fs/fuse/file.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/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/compat.h>
  17
  18 static const struct file_operations fuse_direct_io_file_operations;
  19
  20 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  21                           int opcode, struct fuse_open_out *outargp)
  22 {
  23         struct fuse_open_in inarg;
  24         struct fuse_req *req;
  25         int err;
  26
  27         req = fuse_get_req(fc);
  28         if (IS_ERR(req))
  29                 return PTR_ERR(req);
  30
  31         memset(&inarg, 0, sizeof(inarg));
  32         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  33         if (!fc->atomic_o_trunc)
  34                 inarg.flags &= ~O_TRUNC;
  35         req->in.h.opcode = opcode;
  36         req->in.h.nodeid = nodeid;
  37         req->in.numargs = 1;
  38         req->in.args[0].size = sizeof(inarg);
  39         req->in.args[0].value = &inarg;
  40         req->out.numargs = 1;
  41         req->out.args[0].size = sizeof(*outargp);
  42         req->out.args[0].value = outargp;
  43         fuse_request_send(fc, req);
  44         err = req->out.h.error;
  45         fuse_put_request(fc, req);
  46
  47         return err;
  48 }
  49
  50 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  51 {
  52         struct fuse_file *ff;
  53
  54         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  55         if (unlikely(!ff))
  56                 return NULL;
  57
  58         ff->fc = fc;
  59         ff->reserved_req = fuse_request_alloc();
  60         if (unlikely(!ff->reserved_req)) {
  61                 kfree(ff);
  62                 return NULL;
  63         }
  64
  65         INIT_LIST_HEAD(&ff->write_entry);
  66         atomic_set(&ff->count, 0);
  67         RB_CLEAR_NODE(&ff->polled_node);
  68         init_waitqueue_head(&ff->poll_wait);
  69
  70         spin_lock(&fc->lock);
  71         ff->kh = ++fc->khctr;
  72         spin_unlock(&fc->lock);
  73
  74         return ff;
  75 }
  76
  77 void fuse_file_free(struct fuse_file *ff)
  78 {
  79         fuse_request_free(ff->reserved_req);
  80         kfree(ff);
  81 }
  82
  83 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  84 {
  85         atomic_inc(&ff->count);
  86         return ff;
  87 }
  88
  89 static void fuse_release_async(struct work_struct *work)
  90 {
  91         struct fuse_req *req;
  92         struct fuse_conn *fc;
  93         struct path path;
  94
  95         req = container_of(work, struct fuse_req, misc.release.work);
  96         path = req->misc.release.path;
  97         fc = get_fuse_conn(path.dentry->d_inode);
  98
  99         fuse_put_request(fc, req);
 100         path_put(&path);
 101 }
 102
 103 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
 104 {
 105         if (fc->destroy_req) {
 106                 /*
 107                  * If this is a fuseblk mount, then it's possible that
 108                  * releasing the path will result in releasing the
 109                  * super block and sending the DESTROY request.  If
 110                  * the server is single threaded, this would hang.
 111                  * For this reason do the path_put() in a separate
 112                  * thread.
 113                  */
 114                 atomic_inc(&req->count);
 115                 INIT_WORK(&req->misc.release.work, fuse_release_async);
 116                 schedule_work(&req->misc.release.work);
 117         } else {
 118                 path_put(&req->misc.release.path);
 119         }
 120 }
 121
 122 static void fuse_file_put(struct fuse_file *ff, bool sync)
 123 {
 124         if (atomic_dec_and_test(&ff->count)) {
 125                 struct fuse_req *req = ff->reserved_req;
 126
 127                 if (sync) {
 128                         fuse_request_send(ff->fc, req);
 129                         path_put(&req->misc.release.path);
 130                         fuse_put_request(ff->fc, req);
 131                 } else {
 132                         req->end = fuse_release_end;
 133                         fuse_request_send_background(ff->fc, req);
 134                 }
 135                 kfree(ff);
 136         }
 137 }
 138
 139 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 140                  bool isdir)
 141 {
 142         struct fuse_open_out outarg;
 143         struct fuse_file *ff;
 144         int err;
 145         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 146
 147         ff = fuse_file_alloc(fc);
 148         if (!ff)
 149                 return -ENOMEM;
 150
 151         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 152         if (err) {
 153                 fuse_file_free(ff);
 154                 return err;
 155         }
 156
 157         if (isdir)
 158                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 159
 160         ff->fh = outarg.fh;
 161         ff->nodeid = nodeid;
 162         ff->open_flags = outarg.open_flags;
 163         file->private_data = fuse_file_get(ff);
 164
 165         return 0;
 166 }
 167 EXPORT_SYMBOL_GPL(fuse_do_open);
 168
 169 void fuse_finish_open(struct inode *inode, struct file *file)
 170 {
 171         struct fuse_file *ff = file->private_data;
 172         struct fuse_conn *fc = get_fuse_conn(inode);
 173
 174         if (ff->open_flags & FOPEN_DIRECT_IO)
 175                 file->f_op = &fuse_direct_io_file_operations;
 176         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 177                 invalidate_inode_pages2(inode->i_mapping);
 178         if (ff->open_flags & FOPEN_NONSEEKABLE)
 179                 nonseekable_open(inode, file);
 180         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 181                 struct fuse_inode *fi = get_fuse_inode(inode);
 182
 183                 spin_lock(&fc->lock);
 184                 fi->attr_version = ++fc->attr_version;
 185                 i_size_write(inode, 0);
 186                 spin_unlock(&fc->lock);
 187                 fuse_invalidate_attr(inode);
 188         }
 189 }
 190
 191 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 192 {
 193         struct fuse_conn *fc = get_fuse_conn(inode);
 194         int err;
 195
 196         /* VFS checks this, but only _after_ ->open() */
 197         if (file->f_flags & O_DIRECT)
 198                 return -EINVAL;
 199
 200         err = generic_file_open(inode, file);
 201         if (err)
 202                 return err;
 203
 204         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 205         if (err)
 206                 return err;
 207
 208         fuse_finish_open(inode, file);
 209
 210         return 0;
 211 }
 212
 213 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 214 {
 215         struct fuse_conn *fc = ff->fc;
 216         struct fuse_req *req = ff->reserved_req;
 217         struct fuse_release_in *inarg = &req->misc.release.in;
 218
 219         spin_lock(&fc->lock);
 220         list_del(&ff->write_entry);
 221         if (!RB_EMPTY_NODE(&ff->polled_node))
 222                 rb_erase(&ff->polled_node, &fc->polled_files);
 223         spin_unlock(&fc->lock);
 224
 225         wake_up_interruptible_all(&ff->poll_wait);
 226
 227         inarg->fh = ff->fh;
 228         inarg->flags = flags;
 229         req->in.h.opcode = opcode;
 230         req->in.h.nodeid = ff->nodeid;
 231         req->in.numargs = 1;
 232         req->in.args[0].size = sizeof(struct fuse_release_in);
 233         req->in.args[0].value = inarg;
 234 }
 235
 236 void fuse_release_common(struct file *file, int opcode)
 237 {
 238         struct fuse_file *ff;
 239         struct fuse_req *req;
 240
 241         ff = file->private_data;
 242         if (unlikely(!ff))
 243                 return;
 244
 245         req = ff->reserved_req;
 246         fuse_prepare_release(ff, file->f_flags, opcode);
 247
 248         if (ff->flock) {
 249                 struct fuse_release_in *inarg = &req->misc.release.in;
 250                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 251                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 252                                                        (fl_owner_t) file);
 253         }
 254         /* Hold vfsmount and dentry until release is finished */
 255         path_get(&file->f_path);
 256         req->misc.release.path = file->f_path;
 257
 258         /*
 259          * Normally this will send the RELEASE request, however if
 260          * some asynchronous READ or WRITE requests are outstanding,
 261          * the sending will be delayed.
 262          *
 263          * Make the release synchronous if this is a fuseblk mount,
 264          * synchronous RELEASE is allowed (and desirable) in this case
 265          * because the server can be trusted not to screw up.
