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/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/pipe_fs_i.h>
  20 #include <linux/swap.h>
  21 #include <linux/splice.h>
  22
  23 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  24 MODULE_ALIAS("devname:fuse");
  25
  26 static struct kmem_cache *fuse_req_cachep;
  27
  28 static struct fuse_conn *fuse_get_conn(struct file *file)
  29 {
  30         /*
  31          * Lockless access is OK, because file->private data is set
  32          * once during mount and is valid until the file is released.
  33          */
  34         return file->private_data;
  35 }
  36
  37 static void fuse_request_init(struct fuse_req *req)
  38 {
  39         memset(req, 0, sizeof(*req));
  40         INIT_LIST_HEAD(&req->list);
  41         INIT_LIST_HEAD(&req->intr_entry);
  42         init_waitqueue_head(&req->waitq);
  43         atomic_set(&req->count, 1);
  44 }
  45
  46 struct fuse_req *fuse_request_alloc(void)
  47 {
  48         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_KERNEL);
  49         if (req)
  50                 fuse_request_init(req);
  51         return req;
  52 }
  53 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  54
  55 struct fuse_req *fuse_request_alloc_nofs(void)
  56 {
  57         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_NOFS);
  58         if (req)
  59                 fuse_request_init(req);
  60         return req;
  61 }
  62
  63 void fuse_request_free(struct fuse_req *req)
  64 {
  65         kmem_cache_free(fuse_req_cachep, req);
  66 }
  67
  68 static void block_sigs(sigset_t *oldset)
  69 {
  70         sigset_t mask;
  71
  72         siginitsetinv(&mask, sigmask(SIGKILL));
  73         sigprocmask(SIG_BLOCK, &mask, oldset);
  74 }
  75
  76 static void restore_sigs(sigset_t *oldset)
  77 {
  78         sigprocmask(SIG_SETMASK, oldset, NULL);
  79 }
  80
  81 static void __fuse_get_request(struct fuse_req *req)
  82 {
  83         atomic_inc(&req->count);
  84 }
  85
  86 /* Must be called with > 1 refcount */
  87 static void __fuse_put_request(struct fuse_req *req)
  88 {
  89         BUG_ON(atomic_read(&req->count) < 2);
  90         atomic_dec(&req->count);
  91 }
  92
  93 static void fuse_req_init_context(struct fuse_req *req)
  94 {
  95         req->in.h.uid = current_fsuid();
  96         req->in.h.gid = current_fsgid();
  97         req->in.h.pid = current->pid;
  98 }
  99
 100 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
 101 {
 102         struct fuse_req *req;
 103         sigset_t oldset;
 104         int intr;
 105         int err;
 106
 107         atomic_inc(&fc->num_waiting);
 108         block_sigs(&oldset);
 109         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
 110         restore_sigs(&oldset);
 111         err = -EINTR;
 112         if (intr)
 113                 goto out;
 114
 115         err = -ENOTCONN;
 116         if (!fc->connected)
 117                 goto out;
 118
 119         req = fuse_request_alloc();
 120         err = -ENOMEM;
 121         if (!req)
 122                 goto out;
 123
 124         fuse_req_init_context(req);
 125         req->waiting = 1;
 126         return req;
 127
 128  out:
 129         atomic_dec(&fc->num_waiting);
 130         return ERR_PTR(err);
 131 }
 132 EXPORT_SYMBOL_GPL(fuse_get_req);
 133
 134 /*
 135  * Return request in fuse_file->reserved_req.  However that may
 136  * currently be in use.  If that is the case, wait for it to become
 137  * available.
 138  */
 139 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 140                                          struct file *file)
 141 {
 142         struct fuse_req *req = NULL;
 143         struct fuse_file *ff = file->private_data;
 144
 145         do {
 146                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 147                 spin_lock(&fc->lock);
 148                 if (ff->reserved_req) {
 149                         req = ff->reserved_req;
 150                         ff->reserved_req = NULL;
 151                         req->stolen_file = get_file(file);
 152                 }
 153                 spin_unlock(&fc->lock);
 154         } while (!req);
 155
 156         return req;
 157 }
 158
 159 /*
 160  * Put stolen request back into fuse_file->reserved_req
 161  */
 162 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 163 {
 164         struct file *file = req->stolen_file;
 165         struct fuse_file *ff = file->private_data;
 166
 167         spin_lock(&fc->lock);
 168         fuse_request_init(req);
 169         BUG_ON(ff->reserved_req);
 170         ff->reserved_req = req;
 171         wake_up_all(&fc->reserved_req_waitq);
 172         spin_unlock(&fc->lock);
 173         fput(file);
 174 }
 175
 176 /*
 177  * Gets a requests for a file operation, always succeeds
 178  *
 179  * This is used for sending the FLUSH request, which must get to
 180  * userspace, due to POSIX locks which may need to be unlocked.
 181  *
 182  * If allocation fails due to OOM, use the reserved request in
 183  * fuse_file.
 184  *
 185  * This is very unlikely to deadlock accidentally, since the
 186  * filesystem should not have it's own file open.  If deadlock is
 187  * intentional, it can still be broken by "aborting" the filesystem.
 188  */
 189 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 190 {
 191         struct fuse_req *req;
 192
 193         atomic_inc(&fc->num_waiting);
 194         wait_event(fc->blocked_waitq, !fc->blocked);
 195         req = fuse_request_alloc();
 196         if (!req)
 197                 req = get_reserved_req(fc, file);
 198
 199         fuse_req_init_context(req);
 200         req->waiting = 1;
 201         return req;
 202 }
 203
 204 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 205 {
 206         if (atomic_dec_and_test(&req->count)) {
 207                 if (req->waiting)
 208                         atomic_dec(&fc->num_waiting);
 209
 210                 if (req->stolen_file)
 211                         put_reserved_req(fc, req);
 212                 else
 213                         fuse_request_free(req);
 214         }
 215 }
 216 EXPORT_SYMBOL_GPL(fuse_put_request);
 217
 218 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 219 {
 220         unsigned nbytes = 0;
 221         unsigned i;
 222
 223         for (i = 0; i < numargs; i++)
 224                 nbytes += args[i].size;
 225
 226         return nbytes;
 227 }
 228
 229 static u64 fuse_get_unique(struct fuse_conn *fc)
 230 {
 231         fc->reqctr++;
 232         /* zero is special */
 233         if (fc->reqctr == 0)
 234                 fc->reqctr = 1;
 235
 236         return fc->reqctr;
 237 }
 238
 239 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 240 {
 241         req->in.h.len = sizeof(struct fuse_in_header) +
 242                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 243         list_add_tail(&req->list, &fc->pending);
 244         req->state = FUSE_REQ_PENDING;
 245         if (!req->waiting) {
 246                 req->waiting = 1;
 247                 atomic_inc(&fc->num_waiting);
 248         }
 249         wake_up(&fc->waitq);
 250         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 251 }
 252
 253 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 254                        u64 nodeid, u64 nlookup)
 255 {
 256         forget->forget_one.nodeid = nodeid;
 257         forget->forget_one.nlookup = nlookup;
 258
 259         spin_lock(&fc->lock);
 260         if (fc->connected) {
 261                 fc->forget_list_tail->next = forget;
 262                 fc->forget_list_tail = forget;
 263                 wake_up(&fc->waitq);
 264                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 265         } else {
 266                 kfree(forget);
 267         }
 268         spin_unlock(&fc->lock);
 269 }
 270
 271 static void flush_bg_queue(struct fuse_conn *fc)
 272 {
 273         while (fc->active_background < fc->max_background &&
 274                !list_empty(&fc->bg_queue)) {
 275                 struct fuse_req *req;
 276
 277                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 278                 list_del(&req->list);
 279                 fc->active_background++;
 280                 req->in.h.unique = fuse_get_unique(fc);
 281                 queue_request(fc, req);
