fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19
  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         fuse_put_request(fc, req);
 297 }
 298
 299 static void wait_answer_interruptible(struct fuse_conn *fc,
 300                                       struct fuse_req *req)
 301         __releases(fc->lock) __acquires(fc->lock)
 302 {
 303         if (signal_pending(current))
 304                 return;
 305
 306         spin_unlock(&fc->lock);
 307         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 308         spin_lock(&fc->lock);
 309 }
 310
 311 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 312 {
 313         list_add_tail(&req->intr_entry, &fc->interrupts);
 314         wake_up(&fc->waitq);
 315         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 316 }
 317
 318 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 319         __releases(fc->lock) __acquires(fc->lock)
 320 {
 321         if (!fc->no_interrupt) {
 322                 /* Any signal may interrupt this */
 323                 wait_answer_interruptible(fc, req);
 324
 325                 if (req->aborted)
 326                         goto aborted;
 327                 if (req->state == FUSE_REQ_FINISHED)
 328                         return;
 329
 330                 req->interrupted = 1;
 331                 if (req->state == FUSE_REQ_SENT)
 332                         queue_interrupt(fc, req);
 333         }
 334
 335         if (!req->force) {
 336                 sigset_t oldset;
 337
 338                 /* Only fatal signals may interrupt this */
 339                 block_sigs(&oldset);
 340                 wait_answer_interruptible(fc, req);
 341                 restore_sigs(&oldset);
 342
 343                 if (req->aborted)
 344                         goto aborted;
 345                 if (req->state == FUSE_REQ_FINISHED)
 346                         return;
 347
 348                 /* Request is not yet in userspace, bail out */
 349                 if (req->state == FUSE_REQ_PENDING) {
 350                         list_del(&req->list);
 351                         __fuse_put_request(req);
 352                         req->out.h.error = -EINTR;
 353                         return;
 354                 }
 355         }
 356
 357         /*
 358          * Either request is already in userspace, or it was forced.
 359          * Wait it out.
 360          */
 361         spin_unlock(&fc->lock);
 362         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 363         spin_lock(&fc->lock);
 364
 365         if (!req->aborted)
 366                 return;
 367
 368  aborted:
 369         BUG_ON(req->state != FUSE_REQ_FINISHED);
 370         if (req->locked) {
 371                 /* This is uninterruptible sleep, because data is
 372                    being copied to/from the buffers of req.  During
 373                    locked state, there mustn't be any filesystem
 374                    operation (e.g. page fault), since that could lead
 375                    to deadlock */
 376                 spin_unlock(&fc->lock);
 377                 wait_event(req->waitq, !req->locked);
 378                 spin_lock(&fc->lock);
 379         }
 380 }
 381
 382 void request_send(struct fuse_conn *fc, struct fuse_req *req)
 383 {
 384         req->isreply = 1;
 385         spin_lock(&fc->lock);
 386         if (!fc->connected)
 387                 req->out.h.error = -ENOTCONN;
 388         else if (fc->conn_error)
 389                 req->out.h.error = -ECONNREFUSED;
 390         else {
 391                 queue_request(fc, req);
 392                 /* acquire extra reference, since request is still needed
 393                    after request_end() */
 394                 __fuse_get_request(req);
 395
 396                 request_wait_answer(fc, req);
 397         }
 398         spin_unlock(&fc->lock);
 399 }
 400
 401 static void request_send_nowait_locked(struct fuse_conn *fc,
 402                                        struct fuse_req *req)
 403 {
 404         req->background = 1;
 405         fc->num_background++;
 406         if (fc->num_background == FUSE_MAX_BACKGROUND)
 407                 fc->blocked = 1;
 408         if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 409                 set_bdi_congested(&fc->bdi, READ);
 410                 set_bdi_congested(&fc->bdi, WRITE);
 411         }
 412         list_add_tail(&req->list, &fc->bg_queue);
 413         flush_bg_queue(fc);
 414 }
 415
 416 static void request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 417 {
 418         spin_lock(&fc->lock);
 419         if (fc->connected) {
 420                 request_send_nowait_locked(fc, req);
 421                 spin_unlock(&fc->lock);
 422         } else {
 423                 req->out.h.error = -ENOTCONN;
 424                 request_end(fc, req);
 425         }
 426 }
 427
 428 void request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 429 {
 430         req->isreply = 0;
 431         request_send_nowait(fc, req);
 432 }
 433
 434 void request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 435 {
 436         req->isreply = 1;
 437         request_send_nowait(fc, req);
 438 }
 439
 440 /*
 441  * Called under fc->lock
 442  *
 443  * fc->connected must have been checked previously
 444  */
 445 void request_send_background_locked(struct fuse_conn *fc, struct fuse_req *req)
 446 {
 447         req->isreply = 1;
 448         request_send_nowait_locked(fc, req);
 449 }
 450
 451 /*
 452  * Lock the request.  Up to the next unlock_request() there mustn't be
 453  * anything that could cause a page-fault.  If the request was already
 454  * aborted bail out.
