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
  46         if (fm->fc->handle_killpriv_v2 &&
  47             (inarg.flags & O_TRUNC) && !capable(CAP_FSETID)) {
  48                 inarg.open_flags |= FUSE_OPEN_KILL_SUIDGID;
  49         }
  50
  51         args.opcode = opcode;
  52         args.nodeid = nodeid;
  53         args.in_numargs = 1;
  54         args.in_args[0].size = sizeof(inarg);
  55         args.in_args[0].value = &inarg;
  56         args.out_numargs = 1;
  57         args.out_args[0].size = sizeof(*outargp);
  58         args.out_args[0].value = outargp;
  59
  60         return fuse_simple_request(fm, &args);
  61 }
  62
  63 struct fuse_release_args {
  64         struct fuse_args args;
  65         struct fuse_release_in inarg;
  66         struct inode *inode;
  67 };
  68
  69 struct fuse_file *fuse_file_alloc(struct fuse_mount *fm)
  70 {
  71         struct fuse_file *ff;
  72
  73         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
  74         if (unlikely(!ff))
  75                 return NULL;
  76
  77         ff->fm = fm;
  78         ff->release_args = kzalloc(sizeof(*ff->release_args),
  79                                    GFP_KERNEL_ACCOUNT);
  80         if (!ff->release_args) {
  81                 kfree(ff);
  82                 return NULL;
  83         }
  84
  85         INIT_LIST_HEAD(&ff->write_entry);
  86         mutex_init(&ff->readdir.lock);
  87         refcount_set(&ff->count, 1);
  88         RB_CLEAR_NODE(&ff->polled_node);
  89         init_waitqueue_head(&ff->poll_wait);
  90
  91         ff->kh = atomic64_inc_return(&fm->fc->khctr);
  92
  93         return ff;
  94 }
  95
  96 void fuse_file_free(struct fuse_file *ff)
  97 {
  98         kfree(ff->release_args);
  99         mutex_destroy(&ff->readdir.lock);
 100         kfree(ff);
 101 }
 102
 103 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
 104 {
 105         refcount_inc(&ff->count);
 106         return ff;
 107 }
 108
 109 static void fuse_release_end(struct fuse_mount *fm, struct fuse_args *args,
 110                              int error)
 111 {
 112         struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
 113
 114         iput(ra->inode);
 115         kfree(ra);
 116 }
 117
 118 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
 119 {
 120         if (refcount_dec_and_test(&ff->count)) {
 121                 struct fuse_args *args = &ff->release_args->args;
 122
 123                 if (isdir ? ff->fm->fc->no_opendir : ff->fm->fc->no_open) {
 124                         /* Do nothing when client does not implement 'open' */
 125                         fuse_release_end(ff->fm, args, 0);
 126                 } else if (sync) {
 127                         fuse_simple_request(ff->fm, args);
 128                         fuse_release_end(ff->fm, args, 0);
 129                 } else {
 130                         args->end = fuse_release_end;
 131                         if (fuse_simple_background(ff->fm, args,
 132                                                    GFP_KERNEL | __GFP_NOFAIL))
 133                                 fuse_release_end(ff->fm, args, -ENOTCONN);
 134                 }
 135                 kfree(ff);
 136         }
 137 }
 138
 139 int fuse_do_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
 140                  bool isdir)
 141 {
 142         struct fuse_conn *fc = fm->fc;
 143         struct fuse_file *ff;
 144         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 145
 146         ff = fuse_file_alloc(fm);
 147         if (!ff)
 148                 return -ENOMEM;
 149
 150         ff->fh = 0;
 151         /* Default for no-open */
 152         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
 153         if (isdir ? !fc->no_opendir : !fc->no_open) {
 154                 struct fuse_open_out outarg;
 155                 int err;
 156
 157                 err = fuse_send_open(fm, nodeid, file, opcode, &outarg);
 158                 if (!err) {
 159                         ff->fh = outarg.fh;
 160                         ff->open_flags = outarg.open_flags;
 161
 162                 } else if (err != -ENOSYS) {
 163                         fuse_file_free(ff);
 164                         return err;
 165                 } else {
 166                         if (isdir)
 167                                 fc->no_opendir = 1;
 168                         else
 169                                 fc->no_open = 1;
 170                 }
 171         }
 172
 173         if (isdir)
 174                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 175
 176         ff->nodeid = nodeid;
 177         file->private_data = ff;
 178
 179         return 0;
 180 }
 181 EXPORT_SYMBOL_GPL(fuse_do_open);
 182
 183 static void fuse_link_write_file(struct file *file)
 184 {
 185         struct inode *inode = file_inode(file);
 186         struct fuse_inode *fi = get_fuse_inode(inode);
 187         struct fuse_file *ff = file->private_data;
 188         /*
 189          * file may be written through mmap, so chain it onto the
 190          * inodes's write_file list
 191          */
 192         spin_lock(&fi->lock);
 193         if (list_empty(&ff->write_entry))
 194                 list_add(&ff->write_entry, &fi->write_files);
 195         spin_unlock(&fi->lock);
 196 }
 197
 198 void fuse_finish_open(struct inode *inode, struct file *file)
 199 {
 200         struct fuse_file *ff = file->private_data;
 201         struct fuse_conn *fc = get_fuse_conn(inode);
 202
 203         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 204                 invalidate_inode_pages2(inode->i_mapping);
 205         if (ff->open_flags & FOPEN_STREAM)
 206                 stream_open(inode, file);
 207         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 208                 nonseekable_open(inode, file);
 209         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 210                 struct fuse_inode *fi = get_fuse_inode(inode);
 211
 212                 spin_lock(&fi->lock);
 213                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 214                 i_size_write(inode, 0);
 215                 spin_unlock(&fi->lock);
 216                 fuse_invalidate_attr(inode);
 217                 if (fc->writeback_cache)
 218                         file_update_time(file);
 219         }
 220         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 221                 fuse_link_write_file(file);
 222 }
 223
 224 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 225 {
 226         struct fuse_mount *fm = get_fuse_mount(inode);
 227         struct fuse_conn *fc = fm->fc;
 228         int err;
 229         bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
 230                           fc->atomic_o_trunc &&
 231                           fc->writeback_cache;
 232         bool dax_truncate = (file->f_flags & O_TRUNC) &&
 233                           fc->atomic_o_trunc && FUSE_IS_DAX(inode);
 234
 235         if (fuse_is_bad(inode))
 236                 return -EIO;
 237
 238         err = generic_file_open(inode, file);
 239         if (err)
 240                 return err;
 241
 242         if (is_wb_truncate || dax_truncate) {
 243                 inode_lock(inode);
 244                 fuse_set_nowrite(inode);
 245         }
 246
 247         if (dax_truncate) {
 248                 down_write(&get_fuse_inode(inode)->i_mmap_sem);
 249                 err = fuse_dax_break_layouts(inode, 0, 0);
 250                 if (err)
 251                         goto out;
 252         }
 253
 254         err = fuse_do_open(fm, get_node_id(inode), file, isdir);
 255         if (!err)
 256                 fuse_finish_open(inode, file);
 257
 258 out:
 259         if (dax_truncate)
 260                 up_write(&get_fuse_inode(inode)->i_mmap_sem);
 261
 262         if (is_wb_truncate | dax_truncate) {
 263                 fuse_release_nowrite(inode);
 264                 inode_unlock(inode);
 265         }
 266
 267         return err;
 268 }
 269
 270 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 271                                  int flags, int opcode)
 272 {
 273         struct fuse_conn *fc = ff->fm->fc;
 274         struct fuse_release_args *ra = ff->release_args;
 275
 276         /* Inode is NULL on error path of fuse_create_open() */
 277         if (likely(fi)) {
 278                 spin_lock(&fi->lock);
 279                 list_del(&ff->write_entry);
 280                 spin_unlock(&fi->lock);
 281         }
 282         spin_lock(&fc->lock);
 283         if (!RB_EMPTY_NODE(&ff->polled_node))
 284                 rb_erase(&ff->polled_node, &fc->polled_files);
 285         spin_unlock(&fc->lock);
 286
 287         wake_up_interruptible_all(&ff->poll_wait);
 288
 289         ra->inarg.fh = ff->fh;
 290         ra->inarg.flags = flags;
 291         ra->args.in_numargs = 1;
 292         ra->args.in_args[0].size = sizeof(struct fuse_release_in);
 293         ra->args.in_args[0].value = &ra->inarg;
 294         ra->args.opcode = opcode;
 295         ra->args.nodeid = ff->nodeid;
 296         ra->args.force = true;
 297         ra->args.nocreds = true;
 298 }
 299
 300 void fuse_release_common(struct file *file, bool isdir)
 301 {
 302         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 303         struct fuse_file *ff = file->private_data;
 304         struct fuse_release_args *ra = ff->release_args;
 305         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 306
 307         fuse_prepare_release(fi, ff, file->f_flags, opcode);
 308
 309         if (ff->flock) {
 310                 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 311                 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fm->fc,
 312                                                           (fl_owner_t) file);
 313         }
 314         /* Hold inode until release is finished */
 315         ra->inode = igrab(file_inode(file));
 316
 317         /*
 318          * Normally this will send the RELEASE request, however if
 319          * some asynchronous READ or WRITE requests are outstanding,
 320          * the sending will be delayed.
 321          *
 322          * Make the release synchronous if this is a fuseblk mount,
 323          * synchronous RELEASE is allowed (and desirable) in this case
 324          * because the server can be trusted not to screw up.
 325          */
 326         fuse_file_put(ff, ff->fm->fc->destroy, isdir);
 327 }
 328
 329 static int fuse_open(struct inode *inode, struct file *file)
 330 {
 331         return fuse_open_common(inode, file, false);
 332 }
 333
 334 static int fuse_release(struct inode *inode, struct file *file)
 335 {
 336         struct fuse_conn *fc = get_fuse_conn(inode);
 337
 338         /* see fuse_vma_close() for !writeback_cache case */
 339         if (fc->writeback_cache)
 340                 write_inode_now(inode, 1);
 341
 342         fuse_release_common(file, false);
 343
 344         /* return value is ignored by VFS */
 345         return 0;
 346 }
 347
 348 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
 349 {
 350         WARN_ON(refcount_read(&ff->count) > 1);
 351         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 352         /*
 353          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 354          * synchronous, we are fine with not doing igrab() here"
 355          */
 356         fuse_file_put(ff, true, false);
 357 }
 358 EXPORT_SYMBOL_GPL(fuse_sync_release);
 359
 360 /*
 361  * Scramble the ID space with XTEA, so that the value of the files_struct
 362  * pointer is not exposed to userspace.
 363  */
 364 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 365 {
 366         u32 *k = fc->scramble_key;
 367         u64 v = (unsigned long) id;
 368         u32 v0 = v;
 369         u32 v1 = v >> 32;
 370         u32 sum = 0;
 371         int i;
 372
 373         for (i = 0; i < 32; i++) {
 374                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 375                 sum += 0x9E3779B9;
 376                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 377         }
 378
 379         return (u64) v0 + ((u64) v1 << 32);
 380 }
 381
 382 struct fuse_writepage_args {
 383         struct fuse_io_args ia;
 384         struct rb_node writepages_entry;
 385         struct list_head queue_entry;
 386         struct fuse_writepage_args *next;
 387         struct inode *inode;
 388 };
 389
 390 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
 391                                             pgoff_t idx_from, pgoff_t idx_to)
 392 {
 393         struct rb_node *n;
 394
 395         n = fi->writepages.rb_node;
 396
 397         while (n) {
 398                 struct fuse_writepage_args *wpa;
 399                 pgoff_t curr_index;
 400
 401                 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
 402                 WARN_ON(get_fuse_inode(wpa->inode) != fi);
 403                 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
 404                 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
 405                         n = n->rb_right;
 406                 else if (idx_to < curr_index)
 407                         n = n->rb_left;
 408                 else
 409                         return wpa;
 410         }
 411         return NULL;
 412 }
 413
 414 /*
 415  * Check if any page in a range is under writeback
 416  *
 417  * This is currently done by walking the list of writepage requests
 418  * for the inode, which can be pretty inefficient.
 419  */
 420 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 421                                    pgoff_t idx_to)
 422 {
 423         struct fuse_inode *fi = get_fuse_inode(inode);
 424         bool found;
 425
 426         spin_lock(&fi->lock);
 427         found = fuse_find_writeback(fi, idx_from, idx_to);
 428         spin_unlock(&fi->lock);
 429
 430         return found;
 431 }
 432
 433 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 434 {
 435         return fuse_range_is_writeback(inode, index, index);
 436 }
 437
 438 /*
 439  * Wait for page writeback to be completed.
 440  *
 441  * Since fuse doesn't rely on the VM writeback tracking, this has to
 442  * use some other means.
 443  */
 444 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 445 {
 446         struct fuse_inode *fi = get_fuse_inode(inode);
 447
 448         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 449 }
 450
 451 /*
 452  * Wait for all pending writepages on the inode to finish.
 453  *
 454  * This is currently done by blocking further writes with FUSE_NOWRITE
 455  * and waiting for all sent writes to complete.
 456  *
 457  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 458  * could conflict with truncation.