 282         }
 283 }
 284
 285 /*
 286  * This function is called when a request is finished.  Either a reply
 287  * has arrived or it was aborted (and not yet sent) or some error
 288  * occurred during communication with userspace, or the device file
 289  * was closed.  The requester thread is woken up (if still waiting),
 290  * the 'end' callback is called if given, else the reference to the
 291  * request is released
 292  *
 293  * Called with fc->lock, unlocks it
 294  */
 295 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 296 __releases(fc->lock)
 297 {
 298         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 299         req->end = NULL;
 300         list_del(&req->list);
 301         list_del(&req->intr_entry);
 302         req->state = FUSE_REQ_FINISHED;
 303         if (req->background) {
 304                 if (fc->num_background == fc->max_background) {
 305                         fc->blocked = 0;
 306                         wake_up_all(&fc->blocked_waitq);
 307                 }
 308                 if (fc->num_background == fc->congestion_threshold &&
 309                     fc->connected && fc->bdi_initialized) {
 310                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 311                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 312                 }
 313                 fc->num_background--;
 314                 fc->active_background--;
 315                 flush_bg_queue(fc);
 316         }
 317         spin_unlock(&fc->lock);
 318         wake_up(&req->waitq);
 319         if (end)
 320                 end(fc, req);
 321         fuse_put_request(fc, req);
 322 }
 323
 324 static void wait_answer_interruptible(struct fuse_conn *fc,
 325                                       struct fuse_req *req)
 326 __releases(fc->lock)
 327 __acquires(fc->lock)
 328 {
 329         if (signal_pending(current))
 330                 return;
 331
 332         spin_unlock(&fc->lock);
 333         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 334         spin_lock(&fc->lock);
 335 }
 336
 337 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 338 {
 339         list_add_tail(&req->intr_entry, &fc->interrupts);
 340         wake_up(&fc->waitq);
 341         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 342 }
 343
 344 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 345 __releases(fc->lock)
 346 __acquires(fc->lock)
 347 {
 348         if (!fc->no_interrupt) {
 349                 /* Any signal may interrupt this */
 350                 wait_answer_interruptible(fc, req);
 351
 352                 if (req->aborted)
 353                         goto aborted;
 354                 if (req->state == FUSE_REQ_FINISHED)
 355                         return;
 356
 357                 req->interrupted = 1;
 358                 if (req->state == FUSE_REQ_SENT)
 359                         queue_interrupt(fc, req);
 360         }
 361
 362         if (!req->force) {
 363                 sigset_t oldset;
 364
 365                 /* Only fatal signals may interrupt this */
 366                 block_sigs(&oldset);
 367                 wait_answer_interruptible(fc, req);
 368                 restore_sigs(&oldset);
 369
 370                 if (req->aborted)
 371                         goto aborted;
 372                 if (req->state == FUSE_REQ_FINISHED)
 373                         return;
 374
 375                 /* Request is not yet in userspace, bail out */
 376                 if (req->state == FUSE_REQ_PENDING) {
 377                         list_del(&req->list);
 378                         __fuse_put_request(req);
 379                         req->out.h.error = -EINTR;
 380                         return;
 381                 }
 382         }
 383
 384         /*
 385          * Either request is already in userspace, or it was forced.
 386          * Wait it out.
 387          */
 388         spin_unlock(&fc->lock);
 389         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 390         spin_lock(&fc->lock);
 391
 392         if (!req->aborted)
 393                 return;
 394
 395  aborted:
 396         BUG_ON(req->state != FUSE_REQ_FINISHED);
 397         if (req->locked) {
 398                 /* This is uninterruptible sleep, because data is
 399                    being copied to/from the buffers of req.  During
 400                    locked state, there mustn't be any filesystem
 401                    operation (e.g. page fault), since that could lead
 402                    to deadlock */
 403                 spin_unlock(&fc->lock);
 404                 wait_event(req->waitq, !req->locked);
 405                 spin_lock(&fc->lock);
 406         }
 407 }
 408
 409 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 410 {
 411         req->isreply = 1;
 412         spin_lock(&fc->lock);
 413         if (!fc->connected)
 414                 req->out.h.error = -ENOTCONN;
 415         else if (fc->conn_error)
 416                 req->out.h.error = -ECONNREFUSED;
 417         else {
 418                 req->in.h.unique = fuse_get_unique(fc);
 419                 queue_request(fc, req);
 420                 /* acquire extra reference, since request is still needed
 421                    after request_end() */
 422                 __fuse_get_request(req);
 423
 424                 request_wait_answer(fc, req);
 425         }
 426         spin_unlock(&fc->lock);
 427 }
 428 EXPORT_SYMBOL_GPL(fuse_request_send);
 429
 430 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 431                                             struct fuse_req *req)
 432 {
 433         req->background = 1;
 434         fc->num_background++;
 435         if (fc->num_background == fc->max_background)
 436                 fc->blocked = 1;
 437         if (fc->num_background == fc->congestion_threshold &&
 438             fc->bdi_initialized) {
 439                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 440                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 441         }
 442         list_add_tail(&req->list, &fc->bg_queue);
 443         flush_bg_queue(fc);
 444 }
 445
 446 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 447 {
 448         spin_lock(&fc->lock);
 449         if (fc->connected) {
 450                 fuse_request_send_nowait_locked(fc, req);
 451                 spin_unlock(&fc->lock);
 452         } else {
 453                 req->out.h.error = -ENOTCONN;
 454                 request_end(fc, req);
 455         }
 456 }
 457
 458 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 459 {
 460         req->isreply = 1;
 461         fuse_request_send_nowait(fc, req);
 462 }
 463 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 464
 465 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 466                                           struct fuse_req *req, u64 unique)
 467 {
 468         int err = -ENODEV;
 469
 470         req->isreply = 0;
 471         req->in.h.unique = unique;
 472         spin_lock(&fc->lock);
 473         if (fc->connected) {
 474                 queue_request(fc, req);
 475                 err = 0;
 476         }
 477         spin_unlock(&fc->lock);
 478
 479         return err;
 480 }
 481
 482 /*
 483  * Called under fc->lock
 484  *
 485  * fc->connected must have been checked previously
 486  */
 487 void fuse_request_send_background_locked(struct fuse_conn *fc,
 488                                          struct fuse_req *req)
 489 {
 490         req->isreply = 1;
 491         fuse_request_send_nowait_locked(fc, req);
 492 }
 493
 494 /*
 495  * Lock the request.  Up to the next unlock_request() there mustn't be
 496  * anything that could cause a page-fault.  If the request was already
 497  * aborted bail out.
 498  */
 499 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 500 {
 501         int err = 0;
 502         if (req) {
 503                 spin_lock(&fc->lock);
 504                 if (req->aborted)
 505                         err = -ENOENT;
 506                 else
 507                         req->locked = 1;
 508                 spin_unlock(&fc->lock);
 509         }
 510         return err;
 511 }
 512
 513 /*
 514  * Unlock request.  If it was aborted during being locked, the
 515  * requester thread is currently waiting for it to be unlocked, so
 516  * wake it up.
