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 is_wb_truncate = (file->f_flags & O_TRUNC) &&
 221                           fc->atomic_o_trunc &&
 222                           fc->writeback_cache;
 223
 224         err = generic_file_open(inode, file);
 225         if (err)
 226                 return err;
 227
 228         if (is_wb_truncate) {
 229                 inode_lock(inode);
 230                 fuse_set_nowrite(inode);
 231         }
 232
 233         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 234
 235         if (!err)
 236                 fuse_finish_open(inode, file);
 237
 238         if (is_wb_truncate) {
 239                 fuse_release_nowrite(inode);
 240                 inode_unlock(inode);
 241         }
 242
 243         return err;
 244 }
 245
 246 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 247                                  int flags, int opcode)
 248 {
 249         struct fuse_conn *fc = ff->fc;
 250         struct fuse_release_args *ra = ff->release_args;
 251
 252         /* Inode is NULL on error path of fuse_create_open() */
 253         if (likely(fi)) {
 254                 spin_lock(&fi->lock);
 255                 list_del(&ff->write_entry);
 256                 spin_unlock(&fi->lock);
 257         }
 258         spin_lock(&fc->lock);
 259         if (!RB_EMPTY_NODE(&ff->polled_node))
 260                 rb_erase(&ff->polled_node, &fc->polled_files);
 261         spin_unlock(&fc->lock);
 262
 263         wake_up_interruptible_all(&ff->poll_wait);
 264
 265         ra->inarg.fh = ff->fh;
 266         ra->inarg.flags = flags;
 267         ra->args.in_numargs = 1;
 268         ra->args.in_args[0].size = sizeof(struct fuse_release_in);
 269         ra->args.in_args[0].value = &ra->inarg;
 270         ra->args.opcode = opcode;
 271         ra->args.nodeid = ff->nodeid;
 272         ra->args.force = true;
 273         ra->args.nocreds = true;
 274 }
 275
 276 void fuse_release_common(struct file *file, bool isdir)
 277 {
 278         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 279         struct fuse_file *ff = file->private_data;
 280         struct fuse_release_args *ra = ff->release_args;
 281         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 282
 283         fuse_prepare_release(fi, ff, file->f_flags, opcode);
 284
 285         if (ff->flock) {
 286                 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 287                 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fc,
 288                                                           (fl_owner_t) file);
 289         }
 290         /* Hold inode until release is finished */
 291         ra->inode = igrab(file_inode(file));
 292
 293         /*
 294          * Normally this will send the RELEASE request, however if
 295          * some asynchronous READ or WRITE requests are outstanding,
 296          * the sending will be delayed.
 297          *
 298          * Make the release synchronous if this is a fuseblk mount,
 299          * synchronous RELEASE is allowed (and desirable) in this case
 300          * because the server can be trusted not to screw up.
 301          */
 302         fuse_file_put(ff, ff->fc->destroy, isdir);
 303 }
 304
 305 static int fuse_open(struct inode *inode, struct file *file)
 306 {
 307         return fuse_open_common(inode, file, false);
 308 }
 309
 310 static int fuse_release(struct inode *inode, struct file *file)
 311 {
 312         struct fuse_conn *fc = get_fuse_conn(inode);
 313
 314         /* see fuse_vma_close() for !writeback_cache case */
 315         if (fc->writeback_cache)
 316                 write_inode_now(inode, 1);
 317
 318         fuse_release_common(file, false);
 319
 320         /* return value is ignored by VFS */
 321         return 0;
 322 }
 323
 324 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
 325 {
 326         WARN_ON(refcount_read(&ff->count) > 1);
 327         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 328         /*
 329          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 330          * synchronous, we are fine with not doing igrab() here"
 331          */
 332         fuse_file_put(ff, true, false);
 333 }
 334 EXPORT_SYMBOL_GPL(fuse_sync_release);
 335
 336 /*
 337  * Scramble the ID space with XTEA, so that the value of the files_struct
 338  * pointer is not exposed to userspace.
 339  */
 340 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 341 {
 342         u32 *k = fc->scramble_key;
 343         u64 v = (unsigned long) id;
 344         u32 v0 = v;
 345         u32 v1 = v >> 32;
 346         u32 sum = 0;
 347         int i;
 348
 349         for (i = 0; i < 32; i++) {
 350                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 351                 sum += 0x9E3779B9;
 352                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 353         }
 354
 355         return (u64) v0 + ((u64) v1 << 32);
 356 }
 357
 358 struct fuse_writepage_args {
 359         struct fuse_io_args ia;
 360         struct rb_node writepages_entry;
 361         struct list_head queue_entry;
 362         struct fuse_writepage_args *next;
 363         struct inode *inode;
 364 };
 365
 366 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
 367                                             pgoff_t idx_from, pgoff_t idx_to)
 368 {
 369         struct rb_node *n;
 370
 371         n = fi->writepages.rb_node;
 372
 373         while (n) {
 374                 struct fuse_writepage_args *wpa;
 375                 pgoff_t curr_index;
 376
 377                 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
 378                 WARN_ON(get_fuse_inode(wpa->inode) != fi);
 379                 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
 380                 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
 381                         n = n->rb_right;
 382                 else if (idx_to < curr_index)
 383                         n = n->rb_left;
 384                 else
 385                         return wpa;
 386         }
 387         return NULL;
 388 }
 389
 390 /*
 391  * Check if any page in a range is under writeback
 392  *
 393  * This is currently done by walking the list of writepage requests
 394  * for the inode, which can be pretty inefficient.
 395  */
 396 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 397                                    pgoff_t idx_to)
 398 {
 399         struct fuse_inode *fi = get_fuse_inode(inode);
 400         bool found;
 401
 402         spin_lock(&fi->lock);
 403         found = fuse_find_writeback(fi, idx_from, idx_to);
 404         spin_unlock(&fi->lock);
 405
 406         return found;
 407 }
 408
 409 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 410 {
 411         return fuse_range_is_writeback(inode, index, index);
 412 }
 413
 414 /*
 415  * Wait for page writeback to be completed.
 416  *
 417  * Since fuse doesn't rely on the VM writeback tracking, this has to
 418  * use some other means.
 419  */
 420 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 421 {
 422         struct fuse_inode *fi = get_fuse_inode(inode);
 423
 424         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 425 }
 426
 427 /*
 428  * Wait for all pending writepages on the inode to finish.
 429  *
 430  * This is currently done by blocking further writes with FUSE_NOWRITE
 431  * and waiting for all sent writes to complete.
 432  *
 433  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 434  * could conflict with truncation.
 435  */
 436 static void fuse_sync_writes(struct inode *inode)
 437 {
 438         fuse_set_nowrite(inode);
 439         fuse_release_nowrite(inode);
 440 }
 441
 442 static int fuse_flush(struct file *file, fl_owner_t id)
 443 {
 444         struct inode *inode = file_inode(file);
 445         struct fuse_conn *fc = get_fuse_conn(inode);
 446         struct fuse_file *ff = file->private_data;
 447         struct fuse_flush_in inarg;
 448         FUSE_ARGS(args);
 449         int err;
 450
 451         if (is_bad_inode(inode))
 452                 return -EIO;
 453
 454         err = write_inode_now(inode, 1);
 455         if (err)
 456                 return err;
 457
 458         inode_lock(inode);
 459         fuse_sync_writes(inode);
 460         inode_unlock(inode);
 461
 462         err = filemap_check_errors(file->f_mapping);
 463         if (err)
 464                 return err;
 465
 466         err = 0;
 467         if (fc->no_flush)
 468                 goto inval_attr_out;
 469
 470         memset(&inarg, 0, sizeof(inarg));
 471         inarg.fh = ff->fh;
 472         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 473         args.opcode = FUSE_FLUSH;
 474         args.nodeid = get_node_id(inode);
 475         args.in_numargs = 1;
 476         args.in_args[0].size = sizeof(inarg);
 477         args.in_args[0].value = &inarg;
 478         args.force = true;
 479
 480         err = fuse_simple_request(fc, &args);
 481         if (err == -ENOSYS) {
 482                 fc->no_flush = 1;
 483                 err = 0;
 484         }
 485
 486 inval_attr_out:
 487         /*
 488          * In memory i_blocks is not maintained by fuse, if writeback cache is
 489          * enabled, i_blocks from cached attr may not be accurate.
 490          */
 491         if (!err && fc->writeback_cache)
 492                 fuse_invalidate_attr(inode);
 493         return err;
 494 }
 495
 496 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 497                       int datasync, int opcode)
 498 {
 499         struct inode *inode = file->f_mapping->host;
 500         struct fuse_conn *fc = get_fuse_conn(inode);
 501         struct fuse_file *ff = file->private_data;
 502         FUSE_ARGS(args);
 503         struct fuse_fsync_in inarg;
 504
 505         memset(&inarg, 0, sizeof(inarg));
 506         inarg.fh = ff->fh;
 507         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 508         args.opcode = opcode;
 509         args.nodeid = get_node_id(inode);
 510         args.in_numargs = 1;
 511         args.in_args[0].size = sizeof(inarg);
 512         args.in_args[0].value = &inarg;
 513         return fuse_simple_request(fc, &args);
 514 }
 515
 516 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 517                       int datasync)
 518 {
 519         struct inode *inode = file->f_mapping->host;
 520         struct fuse_conn *fc = get_fuse_conn(inode);
 521         int err;
 522
 523         if (is_bad_inode(inode))
 524                 return -EIO;
 525
 526         inode_lock(inode);
 527
 528         /*
 529          * Start writeback against all dirty pages of the inode, then
 530          * wait for all outstanding writes, before sending the FSYNC
 531          * request.
 532          */
 533         err = file_write_and_wait_range(file, start, end);
 534         if (err)
 535                 goto out;
 536
 537         fuse_sync_writes(inode);
 538
 539         /*
 540          * Due to implementation of fuse writeback
 541          * file_write_and_wait_range() does not catch errors.
 542          * We have to do this directly after fuse_sync_writes()
 543          */
 544         err = file_check_and_advance_wb_err(file);
 545         if (err)
 546                 goto out;
 547
 548         err = sync_inode_metadata(inode, 1);
 549         if (err)
 550                 goto out;
 551
 552         if (fc->no_fsync)
 553                 goto out;
 554
 555         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 556         if (err == -ENOSYS) {
 557                 fc->no_fsync = 1;
 558                 err = 0;
 559         }
 560 out:
 561         inode_unlock(inode);
 562
 563         return err;
 564 }
 565
 566 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
 567                          size_t count, int opcode)
 568 {
 569         struct fuse_file *ff = file->private_data;
 570         struct fuse_args *args = &ia->ap.args;
 571
 572         ia->read.in.fh = ff->fh;
 573         ia->read.in.offset = pos;
 574         ia->read.in.size = count;
 575         ia->read.in.flags = file->f_flags;
 576         args->opcode = opcode;
 577         args->nodeid = ff->nodeid;
 578         args->in_numargs = 1;
 579         args->in_args[0].size = sizeof(ia->read.in);
 580         args->in_args[0].value = &ia->read.in;
 581         args->out_argvar = true;
 582         args->out_numargs = 1;
 583         args->out_args[0].size = count;
 584 }
 585
 586 static void fuse_release_user_pages(struct fuse_args_pages *ap,
 587                                     bool should_dirty)
 588 {
 589         unsigned int i;
 590
 591         for (i = 0; i < ap->num_pages; i++) {
 592                 if (should_dirty)
 593                         set_page_dirty_lock(ap->pages[i]);
 594                 put_page(ap->pages[i]);
 595         }
 596 }
 597
 598 static void fuse_io_release(struct kref *kref)
 599 {
 600         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 601 }
 602
 603 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 604 {
 605         if (io->err)
 606                 return io->err;
 607
 608         if (io->bytes >= 0 && io->write)
 609                 return -EIO;
 610
 611         return io->bytes < 0 ? io->size : io->bytes;
 612 }
 613
 614 /**
 615  * In case of short read, the caller sets 'pos' to the position of
 616  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 617  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 618  *
 619  * An example:
 620  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 621  * both submitted asynchronously. The first of them was ACKed by userspace as
 622  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 623  * second request was ACKed as short, e.g. only 1K was read, resulting in
 624  * pos == 33K.
 625  *
 626  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 627  * will be equal to the length of the longest contiguous fragment of
 628  * transferred data starting from the beginning of IO request.
