fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   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 lock_inode = (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 (lock_inode)
 229                 inode_lock(inode);
 230
 231         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 232
 233         if (!err)
 234                 fuse_finish_open(inode, file);
 235
 236         if (lock_inode)
 237                 inode_unlock(inode);
 238
 239         return err;
 240 }
 241
 242 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 243                                  int flags, int opcode)
 244 {
 245         struct fuse_conn *fc = ff->fc;
 246         struct fuse_release_args *ra = ff->release_args;
 247
 248         /* Inode is NULL on error path of fuse_create_open() */
 249         if (likely(fi)) {
 250                 spin_lock(&fi->lock);
 251                 list_del(&ff->write_entry);
 252                 spin_unlock(&fi->lock);
 253         }
 254         spin_lock(&fc->lock);
 255         if (!RB_EMPTY_NODE(&ff->polled_node))
 256                 rb_erase(&ff->polled_node, &fc->polled_files);
 257         spin_unlock(&fc->lock);
 258
 259         wake_up_interruptible_all(&ff->poll_wait);
 260
 261         ra->inarg.fh = ff->fh;
 262         ra->inarg.flags = flags;
 263         ra->args.in_numargs = 1;
 264         ra->args.in_args[0].size = sizeof(struct fuse_release_in);
 265         ra->args.in_args[0].value = &ra->inarg;
 266         ra->args.opcode = opcode;
 267         ra->args.nodeid = ff->nodeid;
 268         ra->args.force = true;
 269         ra->args.nocreds = true;
 270 }
 271
 272 void fuse_release_common(struct file *file, bool isdir)
 273 {
 274         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 275         struct fuse_file *ff = file->private_data;
 276         struct fuse_release_args *ra = ff->release_args;
 277         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 278
 279         fuse_prepare_release(fi, ff, file->f_flags, opcode);
 280
 281         if (ff->flock) {
 282                 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 283                 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fc,
 284                                                           (fl_owner_t) file);
 285         }
 286         /* Hold inode until release is finished */
 287         ra->inode = igrab(file_inode(file));
 288
 289         /*
 290          * Normally this will send the RELEASE request, however if
 291          * some asynchronous READ or WRITE requests are outstanding,
 292          * the sending will be delayed.
 293          *
 294          * Make the release synchronous if this is a fuseblk mount,
 295          * synchronous RELEASE is allowed (and desirable) in this case
 296          * because the server can be trusted not to screw up.
 297          */
 298         fuse_file_put(ff, ff->fc->destroy, isdir);
 299 }
 300
 301 static int fuse_open(struct inode *inode, struct file *file)
 302 {
 303         return fuse_open_common(inode, file, false);
 304 }
 305
 306 static int fuse_release(struct inode *inode, struct file *file)
 307 {
 308         struct fuse_conn *fc = get_fuse_conn(inode);
 309
 310         /* see fuse_vma_close() for !writeback_cache case */
 311         if (fc->writeback_cache)
 312                 write_inode_now(inode, 1);
 313
 314         fuse_release_common(file, false);
 315
 316         /* return value is ignored by VFS */
 317         return 0;
 318 }
 319
 320 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
 321 {
 322         WARN_ON(refcount_read(&ff->count) > 1);
 323         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 324         /*
 325          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 326          * synchronous, we are fine with not doing igrab() here"
 327          */
 328         fuse_file_put(ff, true, false);
 329 }
 330 EXPORT_SYMBOL_GPL(fuse_sync_release);
 331
 332 /*
 333  * Scramble the ID space with XTEA, so that the value of the files_struct
 334  * pointer is not exposed to userspace.
 335  */
 336 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 337 {
 338         u32 *k = fc->scramble_key;
 339         u64 v = (unsigned long) id;
 340         u32 v0 = v;
 341         u32 v1 = v >> 32;
 342         u32 sum = 0;
 343         int i;
 344
 345         for (i = 0; i < 32; i++) {
 346                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 347                 sum += 0x9E3779B9;
 348                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 349         }
 350
 351         return (u64) v0 + ((u64) v1 << 32);
 352 }
 353
 354 struct fuse_writepage_args {
 355         struct fuse_io_args ia;
 356         struct list_head writepages_entry;
 357         struct list_head queue_entry;
 358         struct fuse_writepage_args *next;
 359         struct inode *inode;
 360 };
 361
 362 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
 363                                             pgoff_t idx_from, pgoff_t idx_to)
 364 {
 365         struct fuse_writepage_args *wpa;
 366
 367         list_for_each_entry(wpa, &fi->writepages, writepages_entry) {
 368                 pgoff_t curr_index;
 369
 370                 WARN_ON(get_fuse_inode(wpa->inode) != fi);
 371                 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
 372                 if (idx_from < curr_index + wpa->ia.ap.num_pages &&
 373                     curr_index <= idx_to) {
 374                         return wpa;
 375                 }
 376         }
 377         return NULL;
 378 }
 379
 380 /*
 381  * Check if any page in a range is under writeback
 382  *
 383  * This is currently done by walking the list of writepage requests
 384  * for the inode, which can be pretty inefficient.
 385  */
 386 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 387                                    pgoff_t idx_to)
 388 {
 389         struct fuse_inode *fi = get_fuse_inode(inode);
 390         bool found;
 391
 392         spin_lock(&fi->lock);
 393         found = fuse_find_writeback(fi, idx_from, idx_to);
 394         spin_unlock(&fi->lock);
 395
 396         return found;
 397 }
 398
 399 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 400 {
 401         return fuse_range_is_writeback(inode, index, index);
 402 }
 403
 404 /*
 405  * Wait for page writeback to be completed.
 406  *
 407  * Since fuse doesn't rely on the VM writeback tracking, this has to
 408  * use some other means.
 409  */
 410 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 411 {
 412         struct fuse_inode *fi = get_fuse_inode(inode);
 413
 414         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 415 }
 416
 417 /*
 418  * Wait for all pending writepages on the inode to finish.
 419  *
 420  * This is currently done by blocking further writes with FUSE_NOWRITE
 421  * and waiting for all sent writes to complete.
 422  *
 423  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 424  * could conflict with truncation.
 425  */
 426 static void fuse_sync_writes(struct inode *inode)
 427 {
 428         fuse_set_nowrite(inode);
 429         fuse_release_nowrite(inode);
 430 }
 431
 432 static int fuse_flush(struct file *file, fl_owner_t id)
 433 {
 434         struct inode *inode = file_inode(file);
 435         struct fuse_conn *fc = get_fuse_conn(inode);
 436         struct fuse_file *ff = file->private_data;
 437         struct fuse_flush_in inarg;
 438         FUSE_ARGS(args);
 439         int err;
 440
 441         if (is_bad_inode(inode))
 442                 return -EIO;
 443
 444         if (fc->no_flush)
 445                 return 0;
 446
 447         err = write_inode_now(inode, 1);
 448         if (err)
 449                 return err;
 450
 451         inode_lock(inode);
 452         fuse_sync_writes(inode);
 453         inode_unlock(inode);
 454
 455         err = filemap_check_errors(file->f_mapping);
 456         if (err)
 457                 return err;
 458
 459         memset(&inarg, 0, sizeof(inarg));
 460         inarg.fh = ff->fh;
 461         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 462         args.opcode = FUSE_FLUSH;
 463         args.nodeid = get_node_id(inode);
 464         args.in_numargs = 1;
 465         args.in_args[0].size = sizeof(inarg);
 466         args.in_args[0].value = &inarg;
 467         args.force = true;
 468
 469         err = fuse_simple_request(fc, &args);
 470         if (err == -ENOSYS) {
 471                 fc->no_flush = 1;
 472                 err = 0;
 473         }
 474         return err;
 475 }
 476
 477 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 478                       int datasync, int opcode)
 479 {
 480         struct inode *inode = file->f_mapping->host;
 481         struct fuse_conn *fc = get_fuse_conn(inode);
 482         struct fuse_file *ff = file->private_data;
 483         FUSE_ARGS(args);
 484         struct fuse_fsync_in inarg;
 485
 486         memset(&inarg, 0, sizeof(inarg));
 487         inarg.fh = ff->fh;
 488         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 489         args.opcode = opcode;
 490         args.nodeid = get_node_id(inode);
 491         args.in_numargs = 1;
 492         args.in_args[0].size = sizeof(inarg);
 493         args.in_args[0].value = &inarg;
 494         return fuse_simple_request(fc, &args);
 495 }
 496
 497 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 498                       int datasync)
 499 {
 500         struct inode *inode = file->f_mapping->host;
 501         struct fuse_conn *fc = get_fuse_conn(inode);
 502         int err;
 503
 504         if (is_bad_inode(inode))
 505                 return -EIO;
 506
 507         inode_lock(inode);
 508
 509         /*
 510          * Start writeback against all dirty pages of the inode, then
 511          * wait for all outstanding writes, before sending the FSYNC
 512          * request.
 513          */
 514         err = file_write_and_wait_range(file, start, end);
 515         if (err)
 516                 goto out;
 517
 518         fuse_sync_writes(inode);
 519
 520         /*
 521          * Due to implementation of fuse writeback
 522          * file_write_and_wait_range() does not catch errors.
 523          * We have to do this directly after fuse_sync_writes()
 524          */
 525         err = file_check_and_advance_wb_err(file);
 526         if (err)
 527                 goto out;
 528
 529         err = sync_inode_metadata(inode, 1);
 530         if (err)
 531                 goto out;
 532
 533         if (fc->no_fsync)
 534                 goto out;
 535
 536         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 537         if (err == -ENOSYS) {
 538                 fc->no_fsync = 1;
 539                 err = 0;
 540         }
 541 out:
 542         inode_unlock(inode);
 543
 544         return err;
 545 }
 546
 547 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
 548                          size_t count, int opcode)
 549 {
 550         struct fuse_file *ff = file->private_data;
 551         struct fuse_args *args = &ia->ap.args;
 552
 553         ia->read.in.fh = ff->fh;
 554         ia->read.in.offset = pos;
 555         ia->read.in.size = count;
 556         ia->read.in.flags = file->f_flags;
 557         args->opcode = opcode;
 558         args->nodeid = ff->nodeid;
 559         args->in_numargs = 1;
 560         args->in_args[0].size = sizeof(ia->read.in);
 561         args->in_args[0].value = &ia->read.in;
 562         args->out_argvar = true;
 563         args->out_numargs = 1;
 564         args->out_args[0].size = count;
 565 }
 566
 567 static void fuse_release_user_pages(struct fuse_args_pages *ap,
 568                                     bool should_dirty)
 569 {
 570         unsigned int i;
 571
 572         for (i = 0; i < ap->num_pages; i++) {
 573                 if (should_dirty)
 574                         set_page_dirty_lock(ap->pages[i]);
 575                 put_page(ap->pages[i]);
 576         }
 577 }
 578
 579 static void fuse_io_release(struct kref *kref)
 580 {
 581         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 582 }
 583
 584 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 585 {
 586         if (io->err)
 587                 return io->err;
 588
 589         if (io->bytes >= 0 && io->write)
 590                 return -EIO;
 591
 592         return io->bytes < 0 ? io->size : io->bytes;
 593 }
 594
 595 /**
 596  * In case of short read, the caller sets 'pos' to the position of
 597  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 598  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 599  *
 600  * An example:
 601  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 602  * both submitted asynchronously. The first of them was ACKed by userspace as
 603  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 604  * second request was ACKed as short, e.g. only 1K was read, resulting in
 605  * pos == 33K.
 606  *
 607  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 608  * will be equal to the length of the longest contiguous fragment of
 609  * transferred data starting from the beginning of IO request.