 459  */
 460 static void fuse_sync_writes(struct inode *inode)
 461 {
 462         fuse_set_nowrite(inode);
 463         fuse_release_nowrite(inode);
 464 }
 465
 466 static int fuse_flush(struct file *file, fl_owner_t id)
 467 {
 468         struct inode *inode = file_inode(file);
 469         struct fuse_mount *fm = get_fuse_mount(inode);
 470         struct fuse_file *ff = file->private_data;
 471         struct fuse_flush_in inarg;
 472         FUSE_ARGS(args);
 473         int err;
 474
 475         if (fuse_is_bad(inode))
 476                 return -EIO;
 477
 478         err = write_inode_now(inode, 1);
 479         if (err)
 480                 return err;
 481
 482         inode_lock(inode);
 483         fuse_sync_writes(inode);
 484         inode_unlock(inode);
 485
 486         err = filemap_check_errors(file->f_mapping);
 487         if (err)
 488                 return err;
 489
 490         err = 0;
 491         if (fm->fc->no_flush)
 492                 goto inval_attr_out;
 493
 494         memset(&inarg, 0, sizeof(inarg));
 495         inarg.fh = ff->fh;
 496         inarg.lock_owner = fuse_lock_owner_id(fm->fc, id);
 497         args.opcode = FUSE_FLUSH;
 498         args.nodeid = get_node_id(inode);
 499         args.in_numargs = 1;
 500         args.in_args[0].size = sizeof(inarg);
 501         args.in_args[0].value = &inarg;
 502         args.force = true;
 503
 504         err = fuse_simple_request(fm, &args);
 505         if (err == -ENOSYS) {
 506                 fm->fc->no_flush = 1;
 507                 err = 0;
 508         }
 509
 510 inval_attr_out:
 511         /*
 512          * In memory i_blocks is not maintained by fuse, if writeback cache is
 513          * enabled, i_blocks from cached attr may not be accurate.
 514          */
 515         if (!err && fm->fc->writeback_cache)
 516                 fuse_invalidate_attr(inode);
 517         return err;
 518 }
 519
 520 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 521                       int datasync, int opcode)
 522 {
 523         struct inode *inode = file->f_mapping->host;
 524         struct fuse_mount *fm = get_fuse_mount(inode);
 525         struct fuse_file *ff = file->private_data;
 526         FUSE_ARGS(args);
 527         struct fuse_fsync_in inarg;
 528
 529         memset(&inarg, 0, sizeof(inarg));
 530         inarg.fh = ff->fh;
 531         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 532         args.opcode = opcode;
 533         args.nodeid = get_node_id(inode);
 534         args.in_numargs = 1;
 535         args.in_args[0].size = sizeof(inarg);
 536         args.in_args[0].value = &inarg;
 537         return fuse_simple_request(fm, &args);
 538 }
 539
 540 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 541                       int datasync)
 542 {
 543         struct inode *inode = file->f_mapping->host;
 544         struct fuse_conn *fc = get_fuse_conn(inode);
 545         int err;
 546
 547         if (fuse_is_bad(inode))
 548                 return -EIO;
 549
 550         inode_lock(inode);
 551
 552         /*
 553          * Start writeback against all dirty pages of the inode, then
 554          * wait for all outstanding writes, before sending the FSYNC
 555          * request.
 556          */
 557         err = file_write_and_wait_range(file, start, end);
 558         if (err)
 559                 goto out;
 560
 561         fuse_sync_writes(inode);
 562
 563         /*
 564          * Due to implementation of fuse writeback
 565          * file_write_and_wait_range() does not catch errors.
 566          * We have to do this directly after fuse_sync_writes()
 567          */
 568         err = file_check_and_advance_wb_err(file);
 569         if (err)
 570                 goto out;
 571
 572         err = sync_inode_metadata(inode, 1);
 573         if (err)
 574                 goto out;
 575
 576         if (fc->no_fsync)
 577                 goto out;
 578
 579         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 580         if (err == -ENOSYS) {
 581                 fc->no_fsync = 1;
 582                 err = 0;
 583         }
 584 out:
 585         inode_unlock(inode);
 586
 587         return err;
 588 }
 589
 590 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
 591                          size_t count, int opcode)
 592 {
 593         struct fuse_file *ff = file->private_data;
 594         struct fuse_args *args = &ia->ap.args;
 595
 596         ia->read.in.fh = ff->fh;
 597         ia->read.in.offset = pos;
 598         ia->read.in.size = count;
 599         ia->read.in.flags = file->f_flags;
 600         args->opcode = opcode;
 601         args->nodeid = ff->nodeid;
 602         args->in_numargs = 1;
 603         args->in_args[0].size = sizeof(ia->read.in);
 604         args->in_args[0].value = &ia->read.in;
 605         args->out_argvar = true;
 606         args->out_numargs = 1;
 607         args->out_args[0].size = count;
 608 }
 609
 610 static void fuse_release_user_pages(struct fuse_args_pages *ap,
 611                                     bool should_dirty)
 612 {
 613         unsigned int i;
 614
 615         for (i = 0; i < ap->num_pages; i++) {
 616                 if (should_dirty)
 617                         set_page_dirty_lock(ap->pages[i]);
 618                 put_page(ap->pages[i]);
 619         }
 620 }
 621
 622 static void fuse_io_release(struct kref *kref)
 623 {
 624         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 625 }
 626
 627 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 628 {
 629         if (io->err)
 630                 return io->err;
 631
 632         if (io->bytes >= 0 && io->write)
 633                 return -EIO;
 634
 635         return io->bytes < 0 ? io->size : io->bytes;
 636 }
 637
 638 /**
 639  * In case of short read, the caller sets 'pos' to the position of
 640  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 641  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 642  *
 643  * An example:
 644  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 645  * both submitted asynchronously. The first of them was ACKed by userspace as
 646  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 647  * second request was ACKed as short, e.g. only 1K was read, resulting in
 648  * pos == 33K.
 649  *
 650  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 651  * will be equal to the length of the longest contiguous fragment of
 652  * transferred data starting from the beginning of IO request.
 653  */
 654 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 655 {
 656         int left;
 657
 658         spin_lock(&io->lock);
 659         if (err)
 660                 io->err = io->err ? : err;
 661         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 662                 io->bytes = pos;
 663
 664         left = --io->reqs;
 665         if (!left && io->blocking)
 666                 complete(io->done);
 667         spin_unlock(&io->lock);
 668
 669         if (!left && !io->blocking) {
 670                 ssize_t res = fuse_get_res_by_io(io);
 671
 672                 if (res >= 0) {
 673                         struct inode *inode = file_inode(io->iocb->ki_filp);
 674                         struct fuse_conn *fc = get_fuse_conn(inode);
 675                         struct fuse_inode *fi = get_fuse_inode(inode);
 676
 677                         spin_lock(&fi->lock);
 678                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 679                         spin_unlock(&fi->lock);
 680                 }
 681
 682                 io->iocb->ki_complete(io->iocb, res, 0);
 683         }
 684
 685         kref_put(&io->refcnt, fuse_io_release);
 686 }
 687
 688 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
 689                                           unsigned int npages)
 690 {
 691         struct fuse_io_args *ia;
 692
 693         ia = kzalloc(sizeof(*ia), GFP_KERNEL);
 694         if (ia) {
 695                 ia->io = io;
 696                 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
 697                                                 &ia->ap.descs);
 698                 if (!ia->ap.pages) {
 699                         kfree(ia);
 700                         ia = NULL;
 701                 }
 702         }
 703         return ia;
 704 }
 705
 706 static void fuse_io_free(struct fuse_io_args *ia)
 707 {
 708         kfree(ia->ap.pages);
 709         kfree(ia);
 710 }
 711
 712 static void fuse_aio_complete_req(struct fuse_mount *fm, struct fuse_args *args,
 713                                   int err)
 714 {
 715         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 716         struct fuse_io_priv *io = ia->io;
 717         ssize_t pos = -1;
 718
 719         fuse_release_user_pages(&ia->ap, io->should_dirty);
 720
 721         if (err) {
 722                 /* Nothing */
 723         } else if (io->write) {
 724                 if (ia->write.out.size > ia->write.in.size) {
 725                         err = -EIO;
 726                 } else if (ia->write.in.size != ia->write.out.size) {
 727                         pos = ia->write.in.offset - io->offset +
 728                                 ia->write.out.size;
 729                 }
 730         } else {
 731                 u32 outsize = args->out_args[0].size;
 732
 733                 if (ia->read.in.size != outsize)
 734                         pos = ia->read.in.offset - io->offset + outsize;
 735         }
 736
 737         fuse_aio_complete(io, err, pos);
 738         fuse_io_free(ia);
 739 }
 740
 741 static ssize_t fuse_async_req_send(struct fuse_mount *fm,
 742                                    struct fuse_io_args *ia, size_t num_bytes)
 743 {
 744         ssize_t err;
 745         struct fuse_io_priv *io = ia->io;
 746
 747         spin_lock(&io->lock);
 748         kref_get(&io->refcnt);
 749         io->size += num_bytes;
 750         io->reqs++;
 751         spin_unlock(&io->lock);
 752
 753         ia->ap.args.end = fuse_aio_complete_req;
 754         ia->ap.args.may_block = io->should_dirty;
 755         err = fuse_simple_background(fm, &ia->ap.args, GFP_KERNEL);
 756         if (err)
 757                 fuse_aio_complete_req(fm, &ia->ap.args, err);
 758
 759         return num_bytes;
 760 }
 761
 762 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
 763                               fl_owner_t owner)
 764 {
 765         struct file *file = ia->io->iocb->ki_filp;
 766         struct fuse_file *ff = file->private_data;
 767         struct fuse_mount *fm = ff->fm;
 768
 769         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 770         if (owner != NULL) {
 771                 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
 772                 ia->read.in.lock_owner = fuse_lock_owner_id(fm->fc, owner);
 773         }
 774
 775         if (ia->io->async)
 776                 return fuse_async_req_send(fm, ia, count);
 777
 778         return fuse_simple_request(fm, &ia->ap.args);
 779 }
 780
 781 static void fuse_read_update_size(struct inode *inode, loff_t size,
 782                                   u64 attr_ver)
 783 {
 784         struct fuse_conn *fc = get_fuse_conn(inode);
 785         struct fuse_inode *fi = get_fuse_inode(inode);
 786
 787         spin_lock(&fi->lock);
 788         if (attr_ver == fi->attr_version && size < inode->i_size &&
 789             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 790                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 791                 i_size_write(inode, size);
 792         }
 793         spin_unlock(&fi->lock);
 794 }
 795
 796 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
 797                             struct fuse_args_pages *ap)
 798 {
 799         struct fuse_conn *fc = get_fuse_conn(inode);
 800
 801         if (fc->writeback_cache) {
 802                 /*
 803                  * A hole in a file. Some data after the hole are in page cache,
 804                  * but have not reached the client fs yet. So, the hole is not
 805                  * present there.
 806                  */
 807                 int i;
 808                 int start_idx = num_read >> PAGE_SHIFT;
 809                 size_t off = num_read & (PAGE_SIZE - 1);
 810
 811                 for (i = start_idx; i < ap->num_pages; i++) {
 812                         zero_user_segment(ap->pages[i], off, PAGE_SIZE);
 813                         off = 0;
 814                 }
 815         } else {
 816                 loff_t pos = page_offset(ap->pages[0]) + num_read;
 817                 fuse_read_update_size(inode, pos, attr_ver);
 818         }
 819 }
 820
 821 static int fuse_do_readpage(struct file *file, struct page *page)
 822 {
 823         struct inode *inode = page->mapping->host;
 824         struct fuse_mount *fm = get_fuse_mount(inode);
 825         loff_t pos = page_offset(page);
 826         struct fuse_page_desc desc = { .length = PAGE_SIZE };
 827         struct fuse_io_args ia = {
 828                 .ap.args.page_zeroing = true,
 829                 .ap.args.out_pages = true,
 830                 .ap.num_pages = 1,
 831                 .ap.pages = &page,
 832                 .ap.descs = &desc,
 833         };
 834         ssize_t res;
 835         u64 attr_ver;
 836
 837         /*
 838          * Page writeback can extend beyond the lifetime of the
 839          * page-cache page, so make sure we read a properly synced
 840          * page.
