fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/sched/signal.h>
  16 #include <linux/module.h>
  17 #include <linux/compat.h>
  18 #include <linux/swap.h>
  19 #include <linux/falloc.h>
  20 #include <linux/uio.h>
  21 #include <linux/fs.h>
  22
  23 static struct page **fuse_pages_alloc(unsigned int npages, gfp_t flags,
  24                                       struct fuse_page_desc **desc)
  25 {
  26         struct page **pages;
  27
  28         pages = kzalloc(npages * (sizeof(struct page *) +
  29                                   sizeof(struct fuse_page_desc)), flags);
  30         *desc = (void *) (pages + npages);
  31
  32         return pages;
  33 }
  34
  35 static int fuse_send_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
  36                           int opcode, struct fuse_open_out *outargp)
  37 {
  38         struct fuse_open_in inarg;
  39         FUSE_ARGS(args);
  40
  41         memset(&inarg, 0, sizeof(inarg));
  42         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  43         if (!fm->fc->atomic_o_trunc)
  44                 inarg.flags &= ~O_TRUNC;
  45         args.opcode = opcode;
  46         args.nodeid = nodeid;
  47         args.in_numargs = 1;
  48         args.in_args[0].size = sizeof(inarg);
  49         args.in_args[0].value = &inarg;
  50         args.out_numargs = 1;
  51         args.out_args[0].size = sizeof(*outargp);
  52         args.out_args[0].value = outargp;
  53
  54         return fuse_simple_request(fm, &args);
  55 }
  56
  57 struct fuse_release_args {
  58         struct fuse_args args;
  59         struct fuse_release_in inarg;
  60         struct inode *inode;
  61 };
  62
  63 struct fuse_file *fuse_file_alloc(struct fuse_mount *fm)
  64 {
  65         struct fuse_file *ff;
  66
  67         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
  68         if (unlikely(!ff))
  69                 return NULL;
  70
  71         ff->fm = fm;
  72         ff->release_args = kzalloc(sizeof(*ff->release_args),
  73                                    GFP_KERNEL_ACCOUNT);
  74         if (!ff->release_args) {
  75                 kfree(ff);
  76                 return NULL;
  77         }
  78
  79         INIT_LIST_HEAD(&ff->write_entry);
  80         mutex_init(&ff->readdir.lock);
  81         refcount_set(&ff->count, 1);
  82         RB_CLEAR_NODE(&ff->polled_node);
  83         init_waitqueue_head(&ff->poll_wait);
  84
  85         ff->kh = atomic64_inc_return(&fm->fc->khctr);
  86
  87         return ff;
  88 }
  89
  90 void fuse_file_free(struct fuse_file *ff)
  91 {
  92         kfree(ff->release_args);
  93         mutex_destroy(&ff->readdir.lock);
  94         kfree(ff);
  95 }
  96
  97 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  98 {
  99         refcount_inc(&ff->count);
 100         return ff;
 101 }
 102
 103 static void fuse_release_end(struct fuse_mount *fm, struct fuse_args *args,
 104                              int error)
 105 {
 106         struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
 107
 108         iput(ra->inode);
 109         kfree(ra);
 110 }
 111
 112 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
 113 {
 114         if (refcount_dec_and_test(&ff->count)) {
 115                 struct fuse_args *args = &ff->release_args->args;
 116
 117                 if (isdir ? ff->fm->fc->no_opendir : ff->fm->fc->no_open) {
 118                         /* Do nothing when client does not implement 'open' */
 119                         fuse_release_end(ff->fm, args, 0);
 120                 } else if (sync) {
 121                         fuse_simple_request(ff->fm, args);
 122                         fuse_release_end(ff->fm, args, 0);
 123                 } else {
 124                         args->end = fuse_release_end;
 125                         if (fuse_simple_background(ff->fm, args,
 126                                                    GFP_KERNEL | __GFP_NOFAIL))
 127                                 fuse_release_end(ff->fm, args, -ENOTCONN);
 128                 }
 129                 kfree(ff);
 130         }
 131 }
 132
 133 int fuse_do_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
 134                  bool isdir)
 135 {
 136         struct fuse_conn *fc = fm->fc;
 137         struct fuse_file *ff;
 138         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 139
 140         ff = fuse_file_alloc(fm);
 141         if (!ff)
 142                 return -ENOMEM;
 143
 144         ff->fh = 0;
 145         /* Default for no-open */
 146         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
 147         if (isdir ? !fc->no_opendir : !fc->no_open) {
 148                 struct fuse_open_out outarg;
 149                 int err;
 150
 151                 err = fuse_send_open(fm, nodeid, file, opcode, &outarg);
 152                 if (!err) {
 153                         ff->fh = outarg.fh;
 154                         ff->open_flags = outarg.open_flags;
 155
 156                 } else if (err != -ENOSYS) {
 157                         fuse_file_free(ff);
 158                         return err;
 159                 } else {
 160                         if (isdir)
 161                                 fc->no_opendir = 1;
 162                         else
 163                                 fc->no_open = 1;
 164                 }
 165         }
 166
 167         if (isdir)
 168                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 169
 170         ff->nodeid = nodeid;
 171         file->private_data = ff;
 172
 173         return 0;
 174 }
 175 EXPORT_SYMBOL_GPL(fuse_do_open);
 176
 177 static void fuse_link_write_file(struct file *file)
 178 {
 179         struct inode *inode = file_inode(file);
 180         struct fuse_inode *fi = get_fuse_inode(inode);
 181         struct fuse_file *ff = file->private_data;
 182         /*
 183          * file may be written through mmap, so chain it onto the
 184          * inodes's write_file list
 185          */
 186         spin_lock(&fi->lock);
 187         if (list_empty(&ff->write_entry))
 188                 list_add(&ff->write_entry, &fi->write_files);
 189         spin_unlock(&fi->lock);
 190 }
 191
 192 void fuse_finish_open(struct inode *inode, struct file *file)
 193 {
 194         struct fuse_file *ff = file->private_data;
 195         struct fuse_conn *fc = get_fuse_conn(inode);
 196
 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
 202         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 203                 struct fuse_inode *fi = get_fuse_inode(inode);
 204
 205                 spin_lock(&fi->lock);
 206                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 207                 i_size_write(inode, 0);
 208                 spin_unlock(&fi->lock);
 209                 fuse_invalidate_attr(inode);
 210                 if (fc->writeback_cache)
 211                         file_update_time(file);
 212         }
 213         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 214                 fuse_link_write_file(file);
 215 }
 216
 217 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 218 {
 219         struct fuse_mount *fm = get_fuse_mount(inode);
 220         struct fuse_conn *fc = fm->fc;
 221         int err;
 222         bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
 223                           fc->atomic_o_trunc &&
 224                           fc->writeback_cache;
 225         bool dax_truncate = (file->f_flags & O_TRUNC) &&
 226                           fc->atomic_o_trunc && FUSE_IS_DAX(inode);
 227
 228         if (fuse_is_bad(inode))
 229                 return -EIO;
 230
 231         err = generic_file_open(inode, file);
 232         if (err)
 233                 return err;
 234
 235         if (is_wb_truncate || dax_truncate)
 236                 inode_lock(inode);
 237
 238         if (dax_truncate) {
 239                 down_write(&get_fuse_inode(inode)->i_mmap_sem);
 240                 err = fuse_dax_break_layouts(inode, 0, 0);
 241                 if (err)
 242                         goto out_inode_unlock;
 243         }
 244
 245         if (is_wb_truncate || dax_truncate)
 246                 fuse_set_nowrite(inode);
 247
 248         err = fuse_do_open(fm, get_node_id(inode), file, isdir);
 249         if (!err)
 250                 fuse_finish_open(inode, file);
 251
 252         if (is_wb_truncate || dax_truncate)
 253                 fuse_release_nowrite(inode);
 254         if (!err) {
 255                 struct fuse_file *ff = file->private_data;
 256
 257                 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC))
 258                         truncate_pagecache(inode, 0);
 259                 else if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 260                         invalidate_inode_pages2(inode->i_mapping);
 261         }
 262         if (dax_truncate)
 263                 up_write(&get_fuse_inode(inode)->i_mmap_sem);
 264
 265 out_inode_unlock:
 266         if (is_wb_truncate || dax_truncate)
 267                 inode_unlock(inode);
 268
 269         return err;
 270 }
 271
 272 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 273                                  int flags, int opcode)
 274 {
 275         struct fuse_conn *fc = ff->fm->fc;
 276         struct fuse_release_args *ra = ff->release_args;
 277
 278         /* Inode is NULL on error path of fuse_create_open() */
 279         if (likely(fi)) {
 280                 spin_lock(&fi->lock);
 281                 list_del(&ff->write_entry);
 282                 spin_unlock(&fi->lock);
 283         }
 284         spin_lock(&fc->lock);
 285         if (!RB_EMPTY_NODE(&ff->polled_node))
 286                 rb_erase(&ff->polled_node, &fc->polled_files);
 287         spin_unlock(&fc->lock);
 288
 289         wake_up_interruptible_all(&ff->poll_wait);
 290
 291         ra->inarg.fh = ff->fh;
 292         ra->inarg.flags = flags;
 293         ra->args.in_numargs = 1;
 294         ra->args.in_args[0].size = sizeof(struct fuse_release_in);
 295         ra->args.in_args[0].value = &ra->inarg;
 296         ra->args.opcode = opcode;
 297         ra->args.nodeid = ff->nodeid;
 298         ra->args.force = true;
 299         ra->args.nocreds = true;
 300 }
 301
 302 void fuse_release_common(struct file *file, bool isdir)
 303 {
 304         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 305         struct fuse_file *ff = file->private_data;
 306         struct fuse_release_args *ra = ff->release_args;
 307         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 308
 309         fuse_prepare_release(fi, ff, file->f_flags, opcode);
 310
 311         if (ff->flock) {
 312                 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 313                 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fm->fc,
 314                                                           (fl_owner_t) file);
 315         }
 316         /* Hold inode until release is finished */
 317         ra->inode = igrab(file_inode(file));
 318
 319         /*
 320          * Normally this will send the RELEASE request, however if
 321          * some asynchronous READ or WRITE requests are outstanding,
 322          * the sending will be delayed.
 323          *
 324          * Make the release synchronous if this is a fuseblk mount,
 325          * synchronous RELEASE is allowed (and desirable) in this case
 326          * because the server can be trusted not to screw up.
 327          */
 328         fuse_file_put(ff, ff->fm->fc->destroy, isdir);
 329 }
 330
 331 static int fuse_open(struct inode *inode, struct file *file)
 332 {
 333         return fuse_open_common(inode, file, false);
 334 }
 335
 336 static int fuse_release(struct inode *inode, struct file *file)
 337 {
 338         struct fuse_conn *fc = get_fuse_conn(inode);
 339
 340         /* see fuse_vma_close() for !writeback_cache case */
 341         if (fc->writeback_cache)
 342                 write_inode_now(inode, 1);
 343
 344         fuse_release_common(file, false);
 345
 346         /* return value is ignored by VFS */
 347         return 0;
 348 }
 349
 350 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
 351 {
 352         WARN_ON(refcount_read(&ff->count) > 1);
 353         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 354         /*
 355          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 356          * synchronous, we are fine with not doing igrab() here"
 357          */
 358         fuse_file_put(ff, true, false);
 359 }
 360 EXPORT_SYMBOL_GPL(fuse_sync_release);
 361
 362 /*
 363  * Scramble the ID space with XTEA, so that the value of the files_struct
 364  * pointer is not exposed to userspace.
 365  */
 366 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 367 {
 368         u32 *k = fc->scramble_key;
 369         u64 v = (unsigned long) id;
 370         u32 v0 = v;
 371         u32 v1 = v >> 32;
 372         u32 sum = 0;
 373         int i;
 374
 375         for (i = 0; i < 32; i++) {
 376                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 377                 sum += 0x9E3779B9;
 378                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 379         }
 380
 381         return (u64) v0 + ((u64) v1 << 32);
 382 }
 383
 384 struct fuse_writepage_args {
 385         struct fuse_io_args ia;
 386         struct rb_node writepages_entry;
 387         struct list_head queue_entry;
 388         struct fuse_writepage_args *next;
 389         struct inode *inode;
 390 };
 391
 392 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
 393                                             pgoff_t idx_from, pgoff_t idx_to)
 394 {
 395         struct rb_node *n;
 396
 397         n = fi->writepages.rb_node;
 398
 399         while (n) {
 400                 struct fuse_writepage_args *wpa;
 401                 pgoff_t curr_index;
 402
 403                 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
 404                 WARN_ON(get_fuse_inode(wpa->inode) != fi);
 405                 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
 406                 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
 407                         n = n->rb_right;
 408                 else if (idx_to < curr_index)
 409                         n = n->rb_left;
 410                 else
 411                         return wpa;
 412         }
 413         return NULL;
 414 }
 415
 416 /*
 417  * Check if any page in a range is under writeback
 418  *
 419  * This is currently done by walking the list of writepage requests
 420  * for the inode, which can be pretty inefficient.
 421  */
 422 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 423                                    pgoff_t idx_to)
 424 {
 425         struct fuse_inode *fi = get_fuse_inode(inode);
 426         bool found;
 427
 428         spin_lock(&fi->lock);
 429         found = fuse_find_writeback(fi, idx_from, idx_to);
 430         spin_unlock(&fi->lock);
 431
 432         return found;
 433 }
 434
 435 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 436 {
 437         return fuse_range_is_writeback(inode, index, index);
 438 }
 439
 440 /*
 441  * Wait for page writeback to be completed.
 442  *
 443  * Since fuse doesn't rely on the VM writeback tracking, this has to
 444  * use some other means.
 445  */
 446 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 447 {
 448         struct fuse_inode *fi = get_fuse_inode(inode);
 449
 450         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 451 }
 452
 453 /*
 454  * Wait for all pending writepages on the inode to finish.
