fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/sched/signal.h>
  16 #include <linux/module.h>
  17 #include <linux/swap.h>
  18 #include <linux/falloc.h>
  19 #include <linux/uio.h>
  20 #include <linux/fs.h>
  21 #include <linux/filelock.h>
  22
  23 static int fuse_send_open(struct fuse_mount *fm, u64 nodeid,
  24                           unsigned int open_flags, int opcode,
  25                           struct fuse_open_out *outargp)
  26 {
  27         struct fuse_open_in inarg;
  28         FUSE_ARGS(args);
  29
  30         memset(&inarg, 0, sizeof(inarg));
  31         inarg.flags = open_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  32         if (!fm->fc->atomic_o_trunc)
  33                 inarg.flags &= ~O_TRUNC;
  34
  35         if (fm->fc->handle_killpriv_v2 &&
  36             (inarg.flags & O_TRUNC) && !capable(CAP_FSETID)) {
  37                 inarg.open_flags |= FUSE_OPEN_KILL_SUIDGID;
  38         }
  39
  40         args.opcode = opcode;
  41         args.nodeid = nodeid;
  42         args.in_numargs = 1;
  43         args.in_args[0].size = sizeof(inarg);
  44         args.in_args[0].value = &inarg;
  45         args.out_numargs = 1;
  46         args.out_args[0].size = sizeof(*outargp);
  47         args.out_args[0].value = outargp;
  48
  49         return fuse_simple_request(fm, &args);
  50 }
  51
  52 struct fuse_release_args {
  53         struct fuse_args args;
  54         struct fuse_release_in inarg;
  55         struct inode *inode;
  56 };
  57
  58 struct fuse_file *fuse_file_alloc(struct fuse_mount *fm)
  59 {
  60         struct fuse_file *ff;
  61
  62         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
  63         if (unlikely(!ff))
  64                 return NULL;
  65
  66         ff->fm = fm;
  67         ff->release_args = kzalloc(sizeof(*ff->release_args),
  68                                    GFP_KERNEL_ACCOUNT);
  69         if (!ff->release_args) {
  70                 kfree(ff);
  71                 return NULL;
  72         }
  73
  74         INIT_LIST_HEAD(&ff->write_entry);
  75         mutex_init(&ff->readdir.lock);
  76         refcount_set(&ff->count, 1);
  77         RB_CLEAR_NODE(&ff->polled_node);
  78         init_waitqueue_head(&ff->poll_wait);
  79
  80         ff->kh = atomic64_inc_return(&fm->fc->khctr);
  81
  82         return ff;
  83 }
  84
  85 void fuse_file_free(struct fuse_file *ff)
  86 {
  87         kfree(ff->release_args);
  88         mutex_destroy(&ff->readdir.lock);
  89         kfree(ff);
  90 }
  91
  92 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  93 {
  94         refcount_inc(&ff->count);
  95         return ff;
  96 }
  97
  98 static void fuse_release_end(struct fuse_mount *fm, struct fuse_args *args,
  99                              int error)
 100 {
 101         struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
 102
 103         iput(ra->inode);
 104         kfree(ra);
 105 }
 106
 107 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
 108 {
 109         if (refcount_dec_and_test(&ff->count)) {
 110                 struct fuse_args *args = &ff->release_args->args;
 111
 112                 if (isdir ? ff->fm->fc->no_opendir : ff->fm->fc->no_open) {
 113                         /* Do nothing when client does not implement 'open' */
 114                         fuse_release_end(ff->fm, args, 0);
 115                 } else if (sync) {
 116                         fuse_simple_request(ff->fm, args);
 117                         fuse_release_end(ff->fm, args, 0);
 118                 } else {
 119                         args->end = fuse_release_end;
 120                         if (fuse_simple_background(ff->fm, args,
 121                                                    GFP_KERNEL | __GFP_NOFAIL))
 122                                 fuse_release_end(ff->fm, args, -ENOTCONN);
 123                 }
 124                 kfree(ff);
 125         }
 126 }
 127
 128 struct fuse_file *fuse_file_open(struct fuse_mount *fm, u64 nodeid,
 129                                  unsigned int open_flags, bool isdir)
 130 {
 131         struct fuse_conn *fc = fm->fc;
 132         struct fuse_file *ff;
 133         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 134
 135         ff = fuse_file_alloc(fm);
 136         if (!ff)
 137                 return ERR_PTR(-ENOMEM);
 138
 139         ff->fh = 0;
 140         /* Default for no-open */
 141         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
 142         if (isdir ? !fc->no_opendir : !fc->no_open) {
 143                 struct fuse_open_out outarg;
 144                 int err;
 145
 146                 err = fuse_send_open(fm, nodeid, open_flags, opcode, &outarg);
 147                 if (!err) {
 148                         ff->fh = outarg.fh;
 149                         ff->open_flags = outarg.open_flags;
 150
 151                 } else if (err != -ENOSYS) {
 152                         fuse_file_free(ff);
 153                         return ERR_PTR(err);
 154                 } else {
 155                         if (isdir)
 156                                 fc->no_opendir = 1;
 157                         else
 158                                 fc->no_open = 1;
 159                 }
 160         }
 161
 162         if (isdir)
 163                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 164
 165         ff->nodeid = nodeid;
 166
 167         return ff;
 168 }
 169
 170 int fuse_do_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
 171                  bool isdir)
 172 {
 173         struct fuse_file *ff = fuse_file_open(fm, nodeid, file->f_flags, isdir);
 174
 175         if (!IS_ERR(ff))
 176                 file->private_data = ff;
 177
 178         return PTR_ERR_OR_ZERO(ff);
 179 }
 180 EXPORT_SYMBOL_GPL(fuse_do_open);
 181
 182 static void fuse_link_write_file(struct file *file)
 183 {
 184         struct inode *inode = file_inode(file);
 185         struct fuse_inode *fi = get_fuse_inode(inode);
 186         struct fuse_file *ff = file->private_data;
 187         /*
 188          * file may be written through mmap, so chain it onto the
 189          * inodes's write_file list
 190          */
 191         spin_lock(&fi->lock);
 192         if (list_empty(&ff->write_entry))
 193                 list_add(&ff->write_entry, &fi->write_files);
 194         spin_unlock(&fi->lock);
 195 }
 196
 197 void fuse_finish_open(struct inode *inode, struct file *file)
 198 {
 199         struct fuse_file *ff = file->private_data;
 200         struct fuse_conn *fc = get_fuse_conn(inode);
 201
 202         if (ff->open_flags & FOPEN_STREAM)
 203                 stream_open(inode, file);
 204         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 205                 nonseekable_open(inode, file);
 206
 207         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 208                 struct fuse_inode *fi = get_fuse_inode(inode);
 209
 210                 spin_lock(&fi->lock);
 211                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 212                 i_size_write(inode, 0);
 213                 spin_unlock(&fi->lock);
 214                 file_update_time(file);
 215                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
 216         }
 217         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 218                 fuse_link_write_file(file);
 219 }
 220
 221 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 222 {
 223         struct fuse_mount *fm = get_fuse_mount(inode);
 224         struct fuse_conn *fc = fm->fc;
 225         int err;
 226         bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
 227                           fc->atomic_o_trunc &&
 228                           fc->writeback_cache;
 229         bool dax_truncate = (file->f_flags & O_TRUNC) &&
 230                           fc->atomic_o_trunc && FUSE_IS_DAX(inode);
 231
 232         if (fuse_is_bad(inode))
 233                 return -EIO;
 234
 235         err = generic_file_open(inode, file);
 236         if (err)
 237                 return err;
 238
 239         if (is_wb_truncate || dax_truncate)
 240                 inode_lock(inode);
 241
 242         if (dax_truncate) {
 243                 filemap_invalidate_lock(inode->i_mapping);
 244                 err = fuse_dax_break_layouts(inode, 0, 0);
 245                 if (err)
 246                         goto out_inode_unlock;
 247         }
 248
 249         if (is_wb_truncate || dax_truncate)
 250                 fuse_set_nowrite(inode);
 251
 252         err = fuse_do_open(fm, get_node_id(inode), file, isdir);
 253         if (!err)
 254                 fuse_finish_open(inode, file);
 255
 256         if (is_wb_truncate || dax_truncate)
 257                 fuse_release_nowrite(inode);
 258         if (!err) {
 259                 struct fuse_file *ff = file->private_data;
 260
 261                 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC))
 262                         truncate_pagecache(inode, 0);
 263                 else if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 264                         invalidate_inode_pages2(inode->i_mapping);
 265         }
 266         if (dax_truncate)
 267                 filemap_invalidate_unlock(inode->i_mapping);
 268 out_inode_unlock:
 269         if (is_wb_truncate || dax_truncate)
 270                 inode_unlock(inode);
 271
 272         return err;
 273 }
 274
 275 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 276                                  unsigned int flags, int opcode)
 277 {
 278         struct fuse_conn *fc = ff->fm->fc;
 279         struct fuse_release_args *ra = ff->release_args;
 280
 281         /* Inode is NULL on error path of fuse_create_open() */
 282         if (likely(fi)) {
 283                 spin_lock(&fi->lock);
 284                 list_del(&ff->write_entry);
 285                 spin_unlock(&fi->lock);
 286         }
 287         spin_lock(&fc->lock);
 288         if (!RB_EMPTY_NODE(&ff->polled_node))
 289                 rb_erase(&ff->polled_node, &fc->polled_files);
 290         spin_unlock(&fc->lock);
 291
 292         wake_up_interruptible_all(&ff->poll_wait);
 293
 294         ra->inarg.fh = ff->fh;
 295         ra->inarg.flags = flags;
 296         ra->args.in_numargs = 1;
 297         ra->args.in_args[0].size = sizeof(struct fuse_release_in);
 298         ra->args.in_args[0].value = &ra->inarg;
 299         ra->args.opcode = opcode;
 300         ra->args.nodeid = ff->nodeid;
 301         ra->args.force = true;
 302         ra->args.nocreds = true;
 303 }
 304
 305 void fuse_file_release(struct inode *inode, struct fuse_file *ff,
 306                        unsigned int open_flags, fl_owner_t id, bool isdir)
 307 {
 308         struct fuse_inode *fi = get_fuse_inode(inode);
 309         struct fuse_release_args *ra = ff->release_args;
 310         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 311
 312         fuse_prepare_release(fi, ff, open_flags, opcode);
 313
 314         if (ff->flock) {
 315                 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 316                 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fm->fc, id);
 317         }
 318         /* Hold inode until release is finished */
 319         ra->inode = igrab(inode);
 320
 321         /*
 322          * Normally this will send the RELEASE request, however if
 323          * some asynchronous READ or WRITE requests are outstanding,
 324          * the sending will be delayed.
 325          *
 326          * Make the release synchronous if this is a fuseblk mount,
 327          * synchronous RELEASE is allowed (and desirable) in this case
 328          * because the server can be trusted not to screw up.
 329          */
 330         fuse_file_put(ff, ff->fm->fc->destroy, isdir);
 331 }
 332
 333 void fuse_release_common(struct file *file, bool isdir)
 334 {
 335         fuse_file_release(file_inode(file), file->private_data, file->f_flags,
 336                           (fl_owner_t) file, isdir);
 337 }
 338
 339 static int fuse_open(struct inode *inode, struct file *file)
 340 {
 341         return fuse_open_common(inode, file, false);
 342 }
 343
 344 static int fuse_release(struct inode *inode, struct file *file)
 345 {
 346         struct fuse_conn *fc = get_fuse_conn(inode);
 347
 348         /*
 349          * Dirty pages might remain despite write_inode_now() call from
 350          * fuse_flush() due to writes racing with the close.
 351          */
 352         if (fc->writeback_cache)
 353                 write_inode_now(inode, 1);
 354
 355         fuse_release_common(file, false);
 356
 357         /* return value is ignored by VFS */
 358         return 0;
 359 }
 360
 361 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff,
 362                        unsigned int flags)
 363 {
 364         WARN_ON(refcount_read(&ff->count) > 1);
 365         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 366         /*
 367          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 368          * synchronous, we are fine with not doing igrab() here"
 369          */
 370         fuse_file_put(ff, true, false);
 371 }
 372 EXPORT_SYMBOL_GPL(fuse_sync_release);
 373
 374 /*
 375  * Scramble the ID space with XTEA, so that the value of the files_struct
 376  * pointer is not exposed to userspace.