 610  */
 611 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 612 {
 613         int left;
 614
 615         spin_lock(&io->lock);
 616         if (err)
 617                 io->err = io->err ? : err;
 618         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 619                 io->bytes = pos;
 620
 621         left = --io->reqs;
 622         if (!left && io->blocking)
 623                 complete(io->done);
 624         spin_unlock(&io->lock);
 625
 626         if (!left && !io->blocking) {
 627                 ssize_t res = fuse_get_res_by_io(io);
 628
 629                 if (res >= 0) {
 630                         struct inode *inode = file_inode(io->iocb->ki_filp);
 631                         struct fuse_conn *fc = get_fuse_conn(inode);
 632                         struct fuse_inode *fi = get_fuse_inode(inode);
 633
 634                         spin_lock(&fi->lock);
 635                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 636                         spin_unlock(&fi->lock);
 637                 }
 638
 639                 io->iocb->ki_complete(io->iocb, res, 0);
 640         }
 641
 642         kref_put(&io->refcnt, fuse_io_release);
 643 }
 644
 645 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
 646                                           unsigned int npages)
 647 {
 648         struct fuse_io_args *ia;
 649
 650         ia = kzalloc(sizeof(*ia), GFP_KERNEL);
 651         if (ia) {
 652                 ia->io = io;
 653                 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
 654                                                 &ia->ap.descs);
 655                 if (!ia->ap.pages) {
 656                         kfree(ia);
 657                         ia = NULL;
 658                 }
 659         }
 660         return ia;
 661 }
 662
 663 static void fuse_io_free(struct fuse_io_args *ia)
 664 {
 665         kfree(ia->ap.pages);
 666         kfree(ia);
 667 }
 668
 669 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_args *args,
 670                                   int err)
 671 {
 672         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 673         struct fuse_io_priv *io = ia->io;
 674         ssize_t pos = -1;
 675
 676         fuse_release_user_pages(&ia->ap, io->should_dirty);
 677
 678         if (err) {
 679                 /* Nothing */
 680         } else if (io->write) {
 681                 if (ia->write.out.size > ia->write.in.size) {
 682                         err = -EIO;
 683                 } else if (ia->write.in.size != ia->write.out.size) {
 684                         pos = ia->write.in.offset - io->offset +
 685                                 ia->write.out.size;
 686                 }
 687         } else {
 688                 u32 outsize = args->out_args[0].size;
 689
 690                 if (ia->read.in.size != outsize)
 691                         pos = ia->read.in.offset - io->offset + outsize;
 692         }
 693
 694         fuse_aio_complete(io, err, pos);
 695         fuse_io_free(ia);
 696 }
 697
 698 static ssize_t fuse_async_req_send(struct fuse_conn *fc,
 699                                    struct fuse_io_args *ia, size_t num_bytes)
 700 {
 701         ssize_t err;
 702         struct fuse_io_priv *io = ia->io;
 703
 704         spin_lock(&io->lock);
 705         kref_get(&io->refcnt);
 706         io->size += num_bytes;
 707         io->reqs++;
 708         spin_unlock(&io->lock);
 709
 710         ia->ap.args.end = fuse_aio_complete_req;
 711         err = fuse_simple_background(fc, &ia->ap.args, GFP_KERNEL);
 712
 713         return err ?: num_bytes;
 714 }
 715
 716 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
 717                               fl_owner_t owner)
 718 {
 719         struct file *file = ia->io->iocb->ki_filp;
 720         struct fuse_file *ff = file->private_data;
 721         struct fuse_conn *fc = ff->fc;
 722
 723         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 724         if (owner != NULL) {
 725                 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
 726                 ia->read.in.lock_owner = fuse_lock_owner_id(fc, owner);
 727         }
 728
 729         if (ia->io->async)
 730                 return fuse_async_req_send(fc, ia, count);
 731
 732         return fuse_simple_request(fc, &ia->ap.args);
 733 }
 734
 735 static void fuse_read_update_size(struct inode *inode, loff_t size,
 736                                   u64 attr_ver)
 737 {
 738         struct fuse_conn *fc = get_fuse_conn(inode);
 739         struct fuse_inode *fi = get_fuse_inode(inode);
 740
 741         spin_lock(&fi->lock);
 742         if (attr_ver == fi->attr_version && size < inode->i_size &&
 743             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 744                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 745                 i_size_write(inode, size);
 746         }
 747         spin_unlock(&fi->lock);
 748 }
 749
 750 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
 751                             struct fuse_args_pages *ap)
 752 {
 753         struct fuse_conn *fc = get_fuse_conn(inode);
 754
 755         if (fc->writeback_cache) {
 756                 /*
 757                  * A hole in a file. Some data after the hole are in page cache,
 758                  * but have not reached the client fs yet. So, the hole is not
 759                  * present there.
 760                  */
 761                 int i;
 762                 int start_idx = num_read >> PAGE_SHIFT;
 763                 size_t off = num_read & (PAGE_SIZE - 1);
 764
 765                 for (i = start_idx; i < ap->num_pages; i++) {
 766                         zero_user_segment(ap->pages[i], off, PAGE_SIZE);
 767                         off = 0;
 768                 }
 769         } else {
 770                 loff_t pos = page_offset(ap->pages[0]) + num_read;
 771                 fuse_read_update_size(inode, pos, attr_ver);
 772         }
 773 }
 774
 775 static int fuse_do_readpage(struct file *file, struct page *page)
 776 {
 777         struct inode *inode = page->mapping->host;
 778         struct fuse_conn *fc = get_fuse_conn(inode);
 779         loff_t pos = page_offset(page);
 780         struct fuse_page_desc desc = { .length = PAGE_SIZE };
 781         struct fuse_io_args ia = {
 782                 .ap.args.page_zeroing = true,
 783                 .ap.args.out_pages = true,
 784                 .ap.num_pages = 1,
 785                 .ap.pages = &page,
 786                 .ap.descs = &desc,
 787         };
 788         ssize_t res;
 789         u64 attr_ver;
 790
 791         /*
 792          * Page writeback can extend beyond the lifetime of the
 793          * page-cache page, so make sure we read a properly synced
 794          * page.
 795          */
 796         fuse_wait_on_page_writeback(inode, page->index);
 797
 798         attr_ver = fuse_get_attr_version(fc);
 799
 800         fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
 801         res = fuse_simple_request(fc, &ia.ap.args);
 802         if (res < 0)
 803                 return res;
 804         /*
 805          * Short read means EOF.  If file size is larger, truncate it
 806          */
 807         if (res < desc.length)
 808                 fuse_short_read(inode, attr_ver, res, &ia.ap);
 809
 810         SetPageUptodate(page);
 811
 812         return 0;
 813 }
 814
 815 static int fuse_readpage(struct file *file, struct page *page)
 816 {
 817         struct inode *inode = page->mapping->host;
 818         int err;
 819
 820         err = -EIO;
 821         if (is_bad_inode(inode))
 822                 goto out;
 823
 824         err = fuse_do_readpage(file, page);
 825         fuse_invalidate_atime(inode);
 826  out:
 827         unlock_page(page);
 828         return err;
 829 }
 830
 831 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_args *args,
 832                                int err)
 833 {
 834         int i;
 835         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 836         struct fuse_args_pages *ap = &ia->ap;
 837         size_t count = ia->read.in.size;
 838         size_t num_read = args->out_args[0].size;
 839         struct address_space *mapping = NULL;
 840
 841         for (i = 0; mapping == NULL && i < ap->num_pages; i++)
 842                 mapping = ap->pages[i]->mapping;
 843
 844         if (mapping) {
 845                 struct inode *inode = mapping->host;
 846
 847                 /*
 848                  * Short read means EOF. If file size is larger, truncate it
 849                  */
 850                 if (!err && num_read < count)
 851                         fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
 852
 853                 fuse_invalidate_atime(inode);
 854         }
 855
 856         for (i = 0; i < ap->num_pages; i++) {
 857                 struct page *page = ap->pages[i];
 858
 859                 if (!err)
 860                         SetPageUptodate(page);
 861                 else
 862                         SetPageError(page);
 863                 unlock_page(page);
 864                 put_page(page);
 865         }
 866         if (ia->ff)
 867                 fuse_file_put(ia->ff, false, false);
 868
 869         fuse_io_free(ia);
 870 }
 871
 872 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
 873 {
 874         struct fuse_file *ff = file->private_data;
 875         struct fuse_conn *fc = ff->fc;
 876         struct fuse_args_pages *ap = &ia->ap;
 877         loff_t pos = page_offset(ap->pages[0]);
 878         size_t count = ap->num_pages << PAGE_SHIFT;
 879         int err;
 880
 881         ap->args.out_pages = true;
 882         ap->args.page_zeroing = true;
 883         ap->args.page_replace = true;
 884         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 885         ia->read.attr_ver = fuse_get_attr_version(fc);
 886         if (fc->async_read) {
 887                 ia->ff = fuse_file_get(ff);
 888                 ap->args.end = fuse_readpages_end;
 889                 err = fuse_simple_background(fc, &ap->args, GFP_KERNEL);
 890                 if (!err)
 891                         return;
 892         } else {
 893                 err = fuse_simple_request(fc, &ap->args);
 894         }
 895         fuse_readpages_end(fc, &ap->args, err);
 896 }
 897
 898 struct fuse_fill_data {
 899         struct fuse_io_args *ia;
 900         struct file *file;
 901         struct inode *inode;
 902         unsigned int nr_pages;
 903         unsigned int max_pages;
 904 };
 905
 906 static int fuse_readpages_fill(void *_data, struct page *page)
 907 {
 908         struct fuse_fill_data *data = _data;
 909         struct fuse_io_args *ia = data->ia;
 910         struct fuse_args_pages *ap = &ia->ap;
 911         struct inode *inode = data->inode;
 912         struct fuse_conn *fc = get_fuse_conn(inode);
 913
 914         fuse_wait_on_page_writeback(inode, page->index);
 915
 916         if (ap->num_pages &&
 917             (ap->num_pages == fc->max_pages ||
 918              (ap->num_pages + 1) * PAGE_SIZE > fc->max_read ||
 919              ap->pages[ap->num_pages - 1]->index + 1 != page->index)) {
 920                 data->max_pages = min_t(unsigned int, data->nr_pages,
 921                                         fc->max_pages);
 922                 fuse_send_readpages(ia, data->file);
 923                 data->ia = ia = fuse_io_alloc(NULL, data->max_pages);
 924                 if (!ia) {
 925                         unlock_page(page);
 926                         return -ENOMEM;
 927                 }
 928                 ap = &ia->ap;
 929         }
 930
 931         if (WARN_ON(ap->num_pages >= data->max_pages)) {
 932                 unlock_page(page);
 933                 fuse_io_free(ia);
 934                 return -EIO;
 935         }
 936
 937         get_page(page);
 938         ap->pages[ap->num_pages] = page;
 939         ap->descs[ap->num_pages].length = PAGE_SIZE;
 940         ap->num_pages++;
 941         data->nr_pages--;
 942         return 0;
 943 }
 944
 945 static int fuse_readpages(struct file *file, struct address_space *mapping,
 946                           struct list_head *pages, unsigned nr_pages)
 947 {
 948         struct inode *inode = mapping->host;
 949         struct fuse_conn *fc = get_fuse_conn(inode);
 950         struct fuse_fill_data data;
 951         int err;
 952
 953         err = -EIO;
 954         if (is_bad_inode(inode))
 955                 goto out;
 956
 957         data.file = file;
 958         data.inode = inode;
 959         data.nr_pages = nr_pages;
 960         data.max_pages = min_t(unsigned int, nr_pages, fc->max_pages);
 961 ;
 962         data.ia = fuse_io_alloc(NULL, data.max_pages);
 963         err = -ENOMEM;
 964         if (!data.ia)
 965                 goto out;
 966
 967         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 968         if (!err) {
 969                 if (data.ia->ap.num_pages)
 970                         fuse_send_readpages(data.ia, file);
 971                 else
 972                         fuse_io_free(data.ia);
 973         }
 974 out:
 975         return err;
 976 }
 977
 978 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
 979 {
 980         struct inode *inode = iocb->ki_filp->f_mapping->host;
 981         struct fuse_conn *fc = get_fuse_conn(inode);
 982
 983         /*
 984          * In auto invalidate mode, always update attributes on read.
 985          * Otherwise, only update if we attempt to read past EOF (to ensure
 986          * i_size is up to date).
