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                     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         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)
 303 __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)
 322 __acquires(&fc->lock)
 323 {
 324         if (!fc->no_interrupt) {
 325                 /* Any signal may interrupt this */
 326                 wait_answer_interruptible(fc, req);
 327
 328                 if (req->aborted)
 329                         goto aborted;
 330                 if (req->state == FUSE_REQ_FINISHED)
 331                         return;
 332
 333                 req->interrupted = 1;
 334                 if (req->state == FUSE_REQ_SENT)
 335                         queue_interrupt(fc, req);
 336         }
 337
 338         if (!req->force) {
 339                 sigset_t oldset;
 340
 341                 /* Only fatal signals may interrupt this */
 342                 block_sigs(&oldset);
 343                 wait_answer_interruptible(fc, req);
 344                 restore_sigs(&oldset);
 345
 346                 if (req->aborted)
 347                         goto aborted;
 348                 if (req->state == FUSE_REQ_FINISHED)
 349                         return;
 350
 351                 /* Request is not yet in userspace, bail out */
 352                 if (req->state == FUSE_REQ_PENDING) {
 353                         list_del(&req->list);
 354                         __fuse_put_request(req);
 355                         req->out.h.error = -EINTR;
 356                         return;
 357                 }
 358         }
 359
 360         /*
 361          * Either request is already in userspace, or it was forced.
 362          * Wait it out.
 363          */
 364         spin_unlock(&fc->lock);
 365         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 366         spin_lock(&fc->lock);
 367
 368         if (!req->aborted)
 369                 return;
 370
 371  aborted:
 372         BUG_ON(req->state != FUSE_REQ_FINISHED);
 373         if (req->locked) {
 374                 /* This is uninterruptible sleep, because data is
 375                    being copied to/from the buffers of req.  During
 376                    locked state, there mustn't be any filesystem
 377                    operation (e.g. page fault), since that could lead
 378                    to deadlock */
 379                 spin_unlock(&fc->lock);
 380                 wait_event(req->waitq, !req->locked);
 381                 spin_lock(&fc->lock);
 382         }
 383 }
 384
 385 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 386 {
 387         req->isreply = 1;
 388         spin_lock(&fc->lock);
 389         if (!fc->connected)
 390                 req->out.h.error = -ENOTCONN;
 391         else if (fc->conn_error)
 392                 req->out.h.error = -ECONNREFUSED;
 393         else {
 394                 queue_request(fc, req);
 395                 /* acquire extra reference, since request is still needed
 396                    after request_end() */
 397                 __fuse_get_request(req);
 398
 399                 request_wait_answer(fc, req);
 400         }
 401         spin_unlock(&fc->lock);
 402 }
 403
 404 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 405                                             struct fuse_req *req)
 406 {
 407         req->background = 1;
 408         fc->num_background++;
 409         if (fc->num_background == FUSE_MAX_BACKGROUND)
 410                 fc->blocked = 1;
 411         if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 412                 set_bdi_congested(&fc->bdi, READ);
 413                 set_bdi_congested(&fc->bdi, WRITE);
 414         }
 415         list_add_tail(&req->list, &fc->bg_queue);
 416         flush_bg_queue(fc);
 417 }
 418
 419 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 420 {
 421         spin_lock(&fc->lock);
 422         if (fc->connected) {
 423                 fuse_request_send_nowait_locked(fc, req);
 424                 spin_unlock(&fc->lock);
 425         } else {
 426                 req->out.h.error = -ENOTCONN;
 427                 request_end(fc, req);
 428         }
 429 }
 430
 431 void fuse_request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 432 {
 433         req->isreply = 0;
 434         fuse_request_send_nowait(fc, req);
 435 }
 436
 437 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 438 {
 439         req->isreply = 1;
 440         fuse_request_send_nowait(fc, req);
 441 }
 442
 443 /*
 444  * Called under fc->lock
 445  *
 446  * fc->connected must have been checked previously
 447  */
 448 void fuse_request_send_background_locked(struct fuse_conn *fc,
 449                                          struct fuse_req *req)
 450 {
 451         req->isreply = 1;
 452         fuse_request_send_nowait_locked(fc, req);
 453 }
 454
 455 /*
 456  * Lock the request.  Up to the next unlock_request() there mustn't be
 457  * anything that could cause a page-fault.  If the request was already
 458  * aborted bail out.