 377  */
 378 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 379 {
 380         u32 *k = fc->scramble_key;
 381         u64 v = (unsigned long) id;
 382         u32 v0 = v;
 383         u32 v1 = v >> 32;
 384         u32 sum = 0;
 385         int i;
 386
 387         for (i = 0; i < 32; i++) {
 388                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 389                 sum += 0x9E3779B9;
 390                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 391         }
 392
 393         return (u64) v0 + ((u64) v1 << 32);
 394 }
 395
 396 struct fuse_writepage_args {
 397         struct fuse_io_args ia;
 398         struct rb_node writepages_entry;
 399         struct list_head queue_entry;
 400         struct fuse_writepage_args *next;
 401         struct inode *inode;
 402         struct fuse_sync_bucket *bucket;
 403 };
 404
 405 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
 406                                             pgoff_t idx_from, pgoff_t idx_to)
 407 {
 408         struct rb_node *n;
 409
 410         n = fi->writepages.rb_node;
 411
 412         while (n) {
 413                 struct fuse_writepage_args *wpa;
 414                 pgoff_t curr_index;
 415
 416                 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
 417                 WARN_ON(get_fuse_inode(wpa->inode) != fi);
 418                 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
 419                 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
 420                         n = n->rb_right;
 421                 else if (idx_to < curr_index)
 422                         n = n->rb_left;
 423                 else
 424                         return wpa;
 425         }
 426         return NULL;
 427 }
 428
 429 /*
 430  * Check if any page in a range is under writeback
 431  *
 432  * This is currently done by walking the list of writepage requests
 433  * for the inode, which can be pretty inefficient.
 434  */
 435 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 436                                    pgoff_t idx_to)
 437 {
 438         struct fuse_inode *fi = get_fuse_inode(inode);
 439         bool found;
 440
 441         spin_lock(&fi->lock);
 442         found = fuse_find_writeback(fi, idx_from, idx_to);
 443         spin_unlock(&fi->lock);
 444
 445         return found;
 446 }
 447
 448 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 449 {
 450         return fuse_range_is_writeback(inode, index, index);
 451 }
 452
 453 /*
 454  * Wait for page writeback to be completed.
 455  *
 456  * Since fuse doesn't rely on the VM writeback tracking, this has to
 457  * use some other means.
 458  */
 459 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 460 {
 461         struct fuse_inode *fi = get_fuse_inode(inode);
 462
 463         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 464 }
 465
 466 /*
 467  * Wait for all pending writepages on the inode to finish.
 468  *
 469  * This is currently done by blocking further writes with FUSE_NOWRITE
 470  * and waiting for all sent writes to complete.
 471  *
 472  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 473  * could conflict with truncation.
 474  */
 475 static void fuse_sync_writes(struct inode *inode)
 476 {
 477         fuse_set_nowrite(inode);
 478         fuse_release_nowrite(inode);
 479 }
 480
 481 static int fuse_flush(struct file *file, fl_owner_t id)
 482 {
 483         struct inode *inode = file_inode(file);
 484         struct fuse_mount *fm = get_fuse_mount(inode);
 485         struct fuse_file *ff = file->private_data;
 486         struct fuse_flush_in inarg;
 487         FUSE_ARGS(args);
 488         int err;
 489
 490         if (fuse_is_bad(inode))
 491                 return -EIO;
 492
 493         if (ff->open_flags & FOPEN_NOFLUSH && !fm->fc->writeback_cache)
 494                 return 0;
 495
 496         err = write_inode_now(inode, 1);
 497         if (err)
 498                 return err;
 499
 500         inode_lock(inode);
 501         fuse_sync_writes(inode);
 502         inode_unlock(inode);
 503
 504         err = filemap_check_errors(file->f_mapping);
 505         if (err)
 506                 return err;
 507
 508         err = 0;
 509         if (fm->fc->no_flush)
 510                 goto inval_attr_out;
 511
 512         memset(&inarg, 0, sizeof(inarg));
 513         inarg.fh = ff->fh;
 514         inarg.lock_owner = fuse_lock_owner_id(fm->fc, id);
 515         args.opcode = FUSE_FLUSH;
 516         args.nodeid = get_node_id(inode);
 517         args.in_numargs = 1;
 518         args.in_args[0].size = sizeof(inarg);
 519         args.in_args[0].value = &inarg;
 520         args.force = true;
 521
 522         err = fuse_simple_request(fm, &args);
 523         if (err == -ENOSYS) {
 524                 fm->fc->no_flush = 1;
 525                 err = 0;
 526         }
 527
 528 inval_attr_out:
 529         /*
 530          * In memory i_blocks is not maintained by fuse, if writeback cache is
 531          * enabled, i_blocks from cached attr may not be accurate.
 532          */
 533         if (!err && fm->fc->writeback_cache)
 534                 fuse_invalidate_attr_mask(inode, STATX_BLOCKS);
 535         return err;
 536 }
 537
 538 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 539                       int datasync, int opcode)
 540 {
 541         struct inode *inode = file->f_mapping->host;
 542         struct fuse_mount *fm = get_fuse_mount(inode);
 543         struct fuse_file *ff = file->private_data;
 544         FUSE_ARGS(args);
 545         struct fuse_fsync_in inarg;
 546
 547         memset(&inarg, 0, sizeof(inarg));
 548         inarg.fh = ff->fh;
 549         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 550         args.opcode = opcode;
 551         args.nodeid = get_node_id(inode);
 552         args.in_numargs = 1;
 553         args.in_args[0].size = sizeof(inarg);
 554         args.in_args[0].value = &inarg;
 555         return fuse_simple_request(fm, &args);
 556 }
 557
 558 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 559                       int datasync)
 560 {
 561         struct inode *inode = file->f_mapping->host;
 562         struct fuse_conn *fc = get_fuse_conn(inode);
 563         int err;
 564
 565         if (fuse_is_bad(inode))
 566                 return -EIO;
 567
 568         inode_lock(inode);
 569
 570         /*
 571          * Start writeback against all dirty pages of the inode, then
 572          * wait for all outstanding writes, before sending the FSYNC
 573          * request.
 574          */
 575         err = file_write_and_wait_range(file, start, end);
 576         if (err)
 577                 goto out;
 578
 579         fuse_sync_writes(inode);
 580
 581         /*
 582          * Due to implementation of fuse writeback
 583          * file_write_and_wait_range() does not catch errors.
 584          * We have to do this directly after fuse_sync_writes()
 585          */
 586         err = file_check_and_advance_wb_err(file);
 587         if (err)
 588                 goto out;
 589
 590         err = sync_inode_metadata(inode, 1);
 591         if (err)
 592                 goto out;
 593
 594         if (fc->no_fsync)
 595                 goto out;
 596
 597         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 598         if (err == -ENOSYS) {
 599                 fc->no_fsync = 1;
 600                 err = 0;
 601         }
 602 out:
 603         inode_unlock(inode);
 604
 605         return err;
 606 }
 607
 608 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
 609                          size_t count, int opcode)
 610 {
 611         struct fuse_file *ff = file->private_data;
 612         struct fuse_args *args = &ia->ap.args;
 613
 614         ia->read.in.fh = ff->fh;
 615         ia->read.in.offset = pos;
 616         ia->read.in.size = count;
 617         ia->read.in.flags = file->f_flags;
 618         args->opcode = opcode;
 619         args->nodeid = ff->nodeid;
 620         args->in_numargs = 1;
 621         args->in_args[0].size = sizeof(ia->read.in);
 622         args->in_args[0].value = &ia->read.in;
 623         args->out_argvar = true;
 624         args->out_numargs = 1;
 625         args->out_args[0].size = count;
 626 }
 627
 628 static void fuse_release_user_pages(struct fuse_args_pages *ap,
 629                                     bool should_dirty)
 630 {
 631         unsigned int i;
 632
 633         for (i = 0; i < ap->num_pages; i++) {
 634                 if (should_dirty)
 635                         set_page_dirty_lock(ap->pages[i]);
 636                 put_page(ap->pages[i]);
 637         }
 638 }
 639
 640 static void fuse_io_release(struct kref *kref)
 641 {
 642         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 643 }
 644
 645 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 646 {
 647         if (io->err)
 648                 return io->err;
 649
 650         if (io->bytes >= 0 && io->write)
 651                 return -EIO;
 652
 653         return io->bytes < 0 ? io->size : io->bytes;
 654 }
 655
 656 /*
 657  * In case of short read, the caller sets 'pos' to the position of
 658  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 659  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 660  *
 661  * An example:
 662  * User requested DIO read of 64K. It was split into two 32K fuse requests,
 663  * both submitted asynchronously. The first of them was ACKed by userspace as
 664  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 665  * second request was ACKed as short, e.g. only 1K was read, resulting in
 666  * pos == 33K.
 667  *
 668  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 669  * will be equal to the length of the longest contiguous fragment of
 670  * transferred data starting from the beginning of IO request.
 671  */
 672 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 673 {
 674         int left;
 675
 676         spin_lock(&io->lock);
 677         if (err)
 678                 io->err = io->err ? : err;
 679         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 680                 io->bytes = pos;
 681
 682         left = --io->reqs;
 683         if (!left && io->blocking)
 684                 complete(io->done);
 685         spin_unlock(&io->lock);
 686
 687         if (!left && !io->blocking) {
 688                 ssize_t res = fuse_get_res_by_io(io);
 689
 690                 if (res >= 0) {
 691                         struct inode *inode = file_inode(io->iocb->ki_filp);
 692                         struct fuse_conn *fc = get_fuse_conn(inode);
 693                         struct fuse_inode *fi = get_fuse_inode(inode);
 694
 695                         spin_lock(&fi->lock);
 696                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 697                         spin_unlock(&fi->lock);
 698                 }
 699
 700                 io->iocb->ki_complete(io->iocb, res);
 701         }
 702
 703         kref_put(&io->refcnt, fuse_io_release);
 704 }
 705
 706 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
 707                                           unsigned int npages)
 708 {
 709         struct fuse_io_args *ia;
 710
 711         ia = kzalloc(sizeof(*ia), GFP_KERNEL);
 712         if (ia) {
 713                 ia->io = io;
 714                 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
 715                                                 &ia->ap.descs);
 716                 if (!ia->ap.pages) {
 717                         kfree(ia);
 718                         ia = NULL;
 719                 }
 720         }
 721         return ia;
 722 }
 723
 724 static void fuse_io_free(struct fuse_io_args *ia)
 725 {
 726         kfree(ia->ap.pages);
 727         kfree(ia);
 728 }
 729
 730 static void fuse_aio_complete_req(struct fuse_mount *fm, struct fuse_args *args,
 731                                   int err)
 732 {
 733         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 734         struct fuse_io_priv *io = ia->io;
 735         ssize_t pos = -1;
 736
 737         fuse_release_user_pages(&ia->ap, io->should_dirty);
 738
 739         if (err) {
 740                 /* Nothing */
 741         } else if (io->write) {
 742                 if (ia->write.out.size > ia->write.in.size) {
 743                         err = -EIO;
 744                 } else if (ia->write.in.size != ia->write.out.size) {
 745                         pos = ia->write.in.offset - io->offset +
 746                                 ia->write.out.size;
 747                 }
 748         } else {
 749                 u32 outsize = args->out_args[0].size;
 750
 751                 if (ia->read.in.size != outsize)
 752                         pos = ia->read.in.offset - io->offset + outsize;
 753         }
 754
 755         fuse_aio_complete(io, err, pos);
 756         fuse_io_free(ia);
 757 }
 758
 759 static ssize_t fuse_async_req_send(struct fuse_mount *fm,
 760                                    struct fuse_io_args *ia, size_t num_bytes)
 761 {
 762         ssize_t err;
 763         struct fuse_io_priv *io = ia->io;
 764
 765         spin_lock(&io->lock);
 766         kref_get(&io->refcnt);
 767         io->size += num_bytes;
 768         io->reqs++;
 769         spin_unlock(&io->lock);
 770
 771         ia->ap.args.end = fuse_aio_complete_req;
 772         ia->ap.args.may_block = io->should_dirty;
 773         err = fuse_simple_background(fm, &ia->ap.args, GFP_KERNEL);
 774         if (err)
 775                 fuse_aio_complete_req(fm, &ia->ap.args, err);
 776
 777         return num_bytes;
 778 }
 779
 780 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
 781                               fl_owner_t owner)
 782 {
 783         struct file *file = ia->io->iocb->ki_filp;
 784         struct fuse_file *ff = file->private_data;
 785         struct fuse_mount *fm = ff->fm;
 786
 787         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 788         if (owner != NULL) {
 789                 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
 790                 ia->read.in.lock_owner = fuse_lock_owner_id(fm->fc, owner);
 791         }
 792
 793         if (ia->io->async)
 794                 return fuse_async_req_send(fm, ia, count);
 795
 796         return fuse_simple_request(fm, &ia->ap.args);
 797 }
 798
 799 static void fuse_read_update_size(struct inode *inode, loff_t size,
 800                                   u64 attr_ver)
 801 {
 802         struct fuse_conn *fc = get_fuse_conn(inode);
 803         struct fuse_inode *fi = get_fuse_inode(inode);
 804
 805         spin_lock(&fi->lock);
 806         if (attr_ver >= fi->attr_version && size < inode->i_size &&
 807             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 808                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 809                 i_size_write(inode, size);
 810         }
 811         spin_unlock(&fi->lock);
 812 }
 813
 814 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
 815                             struct fuse_args_pages *ap)
 816 {
 817         struct fuse_conn *fc = get_fuse_conn(inode);
 818
 819         /*
 820          * If writeback_cache is enabled, a short read means there's a hole in
 821          * the file.  Some data after the hole is in page cache, but has not
 822          * reached the client fs yet.  So the hole is not present there.