 841          */
 842         fuse_wait_on_page_writeback(inode, page->index);
 843
 844         attr_ver = fuse_get_attr_version(fm->fc);
 845
 846         /* Don't overflow end offset */
 847         if (pos + (desc.length - 1) == LLONG_MAX)
 848                 desc.length--;
 849
 850         fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
 851         res = fuse_simple_request(fm, &ia.ap.args);
 852         if (res < 0)
 853                 return res;
 854         /*
 855          * Short read means EOF.  If file size is larger, truncate it
 856          */
 857         if (res < desc.length)
 858                 fuse_short_read(inode, attr_ver, res, &ia.ap);
 859
 860         SetPageUptodate(page);
 861
 862         return 0;
 863 }
 864
 865 static int fuse_readpage(struct file *file, struct page *page)
 866 {
 867         struct inode *inode = page->mapping->host;
 868         int err;
 869
 870         err = -EIO;
 871         if (fuse_is_bad(inode))
 872                 goto out;
 873
 874         err = fuse_do_readpage(file, page);
 875         fuse_invalidate_atime(inode);
 876  out:
 877         unlock_page(page);
 878         return err;
 879 }
 880
 881 static void fuse_readpages_end(struct fuse_mount *fm, struct fuse_args *args,
 882                                int err)
 883 {
 884         int i;
 885         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 886         struct fuse_args_pages *ap = &ia->ap;
 887         size_t count = ia->read.in.size;
 888         size_t num_read = args->out_args[0].size;
 889         struct address_space *mapping = NULL;
 890
 891         for (i = 0; mapping == NULL && i < ap->num_pages; i++)
 892                 mapping = ap->pages[i]->mapping;
 893
 894         if (mapping) {
 895                 struct inode *inode = mapping->host;
 896
 897                 /*
 898                  * Short read means EOF. If file size is larger, truncate it
 899                  */
 900                 if (!err && num_read < count)
 901                         fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
 902
 903                 fuse_invalidate_atime(inode);
 904         }
 905
 906         for (i = 0; i < ap->num_pages; i++) {
 907                 struct page *page = ap->pages[i];
 908
 909                 if (!err)
 910                         SetPageUptodate(page);
 911                 else
 912                         SetPageError(page);
 913                 unlock_page(page);
 914                 put_page(page);
 915         }
 916         if (ia->ff)
 917                 fuse_file_put(ia->ff, false, false);
 918
 919         fuse_io_free(ia);
 920 }
 921
 922 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
 923 {
 924         struct fuse_file *ff = file->private_data;
 925         struct fuse_mount *fm = ff->fm;
 926         struct fuse_args_pages *ap = &ia->ap;
 927         loff_t pos = page_offset(ap->pages[0]);
 928         size_t count = ap->num_pages << PAGE_SHIFT;
 929         ssize_t res;
 930         int err;
 931
 932         ap->args.out_pages = true;
 933         ap->args.page_zeroing = true;
 934         ap->args.page_replace = true;
 935
 936         /* Don't overflow end offset */
 937         if (pos + (count - 1) == LLONG_MAX) {
 938                 count--;
 939                 ap->descs[ap->num_pages - 1].length--;
 940         }
 941         WARN_ON((loff_t) (pos + count) < 0);
 942
 943         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 944         ia->read.attr_ver = fuse_get_attr_version(fm->fc);
 945         if (fm->fc->async_read) {
 946                 ia->ff = fuse_file_get(ff);
 947                 ap->args.end = fuse_readpages_end;
 948                 err = fuse_simple_background(fm, &ap->args, GFP_KERNEL);
 949                 if (!err)
 950                         return;
 951         } else {
 952                 res = fuse_simple_request(fm, &ap->args);
 953                 err = res < 0 ? res : 0;
 954         }
 955         fuse_readpages_end(fm, &ap->args, err);
 956 }
 957
 958 static void fuse_readahead(struct readahead_control *rac)
 959 {
 960         struct inode *inode = rac->mapping->host;
 961         struct fuse_conn *fc = get_fuse_conn(inode);
 962         unsigned int i, max_pages, nr_pages = 0;
 963
 964         if (fuse_is_bad(inode))
 965                 return;
 966
 967         max_pages = min_t(unsigned int, fc->max_pages,
 968                         fc->max_read / PAGE_SIZE);
 969
 970         for (;;) {
 971                 struct fuse_io_args *ia;
 972                 struct fuse_args_pages *ap;
 973
 974                 nr_pages = readahead_count(rac) - nr_pages;
 975                 if (nr_pages > max_pages)
 976                         nr_pages = max_pages;
 977                 if (nr_pages == 0)
 978                         break;
 979                 ia = fuse_io_alloc(NULL, nr_pages);
 980                 if (!ia)
 981                         return;
 982                 ap = &ia->ap;
 983                 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
 984                 for (i = 0; i < nr_pages; i++) {
 985                         fuse_wait_on_page_writeback(inode,
 986                                                     readahead_index(rac) + i);
 987                         ap->descs[i].length = PAGE_SIZE;
 988                 }
 989                 ap->num_pages = nr_pages;
 990                 fuse_send_readpages(ia, rac->file);
 991         }
 992 }
 993
 994 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
 995 {
 996         struct inode *inode = iocb->ki_filp->f_mapping->host;
 997         struct fuse_conn *fc = get_fuse_conn(inode);
 998
 999         /*
1000          * In auto invalidate mode, always update attributes on read.
1001          * Otherwise, only update if we attempt to read past EOF (to ensure
1002          * i_size is up to date).
1003          */
1004         if (fc->auto_inval_data ||
1005             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
1006                 int err;
1007                 err = fuse_update_attributes(inode, iocb->ki_filp);
1008                 if (err)
1009                         return err;
1010         }
1011
1012         return generic_file_read_iter(iocb, to);
1013 }
1014
1015 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1016                                  loff_t pos, size_t count)
1017 {
1018         struct fuse_args *args = &ia->ap.args;
1019
1020         ia->write.in.fh = ff->fh;
1021         ia->write.in.offset = pos;
1022         ia->write.in.size = count;
1023         args->opcode = FUSE_WRITE;
1024         args->nodeid = ff->nodeid;
1025         args->in_numargs = 2;
1026         if (ff->fm->fc->minor < 9)
1027                 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1028         else
1029                 args->in_args[0].size = sizeof(ia->write.in);
1030         args->in_args[0].value = &ia->write.in;
1031         args->in_args[1].size = count;
1032         args->out_numargs = 1;
1033         args->out_args[0].size = sizeof(ia->write.out);
1034         args->out_args[0].value = &ia->write.out;
1035 }
1036
1037 static unsigned int fuse_write_flags(struct kiocb *iocb)
1038 {
1039         unsigned int flags = iocb->ki_filp->f_flags;
1040
1041         if (iocb->ki_flags & IOCB_DSYNC)
1042                 flags |= O_DSYNC;
1043         if (iocb->ki_flags & IOCB_SYNC)
1044                 flags |= O_SYNC;
1045
1046         return flags;
1047 }
1048
1049 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1050                                size_t count, fl_owner_t owner)
1051 {
1052         struct kiocb *iocb = ia->io->iocb;
1053         struct file *file = iocb->ki_filp;
1054         struct fuse_file *ff = file->private_data;
1055         struct fuse_mount *fm = ff->fm;
1056         struct fuse_write_in *inarg = &ia->write.in;
1057         ssize_t err;
1058
1059         fuse_write_args_fill(ia, ff, pos, count);
1060         inarg->flags = fuse_write_flags(iocb);
1061         if (owner != NULL) {
1062                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1063                 inarg->lock_owner = fuse_lock_owner_id(fm->fc, owner);
1064         }
1065
1066         if (ia->io->async)
1067                 return fuse_async_req_send(fm, ia, count);
1068
1069         err = fuse_simple_request(fm, &ia->ap.args);
1070         if (!err && ia->write.out.size > count)
1071                 err = -EIO;
1072
1073         return err ?: ia->write.out.size;
1074 }
1075
1076 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1077 {
1078         struct fuse_conn *fc = get_fuse_conn(inode);
1079         struct fuse_inode *fi = get_fuse_inode(inode);
1080         bool ret = false;
1081
1082         spin_lock(&fi->lock);
1083         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1084         if (pos > inode->i_size) {
1085                 i_size_write(inode, pos);
1086                 ret = true;
1087         }
1088         spin_unlock(&fi->lock);
1089
1090         return ret;
1091 }
1092
1093 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1094                                      struct kiocb *iocb, struct inode *inode,
1095                                      loff_t pos, size_t count)
1096 {
1097         struct fuse_args_pages *ap = &ia->ap;
1098         struct file *file = iocb->ki_filp;
1099         struct fuse_file *ff = file->private_data;
1100         struct fuse_mount *fm = ff->fm;
1101         unsigned int offset, i;
1102         int err;
1103
1104         for (i = 0; i < ap->num_pages; i++)
1105                 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1106
1107         fuse_write_args_fill(ia, ff, pos, count);
1108         ia->write.in.flags = fuse_write_flags(iocb);
1109         if (fm->fc->handle_killpriv_v2 && !capable(CAP_FSETID))
1110                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1111
1112         err = fuse_simple_request(fm, &ap->args);
1113         if (!err && ia->write.out.size > count)
1114                 err = -EIO;
1115
1116         offset = ap->descs[0].offset;
1117         count = ia->write.out.size;
1118         for (i = 0; i < ap->num_pages; i++) {
1119                 struct page *page = ap->pages[i];
1120
1121                 if (!err && !offset && count >= PAGE_SIZE)
1122                         SetPageUptodate(page);
1123
1124                 if (count > PAGE_SIZE - offset)
1125                         count -= PAGE_SIZE - offset;
1126                 else
1127                         count = 0;
1128                 offset = 0;
1129
1130                 unlock_page(page);
1131                 put_page(page);
1132         }
1133
1134         return err;
1135 }
1136
1137 static ssize_t fuse_fill_write_pages(struct fuse_args_pages *ap,
1138                                      struct address_space *mapping,
1139                                      struct iov_iter *ii, loff_t pos,
1140                                      unsigned int max_pages)
1141 {
1142         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1143         unsigned offset = pos & (PAGE_SIZE - 1);
1144         size_t count = 0;
1145         int err;
1146
1147         ap->args.in_pages = true;
1148         ap->descs[0].offset = offset;
1149
1150         do {
1151                 size_t tmp;
1152                 struct page *page;
1153                 pgoff_t index = pos >> PAGE_SHIFT;
1154                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1155                                      iov_iter_count(ii));
1156
1157                 bytes = min_t(size_t, bytes, fc->max_write - count);
1158
1159  again:
1160                 err = -EFAULT;
1161                 if (iov_iter_fault_in_readable(ii, bytes))
1162                         break;
1163
1164                 err = -ENOMEM;
1165                 page = grab_cache_page_write_begin(mapping, index, 0);
1166                 if (!page)
1167                         break;
1168
1169                 if (mapping_writably_mapped(mapping))
1170                         flush_dcache_page(page);
1171
1172                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1173                 flush_dcache_page(page);
1174
1175                 iov_iter_advance(ii, tmp);
1176                 if (!tmp) {
1177                         unlock_page(page);
1178                         put_page(page);
1179                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1180                         goto again;
1181                 }
1182
1183                 err = 0;
1184                 ap->pages[ap->num_pages] = page;
1185                 ap->descs[ap->num_pages].length = tmp;
1186                 ap->num_pages++;
1187
1188                 count += tmp;
1189                 pos += tmp;
1190                 offset += tmp;
1191                 if (offset == PAGE_SIZE)
1192                         offset = 0;
1193
1194                 if (!fc->big_writes)
1195                         break;
1196         } while (iov_iter_count(ii) && count < fc->max_write &&
1197                  ap->num_pages < max_pages && offset == 0);
1198
1199         return count > 0 ? count : err;
1200 }
1201
1202 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1203                                      unsigned int max_pages)
1204 {
1205         return min_t(unsigned int,
1206                      ((pos + len - 1) >> PAGE_SHIFT) -
1207                      (pos >> PAGE_SHIFT) + 1,
1208                      max_pages);
1209 }
1210
1211 static ssize_t fuse_perform_write(struct kiocb *iocb,
1212                                   struct address_space *mapping,
1213                                   struct iov_iter *ii, loff_t pos)
1214 {
1215         struct inode *inode = mapping->host;
1216         struct fuse_conn *fc = get_fuse_conn(inode);
1217         struct fuse_inode *fi = get_fuse_inode(inode);
1218         int err = 0;
1219         ssize_t res = 0;
1220
1221         if (inode->i_size < pos + iov_iter_count(ii))
1222                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1223
1224         do {
1225                 ssize_t count;
1226                 struct fuse_io_args ia = {};
1227                 struct fuse_args_pages *ap = &ia.ap;
1228                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1229                                                       fc->max_pages);
1230
1231                 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1232                 if (!ap->pages) {
1233                         err = -ENOMEM;
1234                         break;
1235                 }
1236
1237                 count = fuse_fill_write_pages(ap, mapping, ii, pos, nr_pages);
1238                 if (count <= 0) {
1239                         err = count;
1240                 } else {
1241                         err = fuse_send_write_pages(&ia, iocb, inode,
1242                                                     pos, count);
1243                         if (!err) {
1244                                 size_t num_written = ia.write.out.size;
1245
1246                                 res += num_written;
1247                                 pos += num_written;
1248
1249                                 /* break out of the loop on short write */
1250                                 if (num_written != count)
1251                                         err = -EIO;
1252                         }
1253                 }
1254                 kfree(ap->pages);
1255         } while (!err && iov_iter_count(ii));
1256
1257         if (res > 0)
1258                 fuse_write_update_size(inode, pos);
1259
1260         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1261         fuse_invalidate_attr(inode);
1262
1263         return res > 0 ? res : err;
1264 }
1265
1266 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1267 {
1268         struct file *file = iocb->ki_filp;
1269         struct address_space *mapping = file->f_mapping;
1270         ssize_t written = 0;
1271         ssize_t written_buffered = 0;
1272         struct inode *inode = mapping->host;
1273         ssize_t err;
1274         struct fuse_conn *fc = get_fuse_conn(inode);
1275         loff_t endbyte = 0;
1276
1277         if (fc->writeback_cache) {
1278                 /* Update size (EOF optimization) and mode (SUID clearing) */
1279                 err = fuse_update_attributes(mapping->host, file);
1280                 if (err)
1281                         return err;
1282
1283                 if (fc->handle_killpriv_v2 &&
1284                     should_remove_suid(file_dentry(file))) {
1285                         goto writethrough;
1286                 }
1287
1288                 return generic_file_write_iter(iocb, from);
1289         }
1290
1291 writethrough:
1292         inode_lock(inode);
1293
1294         /* We can write back this queue in page reclaim */
1295         current->backing_dev_info = inode_to_bdi(inode);
1296
1297         err = generic_write_checks(iocb, from);
1298         if (err <= 0)
1299                 goto out;
1300
1301         err = file_remove_privs(file);
1302         if (err)
1303                 goto out;
1304
1305         err = file_update_time(file);
1306         if (err)
1307                 goto out;
1308
1309         if (iocb->ki_flags & IOCB_DIRECT) {
1310                 loff_t pos = iocb->ki_pos;
1311                 written = generic_file_direct_write(iocb, from);