 455  *
 456  * This is currently done by blocking further writes with FUSE_NOWRITE
 457  * and waiting for all sent writes to complete.
 458  *
 459  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 460  * could conflict with truncation.
 461  */
 462 static void fuse_sync_writes(struct inode *inode)
 463 {
 464         fuse_set_nowrite(inode);
 465         fuse_release_nowrite(inode);
 466 }
 467
 468 static int fuse_flush(struct file *file, fl_owner_t id)
 469 {
 470         struct inode *inode = file_inode(file);
 471         struct fuse_mount *fm = get_fuse_mount(inode);
 472         struct fuse_file *ff = file->private_data;
 473         struct fuse_flush_in inarg;
 474         FUSE_ARGS(args);
 475         int err;
 476
 477         if (fuse_is_bad(inode))
 478                 return -EIO;
 479
 480         err = write_inode_now(inode, 1);
 481         if (err)
 482                 return err;
 483
 484         inode_lock(inode);
 485         fuse_sync_writes(inode);
 486         inode_unlock(inode);
 487
 488         err = filemap_check_errors(file->f_mapping);
 489         if (err)
 490                 return err;
 491
 492         err = 0;
 493         if (fm->fc->no_flush)
 494                 goto inval_attr_out;
 495
 496         memset(&inarg, 0, sizeof(inarg));
 497         inarg.fh = ff->fh;
 498         inarg.lock_owner = fuse_lock_owner_id(fm->fc, id);
 499         args.opcode = FUSE_FLUSH;
 500         args.nodeid = get_node_id(inode);
 501         args.in_numargs = 1;
 502         args.in_args[0].size = sizeof(inarg);
 503         args.in_args[0].value = &inarg;
 504         args.force = true;
 505
 506         err = fuse_simple_request(fm, &args);
 507         if (err == -ENOSYS) {
 508                 fm->fc->no_flush = 1;
 509                 err = 0;
 510         }
 511
 512 inval_attr_out:
 513         /*
 514          * In memory i_blocks is not maintained by fuse, if writeback cache is
 515          * enabled, i_blocks from cached attr may not be accurate.
 516          */
 517         if (!err && fm->fc->writeback_cache)
 518                 fuse_invalidate_attr(inode);
 519         return err;
 520 }
 521
 522 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 523                       int datasync, int opcode)
 524 {
 525         struct inode *inode = file->f_mapping->host;
 526         struct fuse_mount *fm = get_fuse_mount(inode);
 527         struct fuse_file *ff = file->private_data;
 528         FUSE_ARGS(args);
 529         struct fuse_fsync_in inarg;
 530
 531         memset(&inarg, 0, sizeof(inarg));
 532         inarg.fh = ff->fh;
 533         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 534         args.opcode = opcode;
 535         args.nodeid = get_node_id(inode);
 536         args.in_numargs = 1;
 537         args.in_args[0].size = sizeof(inarg);
 538         args.in_args[0].value = &inarg;
 539         return fuse_simple_request(fm, &args);
 540 }
 541
 542 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 543                       int datasync)
 544 {
 545         struct inode *inode = file->f_mapping->host;
 546         struct fuse_conn *fc = get_fuse_conn(inode);
 547         int err;
 548
 549         if (fuse_is_bad(inode))
 550                 return -EIO;
 551
 552         inode_lock(inode);
 553
 554         /*
 555          * Start writeback against all dirty pages of the inode, then
 556          * wait for all outstanding writes, before sending the FSYNC
 557          * request.
 558          */
 559         err = file_write_and_wait_range(file, start, end);
 560         if (err)
 561                 goto out;
 562
 563         fuse_sync_writes(inode);
 564
 565         /*
 566          * Due to implementation of fuse writeback
 567          * file_write_and_wait_range() does not catch errors.
 568          * We have to do this directly after fuse_sync_writes()
 569          */
 570         err = file_check_and_advance_wb_err(file);
 571         if (err)
 572                 goto out;
 573
 574         err = sync_inode_metadata(inode, 1);
 575         if (err)
 576                 goto out;
 577
 578         if (fc->no_fsync)
 579                 goto out;
 580
 581         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 582         if (err == -ENOSYS) {
 583                 fc->no_fsync = 1;
 584                 err = 0;
 585         }
 586 out:
 587         inode_unlock(inode);
 588
 589         return err;
 590 }
 591
 592 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
 593                          size_t count, int opcode)
 594 {
 595         struct fuse_file *ff = file->private_data;
 596         struct fuse_args *args = &ia->ap.args;
 597
 598         ia->read.in.fh = ff->fh;
 599         ia->read.in.offset = pos;
 600         ia->read.in.size = count;
 601         ia->read.in.flags = file->f_flags;
 602         args->opcode = opcode;
 603         args->nodeid = ff->nodeid;
 604         args->in_numargs = 1;
 605         args->in_args[0].size = sizeof(ia->read.in);
 606         args->in_args[0].value = &ia->read.in;
 607         args->out_argvar = true;
 608         args->out_numargs = 1;
 609         args->out_args[0].size = count;
 610 }
 611
 612 static void fuse_release_user_pages(struct fuse_args_pages *ap,
 613                                     bool should_dirty)
 614 {
 615         unsigned int i;
 616
 617         for (i = 0; i < ap->num_pages; i++) {
 618                 if (should_dirty)
 619                         set_page_dirty_lock(ap->pages[i]);
 620                 put_page(ap->pages[i]);
 621         }
 622 }
 623
 624 static void fuse_io_release(struct kref *kref)
 625 {
 626         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 627 }
 628
 629 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 630 {
 631         if (io->err)
 632                 return io->err;
 633
 634         if (io->bytes >= 0 && io->write)
 635                 return -EIO;
 636
 637         return io->bytes < 0 ? io->size : io->bytes;
 638 }
 639
 640 /**
 641  * In case of short read, the caller sets 'pos' to the position of
 642  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 643  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 644  *
 645  * An example:
 646  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 647  * both submitted asynchronously. The first of them was ACKed by userspace as
 648  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 649  * second request was ACKed as short, e.g. only 1K was read, resulting in
 650  * pos == 33K.
 651  *
 652  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 653  * will be equal to the length of the longest contiguous fragment of
 654  * transferred data starting from the beginning of IO request.
 655  */
 656 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 657 {
 658         int left;
 659
 660         spin_lock(&io->lock);
 661         if (err)
 662                 io->err = io->err ? : err;
 663         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 664                 io->bytes = pos;
 665
 666         left = --io->reqs;
 667         if (!left && io->blocking)
 668                 complete(io->done);
 669         spin_unlock(&io->lock);
 670
 671         if (!left && !io->blocking) {
 672                 ssize_t res = fuse_get_res_by_io(io);
 673
 674                 if (res >= 0) {
 675                         struct inode *inode = file_inode(io->iocb->ki_filp);
 676                         struct fuse_conn *fc = get_fuse_conn(inode);
 677                         struct fuse_inode *fi = get_fuse_inode(inode);
 678
 679                         spin_lock(&fi->lock);
 680                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 681                         spin_unlock(&fi->lock);
 682                 }
 683
 684                 io->iocb->ki_complete(io->iocb, res, 0);
 685         }
 686
 687         kref_put(&io->refcnt, fuse_io_release);
 688 }
 689
 690 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
 691                                           unsigned int npages)
 692 {
 693         struct fuse_io_args *ia;
 694
 695         ia = kzalloc(sizeof(*ia), GFP_KERNEL);
 696         if (ia) {
 697                 ia->io = io;
 698                 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
 699                                                 &ia->ap.descs);
 700                 if (!ia->ap.pages) {
 701                         kfree(ia);
 702                         ia = NULL;
 703                 }
 704         }
 705         return ia;
 706 }
 707
 708 static void fuse_io_free(struct fuse_io_args *ia)
 709 {
 710         kfree(ia->ap.pages);
 711         kfree(ia);
 712 }
 713
 714 static void fuse_aio_complete_req(struct fuse_mount *fm, struct fuse_args *args,
 715                                   int err)
 716 {
 717         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 718         struct fuse_io_priv *io = ia->io;
 719         ssize_t pos = -1;
 720
 721         fuse_release_user_pages(&ia->ap, io->should_dirty);
 722
 723         if (err) {
 724                 /* Nothing */
 725         } else if (io->write) {
 726                 if (ia->write.out.size > ia->write.in.size) {
 727                         err = -EIO;
 728                 } else if (ia->write.in.size != ia->write.out.size) {
 729                         pos = ia->write.in.offset - io->offset +
 730                                 ia->write.out.size;
 731                 }
 732         } else {
 733                 u32 outsize = args->out_args[0].size;
 734
 735                 if (ia->read.in.size != outsize)
 736                         pos = ia->read.in.offset - io->offset + outsize;
 737         }
 738
 739         fuse_aio_complete(io, err, pos);
 740         fuse_io_free(ia);
 741 }
 742
 743 static ssize_t fuse_async_req_send(struct fuse_mount *fm,
 744                                    struct fuse_io_args *ia, size_t num_bytes)
 745 {
 746         ssize_t err;
 747         struct fuse_io_priv *io = ia->io;
 748
 749         spin_lock(&io->lock);
 750         kref_get(&io->refcnt);
 751         io->size += num_bytes;
 752         io->reqs++;
 753         spin_unlock(&io->lock);
 754
 755         ia->ap.args.end = fuse_aio_complete_req;
 756         ia->ap.args.may_block = io->should_dirty;
 757         err = fuse_simple_background(fm, &ia->ap.args, GFP_KERNEL);
 758         if (err)
 759                 fuse_aio_complete_req(fm, &ia->ap.args, err);
 760
 761         return num_bytes;
 762 }
 763
 764 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
 765                               fl_owner_t owner)
 766 {
 767         struct file *file = ia->io->iocb->ki_filp;
 768         struct fuse_file *ff = file->private_data;
 769         struct fuse_mount *fm = ff->fm;
 770
 771         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 772         if (owner != NULL) {
 773                 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
 774                 ia->read.in.lock_owner = fuse_lock_owner_id(fm->fc, owner);
 775         }
 776
 777         if (ia->io->async)
 778                 return fuse_async_req_send(fm, ia, count);
 779
 780         return fuse_simple_request(fm, &ia->ap.args);
 781 }
 782
 783 static void fuse_read_update_size(struct inode *inode, loff_t size,
 784                                   u64 attr_ver)
 785 {
 786         struct fuse_conn *fc = get_fuse_conn(inode);
 787         struct fuse_inode *fi = get_fuse_inode(inode);
 788
 789         spin_lock(&fi->lock);
 790         if (attr_ver >= fi->attr_version && size < inode->i_size &&
 791             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 792                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 793                 i_size_write(inode, size);
 794         }
 795         spin_unlock(&fi->lock);
 796 }
 797
 798 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
 799                             struct fuse_args_pages *ap)
 800 {
 801         struct fuse_conn *fc = get_fuse_conn(inode);
 802
 803         if (fc->writeback_cache) {
 804                 /*
 805                  * A hole in a file. Some data after the hole are in page cache,
 806                  * but have not reached the client fs yet. So, the hole is not
 807                  * present there.
 808                  */
 809                 int i;
 810                 int start_idx = num_read >> PAGE_SHIFT;
 811                 size_t off = num_read & (PAGE_SIZE - 1);
 812
 813                 for (i = start_idx; i < ap->num_pages; i++) {
 814                         zero_user_segment(ap->pages[i], off, PAGE_SIZE);
 815                         off = 0;
 816                 }
 817         } else {
 818                 loff_t pos = page_offset(ap->pages[0]) + num_read;
 819                 fuse_read_update_size(inode, pos, attr_ver);
 820         }
 821 }
 822
 823 static int fuse_do_readpage(struct file *file, struct page *page)
 824 {
 825         struct inode *inode = page->mapping->host;
 826         struct fuse_mount *fm = get_fuse_mount(inode);
 827         loff_t pos = page_offset(page);
 828         struct fuse_page_desc desc = { .length = PAGE_SIZE };
 829         struct fuse_io_args ia = {
 830                 .ap.args.page_zeroing = true,
 831                 .ap.args.out_pages = true,
 832                 .ap.num_pages = 1,
 833                 .ap.pages = &page,
 834                 .ap.descs = &desc,
 835         };
 836         ssize_t res;
 837         u64 attr_ver;
 838
 839         /*
 840          * Page writeback can extend beyond the lifetime of the
 841          * page-cache page, so make sure we read a properly synced
 842          * page.