 517  */
 518 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 519 {
 520         if (req) {
 521                 spin_lock(&fc->lock);
 522                 req->locked = 0;
 523                 if (req->aborted)
 524                         wake_up(&req->waitq);
 525                 spin_unlock(&fc->lock);
 526         }
 527 }
 528
 529 struct fuse_copy_state {
 530         struct fuse_conn *fc;
 531         int write;
 532         struct fuse_req *req;
 533         const struct iovec *iov;
 534         struct pipe_buffer *pipebufs;
 535         struct pipe_buffer *currbuf;
 536         struct pipe_inode_info *pipe;
 537         unsigned long nr_segs;
 538         unsigned long seglen;
 539         unsigned long addr;
 540         struct page *pg;
 541         void *mapaddr;
 542         void *buf;
 543         unsigned len;
 544         unsigned move_pages:1;
 545 };
 546
 547 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 548                            int write,
 549                            const struct iovec *iov, unsigned long nr_segs)
 550 {
 551         memset(cs, 0, sizeof(*cs));
 552         cs->fc = fc;
 553         cs->write = write;
 554         cs->iov = iov;
 555         cs->nr_segs = nr_segs;
 556 }
 557
 558 /* Unmap and put previous page of userspace buffer */
 559 static void fuse_copy_finish(struct fuse_copy_state *cs)
 560 {
 561         if (cs->currbuf) {
 562                 struct pipe_buffer *buf = cs->currbuf;
 563
 564                 if (!cs->write) {
 565                         buf->ops->unmap(cs->pipe, buf, cs->mapaddr);
 566                 } else {
 567                         kunmap(buf->page);
 568                         buf->len = PAGE_SIZE - cs->len;
 569                 }
 570                 cs->currbuf = NULL;
 571                 cs->mapaddr = NULL;
 572         } else if (cs->mapaddr) {
 573                 kunmap(cs->pg);
 574                 if (cs->write) {
 575                         flush_dcache_page(cs->pg);
 576                         set_page_dirty_lock(cs->pg);
 577                 }
 578                 put_page(cs->pg);
 579                 cs->mapaddr = NULL;
 580         }
 581 }
 582
 583 /*
 584  * Get another pagefull of userspace buffer, and map it to kernel
 585  * address space, and lock request
 586  */
 587 static int fuse_copy_fill(struct fuse_copy_state *cs)
 588 {
 589         unsigned long offset;
 590         int err;
 591
 592         unlock_request(cs->fc, cs->req);
 593         fuse_copy_finish(cs);
 594         if (cs->pipebufs) {
 595                 struct pipe_buffer *buf = cs->pipebufs;
 596
 597                 if (!cs->write) {
 598                         err = buf->ops->confirm(cs->pipe, buf);
 599                         if (err)
 600                                 return err;
 601
 602                         BUG_ON(!cs->nr_segs);
 603                         cs->currbuf = buf;
 604                         cs->mapaddr = buf->ops->map(cs->pipe, buf, 0);
 605                         cs->len = buf->len;
 606                         cs->buf = cs->mapaddr + buf->offset;
 607                         cs->pipebufs++;
 608                         cs->nr_segs--;
 609                 } else {
 610                         struct page *page;
 611
 612                         if (cs->nr_segs == cs->pipe->buffers)
 613                                 return -EIO;
 614
 615                         page = alloc_page(GFP_HIGHUSER);
 616                         if (!page)
 617                                 return -ENOMEM;
 618
 619                         buf->page = page;
 620                         buf->offset = 0;
 621                         buf->len = 0;
 622
 623                         cs->currbuf = buf;
 624                         cs->mapaddr = kmap(page);
 625                         cs->buf = cs->mapaddr;
 626                         cs->len = PAGE_SIZE;
 627                         cs->pipebufs++;
 628                         cs->nr_segs++;
 629                 }
 630         } else {
 631                 if (!cs->seglen) {
 632                         BUG_ON(!cs->nr_segs);
 633                         cs->seglen = cs->iov[0].iov_len;
 634                         cs->addr = (unsigned long) cs->iov[0].iov_base;
 635                         cs->iov++;
 636                         cs->nr_segs--;
 637                 }
 638                 err = get_user_pages_fast(cs->addr, 1, cs->write, &cs->pg);
 639                 if (err < 0)
 640                         return err;
 641                 BUG_ON(err != 1);
 642                 offset = cs->addr % PAGE_SIZE;
 643                 cs->mapaddr = kmap(cs->pg);
 644                 cs->buf = cs->mapaddr + offset;
 645                 cs->len = min(PAGE_SIZE - offset, cs->seglen);
 646                 cs->seglen -= cs->len;
 647                 cs->addr += cs->len;
 648         }
 649
 650         return lock_request(cs->fc, cs->req);
 651 }
 652
 653 /* Do as much copy to/from userspace buffer as we can */
 654 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 655 {
 656         unsigned ncpy = min(*size, cs->len);
 657         if (val) {
 658                 if (cs->write)
 659                         memcpy(cs->buf, *val, ncpy);
 660                 else
 661                         memcpy(*val, cs->buf, ncpy);
 662                 *val += ncpy;
 663         }
 664         *size -= ncpy;
 665         cs->len -= ncpy;
 666         cs->buf += ncpy;
 667         return ncpy;
 668 }
 669
 670 static int fuse_check_page(struct page *page)
 671 {
 672         if (page_mapcount(page) ||
 673             page->mapping != NULL ||
 674             page_count(page) != 1 ||
 675             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 676              ~(1 << PG_locked |
 677                1 << PG_referenced |
 678                1 << PG_uptodate |
 679                1 << PG_lru |
 680                1 << PG_active |
 681                1 << PG_reclaim))) {
 682                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 683                 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);
 684                 return 1;
 685         }
 686         return 0;
 687 }
 688
 689 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 690 {
 691         int err;
 692         struct page *oldpage = *pagep;
 693         struct page *newpage;
 694         struct pipe_buffer *buf = cs->pipebufs;
 695         struct address_space *mapping;
 696         pgoff_t index;
 697
 698         unlock_request(cs->fc, cs->req);
 699         fuse_copy_finish(cs);
 700
 701         err = buf->ops->confirm(cs->pipe, buf);
 702         if (err)
 703                 return err;
 704
 705         BUG_ON(!cs->nr_segs);
 706         cs->currbuf = buf;
 707         cs->len = buf->len;
 708         cs->pipebufs++;
 709         cs->nr_segs--;
 710
 711         if (cs->len != PAGE_SIZE)
 712                 goto out_fallback;
 713
 714         if (buf->ops->steal(cs->pipe, buf) != 0)
 715                 goto out_fallback;
 716
 717         newpage = buf->page;
 718
 719         if (!PageUptodate(newpage))
 720                 SetPageUptodate(newpage);
 721
 722         ClearPageMappedToDisk(newpage);
 723
 724         if (fuse_check_page(newpage) != 0)
 725                 goto out_fallback_unlock;
 726
 727         mapping = oldpage->mapping;
 728         index = oldpage->index;
 729
 730         /*
 731          * This is a new and locked page, it shouldn't be mapped or
 732          * have any special flags on it
 733          */
 734         if (WARN_ON(page_mapped(oldpage)))
 735                 goto out_fallback_unlock;
 736         if (WARN_ON(page_has_private(oldpage)))
 737                 goto out_fallback_unlock;
 738         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 739                 goto out_fallback_unlock;
 740         if (WARN_ON(PageMlocked(oldpage)))
 741                 goto out_fallback_unlock;
 742
 743         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 744         if (err) {
 745                 unlock_page(newpage);
 746                 return err;
 747         }
 748
 749         page_cache_get(newpage);
 750
 751         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 752                 lru_cache_add_file(newpage);
 753
 754         err = 0;
 755         spin_lock(&cs->fc->lock);
 756         if (cs->req->aborted)
 757                 err = -ENOENT;
 758         else
 759                 *pagep = newpage;
 760         spin_unlock(&cs->fc->lock);
 761
 762         if (err) {
 763                 unlock_page(newpage);
 764                 page_cache_release(newpage);
 765                 return err;
 766         }
 767
 768         unlock_page(oldpage);
 769         page_cache_release(oldpage);
 770         cs->len = 0;
 771
 772         return 0;
 773
 774 out_fallback_unlock:
 775         unlock_page(newpage);
 776 out_fallback:
 777         cs->mapaddr = buf->ops->map(cs->pipe, buf, 1);
 778         cs->buf = cs->mapaddr + buf->offset;
 779
 780         err = lock_request(cs->fc, cs->req);
 781         if (err)
 782                 return err;
 783
 784         return 1;
 785 }
 786
 787 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 788                          unsigned offset, unsigned count)
 789 {
 790         struct pipe_buffer *buf;
 791
 792         if (cs->nr_segs == cs->pipe->buffers)
 793                 return -EIO;
 794
 795         unlock_request(cs->fc, cs->req);
 796         fuse_copy_finish(cs);
 797
 798         buf = cs->pipebufs;
 799         page_cache_get(page);
 800         buf->page = page;
 801         buf->offset = offset;
 802         buf->len = count;
 803
 804         cs->pipebufs++;
 805         cs->nr_segs++;
 806         cs->len = 0;
 807
 808         return 0;
 809 }
 810
 811 /*
 812  * Copy a page in the request to/from the userspace buffer.  Must be
 813  * done atomically
 814  */
 815 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 816                           unsigned offset, unsigned count, int zeroing)
 817 {
 818         int err;
 819         struct page *page = *pagep;
 820
 821         if (page && zeroing && count < PAGE_SIZE)
