fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/uio.h>
  15 #include <linux/miscdevice.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19
  20 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  21
  22 static struct kmem_cache *fuse_req_cachep;
  23
  24 static struct fuse_conn *fuse_get_conn(struct file *file)
  25 {
  26         /*
  27          * Lockless access is OK, because file->private data is set
  28          * once during mount and is valid until the file is released.
  29          */
  30         return file->private_data;
  31 }
  32
  33 static void fuse_request_init(struct fuse_req *req)
  34 {
  35         memset(req, 0, sizeof(*req));
  36         INIT_LIST_HEAD(&req->list);
  37         INIT_LIST_HEAD(&req->intr_entry);
  38         init_waitqueue_head(&req->waitq);
  39         atomic_set(&req->count, 1);
  40 }
  41
  42 struct fuse_req *fuse_request_alloc(void)
  43 {
  44         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_KERNEL);
  45         if (req)
  46                 fuse_request_init(req);
  47         return req;
  48 }
  49
  50 struct fuse_req *fuse_request_alloc_nofs(void)
  51 {
  52         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_NOFS);
  53         if (req)
  54                 fuse_request_init(req);
  55         return req;
  56 }
  57
  58 void fuse_request_free(struct fuse_req *req)
  59 {
  60         kmem_cache_free(fuse_req_cachep, req);
  61 }
  62
  63 static void block_sigs(sigset_t *oldset)
  64 {
  65         sigset_t mask;
  66
  67         siginitsetinv(&mask, sigmask(SIGKILL));
  68         sigprocmask(SIG_BLOCK, &mask, oldset);
  69 }
  70
  71 static void restore_sigs(sigset_t *oldset)
  72 {
  73         sigprocmask(SIG_SETMASK, oldset, NULL);
  74 }
  75
  76 static void __fuse_get_request(struct fuse_req *req)
  77 {
  78         atomic_inc(&req->count);
  79 }
  80
  81 /* Must be called with > 1 refcount */
  82 static void __fuse_put_request(struct fuse_req *req)
  83 {
  84         BUG_ON(atomic_read(&req->count) < 2);
  85         atomic_dec(&req->count);
  86 }
  87
  88 static void fuse_req_init_context(struct fuse_req *req)
  89 {
  90         req->in.h.uid = current->fsuid;
  91         req->in.h.gid = current->fsgid;
  92         req->in.h.pid = current->pid;
  93 }
  94
  95 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
  96 {
  97         struct fuse_req *req;
  98         sigset_t oldset;
  99         int intr;
 100         int err;
 101
 102         atomic_inc(&fc->num_waiting);
 103         block_sigs(&oldset);
 104         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
 105         restore_sigs(&oldset);
 106         err = -EINTR;
 107         if (intr)
 108                 goto out;
 109
 110         err = -ENOTCONN;
 111         if (!fc->connected)
 112                 goto out;
 113
 114         req = fuse_request_alloc();
 115         err = -ENOMEM;
 116         if (!req)
 117                 goto out;
 118
 119         fuse_req_init_context(req);
 120         req->waiting = 1;
 121         return req;
 122
 123  out:
 124         atomic_dec(&fc->num_waiting);
 125         return ERR_PTR(err);
 126 }
 127
 128 /*
 129  * Return request in fuse_file->reserved_req.  However that may
 130  * currently be in use.  If that is the case, wait for it to become
 131  * available.
 132  */
 133 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 134                                          struct file *file)
 135 {
 136         struct fuse_req *req = NULL;
 137         struct fuse_file *ff = file->private_data;
 138
 139         do {
 140                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 141                 spin_lock(&fc->lock);
 142                 if (ff->reserved_req) {
 143                         req = ff->reserved_req;
 144                         ff->reserved_req = NULL;
 145                         get_file(file);
 146                         req->stolen_file = file;
 147                 }
 148                 spin_unlock(&fc->lock);
 149         } while (!req);
 150
 151         return req;
 152 }
 153
 154 /*
 155  * Put stolen request back into fuse_file->reserved_req
 156  */
 157 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 158 {
 159         struct file *file = req->stolen_file;
 160         struct fuse_file *ff = file->private_data;
 161
 162         spin_lock(&fc->lock);
 163         fuse_request_init(req);
 164         BUG_ON(ff->reserved_req);
 165         ff->reserved_req = req;
 166         wake_up_all(&fc->reserved_req_waitq);
 167         spin_unlock(&fc->lock);
 168         fput(file);
 169 }
 170
 171 /*
 172  * Gets a requests for a file operation, always succeeds
 173  *
 174  * This is used for sending the FLUSH request, which must get to
 175  * userspace, due to POSIX locks which may need to be unlocked.