 823          */
 824         if (!fc->writeback_cache) {
 825                 loff_t pos = page_offset(ap->pages[0]) + num_read;
 826                 fuse_read_update_size(inode, pos, attr_ver);
 827         }
 828 }
 829
 830 static int fuse_do_readpage(struct file *file, struct page *page)
 831 {
 832         struct inode *inode = page->mapping->host;
 833         struct fuse_mount *fm = get_fuse_mount(inode);
 834         loff_t pos = page_offset(page);
 835         struct fuse_page_desc desc = { .length = PAGE_SIZE };
 836         struct fuse_io_args ia = {
 837                 .ap.args.page_zeroing = true,
 838                 .ap.args.out_pages = true,
 839                 .ap.num_pages = 1,
 840                 .ap.pages = &page,
 841                 .ap.descs = &desc,
 842         };
 843         ssize_t res;
 844         u64 attr_ver;
 845
 846         /*
 847          * Page writeback can extend beyond the lifetime of the
 848          * page-cache page, so make sure we read a properly synced
 849          * page.
 850          */
 851         fuse_wait_on_page_writeback(inode, page->index);
 852
 853         attr_ver = fuse_get_attr_version(fm->fc);
 854
 855         /* Don't overflow end offset */
 856         if (pos + (desc.length - 1) == LLONG_MAX)
 857                 desc.length--;
 858
 859         fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
 860         res = fuse_simple_request(fm, &ia.ap.args);
 861         if (res < 0)
 862                 return res;
 863         /*
 864          * Short read means EOF.  If file size is larger, truncate it
 865          */
 866         if (res < desc.length)
 867                 fuse_short_read(inode, attr_ver, res, &ia.ap);
 868
 869         SetPageUptodate(page);
 870
 871         return 0;
 872 }
 873
 874 static int fuse_read_folio(struct file *file, struct folio *folio)
 875 {
 876         struct page *page = &folio->page;
 877         struct inode *inode = page->mapping->host;
 878         int err;
 879
 880         err = -EIO;
 881         if (fuse_is_bad(inode))
 882                 goto out;
 883
 884         err = fuse_do_readpage(file, page);
 885         fuse_invalidate_atime(inode);
 886  out:
 887         unlock_page(page);
 888         return err;
 889 }
 890
 891 static void fuse_readpages_end(struct fuse_mount *fm, struct fuse_args *args,
 892                                int err)
 893 {
 894         int i;
 895         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 896         struct fuse_args_pages *ap = &ia->ap;
 897         size_t count = ia->read.in.size;
 898         size_t num_read = args->out_args[0].size;
 899         struct address_space *mapping = NULL;
 900
 901         for (i = 0; mapping == NULL && i < ap->num_pages; i++)
 902                 mapping = ap->pages[i]->mapping;
 903
 904         if (mapping) {
 905                 struct inode *inode = mapping->host;
 906
 907                 /*
 908                  * Short read means EOF. If file size is larger, truncate it
 909                  */
 910                 if (!err && num_read < count)
 911                         fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
 912
 913                 fuse_invalidate_atime(inode);
 914         }
 915
 916         for (i = 0; i < ap->num_pages; i++) {
 917                 struct page *page = ap->pages[i];
 918
 919                 if (!err)
 920                         SetPageUptodate(page);
 921                 else
 922                         SetPageError(page);
 923                 unlock_page(page);
 924                 put_page(page);
 925         }
 926         if (ia->ff)
 927                 fuse_file_put(ia->ff, false, false);
 928
 929         fuse_io_free(ia);
 930 }
 931
 932 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
 933 {
 934         struct fuse_file *ff = file->private_data;
 935         struct fuse_mount *fm = ff->fm;
 936         struct fuse_args_pages *ap = &ia->ap;
 937         loff_t pos = page_offset(ap->pages[0]);
 938         size_t count = ap->num_pages << PAGE_SHIFT;
 939         ssize_t res;
 940         int err;
 941
 942         ap->args.out_pages = true;
 943         ap->args.page_zeroing = true;
 944         ap->args.page_replace = true;
 945
 946         /* Don't overflow end offset */
 947         if (pos + (count - 1) == LLONG_MAX) {
 948                 count--;
 949                 ap->descs[ap->num_pages - 1].length--;
 950         }
 951         WARN_ON((loff_t) (pos + count) < 0);
 952
 953         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 954         ia->read.attr_ver = fuse_get_attr_version(fm->fc);
 955         if (fm->fc->async_read) {
 956                 ia->ff = fuse_file_get(ff);
 957                 ap->args.end = fuse_readpages_end;
 958                 err = fuse_simple_background(fm, &ap->args, GFP_KERNEL);
 959                 if (!err)
 960                         return;
 961         } else {
 962                 res = fuse_simple_request(fm, &ap->args);
 963                 err = res < 0 ? res : 0;
 964         }
 965         fuse_readpages_end(fm, &ap->args, err);
 966 }
 967
 968 static void fuse_readahead(struct readahead_control *rac)
 969 {
 970         struct inode *inode = rac->mapping->host;
 971         struct fuse_conn *fc = get_fuse_conn(inode);
 972         unsigned int i, max_pages, nr_pages = 0;
 973
 974         if (fuse_is_bad(inode))
 975                 return;
 976
 977         max_pages = min_t(unsigned int, fc->max_pages,
 978                         fc->max_read / PAGE_SIZE);
 979
 980         for (;;) {
 981                 struct fuse_io_args *ia;
 982                 struct fuse_args_pages *ap;
 983
 984                 if (fc->num_background >= fc->congestion_threshold &&
 985                     rac->ra->async_size >= readahead_count(rac))
 986                         /*
 987                          * Congested and only async pages left, so skip the
 988                          * rest.
 989                          */
 990                         break;
 991
 992                 nr_pages = readahead_count(rac) - nr_pages;
 993                 if (nr_pages > max_pages)
 994                         nr_pages = max_pages;
 995                 if (nr_pages == 0)
 996                         break;
 997                 ia = fuse_io_alloc(NULL, nr_pages);
 998                 if (!ia)
 999                         return;
1000                 ap = &ia->ap;
1001                 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
1002                 for (i = 0; i < nr_pages; i++) {
1003                         fuse_wait_on_page_writeback(inode,
1004                                                     readahead_index(rac) + i);
1005                         ap->descs[i].length = PAGE_SIZE;
1006                 }
1007                 ap->num_pages = nr_pages;
1008                 fuse_send_readpages(ia, rac->file);
1009         }
1010 }
1011
1012 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
1013 {
1014         struct inode *inode = iocb->ki_filp->f_mapping->host;
1015         struct fuse_conn *fc = get_fuse_conn(inode);
1016
1017         /*
1018          * In auto invalidate mode, always update attributes on read.
1019          * Otherwise, only update if we attempt to read past EOF (to ensure
1020          * i_size is up to date).
1021          */
1022         if (fc->auto_inval_data ||
1023             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
1024                 int err;
1025                 err = fuse_update_attributes(inode, iocb->ki_filp, STATX_SIZE);
1026                 if (err)
1027                         return err;
1028         }
1029
1030         return generic_file_read_iter(iocb, to);
1031 }
1032
1033 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1034                                  loff_t pos, size_t count)
1035 {
1036         struct fuse_args *args = &ia->ap.args;
1037
1038         ia->write.in.fh = ff->fh;
1039         ia->write.in.offset = pos;
1040         ia->write.in.size = count;
1041         args->opcode = FUSE_WRITE;
1042         args->nodeid = ff->nodeid;
1043         args->in_numargs = 2;
1044         if (ff->fm->fc->minor < 9)
1045                 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1046         else
1047                 args->in_args[0].size = sizeof(ia->write.in);
1048         args->in_args[0].value = &ia->write.in;
1049         args->in_args[1].size = count;
1050         args->out_numargs = 1;
1051         args->out_args[0].size = sizeof(ia->write.out);
1052         args->out_args[0].value = &ia->write.out;
1053 }
1054
1055 static unsigned int fuse_write_flags(struct kiocb *iocb)
1056 {
1057         unsigned int flags = iocb->ki_filp->f_flags;
1058
1059         if (iocb_is_dsync(iocb))
1060                 flags |= O_DSYNC;
1061         if (iocb->ki_flags & IOCB_SYNC)
1062                 flags |= O_SYNC;
1063
1064         return flags;
1065 }
1066
1067 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1068                                size_t count, fl_owner_t owner)
1069 {
1070         struct kiocb *iocb = ia->io->iocb;
1071         struct file *file = iocb->ki_filp;
1072         struct fuse_file *ff = file->private_data;
1073         struct fuse_mount *fm = ff->fm;
1074         struct fuse_write_in *inarg = &ia->write.in;
1075         ssize_t err;
1076
1077         fuse_write_args_fill(ia, ff, pos, count);
1078         inarg->flags = fuse_write_flags(iocb);
1079         if (owner != NULL) {
1080                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1081                 inarg->lock_owner = fuse_lock_owner_id(fm->fc, owner);
1082         }
1083
1084         if (ia->io->async)
1085                 return fuse_async_req_send(fm, ia, count);
1086
1087         err = fuse_simple_request(fm, &ia->ap.args);
1088         if (!err && ia->write.out.size > count)
1089                 err = -EIO;
1090
1091         return err ?: ia->write.out.size;
1092 }
1093
1094 bool fuse_write_update_attr(struct inode *inode, loff_t pos, ssize_t written)
1095 {
1096         struct fuse_conn *fc = get_fuse_conn(inode);
1097         struct fuse_inode *fi = get_fuse_inode(inode);
1098         bool ret = false;
1099
1100         spin_lock(&fi->lock);
1101         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1102         if (written > 0 && pos > inode->i_size) {
1103                 i_size_write(inode, pos);
1104                 ret = true;
1105         }
1106         spin_unlock(&fi->lock);
1107
1108         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
1109
1110         return ret;
1111 }
1112
1113 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1114                                      struct kiocb *iocb, struct inode *inode,
1115                                      loff_t pos, size_t count)
1116 {
1117         struct fuse_args_pages *ap = &ia->ap;
1118         struct file *file = iocb->ki_filp;
1119         struct fuse_file *ff = file->private_data;
1120         struct fuse_mount *fm = ff->fm;
1121         unsigned int offset, i;
1122         bool short_write;
1123         int err;
1124
1125         for (i = 0; i < ap->num_pages; i++)
1126                 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1127
1128         fuse_write_args_fill(ia, ff, pos, count);
1129         ia->write.in.flags = fuse_write_flags(iocb);
1130         if (fm->fc->handle_killpriv_v2 && !capable(CAP_FSETID))
1131                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1132
1133         err = fuse_simple_request(fm, &ap->args);
1134         if (!err && ia->write.out.size > count)
1135                 err = -EIO;
1136
1137         short_write = ia->write.out.size < count;
1138         offset = ap->descs[0].offset;
1139         count = ia->write.out.size;
1140         for (i = 0; i < ap->num_pages; i++) {
1141                 struct page *page = ap->pages[i];
1142
1143                 if (err) {
1144                         ClearPageUptodate(page);
1145                 } else {
1146                         if (count >= PAGE_SIZE - offset)
1147                                 count -= PAGE_SIZE - offset;
1148                         else {
1149                                 if (short_write)
1150                                         ClearPageUptodate(page);
1151                                 count = 0;
1152                         }
1153                         offset = 0;
1154                 }
1155                 if (ia->write.page_locked && (i == ap->num_pages - 1))
1156                         unlock_page(page);
1157                 put_page(page);
1158         }
1159
1160         return err;
1161 }
1162
1163 static ssize_t fuse_fill_write_pages(struct fuse_io_args *ia,
1164                                      struct address_space *mapping,
1165                                      struct iov_iter *ii, loff_t pos,
1166                                      unsigned int max_pages)
1167 {
1168         struct fuse_args_pages *ap = &ia->ap;
1169         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1170         unsigned offset = pos & (PAGE_SIZE - 1);
1171         size_t count = 0;
1172         int err;
1173
1174         ap->args.in_pages = true;
1175         ap->descs[0].offset = offset;
1176
1177         do {
1178                 size_t tmp;
1179                 struct page *page;
1180                 pgoff_t index = pos >> PAGE_SHIFT;
1181                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1182                                      iov_iter_count(ii));
1183
1184                 bytes = min_t(size_t, bytes, fc->max_write - count);
1185
1186  again:
1187                 err = -EFAULT;
1188                 if (fault_in_iov_iter_readable(ii, bytes))
1189                         break;
1190
1191                 err = -ENOMEM;
1192                 page = grab_cache_page_write_begin(mapping, index);
1193                 if (!page)
1194                         break;
1195
1196                 if (mapping_writably_mapped(mapping))
1197                         flush_dcache_page(page);
1198
1199                 tmp = copy_page_from_iter_atomic(page, offset, bytes, ii);
1200                 flush_dcache_page(page);
1201
1202                 if (!tmp) {
1203                         unlock_page(page);
1204                         put_page(page);
1205                         goto again;
1206                 }
1207
1208                 err = 0;
1209                 ap->pages[ap->num_pages] = page;
1210                 ap->descs[ap->num_pages].length = tmp;
1211                 ap->num_pages++;
1212
1213                 count += tmp;
1214                 pos += tmp;
