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 < FUSE_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 == FUSE_MAX_BACKGROUND) {
 284                         fc->blocked = 0;
 285                         wake_up_all(&fc->blocked_waitq);
 286                 }
 287                 if (fc->num_background == FUSE_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 == FUSE_MAX_BACKGROUND)
 414                 fc->blocked = 1;
 415         if (fc->num_background == FUSE_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         down_read(&current->mm->mmap_sem);
 555         err = get_user_pages(current, current->mm, cs->addr, 1, cs->write, 0,
 556                              &cs->pg, NULL);
 557         up_read(&current->mm->mmap_sem);
 558         if (err < 0)
 559                 return err;
 560         BUG_ON(err != 1);
 561         offset = cs->addr % PAGE_SIZE;
 562         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
 563         cs->buf = cs->mapaddr + offset;
 564         cs->len = min(PAGE_SIZE - offset, cs->seglen);
 565         cs->seglen -= cs->len;
 566         cs->addr += cs->len;
 567
 568         return lock_request(cs->fc, cs->req);
 569 }
 570
 571 /* Do as much copy to/from userspace buffer as we can */
 572 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 573 {
 574         unsigned ncpy = min(*size, cs->len);
 575         if (val) {
 576                 if (cs->write)
 577                         memcpy(cs->buf, *val, ncpy);
 578                 else
 579                         memcpy(*val, cs->buf, ncpy);
 580                 *val += ncpy;
 581         }
 582         *size -= ncpy;
 583         cs->len -= ncpy;
 584         cs->buf += ncpy;
 585         return ncpy;
 586 }
 587
 588 /*
 589  * Copy a page in the request to/from the userspace buffer.  Must be
 590  * done atomically
 591  */
 592 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
 593                           unsigned offset, unsigned count, int zeroing)
 594 {
 595         if (page && zeroing && count < PAGE_SIZE) {
 596                 void *mapaddr = kmap_atomic(page, KM_USER1);
 597                 memset(mapaddr, 0, PAGE_SIZE);
 598                 kunmap_atomic(mapaddr, KM_USER1);
 599         }
 600         while (count) {
 601                 if (!cs->len) {
 602                         int err = fuse_copy_fill(cs);
 603                         if (err)
 604                                 return err;
 605                 }
 606                 if (page) {
 607                         void *mapaddr = kmap_atomic(page, KM_USER1);
 608                         void *buf = mapaddr + offset;
 609                         offset += fuse_copy_do(cs, &buf, &count);
 610                         kunmap_atomic(mapaddr, KM_USER1);
 611                 } else
 612                         offset += fuse_copy_do(cs, NULL, &count);
 613         }
 614         if (page && !cs->write)
 615                 flush_dcache_page(page);
 616         return 0;
 617 }
 618
 619 /* Copy pages in the request to/from userspace buffer */
 620 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 621                            int zeroing)
 622 {
 623         unsigned i;
 624         struct fuse_req *req = cs->req;
 625         unsigned offset = req->page_offset;
 626         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 627
 628         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 629                 struct page *page = req->pages[i];
 630                 int err = fuse_copy_page(cs, page, offset, count, zeroing);
 631                 if (err)
 632                         return err;
 633
 634                 nbytes -= count;
 635                 count = min(nbytes, (unsigned) PAGE_SIZE);
 636                 offset = 0;
 637         }
 638         return 0;
 639 }
 640
 641 /* Copy a single argument in the request to/from userspace buffer */
