fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2006  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
  20 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  21
  22 static struct kmem_cache *fuse_req_cachep;
  23
  24 static struct fuse_conn *fuse_get_conn(struct file *file)
  25 {
  26         /*
  27          * Lockless access is OK, because file->private data is set
  28          * once during mount and is valid until the file is released.
  29          */
  30         return file->private_data;
  31 }
  32
  33 static void fuse_request_init(struct fuse_req *req)
  34 {
  35         memset(req, 0, sizeof(*req));
  36         INIT_LIST_HEAD(&req->list);
  37         INIT_LIST_HEAD(&req->intr_entry);
  38         init_waitqueue_head(&req->waitq);
  39         atomic_set(&req->count, 1);
  40 }
  41
  42 struct fuse_req *fuse_request_alloc(void)
  43 {
  44         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_KERNEL);
  45         if (req)
  46                 fuse_request_init(req);
  47         return req;
  48 }
  49
  50 void fuse_request_free(struct fuse_req *req)
  51 {
  52         kmem_cache_free(fuse_req_cachep, req);
  53 }
  54
  55 static void block_sigs(sigset_t *oldset)
  56 {
  57         sigset_t mask;
  58
  59         siginitsetinv(&mask, sigmask(SIGKILL));
  60         sigprocmask(SIG_BLOCK, &mask, oldset);
  61 }
  62
  63 static void restore_sigs(sigset_t *oldset)
  64 {
  65         sigprocmask(SIG_SETMASK, oldset, NULL);
  66 }
  67
  68 static void __fuse_get_request(struct fuse_req *req)
  69 {
  70         atomic_inc(&req->count);
  71 }
  72
  73 /* Must be called with > 1 refcount */
  74 static void __fuse_put_request(struct fuse_req *req)
  75 {
  76         BUG_ON(atomic_read(&req->count) < 2);
  77         atomic_dec(&req->count);
  78 }
  79
  80 static void fuse_req_init_context(struct fuse_req *req)
  81 {
  82         req->in.h.uid = current->fsuid;
  83         req->in.h.gid = current->fsgid;
  84         req->in.h.pid = current->pid;
  85 }
  86
  87 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
  88 {
  89         struct fuse_req *req;
  90         sigset_t oldset;
  91         int intr;
  92         int err;
  93
  94         atomic_inc(&fc->num_waiting);
  95         block_sigs(&oldset);
  96         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
  97         restore_sigs(&oldset);
  98         err = -EINTR;
  99         if (intr)
 100                 goto out;
 101
 102         err = -ENOTCONN;
 103         if (!fc->connected)
 104                 goto out;
 105
 106         req = fuse_request_alloc();
 107         err = -ENOMEM;
 108         if (!req)
 109                 goto out;
 110
 111         fuse_req_init_context(req);
 112         req->waiting = 1;
 113         return req;
 114
 115  out:
 116         atomic_dec(&fc->num_waiting);
 117         return ERR_PTR(err);
 118 }
 119
 120 /*
 121  * Return request in fuse_file->reserved_req.  However that may
 122  * currently be in use.  If that is the case, wait for it to become
 123  * available.
 124  */
 125 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 126                                          struct file *file)
 127 {
 128         struct fuse_req *req = NULL;
 129         struct fuse_file *ff = file->private_data;
 130
 131         do {
 132                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 133                 spin_lock(&fc->lock);
 134                 if (ff->reserved_req) {
 135                         req = ff->reserved_req;
 136                         ff->reserved_req = NULL;
 137                         get_file(file);
 138                         req->stolen_file = file;
 139                 }
 140                 spin_unlock(&fc->lock);
 141         } while (!req);
 142
 143         return req;
 144 }
 145
 146 /*
 147  * Put stolen request back into fuse_file->reserved_req
 148  */
 149 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 150 {
 151         struct file *file = req->stolen_file;
 152         struct fuse_file *ff = file->private_data;
 153
 154         spin_lock(&fc->lock);
 155         fuse_request_init(req);
 156         BUG_ON(ff->reserved_req);
 157         ff->reserved_req = req;
 158         wake_up_all(&fc->reserved_req_waitq);
 159         spin_unlock(&fc->lock);
 160         fput(file);
 161 }
 162
 163 /*
 164  * Gets a requests for a file operation, always succeeds
 165  *
 166  * This is used for sending the FLUSH request, which must get to
 167  * userspace, due to POSIX locks which may need to be unlocked.
