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