1312                 if (written < 0 || !iov_iter_count(from))
1313                         goto out;
1314
1315                 pos += written;
1316
1317                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1318                 if (written_buffered < 0) {
1319                         err = written_buffered;
1320                         goto out;
1321                 }
1322                 endbyte = pos + written_buffered - 1;
1323
1324                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1325                                                    endbyte);
1326                 if (err)
1327                         goto out;
1328
1329                 invalidate_mapping_pages(file->f_mapping,
1330                                          pos >> PAGE_SHIFT,
1331                                          endbyte >> PAGE_SHIFT);
1332
1333                 written += written_buffered;
1334                 iocb->ki_pos = pos + written_buffered;
1335         } else {
1336                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1337                 if (written >= 0)
1338                         iocb->ki_pos += written;
1339         }
1340 out:
1341         current->backing_dev_info = NULL;
1342         inode_unlock(inode);
1343         if (written > 0)
1344                 written = generic_write_sync(iocb, written);
1345
1346         return written ? written : err;
1347 }
1348
1349 static inline void fuse_page_descs_length_init(struct fuse_page_desc *descs,
1350                                                unsigned int index,
1351                                                unsigned int nr_pages)
1352 {
1353         int i;
1354
1355         for (i = index; i < index + nr_pages; i++)
1356                 descs[i].length = PAGE_SIZE - descs[i].offset;
1357 }
1358
1359 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1360 {
1361         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1362 }
1363
1364 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1365                                         size_t max_size)
1366 {
1367         return min(iov_iter_single_seg_count(ii), max_size);
1368 }
1369
1370 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1371                                size_t *nbytesp, int write,
1372                                unsigned int max_pages)
1373 {
1374         size_t nbytes = 0;  /* # bytes already packed in req */
1375         ssize_t ret = 0;
1376
1377         /* Special case for kernel I/O: can copy directly into the buffer */
1378         if (iov_iter_is_kvec(ii)) {
1379                 unsigned long user_addr = fuse_get_user_addr(ii);
1380                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1381
1382                 if (write)
1383                         ap->args.in_args[1].value = (void *) user_addr;
1384                 else
1385                         ap->args.out_args[0].value = (void *) user_addr;
1386
1387                 iov_iter_advance(ii, frag_size);
1388                 *nbytesp = frag_size;
1389                 return 0;
1390         }
1391
1392         while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1393                 unsigned npages;
1394                 size_t start;
1395                 ret = iov_iter_get_pages(ii, &ap->pages[ap->num_pages],
1396                                         *nbytesp - nbytes,
1397                                         max_pages - ap->num_pages,
1398                                         &start);
1399                 if (ret < 0)
1400                         break;
1401
1402                 iov_iter_advance(ii, ret);
1403                 nbytes += ret;
1404
1405                 ret += start;
1406                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1407
1408                 ap->descs[ap->num_pages].offset = start;
1409                 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1410
1411                 ap->num_pages += npages;
1412                 ap->descs[ap->num_pages - 1].length -=
1413                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1414         }
1415
1416         if (write)
1417                 ap->args.in_pages = true;
1418         else
1419                 ap->args.out_pages = true;
1420
1421         *nbytesp = nbytes;
1422
1423         return ret < 0 ? ret : 0;
1424 }
1425
1426 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1427                        loff_t *ppos, int flags)
1428 {
1429         int write = flags & FUSE_DIO_WRITE;
1430         int cuse = flags & FUSE_DIO_CUSE;
1431         struct file *file = io->iocb->ki_filp;
1432         struct inode *inode = file->f_mapping->host;
1433         struct fuse_file *ff = file->private_data;
1434         struct fuse_conn *fc = ff->fm->fc;
1435         size_t nmax = write ? fc->max_write : fc->max_read;
1436         loff_t pos = *ppos;
1437         size_t count = iov_iter_count(iter);
1438         pgoff_t idx_from = pos >> PAGE_SHIFT;
1439         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1440         ssize_t res = 0;
1441         int err = 0;
1442         struct fuse_io_args *ia;
1443         unsigned int max_pages;
1444
1445         max_pages = iov_iter_npages(iter, fc->max_pages);
1446         ia = fuse_io_alloc(io, max_pages);
1447         if (!ia)
1448                 return -ENOMEM;
1449
1450         ia->io = io;
1451         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1452                 if (!write)
1453                         inode_lock(inode);
1454                 fuse_sync_writes(inode);
1455                 if (!write)
1456                         inode_unlock(inode);
1457         }
1458
1459         io->should_dirty = !write && iter_is_iovec(iter);
1460         while (count) {
1461                 ssize_t nres;
1462                 fl_owner_t owner = current->files;
1463                 size_t nbytes = min(count, nmax);
1464
1465                 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1466                                           max_pages);
1467                 if (err && !nbytes)
1468                         break;
1469
1470                 if (write) {
1471                         if (!capable(CAP_FSETID))
1472                                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1473
1474                         nres = fuse_send_write(ia, pos, nbytes, owner);
1475                 } else {
1476                         nres = fuse_send_read(ia, pos, nbytes, owner);
1477                 }
1478
1479                 if (!io->async || nres < 0) {
1480                         fuse_release_user_pages(&ia->ap, io->should_dirty);
1481                         fuse_io_free(ia);
1482                 }
1483                 ia = NULL;
1484                 if (nres < 0) {
1485                         iov_iter_revert(iter, nbytes);
1486                         err = nres;
1487                         break;
1488                 }
1489                 WARN_ON(nres > nbytes);
1490
1491                 count -= nres;
1492                 res += nres;
1493                 pos += nres;
1494                 if (nres != nbytes) {
1495                         iov_iter_revert(iter, nbytes - nres);
1496                         break;
1497                 }
1498                 if (count) {
1499                         max_pages = iov_iter_npages(iter, fc->max_pages);
1500                         ia = fuse_io_alloc(io, max_pages);
1501                         if (!ia)
1502                                 break;
1503                 }
1504         }
1505         if (ia)
1506                 fuse_io_free(ia);
1507         if (res > 0)
1508                 *ppos = pos;
1509
1510         return res > 0 ? res : err;
1511 }
1512 EXPORT_SYMBOL_GPL(fuse_direct_io);
1513
1514 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1515                                   struct iov_iter *iter,
1516                                   loff_t *ppos)
1517 {
1518         ssize_t res;
1519         struct inode *inode = file_inode(io->iocb->ki_filp);
1520
1521         res = fuse_direct_io(io, iter, ppos, 0);
1522
1523         fuse_invalidate_atime(inode);
1524
1525         return res;
1526 }
1527
1528 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1529
1530 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1531 {
1532         ssize_t res;
1533
1534         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1535                 res = fuse_direct_IO(iocb, to);
1536         } else {
1537                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1538
1539                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1540         }
1541
1542         return res;
1543 }
1544
1545 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1546 {
1547         struct inode *inode = file_inode(iocb->ki_filp);
1548         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1549         ssize_t res;
1550
1551         /* Don't allow parallel writes to the same file */
1552         inode_lock(inode);
1553         res = generic_write_checks(iocb, from);
1554         if (res > 0) {
1555                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1556                         res = fuse_direct_IO(iocb, from);
1557                 } else {
1558                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1559                                              FUSE_DIO_WRITE);
1560                 }
1561         }
1562         fuse_invalidate_attr(inode);
1563         if (res > 0)
1564                 fuse_write_update_size(inode, iocb->ki_pos);
1565         inode_unlock(inode);
1566
1567         return res;
1568 }
1569
1570 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1571 {
1572         struct file *file = iocb->ki_filp;
1573         struct fuse_file *ff = file->private_data;
1574         struct inode *inode = file_inode(file);
1575
1576         if (fuse_is_bad(inode))
1577                 return -EIO;
1578
1579         if (FUSE_IS_DAX(inode))
1580                 return fuse_dax_read_iter(iocb, to);
1581
1582         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1583                 return fuse_cache_read_iter(iocb, to);
1584         else
1585                 return fuse_direct_read_iter(iocb, to);
1586 }
1587
1588 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1589 {
1590         struct file *file = iocb->ki_filp;
1591         struct fuse_file *ff = file->private_data;
1592         struct inode *inode = file_inode(file);
1593
1594         if (fuse_is_bad(inode))
1595                 return -EIO;
1596
1597         if (FUSE_IS_DAX(inode))
1598                 return fuse_dax_write_iter(iocb, from);
1599
1600         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1601                 return fuse_cache_write_iter(iocb, from);
1602         else
1603                 return fuse_direct_write_iter(iocb, from);
1604 }
1605
1606 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1607 {
1608         struct fuse_args_pages *ap = &wpa->ia.ap;
1609         int i;
1610
1611         for (i = 0; i < ap->num_pages; i++)
1612                 __free_page(ap->pages[i]);
1613
1614         if (wpa->ia.ff)
1615                 fuse_file_put(wpa->ia.ff, false, false);
1616
1617         kfree(ap->pages);
1618         kfree(wpa);
1619 }
1620
1621 static void fuse_writepage_finish(struct fuse_mount *fm,
1622                                   struct fuse_writepage_args *wpa)
1623 {
1624         struct fuse_args_pages *ap = &wpa->ia.ap;
1625         struct inode *inode = wpa->inode;
1626         struct fuse_inode *fi = get_fuse_inode(inode);
1627         struct backing_dev_info *bdi = inode_to_bdi(inode);
1628         int i;
1629
1630         for (i = 0; i < ap->num_pages; i++) {
1631                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1632                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1633                 wb_writeout_inc(&bdi->wb);
1634         }
1635         wake_up(&fi->page_waitq);
1636 }
1637
1638 /* Called under fi->lock, may release and reacquire it */
1639 static void fuse_send_writepage(struct fuse_mount *fm,
1640                                 struct fuse_writepage_args *wpa, loff_t size)
1641 __releases(fi->lock)
1642 __acquires(fi->lock)
1643 {
1644         struct fuse_writepage_args *aux, *next;
1645         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1646         struct fuse_write_in *inarg = &wpa->ia.write.in;
1647         struct fuse_args *args = &wpa->ia.ap.args;
1648         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1649         int err;
1650
1651         fi->writectr++;
1652         if (inarg->offset + data_size <= size) {
1653                 inarg->size = data_size;
1654         } else if (inarg->offset < size) {
1655                 inarg->size = size - inarg->offset;
1656         } else {
1657                 /* Got truncated off completely */
1658                 goto out_free;
1659         }
1660
1661         args->in_args[1].size = inarg->size;
1662         args->force = true;
1663         args->nocreds = true;
1664
1665         err = fuse_simple_background(fm, args, GFP_ATOMIC);
1666         if (err == -ENOMEM) {
1667                 spin_unlock(&fi->lock);
1668                 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1669                 spin_lock(&fi->lock);
1670         }
1671
1672         /* Fails on broken connection only */
1673         if (unlikely(err))
1674                 goto out_free;
1675
1676         return;
1677
1678  out_free:
1679         fi->writectr--;
1680         rb_erase(&wpa->writepages_entry, &fi->writepages);
1681         fuse_writepage_finish(fm, wpa);
1682         spin_unlock(&fi->lock);
1683
1684         /* After fuse_writepage_finish() aux request list is private */
1685         for (aux = wpa->next; aux; aux = next) {
1686                 next = aux->next;
1687                 aux->next = NULL;
1688                 fuse_writepage_free(aux);
1689         }
1690
1691         fuse_writepage_free(wpa);
1692         spin_lock(&fi->lock);
1693 }
1694
1695 /*
1696  * If fi->writectr is positive (no truncate or fsync going on) send
1697  * all queued writepage requests.
1698  *
1699  * Called with fi->lock
1700  */
1701 void fuse_flush_writepages(struct inode *inode)
1702 __releases(fi->lock)
1703 __acquires(fi->lock)
1704 {
1705         struct fuse_mount *fm = get_fuse_mount(inode);
1706         struct fuse_inode *fi = get_fuse_inode(inode);
1707         loff_t crop = i_size_read(inode);
1708         struct fuse_writepage_args *wpa;
1709
1710         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1711                 wpa = list_entry(fi->queued_writes.next,
1712                                  struct fuse_writepage_args, queue_entry);
1713                 list_del_init(&wpa->queue_entry);
1714                 fuse_send_writepage(fm, wpa, crop);
1715         }
1716 }
1717
1718 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1719                                                 struct fuse_writepage_args *wpa)
1720 {
1721         pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1722         pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1723         struct rb_node **p = &root->rb_node;
1724         struct rb_node  *parent = NULL;
1725
1726         WARN_ON(!wpa->ia.ap.num_pages);
1727         while (*p) {
1728                 struct fuse_writepage_args *curr;
1729                 pgoff_t curr_index;
1730
1731                 parent = *p;
1732                 curr = rb_entry(parent, struct fuse_writepage_args,
1733                                 writepages_entry);
1734                 WARN_ON(curr->inode != wpa->inode);
1735                 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1736
1737                 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1738                         p = &(*p)->rb_right;
1739                 else if (idx_to < curr_index)
1740                         p = &(*p)->rb_left;
1741                 else
1742                         return curr;
1743         }
1744
1745         rb_link_node(&wpa->writepages_entry, parent, p);
1746         rb_insert_color(&wpa->writepages_entry, root);
1747         return NULL;
1748 }
1749
1750 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1751 {
1752         WARN_ON(fuse_insert_writeback(root, wpa));
1753 }
1754
1755 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1756                                int error)
1757 {
1758         struct fuse_writepage_args *wpa =
1759                 container_of(args, typeof(*wpa), ia.ap.args);
1760         struct inode *inode = wpa->inode;
1761         struct fuse_inode *fi = get_fuse_inode(inode);
1762
1763         mapping_set_error(inode->i_mapping, error);
1764         spin_lock(&fi->lock);
1765         rb_erase(&wpa->writepages_entry, &fi->writepages);
1766         while (wpa->next) {
1767                 struct fuse_mount *fm = get_fuse_mount(inode);
1768                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1769                 struct fuse_writepage_args *next = wpa->next;
1770
1771                 wpa->next = next->next;
1772                 next->next = NULL;
1773                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1774                 tree_insert(&fi->writepages, next);
1775
1776                 /*
1777                  * Skip fuse_flush_writepages() to make it easy to crop requests
1778                  * based on primary request size.
