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,
1247                                   struct address_space *mapping,
1248                                   struct iov_iter *ii, loff_t pos)
1249 {
1250         struct inode *inode = mapping->host;
1251         struct fuse_conn *fc = get_fuse_conn(inode);
1252         struct fuse_inode *fi = get_fuse_inode(inode);
1253         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         return res > 0 ? res : err;
1296 }
1297
1298 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1299 {
1300         struct file *file = iocb->ki_filp;
1301         struct address_space *mapping = file->f_mapping;
1302         ssize_t written = 0;
1303         ssize_t written_buffered = 0;
1304         struct inode *inode = mapping->host;
1305         ssize_t err;
1306         struct fuse_conn *fc = get_fuse_conn(inode);
1307         loff_t endbyte = 0;
1308
1309         if (fc->writeback_cache) {
1310                 /* Update size (EOF optimization) and mode (SUID clearing) */
1311                 err = fuse_update_attributes(mapping->host, file,
1312                                              STATX_SIZE | STATX_MODE);
1313                 if (err)
1314                         return err;
1315
1316                 if (fc->handle_killpriv_v2 &&
1317                     setattr_should_drop_suidgid(&nop_mnt_idmap,
1318                                                 file_inode(file))) {
1319                         goto writethrough;
1320                 }
1321
1322                 return generic_file_write_iter(iocb, from);
1323         }
1324
1325 writethrough:
1326         inode_lock(inode);
1327
1328         /* We can write back this queue in page reclaim */
1329         current->backing_dev_info = inode_to_bdi(inode);
1330
1331         err = generic_write_checks(iocb, from);
1332         if (err <= 0)
1333                 goto out;
1334
1335         err = file_remove_privs(file);
1336         if (err)
1337                 goto out;
1338
1339         err = file_update_time(file);
1340         if (err)
1341                 goto out;
1342
1343         if (iocb->ki_flags & IOCB_DIRECT) {
1344                 loff_t pos = iocb->ki_pos;
1345                 written = generic_file_direct_write(iocb, from);
1346                 if (written < 0 || !iov_iter_count(from))
1347                         goto out;
1348
1349                 pos += written;
1350
1351                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1352                 if (written_buffered < 0) {
1353                         err = written_buffered;
1354                         goto out;
1355                 }
1356                 endbyte = pos + written_buffered - 1;
1357
1358                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1359                                                    endbyte);
1360                 if (err)
1361                         goto out;
1362
1363                 invalidate_mapping_pages(file->f_mapping,
1364                                          pos >> PAGE_SHIFT,
1365                                          endbyte >> PAGE_SHIFT);
1366
1367                 written += written_buffered;
1368                 iocb->ki_pos = pos + written_buffered;
1369         } else {
1370                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1371                 if (written >= 0)
1372                         iocb->ki_pos += written;
1373         }
1374 out:
1375         current->backing_dev_info = NULL;
1376         inode_unlock(inode);
1377         if (written > 0)
1378                 written = generic_write_sync(iocb, written);
1379
1380         return written ? written : err;
1381 }
1382
1383 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1384 {
1385         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1386 }
1387
1388 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1389                                         size_t max_size)
1390 {
1391         return min(iov_iter_single_seg_count(ii), max_size);
1392 }
1393
1394 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1395                                size_t *nbytesp, int write,
1396                                unsigned int max_pages)
1397 {
1398         size_t nbytes = 0;  /* # bytes already packed in req */
1399         ssize_t ret = 0;
1400
1401         /* Special case for kernel I/O: can copy directly into the buffer */
1402         if (iov_iter_is_kvec(ii)) {
1403                 unsigned long user_addr = fuse_get_user_addr(ii);
1404                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1405
1406                 if (write)
1407                         ap->args.in_args[1].value = (void *) user_addr;
1408                 else
1409                         ap->args.out_args[0].value = (void *) user_addr;
1410
1411                 iov_iter_advance(ii, frag_size);
1412                 *nbytesp = frag_size;
1413                 return 0;
1414         }
1415
1416         while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1417                 unsigned npages;
1418                 size_t start;
1419                 ret = iov_iter_get_pages2(ii, &ap->pages[ap->num_pages],
1420                                         *nbytesp - nbytes,
1421                                         max_pages - ap->num_pages,
1422                                         &start);
1423                 if (ret < 0)
1424                         break;
1425
1426                 nbytes += ret;
1427
1428                 ret += start;
1429                 npages = DIV_ROUND_UP(ret, PAGE_SIZE);
1430
1431                 ap->descs[ap->num_pages].offset = start;
1432                 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1433
1434                 ap->num_pages += npages;
1435                 ap->descs[ap->num_pages - 1].length -=
1436                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1437         }
1438
1439         ap->args.user_pages = true;
1440         if (write)
1441                 ap->args.in_pages = true;
1442         else
1443                 ap->args.out_pages = true;
1444
1445         *nbytesp = nbytes;
1446
1447         return ret < 0 ? ret : 0;
1448 }
1449
1450 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1451                        loff_t *ppos, int flags)
1452 {
1453         int write = flags & FUSE_DIO_WRITE;
1454         int cuse = flags & FUSE_DIO_CUSE;
1455         struct file *file = io->iocb->ki_filp;
1456         struct inode *inode = file->f_mapping->host;
1457         struct fuse_file *ff = file->private_data;
1458         struct fuse_conn *fc = ff->fm->fc;
1459         size_t nmax = write ? fc->max_write : fc->max_read;
1460         loff_t pos = *ppos;
1461         size_t count = iov_iter_count(iter);
1462         pgoff_t idx_from = pos >> PAGE_SHIFT;
1463         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1464         ssize_t res = 0;
1465         int err = 0;
1466         struct fuse_io_args *ia;
1467         unsigned int max_pages;
1468
1469         max_pages = iov_iter_npages(iter, fc->max_pages);
1470         ia = fuse_io_alloc(io, max_pages);
1471         if (!ia)
1472                 return -ENOMEM;
1473
1474         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1475                 if (!write)
1476                         inode_lock(inode);
1477                 fuse_sync_writes(inode);
1478                 if (!write)
1479                         inode_unlock(inode);
1480         }
1481
1482         io->should_dirty = !write && user_backed_iter(iter);
1483         while (count) {
1484                 ssize_t nres;
1485                 fl_owner_t owner = current->files;
1486                 size_t nbytes = min(count, nmax);
1487
1488                 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1489                                           max_pages);
1490                 if (err && !nbytes)
1491                         break;
1492
1493                 if (write) {
1494                         if (!capable(CAP_FSETID))
1495                                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1496
1497                         nres = fuse_send_write(ia, pos, nbytes, owner);
1498                 } else {
1499                         nres = fuse_send_read(ia, pos, nbytes, owner);
1500                 }
1501
1502                 if (!io->async || nres < 0) {
1503                         fuse_release_user_pages(&ia->ap, io->should_dirty);
1504                         fuse_io_free(ia);
1505                 }
1506                 ia = NULL;
1507                 if (nres < 0) {
1508                         iov_iter_revert(iter, nbytes);
1509                         err = nres;
1510                         break;
1511                 }
1512                 WARN_ON(nres > nbytes);
1513
1514                 count -= nres;
1515                 res += nres;
1516                 pos += nres;
1517                 if (nres != nbytes) {
1518                         iov_iter_revert(iter, nbytes - nres);
1519                         break;
1520                 }
1521                 if (count) {
1522                         max_pages = iov_iter_npages(iter, fc->max_pages);
1523                         ia = fuse_io_alloc(io, max_pages);
1524                         if (!ia)
1525                                 break;
1526                 }
1527         }
1528         if (ia)
1529                 fuse_io_free(ia);
1530         if (res > 0)
1531                 *ppos = pos;
1532
1533         return res > 0 ? res : err;
1534 }
1535 EXPORT_SYMBOL_GPL(fuse_direct_io);
1536
1537 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1538                                   struct iov_iter *iter,
1539                                   loff_t *ppos)
1540 {
1541         ssize_t res;
1542         struct inode *inode = file_inode(io->iocb->ki_filp);
1543
1544         res = fuse_direct_io(io, iter, ppos, 0);
1545
1546         fuse_invalidate_atime(inode);
1547
1548         return res;
1549 }
1550
1551 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1552
1553 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1554 {
1555         ssize_t res;
1556
1557         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1558                 res = fuse_direct_IO(iocb, to);
1559         } else {
1560                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1561
1562                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1563         }
1564
1565         return res;
1566 }
1567
1568 static bool fuse_direct_write_extending_i_size(struct kiocb *iocb,
1569                                                struct iov_iter *iter)
1570 {
1571         struct inode *inode = file_inode(iocb->ki_filp);
1572
1573         return iocb->ki_pos + iov_iter_count(iter) > i_size_read(inode);
1574 }
1575
1576 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1577 {
1578         struct inode *inode = file_inode(iocb->ki_filp);
1579         struct file *file = iocb->ki_filp;
1580         struct fuse_file *ff = file->private_data;
1581         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1582         ssize_t res;
1583         bool exclusive_lock =
1584                 !(ff->open_flags & FOPEN_PARALLEL_DIRECT_WRITES) ||
1585                 iocb->ki_flags & IOCB_APPEND ||
1586                 fuse_direct_write_extending_i_size(iocb, from);
1587
1588         /*
1589          * Take exclusive lock if
1590          * - Parallel direct writes are disabled - a user space decision
1591          * - Parallel direct writes are enabled and i_size is being extended.