 843          */
 844         fuse_wait_on_page_writeback(inode, page->index);
 845
 846         attr_ver = fuse_get_attr_version(fm->fc);
 847
 848         /* Don't overflow end offset */
 849         if (pos + (desc.length - 1) == LLONG_MAX)
 850                 desc.length--;
 851
 852         fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
 853         res = fuse_simple_request(fm, &ia.ap.args);
 854         if (res < 0)
 855                 return res;
 856         /*
 857          * Short read means EOF.  If file size is larger, truncate it
 858          */
 859         if (res < desc.length)
 860                 fuse_short_read(inode, attr_ver, res, &ia.ap);
 861
 862         SetPageUptodate(page);
 863
 864         return 0;
 865 }
 866
 867 static int fuse_readpage(struct file *file, struct page *page)
 868 {
 869         struct inode *inode = page->mapping->host;
 870         int err;
 871
 872         err = -EIO;
 873         if (fuse_is_bad(inode))
 874                 goto out;
 875
 876         err = fuse_do_readpage(file, page);
 877         fuse_invalidate_atime(inode);
 878  out:
 879         unlock_page(page);
 880         return err;
 881 }
 882
 883 static void fuse_readpages_end(struct fuse_mount *fm, struct fuse_args *args,
 884                                int err)
 885 {
 886         int i;
 887         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 888         struct fuse_args_pages *ap = &ia->ap;
 889         size_t count = ia->read.in.size;
 890         size_t num_read = args->out_args[0].size;
 891         struct address_space *mapping = NULL;
 892
 893         for (i = 0; mapping == NULL && i < ap->num_pages; i++)
 894                 mapping = ap->pages[i]->mapping;
 895
 896         if (mapping) {
 897                 struct inode *inode = mapping->host;
 898
 899                 /*
 900                  * Short read means EOF. If file size is larger, truncate it
 901                  */
 902                 if (!err && num_read < count)
 903                         fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
 904
 905                 fuse_invalidate_atime(inode);
 906         }
 907
 908         for (i = 0; i < ap->num_pages; i++) {
 909                 struct page *page = ap->pages[i];
 910
 911                 if (!err)
 912                         SetPageUptodate(page);
 913                 else
 914                         SetPageError(page);
 915                 unlock_page(page);
 916                 put_page(page);
 917         }
 918         if (ia->ff)
 919                 fuse_file_put(ia->ff, false, false);
 920
 921         fuse_io_free(ia);
 922 }
 923
 924 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
 925 {
 926         struct fuse_file *ff = file->private_data;
 927         struct fuse_mount *fm = ff->fm;
 928         struct fuse_args_pages *ap = &ia->ap;
 929         loff_t pos = page_offset(ap->pages[0]);
 930         size_t count = ap->num_pages << PAGE_SHIFT;
 931         ssize_t res;
 932         int err;
 933
 934         ap->args.out_pages = true;
 935         ap->args.page_zeroing = true;
 936         ap->args.page_replace = true;
 937
 938         /* Don't overflow end offset */
 939         if (pos + (count - 1) == LLONG_MAX) {
 940                 count--;
 941                 ap->descs[ap->num_pages - 1].length--;
 942         }
 943         WARN_ON((loff_t) (pos + count) < 0);
 944
 945         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 946         ia->read.attr_ver = fuse_get_attr_version(fm->fc);
 947         if (fm->fc->async_read) {
 948                 ia->ff = fuse_file_get(ff);
 949                 ap->args.end = fuse_readpages_end;
 950                 err = fuse_simple_background(fm, &ap->args, GFP_KERNEL);
 951                 if (!err)
 952                         return;
 953         } else {
 954                 res = fuse_simple_request(fm, &ap->args);
 955                 err = res < 0 ? res : 0;
 956         }
 957         fuse_readpages_end(fm, &ap->args, err);
 958 }
 959
 960 static void fuse_readahead(struct readahead_control *rac)
 961 {
 962         struct inode *inode = rac->mapping->host;
 963         struct fuse_conn *fc = get_fuse_conn(inode);
 964         unsigned int i, max_pages, nr_pages = 0;
 965
 966         if (fuse_is_bad(inode))
 967                 return;
 968
 969         max_pages = min_t(unsigned int, fc->max_pages,
 970                         fc->max_read / PAGE_SIZE);
 971
 972         for (;;) {
 973                 struct fuse_io_args *ia;
 974                 struct fuse_args_pages *ap;
 975
 976                 nr_pages = readahead_count(rac) - nr_pages;
 977                 if (nr_pages > max_pages)
 978                         nr_pages = max_pages;
 979                 if (nr_pages == 0)
 980                         break;
 981                 ia = fuse_io_alloc(NULL, nr_pages);
 982                 if (!ia)
 983                         return;
 984                 ap = &ia->ap;
 985                 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
 986                 for (i = 0; i < nr_pages; i++) {
 987                         fuse_wait_on_page_writeback(inode,
 988                                                     readahead_index(rac) + i);
 989                         ap->descs[i].length = PAGE_SIZE;
 990                 }
 991                 ap->num_pages = nr_pages;
 992                 fuse_send_readpages(ia, rac->file);
 993         }
 994 }
 995
 996 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
 997 {
 998         struct inode *inode = iocb->ki_filp->f_mapping->host;
 999         struct fuse_conn *fc = get_fuse_conn(inode);
1000
1001         /*
1002          * In auto invalidate mode, always update attributes on read.
1003          * Otherwise, only update if we attempt to read past EOF (to ensure
1004          * i_size is up to date).
1005          */
1006         if (fc->auto_inval_data ||
1007             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
1008                 int err;
1009                 err = fuse_update_attributes(inode, iocb->ki_filp);
1010                 if (err)
1011                         return err;
1012         }
1013
1014         return generic_file_read_iter(iocb, to);
1015 }
1016
1017 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1018                                  loff_t pos, size_t count)
1019 {
1020         struct fuse_args *args = &ia->ap.args;
1021
1022         ia->write.in.fh = ff->fh;
1023         ia->write.in.offset = pos;
1024         ia->write.in.size = count;
1025         args->opcode = FUSE_WRITE;
1026         args->nodeid = ff->nodeid;
1027         args->in_numargs = 2;
1028         if (ff->fm->fc->minor < 9)
1029                 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1030         else
1031                 args->in_args[0].size = sizeof(ia->write.in);
1032         args->in_args[0].value = &ia->write.in;
1033         args->in_args[1].size = count;
1034         args->out_numargs = 1;
1035         args->out_args[0].size = sizeof(ia->write.out);
1036         args->out_args[0].value = &ia->write.out;
1037 }
1038
1039 static unsigned int fuse_write_flags(struct kiocb *iocb)
1040 {
1041         unsigned int flags = iocb->ki_filp->f_flags;
1042
1043         if (iocb->ki_flags & IOCB_DSYNC)
1044                 flags |= O_DSYNC;
1045         if (iocb->ki_flags & IOCB_SYNC)
1046                 flags |= O_SYNC;
1047
1048         return flags;
1049 }
1050
1051 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1052                                size_t count, fl_owner_t owner)
1053 {
1054         struct kiocb *iocb = ia->io->iocb;
1055         struct file *file = iocb->ki_filp;
1056         struct fuse_file *ff = file->private_data;
1057         struct fuse_mount *fm = ff->fm;
1058         struct fuse_write_in *inarg = &ia->write.in;
1059         ssize_t err;
1060
1061         fuse_write_args_fill(ia, ff, pos, count);
1062         inarg->flags = fuse_write_flags(iocb);
1063         if (owner != NULL) {
1064                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1065                 inarg->lock_owner = fuse_lock_owner_id(fm->fc, owner);
1066         }
1067
1068         if (ia->io->async)
1069                 return fuse_async_req_send(fm, ia, count);
1070
1071         err = fuse_simple_request(fm, &ia->ap.args);
1072         if (!err && ia->write.out.size > count)
1073                 err = -EIO;
1074
1075         return err ?: ia->write.out.size;
1076 }
1077
1078 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1079 {
1080         struct fuse_conn *fc = get_fuse_conn(inode);
1081         struct fuse_inode *fi = get_fuse_inode(inode);
1082         bool ret = false;
1083
1084         spin_lock(&fi->lock);
1085         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1086         if (pos > inode->i_size) {
1087                 i_size_write(inode, pos);
1088                 ret = true;
1089         }
1090         spin_unlock(&fi->lock);
1091
1092         return ret;
1093 }
1094
1095 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1096                                      struct kiocb *iocb, struct inode *inode,
1097                                      loff_t pos, size_t count)
1098 {
1099         struct fuse_args_pages *ap = &ia->ap;
1100         struct file *file = iocb->ki_filp;
1101         struct fuse_file *ff = file->private_data;
1102         struct fuse_mount *fm = ff->fm;
1103         unsigned int offset, i;
1104         bool short_write;
1105         int err;
1106
1107         for (i = 0; i < ap->num_pages; i++)
1108                 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1109
1110         fuse_write_args_fill(ia, ff, pos, count);
1111         ia->write.in.flags = fuse_write_flags(iocb);
1112
1113         err = fuse_simple_request(fm, &ap->args);
1114         if (!err && ia->write.out.size > count)
1115                 err = -EIO;
1116
1117         short_write = ia->write.out.size < count;
1118         offset = ap->descs[0].offset;
1119         count = ia->write.out.size;
1120         for (i = 0; i < ap->num_pages; i++) {
1121                 struct page *page = ap->pages[i];
1122
1123                 if (err) {
1124                         ClearPageUptodate(page);
1125                 } else {
1126                         if (count >= PAGE_SIZE - offset)
1127                                 count -= PAGE_SIZE - offset;
1128                         else {
1129                                 if (short_write)
1130                                         ClearPageUptodate(page);
1131                                 count = 0;
1132                         }
1133                         offset = 0;
1134                 }
1135                 if (ia->write.page_locked && (i == ap->num_pages - 1))
1136                         unlock_page(page);
1137                 put_page(page);
1138         }
1139
1140         return err;
1141 }
1142
1143 static ssize_t fuse_fill_write_pages(struct fuse_io_args *ia,
1144                                      struct address_space *mapping,
1145                                      struct iov_iter *ii, loff_t pos,
1146                                      unsigned int max_pages)
1147 {
1148         struct fuse_args_pages *ap = &ia->ap;
1149         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1150         unsigned offset = pos & (PAGE_SIZE - 1);
1151         size_t count = 0;
1152         int err;
1153
1154         ap->args.in_pages = true;
1155         ap->descs[0].offset = offset;
1156
1157         do {
1158                 size_t tmp;
1159                 struct page *page;
1160                 pgoff_t index = pos >> PAGE_SHIFT;
1161                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1162                                      iov_iter_count(ii));
1163
1164                 bytes = min_t(size_t, bytes, fc->max_write - count);
1165
1166  again:
1167                 err = -EFAULT;
1168                 if (iov_iter_fault_in_readable(ii, bytes))
1169                         break;
1170
1171                 err = -ENOMEM;
1172                 page = grab_cache_page_write_begin(mapping, index, 0);
1173                 if (!page)
1174                         break;
1175
1176                 if (mapping_writably_mapped(mapping))
1177                         flush_dcache_page(page);
1178
1179                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1180                 flush_dcache_page(page);
1181
1182                 iov_iter_advance(ii, tmp);
1183                 if (!tmp) {
1184                         unlock_page(page);
1185                         put_page(page);
1186                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1187                         goto again;
1188                 }
1189
1190                 err = 0;
1191                 ap->pages[ap->num_pages] = page;
1192                 ap->descs[ap->num_pages].length = tmp;
1193                 ap->num_pages++;
1194
1195                 count += tmp;
1196                 pos += tmp;
1197                 offset += tmp;
1198                 if (offset == PAGE_SIZE)
1199                         offset = 0;
1200
1201                 /* If we copied full page, mark it uptodate */
1202                 if (tmp == PAGE_SIZE)
1203                         SetPageUptodate(page);
1204
1205                 if (PageUptodate(page)) {
1206                         unlock_page(page);
1207                 } else {
1208                         ia->write.page_locked = true;
1209                         break;
1210                 }
1211                 if (!fc->big_writes)
1212                         break;
1213         } while (iov_iter_count(ii) && count < fc->max_write &&
1214                  ap->num_pages < max_pages && offset == 0);
1215
1216         return count > 0 ? count : err;
1217 }
1218
1219 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1220                                      unsigned int max_pages)
1221 {
1222         return min_t(unsigned int,
1223                      ((pos + len - 1) >> PAGE_SHIFT) -
1224                      (pos >> PAGE_SHIFT) + 1,
1225                      max_pages);
1226 }
1227
1228 static ssize_t fuse_perform_write(struct kiocb *iocb,
1229                                   struct address_space *mapping,
1230                                   struct iov_iter *ii, loff_t pos)
1231 {
1232         struct inode *inode = mapping->host;
1233         struct fuse_conn *fc = get_fuse_conn(inode);
1234         struct fuse_inode *fi = get_fuse_inode(inode);
1235         int err = 0;
1236         ssize_t res = 0;
1237
1238         if (inode->i_size < pos + iov_iter_count(ii))
1239                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1240
1241         do {
1242                 ssize_t count;
1243                 struct fuse_io_args ia = {};
1244                 struct fuse_args_pages *ap = &ia.ap;
1245                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1246                                                       fc->max_pages);
1247
1248                 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1249                 if (!ap->pages) {
1250                         err = -ENOMEM;
1251                         break;
1252                 }
1253
1254                 count = fuse_fill_write_pages(&ia, mapping, ii, pos, nr_pages);
1255                 if (count <= 0) {
1256                         err = count;
1257                 } else {
1258                         err = fuse_send_write_pages(&ia, iocb, inode,
1259                                                     pos, count);
1260                         if (!err) {
1261                                 size_t num_written = ia.write.out.size;
1262
1263                                 res += num_written;
1264                                 pos += num_written;
1265
1266                                 /* break out of the loop on short write */
1267                                 if (num_written != count)
1268                                         err = -EIO;
1269                         }
1270                 }
1271                 kfree(ap->pages);
1272         } while (!err && iov_iter_count(ii));
1273
1274         if (res > 0)
1275                 fuse_write_update_size(inode, pos);
1276
1277         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1278         fuse_invalidate_attr(inode);
1279
1280         return res > 0 ? res : err;
1281 }
1282
1283 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1284 {
1285         struct file *file = iocb->ki_filp;
1286         struct address_space *mapping = file->f_mapping;
1287         ssize_t written = 0;
1288         ssize_t written_buffered = 0;
1289         struct inode *inode = mapping->host;
1290         ssize_t err;
1291         loff_t endbyte = 0;
1292
1293         if (get_fuse_conn(inode)->writeback_cache) {
1294                 /* Update size (EOF optimization) and mode (SUID clearing) */
1295                 err = fuse_update_attributes(mapping->host, file);
1296                 if (err)
1297                         return err;
1298
1299                 return generic_file_write_iter(iocb, from);
1300         }
1301
1302         inode_lock(inode);
1303
1304         /* We can write back this queue in page reclaim */
1305         current->backing_dev_info = inode_to_bdi(inode);
1306
1307         err = generic_write_checks(iocb, from);
