fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   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/swap.h>
  18 #include <linux/falloc.h>
  19 #include <linux/uio.h>
  20 #include <linux/fs.h>
  21 #include <linux/filelock.h>
  22 #include <linux/splice.h>
  23
  24 static int fuse_send_open(struct fuse_mount *fm, u64 nodeid,
  25                           unsigned int open_flags, int opcode,
  26                           struct fuse_open_out *outargp)
  27 {
  28         struct fuse_open_in inarg;
  29         FUSE_ARGS(args);
  30
  31         memset(&inarg, 0, sizeof(inarg));
  32         inarg.flags = open_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  33         if (!fm->fc->atomic_o_trunc)
  34                 inarg.flags &= ~O_TRUNC;
  35
  36         if (fm->fc->handle_killpriv_v2 &&
  37             (inarg.flags & O_TRUNC) && !capable(CAP_FSETID)) {
  38                 inarg.open_flags |= FUSE_OPEN_KILL_SUIDGID;
  39         }
  40
  41         args.opcode = opcode;
  42         args.nodeid = nodeid;
  43         args.in_numargs = 1;
  44         args.in_args[0].size = sizeof(inarg);
  45         args.in_args[0].value = &inarg;
  46         args.out_numargs = 1;
  47         args.out_args[0].size = sizeof(*outargp);
  48         args.out_args[0].value = outargp;
  49
  50         return fuse_simple_request(fm, &args);
  51 }
  52
  53 struct fuse_release_args {
  54         struct fuse_args args;
  55         struct fuse_release_in inarg;
  56         struct inode *inode;
  57 };
  58
  59 struct fuse_file *fuse_file_alloc(struct fuse_mount *fm)
  60 {
  61         struct fuse_file *ff;
  62
  63         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
  64         if (unlikely(!ff))
  65                 return NULL;
  66
  67         ff->fm = fm;
  68         ff->release_args = kzalloc(sizeof(*ff->release_args),
  69                                    GFP_KERNEL_ACCOUNT);
  70         if (!ff->release_args) {
  71                 kfree(ff);
  72                 return NULL;
  73         }
  74
  75         INIT_LIST_HEAD(&ff->write_entry);
  76         mutex_init(&ff->readdir.lock);
  77         refcount_set(&ff->count, 1);
  78         RB_CLEAR_NODE(&ff->polled_node);
  79         init_waitqueue_head(&ff->poll_wait);
  80
  81         ff->kh = atomic64_inc_return(&fm->fc->khctr);
  82
  83         return ff;
  84 }
  85
  86 void fuse_file_free(struct fuse_file *ff)
  87 {
  88         kfree(ff->release_args);
  89         mutex_destroy(&ff->readdir.lock);
  90         kfree(ff);
  91 }
  92
  93 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  94 {
  95         refcount_inc(&ff->count);
  96         return ff;
  97 }
  98
  99 static void fuse_release_end(struct fuse_mount *fm, struct fuse_args *args,
 100                              int error)
 101 {
 102         struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
 103
 104         iput(ra->inode);
 105         kfree(ra);
 106 }
 107
 108 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
 109 {
 110         if (refcount_dec_and_test(&ff->count)) {
 111                 struct fuse_args *args = &ff->release_args->args;
 112
 113                 if (isdir ? ff->fm->fc->no_opendir : ff->fm->fc->no_open) {
 114                         /* Do nothing when client does not implement 'open' */
 115                         fuse_release_end(ff->fm, args, 0);
 116                 } else if (sync) {
 117                         fuse_simple_request(ff->fm, args);
 118                         fuse_release_end(ff->fm, args, 0);
 119                 } else {
 120                         args->end = fuse_release_end;
 121                         if (fuse_simple_background(ff->fm, args,
 122                                                    GFP_KERNEL | __GFP_NOFAIL))
 123                                 fuse_release_end(ff->fm, args, -ENOTCONN);
 124                 }
 125                 kfree(ff);
 126         }
 127 }
 128
 129 struct fuse_file *fuse_file_open(struct fuse_mount *fm, u64 nodeid,
 130                                  unsigned int open_flags, bool isdir)
 131 {
 132         struct fuse_conn *fc = fm->fc;
 133         struct fuse_file *ff;
 134         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 135
 136         ff = fuse_file_alloc(fm);
 137         if (!ff)
 138                 return ERR_PTR(-ENOMEM);
 139
 140         ff->fh = 0;
 141         /* Default for no-open */
 142         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
 143         if (isdir ? !fc->no_opendir : !fc->no_open) {
 144                 struct fuse_open_out outarg;
 145                 int err;
 146
 147                 err = fuse_send_open(fm, nodeid, open_flags, opcode, &outarg);
 148                 if (!err) {
 149                         ff->fh = outarg.fh;
 150                         ff->open_flags = outarg.open_flags;
 151
 152                 } else if (err != -ENOSYS) {
 153                         fuse_file_free(ff);
 154                         return ERR_PTR(err);
 155                 } else {
 156                         if (isdir)
 157                                 fc->no_opendir = 1;
 158                         else
 159                                 fc->no_open = 1;
 160                 }
 161         }
 162
 163         if (isdir)
 164                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 165
 166         ff->nodeid = nodeid;
 167
 168         return ff;
 169 }
 170
 171 int fuse_do_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
 172                  bool isdir)
 173 {
 174         struct fuse_file *ff = fuse_file_open(fm, nodeid, file->f_flags, isdir);
 175
 176         if (!IS_ERR(ff))
 177                 file->private_data = ff;
 178
 179         return PTR_ERR_OR_ZERO(ff);
 180 }
 181 EXPORT_SYMBOL_GPL(fuse_do_open);
 182
 183 static void fuse_link_write_file(struct file *file)
 184 {
 185         struct inode *inode = file_inode(file);
 186         struct fuse_inode *fi = get_fuse_inode(inode);
 187         struct fuse_file *ff = file->private_data;
 188         /*
 189          * file may be written through mmap, so chain it onto the
 190          * inodes's write_file list
 191          */
 192         spin_lock(&fi->lock);
 193         if (list_empty(&ff->write_entry))
 194                 list_add(&ff->write_entry, &fi->write_files);
 195         spin_unlock(&fi->lock);
 196 }
 197
 198 void fuse_finish_open(struct inode *inode, struct file *file)
 199 {
 200         struct fuse_file *ff = file->private_data;
 201         struct fuse_conn *fc = get_fuse_conn(inode);
 202
 203         if (ff->open_flags & FOPEN_STREAM)
 204                 stream_open(inode, file);
 205         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 206                 nonseekable_open(inode, file);
 207
 208         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 209                 struct fuse_inode *fi = get_fuse_inode(inode);
 210
 211                 spin_lock(&fi->lock);
 212                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 213                 i_size_write(inode, 0);
 214                 spin_unlock(&fi->lock);
 215                 file_update_time(file);
 216                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
 217         }
 218         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 219                 fuse_link_write_file(file);
 220 }
 221
 222 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 223 {
 224         struct fuse_mount *fm = get_fuse_mount(inode);
 225         struct fuse_conn *fc = fm->fc;
 226         int err;
 227         bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
 228                           fc->atomic_o_trunc &&
 229                           fc->writeback_cache;
 230         bool dax_truncate = (file->f_flags & O_TRUNC) &&
 231                           fc->atomic_o_trunc && FUSE_IS_DAX(inode);
 232
 233         if (fuse_is_bad(inode))
 234                 return -EIO;
 235
 236         err = generic_file_open(inode, file);
 237         if (err)
 238                 return err;
 239
 240         if (is_wb_truncate || dax_truncate)
 241                 inode_lock(inode);
 242
 243         if (dax_truncate) {
 244                 filemap_invalidate_lock(inode->i_mapping);
 245                 err = fuse_dax_break_layouts(inode, 0, 0);
 246                 if (err)
 247                         goto out_inode_unlock;
 248         }
 249
 250         if (is_wb_truncate || dax_truncate)
 251                 fuse_set_nowrite(inode);
 252
 253         err = fuse_do_open(fm, get_node_id(inode), file, isdir);
 254         if (!err)
 255                 fuse_finish_open(inode, file);
 256
 257         if (is_wb_truncate || dax_truncate)
 258                 fuse_release_nowrite(inode);
 259         if (!err) {
 260                 struct fuse_file *ff = file->private_data;
 261
 262                 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC))
 263                         truncate_pagecache(inode, 0);
 264                 else if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 265                         invalidate_inode_pages2(inode->i_mapping);
 266         }
 267         if (dax_truncate)
 268                 filemap_invalidate_unlock(inode->i_mapping);
 269 out_inode_unlock:
 270         if (is_wb_truncate || dax_truncate)
 271                 inode_unlock(inode);
 272
 273         return err;
 274 }
 275
 276 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 277                                  unsigned int flags, int opcode)
 278 {
 279         struct fuse_conn *fc = ff->fm->fc;
 280         struct fuse_release_args *ra = ff->release_args;
 281
 282         /* Inode is NULL on error path of fuse_create_open() */
 283         if (likely(fi)) {
 284                 spin_lock(&fi->lock);
 285                 list_del(&ff->write_entry);
 286                 spin_unlock(&fi->lock);
 287         }
 288         spin_lock(&fc->lock);
 289         if (!RB_EMPTY_NODE(&ff->polled_node))
 290                 rb_erase(&ff->polled_node, &fc->polled_files);
 291         spin_unlock(&fc->lock);
 292
 293         wake_up_interruptible_all(&ff->poll_wait);
 294
 295         ra->inarg.fh = ff->fh;
 296         ra->inarg.flags = flags;
 297         ra->args.in_numargs = 1;
 298         ra->args.in_args[0].size = sizeof(struct fuse_release_in);
 299         ra->args.in_args[0].value = &ra->inarg;
 300         ra->args.opcode = opcode;
 301         ra->args.nodeid = ff->nodeid;
 302         ra->args.force = true;
 303         ra->args.nocreds = true;
 304 }
 305
 306 void fuse_file_release(struct inode *inode, struct fuse_file *ff,
 307                        unsigned int open_flags, fl_owner_t id, bool isdir)
 308 {
 309         struct fuse_inode *fi = get_fuse_inode(inode);
 310         struct fuse_release_args *ra = ff->release_args;
 311         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 312
 313         fuse_prepare_release(fi, ff, open_flags, opcode);
 314
 315         if (ff->flock) {
 316                 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 317                 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fm->fc, id);
 318         }
 319         /* Hold inode until release is finished */
 320         ra->inode = igrab(inode);
 321
 322         /*
 323          * Normally this will send the RELEASE request, however if
 324          * some asynchronous READ or WRITE requests are outstanding,
 325          * the sending will be delayed.
 326          *
 327          * Make the release synchronous if this is a fuseblk mount,
 328          * synchronous RELEASE is allowed (and desirable) in this case
 329          * because the server can be trusted not to screw up.
 330          */
 331         fuse_file_put(ff, ff->fm->fc->destroy, isdir);
 332 }
 333
 334 void fuse_release_common(struct file *file, bool isdir)
 335 {
 336         fuse_file_release(file_inode(file), file->private_data, file->f_flags,
 337                           (fl_owner_t) file, isdir);
 338 }
 339
 340 static int fuse_open(struct inode *inode, struct file *file)
 341 {
 342         return fuse_open_common(inode, file, false);
 343 }
 344
 345 static int fuse_release(struct inode *inode, struct file *file)
 346 {
 347         struct fuse_conn *fc = get_fuse_conn(inode);
 348
 349         /*
 350          * Dirty pages might remain despite write_inode_now() call from
 351          * fuse_flush() due to writes racing with the close.
 352          */
 353         if (fc->writeback_cache)
 354                 write_inode_now(inode, 1);
 355
 356         fuse_release_common(file, false);
 357
 358         /* return value is ignored by VFS */
 359         return 0;
 360 }
 361
 362 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff,
 363                        unsigned int flags)
 364 {
 365         WARN_ON(refcount_read(&ff->count) > 1);
 366         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 367         /*
 368          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 369          * synchronous, we are fine with not doing igrab() here"
 370          */
 371         fuse_file_put(ff, true, false);
 372 }
 373 EXPORT_SYMBOL_GPL(fuse_sync_release);
 374
 375 /*
 376  * Scramble the ID space with XTEA, so that the value of the files_struct
 377  * pointer is not exposed to userspace.
