SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  50
  51 struct fuse_req *fuse_request_alloc_nofs(void)
  52 {
  53         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_NOFS);
  54         if (req)
  55                 fuse_request_init(req);
  56         return req;
  57 }
  58
  59 void fuse_request_free(struct fuse_req *req)
  60 {
  61         kmem_cache_free(fuse_req_cachep, req);
  62 }
  63
  64 static void block_sigs(sigset_t *oldset)
  65 {
  66         sigset_t mask;
  67
  68         siginitsetinv(&mask, sigmask(SIGKILL));
  69         sigprocmask(SIG_BLOCK, &mask, oldset);
  70 }
  71
  72 static void restore_sigs(sigset_t *oldset)
  73 {
  74         sigprocmask(SIG_SETMASK, oldset, NULL);
  75 }
  76
  77 static void __fuse_get_request(struct fuse_req *req)
  78 {
  79         atomic_inc(&req->count);
  80 }
  81
  82 /* Must be called with > 1 refcount */
  83 static void __fuse_put_request(struct fuse_req *req)
  84 {
  85         BUG_ON(atomic_read(&req->count) < 2);
  86         atomic_dec(&req->count);
  87 }
  88
  89 static void fuse_req_init_context(struct fuse_req *req)
  90 {
  91         req->in.h.uid = current_fsuid();
  92         req->in.h.gid = current_fsgid();
  93         req->in.h.pid = current->pid;
  94 }
  95
  96 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
  97 {
  98         struct fuse_req *req;
  99         sigset_t oldset;
 100         int intr;
 101         int err;
 102
 103         atomic_inc(&fc->num_waiting);
 104         block_sigs(&oldset);
 105         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
 106         restore_sigs(&oldset);
 107         err = -EINTR;
 108         if (intr)
 109                 goto out;
 110
 111         err = -ENOTCONN;
 112         if (!fc->connected)
 113                 goto out;
 114
 115         req = fuse_request_alloc();
 116         err = -ENOMEM;
 117         if (!req)
 118                 goto out;
 119
 120         fuse_req_init_context(req);
 121         req->waiting = 1;
 122         return req;
 123
 124  out:
 125         atomic_dec(&fc->num_waiting);
 126         return ERR_PTR(err);
 127 }
 128 EXPORT_SYMBOL_GPL(fuse_get_req);
 129
 130 /*
 131  * Return request in fuse_file->reserved_req.  However that may
 132  * currently be in use.  If that is the case, wait for it to become
 133  * available.
 134  */
 135 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 136                                          struct file *file)
 137 {
 138         struct fuse_req *req = NULL;
 139         struct fuse_file *ff = file->private_data;
 140
 141         do {
 142                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 143                 spin_lock(&fc->lock);
 144                 if (ff->reserved_req) {
 145                         req = ff->reserved_req;
 146                         ff->reserved_req = NULL;
 147                         get_file(file);
 148                         req->stolen_file = file;
 149                 }
 150                 spin_unlock(&fc->lock);
 151         } while (!req);
 152
 153         return req;
 154 }
 155
 156 /*
 157  * Put stolen request back into fuse_file->reserved_req
 158  */
 159 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 160 {
 161         struct file *file = req->stolen_file;
 162         struct fuse_file *ff = file->private_data;
 163
 164         spin_lock(&fc->lock);
 165         fuse_request_init(req);
 166         BUG_ON(ff->reserved_req);
 167         ff->reserved_req = req;
 168         wake_up_all(&fc->reserved_req_waitq);
 169         spin_unlock(&fc->lock);
 170         fput(file);
 171 }
 172
 173 /*
 174  * Gets a requests for a file operation, always succeeds
 175  *
 176  * This is used for sending the FLUSH request, which must get to
 177  * userspace, due to POSIX locks which may need to be unlocked.
 178  *
 179  * If allocation fails due to OOM, use the reserved request in
 180  * fuse_file.
 181  *
 182  * This is very unlikely to deadlock accidentally, since the
 183  * filesystem should not have it's own file open.  If deadlock is
 184  * intentional, it can still be broken by "aborting" the filesystem.
