fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2006  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19
  20 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  21
  22 static struct kmem_cache *fuse_req_cachep;
  23
  24 static struct fuse_conn *fuse_get_conn(struct file *file)
  25 {
  26         /*
  27          * Lockless access is OK, because file->private data is set
  28          * once during mount and is valid until the file is released.
  29          */
  30         return file->private_data;
  31 }
  32
  33 static void fuse_request_init(struct fuse_req *req)
  34 {
  35         memset(req, 0, sizeof(*req));
  36         INIT_LIST_HEAD(&req->list);
  37         INIT_LIST_HEAD(&req->intr_entry);
  38         init_waitqueue_head(&req->waitq);
  39         atomic_set(&req->count, 1);
  40 }
  41
  42 struct fuse_req *fuse_request_alloc(void)
  43 {
  44         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_KERNEL);
  45         if (req)
  46                 fuse_request_init(req);
  47         return req;
  48 }
  49
  50 void fuse_request_free(struct fuse_req *req)
  51 {
  52         kmem_cache_free(fuse_req_cachep, req);
  53 }
  54
  55 static void block_sigs(sigset_t *oldset)
  56 {
  57         sigset_t mask;
  58
  59         siginitsetinv(&mask, sigmask(SIGKILL));
  60         sigprocmask(SIG_BLOCK, &mask, oldset);
  61 }
  62
  63 static void restore_sigs(sigset_t *oldset)
  64 {
  65         sigprocmask(SIG_SETMASK, oldset, NULL);
  66 }
  67
  68 static void __fuse_get_request(struct fuse_req *req)
  69 {
  70         atomic_inc(&req->count);
  71 }
  72
  73 /* Must be called with > 1 refcount */
  74 static void __fuse_put_request(struct fuse_req *req)
  75 {
  76         BUG_ON(atomic_read(&req->count) < 2);
  77         atomic_dec(&req->count);
  78 }
  79
  80 static void fuse_req_init_context(struct fuse_req *req)
  81 {
  82         req->in.h.uid = current->fsuid;
  83         req->in.h.gid = current->fsgid;
  84         req->in.h.pid = current->pid;
  85 }
  86
  87 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
  88 {
  89         struct fuse_req *req;
  90         sigset_t oldset;
  91         int intr;
  92         int err;
  93
  94         atomic_inc(&fc->num_waiting);
  95         block_sigs(&oldset);
  96         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
  97         restore_sigs(&oldset);
  98         err = -EINTR;
  99         if (intr)
 100                 goto out;
 101
 102         err = -ENOTCONN;
 103         if (!fc->connected)
 104                 goto out;
 105
 106         req = fuse_request_alloc();
 107         err = -ENOMEM;
 108         if (!req)
 109                 goto out;
 110
 111         fuse_req_init_context(req);
 112         req->waiting = 1;
 113         return req;
 114
 115  out:
 116         atomic_dec(&fc->num_waiting);
 117         return ERR_PTR(err);
 118 }
 119
 120 /*
 121  * Return request in fuse_file->reserved_req.  However that may
 122  * currently be in use.  If that is the case, wait for it to become
 123  * available.
 124  */
 125 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 126                                          struct file *file)
 127 {
 128         struct fuse_req *req = NULL;
 129         struct fuse_file *ff = file->private_data;
 130
 131         do {
 132                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 133                 spin_lock(&fc->lock);
 134                 if (ff->reserved_req) {
 135                         req = ff->reserved_req;
 136                         ff->reserved_req = NULL;
 137                         get_file(file);
 138                         req->stolen_file = file;
 139                 }
 140                 spin_unlock(&fc->lock);
 141         } while (!req);
 142
 143         return req;
 144 }
 145
 146 /*
 147  * Put stolen request back into fuse_file->reserved_req
 148  */
 149 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 150 {
 151         struct file *file = req->stolen_file;
 152         struct fuse_file *ff = file->private_data;
 153
 154         spin_lock(&fc->lock);
 155         fuse_request_init(req);
 156         BUG_ON(ff->reserved_req);
 157         ff->reserved_req = req;
 158         wake_up_all(&fc->reserved_req_waitq);
 159         spin_unlock(&fc->lock);
 160         fput(file);
 161 }
 162
 163 /*
 164  * Gets a requests for a file operation, always succeeds
 165  *
 166  * This is used for sending the FLUSH request, which must get to
 167  * userspace, due to POSIX locks which may need to be unlocked.