 378  */
 379 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 380 {
 381         u32 *k = fc->scramble_key;
 382         u64 v = (unsigned long) id;
 383         u32 v0 = v;
 384         u32 v1 = v >> 32;
 385         u32 sum = 0;
 386         int i;
 387
 388         for (i = 0; i < 32; i++) {
 389                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 390                 sum += 0x9E3779B9;
 391                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 392         }
 393
 394         return (u64) v0 + ((u64) v1 << 32);
 395 }
 396
 397 struct fuse_writepage_args {
 398         struct fuse_io_args ia;
 399         struct rb_node writepages_entry;
 400         struct list_head queue_entry;
 401         struct fuse_writepage_args *next;
 402         struct inode *inode;
 403         struct fuse_sync_bucket *bucket;
 404 };
 405
 406 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
 407                                             pgoff_t idx_from, pgoff_t idx_to)
 408 {
 409         struct rb_node *n;
 410
 411         n = fi->writepages.rb_node;
 412
 413         while (n) {
 414                 struct fuse_writepage_args *wpa;
 415                 pgoff_t curr_index;
 416
 417                 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
 418                 WARN_ON(get_fuse_inode(wpa->inode) != fi);
 419                 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
 420                 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
 421                         n = n->rb_right;
 422                 else if (idx_to < curr_index)
 423                         n = n->rb_left;
 424                 else
 425                         return wpa;
 426         }
 427         return NULL;
 428 }
 429
 430 /*
 431  * Check if any page in a range is under writeback
 432  *
 433  * This is currently done by walking the list of writepage requests
 434  * for the inode, which can be pretty inefficient.
 435  */
 436 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 437                                    pgoff_t idx_to)
 438 {
 439         struct fuse_inode *fi = get_fuse_inode(inode);
 440         bool found;
 441
 442         spin_lock(&fi->lock);
 443         found = fuse_find_writeback(fi, idx_from, idx_to);
 444         spin_unlock(&fi->lock);
 445
 446         return found;
 447 }
 448
 449 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 450 {
 451         return fuse_range_is_writeback(inode, index, index);
 452 }
 453
 454 /*
 455  * Wait for page writeback to be completed.
 456  *
 457  * Since fuse doesn't rely on the VM writeback tracking, this has to
 458  * use some other means.
 459  */
 460 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 461 {
 462         struct fuse_inode *fi = get_fuse_inode(inode);
 463
 464         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 465 }
 466
 467 /*
 468  * Wait for all pending writepages on the inode to finish.
 469  *
 470  * This is currently done by blocking further writes with FUSE_NOWRITE
 471  * and waiting for all sent writes to complete.
 472  *
 473  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 474  * could conflict with truncation.
 475  */
 476 static void fuse_sync_writes(struct inode *inode)
 477 {
 478         fuse_set_nowrite(inode);
 479         fuse_release_nowrite(inode);
 480 }
 481
 482 static int fuse_flush(struct file *file, fl_owner_t id)
 483 {
 484         struct inode *inode = file_inode(file);
 485         struct fuse_mount *fm = get_fuse_mount(inode);
 486         struct fuse_file *ff = file->private_data;
 487         struct fuse_flush_in inarg;
 488         FUSE_ARGS(args);
 489         int err;
 490
 491         if (fuse_is_bad(inode))
 492                 return -EIO;
 493
 494         if (ff->open_flags & FOPEN_NOFLUSH && !fm->fc->writeback_cache)
 495                 return 0;
 496
 497         err = write_inode_now(inode, 1);
 498         if (err)
 499                 return err;
 500
 501         inode_lock(inode);
 502         fuse_sync_writes(inode);
 503         inode_unlock(inode);
 504
 505         err = filemap_check_errors(file->f_mapping);
 506         if (err)
 507                 return err;
 508
 509         err = 0;
 510         if (fm->fc->no_flush)
 511                 goto inval_attr_out;
 512
 513         memset(&inarg, 0, sizeof(inarg));
 514         inarg.fh = ff->fh;
 515         inarg.lock_owner = fuse_lock_owner_id(fm->fc, id);
 516         args.opcode = FUSE_FLUSH;
 517         args.nodeid = get_node_id(inode);
 518         args.in_numargs = 1;
 519         args.in_args[0].size = sizeof(inarg);
 520         args.in_args[0].value = &inarg;
 521         args.force = true;
 522
 523         err = fuse_simple_request(fm, &args);
 524         if (err == -ENOSYS) {
 525                 fm->fc->no_flush = 1;
 526                 err = 0;
 527         }
 528
 529 inval_attr_out:
 530         /*
 531          * In memory i_blocks is not maintained by fuse, if writeback cache is
 532          * enabled, i_blocks from cached attr may not be accurate.
 533          */
 534         if (!err && fm->fc->writeback_cache)
 535                 fuse_invalidate_attr_mask(inode, STATX_BLOCKS);
 536         return err;
 537 }
 538
 539 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 540                       int datasync, int opcode)
 541 {
 542         struct inode *inode = file->f_mapping->host;
 543         struct fuse_mount *fm = get_fuse_mount(inode);
 544         struct fuse_file *ff = file->private_data;
 545         FUSE_ARGS(args);
 546         struct fuse_fsync_in inarg;
 547
 548         memset(&inarg, 0, sizeof(inarg));
 549         inarg.fh = ff->fh;
 550         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 551         args.opcode = opcode;
 552         args.nodeid = get_node_id(inode);
 553         args.in_numargs = 1;
 554         args.in_args[0].size = sizeof(inarg);
 555         args.in_args[0].value = &inarg;
 556         return fuse_simple_request(fm, &args);
 557 }
 558
 559 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 560                       int datasync)
 561 {
 562         struct inode *inode = file->f_mapping->host;
 563         struct fuse_conn *fc = get_fuse_conn(inode);
 564         int err;
 565
 566         if (fuse_is_bad(inode))
 567                 return -EIO;
 568
 569         inode_lock(inode);
 570
 571         /*
 572          * Start writeback against all dirty pages of the inode, then
 573          * wait for all outstanding writes, before sending the FSYNC
 574          * request.
 575          */
 576         err = file_write_and_wait_range(file, start, end);
 577         if (err)
 578                 goto out;
 579
 580         fuse_sync_writes(inode);
 581
 582         /*
 583          * Due to implementation of fuse writeback
 584          * file_write_and_wait_range() does not catch errors.
 585          * We have to do this directly after fuse_sync_writes()
 586          */
 587         err = file_check_and_advance_wb_err(file);
 588         if (err)
 589                 goto out;
 590
 591         err = sync_inode_metadata(inode, 1);
 592         if (err)
 593                 goto out;
 594
 595         if (fc->no_fsync)
 596                 goto out;
 597
 598         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 599         if (err == -ENOSYS) {
 600                 fc->no_fsync = 1;
 601                 err = 0;
 602         }
 603 out:
 604         inode_unlock(inode);
 605
 606         return err;
 607 }
 608
 609 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
 610                          size_t count, int opcode)
 611 {
 612         struct fuse_file *ff = file->private_data;
 613         struct fuse_args *args = &ia->ap.args;
 614
 615         ia->read.in.fh = ff->fh;
 616         ia->read.in.offset = pos;
 617         ia->read.in.size = count;
 618         ia->read.in.flags = file->f_flags;
 619         args->opcode = opcode;
 620         args->nodeid = ff->nodeid;
 621         args->in_numargs = 1;
 622         args->in_args[0].size = sizeof(ia->read.in);
 623         args->in_args[0].value = &ia->read.in;
 624         args->out_argvar = true;
 625         args->out_numargs = 1;
 626         args->out_args[0].size = count;
 627 }
 628
 629 static void fuse_release_user_pages(struct fuse_args_pages *ap,
 630                                     bool should_dirty)
 631 {
 632         unsigned int i;
 633
 634         for (i = 0; i < ap->num_pages; i++) {
 635                 if (should_dirty)
 636                         set_page_dirty_lock(ap->pages[i]);
 637                 put_page(ap->pages[i]);
 638         }
 639 }
 640
 641 static void fuse_io_release(struct kref *kref)
 642 {
 643         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 644 }
 645
 646 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 647 {
 648         if (io->err)
 649                 return io->err;
 650
 651         if (io->bytes >= 0 && io->write)
 652                 return -EIO;
 653
 654         return io->bytes < 0 ? io->size : io->bytes;
 655 }
 656
 657 /*
 658  * In case of short read, the caller sets 'pos' to the position of
 659  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 660  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 661  *
 662  * An example:
 663  * User requested DIO read of 64K. It was split into two 32K fuse requests,
 664  * both submitted asynchronously. The first of them was ACKed by userspace as
 665  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 666  * second request was ACKed as short, e.g. only 1K was read, resulting in
 667  * pos == 33K.
 668  *
 669  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 670  * will be equal to the length of the longest contiguous fragment of
 671  * transferred data starting from the beginning of IO request.
 672  */
 673 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 674 {
 675         int left;
 676
 677         spin_lock(&io->lock);
 678         if (err)
 679                 io->err = io->err ? : err;
 680         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 681                 io->bytes = pos;
 682
 683         left = --io->reqs;
 684         if (!left && io->blocking)
 685                 complete(io->done);
 686         spin_unlock(&io->lock);
 687
 688         if (!left && !io->blocking) {
 689                 ssize_t res = fuse_get_res_by_io(io);
 690
 691                 if (res >= 0) {
 692                         struct inode *inode = file_inode(io->iocb->ki_filp);
 693                         struct fuse_conn *fc = get_fuse_conn(inode);
 694                         struct fuse_inode *fi = get_fuse_inode(inode);
 695
 696                         spin_lock(&fi->lock);
 697                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 698                         spin_unlock(&fi->lock);
 699                 }
 700
 701                 io->iocb->ki_complete(io->iocb, res);
 702         }
 703
 704         kref_put(&io->refcnt, fuse_io_release);
 705 }
 706
 707 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
 708                                           unsigned int npages)
 709 {
 710         struct fuse_io_args *ia;
 711
 712         ia = kzalloc(sizeof(*ia), GFP_KERNEL);
 713         if (ia) {
 714                 ia->io = io;
 715                 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
 716                                                 &ia->ap.descs);
 717                 if (!ia->ap.pages) {
 718                         kfree(ia);
 719                         ia = NULL;
 720                 }
 721         }
 722         return ia;
 723 }
 724
 725 static void fuse_io_free(struct fuse_io_args *ia)
 726 {
 727         kfree(ia->ap.pages);
 728         kfree(ia);
 729 }
 730
 731 static void fuse_aio_complete_req(struct fuse_mount *fm, struct fuse_args *args,
 732                                   int err)
 733 {
 734         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 735         struct fuse_io_priv *io = ia->io;
 736         ssize_t pos = -1;
 737
 738         fuse_release_user_pages(&ia->ap, io->should_dirty);
 739
 740         if (err) {
 741                 /* Nothing */
 742         } else if (io->write) {
 743                 if (ia->write.out.size > ia->write.in.size) {
 744                         err = -EIO;
 745                 } else if (ia->write.in.size != ia->write.out.size) {
 746                         pos = ia->write.in.offset - io->offset +
 747                                 ia->write.out.size;
 748                 }
 749         } else {
 750                 u32 outsize = args->out_args[0].size;
 751
 752                 if (ia->read.in.size != outsize)
 753                         pos = ia->read.in.offset - io->offset + outsize;
 754         }
 755
 756         fuse_aio_complete(io, err, pos);
 757         fuse_io_free(ia);
 758 }
 759
 760 static ssize_t fuse_async_req_send(struct fuse_mount *fm,
 761                                    struct fuse_io_args *ia, size_t num_bytes)
 762 {
 763         ssize_t err;
 764         struct fuse_io_priv *io = ia->io;
 765
 766         spin_lock(&io->lock);
 767         kref_get(&io->refcnt);
 768         io->size += num_bytes;
 769         io->reqs++;
 770         spin_unlock(&io->lock);
 771
 772         ia->ap.args.end = fuse_aio_complete_req;
 773         ia->ap.args.may_block = io->should_dirty;
 774         err = fuse_simple_background(fm, &ia->ap.args, GFP_KERNEL);
 775         if (err)
 776                 fuse_aio_complete_req(fm, &ia->ap.args, err);
 777
 778         return num_bytes;
 779 }
 780
 781 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
 782                               fl_owner_t owner)
 783 {
 784         struct file *file = ia->io->iocb->ki_filp;
 785         struct fuse_file *ff = file->private_data;
 786         struct fuse_mount *fm = ff->fm;
 787
 788         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 789         if (owner != NULL) {
 790                 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
 791                 ia->read.in.lock_owner = fuse_lock_owner_id(fm->fc, owner);
 792         }
 793
 794         if (ia->io->async)
 795                 return fuse_async_req_send(fm, ia, count);
 796
 797         return fuse_simple_request(fm, &ia->ap.args);
 798 }
 799
 800 static void fuse_read_update_size(struct inode *inode, loff_t size,
 801                                   u64 attr_ver)
 802 {
 803         struct fuse_conn *fc = get_fuse_conn(inode);
 804         struct fuse_inode *fi = get_fuse_inode(inode);
 805
 806         spin_lock(&fi->lock);
 807         if (attr_ver >= fi->attr_version && size < inode->i_size &&
 808             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 809                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 810                 i_size_write(inode, size);
 811         }
 812         spin_unlock(&fi->lock);
 813 }
 814
 815 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
 816                             struct fuse_args_pages *ap)
 817 {
 818         struct fuse_conn *fc = get_fuse_conn(inode);
 819
 820         /*
 821          * If writeback_cache is enabled, a short read means there's a hole in
 822          * the file.  Some data after the hole is in page cache, but has not
 823          * reached the client fs yet.  So the hole is not present there.
