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