 987          */
 988         if (fc->auto_inval_data ||
 989             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 990                 int err;
 991                 err = fuse_update_attributes(inode, iocb->ki_filp);
 992                 if (err)
 993                         return err;
 994         }
 995
 996         return generic_file_read_iter(iocb, to);
 997 }
 998
 999 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1000                                  loff_t pos, size_t count)
1001 {
1002         struct fuse_args *args = &ia->ap.args;
1003
1004         ia->write.in.fh = ff->fh;
1005         ia->write.in.offset = pos;
1006         ia->write.in.size = count;
1007         args->opcode = FUSE_WRITE;
1008         args->nodeid = ff->nodeid;
1009         args->in_numargs = 2;
1010         if (ff->fc->minor < 9)
1011                 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1012         else
1013                 args->in_args[0].size = sizeof(ia->write.in);
1014         args->in_args[0].value = &ia->write.in;
1015         args->in_args[1].size = count;
1016         args->out_numargs = 1;
1017         args->out_args[0].size = sizeof(ia->write.out);
1018         args->out_args[0].value = &ia->write.out;
1019 }
1020
1021 static unsigned int fuse_write_flags(struct kiocb *iocb)
1022 {
1023         unsigned int flags = iocb->ki_filp->f_flags;
1024
1025         if (iocb->ki_flags & IOCB_DSYNC)
1026                 flags |= O_DSYNC;
1027         if (iocb->ki_flags & IOCB_SYNC)
1028                 flags |= O_SYNC;
1029
1030         return flags;
1031 }
1032
1033 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1034                                size_t count, fl_owner_t owner)
1035 {
1036         struct kiocb *iocb = ia->io->iocb;
1037         struct file *file = iocb->ki_filp;
1038         struct fuse_file *ff = file->private_data;
1039         struct fuse_conn *fc = ff->fc;
1040         struct fuse_write_in *inarg = &ia->write.in;
1041         ssize_t err;
1042
1043         fuse_write_args_fill(ia, ff, pos, count);
1044         inarg->flags = fuse_write_flags(iocb);
1045         if (owner != NULL) {
1046                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1047                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
1048         }
1049
1050         if (ia->io->async)
1051                 return fuse_async_req_send(fc, ia, count);
1052
1053         err = fuse_simple_request(fc, &ia->ap.args);
1054         if (!err && ia->write.out.size > count)
1055                 err = -EIO;
1056
1057         return err ?: ia->write.out.size;
1058 }
1059
1060 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1061 {
1062         struct fuse_conn *fc = get_fuse_conn(inode);
1063         struct fuse_inode *fi = get_fuse_inode(inode);
1064         bool ret = false;
1065
1066         spin_lock(&fi->lock);
1067         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1068         if (pos > inode->i_size) {
1069                 i_size_write(inode, pos);
1070                 ret = true;
1071         }
1072         spin_unlock(&fi->lock);
1073
1074         return ret;
1075 }
1076
1077 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1078                                      struct kiocb *iocb, struct inode *inode,
1079                                      loff_t pos, size_t count)
1080 {
1081         struct fuse_args_pages *ap = &ia->ap;
1082         struct file *file = iocb->ki_filp;
1083         struct fuse_file *ff = file->private_data;
1084         struct fuse_conn *fc = ff->fc;
1085         unsigned int offset, i;
1086         int err;
1087
1088         for (i = 0; i < ap->num_pages; i++)
1089                 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1090
1091         fuse_write_args_fill(ia, ff, pos, count);
1092         ia->write.in.flags = fuse_write_flags(iocb);
1093
1094         err = fuse_simple_request(fc, &ap->args);
1095
1096         offset = ap->descs[0].offset;
1097         count = ia->write.out.size;
1098         for (i = 0; i < ap->num_pages; i++) {
1099                 struct page *page = ap->pages[i];
1100
1101                 if (!err && !offset && count >= PAGE_SIZE)
1102                         SetPageUptodate(page);
1103
1104                 if (count > PAGE_SIZE - offset)
1105                         count -= PAGE_SIZE - offset;
1106                 else
1107                         count = 0;
1108                 offset = 0;
1109
1110                 unlock_page(page);
1111                 put_page(page);
1112         }
1113
1114         return err;
1115 }
1116
1117 static ssize_t fuse_fill_write_pages(struct fuse_args_pages *ap,
1118                                      struct address_space *mapping,
1119                                      struct iov_iter *ii, loff_t pos,
1120                                      unsigned int max_pages)
1121 {
1122         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1123         unsigned offset = pos & (PAGE_SIZE - 1);
1124         size_t count = 0;
1125         int err;
1126
1127         ap->args.in_pages = true;
1128         ap->descs[0].offset = offset;
1129
1130         do {
1131                 size_t tmp;
1132                 struct page *page;
1133                 pgoff_t index = pos >> PAGE_SHIFT;
1134                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1135                                      iov_iter_count(ii));
1136
1137                 bytes = min_t(size_t, bytes, fc->max_write - count);
1138
1139  again:
1140                 err = -EFAULT;
1141                 if (iov_iter_fault_in_readable(ii, bytes))
1142                         break;
1143
1144                 err = -ENOMEM;
1145                 page = grab_cache_page_write_begin(mapping, index, 0);
1146                 if (!page)
1147                         break;
1148
1149                 if (mapping_writably_mapped(mapping))
1150                         flush_dcache_page(page);
1151
1152                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1153                 flush_dcache_page(page);
1154
1155                 iov_iter_advance(ii, tmp);
1156                 if (!tmp) {
1157                         unlock_page(page);
1158                         put_page(page);
1159                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1160                         goto again;
1161                 }
1162
1163                 err = 0;
1164                 ap->pages[ap->num_pages] = page;
1165                 ap->descs[ap->num_pages].length = tmp;
1166                 ap->num_pages++;
1167
1168                 count += tmp;
1169                 pos += tmp;
1170                 offset += tmp;
1171                 if (offset == PAGE_SIZE)
1172                         offset = 0;
1173
1174                 if (!fc->big_writes)
1175                         break;
1176         } while (iov_iter_count(ii) && count < fc->max_write &&
1177                  ap->num_pages < max_pages && offset == 0);
1178
1179         return count > 0 ? count : err;
1180 }
1181
1182 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1183                                      unsigned int max_pages)
1184 {
1185         return min_t(unsigned int,
1186                      ((pos + len - 1) >> PAGE_SHIFT) -
1187                      (pos >> PAGE_SHIFT) + 1,
1188                      max_pages);
1189 }
1190
1191 static ssize_t fuse_perform_write(struct kiocb *iocb,
1192                                   struct address_space *mapping,
1193                                   struct iov_iter *ii, loff_t pos)
1194 {
1195         struct inode *inode = mapping->host;
1196         struct fuse_conn *fc = get_fuse_conn(inode);
1197         struct fuse_inode *fi = get_fuse_inode(inode);
1198         int err = 0;
1199         ssize_t res = 0;
1200
1201         if (inode->i_size < pos + iov_iter_count(ii))
1202                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1203
1204         do {
1205                 ssize_t count;
1206                 struct fuse_io_args ia = {};
1207                 struct fuse_args_pages *ap = &ia.ap;
1208                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1209                                                       fc->max_pages);
1210
1211                 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1212                 if (!ap->pages) {
1213                         err = -ENOMEM;
1214                         break;
1215                 }
1216
1217                 count = fuse_fill_write_pages(ap, mapping, ii, pos, nr_pages);
1218                 if (count <= 0) {
1219                         err = count;
1220                 } else {
1221                         err = fuse_send_write_pages(&ia, iocb, inode,
1222                                                     pos, count);
1223                         if (!err) {
1224                                 size_t num_written = ia.write.out.size;
1225
1226                                 res += num_written;
1227                                 pos += num_written;
1228
1229                                 /* break out of the loop on short write */
1230                                 if (num_written != count)
1231                                         err = -EIO;
1232                         }
1233                 }
1234                 kfree(ap->pages);
1235         } while (!err && iov_iter_count(ii));
1236
1237         if (res > 0)
1238                 fuse_write_update_size(inode, pos);
1239
1240         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1241         fuse_invalidate_attr(inode);
1242
1243         return res > 0 ? res : err;
1244 }
1245
1246 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1247 {
1248         struct file *file = iocb->ki_filp;
1249         struct address_space *mapping = file->f_mapping;
1250         ssize_t written = 0;
1251         ssize_t written_buffered = 0;
1252         struct inode *inode = mapping->host;
1253         ssize_t err;
1254         loff_t endbyte = 0;
1255
1256         if (get_fuse_conn(inode)->writeback_cache) {
1257                 /* Update size (EOF optimization) and mode (SUID clearing) */
1258                 err = fuse_update_attributes(mapping->host, file);
1259                 if (err)
1260                         return err;
1261
1262                 return generic_file_write_iter(iocb, from);
1263         }
1264
1265         inode_lock(inode);
1266
1267         /* We can write back this queue in page reclaim */
1268         current->backing_dev_info = inode_to_bdi(inode);
1269
1270         err = generic_write_checks(iocb, from);
1271         if (err <= 0)
1272                 goto out;
1273
1274         err = file_remove_privs(file);
1275         if (err)
1276                 goto out;
1277
1278         err = file_update_time(file);
1279         if (err)
1280                 goto out;
1281
1282         if (iocb->ki_flags & IOCB_DIRECT) {
1283                 loff_t pos = iocb->ki_pos;
1284                 written = generic_file_direct_write(iocb, from);
1285                 if (written < 0 || !iov_iter_count(from))
1286                         goto out;
1287
1288                 pos += written;
1289
1290                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1291                 if (written_buffered < 0) {
1292                         err = written_buffered;
1293                         goto out;
1294                 }
1295                 endbyte = pos + written_buffered - 1;
1296
1297                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1298                                                    endbyte);
1299                 if (err)
1300                         goto out;
1301
1302                 invalidate_mapping_pages(file->f_mapping,
1303                                          pos >> PAGE_SHIFT,
1304                                          endbyte >> PAGE_SHIFT);
1305
1306                 written += written_buffered;
1307                 iocb->ki_pos = pos + written_buffered;
1308         } else {
1309                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1310                 if (written >= 0)
1311                         iocb->ki_pos += written;
1312         }
1313 out:
1314         current->backing_dev_info = NULL;
1315         inode_unlock(inode);
1316         if (written > 0)
1317                 written = generic_write_sync(iocb, written);
1318
1319         return written ? written : err;
1320 }
1321
1322 static inline void fuse_page_descs_length_init(struct fuse_page_desc *descs,
1323                                                unsigned int index,
1324                                                unsigned int nr_pages)
1325 {
1326         int i;
1327
1328         for (i = index; i < index + nr_pages; i++)
1329                 descs[i].length = PAGE_SIZE - descs[i].offset;
1330 }
1331
1332 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1333 {
1334         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1335 }
1336
1337 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1338                                         size_t max_size)
1339 {
1340         return min(iov_iter_single_seg_count(ii), max_size);
1341 }
1342
1343 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1344                                size_t *nbytesp, int write,
1345                                unsigned int max_pages)
1346 {
1347         size_t nbytes = 0;  /* # bytes already packed in req */
1348         ssize_t ret = 0;
1349
1350         /* Special case for kernel I/O: can copy directly into the buffer */
1351         if (iov_iter_is_kvec(ii)) {
1352                 unsigned long user_addr = fuse_get_user_addr(ii);
1353                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1354
1355                 if (write)
1356                         ap->args.in_args[1].value = (void *) user_addr;
1357                 else
