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/sched/signal.h>
  16 #include <linux/module.h>
  17 #include <linux/compat.h>
  18 #include <linux/swap.h>
  19 #include <linux/falloc.h>
  20 #include <linux/uio.h>
  21
  22 static struct page **fuse_pages_alloc(unsigned int npages, gfp_t flags,
  23                                       struct fuse_page_desc **desc)
  24 {
  25         struct page **pages;
  26
  27         pages = kzalloc(npages * (sizeof(struct page *) +
  28                                   sizeof(struct fuse_page_desc)), flags);
  29         *desc = (void *) (pages + npages);
  30
  31         return pages;
  32 }
  33
  34 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  35                           int opcode, struct fuse_open_out *outargp)
  36 {
  37         struct fuse_open_in inarg;
  38         FUSE_ARGS(args);
  39
  40         memset(&inarg, 0, sizeof(inarg));
  41         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  42         if (!fc->atomic_o_trunc)
  43                 inarg.flags &= ~O_TRUNC;
  44         args.opcode = opcode;
  45         args.nodeid = nodeid;
  46         args.in_numargs = 1;
  47         args.in_args[0].size = sizeof(inarg);
  48         args.in_args[0].value = &inarg;
  49         args.out_numargs = 1;
  50         args.out_args[0].size = sizeof(*outargp);
  51         args.out_args[0].value = outargp;
  52
  53         return fuse_simple_request(fc, &args);
  54 }
  55
  56 struct fuse_release_args {
  57         struct fuse_args args;
  58         struct fuse_release_in inarg;
  59         struct inode *inode;
  60 };
  61
  62 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  63 {
  64         struct fuse_file *ff;
  65
  66         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
  67         if (unlikely(!ff))
  68                 return NULL;
  69
  70         ff->fc = fc;
  71         ff->release_args = kzalloc(sizeof(*ff->release_args),
  72                                    GFP_KERNEL_ACCOUNT);
  73         if (!ff->release_args) {
  74                 kfree(ff);
  75                 return NULL;
  76         }
  77
  78         INIT_LIST_HEAD(&ff->write_entry);
  79         mutex_init(&ff->readdir.lock);
  80         refcount_set(&ff->count, 1);
  81         RB_CLEAR_NODE(&ff->polled_node);
  82         init_waitqueue_head(&ff->poll_wait);
  83
  84         ff->kh = atomic64_inc_return(&fc->khctr);
  85
  86         return ff;
  87 }
  88
  89 void fuse_file_free(struct fuse_file *ff)
  90 {
  91         kfree(ff->release_args);
  92         mutex_destroy(&ff->readdir.lock);
  93         kfree(ff);
  94 }
  95
  96 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  97 {
  98         refcount_inc(&ff->count);
  99         return ff;
 100 }
 101
 102 static void fuse_release_end(struct fuse_conn *fc, struct fuse_args *args,
 103                              int error)
 104 {
 105         struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
 106
 107         iput(ra->inode);
 108         kfree(ra);
 109 }
 110
 111 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
 112 {
 113         if (refcount_dec_and_test(&ff->count)) {
 114                 struct fuse_args *args = &ff->release_args->args;
 115
 116                 if (isdir ? ff->fc->no_opendir : ff->fc->no_open) {
 117                         /* Do nothing when client does not implement 'open' */
 118                         fuse_release_end(ff->fc, args, 0);
 119                 } else if (sync) {
 120                         fuse_simple_request(ff->fc, args);
 121                         fuse_release_end(ff->fc, args, 0);
 122                 } else {
 123                         args->end = fuse_release_end;
 124                         if (fuse_simple_background(ff->fc, args,
 125                                                    GFP_KERNEL | __GFP_NOFAIL))
 126                                 fuse_release_end(ff->fc, args, -ENOTCONN);
 127                 }
 128                 kfree(ff);
 129         }
 130 }
 131
 132 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 133                  bool isdir)
 134 {
 135         struct fuse_file *ff;
 136         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 137
 138         ff = fuse_file_alloc(fc);
 139         if (!ff)
 140                 return -ENOMEM;
 141
 142         ff->fh = 0;
 143         /* Default for no-open */
 144         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
 145         if (isdir ? !fc->no_opendir : !fc->no_open) {
 146                 struct fuse_open_out outarg;
 147                 int err;
 148
 149                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 150                 if (!err) {
 151                         ff->fh = outarg.fh;
 152                         ff->open_flags = outarg.open_flags;
 153
 154                 } else if (err != -ENOSYS) {
 155                         fuse_file_free(ff);
 156                         return err;
 157                 } else {
 158                         if (isdir)
 159                                 fc->no_opendir = 1;
 160                         else
 161                                 fc->no_open = 1;
 162                 }
 163         }
 164
 165         if (isdir)
 166                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 167
 168         ff->nodeid = nodeid;
 169         file->private_data = ff;
 170
 171         return 0;
 172 }
 173 EXPORT_SYMBOL_GPL(fuse_do_open);
 174
 175 static void fuse_link_write_file(struct file *file)
 176 {
 177         struct inode *inode = file_inode(file);
 178         struct fuse_inode *fi = get_fuse_inode(inode);
 179         struct fuse_file *ff = file->private_data;
 180         /*
 181          * file may be written through mmap, so chain it onto the
 182          * inodes's write_file list
 183          */
 184         spin_lock(&fi->lock);
 185         if (list_empty(&ff->write_entry))
 186                 list_add(&ff->write_entry, &fi->write_files);
 187         spin_unlock(&fi->lock);
 188 }
 189
 190 void fuse_finish_open(struct inode *inode, struct file *file)
 191 {
 192         struct fuse_file *ff = file->private_data;
 193         struct fuse_conn *fc = get_fuse_conn(inode);
 194
 195         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 196                 invalidate_inode_pages2(inode->i_mapping);
 197         if (ff->open_flags & FOPEN_STREAM)
 198                 stream_open(inode, file);
 199         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 200                 nonseekable_open(inode, file);
 201         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 202                 struct fuse_inode *fi = get_fuse_inode(inode);
 203
 204                 spin_lock(&fi->lock);
 205                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 206                 i_size_write(inode, 0);
 207                 spin_unlock(&fi->lock);
 208                 fuse_invalidate_attr(inode);
 209                 if (fc->writeback_cache)
 210                         file_update_time(file);
 211         }
 212         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 213                 fuse_link_write_file(file);
 214 }
 215
 216 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 217 {
 218         struct fuse_conn *fc = get_fuse_conn(inode);
 219         int err;
 220         bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
 221                           fc->atomic_o_trunc &&
 222                           fc->writeback_cache;
 223
 224         err = generic_file_open(inode, file);
 225         if (err)
 226                 return err;
 227
 228         if (is_wb_truncate) {
 229                 inode_lock(inode);
 230                 fuse_set_nowrite(inode);
 231         }
 232
 233         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 234
 235         if (!err)
 236                 fuse_finish_open(inode, file);
 237
 238         if (is_wb_truncate) {
 239                 fuse_release_nowrite(inode);
 240                 inode_unlock(inode);
 241         }
 242
 243         return err;
 244 }
 245
 246 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 247                                  int flags, int opcode)
 248 {
 249         struct fuse_conn *fc = ff->fc;
 250         struct fuse_release_args *ra = ff->release_args;
 251
 252         /* Inode is NULL on error path of fuse_create_open() */
 253         if (likely(fi)) {
 254                 spin_lock(&fi->lock);
 255                 list_del(&ff->write_entry);
 256                 spin_unlock(&fi->lock);
 257         }
 258         spin_lock(&fc->lock);
 259         if (!RB_EMPTY_NODE(&ff->polled_node))
 260                 rb_erase(&ff->polled_node, &fc->polled_files);
 261         spin_unlock(&fc->lock);
 262
 263         wake_up_interruptible_all(&ff->poll_wait);
 264
 265         ra->inarg.fh = ff->fh;
 266         ra->inarg.flags = flags;
 267         ra->args.in_numargs = 1;
 268         ra->args.in_args[0].size = sizeof(struct fuse_release_in);
 269         ra->args.in_args[0].value = &ra->inarg;
 270         ra->args.opcode = opcode;
 271         ra->args.nodeid = ff->nodeid;
 272         ra->args.force = true;
 273         ra->args.nocreds = true;
 274 }
 275
 276 void fuse_release_common(struct file *file, bool isdir)
 277 {
 278         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 279         struct fuse_file *ff = file->private_data;
 280         struct fuse_release_args *ra = ff->release_args;
 281         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 282
 283         fuse_prepare_release(fi, ff, file->f_flags, opcode);
 284
 285         if (ff->flock) {
 286                 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 287                 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fc,
 288                                                           (fl_owner_t) file);
 289         }
 290         /* Hold inode until release is finished */
 291         ra->inode = igrab(file_inode(file));
 292
 293         /*
 294          * Normally this will send the RELEASE request, however if
 295          * some asynchronous READ or WRITE requests are outstanding,
 296          * the sending will be delayed.
 297          *
 298          * Make the release synchronous if this is a fuseblk mount,
 299          * synchronous RELEASE is allowed (and desirable) in this case
 300          * because the server can be trusted not to screw up.
 301          */
 302         fuse_file_put(ff, ff->fc->destroy, isdir);
 303 }
 304
 305 static int fuse_open(struct inode *inode, struct file *file)
 306 {
 307         return fuse_open_common(inode, file, false);
 308 }
 309
 310 static int fuse_release(struct inode *inode, struct file *file)
 311 {
 312         struct fuse_conn *fc = get_fuse_conn(inode);
 313
 314         /* see fuse_vma_close() for !writeback_cache case */
 315         if (fc->writeback_cache)
 316                 write_inode_now(inode, 1);
 317
 318         fuse_release_common(file, false);
 319
 320         /* return value is ignored by VFS */
 321         return 0;
 322 }
 323
 324 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
 325 {
 326         WARN_ON(refcount_read(&ff->count) > 1);
 327         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 328         /*
 329          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 330          * synchronous, we are fine with not doing igrab() here"
 331          */
 332         fuse_file_put(ff, true, false);
 333 }
 334 EXPORT_SYMBOL_GPL(fuse_sync_release);
 335
 336 /*
 337  * Scramble the ID space with XTEA, so that the value of the files_struct
 338  * pointer is not exposed to userspace.
 339  */
 340 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 341 {
 342         u32 *k = fc->scramble_key;
 343         u64 v = (unsigned long) id;
 344         u32 v0 = v;
 345         u32 v1 = v >> 32;
 346         u32 sum = 0;
 347         int i;
 348
 349         for (i = 0; i < 32; i++) {
 350                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 351                 sum += 0x9E3779B9;
 352                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 353         }
 354
 355         return (u64) v0 + ((u64) v1 << 32);
 356 }
 357
 358 struct fuse_writepage_args {
 359         struct fuse_io_args ia;
 360         struct list_head writepages_entry;
 361         struct list_head queue_entry;
 362         struct fuse_writepage_args *next;
 363         struct inode *inode;
 364 };
 365
 366 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
 367                                             pgoff_t idx_from, pgoff_t idx_to)
 368 {
 369         struct fuse_writepage_args *wpa;
 370
 371         list_for_each_entry(wpa, &fi->writepages, writepages_entry) {
 372                 pgoff_t curr_index;
 373
 374                 WARN_ON(get_fuse_inode(wpa->inode) != fi);
 375                 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
 376                 if (idx_from < curr_index + wpa->ia.ap.num_pages &&
 377                     curr_index <= idx_to) {
 378                         return wpa;
 379                 }
 380         }
 381         return NULL;
 382 }
 383
 384 /*
 385  * Check if any page in a range is under writeback
 386  *
 387  * This is currently done by walking the list of writepage requests
 388  * for the inode, which can be pretty inefficient.
 389  */
 390 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 391                                    pgoff_t idx_to)
 392 {
 393         struct fuse_inode *fi = get_fuse_inode(inode);
 394         bool found;
 395
 396         spin_lock(&fi->lock);
 397         found = fuse_find_writeback(fi, idx_from, idx_to);
 398         spin_unlock(&fi->lock);
 399
 400         return found;
 401 }
 402
 403 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 404 {
 405         return fuse_range_is_writeback(inode, index, index);
 406 }
 407
 408 /*
 409  * Wait for page writeback to be completed.
 410  *
 411  * Since fuse doesn't rely on the VM writeback tracking, this has to
 412  * use some other means.
 413  */
 414 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 415 {
 416         struct fuse_inode *fi = get_fuse_inode(inode);
 417
 418         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 419 }
 420
 421 /*
 422  * Wait for all pending writepages on the inode to finish.
 423  *
 424  * This is currently done by blocking further writes with FUSE_NOWRITE
 425  * and waiting for all sent writes to complete.
 426  *
 427  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 428  * could conflict with truncation.
