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                     fc->connected) {
 286                         clear_bdi_congested(&fc->bdi, READ);
 287                         clear_bdi_congested(&fc->bdi, WRITE);
 288                 }
 289                 fc->num_background--;
 290                 fc->active_background--;
 291                 flush_bg_queue(fc);
 292         }
 293         spin_unlock(&fc->lock);
 294         wake_up(&req->waitq);
 295         if (end)
 296                 end(fc, req);
 297         else
 298                 fuse_put_request(fc, req);
 299 }
 300
 301 static void wait_answer_interruptible(struct fuse_conn *fc,
 302                                       struct fuse_req *req)
 303         __releases(fc->lock) __acquires(fc->lock)
 304 {
 305         if (signal_pending(current))
 306                 return;
 307
 308         spin_unlock(&fc->lock);
 309         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 310         spin_lock(&fc->lock);
 311 }
 312
 313 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 314 {
 315         list_add_tail(&req->intr_entry, &fc->interrupts);
 316         wake_up(&fc->waitq);
 317         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 318 }
 319
 320 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 321         __releases(fc->lock) __acquires(fc->lock)
 322 {
 323         if (!fc->no_interrupt) {
 324                 /* Any signal may interrupt this */
 325                 wait_answer_interruptible(fc, req);
 326
 327                 if (req->aborted)
 328                         goto aborted;
 329                 if (req->state == FUSE_REQ_FINISHED)
 330                         return;
 331
 332                 req->interrupted = 1;
 333                 if (req->state == FUSE_REQ_SENT)
 334                         queue_interrupt(fc, req);
 335         }
 336
 337         if (!req->force) {
 338                 sigset_t oldset;
 339
 340                 /* Only fatal signals may interrupt this */
 341                 block_sigs(&oldset);
 342                 wait_answer_interruptible(fc, req);
 343                 restore_sigs(&oldset);
 344
 345                 if (req->aborted)
 346                         goto aborted;
 347                 if (req->state == FUSE_REQ_FINISHED)
 348                         return;
 349
 350                 /* Request is not yet in userspace, bail out */
 351                 if (req->state == FUSE_REQ_PENDING) {
 352                         list_del(&req->list);
 353                         __fuse_put_request(req);
 354                         req->out.h.error = -EINTR;
 355                         return;
 356                 }
 357         }
 358
 359         /*
 360          * Either request is already in userspace, or it was forced.
 361          * Wait it out.
 362          */
 363         spin_unlock(&fc->lock);
 364         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 365         spin_lock(&fc->lock);
 366
 367         if (!req->aborted)
 368                 return;
 369
 370  aborted:
 371         BUG_ON(req->state != FUSE_REQ_FINISHED);
 372         if (req->locked) {
 373                 /* This is uninterruptible sleep, because data is
 374                    being copied to/from the buffers of req.  During
 375                    locked state, there mustn't be any filesystem
 376                    operation (e.g. page fault), since that could lead
 377                    to deadlock */
 378                 spin_unlock(&fc->lock);
 379                 wait_event(req->waitq, !req->locked);
 380                 spin_lock(&fc->lock);
 381         }
 382 }
 383
 384 void request_send(struct fuse_conn *fc, struct fuse_req *req)
 385 {
 386         req->isreply = 1;
 387         spin_lock(&fc->lock);
 388         if (!fc->connected)
 389                 req->out.h.error = -ENOTCONN;
 390         else if (fc->conn_error)
 391                 req->out.h.error = -ECONNREFUSED;
 392         else {
 393                 queue_request(fc, req);
 394                 /* acquire extra reference, since request is still needed
 395                    after request_end() */
 396                 __fuse_get_request(req);
 397
 398                 request_wait_answer(fc, req);
 399         }
 400         spin_unlock(&fc->lock);
 401 }
 402
 403 static void request_send_nowait_locked(struct fuse_conn *fc,
 404                                        struct fuse_req *req)
 405 {
 406         req->background = 1;
 407         fc->num_background++;
 408         if (fc->num_background == FUSE_MAX_BACKGROUND)
 409                 fc->blocked = 1;
 410         if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 411                 set_bdi_congested(&fc->bdi, READ);