1358                         ap->args.out_args[0].value = (void *) user_addr;
1359
1360                 iov_iter_advance(ii, frag_size);
1361                 *nbytesp = frag_size;
1362                 return 0;
1363         }
1364
1365         while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1366                 unsigned npages;
1367                 size_t start;
1368                 ret = iov_iter_get_pages(ii, &ap->pages[ap->num_pages],
1369                                         *nbytesp - nbytes,
1370                                         max_pages - ap->num_pages,
1371                                         &start);
1372                 if (ret < 0)
1373                         break;
1374
1375                 iov_iter_advance(ii, ret);
1376                 nbytes += ret;
1377
1378                 ret += start;
1379                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1380
1381                 ap->descs[ap->num_pages].offset = start;
1382                 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1383
1384                 ap->num_pages += npages;
1385                 ap->descs[ap->num_pages - 1].length -=
1386                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1387         }
1388
1389         if (write)
1390                 ap->args.in_pages = 1;
1391         else
1392                 ap->args.out_pages = 1;
1393
1394         *nbytesp = nbytes;
1395
1396         return ret < 0 ? ret : 0;
1397 }
1398
1399 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1400                        loff_t *ppos, int flags)
1401 {
1402         int write = flags & FUSE_DIO_WRITE;
1403         int cuse = flags & FUSE_DIO_CUSE;
1404         struct file *file = io->iocb->ki_filp;
1405         struct inode *inode = file->f_mapping->host;
1406         struct fuse_file *ff = file->private_data;
1407         struct fuse_conn *fc = ff->fc;
1408         size_t nmax = write ? fc->max_write : fc->max_read;
1409         loff_t pos = *ppos;
1410         size_t count = iov_iter_count(iter);
1411         pgoff_t idx_from = pos >> PAGE_SHIFT;
1412         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1413         ssize_t res = 0;
1414         int err = 0;
1415         struct fuse_io_args *ia;
1416         unsigned int max_pages;
1417
1418         max_pages = iov_iter_npages(iter, fc->max_pages);
1419         ia = fuse_io_alloc(io, max_pages);
1420         if (!ia)
1421                 return -ENOMEM;
1422
1423         ia->io = io;
1424         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1425                 if (!write)
1426                         inode_lock(inode);
1427                 fuse_sync_writes(inode);
1428                 if (!write)
1429                         inode_unlock(inode);
1430         }
1431
1432         io->should_dirty = !write && iter_is_iovec(iter);
1433         while (count) {
1434                 ssize_t nres;
1435                 fl_owner_t owner = current->files;
1436                 size_t nbytes = min(count, nmax);
1437
1438                 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1439                                           max_pages);
1440                 if (err && !nbytes)
1441                         break;
1442
1443                 if (write) {
1444                         if (!capable(CAP_FSETID))
1445                                 ia->write.in.write_flags |= FUSE_WRITE_KILL_PRIV;
1446
1447                         nres = fuse_send_write(ia, pos, nbytes, owner);
1448                 } else {
1449                         nres = fuse_send_read(ia, pos, nbytes, owner);
1450                 }
1451
1452                 if (!io->async || nres < 0) {
1453                         fuse_release_user_pages(&ia->ap, io->should_dirty);
1454                         fuse_io_free(ia);
1455                 }
1456                 ia = NULL;
1457                 if (nres < 0) {
1458                         err = nres;
1459                         break;
1460                 }
1461                 WARN_ON(nres > nbytes);
1462
1463                 count -= nres;
1464                 res += nres;
1465                 pos += nres;
1466                 if (nres != nbytes)
1467                         break;
1468                 if (count) {
1469                         max_pages = iov_iter_npages(iter, fc->max_pages);
1470                         ia = fuse_io_alloc(io, max_pages);
1471                         if (!ia)
1472                                 break;
1473                 }
1474         }
1475         if (ia)
1476                 fuse_io_free(ia);
1477         if (res > 0)
1478                 *ppos = pos;
1479
1480         return res > 0 ? res : err;
1481 }
1482 EXPORT_SYMBOL_GPL(fuse_direct_io);
1483
1484 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1485                                   struct iov_iter *iter,
1486                                   loff_t *ppos)
1487 {
1488         ssize_t res;
1489         struct inode *inode = file_inode(io->iocb->ki_filp);
1490
1491         res = fuse_direct_io(io, iter, ppos, 0);
1492
1493         fuse_invalidate_atime(inode);
1494
1495         return res;
1496 }
1497
1498 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1499
1500 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1501 {
1502         ssize_t res;
1503
1504         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1505                 res = fuse_direct_IO(iocb, to);
1506         } else {
1507                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1508
1509                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1510         }
1511
1512         return res;
1513 }
1514
1515 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1516 {
1517         struct inode *inode = file_inode(iocb->ki_filp);
1518         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1519         ssize_t res;
1520
1521         /* Don't allow parallel writes to the same file */
1522         inode_lock(inode);
1523         res = generic_write_checks(iocb, from);
1524         if (res > 0) {
1525                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1526                         res = fuse_direct_IO(iocb, from);
1527                 } else {
1528                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1529                                              FUSE_DIO_WRITE);
1530                 }
1531         }
1532         fuse_invalidate_attr(inode);
1533         if (res > 0)
1534                 fuse_write_update_size(inode, iocb->ki_pos);
1535         inode_unlock(inode);
1536
1537         return res;
1538 }
1539
1540 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1541 {
1542         struct file *file = iocb->ki_filp;
1543         struct fuse_file *ff = file->private_data;
1544
1545         if (is_bad_inode(file_inode(file)))
1546                 return -EIO;
1547
1548         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1549                 return fuse_cache_read_iter(iocb, to);
1550         else
1551                 return fuse_direct_read_iter(iocb, to);
1552 }
1553
1554 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1555 {
1556         struct file *file = iocb->ki_filp;
1557         struct fuse_file *ff = file->private_data;
1558
1559         if (is_bad_inode(file_inode(file)))
1560                 return -EIO;
1561
1562         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1563                 return fuse_cache_write_iter(iocb, from);
1564         else
1565                 return fuse_direct_write_iter(iocb, from);
1566 }
1567
1568 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1569 {
1570         struct fuse_args_pages *ap = &wpa->ia.ap;
1571         int i;
1572
1573         for (i = 0; i < ap->num_pages; i++)
1574                 __free_page(ap->pages[i]);
1575
1576         if (wpa->ia.ff)
1577                 fuse_file_put(wpa->ia.ff, false, false);
1578
1579         kfree(ap->pages);
1580         kfree(wpa);
1581 }
1582
1583 static void fuse_writepage_finish(struct fuse_conn *fc,
1584                                   struct fuse_writepage_args *wpa)
1585 {
1586         struct fuse_args_pages *ap = &wpa->ia.ap;
1587         struct inode *inode = wpa->inode;
1588         struct fuse_inode *fi = get_fuse_inode(inode);
1589         struct backing_dev_info *bdi = inode_to_bdi(inode);
1590         int i;
1591
1592         list_del(&wpa->writepages_entry);
1593         for (i = 0; i < ap->num_pages; i++) {
1594                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1595                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1596                 wb_writeout_inc(&bdi->wb);
1597         }
1598         wake_up(&fi->page_waitq);
1599 }
1600
1601 /* Called under fi->lock, may release and reacquire it */
1602 static void fuse_send_writepage(struct fuse_conn *fc,
1603                                 struct fuse_writepage_args *wpa, loff_t size)
1604 __releases(fi->lock)
1605 __acquires(fi->lock)
1606 {
1607         struct fuse_writepage_args *aux, *next;
1608         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1609         struct fuse_write_in *inarg = &wpa->ia.write.in;
1610         struct fuse_args *args = &wpa->ia.ap.args;
1611         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1612         int err;
1613
1614         fi->writectr++;
1615         if (inarg->offset + data_size <= size) {
1616                 inarg->size = data_size;
1617         } else if (inarg->offset < size) {
1618                 inarg->size = size - inarg->offset;
1619         } else {
1620                 /* Got truncated off completely */
1621                 goto out_free;
1622         }
1623
1624         args->in_args[1].size = inarg->size;
1625         args->force = true;
1626         args->nocreds = true;
1627
1628         err = fuse_simple_background(fc, args, GFP_ATOMIC);
1629         if (err == -ENOMEM) {
1630                 spin_unlock(&fi->lock);
1631                 err = fuse_simple_background(fc, args, GFP_NOFS | __GFP_NOFAIL);
1632                 spin_lock(&fi->lock);
1633         }
1634
1635         /* Fails on broken connection only */
1636         if (unlikely(err))
1637                 goto out_free;
1638
1639         return;
1640
1641  out_free:
1642         fi->writectr--;
1643         fuse_writepage_finish(fc, wpa);
1644         spin_unlock(&fi->lock);
1645
1646         /* After fuse_writepage_finish() aux request list is private */
1647         for (aux = wpa->next; aux; aux = next) {
1648                 next = aux->next;
1649                 aux->next = NULL;
1650                 fuse_writepage_free(aux);
1651         }
1652
1653         fuse_writepage_free(wpa);
1654         spin_lock(&fi->lock);
1655 }
1656
1657 /*
1658  * If fi->writectr is positive (no truncate or fsync going on) send
1659  * all queued writepage requests.
1660  *
1661  * Called with fi->lock
1662  */
1663 void fuse_flush_writepages(struct inode *inode)
1664 __releases(fi->lock)
1665 __acquires(fi->lock)
1666 {
1667         struct fuse_conn *fc = get_fuse_conn(inode);
1668         struct fuse_inode *fi = get_fuse_inode(inode);
1669         loff_t crop = i_size_read(inode);
1670         struct fuse_writepage_args *wpa;
1671
1672         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1673                 wpa = list_entry(fi->queued_writes.next,
1674                                  struct fuse_writepage_args, queue_entry);
1675                 list_del_init(&wpa->queue_entry);
1676                 fuse_send_writepage(fc, wpa, crop);
1677         }
1678 }
1679
1680 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_args *args,
1681                                int error)
1682 {
1683         struct fuse_writepage_args *wpa =
1684                 container_of(args, typeof(*wpa), ia.ap.args);
1685         struct inode *inode = wpa->inode;
1686         struct fuse_inode *fi = get_fuse_inode(inode);
1687
1688         mapping_set_error(inode->i_mapping, error);
1689         spin_lock(&fi->lock);
1690         while (wpa->next) {
1691                 struct fuse_conn *fc = get_fuse_conn(inode);
1692                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1693                 struct fuse_writepage_args *next = wpa->next;
1694
1695                 wpa->next = next->next;
1696                 next->next = NULL;
1697                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1698                 list_add(&next->writepages_entry, &fi->writepages);
1699
1700                 /*
1701                  * Skip fuse_flush_writepages() to make it easy to crop requests
1702                  * based on primary request size.
1703                  *
1704                  * 1st case (trivial): there are no concurrent activities using
1705                  * fuse_set/release_nowrite.  Then we're on safe side because
1706                  * fuse_flush_writepages() would call fuse_send_writepage()
1707                  * anyway.
1708                  *
1709                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1710                  * now for completion of all in-flight requests.  This happens
1711                  * rarely and no more than once per page, so this should be
1712                  * okay.
1713                  *
1714                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1715                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1716                  * that fuse_set_nowrite returned implies that all in-flight
1717                  * requests were completed along with all of their secondary
1718                  * requests.  Further primary requests are blocked by negative
1719                  * writectr.  Hence there cannot be any in-flight requests and
1720                  * no invocations of fuse_writepage_end() while we're in
1721                  * fuse_set_nowrite..fuse_release_nowrite section.