 429  */
 430 static void fuse_sync_writes(struct inode *inode)
 431 {
 432         fuse_set_nowrite(inode);
 433         fuse_release_nowrite(inode);
 434 }
 435
 436 static int fuse_flush(struct file *file, fl_owner_t id)
 437 {
 438         struct inode *inode = file_inode(file);
 439         struct fuse_conn *fc = get_fuse_conn(inode);
 440         struct fuse_file *ff = file->private_data;
 441         struct fuse_flush_in inarg;
 442         FUSE_ARGS(args);
 443         int err;
 444
 445         if (is_bad_inode(inode))
 446                 return -EIO;
 447
 448         if (fc->no_flush)
 449                 return 0;
 450
 451         err = write_inode_now(inode, 1);
 452         if (err)
 453                 return err;
 454
 455         inode_lock(inode);
 456         fuse_sync_writes(inode);
 457         inode_unlock(inode);
 458
 459         err = filemap_check_errors(file->f_mapping);
 460         if (err)
 461                 return err;
 462
 463         memset(&inarg, 0, sizeof(inarg));
 464         inarg.fh = ff->fh;
 465         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 466         args.opcode = FUSE_FLUSH;
 467         args.nodeid = get_node_id(inode);
 468         args.in_numargs = 1;
 469         args.in_args[0].size = sizeof(inarg);
 470         args.in_args[0].value = &inarg;
 471         args.force = true;
 472
 473         err = fuse_simple_request(fc, &args);
 474         if (err == -ENOSYS) {
 475                 fc->no_flush = 1;
 476                 err = 0;
 477         }
 478         return err;
 479 }
 480
 481 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 482                       int datasync, int opcode)
 483 {
 484         struct inode *inode = file->f_mapping->host;
 485         struct fuse_conn *fc = get_fuse_conn(inode);
 486         struct fuse_file *ff = file->private_data;
 487         FUSE_ARGS(args);
 488         struct fuse_fsync_in inarg;
 489
 490         memset(&inarg, 0, sizeof(inarg));
 491         inarg.fh = ff->fh;
 492         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 493         args.opcode = opcode;
 494         args.nodeid = get_node_id(inode);
 495         args.in_numargs = 1;
 496         args.in_args[0].size = sizeof(inarg);
 497         args.in_args[0].value = &inarg;
 498         return fuse_simple_request(fc, &args);
 499 }
 500
 501 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 502                       int datasync)
 503 {
 504         struct inode *inode = file->f_mapping->host;
 505         struct fuse_conn *fc = get_fuse_conn(inode);
 506         int err;
 507
 508         if (is_bad_inode(inode))
 509                 return -EIO;
 510
 511         inode_lock(inode);
 512
 513         /*
 514          * Start writeback against all dirty pages of the inode, then
 515          * wait for all outstanding writes, before sending the FSYNC
 516          * request.
 517          */
 518         err = file_write_and_wait_range(file, start, end);
 519         if (err)
 520                 goto out;
 521
 522         fuse_sync_writes(inode);
 523
 524         /*
 525          * Due to implementation of fuse writeback
 526          * file_write_and_wait_range() does not catch errors.
 527          * We have to do this directly after fuse_sync_writes()
 528          */
 529         err = file_check_and_advance_wb_err(file);
 530         if (err)
 531                 goto out;
 532
 533         err = sync_inode_metadata(inode, 1);
 534         if (err)
 535                 goto out;
 536
 537         if (fc->no_fsync)
 538                 goto out;
 539
 540         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 541         if (err == -ENOSYS) {
 542                 fc->no_fsync = 1;
 543                 err = 0;
 544         }
 545 out:
 546         inode_unlock(inode);
 547
 548         return err;
 549 }
 550
 551 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
 552                          size_t count, int opcode)
 553 {
 554         struct fuse_file *ff = file->private_data;
 555         struct fuse_args *args = &ia->ap.args;
 556
 557         ia->read.in.fh = ff->fh;
 558         ia->read.in.offset = pos;
 559         ia->read.in.size = count;
 560         ia->read.in.flags = file->f_flags;
 561         args->opcode = opcode;
 562         args->nodeid = ff->nodeid;
 563         args->in_numargs = 1;
 564         args->in_args[0].size = sizeof(ia->read.in);
 565         args->in_args[0].value = &ia->read.in;
 566         args->out_argvar = true;
 567         args->out_numargs = 1;
 568         args->out_args[0].size = count;
 569 }
 570
 571 static void fuse_release_user_pages(struct fuse_args_pages *ap,
 572                                     bool should_dirty)
 573 {
 574         unsigned int i;
 575
 576         for (i = 0; i < ap->num_pages; i++) {
 577                 if (should_dirty)
 578                         set_page_dirty_lock(ap->pages[i]);
 579                 put_page(ap->pages[i]);
 580         }
 581 }
 582
 583 static void fuse_io_release(struct kref *kref)
 584 {
 585         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 586 }
 587
 588 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 589 {
 590         if (io->err)
 591                 return io->err;
 592
 593         if (io->bytes >= 0 && io->write)
 594                 return -EIO;
 595
 596         return io->bytes < 0 ? io->size : io->bytes;
 597 }
 598
 599 /**
 600  * In case of short read, the caller sets 'pos' to the position of
 601  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 602  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 603  *
 604  * An example:
 605  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 606  * both submitted asynchronously. The first of them was ACKed by userspace as
 607  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 608  * second request was ACKed as short, e.g. only 1K was read, resulting in
 609  * pos == 33K.
 610  *
 611  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 612  * will be equal to the length of the longest contiguous fragment of
 613  * transferred data starting from the beginning of IO request.
 614  */
 615 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 616 {
 617         int left;
 618
 619         spin_lock(&io->lock);
 620         if (err)
 621                 io->err = io->err ? : err;
 622         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 623                 io->bytes = pos;
 624
 625         left = --io->reqs;
 626         if (!left && io->blocking)
 627                 complete(io->done);
 628         spin_unlock(&io->lock);
 629
 630         if (!left && !io->blocking) {
 631                 ssize_t res = fuse_get_res_by_io(io);
 632
 633                 if (res >= 0) {
 634                         struct inode *inode = file_inode(io->iocb->ki_filp);
 635                         struct fuse_conn *fc = get_fuse_conn(inode);
 636                         struct fuse_inode *fi = get_fuse_inode(inode);
 637
 638                         spin_lock(&fi->lock);
 639                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 640                         spin_unlock(&fi->lock);
 641                 }
 642
 643                 io->iocb->ki_complete(io->iocb, res, 0);
 644         }
 645
 646         kref_put(&io->refcnt, fuse_io_release);
 647 }
 648
 649 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
 650                                           unsigned int npages)
 651 {
 652         struct fuse_io_args *ia;
 653
 654         ia = kzalloc(sizeof(*ia), GFP_KERNEL);
 655         if (ia) {
 656                 ia->io = io;
 657                 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
 658                                                 &ia->ap.descs);
 659                 if (!ia->ap.pages) {
 660                         kfree(ia);
 661                         ia = NULL;
 662                 }
 663         }
 664         return ia;
 665 }
 666
 667 static void fuse_io_free(struct fuse_io_args *ia)
 668 {
 669         kfree(ia->ap.pages);
 670         kfree(ia);
 671 }
 672
 673 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_args *args,
 674                                   int err)
 675 {
 676         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 677         struct fuse_io_priv *io = ia->io;
 678         ssize_t pos = -1;
 679
 680         fuse_release_user_pages(&ia->ap, io->should_dirty);
 681
 682         if (err) {
 683                 /* Nothing */
 684         } else if (io->write) {
 685                 if (ia->write.out.size > ia->write.in.size) {
 686                         err = -EIO;
 687                 } else if (ia->write.in.size != ia->write.out.size) {
 688                         pos = ia->write.in.offset - io->offset +
 689                                 ia->write.out.size;
 690                 }
 691         } else {
 692                 u32 outsize = args->out_args[0].size;
 693
 694                 if (ia->read.in.size != outsize)
 695                         pos = ia->read.in.offset - io->offset + outsize;
 696         }
 697
 698         fuse_aio_complete(io, err, pos);
 699         fuse_io_free(ia);
 700 }
 701
 702 static ssize_t fuse_async_req_send(struct fuse_conn *fc,
 703                                    struct fuse_io_args *ia, size_t num_bytes)
 704 {
 705         ssize_t err;
 706         struct fuse_io_priv *io = ia->io;
 707
 708         spin_lock(&io->lock);
 709         kref_get(&io->refcnt);
 710         io->size += num_bytes;
 711         io->reqs++;
 712         spin_unlock(&io->lock);
 713
 714         ia->ap.args.end = fuse_aio_complete_req;
 715         err = fuse_simple_background(fc, &ia->ap.args, GFP_KERNEL);
 716         if (err)
 717                 fuse_aio_complete_req(fc, &ia->ap.args, err);
 718
 719         return num_bytes;
 720 }
 721
 722 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
 723                               fl_owner_t owner)
 724 {
 725         struct file *file = ia->io->iocb->ki_filp;
 726         struct fuse_file *ff = file->private_data;
 727         struct fuse_conn *fc = ff->fc;
 728
 729         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 730         if (owner != NULL) {
 731                 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
 732                 ia->read.in.lock_owner = fuse_lock_owner_id(fc, owner);
 733         }
 734
 735         if (ia->io->async)
 736                 return fuse_async_req_send(fc, ia, count);
 737
 738         return fuse_simple_request(fc, &ia->ap.args);
 739 }
 740
 741 static void fuse_read_update_size(struct inode *inode, loff_t size,
 742                                   u64 attr_ver)
 743 {
 744         struct fuse_conn *fc = get_fuse_conn(inode);
 745         struct fuse_inode *fi = get_fuse_inode(inode);
 746
 747         spin_lock(&fi->lock);
 748         if (attr_ver == fi->attr_version && size < inode->i_size &&
 749             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 750                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 751                 i_size_write(inode, size);
 752         }
 753         spin_unlock(&fi->lock);
 754 }
 755
 756 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
 757                             struct fuse_args_pages *ap)
 758 {
 759         struct fuse_conn *fc = get_fuse_conn(inode);
 760
 761         if (fc->writeback_cache) {
 762                 /*
 763                  * A hole in a file. Some data after the hole are in page cache,
 764                  * but have not reached the client fs yet. So, the hole is not
 765                  * present there.
 766                  */
 767                 int i;
 768                 int start_idx = num_read >> PAGE_SHIFT;
 769                 size_t off = num_read & (PAGE_SIZE - 1);
 770
 771                 for (i = start_idx; i < ap->num_pages; i++) {
 772                         zero_user_segment(ap->pages[i], off, PAGE_SIZE);
 773                         off = 0;
 774                 }
 775         } else {
 776                 loff_t pos = page_offset(ap->pages[0]) + num_read;
 777                 fuse_read_update_size(inode, pos, attr_ver);
 778         }
 779 }
 780
 781 static int fuse_do_readpage(struct file *file, struct page *page)
 782 {
 783         struct inode *inode = page->mapping->host;
 784         struct fuse_conn *fc = get_fuse_conn(inode);
 785         loff_t pos = page_offset(page);
 786         struct fuse_page_desc desc = { .length = PAGE_SIZE };
 787         struct fuse_io_args ia = {
 788                 .ap.args.page_zeroing = true,
 789                 .ap.args.out_pages = true,
 790                 .ap.num_pages = 1,
 791                 .ap.pages = &page,
 792                 .ap.descs = &desc,
 793         };
 794         ssize_t res;
 795         u64 attr_ver;
 796
 797         /*
 798          * Page writeback can extend beyond the lifetime of the
 799          * page-cache page, so make sure we read a properly synced
 800          * page.
