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 struct fuse_req *fuse_request_alloc_nofs(void)
  51 {
  52         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_NOFS);
  53         if (req)
  54                 fuse_request_init(req);
  55         return req;
  56 }
  57
  58 void fuse_request_free(struct fuse_req *req)
  59 {
  60         kmem_cache_free(fuse_req_cachep, req);
  61 }
  62
  63 static void block_sigs(sigset_t *oldset)
  64 {
  65         sigset_t mask;
  66
  67         siginitsetinv(&mask, sigmask(SIGKILL));
  68         sigprocmask(SIG_BLOCK, &mask, oldset);
  69 }
  70
  71 static void restore_sigs(sigset_t *oldset)
  72 {
  73         sigprocmask(SIG_SETMASK, oldset, NULL);
  74 }
  75
  76 static void __fuse_get_request(struct fuse_req *req)
  77 {
  78         atomic_inc(&req->count);
  79 }
  80
  81 /* Must be called with > 1 refcount */
  82 static void __fuse_put_request(struct fuse_req *req)
  83 {
  84         BUG_ON(atomic_read(&req->count) < 2);
  85         atomic_dec(&req->count);
  86 }
  87
  88 static void fuse_req_init_context(struct fuse_req *req)
  89 {
  90         req->in.h.uid = current->fsuid;
  91         req->in.h.gid = current->fsgid;
  92         req->in.h.pid = current->pid;
  93 }
  94
  95 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
  96 {
  97         struct fuse_req *req;
  98         sigset_t oldset;
  99         int intr;
 100         int err;
 101
 102         atomic_inc(&fc->num_waiting);
 103         block_sigs(&oldset);
 104         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
 105         restore_sigs(&oldset);
 106         err = -EINTR;
 107         if (intr)
 108                 goto out;
 109
 110         err = -ENOTCONN;
 111         if (!fc->connected)
 112                 goto out;
 113
 114         req = fuse_request_alloc();
 115         err = -ENOMEM;
 116         if (!req)
 117                 goto out;
 118
 119         fuse_req_init_context(req);
 120         req->waiting = 1;
 121         return req;
 122
 123  out:
 124         atomic_dec(&fc->num_waiting);
 125         return ERR_PTR(err);
 126 }
 127
 128 /*
 129  * Return request in fuse_file->reserved_req.  However that may
 130  * currently be in use.  If that is the case, wait for it to become
 131  * available.
 132  */
 133 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 134                                          struct file *file)
 135 {
 136         struct fuse_req *req = NULL;
 137         struct fuse_file *ff = file->private_data;
 138
 139         do {
 140                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 141                 spin_lock(&fc->lock);
 142                 if (ff->reserved_req) {
 143                         req = ff->reserved_req;
 144                         ff->reserved_req = NULL;
 145                         get_file(file);
 146                         req->stolen_file = file;
 147                 }
 148                 spin_unlock(&fc->lock);
 149         } while (!req);
 150
 151         return req;
 152 }
 153
 154 /*
 155  * Put stolen request back into fuse_file->reserved_req
 156  */
 157 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 158 {
 159         struct file *file = req->stolen_file;
 160         struct fuse_file *ff = file->private_data;
 161
 162         spin_lock(&fc->lock);
 163         fuse_request_init(req);
 164         BUG_ON(ff->reserved_req);
 165         ff->reserved_req = req;
 166         wake_up_all(&fc->reserved_req_waitq);
 167         spin_unlock(&fc->lock);
 168         fput(file);
 169 }
 170
 171 /*
 172  * Gets a requests for a file operation, always succeeds
 173  *
 174  * This is used for sending the FLUSH request, which must get to
 175  * userspace, due to POSIX locks which may need to be unlocked.
 176  *
 177  * If allocation fails due to OOM, use the reserved request in
 178  * fuse_file.
 179  *
 180  * This is very unlikely to deadlock accidentally, since the
 181  * filesystem should not have it's own file open.  If deadlock is
 182  * intentional, it can still be broken by "aborting" the filesystem.