 266          */
 267         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 268 }
 269
 270 static int fuse_open(struct inode *inode, struct file *file)
 271 {
 272         return fuse_open_common(inode, file, false);
 273 }
 274
 275 static int fuse_release(struct inode *inode, struct file *file)
 276 {
 277         fuse_release_common(file, FUSE_RELEASE);
 278
 279         /* return value is ignored by VFS */
 280         return 0;
 281 }
 282
 283 void fuse_sync_release(struct fuse_file *ff, int flags)
 284 {
 285         WARN_ON(atomic_read(&ff->count) > 1);
 286         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 287         ff->reserved_req->force = 1;
 288         fuse_request_send(ff->fc, ff->reserved_req);
 289         fuse_put_request(ff->fc, ff->reserved_req);
 290         kfree(ff);
 291 }
 292 EXPORT_SYMBOL_GPL(fuse_sync_release);
 293
 294 /*
 295  * Scramble the ID space with XTEA, so that the value of the files_struct
 296  * pointer is not exposed to userspace.
 297  */
 298 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 299 {
 300         u32 *k = fc->scramble_key;
 301         u64 v = (unsigned long) id;
 302         u32 v0 = v;
 303         u32 v1 = v >> 32;
 304         u32 sum = 0;
 305         int i;
 306
 307         for (i = 0; i < 32; i++) {
 308                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 309                 sum += 0x9E3779B9;
 310                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 311         }
 312
 313         return (u64) v0 + ((u64) v1 << 32);
 314 }
 315
 316 /*
 317  * Check if page is under writeback
 318  *
 319  * This is currently done by walking the list of writepage requests
 320  * for the inode, which can be pretty inefficient.
 321  */
 322 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 323 {
 324         struct fuse_conn *fc = get_fuse_conn(inode);
 325         struct fuse_inode *fi = get_fuse_inode(inode);
 326         struct fuse_req *req;
 327         bool found = false;
 328
 329         spin_lock(&fc->lock);
 330         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 331                 pgoff_t curr_index;
 332
 333                 BUG_ON(req->inode != inode);
 334                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 335                 if (curr_index == index) {
 336                         found = true;
 337                         break;
 338                 }
 339         }
 340         spin_unlock(&fc->lock);
 341
 342         return found;
 343 }
 344
 345 /*
 346  * Wait for page writeback to be completed.
 347  *
 348  * Since fuse doesn't rely on the VM writeback tracking, this has to
 349  * use some other means.
 350  */
 351 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 352 {
 353         struct fuse_inode *fi = get_fuse_inode(inode);
 354
 355         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 356         return 0;
 357 }
 358
 359 static int fuse_flush(struct file *file, fl_owner_t id)
 360 {
 361         struct inode *inode = file->f_path.dentry->d_inode;
 362         struct fuse_conn *fc = get_fuse_conn(inode);
 363         struct fuse_file *ff = file->private_data;
 364         struct fuse_req *req;
 365         struct fuse_flush_in inarg;
 366         int err;
 367
 368         if (is_bad_inode(inode))
 369                 return -EIO;
 370
 371         if (fc->no_flush)
 372                 return 0;
 373
 374         req = fuse_get_req_nofail(fc, file);
 375         memset(&inarg, 0, sizeof(inarg));
 376         inarg.fh = ff->fh;
 377         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 378         req->in.h.opcode = FUSE_FLUSH;
 379         req->in.h.nodeid = get_node_id(inode);
 380         req->in.numargs = 1;
 381         req->in.args[0].size = sizeof(inarg);
 382         req->in.args[0].value = &inarg;
 383         req->force = 1;
 384         fuse_request_send(fc, req);
 385         err = req->out.h.error;
 386         fuse_put_request(fc, req);
 387         if (err == -ENOSYS) {
 388                 fc->no_flush = 1;
 389                 err = 0;
 390         }
 391         return err;
 392 }
 393
 394 /*
 395  * Wait for all pending writepages on the inode to finish.
 396  *
 397  * This is currently done by blocking further writes with FUSE_NOWRITE
 398  * and waiting for all sent writes to complete.
 399  *
 400  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 401  * could conflict with truncation.
 402  */
 403 static void fuse_sync_writes(struct inode *inode)
 404 {
 405         fuse_set_nowrite(inode);
 406         fuse_release_nowrite(inode);
 407 }
 408
 409 int fuse_fsync_common(struct file *file, int datasync, int isdir)
 410 {
 411         struct inode *inode = file->f_mapping->host;
 412         struct fuse_conn *fc = get_fuse_conn(inode);
 413         struct fuse_file *ff = file->private_data;
 414         struct fuse_req *req;
 415         struct fuse_fsync_in inarg;
 416         int err;
 417
 418         if (is_bad_inode(inode))
 419                 return -EIO;
 420
 421         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 422                 return 0;
 423
 424         /*
 425          * Start writeback against all dirty pages of the inode, then
 426          * wait for all outstanding writes, before sending the FSYNC
 427          * request.
 428          */
 429         err = write_inode_now(inode, 0);
 430         if (err)
 431                 return err;
 432
 433         fuse_sync_writes(inode);
 434
 435         req = fuse_get_req(fc);
 436         if (IS_ERR(req))
 437                 return PTR_ERR(req);
 438
 439         memset(&inarg, 0, sizeof(inarg));
 440         inarg.fh = ff->fh;
 441         inarg.fsync_flags = datasync ? 1 : 0;
 442         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 443         req->in.h.nodeid = get_node_id(inode);
 444         req->in.numargs = 1;
 445         req->in.args[0].size = sizeof(inarg);
 446         req->in.args[0].value = &inarg;
 447         fuse_request_send(fc, req);
 448         err = req->out.h.error;
 449         fuse_put_request(fc, req);
 450         if (err == -ENOSYS) {
 451                 if (isdir)
 452                         fc->no_fsyncdir = 1;
 453                 else
 454                         fc->no_fsync = 1;
 455                 err = 0;
 456         }
 457         return err;
 458 }
 459
 460 static int fuse_fsync(struct file *file, int datasync)
 461 {
 462         return fuse_fsync_common(file, datasync, 0);
 463 }
 464
 465 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 466                     size_t count, int opcode)
 467 {
 468         struct fuse_read_in *inarg = &req->misc.read.in;
 469         struct fuse_file *ff = file->private_data;
 470
 471         inarg->fh = ff->fh;
 472         inarg->offset = pos;
 473         inarg->size = count;
 474         inarg->flags = file->f_flags;
 475         req->in.h.opcode = opcode;
 476         req->in.h.nodeid = ff->nodeid;
 477         req->in.numargs = 1;
 478         req->in.args[0].size = sizeof(struct fuse_read_in);
 479         req->in.args[0].value = inarg;
 480         req->out.argvar = 1;
 481         req->out.numargs = 1;
 482         req->out.args[0].size = count;
 483 }
 484
 485 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 486                              loff_t pos, size_t count, fl_owner_t owner)
 487 {
 488         struct fuse_file *ff = file->private_data;
 489         struct fuse_conn *fc = ff->fc;
 490
 491         fuse_read_fill(req, file, pos, count, FUSE_READ);
 492         if (owner != NULL) {
 493                 struct fuse_read_in *inarg = &req->misc.read.in;
 494
 495                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 496                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 497         }
 498         fuse_request_send(fc, req);
 499         return req->out.args[0].size;
 500 }
 501
 502 static void fuse_read_update_size(struct inode *inode, loff_t size,
 503                                   u64 attr_ver)
 504 {
 505         struct fuse_conn *fc = get_fuse_conn(inode);
 506         struct fuse_inode *fi = get_fuse_inode(inode);
 507
 508         spin_lock(&fc->lock);
 509         if (attr_ver == fi->attr_version && size < inode->i_size) {
 510                 fi->attr_version = ++fc->attr_version;
 511                 i_size_write(inode, size);
 512         }
 513         spin_unlock(&fc->lock);
 514 }
 515
 516 static int fuse_readpage(struct file *file, struct page *page)
 517 {
 518         struct inode *inode = page->mapping->host;
 519         struct fuse_conn *fc = get_fuse_conn(inode);
 520         struct fuse_req *req;
 521         size_t num_read;
 522         loff_t pos = page_offset(page);
 523         size_t count = PAGE_CACHE_SIZE;
 524         u64 attr_ver;
 525         int err;
 526
 527         err = -EIO;
 528         if (is_bad_inode(inode))
 529                 goto out;
 530
 531         /*
 532          * Page writeback can extend beyond the lifetime of the
 533          * page-cache page, so make sure we read a properly synced
 534          * page.