 629  */
 630 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 631 {
 632         int left;
 633
 634         spin_lock(&io->lock);
 635         if (err)
 636                 io->err = io->err ? : err;
 637         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 638                 io->bytes = pos;
 639
 640         left = --io->reqs;
 641         if (!left && io->blocking)
 642                 complete(io->done);
 643         spin_unlock(&io->lock);
 644
 645         if (!left && !io->blocking) {
 646                 ssize_t res = fuse_get_res_by_io(io);
 647
 648                 if (res >= 0) {
 649                         struct inode *inode = file_inode(io->iocb->ki_filp);
 650                         struct fuse_conn *fc = get_fuse_conn(inode);
 651                         struct fuse_inode *fi = get_fuse_inode(inode);
 652
 653                         spin_lock(&fi->lock);
 654                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 655                         spin_unlock(&fi->lock);
 656                 }
 657
 658                 io->iocb->ki_complete(io->iocb, res, 0);
 659         }
 660
 661         kref_put(&io->refcnt, fuse_io_release);
 662 }
 663
 664 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
 665                                           unsigned int npages)
 666 {
 667         struct fuse_io_args *ia;
 668
 669         ia = kzalloc(sizeof(*ia), GFP_KERNEL);
 670         if (ia) {
 671                 ia->io = io;
 672                 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
 673                                                 &ia->ap.descs);
 674                 if (!ia->ap.pages) {
 675                         kfree(ia);
 676                         ia = NULL;
 677                 }
 678         }
 679         return ia;
 680 }
 681
 682 static void fuse_io_free(struct fuse_io_args *ia)
 683 {
 684         kfree(ia->ap.pages);
 685         kfree(ia);
 686 }
 687
 688 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_args *args,
 689                                   int err)
 690 {
 691         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 692         struct fuse_io_priv *io = ia->io;
 693         ssize_t pos = -1;
 694
 695         fuse_release_user_pages(&ia->ap, io->should_dirty);
 696
 697         if (err) {
 698                 /* Nothing */
 699         } else if (io->write) {
 700                 if (ia->write.out.size > ia->write.in.size) {
 701                         err = -EIO;
 702                 } else if (ia->write.in.size != ia->write.out.size) {
 703                         pos = ia->write.in.offset - io->offset +
 704                                 ia->write.out.size;
 705                 }
 706         } else {
 707                 u32 outsize = args->out_args[0].size;
 708
 709                 if (ia->read.in.size != outsize)
 710                         pos = ia->read.in.offset - io->offset + outsize;
 711         }
 712
 713         fuse_aio_complete(io, err, pos);
 714         fuse_io_free(ia);
 715 }
 716
 717 static ssize_t fuse_async_req_send(struct fuse_conn *fc,
 718                                    struct fuse_io_args *ia, size_t num_bytes)
 719 {
 720         ssize_t err;
 721         struct fuse_io_priv *io = ia->io;
 722
 723         spin_lock(&io->lock);
 724         kref_get(&io->refcnt);
 725         io->size += num_bytes;
 726         io->reqs++;
 727         spin_unlock(&io->lock);
 728
 729         ia->ap.args.end = fuse_aio_complete_req;
 730         ia->ap.args.may_block = io->should_dirty;
 731         err = fuse_simple_background(fc, &ia->ap.args, GFP_KERNEL);
 732         if (err)
 733                 fuse_aio_complete_req(fc, &ia->ap.args, err);
 734
 735         return num_bytes;
 736 }
 737
 738 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
 739                               fl_owner_t owner)
 740 {
 741         struct file *file = ia->io->iocb->ki_filp;
 742         struct fuse_file *ff = file->private_data;
 743         struct fuse_conn *fc = ff->fc;
 744
 745         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 746         if (owner != NULL) {
 747                 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
 748                 ia->read.in.lock_owner = fuse_lock_owner_id(fc, owner);
 749         }
 750
 751         if (ia->io->async)
 752                 return fuse_async_req_send(fc, ia, count);
 753
 754         return fuse_simple_request(fc, &ia->ap.args);
 755 }
 756
 757 static void fuse_read_update_size(struct inode *inode, loff_t size,
 758                                   u64 attr_ver)
 759 {
 760         struct fuse_conn *fc = get_fuse_conn(inode);
 761         struct fuse_inode *fi = get_fuse_inode(inode);
 762
 763         spin_lock(&fi->lock);
 764         if (attr_ver == fi->attr_version && size < inode->i_size &&
 765             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 766                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 767                 i_size_write(inode, size);
 768         }
 769         spin_unlock(&fi->lock);
 770 }
 771
 772 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
 773                             struct fuse_args_pages *ap)
 774 {
 775         struct fuse_conn *fc = get_fuse_conn(inode);
 776
 777         if (fc->writeback_cache) {
 778                 /*
 779                  * A hole in a file. Some data after the hole are in page cache,
 780                  * but have not reached the client fs yet. So, the hole is not
 781                  * present there.
 782                  */
 783                 int i;
 784                 int start_idx = num_read >> PAGE_SHIFT;
 785                 size_t off = num_read & (PAGE_SIZE - 1);
 786
 787                 for (i = start_idx; i < ap->num_pages; i++) {
 788                         zero_user_segment(ap->pages[i], off, PAGE_SIZE);
 789                         off = 0;
 790                 }
 791         } else {
 792                 loff_t pos = page_offset(ap->pages[0]) + num_read;
 793                 fuse_read_update_size(inode, pos, attr_ver);
 794         }
 795 }
 796
 797 static int fuse_do_readpage(struct file *file, struct page *page)
 798 {
 799         struct inode *inode = page->mapping->host;
 800         struct fuse_conn *fc = get_fuse_conn(inode);
 801         loff_t pos = page_offset(page);
 802         struct fuse_page_desc desc = { .length = PAGE_SIZE };
 803         struct fuse_io_args ia = {
 804                 .ap.args.page_zeroing = true,
 805                 .ap.args.out_pages = true,
 806                 .ap.num_pages = 1,
 807                 .ap.pages = &page,
 808                 .ap.descs = &desc,
 809         };
 810         ssize_t res;
 811         u64 attr_ver;
 812
 813         /*
 814          * Page writeback can extend beyond the lifetime of the
 815          * page-cache page, so make sure we read a properly synced
 816          * page.
 817          */
 818         fuse_wait_on_page_writeback(inode, page->index);
 819
 820         attr_ver = fuse_get_attr_version(fc);
 821
 822         /* Don't overflow end offset */
 823         if (pos + (desc.length - 1) == LLONG_MAX)
 824                 desc.length--;
 825
 826         fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
 827         res = fuse_simple_request(fc, &ia.ap.args);
 828         if (res < 0)
 829                 return res;
 830         /*
 831          * Short read means EOF.  If file size is larger, truncate it
 832          */
 833         if (res < desc.length)
 834                 fuse_short_read(inode, attr_ver, res, &ia.ap);
 835
 836         SetPageUptodate(page);
 837
 838         return 0;
 839 }
 840
 841 static int fuse_readpage(struct file *file, struct page *page)
 842 {
 843         struct inode *inode = page->mapping->host;
 844         int err;
 845
 846         err = -EIO;
 847         if (is_bad_inode(inode))
 848                 goto out;
 849
 850         err = fuse_do_readpage(file, page);
 851         fuse_invalidate_atime(inode);
 852  out:
 853         unlock_page(page);
 854         return err;
 855 }
 856
 857 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_args *args,
 858                                int err)
 859 {
 860         int i;
 861         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 862         struct fuse_args_pages *ap = &ia->ap;
 863         size_t count = ia->read.in.size;
 864         size_t num_read = args->out_args[0].size;
 865         struct address_space *mapping = NULL;
 866
 867         for (i = 0; mapping == NULL && i < ap->num_pages; i++)
 868                 mapping = ap->pages[i]->mapping;
 869
 870         if (mapping) {
 871                 struct inode *inode = mapping->host;
 872
 873                 /*
 874                  * Short read means EOF. If file size is larger, truncate it
 875                  */
 876                 if (!err && num_read < count)
 877                         fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
 878
 879                 fuse_invalidate_atime(inode);
 880         }
 881
 882         for (i = 0; i < ap->num_pages; i++) {
 883                 struct page *page = ap->pages[i];
 884
 885                 if (!err)
 886                         SetPageUptodate(page);
 887                 else
 888                         SetPageError(page);
 889                 unlock_page(page);
 890                 put_page(page);
 891         }
 892         if (ia->ff)
 893                 fuse_file_put(ia->ff, false, false);
 894
 895         fuse_io_free(ia);
 896 }
 897
 898 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
 899 {
 900         struct fuse_file *ff = file->private_data;
 901         struct fuse_conn *fc = ff->fc;
 902         struct fuse_args_pages *ap = &ia->ap;
 903         loff_t pos = page_offset(ap->pages[0]);
 904         size_t count = ap->num_pages << PAGE_SHIFT;
 905         ssize_t res;
 906         int err;
 907
 908         ap->args.out_pages = true;
 909         ap->args.page_zeroing = true;
 910         ap->args.page_replace = true;
 911
 912         /* Don't overflow end offset */
 913         if (pos + (count - 1) == LLONG_MAX) {
 914                 count--;
 915                 ap->descs[ap->num_pages - 1].length--;
 916         }
 917         WARN_ON((loff_t) (pos + count) < 0);
 918
 919         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 920         ia->read.attr_ver = fuse_get_attr_version(fc);
 921         if (fc->async_read) {
 922                 ia->ff = fuse_file_get(ff);
 923                 ap->args.end = fuse_readpages_end;
 924                 err = fuse_simple_background(fc, &ap->args, GFP_KERNEL);
 925                 if (!err)
 926                         return;
 927         } else {
 928                 res = fuse_simple_request(fc, &ap->args);
 929                 err = res < 0 ? res : 0;
 930         }
 931         fuse_readpages_end(fc, &ap->args, err);
 932 }
 933
 934 static void fuse_readahead(struct readahead_control *rac)
 935 {
 936         struct inode *inode = rac->mapping->host;
 937         struct fuse_conn *fc = get_fuse_conn(inode);
 938         unsigned int i, max_pages, nr_pages = 0;
 939
 940         if (is_bad_inode(inode))
 941                 return;
 942
 943         max_pages = min_t(unsigned int, fc->max_pages,
 944                         fc->max_read / PAGE_SIZE);
 945
 946         for (;;) {
 947                 struct fuse_io_args *ia;
 948                 struct fuse_args_pages *ap;
 949
 950                 nr_pages = readahead_count(rac) - nr_pages;
 951                 if (nr_pages > max_pages)
 952                         nr_pages = max_pages;
 953                 if (nr_pages == 0)
 954                         break;
 955                 ia = fuse_io_alloc(NULL, nr_pages);
 956                 if (!ia)
 957                         return;
 958                 ap = &ia->ap;
 959                 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
 960                 for (i = 0; i < nr_pages; i++) {
 961                         fuse_wait_on_page_writeback(inode,
 962                                                     readahead_index(rac) + i);
 963                         ap->descs[i].length = PAGE_SIZE;
 964                 }
 965                 ap->num_pages = nr_pages;
 966                 fuse_send_readpages(ia, rac->file);
 967         }
 968 }
 969
 970 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
 971 {
 972         struct inode *inode = iocb->ki_filp->f_mapping->host;
 973         struct fuse_conn *fc = get_fuse_conn(inode);
 974
 975         /*
 976          * In auto invalidate mode, always update attributes on read.
 977          * Otherwise, only update if we attempt to read past EOF (to ensure
 978          * i_size is up to date).