 168  *
 169  * If allocation fails due to OOM, use the reserved request in
 170  * fuse_file.
 171  *
 172  * This is very unlikely to deadlock accidentally, since the
 173  * filesystem should not have it's own file open.  If deadlock is
 174  * intentional, it can still be broken by "aborting" the filesystem.
 175  */
 176 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 177 {
 178         struct fuse_req *req;
 179
 180         atomic_inc(&fc->num_waiting);
 181         wait_event(fc->blocked_waitq, !fc->blocked);
 182         req = fuse_request_alloc();
 183         if (!req)
 184                 req = get_reserved_req(fc, file);
 185
 186         fuse_req_init_context(req);
 187         req->waiting = 1;
 188         return req;
 189 }
 190
 191 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 192 {
 193         if (atomic_dec_and_test(&req->count)) {
 194                 if (req->waiting)
 195                         atomic_dec(&fc->num_waiting);
 196
 197                 if (req->stolen_file)
 198                         put_reserved_req(fc, req);
 199                 else
 200                         fuse_request_free(req);
 201         }
 202 }
 203
 204 /*
 205  * This function is called when a request is finished.  Either a reply
 206  * has arrived or it was aborted (and not yet sent) or some error
 207  * occurred during communication with userspace, or the device file
 208  * was closed.  The requester thread is woken up (if still waiting),
 209  * the 'end' callback is called if given, else the reference to the
 210  * request is released
 211  *
 212  * Called with fc->lock, unlocks it
 213  */
 214 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 215         __releases(fc->lock)
 216 {
 217         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 218         req->end = NULL;
 219         list_del(&req->list);
 220         list_del(&req->intr_entry);
 221         req->state = FUSE_REQ_FINISHED;
 222         if (req->background) {
 223                 if (fc->num_background == FUSE_MAX_BACKGROUND) {
 224                         fc->blocked = 0;
 225                         wake_up_all(&fc->blocked_waitq);
 226                 }
 227                 if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 228                         clear_bdi_congested(&fc->bdi, READ);
 229                         clear_bdi_congested(&fc->bdi, WRITE);
 230                 }
 231                 fc->num_background--;
 232         }
 233         spin_unlock(&fc->lock);
 234         wake_up(&req->waitq);
 235         if (end)
 236                 end(fc, req);
 237         else
 238                 fuse_put_request(fc, req);
 239 }
 240
 241 static void wait_answer_interruptible(struct fuse_conn *fc,
 242                                       struct fuse_req *req)
 243 {
 244         if (signal_pending(current))
 245                 return;
 246
 247         spin_unlock(&fc->lock);
 248         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 249         spin_lock(&fc->lock);
 250 }
 251
 252 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 253 {
 254         list_add_tail(&req->intr_entry, &fc->interrupts);
 255         wake_up(&fc->waitq);
 256         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 257 }
 258
 259 /* Called with fc->lock held.  Releases, and then reacquires it. */
 260 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 261 {
 262         if (!fc->no_interrupt) {
 263                 /* Any signal may interrupt this */
 264                 wait_answer_interruptible(fc, req);
 265
 266                 if (req->aborted)
 267                         goto aborted;
 268                 if (req->state == FUSE_REQ_FINISHED)
 269                         return;
 270
 271                 req->interrupted = 1;
 272                 if (req->state == FUSE_REQ_SENT)
 273                         queue_interrupt(fc, req);
 274         }
 275
 276         if (req->force) {
 277                 spin_unlock(&fc->lock);
 278                 wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 279                 spin_lock(&fc->lock);
 280         } else {
 281                 sigset_t oldset;
 282
 283                 /* Only fatal signals may interrupt this */
 284                 block_sigs(&oldset);
 285                 wait_answer_interruptible(fc, req);
 286                 restore_sigs(&oldset);
 287         }
 288
 289         if (req->aborted)
 290                 goto aborted;
 291         if (req->state == FUSE_REQ_FINISHED)
 292                 return;
 293
 294         req->out.h.error = -EINTR;