1592          *   This might not be needed at all, but needs further investigation.
1593          */
1594         if (exclusive_lock)
1595                 inode_lock(inode);
1596         else {
1597                 inode_lock_shared(inode);
1598
1599                 /* A race with truncate might have come up as the decision for
1600                  * the lock type was done without holding the lock, check again.
1601                  */
1602                 if (fuse_direct_write_extending_i_size(iocb, from)) {
1603                         inode_unlock_shared(inode);
1604                         inode_lock(inode);
1605                         exclusive_lock = true;
1606                 }
1607         }
1608
1609         res = generic_write_checks(iocb, from);
1610         if (res > 0) {
1611                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1612                         res = fuse_direct_IO(iocb, from);
1613                 } else {
1614                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1615                                              FUSE_DIO_WRITE);
1616                         fuse_write_update_attr(inode, iocb->ki_pos, res);
1617                 }
1618         }
1619         if (exclusive_lock)
1620                 inode_unlock(inode);
1621         else
1622                 inode_unlock_shared(inode);
1623
1624         return res;
1625 }
1626
1627 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1628 {
1629         struct file *file = iocb->ki_filp;
1630         struct fuse_file *ff = file->private_data;
1631         struct inode *inode = file_inode(file);
1632
1633         if (fuse_is_bad(inode))
1634                 return -EIO;
1635
1636         if (FUSE_IS_DAX(inode))
1637                 return fuse_dax_read_iter(iocb, to);
1638
1639         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1640                 return fuse_cache_read_iter(iocb, to);
1641         else
1642                 return fuse_direct_read_iter(iocb, to);
1643 }
1644
1645 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1646 {
1647         struct file *file = iocb->ki_filp;
1648         struct fuse_file *ff = file->private_data;
1649         struct inode *inode = file_inode(file);
1650
1651         if (fuse_is_bad(inode))
1652                 return -EIO;
1653
1654         if (FUSE_IS_DAX(inode))
1655                 return fuse_dax_write_iter(iocb, from);
1656
1657         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1658                 return fuse_cache_write_iter(iocb, from);
1659         else
1660                 return fuse_direct_write_iter(iocb, from);
1661 }
1662
1663 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1664 {
1665         struct fuse_args_pages *ap = &wpa->ia.ap;
1666         int i;
1667
1668         if (wpa->bucket)
1669                 fuse_sync_bucket_dec(wpa->bucket);
1670
1671         for (i = 0; i < ap->num_pages; i++)
1672                 __free_page(ap->pages[i]);
1673
1674         if (wpa->ia.ff)
1675                 fuse_file_put(wpa->ia.ff, false, false);
1676
1677         kfree(ap->pages);
1678         kfree(wpa);
1679 }
1680
1681 static void fuse_writepage_finish(struct fuse_mount *fm,
1682                                   struct fuse_writepage_args *wpa)
1683 {
1684         struct fuse_args_pages *ap = &wpa->ia.ap;
1685         struct inode *inode = wpa->inode;
1686         struct fuse_inode *fi = get_fuse_inode(inode);
1687         struct backing_dev_info *bdi = inode_to_bdi(inode);
1688         int i;
1689
1690         for (i = 0; i < ap->num_pages; i++) {
1691                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1692                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1693                 wb_writeout_inc(&bdi->wb);
1694         }
1695         wake_up(&fi->page_waitq);
1696 }
1697
1698 /* Called under fi->lock, may release and reacquire it */
1699 static void fuse_send_writepage(struct fuse_mount *fm,
1700                                 struct fuse_writepage_args *wpa, loff_t size)
1701 __releases(fi->lock)
1702 __acquires(fi->lock)
1703 {
1704         struct fuse_writepage_args *aux, *next;
1705         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1706         struct fuse_write_in *inarg = &wpa->ia.write.in;
1707         struct fuse_args *args = &wpa->ia.ap.args;
1708         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1709         int err;
1710
1711         fi->writectr++;
1712         if (inarg->offset + data_size <= size) {
1713                 inarg->size = data_size;
1714         } else if (inarg->offset < size) {
1715                 inarg->size = size - inarg->offset;
1716         } else {
1717                 /* Got truncated off completely */
1718                 goto out_free;
1719         }
1720
1721         args->in_args[1].size = inarg->size;
1722         args->force = true;
1723         args->nocreds = true;
1724
1725         err = fuse_simple_background(fm, args, GFP_ATOMIC);
1726         if (err == -ENOMEM) {
1727                 spin_unlock(&fi->lock);
1728                 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1729                 spin_lock(&fi->lock);
1730         }
1731
1732         /* Fails on broken connection only */
1733         if (unlikely(err))
1734                 goto out_free;
1735
1736         return;
1737
1738  out_free:
1739         fi->writectr--;
1740         rb_erase(&wpa->writepages_entry, &fi->writepages);
1741         fuse_writepage_finish(fm, wpa);
1742         spin_unlock(&fi->lock);
1743
1744         /* After fuse_writepage_finish() aux request list is private */
1745         for (aux = wpa->next; aux; aux = next) {
1746                 next = aux->next;
1747                 aux->next = NULL;
1748                 fuse_writepage_free(aux);
1749         }
1750
1751         fuse_writepage_free(wpa);
1752         spin_lock(&fi->lock);
1753 }
1754
1755 /*
1756  * If fi->writectr is positive (no truncate or fsync going on) send
1757  * all queued writepage requests.