 168  *
 169  * If allocation fails due to OOM, use the reserved request in
 170  * fuse_file.
 171  *
 172  * This is very unlikely to deadlock accidentally, since the
 173  * filesystem should not have it's own file open.  If deadlock is
 174  * intentional, it can still be broken by "aborting" the filesystem.
 175  */
 176 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 177 {
 178         struct fuse_req *req;
 179
 180         atomic_inc(&fc->num_waiting);
 181         wait_event(fc->blocked_waitq, !fc->blocked);
 182         req = fuse_request_alloc();
 183         if (!req)
 184                 req = get_reserved_req(fc, file);
 185
 186         fuse_req_init_context(req);
 187         req->waiting = 1;
 188         return req;
 189 }
 190
 191 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 192 {
 193         if (atomic_dec_and_test(&req->count)) {
 194                 if (req->waiting)
 195                         atomic_dec(&fc->num_waiting);
 196
 197                 if (req->stolen_file)
 198                         put_reserved_req(fc, req);
 199                 else
 200                         fuse_request_free(req);
 201         }
 202 }
 203
 204 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 205 {
 206         unsigned nbytes = 0;
 207         unsigned i;
 208
 209         for (i = 0; i < numargs; i++)
 210                 nbytes += args[i].size;
 211
 212         return nbytes;
 213 }
 214
 215 static u64 fuse_get_unique(struct fuse_conn *fc)
 216 {
 217         fc->reqctr++;
 218         /* zero is special */
 219         if (fc->reqctr == 0)
 220                 fc->reqctr = 1;
 221
 222         return fc->reqctr;
 223 }
 224
 225 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 226 {
 227         req->in.h.unique = fuse_get_unique(fc);
 228         req->in.h.len = sizeof(struct fuse_in_header) +
 229                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 230         list_add_tail(&req->list, &fc->pending);
 231         req->state = FUSE_REQ_PENDING;
 232         if (!req->waiting) {
 233                 req->waiting = 1;
 234                 atomic_inc(&fc->num_waiting);
 235         }
 236         wake_up(&fc->waitq);
 237         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 238 }
 239
 240 static void flush_bg_queue(struct fuse_conn *fc)
 241 {
 242         while (fc->active_background < FUSE_MAX_BACKGROUND &&
 243                !list_empty(&fc->bg_queue)) {
 244                 struct fuse_req *req;
 245
 246                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 247                 list_del(&req->list);
 248                 fc->active_background++;
 249                 queue_request(fc, req);
 250         }
 251 }
 252
 253 /*
 254  * This function is called when a request is finished.  Either a reply
 255  * has arrived or it was aborted (and not yet sent) or some error
 256  * occurred during communication with userspace, or the device file
 257  * was closed.  The requester thread is woken up (if still waiting),
 258  * the 'end' callback is called if given, else the reference to the
 259  * request is released
 260  *
 261  * Called with fc->lock, unlocks it
 262  */
 263 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 264         __releases(fc->lock)
 265 {
 266         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 267         req->end = NULL;
 268         list_del(&req->list);
 269         list_del(&req->intr_entry);
 270         req->state = FUSE_REQ_FINISHED;
 271         if (req->background) {
 272                 if (fc->num_background == FUSE_MAX_BACKGROUND) {
 273                         fc->blocked = 0;
 274                         wake_up_all(&fc->blocked_waitq);
 275                 }
 276                 if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 277                         clear_bdi_congested(&fc->bdi, READ);
 278                         clear_bdi_congested(&fc->bdi, WRITE);
 279                 }
 280                 fc->num_background--;
 281                 fc->active_background--;
 282                 flush_bg_queue(fc);
 283         }
 284         spin_unlock(&fc->lock);
 285         wake_up(&req->waitq);
 286         if (end)
 287                 end(fc, req);
 288         else
 289                 fuse_put_request(fc, req);
 290 }
 291
 292 static void wait_answer_interruptible(struct fuse_conn *fc,
 293                                       struct fuse_req *req)
 294 {
 295         if (signal_pending(current))
 296                 return;
 297
 298         spin_unlock(&fc->lock);
 299         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 300         spin_lock(&fc->lock);
 301 }
 302
 303 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 304 {
 305         list_add_tail(&req->intr_entry, &fc->interrupts);
 306         wake_up(&fc->waitq);
 307         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 308 }
 309
 310 /* Called with fc->lock held.  Releases, and then reacquires it. */
 311 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 312 {
 313         if (!fc->no_interrupt) {
 314                 /* Any signal may interrupt this */
 315                 wait_answer_interruptible(fc, req);
 316
 317                 if (req->aborted)
 318                         goto aborted;
 319                 if (req->state == FUSE_REQ_FINISHED)
 320                         return;
 321
 322                 req->interrupted = 1;
 323                 if (req->state == FUSE_REQ_SENT)
 324                         queue_interrupt(fc, req);
 325         }
 326
 327         if (!req->force) {
 328                 sigset_t oldset;
 329
 330                 /* Only fatal signals may interrupt this */
 331                 block_sigs(&oldset);
 332                 wait_answer_interruptible(fc, req);
 333                 restore_sigs(&oldset);
 334
 335                 if (req->aborted)
 336                         goto aborted;
 337                 if (req->state == FUSE_REQ_FINISHED)
 338                         return;
 339
 340                 /* Request is not yet in userspace, bail out */
 341                 if (req->state == FUSE_REQ_PENDING) {
 342                         list_del(&req->list);
 343                         __fuse_put_request(req);
 344                         req->out.h.error = -EINTR;
 345                         return;
 346                 }
 347         }
 348
 349         /*
 350          * Either request is already in userspace, or it was forced.