1722                  */
1723                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1724         }
1725         fi->writectr--;
1726         fuse_writepage_finish(fc, wpa);
1727         spin_unlock(&fi->lock);
1728         fuse_writepage_free(wpa);
1729 }
1730
1731 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1732                                                struct fuse_inode *fi)
1733 {
1734         struct fuse_file *ff = NULL;
1735
1736         spin_lock(&fi->lock);
1737         if (!list_empty(&fi->write_files)) {
1738                 ff = list_entry(fi->write_files.next, struct fuse_file,
1739                                 write_entry);
1740                 fuse_file_get(ff);
1741         }
1742         spin_unlock(&fi->lock);
1743
1744         return ff;
1745 }
1746
1747 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1748                                              struct fuse_inode *fi)
1749 {
1750         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1751         WARN_ON(!ff);
1752         return ff;
1753 }
1754
1755 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1756 {
1757         struct fuse_conn *fc = get_fuse_conn(inode);
1758         struct fuse_inode *fi = get_fuse_inode(inode);
1759         struct fuse_file *ff;
1760         int err;
1761
1762         ff = __fuse_write_file_get(fc, fi);
1763         err = fuse_flush_times(inode, ff);
1764         if (ff)
1765                 fuse_file_put(ff, false, false);
1766
1767         return err;
1768 }
1769
1770 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1771 {
1772         struct fuse_writepage_args *wpa;
1773         struct fuse_args_pages *ap;
1774
1775         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1776         if (wpa) {
1777                 ap = &wpa->ia.ap;
1778                 ap->num_pages = 0;
1779                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1780                 if (!ap->pages) {
1781                         kfree(wpa);
1782                         wpa = NULL;
1783                 }
1784         }
1785         return wpa;
1786
1787 }
1788
1789 static int fuse_writepage_locked(struct page *page)
1790 {
1791         struct address_space *mapping = page->mapping;
1792         struct inode *inode = mapping->host;
1793         struct fuse_conn *fc = get_fuse_conn(inode);
1794         struct fuse_inode *fi = get_fuse_inode(inode);
1795         struct fuse_writepage_args *wpa;
1796         struct fuse_args_pages *ap;
1797         struct page *tmp_page;
1798         int error = -ENOMEM;
1799
1800         set_page_writeback(page);
1801
1802         wpa = fuse_writepage_args_alloc();
1803         if (!wpa)
1804                 goto err;
1805         ap = &wpa->ia.ap;
1806
1807         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1808         if (!tmp_page)
1809                 goto err_free;
1810
1811         error = -EIO;
1812         wpa->ia.ff = fuse_write_file_get(fc, fi);
1813         if (!wpa->ia.ff)
1814                 goto err_nofile;
1815
1816         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1817
1818         copy_highpage(tmp_page, page);
1819         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1820         wpa->next = NULL;
1821         ap->args.in_pages = true;
1822         ap->num_pages = 1;
1823         ap->pages[0] = tmp_page;
1824         ap->descs[0].offset = 0;
1825         ap->descs[0].length = PAGE_SIZE;
1826         ap->args.end = fuse_writepage_end;
1827         wpa->inode = inode;
1828
1829         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1830         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1831
1832         spin_lock(&fi->lock);
1833         list_add(&wpa->writepages_entry, &fi->writepages);
1834         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1835         fuse_flush_writepages(inode);
1836         spin_unlock(&fi->lock);
1837
1838         end_page_writeback(page);
1839
1840         return 0;
1841
1842 err_nofile:
1843         __free_page(tmp_page);
1844 err_free:
1845         kfree(wpa);
1846 err:
1847         mapping_set_error(page->mapping, error);
1848         end_page_writeback(page);
1849         return error;
1850 }
1851
1852 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1853 {
1854         int err;
1855
1856         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1857                 /*
1858                  * ->writepages() should be called for sync() and friends.  We
1859                  * should only get here on direct reclaim and then we are
1860                  * allowed to skip a page which is already in flight
1861                  */
1862                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1863
1864                 redirty_page_for_writepage(wbc, page);
1865                 unlock_page(page);
1866                 return 0;
1867         }
1868
1869         err = fuse_writepage_locked(page);
1870         unlock_page(page);
1871
1872         return err;
1873 }
1874
1875 struct fuse_fill_wb_data {
1876         struct fuse_writepage_args *wpa;
1877         struct fuse_file *ff;
1878         struct inode *inode;
1879         struct page **orig_pages;
1880         unsigned int max_pages;
1881 };
1882
1883 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
1884 {
1885         struct fuse_args_pages *ap = &data->wpa->ia.ap;
1886         struct fuse_conn *fc = get_fuse_conn(data->inode);
1887         struct page **pages;
1888         struct fuse_page_desc *descs;
1889         unsigned int npages = min_t(unsigned int,
1890                                     max_t(unsigned int, data->max_pages * 2,
1891                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
1892                                     fc->max_pages);
1893         WARN_ON(npages <= data->max_pages);
1894
1895         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
1896         if (!pages)
1897                 return false;
1898
1899         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
1900         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
1901         kfree(ap->pages);
1902         ap->pages = pages;
1903         ap->descs = descs;
1904         data->max_pages = npages;
1905
1906         return true;
1907 }
1908
1909 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1910 {
1911         struct fuse_writepage_args *wpa = data->wpa;
1912         struct inode *inode = data->inode;
1913         struct fuse_inode *fi = get_fuse_inode(inode);
1914         int num_pages = wpa->ia.ap.num_pages;
1915         int i;
1916
1917         wpa->ia.ff = fuse_file_get(data->ff);
1918         spin_lock(&fi->lock);
1919         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1920         fuse_flush_writepages(inode);
1921         spin_unlock(&fi->lock);
1922
1923         for (i = 0; i < num_pages; i++)
1924                 end_page_writeback(data->orig_pages[i]);
1925 }
1926
1927 /*
1928  * First recheck under fi->lock if the offending offset is still under
1929  * writeback.  If yes, then iterate auxiliary write requests, to see if there's
1930  * one already added for a page at this offset.  If there's none, then insert
1931  * this new request onto the auxiliary list, otherwise reuse the existing one by
1932  * copying the new page contents over to the old temporary page.
1933  */
1934 static bool fuse_writepage_in_flight(struct fuse_writepage_args *new_wpa,
1935                                      struct page *page)
1936 {
1937         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
1938         struct fuse_writepage_args *tmp;
1939         struct fuse_writepage_args *old_wpa;
1940         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
1941
1942         WARN_ON(new_ap->num_pages != 0);
1943
1944         spin_lock(&fi->lock);
1945         list_del(&new_wpa->writepages_entry);
1946         old_wpa = fuse_find_writeback(fi, page->index, page->index);
1947         if (!old_wpa) {
1948                 list_add(&new_wpa->writepages_entry, &fi->writepages);
1949                 spin_unlock(&fi->lock);
1950                 return false;
1951         }
1952
1953         new_ap->num_pages = 1;
1954         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
1955                 pgoff_t curr_index;
1956
1957                 WARN_ON(tmp->inode != new_wpa->inode);
1958                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
1959                 if (curr_index == page->index) {
1960                         WARN_ON(tmp->ia.ap.num_pages != 1);
1961                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
1962                         break;
1963                 }
1964         }
1965
1966         if (!tmp) {
1967                 new_wpa->next = old_wpa->next;
1968                 old_wpa->next = new_wpa;
1969         }
1970
1971         spin_unlock(&fi->lock);
1972
1973         if (tmp) {
1974                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
1975
1976                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1977                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
1978                 wb_writeout_inc(&bdi->wb);
1979                 fuse_writepage_free(new_wpa);
1980         }
1981
1982         return true;
1983 }
1984
1985 static int fuse_writepages_fill(struct page *page,
1986                 struct writeback_control *wbc, void *_data)
1987 {
1988         struct fuse_fill_wb_data *data = _data;
1989         struct fuse_writepage_args *wpa = data->wpa;
1990         struct fuse_args_pages *ap = &wpa->ia.ap;
1991         struct inode *inode = data->inode;
1992         struct fuse_inode *fi = get_fuse_inode(inode);
1993         struct fuse_conn *fc = get_fuse_conn(inode);
1994         struct page *tmp_page;
1995         bool is_writeback;
1996         int err;
1997
1998         if (!data->ff) {
1999                 err = -EIO;
2000                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
2001                 if (!data->ff)
2002                         goto out_unlock;
2003         }
2004
2005         /*
2006          * Being under writeback is unlikely but possible.  For example direct
2007          * read to an mmaped fuse file will set the page dirty twice; once when
2008          * the pages are faulted with get_user_pages(), and then after the read
2009          * completed.
2010          */
2011         is_writeback = fuse_page_is_writeback(inode, page->index);
2012
2013         if (wpa && ap->num_pages &&
2014             (is_writeback || ap->num_pages == fc->max_pages ||
2015              (ap->num_pages + 1) * PAGE_SIZE > fc->max_write ||
2016              data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)) {
2017                 fuse_writepages_send(data);
2018                 data->wpa = NULL;
2019         } else if (wpa && ap->num_pages == data->max_pages) {
2020                 if (!fuse_pages_realloc(data)) {
2021                         fuse_writepages_send(data);
2022                         data->wpa = NULL;
2023                 }
2024         }
2025
2026         err = -ENOMEM;
2027         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2028         if (!tmp_page)
2029                 goto out_unlock;
2030
2031         /*
2032          * The page must not be redirtied until the writeout is completed
2033          * (i.e. userspace has sent a reply to the write request).  Otherwise
2034          * there could be more than one temporary page instance for each real
2035          * page.
2036          *
2037          * This is ensured by holding the page lock in page_mkwrite() while
2038          * checking fuse_page_is_writeback().  We already hold the page lock
2039          * since clear_page_dirty_for_io() and keep it held until we add the
2040          * request to the fi->writepages list and increment ap->num_pages.
2041          * After this fuse_page_is_writeback() will indicate that the page is
2042          * under writeback, so we can release the page lock.
2043          */
2044         if (data->wpa == NULL) {
2045                 struct fuse_inode *fi = get_fuse_inode(inode);
2046
2047                 err = -ENOMEM;
2048                 wpa = fuse_writepage_args_alloc();
2049                 if (!wpa) {
2050                         __free_page(tmp_page);
2051                         goto out_unlock;
2052                 }
2053                 data->max_pages = 1;
2054
2055                 ap = &wpa->ia.ap;
2056                 fuse_write_args_fill(&wpa->ia, data->ff, page_offset(page), 0);
2057                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2058                 wpa->next = NULL;
2059                 ap->args.in_pages = true;
2060                 ap->args.end = fuse_writepage_end;
2061                 ap->num_pages = 0;
2062                 wpa->inode = inode;
2063
2064                 spin_lock(&fi->lock);
2065                 list_add(&wpa->writepages_entry, &fi->writepages);
2066                 spin_unlock(&fi->lock);
2067
2068                 data->wpa = wpa;
2069         }
2070         set_page_writeback(page);
2071
2072         copy_highpage(tmp_page, page);
2073         ap->pages[ap->num_pages] = tmp_page;
2074         ap->descs[ap->num_pages].offset = 0;
2075         ap->descs[ap->num_pages].length = PAGE_SIZE;
2076
2077         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2078         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2079
2080         err = 0;
2081         if (is_writeback && fuse_writepage_in_flight(wpa, page)) {
2082                 end_page_writeback(page);
2083                 data->wpa = NULL;
2084                 goto out_unlock;
2085         }
2086         data->orig_pages[ap->num_pages] = page;
2087
2088         /*
2089          * Protected by fi->lock against concurrent access by
2090          * fuse_page_is_writeback().
2091          */
2092         spin_lock(&fi->lock);
2093         ap->num_pages++;
2094         spin_unlock(&fi->lock);
2095
2096 out_unlock:
2097         unlock_page(page);
2098
2099         return err;
2100 }
2101
2102 static int fuse_writepages(struct address_space *mapping,
2103                            struct writeback_control *wbc)
2104 {
2105         struct inode *inode = mapping->host;
2106         struct fuse_conn *fc = get_fuse_conn(inode);
2107         struct fuse_fill_wb_data data;
2108         int err;
2109
2110         err = -EIO;
2111         if (is_bad_inode(inode))
2112                 goto out;
2113
2114         data.inode = inode;
2115         data.wpa = NULL;
2116         data.ff = NULL;
2117
2118         err = -ENOMEM;
2119         data.orig_pages = kcalloc(fc->max_pages,
2120                                   sizeof(struct page *),
2121                                   GFP_NOFS);
2122         if (!data.orig_pages)
2123                 goto out;
2124
2125         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2126         if (data.wpa) {
2127                 /* Ignore errors if we can write at least one page */
2128                 WARN_ON(!data.wpa->ia.ap.num_pages);
2129                 fuse_writepages_send(&data);
2130                 err = 0;
2131         }
2132         if (data.ff)
2133                 fuse_file_put(data.ff, false, false);
2134
2135         kfree(data.orig_pages);
2136 out:
2137         return err;
2138 }
2139
2140 /*
2141  * It's worthy to make sure that space is reserved on disk for the write,
2142  * but how to implement it without killing performance need more thinking.
2143  */
2144 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2145                 loff_t pos, unsigned len, unsigned flags,
2146                 struct page **pagep, void **fsdata)
2147 {
2148         pgoff_t index = pos >> PAGE_SHIFT;
2149         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2150         struct page *page;
2151         loff_t fsize;
2152         int err = -ENOMEM;
2153
2154         WARN_ON(!fc->writeback_cache);
2155
2156         page = grab_cache_page_write_begin(mapping, index, flags);
2157         if (!page)
2158                 goto error;
2159
2160         fuse_wait_on_page_writeback(mapping->host, page->index);
2161
2162         if (PageUptodate(page) || len == PAGE_SIZE)
2163                 goto success;
2164         /*
2165          * Check if the start this page comes after the end of file, in which
2166          * case the readpage can be optimized away.