 176  *
 177  * If allocation fails due to OOM, use the reserved request in
 178  * fuse_file.
 179  *
 180  * This is very unlikely to deadlock accidentally, since the
 181  * filesystem should not have it's own file open.  If deadlock is
 182  * intentional, it can still be broken by "aborting" the filesystem.
 183  */
 184 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 185 {
 186         struct fuse_req *req;
 187
 188         atomic_inc(&fc->num_waiting);
 189         wait_event(fc->blocked_waitq, !fc->blocked);
 190         req = fuse_request_alloc();
 191         if (!req)
 192                 req = get_reserved_req(fc, file);
 193
 194         fuse_req_init_context(req);
 195         req->waiting = 1;
 196         return req;
 197 }
 198
 199 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 200 {
 201         if (atomic_dec_and_test(&req->count)) {
 202                 if (req->waiting)
 203                         atomic_dec(&fc->num_waiting);
 204
 205                 if (req->stolen_file)
 206                         put_reserved_req(fc, req);
 207                 else
 208                         fuse_request_free(req);
 209         }
 210 }
 211
 212 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 213 {
 214         unsigned nbytes = 0;
 215         unsigned i;
 216
 217         for (i = 0; i < numargs; i++)
 218                 nbytes += args[i].size;
 219
 220         return nbytes;
 221 }
 222
 223 static u64 fuse_get_unique(struct fuse_conn *fc)
 224 {
 225         fc->reqctr++;
 226         /* zero is special */
 227         if (fc->reqctr == 0)
 228                 fc->reqctr = 1;
 229
 230         return fc->reqctr;
 231 }
 232
 233 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 234 {
 235         req->in.h.unique = fuse_get_unique(fc);
 236         req->in.h.len = sizeof(struct fuse_in_header) +
 237                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 238         list_add_tail(&req->list, &fc->pending);
 239         req->state = FUSE_REQ_PENDING;
 240         if (!req->waiting) {
 241                 req->waiting = 1;
 242                 atomic_inc(&fc->num_waiting);
 243         }
 244         wake_up(&fc->waitq);
 245         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 246 }
 247
 248 static void flush_bg_queue(struct fuse_conn *fc)
 249 {
 250         while (fc->active_background < FUSE_MAX_BACKGROUND &&
 251                !list_empty(&fc->bg_queue)) {
 252                 struct fuse_req *req;
 253
 254                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
 255                 list_del(&req->list);
 256                 fc->active_background++;
 257                 queue_request(fc, req);
 258         }
 259 }
 260
 261 /*
 262  * This function is called when a request is finished.  Either a reply
 263  * has arrived or it was aborted (and not yet sent) or some error
 264  * occurred during communication with userspace, or the device file
 265  * was closed.  The requester thread is woken up (if still waiting),
 266  * the 'end' callback is called if given, else the reference to the
 267  * request is released
 268  *
 269  * Called with fc->lock, unlocks it
 270  */
 271 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 272         __releases(fc->lock)
 273 {
 274         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 275         req->end = NULL;
 276         list_del(&req->list);
 277         list_del(&req->intr_entry);
 278         req->state = FUSE_REQ_FINISHED;
 279         if (req->background) {
 280                 if (fc->num_background == FUSE_MAX_BACKGROUND) {
 281                         fc->blocked = 0;
 282                         wake_up_all(&fc->blocked_waitq);
 283                 }
 284                 if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 285                         clear_bdi_congested(&fc->bdi, READ);
 286                         clear_bdi_congested(&fc->bdi, WRITE);
 287                 }
 288                 fc->num_background--;
 289                 fc->active_background--;
 290                 flush_bg_queue(fc);
 291         }
 292         spin_unlock(&fc->lock);
 293         wake_up(&req->waitq);
 294         if (end)
 295                 end(fc, req);
 296         fuse_put_request(fc, req);
 297 }
 298
 299 static void wait_answer_interruptible(struct fuse_conn *fc,
 300                                       struct fuse_req *req)
 301         __releases(fc->lock) __acquires(fc->lock)
 302 {
 303         if (signal_pending(current))
 304                 return;
 305
 306         spin_unlock(&fc->lock);
 307         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
 308         spin_lock(&fc->lock);
 309 }
 310
 311 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
 312 {
 313         list_add_tail(&req->intr_entry, &fc->interrupts);
 314         wake_up(&fc->waitq);
 315         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 316 }
 317
 318 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 319         __releases(fc->lock) __acquires(fc->lock)
 320 {
 321         if (!fc->no_interrupt) {
 322                 /* Any signal may interrupt this */
 323                 wait_answer_interruptible(fc, req);
 324
 325                 if (req->aborted)
 326                         goto aborted;
 327                 if (req->state == FUSE_REQ_FINISHED)
 328                         return;
 329
 330                 req->interrupted = 1;
 331                 if (req->state == FUSE_REQ_SENT)
 332                         queue_interrupt(fc, req);
 333         }
 334
 335         if (!req->force) {
 336                 sigset_t oldset;
 337
 338                 /* Only fatal signals may interrupt this */
 339                 block_sigs(&oldset);
 340                 wait_answer_interruptible(fc, req);
 341                 restore_sigs(&oldset);
 342
 343                 if (req->aborted)
 344                         goto aborted;
 345                 if (req->state == FUSE_REQ_FINISHED)
 346                         return;
 347
 348                 /* Request is not yet in userspace, bail out */
 349                 if (req->state == FUSE_REQ_PENDING) {
 350                         list_del(&req->list);
 351                         __fuse_put_request(req);
 352                         req->out.h.error = -EINTR;
 353                         return;
 354                 }
 355         }
 356
 357         /*
 358          * Either request is already in userspace, or it was forced.