 801          */
 802         fuse_wait_on_page_writeback(inode, page->index);
 803
 804         attr_ver = fuse_get_attr_version(fc);
 805
 806         /* Don't overflow end offset */
 807         if (pos + (desc.length - 1) == LLONG_MAX)
 808                 desc.length--;
 809
 810         fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
 811         res = fuse_simple_request(fc, &ia.ap.args);
 812         if (res < 0)
 813                 return res;
 814         /*
 815          * Short read means EOF.  If file size is larger, truncate it
 816          */
 817         if (res < desc.length)
 818                 fuse_short_read(inode, attr_ver, res, &ia.ap);
 819
 820         SetPageUptodate(page);
 821
 822         return 0;
 823 }
 824
 825 static int fuse_readpage(struct file *file, struct page *page)
 826 {
 827         struct inode *inode = page->mapping->host;
 828         int err;
 829
 830         err = -EIO;
 831         if (is_bad_inode(inode))
 832                 goto out;
 833
 834         err = fuse_do_readpage(file, page);
 835         fuse_invalidate_atime(inode);
 836  out:
 837         unlock_page(page);
 838         return err;
 839 }
 840
 841 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_args *args,
 842                                int err)
 843 {
 844         int i;
 845         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 846         struct fuse_args_pages *ap = &ia->ap;
 847         size_t count = ia->read.in.size;
 848         size_t num_read = args->out_args[0].size;
 849         struct address_space *mapping = NULL;
 850
 851         for (i = 0; mapping == NULL && i < ap->num_pages; i++)
 852                 mapping = ap->pages[i]->mapping;
 853
 854         if (mapping) {
 855                 struct inode *inode = mapping->host;
 856
 857                 /*
 858                  * Short read means EOF. If file size is larger, truncate it
 859                  */
 860                 if (!err && num_read < count)
 861                         fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
 862
 863                 fuse_invalidate_atime(inode);
 864         }
 865
 866         for (i = 0; i < ap->num_pages; i++) {
 867                 struct page *page = ap->pages[i];
 868
 869                 if (!err)
 870                         SetPageUptodate(page);
 871                 else
 872                         SetPageError(page);
 873                 unlock_page(page);
 874                 put_page(page);
 875         }
 876         if (ia->ff)
 877                 fuse_file_put(ia->ff, false, false);
 878
 879         fuse_io_free(ia);
 880 }
 881
 882 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
 883 {
 884         struct fuse_file *ff = file->private_data;
 885         struct fuse_conn *fc = ff->fc;
 886         struct fuse_args_pages *ap = &ia->ap;
 887         loff_t pos = page_offset(ap->pages[0]);
 888         size_t count = ap->num_pages << PAGE_SHIFT;
 889         ssize_t res;
 890         int err;
 891
 892         ap->args.out_pages = true;
 893         ap->args.page_zeroing = true;
 894         ap->args.page_replace = true;
 895
 896         /* Don't overflow end offset */
 897         if (pos + (count - 1) == LLONG_MAX) {
 898                 count--;
 899                 ap->descs[ap->num_pages - 1].length--;
 900         }
 901         WARN_ON((loff_t) (pos + count) < 0);
 902
 903         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 904         ia->read.attr_ver = fuse_get_attr_version(fc);
 905         if (fc->async_read) {
 906                 ia->ff = fuse_file_get(ff);
 907                 ap->args.end = fuse_readpages_end;
 908                 err = fuse_simple_background(fc, &ap->args, GFP_KERNEL);
 909                 if (!err)
 910                         return;
 911         } else {
 912                 res = fuse_simple_request(fc, &ap->args);
 913                 err = res < 0 ? res : 0;
 914         }
 915         fuse_readpages_end(fc, &ap->args, err);
 916 }
 917
 918 static void fuse_readahead(struct readahead_control *rac)
 919 {
 920         struct inode *inode = rac->mapping->host;
 921         struct fuse_conn *fc = get_fuse_conn(inode);
 922         unsigned int i, max_pages, nr_pages = 0;
 923
 924         if (is_bad_inode(inode))
 925                 return;
 926
 927         max_pages = min_t(unsigned int, fc->max_pages,
 928                         fc->max_read / PAGE_SIZE);
 929
 930         for (;;) {
 931                 struct fuse_io_args *ia;
 932                 struct fuse_args_pages *ap;
 933
 934                 nr_pages = readahead_count(rac) - nr_pages;
 935                 if (nr_pages > max_pages)
 936                         nr_pages = max_pages;
 937                 if (nr_pages == 0)
 938                         break;
 939                 ia = fuse_io_alloc(NULL, nr_pages);
 940                 if (!ia)
 941                         return;
 942                 ap = &ia->ap;
 943                 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
 944                 for (i = 0; i < nr_pages; i++) {
 945                         fuse_wait_on_page_writeback(inode,
 946                                                     readahead_index(rac) + i);
 947                         ap->descs[i].length = PAGE_SIZE;
 948                 }
 949                 ap->num_pages = nr_pages;
 950                 fuse_send_readpages(ia, rac->file);
 951         }
 952 }
 953
 954 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
 955 {
 956         struct inode *inode = iocb->ki_filp->f_mapping->host;
 957         struct fuse_conn *fc = get_fuse_conn(inode);
 958
 959         /*
 960          * In auto invalidate mode, always update attributes on read.
 961          * Otherwise, only update if we attempt to read past EOF (to ensure
 962          * i_size is up to date).
 963          */
 964         if (fc->auto_inval_data ||
 965             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 966                 int err;
 967                 err = fuse_update_attributes(inode, iocb->ki_filp);
 968                 if (err)
 969                         return err;
 970         }
 971
 972         return generic_file_read_iter(iocb, to);
 973 }
 974
 975 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
 976                                  loff_t pos, size_t count)
 977 {
 978         struct fuse_args *args = &ia->ap.args;
 979
 980         ia->write.in.fh = ff->fh;
 981         ia->write.in.offset = pos;
 982         ia->write.in.size = count;
 983         args->opcode = FUSE_WRITE;
 984         args->nodeid = ff->nodeid;
 985         args->in_numargs = 2;
 986         if (ff->fc->minor < 9)
 987                 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 988         else
 989                 args->in_args[0].size = sizeof(ia->write.in);
 990         args->in_args[0].value = &ia->write.in;
 991         args->in_args[1].size = count;
 992         args->out_numargs = 1;
 993         args->out_args[0].size = sizeof(ia->write.out);
 994         args->out_args[0].value = &ia->write.out;
 995 }
 996
 997 static unsigned int fuse_write_flags(struct kiocb *iocb)
 998 {
 999         unsigned int flags = iocb->ki_filp->f_flags;
1000
1001         if (iocb->ki_flags & IOCB_DSYNC)
1002                 flags |= O_DSYNC;
1003         if (iocb->ki_flags & IOCB_SYNC)
1004                 flags |= O_SYNC;
1005
1006         return flags;
1007 }
1008
1009 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1010                                size_t count, fl_owner_t owner)
1011 {
1012         struct kiocb *iocb = ia->io->iocb;
1013         struct file *file = iocb->ki_filp;
1014         struct fuse_file *ff = file->private_data;
1015         struct fuse_conn *fc = ff->fc;
1016         struct fuse_write_in *inarg = &ia->write.in;
1017         ssize_t err;
1018
1019         fuse_write_args_fill(ia, ff, pos, count);
1020         inarg->flags = fuse_write_flags(iocb);
1021         if (owner != NULL) {
1022                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1023                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
1024         }
1025
1026         if (ia->io->async)
1027                 return fuse_async_req_send(fc, ia, count);
1028
1029         err = fuse_simple_request(fc, &ia->ap.args);
1030         if (!err && ia->write.out.size > count)
1031                 err = -EIO;
1032
1033         return err ?: ia->write.out.size;
1034 }
1035
1036 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1037 {
1038         struct fuse_conn *fc = get_fuse_conn(inode);
1039         struct fuse_inode *fi = get_fuse_inode(inode);
1040         bool ret = false;
1041
1042         spin_lock(&fi->lock);
1043         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1044         if (pos > inode->i_size) {
1045                 i_size_write(inode, pos);
1046                 ret = true;
1047         }
1048         spin_unlock(&fi->lock);
1049
1050         return ret;
1051 }
1052
1053 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1054                                      struct kiocb *iocb, struct inode *inode,
1055                                      loff_t pos, size_t count)
1056 {
1057         struct fuse_args_pages *ap = &ia->ap;
1058         struct file *file = iocb->ki_filp;
1059         struct fuse_file *ff = file->private_data;
1060         struct fuse_conn *fc = ff->fc;
1061         unsigned int offset, i;
1062         int err;
1063
1064         for (i = 0; i < ap->num_pages; i++)
1065                 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1066
1067         fuse_write_args_fill(ia, ff, pos, count);
1068         ia->write.in.flags = fuse_write_flags(iocb);
1069
1070         err = fuse_simple_request(fc, &ap->args);
1071         if (!err && ia->write.out.size > count)
1072                 err = -EIO;
1073
1074         offset = ap->descs[0].offset;
1075         count = ia->write.out.size;
1076         for (i = 0; i < ap->num_pages; i++) {
1077                 struct page *page = ap->pages[i];
1078
1079                 if (!err && !offset && count >= PAGE_SIZE)
1080                         SetPageUptodate(page);
1081
1082                 if (count > PAGE_SIZE - offset)
1083                         count -= PAGE_SIZE - offset;
1084                 else
1085                         count = 0;
1086                 offset = 0;
1087
1088                 unlock_page(page);
1089                 put_page(page);
1090         }
1091
1092         return err;
1093 }
1094
1095 static ssize_t fuse_fill_write_pages(struct fuse_args_pages *ap,
1096                                      struct address_space *mapping,
1097                                      struct iov_iter *ii, loff_t pos,
1098                                      unsigned int max_pages)
1099 {
1100         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1101         unsigned offset = pos & (PAGE_SIZE - 1);
1102         size_t count = 0;
1103         int err;
1104
1105         ap->args.in_pages = true;
1106         ap->descs[0].offset = offset;
1107
1108         do {
1109                 size_t tmp;
1110                 struct page *page;
1111                 pgoff_t index = pos >> PAGE_SHIFT;
1112                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1113                                      iov_iter_count(ii));
1114
1115                 bytes = min_t(size_t, bytes, fc->max_write - count);
1116
1117  again:
1118                 err = -EFAULT;
1119                 if (iov_iter_fault_in_readable(ii, bytes))
1120                         break;
1121
1122                 err = -ENOMEM;
1123                 page = grab_cache_page_write_begin(mapping, index, 0);
1124                 if (!page)
1125                         break;
1126
1127                 if (mapping_writably_mapped(mapping))
1128                         flush_dcache_page(page);
1129
1130                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1131                 flush_dcache_page(page);
1132
1133                 iov_iter_advance(ii, tmp);
1134                 if (!tmp) {
1135                         unlock_page(page);
1136                         put_page(page);
1137                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1138                         goto again;
1139                 }
1140
1141                 err = 0;
1142                 ap->pages[ap->num_pages] = page;
1143                 ap->descs[ap->num_pages].length = tmp;
1144                 ap->num_pages++;
1145
1146                 count += tmp;
1147                 pos += tmp;
1148                 offset += tmp;
1149                 if (offset == PAGE_SIZE)
1150                         offset = 0;
1151
1152                 if (!fc->big_writes)
1153                         break;
1154         } while (iov_iter_count(ii) && count < fc->max_write &&
1155                  ap->num_pages < max_pages && offset == 0);
1156
1157         return count > 0 ? count : err;
1158 }
1159
1160 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1161                                      unsigned int max_pages)
1162 {
1163         return min_t(unsigned int,
1164                      ((pos + len - 1) >> PAGE_SHIFT) -
1165                      (pos >> PAGE_SHIFT) + 1,
1166                      max_pages);
1167 }
1168
1169 static ssize_t fuse_perform_write(struct kiocb *iocb,
1170                                   struct address_space *mapping,
1171                                   struct iov_iter *ii, loff_t pos)
1172 {
1173         struct inode *inode = mapping->host;
1174         struct fuse_conn *fc = get_fuse_conn(inode);
1175         struct fuse_inode *fi = get_fuse_inode(inode);
1176         int err = 0;
1177         ssize_t res = 0;
1178
1179         if (inode->i_size < pos + iov_iter_count(ii))
1180                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1181
1182         do {
1183                 ssize_t count;
1184                 struct fuse_io_args ia = {};
1185                 struct fuse_args_pages *ap = &ia.ap;
1186                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1187                                                       fc->max_pages);
1188
1189                 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1190                 if (!ap->pages) {
1191                         err = -ENOMEM;
1192                         break;
1193                 }
1194
1195                 count = fuse_fill_write_pages(ap, mapping, ii, pos, nr_pages);
1196                 if (count <= 0) {
1197                         err = count;
1198                 } else {
1199                         err = fuse_send_write_pages(&ia, iocb, inode,
1200                                                     pos, count);
1201                         if (!err) {
1202                                 size_t num_written = ia.write.out.size;
1203
1204                                 res += num_written;
1205                                 pos += num_written;
1206
1207                                 /* break out of the loop on short write */
1208                                 if (num_written != count)
1209                                         err = -EIO;
1210                         }
1211                 }
1212                 kfree(ap->pages);
1213         } while (!err && iov_iter_count(ii));
1214
1215         if (res > 0)
1216                 fuse_write_update_size(inode, pos);
1217
1218         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1219         fuse_invalidate_attr(inode);
1220
1221         return res > 0 ? res : err;
1222 }
1223
1224 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1225 {
1226         struct file *file = iocb->ki_filp;
1227         struct address_space *mapping = file->f_mapping;
1228         ssize_t written = 0;
1229         ssize_t written_buffered = 0;
1230         struct inode *inode = mapping->host;
1231         ssize_t err;
1232         loff_t endbyte = 0;
1233
1234         if (get_fuse_conn(inode)->writeback_cache) {
1235                 /* Update size (EOF optimization) and mode (SUID clearing) */
1236                 err = fuse_update_attributes(mapping->host, file);
1237                 if (err)
1238                         return err;
1239
1240                 return generic_file_write_iter(iocb, from);