1308         if (err <= 0)
1309                 goto out;
1310
1311         err = file_remove_privs(file);
1312         if (err)
1313                 goto out;
1314
1315         err = file_update_time(file);
1316         if (err)
1317                 goto out;
1318
1319         if (iocb->ki_flags & IOCB_DIRECT) {
1320                 loff_t pos = iocb->ki_pos;
1321                 written = generic_file_direct_write(iocb, from);
1322                 if (written < 0 || !iov_iter_count(from))
1323                         goto out;
1324
1325                 pos += written;
1326
1327                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1328                 if (written_buffered < 0) {
1329                         err = written_buffered;
1330                         goto out;
1331                 }
1332                 endbyte = pos + written_buffered - 1;
1333
1334                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1335                                                    endbyte);
1336                 if (err)
1337                         goto out;
1338
1339                 invalidate_mapping_pages(file->f_mapping,
1340                                          pos >> PAGE_SHIFT,
1341                                          endbyte >> PAGE_SHIFT);
1342
1343                 written += written_buffered;
1344                 iocb->ki_pos = pos + written_buffered;
1345         } else {
1346                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1347                 if (written >= 0)
1348                         iocb->ki_pos += written;
1349         }
1350 out:
1351         current->backing_dev_info = NULL;
1352         inode_unlock(inode);
1353         if (written > 0)
1354                 written = generic_write_sync(iocb, written);
1355
1356         return written ? written : err;
1357 }
1358
1359 static inline void fuse_page_descs_length_init(struct fuse_page_desc *descs,
1360                                                unsigned int index,
1361                                                unsigned int nr_pages)
1362 {
1363         int i;
1364
1365         for (i = index; i < index + nr_pages; i++)
1366                 descs[i].length = PAGE_SIZE - descs[i].offset;
1367 }
1368
1369 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1370 {
1371         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1372 }
1373
1374 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1375                                         size_t max_size)
1376 {
1377         return min(iov_iter_single_seg_count(ii), max_size);
1378 }
1379
1380 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1381                                size_t *nbytesp, int write,
1382                                unsigned int max_pages)
1383 {
1384         size_t nbytes = 0;  /* # bytes already packed in req */
1385         ssize_t ret = 0;
1386
1387         /* Special case for kernel I/O: can copy directly into the buffer */
1388         if (iov_iter_is_kvec(ii)) {
1389                 unsigned long user_addr = fuse_get_user_addr(ii);
1390                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1391
1392                 if (write)
1393                         ap->args.in_args[1].value = (void *) user_addr;
1394                 else
1395                         ap->args.out_args[0].value = (void *) user_addr;
1396
1397                 iov_iter_advance(ii, frag_size);
1398                 *nbytesp = frag_size;
1399                 return 0;
1400         }
1401
1402         while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1403                 unsigned npages;
1404                 size_t start;
1405                 ret = iov_iter_get_pages(ii, &ap->pages[ap->num_pages],
1406                                         *nbytesp - nbytes,
1407                                         max_pages - ap->num_pages,
1408                                         &start);
1409                 if (ret < 0)
1410                         break;
1411
1412                 iov_iter_advance(ii, ret);
1413                 nbytes += ret;
1414
1415                 ret += start;
1416                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1417
1418                 ap->descs[ap->num_pages].offset = start;
1419                 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1420
1421                 ap->num_pages += npages;
1422                 ap->descs[ap->num_pages - 1].length -=
1423                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1424         }
1425
1426         ap->args.user_pages = true;
1427         if (write)
1428                 ap->args.in_pages = true;
1429         else
1430                 ap->args.out_pages = true;
1431
1432         *nbytesp = nbytes;
1433
1434         return ret < 0 ? ret : 0;
1435 }
1436
1437 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1438                        loff_t *ppos, int flags)
1439 {
1440         int write = flags & FUSE_DIO_WRITE;
1441         int cuse = flags & FUSE_DIO_CUSE;
1442         struct file *file = io->iocb->ki_filp;
1443         struct inode *inode = file->f_mapping->host;
1444         struct fuse_file *ff = file->private_data;
1445         struct fuse_conn *fc = ff->fm->fc;
1446         size_t nmax = write ? fc->max_write : fc->max_read;
1447         loff_t pos = *ppos;
1448         size_t count = iov_iter_count(iter);
1449         pgoff_t idx_from = pos >> PAGE_SHIFT;
1450         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1451         ssize_t res = 0;
1452         int err = 0;
1453         struct fuse_io_args *ia;
1454         unsigned int max_pages;
1455
1456         max_pages = iov_iter_npages(iter, fc->max_pages);
1457         ia = fuse_io_alloc(io, max_pages);
1458         if (!ia)
1459                 return -ENOMEM;
1460
1461         ia->io = io;
1462         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1463                 if (!write)
1464                         inode_lock(inode);
1465                 fuse_sync_writes(inode);
1466                 if (!write)
1467                         inode_unlock(inode);
1468         }
1469
1470         io->should_dirty = !write && iter_is_iovec(iter);
1471         while (count) {
1472                 ssize_t nres;
1473                 fl_owner_t owner = current->files;
1474                 size_t nbytes = min(count, nmax);
1475
1476                 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1477                                           max_pages);
1478                 if (err && !nbytes)
1479                         break;
1480
1481                 if (write) {
1482                         if (!capable(CAP_FSETID))
1483                                 ia->write.in.write_flags |= FUSE_WRITE_KILL_PRIV;
1484
1485                         nres = fuse_send_write(ia, pos, nbytes, owner);
1486                 } else {
1487                         nres = fuse_send_read(ia, pos, nbytes, owner);
1488                 }
1489
1490                 if (!io->async || nres < 0) {
1491                         fuse_release_user_pages(&ia->ap, io->should_dirty);
1492                         fuse_io_free(ia);
1493                 }
1494                 ia = NULL;
1495                 if (nres < 0) {
1496                         iov_iter_revert(iter, nbytes);
1497                         err = nres;
1498                         break;
1499                 }
1500                 WARN_ON(nres > nbytes);
1501
1502                 count -= nres;
1503                 res += nres;
1504                 pos += nres;
1505                 if (nres != nbytes) {
1506                         iov_iter_revert(iter, nbytes - nres);
1507                         break;
1508                 }
1509                 if (count) {
1510                         max_pages = iov_iter_npages(iter, fc->max_pages);
1511                         ia = fuse_io_alloc(io, max_pages);
1512                         if (!ia)
1513                                 break;
1514                 }
1515         }
1516         if (ia)
1517                 fuse_io_free(ia);
1518         if (res > 0)
1519                 *ppos = pos;
1520
1521         return res > 0 ? res : err;
1522 }
1523 EXPORT_SYMBOL_GPL(fuse_direct_io);
1524
1525 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1526                                   struct iov_iter *iter,
1527                                   loff_t *ppos)
1528 {
1529         ssize_t res;
1530         struct inode *inode = file_inode(io->iocb->ki_filp);
1531
1532         res = fuse_direct_io(io, iter, ppos, 0);
1533
1534         fuse_invalidate_atime(inode);
1535
1536         return res;
1537 }
1538
1539 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1540
1541 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1542 {
1543         ssize_t res;
1544
1545         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1546                 res = fuse_direct_IO(iocb, to);
1547         } else {
1548                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1549
1550                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1551         }
1552
1553         return res;
1554 }
1555
1556 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1557 {
1558         struct inode *inode = file_inode(iocb->ki_filp);
1559         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1560         ssize_t res;
1561
1562         /* Don't allow parallel writes to the same file */
1563         inode_lock(inode);
1564         res = generic_write_checks(iocb, from);
1565         if (res > 0) {
1566                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1567                         res = fuse_direct_IO(iocb, from);
1568                 } else {
1569                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1570                                              FUSE_DIO_WRITE);
1571                 }
1572         }
1573         fuse_invalidate_attr(inode);
1574         if (res > 0)
1575                 fuse_write_update_size(inode, iocb->ki_pos);
1576         inode_unlock(inode);
1577
1578         return res;
1579 }
1580
1581 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1582 {
1583         struct file *file = iocb->ki_filp;
1584         struct fuse_file *ff = file->private_data;
1585         struct inode *inode = file_inode(file);
1586
1587         if (fuse_is_bad(inode))
1588                 return -EIO;
1589
1590         if (FUSE_IS_DAX(inode))
1591                 return fuse_dax_read_iter(iocb, to);
1592
1593         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1594                 return fuse_cache_read_iter(iocb, to);
1595         else
1596                 return fuse_direct_read_iter(iocb, to);
1597 }
1598
1599 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1600 {
1601         struct file *file = iocb->ki_filp;
1602         struct fuse_file *ff = file->private_data;
1603         struct inode *inode = file_inode(file);
1604
1605         if (fuse_is_bad(inode))
1606                 return -EIO;
1607
1608         if (FUSE_IS_DAX(inode))
1609                 return fuse_dax_write_iter(iocb, from);
1610
1611         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1612                 return fuse_cache_write_iter(iocb, from);
1613         else
1614                 return fuse_direct_write_iter(iocb, from);
1615 }
1616
1617 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1618 {
1619         struct fuse_args_pages *ap = &wpa->ia.ap;
1620         int i;
1621
1622         for (i = 0; i < ap->num_pages; i++)
1623                 __free_page(ap->pages[i]);
1624
1625         if (wpa->ia.ff)
1626                 fuse_file_put(wpa->ia.ff, false, false);
1627
1628         kfree(ap->pages);
1629         kfree(wpa);
1630 }
1631
1632 static void fuse_writepage_finish(struct fuse_mount *fm,
1633                                   struct fuse_writepage_args *wpa)
1634 {
1635         struct fuse_args_pages *ap = &wpa->ia.ap;
1636         struct inode *inode = wpa->inode;
1637         struct fuse_inode *fi = get_fuse_inode(inode);
1638         struct backing_dev_info *bdi = inode_to_bdi(inode);
1639         int i;
1640
1641         for (i = 0; i < ap->num_pages; i++) {
1642                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1643                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1644                 wb_writeout_inc(&bdi->wb);
1645         }
1646         wake_up(&fi->page_waitq);
1647 }
1648
1649 /* Called under fi->lock, may release and reacquire it */
1650 static void fuse_send_writepage(struct fuse_mount *fm,
1651                                 struct fuse_writepage_args *wpa, loff_t size)
1652 __releases(fi->lock)
1653 __acquires(fi->lock)
1654 {
1655         struct fuse_writepage_args *aux, *next;
1656         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1657         struct fuse_write_in *inarg = &wpa->ia.write.in;
1658         struct fuse_args *args = &wpa->ia.ap.args;
1659         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1660         int err;
1661
1662         fi->writectr++;
1663         if (inarg->offset + data_size <= size) {
1664                 inarg->size = data_size;
1665         } else if (inarg->offset < size) {
1666                 inarg->size = size - inarg->offset;
1667         } else {
1668                 /* Got truncated off completely */
1669                 goto out_free;
1670         }
1671
1672         args->in_args[1].size = inarg->size;
1673         args->force = true;
1674         args->nocreds = true;
1675
1676         err = fuse_simple_background(fm, args, GFP_ATOMIC);
1677         if (err == -ENOMEM) {
1678                 spin_unlock(&fi->lock);
1679                 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1680                 spin_lock(&fi->lock);
1681         }
1682
1683         /* Fails on broken connection only */
1684         if (unlikely(err))
1685                 goto out_free;
1686
1687         return;
1688
1689  out_free:
1690         fi->writectr--;
1691         rb_erase(&wpa->writepages_entry, &fi->writepages);
1692         fuse_writepage_finish(fm, wpa);
1693         spin_unlock(&fi->lock);
1694
1695         /* After fuse_writepage_finish() aux request list is private */
1696         for (aux = wpa->next; aux; aux = next) {
1697                 next = aux->next;
1698                 aux->next = NULL;
1699                 fuse_writepage_free(aux);
1700         }
1701
1702         fuse_writepage_free(wpa);
1703         spin_lock(&fi->lock);
1704 }
1705
1706 /*
1707  * If fi->writectr is positive (no truncate or fsync going on) send
1708  * all queued writepage requests.
1709  *
1710  * Called with fi->lock
1711  */
1712 void fuse_flush_writepages(struct inode *inode)
1713 __releases(fi->lock)
1714 __acquires(fi->lock)
1715 {
1716         struct fuse_mount *fm = get_fuse_mount(inode);
1717         struct fuse_inode *fi = get_fuse_inode(inode);
1718         loff_t crop = i_size_read(inode);
1719         struct fuse_writepage_args *wpa;
1720
1721         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1722                 wpa = list_entry(fi->queued_writes.next,
1723                                  struct fuse_writepage_args, queue_entry);
1724                 list_del_init(&wpa->queue_entry);
1725                 fuse_send_writepage(fm, wpa, crop);
1726         }
1727 }
1728
1729 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1730                                                 struct fuse_writepage_args *wpa)
1731 {
1732         pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1733         pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1734         struct rb_node **p = &root->rb_node;
1735         struct rb_node  *parent = NULL;
1736
1737         WARN_ON(!wpa->ia.ap.num_pages);
1738         while (*p) {
1739                 struct fuse_writepage_args *curr;
1740                 pgoff_t curr_index;
1741
1742                 parent = *p;
1743                 curr = rb_entry(parent, struct fuse_writepage_args,
1744                                 writepages_entry);
1745                 WARN_ON(curr->inode != wpa->inode);
1746                 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1747
1748                 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1749                         p = &(*p)->rb_right;
1750                 else if (idx_to < curr_index)
1751                         p = &(*p)->rb_left;
1752                 else
1753                         return curr;
1754         }
1755
1756         rb_link_node(&wpa->writepages_entry, parent, p);
1757         rb_insert_color(&wpa->writepages_entry, root);
1758         return NULL;
1759 }
1760
1761 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1762 {
1763         WARN_ON(fuse_insert_writeback(root, wpa));
1764 }
1765
1766 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1767                                int error)
1768 {
1769         struct fuse_writepage_args *wpa =
1770                 container_of(args, typeof(*wpa), ia.ap.args);
1771         struct inode *inode = wpa->inode;
1772         struct fuse_inode *fi = get_fuse_inode(inode);
1773         struct fuse_conn *fc = get_fuse_conn(inode);
1774
1775         mapping_set_error(inode->i_mapping, error);
1776         /*
1777          * A writeback finished and this might have updated mtime/ctime on
1778          * server making local mtime/ctime stale.  Hence invalidate attrs.