 351          * Wait it out.
 352          */
 353         spin_unlock(&fc->lock);
 354         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 355         spin_lock(&fc->lock);
 356
 357         if (!req->aborted)
 358                 return;
 359
 360  aborted:
 361         BUG_ON(req->state != FUSE_REQ_FINISHED);
 362         if (req->locked) {
 363                 /* This is uninterruptible sleep, because data is
 364                    being copied to/from the buffers of req.  During
 365                    locked state, there mustn't be any filesystem
 366                    operation (e.g. page fault), since that could lead
 367                    to deadlock */
 368                 spin_unlock(&fc->lock);
 369                 wait_event(req->waitq, !req->locked);
 370                 spin_lock(&fc->lock);
 371         }
 372 }
 373
 374 void request_send(struct fuse_conn *fc, struct fuse_req *req)
 375 {
 376         req->isreply = 1;
 377         spin_lock(&fc->lock);
 378         if (!fc->connected)
 379                 req->out.h.error = -ENOTCONN;
 380         else if (fc->conn_error)
 381                 req->out.h.error = -ECONNREFUSED;
 382         else {
 383                 queue_request(fc, req);
 384                 /* acquire extra reference, since request is still needed
 385                    after request_end() */
 386                 __fuse_get_request(req);
 387
 388                 request_wait_answer(fc, req);
 389         }
 390         spin_unlock(&fc->lock);
 391 }
 392
 393 static void request_send_nowait_locked(struct fuse_conn *fc,
 394                                        struct fuse_req *req)
 395 {
 396         req->background = 1;
 397         fc->num_background++;
 398         if (fc->num_background == FUSE_MAX_BACKGROUND)
 399                 fc->blocked = 1;
 400         if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 401                 set_bdi_congested(&fc->bdi, READ);
 402                 set_bdi_congested(&fc->bdi, WRITE);
 403         }
 404         list_add_tail(&req->list, &fc->bg_queue);
 405         flush_bg_queue(fc);
 406 }
 407
 408 static void request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 409 {
 410         spin_lock(&fc->lock);
 411         if (fc->connected) {
 412                 request_send_nowait_locked(fc, req);
 413                 spin_unlock(&fc->lock);
 414         } else {
 415                 req->out.h.error = -ENOTCONN;
 416                 request_end(fc, req);
 417         }
 418 }
 419
 420 void request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 421 {
 422         req->isreply = 0;
 423         request_send_nowait(fc, req);
 424 }
 425
 426 void request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 427 {
 428         req->isreply = 1;
 429         request_send_nowait(fc, req);
 430 }
 431
 432 /*
 433  * Lock the request.  Up to the next unlock_request() there mustn't be
 434  * anything that could cause a page-fault.  If the request was already
 435  * aborted bail out.
 436  */
 437 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 438 {
 439         int err = 0;
 440         if (req) {
 441                 spin_lock(&fc->lock);
 442                 if (req->aborted)
 443                         err = -ENOENT;
 444                 else
 445                         req->locked = 1;
 446                 spin_unlock(&fc->lock);
 447         }
 448         return err;
 449 }
 450
 451 /*
 452  * Unlock request.  If it was aborted during being locked, the
 453  * requester thread is currently waiting for it to be unlocked, so
 454  * wake it up.