 183  */
 184 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 185 {
 186         struct fuse_req *req;
 187
 188         atomic_inc(&fc->num_waiting);
 189         wait_event(fc->blocked_waitq, !fc->blocked);
 190         req = fuse_request_alloc();
 191         if (!req)
 192                 req = get_reserved_req(fc, file);
 193
 194         fuse_req_init_context(req);
 195         req->waiting = 1;
 196         return req;
 197 }
 198
 199 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 200 {
 201         if (atomic_dec_and_test(&req->count)) {
 202                 if (req->waiting)
 203                         atomic_dec(&fc->num_waiting);
 204
 205                 if (req->stolen_file)
 206                         put_reserved_req(fc, req);
 207                 else
 208                         fuse_request_free(req);
 209         }
 210 }
 211
 212 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 213 {
 214         unsigned nbytes = 0;
 215         unsigned i;
 216
 217         for (i = 0; i < numargs; i++)
 218                 nbytes += args[i].size;
 219
 220         return nbytes;
 221 }
 222
 223 static u64 fuse_get_unique(struct fuse_conn *fc)
 224 {
 225         fc->reqctr++;
 226         /* zero is special */
 227         if (fc->reqctr == 0)
 228                 fc->reqctr = 1;
 229
 230         return fc->reqctr;
 231 }
 232
 233 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 234 {
 235         req->in.h.unique = fuse_get_unique(fc);
 236         req->in.h.len = sizeof(struct fuse_in_header) +
 237                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 238         list_add_tail(&req->list, &fc->pending);
 239         req->state = FUSE_REQ_PENDING;
 240         if (!req->waiting) {
 241                 req->waiting = 1;
 242                 atomic_inc(&fc->num_waiting);
 243         }
 244         wake_up(&fc->waitq);
 245         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 246 }
 247
 248 static void flush_bg_queue(struct fuse_conn *fc)
 249 {
 250         while (fc->active_background < FUSE_MAX_BACKGROUND &&
 251                !list_empty(&fc->bg_queue)) {
 252                 struct fuse_req *req;
 253
 254                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 255                 list_del(&req->list);
 256                 fc->active_background++;
 257                 queue_request(fc, req);
 258         }
 259 }
 260
 261 /*
 262  * This function is called when a request is finished.  Either a reply
 263  * has arrived or it was aborted (and not yet sent) or some error
 264  * occurred during communication with userspace, or the device file
 265  * was closed.  The requester thread is woken up (if still waiting),
 266  * the 'end' callback is called if given, else the reference to the
 267  * request is released
 268  *
 269  * Called with fc->lock, unlocks it
 270  */
 271 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 272         __releases(fc->lock)
 273 {
 274         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 275         req->end = NULL;
 276         list_del(&req->list);
 277         list_del(&req->intr_entry);
 278         req->state = FUSE_REQ_FINISHED;
 279         if (req->background) {
 280                 if (fc->num_background == FUSE_MAX_BACKGROUND) {
 281                         fc->blocked = 0;
 282                         wake_up_all(&fc->blocked_waitq);
 283                 }
 284                 if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 285                         clear_bdi_congested(&fc->bdi, READ);
 286                         clear_bdi_congested(&fc->bdi, WRITE);
 287                 }
 288                 fc->num_background--;
 289                 fc->active_background--;
 290                 flush_bg_queue(fc);
 291         }
 292         spin_unlock(&fc->lock);
 293         wake_up(&req->waitq);
 294         if (end)
 295                 end(fc, req);
 296         else
 297                 fuse_put_request(fc, req);
 298 }
 299
 300 static void wait_answer_interruptible(struct fuse_conn *fc,
 301                                       struct fuse_req *req)
 302         __releases(fc->lock) __acquires(fc->lock)
 303 {
 304         if (signal_pending(current))
 305                 return;
 306
 307         spin_unlock(&fc->lock);
 308         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 309         spin_lock(&fc->lock);
 310 }
 311
 312 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 313 {
 314         list_add_tail(&req->intr_entry, &fc->interrupts);
 315         wake_up(&fc->waitq);
 316         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 317 }
 318
 319 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 320         __releases(fc->lock) __acquires(fc->lock)
 321 {
 322         if (!fc->no_interrupt) {
 323                 /* Any signal may interrupt this */
 324                 wait_answer_interruptible(fc, req);
 325
 326                 if (req->aborted)
 327                         goto aborted;
 328                 if (req->state == FUSE_REQ_FINISHED)
 329                         return;
 330
 331                 req->interrupted = 1;
 332                 if (req->state == FUSE_REQ_SENT)
 333                         queue_interrupt(fc, req);
 334         }
 335
 336         if (!req->force) {
 337                 sigset_t oldset;
 338
 339                 /* Only fatal signals may interrupt this */
 340                 block_sigs(&oldset);
 341                 wait_answer_interruptible(fc, req);
 342                 restore_sigs(&oldset);
 343
 344                 if (req->aborted)
 345                         goto aborted;
 346                 if (req->state == FUSE_REQ_FINISHED)
 347                         return;
 348
 349                 /* Request is not yet in userspace, bail out */
 350                 if (req->state == FUSE_REQ_PENDING) {
 351                         list_del(&req->list);
 352                         __fuse_put_request(req);
 353                         req->out.h.error = -EINTR;
 354                         return;
 355                 }
 356         }
 357
 358         /*
 359          * Either request is already in userspace, or it was forced.
