the list and copy the rest of the buffer to the request.
1844  * The request is finished by calling request_end()
1845  */
1846 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1847                                  struct fuse_copy_state *cs, size_t nbytes)
1848 {
1849         int err;
1850         struct fuse_conn *fc = fud->fc;
1851         struct fuse_pqueue *fpq = &fud->pq;
1852         struct fuse_req *req;
1853         struct fuse_out_header oh;
1854
1855         if (nbytes < sizeof(struct fuse_out_header))
1856                 return -EINVAL;
1857
1858         err = fuse_copy_one(cs, &oh, sizeof(oh));
1859         if (err)
1860                 goto err_finish;
1861
1862         err = -EINVAL;
1863         if (oh.len != nbytes)
1864                 goto err_finish;
1865
1866         /*
1867          * Zero oh.unique indicates unsolicited notification message
1868          * and error contains notification code.
1869          */
1870         if (!oh.unique) {
1871                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1872                 return err ? err : nbytes;
1873         }
1874
1875         err = -EINVAL;
1876         if (oh.error <= -1000 || oh.error > 0)
1877                 goto err_finish;
1878
1879         spin_lock(&fpq->lock);
1880         err = -ENOENT;
1881         if (!fpq->connected)
1882                 goto err_unlock_pq;
1883
1884         req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1885         if (!req)
1886                 goto err_unlock_pq;
1887
1888         /* Is it an interrupt reply ID? */
1889         if (oh.unique & FUSE_INT_REQ_BIT) {
1890                 __fuse_get_request(req);
1891                 spin_unlock(&fpq->lock);
1892
1893                 err = -EINVAL;
1894                 if (nbytes != sizeof(struct fuse_out_header)) {
1895                         fuse_put_request(fc, req);
1896                         goto err_finish;
1897                 }
1898
1899                 if (oh.error == -ENOSYS)
1900                         fc->no_interrupt = 1;
1901                 else if (oh.error == -EAGAIN)
1902                         queue_interrupt(&fc->iq, req);
1903                 fuse_put_request(fc, req);
1904
1905                 fuse_copy_finish(cs);
1906                 return nbytes;
1907         }
1908
1909         clear_bit(FR_SENT, &req->flags);
1910         list_move(&req->list, &fpq->io);
1911         req->out.h = oh;
1912         set_bit(FR_LOCKED, &req->flags);
1913         spin_unlock(&fpq->lock);
1914         cs->req = req;
1915         if (!req->out.page_replace)
1916                 cs->move_pages = 0;
1917
1918         err = copy_out_args(cs, &req->out, nbytes);
1919         fuse_copy_finish(cs);
1920
1921         spin_lock(&fpq->lock);
1922         clear_bit(FR_LOCKED, &req->flags);
1923         if (!fpq->connected)
1924                 err = -ENOENT;
1925         else if (err)
1926                 req->out.h.error = -EIO;
1927         if (!test_bit(FR_PRIVATE, &req->flags))
1928                 list_del_init(&req->list);
1929         spin_unlock(&fpq->lock);
1930
1931         request_end(fc, req);
1932
1933         return err ? err : nbytes;
1934
1935  err_unlock_pq:
1936         spin_unlock(&fpq->lock);
1937  err_finish:
1938         fuse_copy_finish(cs);
1939         return err;
1940 }
1941
1942 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
1943 {
1944         struct fuse_copy_state cs;
1945         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
1946
1947         if (!fud)
1948                 return -EPERM;
1949
1950         if (!iter_is_iovec(from))
1951                 return -EINVAL;
1952
1953         fuse_copy_init(&cs, 0, from);
1954
1955         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
1956 }
1957
1958 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1959                                      struct file *out, loff_t *ppos,
1960                                      size_t len, unsigned int flags)
1961 {
1962         unsigned nbuf;
1963         unsigned idx;
1964         struct pipe_buffer *bufs;
1965         struct fuse_copy_state cs;