 642 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 643 {
 644         while (size) {
 645                 if (!cs->len) {
 646                         int err = fuse_copy_fill(cs);
 647                         if (err)
 648                                 return err;
 649                 }
 650                 fuse_copy_do(cs, &val, &size);
 651         }
 652         return 0;
 653 }
 654
 655 /* Copy request arguments to/from userspace buffer */
 656 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 657                           unsigned argpages, struct fuse_arg *args,
 658                           int zeroing)
 659 {
 660         int err = 0;
 661         unsigned i;
 662
 663         for (i = 0; !err && i < numargs; i++)  {
 664                 struct fuse_arg *arg = &args[i];
 665                 if (i == numargs - 1 && argpages)
 666                         err = fuse_copy_pages(cs, arg->size, zeroing);
 667                 else
 668                         err = fuse_copy_one(cs, arg->value, arg->size);
 669         }
 670         return err;
 671 }
 672
 673 static int request_pending(struct fuse_conn *fc)
 674 {
 675         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 676 }
 677
 678 /* Wait until a request is available on the pending list */
 679 static void request_wait(struct fuse_conn *fc)
 680 __releases(&fc->lock)
 681 __acquires(&fc->lock)
 682 {
 683         DECLARE_WAITQUEUE(wait, current);
 684
 685         add_wait_queue_exclusive(&fc->waitq, &wait);
 686         while (fc->connected && !request_pending(fc)) {
 687                 set_current_state(TASK_INTERRUPTIBLE);
 688                 if (signal_pending(current))
 689                         break;
 690
 691                 spin_unlock(&fc->lock);
 692                 schedule();
 693                 spin_lock(&fc->lock);
 694         }
 695         set_current_state(TASK_RUNNING);
 696         remove_wait_queue(&fc->waitq, &wait);
 697 }
 698
 699 /*
 700  * Transfer an interrupt request to userspace
 701  *
 702  * Unlike other requests this is assembled on demand, without a need
 703  * to allocate a separate fuse_req structure.
 704  *
 705  * Called with fc->lock held, releases it
 706  */
 707 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
 708                                const struct iovec *iov, unsigned long nr_segs)
 709 __releases(&fc->lock)
 710 {
 711         struct fuse_copy_state cs;
 712         struct fuse_in_header ih;
 713         struct fuse_interrupt_in arg;
 714         unsigned reqsize = sizeof(ih) + sizeof(arg);
 715         int err;
 716
 717         list_del_init(&req->intr_entry);
 718         req->intr_unique = fuse_get_unique(fc);
 719         memset(&ih, 0, sizeof(ih));
 720         memset(&arg, 0, sizeof(arg));
 721         ih.len = reqsize;
 722         ih.opcode = FUSE_INTERRUPT;
 723         ih.unique = req->intr_unique;
 724         arg.unique = req->in.h.unique;
 725
 726         spin_unlock(&fc->lock);
 727         if (iov_length(iov, nr_segs) < reqsize)
 728                 return -EINVAL;
 729
 730         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
 731         err = fuse_copy_one(&cs, &ih, sizeof(ih));
 732         if (!err)
 733                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
 734         fuse_copy_finish(&cs);
 735
 736         return err ? err : reqsize;
 737 }
 738
 739 /*
 740  * Read a single request into the userspace filesystem's buffer.  This
 741  * function waits until a request is available, then removes it from
 742  * the pending list and copies request data to userspace buffer.  If
 743  * no reply is needed (FORGET) or request has been aborted or there
 744  * was an error during the copying then it's finished by calling
 745  * request_end().  Otherwise add it to the processing list, and set
 746  * the 'sent' flag.
