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 #include <linux/file.h>
  23
  24 static int fuse_send_open(struct fuse_mount *fm, u64 nodeid,
  25                           unsigned int open_flags, int opcode,
  26                           struct fuse_open_out *outargp)
  27 {
  28         struct fuse_open_in inarg;
  29         FUSE_ARGS(args);
  30
  31         memset(&inarg, 0, sizeof(inarg));
  32         inarg.flags = open_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  33         if (!fm->fc->atomic_o_trunc)
  34                 inarg.flags &= ~O_TRUNC;
  35
  36         if (fm->fc->handle_killpriv_v2 &&
  37             (inarg.flags & O_TRUNC) && !capable(CAP_FSETID)) {
  38                 inarg.open_flags |= FUSE_OPEN_KILL_SUIDGID;
  39         }
  40
  41         args.opcode = opcode;
  42         args.nodeid = nodeid;
  43         args.in_numargs = 1;
  44         args.in_args[0].size = sizeof(inarg);
  45         args.in_args[0].value = &inarg;
  46         args.out_numargs = 1;
  47         args.out_args[0].size = sizeof(*outargp);
  48         args.out_args[0].value = outargp;
  49
  50         return fuse_simple_request(fm, &args);
  51 }
  52
  53 struct fuse_release_args {
  54         struct fuse_args args;
  55         struct fuse_release_in inarg;
  56         struct inode *inode;
  57 };
  58
  59 struct fuse_file *fuse_file_alloc(struct fuse_mount *fm)
  60 {
  61         struct fuse_file *ff;
  62
  63         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
  64         if (unlikely(!ff))
  65                 return NULL;
  66
  67         ff->fm = fm;
  68         ff->release_args = kzalloc(sizeof(*ff->release_args),
  69                                    GFP_KERNEL_ACCOUNT);
  70         if (!ff->release_args) {
  71                 kfree(ff);
  72                 return NULL;
  73         }
  74
  75         INIT_LIST_HEAD(&ff->write_entry);
  76         mutex_init(&ff->readdir.lock);
  77         refcount_set(&ff->count, 1);
  78         RB_CLEAR_NODE(&ff->polled_node);
  79         init_waitqueue_head(&ff->poll_wait);
  80
  81         ff->kh = atomic64_inc_return(&fm->fc->khctr);
  82
  83         return ff;
  84 }
  85
  86 void fuse_file_free(struct fuse_file *ff)
  87 {
  88         kfree(ff->release_args);
  89         mutex_destroy(&ff->readdir.lock);
  90         kfree(ff);
  91 }
  92
  93 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  94 {
  95         refcount_inc(&ff->count);
  96         return ff;
  97 }
  98
  99 static void fuse_release_end(struct fuse_mount *fm, struct fuse_args *args,
 100                              int error)
 101 {
 102         struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
 103
 104         iput(ra->inode);
 105         kfree(ra);
 106 }
 107
 108 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
 109 {
 110         if (refcount_dec_and_test(&ff->count)) {
 111                 struct fuse_args *args = &ff->release_args->args;
 112
 113                 if (isdir ? ff->fm->fc->no_opendir : ff->fm->fc->no_open) {
 114                         /* Do nothing when client does not implement 'open' */
 115                         fuse_release_end(ff->fm, args, 0);
 116                 } else if (sync) {
 117                         fuse_simple_request(ff->fm, args);
 118                         fuse_release_end(ff->fm, args, 0);
 119                 } else {
 120                         args->end = fuse_release_end;
 121                         if (fuse_simple_background(ff->fm, args,
 122                                                    GFP_KERNEL | __GFP_NOFAIL))
 123                                 fuse_release_end(ff->fm, args, -ENOTCONN);
 124                 }
 125                 kfree(ff);
 126         }
 127 }
 128
 129 struct fuse_file *fuse_file_open(struct fuse_mount *fm, u64 nodeid,
 130                                  unsigned int open_flags, bool isdir)
 131 {
 132         struct fuse_conn *fc = fm->fc;
 133         struct fuse_file *ff;
 134         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 135
 136         ff = fuse_file_alloc(fm);
 137         if (!ff)
 138                 return ERR_PTR(-ENOMEM);
 139
 140         ff->fh = 0;
 141         /* Default for no-open */
 142         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
 143         if (isdir ? !fc->no_opendir : !fc->no_open) {
 144                 struct fuse_open_out outarg;
 145                 int err;
 146
 147                 err = fuse_send_open(fm, nodeid, open_flags, opcode, &outarg);
 148                 if (!err) {
 149                         ff->fh = outarg.fh;
 150                         ff->open_flags = outarg.open_flags;
 151
 152                 } else if (err != -ENOSYS) {
 153                         fuse_file_free(ff);
 154                         return ERR_PTR(err);
 155                 } else {
 156                         if (isdir)
 157                                 fc->no_opendir = 1;
 158                         else
 159                                 fc->no_open = 1;
 160                 }
 161         }
 162
 163         if (isdir)
 164                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 165
 166         ff->nodeid = nodeid;
 167
 168         return ff;
 169 }
 170
 171 int fuse_do_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
 172                  bool isdir)
 173 {
 174         struct fuse_file *ff = fuse_file_open(fm, nodeid, file->f_flags, isdir);
 175
 176         if (!IS_ERR(ff))
 177                 file->private_data = ff;
 178
 179         return PTR_ERR_OR_ZERO(ff);
 180 }
 181 EXPORT_SYMBOL_GPL(fuse_do_open);
 182
 183 static void fuse_link_write_file(struct file *file)
 184 {
 185         struct inode *inode = file_inode(file);
 186         struct fuse_inode *fi = get_fuse_inode(inode);
 187         struct fuse_file *ff = file->private_data;
 188         /*
 189          * file may be written through mmap, so chain it onto the
 190          * inodes's write_file list
 191          */
 192         spin_lock(&fi->lock);
 193         if (list_empty(&ff->write_entry))
 194                 list_add(&ff->write_entry, &fi->write_files);
 195         spin_unlock(&fi->lock);
 196 }
 197
 198 void fuse_finish_open(struct inode *inode, struct file *file)
 199 {
 200         struct fuse_file *ff = file->private_data;
 201         struct fuse_conn *fc = get_fuse_conn(inode);
 202
 203         if (ff->open_flags & FOPEN_STREAM)
 204                 stream_open(inode, file);
 205         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 206                 nonseekable_open(inode, file);
 207
 208         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 209                 struct fuse_inode *fi = get_fuse_inode(inode);
 210
 211                 spin_lock(&fi->lock);
 212                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 213                 i_size_write(inode, 0);
 214                 spin_unlock(&fi->lock);
 215                 file_update_time(file);
 216                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
 217         }
 218         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 219                 fuse_link_write_file(file);
 220 }
 221
 222 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 223 {
 224         struct fuse_mount *fm = get_fuse_mount(inode);
 225         struct fuse_conn *fc = fm->fc;
 226         int err;
 227         bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
 228                           fc->atomic_o_trunc &&
 229                           fc->writeback_cache;
 230         bool dax_truncate = (file->f_flags & O_TRUNC) &&
 231                           fc->atomic_o_trunc && FUSE_IS_DAX(inode);
 232
 233         if (fuse_is_bad(inode))
 234                 return -EIO;
 235
 236         err = generic_file_open(inode, file);
 237         if (err)
 238                 return err;
 239
 240         if (is_wb_truncate || dax_truncate)
 241                 inode_lock(inode);
 242
 243         if (dax_truncate) {
 244                 filemap_invalidate_lock(inode->i_mapping);
 245                 err = fuse_dax_break_layouts(inode, 0, 0);
 246                 if (err)
 247                         goto out_inode_unlock;
 248         }
 249
 250         if (is_wb_truncate || dax_truncate)
 251                 fuse_set_nowrite(inode);
 252
 253         err = fuse_do_open(fm, get_node_id(inode), file, isdir);
 254         if (!err)
 255                 fuse_finish_open(inode, file);
 256
 257         if (is_wb_truncate || dax_truncate)
 258                 fuse_release_nowrite(inode);
 259         if (!err) {
 260                 struct fuse_file *ff = file->private_data;
 261
 262                 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC))
 263                         truncate_pagecache(inode, 0);
 264                 else if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 265                         invalidate_inode_pages2(inode->i_mapping);
 266         }
 267         if (dax_truncate)
 268                 filemap_invalidate_unlock(inode->i_mapping);
 269 out_inode_unlock:
 270         if (is_wb_truncate || dax_truncate)
 271                 inode_unlock(inode);
 272
 273         return err;
 274 }
 275
 276 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 277                                  unsigned int flags, int opcode)
 278 {
 279         struct fuse_conn *fc = ff->fm->fc;
 280         struct fuse_release_args *ra = ff->release_args;
 281
 282         /* Inode is NULL on error path of fuse_create_open() */
 283         if (likely(fi)) {
 284                 spin_lock(&fi->lock);
 285                 list_del(&ff->write_entry);
 286                 spin_unlock(&fi->lock);
 287         }
 288         spin_lock(&fc->lock);
 289         if (!RB_EMPTY_NODE(&ff->polled_node))
 290                 rb_erase(&ff->polled_node, &fc->polled_files);
 291         spin_unlock(&fc->lock);
 292
 293         wake_up_interruptible_all(&ff->poll_wait);
 294
 295         ra->inarg.fh = ff->fh;
 296         ra->inarg.flags = flags;
 297         ra->args.in_numargs = 1;
 298         ra->args.in_args[0].size = sizeof(struct fuse_release_in);
 299         ra->args.in_args[0].value = &ra->inarg;
 300         ra->args.opcode = opcode;
 301         ra->args.nodeid = ff->nodeid;
 302         ra->args.force = true;
 303         ra->args.nocreds = true;
 304 }
 305
 306 void fuse_file_release(struct inode *inode, struct fuse_file *ff,
 307                        unsigned int open_flags, fl_owner_t id, bool isdir)
 308 {
 309         struct fuse_inode *fi = get_fuse_inode(inode);
 310         struct fuse_release_args *ra = ff->release_args;
 311         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 312
 313         fuse_prepare_release(fi, ff, open_flags, opcode);
 314
 315         if (ff->flock) {
 316                 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 317                 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fm->fc, id);
 318         }
 319         /* Hold inode until release is finished */
 320         ra->inode = igrab(inode);
 321
 322         /*
 323          * Normally this will send the RELEASE request, however if
 324          * some asynchronous READ or WRITE requests are outstanding,
 325          * the sending will be delayed.