2167          */
2168         fsize = i_size_read(mapping->host);
2169         if (fsize <= (pos & PAGE_MASK)) {
2170                 size_t off = pos & ~PAGE_MASK;
2171                 if (off)
2172                         zero_user_segment(page, 0, off);
2173                 goto success;
2174         }
2175         err = fuse_do_readpage(file, page);
2176         if (err)
2177                 goto cleanup;
2178 success:
2179         *pagep = page;
2180         return 0;
2181
2182 cleanup:
2183         unlock_page(page);
2184         put_page(page);
2185 error:
2186         return err;
2187 }
2188
2189 static int fuse_write_end(struct file *file, struct address_space *mapping,
2190                 loff_t pos, unsigned len, unsigned copied,
2191                 struct page *page, void *fsdata)
2192 {
2193         struct inode *inode = page->mapping->host;
2194
2195         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2196         if (!copied)
2197                 goto unlock;
2198
2199         if (!PageUptodate(page)) {
2200                 /* Zero any unwritten bytes at the end of the page */
2201                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2202                 if (endoff)
2203                         zero_user_segment(page, endoff, PAGE_SIZE);
2204                 SetPageUptodate(page);
2205         }
2206
2207         fuse_write_update_size(inode, pos + copied);
2208         set_page_dirty(page);
2209
2210 unlock:
2211         unlock_page(page);
2212         put_page(page);
2213
2214         return copied;
2215 }
2216
2217 static int fuse_launder_page(struct page *page)
2218 {
2219         int err = 0;
2220         if (clear_page_dirty_for_io(page)) {
2221                 struct inode *inode = page->mapping->host;
2222                 err = fuse_writepage_locked(page);
2223                 if (!err)
2224                         fuse_wait_on_page_writeback(inode, page->index);
2225         }
2226         return err;
2227 }
2228
2229 /*
2230  * Write back dirty pages now, because there may not be any suitable
2231  * open files later
2232  */
2233 static void fuse_vma_close(struct vm_area_struct *vma)
2234 {
2235         filemap_write_and_wait(vma->vm_file->f_mapping);
2236 }
2237
2238 /*
2239  * Wait for writeback against this page to complete before allowing it
2240  * to be marked dirty again, and hence written back again, possibly
2241  * before the previous writepage completed.
2242  *
2243  * Block here, instead of in ->writepage(), so that the userspace fs
2244  * can only block processes actually operating on the filesystem.
2245  *
2246  * Otherwise unprivileged userspace fs would be able to block
2247  * unrelated:
2248  *
2249  * - page migration
2250  * - sync(2)
2251  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2252  */
2253 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2254 {
2255         struct page *page = vmf->page;
2256         struct inode *inode = file_inode(vmf->vma->vm_file);
2257
2258         file_update_time(vmf->vma->vm_file);
2259         lock_page(page);
2260         if (page->mapping != inode->i_mapping) {
2261                 unlock_page(page);
2262                 return VM_FAULT_NOPAGE;
2263         }
2264
2265         fuse_wait_on_page_writeback(inode, page->index);
2266         return VM_FAULT_LOCKED;
2267 }
2268
2269 static const struct vm_operations_struct fuse_file_vm_ops = {
2270         .close          = fuse_vma_close,
2271         .fault          = filemap_fault,
2272         .map_pages      = filemap_map_pages,
2273         .page_mkwrite   = fuse_page_mkwrite,
2274 };
2275
2276 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2277 {
2278         struct fuse_file *ff = file->private_data;
2279
2280         if (ff->open_flags & FOPEN_DIRECT_IO) {
2281                 /* Can't provide the coherency needed for MAP_SHARED */
2282                 if (vma->vm_flags & VM_MAYSHARE)
2283                         return -ENODEV;
2284
2285                 invalidate_inode_pages2(file->f_mapping);
2286
2287                 return generic_file_mmap(file, vma);
2288         }
2289
2290         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2291                 fuse_link_write_file(file);
2292
2293         file_accessed(file);
2294         vma->vm_ops = &fuse_file_vm_ops;
2295         return 0;
2296 }
2297
2298 static int convert_fuse_file_lock(struct fuse_conn *fc,
2299                                   const struct fuse_file_lock *ffl,
2300                                   struct file_lock *fl)
2301 {
2302         switch (ffl->type) {
2303         case F_UNLCK:
2304                 break;
2305
2306         case F_RDLCK:
2307         case F_WRLCK:
2308                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2309                     ffl->end < ffl->start)
2310                         return -EIO;
2311
2312                 fl->fl_start = ffl->start;
2313                 fl->fl_end = ffl->end;
2314
2315                 /*
2316                  * Convert pid into init's pid namespace.  The locks API will
2317                  * translate it into the caller's pid namespace.
2318                  */
2319                 rcu_read_lock();
2320                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2321                 rcu_read_unlock();
2322                 break;
2323
2324         default:
2325                 return -EIO;
2326         }
2327         fl->fl_type = ffl->type;
2328         return 0;
2329 }
2330
2331 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2332                          const struct file_lock *fl, int opcode, pid_t pid,
2333                          int flock, struct fuse_lk_in *inarg)
2334 {
2335         struct inode *inode = file_inode(file);
2336         struct fuse_conn *fc = get_fuse_conn(inode);
2337         struct fuse_file *ff = file->private_data;
2338
2339         memset(inarg, 0, sizeof(*inarg));
2340         inarg->fh = ff->fh;
2341         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2342         inarg->lk.start = fl->fl_start;
2343         inarg->lk.end = fl->fl_end;
2344         inarg->lk.type = fl->fl_type;
2345         inarg->lk.pid = pid;
2346         if (flock)
2347                 inarg->lk_flags |= FUSE_LK_FLOCK;
2348         args->opcode = opcode;
2349         args->nodeid = get_node_id(inode);
2350         args->in_numargs = 1;
2351         args->in_args[0].size = sizeof(*inarg);
2352         args->in_args[0].value = inarg;
2353 }
2354
2355 static int fuse_getlk(struct file *file, struct file_lock *fl)
2356 {
2357         struct inode *inode = file_inode(file);
2358         struct fuse_conn *fc = get_fuse_conn(inode);
2359         FUSE_ARGS(args);
2360         struct fuse_lk_in inarg;
2361         struct fuse_lk_out outarg;
2362         int err;
2363
2364         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2365         args.out_numargs = 1;
2366         args.out_args[0].size = sizeof(outarg);
2367         args.out_args[0].value = &outarg;
2368         err = fuse_simple_request(fc, &args);
2369         if (!err)
2370                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2371
2372         return err;
2373 }
2374
2375 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2376 {
2377         struct inode *inode = file_inode(file);
2378         struct fuse_conn *fc = get_fuse_conn(inode);
2379         FUSE_ARGS(args);
2380         struct fuse_lk_in inarg;
2381         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2382         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2383         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2384         int err;
2385
2386         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2387                 /* NLM needs asynchronous locks, which we don't support yet */
2388                 return -ENOLCK;
2389         }
2390
2391         /* Unlock on close is handled by the flush method */
2392         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2393                 return 0;
2394
2395         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2396         err = fuse_simple_request(fc, &args);
2397
2398         /* locking is restartable */
2399         if (err == -EINTR)
2400                 err = -ERESTARTSYS;
2401
2402         return err;
2403 }
2404
2405 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2406 {
2407         struct inode *inode = file_inode(file);
2408         struct fuse_conn *fc = get_fuse_conn(inode);
2409         int err;
2410
2411         if (cmd == F_CANCELLK) {
2412                 err = 0;
2413         } else if (cmd == F_GETLK) {
2414                 if (fc->no_lock) {
2415                         posix_test_lock(file, fl);
2416                         err = 0;
2417                 } else
2418                         err = fuse_getlk(file, fl);
2419         } else {
2420                 if (fc->no_lock)
2421                         err = posix_lock_file(file, fl, NULL);
2422                 else
2423                         err = fuse_setlk(file, fl, 0);
2424         }
2425         return err;
2426 }
2427
2428 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2429 {
2430         struct inode *inode = file_inode(file);
2431         struct fuse_conn *fc = get_fuse_conn(inode);
2432         int err;
2433
2434         if (fc->no_flock) {
2435                 err = locks_lock_file_wait(file, fl);
2436         } else {
2437                 struct fuse_file *ff = file->private_data;
2438
2439                 /* emulate flock with POSIX locks */
2440                 ff->flock = true;
2441                 err = fuse_setlk(file, fl, 1);
2442         }
2443
2444         return err;
2445 }
2446
2447 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2448 {
2449         struct inode *inode = mapping->host;
2450         struct fuse_conn *fc = get_fuse_conn(inode);
2451         FUSE_ARGS(args);
2452         struct fuse_bmap_in inarg;
2453         struct fuse_bmap_out outarg;
2454         int err;
2455
2456         if (!inode->i_sb->s_bdev || fc->no_bmap)
2457                 return 0;
2458
2459         memset(&inarg, 0, sizeof(inarg));
2460         inarg.block = block;
2461         inarg.blocksize = inode->i_sb->s_blocksize;
2462         args.opcode = FUSE_BMAP;
2463         args.nodeid = get_node_id(inode);
2464         args.in_numargs = 1;
2465         args.in_args[0].size = sizeof(inarg);
2466         args.in_args[0].value = &inarg;
2467         args.out_numargs = 1;
2468         args.out_args[0].size = sizeof(outarg);
2469         args.out_args[0].value = &outarg;
2470         err = fuse_simple_request(fc, &args);
2471         if (err == -ENOSYS)
2472                 fc->no_bmap = 1;
2473
2474         return err ? 0 : outarg.block;
2475 }
2476
2477 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2478 {
2479         struct inode *inode = file->f_mapping->host;
2480         struct fuse_conn *fc = get_fuse_conn(inode);
2481         struct fuse_file *ff = file->private_data;
2482         FUSE_ARGS(args);
2483         struct fuse_lseek_in inarg = {
2484                 .fh = ff->fh,
2485                 .offset = offset,
2486                 .whence = whence
2487         };
2488         struct fuse_lseek_out outarg;
2489         int err;
2490
2491         if (fc->no_lseek)
2492                 goto fallback;
2493
2494         args.opcode = FUSE_LSEEK;
2495         args.nodeid = ff->nodeid;
2496         args.in_numargs = 1;
2497         args.in_args[0].size = sizeof(inarg);
2498         args.in_args[0].value = &inarg;
2499         args.out_numargs = 1;
2500         args.out_args[0].size = sizeof(outarg);
2501         args.out_args[0].value = &outarg;
2502         err = fuse_simple_request(fc, &args);
2503         if (err) {
2504                 if (err == -ENOSYS) {
2505                         fc->no_lseek = 1;
2506                         goto fallback;
2507                 }
2508                 return err;
2509         }
2510
2511         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2512
2513 fallback:
2514         err = fuse_update_attributes(inode, file);
2515         if (!err)
2516                 return generic_file_llseek(file, offset, whence);
2517         else
2518                 return err;
2519 }
2520
2521 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2522 {
2523         loff_t retval;
2524         struct inode *inode = file_inode(file);
2525
2526         switch (whence) {
2527         case SEEK_SET:
2528         case SEEK_CUR:
2529                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2530                 retval = generic_file_llseek(file, offset, whence);
2531                 break;
2532         case SEEK_END:
2533                 inode_lock(inode);
2534                 retval = fuse_update_attributes(inode, file);
2535                 if (!retval)
2536                         retval = generic_file_llseek(file, offset, whence);
2537                 inode_unlock(inode);
2538                 break;
2539         case SEEK_HOLE:
2540         case SEEK_DATA:
2541                 inode_lock(inode);
2542                 retval = fuse_lseek(file, offset, whence);
2543                 inode_unlock(inode);
2544                 break;
2545         default:
2546                 retval = -EINVAL;
2547         }
2548
2549         return retval;
2550 }
2551
2552 /*
2553  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2554  * ABI was defined to be 'struct iovec' which is different on 32bit
2555  * and 64bit.  Fortunately we can determine which structure the server
2556  * used from the size of the reply.