 455  */
 456 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 457 {
 458         if (req) {
 459                 spin_lock(&fc->lock);
 460                 req->locked = 0;
 461                 if (req->aborted)
 462                         wake_up(&req->waitq);
 463                 spin_unlock(&fc->lock);
 464         }
 465 }
 466
 467 struct fuse_copy_state {
 468         struct fuse_conn *fc;
 469         int write;
 470         struct fuse_req *req;
 471         const struct iovec *iov;
 472         unsigned long nr_segs;
 473         unsigned long seglen;
 474         unsigned long addr;
 475         struct page *pg;
 476         void *mapaddr;
 477         void *buf;
 478         unsigned len;
 479 };
 480
 481 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 482                            int write, struct fuse_req *req,
 483                            const struct iovec *iov, unsigned long nr_segs)
 484 {
 485         memset(cs, 0, sizeof(*cs));
 486         cs->fc = fc;
 487         cs->write = write;
 488         cs->req = req;
 489         cs->iov = iov;
 490         cs->nr_segs = nr_segs;
 491 }
 492
 493 /* Unmap and put previous page of userspace buffer */
 494 static void fuse_copy_finish(struct fuse_copy_state *cs)
 495 {
 496         if (cs->mapaddr) {
 497                 kunmap_atomic(cs->mapaddr, KM_USER0);
 498                 if (cs->write) {
 499                         flush_dcache_page(cs->pg);
 500                         set_page_dirty_lock(cs->pg);
 501                 }
 502                 put_page(cs->pg);
 503                 cs->mapaddr = NULL;
 504         }
 505 }
 506
 507 /*
 508  * Get another pagefull of userspace buffer, and map it to kernel
 509  * address space, and lock request
 510  */
 511 static int fuse_copy_fill(struct fuse_copy_state *cs)
 512 {
 513         unsigned long offset;
 514         int err;
 515
 516         unlock_request(cs->fc, cs->req);
 517         fuse_copy_finish(cs);
 518         if (!cs->seglen) {
 519                 BUG_ON(!cs->nr_segs);
 520                 cs->seglen = cs->iov[0].iov_len;
 521                 cs->addr = (unsigned long) cs->iov[0].iov_base;
 522                 cs->iov ++;
 523                 cs->nr_segs --;
 524         }
 525         down_read(&current->mm->mmap_sem);
 526         err = get_user_pages(current, current->mm, cs->addr, 1, cs->write, 0,
 527                              &cs->pg, NULL);
 528         up_read(&current->mm->mmap_sem);
 529         if (err < 0)
 530                 return err;
 531         BUG_ON(err != 1);
 532         offset = cs->addr % PAGE_SIZE;
 533         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
 534         cs->buf = cs->mapaddr + offset;
 535         cs->len = min(PAGE_SIZE - offset, cs->seglen);
 536         cs->seglen -= cs->len;
 537         cs->addr += cs->len;
 538
 539         return lock_request(cs->fc, cs->req);
 540 }
 541
 542 /* Do as much copy to/from userspace buffer as we can */
 543 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 544 {
 545         unsigned ncpy = min(*size, cs->len);
 546         if (val) {
 547                 if (cs->write)
 548                         memcpy(cs->buf, *val, ncpy);
 549                 else
 550                         memcpy(*val, cs->buf, ncpy);
 551                 *val += ncpy;
 552         }
 553         *size -= ncpy;
 554         cs->len -= ncpy;
 555         cs->buf += ncpy;
 556         return ncpy;
 557 }
 558
 559 /*
 560  * Copy a page in the request to/from the userspace buffer.  Must be
 561  * done atomically
 562  */
 563 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
 564                           unsigned offset, unsigned count, int zeroing)
 565 {
 566         if (page && zeroing && count < PAGE_SIZE) {
 567                 void *mapaddr = kmap_atomic(page, KM_USER1);
 568                 memset(mapaddr, 0, PAGE_SIZE);
 569                 kunmap_atomic(mapaddr, KM_USER1);
 570         }
 571         while (count) {
 572                 int err;
 573                 if (!cs->len && (err = fuse_copy_fill(cs)))
 574                         return err;
 575                 if (page) {
 576                         void *mapaddr = kmap_atomic(page, KM_USER1);
 577                         void *buf = mapaddr + offset;
 578                         offset += fuse_copy_do(cs, &buf, &count);
 579                         kunmap_atomic(mapaddr, KM_USER1);
 580                 } else
 581                         offset += fuse_copy_do(cs, NULL, &count);
 582         }
 583         if (page && !cs->write)
 584                 flush_dcache_page(page);
 585         return 0;