1758  *
1759  * Called with fi->lock
1760  */
1761 void fuse_flush_writepages(struct inode *inode)
1762 __releases(fi->lock)
1763 __acquires(fi->lock)
1764 {
1765         struct fuse_mount *fm = get_fuse_mount(inode);
1766         struct fuse_inode *fi = get_fuse_inode(inode);
1767         loff_t crop = i_size_read(inode);
1768         struct fuse_writepage_args *wpa;
1769
1770         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1771                 wpa = list_entry(fi->queued_writes.next,
1772                                  struct fuse_writepage_args, queue_entry);
1773                 list_del_init(&wpa->queue_entry);
1774                 fuse_send_writepage(fm, wpa, crop);
1775         }
1776 }
1777
1778 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1779                                                 struct fuse_writepage_args *wpa)
1780 {
1781         pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1782         pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1783         struct rb_node **p = &root->rb_node;
1784         struct rb_node  *parent = NULL;
1785
1786         WARN_ON(!wpa->ia.ap.num_pages);
1787         while (*p) {
1788                 struct fuse_writepage_args *curr;
1789                 pgoff_t curr_index;
1790
1791                 parent = *p;
1792                 curr = rb_entry(parent, struct fuse_writepage_args,
1793                                 writepages_entry);
1794                 WARN_ON(curr->inode != wpa->inode);
1795                 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1796
1797                 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1798                         p = &(*p)->rb_right;
1799                 else if (idx_to < curr_index)
1800                         p = &(*p)->rb_left;
1801                 else
1802                         return curr;
1803         }
1804
1805         rb_link_node(&wpa->writepages_entry, parent, p);
1806         rb_insert_color(&wpa->writepages_entry, root);
1807         return NULL;
1808 }
1809
1810 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1811 {
1812         WARN_ON(fuse_insert_writeback(root, wpa));
1813 }
1814
1815 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1816                                int error)
1817 {
1818         struct fuse_writepage_args *wpa =
1819                 container_of(args, typeof(*wpa), ia.ap.args);
1820         struct inode *inode = wpa->inode;
1821         struct fuse_inode *fi = get_fuse_inode(inode);
1822         struct fuse_conn *fc = get_fuse_conn(inode);
1823
1824         mapping_set_error(inode->i_mapping, error);
1825         /*
1826          * A writeback finished and this might have updated mtime/ctime on
1827          * server making local mtime/ctime stale.  Hence invalidate attrs.
1828          * Do this only if writeback_cache is not enabled.  If writeback_cache
1829          * is enabled, we trust local ctime/mtime.
1830          */
1831         if (!fc->writeback_cache)
1832                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODIFY);
1833         spin_lock(&fi->lock);
1834         rb_erase(&wpa->writepages_entry, &fi->writepages);
1835         while (wpa->next) {
1836                 struct fuse_mount *fm = get_fuse_mount(inode);
1837                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1838                 struct fuse_writepage_args *next = wpa->next;
1839
1840                 wpa->next = next->next;
1841                 next->next = NULL;
1842                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1843                 tree_insert(&fi->writepages, next);
1844
1845                 /*
1846                  * Skip fuse_flush_writepages() to make it easy to crop requests
1847                  * based on primary request size.
1848                  *
1849                  * 1st case (trivial): there are no concurrent activities using
1850                  * fuse_set/release_nowrite.  Then we're on safe side because
1851                  * fuse_flush_writepages() would call fuse_send_writepage()
1852                  * anyway.
1853                  *
1854                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1855                  * now for completion of all in-flight requests.  This happens
1856                  * rarely and no more than once per page, so this should be
1857                  * okay.
1858                  *
1859                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1860                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1861                  * that fuse_set_nowrite returned implies that all in-flight
1862                  * requests were completed along with all of their secondary
1863                  * requests.  Further primary requests are blocked by negative
1864                  * writectr.  Hence there cannot be any in-flight requests and
1865                  * no invocations of fuse_writepage_end() while we're in
1866                  * fuse_set_nowrite..fuse_release_nowrite section.
1867                  */
1868                 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1869         }
1870         fi->writectr--;
1871         fuse_writepage_finish(fm, wpa);
1872         spin_unlock(&fi->lock);
1873         fuse_writepage_free(wpa);
1874 }
1875
1876 static struct fuse_file *__fuse_write_file_get(struct fuse_inode *fi)
1877 {
1878         struct fuse_file *ff;
1879
1880         spin_lock(&fi->lock);
1881         ff = list_first_entry_or_null(&fi->write_files, struct fuse_file,
1882                                       write_entry);
1883         if (ff)
1884                 fuse_file_get(ff);
1885         spin_unlock(&fi->lock);
1886
1887         return ff;
1888 }
1889
1890 static struct fuse_file *fuse_write_file_get(struct fuse_inode *fi)
1891 {
1892         struct fuse_file *ff = __fuse_write_file_get(fi);
1893         WARN_ON(!ff);
1894         return ff;
1895 }
1896
1897 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1898 {
1899         struct fuse_inode *fi = get_fuse_inode(inode);
1900         struct fuse_file *ff;
1901         int err;
1902
1903         /*
1904          * Inode is always written before the last reference is dropped and
1905          * hence this should not be reached from reclaim.
1906          *
1907          * Writing back the inode from reclaim can deadlock if the request
1908          * processing itself needs an allocation.  Allocations triggering
1909          * reclaim while serving a request can't be prevented, because it can
1910          * involve any number of unrelated userspace processes.
1911          */
1912         WARN_ON(wbc->for_reclaim);
1913
1914         ff = __fuse_write_file_get(fi);
1915         err = fuse_flush_times(inode, ff);
1916         if (ff)
1917                 fuse_file_put(ff, false, false);
1918
1919         return err;
1920 }
1921
1922 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1923 {
1924         struct fuse_writepage_args *wpa;
1925         struct fuse_args_pages *ap;
1926
1927         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1928         if (wpa) {
1929                 ap = &wpa->ia.ap;
1930                 ap->num_pages = 0;
1931                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1932                 if (!ap->pages) {
1933                         kfree(wpa);
1934                         wpa = NULL;
1935                 }
1936         }
1937         return wpa;
1938
1939 }
1940
1941 static void fuse_writepage_add_to_bucket(struct fuse_conn *fc,
1942                                          struct fuse_writepage_args *wpa)
1943 {
1944         if (!fc->sync_fs)
1945                 return;
1946
1947         rcu_read_lock();
1948         /* Prevent resurrection of dead bucket in unlikely race with syncfs */
1949         do {
1950                 wpa->bucket = rcu_dereference(fc->curr_bucket);
1951         } while (unlikely(!atomic_inc_not_zero(&wpa->bucket->count)));
1952         rcu_read_unlock();
1953 }
1954
1955 static int fuse_writepage_locked(struct page *page)
1956 {
1957         struct address_space *mapping = page->mapping;
1958         struct inode *inode = mapping->host;
1959         struct fuse_conn *fc = get_fuse_conn(inode);
1960         struct fuse_inode *fi = get_fuse_inode(inode);
1961         struct fuse_writepage_args *wpa;
1962         struct fuse_args_pages *ap;
1963         struct page *tmp_page;
1964         int error = -ENOMEM;
1965
1966         set_page_writeback(page);
1967
1968         wpa = fuse_writepage_args_alloc();
1969         if (!wpa)
1970                 goto err;
1971         ap = &wpa->ia.ap;
1972
1973         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1974         if (!tmp_page)
1975                 goto err_free;
1976
1977         error = -EIO;
1978         wpa->ia.ff = fuse_write_file_get(fi);
1979         if (!wpa->ia.ff)
1980                 goto err_nofile;
1981
1982         fuse_writepage_add_to_bucket(fc, wpa);
1983         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1984
1985         copy_highpage(tmp_page, page);
1986         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1987         wpa->next = NULL;
1988         ap->args.in_pages = true;
1989         ap->num_pages = 1;
1990         ap->pages[0] = tmp_page;
1991         ap->descs[0].offset = 0;
1992         ap->descs[0].length = PAGE_SIZE;
1993         ap->args.end = fuse_writepage_end;
1994         wpa->inode = inode;
1995
1996         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1997         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1998
1999         spin_lock(&fi->lock);
2000         tree_insert(&fi->writepages, wpa);
2001         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2002         fuse_flush_writepages(inode);
2003         spin_unlock(&fi->lock);
2004
2005         end_page_writeback(page);
2006
2007         return 0;
2008
2009 err_nofile:
2010         __free_page(tmp_page);