 459  */
 460 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 461 {
 462         int err = 0;
 463         if (req) {
 464                 spin_lock(&fc->lock);
 465                 if (req->aborted)
 466                         err = -ENOENT;
 467                 else
 468                         req->locked = 1;
 469                 spin_unlock(&fc->lock);
 470         }
 471         return err;
 472 }
 473
 474 /*
 475  * Unlock request.  If it was aborted during being locked, the
 476  * requester thread is currently waiting for it to be unlocked, so
 477  * wake it up.
 478  */
 479 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 480 {
 481         if (req) {
 482                 spin_lock(&fc->lock);
 483                 req->locked = 0;
 484                 if (req->aborted)
 485                         wake_up(&req->waitq);
 486                 spin_unlock(&fc->lock);
 487         }
 488 }
 489
 490 struct fuse_copy_state {
 491         struct fuse_conn *fc;
 492         int write;
 493         struct fuse_req *req;
 494         const struct iovec *iov;
 495         unsigned long nr_segs;
 496         unsigned long seglen;
 497         unsigned long addr;
 498         struct page *pg;
 499         void *mapaddr;
 500         void *buf;
 501         unsigned len;
 502 };
 503
 504 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 505                            int write, struct fuse_req *req,
 506                            const struct iovec *iov, unsigned long nr_segs)
 507 {
 508         memset(cs, 0, sizeof(*cs));
 509         cs->fc = fc;
 510         cs->write = write;
 511         cs->req = req;
 512         cs->iov = iov;
 513         cs->nr_segs = nr_segs;
 514 }
 515
 516 /* Unmap and put previous page of userspace buffer */
 517 static void fuse_copy_finish(struct fuse_copy_state *cs)
 518 {
 519         if (cs->mapaddr) {
 520                 kunmap_atomic(cs->mapaddr, KM_USER0);
 521                 if (cs->write) {
 522                         flush_dcache_page(cs->pg);
 523                         set_page_dirty_lock(cs->pg);
 524                 }
 525                 put_page(cs->pg);
 526                 cs->mapaddr = NULL;
 527         }
 528 }
 529
 530 /*
 531  * Get another pagefull of userspace buffer, and map it to kernel
 532  * address space, and lock request
 533  */
 534 static int fuse_copy_fill(struct fuse_copy_state *cs)
 535 {
 536         unsigned long offset;
 537         int err;
 538
 539         unlock_request(cs->fc, cs->req);
 540         fuse_copy_finish(cs);
 541         if (!cs->seglen) {
 542                 BUG_ON(!cs->nr_segs);
 543                 cs->seglen = cs->iov[0].iov_len;
 544                 cs->addr = (unsigned long) cs->iov[0].iov_base;
 545                 cs->iov++;
 546                 cs->nr_segs--;
 547         }
 548         down_read(&current->mm->mmap_sem);
 549         err = get_user_pages(current, current->mm, cs->addr, 1, cs->write, 0,
 550                              &cs->pg, NULL);
 551         up_read(&current->mm->mmap_sem);
 552         if (err < 0)
 553                 return err;
 554         BUG_ON(err != 1);
 555         offset = cs->addr % PAGE_SIZE;
 556         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
 557         cs->buf = cs->mapaddr + offset;
 558         cs->len = min(PAGE_SIZE - offset, cs->seglen);
 559         cs->seglen -= cs->len;
 560         cs->addr += cs->len;
 561
 562         return lock_request(cs->fc, cs->req);
 563 }
 564
 565 /* Do as much copy to/from userspace buffer as we can */
 566 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 567 {
 568         unsigned ncpy = min(*size, cs->len);
 569         if (val) {
 570                 if (cs->write)
 571                         memcpy(cs->buf, *val, ncpy);
 572                 else
 573                         memcpy(*val, cs->buf, ncpy);
 574                 *val += ncpy;
 575         }
 576         *size -= ncpy;
 577         cs->len -= ncpy;
 578         cs->buf += ncpy;
 579         return ncpy;
 580 }
 581
 582 /*
 583  * Copy a page in the request to/from the userspace buffer.  Must be
 584  * done atomically
 585  */