 185  */
 186 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 187 {
 188         struct fuse_req *req;
 189
 190         atomic_inc(&fc->num_waiting);
 191         wait_event(fc->blocked_waitq, !fc->blocked);
 192         req = fuse_request_alloc();
 193         if (!req)
 194                 req = get_reserved_req(fc, file);
 195
 196         fuse_req_init_context(req);
 197         req->waiting = 1;
 198         return req;
 199 }
 200
 201 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 202 {
 203         if (atomic_dec_and_test(&req->count)) {
 204                 if (req->waiting)
 205                         atomic_dec(&fc->num_waiting);
 206
 207                 if (req->stolen_file)
 208                         put_reserved_req(fc, req);
 209                 else
 210                         fuse_request_free(req);
 211         }
 212 }
 213 EXPORT_SYMBOL_GPL(fuse_put_request);
 214
 215 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 216 {
 217         unsigned nbytes = 0;
 218         unsigned i;
 219
 220         for (i = 0; i < numargs; i++)
 221                 nbytes += args[i].size;
 222
 223         return nbytes;
 224 }
 225
 226 static u64 fuse_get_unique(struct fuse_conn *fc)
 227 {
 228         fc->reqctr++;
 229         /* zero is special */
 230         if (fc->reqctr == 0)
 231                 fc->reqctr = 1;
 232
 233         return fc->reqctr;
 234 }
 235
 236 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 237 {
 238         req->in.h.unique = fuse_get_unique(fc);
 239         req->in.h.len = sizeof(struct fuse_in_header) +
 240                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 241         list_add_tail(&req->list, &fc->pending);
 242         req->state = FUSE_REQ_PENDING;
 243         if (!req->waiting) {
 244                 req->waiting = 1;
 245                 atomic_inc(&fc->num_waiting);
 246         }
 247         wake_up(&fc->waitq);
 248         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 249 }
 250
 251 static void flush_bg_queue(struct fuse_conn *fc)
 252 {
 253         while (fc->active_background < fc->max_background &&
 254                !list_empty(&fc->bg_queue)) {
 255                 struct fuse_req *req;
 256
 257                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 258                 list_del(&req->list);
 259                 fc->active_background++;
 260                 queue_request(fc, req);
 261         }
 262 }
 263
 264 /*
 265  * This function is called when a request is finished.  Either a reply
 266  * has arrived or it was aborted (and not yet sent) or some error
 267  * occurred during communication with userspace, or the device file
 268  * was closed.  The requester thread is woken up (if still waiting),
 269  * the 'end' callback is called if given, else the reference to the
 270  * request is released
 271  *
 272  * Called with fc->lock, unlocks it
 273  */
 274 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 275 __releases(&fc->lock)
 276 {
 277         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 278         req->end = NULL;
 279         list_del(&req->list);
 280         list_del(&req->intr_entry);
 281         req->state = FUSE_REQ_FINISHED;
 282         if (req->background) {
 283                 if (fc->num_background == fc->max_background) {
 284                         fc->blocked = 0;
 285                         wake_up_all(&fc->blocked_waitq);
 286                 }
 287                 if (fc->num_background == fc->congestion_threshold &&
 288                     fc->connected && fc->bdi_initialized) {
 289                         clear_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 290                         clear_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 291                 }
 292                 fc->num_background--;
 293                 fc->active_background--;
 294                 flush_bg_queue(fc);
 295         }
 296         spin_unlock(&fc->lock);
 297         wake_up(&req->waitq);
 298         if (end)
 299                 end(fc, req);
 300         fuse_put_request(fc, req);
 301 }
 302
 303 static void wait_answer_interruptible(struct fuse_conn *fc,
 304                                       struct fuse_req *req)
 305 __releases(&fc->lock)
 306 __acquires(&fc->lock)
 307 {
 308         if (signal_pending(current))
 309                 return;
 310
 311         spin_unlock(&fc->lock);
 312         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 313         spin_lock(&fc->lock);
 314 }
 315
 316 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 317 {
 318         list_add_tail(&req->intr_entry, &fc->interrupts);
 319         wake_up(&fc->waitq);
 320         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 321 }
 322
 323 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 324 __releases(&fc->lock)
 325 __acquires(&fc->lock)
 326 {
 327         if (!fc->no_interrupt) {
 328                 /* Any signal may interrupt this */
 329                 wait_answer_interruptible(fc, req);
 330
 331                 if (req->aborted)
 332                         goto aborted;
 333                 if (req->state == FUSE_REQ_FINISHED)
 334                         return;
 335
 336                 req->interrupted = 1;
 337                 if (req->state == FUSE_REQ_SENT)
 338                         queue_interrupt(fc, req);
 339         }
 340
 341         if (!req->force) {
 342                 sigset_t oldset;
 343
 344                 /* Only fatal signals may interrupt this */
 345                 block_sigs(&oldset);
 346                 wait_answer_interruptible(fc, req);
 347                 restore_sigs(&oldset);
 348
 349                 if (req->aborted)
 350                         goto aborted;
 351                 if (req->state == FUSE_REQ_FINISHED)
 352                         return;
 353
 354                 /* Request is not yet in userspace, bail out */
 355                 if (req->state == FUSE_REQ_PENDING) {
 356                         list_del(&req->list);
 357                         __fuse_put_request(req);
 358                         req->out.h.error = -EINTR;
 359                         return;
 360                 }
 361         }
 362
 363         /*
 364          * Either request is already in userspace, or it was forced.
 365          * Wait it out.