 535          */
 536         fuse_wait_on_page_writeback(inode, page->index);
 537
 538         req = fuse_get_req(fc);
 539         err = PTR_ERR(req);
 540         if (IS_ERR(req))
 541                 goto out;
 542
 543         attr_ver = fuse_get_attr_version(fc);
 544
 545         req->out.page_zeroing = 1;
 546         req->out.argpages = 1;
 547         req->num_pages = 1;
 548         req->pages[0] = page;
 549         num_read = fuse_send_read(req, file, pos, count, NULL);
 550         err = req->out.h.error;
 551         fuse_put_request(fc, req);
 552
 553         if (!err) {
 554                 /*
 555                  * Short read means EOF.  If file size is larger, truncate it
 556                  */
 557                 if (num_read < count)
 558                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 559
 560                 SetPageUptodate(page);
 561         }
 562
 563         fuse_invalidate_attr(inode); /* atime changed */
 564  out:
 565         unlock_page(page);
 566         return err;
 567 }
 568
 569 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 570 {
 571         int i;
 572         size_t count = req->misc.read.in.size;
 573         size_t num_read = req->out.args[0].size;
 574         struct address_space *mapping = NULL;
 575
 576         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 577                 mapping = req->pages[i]->mapping;
 578
 579         if (mapping) {
 580                 struct inode *inode = mapping->host;
 581
 582                 /*
 583                  * Short read means EOF. If file size is larger, truncate it
 584                  */
 585                 if (!req->out.h.error && num_read < count) {
 586                         loff_t pos;
 587
 588                         pos = page_offset(req->pages[0]) + num_read;
 589                         fuse_read_update_size(inode, pos,
 590                                               req->misc.read.attr_ver);
 591                 }
 592                 fuse_invalidate_attr(inode); /* atime changed */
 593         }
 594
 595         for (i = 0; i < req->num_pages; i++) {
 596                 struct page *page = req->pages[i];
 597                 if (!req->out.h.error)
 598                         SetPageUptodate(page);
 599                 else
 600                         SetPageError(page);
 601                 unlock_page(page);
 602                 page_cache_release(page);
 603         }
 604         if (req->ff)
 605                 fuse_file_put(req->ff, false);
 606 }
 607
 608 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 609 {
 610         struct fuse_file *ff = file->private_data;
 611         struct fuse_conn *fc = ff->fc;
 612         loff_t pos = page_offset(req->pages[0]);
 613         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 614
 615         req->out.argpages = 1;
 616         req->out.page_zeroing = 1;
 617         req->out.page_replace = 1;
 618         fuse_read_fill(req, file, pos, count, FUSE_READ);
 619         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 620         if (fc->async_read) {
 621                 req->ff = fuse_file_get(ff);
 622                 req->end = fuse_readpages_end;
 623                 fuse_request_send_background(fc, req);
 624         } else {
 625                 fuse_request_send(fc, req);
 626                 fuse_readpages_end(fc, req);
 627                 fuse_put_request(fc, req);
 628         }
 629 }
 630
 631 struct fuse_fill_data {
 632         struct fuse_req *req;
 633         struct file *file;
 634         struct inode *inode;
 635 };
 636
 637 static int fuse_readpages_fill(void *_data, struct page *page)
 638 {
 639         struct fuse_fill_data *data = _data;
 640         struct fuse_req *req = data->req;
 641         struct inode *inode = data->inode;
 642         struct fuse_conn *fc = get_fuse_conn(inode);
 643
 644         fuse_wait_on_page_writeback(inode, page->index);
 645
 646         if (req->num_pages &&
 647             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 648              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 649              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 650                 fuse_send_readpages(req, data->file);
 651                 data->req = req = fuse_get_req(fc);
 652                 if (IS_ERR(req)) {
 653                         unlock_page(page);
 654                         return PTR_ERR(req);
 655                 }
 656         }
 657         page_cache_get(page);
 658         req->pages[req->num_pages] = page;
 659         req->num_pages++;
 660         return 0;
 661 }
 662
 663 static int fuse_readpages(struct file *file, struct address_space *mapping,
 664                           struct list_head *pages, unsigned nr_pages)
 665 {
 666         struct inode *inode = mapping->host;
 667         struct fuse_conn *fc = get_fuse_conn(inode);
 668         struct fuse_fill_data data;
 669         int err;
 670
 671         err = -EIO;
 672         if (is_bad_inode(inode))
 673                 goto out;
 674
 675         data.file = file;
 676         data.inode = inode;
 677         data.req = fuse_get_req(fc);
 678         err = PTR_ERR(data.req);
 679         if (IS_ERR(data.req))
 680                 goto out;
 681
 682         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 683         if (!err) {
 684                 if (data.req->num_pages)
 685                         fuse_send_readpages(data.req, file);
 686                 else
 687                         fuse_put_request(fc, data.req);
 688         }
 689 out:
 690         return err;
 691 }
 692
 693 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 694                                   unsigned long nr_segs, loff_t pos)
 695 {
 696         struct inode *inode = iocb->ki_filp->f_mapping->host;
 697
 698         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
 699                 int err;
 700                 /*
 701                  * If trying to read past EOF, make sure the i_size
 702                  * attribute is up-to-date.
 703                  */
 704                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 705                 if (err)
 706                         return err;
 707         }
 708
 709         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 710 }
 711
 712 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 713                             loff_t pos, size_t count)
 714 {
 715         struct fuse_write_in *inarg = &req->misc.write.in;
 716         struct fuse_write_out *outarg = &req->misc.write.out;
 717
 718         inarg->fh = ff->fh;
 719         inarg->offset = pos;
 720         inarg->size = count;
 721         req->in.h.opcode = FUSE_WRITE;
 722         req->in.h.nodeid = ff->nodeid;
 723         req->in.numargs = 2;
 724         if (ff->fc->minor < 9)
 725                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 726         else
 727                 req->in.args[0].size = sizeof(struct fuse_write_in);
 728         req->in.args[0].value = inarg;
 729         req->in.args[1].size = count;
 730         req->out.numargs = 1;
 731         req->out.args[0].size = sizeof(struct fuse_write_out);
 732         req->out.args[0].value = outarg;
 733 }
 734
 735 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 736                               loff_t pos, size_t count, fl_owner_t owner)
 737 {
 738         struct fuse_file *ff = file->private_data;
 739         struct fuse_conn *fc = ff->fc;
 740         struct fuse_write_in *inarg = &req->misc.write.in;
 741
 742         fuse_write_fill(req, ff, pos, count);
 743         inarg->flags = file->f_flags;
 744         if (owner != NULL) {
 745                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 746                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 747         }
 748         fuse_request_send(fc, req);
 749         return req->misc.write.out.size;
 750 }
 751
 752 void fuse_write_update_size(struct inode *inode, loff_t pos)
 753 {
 754         struct fuse_conn *fc = get_fuse_conn(inode);
 755         struct fuse_inode *fi = get_fuse_inode(inode);
 756
 757         spin_lock(&fc->lock);
 758         fi->attr_version = ++fc->attr_version;
 759         if (pos > inode->i_size)
 760                 i_size_write(inode, pos);
 761         spin_unlock(&fc->lock);
 762 }
 763
 764 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 765                                     struct inode *inode, loff_t pos,
 766                                     size_t count)
 767 {
 768         size_t res;
 769         unsigned offset;
 770         unsigned i;
 771
 772         for (i = 0; i < req->num_pages; i++)
 773                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 774
 775         res = fuse_send_write(req, file, pos, count, NULL);
 776
 777         offset = req->page_offset;
 778         count = res;
 779         for (i = 0; i < req->num_pages; i++) {
 780                 struct page *page = req->pages[i];
 781