 295         req->aborted = 1;
 296
 297  aborted:
 298         if (req->locked) {
 299                 /* This is uninterruptible sleep, because data is
 300                    being copied to/from the buffers of req.  During
 301                    locked state, there mustn't be any filesystem
 302                    operation (e.g. page fault), since that could lead
 303                    to deadlock */
 304                 spin_unlock(&fc->lock);
 305                 wait_event(req->waitq, !req->locked);
 306                 spin_lock(&fc->lock);
 307         }
 308         if (req->state == FUSE_REQ_PENDING) {
 309                 list_del(&req->list);
 310                 __fuse_put_request(req);
 311         } else if (req->state == FUSE_REQ_SENT) {
 312                 spin_unlock(&fc->lock);
 313                 wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 314                 spin_lock(&fc->lock);
 315         }
 316 }
 317
 318 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 319 {
 320         unsigned nbytes = 0;
 321         unsigned i;
 322
 323         for (i = 0; i < numargs; i++)
 324                 nbytes += args[i].size;
 325
 326         return nbytes;
 327 }
 328
 329 static u64 fuse_get_unique(struct fuse_conn *fc)
 330  {
 331         fc->reqctr++;
 332         /* zero is special */
 333         if (fc->reqctr == 0)
 334                 fc->reqctr = 1;
 335
 336         return fc->reqctr;
 337 }
 338
 339 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 340 {
 341         req->in.h.unique = fuse_get_unique(fc);
 342         req->in.h.len = sizeof(struct fuse_in_header) +
 343                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 344         list_add_tail(&req->list, &fc->pending);
 345         req->state = FUSE_REQ_PENDING;
 346         if (!req->waiting) {
 347                 req->waiting = 1;
 348                 atomic_inc(&fc->num_waiting);
 349         }
 350         wake_up(&fc->waitq);
 351         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 352 }
 353
 354 void request_send(struct fuse_conn *fc, struct fuse_req *req)
 355 {
 356         req->isreply = 1;
 357         spin_lock(&fc->lock);
 358         if (!fc->connected)
 359                 req->out.h.error = -ENOTCONN;
 360         else if (fc->conn_error)
 361                 req->out.h.error = -ECONNREFUSED;
 362         else {
 363                 queue_request(fc, req);
 364                 /* acquire extra reference, since request is still needed
 365                    after request_end() */
 366                 __fuse_get_request(req);
 367
 368                 request_wait_answer(fc, req);
 369         }
 370         spin_unlock(&fc->lock);
 371 }
 372
 373 static void request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 374 {
 375         spin_lock(&fc->lock);
 376         if (fc->connected) {
 377                 req->background = 1;
 378                 fc->num_background++;
 379                 if (fc->num_background == FUSE_MAX_BACKGROUND)
 380                         fc->blocked = 1;
 381                 if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 382                         set_bdi_congested(&fc->bdi, READ);
 383                         set_bdi_congested(&fc->bdi, WRITE);
 384                 }
 385
 386                 queue_request(fc, req);
 387                 spin_unlock(&fc->lock);
 388         } else {
 389                 req->out.h.error = -ENOTCONN;
 390                 request_end(fc, req);
 391         }
 392 }
 393
 394 void request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 395 {
 396         req->isreply = 0;
 397         request_send_nowait(fc, req);
 398 }
 399
 400 void request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 401 {
 402         req->isreply = 1;
 403         request_send_nowait(fc, req);
 404 }
 405
 406 /*
 407  * Lock the request.  Up to the next unlock_request() there mustn't be
 408  * anything that could cause a page-fault.  If the request was already
 409  * aborted bail out.
 410  */
 411 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 412 {
 413         int err = 0;
 414         if (req) {
 415                 spin_lock(&fc->lock);
 416                 if (req->aborted)
 417                         err = -ENOENT;
 418                 else
 419                         req->locked = 1;
 420                 spin_unlock(&fc->lock);
 421         }
 422         return err;
 423 }
 424
 425 /*
 426  * Unlock request.  If it was aborted during being locked, the
 427  * requester thread is currently waiting for it to be unlocked, so
 428  * wake it up.