1215                 offset += tmp;
1216                 if (offset == PAGE_SIZE)
1217                         offset = 0;
1218
1219                 /* If we copied full page, mark it uptodate */
1220                 if (tmp == PAGE_SIZE)
1221                         SetPageUptodate(page);
1222
1223                 if (PageUptodate(page)) {
1224                         unlock_page(page);
1225                 } else {
1226                         ia->write.page_locked = true;
1227                         break;
1228                 }
1229                 if (!fc->big_writes)
1230                         break;
1231         } while (iov_iter_count(ii) && count < fc->max_write &&
1232                  ap->num_pages < max_pages && offset == 0);
1233
1234         return count > 0 ? count : err;
1235 }
1236
1237 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1238                                      unsigned int max_pages)
1239 {
1240         return min_t(unsigned int,
1241                      ((pos + len - 1) >> PAGE_SHIFT) -
1242                      (pos >> PAGE_SHIFT) + 1,
1243                      max_pages);
1244 }
1245
1246 static ssize_t fuse_perform_write(struct kiocb *iocb, struct iov_iter *ii)
1247 {
1248         struct address_space *mapping = iocb->ki_filp->f_mapping;
1249         struct inode *inode = mapping->host;
1250         struct fuse_conn *fc = get_fuse_conn(inode);
1251         struct fuse_inode *fi = get_fuse_inode(inode);
1252         loff_t pos = iocb->ki_pos;
1253         int err = 0;
1254         ssize_t res = 0;
1255
1256         if (inode->i_size < pos + iov_iter_count(ii))
1257                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1258
1259         do {
1260                 ssize_t count;
1261                 struct fuse_io_args ia = {};
1262                 struct fuse_args_pages *ap = &ia.ap;
1263                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1264                                                       fc->max_pages);
1265
1266                 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1267                 if (!ap->pages) {
1268                         err = -ENOMEM;
1269                         break;
1270                 }
1271
1272                 count = fuse_fill_write_pages(&ia, mapping, ii, pos, nr_pages);
1273                 if (count <= 0) {
1274                         err = count;
1275                 } else {
1276                         err = fuse_send_write_pages(&ia, iocb, inode,
1277                                                     pos, count);
1278                         if (!err) {
1279                                 size_t num_written = ia.write.out.size;
1280
1281                                 res += num_written;
1282                                 pos += num_written;
1283
1284                                 /* break out of the loop on short write */
1285                                 if (num_written != count)
1286                                         err = -EIO;
1287                         }
1288                 }
1289                 kfree(ap->pages);
1290         } while (!err && iov_iter_count(ii));
1291
1292         fuse_write_update_attr(inode, pos, res);
1293         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1294
1295         if (!res)
1296                 return err;
1297         iocb->ki_pos += res;
1298         return res;
1299 }
1300
1301 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1302 {
1303         struct file *file = iocb->ki_filp;
1304         struct address_space *mapping = file->f_mapping;
1305         ssize_t written = 0;
1306         struct inode *inode = mapping->host;
1307         ssize_t err;
1308         struct fuse_conn *fc = get_fuse_conn(inode);
1309
1310         if (fc->writeback_cache) {
1311                 /* Update size (EOF optimization) and mode (SUID clearing) */
1312                 err = fuse_update_attributes(mapping->host, file,
1313                                              STATX_SIZE | STATX_MODE);
1314                 if (err)
1315                         return err;
1316
1317                 if (fc->handle_killpriv_v2 &&
1318                     setattr_should_drop_suidgid(&nop_mnt_idmap,
1319                                                 file_inode(file))) {
1320                         goto writethrough;
1321                 }
1322
1323                 return generic_file_write_iter(iocb, from);
1324         }
1325
1326 writethrough:
1327         inode_lock(inode);
1328
1329         err = generic_write_checks(iocb, from);
1330         if (err <= 0)
1331                 goto out;
1332
1333         err = file_remove_privs(file);
1334         if (err)
1335                 goto out;
1336
1337         err = file_update_time(file);
1338         if (err)
1339                 goto out;
1340
1341         if (iocb->ki_flags & IOCB_DIRECT) {
1342                 written = generic_file_direct_write(iocb, from);
1343                 if (written < 0 || !iov_iter_count(from))
1344                         goto out;
1345                 written = direct_write_fallback(iocb, from, written,
1346                                 fuse_perform_write(iocb, from));
1347         } else {
1348                 written = fuse_perform_write(iocb, from);
1349         }
1350 out:
1351         inode_unlock(inode);
1352         if (written > 0)
1353                 written = generic_write_sync(iocb, written);
1354
1355         return written ? written : err;
1356 }
1357
1358 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1359 {
1360         return (unsigned long)iter_iov(ii)->iov_base + ii->iov_offset;
1361 }
1362
1363 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1364                                         size_t max_size)
1365 {
1366         return min(iov_iter_single_seg_count(ii), max_size);
1367 }
1368
1369 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1370                                size_t *nbytesp, int write,
1371                                unsigned int max_pages)
1372 {
1373         size_t nbytes = 0;  /* # bytes already packed in req */
1374         ssize_t ret = 0;
1375
1376         /* Special case for kernel I/O: can copy directly into the buffer */
1377         if (iov_iter_is_kvec(ii)) {
1378                 unsigned long user_addr = fuse_get_user_addr(ii);
1379                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1380
1381                 if (write)
1382                         ap->args.in_args[1].value = (void *) user_addr;
1383                 else
1384                         ap->args.out_args[0].value = (void *) user_addr;
1385
1386                 iov_iter_advance(ii, frag_size);
1387                 *nbytesp = frag_size;
1388                 return 0;
1389         }
1390
1391         while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1392                 unsigned npages;
1393                 size_t start;
1394                 ret = iov_iter_get_pages2(ii, &ap->pages[ap->num_pages],
1395                                         *nbytesp - nbytes,
1396                                         max_pages - ap->num_pages,
1397                                         &start);
1398                 if (ret < 0)
1399                         break;
1400
1401                 nbytes += ret;
1402
1403                 ret += start;
1404                 npages = DIV_ROUND_UP(ret, PAGE_SIZE);
1405
1406                 ap->descs[ap->num_pages].offset = start;
1407                 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1408
1409                 ap->num_pages += npages;
1410                 ap->descs[ap->num_pages - 1].length -=
1411                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1412         }
1413
1414         ap->args.user_pages = true;
1415         if (write)
1416                 ap->args.in_pages = true;
1417         else
1418                 ap->args.out_pages = true;
1419
1420         *nbytesp = nbytes;
1421
1422         return ret < 0 ? ret : 0;
1423 }
1424
1425 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1426                        loff_t *ppos, int flags)
1427 {
1428         int write = flags & FUSE_DIO_WRITE;
1429         int cuse = flags & FUSE_DIO_CUSE;
1430         struct file *file = io->iocb->ki_filp;
1431         struct address_space *mapping = file->f_mapping;
1432         struct inode *inode = mapping->host;
1433         struct fuse_file *ff = file->private_data;
1434         struct fuse_conn *fc = ff->fm->fc;
1435         size_t nmax = write ? fc->max_write : fc->max_read;
1436         loff_t pos = *ppos;
1437         size_t count = iov_iter_count(iter);
1438         pgoff_t idx_from = pos >> PAGE_SHIFT;
1439         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1440         ssize_t res = 0;
1441         int err = 0;
1442         struct fuse_io_args *ia;
1443         unsigned int max_pages;
1444         bool fopen_direct_io = ff->open_flags & FOPEN_DIRECT_IO;
1445
1446         max_pages = iov_iter_npages(iter, fc->max_pages);
1447         ia = fuse_io_alloc(io, max_pages);
1448         if (!ia)
1449                 return -ENOMEM;
1450
1451         if (fopen_direct_io && fc->direct_io_allow_mmap) {
1452                 res = filemap_write_and_wait_range(mapping, pos, pos + count - 1);
1453                 if (res) {
1454                         fuse_io_free(ia);
1455                         return res;
1456                 }
1457         }
1458         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1459                 if (!write)
1460                         inode_lock(inode);
1461                 fuse_sync_writes(inode);
1462                 if (!write)
1463                         inode_unlock(inode);
1464         }
1465
1466         if (fopen_direct_io && write) {
1467                 res = invalidate_inode_pages2_range(mapping, idx_from, idx_to);
1468                 if (res) {
1469                         fuse_io_free(ia);
1470                         return res;
1471                 }
1472         }
1473
1474         io->should_dirty = !write && user_backed_iter(iter);
1475         while (count) {
1476                 ssize_t nres;
1477                 fl_owner_t owner = current->files;
1478                 size_t nbytes = min(count, nmax);
1479
1480                 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1481                                           max_pages);
1482                 if (err && !nbytes)
1483                         break;
1484
1485                 if (write) {
1486                         if (!capable(CAP_FSETID))
1487                                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1488
1489                         nres = fuse_send_write(ia, pos, nbytes, owner);
1490                 } else {
1491                         nres = fuse_send_read(ia, pos, nbytes, owner);
1492                 }
1493
1494                 if (!io->async || nres < 0) {
1495                         fuse_release_user_pages(&ia->ap, io->should_dirty);
1496                         fuse_io_free(ia);
1497                 }
1498                 ia = NULL;
1499                 if (nres < 0) {
1500                         iov_iter_revert(iter, nbytes);
1501                         err = nres;
1502                         break;
1503                 }
1504                 WARN_ON(nres > nbytes);
1505
1506                 count -= nres;
1507                 res += nres;
1508                 pos += nres;
1509                 if (nres != nbytes) {
1510                         iov_iter_revert(iter, nbytes - nres);
1511                         break;
1512                 }
1513                 if (count) {
1514                         max_pages = iov_iter_npages(iter, fc->max_pages);
1515                         ia = fuse_io_alloc(io, max_pages);
1516                         if (!ia)
1517                                 break;
1518                 }
1519         }
1520         if (ia)
1521                 fuse_io_free(ia);
1522         if (res > 0)
1523                 *ppos = pos;
1524
1525         return res > 0 ? res : err;
1526 }
1527 EXPORT_SYMBOL_GPL(fuse_direct_io);
1528
1529 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1530                                   struct iov_iter *iter,
1531                                   loff_t *ppos)
1532 {
1533         ssize_t res;
1534         struct inode *inode = file_inode(io->iocb->ki_filp);
1535
1536         res = fuse_direct_io(io, iter, ppos, 0);
1537
1538         fuse_invalidate_atime(inode);
1539
1540         return res;
1541 }
1542
1543 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1544
1545 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1546 {
1547         ssize_t res;
1548
1549         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1550                 res = fuse_direct_IO(iocb, to);
1551         } else {
1552                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1553
1554                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1555         }
1556
1557         return res;
1558 }
1559
1560 static bool fuse_direct_write_extending_i_size(struct kiocb *iocb,
1561                                                struct iov_iter *iter)
1562 {
1563         struct inode *inode = file_inode(iocb->ki_filp);
1564
1565         return iocb->ki_pos + iov_iter_count(iter) > i_size_read(inode);
1566 }
1567
1568 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1569 {
1570         struct inode *inode = file_inode(iocb->ki_filp);
1571         struct file *file = iocb->ki_filp;
1572         struct fuse_file *ff = file->private_data;
1573         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1574         ssize_t res;
1575         bool exclusive_lock =
1576                 !(ff->open_flags & FOPEN_PARALLEL_DIRECT_WRITES) ||
1577                 get_fuse_conn(inode)->direct_io_allow_mmap ||
1578                 iocb->ki_flags & IOCB_APPEND ||
1579                 fuse_direct_write_extending_i_size(iocb, from);
1580
1581         /*
1582          * Take exclusive lock if
1583          * - Parallel direct writes are disabled - a user space decision
1584          * - Parallel direct writes are enabled and i_size is being extended.
1585          * - Shared mmap on direct_io file is supported (FUSE_DIRECT_IO_ALLOW_MMAP).
1586          *   This might not be needed at all, but needs further investigation.
1587          */
1588         if (exclusive_lock)
1589                 inode_lock(inode);
1590         else {
1591                 inode_lock_shared(inode);
1592
1593                 /* A race with truncate might have come up as the decision for
1594                  * the lock type was done without holding the lock, check again.