 326          *
 327          * Make the release synchronous if this is a fuseblk mount,
 328          * synchronous RELEASE is allowed (and desirable) in this case
 329          * because the server can be trusted not to screw up.
 330          */
 331         fuse_file_put(ff, ff->fm->fc->destroy, isdir);
 332 }
 333
 334 void fuse_release_common(struct file *file, bool isdir)
 335 {
 336         fuse_file_release(file_inode(file), file->private_data, file->f_flags,
 337                           (fl_owner_t) file, isdir);
 338 }
 339
 340 static int fuse_open(struct inode *inode, struct file *file)
 341 {
 342         return fuse_open_common(inode, file, false);
 343 }
 344
 345 static int fuse_release(struct inode *inode, struct file *file)
 346 {
 347         struct fuse_conn *fc = get_fuse_conn(inode);
 348
 349         /*
 350          * Dirty pages might remain despite write_inode_now() call from
 351          * fuse_flush() due to writes racing with the close.
 352          */
 353         if (fc->writeback_cache)
 354                 write_inode_now(inode, 1);
 355
 356         fuse_release_common(file, false);
 357
 358         /* return value is ignored by VFS */
 359         return 0;
 360 }
 361
 362 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff,
 363                        unsigned int flags)
 364 {
 365         WARN_ON(refcount_read(&ff->count) > 1);
 366         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 367         /*
 368          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 369          * synchronous, we are fine with not doing igrab() here"
 370          */
 371         fuse_file_put(ff, true, false);
 372 }
 373 EXPORT_SYMBOL_GPL(fuse_sync_release);
 374
 375 /*
 376  * Scramble the ID space with XTEA, so that the value of the files_struct
 377  * pointer is not exposed to userspace.
 378  */
 379 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 380 {
 381         u32 *k = fc->scramble_key;
 382         u64 v = (unsigned long) id;
 383         u32 v0 = v;
 384         u32 v1 = v >> 32;
 385         u32 sum = 0;
 386         int i;
 387
 388         for (i = 0; i < 32; i++) {
 389                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 390                 sum += 0x9E3779B9;
 391                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 392         }
 393
 394         return (u64) v0 + ((u64) v1 << 32);
 395 }
 396
 397 struct fuse_writepage_args {
 398         struct fuse_io_args ia;
 399         struct rb_node writepages_entry;
 400         struct list_head queue_entry;
 401         struct fuse_writepage_args *next;
 402         struct inode *inode;
 403         struct fuse_sync_bucket *bucket;
 404 };
 405
 406 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
 407                                             pgoff_t idx_from, pgoff_t idx_to)
 408 {
 409         struct rb_node *n;
 410
 411         n = fi->writepages.rb_node;
 412
 413         while (n) {
 414                 struct fuse_writepage_args *wpa;
 415                 pgoff_t curr_index;
 416
 417                 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
 418                 WARN_ON(get_fuse_inode(wpa->inode) != fi);
 419                 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
 420                 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
 421                         n = n->rb_right;
 422                 else if (idx_to < curr_index)
 423                         n = n->rb_left;
 424                 else
 425                         return wpa;
 426         }
 427         return NULL;
 428 }
 429
 430 /*
 431  * Check if any page in a range is under writeback
 432  *
 433  * This is currently done by walking the list of writepage requests
 434  * for the inode, which can be pretty inefficient.
 435  */
 436 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 437                                    pgoff_t idx_to)
 438 {
 439         struct fuse_inode *fi = get_fuse_inode(inode);
 440         bool found;
 441
 442         spin_lock(&fi->lock);
 443         found = fuse_find_writeback(fi, idx_from, idx_to);
 444         spin_unlock(&fi->lock);
 445
 446         return found;
 447 }
 448
 449 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 450 {
 451         return fuse_range_is_writeback(inode, index, index);
 452 }
 453
 454 /*
 455  * Wait for page writeback to be completed.
 456  *
 457  * Since fuse doesn't rely on the VM writeback tracking, this has to
 458  * use some other means.
 459  */
 460 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 461 {
 462         struct fuse_inode *fi = get_fuse_inode(inode);
 463
 464         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 465 }
 466
 467 /*
 468  * Wait for all pending writepages on the inode to finish.
 469  *
 470  * This is currently done by blocking further writes with FUSE_NOWRITE
 471  * and waiting for all sent writes to complete.
 472  *
 473  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 474  * could conflict with truncation.
 475  */
 476 static void fuse_sync_writes(struct inode *inode)
 477 {
 478         fuse_set_nowrite(inode);
 479         fuse_release_nowrite(inode);
 480 }
 481
 482 struct fuse_flush_args {
 483         struct fuse_args args;
 484         struct fuse_flush_in inarg;
 485         struct work_struct work;
 486         struct file *file;
 487 };
 488
 489 static int fuse_do_flush(struct fuse_flush_args *fa)
 490 {
 491         int err;
 492         struct inode *inode = file_inode(fa->file);
 493         struct fuse_mount *fm = get_fuse_mount(inode);
 494
 495         err = write_inode_now(inode, 1);
 496         if (err)
 497                 goto out;
 498
 499         inode_lock(inode);
 500         fuse_sync_writes(inode);
 501         inode_unlock(inode);
 502
 503         err = filemap_check_errors(fa->file->f_mapping);
 504         if (err)
 505                 goto out;
 506
 507         err = 0;
 508         if (fm->fc->no_flush)
 509                 goto inval_attr_out;
 510
 511         err = fuse_simple_request(fm, &fa->args);
 512         if (err == -ENOSYS) {
 513                 fm->fc->no_flush = 1;
 514                 err = 0;
 515         }
 516
 517 inval_attr_out:
 518         /*
 519          * In memory i_blocks is not maintained by fuse, if writeback cache is
 520          * enabled, i_blocks from cached attr may not be accurate.