1779                  *
1780                  * 1st case (trivial): there are no concurrent activities using
1781                  * fuse_set/release_nowrite.  Then we're on safe side because
1782                  * fuse_flush_writepages() would call fuse_send_writepage()
1783                  * anyway.
1784                  *
1785                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1786                  * now for completion of all in-flight requests.  This happens
1787                  * rarely and no more than once per page, so this should be
1788                  * okay.
1789                  *
1790                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1791                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1792                  * that fuse_set_nowrite returned implies that all in-flight
1793                  * requests were completed along with all of their secondary
1794                  * requests.  Further primary requests are blocked by negative
1795                  * writectr.  Hence there cannot be any in-flight requests and
1796                  * no invocations of fuse_writepage_end() while we're in
1797                  * fuse_set_nowrite..fuse_release_nowrite section.
1798                  */
1799                 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1800         }
1801         fi->writectr--;
1802         fuse_writepage_finish(fm, wpa);
1803         spin_unlock(&fi->lock);
1804         fuse_writepage_free(wpa);
1805 }
1806
1807 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1808                                                struct fuse_inode *fi)
1809 {
1810         struct fuse_file *ff = NULL;
1811
1812         spin_lock(&fi->lock);
1813         if (!list_empty(&fi->write_files)) {
1814                 ff = list_entry(fi->write_files.next, struct fuse_file,
1815                                 write_entry);
1816                 fuse_file_get(ff);
1817         }
1818         spin_unlock(&fi->lock);
1819
1820         return ff;
1821 }
1822
1823 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1824                                              struct fuse_inode *fi)
1825 {
1826         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1827         WARN_ON(!ff);
1828         return ff;
1829 }
1830
1831 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1832 {
1833         struct fuse_conn *fc = get_fuse_conn(inode);
1834         struct fuse_inode *fi = get_fuse_inode(inode);
1835         struct fuse_file *ff;
1836         int err;
1837
1838         ff = __fuse_write_file_get(fc, fi);
1839         err = fuse_flush_times(inode, ff);
1840         if (ff)
1841                 fuse_file_put(ff, false, false);
1842
1843         return err;
1844 }
1845
1846 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1847 {
1848         struct fuse_writepage_args *wpa;
1849         struct fuse_args_pages *ap;
1850
1851         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1852         if (wpa) {
1853                 ap = &wpa->ia.ap;
1854                 ap->num_pages = 0;
1855                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1856                 if (!ap->pages) {
1857                         kfree(wpa);
1858                         wpa = NULL;
1859                 }
1860         }
1861         return wpa;
1862
1863 }
1864
1865 static int fuse_writepage_locked(struct page *page)
1866 {
1867         struct address_space *mapping = page->mapping;
1868         struct inode *inode = mapping->host;
1869         struct fuse_conn *fc = get_fuse_conn(inode);
1870         struct fuse_inode *fi = get_fuse_inode(inode);
1871         struct fuse_writepage_args *wpa;
1872         struct fuse_args_pages *ap;
1873         struct page *tmp_page;
1874         int error = -ENOMEM;
1875
1876         set_page_writeback(page);
1877
1878         wpa = fuse_writepage_args_alloc();
1879         if (!wpa)
1880                 goto err;
1881         ap = &wpa->ia.ap;
1882
1883         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1884         if (!tmp_page)
1885                 goto err_free;
1886
1887         error = -EIO;
1888         wpa->ia.ff = fuse_write_file_get(fc, fi);
1889         if (!wpa->ia.ff)
1890                 goto err_nofile;
1891
1892         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1893
1894         copy_highpage(tmp_page, page);
1895         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1896         wpa->next = NULL;
1897         ap->args.in_pages = true;
1898         ap->num_pages = 1;
1899         ap->pages[0] = tmp_page;
1900         ap->descs[0].offset = 0;
1901         ap->descs[0].length = PAGE_SIZE;
1902         ap->args.end = fuse_writepage_end;
1903         wpa->inode = inode;
1904
1905         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1906         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1907
1908         spin_lock(&fi->lock);
1909         tree_insert(&fi->writepages, wpa);
1910         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1911         fuse_flush_writepages(inode);
1912         spin_unlock(&fi->lock);
1913
1914         end_page_writeback(page);
1915
1916         return 0;
1917
1918 err_nofile:
1919         __free_page(tmp_page);
1920 err_free:
1921         kfree(wpa);
1922 err:
1923         mapping_set_error(page->mapping, error);
1924         end_page_writeback(page);
1925         return error;
1926 }
1927
1928 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1929 {
1930         int err;
1931
1932         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1933                 /*
1934                  * ->writepages() should be called for sync() and friends.  We
1935                  * should only get here on direct reclaim and then we are
1936                  * allowed to skip a page which is already in flight
1937                  */
1938                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1939
1940                 redirty_page_for_writepage(wbc, page);
1941                 unlock_page(page);
1942                 return 0;
1943         }
1944
1945         err = fuse_writepage_locked(page);
1946         unlock_page(page);
1947
1948         return err;
1949 }
1950
1951 struct fuse_fill_wb_data {
1952         struct fuse_writepage_args *wpa;
1953         struct fuse_file *ff;
1954         struct inode *inode;
1955         struct page **orig_pages;
1956         unsigned int max_pages;
1957 };
1958
1959 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
1960 {
1961         struct fuse_args_pages *ap = &data->wpa->ia.ap;
1962         struct fuse_conn *fc = get_fuse_conn(data->inode);
1963         struct page **pages;
1964         struct fuse_page_desc *descs;
1965         unsigned int npages = min_t(unsigned int,
1966                                     max_t(unsigned int, data->max_pages * 2,
1967                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
1968                                     fc->max_pages);
1969         WARN_ON(npages <= data->max_pages);
1970
1971         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
1972         if (!pages)
1973                 return false;
1974
1975         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
1976         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
1977         kfree(ap->pages);
1978         ap->pages = pages;
1979         ap->descs = descs;
1980         data->max_pages = npages;
1981
1982         return true;
1983 }
1984
1985 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1986 {
1987         struct fuse_writepage_args *wpa = data->wpa;
1988         struct inode *inode = data->inode;
1989         struct fuse_inode *fi = get_fuse_inode(inode);
1990         int num_pages = wpa->ia.ap.num_pages;
1991         int i;
1992
1993         wpa->ia.ff = fuse_file_get(data->ff);
1994         spin_lock(&fi->lock);
1995         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1996         fuse_flush_writepages(inode);
1997         spin_unlock(&fi->lock);
1998
1999         for (i = 0; i < num_pages; i++)
2000                 end_page_writeback(data->orig_pages[i]);
2001 }
2002
2003 /*
2004  * Check under fi->lock if the page is under writeback, and insert it onto the
2005  * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2006  * one already added for a page at this offset.  If there's none, then insert
2007  * this new request onto the auxiliary list, otherwise reuse the existing one by
2008  * swapping the new temp page with the old one.
2009  */
2010 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2011                                struct page *page)
2012 {
2013         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2014         struct fuse_writepage_args *tmp;
2015         struct fuse_writepage_args *old_wpa;
2016         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2017
2018         WARN_ON(new_ap->num_pages != 0);
2019         new_ap->num_pages = 1;
2020
2021         spin_lock(&fi->lock);
2022         old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2023         if (!old_wpa) {
2024                 spin_unlock(&fi->lock);
2025                 return true;
2026         }
2027
2028         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2029                 pgoff_t curr_index;
2030
2031                 WARN_ON(tmp->inode != new_wpa->inode);
2032                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2033                 if (curr_index == page->index) {
2034                         WARN_ON(tmp->ia.ap.num_pages != 1);
2035                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2036                         break;
2037                 }
2038         }
2039
2040         if (!tmp) {
2041                 new_wpa->next = old_wpa->next;
2042                 old_wpa->next = new_wpa;
2043         }
2044
2045         spin_unlock(&fi->lock);
2046
2047         if (tmp) {
2048                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2049
2050                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2051                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2052                 wb_writeout_inc(&bdi->wb);
2053                 fuse_writepage_free(new_wpa);
2054         }
2055
2056         return false;
2057 }
2058
2059 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2060                                      struct fuse_args_pages *ap,
2061                                      struct fuse_fill_wb_data *data)
2062 {
2063         WARN_ON(!ap->num_pages);
2064
2065         /*
2066          * Being under writeback is unlikely but possible.  For example direct
2067          * read to an mmaped fuse file will set the page dirty twice; once when
2068          * the pages are faulted with get_user_pages(), and then after the read
2069          * completed.
2070          */
2071         if (fuse_page_is_writeback(data->inode, page->index))
2072                 return true;
2073
2074         /* Reached max pages */
2075         if (ap->num_pages == fc->max_pages)
2076                 return true;
2077
2078         /* Reached max write bytes */
2079         if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2080                 return true;
2081
2082         /* Discontinuity */
2083         if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2084                 return true;
2085
2086         /* Need to grow the pages array?  If so, did the expansion fail? */
2087         if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2088                 return true;
2089
2090         return false;
2091 }
2092
2093 static int fuse_writepages_fill(struct page *page,
2094                 struct writeback_control *wbc, void *_data)
2095 {
2096         struct fuse_fill_wb_data *data = _data;
2097         struct fuse_writepage_args *wpa = data->wpa;
2098         struct fuse_args_pages *ap = &wpa->ia.ap;
2099         struct inode *inode = data->inode;
2100         struct fuse_inode *fi = get_fuse_inode(inode);
2101         struct fuse_conn *fc = get_fuse_conn(inode);
2102         struct page *tmp_page;
2103         int err;
2104
2105         if (!data->ff) {
2106                 err = -EIO;
2107                 data->ff = fuse_write_file_get(fc, fi);
2108                 if (!data->ff)
2109                         goto out_unlock;
2110         }
2111
2112         if (wpa && fuse_writepage_need_send(fc, page, ap, data)) {
2113                 fuse_writepages_send(data);
2114                 data->wpa = NULL;
2115         }
2116
2117         err = -ENOMEM;
2118         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2119         if (!tmp_page)
2120                 goto out_unlock;
2121
2122         /*
2123          * The page must not be redirtied until the writeout is completed
2124          * (i.e. userspace has sent a reply to the write request).  Otherwise
2125          * there could be more than one temporary page instance for each real
2126          * page.
2127          *
2128          * This is ensured by holding the page lock in page_mkwrite() while
2129          * checking fuse_page_is_writeback().  We already hold the page lock
2130          * since clear_page_dirty_for_io() and keep it held until we add the
2131          * request to the fi->writepages list and increment ap->num_pages.
2132          * After this fuse_page_is_writeback() will indicate that the page is
2133          * under writeback, so we can release the page lock.
2134          */
2135         if (data->wpa == NULL) {
2136                 err = -ENOMEM;
2137                 wpa = fuse_writepage_args_alloc();
2138                 if (!wpa) {
2139                         __free_page(tmp_page);
2140                         goto out_unlock;
2141                 }
2142                 data->max_pages = 1;
2143
2144                 ap = &wpa->ia.ap;
2145                 fuse_write_args_fill(&wpa->ia, data->ff, page_offset(page), 0);
2146                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2147                 wpa->next = NULL;
2148                 ap->args.in_pages = true;
2149                 ap->args.end = fuse_writepage_end;
2150                 ap->num_pages = 0;
2151                 wpa->inode = inode;
2152         }
2153         set_page_writeback(page);
2154
2155         copy_highpage(tmp_page, page);
2156         ap->pages[ap->num_pages] = tmp_page;
2157         ap->descs[ap->num_pages].offset = 0;
2158         ap->descs[ap->num_pages].length = PAGE_SIZE;
2159         data->orig_pages[ap->num_pages] = page;
2160
2161         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2162         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2163
2164         err = 0;
2165         if (data->wpa) {
2166                 /*
2167                  * Protected by fi->lock against concurrent access by
2168                  * fuse_page_is_writeback().
2169                  */
2170                 spin_lock(&fi->lock);
2171                 ap->num_pages++;
2172                 spin_unlock(&fi->lock);
2173         } else if (fuse_writepage_add(wpa, page)) {
2174                 data->wpa = wpa;
2175         } else {
2176                 end_page_writeback(page);
2177         }
2178 out_unlock:
2179         unlock_page(page);
2180
2181         return err;
2182 }
2183
2184 static int fuse_writepages(struct address_space *mapping,
2185                            struct writeback_control *wbc)
2186 {
2187         struct inode *inode = mapping->host;
2188         struct fuse_conn *fc = get_fuse_conn(inode);
2189         struct fuse_fill_wb_data data;
2190         int err;
2191
2192         err = -EIO;
2193         if (fuse_is_bad(inode))
2194                 goto out;
2195
2196         data.inode = inode;
2197         data.wpa = NULL;
2198         data.ff = NULL;
2199
2200         err = -ENOMEM;
2201         data.orig_pages = kcalloc(fc->max_pages,
2202                                   sizeof(struct page *),
2203                                   GFP_NOFS);
2204         if (!data.orig_pages)
2205                 goto out;
2206
2207         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2208         if (data.wpa) {
2209                 WARN_ON(!data.wpa->ia.ap.num_pages);
2210                 fuse_writepages_send(&data);
2211         }
2212         if (data.ff)
2213                 fuse_file_put(data.ff, false, false);
2214
2215         kfree(data.orig_pages);
2216 out:
2217         return err;
2218 }
2219
2220 /*
2221  * It's worthy to make sure that space is reserved on disk for the write,
2222  * but how to implement it without killing performance need more thinking.