 782                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 783                         SetPageUptodate(page);
 784
 785                 if (count > PAGE_CACHE_SIZE - offset)
 786                         count -= PAGE_CACHE_SIZE - offset;
 787                 else
 788                         count = 0;
 789                 offset = 0;
 790
 791                 unlock_page(page);
 792                 page_cache_release(page);
 793         }
 794
 795         return res;
 796 }
 797
 798 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 799                                struct address_space *mapping,
 800                                struct iov_iter *ii, loff_t pos)
 801 {
 802         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 803         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 804         size_t count = 0;
 805         int err;
 806
 807         req->in.argpages = 1;
 808         req->page_offset = offset;
 809
 810         do {
 811                 size_t tmp;
 812                 struct page *page;
 813                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 814                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 815                                      iov_iter_count(ii));
 816
 817                 bytes = min_t(size_t, bytes, fc->max_write - count);
 818
 819  again:
 820                 err = -EFAULT;
 821                 if (iov_iter_fault_in_readable(ii, bytes))
 822                         break;
 823
 824                 err = -ENOMEM;
 825                 page = grab_cache_page_write_begin(mapping, index, 0);
 826                 if (!page)
 827                         break;
 828
 829                 if (mapping_writably_mapped(mapping))
 830                         flush_dcache_page(page);
 831
 832                 pagefault_disable();
 833                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 834                 pagefault_enable();
 835                 flush_dcache_page(page);
 836
 837                 if (!tmp) {
 838                         unlock_page(page);
 839                         page_cache_release(page);
 840                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 841                         goto again;
 842                 }
 843
 844                 err = 0;
 845                 req->pages[req->num_pages] = page;
 846                 req->num_pages++;
 847
 848                 iov_iter_advance(ii, tmp);
 849                 count += tmp;
 850                 pos += tmp;
 851                 offset += tmp;
 852                 if (offset == PAGE_CACHE_SIZE)
 853                         offset = 0;
 854
 855                 if (!fc->big_writes)
 856                         break;
 857         } while (iov_iter_count(ii) && count < fc->max_write &&
 858                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
 859
 860         return count > 0 ? count : err;
 861 }
 862
 863 static ssize_t fuse_perform_write(struct file *file,
 864                                   struct address_space *mapping,
 865                                   struct iov_iter *ii, loff_t pos)
 866 {
 867         struct inode *inode = mapping->host;
 868         struct fuse_conn *fc = get_fuse_conn(inode);
 869         int err = 0;
 870         ssize_t res = 0;
 871
 872         if (is_bad_inode(inode))
 873                 return -EIO;
 874
 875         do {
 876                 struct fuse_req *req;
 877                 ssize_t count;
 878
 879                 req = fuse_get_req(fc);
 880                 if (IS_ERR(req)) {
 881                         err = PTR_ERR(req);
 882                         break;
 883                 }
 884
 885                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 886                 if (count <= 0) {
 887                         err = count;
 888                 } else {
 889                         size_t num_written;
 890
 891                         num_written = fuse_send_write_pages(req, file, inode,
 892                                                             pos, count);
 893                         err = req->out.h.error;
 894                         if (!err) {
 895                                 res += num_written;
 896                                 pos += num_written;
 897
 898                                 /* break out of the loop on short write */
 899                                 if (num_written != count)
 900                                         err = -EIO;
 901                         }
 902                 }
 903                 fuse_put_request(fc, req);
 904         } while (!err && iov_iter_count(ii));
 905
 906         if (res > 0)
 907                 fuse_write_update_size(inode, pos);
 908
 909         fuse_invalidate_attr(inode);
 910
 911         return res > 0 ? res : err;
 912 }
 913
 914 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 915                                    unsigned long nr_segs, loff_t pos)
 916 {
 917         struct file *file = iocb->ki_filp;
 918         struct address_space *mapping = file->f_mapping;
 919         size_t count = 0;
 920         ssize_t written = 0;
 921         struct inode *inode = mapping->host;
 922         ssize_t err;
 923         struct iov_iter i;
 924
 925         WARN_ON(iocb->ki_pos != pos);
 926
 927         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
 928         if (err)
 929                 return err;
 930
 931         mutex_lock(&inode->i_mutex);
 932         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
 933
 934         /* We can write back this queue in page reclaim */
 935         current->backing_dev_info = mapping->backing_dev_info;
 936
 937         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 938         if (err)
 939                 goto out;
 940
 941         if (count == 0)
 942                 goto out;
 943
 944         err = file_remove_suid(file);
 945         if (err)
 946                 goto out;
 947
 948         file_update_time(file);
 949
 950         iov_iter_init(&i, iov, nr_segs, count, 0);
 951         written = fuse_perform_write(file, mapping, &i, pos);
 952         if (written >= 0)
 953                 iocb->ki_pos = pos + written;
 954
 955 out:
 956         current->backing_dev_info = NULL;
 957         mutex_unlock(&inode->i_mutex);
 958
 959         return written ? written : err;
 960 }
 961
 962 static void fuse_release_user_pages(struct fuse_req *req, int write)
 963 {
 964         unsigned i;
 965
 966         for (i = 0; i < req->num_pages; i++) {
 967                 struct page *page = req->pages[i];
 968                 if (write)
 969                         set_page_dirty_lock(page);
 970                 put_page(page);
 971         }
 972 }
 973
 974 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
 975                                size_t *nbytesp, int write)
 976 {
 977         size_t nbytes = *nbytesp;
 978         unsigned long user_addr = (unsigned long) buf;
 979         unsigned offset = user_addr & ~PAGE_MASK;
 980         int npages;
 981
 982         /* Special case for kernel I/O: can copy directly into the buffer */
 983         if (segment_eq(get_fs(), KERNEL_DS)) {
 984                 if (write)
 985                         req->in.args[1].value = (void *) user_addr;
 986                 else
 987                         req->out.args[0].value = (void *) user_addr;
 988
 989                 return 0;
 990         }
 991
 992         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
 993         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
 994         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
 995         npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
 996         if (npages < 0)
 997                 return npages;
 998
 999         req->num_pages = npages;
1000         req->page_offset = offset;
1001
1002         if (write)
1003                 req->in.argpages = 1;
1004         else
1005                 req->out.argpages = 1;
1006
1007         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1008         *nbytesp = min(*nbytesp, nbytes);
1009
1010         return 0;
1011 }
1012
1013 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1014                        size_t count, loff_t *ppos, int write)
1015 {
1016         struct fuse_file *ff = file->private_data;
1017         struct fuse_conn *fc = ff->fc;
1018         size_t nmax = write ? fc->max_write : fc->max_read;
1019         loff_t pos = *ppos;
1020         ssize_t res = 0;
1021         struct fuse_req *req;
1022
1023         req = fuse_get_req(fc);
1024         if (IS_ERR(req))
1025                 return PTR_ERR(req);
1026
1027         while (count) {
1028                 size_t nres;
1029                 fl_owner_t owner = current->files;
1030                 size_t nbytes = min(count, nmax);
1031                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1032                 if (err) {
1033                         res = err;
1034                         break;
1035                 }
1036
1037                 if (write)