 521          */
 522         if (!err && fm->fc->writeback_cache)
 523                 fuse_invalidate_attr_mask(inode, STATX_BLOCKS);
 524
 525 out:
 526         fput(fa->file);
 527         kfree(fa);
 528         return err;
 529 }
 530
 531 static void fuse_flush_async(struct work_struct *work)
 532 {
 533         struct fuse_flush_args *fa = container_of(work, typeof(*fa), work);
 534
 535         fuse_do_flush(fa);
 536 }
 537
 538 static int fuse_flush(struct file *file, fl_owner_t id)
 539 {
 540         struct fuse_flush_args *fa;
 541         struct inode *inode = file_inode(file);
 542         struct fuse_mount *fm = get_fuse_mount(inode);
 543         struct fuse_file *ff = file->private_data;
 544
 545         if (fuse_is_bad(inode))
 546                 return -EIO;
 547
 548         if (ff->open_flags & FOPEN_NOFLUSH && !fm->fc->writeback_cache)
 549                 return 0;
 550
 551         fa = kzalloc(sizeof(*fa), GFP_KERNEL);
 552         if (!fa)
 553                 return -ENOMEM;
 554
 555         fa->inarg.fh = ff->fh;
 556         fa->inarg.lock_owner = fuse_lock_owner_id(fm->fc, id);
 557         fa->args.opcode = FUSE_FLUSH;
 558         fa->args.nodeid = get_node_id(inode);
 559         fa->args.in_numargs = 1;
 560         fa->args.in_args[0].size = sizeof(fa->inarg);
 561         fa->args.in_args[0].value = &fa->inarg;
 562         fa->args.force = true;
 563         fa->file = get_file(file);
 564
 565         /* Don't wait if the task is exiting */
 566         if (current->flags & PF_EXITING) {
 567                 INIT_WORK(&fa->work, fuse_flush_async);
 568                 schedule_work(&fa->work);
 569                 return 0;
 570         }
 571
 572         return fuse_do_flush(fa);
 573 }
 574
 575 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 576                       int datasync, int opcode)
 577 {
 578         struct inode *inode = file->f_mapping->host;
 579         struct fuse_mount *fm = get_fuse_mount(inode);
 580         struct fuse_file *ff = file->private_data;
 581         FUSE_ARGS(args);
 582         struct fuse_fsync_in inarg;
 583
 584         memset(&inarg, 0, sizeof(inarg));
 585         inarg.fh = ff->fh;
 586         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 587         args.opcode = opcode;
 588         args.nodeid = get_node_id(inode);
 589         args.in_numargs = 1;
 590         args.in_args[0].size = sizeof(inarg);
 591         args.in_args[0].value = &inarg;
 592         return fuse_simple_request(fm, &args);
 593 }
 594
 595 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 596                       int datasync)
 597 {
 598         struct inode *inode = file->f_mapping->host;
 599         struct fuse_conn *fc = get_fuse_conn(inode);
 600         int err;
 601
 602         if (fuse_is_bad(inode))
 603                 return -EIO;
 604
 605         inode_lock(inode);
 606
 607         /*
 608          * Start writeback against all dirty pages of the inode, then
 609          * wait for all outstanding writes, before sending the FSYNC
 610          * request.
 611          */
 612         err = file_write_and_wait_range(file, start, end);
 613         if (err)
 614                 goto out;
 615
 616         fuse_sync_writes(inode);
 617
 618         /*
 619          * Due to implementation of fuse writeback
 620          * file_write_and_wait_range() does not catch errors.
 621          * We have to do this directly after fuse_sync_writes()
 622          */
 623         err = file_check_and_advance_wb_err(file);
 624         if (err)
 625                 goto out;
 626
 627         err = sync_inode_metadata(inode, 1);
 628         if (err)
 629                 goto out;
 630
 631         if (fc->no_fsync)
 632                 goto out;
 633
 634         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 635         if (err == -ENOSYS) {
 636                 fc->no_fsync = 1;
 637                 err = 0;
 638         }
 639 out:
 640         inode_unlock(inode);
 641
 642         return err;
 643 }
 644
 645 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
 646                          size_t count, int opcode)
 647 {
 648         struct fuse_file *ff = file->private_data;
 649         struct fuse_args *args = &ia->ap.args;
 650
 651         ia->read.in.fh = ff->fh;
 652         ia->read.in.offset = pos;
 653         ia->read.in.size = count;
 654         ia->read.in.flags = file->f_flags;
 655         args->opcode = opcode;
 656         args->nodeid = ff->nodeid;
 657         args->in_numargs = 1;
 658         args->in_args[0].size = sizeof(ia->read.in);
 659         args->in_args[0].value = &ia->read.in;
 660         args->out_argvar = true;
 661         args->out_numargs = 1;
 662         args->out_args[0].size = count;
 663 }
 664
 665 static void fuse_release_user_pages(struct fuse_args_pages *ap,
 666                                     bool should_dirty)
 667 {
 668         unsigned int i;
 669
 670         for (i = 0; i < ap->num_pages; i++) {
 671                 if (should_dirty)
 672                         set_page_dirty_lock(ap->pages[i]);
 673                 put_page(ap->pages[i]);
 674         }
 675 }
 676
 677 static void fuse_io_release(struct kref *kref)
 678 {
 679         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 680 }
 681
 682 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 683 {
 684         if (io->err)
 685                 return io->err;
 686
 687         if (io->bytes >= 0 && io->write)
 688                 return -EIO;
 689
 690         return io->bytes < 0 ? io->size : io->bytes;
 691 }
 692
 693 /*
 694  * In case of short read, the caller sets 'pos' to the position of
 695  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 696  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 697  *
 698  * An example:
 699  * User requested DIO read of 64K. It was split into two 32K fuse requests,
 700  * both submitted asynchronously. The first of them was ACKed by userspace as
 701  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 702  * second request was ACKed as short, e.g. only 1K was read, resulting in
 703  * pos == 33K.
 704  *
 705  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 706  * will be equal to the length of the longest contiguous fragment of
 707  * transferred data starting from the beginning of IO request.
 708  */
 709 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 710 {
 711         int left;
 712
 713         spin_lock(&io->lock);
 714         if (err)
 715                 io->err = io->err ? : err;
 716         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 717                 io->bytes = pos;
 718
 719         left = --io->reqs;
 720         if (!left && io->blocking)
 721                 complete(io->done);
 722         spin_unlock(&io->lock);
 723
 724         if (!left && !io->blocking) {
 725                 ssize_t res = fuse_get_res_by_io(io);
 726
 727                 if (res >= 0) {
 728                         struct inode *inode = file_inode(io->iocb->ki_filp);
 729                         struct fuse_conn *fc = get_fuse_conn(inode);
 730                         struct fuse_inode *fi = get_fuse_inode(inode);
 731
 732                         spin_lock(&fi->lock);
 733                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 734                         spin_unlock(&fi->lock);
 735                 }
 736
 737                 io->iocb->ki_complete(io->iocb, res);
 738         }
 739
 740         kref_put(&io->refcnt, fuse_io_release);
 741 }
 742
 743 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
 744                                           unsigned int npages)
 745 {
 746         struct fuse_io_args *ia;
 747
 748         ia = kzalloc(sizeof(*ia), GFP_KERNEL);
 749         if (ia) {
 750                 ia->io = io;
 751                 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
 752                                                 &ia->ap.descs);
 753                 if (!ia->ap.pages) {
 754                         kfree(ia);
 755                         ia = NULL;
 756                 }
 757         }
 758         return ia;
 759 }
 760
 761 static void fuse_io_free(struct fuse_io_args *ia)
 762 {
 763         kfree(ia->ap.pages);
 764         kfree(ia);
 765 }
 766
 767 static void fuse_aio_complete_req(struct fuse_mount *fm, struct fuse_args *args,
 768                                   int err)
 769 {
 770         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 771         struct fuse_io_priv *io = ia->io;
 772         ssize_t pos = -1;
 773
 774         fuse_release_user_pages(&ia->ap, io->should_dirty);
 775
 776         if (err) {
 777                 /* Nothing */
 778         } else if (io->write) {
 779                 if (ia->write.out.size > ia->write.in.size) {
 780                         err = -EIO;
 781                 } else if (ia->write.in.size != ia->write.out.size) {
 782                         pos = ia->write.in.offset - io->offset +
 783                                 ia->write.out.size;
 784                 }
 785         } else {
 786                 u32 outsize = args->out_args[0].size;
 787
 788                 if (ia->read.in.size != outsize)
 789                         pos = ia->read.in.offset - io->offset + outsize;
 790         }
 791
 792         fuse_aio_complete(io, err, pos);
 793         fuse_io_free(ia);
 794 }
 795
 796 static ssize_t fuse_async_req_send(struct fuse_mount *fm,
 797                                    struct fuse_io_args *ia, size_t num_bytes)
 798 {
 799         ssize_t err;
 800         struct fuse_io_priv *io = ia->io;
 801
 802         spin_lock(&io->lock);
 803         kref_get(&io->refcnt);
 804         io->size += num_bytes;
 805         io->reqs++;
 806         spin_unlock(&io->lock);
 807
 808         ia->ap.args.end = fuse_aio_complete_req;
 809         ia->ap.args.may_block = io->should_dirty;
 810         err = fuse_simple_background(fm, &ia->ap.args, GFP_KERNEL);
 811         if (err)
 812                 fuse_aio_complete_req(fm, &ia->ap.args, err);
 813
 814         return num_bytes;
 815 }
 816
 817 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
 818                               fl_owner_t owner)
 819 {
 820         struct file *file = ia->io->iocb->ki_filp;
 821         struct fuse_file *ff = file->private_data;
 822         struct fuse_mount *fm = ff->fm;
 823
 824         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 825         if (owner != NULL) {
 826                 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
 827                 ia->read.in.lock_owner = fuse_lock_owner_id(fm->fc, owner);
 828         }
 829
 830         if (ia->io->async)
 831                 return fuse_async_req_send(fm, ia, count);
 832
 833         return fuse_simple_request(fm, &ia->ap.args);
 834 }
 835
 836 static void fuse_read_update_size(struct inode *inode, loff_t size,
 837                                   u64 attr_ver)
 838 {
 839         struct fuse_conn *fc = get_fuse_conn(inode);
 840         struct fuse_inode *fi = get_fuse_inode(inode);
 841
 842         spin_lock(&fi->lock);
 843         if (attr_ver >= fi->attr_version && size < inode->i_size &&
 844             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 845                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 846                 i_size_write(inode, size);
 847         }
 848         spin_unlock(&fi->lock);
 849 }
 850
 851 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
 852                             struct fuse_args_pages *ap)
 853 {
 854         struct fuse_conn *fc = get_fuse_conn(inode);
 855
 856         /*
 857          * If writeback_cache is enabled, a short read means there's a hole in
 858          * the file.  Some data after the hole is in page cache, but has not
 859          * reached the client fs yet.  So the hole is not present there.