2557  */
2558 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2559                                      size_t transferred, unsigned count,
2560                                      bool is_compat)
2561 {
2562 #ifdef CONFIG_COMPAT
2563         if (count * sizeof(struct compat_iovec) == transferred) {
2564                 struct compat_iovec *ciov = src;
2565                 unsigned i;
2566
2567                 /*
2568                  * With this interface a 32bit server cannot support
2569                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2570                  * requests
2571                  */
2572                 if (!is_compat)
2573                         return -EINVAL;
2574
2575                 for (i = 0; i < count; i++) {
2576                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2577                         dst[i].iov_len = ciov[i].iov_len;
2578                 }
2579                 return 0;
2580         }
2581 #endif
2582
2583         if (count * sizeof(struct iovec) != transferred)
2584                 return -EIO;
2585
2586         memcpy(dst, src, transferred);
2587         return 0;
2588 }
2589
2590 /* Make sure iov_length() won't overflow */
2591 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2592                                  size_t count)
2593 {
2594         size_t n;
2595         u32 max = fc->max_pages << PAGE_SHIFT;
2596
2597         for (n = 0; n < count; n++, iov++) {
2598                 if (iov->iov_len > (size_t) max)
2599                         return -ENOMEM;
2600                 max -= iov->iov_len;
2601         }
2602         return 0;
2603 }
2604
2605 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2606                                  void *src, size_t transferred, unsigned count,
2607                                  bool is_compat)
2608 {
2609         unsigned i;
2610         struct fuse_ioctl_iovec *fiov = src;
2611
2612         if (fc->minor < 16) {
2613                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2614                                                  count, is_compat);
2615         }
2616
2617         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2618                 return -EIO;
2619
2620         for (i = 0; i < count; i++) {
2621                 /* Did the server supply an inappropriate value? */
2622                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2623                     fiov[i].len != (unsigned long) fiov[i].len)
2624                         return -EIO;
2625
2626                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2627                 dst[i].iov_len = (size_t) fiov[i].len;
2628
2629 #ifdef CONFIG_COMPAT
2630                 if (is_compat &&
2631                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2632                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2633                         return -EIO;
2634 #endif
2635         }
2636
2637         return 0;
2638 }
2639
2640
2641 /*
2642  * For ioctls, there is no generic way to determine how much memory
2643  * needs to be read and/or written.  Furthermore, ioctls are allowed
2644  * to dereference the passed pointer, so the parameter requires deep
2645  * copying but FUSE has no idea whatsoever about what to copy in or
2646  * out.
2647  *
2648  * This is solved by allowing FUSE server to retry ioctl with
2649  * necessary in/out iovecs.  Let's assume the ioctl implementation
2650  * needs to read in the following structure.
2651  *
2652  * struct a {
2653  *      char    *buf;
2654  *      size_t  buflen;
2655  * }
2656  *
2657  * On the first callout to FUSE server, inarg->in_size and
2658  * inarg->out_size will be NULL; then, the server completes the ioctl
2659  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2660  * the actual iov array to
2661  *
2662  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2663  *
2664  * which tells FUSE to copy in the requested area and retry the ioctl.
2665  * On the second round, the server has access to the structure and
2666  * from that it can tell what to look for next, so on the invocation,
2667  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2668  *
2669  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2670  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2671  *
2672  * FUSE will copy both struct a and the pointed buffer from the
2673  * process doing the ioctl and retry ioctl with both struct a and the
2674  * buffer.
2675  *
2676  * This time, FUSE server has everything it needs and completes ioctl
2677  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2678  *
2679  * Copying data out works the same way.
2680  *
2681  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2682  * automatically initializes in and out iovs by decoding @cmd with
2683  * _IOC_* macros and the server is not allowed to request RETRY.  This
2684  * limits ioctl data transfers to well-formed ioctls and is the forced
2685  * behavior for all FUSE servers.
2686  */
2687 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2688                    unsigned int flags)
2689 {
2690         struct fuse_file *ff = file->private_data;
2691         struct fuse_conn *fc = ff->fc;
2692         struct fuse_ioctl_in inarg = {
2693                 .fh = ff->fh,
2694                 .cmd = cmd,
2695                 .arg = arg,
2696                 .flags = flags
2697         };
2698         struct fuse_ioctl_out outarg;
2699         struct iovec *iov_page = NULL;
2700         struct iovec *in_iov = NULL, *out_iov = NULL;
2701         unsigned int in_iovs = 0, out_iovs = 0, max_pages;
2702         size_t in_size, out_size, c;
2703         ssize_t transferred;
2704         int err, i;
2705         struct iov_iter ii;
2706         struct fuse_args_pages ap = {};
2707
2708 #if BITS_PER_LONG == 32
2709         inarg.flags |= FUSE_IOCTL_32BIT;
2710 #else
2711         if (flags & FUSE_IOCTL_COMPAT) {
2712                 inarg.flags |= FUSE_IOCTL_32BIT;
2713 #ifdef CONFIG_X86_X32
2714                 if (in_x32_syscall())
2715                         inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2716 #endif
2717         }
2718 #endif
2719
2720         /* assume all the iovs returned by client always fits in a page */
2721         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2722
2723         err = -ENOMEM;
2724         ap.pages = fuse_pages_alloc(fc->max_pages, GFP_KERNEL, &ap.descs);
2725         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2726         if (!ap.pages || !iov_page)
2727                 goto out;
2728
2729         fuse_page_descs_length_init(ap.descs, 0, fc->max_pages);
2730
2731         /*
2732          * If restricted, initialize IO parameters as encoded in @cmd.
2733          * RETRY from server is not allowed.
2734          */
2735         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2736                 struct iovec *iov = iov_page;
2737
2738                 iov->iov_base = (void __user *)arg;
2739                 iov->iov_len = _IOC_SIZE(cmd);
2740
2741                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2742                         in_iov = iov;
2743                         in_iovs = 1;
2744                 }
2745
2746                 if (_IOC_DIR(cmd) & _IOC_READ) {
2747                         out_iov = iov;
2748                         out_iovs = 1;
2749                 }
2750         }
2751
2752  retry:
2753         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2754         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2755
2756         /*
2757          * Out data can be used either for actual out data or iovs,
2758          * make sure there always is at least one page.
2759          */
2760         out_size = max_t(size_t, out_size, PAGE_SIZE);
2761         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2762
2763         /* make sure there are enough buffer pages and init request with them */
2764         err = -ENOMEM;
2765         if (max_pages > fc->max_pages)
2766                 goto out;
2767         while (ap.num_pages < max_pages) {
2768                 ap.pages[ap.num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2769                 if (!ap.pages[ap.num_pages])
2770                         goto out;
2771                 ap.num_pages++;
2772         }
2773
2774
2775         /* okay, let's send it to the client */
2776         ap.args.opcode = FUSE_IOCTL;
2777         ap.args.nodeid = ff->nodeid;
2778         ap.args.in_numargs = 1;
2779         ap.args.in_args[0].size = sizeof(inarg);
2780         ap.args.in_args[0].value = &inarg;
2781         if (in_size) {
2782                 ap.args.in_numargs++;
2783                 ap.args.in_args[1].size = in_size;
2784                 ap.args.in_pages = true;
2785
2786                 err = -EFAULT;
2787                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2788                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2789                         c = copy_page_from_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2790                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2791                                 goto out;
2792                 }
2793         }
2794
2795         ap.args.out_numargs = 2;
2796         ap.args.out_args[0].size = sizeof(outarg);
2797         ap.args.out_args[0].value = &outarg;
2798         ap.args.out_args[1].size = out_size;
2799         ap.args.out_pages = true;
2800         ap.args.out_argvar = true;
2801
2802         transferred = fuse_simple_request(fc, &ap.args);
2803         err = transferred;
2804         if (transferred < 0)
2805                 goto out;
2806
2807         /* did it ask for retry? */
2808         if (outarg.flags & FUSE_IOCTL_RETRY) {
2809                 void *vaddr;
2810
2811                 /* no retry if in restricted mode */
2812                 err = -EIO;
2813                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2814                         goto out;
2815
2816                 in_iovs = outarg.in_iovs;
2817                 out_iovs = outarg.out_iovs;
2818
2819                 /*
2820                  * Make sure things are in boundary, separate checks
2821                  * are to protect against overflow.
2822                  */
2823                 err = -ENOMEM;
2824                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2825                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2826                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2827                         goto out;
2828
2829                 vaddr = kmap_atomic(ap.pages[0]);
2830                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2831                                             transferred, in_iovs + out_iovs,
2832                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2833                 kunmap_atomic(vaddr);
2834                 if (err)
2835                         goto out;
2836
2837                 in_iov = iov_page;
2838                 out_iov = in_iov + in_iovs;
2839
2840                 err = fuse_verify_ioctl_iov(fc, in_iov, in_iovs);
2841                 if (err)
2842                         goto out;
2843
2844                 err = fuse_verify_ioctl_iov(fc, out_iov, out_iovs);
2845                 if (err)
2846                         goto out;
2847
2848                 goto retry;
2849         }
2850
2851         err = -EIO;
2852         if (transferred > inarg.out_size)
2853                 goto out;
2854
2855         err = -EFAULT;
2856         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2857         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2858                 c = copy_page_to_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2859                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2860                         goto out;
2861         }
2862         err = 0;
2863  out:
2864         free_page((unsigned long) iov_page);
2865         while (ap.num_pages)
2866                 __free_page(ap.pages[--ap.num_pages]);
2867         kfree(ap.pages);
2868
2869         return err ? err : outarg.result;
2870 }
2871 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2872
2873 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2874                        unsigned long arg, unsigned int flags)
2875 {
2876         struct inode *inode = file_inode(file);
2877         struct fuse_conn *fc = get_fuse_conn(inode);
2878
2879         if (!fuse_allow_current_process(fc))
2880                 return -EACCES;
2881
2882         if (is_bad_inode(inode))
2883                 return -EIO;
2884
2885         return fuse_do_ioctl(file, cmd, arg, flags);
2886 }
2887
2888 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2889                             unsigned long arg)
2890 {
2891         return fuse_ioctl_common(file, cmd, arg, 0);
2892 }
2893
2894 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2895                                    unsigned long arg)
2896 {
2897         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2898 }
2899
2900 /*
2901  * All files which have been polled are linked to RB tree
2902  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2903  * find the matching one.
2904  */
2905 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2906                                               struct rb_node **parent_out)
2907 {
2908         struct rb_node **link = &fc->polled_files.rb_node;
2909         struct rb_node *last = NULL;
2910
2911         while (*link) {
2912                 struct fuse_file *ff;
2913
2914                 last = *link;
2915                 ff = rb_entry(last, struct fuse_file, polled_node);
2916
2917                 if (kh < ff->kh)
2918                         link = &last->rb_left;
2919                 else if (kh > ff->kh)
2920                         link = &last->rb_right;
2921                 else
2922                         return link;
2923         }
2924
2925         if (parent_out)
2926                 *parent_out = last;
2927         return link;
2928 }
2929
2930 /*
2931  * The file is about to be polled.  Make sure it's on the polled_files
2932  * RB tree.  Note that files once added to the polled_files tree are
2933  * not removed before the file is released.  This is because a file
2934  * polled once is likely to be polled again.
2935  */
2936 static void fuse_register_polled_file(struct fuse_conn *fc,
2937                                       struct fuse_file *ff)
2938 {
2939         spin_lock(&fc->lock);
2940         if (RB_EMPTY_NODE(&ff->polled_node)) {
2941                 struct rb_node **link, *uninitialized_var(parent);
2942
2943                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2944                 BUG_ON(*link);
2945                 rb_link_node(&ff->polled_node, parent, link);
2946                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2947         }
2948         spin_unlock(&fc->lock);
2949 }
2950
2951 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2952 {
2953         struct fuse_file *ff = file->private_data;
2954         struct fuse_conn *fc = ff->fc;
2955         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2956         struct fuse_poll_out outarg;
2957         FUSE_ARGS(args);
2958         int err;
2959
2960         if (fc->no_poll)
2961                 return DEFAULT_POLLMASK;
2962
2963         poll_wait(file, &ff->poll_wait, wait);
2964         inarg.events = mangle_poll(poll_requested_events(wait));
2965
2966         /*
2967          * Ask for notification iff there's someone waiting for it.
2968          * The client may ignore the flag and always notify.
2969          */
2970         if (waitqueue_active(&ff->poll_wait)) {
2971                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2972                 fuse_register_polled_file(fc, ff);
2973         }
2974
2975         args.opcode = FUSE_POLL;
2976         args.nodeid = ff->nodeid;
2977         args.in_numargs = 1;
2978         args.in_args[0].size = sizeof(inarg);
2979         args.in_args[0].value = &inarg;
2980         args.out_numargs = 1;
2981         args.out_args[0].size = sizeof(outarg);
2982         args.out_args[0].value = &outarg;
2983         err = fuse_simple_request(fc, &args);
2984
2985         if (!err)
2986                 return demangle_poll(outarg.revents);
2987         if (err == -ENOSYS) {
2988                 fc->no_poll = 1;
2989                 return DEFAULT_POLLMASK;
2990         }
2991         return EPOLLERR;
2992 }
2993 EXPORT_SYMBOL_GPL(fuse_file_poll);
2994
2995 /*
2996  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2997  * wakes up the poll waiters.