 412                 set_bdi_congested(&fc->bdi, WRITE);
 413         }
 414         list_add_tail(&req->list, &fc->bg_queue);
 415         flush_bg_queue(fc);
 416 }
 417
 418 static void request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 419 {
 420         spin_lock(&fc->lock);
 421         if (fc->connected) {
 422                 request_send_nowait_locked(fc, req);
 423                 spin_unlock(&fc->lock);
 424         } else {
 425                 req->out.h.error = -ENOTCONN;
 426                 request_end(fc, req);
 427         }
 428 }
 429
 430 void request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 431 {
 432         req->isreply = 0;
 433         request_send_nowait(fc, req);
 434 }
 435
 436 void request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 437 {
 438         req->isreply = 1;
 439         request_send_nowait(fc, req);
 440 }
 441
 442 /*
 443  * Called under fc->lock
 444  *
 445  * fc->connected must have been checked previously
 446  */
 447 void request_send_background_locked(struct fuse_conn *fc, struct fuse_req *req)
 448 {
 449         req->isreply = 1;
 450         request_send_nowait_locked(fc, req);
 451 }
 452
 453 /*
 454  * Lock the request.  Up to the next unlock_request() there mustn't be
 455  * anything that could cause a page-fault.  If the request was already
 456  * aborted bail out.
 457  */
 458 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 459 {
 460         int err = 0;
 461         if (req) {
 462                 spin_lock(&fc->lock);
 463                 if (req->aborted)
 464                         err = -ENOENT;
 465                 else
 466                         req->locked = 1;
 467                 spin_unlock(&fc->lock);
 468         }
 469         return err;
 470 }
 471
 472 /*
 473  * Unlock request.  If it was aborted during being locked, the
 474  * requester thread is currently waiting for it to be unlocked, so
 475  * wake it up.
 476  */
 477 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 478 {
 479         if (req) {
 480                 spin_lock(&fc->lock);
 481                 req->locked = 0;
 482                 if (req->aborted)
 483                         wake_up(&req->waitq);
 484                 spin_unlock(&fc->lock);
 485         }
 486 }
 487
 488 struct fuse_copy_state {
 489         struct fuse_conn *fc;
 490         int write;
 491         struct fuse_req *req;
 492         const struct iovec *iov;
 493         unsigned long nr_segs;
 494         unsigned long seglen;
 495         unsigned long addr;
 496         struct page *pg;
 497         void *mapaddr;
 498         void *buf;
 499         unsigned len;
 500 };
 501
 502 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 503                            int write, struct fuse_req *req,
 504                            const struct iovec *iov, unsigned long nr_segs)
 505 {
 506         memset(cs, 0, sizeof(*cs));
 507         cs->fc = fc;
 508         cs->write = write;
 509         cs->req = req;
 510         cs->iov = iov;
 511         cs->nr_segs = nr_segs;
 512 }
 513
 514 /* Unmap and put previous page of userspace buffer */
 515 static void fuse_copy_finish(struct fuse_copy_state *cs)
 516 {
 517         if (cs->mapaddr) {
 518                 kunmap_atomic(cs->mapaddr, KM_USER0);
 519                 if (cs->write) {
 520                         flush_dcache_page(cs->pg);
 521                         set_page_dirty_lock(cs->pg);
 522                 }
 523                 put_page(cs->pg);
 524                 cs->mapaddr = NULL;
 525         }
 526 }
 527
 528 /*
 529  * Get another pagefull of userspace buffer, and map it to kernel
 530  * address space, and lock request
 531  */
 532 static int fuse_copy_fill(struct fuse_copy_state *cs)
 533 {
 534         unsigned long offset;
 535         int err;
 536
 537         unlock_request(cs->fc, cs->req);
 538         fuse_copy_finish(cs);
 539         if (!cs->seglen) {
 540                 BUG_ON(!cs->nr_segs);
 541                 cs->seglen = cs->iov[0].iov_len;
 542                 cs->addr = (unsigned long) cs->iov[0].iov_base;
 543                 cs->iov ++;
 544                 cs->nr_segs --;
 545         }
 546         down_read(&current->mm->mmap_sem);
 547         err = get_user_pages(current, current->mm, cs->addr, 1, cs->write, 0,
 548                              &cs->pg, NULL);
 549         up_read(&current->mm->mmap_sem);
 550         if (err < 0)
 551                 return err;
 552         BUG_ON(err != 1);
 553         offset = cs->addr % PAGE_SIZE;
 554         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
 555         cs->buf = cs->mapaddr + offset;
 556         cs->len = min(PAGE_SIZE - offset, cs->seglen);
 557         cs->seglen -= cs->len;