 366          */
 367         spin_unlock(&fc->lock);
 368         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 369         spin_lock(&fc->lock);
 370
 371         if (!req->aborted)
 372                 return;
 373
 374  aborted:
 375         BUG_ON(req->state != FUSE_REQ_FINISHED);
 376         if (req->locked) {
 377                 /* This is uninterruptible sleep, because data is
 378                    being copied to/from the buffers of req.  During
 379                    locked state, there mustn't be any filesystem
 380                    operation (e.g. page fault), since that could lead
 381                    to deadlock */
 382                 spin_unlock(&fc->lock);
 383                 wait_event(req->waitq, !req->locked);
 384                 spin_lock(&fc->lock);
 385         }
 386 }
 387
 388 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 389 {
 390         req->isreply = 1;
 391         spin_lock(&fc->lock);
 392         if (!fc->connected)
 393                 req->out.h.error = -ENOTCONN;
 394         else if (fc->conn_error)
 395                 req->out.h.error = -ECONNREFUSED;
 396         else {
 397                 queue_request(fc, req);
 398                 /* acquire extra reference, since request is still needed
 399                    after request_end() */
 400                 __fuse_get_request(req);
 401
 402                 request_wait_answer(fc, req);
 403         }
 404         spin_unlock(&fc->lock);
 405 }
 406 EXPORT_SYMBOL_GPL(fuse_request_send);
 407
 408 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 409                                             struct fuse_req *req)
 410 {
 411         req->background = 1;
 412         fc->num_background++;
 413         if (fc->num_background == fc->max_background)
 414                 fc->blocked = 1;
 415         if (fc->num_background == fc->congestion_threshold &&
 416             fc->bdi_initialized) {
 417                 set_bdi_congested(&fc->bdi, BLK_RW_SYNC);
 418                 set_bdi_congested(&fc->bdi, BLK_RW_ASYNC);
 419         }
 420         list_add_tail(&req->list, &fc->bg_queue);
 421         flush_bg_queue(fc);
 422 }
 423
 424 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 425 {
 426         spin_lock(&fc->lock);
 427         if (fc->connected) {
 428                 fuse_request_send_nowait_locked(fc, req);
 429                 spin_unlock(&fc->lock);
 430         } else {
 431                 req->out.h.error = -ENOTCONN;
 432                 request_end(fc, req);
 433         }
 434 }
 435
 436 void fuse_request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 437 {
 438         req->isreply = 0;
 439         fuse_request_send_nowait(fc, req);
 440 }
 441
 442 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 443 {
 444         req->isreply = 1;
 445         fuse_request_send_nowait(fc, req);
 446 }
 447 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 448
 449 /*
 450  * Called under fc->lock
 451  *
 452  * fc->connected must have been checked previously
 453  */
 454 void fuse_request_send_background_locked(struct fuse_conn *fc,
 455                                          struct fuse_req *req)
 456 {
 457         req->isreply = 1;
 458         fuse_request_send_nowait_locked(fc, req);
 459 }
 460
 461 /*
 462  * Lock the request.  Up to the next unlock_request() there mustn't be
 463  * anything that could cause a page-fault.  If the request was already
 464  * aborted bail out.
 465  */
 466 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 467 {
 468         int err = 0;
 469         if (req) {
 470                 spin_lock(&fc->lock);
 471                 if (req->aborted)
 472                         err = -ENOENT;
 473                 else
 474                         req->locked = 1;
 475                 spin_unlock(&fc->lock);
 476         }
 477         return err;
 478 }
 479
 480 /*
 481  * Unlock request.  If it was aborted during being locked, the
 482  * requester thread is currently waiting for it to be unlocked, so
 483  * wake it up.