1038                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1039                 else
1040                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1041
1042                 fuse_release_user_pages(req, !write);
1043                 if (req->out.h.error) {
1044                         if (!res)
1045                                 res = req->out.h.error;
1046                         break;
1047                 } else if (nres > nbytes) {
1048                         res = -EIO;
1049                         break;
1050                 }
1051                 count -= nres;
1052                 res += nres;
1053                 pos += nres;
1054                 buf += nres;
1055                 if (nres != nbytes)
1056                         break;
1057                 if (count) {
1058                         fuse_put_request(fc, req);
1059                         req = fuse_get_req(fc);
1060                         if (IS_ERR(req))
1061                                 break;
1062                 }
1063         }
1064         if (!IS_ERR(req))
1065                 fuse_put_request(fc, req);
1066         if (res > 0)
1067                 *ppos = pos;
1068
1069         return res;
1070 }
1071 EXPORT_SYMBOL_GPL(fuse_direct_io);
1072
1073 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1074                                      size_t count, loff_t *ppos)
1075 {
1076         ssize_t res;
1077         struct inode *inode = file->f_path.dentry->d_inode;
1078
1079         if (is_bad_inode(inode))
1080                 return -EIO;
1081
1082         res = fuse_direct_io(file, buf, count, ppos, 0);
1083
1084         fuse_invalidate_attr(inode);
1085
1086         return res;
1087 }
1088
1089 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1090                                  size_t count, loff_t *ppos)
1091 {
1092         struct inode *inode = file->f_path.dentry->d_inode;
1093         ssize_t res;
1094
1095         if (is_bad_inode(inode))
1096                 return -EIO;
1097
1098         /* Don't allow parallel writes to the same file */
1099         mutex_lock(&inode->i_mutex);
1100         res = generic_write_checks(file, ppos, &count, 0);
1101         if (!res) {
1102                 res = fuse_direct_io(file, buf, count, ppos, 1);
1103                 if (res > 0)
1104                         fuse_write_update_size(inode, *ppos);
1105         }
1106         mutex_unlock(&inode->i_mutex);
1107
1108         fuse_invalidate_attr(inode);
1109
1110         return res;
1111 }
1112
1113 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1114 {
1115         __free_page(req->pages[0]);
1116         fuse_file_put(req->ff, false);
1117 }
1118
1119 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1120 {
1121         struct inode *inode = req->inode;
1122         struct fuse_inode *fi = get_fuse_inode(inode);
1123         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1124
1125         list_del(&req->writepages_entry);
1126         dec_bdi_stat(bdi, BDI_WRITEBACK);
1127         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1128         bdi_writeout_inc(bdi);
1129         wake_up(&fi->page_waitq);
1130 }
1131
1132 /* Called under fc->lock, may release and reacquire it */
1133 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1134 __releases(fc->lock)
1135 __acquires(fc->lock)
1136 {
1137         struct fuse_inode *fi = get_fuse_inode(req->inode);
1138         loff_t size = i_size_read(req->inode);
1139         struct fuse_write_in *inarg = &req->misc.write.in;
1140
1141         if (!fc->connected)
1142                 goto out_free;
1143
1144         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1145                 inarg->size = PAGE_CACHE_SIZE;
1146         } else if (inarg->offset < size) {
1147                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1148         } else {
1149                 /* Got truncated off completely */
1150                 goto out_free;
1151         }
1152
1153         req->in.args[1].size = inarg->size;
1154         fi->writectr++;
1155         fuse_request_send_background_locked(fc, req);
1156         return;
1157
1158  out_free:
1159         fuse_writepage_finish(fc, req);
1160         spin_unlock(&fc->lock);
1161         fuse_writepage_free(fc, req);
1162         fuse_put_request(fc, req);
1163         spin_lock(&fc->lock);
1164 }
1165
1166 /*
1167  * If fi->writectr is positive (no truncate or fsync going on) send
1168  * all queued writepage requests.
1169  *
1170  * Called with fc->lock
1171  */
1172 void fuse_flush_writepages(struct inode *inode)
1173 __releases(fc->lock)
1174 __acquires(fc->lock)
1175 {
1176         struct fuse_conn *fc = get_fuse_conn(inode);
1177         struct fuse_inode *fi = get_fuse_inode(inode);
1178         struct fuse_req *req;
1179
1180         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1181                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1182                 list_del_init(&req->list);
1183                 fuse_send_writepage(fc, req);
1184         }
1185 }
1186
1187 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1188 {
1189         struct inode *inode = req->inode;
1190         struct fuse_inode *fi = get_fuse_inode(inode);
1191
1192         mapping_set_error(inode->i_mapping, req->out.h.error);
1193         spin_lock(&fc->lock);
1194         fi->writectr--;
1195         fuse_writepage_finish(fc, req);
1196         spin_unlock(&fc->lock);
1197         fuse_writepage_free(fc, req);
1198 }
1199
1200 static int fuse_writepage_locked(struct page *page)
1201 {
1202         struct address_space *mapping = page->mapping;
1203         struct inode *inode = mapping->host;
1204         struct fuse_conn *fc = get_fuse_conn(inode);
1205         struct fuse_inode *fi = get_fuse_inode(inode);
1206         struct fuse_req *req;
1207         struct fuse_file *ff;
1208         struct page *tmp_page;
1209
1210         set_page_writeback(page);
1211
1212         req = fuse_request_alloc_nofs();
1213         if (!req)
1214                 goto err;
1215
1216         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1217         if (!tmp_page)
1218                 goto err_free;
1219
1220         spin_lock(&fc->lock);
1221         BUG_ON(list_empty(&fi->write_files));
1222         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1223         req->ff = fuse_file_get(ff);
1224         spin_unlock(&fc->lock);
1225
1226         fuse_write_fill(req, ff, page_offset(page), 0);
1227
1228         copy_highpage(tmp_page, page);
1229         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1230         req->in.argpages = 1;
1231         req->num_pages = 1;
1232         req->pages[0] = tmp_page;
1233         req->page_offset = 0;
1234         req->end = fuse_writepage_end;
1235         req->inode = inode;
1236
1237         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1238         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1239         end_page_writeback(page);
1240
1241         spin_lock(&fc->lock);
1242         list_add(&req->writepages_entry, &fi->writepages);
1243         list_add_tail(&req->list, &fi->queued_writes);
1244         fuse_flush_writepages(inode);
1245         spin_unlock(&fc->lock);
1246
1247         return 0;
1248
1249 err_free:
1250         fuse_request_free(req);
1251 err:
1252         end_page_writeback(page);
1253         return -ENOMEM;
1254 }
1255
1256 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1257 {
1258         int err;
1259
1260         err = fuse_writepage_locked(page);
1261         unlock_page(page);
1262
1263         return err;
1264 }
1265
1266 static int fuse_launder_page(struct page *page)
1267 {
1268         int err = 0;
1269         if (clear_page_dirty_for_io(page)) {
1270                 struct inode *inode = page->mapping->host;
1271                 err = fuse_writepage_locked(page);
1272                 if (!err)
1273                         fuse_wait_on_page_writeback(inode, page->index);
1274         }
1275         return err;
1276 }
1277
1278 /*
1279  * Write back dirty pages now, because there may not be any suitable
1280  * open files later
1281  */
1282 static void fuse_vma_close(struct vm_area_struct *vma)
1283 {
1284         filemap_write_and_wait(vma->vm_file->f_mapping);
1285 }
1286
1287 /*
1288  * Wait for writeback against this page to complete before allowing it
1289  * to be marked dirty again, and hence written back again, possibly
1290  * before the previous writepage completed.
1291  *
1292  * Block here, instead of in ->writepage(), so that the userspace fs
1293  * can only block processes actually operating on the filesystem.
1294  *
1295  * Otherwise unprivileged userspace fs would be able to block
1296  * unrelated:
1297  *
1298  * - page migration
1299  * - sync(2)
1300  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1301  */
1302 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1303 {
1304         struct page *page = vmf->page;
1305         /*
1306          * Don't use page->mapping as it may become NULL from a
1307          * concurrent truncate.