1779          * Do this only if writeback_cache is not enabled.  If writeback_cache
1780          * is enabled, we trust local ctime/mtime.
1781          */
1782         if (!fc->writeback_cache)
1783                 fuse_invalidate_attr(inode);
1784         spin_lock(&fi->lock);
1785         rb_erase(&wpa->writepages_entry, &fi->writepages);
1786         while (wpa->next) {
1787                 struct fuse_mount *fm = get_fuse_mount(inode);
1788                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1789                 struct fuse_writepage_args *next = wpa->next;
1790
1791                 wpa->next = next->next;
1792                 next->next = NULL;
1793                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1794                 tree_insert(&fi->writepages, next);
1795
1796                 /*
1797                  * Skip fuse_flush_writepages() to make it easy to crop requests
1798                  * based on primary request size.
1799                  *
1800                  * 1st case (trivial): there are no concurrent activities using
1801                  * fuse_set/release_nowrite.  Then we're on safe side because
1802                  * fuse_flush_writepages() would call fuse_send_writepage()
1803                  * anyway.
1804                  *
1805                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1806                  * now for completion of all in-flight requests.  This happens
1807                  * rarely and no more than once per page, so this should be
1808                  * okay.
1809                  *
1810                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1811                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1812                  * that fuse_set_nowrite returned implies that all in-flight
1813                  * requests were completed along with all of their secondary
1814                  * requests.  Further primary requests are blocked by negative
1815                  * writectr.  Hence there cannot be any in-flight requests and
1816                  * no invocations of fuse_writepage_end() while we're in
1817                  * fuse_set_nowrite..fuse_release_nowrite section.
1818                  */
1819                 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1820         }
1821         fi->writectr--;
1822         fuse_writepage_finish(fm, wpa);
1823         spin_unlock(&fi->lock);
1824         fuse_writepage_free(wpa);
1825 }
1826
1827 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1828                                                struct fuse_inode *fi)
1829 {
1830         struct fuse_file *ff = NULL;
1831
1832         spin_lock(&fi->lock);
1833         if (!list_empty(&fi->write_files)) {
1834                 ff = list_entry(fi->write_files.next, struct fuse_file,
1835                                 write_entry);
1836                 fuse_file_get(ff);
1837         }
1838         spin_unlock(&fi->lock);
1839
1840         return ff;
1841 }
1842
1843 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1844                                              struct fuse_inode *fi)
1845 {
1846         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1847         WARN_ON(!ff);
1848         return ff;
1849 }
1850
1851 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1852 {
1853         struct fuse_conn *fc = get_fuse_conn(inode);
1854         struct fuse_inode *fi = get_fuse_inode(inode);
1855         struct fuse_file *ff;
1856         int err;
1857
1858         /*
1859          * Inode is always written before the last reference is dropped and
1860          * hence this should not be reached from reclaim.
1861          *
1862          * Writing back the inode from reclaim can deadlock if the request
1863          * processing itself needs an allocation.  Allocations triggering
1864          * reclaim while serving a request can't be prevented, because it can
1865          * involve any number of unrelated userspace processes.
1866          */
1867         WARN_ON(wbc->for_reclaim);
1868
1869         ff = __fuse_write_file_get(fc, fi);
1870         err = fuse_flush_times(inode, ff);
1871         if (ff)
1872                 fuse_file_put(ff, false, false);
1873
1874         return err;
1875 }
1876
1877 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1878 {
1879         struct fuse_writepage_args *wpa;
1880         struct fuse_args_pages *ap;
1881
1882         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1883         if (wpa) {
1884                 ap = &wpa->ia.ap;
1885                 ap->num_pages = 0;
1886                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1887                 if (!ap->pages) {
1888                         kfree(wpa);
1889                         wpa = NULL;
1890                 }
1891         }
1892         return wpa;
1893
1894 }
1895
1896 static int fuse_writepage_locked(struct page *page)
1897 {
1898         struct address_space *mapping = page->mapping;
1899         struct inode *inode = mapping->host;
1900         struct fuse_conn *fc = get_fuse_conn(inode);
1901         struct fuse_inode *fi = get_fuse_inode(inode);
1902         struct fuse_writepage_args *wpa;
1903         struct fuse_args_pages *ap;
1904         struct page *tmp_page;
1905         int error = -ENOMEM;
1906
1907         set_page_writeback(page);
1908
1909         wpa = fuse_writepage_args_alloc();
1910         if (!wpa)
1911                 goto err;
1912         ap = &wpa->ia.ap;
1913
1914         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1915         if (!tmp_page)
1916                 goto err_free;
1917
1918         error = -EIO;
1919         wpa->ia.ff = fuse_write_file_get(fc, fi);
1920         if (!wpa->ia.ff)
1921                 goto err_nofile;
1922
1923         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1924
1925         copy_highpage(tmp_page, page);
1926         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1927         wpa->next = NULL;
1928         ap->args.in_pages = true;
1929         ap->num_pages = 1;
1930         ap->pages[0] = tmp_page;
1931         ap->descs[0].offset = 0;
1932         ap->descs[0].length = PAGE_SIZE;
1933         ap->args.end = fuse_writepage_end;
1934         wpa->inode = inode;
1935
1936         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1937         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1938
1939         spin_lock(&fi->lock);
1940         tree_insert(&fi->writepages, wpa);
1941         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1942         fuse_flush_writepages(inode);
1943         spin_unlock(&fi->lock);
1944
1945         end_page_writeback(page);
1946
1947         return 0;
1948
1949 err_nofile:
1950         __free_page(tmp_page);
1951 err_free:
1952         kfree(wpa);
1953 err:
1954         mapping_set_error(page->mapping, error);
1955         end_page_writeback(page);
1956         return error;
1957 }
1958
1959 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1960 {
1961         int err;
1962
1963         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1964                 /*
1965                  * ->writepages() should be called for sync() and friends.  We
1966                  * should only get here on direct reclaim and then we are
1967                  * allowed to skip a page which is already in flight
1968                  */
1969                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1970
1971                 redirty_page_for_writepage(wbc, page);
1972                 unlock_page(page);
1973                 return 0;
1974         }
1975
1976         err = fuse_writepage_locked(page);
1977         unlock_page(page);
1978
1979         return err;
1980 }
1981
1982 struct fuse_fill_wb_data {
1983         struct fuse_writepage_args *wpa;
1984         struct fuse_file *ff;
1985         struct inode *inode;
1986         struct page **orig_pages;
1987         unsigned int max_pages;
1988 };
1989
1990 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
1991 {
1992         struct fuse_args_pages *ap = &data->wpa->ia.ap;
1993         struct fuse_conn *fc = get_fuse_conn(data->inode);
1994         struct page **pages;
1995         struct fuse_page_desc *descs;
1996         unsigned int npages = min_t(unsigned int,
1997                                     max_t(unsigned int, data->max_pages * 2,
1998                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
1999                                     fc->max_pages);
2000         WARN_ON(npages <= data->max_pages);
2001
2002         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
2003         if (!pages)
2004                 return false;
2005
2006         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
2007         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
2008         kfree(ap->pages);
2009         ap->pages = pages;
2010         ap->descs = descs;
2011         data->max_pages = npages;
2012
2013         return true;
2014 }
2015
2016 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
2017 {
2018         struct fuse_writepage_args *wpa = data->wpa;
2019         struct inode *inode = data->inode;
2020         struct fuse_inode *fi = get_fuse_inode(inode);
2021         int num_pages = wpa->ia.ap.num_pages;
2022         int i;
2023
2024         wpa->ia.ff = fuse_file_get(data->ff);
2025         spin_lock(&fi->lock);
2026         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2027         fuse_flush_writepages(inode);
2028         spin_unlock(&fi->lock);
2029
2030         for (i = 0; i < num_pages; i++)
2031                 end_page_writeback(data->orig_pages[i]);
2032 }
2033
2034 /*
2035  * Check under fi->lock if the page is under writeback, and insert it onto the
2036  * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2037  * one already added for a page at this offset.  If there's none, then insert
2038  * this new request onto the auxiliary list, otherwise reuse the existing one by
2039  * swapping the new temp page with the old one.
2040  */
2041 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2042                                struct page *page)
2043 {
2044         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2045         struct fuse_writepage_args *tmp;
2046         struct fuse_writepage_args *old_wpa;
2047         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2048
2049         WARN_ON(new_ap->num_pages != 0);
2050         new_ap->num_pages = 1;
2051
2052         spin_lock(&fi->lock);
2053         old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2054         if (!old_wpa) {
2055                 spin_unlock(&fi->lock);
2056                 return true;
2057         }
2058
2059         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2060                 pgoff_t curr_index;
2061
2062                 WARN_ON(tmp->inode != new_wpa->inode);
2063                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2064                 if (curr_index == page->index) {
2065                         WARN_ON(tmp->ia.ap.num_pages != 1);
2066                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2067                         break;
2068                 }
2069         }
2070
2071         if (!tmp) {
2072                 new_wpa->next = old_wpa->next;
2073                 old_wpa->next = new_wpa;
2074         }
2075
2076         spin_unlock(&fi->lock);
2077
2078         if (tmp) {
2079                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2080
2081                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2082                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2083                 wb_writeout_inc(&bdi->wb);
2084                 fuse_writepage_free(new_wpa);
2085         }
2086
2087         return false;
2088 }
2089
2090 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2091                                      struct fuse_args_pages *ap,
2092                                      struct fuse_fill_wb_data *data)
2093 {
2094         WARN_ON(!ap->num_pages);
2095
2096         /*
2097          * Being under writeback is unlikely but possible.  For example direct
2098          * read to an mmaped fuse file will set the page dirty twice; once when
2099          * the pages are faulted with get_user_pages(), and then after the read
2100          * completed.
2101          */
2102         if (fuse_page_is_writeback(data->inode, page->index))
2103                 return true;
2104
2105         /* Reached max pages */
2106         if (ap->num_pages == fc->max_pages)
2107                 return true;
2108
2109         /* Reached max write bytes */
2110         if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2111                 return true;
2112
2113         /* Discontinuity */
2114         if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2115                 return true;
2116
2117         /* Need to grow the pages array?  If so, did the expansion fail? */
2118         if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2119                 return true;
2120
2121         return false;
2122 }
2123
2124 static int fuse_writepages_fill(struct page *page,
2125                 struct writeback_control *wbc, void *_data)
2126 {
2127         struct fuse_fill_wb_data *data = _data;
2128         struct fuse_writepage_args *wpa = data->wpa;
2129         struct fuse_args_pages *ap = &wpa->ia.ap;
2130         struct inode *inode = data->inode;
2131         struct fuse_inode *fi = get_fuse_inode(inode);
2132         struct fuse_conn *fc = get_fuse_conn(inode);
2133         struct page *tmp_page;
2134         int err;
2135
2136         if (!data->ff) {
2137                 err = -EIO;
2138                 data->ff = fuse_write_file_get(fc, fi);
2139                 if (!data->ff)
2140                         goto out_unlock;
2141         }
2142
2143         if (wpa && fuse_writepage_need_send(fc, page, ap, data)) {
2144                 fuse_writepages_send(data);
2145                 data->wpa = NULL;
2146         }
2147
2148         err = -ENOMEM;
2149         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2150         if (!tmp_page)
2151                 goto out_unlock;
2152
2153         /*
2154          * The page must not be redirtied until the writeout is completed
2155          * (i.e. userspace has sent a reply to the write request).  Otherwise
2156          * there could be more than one temporary page instance for each real
2157          * page.
2158          *
2159          * This is ensured by holding the page lock in page_mkwrite() while
2160          * checking fuse_page_is_writeback().  We already hold the page lock
2161          * since clear_page_dirty_for_io() and keep it held until we add the
2162          * request to the fi->writepages list and increment ap->num_pages.
2163          * After this fuse_page_is_writeback() will indicate that the page is
2164          * under writeback, so we can release the page lock.
2165          */
2166         if (data->wpa == NULL) {
2167                 err = -ENOMEM;
2168                 wpa = fuse_writepage_args_alloc();
2169                 if (!wpa) {
2170                         __free_page(tmp_page);
2171                         goto out_unlock;
2172                 }
2173                 data->max_pages = 1;
2174
2175                 ap = &wpa->ia.ap;
2176                 fuse_write_args_fill(&wpa->ia, data->ff, page_offset(page), 0);
2177                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2178                 wpa->next = NULL;
2179                 ap->args.in_pages = true;
2180                 ap->args.end = fuse_writepage_end;
2181                 ap->num_pages = 0;
2182                 wpa->inode = inode;
2183         }
2184         set_page_writeback(page);
2185
2186         copy_highpage(tmp_page, page);
2187         ap->pages[ap->num_pages] = tmp_page;
2188         ap->descs[ap->num_pages].offset = 0;
2189         ap->descs[ap->num_pages].length = PAGE_SIZE;
2190         data->orig_pages[ap->num_pages] = page;
2191
2192         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2193         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2194
2195         err = 0;
2196         if (data->wpa) {
2197                 /*
2198                  * Protected by fi->lock against concurrent access by
2199                  * fuse_page_is_writeback().