 822                 clear_highpage(page);
 823
 824         while (count) {
 825                 if (cs->write && cs->pipebufs && page) {
 826                         return fuse_ref_page(cs, page, offset, count);
 827                 } else if (!cs->len) {
 828                         if (cs->move_pages && page &&
 829                             offset == 0 && count == PAGE_SIZE) {
 830                                 err = fuse_try_move_page(cs, pagep);
 831                                 if (err <= 0)
 832                                         return err;
 833                         } else {
 834                                 err = fuse_copy_fill(cs);
 835                                 if (err)
 836                                         return err;
 837                         }
 838                 }
 839                 if (page) {
 840                         void *mapaddr = kmap_atomic(page, KM_USER0);
 841                         void *buf = mapaddr + offset;
 842                         offset += fuse_copy_do(cs, &buf, &count);
 843                         kunmap_atomic(mapaddr, KM_USER0);
 844                 } else
 845                         offset += fuse_copy_do(cs, NULL, &count);
 846         }
 847         if (page && !cs->write)
 848                 flush_dcache_page(page);
 849         return 0;
 850 }
 851
 852 /* Copy pages in the request to/from userspace buffer */
 853 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 854                            int zeroing)
 855 {
 856         unsigned i;
 857         struct fuse_req *req = cs->req;
 858         unsigned offset = req->page_offset;
 859         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 860
 861         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 862                 int err;
 863
 864                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
 865                                      zeroing);
 866                 if (err)
 867                         return err;
 868
 869                 nbytes -= count;
 870                 count = min(nbytes, (unsigned) PAGE_SIZE);
 871                 offset = 0;
 872         }
 873         return 0;
 874 }
 875
 876 /* Copy a single argument in the request to/from userspace buffer */
 877 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 878 {
 879         while (size) {
 880                 if (!cs->len) {
 881                         int err = fuse_copy_fill(cs);
 882                         if (err)
 883                                 return err;
 884                 }
 885                 fuse_copy_do(cs, &val, &size);
 886         }
 887         return 0;
 888 }
 889
 890 /* Copy request arguments to/from userspace buffer */
 891 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 892                           unsigned argpages, struct fuse_arg *args,
 893                           int zeroing)
 894 {
 895         int err = 0;
 896         unsigned i;
 897
 898         for (i = 0; !err && i < numargs; i++)  {
 899                 struct fuse_arg *arg = &args[i];
 900                 if (i == numargs - 1 && argpages)
 901                         err = fuse_copy_pages(cs, arg->size, zeroing);
 902                 else
 903                         err = fuse_copy_one(cs, arg->value, arg->size);
 904         }
 905         return err;
 906 }
 907
 908 static int forget_pending(struct fuse_conn *fc)
 909 {
 910         return fc->forget_list_head.next != NULL;
 911 }
 912
 913 static int request_pending(struct fuse_conn *fc)
 914 {
 915         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts) ||
 916                 forget_pending(fc);
 917 }
 918
 919 /* Wait until a request is available on the pending list */
 920 static void request_wait(struct fuse_conn *fc)
 921 __releases(fc->lock)
 922 __acquires(fc->lock)
 923 {
 924         DECLARE_WAITQUEUE(wait, current);
 925
 926         add_wait_queue_exclusive(&fc->waitq, &wait);
 927         while (fc->connected && !request_pending(fc)) {
 928                 set_current_state(TASK_INTERRUPTIBLE);
 929                 if (signal_pending(current))
 930                         break;
 931
 932                 spin_unlock(&fc->lock);
 933                 schedule();
 934                 spin_lock(&fc->lock);
 935         }
 936         set_current_state(TASK_RUNNING);
 937         remove_wait_queue(&fc->waitq, &wait);
 938 }
 939
 940 /*
 941  * Transfer an interrupt request to userspace
 942  *
 943  * Unlike other requests this is assembled on demand, without a need
 944  * to allocate a separate fuse_req structure.
 945  *
 946  * Called with fc->lock held, releases it
 947  */
 948 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_copy_state *cs,
 949                                size_t nbytes, struct fuse_req *req)
 950 __releases(fc->lock)
 951 {
 952         struct fuse_in_header ih;
 953         struct fuse_interrupt_in arg;
 954         unsigned reqsize = sizeof(ih) + sizeof(arg);
 955         int err;
 956
 957         list_del_init(&req->intr_entry);
 958         req->intr_unique = fuse_get_unique(fc);
 959         memset(&ih, 0, sizeof(ih));
 960         memset(&arg, 0, sizeof(arg));
 961         ih.len = reqsize;
 962         ih.opcode = FUSE_INTERRUPT;
 963         ih.unique = req->intr_unique;
 964         arg.unique = req->in.h.unique;
 965
 966         spin_unlock(&fc->lock);
 967         if (nbytes < reqsize)
 968                 return -EINVAL;
 969
 970         err = fuse_copy_one(cs, &ih, sizeof(ih));
 971         if (!err)
 972                 err = fuse_copy_one(cs, &arg, sizeof(arg));
 973         fuse_copy_finish(cs);
 974
 975         return err ? err : reqsize;
 976 }
 977
 978 static struct fuse_forget_link *dequeue_forget(struct fuse_conn *fc,
 979                                                unsigned max,
 980                                                unsigned *countp)
 981 {
 982         struct fuse_forget_link *head = fc->forget_list_head.next;
 983         struct fuse_forget_link **newhead = &head;
 984         unsigned count;
 985
 986         for (count = 0; *newhead != NULL && count < max; count++)
 987                 newhead = &(*newhead)->next;
 988
 989         fc->forget_list_head.next = *newhead;
 990         *newhead = NULL;
 991         if (fc->forget_list_head.next == NULL)
 992                 fc->forget_list_tail = &fc->forget_list_head;
 993
 994         if (countp != NULL)
 995                 *countp = count;
 996
 997         return head;
 998 }
 999
1000 static int fuse_read_single_forget(struct fuse_conn *fc,
1001                                    struct fuse_copy_state *cs,
1002                                    size_t nbytes)
1003 __releases(fc->lock)
1004 {
1005         int err;
1006         struct fuse_forget_link *forget = dequeue_forget(fc, 1, NULL);
1007         struct fuse_forget_in arg = {
1008                 .nlookup = forget->forget_one.nlookup,
1009         };
1010         struct fuse_in_header ih = {
1011                 .opcode = FUSE_FORGET,
1012                 .nodeid = forget->forget_one.nodeid,
1013                 .unique = fuse_get_unique(fc),
1014                 .len = sizeof(ih) + sizeof(arg),
1015         };
1016
1017         spin_unlock(&fc->lock);
1018         kfree(forget);
1019         if (nbytes < ih.len)
1020                 return -EINVAL;
1021
1022         err = fuse_copy_one(cs, &ih, sizeof(ih));
1023         if (!err)
1024                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1025         fuse_copy_finish(cs);
1026
1027         if (err)
1028                 return err;
1029
1030         return ih.len;
1031 }
1032
1033 static int fuse_read_batch_forget(struct fuse_conn *fc,
1034                                    struct fuse_copy_state *cs, size_t nbytes)
1035 __releases(fc->lock)
1036 {
1037         int err;
1038         unsigned max_forgets;
1039         unsigned count;
1040         struct fuse_forget_link *head;
1041         struct fuse_batch_forget_in arg = { .count = 0 };
1042         struct fuse_in_header ih = {
1043                 .opcode = FUSE_BATCH_FORGET,
1044                 .unique = fuse_get_unique(fc),
1045                 .len = sizeof(ih) + sizeof(arg),
1046         };
1047
1048         if (nbytes < ih.len) {
1049                 spin_unlock(&fc->lock);
1050                 return -EINVAL;
1051         }
1052
1053         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1054         head = dequeue_forget(fc, max_forgets, &count);
1055         spin_unlock(&fc->lock);
1056
1057         arg.count = count;
1058         ih.len += count * sizeof(struct fuse_forget_one);
1059         err = fuse_copy_one(cs, &ih, sizeof(ih));
1060         if (!err)
1061                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1062
1063         while (head) {
1064                 struct fuse_forget_link *forget = head;
1065
1066                 if (!err) {
1067                         err = fuse_copy_one(cs, &forget->forget_one,
1068                                             sizeof(forget->forget_one));
1069                 }
1070                 head = forget->next;
1071                 kfree(forget);
1072         }
1073
1074         fuse_copy_finish(cs);
1075
1076         if (err)
1077                 return err;
1078
1079         return ih.len;
1080 }
1081
1082 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_copy_state *cs,
1083                             size_t nbytes)
1084 __releases(fc->lock)
1085 {
1086         if (fc->minor < 16 || fc->forget_list_head.next->next == NULL)
1087                 return fuse_read_single_forget(fc, cs, nbytes);
1088         else
1089                 return fuse_read_batch_forget(fc, cs, nbytes);
1090 }
1091
1092 /*
1093  * Read a single request into the userspace filesystem's buffer.  This
1094  * function waits until a request is available, then removes it from
1095  * the pending list and copies request data to userspace buffer.  If
1096  * no reply is needed (FORGET) or request has been aborted or there
1097  * was an error during the copying then it's finished by calling
1098  * request_end().  Otherwise add it to the processing list, and set
1099  * the 'sent' flag.