 747  */
 748 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
 749                               unsigned long nr_segs, loff_t pos)
 750 {
 751         int err;
 752         struct fuse_req *req;
 753         struct fuse_in *in;
 754         struct fuse_copy_state cs;
 755         unsigned reqsize;
 756         struct file *file = iocb->ki_filp;
 757         struct fuse_conn *fc = fuse_get_conn(file);
 758         if (!fc)
 759                 return -EPERM;
 760
 761  restart:
 762         spin_lock(&fc->lock);
 763         err = -EAGAIN;
 764         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 765             !request_pending(fc))
 766                 goto err_unlock;
 767
 768         request_wait(fc);
 769         err = -ENODEV;
 770         if (!fc->connected)
 771                 goto err_unlock;
 772         err = -ERESTARTSYS;
 773         if (!request_pending(fc))
 774                 goto err_unlock;
 775
 776         if (!list_empty(&fc->interrupts)) {
 777                 req = list_entry(fc->interrupts.next, struct fuse_req,
 778                                  intr_entry);
 779                 return fuse_read_interrupt(fc, req, iov, nr_segs);
 780         }
 781
 782         req = list_entry(fc->pending.next, struct fuse_req, list);
 783         req->state = FUSE_REQ_READING;
 784         list_move(&req->list, &fc->io);
 785
 786         in = &req->in;
 787         reqsize = in->h.len;
 788         /* If request is too large, reply with an error and restart the read */
 789         if (iov_length(iov, nr_segs) < reqsize) {
 790                 req->out.h.error = -EIO;
 791                 /* SETXATTR is special, since it may contain too large data */
 792                 if (in->h.opcode == FUSE_SETXATTR)
 793                         req->out.h.error = -E2BIG;
 794                 request_end(fc, req);
 795                 goto restart;
 796         }
 797         spin_unlock(&fc->lock);
 798         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
 799         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
 800         if (!err)
 801                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
 802                                      (struct fuse_arg *) in->args, 0);
 803         fuse_copy_finish(&cs);
 804         spin_lock(&fc->lock);
 805         req->locked = 0;
 806         if (req->aborted) {
 807                 request_end(fc, req);
 808                 return -ENODEV;
 809         }
 810         if (err) {
 811                 req->out.h.error = -EIO;
 812                 request_end(fc, req);
 813                 return err;
 814         }
 815         if (!req->isreply)
 816                 request_end(fc, req);
 817         else {
 818                 req->state = FUSE_REQ_SENT;
 819                 list_move_tail(&req->list, &fc->processing);
 820                 if (req->interrupted)
 821                         queue_interrupt(fc, req);
 822                 spin_unlock(&fc->lock);
 823         }
 824         return reqsize;
 825
 826  err_unlock:
 827         spin_unlock(&fc->lock);
 828         return err;
 829 }
 830
 831 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
 832                             struct fuse_copy_state *cs)
 833 {
 834         struct fuse_notify_poll_wakeup_out outarg;
 835         int err = -EINVAL;
 836
 837         if (size != sizeof(outarg))
 838                 goto err;
 839
 840         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
 841         if (err)
 842                 goto err;
 843
 844         fuse_copy_finish(cs);
 845         return fuse_notify_poll_wakeup(fc, &outarg);
 846
 847 err:
 848         fuse_copy_finish(cs);
 849         return err;
 850 }
 851
 852 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
 853                                    struct fuse_copy_state *cs)
 854 {
 855         struct fuse_notify_inval_inode_out outarg;
 856         int err = -EINVAL;
 857
 858         if (size != sizeof(outarg))
 859                 goto err;
 860
 861         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
 862         if (err)
 863                 goto err;
 864         fuse_copy_finish(cs);
 865
 866         down_read(&fc->killsb);
 867         err = -ENOENT;
 868         if (!fc->sb)
 869                 goto err_unlock;
 870
 871         err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
 872                                        outarg.off, outarg.len);
 873
 874 err_unlock:
 875         up_read(&fc->killsb);
 876         return err;
 877
 878 err:
 879         fuse_copy_finish(cs);
 880         return err;
 881 }
 882
 883 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