 484  */
 485 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 486 {
 487         if (req) {
 488                 spin_lock(&fc->lock);
 489                 req->locked = 0;
 490                 if (req->aborted)
 491                         wake_up(&req->waitq);
 492                 spin_unlock(&fc->lock);
 493         }
 494 }
 495
 496 struct fuse_copy_state {
 497         struct fuse_conn *fc;
 498         int write;
 499         struct fuse_req *req;
 500         const struct iovec *iov;
 501         unsigned long nr_segs;
 502         unsigned long seglen;
 503         unsigned long addr;
 504         struct page *pg;
 505         void *mapaddr;
 506         void *buf;
 507         unsigned len;
 508 };
 509
 510 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 511                            int write, struct fuse_req *req,
 512                            const struct iovec *iov, unsigned long nr_segs)
 513 {
 514         memset(cs, 0, sizeof(*cs));
 515         cs->fc = fc;
 516         cs->write = write;
 517         cs->req = req;
 518         cs->iov = iov;
 519         cs->nr_segs = nr_segs;
 520 }
 521
 522 /* Unmap and put previous page of userspace buffer */
 523 static void fuse_copy_finish(struct fuse_copy_state *cs)
 524 {
 525         if (cs->mapaddr) {
 526                 kunmap_atomic(cs->mapaddr, KM_USER0);
 527                 if (cs->write) {
 528                         flush_dcache_page(cs->pg);
 529                         set_page_dirty_lock(cs->pg);
 530                 }
 531                 put_page(cs->pg);
 532                 cs->mapaddr = NULL;
 533         }
 534 }
 535
 536 /*
 537  * Get another pagefull of userspace buffer, and map it to kernel
 538  * address space, and lock request
 539  */
 540 static int fuse_copy_fill(struct fuse_copy_state *cs)
 541 {
 542         unsigned long offset;
 543         int err;
 544
 545         unlock_request(cs->fc, cs->req);
 546         fuse_copy_finish(cs);
 547         if (!cs->seglen) {
 548                 BUG_ON(!cs->nr_segs);
 549                 cs->seglen = cs->iov[0].iov_len;
 550                 cs->addr = (unsigned long) cs->iov[0].iov_base;
 551                 cs->iov++;
 552                 cs->nr_segs--;
 553         }
 554         err = get_user_pages_fast(cs->addr, 1, cs->write, &cs->pg);
 555         if (err < 0)
 556                 return err;
 557         BUG_ON(err != 1);
 558         offset = cs->addr % PAGE_SIZE;
 559         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
 560         cs->buf = cs->mapaddr + offset;
 561         cs->len = min(PAGE_SIZE - offset, cs->seglen);
 562         cs->seglen -= cs->len;
 563         cs->addr += cs->len;
 564
 565         return lock_request(cs->fc, cs->req);
 566 }
 567
 568 /* Do as much copy to/from userspace buffer as we can */
 569 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 570 {
 571         unsigned ncpy = min(*size, cs->len);
 572         if (val) {
 573                 if (cs->write)
 574                         memcpy(cs->buf, *val, ncpy);
 575                 else
 576                         memcpy(*val, cs->buf, ncpy);
 577                 *val += ncpy;
 578         }
 579         *size -= ncpy;
 580         cs->len -= ncpy;
 581         cs->buf += ncpy;
 582         return ncpy;
 583 }
 584
 585 /*
 586  * Copy a page in the request to/from the userspace buffer.  Must be
 587  * done atomically
 588  */
 589 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
 590                           unsigned offset, unsigned count, int zeroing)
 591 {
 592         if (page && zeroing && count < PAGE_SIZE) {
 593                 void *mapaddr = kmap_atomic(page, KM_USER1);
 594                 memset(mapaddr, 0, PAGE_SIZE);
 595                 kunmap_atomic(mapaddr, KM_USER1);
 596         }
 597         while (count) {
 598                 if (!cs->len) {
 599                         int err = fuse_copy_fill(cs);
 600                         if (err)
 601                                 return err;
 602                 }
 603                 if (page) {
 604                         void *mapaddr = kmap_atomic(page, KM_USER1);
 605                         void *buf = mapaddr + offset;
 606                         offset += fuse_copy_do(cs, &buf, &count);
 607                         kunmap_atomic(mapaddr, KM_USER1);
 608                 } else
 609                         offset += fuse_copy_do(cs, NULL, &count);
 610         }
 611         if (page && !cs->write)
 612                 flush_dcache_page(page);
 613         return 0;
 614 }
 615
 616 /* Copy pages in the request to/from userspace buffer */
 617 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 618                            int zeroing)
 619 {
 620         unsigned i;
 621         struct fuse_req *req = cs->req;
 622         unsigned offset = req->page_offset;
 623         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 624
 625         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 626                 struct page *page = req->pages[i];
 627                 int err = fuse_copy_page(cs, page, offset, count, zeroing);
 628                 if (err)
 629                         return err;
 630
 631                 nbytes -= count;
 632                 count = min(nbytes, (unsigned) PAGE_SIZE);
 633                 offset = 0;
 634         }
 635         return 0;
 636 }
 637
 638 /* Copy a single argument in the request to/from userspace buffer */
 639 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 640 {
 641         while (size) {
 642                 if (!cs->len) {
 643                         int err = fuse_copy_fill(cs);
 644                         if (err)
 645                                 return err;
 646                 }
 647                 fuse_copy_do(cs, &val, &size);
 648         }
 649         return 0;
 650 }
 651
 652 /* Copy request arguments to/from userspace buffer */
 653 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 654                           unsigned argpages, struct fuse_arg *args,
 655                           int zeroing)
 656 {
 657         int err = 0;
 658         unsigned i;
 659
 660         for (i = 0; !err && i < numargs; i++)  {
 661                 struct fuse_arg *arg = &args[i];
 662                 if (i == numargs - 1 && argpages)
 663                         err = fuse_copy_pages(cs, arg->size, zeroing);
 664                 else
 665                         err = fuse_copy_one(cs, arg->value, arg->size);
 666         }
 667         return err;
 668 }
 669
 670 static int request_pending(struct fuse_conn *fc)
 671 {
 672         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 673 }
 674
 675 /* Wait until a request is available on the pending list */
 676 static void request_wait(struct fuse_conn *fc)
 677 __releases(&fc->lock)
 678 __acquires(&fc->lock)
 679 {
 680         DECLARE_WAITQUEUE(wait, current);
 681
 682         add_wait_queue_exclusive(&fc->waitq, &wait);
 683         while (fc->connected && !request_pending(fc)) {
 684                 set_current_state(TASK_INTERRUPTIBLE);
 685                 if (signal_pending(current))
 686                         break;
 687
 688                 spin_unlock(&fc->lock);
 689                 schedule();
 690                 spin_lock(&fc->lock);
 691         }
 692         set_current_state(TASK_RUNNING);
 693         remove_wait_queue(&fc->waitq, &wait);
 694 }
 695
 696 /*
 697  * Transfer an interrupt request to userspace
 698  *
 699  * Unlike other requests this is assembled on demand, without a need
 700  * to allocate a separate fuse_req structure.