1100  */
1101 static ssize_t fuse_dev_do_read(struct fuse_conn *fc, struct file *file,
1102                                 struct fuse_copy_state *cs, size_t nbytes)
1103 {
1104         int err;
1105         struct fuse_req *req;
1106         struct fuse_in *in;
1107         unsigned reqsize;
1108
1109  restart:
1110         spin_lock(&fc->lock);
1111         err = -EAGAIN;
1112         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
1113             !request_pending(fc))
1114                 goto err_unlock;
1115
1116         request_wait(fc);
1117         err = -ENODEV;
1118         if (!fc->connected)
1119                 goto err_unlock;
1120         err = -ERESTARTSYS;
1121         if (!request_pending(fc))
1122                 goto err_unlock;
1123
1124         if (!list_empty(&fc->interrupts)) {
1125                 req = list_entry(fc->interrupts.next, struct fuse_req,
1126                                  intr_entry);
1127                 return fuse_read_interrupt(fc, cs, nbytes, req);
1128         }
1129
1130         if (forget_pending(fc)) {
1131                 if (list_empty(&fc->pending) || fc->forget_batch-- > 0)
1132                         return fuse_read_forget(fc, cs, nbytes);
1133
1134                 if (fc->forget_batch <= -8)
1135                         fc->forget_batch = 16;
1136         }
1137
1138         req = list_entry(fc->pending.next, struct fuse_req, list);
1139         req->state = FUSE_REQ_READING;
1140         list_move(&req->list, &fc->io);
1141
1142         in = &req->in;
1143         reqsize = in->h.len;
1144         /* If request is too large, reply with an error and restart the read */
1145         if (nbytes < reqsize) {
1146                 req->out.h.error = -EIO;
1147                 /* SETXATTR is special, since it may contain too large data */
1148                 if (in->h.opcode == FUSE_SETXATTR)
1149                         req->out.h.error = -E2BIG;
1150                 request_end(fc, req);
1151                 goto restart;
1152         }
1153         spin_unlock(&fc->lock);
1154         cs->req = req;
1155         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1156         if (!err)
1157                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1158                                      (struct fuse_arg *) in->args, 0);
1159         fuse_copy_finish(cs);
1160         spin_lock(&fc->lock);
1161         req->locked = 0;
1162         if (req->aborted) {
1163                 request_end(fc, req);
1164                 return -ENODEV;
1165         }
1166         if (err) {
1167                 req->out.h.error = -EIO;
1168                 request_end(fc, req);
1169                 return err;
1170         }
1171         if (!req->isreply)
1172                 request_end(fc, req);
1173         else {
1174                 req->state = FUSE_REQ_SENT;
1175                 list_move_tail(&req->list, &fc->processing);
1176                 if (req->interrupted)
1177                         queue_interrupt(fc, req);
1178                 spin_unlock(&fc->lock);
1179         }
1180         return reqsize;
1181
1182  err_unlock:
1183         spin_unlock(&fc->lock);
1184         return err;
1185 }
1186
1187 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
1188                               unsigned long nr_segs, loff_t pos)
1189 {
1190         struct fuse_copy_state cs;
1191         struct file *file = iocb->ki_filp;
1192         struct fuse_conn *fc = fuse_get_conn(file);
1193         if (!fc)
1194                 return -EPERM;
1195
1196         fuse_copy_init(&cs, fc, 1, iov, nr_segs);
1197
1198         return fuse_dev_do_read(fc, file, &cs, iov_length(iov, nr_segs));
1199 }
1200
1201 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1202                                     struct pipe_inode_info *pipe,
1203                                     size_t len, unsigned int flags)
1204 {
1205         int ret;
1206         int page_nr = 0;
1207         int do_wakeup = 0;
1208         struct pipe_buffer *bufs;
1209         struct fuse_copy_state cs;
1210         struct fuse_conn *fc = fuse_get_conn(in);
1211         if (!fc)
1212                 return -EPERM;
1213
1214         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1215         if (!bufs)
1216                 return -ENOMEM;
1217
1218         fuse_copy_init(&cs, fc, 1, NULL, 0);
1219         cs.pipebufs = bufs;
1220         cs.pipe = pipe;
1221         ret = fuse_dev_do_read(fc, in, &cs, len);
1222         if (ret < 0)
1223                 goto out;
1224
1225         ret = 0;
1226         pipe_lock(pipe);
1227
1228         if (!pipe->readers) {
1229                 send_sig(SIGPIPE, current, 0);
1230                 if (!ret)
1231                         ret = -EPIPE;
1232                 goto out_unlock;
1233         }
1234
1235         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1236                 ret = -EIO;
1237                 goto out_unlock;
1238         }
1239
1240         while (page_nr < cs.nr_segs) {
1241                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
1242                 struct pipe_buffer *buf = pipe->bufs + newbuf;
1243
1244                 buf->page = bufs[page_nr].page;
1245                 buf->offset = bufs[page_nr].offset;
1246                 buf->len = bufs[page_nr].len;
1247                 /*
1248                  * Need to be careful about this.  Having buf->ops in module
1249                  * code can Oops if the buffer persists after module unload.