1308          */
1309         struct inode *inode = vma->vm_file->f_mapping->host;
1310
1311         fuse_wait_on_page_writeback(inode, page->index);
1312         return 0;
1313 }
1314
1315 static const struct vm_operations_struct fuse_file_vm_ops = {
1316         .close          = fuse_vma_close,
1317         .fault          = filemap_fault,
1318         .page_mkwrite   = fuse_page_mkwrite,
1319 };
1320
1321 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1322 {
1323         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1324                 struct inode *inode = file->f_dentry->d_inode;
1325                 struct fuse_conn *fc = get_fuse_conn(inode);
1326                 struct fuse_inode *fi = get_fuse_inode(inode);
1327                 struct fuse_file *ff = file->private_data;
1328                 /*
1329                  * file may be written through mmap, so chain it onto the
1330                  * inodes's write_file list
1331                  */
1332                 spin_lock(&fc->lock);
1333                 if (list_empty(&ff->write_entry))
1334                         list_add(&ff->write_entry, &fi->write_files);
1335                 spin_unlock(&fc->lock);
1336         }
1337         file_accessed(file);
1338         vma->vm_ops = &fuse_file_vm_ops;
1339         return 0;
1340 }
1341
1342 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1343 {
1344         /* Can't provide the coherency needed for MAP_SHARED */
1345         if (vma->vm_flags & VM_MAYSHARE)
1346                 return -ENODEV;
1347
1348         invalidate_inode_pages2(file->f_mapping);
1349
1350         return generic_file_mmap(file, vma);
1351 }
1352
1353 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1354                                   struct file_lock *fl)
1355 {
1356         switch (ffl->type) {
1357         case F_UNLCK:
1358                 break;
1359
1360         case F_RDLCK:
1361         case F_WRLCK:
1362                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1363                     ffl->end < ffl->start)
1364                         return -EIO;
1365
1366                 fl->fl_start = ffl->start;
1367                 fl->fl_end = ffl->end;
1368                 fl->fl_pid = ffl->pid;
1369                 break;
1370
1371         default:
1372                 return -EIO;
1373         }
1374         fl->fl_type = ffl->type;
1375         return 0;
1376 }
1377
1378 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1379                          const struct file_lock *fl, int opcode, pid_t pid,
1380                          int flock)
1381 {
1382         struct inode *inode = file->f_path.dentry->d_inode;
1383         struct fuse_conn *fc = get_fuse_conn(inode);
1384         struct fuse_file *ff = file->private_data;
1385         struct fuse_lk_in *arg = &req->misc.lk_in;
1386
1387         arg->fh = ff->fh;
1388         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1389         arg->lk.start = fl->fl_start;
1390         arg->lk.end = fl->fl_end;
1391         arg->lk.type = fl->fl_type;
1392         arg->lk.pid = pid;
1393         if (flock)
1394                 arg->lk_flags |= FUSE_LK_FLOCK;
1395         req->in.h.opcode = opcode;
1396         req->in.h.nodeid = get_node_id(inode);
1397         req->in.numargs = 1;
1398         req->in.args[0].size = sizeof(*arg);
1399         req->in.args[0].value = arg;
1400 }
1401
1402 static int fuse_getlk(struct file *file, struct file_lock *fl)
1403 {
1404         struct inode *inode = file->f_path.dentry->d_inode;
1405         struct fuse_conn *fc = get_fuse_conn(inode);
1406         struct fuse_req *req;
1407         struct fuse_lk_out outarg;
1408         int err;
1409
1410         req = fuse_get_req(fc);
1411         if (IS_ERR(req))
1412                 return PTR_ERR(req);
1413
1414         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1415         req->out.numargs = 1;
1416         req->out.args[0].size = sizeof(outarg);
1417         req->out.args[0].value = &outarg;
1418         fuse_request_send(fc, req);
1419         err = req->out.h.error;
1420         fuse_put_request(fc, req);
1421         if (!err)
1422                 err = convert_fuse_file_lock(&outarg.lk, fl);
1423
1424         return err;
1425 }
1426
1427 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1428 {
1429         struct inode *inode = file->f_path.dentry->d_inode;
1430         struct fuse_conn *fc = get_fuse_conn(inode);
1431         struct fuse_req *req;
1432         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1433         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1434         int err;
1435
1436         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1437                 /* NLM needs asynchronous locks, which we don't support yet */
1438                 return -ENOLCK;
1439         }
1440
1441         /* Unlock on close is handled by the flush method */
1442         if (fl->fl_flags & FL_CLOSE)
1443                 return 0;
1444
1445         req = fuse_get_req(fc);
1446         if (IS_ERR(req))
1447                 return PTR_ERR(req);
1448
1449         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1450         fuse_request_send(fc, req);
1451         err = req->out.h.error;
1452         /* locking is restartable */
1453         if (err == -EINTR)
1454                 err = -ERESTARTSYS;
1455         fuse_put_request(fc, req);
1456         return err;
1457 }
1458
1459 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1460 {
1461         struct inode *inode = file->f_path.dentry->d_inode;
1462         struct fuse_conn *fc = get_fuse_conn(inode);
1463         int err;
1464
1465         if (cmd == F_CANCELLK) {
1466                 err = 0;
1467         } else if (cmd == F_GETLK) {
1468                 if (fc->no_lock) {
1469                         posix_test_lock(file, fl);
1470                         err = 0;
1471                 } else
1472                         err = fuse_getlk(file, fl);
1473         } else {
1474                 if (fc->no_lock)
1475                         err = posix_lock_file(file, fl, NULL);
1476                 else
1477                         err = fuse_setlk(file, fl, 0);
1478         }
1479         return err;
1480 }
1481
1482 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1483 {
1484         struct inode *inode = file->f_path.dentry->d_inode;
1485         struct fuse_conn *fc = get_fuse_conn(inode);
1486         int err;
1487
1488         if (fc->no_flock) {
1489                 err = flock_lock_file_wait(file, fl);
1490         } else {
1491                 struct fuse_file *ff = file->private_data;
1492
1493                 /* emulate flock with POSIX locks */
1494                 fl->fl_owner = (fl_owner_t) file;
1495                 ff->flock = true;
1496                 err = fuse_setlk(file, fl, 1);
1497         }
1498
1499         return err;
1500 }
1501
1502 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1503 {
1504         struct inode *inode = mapping->host;
1505         struct fuse_conn *fc = get_fuse_conn(inode);
1506         struct fuse_req *req;
1507         struct fuse_bmap_in inarg;
1508         struct fuse_bmap_out outarg;
1509         int err;
1510
1511         if (!inode->i_sb->s_bdev || fc->no_bmap)
1512                 return 0;
1513
1514         req = fuse_get_req(fc);
1515         if (IS_ERR(req))
1516                 return 0;
1517
1518         memset(&inarg, 0, sizeof(inarg));
1519         inarg.block = block;
1520         inarg.blocksize = inode->i_sb->s_blocksize;
1521         req->in.h.opcode = FUSE_BMAP;
1522         req->in.h.nodeid = get_node_id(inode);
1523         req->in.numargs = 1;
1524         req->in.args[0].size = sizeof(inarg);
1525         req->in.args[0].value = &inarg;
1526         req->out.numargs = 1;
1527         req->out.args[0].size = sizeof(outarg);
1528         req->out.args[0].value = &outarg;
1529         fuse_request_send(fc, req);
1530         err = req->out.h.error;
1531         fuse_put_request(fc, req);
1532         if (err == -ENOSYS)
1533                 fc->no_bmap = 1;
1534
1535         return err ? 0 : outarg.block;
1536 }
1537
1538 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1539 {
1540         loff_t retval;
1541         struct inode *inode = file->f_path.dentry->d_inode;
1542
1543         mutex_lock(&inode->i_mutex);
1544         switch (origin) {
1545         case SEEK_END:
1546                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1547                 if (retval)
1548                         goto exit;
1549                 offset += i_size_read(inode);
1550                 break;
1551         case SEEK_CUR:
1552                 offset += file->f_pos;
1553         }
1554         retval = -EINVAL;
1555         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1556                 if (offset != file->f_pos) {
1557                         file->f_pos = offset;
1558                         file->f_version = 0;
1559                 }
1560                 retval = offset;
1561         }
1562 exit:
1563         mutex_unlock(&inode->i_mutex);
1564         return retval;
1565 }
1566
1567 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1568                         unsigned int nr_segs, size_t bytes, bool to_user)
1569 {
1570         struct iov_iter ii;
1571         int page_idx = 0;
1572
1573         if (!bytes)
1574                 return 0;
1575
1576         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1577
1578         while (iov_iter_count(&ii)) {
1579                 struct page *page = pages[page_idx++];
1580                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1581                 void *kaddr;
1582
1583                 kaddr = kmap(page);
1584
1585                 while (todo) {
1586                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1587                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1588                         size_t copy = min(todo, iov_len);
1589                         size_t left;
1590
1591                         if (!to_user)
1592                                 left = copy_from_user(kaddr, uaddr, copy);
1593                         else
1594                                 left = copy_to_user(uaddr, kaddr, copy);
1595
1596                         if (unlikely(left))
1597                                 return -EFAULT;
1598
1599                         iov_iter_advance(&ii, copy);
1600                         todo -= copy;
1601                         kaddr += copy;
1602                 }
1603
1604                 kunmap(page);
1605         }
1606
1607         return 0;
1608 }
1609
1610 /*
1611  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1612  * ABI was defined to be 'struct iovec' which is different on 32bit
1613  * and 64bit.  Fortunately we can determine which structure the server
1614  * used from the size of the reply.