 860          */
 861         if (!fc->writeback_cache) {
 862                 loff_t pos = page_offset(ap->pages[0]) + num_read;
 863                 fuse_read_update_size(inode, pos, attr_ver);
 864         }
 865 }
 866
 867 static int fuse_do_readpage(struct file *file, struct page *page)
 868 {
 869         struct inode *inode = page->mapping->host;
 870         struct fuse_mount *fm = get_fuse_mount(inode);
 871         loff_t pos = page_offset(page);
 872         struct fuse_page_desc desc = { .length = PAGE_SIZE };
 873         struct fuse_io_args ia = {
 874                 .ap.args.page_zeroing = true,
 875                 .ap.args.out_pages = true,
 876                 .ap.num_pages = 1,
 877                 .ap.pages = &page,
 878                 .ap.descs = &desc,
 879         };
 880         ssize_t res;
 881         u64 attr_ver;
 882
 883         /*
 884          * Page writeback can extend beyond the lifetime of the
 885          * page-cache page, so make sure we read a properly synced
 886          * page.
 887          */
 888         fuse_wait_on_page_writeback(inode, page->index);
 889
 890         attr_ver = fuse_get_attr_version(fm->fc);
 891
 892         /* Don't overflow end offset */
 893         if (pos + (desc.length - 1) == LLONG_MAX)
 894                 desc.length--;
 895
 896         fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
 897         res = fuse_simple_request(fm, &ia.ap.args);
 898         if (res < 0)
 899                 return res;
 900         /*
 901          * Short read means EOF.  If file size is larger, truncate it
 902          */
 903         if (res < desc.length)
 904                 fuse_short_read(inode, attr_ver, res, &ia.ap);
 905
 906         SetPageUptodate(page);
 907
 908         return 0;
 909 }
 910
 911 static int fuse_read_folio(struct file *file, struct folio *folio)
 912 {
 913         struct page *page = &folio->page;
 914         struct inode *inode = page->mapping->host;
 915         int err;
 916
 917         err = -EIO;
 918         if (fuse_is_bad(inode))
 919                 goto out;
 920
 921         err = fuse_do_readpage(file, page);
 922         fuse_invalidate_atime(inode);
 923  out:
 924         unlock_page(page);
 925         return err;
 926 }
 927
 928 static void fuse_readpages_end(struct fuse_mount *fm, struct fuse_args *args,
 929                                int err)
 930 {
 931         int i;
 932         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 933         struct fuse_args_pages *ap = &ia->ap;
 934         size_t count = ia->read.in.size;
 935         size_t num_read = args->out_args[0].size;
 936         struct address_space *mapping = NULL;
 937
 938         for (i = 0; mapping == NULL && i < ap->num_pages; i++)
 939                 mapping = ap->pages[i]->mapping;
 940
 941         if (mapping) {
 942                 struct inode *inode = mapping->host;
 943
 944                 /*
 945                  * Short read means EOF. If file size is larger, truncate it
 946                  */
 947                 if (!err && num_read < count)
 948                         fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
 949
 950                 fuse_invalidate_atime(inode);
 951         }
 952
 953         for (i = 0; i < ap->num_pages; i++) {
 954                 struct page *page = ap->pages[i];
 955
 956                 if (!err)
 957                         SetPageUptodate(page);
 958                 else
 959                         SetPageError(page);
 960                 unlock_page(page);
 961                 put_page(page);
 962         }
 963         if (ia->ff)
 964                 fuse_file_put(ia->ff, false, false);
 965
 966         fuse_io_free(ia);
 967 }
 968
 969 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
 970 {
 971         struct fuse_file *ff = file->private_data;
 972         struct fuse_mount *fm = ff->fm;
 973         struct fuse_args_pages *ap = &ia->ap;
 974         loff_t pos = page_offset(ap->pages[0]);
 975         size_t count = ap->num_pages << PAGE_SHIFT;
 976         ssize_t res;
 977         int err;
 978
 979         ap->args.out_pages = true;
 980         ap->args.page_zeroing = true;
 981         ap->args.page_replace = true;
 982
 983         /* Don't overflow end offset */
 984         if (pos + (count - 1) == LLONG_MAX) {
 985                 count--;
 986                 ap->descs[ap->num_pages - 1].length--;
 987         }
 988         WARN_ON((loff_t) (pos + count) < 0);
 989
 990         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 991         ia->read.attr_ver = fuse_get_attr_version(fm->fc);
 992         if (fm->fc->async_read) {
 993                 ia->ff = fuse_file_get(ff);
 994                 ap->args.end = fuse_readpages_end;
 995                 err = fuse_simple_background(fm, &ap->args, GFP_KERNEL);
 996                 if (!err)
 997                         return;
 998         } else {
 999                 res = fuse_simple_request(fm, &ap->args);
1000                 err = res < 0 ? res : 0;
1001         }
1002         fuse_readpages_end(fm, &ap->args, err);
1003 }
1004
1005 static void fuse_readahead(struct readahead_control *rac)
1006 {
1007         struct inode *inode = rac->mapping->host;
1008         struct fuse_conn *fc = get_fuse_conn(inode);
1009         unsigned int i, max_pages, nr_pages = 0;
1010
1011         if (fuse_is_bad(inode))
1012                 return;
1013
1014         max_pages = min_t(unsigned int, fc->max_pages,
1015                         fc->max_read / PAGE_SIZE);
1016
1017         for (;;) {
1018                 struct fuse_io_args *ia;
1019                 struct fuse_args_pages *ap;
1020
1021                 if (fc->num_background >= fc->congestion_threshold &&
1022                     rac->ra->async_size >= readahead_count(rac))
1023                         /*
1024                          * Congested and only async pages left, so skip the
1025                          * rest.
1026                          */
1027                         break;
1028
1029                 nr_pages = readahead_count(rac) - nr_pages;
1030                 if (nr_pages > max_pages)
1031                         nr_pages = max_pages;
1032                 if (nr_pages == 0)
1033                         break;
1034                 ia = fuse_io_alloc(NULL, nr_pages);
1035                 if (!ia)
1036                         return;
1037                 ap = &ia->ap;
1038                 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
1039                 for (i = 0; i < nr_pages; i++) {
1040                         fuse_wait_on_page_writeback(inode,
1041                                                     readahead_index(rac) + i);
1042                         ap->descs[i].length = PAGE_SIZE;
1043                 }
1044                 ap->num_pages = nr_pages;
1045                 fuse_send_readpages(ia, rac->file);
1046         }
1047 }
1048
1049 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
1050 {
1051         struct inode *inode = iocb->ki_filp->f_mapping->host;
1052         struct fuse_conn *fc = get_fuse_conn(inode);
1053
1054         /*
1055          * In auto invalidate mode, always update attributes on read.
1056          * Otherwise, only update if we attempt to read past EOF (to ensure
1057          * i_size is up to date).