 884                                    struct fuse_copy_state *cs)
 885 {
 886         struct fuse_notify_inval_entry_out outarg;
 887         int err = -EINVAL;
 888         char buf[FUSE_NAME_MAX+1];
 889         struct qstr name;
 890
 891         if (size < sizeof(outarg))
 892                 goto err;
 893
 894         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
 895         if (err)
 896                 goto err;
 897
 898         err = -ENAMETOOLONG;
 899         if (outarg.namelen > FUSE_NAME_MAX)
 900                 goto err;
 901
 902         name.name = buf;
 903         name.len = outarg.namelen;
 904         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
 905         if (err)
 906                 goto err;
 907         fuse_copy_finish(cs);
 908         buf[outarg.namelen] = 0;
 909         name.hash = full_name_hash(name.name, name.len);
 910
 911         down_read(&fc->killsb);
 912         err = -ENOENT;
 913         if (!fc->sb)
 914                 goto err_unlock;
 915
 916         err = fuse_reverse_inval_entry(fc->sb, outarg.parent, &name);
 917
 918 err_unlock:
 919         up_read(&fc->killsb);
 920         return err;
 921
 922 err:
 923         fuse_copy_finish(cs);
 924         return err;
 925 }
 926
 927 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
 928                        unsigned int size, struct fuse_copy_state *cs)
 929 {
 930         switch (code) {
 931         case FUSE_NOTIFY_POLL:
 932                 return fuse_notify_poll(fc, size, cs);
 933
 934         case FUSE_NOTIFY_INVAL_INODE:
 935                 return fuse_notify_inval_inode(fc, size, cs);
 936
 937         case FUSE_NOTIFY_INVAL_ENTRY:
 938                 return fuse_notify_inval_entry(fc, size, cs);
 939
 940         default:
 941                 fuse_copy_finish(cs);
 942                 return -EINVAL;
 943         }
 944 }
 945
 946 /* Look up request on processing list by unique ID */
 947 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
 948 {
 949         struct list_head *entry;
 950
 951         list_for_each(entry, &fc->processing) {
 952                 struct fuse_req *req;
 953                 req = list_entry(entry, struct fuse_req, list);
 954                 if (req->in.h.unique == unique || req->intr_unique == unique)
 955                         return req;
 956         }
 957         return NULL;
 958 }
 959
 960 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
 961                          unsigned nbytes)
 962 {
 963         unsigned reqsize = sizeof(struct fuse_out_header);
 964
 965         if (out->h.error)
 966                 return nbytes != reqsize ? -EINVAL : 0;
 967
 968         reqsize += len_args(out->numargs, out->args);
 969
 970         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
 971                 return -EINVAL;
 972         else if (reqsize > nbytes) {
 973                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
 974                 unsigned diffsize = reqsize - nbytes;
 975                 if (diffsize > lastarg->size)
 976                         return -EINVAL;
 977                 lastarg->size -= diffsize;
 978         }
 979         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
 980                               out->page_zeroing);
 981 }
 982
 983 /*
 984  * Write a single reply to a request.  First the header is copied from
 985  * the write buffer.  The request is then searched on the processing
 986  * list by the unique ID found in the header.  If found, then remove
 987  * it from the list and copy the rest of the buffer to the request.
 988  * The request is finished by calling request_end()
 989  */
 990 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
 991                                unsigned long nr_segs, loff_t pos)
 992 {
 993         int err;
 994         size_t nbytes = iov_length(iov, nr_segs);
 995         struct fuse_req *req;
 996         struct fuse_out_header oh;
 997         struct fuse_copy_state cs;
 998         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
 999         if (!fc)
1000                 return -EPERM;
1001
1002         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
1003         if (nbytes < sizeof(struct fuse_out_header))
1004                 return -EINVAL;
1005
1006         err = fuse_copy_one(&cs, &oh, sizeof(oh));
1007         if (err)
1008                 goto err_finish;
1009
1010         err = -EINVAL;
1011         if (oh.len != nbytes)
1012                 goto err_finish;
1013
1014         /*
1015          * Zero oh.unique indicates unsolicited notification message
1016          * and error contains notification code.