 979          */
 980         if (fc->auto_inval_data ||
 981             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 982                 int err;
 983                 err = fuse_update_attributes(inode, iocb->ki_filp);
 984                 if (err)
 985                         return err;
 986         }
 987
 988         return generic_file_read_iter(iocb, to);
 989 }
 990
 991 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
 992                                  loff_t pos, size_t count)
 993 {
 994         struct fuse_args *args = &ia->ap.args;
 995
 996         ia->write.in.fh = ff->fh;
 997         ia->write.in.offset = pos;
 998         ia->write.in.size = count;
 999         args->opcode = FUSE_WRITE;
1000         args->nodeid = ff->nodeid;
1001         args->in_numargs = 2;
1002         if (ff->fc->minor < 9)
1003                 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1004         else
1005                 args->in_args[0].size = sizeof(ia->write.in);
1006         args->in_args[0].value = &ia->write.in;
1007         args->in_args[1].size = count;
1008         args->out_numargs = 1;
1009         args->out_args[0].size = sizeof(ia->write.out);
1010         args->out_args[0].value = &ia->write.out;
1011 }
1012
1013 static unsigned int fuse_write_flags(struct kiocb *iocb)
1014 {
1015         unsigned int flags = iocb->ki_filp->f_flags;
1016
1017         if (iocb->ki_flags & IOCB_DSYNC)
1018                 flags |= O_DSYNC;
1019         if (iocb->ki_flags & IOCB_SYNC)
1020                 flags |= O_SYNC;
1021
1022         return flags;
1023 }
1024
1025 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1026                                size_t count, fl_owner_t owner)
1027 {
1028         struct kiocb *iocb = ia->io->iocb;
1029         struct file *file = iocb->ki_filp;
1030         struct fuse_file *ff = file->private_data;
1031         struct fuse_conn *fc = ff->fc;
1032         struct fuse_write_in *inarg = &ia->write.in;
1033         ssize_t err;
1034
1035         fuse_write_args_fill(ia, ff, pos, count);
1036         inarg->flags = fuse_write_flags(iocb);
1037         if (owner != NULL) {
1038                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1039                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
1040         }
1041
1042         if (ia->io->async)
1043                 return fuse_async_req_send(fc, ia, count);
1044
1045         err = fuse_simple_request(fc, &ia->ap.args);
1046         if (!err && ia->write.out.size > count)
1047                 err = -EIO;
1048
1049         return err ?: ia->write.out.size;
1050 }
1051
1052 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1053 {
1054         struct fuse_conn *fc = get_fuse_conn(inode);
1055         struct fuse_inode *fi = get_fuse_inode(inode);
1056         bool ret = false;
1057
1058         spin_lock(&fi->lock);
1059         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1060         if (pos > inode->i_size) {
1061                 i_size_write(inode, pos);
1062                 ret = true;
1063         }
1064         spin_unlock(&fi->lock);
1065
1066         return ret;
1067 }
1068
1069 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1070                                      struct kiocb *iocb, struct inode *inode,
1071                                      loff_t pos, size_t count)
1072 {
1073         struct fuse_args_pages *ap = &ia->ap;
1074         struct file *file = iocb->ki_filp;
1075         struct fuse_file *ff = file->private_data;
1076         struct fuse_conn *fc = ff->fc;
1077         unsigned int offset, i;
1078         int err;
1079
1080         for (i = 0; i < ap->num_pages; i++)
1081                 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1082
1083         fuse_write_args_fill(ia, ff, pos, count);
1084         ia->write.in.flags = fuse_write_flags(iocb);
1085
1086         err = fuse_simple_request(fc, &ap->args);
1087         if (!err && ia->write.out.size > count)
1088                 err = -EIO;
1089
1090         offset = ap->descs[0].offset;
1091         count = ia->write.out.size;
1092         for (i = 0; i < ap->num_pages; i++) {
1093                 struct page *page = ap->pages[i];
1094
1095                 if (!err && !offset && count >= PAGE_SIZE)
1096                         SetPageUptodate(page);
1097
1098                 if (count > PAGE_SIZE - offset)
1099                         count -= PAGE_SIZE - offset;
1100                 else
1101                         count = 0;
1102                 offset = 0;
1103
1104                 unlock_page(page);
1105                 put_page(page);
1106         }
1107
1108         return err;
1109 }
1110
1111 static ssize_t fuse_fill_write_pages(struct fuse_args_pages *ap,
1112                                      struct address_space *mapping,
1113                                      struct iov_iter *ii, loff_t pos,
1114                                      unsigned int max_pages)
1115 {
1116         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1117         unsigned offset = pos & (PAGE_SIZE - 1);
1118         size_t count = 0;
1119         int err;
1120
1121         ap->args.in_pages = true;
1122         ap->descs[0].offset = offset;
1123
1124         do {
1125                 size_t tmp;
1126                 struct page *page;
1127                 pgoff_t index = pos >> PAGE_SHIFT;
1128                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1129                                      iov_iter_count(ii));
1130
1131                 bytes = min_t(size_t, bytes, fc->max_write - count);
1132
1133  again:
1134                 err = -EFAULT;
1135                 if (iov_iter_fault_in_readable(ii, bytes))
1136                         break;
1137
1138                 err = -ENOMEM;
1139                 page = grab_cache_page_write_begin(mapping, index, 0);
1140                 if (!page)
1141                         break;
1142
1143                 if (mapping_writably_mapped(mapping))
1144                         flush_dcache_page(page);
1145
1146                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1147                 flush_dcache_page(page);
1148
1149                 iov_iter_advance(ii, tmp);
1150                 if (!tmp) {
1151                         unlock_page(page);
1152                         put_page(page);
1153                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1154                         goto again;
1155                 }
1156
1157                 err = 0;
1158                 ap->pages[ap->num_pages] = page;
1159                 ap->descs[ap->num_pages].length = tmp;
1160                 ap->num_pages++;
1161
1162                 count += tmp;
1163                 pos += tmp;
1164                 offset += tmp;
1165                 if (offset == PAGE_SIZE)
1166                         offset = 0;
1167
1168                 if (!fc->big_writes)
1169                         break;
1170         } while (iov_iter_count(ii) && count < fc->max_write &&
1171                  ap->num_pages < max_pages && offset == 0);
1172
1173         return count > 0 ? count : err;
1174 }
1175
1176 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1177                                      unsigned int max_pages)
1178 {
1179         return min_t(unsigned int,
1180                      ((pos + len - 1) >> PAGE_SHIFT) -
1181                      (pos >> PAGE_SHIFT) + 1,
1182                      max_pages);
1183 }
1184
1185 static ssize_t fuse_perform_write(struct kiocb *iocb,
1186                                   struct address_space *mapping,
1187                                   struct iov_iter *ii, loff_t pos)
1188 {
1189         struct inode *inode = mapping->host;
1190         struct fuse_conn *fc = get_fuse_conn(inode);
1191         struct fuse_inode *fi = get_fuse_inode(inode);
1192         int err = 0;
1193         ssize_t res = 0;
1194
1195         if (inode->i_size < pos + iov_iter_count(ii))
1196                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1197
1198         do {
1199                 ssize_t count;
1200                 struct fuse_io_args ia = {};
1201                 struct fuse_args_pages *ap = &ia.ap;
1202                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1203                                                       fc->max_pages);
1204
1205                 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1206                 if (!ap->pages) {
1207                         err = -ENOMEM;
1208                         break;
1209                 }
1210
1211                 count = fuse_fill_write_pages(ap, mapping, ii, pos, nr_pages);
1212                 if (count <= 0) {
1213                         err = count;
1214                 } else {
1215                         err = fuse_send_write_pages(&ia, iocb, inode,
1216                                                     pos, count);
1217                         if (!err) {
1218                                 size_t num_written = ia.write.out.size;
1219
1220                                 res += num_written;
1221                                 pos += num_written;
1222
1223                                 /* break out of the loop on short write */
1224                                 if (num_written != count)
1225                                         err = -EIO;
1226                         }
1227                 }
1228                 kfree(ap->pages);
1229         } while (!err && iov_iter_count(ii));
1230
1231         if (res > 0)
1232                 fuse_write_update_size(inode, pos);
1233
1234         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1235         fuse_invalidate_attr(inode);
1236
1237         return res > 0 ? res : err;
1238 }
1239
1240 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1241 {
1242         struct file *file = iocb->ki_filp;
1243         struct address_space *mapping = file->f_mapping;
1244         ssize_t written = 0;
1245         ssize_t written_buffered = 0;
1246         struct inode *inode = mapping->host;
1247         ssize_t err;
1248         loff_t endbyte = 0;
1249
1250         if (get_fuse_conn(inode)->writeback_cache) {
1251                 /* Update size (EOF optimization) and mode (SUID clearing) */
1252                 err = fuse_update_attributes(mapping->host, file);
1253                 if (err)
1254                         return err;
1255
1256                 return generic_file_write_iter(iocb, from);
1257         }
1258
1259         inode_lock(inode);
1260
1261         /* We can write back this queue in page reclaim */
1262         current->backing_dev_info = inode_to_bdi(inode);
1263
1264         err = generic_write_checks(iocb, from);
1265         if (err <= 0)
1266                 goto out;
1267
1268         err = file_remove_privs(file);
1269         if (err)
1270                 goto out;
1271
1272         err = file_update_time(file);
1273         if (err)
1274                 goto out;
1275
1276         if (iocb->ki_flags & IOCB_DIRECT) {
1277                 loff_t pos = iocb->ki_pos;
1278                 written = generic_file_direct_write(iocb, from);
1279                 if (written < 0 || !iov_iter_count(from))
1280                         goto out;
1281
1282                 pos += written;
1283
1284                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1285                 if (written_buffered < 0) {
1286                         err = written_buffered;
1287                         goto out;
1288                 }
1289                 endbyte = pos + written_buffered - 1;
1290
1291                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1292                                                    endbyte);
1293                 if (err)
1294                         goto out;
1295
1296                 invalidate_mapping_pages(file->f_mapping,
1297                                          pos >> PAGE_SHIFT,
1298                                          endbyte >> PAGE_SHIFT);
1299
1300                 written += written_buffered;
1301                 iocb->ki_pos = pos + written_buffered;
1302         } else {
1303                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1304                 if (written >= 0)
1305                         iocb->ki_pos += written;
1306         }
1307 out:
1308         current->backing_dev_info = NULL;
1309         inode_unlock(inode);
1310         if (written > 0)
1311                 written = generic_write_sync(iocb, written);
1312
1313         return written ? written : err;
1314 }
1315
1316 static inline void fuse_page_descs_length_init(struct fuse_page_desc *descs,
1317                                                unsigned int index,
1318                                                unsigned int nr_pages)
1319 {
1320         int i;
1321
1322         for (i = index; i < index + nr_pages; i++)
1323                 descs[i].length = PAGE_SIZE - descs[i].offset;
1324 }
1325
1326 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1327 {
1328         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1329 }
1330
1331 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1332                                         size_t max_size)
1333 {
1334         return min(iov_iter_single_seg_count(ii), max_size);
1335 }
1336
1337 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1338                                size_t *nbytesp, int write,
1339                                unsigned int max_pages)
1340 {
1341         size_t nbytes = 0;  /* # bytes already packed in req */
1342         ssize_t ret = 0;
1343
1344         /* Special case for kernel I/O: can copy directly into the buffer */
1345         if (iov_iter_is_kvec(ii)) {
1346                 unsigned long user_addr = fuse_get_user_addr(ii);
1347                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1348
1349                 if (write)
1350                         ap->args.in_args[1].value = (void *) user_addr;
1351                 else
1352                         ap->args.out_args[0].value = (void *) user_addr;
1353
1354                 iov_iter_advance(ii, frag_size);
1355                 *nbytesp = frag_size;
1356                 return 0;
1357         }
1358
1359         while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1360                 unsigned npages;
1361                 size_t start;
1362                 ret = iov_iter_get_pages(ii, &ap->pages[ap->num_pages],
1363                                         *nbytesp - nbytes,
1364                                         max_pages - ap->num_pages,
1365                                         &start);