1241         }
1242
1243         inode_lock(inode);
1244
1245         /* We can write back this queue in page reclaim */
1246         current->backing_dev_info = inode_to_bdi(inode);
1247
1248         err = generic_write_checks(iocb, from);
1249         if (err <= 0)
1250                 goto out;
1251
1252         err = file_remove_privs(file);
1253         if (err)
1254                 goto out;
1255
1256         err = file_update_time(file);
1257         if (err)
1258                 goto out;
1259
1260         if (iocb->ki_flags & IOCB_DIRECT) {
1261                 loff_t pos = iocb->ki_pos;
1262                 written = generic_file_direct_write(iocb, from);
1263                 if (written < 0 || !iov_iter_count(from))
1264                         goto out;
1265
1266                 pos += written;
1267
1268                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1269                 if (written_buffered < 0) {
1270                         err = written_buffered;
1271                         goto out;
1272                 }
1273                 endbyte = pos + written_buffered - 1;
1274
1275                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1276                                                    endbyte);
1277                 if (err)
1278                         goto out;
1279
1280                 invalidate_mapping_pages(file->f_mapping,
1281                                          pos >> PAGE_SHIFT,
1282                                          endbyte >> PAGE_SHIFT);
1283
1284                 written += written_buffered;
1285                 iocb->ki_pos = pos + written_buffered;
1286         } else {
1287                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1288                 if (written >= 0)
1289                         iocb->ki_pos += written;
1290         }
1291 out:
1292         current->backing_dev_info = NULL;
1293         inode_unlock(inode);
1294         if (written > 0)
1295                 written = generic_write_sync(iocb, written);
1296
1297         return written ? written : err;
1298 }
1299
1300 static inline void fuse_page_descs_length_init(struct fuse_page_desc *descs,
1301                                                unsigned int index,
1302                                                unsigned int nr_pages)
1303 {
1304         int i;
1305
1306         for (i = index; i < index + nr_pages; i++)
1307                 descs[i].length = PAGE_SIZE - descs[i].offset;
1308 }
1309
1310 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1311 {
1312         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1313 }
1314
1315 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1316                                         size_t max_size)
1317 {
1318         return min(iov_iter_single_seg_count(ii), max_size);
1319 }
1320
1321 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1322                                size_t *nbytesp, int write,
1323                                unsigned int max_pages)
1324 {
1325         size_t nbytes = 0;  /* # bytes already packed in req */
1326         ssize_t ret = 0;
1327
1328         /* Special case for kernel I/O: can copy directly into the buffer */
1329         if (iov_iter_is_kvec(ii)) {
1330                 unsigned long user_addr = fuse_get_user_addr(ii);
1331                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1332
1333                 if (write)
1334                         ap->args.in_args[1].value = (void *) user_addr;
1335                 else
1336                         ap->args.out_args[0].value = (void *) user_addr;
1337
1338                 iov_iter_advance(ii, frag_size);
1339                 *nbytesp = frag_size;
1340                 return 0;
1341         }
1342
1343         while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1344                 unsigned npages;
1345                 size_t start;
1346                 ret = iov_iter_get_pages(ii, &ap->pages[ap->num_pages],
1347                                         *nbytesp - nbytes,
1348                                         max_pages - ap->num_pages,
1349                                         &start);
1350                 if (ret < 0)
1351                         break;
1352
1353                 iov_iter_advance(ii, ret);
1354                 nbytes += ret;
1355
1356                 ret += start;
1357                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1358
1359                 ap->descs[ap->num_pages].offset = start;
1360                 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1361
1362                 ap->num_pages += npages;
1363                 ap->descs[ap->num_pages - 1].length -=
1364                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1365         }
1366
1367         if (write)
1368                 ap->args.in_pages = true;
1369         else
1370                 ap->args.out_pages = true;
1371
1372         *nbytesp = nbytes;
1373
1374         return ret < 0 ? ret : 0;
1375 }
1376
1377 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1378                        loff_t *ppos, int flags)
1379 {
1380         int write = flags & FUSE_DIO_WRITE;
1381         int cuse = flags & FUSE_DIO_CUSE;
1382         struct file *file = io->iocb->ki_filp;
1383         struct inode *inode = file->f_mapping->host;
1384         struct fuse_file *ff = file->private_data;
1385         struct fuse_conn *fc = ff->fc;
1386         size_t nmax = write ? fc->max_write : fc->max_read;
1387         loff_t pos = *ppos;
1388         size_t count = iov_iter_count(iter);
1389         pgoff_t idx_from = pos >> PAGE_SHIFT;
1390         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1391         ssize_t res = 0;
1392         int err = 0;
1393         struct fuse_io_args *ia;
1394         unsigned int max_pages;
1395
1396         max_pages = iov_iter_npages(iter, fc->max_pages);
1397         ia = fuse_io_alloc(io, max_pages);
1398         if (!ia)
1399                 return -ENOMEM;
1400
1401         ia->io = io;
1402         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1403                 if (!write)
1404                         inode_lock(inode);
1405                 fuse_sync_writes(inode);
1406                 if (!write)
1407                         inode_unlock(inode);
1408         }
1409
1410         io->should_dirty = !write && iter_is_iovec(iter);
1411         while (count) {
1412                 ssize_t nres;
1413                 fl_owner_t owner = current->files;
1414                 size_t nbytes = min(count, nmax);
1415
1416                 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1417                                           max_pages);
1418                 if (err && !nbytes)
1419                         break;
1420
1421                 if (write) {
1422                         if (!capable(CAP_FSETID))
1423                                 ia->write.in.write_flags |= FUSE_WRITE_KILL_PRIV;
1424
1425                         nres = fuse_send_write(ia, pos, nbytes, owner);
1426                 } else {
1427                         nres = fuse_send_read(ia, pos, nbytes, owner);
1428                 }
1429
1430                 if (!io->async || nres < 0) {
1431                         fuse_release_user_pages(&ia->ap, io->should_dirty);
1432                         fuse_io_free(ia);
1433                 }
1434                 ia = NULL;
1435                 if (nres < 0) {
1436                         iov_iter_revert(iter, nbytes);
1437                         err = nres;
1438                         break;
1439                 }
1440                 WARN_ON(nres > nbytes);
1441
1442                 count -= nres;
1443                 res += nres;
1444                 pos += nres;
1445                 if (nres != nbytes) {
1446                         iov_iter_revert(iter, nbytes - nres);
1447                         break;
1448                 }
1449                 if (count) {
1450                         max_pages = iov_iter_npages(iter, fc->max_pages);
1451                         ia = fuse_io_alloc(io, max_pages);
1452                         if (!ia)
1453                                 break;
1454                 }
1455         }
1456         if (ia)
1457                 fuse_io_free(ia);
1458         if (res > 0)
1459                 *ppos = pos;
1460
1461         return res > 0 ? res : err;
1462 }
1463 EXPORT_SYMBOL_GPL(fuse_direct_io);
1464
1465 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1466                                   struct iov_iter *iter,
1467                                   loff_t *ppos)
1468 {
1469         ssize_t res;
1470         struct inode *inode = file_inode(io->iocb->ki_filp);
1471
1472         res = fuse_direct_io(io, iter, ppos, 0);
1473
1474         fuse_invalidate_atime(inode);
1475
1476         return res;
1477 }
1478
1479 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1480
1481 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1482 {
1483         ssize_t res;
1484
1485         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1486                 res = fuse_direct_IO(iocb, to);
1487         } else {
1488                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1489
1490                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1491         }
1492
1493         return res;
1494 }
1495
1496 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1497 {
1498         struct inode *inode = file_inode(iocb->ki_filp);
1499         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1500         ssize_t res;
1501
1502         /* Don't allow parallel writes to the same file */
1503         inode_lock(inode);
1504         res = generic_write_checks(iocb, from);
1505         if (res > 0) {
1506                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1507                         res = fuse_direct_IO(iocb, from);
1508                 } else {
1509                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1510                                              FUSE_DIO_WRITE);
1511                 }
1512         }
1513         fuse_invalidate_attr(inode);
1514         if (res > 0)
1515                 fuse_write_update_size(inode, iocb->ki_pos);
1516         inode_unlock(inode);
1517
1518         return res;
1519 }
1520
1521 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1522 {
1523         struct file *file = iocb->ki_filp;
1524         struct fuse_file *ff = file->private_data;
1525
1526         if (is_bad_inode(file_inode(file)))
1527                 return -EIO;
1528
1529         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1530                 return fuse_cache_read_iter(iocb, to);
1531         else
1532                 return fuse_direct_read_iter(iocb, to);
1533 }
1534
1535 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1536 {
1537         struct file *file = iocb->ki_filp;
1538         struct fuse_file *ff = file->private_data;
1539
1540         if (is_bad_inode(file_inode(file)))
1541                 return -EIO;
1542
1543         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1544                 return fuse_cache_write_iter(iocb, from);
1545         else
1546                 return fuse_direct_write_iter(iocb, from);
1547 }
1548
1549 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1550 {
1551         struct fuse_args_pages *ap = &wpa->ia.ap;
1552         int i;
1553
1554         for (i = 0; i < ap->num_pages; i++)
1555                 __free_page(ap->pages[i]);
1556
1557         if (wpa->ia.ff)
1558                 fuse_file_put(wpa->ia.ff, false, false);
1559
1560         kfree(ap->pages);
1561         kfree(wpa);
1562 }
1563
1564 static void fuse_writepage_finish(struct fuse_conn *fc,
1565                                   struct fuse_writepage_args *wpa)
1566 {
1567         struct fuse_args_pages *ap = &wpa->ia.ap;
1568         struct inode *inode = wpa->inode;
1569         struct fuse_inode *fi = get_fuse_inode(inode);
1570         struct backing_dev_info *bdi = inode_to_bdi(inode);
1571         int i;
1572
1573         list_del(&wpa->writepages_entry);
1574         for (i = 0; i < ap->num_pages; i++) {
1575                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1576                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1577                 wb_writeout_inc(&bdi->wb);
1578         }
1579         wake_up(&fi->page_waitq);
1580 }
1581
1582 /* Called under fi->lock, may release and reacquire it */
1583 static void fuse_send_writepage(struct fuse_conn *fc,
1584                                 struct fuse_writepage_args *wpa, loff_t size)
1585 __releases(fi->lock)
1586 __acquires(fi->lock)
1587 {
1588         struct fuse_writepage_args *aux, *next;
1589         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1590         struct fuse_write_in *inarg = &wpa->ia.write.in;
1591         struct fuse_args *args = &wpa->ia.ap.args;
1592         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1593         int err;
1594
1595         fi->writectr++;
1596         if (inarg->offset + data_size <= size) {
1597                 inarg->size = data_size;
1598         } else if (inarg->offset < size) {
1599                 inarg->size = size - inarg->offset;
1600         } else {
1601                 /* Got truncated off completely */
1602                 goto out_free;
1603         }
1604
1605         args->in_args[1].size = inarg->size;
1606         args->force = true;
1607         args->nocreds = true;
1608
1609         err = fuse_simple_background(fc, args, GFP_ATOMIC);
1610         if (err == -ENOMEM) {
1611                 spin_unlock(&fi->lock);
1612                 err = fuse_simple_background(fc, args, GFP_NOFS | __GFP_NOFAIL);
1613                 spin_lock(&fi->lock);
1614         }
1615
1616         /* Fails on broken connection only */
1617         if (unlikely(err))
1618                 goto out_free;
1619
1620         return;
1621
1622  out_free:
1623         fi->writectr--;
1624         fuse_writepage_finish(fc, wpa);
1625         spin_unlock(&fi->lock);
1626
1627         /* After fuse_writepage_finish() aux request list is private */
1628         for (aux = wpa->next; aux; aux = next) {
1629                 next = aux->next;
1630                 aux->next = NULL;
1631                 fuse_writepage_free(aux);
1632         }
1633
1634         fuse_writepage_free(wpa);
1635         spin_lock(&fi->lock);
1636 }
1637
1638 /*
1639  * If fi->writectr is positive (no truncate or fsync going on) send
1640  * all queued writepage requests.
1641  *
1642  * Called with fi->lock
1643  */
1644 void fuse_flush_writepages(struct inode *inode)
1645 __releases(fi->lock)
1646 __acquires(fi->lock)
1647 {
1648         struct fuse_conn *fc = get_fuse_conn(inode);
1649         struct fuse_inode *fi = get_fuse_inode(inode);
1650         loff_t crop = i_size_read(inode);
1651         struct fuse_writepage_args *wpa;
1652
1653         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1654                 wpa = list_entry(fi->queued_writes.next,
1655                                  struct fuse_writepage_args, queue_entry);
1656                 list_del_init(&wpa->queue_entry);
1657                 fuse_send_writepage(fc, wpa, crop);
1658         }
1659 }
1660
1661 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_args *args,
1662                                int error)
1663 {
1664         struct fuse_writepage_args *wpa =
1665                 container_of(args, typeof(*wpa), ia.ap.args);
1666         struct inode *inode = wpa->inode;
1667         struct fuse_inode *fi = get_fuse_inode(inode);
1668
1669         mapping_set_error(inode->i_mapping, error);
1670         spin_lock(&fi->lock);
1671         while (wpa->next) {
1672                 struct fuse_conn *fc = get_fuse_conn(inode);
1673                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1674                 struct fuse_writepage_args *next = wpa->next;
1675
1676                 wpa->next = next->next;
1677                 next->next = NULL;
1678                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1679                 list_add(&next->writepages_entry, &fi->writepages);
1680
1681                 /*
1682                  * Skip fuse_flush_writepages() to make it easy to crop requests
1683                  * based on primary request size.