 586 }
 587
 588 /* Copy pages in the request to/from userspace buffer */
 589 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 590                            int zeroing)
 591 {
 592         unsigned i;
 593         struct fuse_req *req = cs->req;
 594         unsigned offset = req->page_offset;
 595         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 596
 597         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 598                 struct page *page = req->pages[i];
 599                 int err = fuse_copy_page(cs, page, offset, count, zeroing);
 600                 if (err)
 601                         return err;
 602
 603                 nbytes -= count;
 604                 count = min(nbytes, (unsigned) PAGE_SIZE);
 605                 offset = 0;
 606         }
 607         return 0;
 608 }
 609
 610 /* Copy a single argument in the request to/from userspace buffer */
 611 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 612 {
 613         while (size) {
 614                 int err;
 615                 if (!cs->len && (err = fuse_copy_fill(cs)))
 616                         return err;
 617                 fuse_copy_do(cs, &val, &size);
 618         }
 619         return 0;
 620 }
 621
 622 /* Copy request arguments to/from userspace buffer */
 623 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 624                           unsigned argpages, struct fuse_arg *args,
 625                           int zeroing)
 626 {
 627         int err = 0;
 628         unsigned i;
 629
 630         for (i = 0; !err && i < numargs; i++)  {
 631                 struct fuse_arg *arg = &args[i];
 632                 if (i == numargs - 1 && argpages)
 633                         err = fuse_copy_pages(cs, arg->size, zeroing);
 634                 else
 635                         err = fuse_copy_one(cs, arg->value, arg->size);
 636         }
 637         return err;
 638 }
 639
 640 static int request_pending(struct fuse_conn *fc)
 641 {
 642         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 643 }
 644
 645 /* Wait until a request is available on the pending list */
 646 static void request_wait(struct fuse_conn *fc)
 647 {
 648         DECLARE_WAITQUEUE(wait, current);
 649
 650         add_wait_queue_exclusive(&fc->waitq, &wait);
 651         while (fc->connected && !request_pending(fc)) {
 652                 set_current_state(TASK_INTERRUPTIBLE);
 653                 if (signal_pending(current))
 654                         break;
 655
 656                 spin_unlock(&fc->lock);
 657                 schedule();
 658                 spin_lock(&fc->lock);
 659         }
 660         set_current_state(TASK_RUNNING);
 661         remove_wait_queue(&fc->waitq, &wait);
 662 }
 663
 664 /*
 665  * Transfer an interrupt request to userspace
 666  *
 667  * Unlike other requests this is assembled on demand, without a need
 668  * to allocate a separate fuse_req structure.
 669  *
 670  * Called with fc->lock held, releases it
 671  */
 672 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
 673                                const struct iovec *iov, unsigned long nr_segs)
 674         __releases(fc->lock)
 675 {
 676         struct fuse_copy_state cs;
 677         struct fuse_in_header ih;
 678         struct fuse_interrupt_in arg;
 679         unsigned reqsize = sizeof(ih) + sizeof(arg);
 680         int err;
 681
 682         list_del_init(&req->intr_entry);
 683         req->intr_unique = fuse_get_unique(fc);
 684         memset(&ih, 0, sizeof(ih));
 685         memset(&arg, 0, sizeof(arg));
 686         ih.len = reqsize;
 687         ih.opcode = FUSE_INTERRUPT;
 688         ih.unique = req->intr_unique;
 689         arg.unique = req->in.h.unique;
 690
 691         spin_unlock(&fc->lock);
 692         if (iov_length(iov, nr_segs) < reqsize)
 693                 return -EINVAL;
 694
 695         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
 696         err = fuse_copy_one(&cs, &ih, sizeof(ih));
 697         if (!err)
 698                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
 699         fuse_copy_finish(&cs);
 700
 701         return err ? err : reqsize;
 702 }
 703
 704 /*
 705  * Read a single request into the userspace filesystem's buffer.  This
 706  * function waits until a request is available, then removes it from
 707  * the pending list and copies request data to userspace buffer.  If
 708  * no reply is needed (FORGET) or request has been aborted or there
 709  * was an error during the copying then it's finished by calling
 710  * request_end().  Otherwise add it to the processing list, and set
 711  * the 'sent' flag.