 429  */
 430 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 431 {
 432         if (req) {
 433                 spin_lock(&fc->lock);
 434                 req->locked = 0;
 435                 if (req->aborted)
 436                         wake_up(&req->waitq);
 437                 spin_unlock(&fc->lock);
 438         }
 439 }
 440
 441 struct fuse_copy_state {
 442         struct fuse_conn *fc;
 443         int write;
 444         struct fuse_req *req;
 445         const struct iovec *iov;
 446         unsigned long nr_segs;
 447         unsigned long seglen;
 448         unsigned long addr;
 449         struct page *pg;
 450         void *mapaddr;
 451         void *buf;
 452         unsigned len;
 453 };
 454
 455 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 456                            int write, struct fuse_req *req,
 457                            const struct iovec *iov, unsigned long nr_segs)
 458 {
 459         memset(cs, 0, sizeof(*cs));
 460         cs->fc = fc;
 461         cs->write = write;
 462         cs->req = req;
 463         cs->iov = iov;
 464         cs->nr_segs = nr_segs;
 465 }
 466
 467 /* Unmap and put previous page of userspace buffer */
 468 static void fuse_copy_finish(struct fuse_copy_state *cs)
 469 {
 470         if (cs->mapaddr) {
 471                 kunmap_atomic(cs->mapaddr, KM_USER0);
 472                 if (cs->write) {
 473                         flush_dcache_page(cs->pg);
 474                         set_page_dirty_lock(cs->pg);
 475                 }
 476                 put_page(cs->pg);
 477                 cs->mapaddr = NULL;
 478         }
 479 }
 480
 481 /*
 482  * Get another pagefull of userspace buffer, and map it to kernel
 483  * address space, and lock request
 484  */
 485 static int fuse_copy_fill(struct fuse_copy_state *cs)
 486 {
 487         unsigned long offset;
 488         int err;
 489
 490         unlock_request(cs->fc, cs->req);
 491         fuse_copy_finish(cs);
 492         if (!cs->seglen) {
 493                 BUG_ON(!cs->nr_segs);
 494                 cs->seglen = cs->iov[0].iov_len;
 495                 cs->addr = (unsigned long) cs->iov[0].iov_base;
 496                 cs->iov ++;
 497                 cs->nr_segs --;
 498         }
 499         down_read(&current->mm->mmap_sem);
 500         err = get_user_pages(current, current->mm, cs->addr, 1, cs->write, 0,
 501                              &cs->pg, NULL);
 502         up_read(&current->mm->mmap_sem);
 503         if (err < 0)
 504                 return err;
 505         BUG_ON(err != 1);
 506         offset = cs->addr % PAGE_SIZE;
 507         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
 508         cs->buf = cs->mapaddr + offset;
 509         cs->len = min(PAGE_SIZE - offset, cs->seglen);
 510         cs->seglen -= cs->len;
 511         cs->addr += cs->len;
 512
 513         return lock_request(cs->fc, cs->req);
 514 }
 515
 516 /* Do as much copy to/from userspace buffer as we can */
 517 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 518 {
 519         unsigned ncpy = min(*size, cs->len);
 520         if (val) {
 521                 if (cs->write)
 522                         memcpy(cs->buf, *val, ncpy);
 523                 else
 524                         memcpy(*val, cs->buf, ncpy);
 525                 *val += ncpy;
 526         }
 527         *size -= ncpy;
 528         cs->len -= ncpy;
 529         cs->buf += ncpy;
 530         return ncpy;
 531 }
 532
 533 /*
 534  * Copy a page in the request to/from the userspace buffer.  Must be
 535  * done atomically
 536  */
 537 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
 538                           unsigned offset, unsigned count, int zeroing)
 539 {
 540         if (page && zeroing && count < PAGE_SIZE) {
 541                 void *mapaddr = kmap_atomic(page, KM_USER1);
 542                 memset(mapaddr, 0, PAGE_SIZE);
 543                 kunmap_atomic(mapaddr, KM_USER1);
 544         }
 545         while (count) {
 546                 int err;
 547                 if (!cs->len && (err = fuse_copy_fill(cs)))
 548                         return err;
 549                 if (page) {
 550                         void *mapaddr = kmap_atomic(page, KM_USER1);
 551                         void *buf = mapaddr + offset;
 552                         offset += fuse_copy_do(cs, &buf, &count);
 553                         kunmap_atomic(mapaddr, KM_USER1);
 554                 } else
 555                         offset += fuse_copy_do(cs, NULL, &count);
 556         }
 557         if (page && !cs->write)
 558                 flush_dcache_page(page);
 559         return 0;