1684                  *
1685                  * 1st case (trivial): there are no concurrent activities using
1686                  * fuse_set/release_nowrite.  Then we're on safe side because
1687                  * fuse_flush_writepages() would call fuse_send_writepage()
1688                  * anyway.
1689                  *
1690                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1691                  * now for completion of all in-flight requests.  This happens
1692                  * rarely and no more than once per page, so this should be
1693                  * okay.
1694                  *
1695                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1696                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1697                  * that fuse_set_nowrite returned implies that all in-flight
1698                  * requests were completed along with all of their secondary
1699                  * requests.  Further primary requests are blocked by negative
1700                  * writectr.  Hence there cannot be any in-flight requests and
1701                  * no invocations of fuse_writepage_end() while we're in
1702                  * fuse_set_nowrite..fuse_release_nowrite section.
1703                  */
1704                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1705         }
1706         fi->writectr--;
1707         fuse_writepage_finish(fc, wpa);
1708         spin_unlock(&fi->lock);
1709         fuse_writepage_free(wpa);
1710 }
1711
1712 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1713                                                struct fuse_inode *fi)
1714 {
1715         struct fuse_file *ff = NULL;
1716
1717         spin_lock(&fi->lock);
1718         if (!list_empty(&fi->write_files)) {
1719                 ff = list_entry(fi->write_files.next, struct fuse_file,
1720                                 write_entry);
1721                 fuse_file_get(ff);
1722         }
1723         spin_unlock(&fi->lock);
1724
1725         return ff;
1726 }
1727
1728 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1729                                              struct fuse_inode *fi)
1730 {
1731         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1732         WARN_ON(!ff);
1733         return ff;
1734 }
1735
1736 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1737 {
1738         struct fuse_conn *fc = get_fuse_conn(inode);
1739         struct fuse_inode *fi = get_fuse_inode(inode);
1740         struct fuse_file *ff;
1741         int err;
1742
1743         ff = __fuse_write_file_get(fc, fi);
1744         err = fuse_flush_times(inode, ff);
1745         if (ff)
1746                 fuse_file_put(ff, false, false);
1747
1748         return err;
1749 }
1750
1751 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1752 {
1753         struct fuse_writepage_args *wpa;
1754         struct fuse_args_pages *ap;
1755
1756         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1757         if (wpa) {
1758                 ap = &wpa->ia.ap;
1759                 ap->num_pages = 0;
1760                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1761                 if (!ap->pages) {
1762                         kfree(wpa);
1763                         wpa = NULL;
1764                 }
1765         }
1766         return wpa;
1767
1768 }
1769
1770 static int fuse_writepage_locked(struct page *page)
1771 {
1772         struct address_space *mapping = page->mapping;
1773         struct inode *inode = mapping->host;
1774         struct fuse_conn *fc = get_fuse_conn(inode);
1775         struct fuse_inode *fi = get_fuse_inode(inode);
1776         struct fuse_writepage_args *wpa;
1777         struct fuse_args_pages *ap;
1778         struct page *tmp_page;
1779         int error = -ENOMEM;
1780
1781         set_page_writeback(page);
1782
1783         wpa = fuse_writepage_args_alloc();
1784         if (!wpa)
1785                 goto err;
1786         ap = &wpa->ia.ap;
1787
1788         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1789         if (!tmp_page)
1790                 goto err_free;
1791
1792         error = -EIO;
1793         wpa->ia.ff = fuse_write_file_get(fc, fi);
1794         if (!wpa->ia.ff)
1795                 goto err_nofile;
1796
1797         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1798
1799         copy_highpage(tmp_page, page);
1800         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1801         wpa->next = NULL;
1802         ap->args.in_pages = true;
1803         ap->num_pages = 1;
1804         ap->pages[0] = tmp_page;
1805         ap->descs[0].offset = 0;
1806         ap->descs[0].length = PAGE_SIZE;
1807         ap->args.end = fuse_writepage_end;
1808         wpa->inode = inode;
1809
1810         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1811         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1812
1813         spin_lock(&fi->lock);
1814         list_add(&wpa->writepages_entry, &fi->writepages);
1815         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1816         fuse_flush_writepages(inode);
1817         spin_unlock(&fi->lock);
1818
1819         end_page_writeback(page);
1820
1821         return 0;
1822
1823 err_nofile:
1824         __free_page(tmp_page);
1825 err_free:
1826         kfree(wpa);
1827 err:
1828         mapping_set_error(page->mapping, error);
1829         end_page_writeback(page);
1830         return error;
1831 }
1832
1833 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1834 {
1835         int err;
1836
1837         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1838                 /*
1839                  * ->writepages() should be called for sync() and friends.  We
1840                  * should only get here on direct reclaim and then we are
1841                  * allowed to skip a page which is already in flight
1842                  */
1843                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1844
1845                 redirty_page_for_writepage(wbc, page);
1846                 unlock_page(page);
1847                 return 0;
1848         }
1849
1850         err = fuse_writepage_locked(page);
1851         unlock_page(page);
1852
1853         return err;
1854 }
1855
1856 struct fuse_fill_wb_data {
1857         struct fuse_writepage_args *wpa;
1858         struct fuse_file *ff;
1859         struct inode *inode;
1860         struct page **orig_pages;
1861         unsigned int max_pages;
1862 };
1863
1864 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
1865 {
1866         struct fuse_args_pages *ap = &data->wpa->ia.ap;
1867         struct fuse_conn *fc = get_fuse_conn(data->inode);
1868         struct page **pages;
1869         struct fuse_page_desc *descs;
1870         unsigned int npages = min_t(unsigned int,
1871                                     max_t(unsigned int, data->max_pages * 2,
1872                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
1873                                     fc->max_pages);
1874         WARN_ON(npages <= data->max_pages);
1875
1876         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
1877         if (!pages)
1878                 return false;
1879
1880         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
1881         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
1882         kfree(ap->pages);
1883         ap->pages = pages;
1884         ap->descs = descs;
1885         data->max_pages = npages;
1886
1887         return true;
1888 }
1889
1890 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1891 {
1892         struct fuse_writepage_args *wpa = data->wpa;
1893         struct inode *inode = data->inode;
1894         struct fuse_inode *fi = get_fuse_inode(inode);
1895         int num_pages = wpa->ia.ap.num_pages;
1896         int i;
1897
1898         wpa->ia.ff = fuse_file_get(data->ff);
1899         spin_lock(&fi->lock);
1900         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1901         fuse_flush_writepages(inode);
1902         spin_unlock(&fi->lock);
1903
1904         for (i = 0; i < num_pages; i++)
1905                 end_page_writeback(data->orig_pages[i]);
1906 }
1907
1908 /*
1909  * First recheck under fi->lock if the offending offset is still under
1910  * writeback.  If yes, then iterate auxiliary write requests, to see if there's
1911  * one already added for a page at this offset.  If there's none, then insert
1912  * this new request onto the auxiliary list, otherwise reuse the existing one by
1913  * copying the new page contents over to the old temporary page.
1914  */
1915 static bool fuse_writepage_in_flight(struct fuse_writepage_args *new_wpa,
1916                                      struct page *page)
1917 {
1918         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
1919         struct fuse_writepage_args *tmp;
1920         struct fuse_writepage_args *old_wpa;
1921         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
1922
1923         WARN_ON(new_ap->num_pages != 0);
1924
1925         spin_lock(&fi->lock);
1926         list_del(&new_wpa->writepages_entry);
1927         old_wpa = fuse_find_writeback(fi, page->index, page->index);
1928         if (!old_wpa) {
1929                 list_add(&new_wpa->writepages_entry, &fi->writepages);
1930                 spin_unlock(&fi->lock);
1931                 return false;
1932         }
1933
1934         new_ap->num_pages = 1;
1935         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
1936                 pgoff_t curr_index;
1937
1938                 WARN_ON(tmp->inode != new_wpa->inode);
1939                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
1940                 if (curr_index == page->index) {
1941                         WARN_ON(tmp->ia.ap.num_pages != 1);
1942                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
1943                         break;
1944                 }
1945         }
1946
1947         if (!tmp) {
1948                 new_wpa->next = old_wpa->next;
1949                 old_wpa->next = new_wpa;
1950         }
1951
1952         spin_unlock(&fi->lock);
1953
1954         if (tmp) {
1955                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
1956
1957                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1958                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
1959                 wb_writeout_inc(&bdi->wb);
1960                 fuse_writepage_free(new_wpa);
1961         }
1962
1963         return true;
1964 }
1965
1966 static int fuse_writepages_fill(struct page *page,
1967                 struct writeback_control *wbc, void *_data)
1968 {
1969         struct fuse_fill_wb_data *data = _data;
1970         struct fuse_writepage_args *wpa = data->wpa;
1971         struct fuse_args_pages *ap = &wpa->ia.ap;
1972         struct inode *inode = data->inode;
1973         struct fuse_inode *fi = get_fuse_inode(inode);
1974         struct fuse_conn *fc = get_fuse_conn(inode);
1975         struct page *tmp_page;
1976         bool is_writeback;
1977         int err;
1978
1979         if (!data->ff) {
1980                 err = -EIO;
1981                 data->ff = fuse_write_file_get(fc, fi);
1982                 if (!data->ff)
1983                         goto out_unlock;
1984         }
1985
1986         /*
1987          * Being under writeback is unlikely but possible.  For example direct
1988          * read to an mmaped fuse file will set the page dirty twice; once when
1989          * the pages are faulted with get_user_pages(), and then after the read
1990          * completed.
1991          */
1992         is_writeback = fuse_page_is_writeback(inode, page->index);
1993
1994         if (wpa && ap->num_pages &&
1995             (is_writeback || ap->num_pages == fc->max_pages ||
1996              (ap->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1997              data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)) {
1998                 fuse_writepages_send(data);
1999                 data->wpa = NULL;
2000         } else if (wpa && ap->num_pages == data->max_pages) {
2001                 if (!fuse_pages_realloc(data)) {
2002                         fuse_writepages_send(data);
2003                         data->wpa = NULL;
2004                 }
2005         }
2006
2007         err = -ENOMEM;
2008         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2009         if (!tmp_page)
2010                 goto out_unlock;
2011
2012         /*
2013          * The page must not be redirtied until the writeout is completed
2014          * (i.e. userspace has sent a reply to the write request).  Otherwise
2015          * there could be more than one temporary page instance for each real
2016          * page.
2017          *
2018          * This is ensured by holding the page lock in page_mkwrite() while
2019          * checking fuse_page_is_writeback().  We already hold the page lock
2020          * since clear_page_dirty_for_io() and keep it held until we add the
2021          * request to the fi->writepages list and increment ap->num_pages.
2022          * After this fuse_page_is_writeback() will indicate that the page is
2023          * under writeback, so we can release the page lock.
2024          */
2025         if (data->wpa == NULL) {
2026                 err = -ENOMEM;
2027                 wpa = fuse_writepage_args_alloc();
2028                 if (!wpa) {
2029                         __free_page(tmp_page);
2030                         goto out_unlock;
2031                 }
2032                 data->max_pages = 1;
2033
2034                 ap = &wpa->ia.ap;
2035                 fuse_write_args_fill(&wpa->ia, data->ff, page_offset(page), 0);
2036                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2037                 wpa->next = NULL;
2038                 ap->args.in_pages = true;
2039                 ap->args.end = fuse_writepage_end;
2040                 ap->num_pages = 0;
2041                 wpa->inode = inode;
2042
2043                 spin_lock(&fi->lock);
2044                 list_add(&wpa->writepages_entry, &fi->writepages);
2045                 spin_unlock(&fi->lock);
2046
2047                 data->wpa = wpa;
2048         }
2049         set_page_writeback(page);
2050
2051         copy_highpage(tmp_page, page);
2052         ap->pages[ap->num_pages] = tmp_page;
2053         ap->descs[ap->num_pages].offset = 0;
2054         ap->descs[ap->num_pages].length = PAGE_SIZE;
2055
2056         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2057         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2058
2059         err = 0;
2060         if (is_writeback && fuse_writepage_in_flight(wpa, page)) {
2061                 end_page_writeback(page);
2062                 data->wpa = NULL;
2063                 goto out_unlock;
2064         }
2065         data->orig_pages[ap->num_pages] = page;
2066
2067         /*
2068          * Protected by fi->lock against concurrent access by
2069          * fuse_page_is_writeback().