2200                  */
2201                 spin_lock(&fi->lock);
2202                 ap->num_pages++;
2203                 spin_unlock(&fi->lock);
2204         } else if (fuse_writepage_add(wpa, page)) {
2205                 data->wpa = wpa;
2206         } else {
2207                 end_page_writeback(page);
2208         }
2209 out_unlock:
2210         unlock_page(page);
2211
2212         return err;
2213 }
2214
2215 static int fuse_writepages(struct address_space *mapping,
2216                            struct writeback_control *wbc)
2217 {
2218         struct inode *inode = mapping->host;
2219         struct fuse_conn *fc = get_fuse_conn(inode);
2220         struct fuse_fill_wb_data data;
2221         int err;
2222
2223         err = -EIO;
2224         if (fuse_is_bad(inode))
2225                 goto out;
2226
2227         data.inode = inode;
2228         data.wpa = NULL;
2229         data.ff = NULL;
2230
2231         err = -ENOMEM;
2232         data.orig_pages = kcalloc(fc->max_pages,
2233                                   sizeof(struct page *),
2234                                   GFP_NOFS);
2235         if (!data.orig_pages)
2236                 goto out;
2237
2238         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2239         if (data.wpa) {
2240                 WARN_ON(!data.wpa->ia.ap.num_pages);
2241                 fuse_writepages_send(&data);
2242         }
2243         if (data.ff)
2244                 fuse_file_put(data.ff, false, false);
2245
2246         kfree(data.orig_pages);
2247 out:
2248         return err;
2249 }
2250
2251 /*
2252  * It's worthy to make sure that space is reserved on disk for the write,
2253  * but how to implement it without killing performance need more thinking.
2254  */
2255 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2256                 loff_t pos, unsigned len, unsigned flags,
2257                 struct page **pagep, void **fsdata)
2258 {
2259         pgoff_t index = pos >> PAGE_SHIFT;
2260         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2261         struct page *page;
2262         loff_t fsize;
2263         int err = -ENOMEM;
2264
2265         WARN_ON(!fc->writeback_cache);
2266
2267         page = grab_cache_page_write_begin(mapping, index, flags);
2268         if (!page)
2269                 goto error;
2270
2271         fuse_wait_on_page_writeback(mapping->host, page->index);
2272
2273         if (PageUptodate(page) || len == PAGE_SIZE)
2274                 goto success;
2275         /*
2276          * Check if the start this page comes after the end of file, in which
2277          * case the readpage can be optimized away.
2278          */
2279         fsize = i_size_read(mapping->host);
2280         if (fsize <= (pos & PAGE_MASK)) {
2281                 size_t off = pos & ~PAGE_MASK;
2282                 if (off)
2283                         zero_user_segment(page, 0, off);
2284                 goto success;
2285         }
2286         err = fuse_do_readpage(file, page);
2287         if (err)
2288                 goto cleanup;
2289 success:
2290         *pagep = page;
2291         return 0;
2292
2293 cleanup:
2294         unlock_page(page);
2295         put_page(page);
2296 error:
2297         return err;
2298 }
2299
2300 static int fuse_write_end(struct file *file, struct address_space *mapping,
2301                 loff_t pos, unsigned len, unsigned copied,
2302                 struct page *page, void *fsdata)
2303 {
2304         struct inode *inode = page->mapping->host;
2305
2306         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2307         if (!copied)
2308                 goto unlock;
2309
2310         if (!PageUptodate(page)) {
2311                 /* Zero any unwritten bytes at the end of the page */
2312                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2313                 if (endoff)
2314                         zero_user_segment(page, endoff, PAGE_SIZE);
2315                 SetPageUptodate(page);
2316         }
2317
2318         fuse_write_update_size(inode, pos + copied);
2319         set_page_dirty(page);
2320
2321 unlock:
2322         unlock_page(page);
2323         put_page(page);
2324
2325         return copied;
2326 }
2327
2328 static int fuse_launder_page(struct page *page)
2329 {
2330         int err = 0;
2331         if (clear_page_dirty_for_io(page)) {
2332                 struct inode *inode = page->mapping->host;
2333                 err = fuse_writepage_locked(page);
2334                 if (!err)
2335                         fuse_wait_on_page_writeback(inode, page->index);
2336         }
2337         return err;
2338 }
2339
2340 /*
2341  * Write back dirty pages now, because there may not be any suitable
2342  * open files later
2343  */
2344 static void fuse_vma_close(struct vm_area_struct *vma)
2345 {
2346         filemap_write_and_wait(vma->vm_file->f_mapping);
2347 }
2348
2349 /*
2350  * Wait for writeback against this page to complete before allowing it
2351  * to be marked dirty again, and hence written back again, possibly
2352  * before the previous writepage completed.
2353  *
2354  * Block here, instead of in ->writepage(), so that the userspace fs
2355  * can only block processes actually operating on the filesystem.
2356  *
2357  * Otherwise unprivileged userspace fs would be able to block
2358  * unrelated:
2359  *
2360  * - page migration
2361  * - sync(2)
2362  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2363  */
2364 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2365 {
2366         struct page *page = vmf->page;
2367         struct inode *inode = file_inode(vmf->vma->vm_file);
2368
2369         file_update_time(vmf->vma->vm_file);
2370         lock_page(page);
2371         if (page->mapping != inode->i_mapping) {
2372                 unlock_page(page);
2373                 return VM_FAULT_NOPAGE;
2374         }
2375
2376         fuse_wait_on_page_writeback(inode, page->index);
2377         return VM_FAULT_LOCKED;
2378 }
2379
2380 static const struct vm_operations_struct fuse_file_vm_ops = {
2381         .close          = fuse_vma_close,
2382         .fault          = filemap_fault,
2383         .map_pages      = filemap_map_pages,
2384         .page_mkwrite   = fuse_page_mkwrite,
2385 };
2386
2387 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2388 {
2389         struct fuse_file *ff = file->private_data;
2390
2391         /* DAX mmap is superior to direct_io mmap */
2392         if (FUSE_IS_DAX(file_inode(file)))
2393                 return fuse_dax_mmap(file, vma);
2394
2395         if (ff->open_flags & FOPEN_DIRECT_IO) {
2396                 /* Can't provide the coherency needed for MAP_SHARED */
2397                 if (vma->vm_flags & VM_MAYSHARE)
2398                         return -ENODEV;
2399
2400                 invalidate_inode_pages2(file->f_mapping);
2401
2402                 return generic_file_mmap(file, vma);
2403         }
2404
2405         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2406                 fuse_link_write_file(file);
2407
2408         file_accessed(file);
2409         vma->vm_ops = &fuse_file_vm_ops;
2410         return 0;
2411 }
2412
2413 static int convert_fuse_file_lock(struct fuse_conn *fc,
2414                                   const struct fuse_file_lock *ffl,
2415                                   struct file_lock *fl)
2416 {
2417         switch (ffl->type) {
2418         case F_UNLCK:
2419                 break;
2420
2421         case F_RDLCK:
2422         case F_WRLCK:
2423                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2424                     ffl->end < ffl->start)
2425                         return -EIO;
2426
2427                 fl->fl_start = ffl->start;
2428                 fl->fl_end = ffl->end;
2429
2430                 /*
2431                  * Convert pid into init's pid namespace.  The locks API will
2432                  * translate it into the caller's pid namespace.
2433                  */
2434                 rcu_read_lock();
2435                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2436                 rcu_read_unlock();
2437                 break;
2438
2439         default:
2440                 return -EIO;
2441         }
2442         fl->fl_type = ffl->type;
2443         return 0;
2444 }
2445
2446 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2447                          const struct file_lock *fl, int opcode, pid_t pid,
2448                          int flock, struct fuse_lk_in *inarg)
2449 {
2450         struct inode *inode = file_inode(file);
2451         struct fuse_conn *fc = get_fuse_conn(inode);
2452         struct fuse_file *ff = file->private_data;
2453
2454         memset(inarg, 0, sizeof(*inarg));
2455         inarg->fh = ff->fh;
2456         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2457         inarg->lk.start = fl->fl_start;
2458         inarg->lk.end = fl->fl_end;
2459         inarg->lk.type = fl->fl_type;
2460         inarg->lk.pid = pid;
2461         if (flock)
2462                 inarg->lk_flags |= FUSE_LK_FLOCK;
2463         args->opcode = opcode;
2464         args->nodeid = get_node_id(inode);
2465         args->in_numargs = 1;
2466         args->in_args[0].size = sizeof(*inarg);
2467         args->in_args[0].value = inarg;
2468 }
2469
2470 static int fuse_getlk(struct file *file, struct file_lock *fl)
2471 {
2472         struct inode *inode = file_inode(file);
2473         struct fuse_mount *fm = get_fuse_mount(inode);
2474         FUSE_ARGS(args);
2475         struct fuse_lk_in inarg;
2476         struct fuse_lk_out outarg;
2477         int err;
2478
2479         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2480         args.out_numargs = 1;
2481         args.out_args[0].size = sizeof(outarg);
2482         args.out_args[0].value = &outarg;
2483         err = fuse_simple_request(fm, &args);
2484         if (!err)
2485                 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2486
2487         return err;
2488 }
2489
2490 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2491 {
2492         struct inode *inode = file_inode(file);
2493         struct fuse_mount *fm = get_fuse_mount(inode);
2494         FUSE_ARGS(args);
2495         struct fuse_lk_in inarg;
2496         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2497         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2498         pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2499         int err;
2500
2501         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2502                 /* NLM needs asynchronous locks, which we don't support yet */
2503                 return -ENOLCK;
2504         }
2505
2506         /* Unlock on close is handled by the flush method */
2507         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2508                 return 0;
2509
2510         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2511         err = fuse_simple_request(fm, &args);
2512
2513         /* locking is restartable */
2514         if (err == -EINTR)
2515                 err = -ERESTARTSYS;
2516
2517         return err;
2518 }
2519
2520 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2521 {
2522         struct inode *inode = file_inode(file);
2523         struct fuse_conn *fc = get_fuse_conn(inode);
2524         int err;
2525
2526         if (cmd == F_CANCELLK) {
2527                 err = 0;
2528         } else if (cmd == F_GETLK) {
2529                 if (fc->no_lock) {
2530                         posix_test_lock(file, fl);
2531                         err = 0;
2532                 } else
2533                         err = fuse_getlk(file, fl);
2534         } else {
2535                 if (fc->no_lock)
2536                         err = posix_lock_file(file, fl, NULL);
2537                 else
2538                         err = fuse_setlk(file, fl, 0);
2539         }
2540         return err;
2541 }
2542
2543 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2544 {
2545         struct inode *inode = file_inode(file);
2546         struct fuse_conn *fc = get_fuse_conn(inode);
2547         int err;
2548
2549         if (fc->no_flock) {
2550                 err = locks_lock_file_wait(file, fl);
2551         } else {
2552                 struct fuse_file *ff = file->private_data;
2553
2554                 /* emulate flock with POSIX locks */
2555                 ff->flock = true;
2556                 err = fuse_setlk(file, fl, 1);
2557         }
2558
2559         return err;
2560 }
2561
2562 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2563 {
2564         struct inode *inode = mapping->host;
2565         struct fuse_mount *fm = get_fuse_mount(inode);
2566         FUSE_ARGS(args);
2567         struct fuse_bmap_in inarg;
2568         struct fuse_bmap_out outarg;
2569         int err;
2570
2571         if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2572                 return 0;
2573
2574         memset(&inarg, 0, sizeof(inarg));
2575         inarg.block = block;
2576         inarg.blocksize = inode->i_sb->s_blocksize;
2577         args.opcode = FUSE_BMAP;
2578         args.nodeid = get_node_id(inode);
2579         args.in_numargs = 1;
2580         args.in_args[0].size = sizeof(inarg);
2581         args.in_args[0].value = &inarg;
2582         args.out_numargs = 1;
2583         args.out_args[0].size = sizeof(outarg);
2584         args.out_args[0].value = &outarg;
2585         err = fuse_simple_request(fm, &args);
2586         if (err == -ENOSYS)
2587                 fm->fc->no_bmap = 1;
2588
2589         return err ? 0 : outarg.block;
2590 }
2591
2592 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2593 {
2594         struct inode *inode = file->f_mapping->host;
2595         struct fuse_mount *fm = get_fuse_mount(inode);
2596         struct fuse_file *ff = file->private_data;
2597         FUSE_ARGS(args);
2598         struct fuse_lseek_in inarg = {
2599                 .fh = ff->fh,
2600                 .offset = offset,
2601                 .whence = whence
2602         };
2603         struct fuse_lseek_out outarg;
2604         int err;
2605
2606         if (fm->fc->no_lseek)
2607                 goto fallback;
2608
2609         args.opcode = FUSE_LSEEK;
2610         args.nodeid = ff->nodeid;
2611         args.in_numargs = 1;
2612         args.in_args[0].size = sizeof(inarg);
2613         args.in_args[0].value = &inarg;
2614         args.out_numargs = 1;
2615         args.out_args[0].size = sizeof(outarg);
2616         args.out_args[0].value = &outarg;
2617         err = fuse_simple_request(fm, &args);
2618         if (err) {
2619                 if (err == -ENOSYS) {
2620                         fm->fc->no_lseek = 1;
2621                         goto fallback;
2622                 }
2623                 return err;
2624         }
2625
2626         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2627
2628 fallback:
2629         err = fuse_update_attributes(inode, file);
2630         if (!err)
2631                 return generic_file_llseek(file, offset, whence);
2632         else
2633                 return err;
2634 }
2635
2636 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2637 {
2638         loff_t retval;
2639         struct inode *inode = file_inode(file);
2640
2641         switch (whence) {
2642         case SEEK_SET:
2643         case SEEK_CUR:
2644                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2645                 retval = generic_file_llseek(file, offset, whence);
2646                 break;
2647         case SEEK_END:
2648                 inode_lock(inode);
2649                 retval = fuse_update_attributes(inode, file);
2650                 if (!retval)
2651                         retval = generic_file_llseek(file, offset, whence);
2652                 inode_unlock(inode);
2653                 break;
2654         case SEEK_HOLE:
2655         case SEEK_DATA:
2656                 inode_lock(inode);
2657                 retval = fuse_lseek(file, offset, whence);
2658                 inode_unlock(inode);
2659                 break;
2660         default:
2661                 retval = -EINVAL;
2662         }
2663
2664         return retval;
2665 }
2666
2667 /*
2668  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2669  * ABI was defined to be 'struct iovec' which is different on 32bit
2670  * and 64bit.  Fortunately we can determine which structure the server
2671  * used from the size of the reply.