 359          * Wait it out.
 360          */
 361         spin_unlock(&fc->lock);
 362         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
 363         spin_lock(&fc->lock);
 364
 365         if (!req->aborted)
 366                 return;
 367
 368  aborted:
 369         BUG_ON(req->state != FUSE_REQ_FINISHED);
 370         if (req->locked) {
 371                 /* This is uninterruptible sleep, because data is
 372                    being copied to/from the buffers of req.  During
 373                    locked state, there mustn't be any filesystem
 374                    operation (e.g. page fault), since that could lead
 375                    to deadlock */
 376                 spin_unlock(&fc->lock);
 377                 wait_event(req->waitq, !req->locked);
 378                 spin_lock(&fc->lock);
 379         }
 380 }
 381
 382 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 383 {
 384         req->isreply = 1;
 385         spin_lock(&fc->lock);
 386         if (!fc->connected)
 387                 req->out.h.error = -ENOTCONN;
 388         else if (fc->conn_error)
 389                 req->out.h.error = -ECONNREFUSED;
 390         else {
 391                 queue_request(fc, req);
 392                 /* acquire extra reference, since request is still needed
 393                    after request_end() */
 394                 __fuse_get_request(req);
 395
 396                 request_wait_answer(fc, req);
 397         }
 398         spin_unlock(&fc->lock);
 399 }
 400
 401 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
 402                                             struct fuse_req *req)
 403 {
 404         req->background = 1;
 405         fc->num_background++;
 406         if (fc->num_background == FUSE_MAX_BACKGROUND)
 407                 fc->blocked = 1;
 408         if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
 409                 set_bdi_congested(&fc->bdi, READ);
 410                 set_bdi_congested(&fc->bdi, WRITE);
 411         }
 412         list_add_tail(&req->list, &fc->bg_queue);
 413         flush_bg_queue(fc);
 414 }
 415
 416 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 417 {
 418         spin_lock(&fc->lock);
 419         if (fc->connected) {
 420                 fuse_request_send_nowait_locked(fc, req);
 421                 spin_unlock(&fc->lock);
 422         } else {
 423                 req->out.h.error = -ENOTCONN;
 424                 request_end(fc, req);
 425         }
 426 }
 427
 428 void fuse_request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 429 {
 430         req->isreply = 0;
 431         fuse_request_send_nowait(fc, req);
 432 }
 433
 434 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 435 {
 436         req->isreply = 1;
 437         fuse_request_send_nowait(fc, req);
 438 }
 439
 440 /*
 441  * Called under fc->lock
 442  *
 443  * fc->connected must have been checked previously
 444  */
 445 void fuse_request_send_background_locked(struct fuse_conn *fc,
 446                                          struct fuse_req *req)
 447 {
 448         req->isreply = 1;
 449         fuse_request_send_nowait_locked(fc, req);
 450 }
 451
 452 /*
 453  * Lock the request.  Up to the next unlock_request() there mustn't be
 454  * anything that could cause a page-fault.  If the request was already
 455  * aborted bail out.
 456  */
 457 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 458 {
 459         int err = 0;
 460         if (req) {
 461                 spin_lock(&fc->lock);
 462                 if (req->aborted)
 463                         err = -ENOENT;
 464                 else
 465                         req->locked = 1;
 466                 spin_unlock(&fc->lock);
 467         }
 468         return err;
 469 }
 470
 471 /*
 472  * Unlock request.  If it was aborted during being locked, the
 473  * requester thread is currently waiting for it to be unlocked, so
 474  * wake it up.