2011 err_free:
2012         kfree(wpa);
2013 err:
2014         mapping_set_error(page->mapping, error);
2015         end_page_writeback(page);
2016         return error;
2017 }
2018
2019 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
2020 {
2021         struct fuse_conn *fc = get_fuse_conn(page->mapping->host);
2022         int err;
2023
2024         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
2025                 /*
2026                  * ->writepages() should be called for sync() and friends.  We
2027                  * should only get here on direct reclaim and then we are
2028                  * allowed to skip a page which is already in flight
2029                  */
2030                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
2031
2032                 redirty_page_for_writepage(wbc, page);
2033                 unlock_page(page);
2034                 return 0;
2035         }
2036
2037         if (wbc->sync_mode == WB_SYNC_NONE &&
2038             fc->num_background >= fc->congestion_threshold)
2039                 return AOP_WRITEPAGE_ACTIVATE;
2040
2041         err = fuse_writepage_locked(page);
2042         unlock_page(page);
2043
2044         return err;
2045 }
2046
2047 struct fuse_fill_wb_data {
2048         struct fuse_writepage_args *wpa;
2049         struct fuse_file *ff;
2050         struct inode *inode;
2051         struct page **orig_pages;
2052         unsigned int max_pages;
2053 };
2054
2055 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
2056 {
2057         struct fuse_args_pages *ap = &data->wpa->ia.ap;
2058         struct fuse_conn *fc = get_fuse_conn(data->inode);
2059         struct page **pages;
2060         struct fuse_page_desc *descs;
2061         unsigned int npages = min_t(unsigned int,
2062                                     max_t(unsigned int, data->max_pages * 2,
2063                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
2064                                     fc->max_pages);
2065         WARN_ON(npages <= data->max_pages);
2066
2067         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
2068         if (!pages)
2069                 return false;
2070
2071         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
2072         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
2073         kfree(ap->pages);
2074         ap->pages = pages;
2075         ap->descs = descs;
2076         data->max_pages = npages;
2077
2078         return true;
2079 }
2080
2081 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
2082 {
2083         struct fuse_writepage_args *wpa = data->wpa;
2084         struct inode *inode = data->inode;
2085         struct fuse_inode *fi = get_fuse_inode(inode);
2086         int num_pages = wpa->ia.ap.num_pages;
2087         int i;
2088
2089         wpa->ia.ff = fuse_file_get(data->ff);
2090         spin_lock(&fi->lock);
2091         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2092         fuse_flush_writepages(inode);
2093         spin_unlock(&fi->lock);
2094
2095         for (i = 0; i < num_pages; i++)
2096                 end_page_writeback(data->orig_pages[i]);
2097 }
2098
2099 /*
2100  * Check under fi->lock if the page is under writeback, and insert it onto the
2101  * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2102  * one already added for a page at this offset.  If there's none, then insert
2103  * this new request onto the auxiliary list, otherwise reuse the existing one by
2104  * swapping the new temp page with the old one.
2105  */
2106 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2107                                struct page *page)
2108 {
2109         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2110         struct fuse_writepage_args *tmp;
2111         struct fuse_writepage_args *old_wpa;
2112         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2113
2114         WARN_ON(new_ap->num_pages != 0);
2115         new_ap->num_pages = 1;
2116
2117         spin_lock(&fi->lock);
2118         old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2119         if (!old_wpa) {
2120                 spin_unlock(&fi->lock);
2121                 return true;
2122         }
2123
2124         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2125                 pgoff_t curr_index;
2126
2127                 WARN_ON(tmp->inode != new_wpa->inode);
2128                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2129                 if (curr_index == page->index) {
2130                         WARN_ON(tmp->ia.ap.num_pages != 1);
2131                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2132                         break;
2133                 }
2134         }
2135
2136         if (!tmp) {
2137                 new_wpa->next = old_wpa->next;
2138                 old_wpa->next = new_wpa;
2139         }
2140
2141         spin_unlock(&fi->lock);
2142
2143         if (tmp) {
2144                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2145
2146                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2147                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2148                 wb_writeout_inc(&bdi->wb);
2149                 fuse_writepage_free(new_wpa);
2150         }
2151
2152         return false;
2153 }
2154
2155 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2156                                      struct fuse_args_pages *ap,
2157                                      struct fuse_fill_wb_data *data)
2158 {
2159         WARN_ON(!ap->num_pages);
2160
2161         /*
2162          * Being under writeback is unlikely but possible.  For example direct
2163          * read to an mmaped fuse file will set the page dirty twice; once when
2164          * the pages are faulted with get_user_pages(), and then after the read
2165          * completed.
2166          */
2167         if (fuse_page_is_writeback(data->inode, page->index))
2168                 return true;
2169
2170         /* Reached max pages */
2171         if (ap->num_pages == fc->max_pages)
2172                 return true;
2173
2174         /* Reached max write bytes */
2175         if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2176                 return true;
2177
2178         /* Discontinuity */
2179         if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2180                 return true;
2181
2182         /* Need to grow the pages array?  If so, did the expansion fail? */
2183         if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2184                 return true;
2185
2186         return false;
2187 }
2188
2189 static int fuse_writepages_fill(struct folio *folio,
2190                 struct writeback_control *wbc, void *_data)
2191 {
2192         struct fuse_fill_wb_data *data = _data;
2193         struct fuse_writepage_args *wpa = data->wpa;
2194         struct fuse_args_pages *ap = &wpa->ia.ap;
2195         struct inode *inode = data->inode;
2196         struct fuse_inode *fi = get_fuse_inode(inode);
2197         struct fuse_conn *fc = get_fuse_conn(inode);
2198         struct page *tmp_page;
2199         int err;
2200
2201         if (!data->ff) {
2202                 err = -EIO;
2203                 data->ff = fuse_write_file_get(fi);
2204                 if (!data->ff)
2205                         goto out_unlock;
2206         }
2207
2208         if (wpa && fuse_writepage_need_send(fc, &folio->page, ap, data)) {
2209                 fuse_writepages_send(data);
2210                 data->wpa = NULL;
2211         }
2212
2213         err = -ENOMEM;
2214         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2215         if (!tmp_page)
2216                 goto out_unlock;
2217
2218         /*
2219          * The page must not be redirtied until the writeout is completed
2220          * (i.e. userspace has sent a reply to the write request).  Otherwise
2221          * there could be more than one temporary page instance for each real
2222          * page.
2223          *
2224          * This is ensured by holding the page lock in page_mkwrite() while
2225          * checking fuse_page_is_writeback().  We already hold the page lock
2226          * since clear_page_dirty_for_io() and keep it held until we add the
2227          * request to the fi->writepages list and increment ap->num_pages.
2228          * After this fuse_page_is_writeback() will indicate that the page is
2229          * under writeback, so we can release the page lock.
2230          */
2231         if (data->wpa == NULL) {
2232                 err = -ENOMEM;
2233                 wpa = fuse_writepage_args_alloc();
2234                 if (!wpa) {
2235                         __free_page(tmp_page);
2236                         goto out_unlock;
2237                 }
2238                 fuse_writepage_add_to_bucket(fc, wpa);
2239
2240                 data->max_pages = 1;
2241
2242                 ap = &wpa->ia.ap;
2243                 fuse_write_args_fill(&wpa->ia, data->ff, folio_pos(folio), 0);
2244                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2245                 wpa->next = NULL;
2246                 ap->args.in_pages = true;
2247                 ap->args.end = fuse_writepage_end;
2248                 ap->num_pages = 0;
2249                 wpa->inode = inode;
2250         }
2251         folio_start_writeback(folio);
2252
2253         copy_highpage(tmp_page, &folio->page);
2254         ap->pages[ap->num_pages] = tmp_page;
2255         ap->descs[ap->num_pages].offset = 0;
2256         ap->descs[ap->num_pages].length = PAGE_SIZE;
2257         data->orig_pages[ap->num_pages] = &folio->page;
2258
2259         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2260         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2261
2262         err = 0;
2263         if (data->wpa) {
2264                 /*
2265                  * Protected by fi->lock against concurrent access by
2266                  * fuse_page_is_writeback().