1366                 if (ret < 0)
1367                         break;
1368
1369                 iov_iter_advance(ii, ret);
1370                 nbytes += ret;
1371
1372                 ret += start;
1373                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1374
1375                 ap->descs[ap->num_pages].offset = start;
1376                 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1377
1378                 ap->num_pages += npages;
1379                 ap->descs[ap->num_pages - 1].length -=
1380                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1381         }
1382
1383         if (write)
1384                 ap->args.in_pages = true;
1385         else
1386                 ap->args.out_pages = true;
1387
1388         *nbytesp = nbytes;
1389
1390         return ret < 0 ? ret : 0;
1391 }
1392
1393 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1394                        loff_t *ppos, int flags)
1395 {
1396         int write = flags & FUSE_DIO_WRITE;
1397         int cuse = flags & FUSE_DIO_CUSE;
1398         struct file *file = io->iocb->ki_filp;
1399         struct inode *inode = file->f_mapping->host;
1400         struct fuse_file *ff = file->private_data;
1401         struct fuse_conn *fc = ff->fc;
1402         size_t nmax = write ? fc->max_write : fc->max_read;
1403         loff_t pos = *ppos;
1404         size_t count = iov_iter_count(iter);
1405         pgoff_t idx_from = pos >> PAGE_SHIFT;
1406         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1407         ssize_t res = 0;
1408         int err = 0;
1409         struct fuse_io_args *ia;
1410         unsigned int max_pages;
1411
1412         max_pages = iov_iter_npages(iter, fc->max_pages);
1413         ia = fuse_io_alloc(io, max_pages);
1414         if (!ia)
1415                 return -ENOMEM;
1416
1417         ia->io = io;
1418         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1419                 if (!write)
1420                         inode_lock(inode);
1421                 fuse_sync_writes(inode);
1422                 if (!write)
1423                         inode_unlock(inode);
1424         }
1425
1426         io->should_dirty = !write && iter_is_iovec(iter);
1427         while (count) {
1428                 ssize_t nres;
1429                 fl_owner_t owner = current->files;
1430                 size_t nbytes = min(count, nmax);
1431
1432                 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1433                                           max_pages);
1434                 if (err && !nbytes)
1435                         break;
1436
1437                 if (write) {
1438                         if (!capable(CAP_FSETID))
1439                                 ia->write.in.write_flags |= FUSE_WRITE_KILL_PRIV;
1440
1441                         nres = fuse_send_write(ia, pos, nbytes, owner);
1442                 } else {
1443                         nres = fuse_send_read(ia, pos, nbytes, owner);
1444                 }
1445
1446                 if (!io->async || nres < 0) {
1447                         fuse_release_user_pages(&ia->ap, io->should_dirty);
1448                         fuse_io_free(ia);
1449                 }
1450                 ia = NULL;
1451                 if (nres < 0) {
1452                         iov_iter_revert(iter, nbytes);
1453                         err = nres;
1454                         break;
1455                 }
1456                 WARN_ON(nres > nbytes);
1457
1458                 count -= nres;
1459                 res += nres;
1460                 pos += nres;
1461                 if (nres != nbytes) {
1462                         iov_iter_revert(iter, nbytes - nres);
1463                         break;
1464                 }
1465                 if (count) {
1466                         max_pages = iov_iter_npages(iter, fc->max_pages);
1467                         ia = fuse_io_alloc(io, max_pages);
1468                         if (!ia)
1469                                 break;
1470                 }
1471         }
1472         if (ia)
1473                 fuse_io_free(ia);
1474         if (res > 0)
1475                 *ppos = pos;
1476
1477         return res > 0 ? res : err;
1478 }
1479 EXPORT_SYMBOL_GPL(fuse_direct_io);
1480
1481 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1482                                   struct iov_iter *iter,
1483                                   loff_t *ppos)
1484 {
1485         ssize_t res;
1486         struct inode *inode = file_inode(io->iocb->ki_filp);
1487
1488         res = fuse_direct_io(io, iter, ppos, 0);
1489
1490         fuse_invalidate_atime(inode);
1491
1492         return res;
1493 }
1494
1495 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1496
1497 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1498 {
1499         ssize_t res;
1500
1501         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1502                 res = fuse_direct_IO(iocb, to);
1503         } else {
1504                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1505
1506                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1507         }
1508
1509         return res;
1510 }
1511
1512 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1513 {
1514         struct inode *inode = file_inode(iocb->ki_filp);
1515         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1516         ssize_t res;
1517
1518         /* Don't allow parallel writes to the same file */
1519         inode_lock(inode);
1520         res = generic_write_checks(iocb, from);
1521         if (res > 0) {
1522                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1523                         res = fuse_direct_IO(iocb, from);
1524                 } else {
1525                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1526                                              FUSE_DIO_WRITE);
1527                 }
1528         }
1529         fuse_invalidate_attr(inode);
1530         if (res > 0)
1531                 fuse_write_update_size(inode, iocb->ki_pos);
1532         inode_unlock(inode);
1533
1534         return res;
1535 }
1536
1537 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1538 {
1539         struct file *file = iocb->ki_filp;
1540         struct fuse_file *ff = file->private_data;
1541
1542         if (is_bad_inode(file_inode(file)))
1543                 return -EIO;
1544
1545         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1546                 return fuse_cache_read_iter(iocb, to);
1547         else
1548                 return fuse_direct_read_iter(iocb, to);
1549 }
1550
1551 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1552 {
1553         struct file *file = iocb->ki_filp;
1554         struct fuse_file *ff = file->private_data;
1555
1556         if (is_bad_inode(file_inode(file)))
1557                 return -EIO;
1558
1559         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1560                 return fuse_cache_write_iter(iocb, from);
1561         else
1562                 return fuse_direct_write_iter(iocb, from);
1563 }
1564
1565 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1566 {
1567         struct fuse_args_pages *ap = &wpa->ia.ap;
1568         int i;
1569
1570         for (i = 0; i < ap->num_pages; i++)
1571                 __free_page(ap->pages[i]);
1572
1573         if (wpa->ia.ff)
1574                 fuse_file_put(wpa->ia.ff, false, false);
1575
1576         kfree(ap->pages);
1577         kfree(wpa);
1578 }
1579
1580 static void fuse_writepage_finish(struct fuse_conn *fc,
1581                                   struct fuse_writepage_args *wpa)
1582 {
1583         struct fuse_args_pages *ap = &wpa->ia.ap;
1584         struct inode *inode = wpa->inode;
1585         struct fuse_inode *fi = get_fuse_inode(inode);
1586         struct backing_dev_info *bdi = inode_to_bdi(inode);
1587         int i;
1588
1589         rb_erase(&wpa->writepages_entry, &fi->writepages);
1590         for (i = 0; i < ap->num_pages; i++) {
1591                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1592                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1593                 wb_writeout_inc(&bdi->wb);
1594         }
1595         wake_up(&fi->page_waitq);
1596 }
1597
1598 /* Called under fi->lock, may release and reacquire it */
1599 static void fuse_send_writepage(struct fuse_conn *fc,
1600                                 struct fuse_writepage_args *wpa, loff_t size)
1601 __releases(fi->lock)
1602 __acquires(fi->lock)
1603 {
1604         struct fuse_writepage_args *aux, *next;
1605         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1606         struct fuse_write_in *inarg = &wpa->ia.write.in;
1607         struct fuse_args *args = &wpa->ia.ap.args;
1608         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1609         int err;
1610
1611         fi->writectr++;
1612         if (inarg->offset + data_size <= size) {
1613                 inarg->size = data_size;
1614         } else if (inarg->offset < size) {
1615                 inarg->size = size - inarg->offset;
1616         } else {
1617                 /* Got truncated off completely */
1618                 goto out_free;
1619         }
1620
1621         args->in_args[1].size = inarg->size;
1622         args->force = true;
1623         args->nocreds = true;
1624
1625         err = fuse_simple_background(fc, args, GFP_ATOMIC);
1626         if (err == -ENOMEM) {
1627                 spin_unlock(&fi->lock);
1628                 err = fuse_simple_background(fc, args, GFP_NOFS | __GFP_NOFAIL);
1629                 spin_lock(&fi->lock);
1630         }
1631
1632         /* Fails on broken connection only */
1633         if (unlikely(err))
1634                 goto out_free;
1635
1636         return;
1637
1638  out_free:
1639         fi->writectr--;
1640         fuse_writepage_finish(fc, wpa);
1641         spin_unlock(&fi->lock);
1642
1643         /* After fuse_writepage_finish() aux request list is private */
1644         for (aux = wpa->next; aux; aux = next) {
1645                 next = aux->next;
1646                 aux->next = NULL;
1647                 fuse_writepage_free(aux);
1648         }
1649
1650         fuse_writepage_free(wpa);
1651         spin_lock(&fi->lock);
1652 }
1653
1654 /*
1655  * If fi->writectr is positive (no truncate or fsync going on) send
1656  * all queued writepage requests.
1657  *
1658  * Called with fi->lock
1659  */
1660 void fuse_flush_writepages(struct inode *inode)
1661 __releases(fi->lock)
1662 __acquires(fi->lock)
1663 {
1664         struct fuse_conn *fc = get_fuse_conn(inode);
1665         struct fuse_inode *fi = get_fuse_inode(inode);
1666         loff_t crop = i_size_read(inode);
1667         struct fuse_writepage_args *wpa;
1668
1669         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1670                 wpa = list_entry(fi->queued_writes.next,
1671                                  struct fuse_writepage_args, queue_entry);
1672                 list_del_init(&wpa->queue_entry);
1673                 fuse_send_writepage(fc, wpa, crop);
1674         }
1675 }
1676
1677 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1678 {
1679         pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1680         pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1681         struct rb_node **p = &root->rb_node;
1682         struct rb_node  *parent = NULL;
1683
1684         WARN_ON(!wpa->ia.ap.num_pages);
1685         while (*p) {
1686                 struct fuse_writepage_args *curr;
1687                 pgoff_t curr_index;
1688
1689                 parent = *p;
1690                 curr = rb_entry(parent, struct fuse_writepage_args,
1691                                 writepages_entry);
1692                 WARN_ON(curr->inode != wpa->inode);
1693                 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1694
1695                 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1696                         p = &(*p)->rb_right;
1697                 else if (idx_to < curr_index)
1698                         p = &(*p)->rb_left;
1699                 else
1700                         return (void) WARN_ON(true);
1701         }
1702
1703         rb_link_node(&wpa->writepages_entry, parent, p);
1704         rb_insert_color(&wpa->writepages_entry, root);
1705 }
1706
1707 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_args *args,
1708                                int error)
1709 {
1710         struct fuse_writepage_args *wpa =
1711                 container_of(args, typeof(*wpa), ia.ap.args);
1712         struct inode *inode = wpa->inode;
1713         struct fuse_inode *fi = get_fuse_inode(inode);
1714
1715         mapping_set_error(inode->i_mapping, error);
1716         spin_lock(&fi->lock);
1717         while (wpa->next) {
1718                 struct fuse_conn *fc = get_fuse_conn(inode);
1719                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1720                 struct fuse_writepage_args *next = wpa->next;
1721
1722                 wpa->next = next->next;
1723                 next->next = NULL;
1724                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1725                 tree_insert(&fi->writepages, next);
1726
1727                 /*
1728                  * Skip fuse_flush_writepages() to make it easy to crop requests
1729                  * based on primary request size.
1730                  *
1731                  * 1st case (trivial): there are no concurrent activities using
1732                  * fuse_set/release_nowrite.  Then we're on safe side because
1733                  * fuse_flush_writepages() would call fuse_send_writepage()
1734                  * anyway.
1735                  *
1736                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1737                  * now for completion of all in-flight requests.  This happens
1738                  * rarely and no more than once per page, so this should be
1739                  * okay.
1740                  *
1741                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1742                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1743                  * that fuse_set_nowrite returned implies that all in-flight
1744                  * requests were completed along with all of their secondary
1745                  * requests.  Further primary requests are blocked by negative
1746                  * writectr.  Hence there cannot be any in-flight requests and
1747                  * no invocations of fuse_writepage_end() while we're in
1748                  * fuse_set_nowrite..fuse_release_nowrite section.
