fs/fuse/file.c

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