 360          * Wait it out.
 361          */
 362         spin_unlock(&fc->lock);
 363         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 364         spin_lock(&fc->lock);
 365
 366         if (!req->aborted)
 367                 return;
 368
 369  aborted:
 370         BUG_ON(req->state != FUSE_REQ_FINISHED);
 371         if (req->locked) {
 372                 /* This is uninterruptible sleep, because data is
 373                    being copied to/from the buffers of req.  During
 374                    locked state, there mustn't be any filesystem
 375                    operation (e.g. page fault), since that could lead
 376                    to deadlock */
 377                 spin_unlock(&fc->lock);
 378                 wait_event(req->waitq, !req->locked);
 379                 spin_lock(&fc->lock);
 380         }
 381 }
 382
 383 void request_send(struct fuse_conn *fc, struct fuse_req *req)
 384 {
 385         req->isreply = 1;
 386         spin_lock(&fc->lock);
 387         if (!fc->connected)
 388                 req->out.h.error = -ENOTCONN;
 389         else if (fc->conn_error)
 390                 req->out.h.error = -ECONNREFUSED;
 391         else {
 392                 queue_request(fc, req);
 393                 /* acquire extra reference, since request is still needed
 394                    after request_end() */
 395                 __fuse_get_request(req);
 396
 397                 request_wait_answer(fc, req);
 398         }
 399         spin_unlock(&fc->lock);
 400 }
 401
 402 static void request_send_nowait_locked(struct fuse_conn *fc,
 403                                        struct fuse_req *req)
 404 {
 405         req->background = 1;
 406         fc->num_background++;
 407         if (fc->num_background == FUSE_MAX_BACKGROUND)
 408                 fc->blocked = 1;
 409         if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 410                 set_bdi_congested(&fc->bdi, READ);
 411                 set_bdi_congested(&fc->bdi, WRITE);
 412         }
 413         list_add_tail(&req->list, &fc->bg_queue);
 414         flush_bg_queue(fc);
 415 }
 416
 417 static void request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 418 {
 419         spin_lock(&fc->lock);
 420         if (fc->connected) {
 421                 request_send_nowait_locked(fc, req);
 422                 spin_unlock(&fc->lock);
 423         } else {
 424                 req->out.h.error = -ENOTCONN;
 425                 request_end(fc, req);
 426         }
 427 }
 428
 429 void request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 430 {
 431         req->isreply = 0;
 432         request_send_nowait(fc, req);
 433 }
 434
 435 void request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 436 {
 437         req->isreply = 1;
 438         request_send_nowait(fc, req);
 439 }
 440
 441 /*
 442  * Called under fc->lock
 443  *
 444  * fc->connected must have been checked previously
 445  */
 446 void request_send_background_locked(struct fuse_conn *fc, struct fuse_req *req)
 447 {
 448         req->isreply = 1;
 449         request_send_nowait_locked(fc, req);
 450 }
 451
 452 /*
 453  * Lock the request.  Up to the next unlock_request() there mustn't be
 454  * anything that could cause a page-fault.  If the request was already
 455  * aborted bail out.
 456  */
 457 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 458 {
 459         int err = 0;
 460         if (req) {
 461                 spin_lock(&fc->lock);
 462                 if (req->aborted)
 463                         err = -ENOENT;
 464                 else
 465                         req->locked = 1;
 466                 spin_unlock(&fc->lock);
 467         }
 468         return err;
 469 }
 470
 471 /*
 472  * Unlock request.  If it was aborted during being locked, the
 473  * requester thread is currently waiting for it to be unlocked, so
 474  * wake it up.