 586 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
 587                           unsigned offset, unsigned count, int zeroing)
 588 {
 589         if (page && zeroing && count < PAGE_SIZE) {
 590                 void *mapaddr = kmap_atomic(page, KM_USER1);
 591                 memset(mapaddr, 0, PAGE_SIZE);
 592                 kunmap_atomic(mapaddr, KM_USER1);
 593         }
 594         while (count) {
 595                 if (!cs->len) {
 596                         int err = fuse_copy_fill(cs);
 597                         if (err)
 598                                 return err;
 599                 }
 600                 if (page) {
 601                         void *mapaddr = kmap_atomic(page, KM_USER1);
 602                         void *buf = mapaddr + offset;
 603                         offset += fuse_copy_do(cs, &buf, &count);
 604                         kunmap_atomic(mapaddr, KM_USER1);
 605                 } else
 606                         offset += fuse_copy_do(cs, NULL, &count);
 607         }
 608         if (page && !cs->write)
 609                 flush_dcache_page(page);
 610         return 0;
 611 }
 612
 613 /* Copy pages in the request to/from userspace buffer */
 614 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 615                            int zeroing)
 616 {
 617         unsigned i;
 618         struct fuse_req *req = cs->req;
 619         unsigned offset = req->page_offset;
 620         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 621
 622         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 623                 struct page *page = req->pages[i];
 624                 int err = fuse_copy_page(cs, page, offset, count, zeroing);
 625                 if (err)
 626                         return err;
 627
 628                 nbytes -= count;
 629                 count = min(nbytes, (unsigned) PAGE_SIZE);
 630                 offset = 0;
 631         }
 632         return 0;
 633 }
 634
 635 /* Copy a single argument in the request to/from userspace buffer */
 636 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 637 {
 638         while (size) {
 639                 if (!cs->len) {
 640                         int err = fuse_copy_fill(cs);
 641                         if (err)
 642                                 return err;
 643                 }
 644                 fuse_copy_do(cs, &val, &size);
 645         }
 646         return 0;
 647 }
 648
 649 /* Copy request arguments to/from userspace buffer */
 650 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 651                           unsigned argpages, struct fuse_arg *args,
 652                           int zeroing)
 653 {
 654         int err = 0;
 655         unsigned i;
 656
 657         for (i = 0; !err && i < numargs; i++)  {
 658                 struct fuse_arg *arg = &args[i];
 659                 if (i == numargs - 1 && argpages)
 660                         err = fuse_copy_pages(cs, arg->size, zeroing);
 661                 else
 662                         err = fuse_copy_one(cs, arg->value, arg->size);
 663         }
 664         return err;
 665 }
 666
 667 static int request_pending(struct fuse_conn *fc)
 668 {
 669         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 670 }
 671
 672 /* Wait until a request is available on the pending list */
 673 static void request_wait(struct fuse_conn *fc)
 674 __releases(&fc->lock)
 675 __acquires(&fc->lock)
 676 {
 677         DECLARE_WAITQUEUE(wait, current);
 678
 679         add_wait_queue_exclusive(&fc->waitq, &wait);
 680         while (fc->connected && !request_pending(fc)) {
 681                 set_current_state(TASK_INTERRUPTIBLE);
 682                 if (signal_pending(current))
 683                         break;
 684
 685                 spin_unlock(&fc->lock);
 686                 schedule();
 687                 spin_lock(&fc->lock);
 688         }
 689         set_current_state(TASK_RUNNING);
 690         remove_wait_queue(&fc->waitq, &wait);
 691 }
 692
 693 /*
 694  * Transfer an interrupt request to userspace
 695  *
 696  * Unlike other requests this is assembled on demand, without a need
 697  * to allocate a separate fuse_req structure.