 824          */
 825         if (!fc->writeback_cache) {
 826                 loff_t pos = page_offset(ap->pages[0]) + num_read;
 827                 fuse_read_update_size(inode, pos, attr_ver);
 828         }
 829 }
 830
 831 static int fuse_do_readpage(struct file *file, struct page *page)
 832 {
 833         struct inode *inode = page->mapping->host;
 834         struct fuse_mount *fm = get_fuse_mount(inode);
 835         loff_t pos = page_offset(page);
 836         struct fuse_page_desc desc = { .length = PAGE_SIZE };
 837         struct fuse_io_args ia = {
 838                 .ap.args.page_zeroing = true,
 839                 .ap.args.out_pages = true,
 840                 .ap.num_pages = 1,
 841                 .ap.pages = &page,
 842                 .ap.descs = &desc,
 843         };
 844         ssize_t res;
 845         u64 attr_ver;
 846
 847         /*
 848          * Page writeback can extend beyond the lifetime of the
 849          * page-cache page, so make sure we read a properly synced
 850          * page.
 851          */
 852         fuse_wait_on_page_writeback(inode, page->index);
 853
 854         attr_ver = fuse_get_attr_version(fm->fc);
 855
 856         /* Don't overflow end offset */
 857         if (pos + (desc.length - 1) == LLONG_MAX)
 858                 desc.length--;
 859
 860         fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
 861         res = fuse_simple_request(fm, &ia.ap.args);
 862         if (res < 0)
 863                 return res;
 864         /*
 865          * Short read means EOF.  If file size is larger, truncate it
 866          */
 867         if (res < desc.length)
 868                 fuse_short_read(inode, attr_ver, res, &ia.ap);
 869
 870         SetPageUptodate(page);
 871
 872         return 0;
 873 }
 874
 875 static int fuse_read_folio(struct file *file, struct folio *folio)
 876 {
 877         struct page *page = &folio->page;
 878         struct inode *inode = page->mapping->host;
 879         int err;
 880
 881         err = -EIO;
 882         if (fuse_is_bad(inode))
 883                 goto out;
 884
 885         err = fuse_do_readpage(file, page);
 886         fuse_invalidate_atime(inode);
 887  out:
 888         unlock_page(page);
 889         return err;
 890 }
 891
 892 static void fuse_readpages_end(struct fuse_mount *fm, struct fuse_args *args,
 893                                int err)
 894 {
 895         int i;
 896         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 897         struct fuse_args_pages *ap = &ia->ap;
 898         size_t count = ia->read.in.size;
 899         size_t num_read = args->out_args[0].size;
 900         struct address_space *mapping = NULL;
 901
 902         for (i = 0; mapping == NULL && i < ap->num_pages; i++)
 903                 mapping = ap->pages[i]->mapping;
 904
 905         if (mapping) {
 906                 struct inode *inode = mapping->host;
 907
 908                 /*
 909                  * Short read means EOF. If file size is larger, truncate it
 910                  */
 911                 if (!err && num_read < count)
 912                         fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
 913
 914                 fuse_invalidate_atime(inode);
 915         }
 916
 917         for (i = 0; i < ap->num_pages; i++) {
 918                 struct page *page = ap->pages[i];
 919
 920                 if (!err)
 921                         SetPageUptodate(page);
 922                 else
 923                         SetPageError(page);
 924                 unlock_page(page);
 925                 put_page(page);
 926         }
 927         if (ia->ff)
 928                 fuse_file_put(ia->ff, false, false);
 929
 930         fuse_io_free(ia);
 931 }
 932
 933 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
 934 {
 935         struct fuse_file *ff = file->private_data;
 936         struct fuse_mount *fm = ff->fm;
 937         struct fuse_args_pages *ap = &ia->ap;
 938         loff_t pos = page_offset(ap->pages[0]);
 939         size_t count = ap->num_pages << PAGE_SHIFT;
 940         ssize_t res;
 941         int err;
 942
 943         ap->args.out_pages = true;
 944         ap->args.page_zeroing = true;
 945         ap->args.page_replace = true;
 946
 947         /* Don't overflow end offset */
 948         if (pos + (count - 1) == LLONG_MAX) {
 949                 count--;
 950                 ap->descs[ap->num_pages - 1].length--;
 951         }
 952         WARN_ON((loff_t) (pos + count) < 0);
 953
 954         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 955         ia->read.attr_ver = fuse_get_attr_version(fm->fc);
 956         if (fm->fc->async_read) {
 957                 ia->ff = fuse_file_get(ff);
 958                 ap->args.end = fuse_readpages_end;
 959                 err = fuse_simple_background(fm, &ap->args, GFP_KERNEL);
 960                 if (!err)
 961                         return;
 962         } else {
 963                 res = fuse_simple_request(fm, &ap->args);
 964                 err = res < 0 ? res : 0;
 965         }
 966         fuse_readpages_end(fm, &ap->args, err);
 967 }
 968
 969 static void fuse_readahead(struct readahead_control *rac)
 970 {
 971         struct inode *inode = rac->mapping->host;
 972         struct fuse_conn *fc = get_fuse_conn(inode);
 973         unsigned int i, max_pages, nr_pages = 0;
 974
 975         if (fuse_is_bad(inode))
 976                 return;
 977
 978         max_pages = min_t(unsigned int, fc->max_pages,
 979                         fc->max_read / PAGE_SIZE);
 980
 981         for (;;) {
 982                 struct fuse_io_args *ia;
 983                 struct fuse_args_pages *ap;
 984
 985                 if (fc->num_background >= fc->congestion_threshold &&
 986                     rac->ra->async_size >= readahead_count(rac))
 987                         /*
 988                          * Congested and only async pages left, so skip the
 989                          * rest.
 990                          */
 991                         break;
 992
 993                 nr_pages = readahead_count(rac) - nr_pages;
 994                 if (nr_pages > max_pages)
 995                         nr_pages = max_pages;
 996                 if (nr_pages == 0)
 997                         break;
 998                 ia = fuse_io_alloc(NULL, nr_pages);
 999                 if (!ia)
1000                         return;
1001                 ap = &ia->ap;
1002                 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
1003                 for (i = 0; i < nr_pages; i++) {
1004                         fuse_wait_on_page_writeback(inode,
1005                                                     readahead_index(rac) + i);
1006                         ap->descs[i].length = PAGE_SIZE;
1007                 }
1008                 ap->num_pages = nr_pages;
1009                 fuse_send_readpages(ia, rac->file);
1010         }
1011 }
1012
1013 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
1014 {
1015         struct inode *inode = iocb->ki_filp->f_mapping->host;
1016         struct fuse_conn *fc = get_fuse_conn(inode);
1017
1018         /*
1019          * In auto invalidate mode, always update attributes on read.
1020          * Otherwise, only update if we attempt to read past EOF (to ensure
1021          * i_size is up to date).
1022          */
1023         if (fc->auto_inval_data ||
1024             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
1025                 int err;
1026                 err = fuse_update_attributes(inode, iocb->ki_filp, STATX_SIZE);
1027                 if (err)
1028                         return err;
1029         }
1030
1031         return generic_file_read_iter(iocb, to);
1032 }
1033
1034 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1035                                  loff_t pos, size_t count)
1036 {
1037         struct fuse_args *args = &ia->ap.args;
1038
1039         ia->write.in.fh = ff->fh;
1040         ia->write.in.offset = pos;
1041         ia->write.in.size = count;
1042         args->opcode = FUSE_WRITE;
1043         args->nodeid = ff->nodeid;
1044         args->in_numargs = 2;
1045         if (ff->fm->fc->minor < 9)
1046                 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1047         else
1048                 args->in_args[0].size = sizeof(ia->write.in);
1049         args->in_args[0].value = &ia->write.in;
1050         args->in_args[1].size = count;
1051         args->out_numargs = 1;
1052         args->out_args[0].size = sizeof(ia->write.out);
1053         args->out_args[0].value = &ia->write.out;
1054 }
1055
1056 static unsigned int fuse_write_flags(struct kiocb *iocb)
1057 {
1058         unsigned int flags = iocb->ki_filp->f_flags;
1059
1060         if (iocb_is_dsync(iocb))
1061                 flags |= O_DSYNC;
1062         if (iocb->ki_flags & IOCB_SYNC)
1063                 flags |= O_SYNC;
1064
1065         return flags;
1066 }
1067
1068 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1069                                size_t count, fl_owner_t owner)
1070 {
1071         struct kiocb *iocb = ia->io->iocb;
1072         struct file *file = iocb->ki_filp;
1073         struct fuse_file *ff = file->private_data;
1074         struct fuse_mount *fm = ff->fm;
1075         struct fuse_write_in *inarg = &ia->write.in;
1076         ssize_t err;
1077
1078         fuse_write_args_fill(ia, ff, pos, count);
1079         inarg->flags = fuse_write_flags(iocb);
1080         if (owner != NULL) {
1081                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1082                 inarg->lock_owner = fuse_lock_owner_id(fm->fc, owner);
1083         }
1084
1085         if (ia->io->async)
1086                 return fuse_async_req_send(fm, ia, count);
1087
1088         err = fuse_simple_request(fm, &ia->ap.args);
1089         if (!err && ia->write.out.size > count)
1090                 err = -EIO;
1091
1092         return err ?: ia->write.out.size;
1093 }
1094
1095 bool fuse_write_update_attr(struct inode *inode, loff_t pos, ssize_t written)
1096 {
1097         struct fuse_conn *fc = get_fuse_conn(inode);
1098         struct fuse_inode *fi = get_fuse_inode(inode);
1099         bool ret = false;
1100
1101         spin_lock(&fi->lock);
1102         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1103         if (written > 0 && pos > inode->i_size) {
1104                 i_size_write(inode, pos);
1105                 ret = true;
1106         }
1107         spin_unlock(&fi->lock);
1108
1109         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
1110
1111         return ret;
1112 }
1113
1114 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1115                                      struct kiocb *iocb, struct inode *inode,
1116                                      loff_t pos, size_t count)
1117 {
1118         struct fuse_args_pages *ap = &ia->ap;
1119         struct file *file = iocb->ki_filp;
1120         struct fuse_file *ff = file->private_data;
1121         struct fuse_mount *fm = ff->fm;
1122         unsigned int offset, i;
1123         bool short_write;
1124         int err;
1125
1126         for (i = 0; i < ap->num_pages; i++)
1127                 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1128
1129         fuse_write_args_fill(ia, ff, pos, count);
1130         ia->write.in.flags = fuse_write_flags(iocb);
1131         if (fm->fc->handle_killpriv_v2 && !capable(CAP_FSETID))
1132                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1133
1134         err = fuse_simple_request(fm, &ap->args);
1135         if (!err && ia->write.out.size > count)
1136                 err = -EIO;
1137
1138         short_write = ia->write.out.size < count;
1139         offset = ap->descs[0].offset;
1140         count = ia->write.out.size;
1141         for (i = 0; i < ap->num_pages; i++) {
1142                 struct page *page = ap->pages[i];
1143
1144                 if (err) {
1145                         ClearPageUptodate(page);
1146                 } else {
1147                         if (count >= PAGE_SIZE - offset)
1148                                 count -= PAGE_SIZE - offset;
1149                         else {
1150                                 if (short_write)
1151                                         ClearPageUptodate(page);
1152                                 count = 0;
1153                         }
1154                         offset = 0;
1155                 }
1156                 if (ia->write.page_locked && (i == ap->num_pages - 1))
1157                         unlock_page(page);
1158                 put_page(page);
1159         }
1160
1161         return err;
1162 }
1163
1164 static ssize_t fuse_fill_write_pages(struct fuse_io_args *ia,
1165                                      struct address_space *mapping,
1166                                      struct iov_iter *ii, loff_t pos,
1167                                      unsigned int max_pages)
1168 {
1169         struct fuse_args_pages *ap = &ia->ap;
1170         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1171         unsigned offset = pos & (PAGE_SIZE - 1);
1172         size_t count = 0;
1173         int err;
1174
1175         ap->args.in_pages = true;
1176         ap->descs[0].offset = offset;
1177
1178         do {
1179                 size_t tmp;
1180                 struct page *page;
1181                 pgoff_t index = pos >> PAGE_SHIFT;
1182                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1183                                      iov_iter_count(ii));
1184
1185                 bytes = min_t(size_t, bytes, fc->max_write - count);
1186
1187  again:
1188                 err = -EFAULT;
1189                 if (fault_in_iov_iter_readable(ii, bytes))
1190                         break;
1191
1192                 err = -ENOMEM;
1193                 page = grab_cache_page_write_begin(mapping, index);
1194                 if (!page)
1195                         break;
1196
1197                 if (mapping_writably_mapped(mapping))
1198                         flush_dcache_page(page);
1199
1200                 tmp = copy_page_from_iter_atomic(page, offset, bytes, ii);
1201                 flush_dcache_page(page);
1202
1203                 if (!tmp) {
1204                         unlock_page(page);
1205                         put_page(page);
1206                         goto again;
1207                 }
1208
1209                 err = 0;
1210                 ap->pages[ap->num_pages] = page;
1211                 ap->descs[ap->num_pages].length = tmp;
1212                 ap->num_pages++;
1213
1214                 count += tmp;
1215                 pos += tmp;