2267                  */
2268                 spin_lock(&fi->lock);
2269                 ap->num_pages++;
2270                 spin_unlock(&fi->lock);
2271         } else if (fuse_writepage_add(wpa, &folio->page)) {
2272                 data->wpa = wpa;
2273         } else {
2274                 folio_end_writeback(folio);
2275         }
2276 out_unlock:
2277         folio_unlock(folio);
2278
2279         return err;
2280 }
2281
2282 static int fuse_writepages(struct address_space *mapping,
2283                            struct writeback_control *wbc)
2284 {
2285         struct inode *inode = mapping->host;
2286         struct fuse_conn *fc = get_fuse_conn(inode);
2287         struct fuse_fill_wb_data data;
2288         int err;
2289
2290         err = -EIO;
2291         if (fuse_is_bad(inode))
2292                 goto out;
2293
2294         if (wbc->sync_mode == WB_SYNC_NONE &&
2295             fc->num_background >= fc->congestion_threshold)
2296                 return 0;
2297
2298         data.inode = inode;
2299         data.wpa = NULL;
2300         data.ff = NULL;
2301
2302         err = -ENOMEM;
2303         data.orig_pages = kcalloc(fc->max_pages,
2304                                   sizeof(struct page *),
2305                                   GFP_NOFS);
2306         if (!data.orig_pages)
2307                 goto out;
2308
2309         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2310         if (data.wpa) {
2311                 WARN_ON(!data.wpa->ia.ap.num_pages);
2312                 fuse_writepages_send(&data);
2313         }
2314         if (data.ff)
2315                 fuse_file_put(data.ff, false, false);
2316
2317         kfree(data.orig_pages);
2318 out:
2319         return err;
2320 }
2321
2322 /*
2323  * It's worthy to make sure that space is reserved on disk for the write,
2324  * but how to implement it without killing performance need more thinking.
2325  */
2326 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2327                 loff_t pos, unsigned len, struct page **pagep, void **fsdata)
2328 {
2329         pgoff_t index = pos >> PAGE_SHIFT;
2330         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2331         struct page *page;
2332         loff_t fsize;
2333         int err = -ENOMEM;
2334
2335         WARN_ON(!fc->writeback_cache);
2336
2337         page = grab_cache_page_write_begin(mapping, index);
2338         if (!page)
2339                 goto error;
2340
2341         fuse_wait_on_page_writeback(mapping->host, page->index);
2342
2343         if (PageUptodate(page) || len == PAGE_SIZE)
2344                 goto success;
2345         /*
2346          * Check if the start this page comes after the end of file, in which
2347          * case the readpage can be optimized away.
2348          */
2349         fsize = i_size_read(mapping->host);
2350         if (fsize <= (pos & PAGE_MASK)) {
2351                 size_t off = pos & ~PAGE_MASK;
2352                 if (off)
2353                         zero_user_segment(page, 0, off);
2354                 goto success;
2355         }
2356         err = fuse_do_readpage(file, page);
2357         if (err)
2358                 goto cleanup;
2359 success:
2360         *pagep = page;
2361         return 0;
2362
2363 cleanup:
2364         unlock_page(page);
2365         put_page(page);
2366 error:
2367         return err;
2368 }
2369
2370 static int fuse_write_end(struct file *file, struct address_space *mapping,
2371                 loff_t pos, unsigned len, unsigned copied,
2372                 struct page *page, void *fsdata)
2373 {
2374         struct inode *inode = page->mapping->host;
2375
2376         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2377         if (!copied)
2378                 goto unlock;
2379
2380         pos += copied;
2381         if (!PageUptodate(page)) {
2382                 /* Zero any unwritten bytes at the end of the page */
2383                 size_t endoff = pos & ~PAGE_MASK;
2384                 if (endoff)
2385                         zero_user_segment(page, endoff, PAGE_SIZE);
2386                 SetPageUptodate(page);
2387         }
2388
2389         if (pos > inode->i_size)
2390                 i_size_write(inode, pos);
2391
2392         set_page_dirty(page);
2393
2394 unlock:
2395         unlock_page(page);
2396         put_page(page);
2397
2398         return copied;
2399 }
2400
2401 static int fuse_launder_folio(struct folio *folio)
2402 {
2403         int err = 0;
2404         if (folio_clear_dirty_for_io(folio)) {
2405                 struct inode *inode = folio->mapping->host;
2406
2407                 /* Serialize with pending writeback for the same page */
2408                 fuse_wait_on_page_writeback(inode, folio->index);
2409                 err = fuse_writepage_locked(&folio->page);
2410                 if (!err)
2411                         fuse_wait_on_page_writeback(inode, folio->index);
2412         }
2413         return err;
2414 }
2415
2416 /*
2417  * Write back dirty data/metadata now (there may not be any suitable
2418  * open files later for data)
2419  */
2420 static void fuse_vma_close(struct vm_area_struct *vma)
2421 {
2422         int err;
2423
2424         err = write_inode_now(vma->vm_file->f_mapping->host, 1);
2425         mapping_set_error(vma->vm_file->f_mapping, err);
2426 }
2427
2428 /*
2429  * Wait for writeback against this page to complete before allowing it
2430  * to be marked dirty again, and hence written back again, possibly
2431  * before the previous writepage completed.
2432  *
2433  * Block here, instead of in ->writepage(), so that the userspace fs
2434  * can only block processes actually operating on the filesystem.
2435  *
2436  * Otherwise unprivileged userspace fs would be able to block
2437  * unrelated:
2438  *
2439  * - page migration
2440  * - sync(2)
2441  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2442  */
2443 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2444 {
2445         struct page *page = vmf->page;
2446         struct inode *inode = file_inode(vmf->vma->vm_file);
2447
2448         file_update_time(vmf->vma->vm_file);
2449         lock_page(page);
2450         if (page->mapping != inode->i_mapping) {
2451                 unlock_page(page);
2452                 return VM_FAULT_NOPAGE;
2453         }
2454
2455         fuse_wait_on_page_writeback(inode, page->index);
2456         return VM_FAULT_LOCKED;
2457 }
2458
2459 static const struct vm_operations_struct fuse_file_vm_ops = {
2460         .close          = fuse_vma_close,
2461         .fault          = filemap_fault,
2462         .map_pages      = filemap_map_pages,
2463         .page_mkwrite   = fuse_page_mkwrite,
2464 };
2465
2466 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2467 {
2468         struct fuse_file *ff = file->private_data;
2469
2470         /* DAX mmap is superior to direct_io mmap */
2471         if (FUSE_IS_DAX(file_inode(file)))
2472                 return fuse_dax_mmap(file, vma);
2473
2474         if (ff->open_flags & FOPEN_DIRECT_IO) {
2475                 /* Can't provide the coherency needed for MAP_SHARED */
2476                 if (vma->vm_flags & VM_MAYSHARE)
2477                         return -ENODEV;
2478
2479                 invalidate_inode_pages2(file->f_mapping);
2480
2481                 return generic_file_mmap(file, vma);
2482         }
2483
2484         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2485                 fuse_link_write_file(file);
2486
2487         file_accessed(file);
2488         vma->vm_ops = &fuse_file_vm_ops;
2489         return 0;
2490 }
2491
2492 static int convert_fuse_file_lock(struct fuse_conn *fc,
2493                                   const struct fuse_file_lock *ffl,
2494                                   struct file_lock *fl)
2495 {
2496         switch (ffl->type) {
2497         case F_UNLCK:
2498                 break;
2499
2500         case F_RDLCK:
2501         case F_WRLCK:
2502                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2503                     ffl->end < ffl->start)
2504                         return -EIO;
2505
2506                 fl->fl_start = ffl->start;
2507                 fl->fl_end = ffl->end;
2508
2509                 /*
2510                  * Convert pid into init's pid namespace.  The locks API will
2511                  * translate it into the caller's pid namespace.