2672  */
2673 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2674                                      size_t transferred, unsigned count,
2675                                      bool is_compat)
2676 {
2677 #ifdef CONFIG_COMPAT
2678         if (count * sizeof(struct compat_iovec) == transferred) {
2679                 struct compat_iovec *ciov = src;
2680                 unsigned i;
2681
2682                 /*
2683                  * With this interface a 32bit server cannot support
2684                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2685                  * requests
2686                  */
2687                 if (!is_compat)
2688                         return -EINVAL;
2689
2690                 for (i = 0; i < count; i++) {
2691                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2692                         dst[i].iov_len = ciov[i].iov_len;
2693                 }
2694                 return 0;
2695         }
2696 #endif
2697
2698         if (count * sizeof(struct iovec) != transferred)
2699                 return -EIO;
2700
2701         memcpy(dst, src, transferred);
2702         return 0;
2703 }
2704
2705 /* Make sure iov_length() won't overflow */
2706 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2707                                  size_t count)
2708 {
2709         size_t n;
2710         u32 max = fc->max_pages << PAGE_SHIFT;
2711
2712         for (n = 0; n < count; n++, iov++) {
2713                 if (iov->iov_len > (size_t) max)
2714                         return -ENOMEM;
2715                 max -= iov->iov_len;
2716         }
2717         return 0;
2718 }
2719
2720 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2721                                  void *src, size_t transferred, unsigned count,
2722                                  bool is_compat)
2723 {
2724         unsigned i;
2725         struct fuse_ioctl_iovec *fiov = src;
2726
2727         if (fc->minor < 16) {
2728                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2729                                                  count, is_compat);
2730         }
2731
2732         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2733                 return -EIO;
2734
2735         for (i = 0; i < count; i++) {
2736                 /* Did the server supply an inappropriate value? */
2737                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2738                     fiov[i].len != (unsigned long) fiov[i].len)
2739                         return -EIO;
2740
2741                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2742                 dst[i].iov_len = (size_t) fiov[i].len;
2743
2744 #ifdef CONFIG_COMPAT
2745                 if (is_compat &&
2746                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2747                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2748                         return -EIO;
2749 #endif
2750         }
2751
2752         return 0;
2753 }
2754
2755
2756 /*
2757  * For ioctls, there is no generic way to determine how much memory
2758  * needs to be read and/or written.  Furthermore, ioctls are allowed
2759  * to dereference the passed pointer, so the parameter requires deep
2760  * copying but FUSE has no idea whatsoever about what to copy in or
2761  * out.
2762  *
2763  * This is solved by allowing FUSE server to retry ioctl with
2764  * necessary in/out iovecs.  Let's assume the ioctl implementation
2765  * needs to read in the following structure.
2766  *
2767  * struct a {
2768  *      char    *buf;
2769  *      size_t  buflen;
2770  * }
2771  *
2772  * On the first callout to FUSE server, inarg->in_size and
2773  * inarg->out_size will be NULL; then, the server completes the ioctl
2774  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2775  * the actual iov array to
2776  *
2777  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2778  *
2779  * which tells FUSE to copy in the requested area and retry the ioctl.
2780  * On the second round, the server has access to the structure and
2781  * from that it can tell what to look for next, so on the invocation,
2782  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2783  *
2784  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2785  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2786  *
2787  * FUSE will copy both struct a and the pointed buffer from the
2788  * process doing the ioctl and retry ioctl with both struct a and the
2789  * buffer.
2790  *
2791  * This time, FUSE server has everything it needs and completes ioctl
2792  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2793  *
2794  * Copying data out works the same way.
2795  *
2796  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2797  * automatically initializes in and out iovs by decoding @cmd with
2798  * _IOC_* macros and the server is not allowed to request RETRY.  This
2799  * limits ioctl data transfers to well-formed ioctls and is the forced
2800  * behavior for all FUSE servers.
2801  */
2802 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2803                    unsigned int flags)
2804 {
2805         struct fuse_file *ff = file->private_data;
2806         struct fuse_mount *fm = ff->fm;
2807         struct fuse_ioctl_in inarg = {
2808                 .fh = ff->fh,
2809                 .cmd = cmd,
2810                 .arg = arg,
2811                 .flags = flags
2812         };
2813         struct fuse_ioctl_out outarg;
2814         struct iovec *iov_page = NULL;
2815         struct iovec *in_iov = NULL, *out_iov = NULL;
2816         unsigned int in_iovs = 0, out_iovs = 0, max_pages;
2817         size_t in_size, out_size, c;
2818         ssize_t transferred;
2819         int err, i;
2820         struct iov_iter ii;
2821         struct fuse_args_pages ap = {};
2822
2823 #if BITS_PER_LONG == 32
2824         inarg.flags |= FUSE_IOCTL_32BIT;
2825 #else
2826         if (flags & FUSE_IOCTL_COMPAT) {
2827                 inarg.flags |= FUSE_IOCTL_32BIT;
2828 #ifdef CONFIG_X86_X32
2829                 if (in_x32_syscall())
2830                         inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2831 #endif
2832         }
2833 #endif
2834
2835         /* assume all the iovs returned by client always fits in a page */
2836         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2837
2838         err = -ENOMEM;
2839         ap.pages = fuse_pages_alloc(fm->fc->max_pages, GFP_KERNEL, &ap.descs);
2840         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2841         if (!ap.pages || !iov_page)
2842                 goto out;
2843
2844         fuse_page_descs_length_init(ap.descs, 0, fm->fc->max_pages);
2845
2846         /*
2847          * If restricted, initialize IO parameters as encoded in @cmd.
2848          * RETRY from server is not allowed.
2849          */
2850         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2851                 struct iovec *iov = iov_page;
2852
2853                 iov->iov_base = (void __user *)arg;
2854
2855                 switch (cmd) {
2856                 case FS_IOC_GETFLAGS:
2857                 case FS_IOC_SETFLAGS:
2858                         iov->iov_len = sizeof(int);
2859                         break;
2860                 default:
2861                         iov->iov_len = _IOC_SIZE(cmd);
2862                         break;
2863                 }
2864
2865                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2866                         in_iov = iov;
2867                         in_iovs = 1;
2868                 }
2869
2870                 if (_IOC_DIR(cmd) & _IOC_READ) {
2871                         out_iov = iov;
2872                         out_iovs = 1;
2873                 }
2874         }
2875
2876  retry:
2877         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2878         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2879
2880         /*
2881          * Out data can be used either for actual out data or iovs,
2882          * make sure there always is at least one page.
2883          */
2884         out_size = max_t(size_t, out_size, PAGE_SIZE);
2885         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2886
2887         /* make sure there are enough buffer pages and init request with them */
2888         err = -ENOMEM;
2889         if (max_pages > fm->fc->max_pages)
2890                 goto out;
2891         while (ap.num_pages < max_pages) {
2892                 ap.pages[ap.num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2893                 if (!ap.pages[ap.num_pages])
2894                         goto out;
2895                 ap.num_pages++;
2896         }
2897
2898
2899         /* okay, let's send it to the client */
2900         ap.args.opcode = FUSE_IOCTL;
2901         ap.args.nodeid = ff->nodeid;
2902         ap.args.in_numargs = 1;
2903         ap.args.in_args[0].size = sizeof(inarg);
2904         ap.args.in_args[0].value = &inarg;
2905         if (in_size) {
2906                 ap.args.in_numargs++;
2907                 ap.args.in_args[1].size = in_size;
2908                 ap.args.in_pages = true;
2909
2910                 err = -EFAULT;
2911                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2912                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2913                         c = copy_page_from_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2914                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2915                                 goto out;
2916                 }
2917         }
2918
2919         ap.args.out_numargs = 2;
2920         ap.args.out_args[0].size = sizeof(outarg);
2921         ap.args.out_args[0].value = &outarg;
2922         ap.args.out_args[1].size = out_size;
2923         ap.args.out_pages = true;
2924         ap.args.out_argvar = true;
2925
2926         transferred = fuse_simple_request(fm, &ap.args);
2927         err = transferred;
2928         if (transferred < 0)
2929                 goto out;
2930
2931         /* did it ask for retry? */
2932         if (outarg.flags & FUSE_IOCTL_RETRY) {
2933                 void *vaddr;
2934
2935                 /* no retry if in restricted mode */
2936                 err = -EIO;
2937                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2938                         goto out;
2939
2940                 in_iovs = outarg.in_iovs;
2941                 out_iovs = outarg.out_iovs;
2942
2943                 /*
2944                  * Make sure things are in boundary, separate checks
2945                  * are to protect against overflow.
2946                  */
2947                 err = -ENOMEM;
2948                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2949                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2950                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2951                         goto out;
2952
2953                 vaddr = kmap_atomic(ap.pages[0]);
2954                 err = fuse_copy_ioctl_iovec(fm->fc, iov_page, vaddr,
2955                                             transferred, in_iovs + out_iovs,
2956                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2957                 kunmap_atomic(vaddr);
2958                 if (err)
2959                         goto out;
2960
2961                 in_iov = iov_page;
2962                 out_iov = in_iov + in_iovs;
2963
2964                 err = fuse_verify_ioctl_iov(fm->fc, in_iov, in_iovs);
2965                 if (err)
2966                         goto out;
2967
2968                 err = fuse_verify_ioctl_iov(fm->fc, out_iov, out_iovs);
2969                 if (err)
2970                         goto out;
2971
2972                 goto retry;
2973         }
2974
2975         err = -EIO;
2976         if (transferred > inarg.out_size)
2977                 goto out;
2978
2979         err = -EFAULT;
2980         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2981         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2982                 c = copy_page_to_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2983                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2984                         goto out;
2985         }
2986         err = 0;
2987  out:
2988         free_page((unsigned long) iov_page);
2989         while (ap.num_pages)
2990                 __free_page(ap.pages[--ap.num_pages]);
2991         kfree(ap.pages);
2992
2993         return err ? err : outarg.result;
2994 }
2995 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2996
2997 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2998                        unsigned long arg, unsigned int flags)
2999 {
3000         struct inode *inode = file_inode(file);
3001         struct fuse_conn *fc = get_fuse_conn(inode);
3002
3003         if (!fuse_allow_current_process(fc))
3004                 return -EACCES;
3005
3006         if (fuse_is_bad(inode))
3007                 return -EIO;
3008
3009         return fuse_do_ioctl(file, cmd, arg, flags);
3010 }
3011
3012 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
3013                             unsigned long arg)
3014 {
3015         return fuse_ioctl_common(file, cmd, arg, 0);
3016 }
3017
3018 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
3019                                    unsigned long arg)
3020 {
3021         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
3022 }
3023
3024 /*
3025  * All files which have been polled are linked to RB tree
3026  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
3027  * find the matching one.
3028  */
3029 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
3030                                               struct rb_node **parent_out)
3031 {
3032         struct rb_node **link = &fc->polled_files.rb_node;
3033         struct rb_node *last = NULL;
3034
3035         while (*link) {
3036                 struct fuse_file *ff;
3037
3038                 last = *link;
3039                 ff = rb_entry(last, struct fuse_file, polled_node);
3040
3041                 if (kh < ff->kh)
3042                         link = &last->rb_left;
3043                 else if (kh > ff->kh)
3044                         link = &last->rb_right;
3045                 else
3046                         return link;
3047         }
3048
3049         if (parent_out)
3050                 *parent_out = last;
3051         return link;
3052 }
3053
3054 /*
3055  * The file is about to be polled.  Make sure it's on the polled_files
3056  * RB tree.  Note that files once added to the polled_files tree are
3057  * not removed before the file is released.  This is because a file
3058  * polled once is likely to be polled again.
3059  */
3060 static void fuse_register_polled_file(struct fuse_conn *fc,
3061                                       struct fuse_file *ff)
3062 {
3063         spin_lock(&fc->lock);
3064         if (RB_EMPTY_NODE(&ff->polled_node)) {
3065                 struct rb_node **link, *parent;
3066
3067                 link = fuse_find_polled_node(fc, ff->kh, &parent);
3068                 BUG_ON(*link);
3069                 rb_link_node(&ff->polled_node, parent, link);
3070                 rb_insert_color(&ff->polled_node, &fc->polled_files);
3071         }
3072         spin_unlock(&fc->lock);
3073 }
3074
3075 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
3076 {
3077         struct fuse_file *ff = file->private_data;
3078         struct fuse_mount *fm = ff->fm;
3079         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
3080         struct fuse_poll_out outarg;
3081         FUSE_ARGS(args);
3082         int err;
3083
3084         if (fm->fc->no_poll)
3085                 return DEFAULT_POLLMASK;
3086
3087         poll_wait(file, &ff->poll_wait, wait);
3088         inarg.events = mangle_poll(poll_requested_events(wait));
3089
3090         /*
3091          * Ask for notification iff there's someone waiting for it.
3092          * The client may ignore the flag and always notify.
3093          */
3094         if (waitqueue_active(&ff->poll_wait)) {
3095                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
3096                 fuse_register_polled_file(fm->fc, ff);
3097         }
3098
3099         args.opcode = FUSE_POLL;
3100         args.nodeid = ff->nodeid;
3101         args.in_numargs = 1;
3102         args.in_args[0].size = sizeof(inarg);
3103         args.in_args[0].value = &inarg;
3104         args.out_numargs = 1;
3105         args.out_args[0].size = sizeof(outarg);
3106         args.out_args[0].value = &outarg;
3107         err = fuse_simple_request(fm, &args);
3108
3109         if (!err)
3110                 return demangle_poll(outarg.revents);
3111         if (err == -ENOSYS) {
3112                 fm->fc->no_poll = 1;
3113                 return DEFAULT_POLLMASK;
3114         }
3115         return EPOLLERR;
3116 }
3117 EXPORT_SYMBOL_GPL(fuse_file_poll);
3118
3119 /*
3120  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
3121  * wakes up the poll waiters.