1058          */
1059         if (fc->auto_inval_data ||
1060             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
1061                 int err;
1062                 err = fuse_update_attributes(inode, iocb->ki_filp, STATX_SIZE);
1063                 if (err)
1064                         return err;
1065         }
1066
1067         return generic_file_read_iter(iocb, to);
1068 }
1069
1070 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1071                                  loff_t pos, size_t count)
1072 {
1073         struct fuse_args *args = &ia->ap.args;
1074
1075         ia->write.in.fh = ff->fh;
1076         ia->write.in.offset = pos;
1077         ia->write.in.size = count;
1078         args->opcode = FUSE_WRITE;
1079         args->nodeid = ff->nodeid;
1080         args->in_numargs = 2;
1081         if (ff->fm->fc->minor < 9)
1082                 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1083         else
1084                 args->in_args[0].size = sizeof(ia->write.in);
1085         args->in_args[0].value = &ia->write.in;
1086         args->in_args[1].size = count;
1087         args->out_numargs = 1;
1088         args->out_args[0].size = sizeof(ia->write.out);
1089         args->out_args[0].value = &ia->write.out;
1090 }
1091
1092 static unsigned int fuse_write_flags(struct kiocb *iocb)
1093 {
1094         unsigned int flags = iocb->ki_filp->f_flags;
1095
1096         if (iocb_is_dsync(iocb))
1097                 flags |= O_DSYNC;
1098         if (iocb->ki_flags & IOCB_SYNC)
1099                 flags |= O_SYNC;
1100
1101         return flags;
1102 }
1103
1104 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1105                                size_t count, fl_owner_t owner)
1106 {
1107         struct kiocb *iocb = ia->io->iocb;
1108         struct file *file = iocb->ki_filp;
1109         struct fuse_file *ff = file->private_data;
1110         struct fuse_mount *fm = ff->fm;
1111         struct fuse_write_in *inarg = &ia->write.in;
1112         ssize_t err;
1113
1114         fuse_write_args_fill(ia, ff, pos, count);
1115         inarg->flags = fuse_write_flags(iocb);
1116         if (owner != NULL) {
1117                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1118                 inarg->lock_owner = fuse_lock_owner_id(fm->fc, owner);
1119         }
1120
1121         if (ia->io->async)
1122                 return fuse_async_req_send(fm, ia, count);
1123
1124         err = fuse_simple_request(fm, &ia->ap.args);
1125         if (!err && ia->write.out.size > count)
1126                 err = -EIO;
1127
1128         return err ?: ia->write.out.size;
1129 }
1130
1131 bool fuse_write_update_attr(struct inode *inode, loff_t pos, ssize_t written)
1132 {
1133         struct fuse_conn *fc = get_fuse_conn(inode);
1134         struct fuse_inode *fi = get_fuse_inode(inode);
1135         bool ret = false;
1136
1137         spin_lock(&fi->lock);
1138         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1139         if (written > 0 && pos > inode->i_size) {
1140                 i_size_write(inode, pos);
1141                 ret = true;
1142         }
1143         spin_unlock(&fi->lock);
1144
1145         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
1146
1147         return ret;
1148 }
1149
1150 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1151                                      struct kiocb *iocb, struct inode *inode,
1152                                      loff_t pos, size_t count)
1153 {
1154         struct fuse_args_pages *ap = &ia->ap;
1155         struct file *file = iocb->ki_filp;
1156         struct fuse_file *ff = file->private_data;
1157         struct fuse_mount *fm = ff->fm;
1158         unsigned int offset, i;
1159         bool short_write;
1160         int err;
1161
1162         for (i = 0; i < ap->num_pages; i++)
1163                 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1164
1165         fuse_write_args_fill(ia, ff, pos, count);
1166         ia->write.in.flags = fuse_write_flags(iocb);
1167         if (fm->fc->handle_killpriv_v2 && !capable(CAP_FSETID))
1168                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1169
1170         err = fuse_simple_request(fm, &ap->args);
1171         if (!err && ia->write.out.size > count)
1172                 err = -EIO;
1173
1174         short_write = ia->write.out.size < count;
1175         offset = ap->descs[0].offset;
1176         count = ia->write.out.size;
1177         for (i = 0; i < ap->num_pages; i++) {
1178                 struct page *page = ap->pages[i];
1179
1180                 if (err) {
1181                         ClearPageUptodate(page);
1182                 } else {
1183                         if (count >= PAGE_SIZE - offset)
1184                                 count -= PAGE_SIZE - offset;
1185                         else {
1186                                 if (short_write)
1187                                         ClearPageUptodate(page);
1188                                 count = 0;
1189                         }
1190                         offset = 0;
1191                 }
1192                 if (ia->write.page_locked && (i == ap->num_pages - 1))
1193                         unlock_page(page);
1194                 put_page(page);
1195         }
1196
1197         return err;
1198 }
1199
1200 static ssize_t fuse_fill_write_pages(struct fuse_io_args *ia,
1201                                      struct address_space *mapping,
1202                                      struct iov_iter *ii, loff_t pos,
1203                                      unsigned int max_pages)
1204 {
1205         struct fuse_args_pages *ap = &ia->ap;
1206         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1207         unsigned offset = pos & (PAGE_SIZE - 1);
1208         size_t count = 0;
1209         int err;
1210
1211         ap->args.in_pages = true;
1212         ap->descs[0].offset = offset;
1213
1214         do {
1215                 size_t tmp;
1216                 struct page *page;
1217                 pgoff_t index = pos >> PAGE_SHIFT;
1218                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1219                                      iov_iter_count(ii));
1220
1221                 bytes = min_t(size_t, bytes, fc->max_write - count);
1222
1223  again:
1224                 err = -EFAULT;
1225                 if (fault_in_iov_iter_readable(ii, bytes))
1226                         break;
1227
1228                 err = -ENOMEM;
1229                 page = grab_cache_page_write_begin(mapping, index);
1230                 if (!page)
1231                         break;
1232
1233                 if (mapping_writably_mapped(mapping))
1234                         flush_dcache_page(page);
1235
1236                 tmp = copy_page_from_iter_atomic(page, offset, bytes, ii);
1237                 flush_dcache_page(page);
1238
1239                 if (!tmp) {
1240                         unlock_page(page);
1241                         put_page(page);
1242                         goto again;
1243                 }
1244
1245                 err = 0;
1246                 ap->pages[ap->num_pages] = page;
1247                 ap->descs[ap->num_pages].length = tmp;
1248                 ap->num_pages++;
1249
1250                 count += tmp;
1251                 pos += tmp;
1252                 offset += tmp;
1253                 if (offset == PAGE_SIZE)
1254                         offset = 0;
1255
1256                 /* If we copied full page, mark it uptodate */
1257                 if (tmp == PAGE_SIZE)
1258                         SetPageUptodate(page);
1259
1260                 if (PageUptodate(page)) {
1261                         unlock_page(page);
1262                 } else {
1263                         ia->write.page_locked = true;
1264                         break;
1265                 }
1266                 if (!fc->big_writes)
1267                         break;
1268         } while (iov_iter_count(ii) && count < fc->max_write &&
1269                  ap->num_pages < max_pages && offset == 0);
1270
1271         return count > 0 ? count : err;
1272 }
1273
1274 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1275                                      unsigned int max_pages)
1276 {
1277         return min_t(unsigned int,
1278                      ((pos + len - 1) >> PAGE_SHIFT) -
1279                      (pos >> PAGE_SHIFT) + 1,
1280                      max_pages);
1281 }
1282
1283 static ssize_t fuse_perform_write(struct kiocb *iocb, struct iov_iter *ii)
1284 {
1285         struct address_space *mapping = iocb->ki_filp->f_mapping;
1286         struct inode *inode = mapping->host;
1287         struct fuse_conn *fc = get_fuse_conn(inode);
1288         struct fuse_inode *fi = get_fuse_inode(inode);
1289         loff_t pos = iocb->ki_pos;
1290         int err = 0;
1291         ssize_t res = 0;
1292
1293         if (inode->i_size < pos + iov_iter_count(ii))
1294                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1295
1296         do {
1297                 ssize_t count;
1298                 struct fuse_io_args ia = {};
1299                 struct fuse_args_pages *ap = &ia.ap;
1300                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1301                                                       fc->max_pages);
1302
1303                 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1304                 if (!ap->pages) {
1305                         err = -ENOMEM;
1306                         break;
1307                 }
1308
1309                 count = fuse_fill_write_pages(&ia, mapping, ii, pos, nr_pages);
1310                 if (count <= 0) {
1311                         err = count;
1312                 } else {
1313                         err = fuse_send_write_pages(&ia, iocb, inode,
1314                                                     pos, count);
1315                         if (!err) {
1316                                 size_t num_written = ia.write.out.size;
1317
1318                                 res += num_written;
1319                                 pos += num_written;
1320
1321                                 /* break out of the loop on short write */
1322                                 if (num_written != count)
1323                                         err = -EIO;
1324                         }
1325                 }
1326                 kfree(ap->pages);
1327         } while (!err && iov_iter_count(ii));
1328
1329         fuse_write_update_attr(inode, pos, res);
1330         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1331