1749                  */
1750                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1751         }
1752         fi->writectr--;
1753         fuse_writepage_finish(fc, wpa);
1754         spin_unlock(&fi->lock);
1755         fuse_writepage_free(wpa);
1756 }
1757
1758 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1759                                                struct fuse_inode *fi)
1760 {
1761         struct fuse_file *ff = NULL;
1762
1763         spin_lock(&fi->lock);
1764         if (!list_empty(&fi->write_files)) {
1765                 ff = list_entry(fi->write_files.next, struct fuse_file,
1766                                 write_entry);
1767                 fuse_file_get(ff);
1768         }
1769         spin_unlock(&fi->lock);
1770
1771         return ff;
1772 }
1773
1774 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1775                                              struct fuse_inode *fi)
1776 {
1777         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1778         WARN_ON(!ff);
1779         return ff;
1780 }
1781
1782 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1783 {
1784         struct fuse_conn *fc = get_fuse_conn(inode);
1785         struct fuse_inode *fi = get_fuse_inode(inode);
1786         struct fuse_file *ff;
1787         int err;
1788
1789         ff = __fuse_write_file_get(fc, fi);
1790         err = fuse_flush_times(inode, ff);
1791         if (ff)
1792                 fuse_file_put(ff, false, false);
1793
1794         return err;
1795 }
1796
1797 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1798 {
1799         struct fuse_writepage_args *wpa;
1800         struct fuse_args_pages *ap;
1801
1802         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1803         if (wpa) {
1804                 ap = &wpa->ia.ap;
1805                 ap->num_pages = 0;
1806                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1807                 if (!ap->pages) {
1808                         kfree(wpa);
1809                         wpa = NULL;
1810                 }
1811         }
1812         return wpa;
1813
1814 }
1815
1816 static int fuse_writepage_locked(struct page *page)
1817 {
1818         struct address_space *mapping = page->mapping;
1819         struct inode *inode = mapping->host;
1820         struct fuse_conn *fc = get_fuse_conn(inode);
1821         struct fuse_inode *fi = get_fuse_inode(inode);
1822         struct fuse_writepage_args *wpa;
1823         struct fuse_args_pages *ap;
1824         struct page *tmp_page;
1825         int error = -ENOMEM;
1826
1827         set_page_writeback(page);
1828
1829         wpa = fuse_writepage_args_alloc();
1830         if (!wpa)
1831                 goto err;
1832         ap = &wpa->ia.ap;
1833
1834         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1835         if (!tmp_page)
1836                 goto err_free;
1837
1838         error = -EIO;
1839         wpa->ia.ff = fuse_write_file_get(fc, fi);
1840         if (!wpa->ia.ff)
1841                 goto err_nofile;
1842
1843         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1844
1845         copy_highpage(tmp_page, page);
1846         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1847         wpa->next = NULL;
1848         ap->args.in_pages = true;
1849         ap->num_pages = 1;
1850         ap->pages[0] = tmp_page;
1851         ap->descs[0].offset = 0;
1852         ap->descs[0].length = PAGE_SIZE;
1853         ap->args.end = fuse_writepage_end;
1854         wpa->inode = inode;
1855
1856         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1857         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1858
1859         spin_lock(&fi->lock);
1860         tree_insert(&fi->writepages, wpa);
1861         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1862         fuse_flush_writepages(inode);
1863         spin_unlock(&fi->lock);
1864
1865         end_page_writeback(page);
1866
1867         return 0;
1868
1869 err_nofile:
1870         __free_page(tmp_page);
1871 err_free:
1872         kfree(wpa);
1873 err:
1874         mapping_set_error(page->mapping, error);
1875         end_page_writeback(page);
1876         return error;
1877 }
1878
1879 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1880 {
1881         int err;
1882
1883         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1884                 /*
1885                  * ->writepages() should be called for sync() and friends.  We
1886                  * should only get here on direct reclaim and then we are
1887                  * allowed to skip a page which is already in flight
1888                  */
1889                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1890
1891                 redirty_page_for_writepage(wbc, page);
1892                 unlock_page(page);
1893                 return 0;
1894         }
1895
1896         err = fuse_writepage_locked(page);
1897         unlock_page(page);
1898
1899         return err;
1900 }
1901
1902 struct fuse_fill_wb_data {
1903         struct fuse_writepage_args *wpa;
1904         struct fuse_file *ff;
1905         struct inode *inode;
1906         struct page **orig_pages;
1907         unsigned int max_pages;
1908 };
1909
1910 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
1911 {
1912         struct fuse_args_pages *ap = &data->wpa->ia.ap;
1913         struct fuse_conn *fc = get_fuse_conn(data->inode);
1914         struct page **pages;
1915         struct fuse_page_desc *descs;
1916         unsigned int npages = min_t(unsigned int,
1917                                     max_t(unsigned int, data->max_pages * 2,
1918                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
1919                                     fc->max_pages);
1920         WARN_ON(npages <= data->max_pages);
1921
1922         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
1923         if (!pages)
1924                 return false;
1925
1926         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
1927         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
1928         kfree(ap->pages);
1929         ap->pages = pages;
1930         ap->descs = descs;
1931         data->max_pages = npages;
1932
1933         return true;
1934 }
1935
1936 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1937 {
1938         struct fuse_writepage_args *wpa = data->wpa;
1939         struct inode *inode = data->inode;
1940         struct fuse_inode *fi = get_fuse_inode(inode);
1941         int num_pages = wpa->ia.ap.num_pages;
1942         int i;
1943
1944         wpa->ia.ff = fuse_file_get(data->ff);
1945         spin_lock(&fi->lock);
1946         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1947         fuse_flush_writepages(inode);
1948         spin_unlock(&fi->lock);
1949
1950         for (i = 0; i < num_pages; i++)
1951                 end_page_writeback(data->orig_pages[i]);
1952 }
1953
1954 /*
1955  * First recheck under fi->lock if the offending offset is still under
1956  * writeback.  If yes, then iterate auxiliary write requests, to see if there's
1957  * one already added for a page at this offset.  If there's none, then insert
1958  * this new request onto the auxiliary list, otherwise reuse the existing one by
1959  * copying the new page contents over to the old temporary page.
1960  */
1961 static bool fuse_writepage_in_flight(struct fuse_writepage_args *new_wpa,
1962                                      struct page *page)
1963 {
1964         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
1965         struct fuse_writepage_args *tmp;
1966         struct fuse_writepage_args *old_wpa;
1967         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
1968
1969         WARN_ON(new_ap->num_pages != 0);
1970
1971         spin_lock(&fi->lock);
1972         rb_erase(&new_wpa->writepages_entry, &fi->writepages);
1973         old_wpa = fuse_find_writeback(fi, page->index, page->index);
1974         if (!old_wpa) {
1975                 tree_insert(&fi->writepages, new_wpa);
1976                 spin_unlock(&fi->lock);
1977                 return false;
1978         }
1979
1980         new_ap->num_pages = 1;
1981         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
1982                 pgoff_t curr_index;
1983
1984                 WARN_ON(tmp->inode != new_wpa->inode);
1985                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
1986                 if (curr_index == page->index) {
1987                         WARN_ON(tmp->ia.ap.num_pages != 1);
1988                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
1989                         break;
1990                 }
1991         }
1992
1993         if (!tmp) {
1994                 new_wpa->next = old_wpa->next;
1995                 old_wpa->next = new_wpa;
1996         }
1997
1998         spin_unlock(&fi->lock);
1999
2000         if (tmp) {
2001                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2002
2003                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2004                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2005                 wb_writeout_inc(&bdi->wb);
2006                 fuse_writepage_free(new_wpa);
2007         }
2008
2009         return true;
2010 }
2011
2012 static int fuse_writepages_fill(struct page *page,
2013                 struct writeback_control *wbc, void *_data)
2014 {
2015         struct fuse_fill_wb_data *data = _data;
2016         struct fuse_writepage_args *wpa = data->wpa;
2017         struct fuse_args_pages *ap = &wpa->ia.ap;
2018         struct inode *inode = data->inode;
2019         struct fuse_inode *fi = get_fuse_inode(inode);
2020         struct fuse_conn *fc = get_fuse_conn(inode);
2021         struct page *tmp_page;
2022         bool is_writeback;
2023         int err;
2024
2025         if (!data->ff) {
2026                 err = -EIO;
2027                 data->ff = fuse_write_file_get(fc, fi);
2028                 if (!data->ff)
2029                         goto out_unlock;
2030         }
2031
2032         /*
2033          * Being under writeback is unlikely but possible.  For example direct
2034          * read to an mmaped fuse file will set the page dirty twice; once when
2035          * the pages are faulted with get_user_pages(), and then after the read
2036          * completed.
2037          */
2038         is_writeback = fuse_page_is_writeback(inode, page->index);
2039
2040         if (wpa && ap->num_pages &&
2041             (is_writeback || ap->num_pages == fc->max_pages ||
2042              (ap->num_pages + 1) * PAGE_SIZE > fc->max_write ||
2043              data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)) {
2044                 fuse_writepages_send(data);
2045                 data->wpa = NULL;
2046         } else if (wpa && ap->num_pages == data->max_pages) {
2047                 if (!fuse_pages_realloc(data)) {
2048                         fuse_writepages_send(data);
2049                         data->wpa = NULL;
2050                 }
2051         }
2052
2053         err = -ENOMEM;
2054         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2055         if (!tmp_page)
2056                 goto out_unlock;
2057
2058         /*
2059          * The page must not be redirtied until the writeout is completed
2060          * (i.e. userspace has sent a reply to the write request).  Otherwise
2061          * there could be more than one temporary page instance for each real
2062          * page.
2063          *
2064          * This is ensured by holding the page lock in page_mkwrite() while
2065          * checking fuse_page_is_writeback().  We already hold the page lock
2066          * since clear_page_dirty_for_io() and keep it held until we add the
2067          * request to the fi->writepages list and increment ap->num_pages.
2068          * After this fuse_page_is_writeback() will indicate that the page is
2069          * under writeback, so we can release the page lock.
2070          */
2071         if (data->wpa == NULL) {
2072                 err = -ENOMEM;
2073                 wpa = fuse_writepage_args_alloc();
2074                 if (!wpa) {
2075                         __free_page(tmp_page);
2076                         goto out_unlock;
2077                 }
2078                 data->max_pages = 1;
2079
2080                 ap = &wpa->ia.ap;
2081                 fuse_write_args_fill(&wpa->ia, data->ff, page_offset(page), 0);
2082                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2083                 wpa->next = NULL;
2084                 ap->args.in_pages = true;
2085                 ap->args.end = fuse_writepage_end;
2086                 ap->num_pages = 0;
2087                 wpa->inode = inode;
2088
2089                 spin_lock(&fi->lock);
2090                 tree_insert(&fi->writepages, wpa);
2091                 spin_unlock(&fi->lock);
2092
2093                 data->wpa = wpa;
2094         }
2095         set_page_writeback(page);
2096
2097         copy_highpage(tmp_page, page);
2098         ap->pages[ap->num_pages] = tmp_page;
2099         ap->descs[ap->num_pages].offset = 0;
2100         ap->descs[ap->num_pages].length = PAGE_SIZE;
2101
2102         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2103         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2104
2105         err = 0;
2106         if (is_writeback && fuse_writepage_in_flight(wpa, page)) {
2107                 end_page_writeback(page);
2108                 data->wpa = NULL;
2109                 goto out_unlock;
2110         }
2111         data->orig_pages[ap->num_pages] = page;
2112
2113         /*
2114          * Protected by fi->lock against concurrent access by
2115          * fuse_page_is_writeback().
2116          */
2117         spin_lock(&fi->lock);
2118         ap->num_pages++;
2119         spin_unlock(&fi->lock);
2120
2121 out_unlock:
2122         unlock_page(page);
2123
2124         return err;
2125 }
2126
2127 static int fuse_writepages(struct address_space *mapping,
2128                            struct writeback_control *wbc)
2129 {
2130         struct inode *inode = mapping->host;
2131         struct fuse_conn *fc = get_fuse_conn(inode);
2132         struct fuse_fill_wb_data data;
2133         int err;
2134
2135         err = -EIO;
2136         if (is_bad_inode(inode))
2137                 goto out;
2138
2139         data.inode = inode;
2140         data.wpa = NULL;
2141         data.ff = NULL;
2142
2143         err = -ENOMEM;
2144         data.orig_pages = kcalloc(fc->max_pages,
2145                                   sizeof(struct page *),
2146                                   GFP_NOFS);
2147         if (!data.orig_pages)
2148                 goto out;
2149
2150         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2151         if (data.wpa) {
2152                 /* Ignore errors if we can write at least one page */
2153                 WARN_ON(!data.wpa->ia.ap.num_pages);
2154                 fuse_writepages_send(&data);
2155                 err = 0;
2156         }
2157         if (data.ff)
2158                 fuse_file_put(data.ff, false, false);
2159
2160         kfree(data.orig_pages);
2161 out:
2162         return err;
2163 }
2164
2165 /*
2166  * It's worthy to make sure that space is reserved on disk for the write,
2167  * but how to implement it without killing performance need more thinking.
2168  */
2169 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2170                 loff_t pos, unsigned len, unsigned flags,
2171                 struct page **pagep, void **fsdata)
2172 {
2173         pgoff_t index = pos >> PAGE_SHIFT;
2174         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2175         struct page *page;
2176         loff_t fsize;
2177         int err = -ENOMEM;
2178
2179         WARN_ON(!fc->writeback_cache);
2180
2181         page = grab_cache_page_write_begin(mapping, index, flags);
2182         if (!page)
2183                 goto error;
2184
2185         fuse_wait_on_page_writeback(mapping->host, page->index);
2186
2187         if (PageUptodate(page) || len == PAGE_SIZE)
2188                 goto success;
2189         /*
2190          * Check if the start this page comes after the end of file, in which
2191          * case the readpage can be optimized away.