1595                  */
1596                 if (fuse_direct_write_extending_i_size(iocb, from)) {
1597                         inode_unlock_shared(inode);
1598                         inode_lock(inode);
1599                         exclusive_lock = true;
1600                 }
1601         }
1602
1603         res = generic_write_checks(iocb, from);
1604         if (res > 0) {
1605                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1606                         res = fuse_direct_IO(iocb, from);
1607                 } else {
1608                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1609                                              FUSE_DIO_WRITE);
1610                         fuse_write_update_attr(inode, iocb->ki_pos, res);
1611                 }
1612         }
1613         if (exclusive_lock)
1614                 inode_unlock(inode);
1615         else
1616                 inode_unlock_shared(inode);
1617
1618         return res;
1619 }
1620
1621 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1622 {
1623         struct file *file = iocb->ki_filp;
1624         struct fuse_file *ff = file->private_data;
1625         struct inode *inode = file_inode(file);
1626
1627         if (fuse_is_bad(inode))
1628                 return -EIO;
1629
1630         if (FUSE_IS_DAX(inode))
1631                 return fuse_dax_read_iter(iocb, to);
1632
1633         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1634                 return fuse_cache_read_iter(iocb, to);
1635         else
1636                 return fuse_direct_read_iter(iocb, to);
1637 }
1638
1639 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1640 {
1641         struct file *file = iocb->ki_filp;
1642         struct fuse_file *ff = file->private_data;
1643         struct inode *inode = file_inode(file);
1644
1645         if (fuse_is_bad(inode))
1646                 return -EIO;
1647
1648         if (FUSE_IS_DAX(inode))
1649                 return fuse_dax_write_iter(iocb, from);
1650
1651         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1652                 return fuse_cache_write_iter(iocb, from);
1653         else
1654                 return fuse_direct_write_iter(iocb, from);
1655 }
1656
1657 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1658 {
1659         struct fuse_args_pages *ap = &wpa->ia.ap;
1660         int i;
1661
1662         if (wpa->bucket)
1663                 fuse_sync_bucket_dec(wpa->bucket);
1664
1665         for (i = 0; i < ap->num_pages; i++)
1666                 __free_page(ap->pages[i]);
1667
1668         if (wpa->ia.ff)
1669                 fuse_file_put(wpa->ia.ff, false, false);
1670
1671         kfree(ap->pages);
1672         kfree(wpa);
1673 }
1674
1675 static void fuse_writepage_finish(struct fuse_mount *fm,
1676                                   struct fuse_writepage_args *wpa)
1677 {
1678         struct fuse_args_pages *ap = &wpa->ia.ap;
1679         struct inode *inode = wpa->inode;
1680         struct fuse_inode *fi = get_fuse_inode(inode);
1681         struct backing_dev_info *bdi = inode_to_bdi(inode);
1682         int i;
1683
1684         for (i = 0; i < ap->num_pages; i++) {
1685                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1686                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1687                 wb_writeout_inc(&bdi->wb);
1688         }
1689         wake_up(&fi->page_waitq);
1690 }
1691
1692 /* Called under fi->lock, may release and reacquire it */
1693 static void fuse_send_writepage(struct fuse_mount *fm,
1694                                 struct fuse_writepage_args *wpa, loff_t size)
1695 __releases(fi->lock)
1696 __acquires(fi->lock)
1697 {
1698         struct fuse_writepage_args *aux, *next;
1699         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1700         struct fuse_write_in *inarg = &wpa->ia.write.in;
1701         struct fuse_args *args = &wpa->ia.ap.args;
1702         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1703         int err;
1704
1705         fi->writectr++;
1706         if (inarg->offset + data_size <= size) {
1707                 inarg->size = data_size;
1708         } else if (inarg->offset < size) {
1709                 inarg->size = size - inarg->offset;
1710         } else {
1711                 /* Got truncated off completely */
1712                 goto out_free;
1713         }
1714
1715         args->in_args[1].size = inarg->size;
1716         args->force = true;
1717         args->nocreds = true;
1718
1719         err = fuse_simple_background(fm, args, GFP_ATOMIC);
1720         if (err == -ENOMEM) {
1721                 spin_unlock(&fi->lock);
1722                 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1723                 spin_lock(&fi->lock);
1724         }
1725
1726         /* Fails on broken connection only */
1727         if (unlikely(err))
1728                 goto out_free;
1729
1730         return;
1731
1732  out_free:
1733         fi->writectr--;
1734         rb_erase(&wpa->writepages_entry, &fi->writepages);
1735         fuse_writepage_finish(fm, wpa);
1736         spin_unlock(&fi->lock);
1737
1738         /* After fuse_writepage_finish() aux request list is private */
1739         for (aux = wpa->next; aux; aux = next) {
1740                 next = aux->next;
1741                 aux->next = NULL;
1742                 fuse_writepage_free(aux);
1743         }
1744
1745         fuse_writepage_free(wpa);
1746         spin_lock(&fi->lock);
1747 }
1748
1749 /*
1750  * If fi->writectr is positive (no truncate or fsync going on) send
1751  * all queued writepage requests.
1752  *
1753  * Called with fi->lock
1754  */
1755 void fuse_flush_writepages(struct inode *inode)
1756 __releases(fi->lock)
1757 __acquires(fi->lock)
1758 {
1759         struct fuse_mount *fm = get_fuse_mount(inode);
1760         struct fuse_inode *fi = get_fuse_inode(inode);
1761         loff_t crop = i_size_read(inode);
1762         struct fuse_writepage_args *wpa;
1763
1764         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1765                 wpa = list_entry(fi->queued_writes.next,
1766                                  struct fuse_writepage_args, queue_entry);
1767                 list_del_init(&wpa->queue_entry);
1768                 fuse_send_writepage(fm, wpa, crop);
1769         }
1770 }
1771
1772 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1773                                                 struct fuse_writepage_args *wpa)
1774 {
1775         pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1776         pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1777         struct rb_node **p = &root->rb_node;
1778         struct rb_node  *parent = NULL;
1779
1780         WARN_ON(!wpa->ia.ap.num_pages);
1781         while (*p) {
1782                 struct fuse_writepage_args *curr;
1783                 pgoff_t curr_index;
1784
1785                 parent = *p;
1786                 curr = rb_entry(parent, struct fuse_writepage_args,
1787                                 writepages_entry);
1788                 WARN_ON(curr->inode != wpa->inode);
1789                 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1790
1791                 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1792                         p = &(*p)->rb_right;
1793                 else if (idx_to < curr_index)
1794                         p = &(*p)->rb_left;
1795                 else
1796                         return curr;
1797         }
1798
1799         rb_link_node(&wpa->writepages_entry, parent, p);
1800         rb_insert_color(&wpa->writepages_entry, root);
1801         return NULL;
1802 }
1803
1804 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1805 {
1806         WARN_ON(fuse_insert_writeback(root, wpa));
1807 }
1808
1809 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1810                                int error)
1811 {
1812         struct fuse_writepage_args *wpa =
1813                 container_of(args, typeof(*wpa), ia.ap.args);
1814         struct inode *inode = wpa->inode;
1815         struct fuse_inode *fi = get_fuse_inode(inode);
1816         struct fuse_conn *fc = get_fuse_conn(inode);
1817
1818         mapping_set_error(inode->i_mapping, error);
1819         /*
1820          * A writeback finished and this might have updated mtime/ctime on
1821          * server making local mtime/ctime stale.  Hence invalidate attrs.
1822          * Do this only if writeback_cache is not enabled.  If writeback_cache
1823          * is enabled, we trust local ctime/mtime.
1824          */
1825         if (!fc->writeback_cache)
1826                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODIFY);
1827         spin_lock(&fi->lock);
1828         rb_erase(&wpa->writepages_entry, &fi->writepages);
1829         while (wpa->next) {
1830                 struct fuse_mount *fm = get_fuse_mount(inode);
1831                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1832                 struct fuse_writepage_args *next = wpa->next;
1833
1834                 wpa->next = next->next;
1835                 next->next = NULL;
1836                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1837                 tree_insert(&fi->writepages, next);
1838
1839                 /*
1840                  * Skip fuse_flush_writepages() to make it easy to crop requests
1841                  * based on primary request size.
1842                  *
1843                  * 1st case (trivial): there are no concurrent activities using
1844                  * fuse_set/release_nowrite.  Then we're on safe side because
1845                  * fuse_flush_writepages() would call fuse_send_writepage()
1846                  * anyway.
1847                  *
1848                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1849                  * now for completion of all in-flight requests.  This happens
1850                  * rarely and no more than once per page, so this should be
1851                  * okay.
1852                  *
1853                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1854                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1855                  * that fuse_set_nowrite returned implies that all in-flight
1856                  * requests were completed along with all of their secondary
1857                  * requests.  Further primary requests are blocked by negative
1858                  * writectr.  Hence there cannot be any in-flight requests and
1859                  * no invocations of fuse_writepage_end() while we're in
1860                  * fuse_set_nowrite..fuse_release_nowrite section.
1861                  */
1862                 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1863         }
1864         fi->writectr--;
1865         fuse_writepage_finish(fm, wpa);
1866         spin_unlock(&fi->lock);
1867         fuse_writepage_free(wpa);
1868 }
1869
1870 static struct fuse_file *__fuse_write_file_get(struct fuse_inode *fi)
1871 {
1872         struct fuse_file *ff;
1873
1874         spin_lock(&fi->lock);
1875         ff = list_first_entry_or_null(&fi->write_files, struct fuse_file,
1876                                       write_entry);
1877         if (ff)
1878                 fuse_file_get(ff);
1879         spin_unlock(&fi->lock);
1880
1881         return ff;
1882 }
1883
1884 static struct fuse_file *fuse_write_file_get(struct fuse_inode *fi)
1885 {
1886         struct fuse_file *ff = __fuse_write_file_get(fi);
1887         WARN_ON(!ff);
1888         return ff;
1889 }
1890
1891 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1892 {
1893         struct fuse_inode *fi = get_fuse_inode(inode);
1894         struct fuse_file *ff;
1895         int err;
1896
1897         /*
1898          * Inode is always written before the last reference is dropped and
1899          * hence this should not be reached from reclaim.
1900          *
1901          * Writing back the inode from reclaim can deadlock if the request
1902          * processing itself needs an allocation.  Allocations triggering
1903          * reclaim while serving a request can't be prevented, because it can
1904          * involve any number of unrelated userspace processes.
1905          */
1906         WARN_ON(wbc->for_reclaim);
1907
1908         ff = __fuse_write_file_get(fi);
1909         err = fuse_flush_times(inode, ff);
1910         if (ff)
1911                 fuse_file_put(ff, false, false);
1912
1913         return err;
1914 }
1915
1916 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1917 {
1918         struct fuse_writepage_args *wpa;
1919         struct fuse_args_pages *ap;
1920
1921         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1922         if (wpa) {
1923                 ap = &wpa->ia.ap;
1924                 ap->num_pages = 0;
1925                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1926                 if (!ap->pages) {
1927                         kfree(wpa);
1928                         wpa = NULL;
1929                 }
1930         }
1931         return wpa;
1932
1933 }
1934
1935 static void fuse_writepage_add_to_bucket(struct fuse_conn *fc,
1936                                          struct fuse_writepage_args *wpa)
1937 {
1938         if (!fc->sync_fs)
1939                 return;
1940
1941         rcu_read_lock();
1942         /* Prevent resurrection of dead bucket in unlikely race with syncfs */
1943         do {
1944                 wpa->bucket = rcu_dereference(fc->curr_bucket);
1945         } while (unlikely(!atomic_inc_not_zero(&wpa->bucket->count)));
1946         rcu_read_unlock();
1947 }
1948
1949 static int fuse_writepage_locked(struct page *page)
1950 {
1951         struct address_space *mapping = page->mapping;
1952         struct inode *inode = mapping->host;
1953         struct fuse_conn *fc = get_fuse_conn(inode);
1954         struct fuse_inode *fi = get_fuse_inode(inode);
1955         struct fuse_writepage_args *wpa;
1956         struct fuse_args_pages *ap;
1957         struct page *tmp_page;
1958         int error = -ENOMEM;
1959
1960         set_page_writeback(page);
1961
1962         wpa = fuse_writepage_args_alloc();
1963         if (!wpa)
1964                 goto err;
1965         ap = &wpa->ia.ap;
1966
1967         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1968         if (!tmp_page)
1969                 goto err_free;
1970
1971         error = -EIO;
1972         wpa->ia.ff = fuse_write_file_get(fi);
1973         if (!wpa->ia.ff)
1974                 goto err_nofile;
1975
1976         fuse_writepage_add_to_bucket(fc, wpa);
1977         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1978
1979         copy_highpage(tmp_page, page);
1980         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1981         wpa->next = NULL;
1982         ap->args.in_pages = true;
1983         ap->num_pages = 1;
1984         ap->pages[0] = tmp_page;
1985         ap->descs[0].offset = 0;
1986         ap->descs[0].length = PAGE_SIZE;
1987         ap->args.end = fuse_writepage_end;
1988         wpa->inode = inode;
1989
1990         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1991         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1992
1993         spin_lock(&fi->lock);
1994         tree_insert(&fi->writepages, wpa);
1995         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1996         fuse_flush_writepages(inode);
1997         spin_unlock(&fi->lock);
1998
1999         end_page_writeback(page);
2000
2001         return 0;
2002
2003 err_nofile:
2004         __free_page(tmp_page);
2005 err_free:
2006         kfree(wpa);
2007 err:
2008         mapping_set_error(page->mapping, error);
2009         end_page_writeback(page);
2010         return error;
2011 }
2012
2013 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
2014 {
2015         struct fuse_conn *fc = get_fuse_conn(page->mapping->host);
2016         int err;
2017
2018         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
2019                 /*
2020                  * ->writepages() should be called for sync() and friends.  We
2021                  * should only get here on direct reclaim and then we are
2022                  * allowed to skip a page which is already in flight
2023                  */
2024                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
2025
2026                 redirty_page_for_writepage(wbc, page);
2027                 unlock_page(page);
2028                 return 0;
2029         }
2030
2031         if (wbc->sync_mode == WB_SYNC_NONE &&
2032             fc->num_background >= fc->congestion_threshold)
2033                 return AOP_WRITEPAGE_ACTIVATE;
2034
2035         err = fuse_writepage_locked(page);
2036         unlock_page(page);
2037
2038         return err;
2039 }
2040
2041 struct fuse_fill_wb_data {
2042         struct fuse_writepage_args *wpa;
2043         struct fuse_file *ff;
2044         struct inode *inode;
2045         struct page **orig_pages;
2046         unsigned int max_pages;
2047 };
2048
2049 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
2050 {
2051         struct fuse_args_pages *ap = &data->wpa->ia.ap;
2052         struct fuse_conn *fc = get_fuse_conn(data->inode);
2053         struct page **pages;
2054         struct fuse_page_desc *descs;
2055         unsigned int npages = min_t(unsigned int,
2056                                     max_t(unsigned int, data->max_pages * 2,
2057                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
2058                                     fc->max_pages);
2059         WARN_ON(npages <= data->max_pages);
2060
2061         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
2062         if (!pages)
2063                 return false;
2064
2065         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
2066         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
2067         kfree(ap->pages);
2068         ap->pages = pages;
2069         ap->descs = descs;
2070         data->max_pages = npages;
2071
2072         return true;
2073 }
2074
2075 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
2076 {
2077         struct fuse_writepage_args *wpa = data->wpa;
2078         struct inode *inode = data->inode;
2079         struct fuse_inode *fi = get_fuse_inode(inode);
2080         int num_pages = wpa->ia.ap.num_pages;
2081         int i;
2082
2083         wpa->ia.ff = fuse_file_get(data->ff);
2084         spin_lock(&fi->lock);
2085         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2086         fuse_flush_writepages(inode);
2087         spin_unlock(&fi->lock);
2088
2089         for (i = 0; i < num_pages; i++)
2090                 end_page_writeback(data->orig_pages[i]);
2091 }
2092
2093 /*
2094  * Check under fi->lock if the page is under writeback, and insert it onto the
2095  * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2096  * one already added for a page at this offset.  If there's none, then insert
2097  * this new request onto the auxiliary list, otherwise reuse the existing one by
2098  * swapping the new temp page with the old one.