1216                 offset += tmp;
1217                 if (offset == PAGE_SIZE)
1218                         offset = 0;
1219
1220                 /* If we copied full page, mark it uptodate */
1221                 if (tmp == PAGE_SIZE)
1222                         SetPageUptodate(page);
1223
1224                 if (PageUptodate(page)) {
1225                         unlock_page(page);
1226                 } else {
1227                         ia->write.page_locked = true;
1228                         break;
1229                 }
1230                 if (!fc->big_writes)
1231                         break;
1232         } while (iov_iter_count(ii) && count < fc->max_write &&
1233                  ap->num_pages < max_pages && offset == 0);
1234
1235         return count > 0 ? count : err;
1236 }
1237
1238 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1239                                      unsigned int max_pages)
1240 {
1241         return min_t(unsigned int,
1242                      ((pos + len - 1) >> PAGE_SHIFT) -
1243                      (pos >> PAGE_SHIFT) + 1,
1244                      max_pages);
1245 }
1246
1247 static ssize_t fuse_perform_write(struct kiocb *iocb, struct iov_iter *ii)
1248 {
1249         struct address_space *mapping = iocb->ki_filp->f_mapping;
1250         struct inode *inode = mapping->host;
1251         struct fuse_conn *fc = get_fuse_conn(inode);
1252         struct fuse_inode *fi = get_fuse_inode(inode);
1253         loff_t pos = iocb->ki_pos;
1254         int err = 0;
1255         ssize_t res = 0;
1256
1257         if (inode->i_size < pos + iov_iter_count(ii))
1258                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1259
1260         do {
1261                 ssize_t count;
1262                 struct fuse_io_args ia = {};
1263                 struct fuse_args_pages *ap = &ia.ap;
1264                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1265                                                       fc->max_pages);
1266
1267                 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1268                 if (!ap->pages) {
1269                         err = -ENOMEM;
1270                         break;
1271                 }
1272
1273                 count = fuse_fill_write_pages(&ia, mapping, ii, pos, nr_pages);
1274                 if (count <= 0) {
1275                         err = count;
1276                 } else {
1277                         err = fuse_send_write_pages(&ia, iocb, inode,
1278                                                     pos, count);
1279                         if (!err) {
1280                                 size_t num_written = ia.write.out.size;
1281
1282                                 res += num_written;
1283                                 pos += num_written;
1284
1285                                 /* break out of the loop on short write */
1286                                 if (num_written != count)
1287                                         err = -EIO;
1288                         }
1289                 }
1290                 kfree(ap->pages);
1291         } while (!err && iov_iter_count(ii));
1292
1293         fuse_write_update_attr(inode, pos, res);
1294         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1295
1296         if (!res)
1297                 return err;
1298         iocb->ki_pos += res;
1299         return res;
1300 }
1301
1302 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1303 {
1304         struct file *file = iocb->ki_filp;
1305         struct address_space *mapping = file->f_mapping;
1306         ssize_t written = 0;
1307         struct inode *inode = mapping->host;
1308         ssize_t err;
1309         struct fuse_conn *fc = get_fuse_conn(inode);
1310
1311         if (fc->writeback_cache) {
1312                 /* Update size (EOF optimization) and mode (SUID clearing) */
1313                 err = fuse_update_attributes(mapping->host, file,
1314                                              STATX_SIZE | STATX_MODE);
1315                 if (err)
1316                         return err;
1317
1318                 if (fc->handle_killpriv_v2 &&
1319                     setattr_should_drop_suidgid(&nop_mnt_idmap,
1320                                                 file_inode(file))) {
1321                         goto writethrough;
1322                 }
1323
1324                 return generic_file_write_iter(iocb, from);
1325         }
1326
1327 writethrough:
1328         inode_lock(inode);
1329
1330         err = generic_write_checks(iocb, from);
1331         if (err <= 0)
1332                 goto out;
1333
1334         err = file_remove_privs(file);
1335         if (err)
1336                 goto out;
1337
1338         err = file_update_time(file);
1339         if (err)
1340                 goto out;
1341
1342         if (iocb->ki_flags & IOCB_DIRECT) {
1343                 written = generic_file_direct_write(iocb, from);
1344                 if (written < 0 || !iov_iter_count(from))
1345                         goto out;
1346                 written = direct_write_fallback(iocb, from, written,
1347                                 fuse_perform_write(iocb, from));
1348         } else {
1349                 written = fuse_perform_write(iocb, from);
1350         }
1351 out:
1352         inode_unlock(inode);
1353         if (written > 0)
1354                 written = generic_write_sync(iocb, written);
1355
1356         return written ? written : err;
1357 }
1358
1359 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1360 {
1361         return (unsigned long)iter_iov(ii)->iov_base + ii->iov_offset;
1362 }
1363
1364 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1365                                         size_t max_size)
1366 {
1367         return min(iov_iter_single_seg_count(ii), max_size);
1368 }
1369
1370 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1371                                size_t *nbytesp, int write,
1372                                unsigned int max_pages)
1373 {
1374         size_t nbytes = 0;  /* # bytes already packed in req */
1375         ssize_t ret = 0;
1376
1377         /* Special case for kernel I/O: can copy directly into the buffer */
1378         if (iov_iter_is_kvec(ii)) {
1379                 unsigned long user_addr = fuse_get_user_addr(ii);
1380                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1381
1382                 if (write)
1383                         ap->args.in_args[1].value = (void *) user_addr;
1384                 else
1385                         ap->args.out_args[0].value = (void *) user_addr;
1386
1387                 iov_iter_advance(ii, frag_size);
1388                 *nbytesp = frag_size;
1389                 return 0;
1390         }
1391
1392         while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1393                 unsigned npages;
1394                 size_t start;
1395                 ret = iov_iter_get_pages2(ii, &ap->pages[ap->num_pages],
1396                                         *nbytesp - nbytes,
1397                                         max_pages - ap->num_pages,
1398                                         &start);
1399                 if (ret < 0)
1400                         break;
1401
1402                 nbytes += ret;
1403
1404                 ret += start;
1405                 npages = DIV_ROUND_UP(ret, PAGE_SIZE);
1406
1407                 ap->descs[ap->num_pages].offset = start;
1408                 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1409
1410                 ap->num_pages += npages;
1411                 ap->descs[ap->num_pages - 1].length -=
1412                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1413         }
1414
1415         ap->args.user_pages = true;
1416         if (write)
1417                 ap->args.in_pages = true;
1418         else
1419                 ap->args.out_pages = true;
1420
1421         *nbytesp = nbytes;
1422
1423         return ret < 0 ? ret : 0;
1424 }
1425
1426 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1427                        loff_t *ppos, int flags)
1428 {
1429         int write = flags & FUSE_DIO_WRITE;
1430         int cuse = flags & FUSE_DIO_CUSE;
1431         struct file *file = io->iocb->ki_filp;
1432         struct address_space *mapping = file->f_mapping;
1433         struct inode *inode = mapping->host;
1434         struct fuse_file *ff = file->private_data;
1435         struct fuse_conn *fc = ff->fm->fc;
1436         size_t nmax = write ? fc->max_write : fc->max_read;
1437         loff_t pos = *ppos;
1438         size_t count = iov_iter_count(iter);
1439         pgoff_t idx_from = pos >> PAGE_SHIFT;
1440         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1441         ssize_t res = 0;
1442         int err = 0;
1443         struct fuse_io_args *ia;
1444         unsigned int max_pages;
1445         bool fopen_direct_io = ff->open_flags & FOPEN_DIRECT_IO;
1446
1447         max_pages = iov_iter_npages(iter, fc->max_pages);
1448         ia = fuse_io_alloc(io, max_pages);
1449         if (!ia)
1450                 return -ENOMEM;
1451
1452         if (fopen_direct_io && fc->direct_io_allow_mmap) {
1453                 res = filemap_write_and_wait_range(mapping, pos, pos + count - 1);
1454                 if (res) {
1455                         fuse_io_free(ia);
1456                         return res;
1457                 }
1458         }
1459         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1460                 if (!write)
1461                         inode_lock(inode);
1462                 fuse_sync_writes(inode);
1463                 if (!write)
1464                         inode_unlock(inode);
1465         }
1466
1467         if (fopen_direct_io && write) {
1468                 res = invalidate_inode_pages2_range(mapping, idx_from, idx_to);
1469                 if (res) {
1470                         fuse_io_free(ia);
1471                         return res;
1472                 }
1473         }
1474
1475         io->should_dirty = !write && user_backed_iter(iter);
1476         while (count) {
1477                 ssize_t nres;
1478                 fl_owner_t owner = current->files;
1479                 size_t nbytes = min(count, nmax);
1480
1481                 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1482                                           max_pages);
1483                 if (err && !nbytes)
1484                         break;
1485
1486                 if (write) {
1487                         if (!capable(CAP_FSETID))
1488                                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1489
1490                         nres = fuse_send_write(ia, pos, nbytes, owner);
1491                 } else {
1492                         nres = fuse_send_read(ia, pos, nbytes, owner);
1493                 }
1494
1495                 if (!io->async || nres < 0) {
1496                         fuse_release_user_pages(&ia->ap, io->should_dirty);
1497                         fuse_io_free(ia);
1498                 }
1499                 ia = NULL;
1500                 if (nres < 0) {
1501                         iov_iter_revert(iter, nbytes);
1502                         err = nres;
1503                         break;
1504                 }
1505                 WARN_ON(nres > nbytes);
1506
1507                 count -= nres;
1508                 res += nres;
1509                 pos += nres;
1510                 if (nres != nbytes) {
1511                         iov_iter_revert(iter, nbytes - nres);
1512                         break;
1513                 }
1514                 if (count) {
1515                         max_pages = iov_iter_npages(iter, fc->max_pages);
1516                         ia = fuse_io_alloc(io, max_pages);
1517                         if (!ia)
1518                                 break;
1519                 }
1520         }
1521         if (ia)
1522                 fuse_io_free(ia);
1523         if (res > 0)
1524                 *ppos = pos;
1525
1526         return res > 0 ? res : err;
1527 }
1528 EXPORT_SYMBOL_GPL(fuse_direct_io);
1529
1530 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1531                                   struct iov_iter *iter,
1532                                   loff_t *ppos)
1533 {
1534         ssize_t res;
1535         struct inode *inode = file_inode(io->iocb->ki_filp);
1536
1537         res = fuse_direct_io(io, iter, ppos, 0);
1538
1539         fuse_invalidate_atime(inode);
1540
1541         return res;
1542 }
1543
1544 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1545
1546 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1547 {
1548         ssize_t res;
1549
1550         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1551                 res = fuse_direct_IO(iocb, to);
1552         } else {
1553                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1554
1555                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1556         }
1557
1558         return res;
1559 }
1560
1561 static bool fuse_direct_write_extending_i_size(struct kiocb *iocb,
1562                                                struct iov_iter *iter)
1563 {
1564         struct inode *inode = file_inode(iocb->ki_filp);
1565
1566         return iocb->ki_pos + iov_iter_count(iter) > i_size_read(inode);
1567 }
1568
1569 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1570 {
1571         struct inode *inode = file_inode(iocb->ki_filp);
1572         struct file *file = iocb->ki_filp;
1573         struct fuse_file *ff = file->private_data;
1574         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1575         ssize_t res;
1576         bool exclusive_lock =
1577                 !(ff->open_flags & FOPEN_PARALLEL_DIRECT_WRITES) ||
1578                 get_fuse_conn(inode)->direct_io_allow_mmap ||
1579                 iocb->ki_flags & IOCB_APPEND ||
1580                 fuse_direct_write_extending_i_size(iocb, from);
1581
1582         /*
1583          * Take exclusive lock if
1584          * - Parallel direct writes are disabled - a user space decision
1585          * - Parallel direct writes are enabled and i_size is being extended.
1586          * - Shared mmap on direct_io file is supported (FUSE_DIRECT_IO_ALLOW_MMAP).
1587          *   This might not be needed at all, but needs further investigation.
1588          */
1589         if (exclusive_lock)
1590                 inode_lock(inode);
1591         else {
1592                 inode_lock_shared(inode);
1593
1594                 /* A race with truncate might have come up as the decision for
1595                  * the lock type was done without holding the lock, check again.