 455  */
 456 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 457 {
 458         int err = 0;
 459         if (req) {
 460                 spin_lock(&fc->lock);
 461                 if (req->aborted)
 462                         err = -ENOENT;
 463                 else
 464                         req->locked = 1;
 465                 spin_unlock(&fc->lock);
 466         }
 467         return err;
 468 }
 469
 470 /*
 471  * Unlock request.  If it was aborted during being locked, the
 472  * requester thread is currently waiting for it to be unlocked, so
 473  * wake it up.
 474  */
 475 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 476 {
 477         if (req) {
 478                 spin_lock(&fc->lock);
 479                 req->locked = 0;
 480                 if (req->aborted)
 481                         wake_up(&req->waitq);
 482                 spin_unlock(&fc->lock);
 483         }
 484 }
 485
 486 struct fuse_copy_state {
 487         struct fuse_conn *fc;
 488         int write;
 489         struct fuse_req *req;
 490         const struct iovec *iov;
 491         unsigned long nr_segs;
 492         unsigned long seglen;
 493         unsigned long addr;
 494         struct page *pg;
 495         void *mapaddr;
 496         void *buf;
 497         unsigned len;
 498 };
 499
 500 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 501                            int write, struct fuse_req *req,
 502                            const struct iovec *iov, unsigned long nr_segs)
 503 {
 504         memset(cs, 0, sizeof(*cs));
 505         cs->fc = fc;
 506         cs->write = write;
 507         cs->req = req;
 508         cs->iov = iov;
 509         cs->nr_segs = nr_segs;
 510 }
 511
 512 /* Unmap and put previous page of userspace buffer */
 513 static void fuse_copy_finish(struct fuse_copy_state *cs)
 514 {
 515         if (cs->mapaddr) {
 516                 kunmap_atomic(cs->mapaddr, KM_USER0);
 517                 if (cs->write) {
 518                         flush_dcache_page(cs->pg);
 519                         set_page_dirty_lock(cs->pg);
 520                 }
 521                 put_page(cs->pg);
 522                 cs->mapaddr = NULL;
 523         }
 524 }
 525
 526 /*
 527  * Get another pagefull of userspace buffer, and map it to kernel
 528  * address space, and lock request
 529  */
 530 static int fuse_copy_fill(struct fuse_copy_state *cs)
 531 {
 532         unsigned long offset;
 533         int err;
 534
 535         unlock_request(cs->fc, cs->req);
 536         fuse_copy_finish(cs);
 537         if (!cs->seglen) {
 538                 BUG_ON(!cs->nr_segs);
 539                 cs->seglen = cs->iov[0].iov_len;
 540                 cs->addr = (unsigned long) cs->iov[0].iov_base;
 541                 cs->iov++;
 542                 cs->nr_segs--;
 543         }
 544         down_read(&current->mm->mmap_sem);
 545         err = get_user_pages(current, current->mm, cs->addr, 1, cs->write, 0,
 546                              &cs->pg, NULL);
 547         up_read(&current->mm->mmap_sem);
 548         if (err < 0)
 549                 return err;
 550         BUG_ON(err != 1);
 551         offset = cs->addr % PAGE_SIZE;
 552         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
 553         cs->buf = cs->mapaddr + offset;
 554         cs->len = min(PAGE_SIZE - offset, cs->seglen);
 555         cs->seglen -= cs->len;
 556         cs->addr += cs->len;
 557
 558         return lock_request(cs->fc, cs->req);
 559 }
 560
 561 /* Do as much copy to/from userspace buffer as we can */
 562 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 563 {
 564         unsigned ncpy = min(*size, cs->len);
 565         if (val) {
 566                 if (cs->write)
 567                         memcpy(cs->buf, *val, ncpy);
 568                 else
 569                         memcpy(*val, cs->buf, ncpy);
 570                 *val += ncpy;
 571         }
 572         *size -= ncpy;
 573         cs->len -= ncpy;
 574         cs->buf += ncpy;
 575         return ncpy;
 576 }
 577
 578 /*
 579  * Copy a page in the request to/from the userspace buffer.  Must be
 580  * done atomically
 581  */
 582 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
 583                           unsigned offset, unsigned count, int zeroing)