 701  *
 702  * Called with fc->lock held, releases it
 703  */
 704 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
 705                                const struct iovec *iov, unsigned long nr_segs)
 706 __releases(&fc->lock)
 707 {
 708         struct fuse_copy_state cs;
 709         struct fuse_in_header ih;
 710         struct fuse_interrupt_in arg;
 711         unsigned reqsize = sizeof(ih) + sizeof(arg);
 712         int err;
 713
 714         list_del_init(&req->intr_entry);
 715         req->intr_unique = fuse_get_unique(fc);
 716         memset(&ih, 0, sizeof(ih));
 717         memset(&arg, 0, sizeof(arg));
 718         ih.len = reqsize;
 719         ih.opcode = FUSE_INTERRUPT;
 720         ih.unique = req->intr_unique;
 721         arg.unique = req->in.h.unique;
 722
 723         spin_unlock(&fc->lock);
 724         if (iov_length(iov, nr_segs) < reqsize)
 725                 return -EINVAL;
 726
 727         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
 728         err = fuse_copy_one(&cs, &ih, sizeof(ih));
 729         if (!err)
 730                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
 731         fuse_copy_finish(&cs);
 732
 733         return err ? err : reqsize;
 734 }
 735
 736 /*
 737  * Read a single request into the userspace filesystem's buffer.  This
 738  * function waits until a request is available, then removes it from
 739  * the pending list and copies request data to userspace buffer.  If
 740  * no reply is needed (FORGET) or request has been aborted or there
 741  * was an error during the copying then it's finished by calling
 742  * request_end().  Otherwise add it to the processing list, and set
 743  * the 'sent' flag.
 744  */
 745 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
 746                               unsigned long nr_segs, loff_t pos)
 747 {
 748         int err;
 749         struct fuse_req *req;
 750         struct fuse_in *in;
 751         struct fuse_copy_state cs;
 752         unsigned reqsize;
 753         struct file *file = iocb->ki_filp;
 754         struct fuse_conn *fc = fuse_get_conn(file);
 755         if (!fc)
 756                 return -EPERM;
 757
 758  restart:
 759         spin_lock(&fc->lock);
 760         err = -EAGAIN;
 761         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 762             !request_pending(fc))
 763                 goto err_unlock;
 764
 765         request_wait(fc);
 766         err = -ENODEV;
 767         if (!fc->connected)
 768                 goto err_unlock;
 769         err = -ERESTARTSYS;
 770         if (!request_pending(fc))
 771                 goto err_unlock;
 772
 773         if (!list_empty(&fc->interrupts)) {
 774                 req = list_entry(fc->interrupts.next, struct fuse_req,
 775                                  intr_entry);
 776                 return fuse_read_interrupt(fc, req, iov, nr_segs);
 777         }
 778
 779         req = list_entry(fc->pending.next, struct fuse_req, list);
 780         req->state = FUSE_REQ_READING;
 781         list_move(&req->list, &fc->io);
 782
 783         in = &req->in;
 784         reqsize = in->h.len;
 785         /* If request is too large, reply with an error and restart the read */
 786         if (iov_length(iov, nr_segs) < reqsize) {
 787                 req->out.h.error = -EIO;
 788                 /* SETXATTR is special, since it may contain too large data */
 789                 if (in->h.opcode == FUSE_SETXATTR)
 790                         req->out.h.error = -E2BIG;
 791                 request_end(fc, req);
 792                 goto restart;
 793         }
 794         spin_unlock(&fc->lock);
 795         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
 796         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
 797         if (!err)
 798                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
 799                                      (struct fuse_arg *) in->args, 0);
 800         fuse_copy_finish(&cs);
 801         spin_lock(&fc->lock);
 802         req->locked = 0;
 803         if (req->aborted) {
 804                 request_end(fc, req);
 805                 return -ENODEV;
 806         }
 807         if (err) {
 808                 req->out.h.error = -EIO;
 809                 request_end(fc, req);
 810                 return err;
 811         }
 812         if (!req->isreply)
 813                 request_end(fc, req);
 814         else {
 815                 req->state = FUSE_REQ_SENT;
 816                 list_move_tail(&req->list, &fc->processing);
 817                 if (req->interrupted)
 818                         queue_interrupt(fc, req);
 819                 spin_unlock(&fc->lock);
 820         }
 821         return reqsize;
 822
 823  err_unlock:
 824         spin_unlock(&fc->lock);
 825         return err;
 826 }
 827
 828 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
 829                             struct fuse_copy_state *cs)
 830 {
 831         struct fuse_notify_poll_wakeup_out outarg;
 832         int err = -EINVAL;