2070          */
2071         spin_lock(&fi->lock);
2072         ap->num_pages++;
2073         spin_unlock(&fi->lock);
2074
2075 out_unlock:
2076         unlock_page(page);
2077
2078         return err;
2079 }
2080
2081 static int fuse_writepages(struct address_space *mapping,
2082                            struct writeback_control *wbc)
2083 {
2084         struct inode *inode = mapping->host;
2085         struct fuse_conn *fc = get_fuse_conn(inode);
2086         struct fuse_fill_wb_data data;
2087         int err;
2088
2089         err = -EIO;
2090         if (is_bad_inode(inode))
2091                 goto out;
2092
2093         data.inode = inode;
2094         data.wpa = NULL;
2095         data.ff = NULL;
2096
2097         err = -ENOMEM;
2098         data.orig_pages = kcalloc(fc->max_pages,
2099                                   sizeof(struct page *),
2100                                   GFP_NOFS);
2101         if (!data.orig_pages)
2102                 goto out;
2103
2104         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2105         if (data.wpa) {
2106                 /* Ignore errors if we can write at least one page */
2107                 WARN_ON(!data.wpa->ia.ap.num_pages);
2108                 fuse_writepages_send(&data);
2109                 err = 0;
2110         }
2111         if (data.ff)
2112                 fuse_file_put(data.ff, false, false);
2113
2114         kfree(data.orig_pages);
2115 out:
2116         return err;
2117 }
2118
2119 /*
2120  * It's worthy to make sure that space is reserved on disk for the write,
2121  * but how to implement it without killing performance need more thinking.
2122  */
2123 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2124                 loff_t pos, unsigned len, unsigned flags,
2125                 struct page **pagep, void **fsdata)
2126 {
2127         pgoff_t index = pos >> PAGE_SHIFT;
2128         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2129         struct page *page;
2130         loff_t fsize;
2131         int err = -ENOMEM;
2132
2133         WARN_ON(!fc->writeback_cache);
2134
2135         page = grab_cache_page_write_begin(mapping, index, flags);
2136         if (!page)
2137                 goto error;
2138
2139         fuse_wait_on_page_writeback(mapping->host, page->index);
2140
2141         if (PageUptodate(page) || len == PAGE_SIZE)
2142                 goto success;
2143         /*
2144          * Check if the start this page comes after the end of file, in which
2145          * case the readpage can be optimized away.
2146          */
2147         fsize = i_size_read(mapping->host);
2148         if (fsize <= (pos & PAGE_MASK)) {
2149                 size_t off = pos & ~PAGE_MASK;
2150                 if (off)
2151                         zero_user_segment(page, 0, off);
2152                 goto success;
2153         }
2154         err = fuse_do_readpage(file, page);
2155         if (err)
2156                 goto cleanup;
2157 success:
2158         *pagep = page;
2159         return 0;
2160
2161 cleanup:
2162         unlock_page(page);
2163         put_page(page);
2164 error:
2165         return err;
2166 }
2167
2168 static int fuse_write_end(struct file *file, struct address_space *mapping,
2169                 loff_t pos, unsigned len, unsigned copied,
2170                 struct page *page, void *fsdata)
2171 {
2172         struct inode *inode = page->mapping->host;
2173
2174         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2175         if (!copied)
2176                 goto unlock;
2177
2178         if (!PageUptodate(page)) {
2179                 /* Zero any unwritten bytes at the end of the page */
2180                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2181                 if (endoff)
2182                         zero_user_segment(page, endoff, PAGE_SIZE);
2183                 SetPageUptodate(page);
2184         }
2185
2186         fuse_write_update_size(inode, pos + copied);
2187         set_page_dirty(page);
2188
2189 unlock:
2190         unlock_page(page);
2191         put_page(page);
2192
2193         return copied;
2194 }
2195
2196 static int fuse_launder_page(struct page *page)
2197 {
2198         int err = 0;
2199         if (clear_page_dirty_for_io(page)) {
2200                 struct inode *inode = page->mapping->host;
2201                 err = fuse_writepage_locked(page);
2202                 if (!err)
2203                         fuse_wait_on_page_writeback(inode, page->index);
2204         }
2205         return err;
2206 }
2207
2208 /*
2209  * Write back dirty pages now, because there may not be any suitable
2210  * open files later
2211  */
2212 static void fuse_vma_close(struct vm_area_struct *vma)
2213 {
2214         filemap_write_and_wait(vma->vm_file->f_mapping);
2215 }
2216
2217 /*
2218  * Wait for writeback against this page to complete before allowing it
2219  * to be marked dirty again, and hence written back again, possibly
2220  * before the previous writepage completed.
2221  *
2222  * Block here, instead of in ->writepage(), so that the userspace fs
2223  * can only block processes actually operating on the filesystem.
2224  *
2225  * Otherwise unprivileged userspace fs would be able to block
2226  * unrelated:
2227  *
2228  * - page migration
2229  * - sync(2)
2230  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2231  */
2232 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2233 {
2234         struct page *page = vmf->page;
2235         struct inode *inode = file_inode(vmf->vma->vm_file);
2236
2237         file_update_time(vmf->vma->vm_file);
2238         lock_page(page);
2239         if (page->mapping != inode->i_mapping) {
2240                 unlock_page(page);
2241                 return VM_FAULT_NOPAGE;
2242         }
2243
2244         fuse_wait_on_page_writeback(inode, page->index);
2245         return VM_FAULT_LOCKED;
2246 }
2247
2248 static const struct vm_operations_struct fuse_file_vm_ops = {
2249         .close          = fuse_vma_close,
2250         .fault          = filemap_fault,
2251         .map_pages      = filemap_map_pages,
2252         .page_mkwrite   = fuse_page_mkwrite,
2253 };
2254
2255 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2256 {
2257         struct fuse_file *ff = file->private_data;
2258
2259         if (ff->open_flags & FOPEN_DIRECT_IO) {
2260                 /* Can't provide the coherency needed for MAP_SHARED */
2261                 if (vma->vm_flags & VM_MAYSHARE)
2262                         return -ENODEV;
2263
2264                 invalidate_inode_pages2(file->f_mapping);
2265
2266                 return generic_file_mmap(file, vma);
2267         }
2268
2269         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2270                 fuse_link_write_file(file);
2271
2272         file_accessed(file);
2273         vma->vm_ops = &fuse_file_vm_ops;
2274         return 0;
2275 }
2276
2277 static int convert_fuse_file_lock(struct fuse_conn *fc,
2278                                   const struct fuse_file_lock *ffl,
2279                                   struct file_lock *fl)
2280 {
2281         switch (ffl->type) {
2282         case F_UNLCK:
2283                 break;
2284
2285         case F_RDLCK:
2286         case F_WRLCK:
2287                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2288                     ffl->end < ffl->start)
2289                         return -EIO;
2290
2291                 fl->fl_start = ffl->start;
2292                 fl->fl_end = ffl->end;
2293
2294                 /*
2295                  * Convert pid into init's pid namespace.  The locks API will
2296                  * translate it into the caller's pid namespace.
2297                  */
2298                 rcu_read_lock();
2299                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2300                 rcu_read_unlock();
2301                 break;
2302
2303         default:
2304                 return -EIO;
2305         }
2306         fl->fl_type = ffl->type;
2307         return 0;
2308 }
2309
2310 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2311                          const struct file_lock *fl, int opcode, pid_t pid,
2312                          int flock, struct fuse_lk_in *inarg)
2313 {
2314         struct inode *inode = file_inode(file);
2315         struct fuse_conn *fc = get_fuse_conn(inode);
2316         struct fuse_file *ff = file->private_data;
2317
2318         memset(inarg, 0, sizeof(*inarg));
2319         inarg->fh = ff->fh;
2320         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2321         inarg->lk.start = fl->fl_start;
2322         inarg->lk.end = fl->fl_end;
2323         inarg->lk.type = fl->fl_type;
2324         inarg->lk.pid = pid;
2325         if (flock)
2326                 inarg->lk_flags |= FUSE_LK_FLOCK;
2327         args->opcode = opcode;
2328         args->nodeid = get_node_id(inode);
2329         args->in_numargs = 1;
2330         args->in_args[0].size = sizeof(*inarg);
2331         args->in_args[0].value = inarg;
2332 }
2333
2334 static int fuse_getlk(struct file *file, struct file_lock *fl)
2335 {
2336         struct inode *inode = file_inode(file);
2337         struct fuse_conn *fc = get_fuse_conn(inode);
2338         FUSE_ARGS(args);
2339         struct fuse_lk_in inarg;
2340         struct fuse_lk_out outarg;
2341         int err;
2342
2343         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2344         args.out_numargs = 1;
2345         args.out_args[0].size = sizeof(outarg);
2346         args.out_args[0].value = &outarg;
2347         err = fuse_simple_request(fc, &args);
2348         if (!err)
2349                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2350
2351         return err;
2352 }
2353
2354 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2355 {
2356         struct inode *inode = file_inode(file);
2357         struct fuse_conn *fc = get_fuse_conn(inode);
2358         FUSE_ARGS(args);
2359         struct fuse_lk_in inarg;
2360         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2361         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2362         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2363         int err;
2364
2365         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2366                 /* NLM needs asynchronous locks, which we don't support yet */
2367                 return -ENOLCK;
2368         }
2369
2370         /* Unlock on close is handled by the flush method */
2371         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2372                 return 0;
2373
2374         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2375         err = fuse_simple_request(fc, &args);
2376
2377         /* locking is restartable */
2378         if (err == -EINTR)
2379                 err = -ERESTARTSYS;
2380
2381         return err;
2382 }
2383
2384 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2385 {
2386         struct inode *inode = file_inode(file);
2387         struct fuse_conn *fc = get_fuse_conn(inode);
2388         int err;
2389
2390         if (cmd == F_CANCELLK) {
2391                 err = 0;
2392         } else if (cmd == F_GETLK) {
2393                 if (fc->no_lock) {
2394                         posix_test_lock(file, fl);
2395                         err = 0;
2396                 } else
2397                         err = fuse_getlk(file, fl);
2398         } else {
2399                 if (fc->no_lock)
2400                         err = posix_lock_file(file, fl, NULL);
2401                 else
2402                         err = fuse_setlk(file, fl, 0);
2403         }
2404         return err;
2405 }
2406
2407 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2408 {
2409         struct inode *inode = file_inode(file);
2410         struct fuse_conn *fc = get_fuse_conn(inode);
2411         int err;
2412
2413         if (fc->no_flock) {
2414                 err = locks_lock_file_wait(file, fl);
2415         } else {
2416                 struct fuse_file *ff = file->private_data;
2417
2418                 /* emulate flock with POSIX locks */
2419                 ff->flock = true;
2420                 err = fuse_setlk(file, fl, 1);
2421         }
2422
2423         return err;
2424 }
2425
2426 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2427 {
2428         struct inode *inode = mapping->host;
2429         struct fuse_conn *fc = get_fuse_conn(inode);
2430         FUSE_ARGS(args);
2431         struct fuse_bmap_in inarg;
2432         struct fuse_bmap_out outarg;
2433         int err;
2434
2435         if (!inode->i_sb->s_bdev || fc->no_bmap)
2436                 return 0;
2437
2438         memset(&inarg, 0, sizeof(inarg));
2439         inarg.block = block;
2440         inarg.blocksize = inode->i_sb->s_blocksize;
2441         args.opcode = FUSE_BMAP;
2442         args.nodeid = get_node_id(inode);
2443         args.in_numargs = 1;
2444         args.in_args[0].size = sizeof(inarg);
2445         args.in_args[0].value = &inarg;
2446         args.out_numargs = 1;
2447         args.out_args[0].size = sizeof(outarg);
2448         args.out_args[0].value = &outarg;
2449         err = fuse_simple_request(fc, &args);
2450         if (err == -ENOSYS)
2451                 fc->no_bmap = 1;
2452
2453         return err ? 0 : outarg.block;
2454 }
2455
2456 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2457 {
2458         struct inode *inode = file->f_mapping->host;
2459         struct fuse_conn *fc = get_fuse_conn(inode);
2460         struct fuse_file *ff = file->private_data;
2461         FUSE_ARGS(args);
2462         struct fuse_lseek_in inarg = {
2463                 .fh = ff->fh,
2464                 .offset = offset,
2465                 .whence = whence
2466         };
2467         struct fuse_lseek_out outarg;
2468         int err;
2469
2470         if (fc->no_lseek)
2471                 goto fallback;
2472
2473         args.opcode = FUSE_LSEEK;
2474         args.nodeid = ff->nodeid;
2475         args.in_numargs = 1;
2476         args.in_args[0].size = sizeof(inarg);
2477         args.in_args[0].value = &inarg;
2478         args.out_numargs = 1;
2479         args.out_args[0].size = sizeof(outarg);
2480         args.out_args[0].value = &outarg;
2481         err = fuse_simple_request(fc, &args);
2482         if (err) {
2483                 if (err == -ENOSYS) {
2484                         fc->no_lseek = 1;
2485                         goto fallback;
2486                 }
2487                 return err;
2488         }
2489
2490         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2491
2492 fallback:
2493         err = fuse_update_attributes(inode, file);
2494         if (!err)
2495                 return generic_file_llseek(file, offset, whence);
2496         else
2497                 return err;
2498 }
2499
2500 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2501 {
2502         loff_t retval;
2503         struct inode *inode = file_inode(file);
2504
2505         switch (whence) {
2506         case SEEK_SET:
2507         case SEEK_CUR:
2508                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2509                 retval = generic_file_llseek(file, offset, whence);
2510                 break;
2511         case SEEK_END:
2512                 inode_lock(inode);
2513                 retval = fuse_update_attributes(inode, file);
2514                 if (!retval)
2515                         retval = generic_file_llseek(file, offset, whence);
2516                 inode_unlock(inode);
2517                 break;
2518         case SEEK_HOLE:
2519         case SEEK_DATA:
2520                 inode_lock(inode);
2521                 retval = fuse_lseek(file, offset, whence);
2522                 inode_unlock(inode);
2523                 break;
2524         default:
2525                 retval = -EINVAL;
2526         }
2527
2528         return retval;
2529 }
2530
2531 /*
2532  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2533  * ABI was defined to be 'struct iovec' which is different on 32bit
2534  * and 64bit.  Fortunately we can determine which structure the server
2535  * used from the size of the reply.