 475  */
 476 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 477 {
 478         if (req) {
 479                 spin_lock(&fc->lock);
 480                 req->locked = 0;
 481                 if (req->aborted)
 482                         wake_up(&req->waitq);
 483                 spin_unlock(&fc->lock);
 484         }
 485 }
 486
 487 struct fuse_copy_state {
 488         struct fuse_conn *fc;
 489         int write;
 490         struct fuse_req *req;
 491         const struct iovec *iov;
 492         unsigned long nr_segs;
 493         unsigned long seglen;
 494         unsigned long addr;
 495         struct page *pg;
 496         void *mapaddr;
 497         void *buf;
 498         unsigned len;
 499 };
 500
 501 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 502                            int write, struct fuse_req *req,
 503                            const struct iovec *iov, unsigned long nr_segs)
 504 {
 505         memset(cs, 0, sizeof(*cs));
 506         cs->fc = fc;
 507         cs->write = write;
 508         cs->req = req;
 509         cs->iov = iov;
 510         cs->nr_segs = nr_segs;
 511 }
 512
 513 /* Unmap and put previous page of userspace buffer */
 514 static void fuse_copy_finish(struct fuse_copy_state *cs)
 515 {
 516         if (cs->mapaddr) {
 517                 kunmap_atomic(cs->mapaddr, KM_USER0);
 518                 if (cs->write) {
 519                         flush_dcache_page(cs->pg);
 520                         set_page_dirty_lock(cs->pg);
 521                 }
 522                 put_page(cs->pg);
 523                 cs->mapaddr = NULL;
 524         }
 525 }
 526
 527 /*
 528  * Get another pagefull of userspace buffer, and map it to kernel
 529  * address space, and lock request
 530  */
 531 static int fuse_copy_fill(struct fuse_copy_state *cs)
 532 {
 533         unsigned long offset;
 534         int err;
 535
 536         unlock_request(cs->fc, cs->req);
 537         fuse_copy_finish(cs);
 538         if (!cs->seglen) {
 539                 BUG_ON(!cs->nr_segs);
 540                 cs->seglen = cs->iov[0].iov_len;
 541                 cs->addr = (unsigned long) cs->iov[0].iov_base;
 542                 cs->iov ++;
 543                 cs->nr_segs --;
 544         }
 545         down_read(&current->mm->mmap_sem);
 546         err = get_user_pages(current, current->mm, cs->addr, 1, cs->write, 0,
 547                              &cs->pg, NULL);
 548         up_read(&current->mm->mmap_sem);
 549         if (err < 0)
 550                 return err;
 551         BUG_ON(err != 1);
 552         offset = cs->addr % PAGE_SIZE;
 553         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
 554         cs->buf = cs->mapaddr + offset;
 555         cs->len = min(PAGE_SIZE - offset, cs->seglen);
 556         cs->seglen -= cs->len;
 557         cs->addr += cs->len;
 558
 559         return lock_request(cs->fc, cs->req);
 560 }
 561
 562 /* Do as much copy to/from userspace buffer as we can */
 563 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 564 {
 565         unsigned ncpy = min(*size, cs->len);
 566         if (val) {
 567                 if (cs->write)
 568                         memcpy(cs->buf, *val, ncpy);
 569                 else
 570                         memcpy(*val, cs->buf, ncpy);
 571                 *val += ncpy;
 572         }
 573         *size -= ncpy;
 574         cs->len -= ncpy;
 575         cs->buf += ncpy;
 576         return ncpy;
 577 }
 578
 579 /*
 580  * Copy a page in the request to/from the userspace buffer.  Must be
 581  * done atomically
 582  */
 583 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
 584                           unsigned offset, unsigned count, int zeroing)
 585 {
 586         if (page && zeroing && count < PAGE_SIZE) {
 587                 void *mapaddr = kmap_atomic(page, KM_USER1);
 588                 memset(mapaddr, 0, PAGE_SIZE);
 589                 kunmap_atomic(mapaddr, KM_USER1);
 590         }
 591         while (count) {
 592                 int err;
 593                 if (!cs->len && (err = fuse_copy_fill(cs)))
 594                         return err;
 595                 if (page) {
 596                         void *mapaddr = kmap_atomic(page, KM_USER1);
 597                         void *buf = mapaddr + offset;
 598                         offset += fuse_copy_do(cs, &buf, &count);
 599                         kunmap_atomic(mapaddr, KM_USER1);
 600                 } else
 601                         offset += fuse_copy_do(cs, NULL, &count);
 602         }
 603         if (page && !cs->write)
 604                 flush_dcache_page(page);
 605         return 0;