 833
 834         if (size != sizeof(outarg))
 835                 goto err;
 836
 837         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
 838         if (err)
 839                 goto err;
 840
 841         fuse_copy_finish(cs);
 842         return fuse_notify_poll_wakeup(fc, &outarg);
 843
 844 err:
 845         fuse_copy_finish(cs);
 846         return err;
 847 }
 848
 849 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
 850                                    struct fuse_copy_state *cs)
 851 {
 852         struct fuse_notify_inval_inode_out outarg;
 853         int err = -EINVAL;
 854
 855         if (size != sizeof(outarg))
 856                 goto err;
 857
 858         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
 859         if (err)
 860                 goto err;
 861         fuse_copy_finish(cs);
 862
 863         down_read(&fc->killsb);
 864         err = -ENOENT;
 865         if (fc->sb) {
 866                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
 867                                                outarg.off, outarg.len);
 868         }
 869         up_read(&fc->killsb);
 870         return err;
 871
 872 err:
 873         fuse_copy_finish(cs);
 874         return err;
 875 }
 876
 877 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
 878                                    struct fuse_copy_state *cs)
 879 {
 880         struct fuse_notify_inval_entry_out outarg;
 881         int err = -ENOMEM;
 882         char *buf;
 883         struct qstr name;
 884
 885         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
 886         if (!buf)
 887                 goto err;
 888
 889         err = -EINVAL;
 890         if (size < sizeof(outarg))
 891                 goto err;
 892
 893         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
 894         if (err)
 895                 goto err;
 896
 897         err = -ENAMETOOLONG;
 898         if (outarg.namelen > FUSE_NAME_MAX)
 899                 goto err;
 900
 901         name.name = buf;
 902         name.len = outarg.namelen;
 903         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
 904         if (err)
 905                 goto err;
 906         fuse_copy_finish(cs);
 907         buf[outarg.namelen] = 0;
 908         name.hash = full_name_hash(name.name, name.len);
 909
 910         down_read(&fc->killsb);
 911         err = -ENOENT;
 912         if (fc->sb)
 913                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, &name);
 914         up_read(&fc->killsb);
 915         kfree(buf);
 916         return err;
 917
 918 err:
 919         kfree(buf);
 920         fuse_copy_finish(cs);
 921         return err;
 922 }
 923
 924 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
 925                        unsigned int size, struct fuse_copy_state *cs)
 926 {
 927         switch (code) {
 928         case FUSE_NOTIFY_POLL:
 929                 return fuse_notify_poll(fc, size, cs);
 930
 931         case FUSE_NOTIFY_INVAL_INODE:
 932                 return fuse_notify_inval_inode(fc, size, cs);
 933
 934         case FUSE_NOTIFY_INVAL_ENTRY:
 935                 return fuse_notify_inval_entry(fc, size, cs);
 936
 937         default:
 938                 fuse_copy_finish(cs);
 939                 return -EINVAL;
 940         }
 941 }
 942
 943 /* Look up request on processing list by unique ID */
 944 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
 945 {
 946         struct list_head *entry;
 947
 948         list_for_each(entry, &fc->processing) {
 949                 struct fuse_req *req;
 950                 req = list_entry(entry, struct fuse_req, list);
 951                 if (req->in.h.unique == unique || req->intr_unique == unique)
 952                         return req;
 953         }
 954         return NULL;
 955 }
 956
 957 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
 958                          unsigned nbytes)
 959 {
 960         unsigned reqsize = sizeof(struct fuse_out_header);
 961
 962         if (out->h.error)
 963                 return nbytes != reqsize ? -EINVAL : 0;
 964
 965         reqsize += len_args(out->numargs, out->args);
 966
 967         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
 968                 return -EINVAL;
 969         else if (reqsize > nbytes) {
 970                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
 971                 unsigned diffsize = reqsize - nbytes;
 972                 if (diffsize > lastarg->size)
 973                         return -EINVAL;
 974                 lastarg->size -= diffsize;
 975         }
 976         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
 977                               out->page_zeroing);
 978 }
 979
 980 /*
 981  * Write a single reply to a request.  First the header is copied from
 982  * the write buffer.  The request is then searched on the processing
 983  * list by the unique ID found in the header.  If found, then remove
 984  * it from the list and copy the rest of the buffer to the request.