1596                  */
1597                 if (fuse_direct_write_extending_i_size(iocb, from)) {
1598                         inode_unlock_shared(inode);
1599                         inode_lock(inode);
1600                         exclusive_lock = true;
1601                 }
1602         }
1603
1604         res = generic_write_checks(iocb, from);
1605         if (res > 0) {
1606                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1607                         res = fuse_direct_IO(iocb, from);
1608                 } else {
1609                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1610                                              FUSE_DIO_WRITE);
1611                         fuse_write_update_attr(inode, iocb->ki_pos, res);
1612                 }
1613         }
1614         if (exclusive_lock)
1615                 inode_unlock(inode);
1616         else
1617                 inode_unlock_shared(inode);
1618
1619         return res;
1620 }
1621
1622 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1623 {
1624         struct file *file = iocb->ki_filp;
1625         struct fuse_file *ff = file->private_data;
1626         struct inode *inode = file_inode(file);
1627
1628         if (fuse_is_bad(inode))
1629                 return -EIO;
1630
1631         if (FUSE_IS_DAX(inode))
1632                 return fuse_dax_read_iter(iocb, to);
1633
1634         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1635                 return fuse_cache_read_iter(iocb, to);
1636         else
1637                 return fuse_direct_read_iter(iocb, to);
1638 }
1639
1640 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1641 {
1642         struct file *file = iocb->ki_filp;
1643         struct fuse_file *ff = file->private_data;
1644         struct inode *inode = file_inode(file);
1645
1646         if (fuse_is_bad(inode))
1647                 return -EIO;
1648
1649         if (FUSE_IS_DAX(inode))
1650                 return fuse_dax_write_iter(iocb, from);
1651
1652         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1653                 return fuse_cache_write_iter(iocb, from);
1654         else
1655                 return fuse_direct_write_iter(iocb, from);
1656 }
1657
1658 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1659 {
1660         struct fuse_args_pages *ap = &wpa->ia.ap;
1661         int i;
1662
1663         if (wpa->bucket)
1664                 fuse_sync_bucket_dec(wpa->bucket);
1665
1666         for (i = 0; i < ap->num_pages; i++)
1667                 __free_page(ap->pages[i]);
1668
1669         if (wpa->ia.ff)
1670                 fuse_file_put(wpa->ia.ff, false, false);
1671
1672         kfree(ap->pages);
1673         kfree(wpa);
1674 }
1675
1676 static void fuse_writepage_finish(struct fuse_mount *fm,
1677                                   struct fuse_writepage_args *wpa)
1678 {
1679         struct fuse_args_pages *ap = &wpa->ia.ap;
1680         struct inode *inode = wpa->inode;
1681         struct fuse_inode *fi = get_fuse_inode(inode);
1682         struct backing_dev_info *bdi = inode_to_bdi(inode);
1683         int i;
1684
1685         for (i = 0; i < ap->num_pages; i++) {
1686                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1687                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1688                 wb_writeout_inc(&bdi->wb);
1689         }
1690         wake_up(&fi->page_waitq);
1691 }
1692
1693 /* Called under fi->lock, may release and reacquire it */
1694 static void fuse_send_writepage(struct fuse_mount *fm,
1695                                 struct fuse_writepage_args *wpa, loff_t size)
1696 __releases(fi->lock)
1697 __acquires(fi->lock)
1698 {
1699         struct fuse_writepage_args *aux, *next;
1700         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1701         struct fuse_write_in *inarg = &wpa->ia.write.in;
1702         struct fuse_args *args = &wpa->ia.ap.args;
1703         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1704         int err;
1705
1706         fi->writectr++;
1707         if (inarg->offset + data_size <= size) {
1708                 inarg->size = data_size;
1709         } else if (inarg->offset < size) {
1710                 inarg->size = size - inarg->offset;
1711         } else {
1712                 /* Got truncated off completely */
1713                 goto out_free;
1714         }
1715
1716         args->in_args[1].size = inarg->size;
1717         args->force = true;
1718         args->nocreds = true;
1719
1720         err = fuse_simple_background(fm, args, GFP_ATOMIC);
1721         if (err == -ENOMEM) {
1722                 spin_unlock(&fi->lock);
1723                 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1724                 spin_lock(&fi->lock);
1725         }
1726
1727         /* Fails on broken connection only */
1728         if (unlikely(err))
1729                 goto out_free;
1730
1731         return;
1732
1733  out_free:
1734         fi->writectr--;
1735         rb_erase(&wpa->writepages_entry, &fi->writepages);
1736         fuse_writepage_finish(fm, wpa);
1737         spin_unlock(&fi->lock);
1738
1739         /* After fuse_writepage_finish() aux request list is private */
1740         for (aux = wpa->next; aux; aux = next) {
1741                 next = aux->next;
1742                 aux->next = NULL;
1743                 fuse_writepage_free(aux);
1744         }
1745
1746         fuse_writepage_free(wpa);
1747         spin_lock(&fi->lock);
1748 }
1749
1750 /*
1751  * If fi->writectr is positive (no truncate or fsync going on) send
1752  * all queued writepage requests.
1753  *
1754  * Called with fi->lock
1755  */
1756 void fuse_flush_writepages(struct inode *inode)
1757 __releases(fi->lock)
1758 __acquires(fi->lock)
1759 {
1760         struct fuse_mount *fm = get_fuse_mount(inode);
1761         struct fuse_inode *fi = get_fuse_inode(inode);
1762         loff_t crop = i_size_read(inode);
1763         struct fuse_writepage_args *wpa;
1764
1765         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1766                 wpa = list_entry(fi->queued_writes.next,
1767                                  struct fuse_writepage_args, queue_entry);
1768                 list_del_init(&wpa->queue_entry);
1769                 fuse_send_writepage(fm, wpa, crop);
1770         }
1771 }
1772
1773 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1774                                                 struct fuse_writepage_args *wpa)
1775 {
1776         pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1777         pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1778         struct rb_node **p = &root->rb_node;
1779         struct rb_node  *parent = NULL;
1780
1781         WARN_ON(!wpa->ia.ap.num_pages);
1782         while (*p) {
1783                 struct fuse_writepage_args *curr;
1784                 pgoff_t curr_index;
1785
1786                 parent = *p;
1787                 curr = rb_entry(parent, struct fuse_writepage_args,
1788                                 writepages_entry);
1789                 WARN_ON(curr->inode != wpa->inode);
1790                 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1791
1792                 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1793                         p = &(*p)->rb_right;
1794                 else if (idx_to < curr_index)
1795                         p = &(*p)->rb_left;
1796                 else
1797                         return curr;
1798         }
1799
1800         rb_link_node(&wpa->writepages_entry, parent, p);
1801         rb_insert_color(&wpa->writepages_entry, root);
1802         return NULL;
1803 }
1804
1805 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1806 {
1807         WARN_ON(fuse_insert_writeback(root, wpa));
1808 }
1809
1810 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1811                                int error)
1812 {
1813         struct fuse_writepage_args *wpa =
1814                 container_of(args, typeof(*wpa), ia.ap.args);
1815         struct inode *inode = wpa->inode;
1816         struct fuse_inode *fi = get_fuse_inode(inode);
1817         struct fuse_conn *fc = get_fuse_conn(inode);
1818
1819         mapping_set_error(inode->i_mapping, error);
1820         /*
1821          * A writeback finished and this might have updated mtime/ctime on
1822          * server making local mtime/ctime stale.  Hence invalidate attrs.
1823          * Do this only if writeback_cache is not enabled.  If writeback_cache
1824          * is enabled, we trust local ctime/mtime.
1825          */
1826         if (!fc->writeback_cache)
1827                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODIFY);
1828         spin_lock(&fi->lock);
1829         rb_erase(&wpa->writepages_entry, &fi->writepages);
1830         while (wpa->next) {
1831                 struct fuse_mount *fm = get_fuse_mount(inode);
1832                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1833                 struct fuse_writepage_args *next = wpa->next;
1834
1835                 wpa->next = next->next;
1836                 next->next = NULL;
1837                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1838                 tree_insert(&fi->writepages, next);
1839
1840                 /*
1841                  * Skip fuse_flush_writepages() to make it easy to crop requests
1842                  * based on primary request size.
1843                  *
1844                  * 1st case (trivial): there are no concurrent activities using
1845                  * fuse_set/release_nowrite.  Then we're on safe side because
1846                  * fuse_flush_writepages() would call fuse_send_writepage()
1847                  * anyway.
1848                  *
1849                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1850                  * now for completion of all in-flight requests.  This happens
1851                  * rarely and no more than once per page, so this should be
1852                  * okay.
1853                  *
1854                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1855                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1856                  * that fuse_set_nowrite returned implies that all in-flight
1857                  * requests were completed along with all of their secondary
1858                  * requests.  Further primary requests are blocked by negative
1859                  * writectr.  Hence there cannot be any in-flight requests and
1860                  * no invocations of fuse_writepage_end() while we're in
1861                  * fuse_set_nowrite..fuse_release_nowrite section.
1862                  */
1863                 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1864         }
1865         fi->writectr--;
1866         fuse_writepage_finish(fm, wpa);
1867         spin_unlock(&fi->lock);
1868         fuse_writepage_free(wpa);
1869 }
1870
1871 static struct fuse_file *__fuse_write_file_get(struct fuse_inode *fi)
1872 {
1873         struct fuse_file *ff;
1874
1875         spin_lock(&fi->lock);
1876         ff = list_first_entry_or_null(&fi->write_files, struct fuse_file,
1877                                       write_entry);
1878         if (ff)
1879                 fuse_file_get(ff);
1880         spin_unlock(&fi->lock);
1881
1882         return ff;
1883 }
1884
1885 static struct fuse_file *fuse_write_file_get(struct fuse_inode *fi)
1886 {
1887         struct fuse_file *ff = __fuse_write_file_get(fi);
1888         WARN_ON(!ff);
1889         return ff;
1890 }
1891
1892 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1893 {
1894         struct fuse_inode *fi = get_fuse_inode(inode);
1895         struct fuse_file *ff;
1896         int err;
1897
1898         /*
1899          * Inode is always written before the last reference is dropped and
1900          * hence this should not be reached from reclaim.
1901          *
1902          * Writing back the inode from reclaim can deadlock if the request
1903          * processing itself needs an allocation.  Allocations triggering
1904          * reclaim while serving a request can't be prevented, because it can
1905          * involve any number of unrelated userspace processes.
1906          */
1907         WARN_ON(wbc->for_reclaim);
1908
1909         ff = __fuse_write_file_get(fi);
1910         err = fuse_flush_times(inode, ff);
1911         if (ff)
1912                 fuse_file_put(ff, false, false);
1913
1914         return err;
1915 }
1916
1917 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1918 {
1919         struct fuse_writepage_args *wpa;
1920         struct fuse_args_pages *ap;
1921
1922         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1923         if (wpa) {
1924                 ap = &wpa->ia.ap;
1925                 ap->num_pages = 0;
1926                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1927                 if (!ap->pages) {
1928                         kfree(wpa);
1929                         wpa = NULL;
1930                 }
1931         }
1932         return wpa;
1933
1934 }
1935
1936 static void fuse_writepage_add_to_bucket(struct fuse_conn *fc,
1937                                          struct fuse_writepage_args *wpa)
1938 {
1939         if (!fc->sync_fs)
1940                 return;
1941
1942         rcu_read_lock();
1943         /* Prevent resurrection of dead bucket in unlikely race with syncfs */
1944         do {
1945                 wpa->bucket = rcu_dereference(fc->curr_bucket);
1946         } while (unlikely(!atomic_inc_not_zero(&wpa->bucket->count)));
1947         rcu_read_unlock();
1948 }
1949
1950 static int fuse_writepage_locked(struct page *page)
1951 {
1952         struct address_space *mapping = page->mapping;
1953         struct inode *inode = mapping->host;
1954         struct fuse_conn *fc = get_fuse_conn(inode);
1955         struct fuse_inode *fi = get_fuse_inode(inode);
1956         struct fuse_writepage_args *wpa;
1957         struct fuse_args_pages *ap;
1958         struct page *tmp_page;
1959         int error = -ENOMEM;
1960
1961         set_page_writeback(page);
1962
1963         wpa = fuse_writepage_args_alloc();
1964         if (!wpa)
1965                 goto err;
1966         ap = &wpa->ia.ap;
1967
1968         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1969         if (!tmp_page)
1970                 goto err_free;
1971
1972         error = -EIO;
1973         wpa->ia.ff = fuse_write_file_get(fi);
1974         if (!wpa->ia.ff)
1975                 goto err_nofile;
1976
1977         fuse_writepage_add_to_bucket(fc, wpa);
1978         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1979
1980         copy_highpage(tmp_page, page);
1981         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1982         wpa->next = NULL;
1983         ap->args.in_pages = true;
1984         ap->num_pages = 1;
1985         ap->pages[0] = tmp_page;
1986         ap->descs[0].offset = 0;
1987         ap->descs[0].length = PAGE_SIZE;
1988         ap->args.end = fuse_writepage_end;
1989         wpa->inode = inode;
1990
1991         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1992         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1993
1994         spin_lock(&fi->lock);
1995         tree_insert(&fi->writepages, wpa);
1996         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1997         fuse_flush_writepages(inode);
1998         spin_unlock(&fi->lock);
1999
2000         end_page_writeback(page);
2001
2002         return 0;
2003
2004 err_nofile:
2005         __free_page(tmp_page);
2006 err_free:
2007         kfree(wpa);
2008 err:
2009         mapping_set_error(page->mapping, error);
2010         end_page_writeback(page);
2011         return error;
2012 }
2013
2014 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
2015 {
2016         struct fuse_conn *fc = get_fuse_conn(page->mapping->host);
2017         int err;
2018
2019         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
2020                 /*
2021                  * ->writepages() should be called for sync() and friends.  We
2022                  * should only get here on direct reclaim and then we are
2023                  * allowed to skip a page which is already in flight
2024                  */
2025                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
2026
2027                 redirty_page_for_writepage(wbc, page);
2028                 unlock_page(page);
2029                 return 0;
2030         }
2031
2032         if (wbc->sync_mode == WB_SYNC_NONE &&
2033             fc->num_background >= fc->congestion_threshold)
2034                 return AOP_WRITEPAGE_ACTIVATE;
2035
2036         err = fuse_writepage_locked(page);
2037         unlock_page(page);
2038
2039         return err;
2040 }
2041
2042 struct fuse_fill_wb_data {
2043         struct fuse_writepage_args *wpa;
2044         struct fuse_file *ff;
2045         struct inode *inode;
2046         struct page **orig_pages;
2047         unsigned int max_pages;
2048 };
2049
2050 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
2051 {
2052         struct fuse_args_pages *ap = &data->wpa->ia.ap;
2053         struct fuse_conn *fc = get_fuse_conn(data->inode);
2054         struct page **pages;
2055         struct fuse_page_desc *descs;
2056         unsigned int npages = min_t(unsigned int,
2057                                     max_t(unsigned int, data->max_pages * 2,
2058                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
2059                                     fc->max_pages);
2060         WARN_ON(npages <= data->max_pages);
2061
2062         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
2063         if (!pages)
2064                 return false;
2065
2066         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
2067         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
2068         kfree(ap->pages);
2069         ap->pages = pages;
2070         ap->descs = descs;
2071         data->max_pages = npages;
2072
2073         return true;
2074 }
2075
2076 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
2077 {
2078         struct fuse_writepage_args *wpa = data->wpa;
2079         struct inode *inode = data->inode;
2080         struct fuse_inode *fi = get_fuse_inode(inode);
2081         int num_pages = wpa->ia.ap.num_pages;
2082         int i;
2083
2084         wpa->ia.ff = fuse_file_get(data->ff);
2085         spin_lock(&fi->lock);
2086         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2087         fuse_flush_writepages(inode);
2088         spin_unlock(&fi->lock);
2089
2090         for (i = 0; i < num_pages; i++)
2091                 end_page_writeback(data->orig_pages[i]);
2092 }
2093
2094 /*
2095  * Check under fi->lock if the page is under writeback, and insert it onto the
2096  * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2097  * one already added for a page at this offset.  If there's none, then insert
2098  * this new request onto the auxiliary list, otherwise reuse the existing one by
2099  * swapping the new temp page with the old one.