 584 {
 585         if (page && zeroing && count < PAGE_SIZE) {
 586                 void *mapaddr = kmap_atomic(page, KM_USER1);
 587                 memset(mapaddr, 0, PAGE_SIZE);
 588                 kunmap_atomic(mapaddr, KM_USER1);
 589         }
 590         while (count) {
 591                 if (!cs->len) {
 592                         int err = fuse_copy_fill(cs);
 593                         if (err)
 594                                 return err;
 595                 }
 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                 if (!cs->len) {
 636                         int err = fuse_copy_fill(cs);
 637                         if (err)
 638                                 return err;
 639                 }
 640                 fuse_copy_do(cs, &val, &size);
 641         }
 642         return 0;
 643 }
 644
 645 /* Copy request arguments to/from userspace buffer */
 646 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 647                           unsigned argpages, struct fuse_arg *args,
 648                           int zeroing)
 649 {
 650         int err = 0;
 651         unsigned i;
 652
 653         for (i = 0; !err && i < numargs; i++)  {
 654                 struct fuse_arg *arg = &args[i];
 655                 if (i == numargs - 1 && argpages)
 656                         err = fuse_copy_pages(cs, arg->size, zeroing);
 657                 else
 658                         err = fuse_copy_one(cs, arg->value, arg->size);
 659         }
 660         return err;
 661 }
 662
 663 static int request_pending(struct fuse_conn *fc)
 664 {
 665         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 666 }
 667
 668 /* Wait until a request is available on the pending list */
 669 static void request_wait(struct fuse_conn *fc)
 670 {
 671         DECLARE_WAITQUEUE(wait, current);
 672
 673         add_wait_queue_exclusive(&fc->waitq, &wait);
 674         while (fc->connected && !request_pending(fc)) {
 675                 set_current_state(TASK_INTERRUPTIBLE);
 676                 if (signal_pending(current))
 677                         break;
 678
 679                 spin_unlock(&fc->lock);
 680                 schedule();
 681                 spin_lock(&fc->lock);
 682         }
 683         set_current_state(TASK_RUNNING);
 684         remove_wait_queue(&fc->waitq, &wait);
 685 }
 686
 687 /*
 688  * Transfer an interrupt request to userspace
 689  *
 690  * Unlike other requests this is assembled on demand, without a need
 691  * to allocate a separate fuse_req structure.
 692  *
 693  * Called with fc->lock held, releases it
 694  */
 695 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
 696                                const struct iovec *iov, unsigned long nr_segs)
 697         __releases(fc->lock)
 698 {
 699         struct fuse_copy_state cs;
 700         struct fuse_in_header ih;
 701         struct fuse_interrupt_in arg;
 702         unsigned reqsize = sizeof(ih) + sizeof(arg);
 703         int err;
 704
 705         list_del_init(&req->intr_entry);
 706         req->intr_unique = fuse_get_unique(fc);
 707         memset(&ih, 0, sizeof(ih));
 708         memset(&arg, 0, sizeof(arg));
 709         ih.len = reqsize;
 710         ih.opcode = FUSE_INTERRUPT;
 711         ih.unique = req->intr_unique;
 712         arg.unique = req->in.h.unique;
 713
 714         spin_unlock(&fc->lock);
 715         if (iov_length(iov, nr_segs) < reqsize)
 716                 return -EINVAL;
 717
 718         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
 719         err = fuse_copy_one(&cs, &ih, sizeof(ih));
 720         if (!err)
 721                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
 722         fuse_copy_finish(&cs);
 723
 724         return err ? err : reqsize;
 725 }
 726
 727 /*
 728  * Read a single request into the userspace filesystem's buffer.  This
 729  * function waits until a request is available, then removes it from
 730  * the pending list and copies request data to userspace buffer.  If
 731  * no reply is needed (FORGET) or request has been aborted or there
 732  * was an error during the copying then it's finished by calling
 733  * request_end().  Otherwise add it to the processing list, and set
 734  * the 'sent' flag.