1332         if (!res)
1333                 return err;
1334         iocb->ki_pos += res;
1335         return res;
1336 }
1337
1338 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1339 {
1340         struct file *file = iocb->ki_filp;
1341         struct address_space *mapping = file->f_mapping;
1342         ssize_t written = 0;
1343         struct inode *inode = mapping->host;
1344         ssize_t err;
1345         struct fuse_conn *fc = get_fuse_conn(inode);
1346
1347         if (fc->writeback_cache) {
1348                 /* Update size (EOF optimization) and mode (SUID clearing) */
1349                 err = fuse_update_attributes(mapping->host, file,
1350                                              STATX_SIZE | STATX_MODE);
1351                 if (err)
1352                         return err;
1353
1354                 if (fc->handle_killpriv_v2 &&
1355                     setattr_should_drop_suidgid(&nop_mnt_idmap,
1356                                                 file_inode(file))) {
1357                         goto writethrough;
1358                 }
1359
1360                 return generic_file_write_iter(iocb, from);
1361         }
1362
1363 writethrough:
1364         inode_lock(inode);
1365
1366         err = generic_write_checks(iocb, from);
1367         if (err <= 0)
1368                 goto out;
1369
1370         err = file_remove_privs(file);
1371         if (err)
1372                 goto out;
1373
1374         err = file_update_time(file);
1375         if (err)
1376                 goto out;
1377
1378         if (iocb->ki_flags & IOCB_DIRECT) {
1379                 written = generic_file_direct_write(iocb, from);
1380                 if (written < 0 || !iov_iter_count(from))
1381                         goto out;
1382                 written = direct_write_fallback(iocb, from, written,
1383                                 fuse_perform_write(iocb, from));
1384         } else {
1385                 written = fuse_perform_write(iocb, from);
1386         }
1387 out:
1388         inode_unlock(inode);
1389         if (written > 0)
1390                 written = generic_write_sync(iocb, written);
1391
1392         return written ? written : err;
1393 }
1394
1395 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1396 {
1397         return (unsigned long)iter_iov(ii)->iov_base + ii->iov_offset;
1398 }
1399
1400 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1401                                         size_t max_size)
1402 {
1403         return min(iov_iter_single_seg_count(ii), max_size);
1404 }
1405
1406 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1407                                size_t *nbytesp, int write,
1408                                unsigned int max_pages)
1409 {
1410         size_t nbytes = 0;  /* # bytes already packed in req */
1411         ssize_t ret = 0;
1412
1413         /* Special case for kernel I/O: can copy directly into the buffer */
1414         if (iov_iter_is_kvec(ii)) {
1415                 unsigned long user_addr = fuse_get_user_addr(ii);
1416                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1417
1418                 if (write)
1419                         ap->args.in_args[1].value = (void *) user_addr;
1420                 else
1421                         ap->args.out_args[0].value = (void *) user_addr;
1422
1423                 iov_iter_advance(ii, frag_size);
1424                 *nbytesp = frag_size;
1425                 return 0;
1426         }
1427
1428         while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1429                 unsigned npages;
1430                 size_t start;
1431                 ret = iov_iter_get_pages2(ii, &ap->pages[ap->num_pages],
1432                                         *nbytesp - nbytes,
1433                                         max_pages - ap->num_pages,
1434                                         &start);
1435                 if (ret < 0)
1436                         break;
1437
1438                 nbytes += ret;
1439
1440                 ret += start;
1441                 npages = DIV_ROUND_UP(ret, PAGE_SIZE);
1442
1443                 ap->descs[ap->num_pages].offset = start;
1444                 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1445
1446                 ap->num_pages += npages;
1447                 ap->descs[ap->num_pages - 1].length -=
1448                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1449         }
1450
1451         ap->args.user_pages = true;
1452         if (write)
1453                 ap->args.in_pages = true;
1454         else
1455                 ap->args.out_pages = true;
1456
1457         *nbytesp = nbytes;
1458
1459         return ret < 0 ? ret : 0;
1460 }
1461
1462 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1463                        loff_t *ppos, int flags)
1464 {
1465         int write = flags & FUSE_DIO_WRITE;
1466         int cuse = flags & FUSE_DIO_CUSE;
1467         struct file *file = io->iocb->ki_filp;
1468         struct inode *inode = file->f_mapping->host;
1469         struct fuse_file *ff = file->private_data;
1470         struct fuse_conn *fc = ff->fm->fc;
1471         size_t nmax = write ? fc->max_write : fc->max_read;
1472         loff_t pos = *ppos;
1473         size_t count = iov_iter_count(iter);
1474         pgoff_t idx_from = pos >> PAGE_SHIFT;
1475         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1476         ssize_t res = 0;
1477         int err = 0;
1478         struct fuse_io_args *ia;
1479         unsigned int max_pages;
1480
1481         max_pages = iov_iter_npages(iter, fc->max_pages);
1482         ia = fuse_io_alloc(io, max_pages);
1483         if (!ia)
1484                 return -ENOMEM;
1485
1486         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1487                 if (!write)
1488                         inode_lock(inode);
1489                 fuse_sync_writes(inode);
1490                 if (!write)
1491                         inode_unlock(inode);
1492         }
1493
1494         io->should_dirty = !write && user_backed_iter(iter);
1495         while (count) {
1496                 ssize_t nres;
1497                 fl_owner_t owner = current->files;
1498                 size_t nbytes = min(count, nmax);
1499
1500                 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1501                                           max_pages);
1502                 if (err && !nbytes)
1503                         break;
1504
1505                 if (write) {
1506                         if (!capable(CAP_FSETID))
1507                                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1508
1509                         nres = fuse_send_write(ia, pos, nbytes, owner);
1510                 } else {
1511                         nres = fuse_send_read(ia, pos, nbytes, owner);
1512                 }
1513
1514                 if (!io->async || nres < 0) {
1515                         fuse_release_user_pages(&ia->ap, io->should_dirty);
1516                         fuse_io_free(ia);
1517                 }
1518                 ia = NULL;
1519                 if (nres < 0) {
1520                         iov_iter_revert(iter, nbytes);
1521                         err = nres;
1522                         break;
1523                 }
1524                 WARN_ON(nres > nbytes);
1525
1526                 count -= nres;
1527                 res += nres;
1528                 pos += nres;
1529                 if (nres != nbytes) {
1530                         iov_iter_revert(iter, nbytes - nres);
1531                         break;
1532                 }
1533                 if (count) {
1534                         max_pages = iov_iter_npages(iter, fc->max_pages);
1535                         ia = fuse_io_alloc(io, max_pages);
1536                         if (!ia)
1537                                 break;
1538                 }
1539         }
1540         if (ia)
1541                 fuse_io_free(ia);
1542         if (res > 0)
1543                 *ppos = pos;
1544
1545         return res > 0 ? res : err;
1546 }
1547 EXPORT_SYMBOL_GPL(fuse_direct_io);
1548
1549 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1550                                   struct iov_iter *iter,
1551                                   loff_t *ppos)
1552 {
1553         ssize_t res;
1554         struct inode *inode = file_inode(io->iocb->ki_filp);
1555
1556         res = fuse_direct_io(io, iter, ppos, 0);
1557
1558         fuse_invalidate_atime(inode);
1559
1560         return res;
1561 }
1562
1563 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1564
1565 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1566 {
1567         ssize_t res;
1568
1569         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1570                 res = fuse_direct_IO(iocb, to);
1571         } else {
1572                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1573
1574                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1575         }
1576
1577         return res;
1578 }
1579
1580 static bool fuse_direct_write_extending_i_size(struct kiocb *iocb,
1581                                                struct iov_iter *iter)
1582 {
1583         struct inode *inode = file_inode(iocb->ki_filp);
1584
1585         return iocb->ki_pos + iov_iter_count(iter) > i_size_read(inode);
1586 }
1587
1588 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1589 {
1590         struct inode *inode = file_inode(iocb->ki_filp);
1591         struct file *file = iocb->ki_filp;
1592         struct fuse_file *ff = file->private_data;
1593         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1594         ssize_t res;
1595         bool exclusive_lock =
1596                 !(ff->open_flags & FOPEN_PARALLEL_DIRECT_WRITES) ||
1597                 iocb->ki_flags & IOCB_APPEND ||
1598                 fuse_direct_write_extending_i_size(iocb, from);
1599
1600         /*
1601          * Take exclusive lock if
1602          * - Parallel direct writes are disabled - a user space decision
1603          * - Parallel direct writes are enabled and i_size is being extended.
1604          *   This might not be needed at all, but needs further investigation.
1605          */
1606         if (exclusive_lock)
1607                 inode_lock(inode);
1608         else {
1609                 inode_lock_shared(inode);
1610
1611                 /* A race with truncate might have come up as the decision for
1612                  * the lock type was done without holding the lock, check again.