2100  */
2101 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2102                                struct page *page)
2103 {
2104         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2105         struct fuse_writepage_args *tmp;
2106         struct fuse_writepage_args *old_wpa;
2107         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2108
2109         WARN_ON(new_ap->num_pages != 0);
2110         new_ap->num_pages = 1;
2111
2112         spin_lock(&fi->lock);
2113         old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2114         if (!old_wpa) {
2115                 spin_unlock(&fi->lock);
2116                 return true;
2117         }
2118
2119         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2120                 pgoff_t curr_index;
2121
2122                 WARN_ON(tmp->inode != new_wpa->inode);
2123                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2124                 if (curr_index == page->index) {
2125                         WARN_ON(tmp->ia.ap.num_pages != 1);
2126                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2127                         break;
2128                 }
2129         }
2130
2131         if (!tmp) {
2132                 new_wpa->next = old_wpa->next;
2133                 old_wpa->next = new_wpa;
2134         }
2135
2136         spin_unlock(&fi->lock);
2137
2138         if (tmp) {
2139                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2140
2141                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2142                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2143                 wb_writeout_inc(&bdi->wb);
2144                 fuse_writepage_free(new_wpa);
2145         }
2146
2147         return false;
2148 }
2149
2150 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2151                                      struct fuse_args_pages *ap,
2152                                      struct fuse_fill_wb_data *data)
2153 {
2154         WARN_ON(!ap->num_pages);
2155
2156         /*
2157          * Being under writeback is unlikely but possible.  For example direct
2158          * read to an mmaped fuse file will set the page dirty twice; once when
2159          * the pages are faulted with get_user_pages(), and then after the read
2160          * completed.
2161          */
2162         if (fuse_page_is_writeback(data->inode, page->index))
2163                 return true;
2164
2165         /* Reached max pages */
2166         if (ap->num_pages == fc->max_pages)
2167                 return true;
2168
2169         /* Reached max write bytes */
2170         if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2171                 return true;
2172
2173         /* Discontinuity */
2174         if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2175                 return true;
2176
2177         /* Need to grow the pages array?  If so, did the expansion fail? */
2178         if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2179                 return true;
2180
2181         return false;
2182 }
2183
2184 static int fuse_writepages_fill(struct folio *folio,
2185                 struct writeback_control *wbc, void *_data)
2186 {
2187         struct fuse_fill_wb_data *data = _data;
2188         struct fuse_writepage_args *wpa = data->wpa;
2189         struct fuse_args_pages *ap = &wpa->ia.ap;
2190         struct inode *inode = data->inode;
2191         struct fuse_inode *fi = get_fuse_inode(inode);
2192         struct fuse_conn *fc = get_fuse_conn(inode);
2193         struct page *tmp_page;
2194         int err;
2195
2196         if (!data->ff) {
2197                 err = -EIO;
2198                 data->ff = fuse_write_file_get(fi);
2199                 if (!data->ff)
2200                         goto out_unlock;
2201         }
2202
2203         if (wpa && fuse_writepage_need_send(fc, &folio->page, ap, data)) {
2204                 fuse_writepages_send(data);
2205                 data->wpa = NULL;
2206         }
2207
2208         err = -ENOMEM;
2209         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2210         if (!tmp_page)
2211                 goto out_unlock;
2212
2213         /*
2214          * The page must not be redirtied until the writeout is completed
2215          * (i.e. userspace has sent a reply to the write request).  Otherwise
2216          * there could be more than one temporary page instance for each real
2217          * page.
2218          *
2219          * This is ensured by holding the page lock in page_mkwrite() while
2220          * checking fuse_page_is_writeback().  We already hold the page lock
2221          * since clear_page_dirty_for_io() and keep it held until we add the
2222          * request to the fi->writepages list and increment ap->num_pages.
2223          * After this fuse_page_is_writeback() will indicate that the page is
2224          * under writeback, so we can release the page lock.
2225          */
2226         if (data->wpa == NULL) {
2227                 err = -ENOMEM;
2228                 wpa = fuse_writepage_args_alloc();
2229                 if (!wpa) {
2230                         __free_page(tmp_page);
2231                         goto out_unlock;
2232                 }
2233                 fuse_writepage_add_to_bucket(fc, wpa);
2234
2235                 data->max_pages = 1;
2236
2237                 ap = &wpa->ia.ap;
2238                 fuse_write_args_fill(&wpa->ia, data->ff, folio_pos(folio), 0);
2239                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2240                 wpa->next = NULL;
2241                 ap->args.in_pages = true;
2242                 ap->args.end = fuse_writepage_end;
2243                 ap->num_pages = 0;
2244                 wpa->inode = inode;
2245         }
2246         folio_start_writeback(folio);
2247
2248         copy_highpage(tmp_page, &folio->page);
2249         ap->pages[ap->num_pages] = tmp_page;
2250         ap->descs[ap->num_pages].offset = 0;
2251         ap->descs[ap->num_pages].length = PAGE_SIZE;
2252         data->orig_pages[ap->num_pages] = &folio->page;
2253
2254         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2255         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2256
2257         err = 0;
2258         if (data->wpa) {
2259                 /*
2260                  * Protected by fi->lock against concurrent access by
2261                  * fuse_page_is_writeback().
2262                  */
2263                 spin_lock(&fi->lock);
2264                 ap->num_pages++;
2265                 spin_unlock(&fi->lock);
2266         } else if (fuse_writepage_add(wpa, &folio->page)) {
2267                 data->wpa = wpa;
2268         } else {
2269                 folio_end_writeback(folio);
2270         }
2271 out_unlock:
2272         folio_unlock(folio);
2273
2274         return err;
2275 }
2276
2277 static int fuse_writepages(struct address_space *mapping,
2278                            struct writeback_control *wbc)
2279 {
2280         struct inode *inode = mapping->host;
2281         struct fuse_conn *fc = get_fuse_conn(inode);
2282         struct fuse_fill_wb_data data;
2283         int err;
2284
2285         err = -EIO;
2286         if (fuse_is_bad(inode))
2287                 goto out;
2288
2289         if (wbc->sync_mode == WB_SYNC_NONE &&
2290             fc->num_background >= fc->congestion_threshold)
2291                 return 0;
2292
2293         data.inode = inode;
2294         data.wpa = NULL;
2295         data.ff = NULL;
2296
2297         err = -ENOMEM;
2298         data.orig_pages = kcalloc(fc->max_pages,
2299                                   sizeof(struct page *),
2300                                   GFP_NOFS);
2301         if (!data.orig_pages)
2302                 goto out;
2303
2304         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2305         if (data.wpa) {
2306                 WARN_ON(!data.wpa->ia.ap.num_pages);
2307                 fuse_writepages_send(&data);
2308         }
2309         if (data.ff)
2310                 fuse_file_put(data.ff, false, false);
2311
2312         kfree(data.orig_pages);
2313 out:
2314         return err;
2315 }
2316
2317 /*
2318  * It's worthy to make sure that space is reserved on disk for the write,
2319  * but how to implement it without killing performance need more thinking.
2320  */
2321 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2322                 loff_t pos, unsigned len, struct page **pagep, void **fsdata)
2323 {
2324         pgoff_t index = pos >> PAGE_SHIFT;
2325         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2326         struct page *page;
2327         loff_t fsize;
2328         int err = -ENOMEM;
2329
2330         WARN_ON(!fc->writeback_cache);
2331
2332         page = grab_cache_page_write_begin(mapping, index);
2333         if (!page)
2334                 goto error;
2335
2336         fuse_wait_on_page_writeback(mapping->host, page->index);
2337
2338         if (PageUptodate(page) || len == PAGE_SIZE)
2339                 goto success;
2340         /*
2341          * Check if the start this page comes after the end of file, in which
2342          * case the readpage can be optimized away.
2343          */
2344         fsize = i_size_read(mapping->host);
2345         if (fsize <= (pos & PAGE_MASK)) {
2346                 size_t off = pos & ~PAGE_MASK;
2347                 if (off)
2348                         zero_user_segment(page, 0, off);
2349                 goto success;
2350         }
2351         err = fuse_do_readpage(file, page);
2352         if (err)
2353                 goto cleanup;
2354 success:
2355         *pagep = page;
2356         return 0;
2357
2358 cleanup:
2359         unlock_page(page);
2360         put_page(page);
2361 error:
2362         return err;
2363 }
2364
2365 static int fuse_write_end(struct file *file, struct address_space *mapping,
2366                 loff_t pos, unsigned len, unsigned copied,
2367                 struct page *page, void *fsdata)
2368 {
2369         struct inode *inode = page->mapping->host;
2370
2371         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2372         if (!copied)
2373                 goto unlock;
2374
2375         pos += copied;
2376         if (!PageUptodate(page)) {
2377                 /* Zero any unwritten bytes at the end of the page */
2378                 size_t endoff = pos & ~PAGE_MASK;
2379                 if (endoff)
2380                         zero_user_segment(page, endoff, PAGE_SIZE);
2381                 SetPageUptodate(page);
2382         }
2383
2384         if (pos > inode->i_size)
2385                 i_size_write(inode, pos);
2386
2387         set_page_dirty(page);
2388
2389 unlock:
2390         unlock_page(page);
2391         put_page(page);
2392
2393         return copied;
2394 }
2395
2396 static int fuse_launder_folio(struct folio *folio)
2397 {
2398         int err = 0;
2399         if (folio_clear_dirty_for_io(folio)) {
2400                 struct inode *inode = folio->mapping->host;
2401
2402                 /* Serialize with pending writeback for the same page */
2403                 fuse_wait_on_page_writeback(inode, folio->index);
2404                 err = fuse_writepage_locked(&folio->page);
2405                 if (!err)
2406                         fuse_wait_on_page_writeback(inode, folio->index);
2407         }
2408         return err;
2409 }
2410
2411 /*
2412  * Write back dirty data/metadata now (there may not be any suitable
2413  * open files later for data)
2414  */
2415 static void fuse_vma_close(struct vm_area_struct *vma)
2416 {
2417         int err;
2418
2419         err = write_inode_now(vma->vm_file->f_mapping->host, 1);
2420         mapping_set_error(vma->vm_file->f_mapping, err);
2421 }
2422
2423 /*
2424  * Wait for writeback against this page to complete before allowing it
2425  * to be marked dirty again, and hence written back again, possibly
2426  * before the previous writepage completed.
2427  *
2428  * Block here, instead of in ->writepage(), so that the userspace fs
2429  * can only block processes actually operating on the filesystem.