 735  */
 736 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
 737                               unsigned long nr_segs, loff_t pos)
 738 {
 739         int err;
 740         struct fuse_req *req;
 741         struct fuse_in *in;
 742         struct fuse_copy_state cs;
 743         unsigned reqsize;
 744         struct file *file = iocb->ki_filp;
 745         struct fuse_conn *fc = fuse_get_conn(file);
 746         if (!fc)
 747                 return -EPERM;
 748
 749  restart:
 750         spin_lock(&fc->lock);
 751         err = -EAGAIN;
 752         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 753             !request_pending(fc))
 754                 goto err_unlock;
 755
 756         request_wait(fc);
 757         err = -ENODEV;
 758         if (!fc->connected)
 759                 goto err_unlock;
 760         err = -ERESTARTSYS;
 761         if (!request_pending(fc))
 762                 goto err_unlock;
 763
 764         if (!list_empty(&fc->interrupts)) {
 765                 req = list_entry(fc->interrupts.next, struct fuse_req,
 766                                  intr_entry);
 767                 return fuse_read_interrupt(fc, req, iov, nr_segs);
 768         }
 769
 770         req = list_entry(fc->pending.next, struct fuse_req, list);
 771         req->state = FUSE_REQ_READING;
 772         list_move(&req->list, &fc->io);
 773
 774         in = &req->in;
 775         reqsize = in->h.len;
 776         /* If request is too large, reply with an error and restart the read */
 777         if (iov_length(iov, nr_segs) < reqsize) {
 778                 req->out.h.error = -EIO;
 779                 /* SETXATTR is special, since it may contain too large data */
 780                 if (in->h.opcode == FUSE_SETXATTR)
 781                         req->out.h.error = -E2BIG;
 782                 request_end(fc, req);
 783                 goto restart;
 784         }
 785         spin_unlock(&fc->lock);
 786         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
 787         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
 788         if (!err)
 789                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
 790                                      (struct fuse_arg *) in->args, 0);
 791         fuse_copy_finish(&cs);
 792         spin_lock(&fc->lock);
 793         req->locked = 0;
 794         if (req->aborted) {
 795                 request_end(fc, req);
 796                 return -ENODEV;
 797         }
 798         if (err) {
 799                 req->out.h.error = -EIO;
 800                 request_end(fc, req);
 801                 return err;
 802         }
 803         if (!req->isreply)
 804                 request_end(fc, req);
 805         else {
 806                 req->state = FUSE_REQ_SENT;
 807                 list_move_tail(&req->list, &fc->processing);
 808                 if (req->interrupted)
 809                         queue_interrupt(fc, req);
 810                 spin_unlock(&fc->lock);
 811         }
 812         return reqsize;
 813
 814  err_unlock:
 815         spin_unlock(&fc->lock);
 816         return err;
 817 }
 818
 819 /* Look up request on processing list by unique ID */
 820 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
 821 {
 822         struct list_head *entry;
 823
 824         list_for_each(entry, &fc->processing) {
 825                 struct fuse_req *req;
 826                 req = list_entry(entry, struct fuse_req, list);
 827                 if (req->in.h.unique == unique || req->intr_unique == unique)
 828                         return req;
 829         }
 830         return NULL;
 831 }
 832
 833 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
 834                          unsigned nbytes)
 835 {
 836         unsigned reqsize = sizeof(struct fuse_out_header);
 837
 838         if (out->h.error)
 839                 return nbytes != reqsize ? -EINVAL : 0;
 840
 841         reqsize += len_args(out->numargs, out->args);
 842
 843         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
 844                 return -EINVAL;
 845         else if (reqsize > nbytes) {
 846                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
 847                 unsigned diffsize = reqsize - nbytes;
 848                 if (diffsize > lastarg->size)
 849                         return -EINVAL;
 850                 lastarg->size -= diffsize;
 851         }
 852         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
 853                               out->page_zeroing);
 854 }
 855
 856 /*
 857  * Write a single reply to a request.  First the header is copied from
 858  * the write buffer.  The request is then searched on the processing
 859  * list by the unique ID found in the header.  If found, then remove
 860  * it from the list and copy the rest of the buffer to the request.