2223  */
2224 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2225                 loff_t pos, unsigned len, unsigned flags,
2226                 struct page **pagep, void **fsdata)
2227 {
2228         pgoff_t index = pos >> PAGE_SHIFT;
2229         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2230         struct page *page;
2231         loff_t fsize;
2232         int err = -ENOMEM;
2233
2234         WARN_ON(!fc->writeback_cache);
2235
2236         page = grab_cache_page_write_begin(mapping, index, flags);
2237         if (!page)
2238                 goto error;
2239
2240         fuse_wait_on_page_writeback(mapping->host, page->index);
2241
2242         if (PageUptodate(page) || len == PAGE_SIZE)
2243                 goto success;
2244         /*
2245          * Check if the start this page comes after the end of file, in which
2246          * case the readpage can be optimized away.
2247          */
2248         fsize = i_size_read(mapping->host);
2249         if (fsize <= (pos & PAGE_MASK)) {
2250                 size_t off = pos & ~PAGE_MASK;
2251                 if (off)
2252                         zero_user_segment(page, 0, off);
2253                 goto success;
2254         }
2255         err = fuse_do_readpage(file, page);
2256         if (err)
2257                 goto cleanup;
2258 success:
2259         *pagep = page;
2260         return 0;
2261
2262 cleanup:
2263         unlock_page(page);
2264         put_page(page);
2265 error:
2266         return err;
2267 }
2268
2269 static int fuse_write_end(struct file *file, struct address_space *mapping,
2270                 loff_t pos, unsigned len, unsigned copied,
2271                 struct page *page, void *fsdata)
2272 {
2273         struct inode *inode = page->mapping->host;
2274
2275         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2276         if (!copied)
2277                 goto unlock;
2278
2279         if (!PageUptodate(page)) {
2280                 /* Zero any unwritten bytes at the end of the page */
2281                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2282                 if (endoff)
2283                         zero_user_segment(page, endoff, PAGE_SIZE);
2284                 SetPageUptodate(page);
2285         }
2286
2287         fuse_write_update_size(inode, pos + copied);
2288         set_page_dirty(page);
2289
2290 unlock:
2291         unlock_page(page);
2292         put_page(page);
2293
2294         return copied;
2295 }
2296
2297 static int fuse_launder_page(struct page *page)
2298 {
2299         int err = 0;
2300         if (clear_page_dirty_for_io(page)) {
2301                 struct inode *inode = page->mapping->host;
2302
2303                 /* Serialize with pending writeback for the same page */
2304                 fuse_wait_on_page_writeback(inode, page->index);
2305                 err = fuse_writepage_locked(page);
2306                 if (!err)
2307                         fuse_wait_on_page_writeback(inode, page->index);
2308         }
2309         return err;
2310 }
2311
2312 /*
2313  * Write back dirty pages now, because there may not be any suitable
2314  * open files later
2315  */
2316 static void fuse_vma_close(struct vm_area_struct *vma)
2317 {
2318         filemap_write_and_wait(vma->vm_file->f_mapping);
2319 }
2320
2321 /*
2322  * Wait for writeback against this page to complete before allowing it
2323  * to be marked dirty again, and hence written back again, possibly
2324  * before the previous writepage completed.
2325  *
2326  * Block here, instead of in ->writepage(), so that the userspace fs
2327  * can only block processes actually operating on the filesystem.
2328  *
2329  * Otherwise unprivileged userspace fs would be able to block
2330  * unrelated:
2331  *
2332  * - page migration
2333  * - sync(2)
2334  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2335  */
2336 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2337 {
2338         struct page *page = vmf->page;
2339         struct inode *inode = file_inode(vmf->vma->vm_file);
2340
2341         file_update_time(vmf->vma->vm_file);
2342         lock_page(page);
2343         if (page->mapping != inode->i_mapping) {
2344                 unlock_page(page);
2345                 return VM_FAULT_NOPAGE;
2346         }
2347
2348         fuse_wait_on_page_writeback(inode, page->index);
2349         return VM_FAULT_LOCKED;
2350 }
2351
2352 static const struct vm_operations_struct fuse_file_vm_ops = {
2353         .close          = fuse_vma_close,
2354         .fault          = filemap_fault,
2355         .map_pages      = filemap_map_pages,
2356         .page_mkwrite   = fuse_page_mkwrite,
2357 };
2358
2359 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2360 {
2361         struct fuse_file *ff = file->private_data;
2362
2363         /* DAX mmap is superior to direct_io mmap */
2364         if (FUSE_IS_DAX(file_inode(file)))
2365                 return fuse_dax_mmap(file, vma);
2366
2367         if (ff->open_flags & FOPEN_DIRECT_IO) {
2368                 /* Can't provide the coherency needed for MAP_SHARED */
2369                 if (vma->vm_flags & VM_MAYSHARE)
2370                         return -ENODEV;
2371
2372                 invalidate_inode_pages2(file->f_mapping);
2373
2374                 return generic_file_mmap(file, vma);
2375         }
2376
2377         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2378                 fuse_link_write_file(file);
2379
2380         file_accessed(file);
2381         vma->vm_ops = &fuse_file_vm_ops;
2382         return 0;
2383 }
2384
2385 static int convert_fuse_file_lock(struct fuse_conn *fc,
2386                                   const struct fuse_file_lock *ffl,
2387                                   struct file_lock *fl)
2388 {
2389         switch (ffl->type) {
2390         case F_UNLCK:
2391                 break;
2392
2393         case F_RDLCK:
2394         case F_WRLCK:
2395                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2396                     ffl->end < ffl->start)
2397                         return -EIO;
2398
2399                 fl->fl_start = ffl->start;
2400                 fl->fl_end = ffl->end;
2401
2402                 /*
2403                  * Convert pid into init's pid namespace.  The locks API will
2404                  * translate it into the caller's pid namespace.
2405                  */
2406                 rcu_read_lock();
2407                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2408                 rcu_read_unlock();
2409                 break;
2410
2411         default:
2412                 return -EIO;
2413         }
2414         fl->fl_type = ffl->type;
2415         return 0;
2416 }
2417
2418 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2419                          const struct file_lock *fl, int opcode, pid_t pid,
2420                          int flock, struct fuse_lk_in *inarg)
2421 {
2422         struct inode *inode = file_inode(file);
2423         struct fuse_conn *fc = get_fuse_conn(inode);
2424         struct fuse_file *ff = file->private_data;
2425
2426         memset(inarg, 0, sizeof(*inarg));
2427         inarg->fh = ff->fh;
2428         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2429         inarg->lk.start = fl->fl_start;
2430         inarg->lk.end = fl->fl_end;
2431         inarg->lk.type = fl->fl_type;
2432         inarg->lk.pid = pid;
2433         if (flock)
2434                 inarg->lk_flags |= FUSE_LK_FLOCK;
2435         args->opcode = opcode;
2436         args->nodeid = get_node_id(inode);
2437         args->in_numargs = 1;
2438         args->in_args[0].size = sizeof(*inarg);
2439         args->in_args[0].value = inarg;
2440 }
2441
2442 static int fuse_getlk(struct file *file, struct file_lock *fl)
2443 {
2444         struct inode *inode = file_inode(file);
2445         struct fuse_mount *fm = get_fuse_mount(inode);
2446         FUSE_ARGS(args);
2447         struct fuse_lk_in inarg;
2448         struct fuse_lk_out outarg;
2449         int err;
2450
2451         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2452         args.out_numargs = 1;
2453         args.out_args[0].size = sizeof(outarg);
2454         args.out_args[0].value = &outarg;
2455         err = fuse_simple_request(fm, &args);
2456         if (!err)
2457                 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2458
2459         return err;
2460 }
2461
2462 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2463 {
2464         struct inode *inode = file_inode(file);
2465         struct fuse_mount *fm = get_fuse_mount(inode);
2466         FUSE_ARGS(args);
2467         struct fuse_lk_in inarg;
2468         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2469         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2470         pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2471         int err;
2472
2473         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2474                 /* NLM needs asynchronous locks, which we don't support yet */
2475                 return -ENOLCK;
2476         }
2477
2478         /* Unlock on close is handled by the flush method */
2479         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2480                 return 0;
2481
2482         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2483         err = fuse_simple_request(fm, &args);
2484
2485         /* locking is restartable */
2486         if (err == -EINTR)
2487                 err = -ERESTARTSYS;
2488
2489         return err;
2490 }
2491
2492 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2493 {
2494         struct inode *inode = file_inode(file);
2495         struct fuse_conn *fc = get_fuse_conn(inode);
2496         int err;
2497
2498         if (cmd == F_CANCELLK) {
2499                 err = 0;
2500         } else if (cmd == F_GETLK) {
2501                 if (fc->no_lock) {
2502                         posix_test_lock(file, fl);
2503                         err = 0;
2504                 } else
2505                         err = fuse_getlk(file, fl);
2506         } else {
2507                 if (fc->no_lock)
2508                         err = posix_lock_file(file, fl, NULL);
2509                 else
2510                         err = fuse_setlk(file, fl, 0);
2511         }
2512         return err;
2513 }
2514
2515 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2516 {
2517         struct inode *inode = file_inode(file);
2518         struct fuse_conn *fc = get_fuse_conn(inode);
2519         int err;
2520
2521         if (fc->no_flock) {
2522                 err = locks_lock_file_wait(file, fl);
2523         } else {
2524                 struct fuse_file *ff = file->private_data;
2525
2526                 /* emulate flock with POSIX locks */
2527                 ff->flock = true;
2528                 err = fuse_setlk(file, fl, 1);
2529         }
2530
2531         return err;
2532 }
2533
2534 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2535 {
2536         struct inode *inode = mapping->host;
2537         struct fuse_mount *fm = get_fuse_mount(inode);
2538         FUSE_ARGS(args);
2539         struct fuse_bmap_in inarg;
2540         struct fuse_bmap_out outarg;
2541         int err;
2542
2543         if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2544                 return 0;
2545
2546         memset(&inarg, 0, sizeof(inarg));
2547         inarg.block = block;
2548         inarg.blocksize = inode->i_sb->s_blocksize;
2549         args.opcode = FUSE_BMAP;
2550         args.nodeid = get_node_id(inode);
2551         args.in_numargs = 1;
2552         args.in_args[0].size = sizeof(inarg);
2553         args.in_args[0].value = &inarg;
2554         args.out_numargs = 1;
2555         args.out_args[0].size = sizeof(outarg);
2556         args.out_args[0].value = &outarg;
2557         err = fuse_simple_request(fm, &args);
2558         if (err == -ENOSYS)
2559                 fm->fc->no_bmap = 1;
2560
2561         return err ? 0 : outarg.block;
2562 }
2563
2564 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2565 {
2566         struct inode *inode = file->f_mapping->host;
2567         struct fuse_mount *fm = get_fuse_mount(inode);
2568         struct fuse_file *ff = file->private_data;
2569         FUSE_ARGS(args);
2570         struct fuse_lseek_in inarg = {
2571                 .fh = ff->fh,
2572                 .offset = offset,
2573                 .whence = whence
2574         };
2575         struct fuse_lseek_out outarg;
2576         int err;
2577
2578         if (fm->fc->no_lseek)
2579                 goto fallback;
2580
2581         args.opcode = FUSE_LSEEK;
2582         args.nodeid = ff->nodeid;
2583         args.in_numargs = 1;
2584         args.in_args[0].size = sizeof(inarg);
2585         args.in_args[0].value = &inarg;
2586         args.out_numargs = 1;
2587         args.out_args[0].size = sizeof(outarg);
2588         args.out_args[0].value = &outarg;
2589         err = fuse_simple_request(fm, &args);
2590         if (err) {
2591                 if (err == -ENOSYS) {
2592                         fm->fc->no_lseek = 1;
2593                         goto fallback;
2594                 }
2595                 return err;
2596         }
2597
2598         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2599
2600 fallback:
2601         err = fuse_update_attributes(inode, file);
2602         if (!err)
2603                 return generic_file_llseek(file, offset, whence);
2604         else
2605                 return err;
2606 }
2607
2608 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2609 {
2610         loff_t retval;
2611         struct inode *inode = file_inode(file);
2612
2613         switch (whence) {
2614         case SEEK_SET:
2615         case SEEK_CUR:
2616                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2617                 retval = generic_file_llseek(file, offset, whence);
2618                 break;
2619         case SEEK_END:
2620                 inode_lock(inode);
2621                 retval = fuse_update_attributes(inode, file);
2622                 if (!retval)
2623                         retval = generic_file_llseek(file, offset, whence);
2624                 inode_unlock(inode);
2625                 break;
2626         case SEEK_HOLE:
2627         case SEEK_DATA:
2628                 inode_lock(inode);
2629                 retval = fuse_lseek(file, offset, whence);
2630                 inode_unlock(inode);
2631                 break;
2632         default:
2633                 retval = -EINVAL;
2634         }
2635
2636         return retval;
2637 }
2638
2639 /*
2640  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2641  * ABI was defined to be 'struct iovec' which is different on 32bit
2642  * and 64bit.  Fortunately we can determine which structure the server
2643  * used from the size of the reply.