2430  *
2431  * Otherwise unprivileged userspace fs would be able to block
2432  * unrelated:
2433  *
2434  * - page migration
2435  * - sync(2)
2436  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2437  */
2438 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2439 {
2440         struct page *page = vmf->page;
2441         struct inode *inode = file_inode(vmf->vma->vm_file);
2442
2443         file_update_time(vmf->vma->vm_file);
2444         lock_page(page);
2445         if (page->mapping != inode->i_mapping) {
2446                 unlock_page(page);
2447                 return VM_FAULT_NOPAGE;
2448         }
2449
2450         fuse_wait_on_page_writeback(inode, page->index);
2451         return VM_FAULT_LOCKED;
2452 }
2453
2454 static const struct vm_operations_struct fuse_file_vm_ops = {
2455         .close          = fuse_vma_close,
2456         .fault          = filemap_fault,
2457         .map_pages      = filemap_map_pages,
2458         .page_mkwrite   = fuse_page_mkwrite,
2459 };
2460
2461 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2462 {
2463         struct fuse_file *ff = file->private_data;
2464         struct fuse_conn *fc = ff->fm->fc;
2465
2466         /* DAX mmap is superior to direct_io mmap */
2467         if (FUSE_IS_DAX(file_inode(file)))
2468                 return fuse_dax_mmap(file, vma);
2469
2470         if (ff->open_flags & FOPEN_DIRECT_IO) {
2471                 /* Can't provide the coherency needed for MAP_SHARED
2472                  * if FUSE_DIRECT_IO_ALLOW_MMAP isn't set.
2473                  */
2474                 if ((vma->vm_flags & VM_MAYSHARE) && !fc->direct_io_allow_mmap)
2475                         return -ENODEV;
2476
2477                 invalidate_inode_pages2(file->f_mapping);
2478
2479                 return generic_file_mmap(file, vma);
2480         }
2481
2482         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2483                 fuse_link_write_file(file);
2484
2485         file_accessed(file);
2486         vma->vm_ops = &fuse_file_vm_ops;
2487         return 0;
2488 }
2489
2490 static int convert_fuse_file_lock(struct fuse_conn *fc,
2491                                   const struct fuse_file_lock *ffl,
2492                                   struct file_lock *fl)
2493 {
2494         switch (ffl->type) {
2495         case F_UNLCK:
2496                 break;
2497
2498         case F_RDLCK:
2499         case F_WRLCK:
2500                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2501                     ffl->end < ffl->start)
2502                         return -EIO;
2503
2504                 fl->fl_start = ffl->start;
2505                 fl->fl_end = ffl->end;
2506
2507                 /*
2508                  * Convert pid into init's pid namespace.  The locks API will
2509                  * translate it into the caller's pid namespace.
2510                  */
2511                 rcu_read_lock();
2512                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2513                 rcu_read_unlock();
2514                 break;
2515
2516         default:
2517                 return -EIO;
2518         }
2519         fl->fl_type = ffl->type;
2520         return 0;
2521 }
2522
2523 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2524                          const struct file_lock *fl, int opcode, pid_t pid,
2525                          int flock, struct fuse_lk_in *inarg)
2526 {
2527         struct inode *inode = file_inode(file);
2528         struct fuse_conn *fc = get_fuse_conn(inode);
2529         struct fuse_file *ff = file->private_data;
2530
2531         memset(inarg, 0, sizeof(*inarg));
2532         inarg->fh = ff->fh;
2533         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2534         inarg->lk.start = fl->fl_start;
2535         inarg->lk.end = fl->fl_end;
2536         inarg->lk.type = fl->fl_type;
2537         inarg->lk.pid = pid;
2538         if (flock)
2539                 inarg->lk_flags |= FUSE_LK_FLOCK;
2540         args->opcode = opcode;
2541         args->nodeid = get_node_id(inode);
2542         args->in_numargs = 1;
2543         args->in_args[0].size = sizeof(*inarg);
2544         args->in_args[0].value = inarg;
2545 }
2546
2547 static int fuse_getlk(struct file *file, struct file_lock *fl)
2548 {
2549         struct inode *inode = file_inode(file);
2550         struct fuse_mount *fm = get_fuse_mount(inode);
2551         FUSE_ARGS(args);
2552         struct fuse_lk_in inarg;
2553         struct fuse_lk_out outarg;
2554         int err;
2555
2556         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2557         args.out_numargs = 1;
2558         args.out_args[0].size = sizeof(outarg);
2559         args.out_args[0].value = &outarg;
2560         err = fuse_simple_request(fm, &args);
2561         if (!err)
2562                 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2563
2564         return err;
2565 }
2566
2567 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2568 {
2569         struct inode *inode = file_inode(file);
2570         struct fuse_mount *fm = get_fuse_mount(inode);
2571         FUSE_ARGS(args);
2572         struct fuse_lk_in inarg;
2573         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2574         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2575         pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2576         int err;
2577
2578         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2579                 /* NLM needs asynchronous locks, which we don't support yet */
2580                 return -ENOLCK;
2581         }
2582
2583         /* Unlock on close is handled by the flush method */
2584         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2585                 return 0;
2586
2587         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2588         err = fuse_simple_request(fm, &args);
2589
2590         /* locking is restartable */
2591         if (err == -EINTR)
2592                 err = -ERESTARTSYS;
2593
2594         return err;
2595 }
2596
2597 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2598 {
2599         struct inode *inode = file_inode(file);
2600         struct fuse_conn *fc = get_fuse_conn(inode);
2601         int err;
2602
2603         if (cmd == F_CANCELLK) {
2604                 err = 0;
2605         } else if (cmd == F_GETLK) {
2606                 if (fc->no_lock) {
2607                         posix_test_lock(file, fl);
2608                         err = 0;
2609                 } else
2610                         err = fuse_getlk(file, fl);
2611         } else {
2612                 if (fc->no_lock)
2613                         err = posix_lock_file(file, fl, NULL);
2614                 else
2615                         err = fuse_setlk(file, fl, 0);
2616         }
2617         return err;
2618 }
2619
2620 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2621 {
2622         struct inode *inode = file_inode(file);
2623         struct fuse_conn *fc = get_fuse_conn(inode);
2624         int err;
2625
2626         if (fc->no_flock) {
2627                 err = locks_lock_file_wait(file, fl);
2628         } else {
2629                 struct fuse_file *ff = file->private_data;
2630
2631                 /* emulate flock with POSIX locks */
2632                 ff->flock = true;
2633                 err = fuse_setlk(file, fl, 1);
2634         }
2635
2636         return err;
2637 }
2638
2639 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2640 {
2641         struct inode *inode = mapping->host;
2642         struct fuse_mount *fm = get_fuse_mount(inode);
2643         FUSE_ARGS(args);
2644         struct fuse_bmap_in inarg;
2645         struct fuse_bmap_out outarg;
2646         int err;
2647
2648         if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2649                 return 0;
2650
2651         memset(&inarg, 0, sizeof(inarg));
2652         inarg.block = block;
2653         inarg.blocksize = inode->i_sb->s_blocksize;
2654         args.opcode = FUSE_BMAP;
2655         args.nodeid = get_node_id(inode);
2656         args.in_numargs = 1;
2657         args.in_args[0].size = sizeof(inarg);
2658         args.in_args[0].value = &inarg;
2659         args.out_numargs = 1;
2660         args.out_args[0].size = sizeof(outarg);
2661         args.out_args[0].value = &outarg;
2662         err = fuse_simple_request(fm, &args);
2663         if (err == -ENOSYS)
2664                 fm->fc->no_bmap = 1;
2665
2666         return err ? 0 : outarg.block;
2667 }
2668
2669 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2670 {
2671         struct inode *inode = file->f_mapping->host;
2672         struct fuse_mount *fm = get_fuse_mount(inode);
2673         struct fuse_file *ff = file->private_data;
2674         FUSE_ARGS(args);
2675         struct fuse_lseek_in inarg = {
2676                 .fh = ff->fh,
2677                 .offset = offset,
2678                 .whence = whence
2679         };
2680         struct fuse_lseek_out outarg;
2681         int err;
2682
2683         if (fm->fc->no_lseek)
2684                 goto fallback;
2685
2686         args.opcode = FUSE_LSEEK;
2687         args.nodeid = ff->nodeid;
2688         args.in_numargs = 1;
2689         args.in_args[0].size = sizeof(inarg);
2690         args.in_args[0].value = &inarg;
2691         args.out_numargs = 1;
2692         args.out_args[0].size = sizeof(outarg);
2693         args.out_args[0].value = &outarg;
2694         err = fuse_simple_request(fm, &args);
2695         if (err) {
2696                 if (err == -ENOSYS) {
2697                         fm->fc->no_lseek = 1;
2698                         goto fallback;
2699                 }
2700                 return err;
2701         }
2702
2703         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2704
2705 fallback:
2706         err = fuse_update_attributes(inode, file, STATX_SIZE);
2707         if (!err)
2708                 return generic_file_llseek(file, offset, whence);
2709         else
2710                 return err;
2711 }
2712
2713 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2714 {
2715         loff_t retval;
2716         struct inode *inode = file_inode(file);
2717
2718         switch (whence) {
2719         case SEEK_SET:
2720         case SEEK_CUR:
2721                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2722                 retval = generic_file_llseek(file, offset, whence);
2723                 break;
2724         case SEEK_END:
2725                 inode_lock(inode);
2726                 retval = fuse_update_attributes(inode, file, STATX_SIZE);
2727                 if (!retval)
2728                         retval = generic_file_llseek(file, offset, whence);
2729                 inode_unlock(inode);
2730                 break;
2731         case SEEK_HOLE:
2732         case SEEK_DATA:
2733                 inode_lock(inode);
2734                 retval = fuse_lseek(file, offset, whence);
2735                 inode_unlock(inode);
2736                 break;
2737         default:
2738                 retval = -EINVAL;
2739         }
2740
2741         return retval;
2742 }
2743
2744 /*
2745  * All files which have been polled are linked to RB tree
2746  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2747  * find the matching one.
2748  */
2749 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2750                                               struct rb_node **parent_out)
2751 {
2752         struct rb_node **link = &fc->polled_files.rb_node;
2753         struct rb_node *last = NULL;
2754
2755         while (*link) {
2756                 struct fuse_file *ff;
2757
2758                 last = *link;
2759                 ff = rb_entry(last, struct fuse_file, polled_node);
2760
2761                 if (kh < ff->kh)
2762                         link = &last->rb_left;
2763                 else if (kh > ff->kh)
2764                         link = &last->rb_right;
2765                 else
2766                         return link;
2767         }
2768
2769         if (parent_out)
2770                 *parent_out = last;
2771         return link;
2772 }
2773
2774 /*
2775  * The file is about to be polled.  Make sure it's on the polled_files
2776  * RB tree.  Note that files once added to the polled_files tree are
2777  * not removed before the file is released.  This is because a file
2778  * polled once is likely to be polled again.
2779  */
2780 static void fuse_register_polled_file(struct fuse_conn *fc,
2781                                       struct fuse_file *ff)
2782 {
2783         spin_lock(&fc->lock);
2784         if (RB_EMPTY_NODE(&ff->polled_node)) {
2785                 struct rb_node **link, *parent;
2786
2787                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2788                 BUG_ON(*link);
2789                 rb_link_node(&ff->polled_node, parent, link);
2790                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2791         }
2792         spin_unlock(&fc->lock);
2793 }
2794
2795 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2796 {
2797         struct fuse_file *ff = file->private_data;
2798         struct fuse_mount *fm = ff->fm;
2799         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2800         struct fuse_poll_out outarg;
2801         FUSE_ARGS(args);
2802         int err;
2803
2804         if (fm->fc->no_poll)
2805                 return DEFAULT_POLLMASK;
2806
2807         poll_wait(file, &ff->poll_wait, wait);
2808         inarg.events = mangle_poll(poll_requested_events(wait));
2809
2810         /*
2811          * Ask for notification iff there's someone waiting for it.
2812          * The client may ignore the flag and always notify.
2813          */
2814         if (waitqueue_active(&ff->poll_wait)) {
2815                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2816                 fuse_register_polled_file(fm->fc, ff);
2817         }
2818
2819         args.opcode = FUSE_POLL;
2820         args.nodeid = ff->nodeid;
2821         args.in_numargs = 1;
2822         args.in_args[0].size = sizeof(inarg);
2823         args.in_args[0].value = &inarg;
2824         args.out_numargs = 1;
2825         args.out_args[0].size = sizeof(outarg);
2826         args.out_args[0].value = &outarg;
2827         err = fuse_simple_request(fm, &args);
2828
2829         if (!err)
2830                 return demangle_poll(outarg.revents);
2831         if (err == -ENOSYS) {
2832                 fm->fc->no_poll = 1;
2833                 return DEFAULT_POLLMASK;
2834         }
2835         return EPOLLERR;
2836 }
2837 EXPORT_SYMBOL_GPL(fuse_file_poll);
2838
2839 /*
2840  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2841  * wakes up the poll waiters.