1615  */
1616 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1617                                      size_t transferred, unsigned count,
1618                                      bool is_compat)
1619 {
1620 #ifdef CONFIG_COMPAT
1621         if (count * sizeof(struct compat_iovec) == transferred) {
1622                 struct compat_iovec *ciov = src;
1623                 unsigned i;
1624
1625                 /*
1626                  * With this interface a 32bit server cannot support
1627                  * non-compat (i.e. ones coming from 64bit apps) ioctl
1628                  * requests
1629                  */
1630                 if (!is_compat)
1631                         return -EINVAL;
1632
1633                 for (i = 0; i < count; i++) {
1634                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1635                         dst[i].iov_len = ciov[i].iov_len;
1636                 }
1637                 return 0;
1638         }
1639 #endif
1640
1641         if (count * sizeof(struct iovec) != transferred)
1642                 return -EIO;
1643
1644         memcpy(dst, src, transferred);
1645         return 0;
1646 }
1647
1648 /* Make sure iov_length() won't overflow */
1649 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1650 {
1651         size_t n;
1652         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1653
1654         for (n = 0; n < count; n++) {
1655                 if (iov->iov_len > (size_t) max)
1656                         return -ENOMEM;
1657                 max -= iov->iov_len;
1658         }
1659         return 0;
1660 }
1661
1662 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1663                                  void *src, size_t transferred, unsigned count,
1664                                  bool is_compat)
1665 {
1666         unsigned i;
1667         struct fuse_ioctl_iovec *fiov = src;
1668
1669         if (fc->minor < 16) {
1670                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1671                                                  count, is_compat);
1672         }
1673
1674         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1675                 return -EIO;
1676
1677         for (i = 0; i < count; i++) {
1678                 /* Did the server supply an inappropriate value? */
1679                 if (fiov[i].base != (unsigned long) fiov[i].base ||
1680                     fiov[i].len != (unsigned long) fiov[i].len)
1681                         return -EIO;
1682
1683                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1684                 dst[i].iov_len = (size_t) fiov[i].len;
1685
1686 #ifdef CONFIG_COMPAT
1687                 if (is_compat &&
1688                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1689                      (compat_size_t) dst[i].iov_len != fiov[i].len))
1690                         return -EIO;
1691 #endif
1692         }
1693
1694         return 0;
1695 }
1696
1697
1698 /*
1699  * For ioctls, there is no generic way to determine how much memory
1700  * needs to be read and/or written.  Furthermore, ioctls are allowed
1701  * to dereference the passed pointer, so the parameter requires deep
1702  * copying but FUSE has no idea whatsoever about what to copy in or
1703  * out.
1704  *
1705  * This is solved by allowing FUSE server to retry ioctl with
1706  * necessary in/out iovecs.  Let's assume the ioctl implementation
1707  * needs to read in the following structure.
1708  *
1709  * struct a {
1710  *      char    *buf;
1711  *      size_t  buflen;
1712  * }
1713  *
1714  * On the first callout to FUSE server, inarg->in_size and
1715  * inarg->out_size will be NULL; then, the server completes the ioctl
1716  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1717  * the actual iov array to
1718  *
1719  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1720  *
1721  * which tells FUSE to copy in the requested area and retry the ioctl.
1722  * On the second round, the server has access to the structure and
1723  * from that it can tell what to look for next, so on the invocation,
1724  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1725  *
1726  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1727  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1728  *
1729  * FUSE will copy both struct a and the pointed buffer from the
1730  * process doing the ioctl and retry ioctl with both struct a and the
1731  * buffer.
1732  *
1733  * This time, FUSE server has everything it needs and completes ioctl
1734  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1735  *
1736  * Copying data out works the same way.
1737  *
1738  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1739  * automatically initializes in and out iovs by decoding @cmd with
1740  * _IOC_* macros and the server is not allowed to request RETRY.  This
1741  * limits ioctl data transfers to well-formed ioctls and is the forced
1742  * behavior for all FUSE servers.
1743  */
1744 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1745                    unsigned int flags)
1746 {
1747         struct fuse_file *ff = file->private_data;
1748         struct fuse_conn *fc = ff->fc;
1749         struct fuse_ioctl_in inarg = {
1750                 .fh = ff->fh,
1751                 .cmd = cmd,
1752                 .arg = arg,
1753                 .flags = flags
1754         };
1755         struct fuse_ioctl_out outarg;
1756         struct fuse_req *req = NULL;
1757         struct page **pages = NULL;
1758         struct iovec *iov_page = NULL;
1759         struct iovec *in_iov = NULL, *out_iov = NULL;
1760         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1761         size_t in_size, out_size, transferred;
1762         int err;
1763
1764 #if BITS_PER_LONG == 32
1765         inarg.flags |= FUSE_IOCTL_32BIT;
1766 #else
1767         if (flags & FUSE_IOCTL_COMPAT)
1768                 inarg.flags |= FUSE_IOCTL_32BIT;
1769 #endif
1770
1771         /* assume all the iovs returned by client always fits in a page */
1772         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1773
1774         err = -ENOMEM;
1775         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1776         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1777         if (!pages || !iov_page)
1778                 goto out;
1779
1780         /*
1781          * If restricted, initialize IO parameters as encoded in @cmd.
1782          * RETRY from server is not allowed.
1783          */
1784         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1785                 struct iovec *iov = iov_page;
1786
1787                 iov->iov_base = (void __user *)arg;
1788                 iov->iov_len = _IOC_SIZE(cmd);
1789
1790                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1791                         in_iov = iov;
1792                         in_iovs = 1;
1793                 }
1794
1795                 if (_IOC_DIR(cmd) & _IOC_READ) {
1796                         out_iov = iov;
1797                         out_iovs = 1;
1798                 }
1799         }
1800
1801  retry:
1802         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1803         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1804
1805         /*
1806          * Out data can be used either for actual out data or iovs,
1807          * make sure there always is at least one page.
1808          */
1809         out_size = max_t(size_t, out_size, PAGE_SIZE);
1810         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1811
1812         /* make sure there are enough buffer pages and init request with them */
1813         err = -ENOMEM;
1814         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1815                 goto out;
1816         while (num_pages < max_pages) {
1817                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1818                 if (!pages[num_pages])
1819                         goto out;
1820                 num_pages++;
1821         }
1822
1823         req = fuse_get_req(fc);
1824         if (IS_ERR(req)) {
1825                 err = PTR_ERR(req);
1826                 req = NULL;
1827                 goto out;
1828         }
1829         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1830         req->num_pages = num_pages;
1831
1832         /* okay, let's send it to the client */
1833         req->in.h.opcode = FUSE_IOCTL;
1834         req->in.h.nodeid = ff->nodeid;
1835         req->in.numargs = 1;
1836         req->in.args[0].size = sizeof(inarg);
1837         req->in.args[0].value = &inarg;
1838         if (in_size) {
1839                 req->in.numargs++;
1840                 req->in.args[1].size = in_size;
1841                 req->in.argpages = 1;
1842
1843                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1844                                            false);
1845                 if (err)
1846                         goto out;
1847         }
1848
1849         req->out.numargs = 2;
1850         req->out.args[0].size = sizeof(outarg);
1851         req->out.args[0].value = &outarg;
1852         req->out.args[1].size = out_size;
1853         req->out.argpages = 1;
1854         req->out.argvar = 1;
1855
1856         fuse_request_send(fc, req);
1857         err = req->out.h.error;
1858         transferred = req->out.args[1].size;
1859         fuse_put_request(fc, req);
1860         req = NULL;
1861         if (err)
1862                 goto out;
1863
1864         /* did it ask for retry? */
1865         if (outarg.flags & FUSE_IOCTL_RETRY) {
1866                 void *vaddr;
1867
1868                 /* no retry if in restricted mode */
1869                 err = -EIO;
1870                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1871                         goto out;
1872
1873                 in_iovs = outarg.in_iovs;
1874                 out_iovs = outarg.out_iovs;
1875
1876                 /*
1877                  * Make sure things are in boundary, separate checks
1878                  * are to protect against overflow.