2644  */
2645 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2646                                      size_t transferred, unsigned count,
2647                                      bool is_compat)
2648 {
2649 #ifdef CONFIG_COMPAT
2650         if (count * sizeof(struct compat_iovec) == transferred) {
2651                 struct compat_iovec *ciov = src;
2652                 unsigned i;
2653
2654                 /*
2655                  * With this interface a 32bit server cannot support
2656                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2657                  * requests
2658                  */
2659                 if (!is_compat)
2660                         return -EINVAL;
2661
2662                 for (i = 0; i < count; i++) {
2663                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2664                         dst[i].iov_len = ciov[i].iov_len;
2665                 }
2666                 return 0;
2667         }
2668 #endif
2669
2670         if (count * sizeof(struct iovec) != transferred)
2671                 return -EIO;
2672
2673         memcpy(dst, src, transferred);
2674         return 0;
2675 }
2676
2677 /* Make sure iov_length() won't overflow */
2678 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2679                                  size_t count)
2680 {
2681         size_t n;
2682         u32 max = fc->max_pages << PAGE_SHIFT;
2683
2684         for (n = 0; n < count; n++, iov++) {
2685                 if (iov->iov_len > (size_t) max)
2686                         return -ENOMEM;
2687                 max -= iov->iov_len;
2688         }
2689         return 0;
2690 }
2691
2692 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2693                                  void *src, size_t transferred, unsigned count,
2694                                  bool is_compat)
2695 {
2696         unsigned i;
2697         struct fuse_ioctl_iovec *fiov = src;
2698
2699         if (fc->minor < 16) {
2700                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2701                                                  count, is_compat);
2702         }
2703
2704         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2705                 return -EIO;
2706
2707         for (i = 0; i < count; i++) {
2708                 /* Did the server supply an inappropriate value? */
2709                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2710                     fiov[i].len != (unsigned long) fiov[i].len)
2711                         return -EIO;
2712
2713                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2714                 dst[i].iov_len = (size_t) fiov[i].len;
2715
2716 #ifdef CONFIG_COMPAT
2717                 if (is_compat &&
2718                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2719                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2720                         return -EIO;
2721 #endif
2722         }
2723
2724         return 0;
2725 }
2726
2727
2728 /*
2729  * For ioctls, there is no generic way to determine how much memory
2730  * needs to be read and/or written.  Furthermore, ioctls are allowed
2731  * to dereference the passed pointer, so the parameter requires deep
2732  * copying but FUSE has no idea whatsoever about what to copy in or
2733  * out.
2734  *
2735  * This is solved by allowing FUSE server to retry ioctl with
2736  * necessary in/out iovecs.  Let's assume the ioctl implementation
2737  * needs to read in the following structure.
2738  *
2739  * struct a {
2740  *      char    *buf;
2741  *      size_t  buflen;
2742  * }
2743  *
2744  * On the first callout to FUSE server, inarg->in_size and
2745  * inarg->out_size will be NULL; then, the server completes the ioctl
2746  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2747  * the actual iov array to
2748  *
2749  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2750  *
2751  * which tells FUSE to copy in the requested area and retry the ioctl.
2752  * On the second round, the server has access to the structure and
2753  * from that it can tell what to look for next, so on the invocation,
2754  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2755  *
2756  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2757  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2758  *
2759  * FUSE will copy both struct a and the pointed buffer from the
2760  * process doing the ioctl and retry ioctl with both struct a and the
2761  * buffer.
2762  *
2763  * This time, FUSE server has everything it needs and completes ioctl
2764  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2765  *
2766  * Copying data out works the same way.
2767  *
2768  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2769  * automatically initializes in and out iovs by decoding @cmd with
2770  * _IOC_* macros and the server is not allowed to request RETRY.  This
2771  * limits ioctl data transfers to well-formed ioctls and is the forced
2772  * behavior for all FUSE servers.
2773  */
2774 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2775                    unsigned int flags)
2776 {
2777         struct fuse_file *ff = file->private_data;
2778         struct fuse_mount *fm = ff->fm;
2779         struct fuse_ioctl_in inarg = {
2780                 .fh = ff->fh,
2781                 .cmd = cmd,
2782                 .arg = arg,
2783                 .flags = flags
2784         };
2785         struct fuse_ioctl_out outarg;
2786         struct iovec *iov_page = NULL;
2787         struct iovec *in_iov = NULL, *out_iov = NULL;
2788         unsigned int in_iovs = 0, out_iovs = 0, max_pages;
2789         size_t in_size, out_size, c;
2790         ssize_t transferred;
2791         int err, i;
2792         struct iov_iter ii;
2793         struct fuse_args_pages ap = {};
2794
2795 #if BITS_PER_LONG == 32
2796         inarg.flags |= FUSE_IOCTL_32BIT;
2797 #else
2798         if (flags & FUSE_IOCTL_COMPAT) {
2799                 inarg.flags |= FUSE_IOCTL_32BIT;
2800 #ifdef CONFIG_X86_X32
2801                 if (in_x32_syscall())
2802                         inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2803 #endif
2804         }
2805 #endif
2806
2807         /* assume all the iovs returned by client always fits in a page */
2808         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2809
2810         err = -ENOMEM;
2811         ap.pages = fuse_pages_alloc(fm->fc->max_pages, GFP_KERNEL, &ap.descs);
2812         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2813         if (!ap.pages || !iov_page)
2814                 goto out;
2815
2816         fuse_page_descs_length_init(ap.descs, 0, fm->fc->max_pages);
2817
2818         /*
2819          * If restricted, initialize IO parameters as encoded in @cmd.
2820          * RETRY from server is not allowed.
2821          */
2822         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2823                 struct iovec *iov = iov_page;
2824
2825                 iov->iov_base = (void __user *)arg;
2826
2827                 switch (cmd) {
2828                 case FS_IOC_GETFLAGS:
2829                 case FS_IOC_SETFLAGS:
2830                         iov->iov_len = sizeof(int);
2831                         break;
2832                 default:
2833                         iov->iov_len = _IOC_SIZE(cmd);
2834                         break;
2835                 }
2836
2837                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2838                         in_iov = iov;
2839                         in_iovs = 1;
2840                 }
2841
2842                 if (_IOC_DIR(cmd) & _IOC_READ) {
2843                         out_iov = iov;
2844                         out_iovs = 1;
2845                 }
2846         }
2847
2848  retry:
2849         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2850         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2851
2852         /*
2853          * Out data can be used either for actual out data or iovs,
2854          * make sure there always is at least one page.
2855          */
2856         out_size = max_t(size_t, out_size, PAGE_SIZE);
2857         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2858
2859         /* make sure there are enough buffer pages and init request with them */
2860         err = -ENOMEM;
2861         if (max_pages > fm->fc->max_pages)
2862                 goto out;
2863         while (ap.num_pages < max_pages) {
2864                 ap.pages[ap.num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2865                 if (!ap.pages[ap.num_pages])
2866                         goto out;
2867                 ap.num_pages++;
2868         }
2869
2870
2871         /* okay, let's send it to the client */
2872         ap.args.opcode = FUSE_IOCTL;
2873         ap.args.nodeid = ff->nodeid;
2874         ap.args.in_numargs = 1;
2875         ap.args.in_args[0].size = sizeof(inarg);
2876         ap.args.in_args[0].value = &inarg;
2877         if (in_size) {
2878                 ap.args.in_numargs++;
2879                 ap.args.in_args[1].size = in_size;
2880                 ap.args.in_pages = true;
2881
2882                 err = -EFAULT;
2883                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2884                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2885                         c = copy_page_from_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2886                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2887                                 goto out;
2888                 }
2889         }
2890
2891         ap.args.out_numargs = 2;
2892         ap.args.out_args[0].size = sizeof(outarg);
2893         ap.args.out_args[0].value = &outarg;
2894         ap.args.out_args[1].size = out_size;
2895         ap.args.out_pages = true;
2896         ap.args.out_argvar = true;
2897
2898         transferred = fuse_simple_request(fm, &ap.args);
2899         err = transferred;
2900         if (transferred < 0)
2901                 goto out;
2902
2903         /* did it ask for retry? */
2904         if (outarg.flags & FUSE_IOCTL_RETRY) {
2905                 void *vaddr;
2906
2907                 /* no retry if in restricted mode */
2908                 err = -EIO;
2909                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2910                         goto out;
2911
2912                 in_iovs = outarg.in_iovs;
2913                 out_iovs = outarg.out_iovs;
2914
2915                 /*
2916                  * Make sure things are in boundary, separate checks
2917                  * are to protect against overflow.
2918                  */
2919                 err = -ENOMEM;
2920                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2921                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2922                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2923                         goto out;
2924
2925                 vaddr = kmap_atomic(ap.pages[0]);
2926                 err = fuse_copy_ioctl_iovec(fm->fc, iov_page, vaddr,
2927                                             transferred, in_iovs + out_iovs,
2928                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2929                 kunmap_atomic(vaddr);
2930                 if (err)
2931                         goto out;
2932
2933                 in_iov = iov_page;
2934                 out_iov = in_iov + in_iovs;
2935
2936                 err = fuse_verify_ioctl_iov(fm->fc, in_iov, in_iovs);
2937                 if (err)
2938                         goto out;
2939
2940                 err = fuse_verify_ioctl_iov(fm->fc, out_iov, out_iovs);
2941                 if (err)
2942                         goto out;
2943
2944                 goto retry;
2945         }
2946
2947         err = -EIO;
2948         if (transferred > inarg.out_size)
2949                 goto out;
2950
2951         err = -EFAULT;
2952         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2953         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2954                 c = copy_page_to_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2955                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2956                         goto out;
2957         }
2958         err = 0;
2959  out:
2960         free_page((unsigned long) iov_page);
2961         while (ap.num_pages)
2962                 __free_page(ap.pages[--ap.num_pages]);
2963         kfree(ap.pages);
2964
2965         return err ? err : outarg.result;
2966 }
2967 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2968
2969 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2970                        unsigned long arg, unsigned int flags)
2971 {
2972         struct inode *inode = file_inode(file);
2973         struct fuse_conn *fc = get_fuse_conn(inode);
2974
2975         if (!fuse_allow_current_process(fc))
2976                 return -EACCES;
2977
2978         if (fuse_is_bad(inode))
2979                 return -EIO;
2980
2981         return fuse_do_ioctl(file, cmd, arg, flags);
2982 }
2983
2984 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2985                             unsigned long arg)
2986 {
2987         return fuse_ioctl_common(file, cmd, arg, 0);
2988 }
2989
2990 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2991                                    unsigned long arg)
2992 {
2993         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2994 }
2995
2996 /*
2997  * All files which have been polled are linked to RB tree
2998  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2999  * find the matching one.
3000  */
3001 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
3002                                               struct rb_node **parent_out)
3003 {
3004         struct rb_node **link = &fc->polled_files.rb_node;
3005         struct rb_node *last = NULL;
3006
3007         while (*link) {
3008                 struct fuse_file *ff;
3009
3010                 last = *link;
3011                 ff = rb_entry(last, struct fuse_file, polled_node);
3012
3013                 if (kh < ff->kh)
3014                         link = &last->rb_left;
3015                 else if (kh > ff->kh)
3016                         link = &last->rb_right;
3017                 else
3018                         return link;
3019         }
3020
3021         if (parent_out)
3022                 *parent_out = last;
3023         return link;
3024 }
3025
3026 /*
3027  * The file is about to be polled.  Make sure it's on the polled_files
3028  * RB tree.  Note that files once added to the polled_files tree are
3029  * not removed before the file is released.  This is because a file
3030  * polled once is likely to be polled again.
3031  */
3032 static void fuse_register_polled_file(struct fuse_conn *fc,
3033                                       struct fuse_file *ff)
3034 {
3035         spin_lock(&fc->lock);
3036         if (RB_EMPTY_NODE(&ff->polled_node)) {
3037                 struct rb_node **link, *parent;
3038
3039                 link = fuse_find_polled_node(fc, ff->kh, &parent);
3040                 BUG_ON(*link);
3041                 rb_link_node(&ff->polled_node, parent, link);
3042                 rb_insert_color(&ff->polled_node, &fc->polled_files);
3043         }
3044         spin_unlock(&fc->lock);
3045 }
3046
3047 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
3048 {
3049         struct fuse_file *ff = file->private_data;
3050         struct fuse_mount *fm = ff->fm;
3051         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
3052         struct fuse_poll_out outarg;
3053         FUSE_ARGS(args);
3054         int err;
3055
3056         if (fm->fc->no_poll)
3057                 return DEFAULT_POLLMASK;
3058
3059         poll_wait(file, &ff->poll_wait, wait);
3060         inarg.events = mangle_poll(poll_requested_events(wait));
3061
3062         /*
3063          * Ask for notification iff there's someone waiting for it.
3064          * The client may ignore the flag and always notify.
3065          */
3066         if (waitqueue_active(&ff->poll_wait)) {
3067                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
3068                 fuse_register_polled_file(fm->fc, ff);
3069         }
3070
3071         args.opcode = FUSE_POLL;
3072         args.nodeid = ff->nodeid;
3073         args.in_numargs = 1;
3074         args.in_args[0].size = sizeof(inarg);
3075         args.in_args[0].value = &inarg;
3076         args.out_numargs = 1;
3077         args.out_args[0].size = sizeof(outarg);
3078         args.out_args[0].value = &outarg;
3079         err = fuse_simple_request(fm, &args);
3080
3081         if (!err)
3082                 return demangle_poll(outarg.revents);
3083         if (err == -ENOSYS) {
3084                 fm->fc->no_poll = 1;
3085                 return DEFAULT_POLLMASK;
3086         }
3087         return EPOLLERR;
3088 }
3089 EXPORT_SYMBOL_GPL(fuse_file_poll);
3090
3091 /*
3092  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
3093  * wakes up the poll waiters.