 558         cs->addr += cs->len;
 559
 560         return lock_request(cs->fc, cs->req);
 561 }
 562
 563 /* Do as much copy to/from userspace buffer as we can */
 564 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 565 {
 566         unsigned ncpy = min(*size, cs->len);
 567         if (val) {
 568                 if (cs->write)
 569                         memcpy(cs->buf, *val, ncpy);
 570                 else
 571                         memcpy(*val, cs->buf, ncpy);
 572                 *val += ncpy;
 573         }
 574         *size -= ncpy;
 575         cs->len -= ncpy;
 576         cs->buf += ncpy;
 577         return ncpy;
 578 }
 579
 580 /*
 581  * Copy a page in the request to/from the userspace buffer.  Must be
 582  * done atomically
 583  */
 584 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
 585                           unsigned offset, unsigned count, int zeroing)
 586 {
 587         if (page && zeroing && count < PAGE_SIZE) {
 588                 void *mapaddr = kmap_atomic(page, KM_USER1);
 589                 memset(mapaddr, 0, PAGE_SIZE);
 590                 kunmap_atomic(mapaddr, KM_USER1);
 591         }
 592         while (count) {
 593                 int err;
 594                 if (!cs->len && (err = fuse_copy_fill(cs)))
 595                         return err;
 596                 if (page) {
 597                         void *mapaddr = kmap_atomic(page, KM_USER1);
 598                         void *buf = mapaddr + offset;
 599                         offset += fuse_copy_do(cs, &buf, &count);
 600                         kunmap_atomic(mapaddr, KM_USER1);
 601                 } else
 602                         offset += fuse_copy_do(cs, NULL, &count);
 603         }
 604         if (page && !cs->write)
 605                 flush_dcache_page(page);
 606         return 0;
 607 }
 608
 609 /* Copy pages in the request to/from userspace buffer */
 610 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 611                            int zeroing)
 612 {
 613         unsigned i;
 614         struct fuse_req *req = cs->req;
 615         unsigned offset = req->page_offset;
 616         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 617
 618         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 619                 struct page *page = req->pages[i];
 620                 int err = fuse_copy_page(cs, page, offset, count, zeroing);
 621                 if (err)
 622                         return err;
 623
 624                 nbytes -= count;
 625                 count = min(nbytes, (unsigned) PAGE_SIZE);
 626                 offset = 0;
 627         }
 628         return 0;
 629 }
 630
 631 /* Copy a single argument in the request to/from userspace buffer */
 632 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 633 {
 634         while (size) {
 635                 int err;
 636                 if (!cs->len && (err = fuse_copy_fill(cs)))
 637                         return err;
 638                 fuse_copy_do(cs, &val, &size);
 639         }
 640         return 0;
 641 }
 642
 643 /* Copy request arguments to/from userspace buffer */
 644 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 645                           unsigned argpages, struct fuse_arg *args,
 646                           int zeroing)
 647 {
 648         int err = 0;
 649         unsigned i;
 650
 651         for (i = 0; !err && i < numargs; i++)  {
 652                 struct fuse_arg *arg = &args[i];
 653                 if (i == numargs - 1 && argpages)
 654                         err = fuse_copy_pages(cs, arg->size, zeroing);
 655                 else
 656                         err = fuse_copy_one(cs, arg->value, arg->size);
 657         }
 658         return err;
 659 }
 660
 661 static int request_pending(struct fuse_conn *fc)
 662 {
 663         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 664 }
 665
 666 /* Wait until a request is available on the pending list */
 667 static void request_wait(struct fuse_conn *fc)
 668 {
 669         DECLARE_WAITQUEUE(wait, current);
 670
 671         add_wait_queue_exclusive(&fc->waitq, &wait);
 672         while (fc->connected && !request_pending(fc)) {
 673                 set_current_state(TASK_INTERRUPTIBLE);
 674                 if (signal_pending(current))
 675                         break;
 676
 677                 spin_unlock(&fc->lock);
 678                 schedule();
 679                 spin_lock(&fc->lock);
 680         }
 681         set_current_state(TASK_RUNNING);
 682         remove_wait_queue(&fc->waitq, &wait);
 683 }
 684
 685 /*
 686  * Transfer an interrupt request to userspace
 687  *
 688  * Unlike other requests this is assembled on demand, without a need
 689  * to allocate a separate fuse_req structure.