1966         struct fuse_dev *fud;
1967         size_t rem;
1968         ssize_t ret;
1969
1970         fud = fuse_get_dev(out);
1971         if (!fud)
1972                 return -EPERM;
1973
1974         pipe_lock(pipe);
1975
1976         bufs = kvmalloc_array(pipe->nrbufs, sizeof(struct pipe_buffer),
1977                               GFP_KERNEL);
1978         if (!bufs) {
1979                 pipe_unlock(pipe);
1980                 return -ENOMEM;
1981         }
1982
1983         nbuf = 0;
1984         rem = 0;
1985         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1986                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1987
1988         ret = -EINVAL;
1989         if (rem < len) {
1990                 pipe_unlock(pipe);
1991                 goto out;
1992         }
1993
1994         rem = len;
1995         while (rem) {
1996                 struct pipe_buffer *ibuf;
1997                 struct pipe_buffer *obuf;
1998
1999                 BUG_ON(nbuf >= pipe->buffers);
2000                 BUG_ON(!pipe->nrbufs);
2001                 ibuf = &pipe->bufs[pipe->curbuf];
2002                 obuf = &bufs[nbuf];
2003
2004                 if (rem >= ibuf->len) {
2005                         *obuf = *ibuf;
2006                         ibuf->ops = NULL;
2007                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2008                         pipe->nrbufs--;
2009                 } else {
2010                         pipe_buf_get(pipe, ibuf);
2011                         *obuf = *ibuf;
2012                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2013                         obuf->len = rem;
2014                         ibuf->offset += obuf->len;
2015                         ibuf->len -= obuf->len;
2016                 }
2017                 nbuf++;
2018                 rem -= obuf->len;
2019         }
2020         pipe_unlock(pipe);
2021
2022         fuse_copy_init(&cs, 0, NULL);
2023         cs.pipebufs = bufs;
2024         cs.nr_segs = nbuf;
2025         cs.pipe = pipe;
2026
2027         if (flags & SPLICE_F_MOVE)
2028                 cs.move_pages = 1;
2029
2030         ret = fuse_dev_do_write(fud, &cs, len);
2031
2032         for (idx = 0; idx < nbuf; idx++)
2033                 pipe_buf_release(pipe, &bufs[idx]);
2034
2035 out:
2036         kvfree(bufs);
2037         return ret;
2038 }
2039
2040 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2041 {
2042         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2043         struct fuse_iqueue *fiq;
2044         struct fuse_dev *fud = fuse_get_dev(file);
2045
2046         if (!fud)
2047                 return EPOLLERR;
2048
2049         fiq = &fud->fc->iq;
2050         poll_wait(file, &fiq->waitq, wait);
2051
2052         spin_lock(&fiq->waitq.lock);
2053         if (!fiq->connected)
2054                 mask = EPOLLERR;
2055         else if (request_pending(fiq))
2056                 mask |= EPOLLIN | EPOLLRDNORM;
2057         spin_unlock(&fiq->waitq.lock);
2058
2059         return mask;
2060 }
2061
2062 /*
2063  * Abort all requests on the given list (pending or processing)
2064  *
2065  * This function releases and reacquires fc->lock
2066  */
2067 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2068 {
2069         while (!list_empty(head)) {
2070                 struct fuse_req *req;
2071                 req = list_entry(head->next, struct fuse_req, list);
2072                 req->out.h.error = -ECONNABORTED;
2073                 clear_bit(FR_SENT, &req->flags);
2074                 list_del_init(&req->list);
2075                 request_end(fc, req);
2076         }
2077 }
2078
2079 static void end_polls(struct fuse_conn *fc)
2080 {
2081         struct rb_node *p;
2082
2083         p = rb_first(&fc->polled_files);
2084
2085         while (p) {
2086                 struct fuse_file *ff;
2087                 ff = rb_entry(p, struct fuse_file, polled_node);
2088                 wake_up_interruptible_all(&ff->poll_wait);
2089
2090                 p = rb_next(p);
2091         }
2092 }
2093
2094 /*
2095  * Abort all requests.