1250                  */
1251                 buf->ops = &nosteal_pipe_buf_ops;
1252
1253                 pipe->nrbufs++;
1254                 page_nr++;
1255                 ret += buf->len;
1256
1257                 if (pipe->inode)
1258                         do_wakeup = 1;
1259         }
1260
1261 out_unlock:
1262         pipe_unlock(pipe);
1263
1264         if (do_wakeup) {
1265                 smp_mb();
1266                 if (waitqueue_active(&pipe->wait))
1267                         wake_up_interruptible(&pipe->wait);
1268                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1269         }
1270
1271 out:
1272         for (; page_nr < cs.nr_segs; page_nr++)
1273                 page_cache_release(bufs[page_nr].page);
1274
1275         kfree(bufs);
1276         return ret;
1277 }
1278
1279 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1280                             struct fuse_copy_state *cs)
1281 {
1282         struct fuse_notify_poll_wakeup_out outarg;
1283         int err = -EINVAL;
1284
1285         if (size != sizeof(outarg))
1286                 goto err;
1287
1288         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1289         if (err)
1290                 goto err;
1291
1292         fuse_copy_finish(cs);
1293         return fuse_notify_poll_wakeup(fc, &outarg);
1294
1295 err:
1296         fuse_copy_finish(cs);
1297         return err;
1298 }
1299
1300 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1301                                    struct fuse_copy_state *cs)
1302 {
1303         struct fuse_notify_inval_inode_out outarg;
1304         int err = -EINVAL;
1305
1306         if (size != sizeof(outarg))
1307                 goto err;
1308
1309         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1310         if (err)
1311                 goto err;
1312         fuse_copy_finish(cs);
1313
1314         down_read(&fc->killsb);
1315         err = -ENOENT;
1316         if (fc->sb) {
1317                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1318                                                outarg.off, outarg.len);
1319         }
1320         up_read(&fc->killsb);
1321         return err;
1322
1323 err:
1324         fuse_copy_finish(cs);
1325         return err;
1326 }
1327
1328 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1329                                    struct fuse_copy_state *cs)
1330 {
1331         struct fuse_notify_inval_entry_out outarg;
1332         int err = -ENOMEM;
1333         char *buf;
1334         struct qstr name;
1335
1336         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1337         if (!buf)
1338                 goto err;
1339
1340         err = -EINVAL;
1341         if (size < sizeof(outarg))
1342                 goto err;
1343
1344         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1345         if (err)
1346                 goto err;
1347
1348         err = -ENAMETOOLONG;
1349         if (outarg.namelen > FUSE_NAME_MAX)
1350                 goto err;
1351
1352         err = -EINVAL;
1353         if (size != sizeof(outarg) + outarg.namelen + 1)
1354                 goto err;
1355
1356         name.name = buf;
1357         name.len = outarg.namelen;
1358         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1359         if (err)
1360                 goto err;
1361         fuse_copy_finish(cs);
1362         buf[outarg.namelen] = 0;
1363         name.hash = full_name_hash(name.name, name.len);
1364
1365         down_read(&fc->killsb);
1366         err = -ENOENT;
1367         if (fc->sb)
1368                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, &name);
1369         up_read(&fc->killsb);
1370         kfree(buf);
1371         return err;
1372
1373 err:
1374         kfree(buf);
1375         fuse_copy_finish(cs);
1376         return err;
1377 }
1378
1379 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1380                              struct fuse_copy_state *cs)
1381 {
1382         struct fuse_notify_store_out outarg;
1383         struct inode *inode;
1384         struct address_space *mapping;
1385         u64 nodeid;
1386         int err;
1387         pgoff_t index;
1388         unsigned int offset;
1389         unsigned int num;
1390         loff_t file_size;
1391         loff_t end;
1392
1393         err = -EINVAL;
1394         if (size < sizeof(outarg))
1395                 goto out_finish;
1396
1397         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1398         if (err)
1399                 goto out_finish;
1400
1401         err = -EINVAL;
1402         if (size - sizeof(outarg) != outarg.size)
1403                 goto out_finish;
1404
1405         nodeid = outarg.nodeid;
1406
1407         down_read(&fc->killsb);
1408
1409         err = -ENOENT;
1410         if (!fc->sb)
1411                 goto out_up_killsb;
1412
1413         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1414         if (!inode)
1415                 goto out_up_killsb;
1416
1417         mapping = inode->i_mapping;
1418         index = outarg.offset >> PAGE_CACHE_SHIFT;
1419         offset = outarg.offset & ~PAGE_CACHE_MASK;
1420         file_size = i_size_read(inode);
1421         end = outarg.offset + outarg.size;
1422         if (end > file_size) {
1423                 file_size = end;
1424                 fuse_write_update_size(inode, file_size);
1425         }
1426
1427         num = outarg.size;
1428         while (num) {
1429                 struct page *page;
1430                 unsigned int this_num;
1431
1432                 err = -ENOMEM;
1433                 page = find_or_create_page(mapping, index,
1434                                            mapping_gfp_mask(mapping));
1435                 if (!page)
1436                         goto out_iput;
1437
1438                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1439                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1440                 if (!err && offset == 0 && (num != 0 || file_size == end))
1441                         SetPageUptodate(page);
1442                 unlock_page(page);
1443                 page_cache_release(page);
1444
1445                 if (err)
1446                         goto out_iput;
1447
1448                 num -= this_num;
1449                 offset = 0;
1450                 index++;
1451         }
1452
1453         err = 0;
1454
1455 out_iput:
1456         iput(inode);
1457 out_up_killsb:
1458         up_read(&fc->killsb);
1459 out_finish:
1460         fuse_copy_finish(cs);
1461         return err;
1462 }
1463
1464 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1465 {
1466         release_pages(req->pages, req->num_pages, 0);
1467 }
1468
1469 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1470                          struct fuse_notify_retrieve_out *outarg)
1471 {
1472         int err;
1473         struct address_space *mapping = inode->i_mapping;
1474         struct fuse_req *req;
1475         pgoff_t index;
1476         loff_t file_size;
1477         unsigned int num;
1478         unsigned int offset;
1479         size_t total_len = 0;
1480
1481         req = fuse_get_req(fc);
1482         if (IS_ERR(req))
1483                 return PTR_ERR(req);
1484
1485         offset = outarg->offset & ~PAGE_CACHE_MASK;
1486
1487         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1488         req->in.h.nodeid = outarg->nodeid;
1489         req->in.numargs = 2;
1490         req->in.argpages = 1;
1491         req->page_offset = offset;
1492         req->end = fuse_retrieve_end;
1493
1494         index = outarg->offset >> PAGE_CACHE_SHIFT;
1495         file_size = i_size_read(inode);
1496         num = outarg->size;
1497         if (outarg->offset > file_size)
1498                 num = 0;
1499         else if (outarg->offset + num > file_size)
1500                 num = file_size - outarg->offset;
1501
1502         while (num && req->num_pages < FUSE_MAX_PAGES_PER_REQ) {
1503                 struct page *page;
1504                 unsigned int this_num;
1505
1506                 page = find_get_page(mapping, index);
1507                 if (!page)
1508                         break;
1509
1510                 this_num = min_t(unsigned, num, PAGE_CACHE_SIZE - offset);
1511                 req->pages[req->num_pages] = page;
1512                 req->num_pages++;
1513
1514                 offset = 0;
1515                 num -= this_num;
1516                 total_len += this_num;
1517                 index++;
1518         }
1519         req->misc.retrieve_in.offset = outarg->offset;
1520         req->misc.retrieve_in.size = total_len;
1521         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1522         req->in.args[0].value = &req->misc.retrieve_in;
1523         req->in.args[1].size = total_len;
1524
1525         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1526         if (err)
1527                 fuse_retrieve_end(fc, req);
1528
1529         return err;
1530 }
1531
1532 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1533                                 struct fuse_copy_state *cs)
1534 {
1535         struct fuse_notify_retrieve_out outarg;
1536         struct inode *inode;
1537         int err;
1538
1539         err = -EINVAL;
1540         if (size != sizeof(outarg))
1541                 goto copy_finish;
1542
1543         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1544         if (err)
1545                 goto copy_finish;
1546
1547         fuse_copy_finish(cs);
1548
1549         down_read(&fc->killsb);
1550         err = -ENOENT;
1551         if (fc->sb) {
1552                 u64 nodeid = outarg.nodeid;
1553
1554                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1555                 if (inode) {
1556                         err = fuse_retrieve(fc, inode, &outarg);
1557                         iput(inode);
1558                 }
1559         }
1560         up_read(&fc->killsb);
1561
1562         return err;
1563
1564 copy_finish:
1565         fuse_copy_finish(cs);
1566         return err;
1567 }
1568
1569 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1570                        unsigned int size, struct fuse_copy_state *cs)