 560 }
 561
 562 /* Copy pages in the request to/from userspace buffer */
 563 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 564                            int zeroing)
 565 {
 566         unsigned i;
 567         struct fuse_req *req = cs->req;
 568         unsigned offset = req->page_offset;
 569         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 570
 571         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 572                 struct page *page = req->pages[i];
 573                 int err = fuse_copy_page(cs, page, offset, count, zeroing);
 574                 if (err)
 575                         return err;
 576
 577                 nbytes -= count;
 578                 count = min(nbytes, (unsigned) PAGE_SIZE);
 579                 offset = 0;
 580         }
 581         return 0;
 582 }
 583
 584 /* Copy a single argument in the request to/from userspace buffer */
 585 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 586 {
 587         while (size) {
 588                 int err;
 589                 if (!cs->len && (err = fuse_copy_fill(cs)))
 590                         return err;
 591                 fuse_copy_do(cs, &val, &size);
 592         }
 593         return 0;
 594 }
 595
 596 /* Copy request arguments to/from userspace buffer */
 597 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 598                           unsigned argpages, struct fuse_arg *args,
 599                           int zeroing)
 600 {
 601         int err = 0;
 602         unsigned i;
 603
 604         for (i = 0; !err && i < numargs; i++)  {
 605                 struct fuse_arg *arg = &args[i];
 606                 if (i == numargs - 1 && argpages)
 607                         err = fuse_copy_pages(cs, arg->size, zeroing);
 608                 else
 609                         err = fuse_copy_one(cs, arg->value, arg->size);
 610         }
 611         return err;
 612 }
 613
 614 static int request_pending(struct fuse_conn *fc)
 615 {
 616         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 617 }
 618
 619 /* Wait until a request is available on the pending list */
 620 static void request_wait(struct fuse_conn *fc)
 621 {
 622         DECLARE_WAITQUEUE(wait, current);
 623
 624         add_wait_queue_exclusive(&fc->waitq, &wait);
 625         while (fc->connected && !request_pending(fc)) {
 626                 set_current_state(TASK_INTERRUPTIBLE);
 627                 if (signal_pending(current))
 628                         break;
 629
 630                 spin_unlock(&fc->lock);
 631                 schedule();
 632                 spin_lock(&fc->lock);
 633         }
 634         set_current_state(TASK_RUNNING);
 635         remove_wait_queue(&fc->waitq, &wait);
 636 }
 637
 638 /*
 639  * Transfer an interrupt request to userspace
 640  *
 641  * Unlike other requests this is assembled on demand, without a need
 642  * to allocate a separate fuse_req structure.
 643  *
 644  * Called with fc->lock held, releases it
 645  */
 646 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
 647                                const struct iovec *iov, unsigned long nr_segs)
 648         __releases(fc->lock)
 649 {
 650         struct fuse_copy_state cs;
 651         struct fuse_in_header ih;
 652         struct fuse_interrupt_in arg;
 653         unsigned reqsize = sizeof(ih) + sizeof(arg);
 654         int err;
 655
 656         list_del_init(&req->intr_entry);
 657         req->intr_unique = fuse_get_unique(fc);
 658         memset(&ih, 0, sizeof(ih));
 659         memset(&arg, 0, sizeof(arg));
 660         ih.len = reqsize;
 661         ih.opcode = FUSE_INTERRUPT;
 662         ih.unique = req->intr_unique;
 663         arg.unique = req->in.h.unique;
 664
 665         spin_unlock(&fc->lock);
 666         if (iov_length(iov, nr_segs) < reqsize)
 667                 return -EINVAL;
 668
 669         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
 670         err = fuse_copy_one(&cs, &ih, sizeof(ih));
 671         if (!err)
 672                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
 673         fuse_copy_finish(&cs);
 674
 675         return err ? err : reqsize;
 676 }
 677
 678 /*
 679  * Read a single request into the userspace filesystem's buffer.  This
 680  * function waits until a request is available, then removes it from
 681  * the pending list and copies request data to userspace buffer.  If
 682  * no reply is needed (FORGET) or request has been aborted or there
 683  * was an error during the copying then it's finished by calling
 684  * request_end().  Otherwise add it to the processing list, and set
 685  * the 'sent' flag.