3122  */
3123 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
3124                             struct fuse_notify_poll_wakeup_out *outarg)
3125 {
3126         u64 kh = outarg->kh;
3127         struct rb_node **link;
3128
3129         spin_lock(&fc->lock);
3130
3131         link = fuse_find_polled_node(fc, kh, NULL);
3132         if (*link) {
3133                 struct fuse_file *ff;
3134
3135                 ff = rb_entry(*link, struct fuse_file, polled_node);
3136                 wake_up_interruptible_sync(&ff->poll_wait);
3137         }
3138
3139         spin_unlock(&fc->lock);
3140         return 0;
3141 }
3142
3143 static void fuse_do_truncate(struct file *file)
3144 {
3145         struct inode *inode = file->f_mapping->host;
3146         struct iattr attr;
3147
3148         attr.ia_valid = ATTR_SIZE;
3149         attr.ia_size = i_size_read(inode);
3150
3151         attr.ia_file = file;
3152         attr.ia_valid |= ATTR_FILE;
3153
3154         fuse_do_setattr(file_dentry(file), &attr, file);
3155 }
3156
3157 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
3158 {
3159         return round_up(off, fc->max_pages << PAGE_SHIFT);
3160 }
3161
3162 static ssize_t
3163 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3164 {
3165         DECLARE_COMPLETION_ONSTACK(wait);
3166         ssize_t ret = 0;
3167         struct file *file = iocb->ki_filp;
3168         struct fuse_file *ff = file->private_data;
3169         loff_t pos = 0;
3170         struct inode *inode;
3171         loff_t i_size;
3172         size_t count = iov_iter_count(iter), shortened = 0;
3173         loff_t offset = iocb->ki_pos;
3174         struct fuse_io_priv *io;
3175
3176         pos = offset;
3177         inode = file->f_mapping->host;
3178         i_size = i_size_read(inode);
3179
3180         if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
3181                 return 0;
3182
3183         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
3184         if (!io)
3185                 return -ENOMEM;
3186         spin_lock_init(&io->lock);
3187         kref_init(&io->refcnt);
3188         io->reqs = 1;
3189         io->bytes = -1;
3190         io->size = 0;
3191         io->offset = offset;
3192         io->write = (iov_iter_rw(iter) == WRITE);
3193         io->err = 0;
3194         /*
3195          * By default, we want to optimize all I/Os with async request
3196          * submission to the client filesystem if supported.
3197          */
3198         io->async = ff->fm->fc->async_dio;
3199         io->iocb = iocb;
3200         io->blocking = is_sync_kiocb(iocb);
3201
3202         /* optimization for short read */
3203         if (io->async && !io->write && offset + count > i_size) {
3204                 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
3205                 shortened = count - iov_iter_count(iter);
3206                 count -= shortened;
3207         }
3208
3209         /*
3210          * We cannot asynchronously extend the size of a file.
3211          * In such case the aio will behave exactly like sync io.
3212          */
3213         if ((offset + count > i_size) && io->write)
3214                 io->blocking = true;
3215
3216         if (io->async && io->blocking) {
3217                 /*
3218                  * Additional reference to keep io around after
3219                  * calling fuse_aio_complete()
3220                  */
3221                 kref_get(&io->refcnt);
3222                 io->done = &wait;
3223         }
3224
3225         if (iov_iter_rw(iter) == WRITE) {
3226                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
3227                 fuse_invalidate_attr(inode);
3228         } else {
3229                 ret = __fuse_direct_read(io, iter, &pos);
3230         }
3231         iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
3232
3233         if (io->async) {
3234                 bool blocking = io->blocking;
3235
3236                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3237
3238                 /* we have a non-extending, async request, so return */
3239                 if (!blocking)
3240                         return -EIOCBQUEUED;
3241
3242                 wait_for_completion(&wait);
3243                 ret = fuse_get_res_by_io(io);
3244         }
3245
3246         kref_put(&io->refcnt, fuse_io_release);
3247
3248         if (iov_iter_rw(iter) == WRITE) {
3249                 if (ret > 0)
3250                         fuse_write_update_size(inode, pos);
3251                 else if (ret < 0 && offset + count > i_size)
3252                         fuse_do_truncate(file);
3253         }
3254
3255         return ret;
3256 }
3257
3258 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3259 {
3260         int err = filemap_write_and_wait_range(inode->i_mapping, start, LLONG_MAX);
3261
3262         if (!err)
3263                 fuse_sync_writes(inode);
3264
3265         return err;
3266 }
3267
3268 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3269                                 loff_t length)
3270 {
3271         struct fuse_file *ff = file->private_data;
3272         struct inode *inode = file_inode(file);
3273         struct fuse_inode *fi = get_fuse_inode(inode);
3274         struct fuse_mount *fm = ff->fm;
3275         FUSE_ARGS(args);
3276         struct fuse_fallocate_in inarg = {
3277                 .fh = ff->fh,
3278                 .offset = offset,
3279                 .length = length,
3280                 .mode = mode
3281         };
3282         int err;
3283         bool block_faults = FUSE_IS_DAX(inode) &&
3284                 (!(mode & FALLOC_FL_KEEP_SIZE) ||
3285                  (mode & FALLOC_FL_PUNCH_HOLE));
3286
3287         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3288                 return -EOPNOTSUPP;
3289
3290         if (fm->fc->no_fallocate)
3291                 return -EOPNOTSUPP;
3292
3293         inode_lock(inode);
3294         if (block_faults) {
3295                 down_write(&fi->i_mmap_sem);
3296                 err = fuse_dax_break_layouts(inode, 0, 0);
3297                 if (err)
3298                         goto out;
3299         }
3300
3301         if (mode & FALLOC_FL_PUNCH_HOLE) {
3302                 loff_t endbyte = offset + length - 1;
3303
3304                 err = fuse_writeback_range(inode, offset, endbyte);
3305                 if (err)
3306                         goto out;
3307         }
3308
3309         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3310             offset + length > i_size_read(inode)) {
3311                 err = inode_newsize_ok(inode, offset + length);
3312                 if (err)
3313                         goto out;
3314         }
3315
3316         err = file_modified(file);
3317         if (err)
3318                 goto out;
3319
3320         if (!(mode & FALLOC_FL_KEEP_SIZE))
3321                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3322
3323         args.opcode = FUSE_FALLOCATE;
3324         args.nodeid = ff->nodeid;
3325         args.in_numargs = 1;
3326         args.in_args[0].size = sizeof(inarg);
3327         args.in_args[0].value = &inarg;
3328         err = fuse_simple_request(fm, &args);
3329         if (err == -ENOSYS) {
3330                 fm->fc->no_fallocate = 1;
3331                 err = -EOPNOTSUPP;
3332         }
3333         if (err)
3334                 goto out;
3335
3336         /* we could have extended the file */
3337         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3338                 bool changed = fuse_write_update_size(inode, offset + length);
3339
3340                 if (changed && fm->fc->writeback_cache)
3341                         file_update_time(file);
3342         }
3343
3344         if (mode & FALLOC_FL_PUNCH_HOLE)
3345                 truncate_pagecache_range(inode, offset, offset + length - 1);
3346
3347         fuse_invalidate_attr(inode);
3348
3349 out:
3350         if (!(mode & FALLOC_FL_KEEP_SIZE))
3351                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3352
3353         if (block_faults)
3354                 up_write(&fi->i_mmap_sem);
3355
3356         inode_unlock(inode);
3357
3358         fuse_flush_time_update(inode);
3359
3360         return err;
3361 }
3362
3363 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3364                                       struct file *file_out, loff_t pos_out,
3365                                       size_t len, unsigned int flags)
3366 {
3367         struct fuse_file *ff_in = file_in->private_data;
3368         struct fuse_file *ff_out = file_out->private_data;
3369         struct inode *inode_in = file_inode(file_in);
3370         struct inode *inode_out = file_inode(file_out);
3371         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3372         struct fuse_mount *fm = ff_in->fm;
3373         struct fuse_conn *fc = fm->fc;
3374         FUSE_ARGS(args);
3375         struct fuse_copy_file_range_in inarg = {
3376                 .fh_in = ff_in->fh,
3377                 .off_in = pos_in,
3378                 .nodeid_out = ff_out->nodeid,
3379                 .fh_out = ff_out->fh,
3380                 .off_out = pos_out,
3381                 .len = len,
3382                 .flags = flags
3383         };
3384         struct fuse_write_out outarg;
3385         ssize_t err;
3386         /* mark unstable when write-back is not used, and file_out gets
3387          * extended */
3388         bool is_unstable = (!fc->writeback_cache) &&
3389                            ((pos_out + len) > inode_out->i_size);
3390
3391         if (fc->no_copy_file_range)
3392                 return -EOPNOTSUPP;
3393
3394         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3395                 return -EXDEV;
3396
3397         inode_lock(inode_in);
3398         err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3399         inode_unlock(inode_in);
3400         if (err)
3401                 return err;
3402
3403         inode_lock(inode_out);
3404
3405         err = file_modified(file_out);
3406         if (err)
3407                 goto out;
3408
3409         /*
3410          * Write out dirty pages in the destination file before sending the COPY
3411          * request to userspace.  After the request is completed, truncate off
3412          * pages (including partial ones) from the cache that have been copied,
3413          * since these contain stale data at that point.
3414          *
3415          * This should be mostly correct, but if the COPY writes to partial
3416          * pages (at the start or end) and the parts not covered by the COPY are
3417          * written through a memory map after calling fuse_writeback_range(),
3418          * then these partial page modifications will be lost on truncation.
3419          *
3420          * It is unlikely that someone would rely on such mixed style
3421          * modifications.  Yet this does give less guarantees than if the
3422          * copying was performed with write(2).
3423          *
3424          * To fix this a i_mmap_sem style lock could be used to prevent new
3425          * faults while the copy is ongoing.
3426          */
3427         err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3428         if (err)
3429                 goto out;
3430
3431         if (is_unstable)
3432                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3433
3434         args.opcode = FUSE_COPY_FILE_RANGE;
3435         args.nodeid = ff_in->nodeid;
3436         args.in_numargs = 1;
3437         args.in_args[0].size = sizeof(inarg);
3438         args.in_args[0].value = &inarg;
3439         args.out_numargs = 1;
3440         args.out_args[0].size = sizeof(outarg);
3441         args.out_args[0].value = &outarg;
3442         err = fuse_simple_request(fm, &args);
3443         if (err == -ENOSYS) {
3444                 fc->no_copy_file_range = 1;
3445                 err = -EOPNOTSUPP;
3446         }
3447         if (err)
3448                 goto out;
3449
3450         truncate_inode_pages_range(inode_out->i_mapping,
3451                                    ALIGN_DOWN(pos_out, PAGE_SIZE),
3452                                    ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3453
3454         if (fc->writeback_cache) {
3455                 fuse_write_update_size(inode_out, pos_out + outarg.size);
3456                 file_update_time(file_out);
3457         }
3458
3459         fuse_invalidate_attr(inode_out);
3460
3461         err = outarg.size;
3462 out:
3463         if (is_unstable)
3464                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3465
3466         inode_unlock(inode_out);
3467         file_accessed(file_in);
3468
3469         fuse_flush_time_update(inode_out);
3470
3471         return err;
3472 }
3473
3474 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3475                                     struct file *dst_file, loff_t dst_off,
3476                                     size_t len, unsigned int flags)
3477 {
3478         ssize_t ret;
3479
3480         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3481                                      len, flags);
3482
3483         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3484                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3485                                               dst_off, len, flags);
3486         return ret;
3487 }
3488
3489 static const struct file_operations fuse_file_operations = {
3490         .llseek         = fuse_file_llseek,
3491         .read_iter      = fuse_file_read_iter,
3492         .write_iter     = fuse_file_write_iter,
3493         .mmap           = fuse_file_mmap,
3494         .open           = fuse_open,
3495         .flush          = fuse_flush,
3496         .release        = fuse_release,
3497         .fsync          = fuse_fsync,
3498         .lock           = fuse_file_lock,
3499         .get_unmapped_area = thp_get_unmapped_area,
3500         .flock          = fuse_file_flock,
3501         .splice_read    = generic_file_splice_read,
3502         .splice_write   = iter_file_splice_write,
3503         .unlocked_ioctl = fuse_file_ioctl,
3504         .compat_ioctl   = fuse_file_compat_ioctl,
3505         .poll           = fuse_file_poll,
3506         .fallocate      = fuse_file_fallocate,
3507         .copy_file_range = fuse_copy_file_range,
3508 };
3509
3510 static const struct address_space_operations fuse_file_aops  = {
3511         .readpage       = fuse_readpage,
3512         .readahead      = fuse_readahead,
3513         .writepage      = fuse_writepage,
3514         .writepages     = fuse_writepages,
3515         .launder_page   = fuse_launder_page,
3516         .set_page_dirty = __set_page_dirty_nobuffers,
3517         .bmap           = fuse_bmap,
3518         .direct_IO      = fuse_direct_IO,
3519         .write_begin    = fuse_write_begin,
3520         .write_end      = fuse_write_end,
3521 };
3522
3523 void fuse_init_file_inode(struct inode *inode)
3524 {
3525         struct fuse_inode *fi = get_fuse_inode(inode);
3526
3527         inode->i_fop = &fuse_file_operations;
3528         inode->i_data.a_ops = &fuse_file_aops;
3529
3530         INIT_LIST_HEAD(&fi->write_files);
3531         INIT_LIST_HEAD(&fi->queued_writes);
3532         fi->writectr = 0;
3533         init_waitqueue_head(&fi->page_waitq);
3534         fi->writepages = RB_ROOT;
3535
3536         if (IS_ENABLED(CONFIG_FUSE_DAX))
3537                 fuse_dax_inode_init(inode);
3538 }