2512                  */
2513                 rcu_read_lock();
2514                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2515                 rcu_read_unlock();
2516                 break;
2517
2518         default:
2519                 return -EIO;
2520         }
2521         fl->fl_type = ffl->type;
2522         return 0;
2523 }
2524
2525 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2526                          const struct file_lock *fl, int opcode, pid_t pid,
2527                          int flock, struct fuse_lk_in *inarg)
2528 {
2529         struct inode *inode = file_inode(file);
2530         struct fuse_conn *fc = get_fuse_conn(inode);
2531         struct fuse_file *ff = file->private_data;
2532
2533         memset(inarg, 0, sizeof(*inarg));
2534         inarg->fh = ff->fh;
2535         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2536         inarg->lk.start = fl->fl_start;
2537         inarg->lk.end = fl->fl_end;
2538         inarg->lk.type = fl->fl_type;
2539         inarg->lk.pid = pid;
2540         if (flock)
2541                 inarg->lk_flags |= FUSE_LK_FLOCK;
2542         args->opcode = opcode;
2543         args->nodeid = get_node_id(inode);
2544         args->in_numargs = 1;
2545         args->in_args[0].size = sizeof(*inarg);
2546         args->in_args[0].value = inarg;
2547 }
2548
2549 static int fuse_getlk(struct file *file, struct file_lock *fl)
2550 {
2551         struct inode *inode = file_inode(file);
2552         struct fuse_mount *fm = get_fuse_mount(inode);
2553         FUSE_ARGS(args);
2554         struct fuse_lk_in inarg;
2555         struct fuse_lk_out outarg;
2556         int err;
2557
2558         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2559         args.out_numargs = 1;
2560         args.out_args[0].size = sizeof(outarg);
2561         args.out_args[0].value = &outarg;
2562         err = fuse_simple_request(fm, &args);
2563         if (!err)
2564                 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2565
2566         return err;
2567 }
2568
2569 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2570 {
2571         struct inode *inode = file_inode(file);
2572         struct fuse_mount *fm = get_fuse_mount(inode);
2573         FUSE_ARGS(args);
2574         struct fuse_lk_in inarg;
2575         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2576         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2577         pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2578         int err;
2579
2580         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2581                 /* NLM needs asynchronous locks, which we don't support yet */
2582                 return -ENOLCK;
2583         }
2584
2585         /* Unlock on close is handled by the flush method */
2586         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2587                 return 0;
2588
2589         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2590         err = fuse_simple_request(fm, &args);
2591
2592         /* locking is restartable */
2593         if (err == -EINTR)
2594                 err = -ERESTARTSYS;
2595
2596         return err;
2597 }
2598
2599 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2600 {
2601         struct inode *inode = file_inode(file);
2602         struct fuse_conn *fc = get_fuse_conn(inode);
2603         int err;
2604
2605         if (cmd == F_CANCELLK) {
2606                 err = 0;
2607         } else if (cmd == F_GETLK) {
2608                 if (fc->no_lock) {
2609                         posix_test_lock(file, fl);
2610                         err = 0;
2611                 } else
2612                         err = fuse_getlk(file, fl);
2613         } else {
2614                 if (fc->no_lock)
2615                         err = posix_lock_file(file, fl, NULL);
2616                 else
2617                         err = fuse_setlk(file, fl, 0);
2618         }
2619         return err;
2620 }
2621
2622 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2623 {
2624         struct inode *inode = file_inode(file);
2625         struct fuse_conn *fc = get_fuse_conn(inode);
2626         int err;
2627
2628         if (fc->no_flock) {
2629                 err = locks_lock_file_wait(file, fl);
2630         } else {
2631                 struct fuse_file *ff = file->private_data;
2632
2633                 /* emulate flock with POSIX locks */
2634                 ff->flock = true;
2635                 err = fuse_setlk(file, fl, 1);
2636         }
2637
2638         return err;
2639 }
2640
2641 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2642 {
2643         struct inode *inode = mapping->host;
2644         struct fuse_mount *fm = get_fuse_mount(inode);
2645         FUSE_ARGS(args);
2646         struct fuse_bmap_in inarg;
2647         struct fuse_bmap_out outarg;
2648         int err;
2649
2650         if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2651                 return 0;
2652
2653         memset(&inarg, 0, sizeof(inarg));
2654         inarg.block = block;
2655         inarg.blocksize = inode->i_sb->s_blocksize;
2656         args.opcode = FUSE_BMAP;
2657         args.nodeid = get_node_id(inode);
2658         args.in_numargs = 1;
2659         args.in_args[0].size = sizeof(inarg);
2660         args.in_args[0].value = &inarg;
2661         args.out_numargs = 1;
2662         args.out_args[0].size = sizeof(outarg);
2663         args.out_args[0].value = &outarg;
2664         err = fuse_simple_request(fm, &args);
2665         if (err == -ENOSYS)
2666                 fm->fc->no_bmap = 1;
2667
2668         return err ? 0 : outarg.block;
2669 }
2670
2671 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2672 {
2673         struct inode *inode = file->f_mapping->host;
2674         struct fuse_mount *fm = get_fuse_mount(inode);
2675         struct fuse_file *ff = file->private_data;
2676         FUSE_ARGS(args);
2677         struct fuse_lseek_in inarg = {
2678                 .fh = ff->fh,
2679                 .offset = offset,
2680                 .whence = whence
2681         };
2682         struct fuse_lseek_out outarg;
2683         int err;
2684
2685         if (fm->fc->no_lseek)
2686                 goto fallback;
2687
2688         args.opcode = FUSE_LSEEK;
2689         args.nodeid = ff->nodeid;
2690         args.in_numargs = 1;
2691         args.in_args[0].size = sizeof(inarg);
2692         args.in_args[0].value = &inarg;
2693         args.out_numargs = 1;
2694         args.out_args[0].size = sizeof(outarg);
2695         args.out_args[0].value = &outarg;
2696         err = fuse_simple_request(fm, &args);
2697         if (err) {
2698                 if (err == -ENOSYS) {
2699                         fm->fc->no_lseek = 1;
2700                         goto fallback;
2701                 }
2702                 return err;
2703         }
2704
2705         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2706
2707 fallback:
2708         err = fuse_update_attributes(inode, file, STATX_SIZE);
2709         if (!err)
2710                 return generic_file_llseek(file, offset, whence);
2711         else
2712                 return err;
2713 }
2714
2715 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2716 {
2717         loff_t retval;
2718         struct inode *inode = file_inode(file);
2719
2720         switch (whence) {
2721         case SEEK_SET:
2722         case SEEK_CUR:
2723                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2724                 retval = generic_file_llseek(file, offset, whence);
2725                 break;
2726         case SEEK_END:
2727                 inode_lock(inode);
2728                 retval = fuse_update_attributes(inode, file, STATX_SIZE);
2729                 if (!retval)
2730                         retval = generic_file_llseek(file, offset, whence);
2731                 inode_unlock(inode);
2732                 break;
2733         case SEEK_HOLE:
2734         case SEEK_DATA:
2735                 inode_lock(inode);
2736                 retval = fuse_lseek(file, offset, whence);
2737                 inode_unlock(inode);
2738                 break;
2739         default:
2740                 retval = -EINVAL;
2741         }
2742
2743         return retval;
2744 }
2745
2746 /*
2747  * All files which have been polled are linked to RB tree
2748  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2749  * find the matching one.
2750  */
2751 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2752                                               struct rb_node **parent_out)
2753 {
2754         struct rb_node **link = &fc->polled_files.rb_node;
2755         struct rb_node *last = NULL;
2756
2757         while (*link) {
2758                 struct fuse_file *ff;
2759
2760                 last = *link;
2761                 ff = rb_entry(last, struct fuse_file, polled_node);
2762
2763                 if (kh < ff->kh)
2764                         link = &last->rb_left;
2765                 else if (kh > ff->kh)
2766                         link = &last->rb_right;
2767                 else
2768                         return link;
2769         }
2770
2771         if (parent_out)
2772                 *parent_out = last;
2773         return link;
2774 }
2775
2776 /*
2777  * The file is about to be polled.  Make sure it's on the polled_files
2778  * RB tree.  Note that files once added to the polled_files tree are
2779  * not removed before the file is released.  This is because a file
2780  * polled once is likely to be polled again.