 698  *
 699  * Called with fc->lock held, releases it
 700  */
 701 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
 702                                const struct iovec *iov, unsigned long nr_segs)
 703 __releases(&fc->lock)
 704 {
 705         struct fuse_copy_state cs;
 706         struct fuse_in_header ih;
 707         struct fuse_interrupt_in arg;
 708         unsigned reqsize = sizeof(ih) + sizeof(arg);
 709         int err;
 710
 711         list_del_init(&req->intr_entry);
 712         req->intr_unique = fuse_get_unique(fc);
 713         memset(&ih, 0, sizeof(ih));
 714         memset(&arg, 0, sizeof(arg));
 715         ih.len = reqsize;
 716         ih.opcode = FUSE_INTERRUPT;
 717         ih.unique = req->intr_unique;
 718         arg.unique = req->in.h.unique;
 719
 720         spin_unlock(&fc->lock);
 721         if (iov_length(iov, nr_segs) < reqsize)
 722                 return -EINVAL;
 723
 724         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
 725         err = fuse_copy_one(&cs, &ih, sizeof(ih));
 726         if (!err)
 727                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
 728         fuse_copy_finish(&cs);
 729
 730         return err ? err : reqsize;
 731 }
 732
 733 /*
 734  * Read a single request into the userspace filesystem's buffer.  This
 735  * function waits until a request is available, then removes it from
 736  * the pending list and copies request data to userspace buffer.  If
 737  * no reply is needed (FORGET) or request has been aborted or there
 738  * was an error during the copying then it's finished by calling
 739  * request_end().  Otherwise add it to the processing list, and set
 740  * the 'sent' flag.
 741  */
 742 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
 743                               unsigned long nr_segs, loff_t pos)
 744 {
 745         int err;
 746         struct fuse_req *req;
 747         struct fuse_in *in;
 748         struct fuse_copy_state cs;
 749         unsigned reqsize;
 750         struct file *file = iocb->ki_filp;
 751         struct fuse_conn *fc = fuse_get_conn(file);
 752         if (!fc)
 753                 return -EPERM;
 754
 755  restart:
 756         spin_lock(&fc->lock);
 757         err = -EAGAIN;
 758         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 759             !request_pending(fc))
 760                 goto err_unlock;
 761
 762         request_wait(fc);
 763         err = -ENODEV;
 764         if (!fc->connected)
 765                 goto err_unlock;
 766         err = -ERESTARTSYS;
 767         if (!request_pending(fc))
 768                 goto err_unlock;
 769
 770         if (!list_empty(&fc->interrupts)) {
 771                 req = list_entry(fc->interrupts.next, struct fuse_req,
 772                                  intr_entry);
 773                 return fuse_read_interrupt(fc, req, iov, nr_segs);
 774         }
 775
 776         req = list_entry(fc->pending.next, struct fuse_req, list);
 777         req->state = FUSE_REQ_READING;
 778         list_move(&req->list, &fc->io);
 779
 780         in = &req->in;
 781         reqsize = in->h.len;
 782         /* If request is too large, reply with an error and restart the read */
 783         if (iov_length(iov, nr_segs) < reqsize) {
 784                 req->out.h.error = -EIO;
 785                 /* SETXATTR is special, since it may contain too large data */
 786                 if (in->h.opcode == FUSE_SETXATTR)
 787                         req->out.h.error = -E2BIG;
 788                 request_end(fc, req);
 789                 goto restart;
 790         }
 791         spin_unlock(&fc->lock);
 792         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
 793         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
 794         if (!err)
 795                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
 796                                      (struct fuse_arg *) in->args, 0);
 797         fuse_copy_finish(&cs);
 798         spin_lock(&fc->lock);
 799         req->locked = 0;
 800         if (req->aborted) {
 801                 request_end(fc, req);
 802                 return -ENODEV;
 803         }
 804         if (err) {
 805                 req->out.h.error = -EIO;
 806                 request_end(fc, req);