2192          */
2193         fsize = i_size_read(mapping->host);
2194         if (fsize <= (pos & PAGE_MASK)) {
2195                 size_t off = pos & ~PAGE_MASK;
2196                 if (off)
2197                         zero_user_segment(page, 0, off);
2198                 goto success;
2199         }
2200         err = fuse_do_readpage(file, page);
2201         if (err)
2202                 goto cleanup;
2203 success:
2204         *pagep = page;
2205         return 0;
2206
2207 cleanup:
2208         unlock_page(page);
2209         put_page(page);
2210 error:
2211         return err;
2212 }
2213
2214 static int fuse_write_end(struct file *file, struct address_space *mapping,
2215                 loff_t pos, unsigned len, unsigned copied,
2216                 struct page *page, void *fsdata)
2217 {
2218         struct inode *inode = page->mapping->host;
2219
2220         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2221         if (!copied)
2222                 goto unlock;
2223
2224         if (!PageUptodate(page)) {
2225                 /* Zero any unwritten bytes at the end of the page */
2226                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2227                 if (endoff)
2228                         zero_user_segment(page, endoff, PAGE_SIZE);
2229                 SetPageUptodate(page);
2230         }
2231
2232         fuse_write_update_size(inode, pos + copied);
2233         set_page_dirty(page);
2234
2235 unlock:
2236         unlock_page(page);
2237         put_page(page);
2238
2239         return copied;
2240 }
2241
2242 static int fuse_launder_page(struct page *page)
2243 {
2244         int err = 0;
2245         if (clear_page_dirty_for_io(page)) {
2246                 struct inode *inode = page->mapping->host;
2247                 err = fuse_writepage_locked(page);
2248                 if (!err)
2249                         fuse_wait_on_page_writeback(inode, page->index);
2250         }
2251         return err;
2252 }
2253
2254 /*
2255  * Write back dirty pages now, because there may not be any suitable
2256  * open files later
2257  */
2258 static void fuse_vma_close(struct vm_area_struct *vma)
2259 {
2260         filemap_write_and_wait(vma->vm_file->f_mapping);
2261 }
2262
2263 /*
2264  * Wait for writeback against this page to complete before allowing it
2265  * to be marked dirty again, and hence written back again, possibly
2266  * before the previous writepage completed.
2267  *
2268  * Block here, instead of in ->writepage(), so that the userspace fs
2269  * can only block processes actually operating on the filesystem.
2270  *
2271  * Otherwise unprivileged userspace fs would be able to block
2272  * unrelated:
2273  *
2274  * - page migration
2275  * - sync(2)
2276  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2277  */
2278 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2279 {
2280         struct page *page = vmf->page;
2281         struct inode *inode = file_inode(vmf->vma->vm_file);
2282
2283         file_update_time(vmf->vma->vm_file);
2284         lock_page(page);
2285         if (page->mapping != inode->i_mapping) {
2286                 unlock_page(page);
2287                 return VM_FAULT_NOPAGE;
2288         }
2289
2290         fuse_wait_on_page_writeback(inode, page->index);
2291         return VM_FAULT_LOCKED;
2292 }
2293
2294 static const struct vm_operations_struct fuse_file_vm_ops = {
2295         .close          = fuse_vma_close,
2296         .fault          = filemap_fault,
2297         .map_pages      = filemap_map_pages,
2298         .page_mkwrite   = fuse_page_mkwrite,
2299 };
2300
2301 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2302 {
2303         struct fuse_file *ff = file->private_data;
2304
2305         if (ff->open_flags & FOPEN_DIRECT_IO) {
2306                 /* Can't provide the coherency needed for MAP_SHARED */
2307                 if (vma->vm_flags & VM_MAYSHARE)
2308                         return -ENODEV;
2309
2310                 invalidate_inode_pages2(file->f_mapping);
2311
2312                 return generic_file_mmap(file, vma);
2313         }
2314
2315         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2316                 fuse_link_write_file(file);
2317
2318         file_accessed(file);
2319         vma->vm_ops = &fuse_file_vm_ops;
2320         return 0;
2321 }
2322
2323 static int convert_fuse_file_lock(struct fuse_conn *fc,
2324                                   const struct fuse_file_lock *ffl,
2325                                   struct file_lock *fl)
2326 {
2327         switch (ffl->type) {
2328         case F_UNLCK:
2329                 break;
2330
2331         case F_RDLCK:
2332         case F_WRLCK:
2333                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2334                     ffl->end < ffl->start)
2335                         return -EIO;
2336
2337                 fl->fl_start = ffl->start;
2338                 fl->fl_end = ffl->end;
2339
2340                 /*
2341                  * Convert pid into init's pid namespace.  The locks API will
2342                  * translate it into the caller's pid namespace.
2343                  */
2344                 rcu_read_lock();
2345                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2346                 rcu_read_unlock();
2347                 break;
2348
2349         default:
2350                 return -EIO;
2351         }
2352         fl->fl_type = ffl->type;
2353         return 0;
2354 }
2355
2356 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2357                          const struct file_lock *fl, int opcode, pid_t pid,
2358                          int flock, struct fuse_lk_in *inarg)
2359 {
2360         struct inode *inode = file_inode(file);
2361         struct fuse_conn *fc = get_fuse_conn(inode);
2362         struct fuse_file *ff = file->private_data;
2363
2364         memset(inarg, 0, sizeof(*inarg));
2365         inarg->fh = ff->fh;
2366         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2367         inarg->lk.start = fl->fl_start;
2368         inarg->lk.end = fl->fl_end;
2369         inarg->lk.type = fl->fl_type;
2370         inarg->lk.pid = pid;
2371         if (flock)
2372                 inarg->lk_flags |= FUSE_LK_FLOCK;
2373         args->opcode = opcode;
2374         args->nodeid = get_node_id(inode);
2375         args->in_numargs = 1;
2376         args->in_args[0].size = sizeof(*inarg);
2377         args->in_args[0].value = inarg;
2378 }
2379
2380 static int fuse_getlk(struct file *file, struct file_lock *fl)
2381 {
2382         struct inode *inode = file_inode(file);
2383         struct fuse_conn *fc = get_fuse_conn(inode);
2384         FUSE_ARGS(args);
2385         struct fuse_lk_in inarg;
2386         struct fuse_lk_out outarg;
2387         int err;
2388
2389         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2390         args.out_numargs = 1;
2391         args.out_args[0].size = sizeof(outarg);
2392         args.out_args[0].value = &outarg;
2393         err = fuse_simple_request(fc, &args);
2394         if (!err)
2395                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2396
2397         return err;
2398 }
2399
2400 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2401 {
2402         struct inode *inode = file_inode(file);
2403         struct fuse_conn *fc = get_fuse_conn(inode);
2404         FUSE_ARGS(args);
2405         struct fuse_lk_in inarg;
2406         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2407         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2408         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2409         int err;
2410
2411         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2412                 /* NLM needs asynchronous locks, which we don't support yet */
2413                 return -ENOLCK;
2414         }
2415
2416         /* Unlock on close is handled by the flush method */
2417         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2418                 return 0;
2419
2420         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2421         err = fuse_simple_request(fc, &args);
2422
2423         /* locking is restartable */
2424         if (err == -EINTR)
2425                 err = -ERESTARTSYS;
2426
2427         return err;
2428 }
2429
2430 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2431 {
2432         struct inode *inode = file_inode(file);
2433         struct fuse_conn *fc = get_fuse_conn(inode);
2434         int err;
2435
2436         if (cmd == F_CANCELLK) {
2437                 err = 0;
2438         } else if (cmd == F_GETLK) {
2439                 if (fc->no_lock) {
2440                         posix_test_lock(file, fl);
2441                         err = 0;
2442                 } else
2443                         err = fuse_getlk(file, fl);
2444         } else {
2445                 if (fc->no_lock)
2446                         err = posix_lock_file(file, fl, NULL);
2447                 else
2448                         err = fuse_setlk(file, fl, 0);
2449         }
2450         return err;
2451 }
2452
2453 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2454 {
2455         struct inode *inode = file_inode(file);
2456         struct fuse_conn *fc = get_fuse_conn(inode);
2457         int err;
2458
2459         if (fc->no_flock) {
2460                 err = locks_lock_file_wait(file, fl);
2461         } else {
2462                 struct fuse_file *ff = file->private_data;
2463
2464                 /* emulate flock with POSIX locks */
2465                 ff->flock = true;
2466                 err = fuse_setlk(file, fl, 1);
2467         }
2468
2469         return err;
2470 }
2471
2472 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2473 {
2474         struct inode *inode = mapping->host;
2475         struct fuse_conn *fc = get_fuse_conn(inode);
2476         FUSE_ARGS(args);
2477         struct fuse_bmap_in inarg;
2478         struct fuse_bmap_out outarg;
2479         int err;
2480
2481         if (!inode->i_sb->s_bdev || fc->no_bmap)
2482                 return 0;
2483
2484         memset(&inarg, 0, sizeof(inarg));
2485         inarg.block = block;
2486         inarg.blocksize = inode->i_sb->s_blocksize;
2487         args.opcode = FUSE_BMAP;
2488         args.nodeid = get_node_id(inode);
2489         args.in_numargs = 1;
2490         args.in_args[0].size = sizeof(inarg);
2491         args.in_args[0].value = &inarg;
2492         args.out_numargs = 1;
2493         args.out_args[0].size = sizeof(outarg);
2494         args.out_args[0].value = &outarg;
2495         err = fuse_simple_request(fc, &args);
2496         if (err == -ENOSYS)
2497                 fc->no_bmap = 1;
2498
2499         return err ? 0 : outarg.block;
2500 }
2501
2502 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2503 {
2504         struct inode *inode = file->f_mapping->host;
2505         struct fuse_conn *fc = get_fuse_conn(inode);
2506         struct fuse_file *ff = file->private_data;
2507         FUSE_ARGS(args);
2508         struct fuse_lseek_in inarg = {
2509                 .fh = ff->fh,
2510                 .offset = offset,
2511                 .whence = whence
2512         };
2513         struct fuse_lseek_out outarg;
2514         int err;
2515
2516         if (fc->no_lseek)
2517                 goto fallback;
2518
2519         args.opcode = FUSE_LSEEK;
2520         args.nodeid = ff->nodeid;
2521         args.in_numargs = 1;
2522         args.in_args[0].size = sizeof(inarg);
2523         args.in_args[0].value = &inarg;
2524         args.out_numargs = 1;
2525         args.out_args[0].size = sizeof(outarg);
2526         args.out_args[0].value = &outarg;
2527         err = fuse_simple_request(fc, &args);
2528         if (err) {
2529                 if (err == -ENOSYS) {
2530                         fc->no_lseek = 1;
2531                         goto fallback;
2532                 }
2533                 return err;
2534         }
2535
2536         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2537
2538 fallback:
2539         err = fuse_update_attributes(inode, file);
2540         if (!err)
2541                 return generic_file_llseek(file, offset, whence);
2542         else
2543                 return err;
2544 }
2545
2546 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2547 {
2548         loff_t retval;
2549         struct inode *inode = file_inode(file);
2550
2551         switch (whence) {
2552         case SEEK_SET:
2553         case SEEK_CUR:
2554                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2555                 retval = generic_file_llseek(file, offset, whence);
2556                 break;
2557         case SEEK_END:
2558                 inode_lock(inode);
2559                 retval = fuse_update_attributes(inode, file);
2560                 if (!retval)
2561                         retval = generic_file_llseek(file, offset, whence);
2562                 inode_unlock(inode);
2563                 break;
2564         case SEEK_HOLE:
2565         case SEEK_DATA:
2566                 inode_lock(inode);
2567                 retval = fuse_lseek(file, offset, whence);
2568                 inode_unlock(inode);
2569                 break;
2570         default:
2571                 retval = -EINVAL;
2572         }
2573
2574         return retval;
2575 }
2576
2577 /*
2578  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2579  * ABI was defined to be 'struct iovec' which is different on 32bit
2580  * and 64bit.  Fortunately we can determine which structure the server
2581  * used from the size of the reply.