2099  */
2100 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2101                                struct page *page)
2102 {
2103         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2104         struct fuse_writepage_args *tmp;
2105         struct fuse_writepage_args *old_wpa;
2106         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2107
2108         WARN_ON(new_ap->num_pages != 0);
2109         new_ap->num_pages = 1;
2110
2111         spin_lock(&fi->lock);
2112         old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2113         if (!old_wpa) {
2114                 spin_unlock(&fi->lock);
2115                 return true;
2116         }
2117
2118         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2119                 pgoff_t curr_index;
2120
2121                 WARN_ON(tmp->inode != new_wpa->inode);
2122                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2123                 if (curr_index == page->index) {
2124                         WARN_ON(tmp->ia.ap.num_pages != 1);
2125                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2126                         break;
2127                 }
2128         }
2129
2130         if (!tmp) {
2131                 new_wpa->next = old_wpa->next;
2132                 old_wpa->next = new_wpa;
2133         }
2134
2135         spin_unlock(&fi->lock);
2136
2137         if (tmp) {
2138                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2139
2140                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2141                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2142                 wb_writeout_inc(&bdi->wb);
2143                 fuse_writepage_free(new_wpa);
2144         }
2145
2146         return false;
2147 }
2148
2149 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2150                                      struct fuse_args_pages *ap,
2151                                      struct fuse_fill_wb_data *data)
2152 {
2153         WARN_ON(!ap->num_pages);
2154
2155         /*
2156          * Being under writeback is unlikely but possible.  For example direct
2157          * read to an mmaped fuse file will set the page dirty twice; once when
2158          * the pages are faulted with get_user_pages(), and then after the read
2159          * completed.
2160          */
2161         if (fuse_page_is_writeback(data->inode, page->index))
2162                 return true;
2163
2164         /* Reached max pages */
2165         if (ap->num_pages == fc->max_pages)
2166                 return true;
2167
2168         /* Reached max write bytes */
2169         if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2170                 return true;
2171
2172         /* Discontinuity */
2173         if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2174                 return true;
2175
2176         /* Need to grow the pages array?  If so, did the expansion fail? */
2177         if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2178                 return true;
2179
2180         return false;
2181 }
2182
2183 static int fuse_writepages_fill(struct folio *folio,
2184                 struct writeback_control *wbc, void *_data)
2185 {
2186         struct fuse_fill_wb_data *data = _data;
2187         struct fuse_writepage_args *wpa = data->wpa;
2188         struct fuse_args_pages *ap = &wpa->ia.ap;
2189         struct inode *inode = data->inode;
2190         struct fuse_inode *fi = get_fuse_inode(inode);
2191         struct fuse_conn *fc = get_fuse_conn(inode);
2192         struct page *tmp_page;
2193         int err;
2194
2195         if (!data->ff) {
2196                 err = -EIO;
2197                 data->ff = fuse_write_file_get(fi);
2198                 if (!data->ff)
2199                         goto out_unlock;
2200         }
2201
2202         if (wpa && fuse_writepage_need_send(fc, &folio->page, ap, data)) {
2203                 fuse_writepages_send(data);
2204                 data->wpa = NULL;
2205         }
2206
2207         err = -ENOMEM;
2208         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2209         if (!tmp_page)
2210                 goto out_unlock;
2211
2212         /*
2213          * The page must not be redirtied until the writeout is completed
2214          * (i.e. userspace has sent a reply to the write request).  Otherwise
2215          * there could be more than one temporary page instance for each real
2216          * page.
2217          *
2218          * This is ensured by holding the page lock in page_mkwrite() while
2219          * checking fuse_page_is_writeback().  We already hold the page lock
2220          * since clear_page_dirty_for_io() and keep it held until we add the
2221          * request to the fi->writepages list and increment ap->num_pages.
2222          * After this fuse_page_is_writeback() will indicate that the page is
2223          * under writeback, so we can release the page lock.
2224          */
2225         if (data->wpa == NULL) {
2226                 err = -ENOMEM;
2227                 wpa = fuse_writepage_args_alloc();
2228                 if (!wpa) {
2229                         __free_page(tmp_page);
2230                         goto out_unlock;
2231                 }
2232                 fuse_writepage_add_to_bucket(fc, wpa);
2233
2234                 data->max_pages = 1;
2235
2236                 ap = &wpa->ia.ap;
2237                 fuse_write_args_fill(&wpa->ia, data->ff, folio_pos(folio), 0);
2238                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2239                 wpa->next = NULL;
2240                 ap->args.in_pages = true;
2241                 ap->args.end = fuse_writepage_end;
2242                 ap->num_pages = 0;
2243                 wpa->inode = inode;
2244         }
2245         folio_start_writeback(folio);
2246
2247         copy_highpage(tmp_page, &folio->page);
2248         ap->pages[ap->num_pages] = tmp_page;
2249         ap->descs[ap->num_pages].offset = 0;
2250         ap->descs[ap->num_pages].length = PAGE_SIZE;
2251         data->orig_pages[ap->num_pages] = &folio->page;
2252
2253         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2254         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2255
2256         err = 0;
2257         if (data->wpa) {
2258                 /*
2259                  * Protected by fi->lock against concurrent access by
2260                  * fuse_page_is_writeback().
2261                  */
2262                 spin_lock(&fi->lock);
2263                 ap->num_pages++;
2264                 spin_unlock(&fi->lock);
2265         } else if (fuse_writepage_add(wpa, &folio->page)) {
2266                 data->wpa = wpa;
2267         } else {
2268                 folio_end_writeback(folio);
2269         }
2270 out_unlock:
2271         folio_unlock(folio);
2272
2273         return err;
2274 }
2275
2276 static int fuse_writepages(struct address_space *mapping,
2277                            struct writeback_control *wbc)
2278 {
2279         struct inode *inode = mapping->host;
2280         struct fuse_conn *fc = get_fuse_conn(inode);
2281         struct fuse_fill_wb_data data;
2282         int err;
2283
2284         err = -EIO;
2285         if (fuse_is_bad(inode))
2286                 goto out;
2287
2288         if (wbc->sync_mode == WB_SYNC_NONE &&
2289             fc->num_background >= fc->congestion_threshold)
2290                 return 0;
2291
2292         data.inode = inode;
2293         data.wpa = NULL;
2294         data.ff = NULL;
2295
2296         err = -ENOMEM;
2297         data.orig_pages = kcalloc(fc->max_pages,
2298                                   sizeof(struct page *),
2299                                   GFP_NOFS);
2300         if (!data.orig_pages)
2301                 goto out;
2302
2303         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2304         if (data.wpa) {
2305                 WARN_ON(!data.wpa->ia.ap.num_pages);
2306                 fuse_writepages_send(&data);
2307         }
2308         if (data.ff)
2309                 fuse_file_put(data.ff, false, false);
2310
2311         kfree(data.orig_pages);
2312 out:
2313         return err;
2314 }
2315
2316 /*
2317  * It's worthy to make sure that space is reserved on disk for the write,
2318  * but how to implement it without killing performance need more thinking.
2319  */
2320 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2321                 loff_t pos, unsigned len, struct page **pagep, void **fsdata)
2322 {
2323         pgoff_t index = pos >> PAGE_SHIFT;
2324         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2325         struct page *page;
2326         loff_t fsize;
2327         int err = -ENOMEM;
2328
2329         WARN_ON(!fc->writeback_cache);
2330
2331         page = grab_cache_page_write_begin(mapping, index);
2332         if (!page)
2333                 goto error;
2334
2335         fuse_wait_on_page_writeback(mapping->host, page->index);
2336
2337         if (PageUptodate(page) || len == PAGE_SIZE)
2338                 goto success;
2339         /*
2340          * Check if the start this page comes after the end of file, in which
2341          * case the readpage can be optimized away.
2342          */
2343         fsize = i_size_read(mapping->host);
2344         if (fsize <= (pos & PAGE_MASK)) {
2345                 size_t off = pos & ~PAGE_MASK;
2346                 if (off)
2347                         zero_user_segment(page, 0, off);
2348                 goto success;
2349         }
2350         err = fuse_do_readpage(file, page);
2351         if (err)
2352                 goto cleanup;
2353 success:
2354         *pagep = page;
2355         return 0;
2356
2357 cleanup:
2358         unlock_page(page);
2359         put_page(page);
2360 error:
2361         return err;
2362 }
2363
2364 static int fuse_write_end(struct file *file, struct address_space *mapping,
2365                 loff_t pos, unsigned len, unsigned copied,
2366                 struct page *page, void *fsdata)
2367 {
2368         struct inode *inode = page->mapping->host;
2369
2370         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2371         if (!copied)
2372                 goto unlock;
2373
2374         pos += copied;
2375         if (!PageUptodate(page)) {
2376                 /* Zero any unwritten bytes at the end of the page */
2377                 size_t endoff = pos & ~PAGE_MASK;
2378                 if (endoff)
2379                         zero_user_segment(page, endoff, PAGE_SIZE);
2380                 SetPageUptodate(page);
2381         }
2382
2383         if (pos > inode->i_size)
2384                 i_size_write(inode, pos);
2385
2386         set_page_dirty(page);
2387
2388 unlock:
2389         unlock_page(page);
2390         put_page(page);
2391
2392         return copied;
2393 }
2394
2395 static int fuse_launder_folio(struct folio *folio)
2396 {
2397         int err = 0;
2398         if (folio_clear_dirty_for_io(folio)) {
2399                 struct inode *inode = folio->mapping->host;
2400
2401                 /* Serialize with pending writeback for the same page */
2402                 fuse_wait_on_page_writeback(inode, folio->index);
2403                 err = fuse_writepage_locked(&folio->page);
2404                 if (!err)
2405                         fuse_wait_on_page_writeback(inode, folio->index);
2406         }
2407         return err;
2408 }
2409
2410 /*
2411  * Write back dirty data/metadata now (there may not be any suitable
2412  * open files later for data)
2413  */
2414 static void fuse_vma_close(struct vm_area_struct *vma)
2415 {
2416         int err;
2417
2418         err = write_inode_now(vma->vm_file->f_mapping->host, 1);
2419         mapping_set_error(vma->vm_file->f_mapping, err);
2420 }
2421
2422 /*
2423  * Wait for writeback against this page to complete before allowing it
2424  * to be marked dirty again, and hence written back again, possibly
2425  * before the previous writepage completed.
2426  *
2427  * Block here, instead of in ->writepage(), so that the userspace fs
2428  * can only block processes actually operating on the filesystem.