 712  */
 713 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
 714                               unsigned long nr_segs, loff_t pos)
 715 {
 716         int err;
 717         struct fuse_req *req;
 718         struct fuse_in *in;
 719         struct fuse_copy_state cs;
 720         unsigned reqsize;
 721         struct file *file = iocb->ki_filp;
 722         struct fuse_conn *fc = fuse_get_conn(file);
 723         if (!fc)
 724                 return -EPERM;
 725
 726  restart:
 727         spin_lock(&fc->lock);
 728         err = -EAGAIN;
 729         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 730             !request_pending(fc))
 731                 goto err_unlock;
 732
 733         request_wait(fc);
 734         err = -ENODEV;
 735         if (!fc->connected)
 736                 goto err_unlock;
 737         err = -ERESTARTSYS;
 738         if (!request_pending(fc))
 739                 goto err_unlock;
 740
 741         if (!list_empty(&fc->interrupts)) {
 742                 req = list_entry(fc->interrupts.next, struct fuse_req,
 743                                  intr_entry);
 744                 return fuse_read_interrupt(fc, req, iov, nr_segs);
 745         }
 746
 747         req = list_entry(fc->pending.next, struct fuse_req, list);
 748         req->state = FUSE_REQ_READING;
 749         list_move(&req->list, &fc->io);
 750
 751         in = &req->in;
 752         reqsize = in->h.len;
 753         /* If request is too large, reply with an error and restart the read */
 754         if (iov_length(iov, nr_segs) < reqsize) {
 755                 req->out.h.error = -EIO;
 756                 /* SETXATTR is special, since it may contain too large data */
 757                 if (in->h.opcode == FUSE_SETXATTR)
 758                         req->out.h.error = -E2BIG;
 759                 request_end(fc, req);
 760                 goto restart;
 761         }
 762         spin_unlock(&fc->lock);
 763         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
 764         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
 765         if (!err)
 766                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
 767                                      (struct fuse_arg *) in->args, 0);
 768         fuse_copy_finish(&cs);
 769         spin_lock(&fc->lock);
 770         req->locked = 0;
 771         if (req->aborted) {
 772                 request_end(fc, req);
 773                 return -ENODEV;
 774         }
 775         if (err) {
 776                 req->out.h.error = -EIO;
 777                 request_end(fc, req);
 778                 return err;
 779         }
 780         if (!req->isreply)
 781                 request_end(fc, req);
 782         else {
 783                 req->state = FUSE_REQ_SENT;
 784                 list_move_tail(&req->list, &fc->processing);
 785                 if (req->interrupted)
 786                         queue_interrupt(fc, req);
 787                 spin_unlock(&fc->lock);
 788         }
 789         return reqsize;
 790
 791  err_unlock:
 792         spin_unlock(&fc->lock);
 793         return err;
 794 }
 795
 796 /* Look up request on processing list by unique ID */
 797 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
 798 {
 799         struct list_head *entry;
 800
 801         list_for_each(entry, &fc->processing) {
 802                 struct fuse_req *req;
 803                 req = list_entry(entry, struct fuse_req, list);
 804                 if (req->in.h.unique == unique || req->intr_unique == unique)
 805                         return req;
 806         }
 807         return NULL;
 808 }
 809
 810 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
 811                          unsigned nbytes)
 812 {
 813         unsigned reqsize = sizeof(struct fuse_out_header);
 814
 815         if (out->h.error)
 816                 return nbytes != reqsize ? -EINVAL : 0;
 817
 818         reqsize += len_args(out->numargs, out->args);
 819
 820         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
 821                 return -EINVAL;
 822         else if (reqsize > nbytes) {
 823                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
 824                 unsigned diffsize = reqsize - nbytes;
 825                 if (diffsize > lastarg->size)
 826                         return -EINVAL;
 827                 lastarg->size -= diffsize;
 828         }
 829         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
 830                               out->page_zeroing);
 831 }
 832
 833 /*
 834  * Write a single reply to a request.  First the header is copied from
 835  * the write buffer.  The request is then searched on the processing
 836  * list by the unique ID found in the header.  If found, then remove
 837  * it from the list and copy the rest of the buffer to the request.