1017          */
1018         if (!oh.unique) {
1019                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), &cs);
1020                 return err ? err : nbytes;
1021         }
1022
1023         err = -EINVAL;
1024         if (oh.error <= -1000 || oh.error > 0)
1025                 goto err_finish;
1026
1027         spin_lock(&fc->lock);
1028         err = -ENOENT;
1029         if (!fc->connected)
1030                 goto err_unlock;
1031
1032         req = request_find(fc, oh.unique);
1033         if (!req)
1034                 goto err_unlock;
1035
1036         if (req->aborted) {
1037                 spin_unlock(&fc->lock);
1038                 fuse_copy_finish(&cs);
1039                 spin_lock(&fc->lock);
1040                 request_end(fc, req);
1041                 return -ENOENT;
1042         }
1043         /* Is it an interrupt reply? */
1044         if (req->intr_unique == oh.unique) {
1045                 err = -EINVAL;
1046                 if (nbytes != sizeof(struct fuse_out_header))
1047                         goto err_unlock;
1048
1049                 if (oh.error == -ENOSYS)
1050                         fc->no_interrupt = 1;
1051                 else if (oh.error == -EAGAIN)
1052                         queue_interrupt(fc, req);
1053
1054                 spin_unlock(&fc->lock);
1055                 fuse_copy_finish(&cs);
1056                 return nbytes;
1057         }
1058
1059         req->state = FUSE_REQ_WRITING;
1060         list_move(&req->list, &fc->io);
1061         req->out.h = oh;
1062         req->locked = 1;
1063         cs.req = req;
1064         spin_unlock(&fc->lock);
1065
1066         err = copy_out_args(&cs, &req->out, nbytes);
1067         fuse_copy_finish(&cs);
1068
1069         spin_lock(&fc->lock);
1070         req->locked = 0;
1071         if (!err) {
1072                 if (req->aborted)
1073                         err = -ENOENT;
1074         } else if (!req->aborted)
1075                 req->out.h.error = -EIO;
1076         request_end(fc, req);
1077
1078         return err ? err : nbytes;
1079
1080  err_unlock:
1081         spin_unlock(&fc->lock);
1082  err_finish:
1083         fuse_copy_finish(&cs);
1084         return err;
1085 }
1086
1087 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
1088 {
1089         unsigned mask = POLLOUT | POLLWRNORM;
1090         struct fuse_conn *fc = fuse_get_conn(file);
1091         if (!fc)
1092                 return POLLERR;
1093
1094         poll_wait(file, &fc->waitq, wait);
1095
1096         spin_lock(&fc->lock);
1097         if (!fc->connected)
1098                 mask = POLLERR;
1099         else if (request_pending(fc))
1100                 mask |= POLLIN | POLLRDNORM;
1101         spin_unlock(&fc->lock);
1102
1103         return mask;
1104 }
1105
1106 /*
1107  * Abort all requests on the given list (pending or processing)
1108  *
1109  * This function releases and reacquires fc->lock
1110  */
1111 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1112 __releases(&fc->lock)
1113 __acquires(&fc->lock)
1114 {
1115         while (!list_empty(head)) {
1116                 struct fuse_req *req;
1117                 req = list_entry(head->next, struct fuse_req, list);
1118                 req->out.h.error = -ECONNABORTED;
1119                 request_end(fc, req);
1120                 spin_lock(&fc->lock);
1121         }
1122 }
1123
1124 /*
1125  * Abort requests under I/O
1126  *
1127  * The requests are set to aborted and finished, and the request
1128  * waiter is woken up.  This will make request_wait_answer() wait
1129  * until the request is unlocked and then return.
1130  *
1131  * If the request is asynchronous, then the end function needs to be
1132  * called after waiting for the request to be unlocked (if it was
1133  * locked).
1134  */
1135 static void end_io_requests(struct fuse_conn *fc)
1136 __releases(&fc->lock)
1137 __acquires(&fc->lock)
1138 {
1139         while (!list_empty(&fc->io)) {
1140                 struct fuse_req *req =
1141                         list_entry(fc->io.next, struct fuse_req, list);
1142                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1143
1144                 req->aborted = 1;
1145                 req->out.h.error = -ECONNABORTED;
1146                 req->state = FUSE_REQ_FINISHED;
1147                 list_del_init(&req->list);
1148                 wake_up(&req->waitq);
1149                 if (end) {
1150                         req->end = NULL;
1151                         __fuse_get_request(req);
1152                         spin_unlock(&fc->lock);
1153                         wait_event(req->waitq, !req->locked);
1154                         end(fc, req);
1155                         fuse_put_request(fc, req);
1156                         spin_lock(&fc->lock);
1157                 }
1158         }
1159 }
1160
1161 /*
1162  * Abort all requests.