2096  *
2097  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2098  * filesystem.
2099  *
2100  * The same effect is usually achievable through killing the filesystem daemon
2101  * and all users of the filesystem.  The exception is the combination of an
2102  * asynchronous request and the tricky deadlock (see
2103  * Documentation/filesystems/fuse.txt).
2104  *
2105  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2106  * requests, they should be finished off immediately.  Locked requests will be
2107  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2108  * requests.  It is possible that some request will finish before we can.  This
2109  * is OK, the request will in that case be removed from the list before we touch
2110  * it.
2111  */
2112 void fuse_abort_conn(struct fuse_conn *fc, bool is_abort)
2113 {
2114         struct fuse_iqueue *fiq = &fc->iq;
2115
2116         spin_lock(&fc->lock);
2117         if (fc->connected) {
2118                 struct fuse_dev *fud;
2119                 struct fuse_req *req, *next;
2120                 LIST_HEAD(to_end);
2121
2122                 /* Background queuing checks fc->connected under bg_lock */
2123                 spin_lock(&fc->bg_lock);
2124                 fc->connected = 0;
2125                 spin_unlock(&fc->bg_lock);
2126
2127                 fc->aborted = is_abort;
2128                 fuse_set_initialized(fc);
2129                 list_for_each_entry(fud, &fc->devices, entry) {
2130                         struct fuse_pqueue *fpq = &fud->pq;
2131
2132                         spin_lock(&fpq->lock);
2133                         fpq->connected = 0;
2134                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2135                                 req->out.h.error = -ECONNABORTED;
2136                                 spin_lock(&req->waitq.lock);
2137                                 set_bit(FR_ABORTED, &req->flags);
2138                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2139                                         set_bit(FR_PRIVATE, &req->flags);
2140                                         __fuse_get_request(req);
2141                                         list_move(&req->list, &to_end);
2142                                 }
2143                                 spin_unlock(&req->waitq.lock);
2144                         }
2145                         list_splice_tail_init(&fpq->processing, &to_end);
2146                         spin_unlock(&fpq->lock);
2147                 }
2148                 spin_lock(&fc->bg_lock);
2149                 fc->blocked = 0;
2150                 fc->max_background = UINT_MAX;
2151                 flush_bg_queue(fc);
2152                 spin_unlock(&fc->bg_lock);
2153
2154                 spin_lock(&fiq->waitq.lock);
2155                 fiq->connected = 0;
2156                 list_for_each_entry(req, &fiq->pending, list)
2157                         clear_bit(FR_PENDING, &req->flags);
2158                 list_splice_tail_init(&fiq->pending, &to_end);
2159                 while (forget_pending(fiq))
2160                         kfree(dequeue_forget(fiq, 1, NULL));
2161                 wake_up_all_locked(&fiq->waitq);
2162                 spin_unlock(&fiq->waitq.lock);
2163                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2164                 end_polls(fc);
2165                 wake_up_all(&fc->blocked_waitq);
2166                 spin_unlock(&fc->lock);
2167
2168                 end_requests(fc, &to_end);
2169         } else {
2170                 spin_unlock(&fc->lock);
2171         }
2172 }
2173 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2174
2175 void fuse_wait_aborted(struct fuse_conn *fc)
2176 {
2177         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2178 }
2179
2180 int fuse_dev_release(struct inode *inode, struct file *file)
2181 {
2182         struct fuse_dev *fud = fuse_get_dev(file);
2183
2184         if (fud) {
2185                 struct fuse_conn *fc = fud->fc;
2186                 struct fuse_pqueue *fpq = &fud->pq;
2187                 LIST_HEAD(to_end);
2188
2189                 spin_lock(&fpq->lock);
2190                 WARN_ON(!list_empty(&fpq->io));
2191                 list_splice_init(&fpq->processing, &to_end);
2192                 spin_unlock(&fpq->lock);