 807                 return err;
 808         }
 809         if (!req->isreply)
 810                 request_end(fc, req);
 811         else {
 812                 req->state = FUSE_REQ_SENT;
 813                 list_move_tail(&req->list, &fc->processing);
 814                 if (req->interrupted)
 815                         queue_interrupt(fc, req);
 816                 spin_unlock(&fc->lock);
 817         }
 818         return reqsize;
 819
 820  err_unlock:
 821         spin_unlock(&fc->lock);
 822         return err;
 823 }
 824
 825 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
 826                             struct fuse_copy_state *cs)
 827 {
 828         struct fuse_notify_poll_wakeup_out outarg;
 829         int err = -EINVAL;
 830
 831         if (size != sizeof(outarg))
 832                 goto err;
 833
 834         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
 835         if (err)
 836                 goto err;
 837
 838         fuse_copy_finish(cs);
 839         return fuse_notify_poll_wakeup(fc, &outarg);
 840
 841 err:
 842         fuse_copy_finish(cs);
 843         return err;
 844 }
 845
 846 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
 847                        unsigned int size, struct fuse_copy_state *cs)
 848 {
 849         switch (code) {
 850         case FUSE_NOTIFY_POLL:
 851                 return fuse_notify_poll(fc, size, cs);
 852
 853         default:
 854                 fuse_copy_finish(cs);
 855                 return -EINVAL;
 856         }
 857 }
 858
 859 /* Look up request on processing list by unique ID */
 860 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
 861 {
 862         struct list_head *entry;
 863
 864         list_for_each(entry, &fc->processing) {
 865                 struct fuse_req *req;
 866                 req = list_entry(entry, struct fuse_req, list);
 867                 if (req->in.h.unique == unique || req->intr_unique == unique)
 868                         return req;
 869         }
 870         return NULL;
 871 }
 872
 873 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
 874                          unsigned nbytes)
 875 {
 876         unsigned reqsize = sizeof(struct fuse_out_header);
 877
 878         if (out->h.error)
 879                 return nbytes != reqsize ? -EINVAL : 0;
 880
 881         reqsize += len_args(out->numargs, out->args);
 882
 883         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
 884                 return -EINVAL;
 885         else if (reqsize > nbytes) {
 886                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
 887                 unsigned diffsize = reqsize - nbytes;
 888                 if (diffsize > lastarg->size)
 889                         return -EINVAL;
 890                 lastarg->size -= diffsize;
 891         }
 892         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
 893                               out->page_zeroing);
 894 }
 895
 896 /*
 897  * Write a single reply to a request.  First the header is copied from
 898  * the write buffer.  The request is then searched on the processing
 899  * list by the unique ID found in the header.  If found, then remove
 900  * it from the list and copy the rest of the buffer to the request.
 901  * The request is finished by calling request_end()
 902  */
 903 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
 904                                unsigned long nr_segs, loff_t pos)
 905 {
 906         int err;
 907         unsigned nbytes = iov_length(iov, nr_segs);
 908         struct fuse_req *req;
 909         struct fuse_out_header oh;
 910         struct fuse_copy_state cs;
 911         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
 912         if (!fc)
 913                 return -EPERM;
 914
 915         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
 916         if (nbytes < sizeof(struct fuse_out_header))
 917                 return -EINVAL;
 918
 919         err = fuse_copy_one(&cs, &oh, sizeof(oh));
 920         if (err)
 921                 goto err_finish;
 922
 923         err = -EINVAL;
 924         if (oh.len != nbytes)
 925                 goto err_finish;
 926
 927         /*
 928          * Zero oh.unique indicates unsolicited notification message
 929          * and error contains notification code.