 985  * The request is finished by calling request_end()
 986  */
 987 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
 988                                unsigned long nr_segs, loff_t pos)
 989 {
 990         int err;
 991         size_t nbytes = iov_length(iov, nr_segs);
 992         struct fuse_req *req;
 993         struct fuse_out_header oh;
 994         struct fuse_copy_state cs;
 995         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
 996         if (!fc)
 997                 return -EPERM;
 998
 999         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
1000         if (nbytes < sizeof(struct fuse_out_header))
1001                 return -EINVAL;
1002
1003         err = fuse_copy_one(&cs, &oh, sizeof(oh));
1004         if (err)
1005                 goto err_finish;
1006
1007         err = -EINVAL;
1008         if (oh.len != nbytes)
1009                 goto err_finish;
1010
1011         /*
1012          * Zero oh.unique indicates unsolicited notification message
1013          * and error contains notification code.
1014          */
1015         if (!oh.unique) {
1016                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), &cs);
1017                 return err ? err : nbytes;
1018         }
1019
1020         err = -EINVAL;
1021         if (oh.error <= -1000 || oh.error > 0)
1022                 goto err_finish;
1023
1024         spin_lock(&fc->lock);
1025         err = -ENOENT;
1026         if (!fc->connected)
1027                 goto err_unlock;
1028
1029         req = request_find(fc, oh.unique);
1030         if (!req)
1031                 goto err_unlock;
1032
1033         if (req->aborted) {
1034                 spin_unlock(&fc->lock);
1035                 fuse_copy_finish(&cs);
1036                 spin_lock(&fc->lock);
1037                 request_end(fc, req);
1038                 return -ENOENT;
1039         }
1040         /* Is it an interrupt reply? */
1041         if (req->intr_unique == oh.unique) {
1042                 err = -EINVAL;
1043                 if (nbytes != sizeof(struct fuse_out_header))
1044                         goto err_unlock;
1045
1046                 if (oh.error == -ENOSYS)
1047                         fc->no_interrupt = 1;
1048                 else if (oh.error == -EAGAIN)
1049                         queue_interrupt(fc, req);
1050
1051                 spin_unlock(&fc->lock);
1052                 fuse_copy_finish(&cs);
1053                 return nbytes;
1054         }
1055
1056         req->state = FUSE_REQ_WRITING;
1057         list_move(&req->list, &fc->io);
1058         req->out.h = oh;
1059         req->locked = 1;
1060         cs.req = req;
1061         spin_unlock(&fc->lock);
1062
1063         err = copy_out_args(&cs, &req->out, nbytes);
1064         fuse_copy_finish(&cs);
1065
1066         spin_lock(&fc->lock);
1067         req->locked = 0;
1068         if (!err) {
1069                 if (req->aborted)
1070                         err = -ENOENT;
1071         } else if (!req->aborted)
1072                 req->out.h.error = -EIO;
1073         request_end(fc, req);
1074
1075         return err ? err : nbytes;
1076
1077  err_unlock:
1078         spin_unlock(&fc->lock);
1079  err_finish:
1080         fuse_copy_finish(&cs);
1081         return err;
1082 }
1083
1084 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1085 {
1086         unsigned mask = POLLOUT | POLLWRNORM;
1087         struct fuse_conn *fc = fuse_get_conn(file);
1088         if (!fc)
1089                 return POLLERR;
1090
1091         poll_wait(file, &fc->waitq, wait);
1092
1093         spin_lock(&fc->lock);
1094         if (!fc->connected)
1095                 mask = POLLERR;
1096         else if (request_pending(fc))
1097                 mask |= POLLIN | POLLRDNORM;
1098         spin_unlock(&fc->lock);
1099
1100         return mask;
1101 }
1102
1103 /*
1104  * Abort all requests on the given list (pending or processing)
1105  *
1106  * This function releases and reacquires fc->lock
1107  */
1108 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1109 __releases(&fc->lock)
1110 __acquires(&fc->lock)
1111 {
1112         while (!list_empty(head)) {
1113                 struct fuse_req *req;
1114                 req = list_entry(head->next, struct fuse_req, list);
1115                 req->out.h.error = -ECONNABORTED;
1116                 request_end(fc, req);
1117                 spin_lock(&fc->lock);
1118         }
1119 }
1120
1121 /*
1122  * Abort requests under I/O
1123  *
1124  * The requests are set to aborted and finished, and the request
1125  * waiter is woken up.  This will make request_wait_answer() wait
1126  * until the request is unlocked and then return.
1127  *
1128  * If the request is asynchronous, then the end function needs to be
1129  * called after waiting for the request to be unlocked (if it was
1130  * locked).