2429  *
2430  * Otherwise unprivileged userspace fs would be able to block
2431  * unrelated:
2432  *
2433  * - page migration
2434  * - sync(2)
2435  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2436  */
2437 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2438 {
2439         struct page *page = vmf->page;
2440         struct inode *inode = file_inode(vmf->vma->vm_file);
2441
2442         file_update_time(vmf->vma->vm_file);
2443         lock_page(page);
2444         if (page->mapping != inode->i_mapping) {
2445                 unlock_page(page);
2446                 return VM_FAULT_NOPAGE;
2447         }
2448
2449         fuse_wait_on_page_writeback(inode, page->index);
2450         return VM_FAULT_LOCKED;
2451 }
2452
2453 static const struct vm_operations_struct fuse_file_vm_ops = {
2454         .close          = fuse_vma_close,
2455         .fault          = filemap_fault,
2456         .map_pages      = filemap_map_pages,
2457         .page_mkwrite   = fuse_page_mkwrite,
2458 };
2459
2460 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2461 {
2462         struct fuse_file *ff = file->private_data;
2463         struct fuse_conn *fc = ff->fm->fc;
2464
2465         /* DAX mmap is superior to direct_io mmap */
2466         if (FUSE_IS_DAX(file_inode(file)))
2467                 return fuse_dax_mmap(file, vma);
2468
2469         if (ff->open_flags & FOPEN_DIRECT_IO) {
2470                 /* Can't provide the coherency needed for MAP_SHARED
2471                  * if FUSE_DIRECT_IO_ALLOW_MMAP isn't set.
2472                  */
2473                 if ((vma->vm_flags & VM_MAYSHARE) && !fc->direct_io_allow_mmap)
2474                         return -ENODEV;
2475
2476                 invalidate_inode_pages2(file->f_mapping);
2477
2478                 return generic_file_mmap(file, vma);
2479         }
2480
2481         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2482                 fuse_link_write_file(file);
2483
2484         file_accessed(file);
2485         vma->vm_ops = &fuse_file_vm_ops;
2486         return 0;
2487 }
2488
2489 static int convert_fuse_file_lock(struct fuse_conn *fc,
2490                                   const struct fuse_file_lock *ffl,
2491                                   struct file_lock *fl)
2492 {
2493         switch (ffl->type) {
2494         case F_UNLCK:
2495                 break;
2496
2497         case F_RDLCK:
2498         case F_WRLCK:
2499                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2500                     ffl->end < ffl->start)
2501                         return -EIO;
2502
2503                 fl->fl_start = ffl->start;
2504                 fl->fl_end = ffl->end;
2505
2506                 /*
2507                  * Convert pid into init's pid namespace.  The locks API will
2508                  * translate it into the caller's pid namespace.
2509                  */
2510                 rcu_read_lock();
2511                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2512                 rcu_read_unlock();
2513                 break;
2514
2515         default:
2516                 return -EIO;
2517         }
2518         fl->fl_type = ffl->type;
2519         return 0;
2520 }
2521
2522 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2523                          const struct file_lock *fl, int opcode, pid_t pid,
2524                          int flock, struct fuse_lk_in *inarg)
2525 {
2526         struct inode *inode = file_inode(file);
2527         struct fuse_conn *fc = get_fuse_conn(inode);
2528         struct fuse_file *ff = file->private_data;
2529
2530         memset(inarg, 0, sizeof(*inarg));
2531         inarg->fh = ff->fh;
2532         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2533         inarg->lk.start = fl->fl_start;
2534         inarg->lk.end = fl->fl_end;
2535         inarg->lk.type = fl->fl_type;
2536         inarg->lk.pid = pid;
2537         if (flock)
2538                 inarg->lk_flags |= FUSE_LK_FLOCK;
2539         args->opcode = opcode;
2540         args->nodeid = get_node_id(inode);
2541         args->in_numargs = 1;
2542         args->in_args[0].size = sizeof(*inarg);
2543         args->in_args[0].value = inarg;
2544 }
2545
2546 static int fuse_getlk(struct file *file, struct file_lock *fl)
2547 {
2548         struct inode *inode = file_inode(file);
2549         struct fuse_mount *fm = get_fuse_mount(inode);
2550         FUSE_ARGS(args);
2551         struct fuse_lk_in inarg;
2552         struct fuse_lk_out outarg;
2553         int err;
2554
2555         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2556         args.out_numargs = 1;
2557         args.out_args[0].size = sizeof(outarg);
2558         args.out_args[0].value = &outarg;
2559         err = fuse_simple_request(fm, &args);
2560         if (!err)
2561                 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2562
2563         return err;
2564 }
2565
2566 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2567 {
2568         struct inode *inode = file_inode(file);
2569         struct fuse_mount *fm = get_fuse_mount(inode);
2570         FUSE_ARGS(args);
2571         struct fuse_lk_in inarg;
2572         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2573         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2574         pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2575         int err;
2576
2577         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2578                 /* NLM needs asynchronous locks, which we don't support yet */
2579                 return -ENOLCK;
2580         }
2581
2582         /* Unlock on close is handled by the flush method */
2583         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2584                 return 0;
2585
2586         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2587         err = fuse_simple_request(fm, &args);
2588
2589         /* locking is restartable */
2590         if (err == -EINTR)
2591                 err = -ERESTARTSYS;
2592
2593         return err;
2594 }
2595
2596 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2597 {
2598         struct inode *inode = file_inode(file);
2599         struct fuse_conn *fc = get_fuse_conn(inode);
2600         int err;
2601
2602         if (cmd == F_CANCELLK) {
2603                 err = 0;
2604         } else if (cmd == F_GETLK) {
2605                 if (fc->no_lock) {
2606                         posix_test_lock(file, fl);
2607                         err = 0;
2608                 } else
2609                         err = fuse_getlk(file, fl);
2610         } else {
2611                 if (fc->no_lock)
2612                         err = posix_lock_file(file, fl, NULL);
2613                 else
2614                         err = fuse_setlk(file, fl, 0);
2615         }
2616         return err;
2617 }
2618
2619 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2620 {
2621         struct inode *inode = file_inode(file);
2622         struct fuse_conn *fc = get_fuse_conn(inode);
2623         int err;
2624
2625         if (fc->no_flock) {
2626                 err = locks_lock_file_wait(file, fl);
2627         } else {
2628                 struct fuse_file *ff = file->private_data;
2629
2630                 /* emulate flock with POSIX locks */
2631                 ff->flock = true;
2632                 err = fuse_setlk(file, fl, 1);
2633         }
2634
2635         return err;
2636 }
2637
2638 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2639 {
2640         struct inode *inode = mapping->host;
2641         struct fuse_mount *fm = get_fuse_mount(inode);
2642         FUSE_ARGS(args);
2643         struct fuse_bmap_in inarg;
2644         struct fuse_bmap_out outarg;
2645         int err;
2646
2647         if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2648                 return 0;
2649
2650         memset(&inarg, 0, sizeof(inarg));
2651         inarg.block = block;
2652         inarg.blocksize = inode->i_sb->s_blocksize;
2653         args.opcode = FUSE_BMAP;
2654         args.nodeid = get_node_id(inode);
2655         args.in_numargs = 1;
2656         args.in_args[0].size = sizeof(inarg);
2657         args.in_args[0].value = &inarg;
2658         args.out_numargs = 1;
2659         args.out_args[0].size = sizeof(outarg);
2660         args.out_args[0].value = &outarg;
2661         err = fuse_simple_request(fm, &args);
2662         if (err == -ENOSYS)
2663                 fm->fc->no_bmap = 1;
2664
2665         return err ? 0 : outarg.block;
2666 }
2667
2668 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2669 {
2670         struct inode *inode = file->f_mapping->host;
2671         struct fuse_mount *fm = get_fuse_mount(inode);
2672         struct fuse_file *ff = file->private_data;
2673         FUSE_ARGS(args);
2674         struct fuse_lseek_in inarg = {
2675                 .fh = ff->fh,
2676                 .offset = offset,
2677                 .whence = whence
2678         };
2679         struct fuse_lseek_out outarg;
2680         int err;
2681
2682         if (fm->fc->no_lseek)
2683                 goto fallback;
2684
2685         args.opcode = FUSE_LSEEK;
2686         args.nodeid = ff->nodeid;
2687         args.in_numargs = 1;
2688         args.in_args[0].size = sizeof(inarg);
2689         args.in_args[0].value = &inarg;
2690         args.out_numargs = 1;
2691         args.out_args[0].size = sizeof(outarg);
2692         args.out_args[0].value = &outarg;
2693         err = fuse_simple_request(fm, &args);
2694         if (err) {
2695                 if (err == -ENOSYS) {
2696                         fm->fc->no_lseek = 1;
2697                         goto fallback;
2698                 }
2699                 return err;
2700         }
2701
2702         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2703
2704 fallback:
2705         err = fuse_update_attributes(inode, file, STATX_SIZE);
2706         if (!err)
2707                 return generic_file_llseek(file, offset, whence);
2708         else
2709                 return err;
2710 }
2711
2712 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2713 {
2714         loff_t retval;
2715         struct inode *inode = file_inode(file);
2716
2717         switch (whence) {
2718         case SEEK_SET:
2719         case SEEK_CUR:
2720                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2721                 retval = generic_file_llseek(file, offset, whence);
2722                 break;
2723         case SEEK_END:
2724                 inode_lock(inode);
2725                 retval = fuse_update_attributes(inode, file, STATX_SIZE);
2726                 if (!retval)
2727                         retval = generic_file_llseek(file, offset, whence);
2728                 inode_unlock(inode);
2729                 break;
2730         case SEEK_HOLE:
2731         case SEEK_DATA:
2732                 inode_lock(inode);
2733                 retval = fuse_lseek(file, offset, whence);
2734                 inode_unlock(inode);
2735                 break;
2736         default:
2737                 retval = -EINVAL;
2738         }
2739
2740         return retval;
2741 }
2742
2743 /*
2744  * All files which have been polled are linked to RB tree
2745  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2746  * find the matching one.
2747  */
2748 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2749                                               struct rb_node **parent_out)
2750 {
2751         struct rb_node **link = &fc->polled_files.rb_node;
2752         struct rb_node *last = NULL;
2753
2754         while (*link) {
2755                 struct fuse_file *ff;
2756
2757                 last = *link;
2758                 ff = rb_entry(last, struct fuse_file, polled_node);
2759
2760                 if (kh < ff->kh)
2761                         link = &last->rb_left;
2762                 else if (kh > ff->kh)
2763                         link = &last->rb_right;
2764                 else
2765                         return link;
2766         }
2767
2768         if (parent_out)
2769                 *parent_out = last;
2770         return link;
2771 }
2772
2773 /*
2774  * The file is about to be polled.  Make sure it's on the polled_files
2775  * RB tree.  Note that files once added to the polled_files tree are
2776  * not removed before the file is released.  This is because a file
2777  * polled once is likely to be polled again.
2778  */
2779 static void fuse_register_polled_file(struct fuse_conn *fc,
2780                                       struct fuse_file *ff)
2781 {
2782         spin_lock(&fc->lock);
2783         if (RB_EMPTY_NODE(&ff->polled_node)) {
2784                 struct rb_node **link, *parent;
2785
2786                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2787                 BUG_ON(*link);
2788                 rb_link_node(&ff->polled_node, parent, link);
2789                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2790         }
2791         spin_unlock(&fc->lock);
2792 }
2793
2794 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2795 {
2796         struct fuse_file *ff = file->private_data;
2797         struct fuse_mount *fm = ff->fm;
2798         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2799         struct fuse_poll_out outarg;
2800         FUSE_ARGS(args);
2801         int err;
2802
2803         if (fm->fc->no_poll)
2804                 return DEFAULT_POLLMASK;
2805
2806         poll_wait(file, &ff->poll_wait, wait);
2807         inarg.events = mangle_poll(poll_requested_events(wait));
2808
2809         /*
2810          * Ask for notification iff there's someone waiting for it.
2811          * The client may ignore the flag and always notify.
2812          */
2813         if (waitqueue_active(&ff->poll_wait)) {
2814                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2815                 fuse_register_polled_file(fm->fc, ff);
2816         }
2817
2818         args.opcode = FUSE_POLL;
2819         args.nodeid = ff->nodeid;
2820         args.in_numargs = 1;
2821         args.in_args[0].size = sizeof(inarg);
2822         args.in_args[0].value = &inarg;
2823         args.out_numargs = 1;
2824         args.out_args[0].size = sizeof(outarg);
2825         args.out_args[0].value = &outarg;
2826         err = fuse_simple_request(fm, &args);
2827
2828         if (!err)
2829                 return demangle_poll(outarg.revents);
2830         if (err == -ENOSYS) {
2831                 fm->fc->no_poll = 1;
2832                 return DEFAULT_POLLMASK;
2833         }
2834         return EPOLLERR;
2835 }
2836 EXPORT_SYMBOL_GPL(fuse_file_poll);
2837
2838 /*
2839  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2840  * wakes up the poll waiters.