 930          */
 931         if (!oh.unique) {
 932                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), &cs);
 933                 return err ? err : nbytes;
 934         }
 935
 936         err = -EINVAL;
 937         if (oh.error <= -1000 || oh.error > 0)
 938                 goto err_finish;
 939
 940         spin_lock(&fc->lock);
 941         err = -ENOENT;
 942         if (!fc->connected)
 943                 goto err_unlock;
 944
 945         req = request_find(fc, oh.unique);
 946         if (!req)
 947                 goto err_unlock;
 948
 949         if (req->aborted) {
 950                 spin_unlock(&fc->lock);
 951                 fuse_copy_finish(&cs);
 952                 spin_lock(&fc->lock);
 953                 request_end(fc, req);
 954                 return -ENOENT;
 955         }
 956         /* Is it an interrupt reply? */
 957         if (req->intr_unique == oh.unique) {
 958                 err = -EINVAL;
 959                 if (nbytes != sizeof(struct fuse_out_header))
 960                         goto err_unlock;
 961
 962                 if (oh.error == -ENOSYS)
 963                         fc->no_interrupt = 1;
 964                 else if (oh.error == -EAGAIN)
 965                         queue_interrupt(fc, req);
 966
 967                 spin_unlock(&fc->lock);
 968                 fuse_copy_finish(&cs);
 969                 return nbytes;
 970         }
 971
 972         req->state = FUSE_REQ_WRITING;
 973         list_move(&req->list, &fc->io);
 974         req->out.h = oh;
 975         req->locked = 1;
 976         cs.req = req;
 977         spin_unlock(&fc->lock);
 978
 979         err = copy_out_args(&cs, &req->out, nbytes);
 980         fuse_copy_finish(&cs);
 981
 982         spin_lock(&fc->lock);
 983         req->locked = 0;
 984         if (!err) {
 985                 if (req->aborted)
 986                         err = -ENOENT;
 987         } else if (!req->aborted)
 988                 req->out.h.error = -EIO;
 989         request_end(fc, req);
 990
 991         return err ? err : nbytes;
 992
 993  err_unlock:
 994         spin_unlock(&fc->lock);
 995  err_finish:
 996         fuse_copy_finish(&cs);
 997         return err;
 998 }
 999
1000 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1001 {
1002         unsigned mask = POLLOUT | POLLWRNORM;
1003         struct fuse_conn *fc = fuse_get_conn(file);
1004         if (!fc)
1005                 return POLLERR;
1006
1007         poll_wait(file, &fc->waitq, wait);
1008
1009         spin_lock(&fc->lock);
1010         if (!fc->connected)
1011                 mask = POLLERR;
1012         else if (request_pending(fc))
1013                 mask |= POLLIN | POLLRDNORM;
1014         spin_unlock(&fc->lock);
1015
1016         return mask;
1017 }
1018
1019 /*
1020  * Abort all requests on the given list (pending or processing)
1021  *
1022  * This function releases and reacquires fc->lock
1023  */
1024 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1025 __releases(&fc->lock)
1026 __acquires(&fc->lock)
1027 {
1028         while (!list_empty(head)) {
1029                 struct fuse_req *req;
1030                 req = list_entry(head->next, struct fuse_req, list);
1031                 req->out.h.error = -ECONNABORTED;
1032                 request_end(fc, req);
1033                 spin_lock(&fc->lock);
1034         }
1035 }
1036
1037 /*
1038  * Abort requests under I/O
1039  *
1040  * The requests are set to aborted and finished, and the request
1041  * waiter is woken up.  This will make request_wait_answer() wait
1042  * until the request is unlocked and then return.
1043  *
1044  * If the request is asynchronous, then the end function needs to be
1045  * called after waiting for the request to be unlocked (if it was
1046  * locked).
1047  */
1048 static void end_io_requests(struct fuse_conn *fc)
1049 __releases(&fc->lock)
1050 __acquires(&fc->lock)
1051 {
1052         while (!list_empty(&fc->io)) {
1053                 struct fuse_req *req =
1054                         list_entry(fc->io.next, struct fuse_req, list);
1055                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1056
1057                 req->aborted = 1;
1058                 req->out.h.error = -ECONNABORTED;
1059                 req->state = FUSE_REQ_FINISHED;
1060                 list_del_init(&req->list);
1061                 wake_up(&req->waitq);
1062                 if (end) {
1063                         req->end = NULL;
1064                         __fuse_get_request(req);
1065                         spin_unlock(&fc->lock);
1066                         wait_event(req->waitq, !req->locked);
1067                         end(fc, req);
1068                         fuse_put_request(fc, req);
1069                         spin_lock(&fc->lock);
1070                 }
1071         }
1072 }
1073
1074 /*
1075  * Abort all requests.