 861  * The request is finished by calling request_end()
 862  */
 863 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
 864                                unsigned long nr_segs, loff_t pos)
 865 {
 866         int err;
 867         unsigned nbytes = iov_length(iov, nr_segs);
 868         struct fuse_req *req;
 869         struct fuse_out_header oh;
 870         struct fuse_copy_state cs;
 871         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
 872         if (!fc)
 873                 return -EPERM;
 874
 875         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
 876         if (nbytes < sizeof(struct fuse_out_header))
 877                 return -EINVAL;
 878
 879         err = fuse_copy_one(&cs, &oh, sizeof(oh));
 880         if (err)
 881                 goto err_finish;
 882         err = -EINVAL;
 883         if (!oh.unique || oh.error <= -1000 || oh.error > 0 ||
 884             oh.len != nbytes)
 885                 goto err_finish;
 886
 887         spin_lock(&fc->lock);
 888         err = -ENOENT;
 889         if (!fc->connected)
 890                 goto err_unlock;
 891
 892         req = request_find(fc, oh.unique);
 893         if (!req)
 894                 goto err_unlock;
 895
 896         if (req->aborted) {
 897                 spin_unlock(&fc->lock);
 898                 fuse_copy_finish(&cs);
 899                 spin_lock(&fc->lock);
 900                 request_end(fc, req);
 901                 return -ENOENT;
 902         }
 903         /* Is it an interrupt reply? */
 904         if (req->intr_unique == oh.unique) {
 905                 err = -EINVAL;
 906                 if (nbytes != sizeof(struct fuse_out_header))
 907                         goto err_unlock;
 908
 909                 if (oh.error == -ENOSYS)
 910                         fc->no_interrupt = 1;
 911                 else if (oh.error == -EAGAIN)
 912                         queue_interrupt(fc, req);
 913
 914                 spin_unlock(&fc->lock);
 915                 fuse_copy_finish(&cs);
 916                 return nbytes;
 917         }
 918
 919         req->state = FUSE_REQ_WRITING;
 920         list_move(&req->list, &fc->io);
 921         req->out.h = oh;
 922         req->locked = 1;
 923         cs.req = req;
 924         spin_unlock(&fc->lock);
 925
 926         err = copy_out_args(&cs, &req->out, nbytes);
 927         fuse_copy_finish(&cs);
 928
 929         spin_lock(&fc->lock);
 930         req->locked = 0;
 931         if (!err) {
 932                 if (req->aborted)
 933                         err = -ENOENT;
 934         } else if (!req->aborted)
 935                 req->out.h.error = -EIO;
 936         request_end(fc, req);
 937
 938         return err ? err : nbytes;
 939
 940  err_unlock:
 941         spin_unlock(&fc->lock);
 942  err_finish:
 943         fuse_copy_finish(&cs);
 944         return err;
 945 }
 946
 947 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
 948 {
 949         unsigned mask = POLLOUT | POLLWRNORM;
 950         struct fuse_conn *fc = fuse_get_conn(file);
 951         if (!fc)
 952                 return POLLERR;
 953
 954         poll_wait(file, &fc->waitq, wait);
 955
 956         spin_lock(&fc->lock);
 957         if (!fc->connected)
 958                 mask = POLLERR;
 959         else if (request_pending(fc))
 960                 mask |= POLLIN | POLLRDNORM;
 961         spin_unlock(&fc->lock);
 962
 963         return mask;
 964 }
 965
 966 /*
 967  * Abort all requests on the given list (pending or processing)
 968  *
 969  * This function releases and reacquires fc->lock
 970  */
 971 static void end_requests(struct fuse_conn *fc, struct list_head *head)
 972 {
 973         while (!list_empty(head)) {
 974                 struct fuse_req *req;
 975                 req = list_entry(head->next, struct fuse_req, list);
 976                 req->out.h.error = -ECONNABORTED;
 977                 request_end(fc, req);
 978                 spin_lock(&fc->lock);
 979         }
 980 }
 981
 982 /*
 983  * Abort requests under I/O
 984  *
 985  * The requests are set to aborted and finished, and the request
 986  * waiter is woken up.  This will make request_wait_answer() wait
 987  * until the request is unlocked and then return.
 988  *
 989  * If the request is asynchronous, then the end function needs to be
 990  * called after waiting for the request to be unlocked (if it was
 991  * locked).