 475  */
 476 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 477 {
 478         if (req) {
 479                 spin_lock(&fc->lock);
 480                 req->locked = 0;
 481                 if (req->aborted)
 482                         wake_up(&req->waitq);
 483                 spin_unlock(&fc->lock);
 484         }
 485 }
 486
 487 struct fuse_copy_state {
 488         struct fuse_conn *fc;
 489         int write;
 490         struct fuse_req *req;
 491         const struct iovec *iov;
 492         unsigned long nr_segs;
 493         unsigned long seglen;
 494         unsigned long addr;
 495         struct page *pg;
 496         void *mapaddr;
 497         void *buf;
 498         unsigned len;
 499 };
 500
 501 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
 502                            int write, struct fuse_req *req,
 503                            const struct iovec *iov, unsigned long nr_segs)
 504 {
 505         memset(cs, 0, sizeof(*cs));
 506         cs->fc = fc;
 507         cs->write = write;
 508         cs->req = req;
 509         cs->iov = iov;
 510         cs->nr_segs = nr_segs;
 511 }
 512
 513 /* Unmap and put previous page of userspace buffer */
 514 static void fuse_copy_finish(struct fuse_copy_state *cs)
 515 {
 516         if (cs->mapaddr) {
 517                 kunmap_atomic(cs->mapaddr, KM_USER0);
 518                 if (cs->write) {
 519                         flush_dcache_page(cs->pg);
 520                         set_page_dirty_lock(cs->pg);
 521                 }
 522                 put_page(cs->pg);
 523                 cs->mapaddr = NULL;
 524         }
 525 }
 526
 527 /*
 528  * Get another pagefull of userspace buffer, and map it to kernel
 529  * address space, and lock request
 530  */
 531 static int fuse_copy_fill(struct fuse_copy_state *cs)
 532 {
 533         unsigned long offset;
 534         int err;
 535
 536         unlock_request(cs->fc, cs->req);
 537         fuse_copy_finish(cs);
 538         if (!cs->seglen) {
 539                 BUG_ON(!cs->nr_segs);
 540                 cs->seglen = cs->iov[0].iov_len;
 541                 cs->addr = (unsigned long) cs->iov[0].iov_base;
 542                 cs->iov++;
 543                 cs->nr_segs--;
 544         }
 545         down_read(&current->mm->mmap_sem);
 546         err = get_user_pages(current, current->mm, cs->addr, 1, cs->write, 0,
 547                              &cs->pg, NULL);
 548         up_read(&current->mm->mmap_sem);
 549         if (err < 0)
 550                 return err;
 551         BUG_ON(err != 1);
 552         offset = cs->addr % PAGE_SIZE;
 553         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
 554         cs->buf = cs->mapaddr + offset;
 555         cs->len = min(PAGE_SIZE - offset, cs->seglen);
 556         cs->seglen -= cs->len;
 557         cs->addr += cs->len;
 558
 559         return lock_request(cs->fc, cs->req);
 560 }
 561
 562 /* Do as much copy to/from userspace buffer as we can */
 563 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 564 {
 565         unsigned ncpy = min(*size, cs->len);
 566         if (val) {
 567                 if (cs->write)
 568                         memcpy(cs->buf, *val, ncpy);
 569                 else
 570                         memcpy(*val, cs->buf, ncpy);
 571                 *val += ncpy;
 572         }
 573         *size -= ncpy;
 574         cs->len -= ncpy;
 575         cs->buf += ncpy;
 576         return ncpy;
 577 }
 578
 579 /*
 580  * Copy a page in the request to/from the userspace buffer.  Must be
 581  * done atomically
 582  */
 583 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
 584                           unsigned offset, unsigned count, int zeroing)
 585 {
 586         if (page && zeroing && count < PAGE_SIZE) {
 587                 void *mapaddr = kmap_atomic(page, KM_USER1);
 588                 memset(mapaddr, 0, PAGE_SIZE);
 589                 kunmap_atomic(mapaddr, KM_USER1);
 590         }
 591         while (count) {
 592                 if (!cs->len) {
 593                         int err = fuse_copy_fill(cs);
 594                         if (err)
 595                                 return err;
 596                 }
 597                 if (page) {
 598                         void *mapaddr = kmap_atomic(page, KM_USER1);
 599                         void *buf = mapaddr + offset;
 600                         offset += fuse_copy_do(cs, &buf, &count);
 601                         kunmap_atomic(mapaddr, KM_USER1);
 602                 } else
 603                         offset += fuse_copy_do(cs, NULL, &count);
 604         }
 605         if (page && !cs->write)
 606                 flush_dcache_page(page);
 607         return 0;