1613                  */
1614                 if (fuse_direct_write_extending_i_size(iocb, from)) {
1615                         inode_unlock_shared(inode);
1616                         inode_lock(inode);
1617                         exclusive_lock = true;
1618                 }
1619         }
1620
1621         res = generic_write_checks(iocb, from);
1622         if (res > 0) {
1623                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1624                         res = fuse_direct_IO(iocb, from);
1625                 } else {
1626                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1627                                              FUSE_DIO_WRITE);
1628                         fuse_write_update_attr(inode, iocb->ki_pos, res);
1629                 }
1630         }
1631         if (exclusive_lock)
1632                 inode_unlock(inode);
1633         else
1634                 inode_unlock_shared(inode);
1635
1636         return res;
1637 }
1638
1639 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1640 {
1641         struct file *file = iocb->ki_filp;
1642         struct fuse_file *ff = file->private_data;
1643         struct inode *inode = file_inode(file);
1644
1645         if (fuse_is_bad(inode))
1646                 return -EIO;
1647
1648         if (FUSE_IS_DAX(inode))
1649                 return fuse_dax_read_iter(iocb, to);
1650
1651         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1652                 return fuse_cache_read_iter(iocb, to);
1653         else
1654                 return fuse_direct_read_iter(iocb, to);
1655 }
1656
1657 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1658 {
1659         struct file *file = iocb->ki_filp;
1660         struct fuse_file *ff = file->private_data;
1661         struct inode *inode = file_inode(file);
1662
1663         if (fuse_is_bad(inode))
1664                 return -EIO;
1665
1666         if (FUSE_IS_DAX(inode))
1667                 return fuse_dax_write_iter(iocb, from);
1668
1669         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1670                 return fuse_cache_write_iter(iocb, from);
1671         else
1672                 return fuse_direct_write_iter(iocb, from);
1673 }
1674
1675 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1676 {
1677         struct fuse_args_pages *ap = &wpa->ia.ap;
1678         int i;
1679
1680         if (wpa->bucket)
1681                 fuse_sync_bucket_dec(wpa->bucket);
1682
1683         for (i = 0; i < ap->num_pages; i++)
1684                 __free_page(ap->pages[i]);
1685
1686         if (wpa->ia.ff)
1687                 fuse_file_put(wpa->ia.ff, false, false);
1688
1689         kfree(ap->pages);
1690         kfree(wpa);
1691 }
1692
1693 static void fuse_writepage_finish(struct fuse_mount *fm,
1694                                   struct fuse_writepage_args *wpa)
1695 {
1696         struct fuse_args_pages *ap = &wpa->ia.ap;
1697         struct inode *inode = wpa->inode;
1698         struct fuse_inode *fi = get_fuse_inode(inode);
1699         struct backing_dev_info *bdi = inode_to_bdi(inode);
1700         int i;
1701
1702         for (i = 0; i < ap->num_pages; i++) {
1703                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1704                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1705                 wb_writeout_inc(&bdi->wb);
1706         }
1707         wake_up(&fi->page_waitq);
1708 }
1709
1710 /* Called under fi->lock, may release and reacquire it */
1711 static void fuse_send_writepage(struct fuse_mount *fm,
1712                                 struct fuse_writepage_args *wpa, loff_t size)
1713 __releases(fi->lock)
1714 __acquires(fi->lock)
1715 {
1716         struct fuse_writepage_args *aux, *next;
1717         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1718         struct fuse_write_in *inarg = &wpa->ia.write.in;
1719         struct fuse_args *args = &wpa->ia.ap.args;
1720         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1721         int err;
1722
1723         fi->writectr++;
1724         if (inarg->offset + data_size <= size) {
1725                 inarg->size = data_size;
1726         } else if (inarg->offset < size) {
1727                 inarg->size = size - inarg->offset;
1728         } else {
1729                 /* Got truncated off completely */
1730                 goto out_free;
1731         }
1732
1733         args->in_args[1].size = inarg->size;
1734         args->force = true;
1735         args->nocreds = true;
1736
1737         err = fuse_simple_background(fm, args, GFP_ATOMIC);
1738         if (err == -ENOMEM) {
1739                 spin_unlock(&fi->lock);
1740                 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1741                 spin_lock(&fi->lock);
1742         }
1743
1744         /* Fails on broken connection only */
1745         if (unlikely(err))
1746                 goto out_free;
1747
1748         return;
1749
1750  out_free:
1751         fi->writectr--;
1752         rb_erase(&wpa->writepages_entry, &fi->writepages);
1753         fuse_writepage_finish(fm, wpa);
1754         spin_unlock(&fi->lock);
1755
1756         /* After fuse_writepage_finish() aux request list is private */
1757         for (aux = wpa->next; aux; aux = next) {
1758                 next = aux->next;
1759                 aux->next = NULL;
1760                 fuse_writepage_free(aux);
1761         }
1762
1763         fuse_writepage_free(wpa);
1764         spin_lock(&fi->lock);
1765 }
1766
1767 /*
1768  * If fi->writectr is positive (no truncate or fsync going on) send
1769  * all queued writepage requests.
1770  *
1771  * Called with fi->lock
1772  */
1773 void fuse_flush_writepages(struct inode *inode)
1774 __releases(fi->lock)
1775 __acquires(fi->lock)
1776 {
1777         struct fuse_mount *fm = get_fuse_mount(inode);
1778         struct fuse_inode *fi = get_fuse_inode(inode);
1779         loff_t crop = i_size_read(inode);
1780         struct fuse_writepage_args *wpa;
1781
1782         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1783                 wpa = list_entry(fi->queued_writes.next,
1784                                  struct fuse_writepage_args, queue_entry);
1785                 list_del_init(&wpa->queue_entry);
1786                 fuse_send_writepage(fm, wpa, crop);
1787         }
1788 }
1789
1790 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1791                                                 struct fuse_writepage_args *wpa)
1792 {
1793         pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1794         pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1795         struct rb_node **p = &root->rb_node;
1796         struct rb_node  *parent = NULL;
1797
1798         WARN_ON(!wpa->ia.ap.num_pages);
1799         while (*p) {
1800                 struct fuse_writepage_args *curr;
1801                 pgoff_t curr_index;
1802
1803                 parent = *p;
1804                 curr = rb_entry(parent, struct fuse_writepage_args,
1805                                 writepages_entry);
1806                 WARN_ON(curr->inode != wpa->inode);
1807                 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1808
1809                 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1810                         p = &(*p)->rb_right;
1811                 else if (idx_to < curr_index)
1812                         p = &(*p)->rb_left;
1813                 else
1814                         return curr;
1815         }
1816
1817         rb_link_node(&wpa->writepages_entry, parent, p);
1818         rb_insert_color(&wpa->writepages_entry, root);
1819         return NULL;
1820 }
1821
1822 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1823 {
1824         WARN_ON(fuse_insert_writeback(root, wpa));
1825 }
1826
1827 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1828                                int error)
1829 {
1830         struct fuse_writepage_args *wpa =
1831                 container_of(args, typeof(*wpa), ia.ap.args);
1832         struct inode *inode = wpa->inode;
1833         struct fuse_inode *fi = get_fuse_inode(inode);
1834         struct fuse_conn *fc = get_fuse_conn(inode);
1835
1836         mapping_set_error(inode->i_mapping, error);
1837         /*
1838          * A writeback finished and this might have updated mtime/ctime on
1839          * server making local mtime/ctime stale.  Hence invalidate attrs.
1840          * Do this only if writeback_cache is not enabled.  If writeback_cache
1841          * is enabled, we trust local ctime/mtime.
1842          */
1843         if (!fc->writeback_cache)
1844                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODIFY);
1845         spin_lock(&fi->lock);
1846         rb_erase(&wpa->writepages_entry, &fi->writepages);
1847         while (wpa->next) {
1848                 struct fuse_mount *fm = get_fuse_mount(inode);
1849                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1850                 struct fuse_writepage_args *next = wpa->next;
1851
1852                 wpa->next = next->next;
1853                 next->next = NULL;
1854                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1855                 tree_insert(&fi->writepages, next);
1856
1857                 /*
1858                  * Skip fuse_flush_writepages() to make it easy to crop requests
1859                  * based on primary request size.
1860                  *
1861                  * 1st case (trivial): there are no concurrent activities using
1862                  * fuse_set/release_nowrite.  Then we're on safe side because
1863                  * fuse_flush_writepages() would call fuse_send_writepage()
1864                  * anyway.
1865                  *
1866                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1867                  * now for completion of all in-flight requests.  This happens
1868                  * rarely and no more than once per page, so this should be
1869                  * okay.
1870                  *
1871                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1872                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1873                  * that fuse_set_nowrite returned implies that all in-flight
1874                  * requests were completed along with all of their secondary
1875                  * requests.  Further primary requests are blocked by negative
1876                  * writectr.  Hence there cannot be any in-flight requests and
1877                  * no invocations of fuse_writepage_end() while we're in
1878                  * fuse_set_nowrite..fuse_release_nowrite section.
1879                  */
1880                 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1881         }
1882         fi->writectr--;
1883         fuse_writepage_finish(fm, wpa);
1884         spin_unlock(&fi->lock);
1885         fuse_writepage_free(wpa);
1886 }
1887
1888 static struct fuse_file *__fuse_write_file_get(struct fuse_inode *fi)
1889 {
1890         struct fuse_file *ff;
1891
1892         spin_lock(&fi->lock);
1893         ff = list_first_entry_or_null(&fi->write_files, struct fuse_file,
1894                                       write_entry);
1895         if (ff)
1896                 fuse_file_get(ff);
1897         spin_unlock(&fi->lock);
1898
1899         return ff;
1900 }
1901
1902 static struct fuse_file *fuse_write_file_get(struct fuse_inode *fi)
1903 {
1904         struct fuse_file *ff = __fuse_write_file_get(fi);
1905         WARN_ON(!ff);
1906         return ff;
1907 }
1908
1909 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1910 {
1911         struct fuse_inode *fi = get_fuse_inode(inode);
1912         struct fuse_file *ff;
1913         int err;
1914
1915         /*
1916          * Inode is always written before the last reference is dropped and
1917          * hence this should not be reached from reclaim.