 992  */
 993 static void end_io_requests(struct fuse_conn *fc)
 994         __releases(fc->lock) __acquires(fc->lock)
 995 {
 996         while (!list_empty(&fc->io)) {
 997                 struct fuse_req *req =
 998                         list_entry(fc->io.next, struct fuse_req, list);
 999                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1000
1001                 req->aborted = 1;
1002                 req->out.h.error = -ECONNABORTED;
1003                 req->state = FUSE_REQ_FINISHED;
1004                 list_del_init(&req->list);
1005                 wake_up(&req->waitq);
1006                 if (end) {
1007                         req->end = NULL;
1008                         __fuse_get_request(req);
1009                         spin_unlock(&fc->lock);
1010                         wait_event(req->waitq, !req->locked);
1011                         end(fc, req);
1012                         fuse_put_request(fc, req);
1013                         spin_lock(&fc->lock);
1014                 }
1015         }
1016 }
1017
1018 /*
1019  * Abort all requests.
1020  *
1021  * Emergency exit in case of a malicious or accidental deadlock, or
1022  * just a hung filesystem.
1023  *
1024  * The same effect is usually achievable through killing the
1025  * filesystem daemon and all users of the filesystem.  The exception
1026  * is the combination of an asynchronous request and the tricky
1027  * deadlock (see Documentation/filesystems/fuse.txt).
1028  *
1029  * During the aborting, progression of requests from the pending and
1030  * processing lists onto the io list, and progression of new requests
1031  * onto the pending list is prevented by req->connected being false.
1032  *
1033  * Progression of requests under I/O to the processing list is
1034  * prevented by the req->aborted flag being true for these requests.
1035  * For this reason requests on the io list must be aborted first.
1036  */
1037 void fuse_abort_conn(struct fuse_conn *fc)
1038 {
1039         spin_lock(&fc->lock);
1040         if (fc->connected) {
1041                 fc->connected = 0;
1042                 fc->blocked = 0;
1043                 end_io_requests(fc);
1044                 end_requests(fc, &fc->pending);
1045                 end_requests(fc, &fc->processing);
1046                 wake_up_all(&fc->waitq);
1047                 wake_up_all(&fc->blocked_waitq);
1048                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1049         }
1050         spin_unlock(&fc->lock);
1051 }
1052
1053 static int fuse_dev_release(struct inode *inode, struct file *file)
1054 {
1055         struct fuse_conn *fc = fuse_get_conn(file);
1056         if (fc) {
1057                 spin_lock(&fc->lock);
1058                 fc->connected = 0;
1059                 end_requests(fc, &fc->pending);
1060                 end_requests(fc, &fc->processing);
1061                 spin_unlock(&fc->lock);
1062                 fuse_conn_put(fc);
1063         }
1064
1065         return 0;
1066 }
1067
1068 static int fuse_dev_fasync(int fd, struct file *file, int on)
1069 {
1070         struct fuse_conn *fc = fuse_get_conn(file);
1071         if (!fc)
1072                 return -EPERM;
1073
1074         /* No locking - fasync_helper does its own locking */
1075         return fasync_helper(fd, file, on, &fc->fasync);
1076 }
1077
1078 const struct file_operations fuse_dev_operations = {
1079         .owner          = THIS_MODULE,
1080         .llseek         = no_llseek,
1081         .read           = do_sync_read,
1082         .aio_read       = fuse_dev_read,
1083         .write          = do_sync_write,
1084         .aio_write      = fuse_dev_write,
1085         .poll           = fuse_dev_poll,
1086         .release        = fuse_dev_release,
1087         .fasync         = fuse_dev_fasync,
1088 };
1089
1090 static struct miscdevice fuse_miscdevice = {
1091         .minor = FUSE_MINOR,
1092         .name  = "fuse",
1093         .fops = &fuse_dev_operations,
1094 };
1095
1096 int __init fuse_dev_init(void)
1097 {
1098         int err = -ENOMEM;
1099         fuse_req_cachep = kmem_cache_create("fuse_request",
1100                                             sizeof(struct fuse_req),
1101                                             0, 0, NULL);
1102         if (!fuse_req_cachep)
1103                 goto out;
1104
1105         err = misc_register(&fuse_miscdevice);
1106         if (err)
1107                 goto out_cache_clean;
1108
1109         return 0;
1110
1111  out_cache_clean:
1112         kmem_cache_destroy(fuse_req_cachep);
1113  out:
1114         return err;
1115 }
1116
1117 void fuse_dev_cleanup(void)
1118 {
1119         misc_deregister(&fuse_miscdevice);
1120         kmem_cache_destroy(fuse_req_cachep);
1121 }