2536  */
2537 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2538                                      size_t transferred, unsigned count,
2539                                      bool is_compat)
2540 {
2541 #ifdef CONFIG_COMPAT
2542         if (count * sizeof(struct compat_iovec) == transferred) {
2543                 struct compat_iovec *ciov = src;
2544                 unsigned i;
2545
2546                 /*
2547                  * With this interface a 32bit server cannot support
2548                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2549                  * requests
2550                  */
2551                 if (!is_compat)
2552                         return -EINVAL;
2553
2554                 for (i = 0; i < count; i++) {
2555                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2556                         dst[i].iov_len = ciov[i].iov_len;
2557                 }
2558                 return 0;
2559         }
2560 #endif
2561
2562         if (count * sizeof(struct iovec) != transferred)
2563                 return -EIO;
2564
2565         memcpy(dst, src, transferred);
2566         return 0;
2567 }
2568
2569 /* Make sure iov_length() won't overflow */
2570 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2571                                  size_t count)
2572 {
2573         size_t n;
2574         u32 max = fc->max_pages << PAGE_SHIFT;
2575
2576         for (n = 0; n < count; n++, iov++) {
2577                 if (iov->iov_len > (size_t) max)
2578                         return -ENOMEM;
2579                 max -= iov->iov_len;
2580         }
2581         return 0;
2582 }
2583
2584 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2585                                  void *src, size_t transferred, unsigned count,
2586                                  bool is_compat)
2587 {
2588         unsigned i;
2589         struct fuse_ioctl_iovec *fiov = src;
2590
2591         if (fc->minor < 16) {
2592                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2593                                                  count, is_compat);
2594         }
2595
2596         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2597                 return -EIO;
2598
2599         for (i = 0; i < count; i++) {
2600                 /* Did the server supply an inappropriate value? */
2601                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2602                     fiov[i].len != (unsigned long) fiov[i].len)
2603                         return -EIO;
2604
2605                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2606                 dst[i].iov_len = (size_t) fiov[i].len;
2607
2608 #ifdef CONFIG_COMPAT
2609                 if (is_compat &&
2610                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2611                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2612                         return -EIO;
2613 #endif
2614         }
2615
2616         return 0;
2617 }
2618
2619
2620 /*
2621  * For ioctls, there is no generic way to determine how much memory
2622  * needs to be read and/or written.  Furthermore, ioctls are allowed
2623  * to dereference the passed pointer, so the parameter requires deep
2624  * copying but FUSE has no idea whatsoever about what to copy in or
2625  * out.
2626  *
2627  * This is solved by allowing FUSE server to retry ioctl with
2628  * necessary in/out iovecs.  Let's assume the ioctl implementation
2629  * needs to read in the following structure.
2630  *
2631  * struct a {
2632  *      char    *buf;
2633  *      size_t  buflen;
2634  * }
2635  *
2636  * On the first callout to FUSE server, inarg->in_size and
2637  * inarg->out_size will be NULL; then, the server completes the ioctl
2638  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2639  * the actual iov array to
2640  *
2641  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2642  *
2643  * which tells FUSE to copy in the requested area and retry the ioctl.
2644  * On the second round, the server has access to the structure and
2645  * from that it can tell what to look for next, so on the invocation,
2646  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2647  *
2648  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2649  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2650  *
2651  * FUSE will copy both struct a and the pointed buffer from the
2652  * process doing the ioctl and retry ioctl with both struct a and the
2653  * buffer.
2654  *
2655  * This time, FUSE server has everything it needs and completes ioctl
2656  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2657  *
2658  * Copying data out works the same way.
2659  *
2660  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2661  * automatically initializes in and out iovs by decoding @cmd with
2662  * _IOC_* macros and the server is not allowed to request RETRY.  This
2663  * limits ioctl data transfers to well-formed ioctls and is the forced
2664  * behavior for all FUSE servers.
2665  */
2666 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2667                    unsigned int flags)
2668 {
2669         struct fuse_file *ff = file->private_data;
2670         struct fuse_conn *fc = ff->fc;
2671         struct fuse_ioctl_in inarg = {
2672                 .fh = ff->fh,
2673                 .cmd = cmd,
2674                 .arg = arg,
2675                 .flags = flags
2676         };
2677         struct fuse_ioctl_out outarg;
2678         struct iovec *iov_page = NULL;
2679         struct iovec *in_iov = NULL, *out_iov = NULL;
2680         unsigned int in_iovs = 0, out_iovs = 0, max_pages;
2681         size_t in_size, out_size, c;
2682         ssize_t transferred;
2683         int err, i;
2684         struct iov_iter ii;
2685         struct fuse_args_pages ap = {};
2686
2687 #if BITS_PER_LONG == 32
2688         inarg.flags |= FUSE_IOCTL_32BIT;
2689 #else
2690         if (flags & FUSE_IOCTL_COMPAT) {
2691                 inarg.flags |= FUSE_IOCTL_32BIT;
2692 #ifdef CONFIG_X86_X32
2693                 if (in_x32_syscall())
2694                         inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2695 #endif
2696         }
2697 #endif
2698
2699         /* assume all the iovs returned by client always fits in a page */
2700         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2701
2702         err = -ENOMEM;
2703         ap.pages = fuse_pages_alloc(fc->max_pages, GFP_KERNEL, &ap.descs);
2704         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2705         if (!ap.pages || !iov_page)
2706                 goto out;
2707
2708         fuse_page_descs_length_init(ap.descs, 0, fc->max_pages);
2709
2710         /*
2711          * If restricted, initialize IO parameters as encoded in @cmd.
2712          * RETRY from server is not allowed.
2713          */
2714         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2715                 struct iovec *iov = iov_page;
2716
2717                 iov->iov_base = (void __user *)arg;
2718                 iov->iov_len = _IOC_SIZE(cmd);
2719
2720                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2721                         in_iov = iov;
2722                         in_iovs = 1;
2723                 }
2724
2725                 if (_IOC_DIR(cmd) & _IOC_READ) {
2726                         out_iov = iov;
2727                         out_iovs = 1;
2728                 }
2729         }
2730
2731  retry:
2732         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2733         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2734
2735         /*
2736          * Out data can be used either for actual out data or iovs,
2737          * make sure there always is at least one page.
2738          */
2739         out_size = max_t(size_t, out_size, PAGE_SIZE);
2740         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2741
2742         /* make sure there are enough buffer pages and init request with them */
2743         err = -ENOMEM;
2744         if (max_pages > fc->max_pages)
2745                 goto out;
2746         while (ap.num_pages < max_pages) {
2747                 ap.pages[ap.num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2748                 if (!ap.pages[ap.num_pages])
2749                         goto out;
2750                 ap.num_pages++;
2751         }
2752
2753
2754         /* okay, let's send it to the client */
2755         ap.args.opcode = FUSE_IOCTL;
2756         ap.args.nodeid = ff->nodeid;
2757         ap.args.in_numargs = 1;
2758         ap.args.in_args[0].size = sizeof(inarg);
2759         ap.args.in_args[0].value = &inarg;
2760         if (in_size) {
2761                 ap.args.in_numargs++;
2762                 ap.args.in_args[1].size = in_size;
2763                 ap.args.in_pages = true;
2764
2765                 err = -EFAULT;
2766                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2767                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2768                         c = copy_page_from_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2769                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2770                                 goto out;
2771                 }
2772         }
2773
2774         ap.args.out_numargs = 2;
2775         ap.args.out_args[0].size = sizeof(outarg);
2776         ap.args.out_args[0].value = &outarg;
2777         ap.args.out_args[1].size = out_size;
2778         ap.args.out_pages = true;
2779         ap.args.out_argvar = true;
2780
2781         transferred = fuse_simple_request(fc, &ap.args);
2782         err = transferred;
2783         if (transferred < 0)
2784                 goto out;
2785
2786         /* did it ask for retry? */
2787         if (outarg.flags & FUSE_IOCTL_RETRY) {
2788                 void *vaddr;
2789
2790                 /* no retry if in restricted mode */
2791                 err = -EIO;
2792                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2793                         goto out;
2794
2795                 in_iovs = outarg.in_iovs;
2796                 out_iovs = outarg.out_iovs;
2797
2798                 /*
2799                  * Make sure things are in boundary, separate checks
2800                  * are to protect against overflow.
2801                  */
2802                 err = -ENOMEM;
2803                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2804                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2805                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2806                         goto out;
2807
2808                 vaddr = kmap_atomic(ap.pages[0]);
2809                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2810                                             transferred, in_iovs + out_iovs,
2811                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2812                 kunmap_atomic(vaddr);
2813                 if (err)
2814                         goto out;
2815
2816                 in_iov = iov_page;
2817                 out_iov = in_iov + in_iovs;
2818
2819                 err = fuse_verify_ioctl_iov(fc, in_iov, in_iovs);
2820                 if (err)
2821                         goto out;
2822
2823                 err = fuse_verify_ioctl_iov(fc, out_iov, out_iovs);
2824                 if (err)
2825                         goto out;
2826
2827                 goto retry;
2828         }
2829
2830         err = -EIO;
2831         if (transferred > inarg.out_size)
2832                 goto out;
2833
2834         err = -EFAULT;
2835         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2836         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2837                 c = copy_page_to_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2838                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2839                         goto out;
2840         }
2841         err = 0;
2842  out:
2843         free_page((unsigned long) iov_page);
2844         while (ap.num_pages)
2845                 __free_page(ap.pages[--ap.num_pages]);
2846         kfree(ap.pages);
2847
2848         return err ? err : outarg.result;
2849 }
2850 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2851
2852 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2853                        unsigned long arg, unsigned int flags)
2854 {
2855         struct inode *inode = file_inode(file);
2856         struct fuse_conn *fc = get_fuse_conn(inode);
2857
2858         if (!fuse_allow_current_process(fc))
2859                 return -EACCES;
2860
2861         if (is_bad_inode(inode))
2862                 return -EIO;
2863
2864         return fuse_do_ioctl(file, cmd, arg, flags);
2865 }
2866
2867 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2868                             unsigned long arg)
2869 {
2870         return fuse_ioctl_common(file, cmd, arg, 0);
2871 }
2872
2873 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2874                                    unsigned long arg)
2875 {
2876         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2877 }
2878
2879 /*
2880  * All files which have been polled are linked to RB tree
2881  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2882  * find the matching one.
2883  */
2884 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2885                                               struct rb_node **parent_out)
2886 {
2887         struct rb_node **link = &fc->polled_files.rb_node;
2888         struct rb_node *last = NULL;
2889
2890         while (*link) {
2891                 struct fuse_file *ff;
2892
2893                 last = *link;
2894                 ff = rb_entry(last, struct fuse_file, polled_node);
2895
2896                 if (kh < ff->kh)
2897                         link = &last->rb_left;
2898                 else if (kh > ff->kh)
2899                         link = &last->rb_right;
2900                 else
2901                         return link;
2902         }
2903
2904         if (parent_out)
2905                 *parent_out = last;
2906         return link;
2907 }
2908
2909 /*
2910  * The file is about to be polled.  Make sure it's on the polled_files
2911  * RB tree.  Note that files once added to the polled_files tree are
2912  * not removed before the file is released.  This is because a file
2913  * polled once is likely to be polled again.
2914  */
2915 static void fuse_register_polled_file(struct fuse_conn *fc,
2916                                       struct fuse_file *ff)
2917 {
2918         spin_lock(&fc->lock);
2919         if (RB_EMPTY_NODE(&ff->polled_node)) {
2920                 struct rb_node **link, *uninitialized_var(parent);
2921
2922                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2923                 BUG_ON(*link);
2924                 rb_link_node(&ff->polled_node, parent, link);
2925                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2926         }
2927         spin_unlock(&fc->lock);
2928 }
2929
2930 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2931 {
2932         struct fuse_file *ff = file->private_data;
2933         struct fuse_conn *fc = ff->fc;
2934         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2935         struct fuse_poll_out outarg;
2936         FUSE_ARGS(args);
2937         int err;
2938
2939         if (fc->no_poll)
2940                 return DEFAULT_POLLMASK;
2941
2942         poll_wait(file, &ff->poll_wait, wait);
2943         inarg.events = mangle_poll(poll_requested_events(wait));
2944
2945         /*
2946          * Ask for notification iff there's someone waiting for it.
2947          * The client may ignore the flag and always notify.