 690  *
 691  * Called with fc->lock held, releases it
 692  */
 693 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
 694                                const struct iovec *iov, unsigned long nr_segs)
 695         __releases(fc->lock)
 696 {
 697         struct fuse_copy_state cs;
 698         struct fuse_in_header ih;
 699         struct fuse_interrupt_in arg;
 700         unsigned reqsize = sizeof(ih) + sizeof(arg);
 701         int err;
 702
 703         list_del_init(&req->intr_entry);
 704         req->intr_unique = fuse_get_unique(fc);
 705         memset(&ih, 0, sizeof(ih));
 706         memset(&arg, 0, sizeof(arg));
 707         ih.len = reqsize;
 708         ih.opcode = FUSE_INTERRUPT;
 709         ih.unique = req->intr_unique;
 710         arg.unique = req->in.h.unique;
 711
 712         spin_unlock(&fc->lock);
 713         if (iov_length(iov, nr_segs) < reqsize)
 714                 return -EINVAL;
 715
 716         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
 717         err = fuse_copy_one(&cs, &ih, sizeof(ih));
 718         if (!err)
 719                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
 720         fuse_copy_finish(&cs);
 721
 722         return err ? err : reqsize;
 723 }
 724
 725 /*
 726  * Read a single request into the userspace filesystem's buffer.  This
 727  * function waits until a request is available, then removes it from
 728  * the pending list and copies request data to userspace buffer.  If
 729  * no reply is needed (FORGET) or request has been aborted or there
 730  * was an error during the copying then it's finished by calling
 731  * request_end().  Otherwise add it to the processing list, and set
 732  * the 'sent' flag.
 733  */
 734 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
 735                               unsigned long nr_segs, loff_t pos)
 736 {
 737         int err;
 738         struct fuse_req *req;
 739         struct fuse_in *in;
 740         struct fuse_copy_state cs;
 741         unsigned reqsize;
 742         struct file *file = iocb->ki_filp;
 743         struct fuse_conn *fc = fuse_get_conn(file);
 744         if (!fc)
 745                 return -EPERM;
 746
 747  restart:
 748         spin_lock(&fc->lock);
 749         err = -EAGAIN;
 750         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 751             !request_pending(fc))
 752                 goto err_unlock;
 753
 754         request_wait(fc);
 755         err = -ENODEV;
 756         if (!fc->connected)
 757                 goto err_unlock;
 758         err = -ERESTARTSYS;
 759         if (!request_pending(fc))
 760                 goto err_unlock;
 761
 762         if (!list_empty(&fc->interrupts)) {
 763                 req = list_entry(fc->interrupts.next, struct fuse_req,
 764                                  intr_entry);
 765                 return fuse_read_interrupt(fc, req, iov, nr_segs);
 766         }
 767
 768         req = list_entry(fc->pending.next, struct fuse_req, list);
 769         req->state = FUSE_REQ_READING;
 770         list_move(&req->list, &fc->io);
 771
 772         in = &req->in;
 773         reqsize = in->h.len;
 774         /* If request is too large, reply with an error and restart the read */
 775         if (iov_length(iov, nr_segs) < reqsize) {