2193
2194                 end_requests(fc, &to_end);
2195
2196                 /* Are we the last open device? */
2197                 if (atomic_dec_and_test(&fc->dev_count)) {
2198                         WARN_ON(fc->iq.fasync != NULL);
2199                         fuse_abort_conn(fc, false);
2200                 }
2201                 fuse_dev_free(fud);
2202         }
2203         return 0;
2204 }
2205 EXPORT_SYMBOL_GPL(fuse_dev_release);
2206
2207 static int fuse_dev_fasync(int fd, struct file *file, int on)
2208 {
2209         struct fuse_dev *fud = fuse_get_dev(file);
2210
2211         if (!fud)
2212                 return -EPERM;
2213
2214         /* No locking - fasync_helper does its own locking */
2215         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2216 }
2217
2218 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2219 {
2220         struct fuse_dev *fud;
2221
2222         if (new->private_data)
2223                 return -EINVAL;
2224
2225         fud = fuse_dev_alloc(fc);
2226         if (!fud)
2227                 return -ENOMEM;
2228
2229         new->private_data = fud;
2230         atomic_inc(&fc->dev_count);
2231
2232         return 0;
2233 }
2234
2235 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2236                            unsigned long arg)
2237 {
2238         int err = -ENOTTY;
2239
2240         if (cmd == FUSE_DEV_IOC_CLONE) {
2241                 int oldfd;
2242
2243                 err = -EFAULT;
2244                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2245                         struct file *old = fget(oldfd);
2246
2247                         err = -EINVAL;
2248                         if (old) {
2249                                 struct fuse_dev *fud = NULL;
2250
2251                                 /*
2252                                  * Check against file->f_op because CUSE
2253                                  * uses the same ioctl handler.
2254                                  */
2255                                 if (old->f_op == file->f_op &&
2256                                     old->f_cred->user_ns == file->f_cred->user_ns)
2257                                         fud = fuse_get_dev(old);
2258
2259                                 if (fud) {
2260                                         mutex_lock(&fuse_mutex);
2261                                         err = fuse_device_clone(fud->fc, file);
2262                                         mutex_unlock(&fuse_mutex);
2263                                 }
2264                                 fput(old);
2265                         }
2266                 }
2267         }
2268         return err;
2269 }
2270
2271 const struct file_operations fuse_dev_operations = {
2272         .owner          = THIS_MODULE,
2273         .open           = fuse_dev_open,
2274         .llseek         = no_llseek,
2275         .read_iter      = fuse_dev_read,
2276         .splice_read    = fuse_dev_splice_read,
2277         .write_iter     = fuse_dev_write,
2278         .splice_write   = fuse_dev_splice_write,
2279         .poll           = fuse_dev_poll,
2280         .release        = fuse_dev_release,
2281         .fasync         = fuse_dev_fasync,
2282         .unlocked_ioctl = fuse_dev_ioctl,
2283         .compat_ioctl   = fuse_dev_ioctl,
2284 };
2285 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2286
2287 static struct miscdevice fuse_miscdevice = {
2288         .minor = FUSE_MINOR,
2289         .name  = "fuse",
2290         .fops = &fuse_dev_operations,
2291 };
2292
2293 int __init fuse_dev_init(void)
2294 {
2295         int err = -ENOMEM;
2296         fuse_req_cachep = kmem_cache_create("fuse_request",
2297                                             sizeof(struct fuse_req),
2298                                             0, 0, NULL);
2299         if (!fuse_req_cachep)
2300                 goto out;
2301
2302         err = misc_register(&fuse_miscdevice);
2303         if (err)
2304                 goto out_cache_clean;
2305
2306         return 0;
2307
2308  out_cache_clean:
2309         kmem_cache_destroy(fuse_req_cachep);
2310  out:
2311         return err;
2312 }
2313
2314 void fuse_dev_cleanup(void)
2315 {
2316         misc_deregister(&fuse_miscdevice);
2317         kmem_cache_destroy(fuse_req_cachep);
2318 }