1131  */
1132 static void end_io_requests(struct fuse_conn *fc)
1133 __releases(&fc->lock)
1134 __acquires(&fc->lock)
1135 {
1136         while (!list_empty(&fc->io)) {
1137                 struct fuse_req *req =
1138                         list_entry(fc->io.next, struct fuse_req, list);
1139                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1140
1141                 req->aborted = 1;
1142                 req->out.h.error = -ECONNABORTED;
1143                 req->state = FUSE_REQ_FINISHED;
1144                 list_del_init(&req->list);
1145                 wake_up(&req->waitq);
1146                 if (end) {
1147                         req->end = NULL;
1148                         __fuse_get_request(req);
1149                         spin_unlock(&fc->lock);
1150                         wait_event(req->waitq, !req->locked);
1151                         end(fc, req);
1152                         fuse_put_request(fc, req);
1153                         spin_lock(&fc->lock);
1154                 }
1155         }
1156 }
1157
1158 /*
1159  * Abort all requests.
1160  *
1161  * Emergency exit in case of a malicious or accidental deadlock, or
1162  * just a hung filesystem.
1163  *
1164  * The same effect is usually achievable through killing the
1165  * filesystem daemon and all users of the filesystem.  The exception
1166  * is the combination of an asynchronous request and the tricky
1167  * deadlock (see Documentation/filesystems/fuse.txt).
1168  *
1169  * During the aborting, progression of requests from the pending and
1170  * processing lists onto the io list, and progression of new requests
1171  * onto the pending list is prevented by req->connected being false.
1172  *
1173  * Progression of requests under I/O to the processing list is
1174  * prevented by the req->aborted flag being true for these requests.
1175  * For this reason requests on the io list must be aborted first.
1176  */
1177 void fuse_abort_conn(struct fuse_conn *fc)
1178 {
1179         spin_lock(&fc->lock);
1180         if (fc->connected) {
1181                 fc->connected = 0;
1182                 fc->blocked = 0;
1183                 end_io_requests(fc);
1184                 end_requests(fc, &fc->pending);
1185                 end_requests(fc, &fc->processing);
1186                 wake_up_all(&fc->waitq);
1187                 wake_up_all(&fc->blocked_waitq);
1188                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1189         }
1190         spin_unlock(&fc->lock);
1191 }
1192 EXPORT_SYMBOL_GPL(fuse_abort_conn);
1193
1194 int fuse_dev_release(struct inode *inode, struct file *file)
1195 {
1196         struct fuse_conn *fc = fuse_get_conn(file);
1197         if (fc) {
1198                 spin_lock(&fc->lock);
1199                 fc->connected = 0;
1200                 end_requests(fc, &fc->pending);
1201                 end_requests(fc, &fc->processing);
1202                 spin_unlock(&fc->lock);
1203                 fuse_conn_put(fc);
1204         }
1205
1206         return 0;
1207 }
1208 EXPORT_SYMBOL_GPL(fuse_dev_release);
1209
1210 static int fuse_dev_fasync(int fd, struct file *file, int on)
1211 {
1212         struct fuse_conn *fc = fuse_get_conn(file);
1213         if (!fc)
1214                 return -EPERM;
1215
1216         /* No locking - fasync_helper does its own locking */
1217         return fasync_helper(fd, file, on, &fc->fasync);
1218 }
1219
1220 const struct file_operations fuse_dev_operations = {
1221         .owner          = THIS_MODULE,
1222         .llseek         = no_llseek,
1223         .read           = do_sync_read,
1224         .aio_read       = fuse_dev_read,
1225         .write          = do_sync_write,
1226         .aio_write      = fuse_dev_write,
1227         .poll           = fuse_dev_poll,
1228         .release        = fuse_dev_release,
1229         .fasync         = fuse_dev_fasync,
1230 };
1231 EXPORT_SYMBOL_GPL(fuse_dev_operations);
1232
1233 static struct miscdevice fuse_miscdevice = {
1234         .minor = FUSE_MINOR,
1235         .name  = "fuse",
1236         .fops = &fuse_dev_operations,
1237 };
1238
1239 int __init fuse_dev_init(void)
1240 {
1241         int err = -ENOMEM;
1242         fuse_req_cachep = kmem_cache_create("fuse_request",
1243                                             sizeof(struct fuse_req),
1244                                             0, 0, NULL);
1245         if (!fuse_req_cachep)
1246                 goto out;
1247
1248         err = misc_register(&fuse_miscdevice);
1249         if (err)
1250                 goto out_cache_clean;
1251
1252         return 0;
1253
1254  out_cache_clean:
1255         kmem_cache_destroy(fuse_req_cachep);
1256  out:
1257         return err;
1258 }
1259
1260 void fuse_dev_cleanup(void)
1261 {
1262         misc_deregister(&fuse_miscdevice);
1263         kmem_cache_destroy(fuse_req_cachep);
1264 }