 776                 req->out.h.error = -EIO;
 777                 /* SETXATTR is special, since it may contain too large data */
 778                 if (in->h.opcode == FUSE_SETXATTR)
 779                         req->out.h.error = -E2BIG;
 780                 request_end(fc, req);
 781                 goto restart;
 782         }
 783         spin_unlock(&fc->lock);
 784         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
 785         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
 786         if (!err)
 787                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
 788                                      (struct fuse_arg *) in->args, 0);
 789         fuse_copy_finish(&cs);
 790         spin_lock(&fc->lock);
 791         req->locked = 0;
 792         if (req->aborted) {
 793                 request_end(fc, req);
 794                 return -ENODEV;
 795         }
 796         if (err) {
 797                 req->out.h.error = -EIO;
 798                 request_end(fc, req);
 799                 return err;
 800         }
 801         if (!req->isreply)
 802                 request_end(fc, req);
 803         else {
 804                 req->state = FUSE_REQ_SENT;
 805                 list_move_tail(&req->list, &fc->processing);
 806                 if (req->interrupted)
 807                         queue_interrupt(fc, req);
 808                 spin_unlock(&fc->lock);
 809         }
 810         return reqsize;
 811
 812  err_unlock:
 813         spin_unlock(&fc->lock);
 814         return err;
 815 }
 816
 817 /* Look up request on processing list by unique ID */
 818 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
 819 {
 820         struct list_head *entry;
 821
 822         list_for_each(entry, &fc->processing) {
 823                 struct fuse_req *req;
 824                 req = list_entry(entry, struct fuse_req, list);
 825                 if (req->in.h.unique == unique || req->intr_unique == unique)
 826                         return req;
 827         }
 828         return NULL;
 829 }
 830
 831 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
 832                          unsigned nbytes)
 833 {
 834         unsigned reqsize = sizeof(struct fuse_out_header);
 835
 836         if (out->h.error)
 837                 return nbytes != reqsize ? -EINVAL : 0;
 838
 839         reqsize += len_args(out->numargs, out->args);
 840
 841         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
 842                 return -EINVAL;
 843         else if (reqsize > nbytes) {
 844                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
 845                 unsigned diffsize = reqsize - nbytes;
 846                 if (diffsize > lastarg->size)
 847                         return -EINVAL;
 848                 lastarg->size -= diffsize;
 849         }
 850         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
 851                               out->page_zeroing);
 852 }
 853
 854 /*
 855  * Write a single reply to a request.  First the header is copied from
 856  * the write buffer.  The request is then searched on the processing
 857  * list by the unique ID found in the header.  If found, then remove
 858  * it from the list and copy the rest of the buffer to the request.