 608 }
 609
 610 /* Copy pages in the request to/from userspace buffer */
 611 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 612                            int zeroing)
 613 {
 614         unsigned i;
 615         struct fuse_req *req = cs->req;
 616         unsigned offset = req->page_offset;
 617         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
 618
 619         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
 620                 struct page *page = req->pages[i];
 621                 int err = fuse_copy_page(cs, page, offset, count, zeroing);
 622                 if (err)
 623                         return err;
 624
 625                 nbytes -= count;
 626                 count = min(nbytes, (unsigned) PAGE_SIZE);
 627                 offset = 0;
 628         }
 629         return 0;
 630 }
 631
 632 /* Copy a single argument in the request to/from userspace buffer */
 633 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
 634 {
 635         while (size) {
 636                 if (!cs->len) {
 637                         int err = fuse_copy_fill(cs);
 638                         if (err)
 639                                 return err;
 640                 }
 641                 fuse_copy_do(cs, &val, &size);
 642         }
 643         return 0;
 644 }
 645
 646 /* Copy request arguments to/from userspace buffer */
 647 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
 648                           unsigned argpages, struct fuse_arg *args,
 649                           int zeroing)
 650 {
 651         int err = 0;
 652         unsigned i;
 653
 654         for (i = 0; !err && i < numargs; i++)  {
 655                 struct fuse_arg *arg = &args[i];
 656                 if (i == numargs - 1 && argpages)
 657                         err = fuse_copy_pages(cs, arg->size, zeroing);
 658                 else
 659                         err = fuse_copy_one(cs, arg->value, arg->size);
 660         }
 661         return err;
 662 }
 663
 664 static int request_pending(struct fuse_conn *fc)
 665 {
 666         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
 667 }
 668
 669 /* Wait until a request is available on the pending list */
 670 static void request_wait(struct fuse_conn *fc)
 671 {
 672         DECLARE_WAITQUEUE(wait, current);
 673
 674         add_wait_queue_exclusive(&fc->waitq, &wait);
 675         while (fc->connected && !request_pending(fc)) {
 676                 set_current_state(TASK_INTERRUPTIBLE);
 677                 if (signal_pending(current))
 678                         break;
 679
 680                 spin_unlock(&fc->lock);
 681                 schedule();
 682                 spin_lock(&fc->lock);
 683         }
 684         set_current_state(TASK_RUNNING);
 685         remove_wait_queue(&fc->waitq, &wait);
 686 }
 687
 688 /*
 689  * Transfer an interrupt request to userspace
 690  *
 691  * Unlike other requests this is assembled on demand, without a need
 692  * to allocate a separate fuse_req structure.
 693  *
 694  * Called with fc->lock held, releases it
 695  */
 696 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
 697                                const struct iovec *iov, unsigned long nr_segs)
 698         __releases(fc->lock)
 699 {
 700         struct fuse_copy_state cs;
 701         struct fuse_in_header ih;
 702         struct fuse_interrupt_in arg;
 703         unsigned reqsize = sizeof(ih) + sizeof(arg);
 704         int err;
 705
 706         list_del_init(&req->intr_entry);
 707         req->intr_unique = fuse_get_unique(fc);
 708         memset(&ih, 0, sizeof(ih));
 709         memset(&arg, 0, sizeof(arg));
 710         ih.len = reqsize;
 711         ih.opcode = FUSE_INTERRUPT;
 712         ih.unique = req->intr_unique;
 713         arg.unique = req->in.h.unique;
 714
 715         spin_unlock(&fc->lock);
 716         if (iov_length(iov, nr_segs) < reqsize)
 717                 return -EINVAL;
 718
 719         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
 720         err = fuse_copy_one(&cs, &ih, sizeof(ih));
 721         if (!err)
 722                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
 723         fuse_copy_finish(&cs);
 724
 725         return err ? err : reqsize;
 726 }
 727
 728 /*
 729  * Read a single request into the userspace filesystem's buffer.  This
 730  * function waits until a request is available, then removes it from
 731  * the pending list and copies request data to userspace buffer.  If
 732  * no reply is needed (FORGET) or request has been aborted or there
 733  * was an error during the copying then it's finished by calling
 734  * request_end().  Otherwise add it to the processing list, and set
 735  * the 'sent' flag.