 838  * The request is finished by calling request_end()
 839  */
 840 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
 841                                unsigned long nr_segs, loff_t pos)
 842 {
 843         int err;
 844         unsigned nbytes = iov_length(iov, nr_segs);
 845         struct fuse_req *req;
 846         struct fuse_out_header oh;
 847         struct fuse_copy_state cs;
 848         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
 849         if (!fc)
 850                 return -EPERM;
 851
 852         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
 853         if (nbytes < sizeof(struct fuse_out_header))
 854                 return -EINVAL;
 855
 856         err = fuse_copy_one(&cs, &oh, sizeof(oh));
 857         if (err)
 858                 goto err_finish;
 859         err = -EINVAL;
 860         if (!oh.unique || oh.error <= -1000 || oh.error > 0 ||
 861             oh.len != nbytes)
 862                 goto err_finish;
 863
 864         spin_lock(&fc->lock);
 865         err = -ENOENT;
 866         if (!fc->connected)
 867                 goto err_unlock;
 868
 869         req = request_find(fc, oh.unique);
 870         if (!req)
 871                 goto err_unlock;
 872
 873         if (req->aborted) {
 874                 spin_unlock(&fc->lock);
 875                 fuse_copy_finish(&cs);
 876                 spin_lock(&fc->lock);
 877                 request_end(fc, req);
 878                 return -ENOENT;
 879         }
 880         /* Is it an interrupt reply? */
 881         if (req->intr_unique == oh.unique) {
 882                 err = -EINVAL;
 883                 if (nbytes != sizeof(struct fuse_out_header))
 884                         goto err_unlock;
 885
 886                 if (oh.error == -ENOSYS)
 887                         fc->no_interrupt = 1;
 888                 else if (oh.error == -EAGAIN)
 889                         queue_interrupt(fc, req);
 890
 891                 spin_unlock(&fc->lock);
 892                 fuse_copy_finish(&cs);
 893                 return nbytes;
 894         }
 895
 896         req->state = FUSE_REQ_WRITING;
 897         list_move(&req->list, &fc->io);
 898         req->out.h = oh;
 899         req->locked = 1;
 900         cs.req = req;
 901         spin_unlock(&fc->lock);
 902
 903         err = copy_out_args(&cs, &req->out, nbytes);
 904         fuse_copy_finish(&cs);
 905
 906         spin_lock(&fc->lock);
 907         req->locked = 0;
 908         if (!err) {
 909                 if (req->aborted)
 910                         err = -ENOENT;
 911         } else if (!req->aborted)
 912                 req->out.h.error = -EIO;
 913         request_end(fc, req);
 914
 915         return err ? err : nbytes;
 916
 917  err_unlock:
 918         spin_unlock(&fc->lock);
 919  err_finish:
 920         fuse_copy_finish(&cs);
 921         return err;
 922 }
 923
 924 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
 925 {
 926         unsigned mask = POLLOUT | POLLWRNORM;
 927         struct fuse_conn *fc = fuse_get_conn(file);
 928         if (!fc)
 929                 return POLLERR;
 930
 931         poll_wait(file, &fc->waitq, wait);
 932
 933         spin_lock(&fc->lock);
 934         if (!fc->connected)
 935                 mask = POLLERR;
 936         else if (request_pending(fc))
 937                 mask |= POLLIN | POLLRDNORM;
 938         spin_unlock(&fc->lock);
 939
 940         return mask;
 941 }
 942
 943 /*
 944  * Abort all requests on the given list (pending or processing)
 945  *
 946  * This function releases and reacquires fc->lock
 947  */
 948 static void end_requests(struct fuse_conn *fc, struct list_head *head)
 949 {
 950         while (!list_empty(head)) {
 951                 struct fuse_req *req;
 952                 req = list_entry(head->next, struct fuse_req, list);
 953                 req->out.h.error = -ECONNABORTED;
 954                 request_end(fc, req);
 955                 spin_lock(&fc->lock);
 956         }
 957 }
 958
 959 /*
 960  * Abort requests under I/O
 961  *
 962  * The requests are set to aborted and finished, and the request
 963  * waiter is woken up.  This will make request_wait_answer() wait
 964  * until the request is unlocked and then return.
 965  *
 966  * If the request is asynchronous, then the end function needs to be
 967  * called after waiting for the request to be unlocked (if it was
 968  * locked).