2582  */
2583 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2584                                      size_t transferred, unsigned count,
2585                                      bool is_compat)
2586 {
2587 #ifdef CONFIG_COMPAT
2588         if (count * sizeof(struct compat_iovec) == transferred) {
2589                 struct compat_iovec *ciov = src;
2590                 unsigned i;
2591
2592                 /*
2593                  * With this interface a 32bit server cannot support
2594                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2595                  * requests
2596                  */
2597                 if (!is_compat)
2598                         return -EINVAL;
2599
2600                 for (i = 0; i < count; i++) {
2601                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2602                         dst[i].iov_len = ciov[i].iov_len;
2603                 }
2604                 return 0;
2605         }
2606 #endif
2607
2608         if (count * sizeof(struct iovec) != transferred)
2609                 return -EIO;
2610
2611         memcpy(dst, src, transferred);
2612         return 0;
2613 }
2614
2615 /* Make sure iov_length() won't overflow */
2616 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2617                                  size_t count)
2618 {
2619         size_t n;
2620         u32 max = fc->max_pages << PAGE_SHIFT;
2621
2622         for (n = 0; n < count; n++, iov++) {
2623                 if (iov->iov_len > (size_t) max)
2624                         return -ENOMEM;
2625                 max -= iov->iov_len;
2626         }
2627         return 0;
2628 }
2629
2630 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2631                                  void *src, size_t transferred, unsigned count,
2632                                  bool is_compat)
2633 {
2634         unsigned i;
2635         struct fuse_ioctl_iovec *fiov = src;
2636
2637         if (fc->minor < 16) {
2638                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2639                                                  count, is_compat);
2640         }
2641
2642         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2643                 return -EIO;
2644
2645         for (i = 0; i < count; i++) {
2646                 /* Did the server supply an inappropriate value? */
2647                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2648                     fiov[i].len != (unsigned long) fiov[i].len)
2649                         return -EIO;
2650
2651                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2652                 dst[i].iov_len = (size_t) fiov[i].len;
2653
2654 #ifdef CONFIG_COMPAT
2655                 if (is_compat &&
2656                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2657                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2658                         return -EIO;
2659 #endif
2660         }
2661
2662         return 0;
2663 }
2664
2665
2666 /*
2667  * For ioctls, there is no generic way to determine how much memory
2668  * needs to be read and/or written.  Furthermore, ioctls are allowed
2669  * to dereference the passed pointer, so the parameter requires deep
2670  * copying but FUSE has no idea whatsoever about what to copy in or
2671  * out.
2672  *
2673  * This is solved by allowing FUSE server to retry ioctl with
2674  * necessary in/out iovecs.  Let's assume the ioctl implementation
2675  * needs to read in the following structure.
2676  *
2677  * struct a {
2678  *      char    *buf;
2679  *      size_t  buflen;
2680  * }
2681  *
2682  * On the first callout to FUSE server, inarg->in_size and
2683  * inarg->out_size will be NULL; then, the server completes the ioctl
2684  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2685  * the actual iov array to
2686  *
2687  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2688  *
2689  * which tells FUSE to copy in the requested area and retry the ioctl.
2690  * On the second round, the server has access to the structure and
2691  * from that it can tell what to look for next, so on the invocation,
2692  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2693  *
2694  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2695  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2696  *
2697  * FUSE will copy both struct a and the pointed buffer from the
2698  * process doing the ioctl and retry ioctl with both struct a and the
2699  * buffer.
2700  *
2701  * This time, FUSE server has everything it needs and completes ioctl
2702  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2703  *
2704  * Copying data out works the same way.
2705  *
2706  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2707  * automatically initializes in and out iovs by decoding @cmd with
2708  * _IOC_* macros and the server is not allowed to request RETRY.  This
2709  * limits ioctl data transfers to well-formed ioctls and is the forced
2710  * behavior for all FUSE servers.
2711  */
2712 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2713                    unsigned int flags)
2714 {
2715         struct fuse_file *ff = file->private_data;
2716         struct fuse_conn *fc = ff->fc;
2717         struct fuse_ioctl_in inarg = {
2718                 .fh = ff->fh,
2719                 .cmd = cmd,
2720                 .arg = arg,
2721                 .flags = flags
2722         };
2723         struct fuse_ioctl_out outarg;
2724         struct iovec *iov_page = NULL;
2725         struct iovec *in_iov = NULL, *out_iov = NULL;
2726         unsigned int in_iovs = 0, out_iovs = 0, max_pages;
2727         size_t in_size, out_size, c;
2728         ssize_t transferred;
2729         int err, i;
2730         struct iov_iter ii;
2731         struct fuse_args_pages ap = {};
2732
2733 #if BITS_PER_LONG == 32
2734         inarg.flags |= FUSE_IOCTL_32BIT;
2735 #else
2736         if (flags & FUSE_IOCTL_COMPAT) {
2737                 inarg.flags |= FUSE_IOCTL_32BIT;
2738 #ifdef CONFIG_X86_X32
2739                 if (in_x32_syscall())
2740                         inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2741 #endif
2742         }
2743 #endif
2744
2745         /* assume all the iovs returned by client always fits in a page */
2746         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2747
2748         err = -ENOMEM;
2749         ap.pages = fuse_pages_alloc(fc->max_pages, GFP_KERNEL, &ap.descs);
2750         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2751         if (!ap.pages || !iov_page)
2752                 goto out;
2753
2754         fuse_page_descs_length_init(ap.descs, 0, fc->max_pages);
2755
2756         /*
2757          * If restricted, initialize IO parameters as encoded in @cmd.
2758          * RETRY from server is not allowed.
2759          */
2760         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2761                 struct iovec *iov = iov_page;
2762
2763                 iov->iov_base = (void __user *)arg;
2764                 iov->iov_len = _IOC_SIZE(cmd);
2765
2766                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2767                         in_iov = iov;
2768                         in_iovs = 1;
2769                 }
2770
2771                 if (_IOC_DIR(cmd) & _IOC_READ) {
2772                         out_iov = iov;
2773                         out_iovs = 1;
2774                 }
2775         }
2776
2777  retry:
2778         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2779         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2780
2781         /*
2782          * Out data can be used either for actual out data or iovs,
2783          * make sure there always is at least one page.
2784          */
2785         out_size = max_t(size_t, out_size, PAGE_SIZE);
2786         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2787
2788         /* make sure there are enough buffer pages and init request with them */
2789         err = -ENOMEM;
2790         if (max_pages > fc->max_pages)
2791                 goto out;
2792         while (ap.num_pages < max_pages) {
2793                 ap.pages[ap.num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2794                 if (!ap.pages[ap.num_pages])
2795                         goto out;
2796                 ap.num_pages++;
2797         }
2798
2799
2800         /* okay, let's send it to the client */
2801         ap.args.opcode = FUSE_IOCTL;
2802         ap.args.nodeid = ff->nodeid;
2803         ap.args.in_numargs = 1;
2804         ap.args.in_args[0].size = sizeof(inarg);
2805         ap.args.in_args[0].value = &inarg;
2806         if (in_size) {
2807                 ap.args.in_numargs++;
2808                 ap.args.in_args[1].size = in_size;
2809                 ap.args.in_pages = true;
2810
2811                 err = -EFAULT;
2812                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2813                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2814                         c = copy_page_from_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2815                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2816                                 goto out;
2817                 }
2818         }
2819
2820         ap.args.out_numargs = 2;
2821         ap.args.out_args[0].size = sizeof(outarg);
2822         ap.args.out_args[0].value = &outarg;
2823         ap.args.out_args[1].size = out_size;
2824         ap.args.out_pages = true;
2825         ap.args.out_argvar = true;
2826
2827         transferred = fuse_simple_request(fc, &ap.args);
2828         err = transferred;
2829         if (transferred < 0)
2830                 goto out;
2831
2832         /* did it ask for retry? */
2833         if (outarg.flags & FUSE_IOCTL_RETRY) {
2834                 void *vaddr;
2835
2836                 /* no retry if in restricted mode */
2837                 err = -EIO;
2838                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2839                         goto out;
2840
2841                 in_iovs = outarg.in_iovs;
2842                 out_iovs = outarg.out_iovs;
2843
2844                 /*
2845                  * Make sure things are in boundary, separate checks
2846                  * are to protect against overflow.
2847                  */
2848                 err = -ENOMEM;
2849                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2850                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2851                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2852                         goto out;
2853
2854                 vaddr = kmap_atomic(ap.pages[0]);
2855                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2856                                             transferred, in_iovs + out_iovs,
2857                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2858                 kunmap_atomic(vaddr);
2859                 if (err)
2860                         goto out;
2861
2862                 in_iov = iov_page;
2863                 out_iov = in_iov + in_iovs;
2864
2865                 err = fuse_verify_ioctl_iov(fc, in_iov, in_iovs);
2866                 if (err)
2867                         goto out;
2868
2869                 err = fuse_verify_ioctl_iov(fc, out_iov, out_iovs);
2870                 if (err)
2871                         goto out;
2872
2873                 goto retry;
2874         }
2875
2876         err = -EIO;
2877         if (transferred > inarg.out_size)
2878                 goto out;
2879
2880         err = -EFAULT;
2881         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2882         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2883                 c = copy_page_to_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2884                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2885                         goto out;
2886         }
2887         err = 0;
2888  out:
2889         free_page((unsigned long) iov_page);
2890         while (ap.num_pages)
2891                 __free_page(ap.pages[--ap.num_pages]);
2892         kfree(ap.pages);
2893
2894         return err ? err : outarg.result;
2895 }
2896 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2897
2898 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2899                        unsigned long arg, unsigned int flags)
2900 {
2901         struct inode *inode = file_inode(file);
2902         struct fuse_conn *fc = get_fuse_conn(inode);
2903
2904         if (!fuse_allow_current_process(fc))
2905                 return -EACCES;
2906
2907         if (is_bad_inode(inode))
2908                 return -EIO;
2909
2910         return fuse_do_ioctl(file, cmd, arg, flags);
2911 }
2912
2913 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2914                             unsigned long arg)
2915 {
2916         return fuse_ioctl_common(file, cmd, arg, 0);
2917 }
2918
2919 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2920                                    unsigned long arg)
2921 {
2922         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2923 }
2924
2925 /*
2926  * All files which have been polled are linked to RB tree
2927  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2928  * find the matching one.
2929  */
2930 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2931                                               struct rb_node **parent_out)
2932 {
2933         struct rb_node **link = &fc->polled_files.rb_node;
2934         struct rb_node *last = NULL;
2935
2936         while (*link) {
2937                 struct fuse_file *ff;
2938
2939                 last = *link;
2940                 ff = rb_entry(last, struct fuse_file, polled_node);
2941
2942                 if (kh < ff->kh)
2943                         link = &last->rb_left;
2944                 else if (kh > ff->kh)
2945                         link = &last->rb_right;
2946                 else
2947                         return link;
2948         }
2949
2950         if (parent_out)
2951                 *parent_out = last;
2952         return link;
2953 }
2954
2955 /*
2956  * The file is about to be polled.  Make sure it's on the polled_files
2957  * RB tree.  Note that files once added to the polled_files tree are
2958  * not removed before the file is released.  This is because a file
2959  * polled once is likely to be polled again.
2960  */
2961 static void fuse_register_polled_file(struct fuse_conn *fc,
2962                                       struct fuse_file *ff)
2963 {
2964         spin_lock(&fc->lock);
2965         if (RB_EMPTY_NODE(&ff->polled_node)) {
2966                 struct rb_node **link, *parent;
2967
2968                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2969                 BUG_ON(*link);
2970                 rb_link_node(&ff->polled_node, parent, link);
2971                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2972         }
2973         spin_unlock(&fc->lock);
2974 }
2975
2976 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2977 {
2978         struct fuse_file *ff = file->private_data;
2979         struct fuse_conn *fc = ff->fc;
2980         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2981         struct fuse_poll_out outarg;
2982         FUSE_ARGS(args);
2983         int err;
2984
2985         if (fc->no_poll)
2986                 return DEFAULT_POLLMASK;
2987
2988         poll_wait(file, &ff->poll_wait, wait);
2989         inarg.events = mangle_poll(poll_requested_events(wait));
2990
2991         /*
2992          * Ask for notification iff there's someone waiting for it.
2993          * The client may ignore the flag and always notify.
2994          */
2995         if (waitqueue_active(&ff->poll_wait)) {
2996                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2997                 fuse_register_polled_file(fc, ff);
2998         }
2999
3000         args.opcode = FUSE_POLL;
3001         args.nodeid = ff->nodeid;
3002         args.in_numargs = 1;
3003         args.in_args[0].size = sizeof(inarg);
3004         args.in_args[0].value = &inarg;
3005         args.out_numargs = 1;
3006         args.out_args[0].size = sizeof(outarg);
3007         args.out_args[0].value = &outarg;
3008         err = fuse_simple_request(fc, &args);
3009
3010         if (!err)
3011                 return demangle_poll(outarg.revents);
3012         if (err == -ENOSYS) {
3013                 fc->no_poll = 1;
3014                 return DEFAULT_POLLMASK;
3015         }
3016         return EPOLLERR;
3017 }
3018 EXPORT_SYMBOL_GPL(fuse_file_poll);
3019
3020 /*
3021  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
3022  * wakes up the poll waiters.