2998  */
2999 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
3000                             struct fuse_notify_poll_wakeup_out *outarg)
3001 {
3002         u64 kh = outarg->kh;
3003         struct rb_node **link;
3004
3005         spin_lock(&fc->lock);
3006
3007         link = fuse_find_polled_node(fc, kh, NULL);
3008         if (*link) {
3009                 struct fuse_file *ff;
3010
3011                 ff = rb_entry(*link, struct fuse_file, polled_node);
3012                 wake_up_interruptible_sync(&ff->poll_wait);
3013         }
3014
3015         spin_unlock(&fc->lock);
3016         return 0;
3017 }
3018
3019 static void fuse_do_truncate(struct file *file)
3020 {
3021         struct inode *inode = file->f_mapping->host;
3022         struct iattr attr;
3023
3024         attr.ia_valid = ATTR_SIZE;
3025         attr.ia_size = i_size_read(inode);
3026
3027         attr.ia_file = file;
3028         attr.ia_valid |= ATTR_FILE;
3029
3030         fuse_do_setattr(file_dentry(file), &attr, file);
3031 }
3032
3033 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
3034 {
3035         return round_up(off, fc->max_pages << PAGE_SHIFT);
3036 }
3037
3038 static ssize_t
3039 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3040 {
3041         DECLARE_COMPLETION_ONSTACK(wait);
3042         ssize_t ret = 0;
3043         struct file *file = iocb->ki_filp;
3044         struct fuse_file *ff = file->private_data;
3045         bool async_dio = ff->fc->async_dio;
3046         loff_t pos = 0;
3047         struct inode *inode;
3048         loff_t i_size;
3049         size_t count = iov_iter_count(iter);
3050         loff_t offset = iocb->ki_pos;
3051         struct fuse_io_priv *io;
3052
3053         pos = offset;
3054         inode = file->f_mapping->host;
3055         i_size = i_size_read(inode);
3056
3057         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
3058                 return 0;
3059
3060         /* optimization for short read */
3061         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
3062                 if (offset >= i_size)
3063                         return 0;
3064                 iov_iter_truncate(iter, fuse_round_up(ff->fc, i_size - offset));
3065                 count = iov_iter_count(iter);
3066         }
3067
3068         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
3069         if (!io)
3070                 return -ENOMEM;
3071         spin_lock_init(&io->lock);
3072         kref_init(&io->refcnt);
3073         io->reqs = 1;
3074         io->bytes = -1;
3075         io->size = 0;
3076         io->offset = offset;
3077         io->write = (iov_iter_rw(iter) == WRITE);
3078         io->err = 0;
3079         /*
3080          * By default, we want to optimize all I/Os with async request
3081          * submission to the client filesystem if supported.
3082          */
3083         io->async = async_dio;
3084         io->iocb = iocb;
3085         io->blocking = is_sync_kiocb(iocb);
3086
3087         /*
3088          * We cannot asynchronously extend the size of a file.
3089          * In such case the aio will behave exactly like sync io.
3090          */
3091         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
3092                 io->blocking = true;
3093
3094         if (io->async && io->blocking) {
3095                 /*
3096                  * Additional reference to keep io around after
3097                  * calling fuse_aio_complete()
3098                  */
3099                 kref_get(&io->refcnt);
3100                 io->done = &wait;
3101         }
3102
3103         if (iov_iter_rw(iter) == WRITE) {
3104                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
3105                 fuse_invalidate_attr(inode);
3106         } else {
3107                 ret = __fuse_direct_read(io, iter, &pos);
3108         }
3109
3110         if (io->async) {
3111                 bool blocking = io->blocking;
3112
3113                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3114
3115                 /* we have a non-extending, async request, so return */
3116                 if (!blocking)
3117                         return -EIOCBQUEUED;
3118
3119                 wait_for_completion(&wait);
3120                 ret = fuse_get_res_by_io(io);
3121         }
3122
3123         kref_put(&io->refcnt, fuse_io_release);
3124
3125         if (iov_iter_rw(iter) == WRITE) {
3126                 if (ret > 0)
3127                         fuse_write_update_size(inode, pos);
3128                 else if (ret < 0 && offset + count > i_size)
3129                         fuse_do_truncate(file);
3130         }
3131
3132         return ret;
3133 }
3134
3135 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3136 {
3137         int err = filemap_write_and_wait_range(inode->i_mapping, start, end);
3138
3139         if (!err)
3140                 fuse_sync_writes(inode);
3141
3142         return err;
3143 }
3144
3145 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3146                                 loff_t length)
3147 {
3148         struct fuse_file *ff = file->private_data;
3149         struct inode *inode = file_inode(file);
3150         struct fuse_inode *fi = get_fuse_inode(inode);
3151         struct fuse_conn *fc = ff->fc;
3152         FUSE_ARGS(args);
3153         struct fuse_fallocate_in inarg = {
3154                 .fh = ff->fh,
3155                 .offset = offset,
3156                 .length = length,
3157                 .mode = mode
3158         };
3159         int err;
3160         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
3161                            (mode & FALLOC_FL_PUNCH_HOLE);
3162
3163         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3164                 return -EOPNOTSUPP;
3165
3166         if (fc->no_fallocate)
3167                 return -EOPNOTSUPP;
3168
3169         if (lock_inode) {
3170                 inode_lock(inode);
3171                 if (mode & FALLOC_FL_PUNCH_HOLE) {
3172                         loff_t endbyte = offset + length - 1;
3173
3174                         err = fuse_writeback_range(inode, offset, endbyte);
3175                         if (err)
3176                                 goto out;
3177                 }
3178         }
3179
3180         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3181             offset + length > i_size_read(inode)) {
3182                 err = inode_newsize_ok(inode, offset + length);
3183                 if (err)
3184                         goto out;
3185         }
3186
3187         if (!(mode & FALLOC_FL_KEEP_SIZE))
3188                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3189
3190         args.opcode = FUSE_FALLOCATE;
3191         args.nodeid = ff->nodeid;
3192         args.in_numargs = 1;
3193         args.in_args[0].size = sizeof(inarg);
3194         args.in_args[0].value = &inarg;
3195         err = fuse_simple_request(fc, &args);
3196         if (err == -ENOSYS) {
3197                 fc->no_fallocate = 1;
3198                 err = -EOPNOTSUPP;
3199         }
3200         if (err)
3201                 goto out;
3202
3203         /* we could have extended the file */
3204         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3205                 bool changed = fuse_write_update_size(inode, offset + length);
3206
3207                 if (changed && fc->writeback_cache)
3208                         file_update_time(file);
3209         }
3210
3211         if (mode & FALLOC_FL_PUNCH_HOLE)
3212                 truncate_pagecache_range(inode, offset, offset + length - 1);
3213
3214         fuse_invalidate_attr(inode);
3215
3216 out:
3217         if (!(mode & FALLOC_FL_KEEP_SIZE))
3218                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3219
3220         if (lock_inode)
3221                 inode_unlock(inode);
3222
3223         return err;
3224 }
3225
3226 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3227                                       struct file *file_out, loff_t pos_out,
3228                                       size_t len, unsigned int flags)
3229 {
3230         struct fuse_file *ff_in = file_in->private_data;
3231         struct fuse_file *ff_out = file_out->private_data;
3232         struct inode *inode_in = file_inode(file_in);
3233         struct inode *inode_out = file_inode(file_out);
3234         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3235         struct fuse_conn *fc = ff_in->fc;
3236         FUSE_ARGS(args);
3237         struct fuse_copy_file_range_in inarg = {
3238                 .fh_in = ff_in->fh,
3239                 .off_in = pos_in,
3240                 .nodeid_out = ff_out->nodeid,
3241                 .fh_out = ff_out->fh,
3242                 .off_out = pos_out,
3243                 .len = len,
3244                 .flags = flags
3245         };
3246         struct fuse_write_out outarg;
3247         ssize_t err;
3248         /* mark unstable when write-back is not used, and file_out gets
3249          * extended */
3250         bool is_unstable = (!fc->writeback_cache) &&
3251                            ((pos_out + len) > inode_out->i_size);
3252
3253         if (fc->no_copy_file_range)
3254                 return -EOPNOTSUPP;
3255
3256         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3257                 return -EXDEV;
3258
3259         if (fc->writeback_cache) {
3260                 inode_lock(inode_in);
3261                 err = fuse_writeback_range(inode_in, pos_in, pos_in + len);
3262                 inode_unlock(inode_in);
3263                 if (err)
3264                         return err;
3265         }
3266
3267         inode_lock(inode_out);
3268
3269         err = file_modified(file_out);
3270         if (err)
3271                 goto out;
3272
3273         if (fc->writeback_cache) {
3274                 err = fuse_writeback_range(inode_out, pos_out, pos_out + len);
3275                 if (err)
3276                         goto out;
3277         }
3278
3279         if (is_unstable)
3280                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3281
3282         args.opcode = FUSE_COPY_FILE_RANGE;
3283         args.nodeid = ff_in->nodeid;
3284         args.in_numargs = 1;
3285         args.in_args[0].size = sizeof(inarg);
3286         args.in_args[0].value = &inarg;
3287         args.out_numargs = 1;
3288         args.out_args[0].size = sizeof(outarg);
3289         args.out_args[0].value = &outarg;
3290         err = fuse_simple_request(fc, &args);
3291         if (err == -ENOSYS) {
3292                 fc->no_copy_file_range = 1;
3293                 err = -EOPNOTSUPP;
3294         }
3295         if (err)
3296                 goto out;
3297
3298         if (fc->writeback_cache) {
3299                 fuse_write_update_size(inode_out, pos_out + outarg.size);
3300                 file_update_time(file_out);
3301         }
3302
3303         fuse_invalidate_attr(inode_out);
3304
3305         err = outarg.size;
3306 out:
3307         if (is_unstable)
3308                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3309
3310         inode_unlock(inode_out);
3311         file_accessed(file_in);
3312
3313         return err;
3314 }
3315
3316 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3317                                     struct file *dst_file, loff_t dst_off,
3318                                     size_t len, unsigned int flags)
3319 {
3320         ssize_t ret;
3321
3322         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3323                                      len, flags);
3324
3325         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3326                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3327                                               dst_off, len, flags);
3328         return ret;
3329 }
3330
3331 static const struct file_operations fuse_file_operations = {
3332         .llseek         = fuse_file_llseek,
3333         .read_iter      = fuse_file_read_iter,
3334         .write_iter     = fuse_file_write_iter,
3335         .mmap           = fuse_file_mmap,
3336         .open           = fuse_open,
3337         .flush          = fuse_flush,
3338         .release        = fuse_release,
3339         .fsync          = fuse_fsync,
3340         .lock           = fuse_file_lock,
3341         .flock          = fuse_file_flock,
3342         .splice_read    = generic_file_splice_read,
3343         .splice_write   = iter_file_splice_write,
3344         .unlocked_ioctl = fuse_file_ioctl,
3345         .compat_ioctl   = fuse_file_compat_ioctl,
3346         .poll           = fuse_file_poll,
3347         .fallocate      = fuse_file_fallocate,
3348         .copy_file_range = fuse_copy_file_range,
3349 };
3350
3351 static const struct address_space_operations fuse_file_aops  = {
3352         .readpage       = fuse_readpage,
3353         .writepage      = fuse_writepage,
3354         .writepages     = fuse_writepages,
3355         .launder_page   = fuse_launder_page,
3356         .readpages      = fuse_readpages,
3357         .set_page_dirty = __set_page_dirty_nobuffers,
3358         .bmap           = fuse_bmap,
3359         .direct_IO      = fuse_direct_IO,
3360         .write_begin    = fuse_write_begin,
3361         .write_end      = fuse_write_end,
3362 };
3363
3364 void fuse_init_file_inode(struct inode *inode)
3365 {
3366         struct fuse_inode *fi = get_fuse_inode(inode);
3367
3368         inode->i_fop = &fuse_file_operations;
3369         inode->i_data.a_ops = &fuse_file_aops;
3370
3371         INIT_LIST_HEAD(&fi->write_files);
3372         INIT_LIST_HEAD(&fi->queued_writes);
3373         fi->writectr = 0;
3374         init_waitqueue_head(&fi->page_waitq);
3375         INIT_LIST_HEAD(&fi->writepages);
3376 }