2842  */
2843 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2844                             struct fuse_notify_poll_wakeup_out *outarg)
2845 {
2846         u64 kh = outarg->kh;
2847         struct rb_node **link;
2848
2849         spin_lock(&fc->lock);
2850
2851         link = fuse_find_polled_node(fc, kh, NULL);
2852         if (*link) {
2853                 struct fuse_file *ff;
2854
2855                 ff = rb_entry(*link, struct fuse_file, polled_node);
2856                 wake_up_interruptible_sync(&ff->poll_wait);
2857         }
2858
2859         spin_unlock(&fc->lock);
2860         return 0;
2861 }
2862
2863 static void fuse_do_truncate(struct file *file)
2864 {
2865         struct inode *inode = file->f_mapping->host;
2866         struct iattr attr;
2867
2868         attr.ia_valid = ATTR_SIZE;
2869         attr.ia_size = i_size_read(inode);
2870
2871         attr.ia_file = file;
2872         attr.ia_valid |= ATTR_FILE;
2873
2874         fuse_do_setattr(file_dentry(file), &attr, file);
2875 }
2876
2877 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2878 {
2879         return round_up(off, fc->max_pages << PAGE_SHIFT);
2880 }
2881
2882 static ssize_t
2883 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2884 {
2885         DECLARE_COMPLETION_ONSTACK(wait);
2886         ssize_t ret = 0;
2887         struct file *file = iocb->ki_filp;
2888         struct fuse_file *ff = file->private_data;
2889         loff_t pos = 0;
2890         struct inode *inode;
2891         loff_t i_size;
2892         size_t count = iov_iter_count(iter), shortened = 0;
2893         loff_t offset = iocb->ki_pos;
2894         struct fuse_io_priv *io;
2895
2896         pos = offset;
2897         inode = file->f_mapping->host;
2898         i_size = i_size_read(inode);
2899
2900         if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
2901                 return 0;
2902
2903         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2904         if (!io)
2905                 return -ENOMEM;
2906         spin_lock_init(&io->lock);
2907         kref_init(&io->refcnt);
2908         io->reqs = 1;
2909         io->bytes = -1;
2910         io->size = 0;
2911         io->offset = offset;
2912         io->write = (iov_iter_rw(iter) == WRITE);
2913         io->err = 0;
2914         /*
2915          * By default, we want to optimize all I/Os with async request
2916          * submission to the client filesystem if supported.
2917          */
2918         io->async = ff->fm->fc->async_dio;
2919         io->iocb = iocb;
2920         io->blocking = is_sync_kiocb(iocb);
2921
2922         /* optimization for short read */
2923         if (io->async && !io->write && offset + count > i_size) {
2924                 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
2925                 shortened = count - iov_iter_count(iter);
2926                 count -= shortened;
2927         }
2928
2929         /*
2930          * We cannot asynchronously extend the size of a file.
2931          * In such case the aio will behave exactly like sync io.
2932          */
2933         if ((offset + count > i_size) && io->write)
2934                 io->blocking = true;
2935
2936         if (io->async && io->blocking) {
2937                 /*
2938                  * Additional reference to keep io around after
2939                  * calling fuse_aio_complete()
2940                  */
2941                 kref_get(&io->refcnt);
2942                 io->done = &wait;
2943         }
2944
2945         if (iov_iter_rw(iter) == WRITE) {
2946                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2947                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
2948         } else {
2949                 ret = __fuse_direct_read(io, iter, &pos);
2950         }
2951         iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
2952
2953         if (io->async) {
2954                 bool blocking = io->blocking;
2955
2956                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2957
2958                 /* we have a non-extending, async request, so return */
2959                 if (!blocking)
2960                         return -EIOCBQUEUED;
2961
2962                 wait_for_completion(&wait);
2963                 ret = fuse_get_res_by_io(io);
2964         }
2965
2966         kref_put(&io->refcnt, fuse_io_release);
2967
2968         if (iov_iter_rw(iter) == WRITE) {
2969                 fuse_write_update_attr(inode, pos, ret);
2970                 /* For extending writes we already hold exclusive lock */
2971                 if (ret < 0 && offset + count > i_size)
2972                         fuse_do_truncate(file);
2973         }
2974
2975         return ret;
2976 }
2977
2978 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
2979 {
2980         int err = filemap_write_and_wait_range(inode->i_mapping, start, LLONG_MAX);
2981
2982         if (!err)
2983                 fuse_sync_writes(inode);
2984
2985         return err;
2986 }
2987
2988 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2989                                 loff_t length)
2990 {
2991         struct fuse_file *ff = file->private_data;
2992         struct inode *inode = file_inode(file);
2993         struct fuse_inode *fi = get_fuse_inode(inode);
2994         struct fuse_mount *fm = ff->fm;
2995         FUSE_ARGS(args);
2996         struct fuse_fallocate_in inarg = {
2997                 .fh = ff->fh,
2998                 .offset = offset,
2999                 .length = length,
3000                 .mode = mode
3001         };
3002         int err;
3003         bool block_faults = FUSE_IS_DAX(inode) &&
3004                 (!(mode & FALLOC_FL_KEEP_SIZE) ||
3005                  (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)));
3006
3007         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3008                      FALLOC_FL_ZERO_RANGE))
3009                 return -EOPNOTSUPP;
3010
3011         if (fm->fc->no_fallocate)
3012                 return -EOPNOTSUPP;
3013
3014         inode_lock(inode);
3015         if (block_faults) {
3016                 filemap_invalidate_lock(inode->i_mapping);
3017                 err = fuse_dax_break_layouts(inode, 0, 0);
3018                 if (err)
3019                         goto out;
3020         }
3021
3022         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)) {
3023                 loff_t endbyte = offset + length - 1;
3024
3025                 err = fuse_writeback_range(inode, offset, endbyte);
3026                 if (err)
3027                         goto out;
3028         }
3029
3030         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3031             offset + length > i_size_read(inode)) {
3032                 err = inode_newsize_ok(inode, offset + length);
3033                 if (err)
3034                         goto out;
3035         }
3036
3037         err = file_modified(file);
3038         if (err)
3039                 goto out;
3040
3041         if (!(mode & FALLOC_FL_KEEP_SIZE))
3042                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3043
3044         args.opcode = FUSE_FALLOCATE;
3045         args.nodeid = ff->nodeid;
3046         args.in_numargs = 1;
3047         args.in_args[0].size = sizeof(inarg);
3048         args.in_args[0].value = &inarg;
3049         err = fuse_simple_request(fm, &args);
3050         if (err == -ENOSYS) {
3051                 fm->fc->no_fallocate = 1;
3052                 err = -EOPNOTSUPP;
3053         }
3054         if (err)
3055                 goto out;
3056
3057         /* we could have extended the file */
3058         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3059                 if (fuse_write_update_attr(inode, offset + length, length))
3060                         file_update_time(file);
3061         }
3062
3063         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE))
3064                 truncate_pagecache_range(inode, offset, offset + length - 1);
3065
3066         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
3067
3068 out:
3069         if (!(mode & FALLOC_FL_KEEP_SIZE))
3070                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3071
3072         if (block_faults)
3073                 filemap_invalidate_unlock(inode->i_mapping);
3074
3075         inode_unlock(inode);
3076
3077         fuse_flush_time_update(inode);
3078
3079         return err;
3080 }
3081
3082 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3083                                       struct file *file_out, loff_t pos_out,
3084                                       size_t len, unsigned int flags)
3085 {
3086         struct fuse_file *ff_in = file_in->private_data;
3087         struct fuse_file *ff_out = file_out->private_data;
3088         struct inode *inode_in = file_inode(file_in);
3089         struct inode *inode_out = file_inode(file_out);
3090         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3091         struct fuse_mount *fm = ff_in->fm;
3092         struct fuse_conn *fc = fm->fc;
3093         FUSE_ARGS(args);
3094         struct fuse_copy_file_range_in inarg = {
3095                 .fh_in = ff_in->fh,
3096                 .off_in = pos_in,
3097                 .nodeid_out = ff_out->nodeid,
3098                 .fh_out = ff_out->fh,
3099                 .off_out = pos_out,
3100                 .len = len,
3101                 .flags = flags
3102         };
3103         struct fuse_write_out outarg;
3104         ssize_t err;
3105         /* mark unstable when write-back is not used, and file_out gets
3106          * extended */
3107         bool is_unstable = (!fc->writeback_cache) &&
3108                            ((pos_out + len) > inode_out->i_size);
3109
3110         if (fc->no_copy_file_range)
3111                 return -EOPNOTSUPP;
3112
3113         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3114                 return -EXDEV;
3115
3116         inode_lock(inode_in);
3117         err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3118         inode_unlock(inode_in);
3119         if (err)
3120                 return err;
3121
3122         inode_lock(inode_out);
3123
3124         err = file_modified(file_out);
3125         if (err)
3126                 goto out;
3127
3128         /*
3129          * Write out dirty pages in the destination file before sending the COPY
3130          * request to userspace.  After the request is completed, truncate off
3131          * pages (including partial ones) from the cache that have been copied,
3132          * since these contain stale data at that point.
3133          *
3134          * This should be mostly correct, but if the COPY writes to partial
3135          * pages (at the start or end) and the parts not covered by the COPY are
3136          * written through a memory map after calling fuse_writeback_range(),
3137          * then these partial page modifications will be lost on truncation.
3138          *
3139          * It is unlikely that someone would rely on such mixed style
3140          * modifications.  Yet this does give less guarantees than if the
3141          * copying was performed with write(2).
3142          *
3143          * To fix this a mapping->invalidate_lock could be used to prevent new
3144          * faults while the copy is ongoing.
3145          */
3146         err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3147         if (err)
3148                 goto out;
3149
3150         if (is_unstable)
3151                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3152
3153         args.opcode = FUSE_COPY_FILE_RANGE;
3154         args.nodeid = ff_in->nodeid;
3155         args.in_numargs = 1;
3156         args.in_args[0].size = sizeof(inarg);
3157         args.in_args[0].value = &inarg;
3158         args.out_numargs = 1;
3159         args.out_args[0].size = sizeof(outarg);
3160         args.out_args[0].value = &outarg;
3161         err = fuse_simple_request(fm, &args);
3162         if (err == -ENOSYS) {
3163                 fc->no_copy_file_range = 1;
3164                 err = -EOPNOTSUPP;
3165         }
3166         if (err)
3167                 goto out;
3168
3169         truncate_inode_pages_range(inode_out->i_mapping,
3170                                    ALIGN_DOWN(pos_out, PAGE_SIZE),
3171                                    ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3172
3173         file_update_time(file_out);
3174         fuse_write_update_attr(inode_out, pos_out + outarg.size, outarg.size);
3175
3176         err = outarg.size;
3177 out:
3178         if (is_unstable)
3179                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3180
3181         inode_unlock(inode_out);
3182         file_accessed(file_in);
3183
3184         fuse_flush_time_update(inode_out);
3185
3186         return err;
3187 }
3188
3189 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3190                                     struct file *dst_file, loff_t dst_off,
3191                                     size_t len, unsigned int flags)
3192 {
3193         ssize_t ret;
3194
3195         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3196                                      len, flags);
3197
3198         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3199                 ret = splice_copy_file_range(src_file, src_off, dst_file,
3200                                              dst_off, len);
3201         return ret;
3202 }
3203
3204 static const struct file_operations fuse_file_operations = {
3205         .llseek         = fuse_file_llseek,
3206         .read_iter      = fuse_file_read_iter,
3207         .write_iter     = fuse_file_write_iter,
3208         .mmap           = fuse_file_mmap,
3209         .open           = fuse_open,
3210         .flush          = fuse_flush,
3211         .release        = fuse_release,
3212         .fsync          = fuse_fsync,
3213         .lock           = fuse_file_lock,
3214         .get_unmapped_area = thp_get_unmapped_area,
3215         .flock          = fuse_file_flock,
3216         .splice_read    = filemap_splice_read,
3217         .splice_write   = iter_file_splice_write,
3218         .unlocked_ioctl = fuse_file_ioctl,
3219         .compat_ioctl   = fuse_file_compat_ioctl,
3220         .poll           = fuse_file_poll,
3221         .fallocate      = fuse_file_fallocate,
3222         .copy_file_range = fuse_copy_file_range,
3223 };
3224
3225 static const struct address_space_operations fuse_file_aops  = {
3226         .read_folio     = fuse_read_folio,
3227         .readahead      = fuse_readahead,
3228         .writepage      = fuse_writepage,
3229         .writepages     = fuse_writepages,
3230         .launder_folio  = fuse_launder_folio,
3231         .dirty_folio    = filemap_dirty_folio,
3232         .bmap           = fuse_bmap,
3233         .direct_IO      = fuse_direct_IO,
3234         .write_begin    = fuse_write_begin,
3235         .write_end      = fuse_write_end,
3236 };
3237
3238 void fuse_init_file_inode(struct inode *inode, unsigned int flags)
3239 {
3240         struct fuse_inode *fi = get_fuse_inode(inode);
3241
3242         inode->i_fop = &fuse_file_operations;
3243         inode->i_data.a_ops = &fuse_file_aops;
3244
3245         INIT_LIST_HEAD(&fi->write_files);
3246         INIT_LIST_HEAD(&fi->queued_writes);
3247         fi->writectr = 0;
3248         init_waitqueue_head(&fi->page_waitq);
3249         fi->writepages = RB_ROOT;
3250
3251         if (IS_ENABLED(CONFIG_FUSE_DAX))
3252                 fuse_dax_inode_init(inode, flags);
3253 }