 606 }
 607
 608 /* Copy pages in the request to/from userspace buffer */
 609 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 610                            int zeroing)
 611 {
 612         unsigned i;
 613         struct fuse_req *req = cs->req;
 614         unsigned offset = req->page_offset;
 615         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 616
 617         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 618                 struct page *page = req->pages[i];
 619                 int err = fuse_copy_page(cs, page, offset, count, zeroing);
 620                 if (err)
 621                         return err;
 622
 623                 nbytes -= count;
 624                 count = min(nbytes, (unsigned) PAGE_SIZE);
 625                 offset = 0;
 626         }
 627         return 0;
 628 }
 629
 630 /* Copy a single argument in the request to/from userspace buffer */
 631 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 632 {
 633         while (size) {
 634                 int err;
 635                 if (!cs->len && (err = fuse_copy_fill(cs)))
 636                         return err;
 637                 fuse_copy_do(cs, &val, &size);
 638         }
 639         return 0;
 640 }
 641
 642 /* Copy request arguments to/from userspace buffer */
 643 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 644                           unsigned argpages, struct fuse_arg *args,
 645                           int zeroing)
 646 {
 647         int err = 0;
 648         unsigned i;
 649
 650         for (i = 0; !err && i < numargs; i++)  {
 651                 struct fuse_arg *arg = &args[i];
 652                 if (i == numargs - 1 && argpages)
 653                         err = fuse_copy_pages(cs, arg->size, zeroing);
 654                 else
 655                         err = fuse_copy_one(cs, arg->value, arg->size);
 656         }
 657         return err;
 658 }
 659
 660 static int request_pending(struct fuse_conn *fc)
 661 {
 662         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 663 }
 664
 665 /* Wait until a request is available on the pending list */
 666 static void request_wait(struct fuse_conn *fc)
 667 {
 668         DECLARE_WAITQUEUE(wait, current);
 669
 670         add_wait_queue_exclusive(&fc->waitq, &wait);
 671         while (fc->connected && !request_pending(fc)) {
 672                 set_current_state(TASK_INTERRUPTIBLE);
 673                 if (signal_pending(current))
 674                         break;
 675
 676                 spin_unlock(&fc->lock);
 677                 schedule();
 678                 spin_lock(&fc->lock);
 679         }
 680         set_current_state(TASK_RUNNING);
 681         remove_wait_queue(&fc->waitq, &wait);
 682 }
 683
 684 /*
 685  * Transfer an interrupt request to userspace
 686  *
 687  * Unlike other requests this is assembled on demand, without a need
 688  * to allocate a separate fuse_req structure.
 689  *
 690  * Called with fc->lock held, releases it
 691  */
 692 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
 693                                const struct iovec *iov, unsigned long nr_segs)
 694         __releases(fc->lock)
 695 {
 696         struct fuse_copy_state cs;
 697         struct fuse_in_header ih;
 698         struct fuse_interrupt_in arg;
 699         unsigned reqsize = sizeof(ih) + sizeof(arg);
 700         int err;
 701
 702         list_del_init(&req->intr_entry);
 703         req->intr_unique = fuse_get_unique(fc);
 704         memset(&ih, 0, sizeof(ih));
 705         memset(&arg, 0, sizeof(arg));
 706         ih.len = reqsize;
 707         ih.opcode = FUSE_INTERRUPT;
 708         ih.unique = req->intr_unique;
 709         arg.unique = req->in.h.unique;
 710
 711         spin_unlock(&fc->lock);
 712         if (iov_length(iov, nr_segs) < reqsize)
 713                 return -EINVAL;
 714
 715         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
 716         err = fuse_copy_one(&cs, &ih, sizeof(ih));
 717         if (!err)
 718                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
 719         fuse_copy_finish(&cs);
 720
 721         return err ? err : reqsize;
 722 }
 723
 724 /*
 725  * Read a single request into the userspace filesystem's buffer.  This
 726  * function waits until a request is available, then removes it from
 727  * the pending list and copies request data to userspace buffer.  If
 728  * no reply is needed (FORGET) or request has been aborted or there
 729  * was an error during the copying then it's finished by calling
 730  * request_end().  Otherwise add it to the processing list, and set
 731  * the 'sent' flag.