2841  */
2842 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2843                             struct fuse_notify_poll_wakeup_out *outarg)
2844 {
2845         u64 kh = outarg->kh;
2846         struct rb_node **link;
2847
2848         spin_lock(&fc->lock);
2849
2850         link = fuse_find_polled_node(fc, kh, NULL);
2851         if (*link) {
2852                 struct fuse_file *ff;
2853
2854                 ff = rb_entry(*link, struct fuse_file, polled_node);
2855                 wake_up_interruptible_sync(&ff->poll_wait);
2856         }
2857
2858         spin_unlock(&fc->lock);
2859         return 0;
2860 }
2861
2862 static void fuse_do_truncate(struct file *file)
2863 {
2864         struct inode *inode = file->f_mapping->host;
2865         struct iattr attr;
2866
2867         attr.ia_valid = ATTR_SIZE;
2868         attr.ia_size = i_size_read(inode);
2869
2870         attr.ia_file = file;
2871         attr.ia_valid |= ATTR_FILE;
2872
2873         fuse_do_setattr(file_dentry(file), &attr, file);
2874 }
2875
2876 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2877 {
2878         return round_up(off, fc->max_pages << PAGE_SHIFT);
2879 }
2880
2881 static ssize_t
2882 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2883 {
2884         DECLARE_COMPLETION_ONSTACK(wait);
2885         ssize_t ret = 0;
2886         struct file *file = iocb->ki_filp;
2887         struct fuse_file *ff = file->private_data;
2888         loff_t pos = 0;
2889         struct inode *inode;
2890         loff_t i_size;
2891         size_t count = iov_iter_count(iter), shortened = 0;
2892         loff_t offset = iocb->ki_pos;
2893         struct fuse_io_priv *io;
2894
2895         pos = offset;
2896         inode = file->f_mapping->host;
2897         i_size = i_size_read(inode);
2898
2899         if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
2900                 return 0;
2901
2902         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2903         if (!io)
2904                 return -ENOMEM;
2905         spin_lock_init(&io->lock);
2906         kref_init(&io->refcnt);
2907         io->reqs = 1;
2908         io->bytes = -1;
2909         io->size = 0;
2910         io->offset = offset;
2911         io->write = (iov_iter_rw(iter) == WRITE);
2912         io->err = 0;
2913         /*
2914          * By default, we want to optimize all I/Os with async request
2915          * submission to the client filesystem if supported.
2916          */
2917         io->async = ff->fm->fc->async_dio;
2918         io->iocb = iocb;
2919         io->blocking = is_sync_kiocb(iocb);
2920
2921         /* optimization for short read */
2922         if (io->async && !io->write && offset + count > i_size) {
2923                 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
2924                 shortened = count - iov_iter_count(iter);
2925                 count -= shortened;
2926         }
2927
2928         /*
2929          * We cannot asynchronously extend the size of a file.
2930          * In such case the aio will behave exactly like sync io.
2931          */
2932         if ((offset + count > i_size) && io->write)
2933                 io->blocking = true;
2934
2935         if (io->async && io->blocking) {
2936                 /*
2937                  * Additional reference to keep io around after
2938                  * calling fuse_aio_complete()
2939                  */
2940                 kref_get(&io->refcnt);
2941                 io->done = &wait;
2942         }
2943
2944         if (iov_iter_rw(iter) == WRITE) {
2945                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2946                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
2947         } else {
2948                 ret = __fuse_direct_read(io, iter, &pos);
2949         }
2950         iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
2951
2952         if (io->async) {
2953                 bool blocking = io->blocking;
2954
2955                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2956
2957                 /* we have a non-extending, async request, so return */
2958                 if (!blocking)
2959                         return -EIOCBQUEUED;
2960
2961                 wait_for_completion(&wait);
2962                 ret = fuse_get_res_by_io(io);
2963         }
2964
2965         kref_put(&io->refcnt, fuse_io_release);
2966
2967         if (iov_iter_rw(iter) == WRITE) {
2968                 fuse_write_update_attr(inode, pos, ret);
2969                 /* For extending writes we already hold exclusive lock */
2970                 if (ret < 0 && offset + count > i_size)
2971                         fuse_do_truncate(file);
2972         }
2973
2974         return ret;
2975 }
2976
2977 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
2978 {
2979         int err = filemap_write_and_wait_range(inode->i_mapping, start, LLONG_MAX);
2980
2981         if (!err)
2982                 fuse_sync_writes(inode);
2983
2984         return err;
2985 }
2986
2987 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2988                                 loff_t length)
2989 {
2990         struct fuse_file *ff = file->private_data;
2991         struct inode *inode = file_inode(file);
2992         struct fuse_inode *fi = get_fuse_inode(inode);
2993         struct fuse_mount *fm = ff->fm;
2994         FUSE_ARGS(args);
2995         struct fuse_fallocate_in inarg = {
2996                 .fh = ff->fh,
2997                 .offset = offset,
2998                 .length = length,
2999                 .mode = mode
3000         };
3001         int err;
3002         bool block_faults = FUSE_IS_DAX(inode) &&
3003                 (!(mode & FALLOC_FL_KEEP_SIZE) ||
3004                  (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)));
3005
3006         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3007                      FALLOC_FL_ZERO_RANGE))
3008                 return -EOPNOTSUPP;
3009
3010         if (fm->fc->no_fallocate)
3011                 return -EOPNOTSUPP;
3012
3013         inode_lock(inode);
3014         if (block_faults) {
3015                 filemap_invalidate_lock(inode->i_mapping);
3016                 err = fuse_dax_break_layouts(inode, 0, 0);
3017                 if (err)
3018                         goto out;
3019         }
3020
3021         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)) {
3022                 loff_t endbyte = offset + length - 1;
3023
3024                 err = fuse_writeback_range(inode, offset, endbyte);
3025                 if (err)
3026                         goto out;
3027         }
3028
3029         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3030             offset + length > i_size_read(inode)) {
3031                 err = inode_newsize_ok(inode, offset + length);
3032                 if (err)
3033                         goto out;
3034         }
3035
3036         err = file_modified(file);
3037         if (err)
3038                 goto out;
3039
3040         if (!(mode & FALLOC_FL_KEEP_SIZE))
3041                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3042
3043         args.opcode = FUSE_FALLOCATE;
3044         args.nodeid = ff->nodeid;
3045         args.in_numargs = 1;
3046         args.in_args[0].size = sizeof(inarg);
3047         args.in_args[0].value = &inarg;
3048         err = fuse_simple_request(fm, &args);
3049         if (err == -ENOSYS) {
3050                 fm->fc->no_fallocate = 1;
3051                 err = -EOPNOTSUPP;
3052         }
3053         if (err)
3054                 goto out;
3055
3056         /* we could have extended the file */
3057         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3058                 if (fuse_write_update_attr(inode, offset + length, length))
3059                         file_update_time(file);
3060         }
3061
3062         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE))
3063                 truncate_pagecache_range(inode, offset, offset + length - 1);
3064
3065         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
3066
3067 out:
3068         if (!(mode & FALLOC_FL_KEEP_SIZE))
3069                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3070
3071         if (block_faults)
3072                 filemap_invalidate_unlock(inode->i_mapping);
3073
3074         inode_unlock(inode);
3075
3076         fuse_flush_time_update(inode);
3077
3078         return err;
3079 }
3080
3081 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3082                                       struct file *file_out, loff_t pos_out,
3083                                       size_t len, unsigned int flags)
3084 {
3085         struct fuse_file *ff_in = file_in->private_data;
3086         struct fuse_file *ff_out = file_out->private_data;
3087         struct inode *inode_in = file_inode(file_in);
3088         struct inode *inode_out = file_inode(file_out);
3089         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3090         struct fuse_mount *fm = ff_in->fm;
3091         struct fuse_conn *fc = fm->fc;
3092         FUSE_ARGS(args);
3093         struct fuse_copy_file_range_in inarg = {
3094                 .fh_in = ff_in->fh,
3095                 .off_in = pos_in,
3096                 .nodeid_out = ff_out->nodeid,
3097                 .fh_out = ff_out->fh,
3098                 .off_out = pos_out,
3099                 .len = len,
3100                 .flags = flags
3101         };
3102         struct fuse_write_out outarg;
3103         ssize_t err;
3104         /* mark unstable when write-back is not used, and file_out gets
3105          * extended */
3106         bool is_unstable = (!fc->writeback_cache) &&
3107                            ((pos_out + len) > inode_out->i_size);
3108
3109         if (fc->no_copy_file_range)
3110                 return -EOPNOTSUPP;
3111
3112         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3113                 return -EXDEV;
3114
3115         inode_lock(inode_in);
3116         err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3117         inode_unlock(inode_in);
3118         if (err)
3119                 return err;
3120
3121         inode_lock(inode_out);
3122
3123         err = file_modified(file_out);
3124         if (err)
3125                 goto out;
3126
3127         /*
3128          * Write out dirty pages in the destination file before sending the COPY
3129          * request to userspace.  After the request is completed, truncate off
3130          * pages (including partial ones) from the cache that have been copied,
3131          * since these contain stale data at that point.
3132          *
3133          * This should be mostly correct, but if the COPY writes to partial
3134          * pages (at the start or end) and the parts not covered by the COPY are
3135          * written through a memory map after calling fuse_writeback_range(),
3136          * then these partial page modifications will be lost on truncation.
3137          *
3138          * It is unlikely that someone would rely on such mixed style
3139          * modifications.  Yet this does give less guarantees than if the
3140          * copying was performed with write(2).
3141          *
3142          * To fix this a mapping->invalidate_lock could be used to prevent new
3143          * faults while the copy is ongoing.
3144          */
3145         err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3146         if (err)
3147                 goto out;
3148
3149         if (is_unstable)
3150                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3151
3152         args.opcode = FUSE_COPY_FILE_RANGE;
3153         args.nodeid = ff_in->nodeid;
3154         args.in_numargs = 1;
3155         args.in_args[0].size = sizeof(inarg);
3156         args.in_args[0].value = &inarg;
3157         args.out_numargs = 1;
3158         args.out_args[0].size = sizeof(outarg);
3159         args.out_args[0].value = &outarg;
3160         err = fuse_simple_request(fm, &args);
3161         if (err == -ENOSYS) {
3162                 fc->no_copy_file_range = 1;
3163                 err = -EOPNOTSUPP;
3164         }
3165         if (err)
3166                 goto out;
3167
3168         truncate_inode_pages_range(inode_out->i_mapping,
3169                                    ALIGN_DOWN(pos_out, PAGE_SIZE),
3170                                    ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3171
3172         file_update_time(file_out);
3173         fuse_write_update_attr(inode_out, pos_out + outarg.size, outarg.size);
3174
3175         err = outarg.size;
3176 out:
3177         if (is_unstable)
3178                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3179
3180         inode_unlock(inode_out);
3181         file_accessed(file_in);
3182
3183         fuse_flush_time_update(inode_out);
3184
3185         return err;
3186 }
3187
3188 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3189                                     struct file *dst_file, loff_t dst_off,
3190                                     size_t len, unsigned int flags)
3191 {
3192         ssize_t ret;
3193
3194         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3195                                      len, flags);
3196
3197         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3198                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3199                                               dst_off, len, flags);
3200         return ret;
3201 }
3202
3203 static const struct file_operations fuse_file_operations = {
3204         .llseek         = fuse_file_llseek,
3205         .read_iter      = fuse_file_read_iter,
3206         .write_iter     = fuse_file_write_iter,
3207         .mmap           = fuse_file_mmap,
3208         .open           = fuse_open,
3209         .flush          = fuse_flush,
3210         .release        = fuse_release,
3211         .fsync          = fuse_fsync,
3212         .lock           = fuse_file_lock,
3213         .get_unmapped_area = thp_get_unmapped_area,
3214         .flock          = fuse_file_flock,
3215         .splice_read    = filemap_splice_read,
3216         .splice_write   = iter_file_splice_write,
3217         .unlocked_ioctl = fuse_file_ioctl,
3218         .compat_ioctl   = fuse_file_compat_ioctl,
3219         .poll           = fuse_file_poll,
3220         .fallocate      = fuse_file_fallocate,
3221         .copy_file_range = fuse_copy_file_range,
3222 };
3223
3224 static const struct address_space_operations fuse_file_aops  = {
3225         .read_folio     = fuse_read_folio,
3226         .readahead      = fuse_readahead,
3227         .writepage      = fuse_writepage,
3228         .writepages     = fuse_writepages,
3229         .launder_folio  = fuse_launder_folio,
3230         .dirty_folio    = filemap_dirty_folio,
3231         .bmap           = fuse_bmap,
3232         .direct_IO      = fuse_direct_IO,
3233         .write_begin    = fuse_write_begin,
3234         .write_end      = fuse_write_end,
3235 };
3236
3237 void fuse_init_file_inode(struct inode *inode, unsigned int flags)
3238 {
3239         struct fuse_inode *fi = get_fuse_inode(inode);
3240
3241         inode->i_fop = &fuse_file_operations;
3242         inode->i_data.a_ops = &fuse_file_aops;
3243
3244         INIT_LIST_HEAD(&fi->write_files);
3245         INIT_LIST_HEAD(&fi->queued_writes);
3246         fi->writectr = 0;
3247         init_waitqueue_head(&fi->page_waitq);
3248         fi->writepages = RB_ROOT;
3249
3250         if (IS_ENABLED(CONFIG_FUSE_DAX))
3251                 fuse_dax_inode_init(inode, flags);
3252 }