 859  * The request is finished by calling request_end()
 860  */
 861 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
 862                                unsigned long nr_segs, loff_t pos)
 863 {
 864         int err;
 865         unsigned nbytes = iov_length(iov, nr_segs);
 866         struct fuse_req *req;
 867         struct fuse_out_header oh;
 868         struct fuse_copy_state cs;
 869         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
 870         if (!fc)
 871                 return -EPERM;
 872
 873         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
 874         if (nbytes < sizeof(struct fuse_out_header))
 875                 return -EINVAL;
 876
 877         err = fuse_copy_one(&cs, &oh, sizeof(oh));
 878         if (err)
 879                 goto err_finish;
 880         err = -EINVAL;
 881         if (!oh.unique || oh.error <= -1000 || oh.error > 0 ||
 882             oh.len != nbytes)
 883                 goto err_finish;
 884
 885         spin_lock(&fc->lock);
 886         err = -ENOENT;
 887         if (!fc->connected)
 888                 goto err_unlock;
 889
 890         req = request_find(fc, oh.unique);
 891         if (!req)
 892                 goto err_unlock;
 893
 894         if (req->aborted) {
 895                 spin_unlock(&fc->lock);
 896                 fuse_copy_finish(&cs);
 897                 spin_lock(&fc->lock);
 898                 request_end(fc, req);
 899                 return -ENOENT;
 900         }
 901         /* Is it an interrupt reply? */
 902         if (req->intr_unique == oh.unique) {
 903                 err = -EINVAL;
 904                 if (nbytes != sizeof(struct fuse_out_header))
 905                         goto err_unlock;
 906
 907                 if (oh.error == -ENOSYS)
 908                         fc->no_interrupt = 1;
 909                 else if (oh.error == -EAGAIN)
 910                         queue_interrupt(fc, req);
 911
 912                 spin_unlock(&fc->lock);
 913                 fuse_copy_finish(&cs);
 914                 return nbytes;
 915         }
 916
 917         req->state = FUSE_REQ_WRITING;
 918         list_move(&req->list, &fc->io);
 919         req->out.h = oh;
 920         req->locked = 1;
 921         cs.req = req;
 922         spin_unlock(&fc->lock);
 923
 924         err = copy_out_args(&cs, &req->out, nbytes);
 925         fuse_copy_finish(&cs);
 926
 927         spin_lock(&fc->lock);
 928         req->locked = 0;
 929         if (!err) {
 930                 if (req->aborted)
 931                         err = -ENOENT;
 932         } else if (!req->aborted)
 933                 req->out.h.error = -EIO;
 934         request_end(fc, req);
 935
 936         return err ? err : nbytes;
 937
 938  err_unlock:
 939         spin_unlock(&fc->lock);
 940  err_finish:
 941         fuse_copy_finish(&cs);
 942         return err;
 943 }
 944
 945 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
 946 {
 947         unsigned mask = POLLOUT | POLLWRNORM;
 948         struct fuse_conn *fc = fuse_get_conn(file);
 949         if (!fc)
 950                 return POLLERR;
 951
 952         poll_wait(file, &fc->waitq, wait);
 953
 954         spin_lock(&fc->lock);
 955         if (!fc->connected)
 956                 mask = POLLERR;
 957         else if (request_pending(fc))
 958                 mask |= POLLIN | POLLRDNORM;
 959         spin_unlock(&fc->lock);
 960
 961         return mask;
 962 }
 963
 964 /*
 965  * Abort all requests on the given list (pending or processing)
 966  *
 967  * This function releases and reacquires fc->lock
 968  */
 969 static void end_requests(struct fuse_conn *fc, struct list_head *head)
 970 {
 971         while (!list_empty(head)) {
 972                 struct fuse_req *req;
 973                 req = list_entry(head->next, struct fuse_req, list);
 974                 req->out.h.error = -ECONNABORTED;
 975                 request_end(fc, req);
 976                 spin_lock(&fc->lock);
 977         }
 978 }
 979
 980 /*
 981  * Abort requests under I/O
 982  *
 983  * The requests are set to aborted and finished, and the request
 984  * waiter is woken up.  This will make request_wait_answer() wait
 985  * until the request is unlocked and then return.
 986  *
 987  * If the request is asynchronous, then the end function needs to be
 988  * called after waiting for the request to be unlocked (if it was
 989  * locked).
 990  */
 991 static void end_io_requests(struct fuse_conn *fc)
 992         __releases(fc->lock) __acquires(fc->lock)
 993 {
 994         while (!list_empty(&fc->io)) {
 995                 struct fuse_req *req =
 996                         list_entry(fc->io.next, struct fuse_req, list);
 997                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 998
 999                 req->aborted = 1;
1000                 req->out.h.error = -ECONNABORTED;
1001                 req->state = FUSE_REQ_FINISHED;
1002                 list_del_init(&req->list);
1003                 wake_up(&req->waitq);
1004                 if (end) {
1005                         req->end = NULL;
1006                         /* The end function will consume this reference */
1007                         __fuse_get_request(req);
1008                         spin_unlock(&fc->lock);
1009                         wait_event(req->waitq, !req->locked);
1010                         end(fc, req);
1011                         spin_lock(&fc->lock);
1012                 }
1013         }
1014 }
1015
1016 /*
1017  * Abort all requests.