2781  */
2782 static void fuse_register_polled_file(struct fuse_conn *fc,
2783                                       struct fuse_file *ff)
2784 {
2785         spin_lock(&fc->lock);
2786         if (RB_EMPTY_NODE(&ff->polled_node)) {
2787                 struct rb_node **link, *parent;
2788
2789                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2790                 BUG_ON(*link);
2791                 rb_link_node(&ff->polled_node, parent, link);
2792                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2793         }
2794         spin_unlock(&fc->lock);
2795 }
2796
2797 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2798 {
2799         struct fuse_file *ff = file->private_data;
2800         struct fuse_mount *fm = ff->fm;
2801         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2802         struct fuse_poll_out outarg;
2803         FUSE_ARGS(args);
2804         int err;
2805
2806         if (fm->fc->no_poll)
2807                 return DEFAULT_POLLMASK;
2808
2809         poll_wait(file, &ff->poll_wait, wait);
2810         inarg.events = mangle_poll(poll_requested_events(wait));
2811
2812         /*
2813          * Ask for notification iff there's someone waiting for it.
2814          * The client may ignore the flag and always notify.
2815          */
2816         if (waitqueue_active(&ff->poll_wait)) {
2817                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2818                 fuse_register_polled_file(fm->fc, ff);
2819         }
2820
2821         args.opcode = FUSE_POLL;
2822         args.nodeid = ff->nodeid;
2823         args.in_numargs = 1;
2824         args.in_args[0].size = sizeof(inarg);
2825         args.in_args[0].value = &inarg;
2826         args.out_numargs = 1;
2827         args.out_args[0].size = sizeof(outarg);
2828         args.out_args[0].value = &outarg;
2829         err = fuse_simple_request(fm, &args);
2830
2831         if (!err)
2832                 return demangle_poll(outarg.revents);
2833         if (err == -ENOSYS) {
2834                 fm->fc->no_poll = 1;
2835                 return DEFAULT_POLLMASK;
2836         }
2837         return EPOLLERR;
2838 }
2839 EXPORT_SYMBOL_GPL(fuse_file_poll);
2840
2841 /*
2842  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2843  * wakes up the poll waiters.
2844  */
2845 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2846                             struct fuse_notify_poll_wakeup_out *outarg)
2847 {
2848         u64 kh = outarg->kh;
2849         struct rb_node **link;
2850
2851         spin_lock(&fc->lock);
2852
2853         link = fuse_find_polled_node(fc, kh, NULL);
2854         if (*link) {
2855                 struct fuse_file *ff;
2856
2857                 ff = rb_entry(*link, struct fuse_file, polled_node);
2858                 wake_up_interruptible_sync(&ff->poll_wait);
2859         }
2860
2861         spin_unlock(&fc->lock);
2862         return 0;
2863 }
2864
2865 static void fuse_do_truncate(struct file *file)
2866 {
2867         struct inode *inode = file->f_mapping->host;
2868         struct iattr attr;
2869
2870         attr.ia_valid = ATTR_SIZE;
2871         attr.ia_size = i_size_read(inode);
2872
2873         attr.ia_file = file;
2874         attr.ia_valid |= ATTR_FILE;
2875
2876         fuse_do_setattr(file_dentry(file), &attr, file);
2877 }
2878
2879 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2880 {
2881         return round_up(off, fc->max_pages << PAGE_SHIFT);
2882 }
2883
2884 static ssize_t
2885 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2886 {
2887         DECLARE_COMPLETION_ONSTACK(wait);
2888         ssize_t ret = 0;
2889         struct file *file = iocb->ki_filp;
2890         struct fuse_file *ff = file->private_data;
2891         loff_t pos = 0;
2892         struct inode *inode;
2893         loff_t i_size;
2894         size_t count = iov_iter_count(iter), shortened = 0;
2895         loff_t offset = iocb->ki_pos;
2896         struct fuse_io_priv *io;
2897
2898         pos = offset;
2899         inode = file->f_mapping->host;
2900         i_size = i_size_read(inode);
2901
2902         if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
2903                 return 0;
2904
2905         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2906         if (!io)
2907                 return -ENOMEM;
2908         spin_lock_init(&io->lock);
2909         kref_init(&io->refcnt);
2910         io->reqs = 1;
2911         io->bytes = -1;
2912         io->size = 0;
2913         io->offset = offset;
2914         io->write = (iov_iter_rw(iter) == WRITE);
2915         io->err = 0;
2916         /*
2917          * By default, we want to optimize all I/Os with async request
2918          * submission to the client filesystem if supported.
2919          */
2920         io->async = ff->fm->fc->async_dio;
2921         io->iocb = iocb;
2922         io->blocking = is_sync_kiocb(iocb);
2923
2924         /* optimization for short read */
2925         if (io->async && !io->write && offset + count > i_size) {
2926                 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
2927                 shortened = count - iov_iter_count(iter);
2928                 count -= shortened;
2929         }
2930
2931         /*
2932          * We cannot asynchronously extend the size of a file.
2933          * In such case the aio will behave exactly like sync io.
2934          */
2935         if ((offset + count > i_size) && io->write)
2936                 io->blocking = true;
2937
2938         if (io->async && io->blocking) {
2939                 /*
2940                  * Additional reference to keep io around after
2941                  * calling fuse_aio_complete()
2942                  */
2943                 kref_get(&io->refcnt);
2944                 io->done = &wait;
2945         }
2946
2947         if (iov_iter_rw(iter) == WRITE) {
2948                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2949                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
2950         } else {
2951                 ret = __fuse_direct_read(io, iter, &pos);
2952         }
2953         iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
2954
2955         if (io->async) {
2956                 bool blocking = io->blocking;
2957
2958                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2959
2960                 /* we have a non-extending, async request, so return */
2961                 if (!blocking)
2962                         return -EIOCBQUEUED;
2963
2964                 wait_for_completion(&wait);
2965                 ret = fuse_get_res_by_io(io);
2966         }
2967
2968         kref_put(&io->refcnt, fuse_io_release);
2969
2970         if (iov_iter_rw(iter) == WRITE) {
2971                 fuse_write_update_attr(inode, pos, ret);
2972                 /* For extending writes we already hold exclusive lock */
2973                 if (ret < 0 && offset + count > i_size)
2974                         fuse_do_truncate(file);
2975         }
2976
2977         return ret;
2978 }
2979
2980 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
2981 {
2982         int err = filemap_write_and_wait_range(inode->i_mapping, start, LLONG_MAX);
2983
2984         if (!err)
2985                 fuse_sync_writes(inode);
2986
2987         return err;
2988 }
2989
2990 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2991                                 loff_t length)
2992 {
2993         struct fuse_file *ff = file->private_data;
2994         struct inode *inode = file_inode(file);
2995         struct fuse_inode *fi = get_fuse_inode(inode);
2996         struct fuse_mount *fm = ff->fm;
2997         FUSE_ARGS(args);
2998         struct fuse_fallocate_in inarg = {
2999                 .fh = ff->fh,
3000                 .offset = offset,
3001                 .length = length,
3002                 .mode = mode
3003         };
3004         int err;
3005         bool block_faults = FUSE_IS_DAX(inode) &&
3006                 (!(mode & FALLOC_FL_KEEP_SIZE) ||
3007                  (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)));
3008
3009         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3010                      FALLOC_FL_ZERO_RANGE))
3011                 return -EOPNOTSUPP;
3012
3013         if (fm->fc->no_fallocate)
3014                 return -EOPNOTSUPP;
3015
3016         inode_lock(inode);
3017         if (block_faults) {
3018                 filemap_invalidate_lock(inode->i_mapping);
3019                 err = fuse_dax_break_layouts(inode, 0, 0);
3020                 if (err)
3021                         goto out;
3022         }
3023
3024         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)) {
3025                 loff_t endbyte = offset + length - 1;
3026
3027                 err = fuse_writeback_range(inode, offset, endbyte);
3028                 if (err)
3029                         goto out;
3030         }
3031