 736  */
 737 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
 738                               unsigned long nr_segs, loff_t pos)
 739 {
 740         int err;
 741         struct fuse_req *req;
 742         struct fuse_in *in;
 743         struct fuse_copy_state cs;
 744         unsigned reqsize;
 745         struct file *file = iocb->ki_filp;
 746         struct fuse_conn *fc = fuse_get_conn(file);
 747         if (!fc)
 748                 return -EPERM;
 749
 750  restart:
 751         spin_lock(&fc->lock);
 752         err = -EAGAIN;
 753         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
 754             !request_pending(fc))
 755                 goto err_unlock;
 756
 757         request_wait(fc);
 758         err = -ENODEV;
 759         if (!fc->connected)
 760                 goto err_unlock;
 761         err = -ERESTARTSYS;
 762         if (!request_pending(fc))
 763                 goto err_unlock;
 764
 765         if (!list_empty(&fc->interrupts)) {
 766                 req = list_entry(fc->interrupts.next, struct fuse_req,
 767                                  intr_entry);
 768                 return fuse_read_interrupt(fc, req, iov, nr_segs);
 769         }
 770
 771         req = list_entry(fc->pending.next, struct fuse_req, list);
 772         req->state = FUSE_REQ_READING;
 773         list_move(&req->list, &fc->io);
 774
 775         in = &req->in;
 776         reqsize = in->h.len;
 777         /* If request is too large, reply with an error and restart the read */
 778         if (iov_length(iov, nr_segs) < reqsize) {
 779                 req->out.h.error = -EIO;
 780                 /* SETXATTR is special, since it may contain too large data */
 781                 if (in->h.opcode == FUSE_SETXATTR)
 782                         req->out.h.error = -E2BIG;
 783                 request_end(fc, req);
 784                 goto restart;
 785         }
 786         spin_unlock(&fc->lock);
 787         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
 788         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
 789         if (!err)
 790                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
 791                                      (struct fuse_arg *) in->args, 0);
 792         fuse_copy_finish(&cs);
 793         spin_lock(&fc->lock);
 794         req->locked = 0;
 795         if (req->aborted) {
 796                 request_end(fc, req);
 797                 return -ENODEV;
 798         }
 799         if (err) {
 800                 req->out.h.error = -EIO;
 801                 request_end(fc, req);
 802                 return err;
 803         }
 804         if (!req->isreply)
 805                 request_end(fc, req);
 806         else {
 807                 req->state = FUSE_REQ_SENT;
 808                 list_move_tail(&req->list, &fc->processing);
 809                 if (req->interrupted)
 810                         queue_interrupt(fc, req);
 811                 spin_unlock(&fc->lock);
 812         }
 813         return reqsize;
 814
 815  err_unlock:
 816         spin_unlock(&fc->lock);
 817         return err;
 818 }
 819
 820 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
 821                             struct fuse_copy_state *cs)
 822 {
 823         struct fuse_notify_poll_wakeup_out outarg;
 824         int err;
 825
 826         if (size != sizeof(outarg))
 827                 return -EINVAL;
 828
 829         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
 830         if (err)
 831                 return err;
 832
 833         return fuse_notify_poll_wakeup(fc, &outarg);
 834 }
 835
 836 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
 837                        unsigned int size, struct fuse_copy_state *cs)
 838 {
 839         switch (code) {
 840         case FUSE_NOTIFY_POLL:
 841                 return fuse_notify_poll(fc, size, cs);
 842
 843         default:
 844                 return -EINVAL;
 845         }
 846 }
 847
 848 /* Look up request on processing list by unique ID */
 849 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
 850 {
 851         struct list_head *entry;
 852
 853         list_for_each(entry, &fc->processing) {
 854                 struct fuse_req *req;
 855                 req = list_entry(entry, struct fuse_req, list);
 856                 if (req->in.h.unique == unique || req->intr_unique == unique)
 857                         return req;
 858         }
 859         return NULL;
 860 }
 861
 862 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
 863                          unsigned nbytes)
 864 {
 865         unsigned reqsize = sizeof(struct fuse_out_header);
 866
 867         if (out->h.error)
 868                 return nbytes != reqsize ? -EINVAL : 0;
 869
 870         reqsize += len_args(out->numargs, out->args);
 871
 872         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
 873                 return -EINVAL;
 874         else if (reqsize > nbytes) {
 875                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
 876                 unsigned diffsize = reqsize - nbytes;
 877                 if (diffsize > lastarg->size)
 878                         return -EINVAL;
 879                 lastarg->size -= diffsize;
 880         }
 881         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
 882                               out->page_zeroing);
 883 }
 884
 885 /*
 886  * Write a single reply to a request.  First the header is copied from
 887  * the write buffer.  The request is then searched on the processing
 888  * list by the unique ID found in the header.  If found, then remove
 889  * it from the list and copy the rest of the buffer to the request.