1018  *
1019  * Emergency exit in case of a malicious or accidental deadlock, or
1020  * just a hung filesystem.
1021  *
1022  * The same effect is usually achievable through killing the
1023  * filesystem daemon and all users of the filesystem.  The exception
1024  * is the combination of an asynchronous request and the tricky
1025  * deadlock (see Documentation/filesystems/fuse.txt).
1026  *
1027  * During the aborting, progression of requests from the pending and
1028  * processing lists onto the io list, and progression of new requests
1029  * onto the pending list is prevented by req->connected being false.
1030  *
1031  * Progression of requests under I/O to the processing list is
1032  * prevented by the req->aborted flag being true for these requests.
1033  * For this reason requests on the io list must be aborted first.
1034  */
1035 void fuse_abort_conn(struct fuse_conn *fc)
1036 {
1037         spin_lock(&fc->lock);
1038         if (fc->connected) {
1039                 fc->connected = 0;
1040                 fc->blocked = 0;
1041                 end_io_requests(fc);
1042                 end_requests(fc, &fc->pending);
1043                 end_requests(fc, &fc->processing);
1044                 wake_up_all(&fc->waitq);
1045                 wake_up_all(&fc->blocked_waitq);
1046                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1047         }
1048         spin_unlock(&fc->lock);
1049 }
1050
1051 static int fuse_dev_release(struct inode *inode, struct file *file)
1052 {
1053         struct fuse_conn *fc = fuse_get_conn(file);
1054         if (fc) {
1055                 spin_lock(&fc->lock);
1056                 fc->connected = 0;
1057                 end_requests(fc, &fc->pending);
1058                 end_requests(fc, &fc->processing);
1059                 spin_unlock(&fc->lock);
1060                 fasync_helper(-1, file, 0, &fc->fasync);
1061                 fuse_conn_put(fc);
1062         }
1063
1064         return 0;
1065 }
1066
1067 static int fuse_dev_fasync(int fd, struct file *file, int on)
1068 {
1069         struct fuse_conn *fc = fuse_get_conn(file);
1070         if (!fc)
1071                 return -EPERM;
1072
1073         /* No locking - fasync_helper does its own locking */
1074         return fasync_helper(fd, file, on, &fc->fasync);
1075 }
1076
1077 const struct file_operations fuse_dev_operations = {
1078         .owner          = THIS_MODULE,
1079         .llseek         = no_llseek,
1080         .read           = do_sync_read,
1081         .aio_read       = fuse_dev_read,
1082         .write          = do_sync_write,
1083         .aio_write      = fuse_dev_write,
1084         .poll           = fuse_dev_poll,
1085         .release        = fuse_dev_release,
1086         .fasync         = fuse_dev_fasync,
1087 };
1088
1089 static struct miscdevice fuse_miscdevice = {
1090         .minor = FUSE_MINOR,
1091         .name  = "fuse",
1092         .fops = &fuse_dev_operations,
1093 };
1094
1095 int __init fuse_dev_init(void)
1096 {
1097         int err = -ENOMEM;
1098         fuse_req_cachep = kmem_cache_create("fuse_request",
1099                                             sizeof(struct fuse_req),
1100                                             0, 0, NULL);
1101         if (!fuse_req_cachep)
1102                 goto out;
1103
1104         err = misc_register(&fuse_miscdevice);
1105         if (err)
1106                 goto out_cache_clean;
1107
1108         return 0;
1109
1110  out_cache_clean:
1111         kmem_cache_destroy(fuse_req_cachep);
1112  out:
1113         return err;
1114 }
1115
1116 void fuse_dev_cleanup(void)
1117 {
1118         misc_deregister(&fuse_miscdevice);
1119         kmem_cache_destroy(fuse_req_cachep);
1120 }