3032         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3033             offset + length > i_size_read(inode)) {
3034                 err = inode_newsize_ok(inode, offset + length);
3035                 if (err)
3036                         goto out;
3037         }
3038
3039         err = file_modified(file);
3040         if (err)
3041                 goto out;
3042
3043         if (!(mode & FALLOC_FL_KEEP_SIZE))
3044                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3045
3046         args.opcode = FUSE_FALLOCATE;
3047         args.nodeid = ff->nodeid;
3048         args.in_numargs = 1;
3049         args.in_args[0].size = sizeof(inarg);
3050         args.in_args[0].value = &inarg;
3051         err = fuse_simple_request(fm, &args);
3052         if (err == -ENOSYS) {
3053                 fm->fc->no_fallocate = 1;
3054                 err = -EOPNOTSUPP;
3055         }
3056         if (err)
3057                 goto out;
3058
3059         /* we could have extended the file */
3060         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3061                 if (fuse_write_update_attr(inode, offset + length, length))
3062                         file_update_time(file);
3063         }
3064
3065         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE))
3066                 truncate_pagecache_range(inode, offset, offset + length - 1);
3067
3068         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
3069
3070 out:
3071         if (!(mode & FALLOC_FL_KEEP_SIZE))
3072                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3073
3074         if (block_faults)
3075                 filemap_invalidate_unlock(inode->i_mapping);
3076
3077         inode_unlock(inode);
3078
3079         fuse_flush_time_update(inode);
3080
3081         return err;
3082 }
3083
3084 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3085                                       struct file *file_out, loff_t pos_out,
3086                                       size_t len, unsigned int flags)
3087 {
3088         struct fuse_file *ff_in = file_in->private_data;
3089         struct fuse_file *ff_out = file_out->private_data;
3090         struct inode *inode_in = file_inode(file_in);
3091         struct inode *inode_out = file_inode(file_out);
3092         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3093         struct fuse_mount *fm = ff_in->fm;
3094         struct fuse_conn *fc = fm->fc;
3095         FUSE_ARGS(args);
3096         struct fuse_copy_file_range_in inarg = {
3097                 .fh_in = ff_in->fh,
3098                 .off_in = pos_in,
3099                 .nodeid_out = ff_out->nodeid,
3100                 .fh_out = ff_out->fh,
3101                 .off_out = pos_out,
3102                 .len = len,
3103                 .flags = flags
3104         };
3105         struct fuse_write_out outarg;
3106         ssize_t err;
3107         /* mark unstable when write-back is not used, and file_out gets
3108          * extended */
3109         bool is_unstable = (!fc->writeback_cache) &&
3110                            ((pos_out + len) > inode_out->i_size);
3111
3112         if (fc->no_copy_file_range)
3113                 return -EOPNOTSUPP;
3114
3115         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3116                 return -EXDEV;
3117
3118         inode_lock(inode_in);
3119         err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3120         inode_unlock(inode_in);
3121         if (err)
3122                 return err;
3123
3124         inode_lock(inode_out);
3125
3126         err = file_modified(file_out);
3127         if (err)
3128                 goto out;
3129
3130         /*
3131          * Write out dirty pages in the destination file before sending the COPY
3132          * request to userspace.  After the request is completed, truncate off
3133          * pages (including partial ones) from the cache that have been copied,
3134          * since these contain stale data at that point.
3135          *
3136          * This should be mostly correct, but if the COPY writes to partial
3137          * pages (at the start or end) and the parts not covered by the COPY are
3138          * written through a memory map after calling fuse_writeback_range(),
3139          * then these partial page modifications will be lost on truncation.
3140          *
3141          * It is unlikely that someone would rely on such mixed style
3142          * modifications.  Yet this does give less guarantees than if the
3143          * copying was performed with write(2).
3144          *
3145          * To fix this a mapping->invalidate_lock could be used to prevent new
3146          * faults while the copy is ongoing.
3147          */
3148         err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3149         if (err)
3150                 goto out;
3151
3152         if (is_unstable)
3153                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3154
3155         args.opcode = FUSE_COPY_FILE_RANGE;
3156         args.nodeid = ff_in->nodeid;
3157         args.in_numargs = 1;
3158         args.in_args[0].size = sizeof(inarg);
3159         args.in_args[0].value = &inarg;
3160         args.out_numargs = 1;
3161         args.out_args[0].size = sizeof(outarg);
3162         args.out_args[0].value = &outarg;
3163         err = fuse_simple_request(fm, &args);
3164         if (err == -ENOSYS) {
3165                 fc->no_copy_file_range = 1;
3166                 err = -EOPNOTSUPP;
3167         }
3168         if (err)
3169                 goto out;
3170
3171         truncate_inode_pages_range(inode_out->i_mapping,
3172                                    ALIGN_DOWN(pos_out, PAGE_SIZE),
3173                                    ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3174
3175         file_update_time(file_out);
3176         fuse_write_update_attr(inode_out, pos_out + outarg.size, outarg.size);
3177
3178         err = outarg.size;
3179 out:
3180         if (is_unstable)
3181                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3182
3183         inode_unlock(inode_out);
3184         file_accessed(file_in);
3185
3186         fuse_flush_time_update(inode_out);
3187
3188         return err;
3189 }
3190
3191 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3192                                     struct file *dst_file, loff_t dst_off,
3193                                     size_t len, unsigned int flags)
3194 {
3195         ssize_t ret;
3196
3197         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3198                                      len, flags);
3199
3200         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3201                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3202                                               dst_off, len, flags);
3203         return ret;
3204 }
3205
3206 static const struct file_operations fuse_file_operations = {
3207         .llseek         = fuse_file_llseek,
3208         .read_iter      = fuse_file_read_iter,
3209         .write_iter     = fuse_file_write_iter,
3210         .mmap           = fuse_file_mmap,
3211         .open           = fuse_open,
3212         .flush          = fuse_flush,
3213         .release        = fuse_release,
3214         .fsync          = fuse_fsync,
3215         .lock           = fuse_file_lock,
3216         .get_unmapped_area = thp_get_unmapped_area,
3217         .flock          = fuse_file_flock,
3218         .splice_read    = generic_file_splice_read,
3219         .splice_write   = iter_file_splice_write,
3220         .unlocked_ioctl = fuse_file_ioctl,
3221         .compat_ioctl   = fuse_file_compat_ioctl,
3222         .poll           = fuse_file_poll,
3223         .fallocate      = fuse_file_fallocate,
3224         .copy_file_range = fuse_copy_file_range,
3225 };
3226
3227 static const struct address_space_operations fuse_file_aops  = {
3228         .read_folio     = fuse_read_folio,
3229         .readahead      = fuse_readahead,
3230         .writepage      = fuse_writepage,
3231         .writepages     = fuse_writepages,
3232         .launder_folio  = fuse_launder_folio,
3233         .dirty_folio    = filemap_dirty_folio,
3234         .bmap           = fuse_bmap,
3235         .direct_IO      = fuse_direct_IO,
3236         .write_begin    = fuse_write_begin,
3237         .write_end      = fuse_write_end,
3238 };
3239
3240 void fuse_init_file_inode(struct inode *inode, unsigned int flags)
3241 {
3242         struct fuse_inode *fi = get_fuse_inode(inode);
3243
3244         inode->i_fop = &fuse_file_operations;
3245         inode->i_data.a_ops = &fuse_file_aops;
3246
3247         INIT_LIST_HEAD(&fi->write_files);
3248         INIT_LIST_HEAD(&fi->queued_writes);
3249         fi->writectr = 0;
3250         init_waitqueue_head(&fi->page_waitq);
3251         fi->writepages = RB_ROOT;
3252
3253         if (IS_ENABLED(CONFIG_FUSE_DAX))
3254                 fuse_dax_inode_init(inode, flags);
3255 }