 686  */
 687 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
 688                               unsigned long nr_segs, loff_t pos)
 689 {
 690         int err;
 691         struct fuse_req *req;
 692         struct fuse_in *in;
 693         struct fuse_copy_state cs;
 694         unsigned reqsize;
 695         struct file *file = iocb->ki_filp;
 696         struct fuse_conn *fc = fuse_get_conn(file);
 697         if (!fc)
 698                 return -EPERM;
 699
 700  restart:
 701         spin_lock(&fc->lock);
 702         err = -EAGAIN;
 703         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 704             !request_pending(fc))
 705                 goto err_unlock;
 706
 707         request_wait(fc);
 708         err = -ENODEV;
 709         if (!fc->connected)
 710                 goto err_unlock;
 711         err = -ERESTARTSYS;
 712         if (!request_pending(fc))
 713                 goto err_unlock;
 714
 715         if (!list_empty(&fc->interrupts)) {
 716                 req = list_entry(fc->interrupts.next, struct fuse_req,
 717                                  intr_entry);
 718                 return fuse_read_interrupt(fc, req, iov, nr_segs);
 719         }
 720
 721         req = list_entry(fc->pending.next, struct fuse_req, list);
 722         req->state = FUSE_REQ_READING;
 723         list_move(&req->list, &fc->io);
 724
 725         in = &req->in;
 726         reqsize = in->h.len;
 727         /* If request is too large, reply with an error and restart the read */
 728         if (iov_length(iov, nr_segs) < reqsize) {
 729                 req->out.h.error = -EIO;
 730                 /* SETXATTR is special, since it may contain too large data */
 731                 if (in->h.opcode == FUSE_SETXATTR)
 732                         req->out.h.error = -E2BIG;
 733                 request_end(fc, req);
 734                 goto restart;
 735         }
 736         spin_unlock(&fc->lock);
 737         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
 738         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
 739         if (!err)
 740                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
 741                                      (struct fuse_arg *) in->args, 0);
 742         fuse_copy_finish(&cs);
 743         spin_lock(&fc->lock);
 744         req->locked = 0;
 745         if (!err && req->aborted)
 746                 err = -ENOENT;
 747         if (err) {
 748                 if (!req->aborted)
 749                         req->out.h.error = -EIO;
 750                 request_end(fc, req);
 751                 return err;
 752         }
 753         if (!req->isreply)
 754                 request_end(fc, req);
 755         else {
 756                 req->state = FUSE_REQ_SENT;
 757                 list_move_tail(&req->list, &fc->processing);
 758                 if (req->interrupted)
 759                         queue_interrupt(fc, req);
 760                 spin_unlock(&fc->lock);
 761         }
 762         return reqsize;
 763
 764  err_unlock:
 765         spin_unlock(&fc->lock);
 766         return err;
 767 }
 768
 769 /* Look up request on processing list by unique ID */
 770 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
 771 {
 772         struct list_head *entry;
 773
 774         list_for_each(entry, &fc->processing) {
 775                 struct fuse_req *req;
 776                 req = list_entry(entry, struct fuse_req, list);
 777                 if (req->in.h.unique == unique || req->intr_unique == unique)
 778                         return req;
 779         }
 780         return NULL;
 781 }
 782
 783 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
 784                          unsigned nbytes)
 785 {
 786         unsigned reqsize = sizeof(struct fuse_out_header);
 787
 788         if (out->h.error)
 789                 return nbytes != reqsize ? -EINVAL : 0;
 790
 791         reqsize += len_args(out->numargs, out->args);
 792
 793         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
 794                 return -EINVAL;
 795         else if (reqsize > nbytes) {
 796                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
 797                 unsigned diffsize = reqsize - nbytes;
 798                 if (diffsize > lastarg->size)
 799                         return -EINVAL;
 800                 lastarg->size -= diffsize;
 801         }
 802         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
 803                               out->page_zeroing);
 804 }
 805
 806 /*
 807  * Write a single reply to a request.  First the header is copied from
 808  * the write buffer.  The request is then searched on the processing
 809  * list by the unique ID found in the header.  If found, then remove
 810  * it from the list and copy the rest of the buffer to the request.