1076  *
1077  * Emergency exit in case of a malicious or accidental deadlock, or
1078  * just a hung filesystem.
1079  *
1080  * The same effect is usually achievable through killing the
1081  * filesystem daemon and all users of the filesystem.  The exception
1082  * is the combination of an asynchronous request and the tricky
1083  * deadlock (see Documentation/filesystems/fuse.txt).
1084  *
1085  * During the aborting, progression of requests from the pending and
1086  * processing lists onto the io list, and progression of new requests
1087  * onto the pending list is prevented by req->connected being false.
1088  *
1089  * Progression of requests under I/O to the processing list is
1090  * prevented by the req->aborted flag being true for these requests.
1091  * For this reason requests on the io list must be aborted first.
1092  */
1093 void fuse_abort_conn(struct fuse_conn *fc)
1094 {
1095         spin_lock(&fc->lock);
1096         if (fc->connected) {
1097                 fc->connected = 0;
1098                 fc->blocked = 0;
1099                 end_io_requests(fc);
1100                 end_requests(fc, &fc->pending);
1101                 end_requests(fc, &fc->processing);
1102                 wake_up_all(&fc->waitq);
1103                 wake_up_all(&fc->blocked_waitq);
1104                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1105         }
1106         spin_unlock(&fc->lock);
1107 }
1108
1109 static int fuse_dev_release(struct inode *inode, struct file *file)
1110 {
1111         struct fuse_conn *fc = fuse_get_conn(file);
1112         if (fc) {
1113                 spin_lock(&fc->lock);
1114                 fc->connected = 0;
1115                 end_requests(fc, &fc->pending);
1116                 end_requests(fc, &fc->processing);
1117                 spin_unlock(&fc->lock);
1118                 fuse_conn_put(fc);
1119         }
1120
1121         return 0;
1122 }
1123
1124 static int fuse_dev_fasync(int fd, struct file *file, int on)
1125 {
1126         struct fuse_conn *fc = fuse_get_conn(file);
1127         if (!fc)
1128                 return -EPERM;
1129
1130         /* No locking - fasync_helper does its own locking */
1131         return fasync_helper(fd, file, on, &fc->fasync);
1132 }
1133
1134 const struct file_operations fuse_dev_operations = {
1135         .owner          = THIS_MODULE,
1136         .llseek         = no_llseek,
1137         .read           = do_sync_read,
1138         .aio_read       = fuse_dev_read,
1139         .write          = do_sync_write,
1140         .aio_write      = fuse_dev_write,
1141         .poll           = fuse_dev_poll,
1142         .release        = fuse_dev_release,
1143         .fasync         = fuse_dev_fasync,
1144 };
1145
1146 static struct miscdevice fuse_miscdevice = {
1147         .minor = FUSE_MINOR,
1148         .name  = "fuse",
1149         .fops = &fuse_dev_operations,
1150 };
1151
1152 int __init fuse_dev_init(void)
1153 {
1154         int err = -ENOMEM;
1155         fuse_req_cachep = kmem_cache_create("fuse_request",
1156                                             sizeof(struct fuse_req),
1157                                             0, 0, NULL);
1158         if (!fuse_req_cachep)
1159                 goto out;
1160
1161         err = misc_register(&fuse_miscdevice);
1162         if (err)
1163                 goto out_cache_clean;
1164
1165         return 0;
1166
1167  out_cache_clean:
1168         kmem_cache_destroy(fuse_req_cachep);
1169  out:
1170         return err;
1171 }
1172
1173 void fuse_dev_cleanup(void)
1174 {
1175         misc_deregister(&fuse_miscdevice);
1176         kmem_cache_destroy(fuse_req_cachep);
1177 }