 732  */
 733 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
 734                               unsigned long nr_segs, loff_t pos)
 735 {
 736         int err;
 737         struct fuse_req *req;
 738         struct fuse_in *in;
 739         struct fuse_copy_state cs;
 740         unsigned reqsize;
 741         struct file *file = iocb->ki_filp;
 742         struct fuse_conn *fc = fuse_get_conn(file);
 743         if (!fc)
 744                 return -EPERM;
 745
 746  restart:
 747         spin_lock(&fc->lock);
 748         err = -EAGAIN;
 749         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 750             !request_pending(fc))
 751                 goto err_unlock;
 752
 753         request_wait(fc);
 754         err = -ENODEV;
 755         if (!fc->connected)
 756                 goto err_unlock;
 757         err = -ERESTARTSYS;
 758         if (!request_pending(fc))
 759                 goto err_unlock;
 760
 761         if (!list_empty(&fc->interrupts)) {
 762                 req = list_entry(fc->interrupts.next, struct fuse_req,
 763                                  intr_entry);
 764                 return fuse_read_interrupt(fc, req, iov, nr_segs);
 765         }
 766
 767         req = list_entry(fc->pending.next, struct fuse_req, list);
 768         req->state = FUSE_REQ_READING;
 769         list_move(&req->list, &fc->io);
 770
 771         in = &req->in;
 772         reqsize = in->h.len;
 773         /* If request is too large, reply with an error and restart the read */
 774         if (iov_length(iov, nr_segs) < reqsize) {
 775                 req->out.h.error = -EIO;
 776                 /* SETXATTR is special, since it may contain too large data */
 777                 if (in->h.opcode == FUSE_SETXATTR)
 778                         req->out.h.error = -E2BIG;
 779                 request_end(fc, req);
 780                 goto restart;
 781         }
 782         spin_unlock(&fc->lock);
 783         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
 784         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
 785         if (!err)
 786                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
 787                                      (struct fuse_arg *) in->args, 0);
 788         fuse_copy_finish(&cs);
 789         spin_lock(&fc->lock);
 790         req->locked = 0;
 791         if (req->aborted) {
 792                 request_end(fc, req);
 793                 return -ENODEV;
 794         }
 795         if (err) {
 796                 req->out.h.error = -EIO;
 797                 request_end(fc, req);
 798                 return err;
 799         }
 800         if (!req->isreply)
 801                 request_end(fc, req);
 802         else {
 803                 req->state = FUSE_REQ_SENT;
 804                 list_move_tail(&req->list, &fc->processing);
 805                 if (req->interrupted)
 806                         queue_interrupt(fc, req);
 807                 spin_unlock(&fc->lock);
 808         }
 809         return reqsize;
 810
 811  err_unlock:
 812         spin_unlock(&fc->lock);
 813         return err;
 814 }
 815
 816 /* Look up request on processing list by unique ID */
 817 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
 818 {
 819         struct list_head *entry;
 820
 821         list_for_each(entry, &fc->processing) {
 822                 struct fuse_req *req;
 823                 req = list_entry(entry, struct fuse_req, list);
 824                 if (req->in.h.unique == unique || req->intr_unique == unique)
 825                         return req;
 826         }
 827         return NULL;
 828 }
 829
 830 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
 831                          unsigned nbytes)
 832 {
 833         unsigned reqsize = sizeof(struct fuse_out_header);
 834
 835         if (out->h.error)
 836                 return nbytes != reqsize ? -EINVAL : 0;
 837
 838         reqsize += len_args(out->numargs, out->args);
 839
 840         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
 841                 return -EINVAL;
 842         else if (reqsize > nbytes) {
 843                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
 844                 unsigned diffsize = reqsize - nbytes;
 845                 if (diffsize > lastarg->size)
 846                         return -EINVAL;
 847                 lastarg->size -= diffsize;
 848         }
 849         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
 850                               out->page_zeroing);
 851 }
 852
 853 /*
 854  * Write a single reply to a request.  First the header is copied from
 855  * the write buffer.  The request is then searched on the processing
 856  * list by the unique ID found in the header.  If found, then remove
 857  * it from the list and copy the rest of the buffer to the request.