 811  * The request is finished by calling request_end()
 812  */
 813 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
 814                                unsigned long nr_segs, loff_t pos)
 815 {
 816         int err;
 817         unsigned nbytes = iov_length(iov, nr_segs);
 818         struct fuse_req *req;
 819         struct fuse_out_header oh;
 820         struct fuse_copy_state cs;
 821         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
 822         if (!fc)
 823                 return -EPERM;
 824
 825         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
 826         if (nbytes < sizeof(struct fuse_out_header))
 827                 return -EINVAL;
 828
 829         err = fuse_copy_one(&cs, &oh, sizeof(oh));
 830         if (err)
 831                 goto err_finish;
 832         err = -EINVAL;
 833         if (!oh.unique || oh.error <= -1000 || oh.error > 0 ||
 834             oh.len != nbytes)
 835                 goto err_finish;
 836
 837         spin_lock(&fc->lock);
 838         err = -ENOENT;
 839         if (!fc->connected)
 840                 goto err_unlock;
 841
 842         req = request_find(fc, oh.unique);
 843         if (!req)
 844                 goto err_unlock;
 845
 846         if (req->aborted) {
 847                 spin_unlock(&fc->lock);
 848                 fuse_copy_finish(&cs);
 849                 spin_lock(&fc->lock);
 850                 request_end(fc, req);
 851                 return -ENOENT;
 852         }
 853         /* Is it an interrupt reply? */
 854         if (req->intr_unique == oh.unique) {
 855                 err = -EINVAL;
 856                 if (nbytes != sizeof(struct fuse_out_header))
 857                         goto err_unlock;
 858
 859                 if (oh.error == -ENOSYS)
 860                         fc->no_interrupt = 1;
 861                 else if (oh.error == -EAGAIN)
 862                         queue_interrupt(fc, req);
 863
 864                 spin_unlock(&fc->lock);
 865                 fuse_copy_finish(&cs);
 866                 return nbytes;
 867         }
 868
 869         req->state = FUSE_REQ_WRITING;
 870         list_move(&req->list, &fc->io);
 871         req->out.h = oh;
 872         req->locked = 1;
 873         cs.req = req;
 874         spin_unlock(&fc->lock);
 875
 876         err = copy_out_args(&cs, &req->out, nbytes);
 877         fuse_copy_finish(&cs);
 878
 879         spin_lock(&fc->lock);
 880         req->locked = 0;
 881         if (!err) {
 882                 if (req->aborted)
 883                         err = -ENOENT;
 884         } else if (!req->aborted)
 885                 req->out.h.error = -EIO;
 886         request_end(fc, req);
 887
 888         return err ? err : nbytes;
 889
 890  err_unlock:
 891         spin_unlock(&fc->lock);
 892  err_finish:
 893         fuse_copy_finish(&cs);
 894         return err;
 895 }
 896
 897 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
 898 {
 899         unsigned mask = POLLOUT | POLLWRNORM;
 900         struct fuse_conn *fc = fuse_get_conn(file);
 901         if (!fc)
 902                 return POLLERR;
 903
 904         poll_wait(file, &fc->waitq, wait);
 905
 906         spin_lock(&fc->lock);
 907         if (!fc->connected)
 908                 mask = POLLERR;
 909         else if (request_pending(fc))
 910                 mask |= POLLIN | POLLRDNORM;
 911         spin_unlock(&fc->lock);
 912
 913         return mask;
 914 }
 915
 916 /*
 917  * Abort all requests on the given list (pending or processing)
 918  *
 919  * This function releases and reacquires fc->lock
 920  */
 921 static void end_requests(struct fuse_conn *fc, struct list_head *head)
 922 {
 923         while (!list_empty(head)) {
 924                 struct fuse_req *req;
 925                 req = list_entry(head->next, struct fuse_req, list);
 926                 req->out.h.error = -ECONNABORTED;
 927                 request_end(fc, req);
 928                 spin_lock(&fc->lock);
 929         }
 930 }
 931
 932 /*
 933  * Abort requests under I/O
 934  *
 935  * The requests are set to aborted and finished, and the request
 936  * waiter is woken up.  This will make request_wait_answer() wait
 937  * until the request is unlocked and then return.
 938  *
 939  * If the request is asynchronous, then the end function needs to be
 940  * called after waiting for the request to be unlocked (if it was
 941  * locked).