 969  */
 970 static void end_io_requests(struct fuse_conn *fc)
 971 {
 972         while (!list_empty(&fc->io)) {
 973                 struct fuse_req *req =
 974                         list_entry(fc->io.next, struct fuse_req, list);
 975                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 976
 977                 req->aborted = 1;
 978                 req->out.h.error = -ECONNABORTED;
 979                 req->state = FUSE_REQ_FINISHED;
 980                 list_del_init(&req->list);
 981                 wake_up(&req->waitq);
 982                 if (end) {
 983                         req->end = NULL;
 984                         /* The end function will consume this reference */
 985                         __fuse_get_request(req);
 986                         spin_unlock(&fc->lock);
 987                         wait_event(req->waitq, !req->locked);
 988                         end(fc, req);
 989                         spin_lock(&fc->lock);
 990                 }
 991         }
 992 }
 993
 994 /*
 995  * Abort all requests.
 996  *
 997  * Emergency exit in case of a malicious or accidental deadlock, or
 998  * just a hung filesystem.
 999  *
1000  * The same effect is usually achievable through killing the
1001  * filesystem daemon and all users of the filesystem.  The exception
1002  * is the combination of an asynchronous request and the tricky
1003  * deadlock (see Documentation/filesystems/fuse.txt).
1004  *
1005  * During the aborting, progression of requests from the pending and
1006  * processing lists onto the io list, and progression of new requests
1007  * onto the pending list is prevented by req->connected being false.
1008  *
1009  * Progression of requests under I/O to the processing list is
1010  * prevented by the req->aborted flag being true for these requests.
1011  * For this reason requests on the io list must be aborted first.
1012  */
1013 void fuse_abort_conn(struct fuse_conn *fc)
1014 {
1015         spin_lock(&fc->lock);
1016         if (fc->connected) {
1017                 fc->connected = 0;
1018                 fc->blocked = 0;
1019                 end_io_requests(fc);
1020                 end_requests(fc, &fc->pending);
1021                 end_requests(fc, &fc->processing);
1022                 wake_up_all(&fc->waitq);
1023                 wake_up_all(&fc->blocked_waitq);
1024                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1025         }
1026         spin_unlock(&fc->lock);
1027 }
1028
1029 static int fuse_dev_release(struct inode *inode, struct file *file)
1030 {
1031         struct fuse_conn *fc = fuse_get_conn(file);
1032         if (fc) {
1033                 spin_lock(&fc->lock);
1034                 fc->connected = 0;
1035                 end_requests(fc, &fc->pending);
1036                 end_requests(fc, &fc->processing);
1037                 spin_unlock(&fc->lock);
1038                 fasync_helper(-1, file, 0, &fc->fasync);
1039                 fuse_conn_put(fc);
1040         }
1041
1042         return 0;
1043 }
1044
1045 static int fuse_dev_fasync(int fd, struct file *file, int on)
1046 {
1047         struct fuse_conn *fc = fuse_get_conn(file);
1048         if (!fc)
1049                 return -EPERM;
1050
1051         /* No locking - fasync_helper does its own locking */
1052         return fasync_helper(fd, file, on, &fc->fasync);
1053 }
1054
1055 const struct file_operations fuse_dev_operations = {
1056         .owner          = THIS_MODULE,
1057         .llseek         = no_llseek,
1058         .read           = do_sync_read,
1059         .aio_read       = fuse_dev_read,
1060         .write          = do_sync_write,
1061         .aio_write      = fuse_dev_write,
1062         .poll           = fuse_dev_poll,
1063         .release        = fuse_dev_release,
1064         .fasync         = fuse_dev_fasync,
1065 };
1066
1067 static struct miscdevice fuse_miscdevice = {
1068         .minor = FUSE_MINOR,
1069         .name  = "fuse",
1070         .fops = &fuse_dev_operations,
1071 };
1072
1073 int __init fuse_dev_init(void)
1074 {
1075         int err = -ENOMEM;
1076         fuse_req_cachep = kmem_cache_create("fuse_request",
1077                                             sizeof(struct fuse_req),
1078                                             0, 0, NULL);
1079         if (!fuse_req_cachep)
1080                 goto out;
1081
1082         err = misc_register(&fuse_miscdevice);
1083         if (err)
1084                 goto out_cache_clean;
1085
1086         return 0;
1087
1088  out_cache_clean:
1089         kmem_cache_destroy(fuse_req_cachep);
1090  out:
1091         return err;
1092 }
1093
1094 void fuse_dev_cleanup(void)
1095 {
1096         misc_deregister(&fuse_miscdevice);
1097         kmem_cache_destroy(fuse_req_cachep);
1098 }