 858  * The request is finished by calling request_end()
 859  */
 860 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
 861                                unsigned long nr_segs, loff_t pos)
 862 {
 863         int err;
 864         unsigned nbytes = iov_length(iov, nr_segs);
 865         struct fuse_req *req;
 866         struct fuse_out_header oh;
 867         struct fuse_copy_state cs;
 868         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
 869         if (!fc)
 870                 return -EPERM;
 871
 872         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
 873         if (nbytes < sizeof(struct fuse_out_header))
 874                 return -EINVAL;
 875
 876         err = fuse_copy_one(&cs, &oh, sizeof(oh));
 877         if (err)
 878                 goto err_finish;
 879         err = -EINVAL;
 880         if (!oh.unique || oh.error <= -1000 || oh.error > 0 ||
 881             oh.len != nbytes)
 882                 goto err_finish;
 883
 884         spin_lock(&fc->lock);
 885         err = -ENOENT;
 886         if (!fc->connected)
 887                 goto err_unlock;
 888
 889         req = request_find(fc, oh.unique);
 890         if (!req)
 891                 goto err_unlock;
 892
 893         if (req->aborted) {
 894                 spin_unlock(&fc->lock);
 895                 fuse_copy_finish(&cs);
 896                 spin_lock(&fc->lock);
 897                 request_end(fc, req);
 898                 return -ENOENT;
 899         }
 900         /* Is it an interrupt reply? */
 901         if (req->intr_unique == oh.unique) {
 902                 err = -EINVAL;
 903                 if (nbytes != sizeof(struct fuse_out_header))
 904                         goto err_unlock;
 905
 906                 if (oh.error == -ENOSYS)
 907                         fc->no_interrupt = 1;
 908                 else if (oh.error == -EAGAIN)
 909                         queue_interrupt(fc, req);
 910
 911                 spin_unlock(&fc->lock);
 912                 fuse_copy_finish(&cs);
 913                 return nbytes;
 914         }
 915
 916         req->state = FUSE_REQ_WRITING;
 917         list_move(&req->list, &fc->io);
 918         req->out.h = oh;
 919         req->locked = 1;
 920         cs.req = req;
 921         spin_unlock(&fc->lock);
 922
 923         err = copy_out_args(&cs, &req->out, nbytes);
 924         fuse_copy_finish(&cs);
 925
 926         spin_lock(&fc->lock);
 927         req->locked = 0;
 928         if (!err) {
 929                 if (req->aborted)
 930                         err = -ENOENT;
 931         } else if (!req->aborted)
 932                 req->out.h.error = -EIO;
 933         request_end(fc, req);
 934
 935         return err ? err : nbytes;
 936
 937  err_unlock:
 938         spin_unlock(&fc->lock);
 939  err_finish:
 940         fuse_copy_finish(&cs);
 941         return err;
 942 }
 943
 944 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
 945 {
 946         unsigned mask = POLLOUT | POLLWRNORM;
 947         struct fuse_conn *fc = fuse_get_conn(file);
 948         if (!fc)
 949                 return POLLERR;
 950
 951         poll_wait(file, &fc->waitq, wait);
 952
 953         spin_lock(&fc->lock);
 954         if (!fc->connected)
 955                 mask = POLLERR;
 956         else if (request_pending(fc))
 957                 mask |= POLLIN | POLLRDNORM;
 958         spin_unlock(&fc->lock);
 959
 960         return mask;
 961 }
 962
 963 /*
 964  * Abort all requests on the given list (pending or processing)
 965  *
 966  * This function releases and reacquires fc->lock
 967  */
 968 static void end_requests(struct fuse_conn *fc, struct list_head *head)
 969 {
 970         while (!list_empty(head)) {
 971                 struct fuse_req *req;
 972                 req = list_entry(head->next, struct fuse_req, list);
 973                 req->out.h.error = -ECONNABORTED;
 974                 request_end(fc, req);
 975                 spin_lock(&fc->lock);
 976         }
 977 }
 978
 979 /*
 980  * Abort requests under I/O
 981  *
 982  * The requests are set to aborted and finished, and the request
 983  * waiter is woken up.  This will make request_wait_answer() wait
 984  * until the request is unlocked and then return.
 985  *
 986  * If the request is asynchronous, then the end function needs to be
 987  * called after waiting for the request to be unlocked (if it was
 988  * locked).