1571 {
1572         /* Don't try to move pages (yet) */
1573         cs->move_pages = 0;
1574
1575         switch (code) {
1576         case FUSE_NOTIFY_POLL:
1577                 return fuse_notify_poll(fc, size, cs);
1578
1579         case FUSE_NOTIFY_INVAL_INODE:
1580                 return fuse_notify_inval_inode(fc, size, cs);
1581
1582         case FUSE_NOTIFY_INVAL_ENTRY:
1583                 return fuse_notify_inval_entry(fc, size, cs);
1584
1585         case FUSE_NOTIFY_STORE:
1586                 return fuse_notify_store(fc, size, cs);
1587
1588         case FUSE_NOTIFY_RETRIEVE:
1589                 return fuse_notify_retrieve(fc, size, cs);
1590
1591         default:
1592                 fuse_copy_finish(cs);
1593                 return -EINVAL;
1594         }
1595 }
1596
1597 /* Look up request on processing list by unique ID */
1598 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
1599 {
1600         struct list_head *entry;
1601
1602         list_for_each(entry, &fc->processing) {
1603                 struct fuse_req *req;
1604                 req = list_entry(entry, struct fuse_req, list);
1605                 if (req->in.h.unique == unique || req->intr_unique == unique)
1606                         return req;
1607         }
1608         return NULL;
1609 }
1610
1611 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1612                          unsigned nbytes)
1613 {
1614         unsigned reqsize = sizeof(struct fuse_out_header);
1615
1616         if (out->h.error)
1617                 return nbytes != reqsize ? -EINVAL : 0;
1618
1619         reqsize += len_args(out->numargs, out->args);
1620
1621         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1622                 return -EINVAL;
1623         else if (reqsize > nbytes) {
1624                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1625                 unsigned diffsize = reqsize - nbytes;
1626                 if (diffsize > lastarg->size)
1627                         return -EINVAL;
1628                 lastarg->size -= diffsize;
1629         }
1630         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1631                               out->page_zeroing);
1632 }
1633
1634 /*
1635  * Write a single reply to a request.  First the header is copied from
1636  * the write buffer.  The request is then searched on the processing
1637  * list by the unique ID found in the header.  If found, then remove
1638  * it from the list and copy the rest of the buffer to the request.
1639  * The request is finished by calling request_end()
1640  */
1641 static ssize_t fuse_dev_do_write(struct fuse_conn *fc,
1642                                  struct fuse_copy_state *cs, size_t nbytes)
1643 {
1644         int err;
1645         struct fuse_req *req;
1646         struct fuse_out_header oh;
1647
1648         if (nbytes < sizeof(struct fuse_out_header))
1649                 return -EINVAL;
1650
1651         err = fuse_copy_one(cs, &oh, sizeof(oh));
1652         if (err)
1653                 goto err_finish;
1654
1655         err = -EINVAL;
1656         if (oh.len != nbytes)
1657                 goto err_finish;
1658
1659         /*
1660          * Zero oh.unique indicates unsolicited notification message
1661          * and error contains notification code.
1662          */
1663         if (!oh.unique) {
1664                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1665                 return err ? err : nbytes;
1666         }
1667
1668         err = -EINVAL;
1669         if (oh.error <= -1000 || oh.error > 0)
1670                 goto err_finish;
1671
1672         spin_lock(&fc->lock);
1673         err = -ENOENT;
1674         if (!fc->connected)
1675                 goto err_unlock;
1676
1677         req = request_find(fc, oh.unique);
1678         if (!req)
1679                 goto err_unlock;
1680
1681         if (req->aborted) {
1682                 spin_unlock(&fc->lock);
1683                 fuse_copy_finish(cs);
1684                 spin_lock(&fc->lock);
1685                 request_end(fc, req);
1686                 return -ENOENT;
1687         }
1688         /* Is it an interrupt reply? */
1689         if (req->intr_unique == oh.unique) {
1690                 err = -EINVAL;
1691                 if (nbytes != sizeof(struct fuse_out_header))
1692                         goto err_unlock;
1693
1694                 if (oh.error == -ENOSYS)
1695                         fc->no_interrupt = 1;
1696                 else if (oh.error == -EAGAIN)
1697                         queue_interrupt(fc, req);
1698
1699                 spin_unlock(&fc->lock);
1700                 fuse_copy_finish(cs);
1701                 return nbytes;
1702         }
1703
1704         req->state = FUSE_REQ_WRITING;
1705         list_move(&req->list, &fc->io);
1706         req->out.h = oh;
1707         req->locked = 1;
1708         cs->req = req;
1709         if (!req->out.page_replace)
1710                 cs->move_pages = 0;
1711         spin_unlock(&fc->lock);
1712
1713         err = copy_out_args(cs, &req->out, nbytes);
1714         fuse_copy_finish(cs);
1715
1716         spin_lock(&fc->lock);
1717         req->locked = 0;
1718         if (!err) {
1719                 if (req->aborted)
1720                         err = -ENOENT;
1721         } else if (!req->aborted)
1722                 req->out.h.error = -EIO;
1723         request_end(fc, req);
1724
1725         return err ? err : nbytes;
1726
1727  err_unlock:
1728         spin_unlock(&fc->lock);
1729  err_finish:
1730         fuse_copy_finish(cs);
1731         return err;
1732 }
1733
1734 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
1735                               unsigned long nr_segs, loff_t pos)
1736 {
1737         struct fuse_copy_state cs;
1738         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
1739         if (!fc)
1740                 return -EPERM;
1741
1742         fuse_copy_init(&cs, fc, 0, iov, nr_segs);
1743
1744         return fuse_dev_do_write(fc, &cs, iov_length(iov, nr_segs));
1745 }
1746
1747 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
1748                                      struct file *out, loff_t *ppos,
1749                                      size_t len, unsigned int flags)
1750 {
1751         unsigned nbuf;
1752         unsigned idx;
1753         struct pipe_buffer *bufs;
1754         struct fuse_copy_state cs;
1755         struct fuse_conn *fc;
1756         size_t rem;
1757         ssize_t ret;
1758
1759         fc = fuse_get_conn(out);
1760         if (!fc)
1761                 return -EPERM;
1762
1763         bufs = kmalloc(pipe->buffers * sizeof(struct pipe_buffer), GFP_KERNEL);
1764         if (!bufs)
1765                 return -ENOMEM;
1766
1767         pipe_lock(pipe);
1768         nbuf = 0;
1769         rem = 0;
1770         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
1771                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
1772
1773         ret = -EINVAL;
1774         if (rem < len) {
1775                 pipe_unlock(pipe);
1776                 goto out;
1777         }
1778
1779         rem = len;
1780         while (rem) {
1781                 struct pipe_buffer *ibuf;
1782                 struct pipe_buffer *obuf;
1783
1784                 BUG_ON(nbuf >= pipe->buffers);
1785                 BUG_ON(!pipe->nrbufs);
1786                 ibuf = &pipe->bufs[pipe->curbuf];
1787                 obuf = &bufs[nbuf];
1788
1789                 if (rem >= ibuf->len) {
1790                         *obuf = *ibuf;
1791                         ibuf->ops = NULL;
1792                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1793                         pipe->nrbufs--;
1794                 } else {
1795                         ibuf->ops->get(pipe, ibuf);
1796                         *obuf = *ibuf;
1797                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1798                         obuf->len = rem;
1799                         ibuf->offset += obuf->len;
1800                         ibuf->len -= obuf->len;
1801                 }
1802                 nbuf++;
1803                 rem -= obuf->len;
1804         }
1805         pipe_unlock(pipe);
1806
1807         fuse_copy_init(&cs, fc, 0, NULL, nbuf);
1808         cs.pipebufs = bufs;
1809         cs.pipe = pipe;
1810
1811         if (flags & SPLICE_F_MOVE)
1812                 cs.move_pages = 1;
1813
1814         ret = fuse_dev_do_write(fc, &cs, len);
1815
1816         for (idx = 0; idx < nbuf; idx++) {
1817                 struct pipe_buffer *buf = &bufs[idx];
1818                 buf->ops->release(pipe, buf);
1819         }
1820 out:
1821         kfree(bufs);
1822         return ret;
1823 }
1824
1825 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1826 {
1827         unsigned mask = POLLOUT | POLLWRNORM;
1828         struct fuse_conn *fc = fuse_get_conn(file);
1829         if (!fc)
1830                 return POLLERR;
1831
1832         poll_wait(file, &fc->waitq, wait);
1833
1834         spin_lock(&fc->lock);
1835         if (!fc->connected)
1836                 mask = POLLERR;
1837         else if (request_pending(fc))
1838                 mask |= POLLIN | POLLRDNORM;
1839         spin_unlock(&fc->lock);
1840
1841         return mask;
1842 }
1843
1844 /*
1845  * Abort all requests on the given list (pending or processing)
1846  *
1847  * This function releases and reacquires fc->lock
1848  */
1849 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1850 __releases(fc->lock)
1851 __acquires(fc->lock)
1852 {
1853         while (!list_empty(head)) {
1854                 struct fuse_req *req;
1855                 req = list_entry(head->next, struct fuse_req, list);
1856                 req->out.h.error = -ECONNABORTED;
1857                 request_end(fc, req);
1858                 spin_lock(&fc->lock);
1859         }
1860 }
1861
1862 /*
1863  * Abort requests under I/O
1864  *
1865  * The requests are set to aborted and finished, and the request
1866  * waiter is woken up.  This will make request_wait_answer() wait
1867  * until the request is unlocked and then return.