3094  */
3095 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
3096                             struct fuse_notify_poll_wakeup_out *outarg)
3097 {
3098         u64 kh = outarg->kh;
3099         struct rb_node **link;
3100
3101         spin_lock(&fc->lock);
3102
3103         link = fuse_find_polled_node(fc, kh, NULL);
3104         if (*link) {
3105                 struct fuse_file *ff;
3106
3107                 ff = rb_entry(*link, struct fuse_file, polled_node);
3108                 wake_up_interruptible_sync(&ff->poll_wait);
3109         }
3110
3111         spin_unlock(&fc->lock);
3112         return 0;
3113 }
3114
3115 static void fuse_do_truncate(struct file *file)
3116 {
3117         struct inode *inode = file->f_mapping->host;
3118         struct iattr attr;
3119
3120         attr.ia_valid = ATTR_SIZE;
3121         attr.ia_size = i_size_read(inode);
3122
3123         attr.ia_file = file;
3124         attr.ia_valid |= ATTR_FILE;
3125
3126         fuse_do_setattr(file_dentry(file), &attr, file);
3127 }
3128
3129 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
3130 {
3131         return round_up(off, fc->max_pages << PAGE_SHIFT);
3132 }
3133
3134 static ssize_t
3135 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3136 {
3137         DECLARE_COMPLETION_ONSTACK(wait);
3138         ssize_t ret = 0;
3139         struct file *file = iocb->ki_filp;
3140         struct fuse_file *ff = file->private_data;
3141         loff_t pos = 0;
3142         struct inode *inode;
3143         loff_t i_size;
3144         size_t count = iov_iter_count(iter), shortened = 0;
3145         loff_t offset = iocb->ki_pos;
3146         struct fuse_io_priv *io;
3147
3148         pos = offset;
3149         inode = file->f_mapping->host;
3150         i_size = i_size_read(inode);
3151
3152         if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
3153                 return 0;
3154
3155         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
3156         if (!io)
3157                 return -ENOMEM;
3158         spin_lock_init(&io->lock);
3159         kref_init(&io->refcnt);
3160         io->reqs = 1;
3161         io->bytes = -1;
3162         io->size = 0;
3163         io->offset = offset;
3164         io->write = (iov_iter_rw(iter) == WRITE);
3165         io->err = 0;
3166         /*
3167          * By default, we want to optimize all I/Os with async request
3168          * submission to the client filesystem if supported.
3169          */
3170         io->async = ff->fm->fc->async_dio;
3171         io->iocb = iocb;
3172         io->blocking = is_sync_kiocb(iocb);
3173
3174         /* optimization for short read */
3175         if (io->async && !io->write && offset + count > i_size) {
3176                 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
3177                 shortened = count - iov_iter_count(iter);
3178                 count -= shortened;
3179         }
3180
3181         /*
3182          * We cannot asynchronously extend the size of a file.
3183          * In such case the aio will behave exactly like sync io.
3184          */
3185         if ((offset + count > i_size) && io->write)
3186                 io->blocking = true;
3187
3188         if (io->async && io->blocking) {
3189                 /*
3190                  * Additional reference to keep io around after
3191                  * calling fuse_aio_complete()
3192                  */
3193                 kref_get(&io->refcnt);
3194                 io->done = &wait;
3195         }
3196
3197         if (iov_iter_rw(iter) == WRITE) {
3198                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
3199                 fuse_invalidate_attr(inode);
3200         } else {
3201                 ret = __fuse_direct_read(io, iter, &pos);
3202         }
3203         iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
3204
3205         if (io->async) {
3206                 bool blocking = io->blocking;
3207
3208                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3209
3210                 /* we have a non-extending, async request, so return */
3211                 if (!blocking)
3212                         return -EIOCBQUEUED;
3213
3214                 wait_for_completion(&wait);
3215                 ret = fuse_get_res_by_io(io);
3216         }
3217
3218         kref_put(&io->refcnt, fuse_io_release);
3219
3220         if (iov_iter_rw(iter) == WRITE) {
3221                 if (ret > 0)
3222                         fuse_write_update_size(inode, pos);
3223                 else if (ret < 0 && offset + count > i_size)
3224                         fuse_do_truncate(file);
3225         }
3226
3227         return ret;
3228 }
3229
3230 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3231 {
3232         int err = filemap_write_and_wait_range(inode->i_mapping, start, end);
3233
3234         if (!err)
3235                 fuse_sync_writes(inode);
3236
3237         return err;
3238 }
3239
3240 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3241                                 loff_t length)
3242 {
3243         struct fuse_file *ff = file->private_data;
3244         struct inode *inode = file_inode(file);
3245         struct fuse_inode *fi = get_fuse_inode(inode);
3246         struct fuse_mount *fm = ff->fm;
3247         FUSE_ARGS(args);
3248         struct fuse_fallocate_in inarg = {
3249                 .fh = ff->fh,
3250                 .offset = offset,
3251                 .length = length,
3252                 .mode = mode
3253         };
3254         int err;
3255         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
3256                            (mode & FALLOC_FL_PUNCH_HOLE);
3257
3258         bool block_faults = FUSE_IS_DAX(inode) && lock_inode;
3259
3260         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3261                 return -EOPNOTSUPP;
3262
3263         if (fm->fc->no_fallocate)
3264                 return -EOPNOTSUPP;
3265
3266         if (lock_inode) {
3267                 inode_lock(inode);
3268                 if (block_faults) {
3269                         down_write(&fi->i_mmap_sem);
3270                         err = fuse_dax_break_layouts(inode, 0, 0);
3271                         if (err)
3272                                 goto out;
3273                 }
3274
3275                 if (mode & FALLOC_FL_PUNCH_HOLE) {
3276                         loff_t endbyte = offset + length - 1;
3277
3278                         err = fuse_writeback_range(inode, offset, endbyte);
3279                         if (err)
3280                                 goto out;
3281                 }
3282         }
3283
3284         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3285             offset + length > i_size_read(inode)) {
3286                 err = inode_newsize_ok(inode, offset + length);
3287                 if (err)
3288                         goto out;
3289         }
3290
3291         if (!(mode & FALLOC_FL_KEEP_SIZE))
3292                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3293
3294         args.opcode = FUSE_FALLOCATE;
3295         args.nodeid = ff->nodeid;
3296         args.in_numargs = 1;
3297         args.in_args[0].size = sizeof(inarg);
3298         args.in_args[0].value = &inarg;
3299         err = fuse_simple_request(fm, &args);
3300         if (err == -ENOSYS) {
3301                 fm->fc->no_fallocate = 1;
3302                 err = -EOPNOTSUPP;
3303         }
3304         if (err)
3305                 goto out;
3306
3307         /* we could have extended the file */
3308         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3309                 bool changed = fuse_write_update_size(inode, offset + length);
3310
3311                 if (changed && fm->fc->writeback_cache)
3312                         file_update_time(file);
3313         }
3314
3315         if (mode & FALLOC_FL_PUNCH_HOLE)
3316                 truncate_pagecache_range(inode, offset, offset + length - 1);
3317
3318         fuse_invalidate_attr(inode);
3319
3320 out:
3321         if (!(mode & FALLOC_FL_KEEP_SIZE))
3322                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3323
3324         if (block_faults)
3325                 up_write(&fi->i_mmap_sem);
3326
3327         if (lock_inode)
3328                 inode_unlock(inode);
3329
3330         return err;
3331 }
3332
3333 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3334                                       struct file *file_out, loff_t pos_out,
3335                                       size_t len, unsigned int flags)
3336 {
3337         struct fuse_file *ff_in = file_in->private_data;
3338         struct fuse_file *ff_out = file_out->private_data;
3339         struct inode *inode_in = file_inode(file_in);
3340         struct inode *inode_out = file_inode(file_out);
3341         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3342         struct fuse_mount *fm = ff_in->fm;
3343         struct fuse_conn *fc = fm->fc;
3344         FUSE_ARGS(args);
3345         struct fuse_copy_file_range_in inarg = {
3346                 .fh_in = ff_in->fh,
3347                 .off_in = pos_in,
3348                 .nodeid_out = ff_out->nodeid,
3349                 .fh_out = ff_out->fh,
3350                 .off_out = pos_out,
3351                 .len = len,
3352                 .flags = flags
3353         };
3354         struct fuse_write_out outarg;
3355         ssize_t err;
3356         /* mark unstable when write-back is not used, and file_out gets
3357          * extended */
3358         bool is_unstable = (!fc->writeback_cache) &&
3359                            ((pos_out + len) > inode_out->i_size);
3360
3361         if (fc->no_copy_file_range)
3362                 return -EOPNOTSUPP;
3363
3364         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3365                 return -EXDEV;
3366
3367         inode_lock(inode_in);
3368         err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3369         inode_unlock(inode_in);
3370         if (err)
3371                 return err;
3372
3373         inode_lock(inode_out);
3374
3375         err = file_modified(file_out);
3376         if (err)
3377                 goto out;
3378
3379         /*
3380          * Write out dirty pages in the destination file before sending the COPY
3381          * request to userspace.  After the request is completed, truncate off
3382          * pages (including partial ones) from the cache that have been copied,
3383          * since these contain stale data at that point.
3384          *
3385          * This should be mostly correct, but if the COPY writes to partial
3386          * pages (at the start or end) and the parts not covered by the COPY are
3387          * written through a memory map after calling fuse_writeback_range(),
3388          * then these partial page modifications will be lost on truncation.
3389          *
3390          * It is unlikely that someone would rely on such mixed style
3391          * modifications.  Yet this does give less guarantees than if the
3392          * copying was performed with write(2).
3393          *
3394          * To fix this a i_mmap_sem style lock could be used to prevent new
3395          * faults while the copy is ongoing.
3396          */
3397         err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3398         if (err)
3399                 goto out;
3400
3401         if (is_unstable)
3402                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3403
3404         args.opcode = FUSE_COPY_FILE_RANGE;
3405         args.nodeid = ff_in->nodeid;
3406         args.in_numargs = 1;
3407         args.in_args[0].size = sizeof(inarg);
3408         args.in_args[0].value = &inarg;
3409         args.out_numargs = 1;
3410         args.out_args[0].size = sizeof(outarg);
3411         args.out_args[0].value = &outarg;
3412         err = fuse_simple_request(fm, &args);
3413         if (err == -ENOSYS) {
3414                 fc->no_copy_file_range = 1;
3415                 err = -EOPNOTSUPP;
3416         }
3417         if (err)
3418                 goto out;
3419
3420         truncate_inode_pages_range(inode_out->i_mapping,
3421                                    ALIGN_DOWN(pos_out, PAGE_SIZE),
3422                                    ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3423
3424         if (fc->writeback_cache) {
3425                 fuse_write_update_size(inode_out, pos_out + outarg.size);
3426                 file_update_time(file_out);
3427         }
3428
3429         fuse_invalidate_attr(inode_out);
3430
3431         err = outarg.size;
3432 out:
3433         if (is_unstable)
3434                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3435
3436         inode_unlock(inode_out);
3437         file_accessed(file_in);
3438
3439         return err;
3440 }
3441
3442 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3443                                     struct file *dst_file, loff_t dst_off,
3444                                     size_t len, unsigned int flags)
3445 {
3446         ssize_t ret;
3447
3448         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3449                                      len, flags);
3450
3451         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3452                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3453                                               dst_off, len, flags);
3454         return ret;
3455 }
3456
3457 static const struct file_operations fuse_file_operations = {
3458         .llseek         = fuse_file_llseek,
3459         .read_iter      = fuse_file_read_iter,
3460         .write_iter     = fuse_file_write_iter,
3461         .mmap           = fuse_file_mmap,
3462         .open           = fuse_open,
3463         .flush          = fuse_flush,
3464         .release        = fuse_release,
3465         .fsync          = fuse_fsync,
3466         .lock           = fuse_file_lock,
3467         .get_unmapped_area = thp_get_unmapped_area,
3468         .flock          = fuse_file_flock,
3469         .splice_read    = generic_file_splice_read,
3470         .splice_write   = iter_file_splice_write,
3471         .unlocked_ioctl = fuse_file_ioctl,
3472         .compat_ioctl   = fuse_file_compat_ioctl,
3473         .poll           = fuse_file_poll,
3474         .fallocate      = fuse_file_fallocate,
3475         .copy_file_range = fuse_copy_file_range,
3476 };
3477
3478 static const struct address_space_operations fuse_file_aops  = {
3479         .readpage       = fuse_readpage,
3480         .readahead      = fuse_readahead,
3481         .writepage      = fuse_writepage,
3482         .writepages     = fuse_writepages,
3483         .launder_page   = fuse_launder_page,
3484         .set_page_dirty = __set_page_dirty_nobuffers,
3485         .bmap           = fuse_bmap,
3486         .direct_IO      = fuse_direct_IO,
3487         .write_begin    = fuse_write_begin,
3488         .write_end      = fuse_write_end,
3489 };
3490
3491 void fuse_init_file_inode(struct inode *inode)
3492 {
3493         struct fuse_inode *fi = get_fuse_inode(inode);
3494
3495         inode->i_fop = &fuse_file_operations;
3496         inode->i_data.a_ops = &fuse_file_aops;
3497
3498         INIT_LIST_HEAD(&fi->write_files);
3499         INIT_LIST_HEAD(&fi->queued_writes);
3500         fi->writectr = 0;
3501         init_waitqueue_head(&fi->page_waitq);
3502         fi->writepages = RB_ROOT;
3503
3504         if (IS_ENABLED(CONFIG_FUSE_DAX))
3505                 fuse_dax_inode_init(inode);
3506 }