2948          */
2949         if (waitqueue_active(&ff->poll_wait)) {
2950                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2951                 fuse_register_polled_file(fc, ff);
2952         }
2953
2954         args.opcode = FUSE_POLL;
2955         args.nodeid = ff->nodeid;
2956         args.in_numargs = 1;
2957         args.in_args[0].size = sizeof(inarg);
2958         args.in_args[0].value = &inarg;
2959         args.out_numargs = 1;
2960         args.out_args[0].size = sizeof(outarg);
2961         args.out_args[0].value = &outarg;
2962         err = fuse_simple_request(fc, &args);
2963
2964         if (!err)
2965                 return demangle_poll(outarg.revents);
2966         if (err == -ENOSYS) {
2967                 fc->no_poll = 1;
2968                 return DEFAULT_POLLMASK;
2969         }
2970         return EPOLLERR;
2971 }
2972 EXPORT_SYMBOL_GPL(fuse_file_poll);
2973
2974 /*
2975  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2976  * wakes up the poll waiters.
2977  */
2978 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2979                             struct fuse_notify_poll_wakeup_out *outarg)
2980 {
2981         u64 kh = outarg->kh;
2982         struct rb_node **link;
2983
2984         spin_lock(&fc->lock);
2985
2986         link = fuse_find_polled_node(fc, kh, NULL);
2987         if (*link) {
2988                 struct fuse_file *ff;
2989
2990                 ff = rb_entry(*link, struct fuse_file, polled_node);
2991                 wake_up_interruptible_sync(&ff->poll_wait);
2992         }
2993
2994         spin_unlock(&fc->lock);
2995         return 0;
2996 }
2997
2998 static void fuse_do_truncate(struct file *file)
2999 {
3000         struct inode *inode = file->f_mapping->host;
3001         struct iattr attr;
3002
3003         attr.ia_valid = ATTR_SIZE;
3004         attr.ia_size = i_size_read(inode);
3005
3006         attr.ia_file = file;
3007         attr.ia_valid |= ATTR_FILE;
3008
3009         fuse_do_setattr(file_dentry(file), &attr, file);
3010 }
3011
3012 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
3013 {
3014         return round_up(off, fc->max_pages << PAGE_SHIFT);
3015 }
3016
3017 static ssize_t
3018 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3019 {
3020         DECLARE_COMPLETION_ONSTACK(wait);
3021         ssize_t ret = 0;
3022         struct file *file = iocb->ki_filp;
3023         struct fuse_file *ff = file->private_data;
3024         bool async_dio = ff->fc->async_dio;
3025         loff_t pos = 0;
3026         struct inode *inode;
3027         loff_t i_size;
3028         size_t count = iov_iter_count(iter);
3029         loff_t offset = iocb->ki_pos;
3030         struct fuse_io_priv *io;
3031
3032         pos = offset;
3033         inode = file->f_mapping->host;
3034         i_size = i_size_read(inode);
3035
3036         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
3037                 return 0;
3038
3039         /* optimization for short read */
3040         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
3041                 if (offset >= i_size)
3042                         return 0;
3043                 iov_iter_truncate(iter, fuse_round_up(ff->fc, i_size - offset));
3044                 count = iov_iter_count(iter);
3045         }
3046
3047         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
3048         if (!io)
3049                 return -ENOMEM;
3050         spin_lock_init(&io->lock);
3051         kref_init(&io->refcnt);
3052         io->reqs = 1;
3053         io->bytes = -1;
3054         io->size = 0;
3055         io->offset = offset;
3056         io->write = (iov_iter_rw(iter) == WRITE);
3057         io->err = 0;
3058         /*
3059          * By default, we want to optimize all I/Os with async request
3060          * submission to the client filesystem if supported.
3061          */
3062         io->async = async_dio;
3063         io->iocb = iocb;
3064         io->blocking = is_sync_kiocb(iocb);
3065
3066         /*
3067          * We cannot asynchronously extend the size of a file.
3068          * In such case the aio will behave exactly like sync io.
3069          */
3070         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
3071                 io->blocking = true;
3072
3073         if (io->async && io->blocking) {
3074                 /*
3075                  * Additional reference to keep io around after
3076                  * calling fuse_aio_complete()
3077                  */
3078                 kref_get(&io->refcnt);
3079                 io->done = &wait;
3080         }
3081
3082         if (iov_iter_rw(iter) == WRITE) {
3083                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
3084                 fuse_invalidate_attr(inode);
3085         } else {
3086                 ret = __fuse_direct_read(io, iter, &pos);
3087         }
3088
3089         if (io->async) {
3090                 bool blocking = io->blocking;
3091
3092                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3093
3094                 /* we have a non-extending, async request, so return */
3095                 if (!blocking)
3096                         return -EIOCBQUEUED;
3097
3098                 wait_for_completion(&wait);
3099                 ret = fuse_get_res_by_io(io);
3100         }
3101
3102         kref_put(&io->refcnt, fuse_io_release);
3103
3104         if (iov_iter_rw(iter) == WRITE) {
3105                 if (ret > 0)
3106                         fuse_write_update_size(inode, pos);
3107                 else if (ret < 0 && offset + count > i_size)
3108                         fuse_do_truncate(file);
3109         }
3110
3111         return ret;
3112 }
3113
3114 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3115 {
3116         int err = filemap_write_and_wait_range(inode->i_mapping, start, end);
3117
3118         if (!err)
3119                 fuse_sync_writes(inode);
3120
3121         return err;
3122 }
3123
3124 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3125                                 loff_t length)
3126 {
3127         struct fuse_file *ff = file->private_data;
3128         struct inode *inode = file_inode(file);
3129         struct fuse_inode *fi = get_fuse_inode(inode);
3130         struct fuse_conn *fc = ff->fc;
3131         FUSE_ARGS(args);
3132         struct fuse_fallocate_in inarg = {
3133                 .fh = ff->fh,
3134                 .offset = offset,
3135                 .length = length,
3136                 .mode = mode
3137         };
3138         int err;
3139         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
3140                            (mode & FALLOC_FL_PUNCH_HOLE);
3141
3142         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3143                 return -EOPNOTSUPP;
3144
3145         if (fc->no_fallocate)
3146                 return -EOPNOTSUPP;
3147
3148         if (lock_inode) {
3149                 inode_lock(inode);
3150                 if (mode & FALLOC_FL_PUNCH_HOLE) {
3151                         loff_t endbyte = offset + length - 1;
3152
3153                         err = fuse_writeback_range(inode, offset, endbyte);
3154                         if (err)
3155                                 goto out;
3156                 }
3157         }
3158
3159         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3160             offset + length > i_size_read(inode)) {
3161                 err = inode_newsize_ok(inode, offset + length);
3162                 if (err)
3163                         goto out;
3164         }
3165
3166         if (!(mode & FALLOC_FL_KEEP_SIZE))
3167                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3168
3169         args.opcode = FUSE_FALLOCATE;
3170         args.nodeid = ff->nodeid;
3171         args.in_numargs = 1;
3172         args.in_args[0].size = sizeof(inarg);
3173         args.in_args[0].value = &inarg;
3174         err = fuse_simple_request(fc, &args);
3175         if (err == -ENOSYS) {
3176                 fc->no_fallocate = 1;
3177                 err = -EOPNOTSUPP;
3178         }
3179         if (err)
3180                 goto out;
3181
3182         /* we could have extended the file */
3183         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3184                 bool changed = fuse_write_update_size(inode, offset + length);
3185
3186                 if (changed && fc->writeback_cache)
3187                         file_update_time(file);
3188         }
3189
3190         if (mode & FALLOC_FL_PUNCH_HOLE)
3191                 truncate_pagecache_range(inode, offset, offset + length - 1);
3192
3193         fuse_invalidate_attr(inode);
3194
3195 out:
3196         if (!(mode & FALLOC_FL_KEEP_SIZE))
3197                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3198
3199         if (lock_inode)
3200                 inode_unlock(inode);
3201
3202         return err;
3203 }
3204
3205 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3206                                       struct file *file_out, loff_t pos_out,
3207                                       size_t len, unsigned int flags)
3208 {
3209         struct fuse_file *ff_in = file_in->private_data;
3210         struct fuse_file *ff_out = file_out->private_data;
3211         struct inode *inode_in = file_inode(file_in);
3212         struct inode *inode_out = file_inode(file_out);
3213         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3214         struct fuse_conn *fc = ff_in->fc;
3215         FUSE_ARGS(args);
3216         struct fuse_copy_file_range_in inarg = {
3217                 .fh_in = ff_in->fh,
3218                 .off_in = pos_in,
3219                 .nodeid_out = ff_out->nodeid,
3220                 .fh_out = ff_out->fh,
3221                 .off_out = pos_out,
3222                 .len = len,
3223                 .flags = flags
3224         };
3225         struct fuse_write_out outarg;
3226         ssize_t err;
3227         /* mark unstable when write-back is not used, and file_out gets
3228          * extended */
3229         bool is_unstable = (!fc->writeback_cache) &&
3230                            ((pos_out + len) > inode_out->i_size);
3231
3232         if (fc->no_copy_file_range)
3233                 return -EOPNOTSUPP;
3234
3235         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3236                 return -EXDEV;
3237
3238         if (fc->writeback_cache) {
3239                 inode_lock(inode_in);
3240                 err = fuse_writeback_range(inode_in, pos_in, pos_in + len);
3241                 inode_unlock(inode_in);
3242                 if (err)
3243                         return err;
3244         }
3245
3246         inode_lock(inode_out);
3247
3248         err = file_modified(file_out);
3249         if (err)
3250                 goto out;
3251
3252         if (fc->writeback_cache) {
3253                 err = fuse_writeback_range(inode_out, pos_out, pos_out + len);
3254                 if (err)
3255                         goto out;
3256         }
3257
3258         if (is_unstable)
3259                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3260
3261         args.opcode = FUSE_COPY_FILE_RANGE;
3262         args.nodeid = ff_in->nodeid;
3263         args.in_numargs = 1;
3264         args.in_args[0].size = sizeof(inarg);
3265         args.in_args[0].value = &inarg;
3266         args.out_numargs = 1;
3267         args.out_args[0].size = sizeof(outarg);
3268         args.out_args[0].value = &outarg;
3269         err = fuse_simple_request(fc, &args);
3270         if (err == -ENOSYS) {
3271                 fc->no_copy_file_range = 1;
3272                 err = -EOPNOTSUPP;
3273         }
3274         if (err)
3275                 goto out;
3276
3277         if (fc->writeback_cache) {
3278                 fuse_write_update_size(inode_out, pos_out + outarg.size);
3279                 file_update_time(file_out);
3280         }
3281
3282         fuse_invalidate_attr(inode_out);
3283
3284         err = outarg.size;
3285 out:
3286         if (is_unstable)
3287                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3288
3289         inode_unlock(inode_out);
3290         file_accessed(file_in);
3291
3292         return err;
3293 }
3294
3295 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3296                                     struct file *dst_file, loff_t dst_off,
3297                                     size_t len, unsigned int flags)
3298 {
3299         ssize_t ret;
3300
3301         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3302                                      len, flags);
3303
3304         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3305                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3306                                               dst_off, len, flags);
3307         return ret;
3308 }
3309
3310 static const struct file_operations fuse_file_operations = {
3311         .llseek         = fuse_file_llseek,
3312         .read_iter      = fuse_file_read_iter,
3313         .write_iter     = fuse_file_write_iter,
3314         .mmap           = fuse_file_mmap,
3315         .open           = fuse_open,
3316         .flush          = fuse_flush,
3317         .release        = fuse_release,
3318         .fsync          = fuse_fsync,
3319         .lock           = fuse_file_lock,
3320         .flock          = fuse_file_flock,
3321         .splice_read    = generic_file_splice_read,
3322         .splice_write   = iter_file_splice_write,
3323         .unlocked_ioctl = fuse_file_ioctl,
3324         .compat_ioctl   = fuse_file_compat_ioctl,
3325         .poll           = fuse_file_poll,
3326         .fallocate      = fuse_file_fallocate,
3327         .copy_file_range = fuse_copy_file_range,
3328 };
3329
3330 static const struct address_space_operations fuse_file_aops  = {
3331         .readpage       = fuse_readpage,
3332         .readahead      = fuse_readahead,
3333         .writepage      = fuse_writepage,
3334         .writepages     = fuse_writepages,
3335         .launder_page   = fuse_launder_page,
3336         .set_page_dirty = __set_page_dirty_nobuffers,
3337         .bmap           = fuse_bmap,
3338         .direct_IO      = fuse_direct_IO,
3339         .write_begin    = fuse_write_begin,
3340         .write_end      = fuse_write_end,
3341 };
3342
3343 void fuse_init_file_inode(struct inode *inode)
3344 {
3345         struct fuse_inode *fi = get_fuse_inode(inode);
3346
3347         inode->i_fop = &fuse_file_operations;
3348         inode->i_data.a_ops = &fuse_file_aops;
3349
3350         INIT_LIST_HEAD(&fi->write_files);
3351         INIT_LIST_HEAD(&fi->queued_writes);
3352         fi->writectr = 0;
3353         init_waitqueue_head(&fi->page_waitq);
3354         INIT_LIST_HEAD(&fi->writepages);
3355 }