3023  */
3024 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
3025                             struct fuse_notify_poll_wakeup_out *outarg)
3026 {
3027         u64 kh = outarg->kh;
3028         struct rb_node **link;
3029
3030         spin_lock(&fc->lock);
3031
3032         link = fuse_find_polled_node(fc, kh, NULL);
3033         if (*link) {
3034                 struct fuse_file *ff;
3035
3036                 ff = rb_entry(*link, struct fuse_file, polled_node);
3037                 wake_up_interruptible_sync(&ff->poll_wait);
3038         }
3039
3040         spin_unlock(&fc->lock);
3041         return 0;
3042 }
3043
3044 static void fuse_do_truncate(struct file *file)
3045 {
3046         struct inode *inode = file->f_mapping->host;
3047         struct iattr attr;
3048
3049         attr.ia_valid = ATTR_SIZE;
3050         attr.ia_size = i_size_read(inode);
3051
3052         attr.ia_file = file;
3053         attr.ia_valid |= ATTR_FILE;
3054
3055         fuse_do_setattr(file_dentry(file), &attr, file);
3056 }
3057
3058 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
3059 {
3060         return round_up(off, fc->max_pages << PAGE_SHIFT);
3061 }
3062
3063 static ssize_t
3064 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3065 {
3066         DECLARE_COMPLETION_ONSTACK(wait);
3067         ssize_t ret = 0;
3068         struct file *file = iocb->ki_filp;
3069         struct fuse_file *ff = file->private_data;
3070         bool async_dio = ff->fc->async_dio;
3071         loff_t pos = 0;
3072         struct inode *inode;
3073         loff_t i_size;
3074         size_t count = iov_iter_count(iter);
3075         loff_t offset = iocb->ki_pos;
3076         struct fuse_io_priv *io;
3077
3078         pos = offset;
3079         inode = file->f_mapping->host;
3080         i_size = i_size_read(inode);
3081
3082         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
3083                 return 0;
3084
3085         /* optimization for short read */
3086         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
3087                 if (offset >= i_size)
3088                         return 0;
3089                 iov_iter_truncate(iter, fuse_round_up(ff->fc, i_size - offset));
3090                 count = iov_iter_count(iter);
3091         }
3092
3093         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
3094         if (!io)
3095                 return -ENOMEM;
3096         spin_lock_init(&io->lock);
3097         kref_init(&io->refcnt);
3098         io->reqs = 1;
3099         io->bytes = -1;
3100         io->size = 0;
3101         io->offset = offset;
3102         io->write = (iov_iter_rw(iter) == WRITE);
3103         io->err = 0;
3104         /*
3105          * By default, we want to optimize all I/Os with async request
3106          * submission to the client filesystem if supported.
3107          */
3108         io->async = async_dio;
3109         io->iocb = iocb;
3110         io->blocking = is_sync_kiocb(iocb);
3111
3112         /*
3113          * We cannot asynchronously extend the size of a file.
3114          * In such case the aio will behave exactly like sync io.
3115          */
3116         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
3117                 io->blocking = true;
3118
3119         if (io->async && io->blocking) {
3120                 /*
3121                  * Additional reference to keep io around after
3122                  * calling fuse_aio_complete()
3123                  */
3124                 kref_get(&io->refcnt);
3125                 io->done = &wait;
3126         }
3127
3128         if (iov_iter_rw(iter) == WRITE) {
3129                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
3130                 fuse_invalidate_attr(inode);
3131         } else {
3132                 ret = __fuse_direct_read(io, iter, &pos);
3133         }
3134
3135         if (io->async) {
3136                 bool blocking = io->blocking;
3137
3138                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3139
3140                 /* we have a non-extending, async request, so return */
3141                 if (!blocking)
3142                         return -EIOCBQUEUED;
3143
3144                 wait_for_completion(&wait);
3145                 ret = fuse_get_res_by_io(io);
3146         }
3147
3148         kref_put(&io->refcnt, fuse_io_release);
3149
3150         if (iov_iter_rw(iter) == WRITE) {
3151                 if (ret > 0)
3152                         fuse_write_update_size(inode, pos);
3153                 else if (ret < 0 && offset + count > i_size)
3154                         fuse_do_truncate(file);
3155         }
3156
3157         return ret;
3158 }
3159
3160 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3161 {
3162         int err = filemap_write_and_wait_range(inode->i_mapping, start, end);
3163
3164         if (!err)
3165                 fuse_sync_writes(inode);
3166
3167         return err;
3168 }
3169
3170 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3171                                 loff_t length)
3172 {
3173         struct fuse_file *ff = file->private_data;
3174         struct inode *inode = file_inode(file);
3175         struct fuse_inode *fi = get_fuse_inode(inode);
3176         struct fuse_conn *fc = ff->fc;
3177         FUSE_ARGS(args);
3178         struct fuse_fallocate_in inarg = {
3179                 .fh = ff->fh,
3180                 .offset = offset,
3181                 .length = length,
3182                 .mode = mode
3183         };
3184         int err;
3185         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
3186                            (mode & FALLOC_FL_PUNCH_HOLE);
3187
3188         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3189                 return -EOPNOTSUPP;
3190
3191         if (fc->no_fallocate)
3192                 return -EOPNOTSUPP;
3193
3194         if (lock_inode) {
3195                 inode_lock(inode);
3196                 if (mode & FALLOC_FL_PUNCH_HOLE) {
3197                         loff_t endbyte = offset + length - 1;
3198
3199                         err = fuse_writeback_range(inode, offset, endbyte);
3200                         if (err)
3201                                 goto out;
3202                 }
3203         }
3204
3205         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3206             offset + length > i_size_read(inode)) {
3207                 err = inode_newsize_ok(inode, offset + length);
3208                 if (err)
3209                         goto out;
3210         }
3211
3212         if (!(mode & FALLOC_FL_KEEP_SIZE))
3213                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3214
3215         args.opcode = FUSE_FALLOCATE;
3216         args.nodeid = ff->nodeid;
3217         args.in_numargs = 1;
3218         args.in_args[0].size = sizeof(inarg);
3219         args.in_args[0].value = &inarg;
3220         err = fuse_simple_request(fc, &args);
3221         if (err == -ENOSYS) {
3222                 fc->no_fallocate = 1;
3223                 err = -EOPNOTSUPP;
3224         }
3225         if (err)
3226                 goto out;
3227
3228         /* we could have extended the file */
3229         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3230                 bool changed = fuse_write_update_size(inode, offset + length);
3231
3232                 if (changed && fc->writeback_cache)
3233                         file_update_time(file);
3234         }
3235
3236         if (mode & FALLOC_FL_PUNCH_HOLE)
3237                 truncate_pagecache_range(inode, offset, offset + length - 1);
3238
3239         fuse_invalidate_attr(inode);
3240
3241 out:
3242         if (!(mode & FALLOC_FL_KEEP_SIZE))
3243                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3244
3245         if (lock_inode)
3246                 inode_unlock(inode);
3247
3248         return err;
3249 }
3250
3251 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3252                                       struct file *file_out, loff_t pos_out,
3253                                       size_t len, unsigned int flags)
3254 {
3255         struct fuse_file *ff_in = file_in->private_data;
3256         struct fuse_file *ff_out = file_out->private_data;
3257         struct inode *inode_in = file_inode(file_in);
3258         struct inode *inode_out = file_inode(file_out);
3259         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3260         struct fuse_conn *fc = ff_in->fc;
3261         FUSE_ARGS(args);
3262         struct fuse_copy_file_range_in inarg = {
3263                 .fh_in = ff_in->fh,
3264                 .off_in = pos_in,
3265                 .nodeid_out = ff_out->nodeid,
3266                 .fh_out = ff_out->fh,
3267                 .off_out = pos_out,
3268                 .len = len,
3269                 .flags = flags
3270         };
3271         struct fuse_write_out outarg;
3272         ssize_t err;
3273         /* mark unstable when write-back is not used, and file_out gets
3274          * extended */
3275         bool is_unstable = (!fc->writeback_cache) &&
3276                            ((pos_out + len) > inode_out->i_size);
3277
3278         if (fc->no_copy_file_range)
3279                 return -EOPNOTSUPP;
3280
3281         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3282                 return -EXDEV;
3283
3284         inode_lock(inode_in);
3285         err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3286         inode_unlock(inode_in);
3287         if (err)
3288                 return err;
3289
3290         inode_lock(inode_out);
3291
3292         err = file_modified(file_out);
3293         if (err)
3294                 goto out;
3295
3296         /*
3297          * Write out dirty pages in the destination file before sending the COPY
3298          * request to userspace.  After the request is completed, truncate off
3299          * pages (including partial ones) from the cache that have been copied,
3300          * since these contain stale data at that point.
3301          *
3302          * This should be mostly correct, but if the COPY writes to partial
3303          * pages (at the start or end) and the parts not covered by the COPY are
3304          * written through a memory map after calling fuse_writeback_range(),
3305          * then these partial page modifications will be lost on truncation.
3306          *
3307          * It is unlikely that someone would rely on such mixed style
3308          * modifications.  Yet this does give less guarantees than if the
3309          * copying was performed with write(2).
3310          *
3311          * To fix this a i_mmap_sem style lock could be used to prevent new
3312          * faults while the copy is ongoing.
3313          */
3314         err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3315         if (err)
3316                 goto out;
3317
3318         if (is_unstable)
3319                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3320
3321         args.opcode = FUSE_COPY_FILE_RANGE;
3322         args.nodeid = ff_in->nodeid;
3323         args.in_numargs = 1;
3324         args.in_args[0].size = sizeof(inarg);
3325         args.in_args[0].value = &inarg;
3326         args.out_numargs = 1;
3327         args.out_args[0].size = sizeof(outarg);
3328         args.out_args[0].value = &outarg;
3329         err = fuse_simple_request(fc, &args);
3330         if (err == -ENOSYS) {
3331                 fc->no_copy_file_range = 1;
3332                 err = -EOPNOTSUPP;
3333         }
3334         if (err)
3335                 goto out;
3336
3337         truncate_inode_pages_range(inode_out->i_mapping,
3338                                    ALIGN_DOWN(pos_out, PAGE_SIZE),
3339                                    ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3340
3341         if (fc->writeback_cache) {
3342                 fuse_write_update_size(inode_out, pos_out + outarg.size);
3343                 file_update_time(file_out);
3344         }
3345
3346         fuse_invalidate_attr(inode_out);
3347
3348         err = outarg.size;
3349 out:
3350         if (is_unstable)
3351                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3352
3353         inode_unlock(inode_out);
3354         file_accessed(file_in);
3355
3356         return err;
3357 }
3358
3359 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3360                                     struct file *dst_file, loff_t dst_off,
3361                                     size_t len, unsigned int flags)
3362 {
3363         ssize_t ret;
3364
3365         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3366                                      len, flags);
3367
3368         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3369                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3370                                               dst_off, len, flags);
3371         return ret;
3372 }
3373
3374 static const struct file_operations fuse_file_operations = {
3375         .llseek         = fuse_file_llseek,
3376         .read_iter      = fuse_file_read_iter,
3377         .write_iter     = fuse_file_write_iter,
3378         .mmap           = fuse_file_mmap,
3379         .open           = fuse_open,
3380         .flush          = fuse_flush,
3381         .release        = fuse_release,
3382         .fsync          = fuse_fsync,
3383         .lock           = fuse_file_lock,
3384         .flock          = fuse_file_flock,
3385         .splice_read    = generic_file_splice_read,
3386         .splice_write   = iter_file_splice_write,
3387         .unlocked_ioctl = fuse_file_ioctl,
3388         .compat_ioctl   = fuse_file_compat_ioctl,
3389         .poll           = fuse_file_poll,
3390         .fallocate      = fuse_file_fallocate,
3391         .copy_file_range = fuse_copy_file_range,
3392 };
3393
3394 static const struct address_space_operations fuse_file_aops  = {
3395         .readpage       = fuse_readpage,
3396         .readahead      = fuse_readahead,
3397         .writepage      = fuse_writepage,
3398         .writepages     = fuse_writepages,
3399         .launder_page   = fuse_launder_page,
3400         .set_page_dirty = __set_page_dirty_nobuffers,
3401         .bmap           = fuse_bmap,
3402         .direct_IO      = fuse_direct_IO,
3403         .write_begin    = fuse_write_begin,
3404         .write_end      = fuse_write_end,
3405 };
3406
3407 void fuse_init_file_inode(struct inode *inode)
3408 {
3409         struct fuse_inode *fi = get_fuse_inode(inode);
3410
3411         inode->i_fop = &fuse_file_operations;
3412         inode->i_data.a_ops = &fuse_file_aops;
3413
3414         INIT_LIST_HEAD(&fi->write_files);
3415         INIT_LIST_HEAD(&fi->queued_writes);
3416         fi->writectr = 0;
3417         init_waitqueue_head(&fi->page_waitq);
3418         fi->writepages = RB_ROOT;
3419 }