fs/fuse/file.c

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