 989  */
 990 static void end_io_requests(struct fuse_conn *fc)
 991         __releases(fc->lock) __acquires(fc->lock)
 992 {
 993         while (!list_empty(&fc->io)) {
 994                 struct fuse_req *req =
 995                         list_entry(fc->io.next, struct fuse_req, list);
 996                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 997
 998                 req->aborted = 1;
 999                 req->out.h.error = -ECONNABORTED;
1000                 req->state = FUSE_REQ_FINISHED;
1001                 list_del_init(&req->list);
1002                 wake_up(&req->waitq);
1003                 if (end) {
1004                         req->end = NULL;
1005                         /* The end function will consume this reference */
1006                         __fuse_get_request(req);
1007                         spin_unlock(&fc->lock);
1008                         wait_event(req->waitq, !req->locked);
1009                         end(fc, req);
1010                         spin_lock(&fc->lock);
1011                 }
1012         }
1013 }
1014
1015 /*
1016  * Abort all requests.
1017  *
1018  * Emergency exit in case of a malicious or accidental deadlock, or
1019  * just a hung filesystem.
1020  *
1021  * The same effect is usually achievable through killing the
1022  * filesystem daemon and all users of the filesystem.  The exception
1023  * is the combination of an asynchronous request and the tricky
1024  * deadlock (see Documentation/filesystems/fuse.txt).
1025  *
1026  * During the aborting, progression of requests from the pending and
1027  * processing lists onto the io list, and progression of new requests
1028  * onto the pending list is prevented by req->connected being false.
1029  *
1030  * Progression of requests under I/O to the processing list is
1031  * prevented by the req->aborted flag being true for these requests.
1032  * For this reason requests on the io list must be aborted first.
1033  */
1034 void fuse_abort_conn(struct fuse_conn *fc)
1035 {
1036         spin_lock(&fc->lock);
1037         if (fc->connected) {
1038                 fc->connected = 0;
1039                 fc->blocked = 0;
1040                 end_io_requests(fc);
1041                 end_requests(fc, &fc->pending);
1042                 end_requests(fc, &fc->processing);
1043                 wake_up_all(&fc->waitq);
1044                 wake_up_all(&fc->blocked_waitq);
1045                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1046         }
1047         spin_unlock(&fc->lock);
1048 }
1049
1050 static int fuse_dev_release(struct inode *inode, struct file *file)
1051 {
1052         struct fuse_conn *fc = fuse_get_conn(file);
1053         if (fc) {
1054                 spin_lock(&fc->lock);
1055                 fc->connected = 0;
1056                 end_requests(fc, &fc->pending);
1057                 end_requests(fc, &fc->processing);
1058                 spin_unlock(&fc->lock);
1059                 fuse_conn_put(fc);
1060         }
1061
1062         return 0;
1063 }
1064
1065 static int fuse_dev_fasync(int fd, struct file *file, int on)
1066 {
1067         struct fuse_conn *fc = fuse_get_conn(file);
1068         if (!fc)
1069                 return -EPERM;
1070
1071         /* No locking - fasync_helper does its own locking */
1072         return fasync_helper(fd, file, on, &fc->fasync);
1073 }
1074
1075 const struct file_operations fuse_dev_operations = {
1076         .owner          = THIS_MODULE,
1077         .llseek         = no_llseek,
1078         .read           = do_sync_read,
1079         .aio_read       = fuse_dev_read,
1080         .write          = do_sync_write,
1081         .aio_write      = fuse_dev_write,
1082         .poll           = fuse_dev_poll,
1083         .release        = fuse_dev_release,
1084         .fasync         = fuse_dev_fasync,
1085 };
1086
1087 static struct miscdevice fuse_miscdevice = {
1088         .minor = FUSE_MINOR,
1089         .name  = "fuse",
1090         .fops = &fuse_dev_operations,
1091 };
1092
1093 int __init fuse_dev_init(void)
1094 {
1095         int err = -ENOMEM;
1096         fuse_req_cachep = kmem_cache_create("fuse_request",
1097                                             sizeof(struct fuse_req),
1098                                             0, 0, NULL);
1099         if (!fuse_req_cachep)
1100                 goto out;
1101
1102         err = misc_register(&fuse_miscdevice);
1103         if (err)
1104                 goto out_cache_clean;
1105
1106         return 0;
1107
1108  out_cache_clean:
1109         kmem_cache_destroy(fuse_req_cachep);
1110  out:
1111         return err;
1112 }
1113
1114 void fuse_dev_cleanup(void)
1115 {
1116         misc_deregister(&fuse_miscdevice);
1117         kmem_cache_destroy(fuse_req_cachep);
1118 }