1879                  */
1880                 err = -ENOMEM;
1881                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1882                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1883                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1884                         goto out;
1885
1886                 vaddr = kmap_atomic(pages[0], KM_USER0);
1887                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1888                                             transferred, in_iovs + out_iovs,
1889                                             (flags & FUSE_IOCTL_COMPAT) != 0);
1890                 kunmap_atomic(vaddr, KM_USER0);
1891                 if (err)
1892                         goto out;
1893
1894                 in_iov = iov_page;
1895                 out_iov = in_iov + in_iovs;
1896
1897                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1898                 if (err)
1899                         goto out;
1900
1901                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1902                 if (err)
1903                         goto out;
1904
1905                 goto retry;
1906         }
1907
1908         err = -EIO;
1909         if (transferred > inarg.out_size)
1910                 goto out;
1911
1912         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1913  out:
1914         if (req)
1915                 fuse_put_request(fc, req);
1916         free_page((unsigned long) iov_page);
1917         while (num_pages)
1918                 __free_page(pages[--num_pages]);
1919         kfree(pages);
1920
1921         return err ? err : outarg.result;
1922 }
1923 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1924
1925 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
1926                                    unsigned long arg, unsigned int flags)
1927 {
1928         struct inode *inode = file->f_dentry->d_inode;
1929         struct fuse_conn *fc = get_fuse_conn(inode);
1930
1931         if (!fuse_allow_task(fc, current))
1932                 return -EACCES;
1933
1934         if (is_bad_inode(inode))
1935                 return -EIO;
1936
1937         return fuse_do_ioctl(file, cmd, arg, flags);
1938 }
1939
1940 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1941                             unsigned long arg)
1942 {
1943         return fuse_file_ioctl_common(file, cmd, arg, 0);
1944 }
1945
1946 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1947                                    unsigned long arg)
1948 {
1949         return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
1950 }
1951
1952 /*
1953  * All files which have been polled are linked to RB tree
1954  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
1955  * find the matching one.
1956  */
1957 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1958                                               struct rb_node **parent_out)
1959 {
1960         struct rb_node **link = &fc->polled_files.rb_node;
1961         struct rb_node *last = NULL;
1962
1963         while (*link) {
1964                 struct fuse_file *ff;
1965
1966                 last = *link;
1967                 ff = rb_entry(last, struct fuse_file, polled_node);
1968
1969                 if (kh < ff->kh)
1970                         link = &last->rb_left;
1971                 else if (kh > ff->kh)
1972                         link = &last->rb_right;
1973                 else
1974                         return link;
1975         }
1976
1977         if (parent_out)
1978                 *parent_out = last;
1979         return link;
1980 }
1981
1982 /*
1983  * The file is about to be polled.  Make sure it's on the polled_files
1984  * RB tree.  Note that files once added to the polled_files tree are
1985  * not removed before the file is released.  This is because a file
1986  * polled once is likely to be polled again.
1987  */
1988 static void fuse_register_polled_file(struct fuse_conn *fc,
1989                                       struct fuse_file *ff)
1990 {
1991         spin_lock(&fc->lock);
1992         if (RB_EMPTY_NODE(&ff->polled_node)) {
1993                 struct rb_node **link, *parent;
1994
1995                 link = fuse_find_polled_node(fc, ff->kh, &parent);
1996                 BUG_ON(*link);
1997                 rb_link_node(&ff->polled_node, parent, link);
1998                 rb_insert_color(&ff->polled_node, &fc->polled_files);
1999         }
2000         spin_unlock(&fc->lock);
2001 }
2002
2003 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2004 {
2005         struct fuse_file *ff = file->private_data;
2006         struct fuse_conn *fc = ff->fc;
2007         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2008         struct fuse_poll_out outarg;
2009         struct fuse_req *req;
2010         int err;
2011
2012         if (fc->no_poll)
2013                 return DEFAULT_POLLMASK;
2014
2015         poll_wait(file, &ff->poll_wait, wait);
2016
2017         /*
2018          * Ask for notification iff there's someone waiting for it.
2019          * The client may ignore the flag and always notify.
2020          */
2021         if (waitqueue_active(&ff->poll_wait)) {
2022                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2023                 fuse_register_polled_file(fc, ff);
2024         }
2025
2026         req = fuse_get_req(fc);
2027         if (IS_ERR(req))
2028                 return POLLERR;
2029
2030         req->in.h.opcode = FUSE_POLL;
2031         req->in.h.nodeid = ff->nodeid;
2032         req->in.numargs = 1;
2033         req->in.args[0].size = sizeof(inarg);
2034         req->in.args[0].value = &inarg;
2035         req->out.numargs = 1;
2036         req->out.args[0].size = sizeof(outarg);
2037         req->out.args[0].value = &outarg;
2038         fuse_request_send(fc, req);
2039         err = req->out.h.error;
2040         fuse_put_request(fc, req);
2041
2042         if (!err)
2043                 return outarg.revents;
2044         if (err == -ENOSYS) {
2045                 fc->no_poll = 1;
2046                 return DEFAULT_POLLMASK;
2047         }
2048         return POLLERR;
2049 }
2050 EXPORT_SYMBOL_GPL(fuse_file_poll);
2051
2052 /*
2053  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2054  * wakes up the poll waiters.
2055  */
2056 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2057                             struct fuse_notify_poll_wakeup_out *outarg)
2058 {
2059         u64 kh = outarg->kh;
2060         struct rb_node **link;
2061
2062         spin_lock(&fc->lock);
2063
2064         link = fuse_find_polled_node(fc, kh, NULL);
2065         if (*link) {
2066                 struct fuse_file *ff;
2067
2068                 ff = rb_entry(*link, struct fuse_file, polled_node);
2069                 wake_up_interruptible_sync(&ff->poll_wait);
2070         }
2071
2072         spin_unlock(&fc->lock);
2073         return 0;
2074 }
2075
2076 static const struct file_operations fuse_file_operations = {
2077         .llseek         = fuse_file_llseek,
2078         .read           = do_sync_read,
2079         .aio_read       = fuse_file_aio_read,
2080         .write          = do_sync_write,
2081         .aio_write      = fuse_file_aio_write,
2082         .mmap           = fuse_file_mmap,
2083         .open           = fuse_open,
2084         .flush          = fuse_flush,
2085         .release        = fuse_release,
2086         .fsync          = fuse_fsync,
2087         .lock           = fuse_file_lock,
2088         .flock          = fuse_file_flock,
2089         .splice_read    = generic_file_splice_read,
2090         .unlocked_ioctl = fuse_file_ioctl,
2091         .compat_ioctl   = fuse_file_compat_ioctl,
2092         .poll           = fuse_file_poll,
2093 };
2094
2095 static const struct file_operations fuse_direct_io_file_operations = {
2096         .llseek         = fuse_file_llseek,
2097         .read           = fuse_direct_read,
2098         .write          = fuse_direct_write,
2099         .mmap           = fuse_direct_mmap,
2100         .open           = fuse_open,
2101         .flush          = fuse_flush,
2102         .release        = fuse_release,
2103         .fsync          = fuse_fsync,
2104         .lock           = fuse_file_lock,
2105         .flock          = fuse_file_flock,
2106         .unlocked_ioctl = fuse_file_ioctl,
2107         .compat_ioctl   = fuse_file_compat_ioctl,
2108         .poll           = fuse_file_poll,
2109         /* no splice_read */
2110 };
2111
2112 static const struct address_space_operations fuse_file_aops  = {
2113         .readpage       = fuse_readpage,
2114         .writepage      = fuse_writepage,
2115         .launder_page   = fuse_launder_page,
2116         .readpages      = fuse_readpages,
2117         .set_page_dirty = __set_page_dirty_nobuffers,
2118         .bmap           = fuse_bmap,
2119 };
2120
2121 void fuse_init_file_inode(struct inode *inode)
2122 {
2123         inode->i_fop = &fuse_file_operations;
2124         inode->i_data.a_ops = &fuse_file_aops;
2125 }