 890  * The request is finished by calling request_end()
 891  */
 892 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
 893                                unsigned long nr_segs, loff_t pos)
 894 {
 895         int err;
 896         unsigned nbytes = iov_length(iov, nr_segs);
 897         struct fuse_req *req;
 898         struct fuse_out_header oh;
 899         struct fuse_copy_state cs;
 900         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
 901         if (!fc)
 902                 return -EPERM;
 903
 904         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
 905         if (nbytes < sizeof(struct fuse_out_header))
 906                 return -EINVAL;
 907
 908         err = fuse_copy_one(&cs, &oh, sizeof(oh));
 909         if (err)
 910                 goto err_finish;
 911
 912         err = -EINVAL;
 913         if (oh.len != nbytes)
 914                 goto err_finish;
 915
 916         /*
 917          * Zero oh.unique indicates unsolicited notification message
 918          * and error contains notification code.
 919          */
 920         if (!oh.unique) {
 921                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), &cs);
 922                 fuse_copy_finish(&cs);
 923                 return err ? err : nbytes;
 924         }
 925
 926         err = -EINVAL;
 927         if (oh.error <= -1000 || oh.error > 0)
 928                 goto err_finish;
 929
 930         spin_lock(&fc->lock);
 931         err = -ENOENT;
 932         if (!fc->connected)
 933                 goto err_unlock;
 934
 935         req = request_find(fc, oh.unique);
 936         if (!req)
 937                 goto err_unlock;
 938
 939         if (req->aborted) {
 940                 spin_unlock(&fc->lock);
 941                 fuse_copy_finish(&cs);
 942                 spin_lock(&fc->lock);
 943                 request_end(fc, req);
 944                 return -ENOENT;
 945         }
 946         /* Is it an interrupt reply? */
 947         if (req->intr_unique == oh.unique) {
 948                 err = -EINVAL;
 949                 if (nbytes != sizeof(struct fuse_out_header))
 950                         goto err_unlock;
 951
 952                 if (oh.error == -ENOSYS)
 953                         fc->no_interrupt = 1;
 954                 else if (oh.error == -EAGAIN)
 955                         queue_interrupt(fc, req);
 956
 957                 spin_unlock(&fc->lock);
 958                 fuse_copy_finish(&cs);
 959                 return nbytes;
 960         }
 961
 962         req->state = FUSE_REQ_WRITING;
 963         list_move(&req->list, &fc->io);
 964         req->out.h = oh;
 965         req->locked = 1;
 966         cs.req = req;
 967         spin_unlock(&fc->lock);
 968
 969         err = copy_out_args(&cs, &req->out, nbytes);
 970         fuse_copy_finish(&cs);
 971
 972         spin_lock(&fc->lock);
 973         req->locked = 0;
 974         if (!err) {
 975                 if (req->aborted)
 976                         err = -ENOENT;
 977         } else if (!req->aborted)
 978                 req->out.h.error = -EIO;
 979         request_end(fc, req);
 980
 981         return err ? err : nbytes;
 982
 983  err_unlock:
 984         spin_unlock(&fc->lock);
 985  err_finish:
 986         fuse_copy_finish(&cs);
 987         return err;
 988 }
 989
 990 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
 991 {
 992         unsigned mask = POLLOUT | POLLWRNORM;
 993         struct fuse_conn *fc = fuse_get_conn(file);
 994         if (!fc)
 995                 return POLLERR;
 996
 997         poll_wait(file, &fc->waitq, wait);
 998
 999         spin_lock(&fc->lock);
1000         if (!fc->connected)
1001                 mask = POLLERR;
1002         else if (request_pending(fc))
1003                 mask |= POLLIN | POLLRDNORM;
1004         spin_unlock(&fc->lock);
1005
1006         return mask;
1007 }
1008
1009 /*
1010  * Abort all requests on the given list (pending or processing)
1011  *
1012  * This function releases and reacquires fc->lock
1013  */
1014 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1015 {
1016         while (!list_empty(head)) {
1017                 struct fuse_req *req;
1018                 req = list_entry(head->next, struct fuse_req, list);
1019                 req->out.h.error = -ECONNABORTED;
1020                 request_end(fc, req);
1021                 spin_lock(&fc->lock);
1022         }
1023 }
1024
1025 /*
1026  * Abort requests under I/O
1027  *
1028  * The requests are set to aborted and finished, and the request
1029  * waiter is woken up.  This will make request_wait_answer() wait
1030  * until the request is unlocked and then return.
1031  *
1032  * If the request is asynchronous, then the end function needs to be
1033  * called after waiting for the request to be unlocked (if it was
1034  * locked).