1868  *
1869  * If the request is asynchronous, then the end function needs to be
1870  * called after waiting for the request to be unlocked (if it was
1871  * locked).
1872  */
1873 static void end_io_requests(struct fuse_conn *fc)
1874 __releases(fc->lock)
1875 __acquires(fc->lock)
1876 {
1877         while (!list_empty(&fc->io)) {
1878                 struct fuse_req *req =
1879                         list_entry(fc->io.next, struct fuse_req, list);
1880                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1881
1882                 req->aborted = 1;
1883                 req->out.h.error = -ECONNABORTED;
1884                 req->state = FUSE_REQ_FINISHED;
1885                 list_del_init(&req->list);
1886                 wake_up(&req->waitq);
1887                 if (end) {
1888                         req->end = NULL;
1889                         __fuse_get_request(req);
1890                         spin_unlock(&fc->lock);
1891                         wait_event(req->waitq, !req->locked);
1892                         end(fc, req);
1893                         fuse_put_request(fc, req);
1894                         spin_lock(&fc->lock);
1895                 }
1896         }
1897 }
1898
1899 static void end_queued_requests(struct fuse_conn *fc)
1900 __releases(fc->lock)
1901 __acquires(fc->lock)
1902 {
1903         fc->max_background = UINT_MAX;
1904         flush_bg_queue(fc);
1905         end_requests(fc, &fc->pending);
1906         end_requests(fc, &fc->processing);
1907         while (forget_pending(fc))
1908                 kfree(dequeue_forget(fc, 1, NULL));
1909 }
1910
1911 static void end_polls(struct fuse_conn *fc)
1912 {
1913         struct rb_node *p;
1914
1915         p = rb_first(&fc->polled_files);
1916
1917         while (p) {
1918                 struct fuse_file *ff;
1919                 ff = rb_entry(p, struct fuse_file, polled_node);
1920                 wake_up_interruptible_all(&ff->poll_wait);
1921
1922                 p = rb_next(p);
1923         }
1924 }
1925
1926 /*
1927  * Abort all requests.
1928  *
1929  * Emergency exit in case of a malicious or accidental deadlock, or
1930  * just a hung filesystem.
1931  *
1932  * The same effect is usually achievable through killing the
1933  * filesystem daemon and all users of the filesystem.  The exception
1934  * is the combination of an asynchronous request and the tricky
1935  * deadlock (see Documentation/filesystems/fuse.txt).
1936  *
1937  * During the aborting, progression of requests from the pending and
1938  * processing lists onto the io list, and progression of new requests
1939  * onto the pending list is prevented by req->connected being false.
1940  *
1941  * Progression of requests under I/O to the processing list is
1942  * prevented by the req->aborted flag being true for these requests.
1943  * For this reason requests on the io list must be aborted first.
1944  */
1945 void fuse_abort_conn(struct fuse_conn *fc)
1946 {
1947         spin_lock(&fc->lock);
1948         if (fc->connected) {
1949                 fc->connected = 0;
1950                 fc->blocked = 0;
1951                 end_io_requests(fc);
1952                 end_queued_requests(fc);
1953                 end_polls(fc);
1954                 wake_up_all(&fc->waitq);
1955                 wake_up_all(&fc->blocked_waitq);
1956                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1957         }
1958         spin_unlock(&fc->lock);
1959 }
1960 EXPORT_SYMBOL_GPL(fuse_abort_conn);
1961
1962 int fuse_dev_release(struct inode *inode, struct file *file)
1963 {
1964         struct fuse_conn *fc = fuse_get_conn(file);
1965         if (fc) {
1966                 spin_lock(&fc->lock);
1967                 fc->connected = 0;
1968                 fc->blocked = 0;
1969                 end_queued_requests(fc);
1970                 end_polls(fc);
1971                 wake_up_all(&fc->blocked_waitq);
1972                 spin_unlock(&fc->lock);
1973                 fuse_conn_put(fc);
1974         }
1975
1976         return 0;
1977 }
1978 EXPORT_SYMBOL_GPL(fuse_dev_release);
1979
1980 static int fuse_dev_fasync(int fd, struct file *file, int on)
1981 {
1982         struct fuse_conn *fc = fuse_get_conn(file);
1983         if (!fc)
1984                 return -EPERM;
1985
1986         /* No locking - fasync_helper does its own locking */
1987         return fasync_helper(fd, file, on, &fc->fasync);
1988 }
1989
1990 const struct file_operations fuse_dev_operations = {
1991         .owner          = THIS_MODULE,
1992         .llseek         = no_llseek,
1993         .read           = do_sync_read,
1994         .aio_read       = fuse_dev_read,
1995         .splice_read    = fuse_dev_splice_read,
1996         .write          = do_sync_write,
1997         .aio_write      = fuse_dev_write,
1998         .splice_write   = fuse_dev_splice_write,
1999         .poll           = fuse_dev_poll,
2000         .release        = fuse_dev_release,
2001         .fasync         = fuse_dev_fasync,
2002 };
2003 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2004
2005 static struct miscdevice fuse_miscdevice = {
2006         .minor = FUSE_MINOR,
2007         .name  = "fuse",
2008         .fops = &fuse_dev_operations,
2009 };
2010
2011 int __init fuse_dev_init(void)
2012 {
2013         int err = -ENOMEM;
2014         fuse_req_cachep = kmem_cache_create("fuse_request",
2015                                             sizeof(struct fuse_req),
2016                                             0, 0, NULL);
2017         if (!fuse_req_cachep)
2018                 goto out;
2019
2020         err = misc_register(&fuse_miscdevice);
2021         if (err)
2022                 goto out_cache_clean;
2023
2024         return 0;
2025
2026  out_cache_clean:
2027         kmem_cache_destroy(fuse_req_cachep);
2028  out:
2029         return err;
2030 }
2031
2032 void fuse_dev_cleanup(void)
2033 {
2034         misc_deregister(&fuse_miscdevice);
2035         kmem_cache_destroy(fuse_req_cachep);
2036 }