1163  *
1164  * Emergency exit in case of a malicious or accidental deadlock, or
1165  * just a hung filesystem.
1166  *
1167  * The same effect is usually achievable through killing the
1168  * filesystem daemon and all users of the filesystem.  The exception
1169  * is the combination of an asynchronous request and the tricky
1170  * deadlock (see Documentation/filesystems/fuse.txt).
1171  *
1172  * During the aborting, progression of requests from the pending and
1173  * processing lists onto the io list, and progression of new requests
1174  * onto the pending list is prevented by req->connected being false.
1175  *
1176  * Progression of requests under I/O to the processing list is
1177  * prevented by the req->aborted flag being true for these requests.
1178  * For this reason requests on the io list must be aborted first.
1179  */
1180 void fuse_abort_conn(struct fuse_conn *fc)
1181 {
1182         spin_lock(&fc->lock);
1183         if (fc->connected) {
1184                 fc->connected = 0;
1185                 fc->blocked = 0;
1186                 end_io_requests(fc);
1187                 end_requests(fc, &fc->pending);
1188                 end_requests(fc, &fc->processing);
1189                 wake_up_all(&fc->waitq);
1190                 wake_up_all(&fc->blocked_waitq);
1191                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1192         }
1193         spin_unlock(&fc->lock);
1194 }
1195 EXPORT_SYMBOL_GPL(fuse_abort_conn);
1196
1197 int fuse_dev_release(struct inode *inode, struct file *file)
1198 {
1199         struct fuse_conn *fc = fuse_get_conn(file);
1200         if (fc) {
1201                 spin_lock(&fc->lock);
1202                 fc->connected = 0;
1203                 end_requests(fc, &fc->pending);
1204                 end_requests(fc, &fc->processing);
1205                 spin_unlock(&fc->lock);
1206                 fuse_conn_put(fc);
1207         }
1208
1209         return 0;
1210 }
1211 EXPORT_SYMBOL_GPL(fuse_dev_release);
1212
1213 static int fuse_dev_fasync(int fd, struct file *file, int on)
1214 {
1215         struct fuse_conn *fc = fuse_get_conn(file);
1216         if (!fc)
1217                 return -EPERM;
1218
1219         /* No locking - fasync_helper does its own locking */
1220         return fasync_helper(fd, file, on, &fc->fasync);
1221 }
1222
1223 const struct file_operations fuse_dev_operations = {
1224         .owner          = THIS_MODULE,
1225         .llseek         = no_llseek,
1226         .read           = do_sync_read,
1227         .aio_read       = fuse_dev_read,
1228         .write          = do_sync_write,
1229         .aio_write      = fuse_dev_write,
1230         .poll           = fuse_dev_poll,
1231         .release        = fuse_dev_release,
1232         .fasync         = fuse_dev_fasync,
1233 };
1234 EXPORT_SYMBOL_GPL(fuse_dev_operations);
1235
1236 static struct miscdevice fuse_miscdevice = {
1237         .minor = FUSE_MINOR,
1238         .name  = "fuse",
1239         .fops = &fuse_dev_operations,
1240 };
1241
1242 int __init fuse_dev_init(void)
1243 {
1244         int err = -ENOMEM;
1245         fuse_req_cachep = kmem_cache_create("fuse_request",
1246                                             sizeof(struct fuse_req),
1247                                             0, 0, NULL);
1248         if (!fuse_req_cachep)
1249                 goto out;
1250
1251         err = misc_register(&fuse_miscdevice);
1252         if (err)
1253                 goto out_cache_clean;
1254
1255         return 0;
1256
1257  out_cache_clean:
1258         kmem_cache_destroy(fuse_req_cachep);
1259  out:
1260         return err;
1261 }
1262
1263 void fuse_dev_cleanup(void)
1264 {
1265         misc_deregister(&fuse_miscdevice);
1266         kmem_cache_destroy(fuse_req_cachep);
1267 }