1035  */
1036 static void end_io_requests(struct fuse_conn *fc)
1037         __releases(fc->lock) __acquires(fc->lock)
1038 {
1039         while (!list_empty(&fc->io)) {
1040                 struct fuse_req *req =
1041                         list_entry(fc->io.next, struct fuse_req, list);
1042                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1043
1044                 req->aborted = 1;
1045                 req->out.h.error = -ECONNABORTED;
1046                 req->state = FUSE_REQ_FINISHED;
1047                 list_del_init(&req->list);
1048                 wake_up(&req->waitq);
1049                 if (end) {
1050                         req->end = NULL;
1051                         __fuse_get_request(req);
1052                         spin_unlock(&fc->lock);
1053                         wait_event(req->waitq, !req->locked);
1054                         end(fc, req);
1055                         fuse_put_request(fc, req);
1056                         spin_lock(&fc->lock);
1057                 }
1058         }
1059 }
1060
1061 /*
1062  * Abort all requests.
1063  *
1064  * Emergency exit in case of a malicious or accidental deadlock, or
1065  * just a hung filesystem.
1066  *
1067  * The same effect is usually achievable through killing the
1068  * filesystem daemon and all users of the filesystem.  The exception
1069  * is the combination of an asynchronous request and the tricky
1070  * deadlock (see Documentation/filesystems/fuse.txt).
1071  *
1072  * During the aborting, progression of requests from the pending and
1073  * processing lists onto the io list, and progression of new requests
1074  * onto the pending list is prevented by req->connected being false.
1075  *
1076  * Progression of requests under I/O to the processing list is
1077  * prevented by the req->aborted flag being true for these requests.
1078  * For this reason requests on the io list must be aborted first.
1079  */
1080 void fuse_abort_conn(struct fuse_conn *fc)
1081 {
1082         spin_lock(&fc->lock);
1083         if (fc->connected) {
1084                 fc->connected = 0;
1085                 fc->blocked = 0;
1086                 end_io_requests(fc);
1087                 end_requests(fc, &fc->pending);
1088                 end_requests(fc, &fc->processing);
1089                 wake_up_all(&fc->waitq);
1090                 wake_up_all(&fc->blocked_waitq);
1091                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1092         }
1093         spin_unlock(&fc->lock);
1094 }
1095
1096 static int fuse_dev_release(struct inode *inode, struct file *file)
1097 {
1098         struct fuse_conn *fc = fuse_get_conn(file);
1099         if (fc) {
1100                 spin_lock(&fc->lock);
1101                 fc->connected = 0;
1102                 end_requests(fc, &fc->pending);
1103                 end_requests(fc, &fc->processing);
1104                 spin_unlock(&fc->lock);
1105                 fuse_conn_put(fc);
1106         }
1107
1108         return 0;
1109 }
1110
1111 static int fuse_dev_fasync(int fd, struct file *file, int on)
1112 {
1113         struct fuse_conn *fc = fuse_get_conn(file);
1114         if (!fc)
1115                 return -EPERM;
1116
1117         /* No locking - fasync_helper does its own locking */
1118         return fasync_helper(fd, file, on, &fc->fasync);
1119 }
1120
1121 const struct file_operations fuse_dev_operations = {
1122         .owner          = THIS_MODULE,
1123         .llseek         = no_llseek,
1124         .read           = do_sync_read,
1125         .aio_read       = fuse_dev_read,
1126         .write          = do_sync_write,
1127         .aio_write      = fuse_dev_write,
1128         .poll           = fuse_dev_poll,
1129         .release        = fuse_dev_release,
1130         .fasync         = fuse_dev_fasync,
1131 };
1132
1133 static struct miscdevice fuse_miscdevice = {
1134         .minor = FUSE_MINOR,
1135         .name  = "fuse",
1136         .fops = &fuse_dev_operations,
1137 };
1138
1139 int __init fuse_dev_init(void)
1140 {
1141         int err = -ENOMEM;
1142         fuse_req_cachep = kmem_cache_create("fuse_request",
1143                                             sizeof(struct fuse_req),
1144                                             0, 0, NULL);
1145         if (!fuse_req_cachep)
1146                 goto out;
1147
1148         err = misc_register(&fuse_miscdevice);
1149         if (err)
1150                 goto out_cache_clean;
1151
1152         return 0;
1153
1154  out_cache_clean:
1155         kmem_cache_destroy(fuse_req_cachep);
1156  out:
1157         return err;
1158 }
1159
1160 void fuse_dev_cleanup(void)
1161 {
1162         misc_deregister(&fuse_miscdevice);
1163         kmem_cache_destroy(fuse_req_cachep);
1164 }