 942  */
 943 static void end_io_requests(struct fuse_conn *fc)
 944 {
 945         while (!list_empty(&fc->io)) {
 946                 struct fuse_req *req =
 947                         list_entry(fc->io.next, struct fuse_req, list);
 948                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 949
 950                 req->aborted = 1;
 951                 req->out.h.error = -ECONNABORTED;
 952                 req->state = FUSE_REQ_FINISHED;
 953                 list_del_init(&req->list);
 954                 wake_up(&req->waitq);
 955                 if (end) {
 956                         req->end = NULL;
 957                         /* The end function will consume this reference */
 958                         __fuse_get_request(req);
 959                         spin_unlock(&fc->lock);
 960                         wait_event(req->waitq, !req->locked);
 961                         end(fc, req);
 962                         spin_lock(&fc->lock);
 963                 }
 964         }
 965 }
 966
 967 /*
 968  * Abort all requests.
 969  *
 970  * Emergency exit in case of a malicious or accidental deadlock, or
 971  * just a hung filesystem.
 972  *
 973  * The same effect is usually achievable through killing the
 974  * filesystem daemon and all users of the filesystem.  The exception
 975  * is the combination of an asynchronous request and the tricky
 976  * deadlock (see Documentation/filesystems/fuse.txt).
 977  *
 978  * During the aborting, progression of requests from the pending and
 979  * processing lists onto the io list, and progression of new requests
 980  * onto the pending list is prevented by req->connected being false.
 981  *
 982  * Progression of requests under I/O to the processing list is
 983  * prevented by the req->aborted flag being true for these requests.
 984  * For this reason requests on the io list must be aborted first.
 985  */
 986 void fuse_abort_conn(struct fuse_conn *fc)
 987 {
 988         spin_lock(&fc->lock);
 989         if (fc->connected) {
 990                 fc->connected = 0;
 991                 fc->blocked = 0;
 992                 end_io_requests(fc);
 993                 end_requests(fc, &fc->pending);
 994                 end_requests(fc, &fc->processing);
 995                 wake_up_all(&fc->waitq);
 996                 wake_up_all(&fc->blocked_waitq);
 997                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 998         }
 999         spin_unlock(&fc->lock);
1000 }
1001
1002 static int fuse_dev_release(struct inode *inode, struct file *file)
1003 {
1004         struct fuse_conn *fc = fuse_get_conn(file);
1005         if (fc) {
1006                 spin_lock(&fc->lock);
1007                 fc->connected = 0;
1008                 end_requests(fc, &fc->pending);
1009                 end_requests(fc, &fc->processing);
1010                 spin_unlock(&fc->lock);
1011                 fasync_helper(-1, file, 0, &fc->fasync);
1012                 fuse_conn_put(fc);
1013         }
1014
1015         return 0;
1016 }
1017
1018 static int fuse_dev_fasync(int fd, struct file *file, int on)
1019 {
1020         struct fuse_conn *fc = fuse_get_conn(file);
1021         if (!fc)
1022                 return -EPERM;
1023
1024         /* No locking - fasync_helper does its own locking */
1025         return fasync_helper(fd, file, on, &fc->fasync);
1026 }
1027
1028 const struct file_operations fuse_dev_operations = {
1029         .owner          = THIS_MODULE,
1030         .llseek         = no_llseek,
1031         .read           = do_sync_read,
1032         .aio_read       = fuse_dev_read,
1033         .write          = do_sync_write,
1034         .aio_write      = fuse_dev_write,
1035         .poll           = fuse_dev_poll,
1036         .release        = fuse_dev_release,
1037         .fasync         = fuse_dev_fasync,
1038 };
1039
1040 static struct miscdevice fuse_miscdevice = {
1041         .minor = FUSE_MINOR,
1042         .name  = "fuse",
1043         .fops = &fuse_dev_operations,
1044 };
1045
1046 int __init fuse_dev_init(void)
1047 {
1048         int err = -ENOMEM;
1049         fuse_req_cachep = kmem_cache_create("fuse_request",
1050                                             sizeof(struct fuse_req),
1051                                             0, 0, NULL);
1052         if (!fuse_req_cachep)
1053                 goto out;
1054
1055         err = misc_register(&fuse_miscdevice);
1056         if (err)
1057                 goto out_cache_clean;
1058
1059         return 0;
1060
1061  out_cache_clean:
1062         kmem_cache_destroy(fuse_req_cachep);
1063  out:
1064         return err;
1065 }
1066
1067 void fuse_dev_cleanup(void)
1068 {
1069         misc_deregister(&fuse_miscdevice);
1070         kmem_cache_destroy(fuse_req_cachep);
1071 }