1918          *
1919          * Writing back the inode from reclaim can deadlock if the request
1920          * processing itself needs an allocation.  Allocations triggering
1921          * reclaim while serving a request can't be prevented, because it can
1922          * involve any number of unrelated userspace processes.
1923          */
1924         WARN_ON(wbc->for_reclaim);
1925
1926         ff = __fuse_write_file_get(fi);
1927         err = fuse_flush_times(inode, ff);
1928         if (ff)
1929                 fuse_file_put(ff, false, false);
1930
1931         return err;
1932 }
1933
1934 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1935 {
1936         struct fuse_writepage_args *wpa;
1937         struct fuse_args_pages *ap;
1938
1939         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1940         if (wpa) {
1941                 ap = &wpa->ia.ap;
1942                 ap->num_pages = 0;
1943                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1944                 if (!ap->pages) {
1945                         kfree(wpa);
1946                         wpa = NULL;
1947                 }
1948         }
1949         return wpa;
1950
1951 }
1952
1953 static void fuse_writepage_add_to_bucket(struct fuse_conn *fc,
1954                                          struct fuse_writepage_args *wpa)
1955 {
1956         if (!fc->sync_fs)
1957                 return;
1958
1959         rcu_read_lock();
1960         /* Prevent resurrection of dead bucket in unlikely race with syncfs */
1961         do {
1962                 wpa->bucket = rcu_dereference(fc->curr_bucket);
1963         } while (unlikely(!atomic_inc_not_zero(&wpa->bucket->count)));
1964         rcu_read_unlock();
1965 }
1966
1967 static int fuse_writepage_locked(struct page *page)
1968 {
1969         struct address_space *mapping = page->mapping;
1970         struct inode *inode = mapping->host;
1971         struct fuse_conn *fc = get_fuse_conn(inode);
1972         struct fuse_inode *fi = get_fuse_inode(inode);
1973         struct fuse_writepage_args *wpa;
1974         struct fuse_args_pages *ap;
1975         struct page *tmp_page;
1976         int error = -ENOMEM;
1977
1978         set_page_writeback(page);
1979
1980         wpa = fuse_writepage_args_alloc();
1981         if (!wpa)
1982                 goto err;
1983         ap = &wpa->ia.ap;
1984
1985         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1986         if (!tmp_page)
1987                 goto err_free;
1988
1989         error = -EIO;
1990         wpa->ia.ff = fuse_write_file_get(fi);
1991         if (!wpa->ia.ff)
1992                 goto err_nofile;
1993
1994         fuse_writepage_add_to_bucket(fc, wpa);
1995         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1996
1997         copy_highpage(tmp_page, page);
1998         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1999         wpa->next = NULL;
2000         ap->args.in_pages = true;
2001         ap->num_pages = 1;
2002         ap->pages[0] = tmp_page;
2003         ap->descs[0].offset = 0;
2004         ap->descs[0].length = PAGE_SIZE;
2005         ap->args.end = fuse_writepage_end;
2006         wpa->inode = inode;
2007
2008         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2009         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2010
2011         spin_lock(&fi->lock);
2012         tree_insert(&fi->writepages, wpa);
2013         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2014         fuse_flush_writepages(inode);
2015         spin_unlock(&fi->lock);
2016
2017         end_page_writeback(page);
2018
2019         return 0;
2020
2021 err_nofile:
2022         __free_page(tmp_page);
2023 err_free:
2024         kfree(wpa);
2025 err:
2026         mapping_set_error(page->mapping, error);
2027         end_page_writeback(page);
2028         return error;
2029 }
2030
2031 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
2032 {
2033         struct fuse_conn *fc = get_fuse_conn(page->mapping->host);
2034         int err;
2035
2036         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
2037                 /*
2038                  * ->writepages() should be called for sync() and friends.  We
2039                  * should only get here on direct reclaim and then we are
2040                  * allowed to skip a page which is already in flight
2041                  */
2042                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
2043
2044                 redirty_page_for_writepage(wbc, page);
2045                 unlock_page(page);
2046                 return 0;
2047         }
2048
2049         if (wbc->sync_mode == WB_SYNC_NONE &&
2050             fc->num_background >= fc->congestion_threshold)
2051                 return AOP_WRITEPAGE_ACTIVATE;
2052
2053         err = fuse_writepage_locked(page);
2054         unlock_page(page);
2055
2056         return err;
2057 }
2058
2059 struct fuse_fill_wb_data {
2060         struct fuse_writepage_args *wpa;
2061         struct fuse_file *ff;
2062         struct inode *inode;
2063         struct page **orig_pages;
2064         unsigned int max_pages;
2065 };
2066
2067 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
2068 {
2069         struct fuse_args_pages *ap = &data->wpa->ia.ap;
2070         struct fuse_conn *fc = get_fuse_conn(data->inode);
2071         struct page **pages;
2072         struct fuse_page_desc *descs;
2073         unsigned int npages = min_t(unsigned int,
2074                                     max_t(unsigned int, data->max_pages * 2,
2075                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
2076                                     fc->max_pages);
2077         WARN_ON(npages <= data->max_pages);
2078
2079         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
2080         if (!pages)
2081                 return false;
2082
2083         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
2084         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
2085         kfree(ap->pages);
2086         ap->pages = pages;
2087         ap->descs = descs;
2088         data->max_pages = npages;
2089
2090         return true;
2091 }
2092
2093 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
2094 {
2095         struct fuse_writepage_args *wpa = data->wpa;
2096         struct inode *inode = data->inode;
2097         struct fuse_inode *fi = get_fuse_inode(inode);
2098         int num_pages = wpa->ia.ap.num_pages;
2099         int i;
2100
2101         wpa->ia.ff = fuse_file_get(data->ff);
2102         spin_lock(&fi->lock);
2103         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2104         fuse_flush_writepages(inode);
2105         spin_unlock(&fi->lock);
2106
2107         for (i = 0; i < num_pages; i++)
2108                 end_page_writeback(data->orig_pages[i]);
2109 }
2110
2111 /*
2112  * Check under fi->lock if the page is under writeback, and insert it onto the
2113  * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2114  * one already added for a page at this offset.  If there's none, then insert
2115  * this new request onto the auxiliary list, otherwise reuse the existing one by
2116  * swapping the new temp page with the old one.
2117  */
2118 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2119                                struct page *page)
2120 {
2121         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2122         struct fuse_writepage_args *tmp;
2123         struct fuse_writepage_args *old_wpa;
2124         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2125
2126         WARN_ON(new_ap->num_pages != 0);
2127         new_ap->num_pages = 1;
2128
2129         spin_lock(&fi->lock);
2130         old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2131         if (!old_wpa) {
2132                 spin_unlock(&fi->lock);
2133                 return true;
2134         }
2135
2136         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2137                 pgoff_t curr_index;
2138
2139                 WARN_ON(tmp->inode != new_wpa->inode);
2140                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2141                 if (curr_index == page->index) {
2142                         WARN_ON(tmp->ia.ap.num_pages != 1);
2143                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2144                         break;
2145                 }
2146         }
2147
2148         if (!tmp) {
2149                 new_wpa->next = old_wpa->next;
2150                 old_wpa->next = new_wpa;
2151         }
2152
2153         spin_unlock(&fi->lock);
2154
2155         if (tmp) {
2156                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2157
2158                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2159                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2160                 wb_writeout_inc(&bdi->wb);
2161                 fuse_writepage_free(new_wpa);
2162         }
2163
2164         return false;
2165 }
2166
2167 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2168                                      struct fuse_args_pages *ap,
2169                                      struct fuse_fill_wb_data *data)
2170 {
2171         WARN_ON(!ap->num_pages);
2172
2173         /*
2174          * Being under writeback is unlikely but possible.  For example direct
2175          * read to an mmaped fuse file will set the page dirty twice; once when
2176          * the pages are faulted with get_user_pages(), and then after the read
2177          * completed.
2178          */
2179         if (fuse_page_is_writeback(data->inode, page->index))
2180                 return true;
2181
2182         /* Reached max pages */
2183         if (ap->num_pages == fc->max_pages)
2184                 return true;
2185
2186         /* Reached max write bytes */
2187         if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2188                 return true;
2189
2190         /* Discontinuity */
2191         if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2192                 return true;
2193
2194         /* Need to grow the pages array?  If so, did the expansion fail? */
2195         if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2196                 return true;
2197
2198         return false;
2199 }
2200
2201 static int fuse_writepages_fill(struct folio *folio,
2202                 struct writeback_control *wbc, void *_data)
2203 {
2204         struct fuse_fill_wb_data *data = _data;
2205         struct fuse_writepage_args *wpa = data->wpa;
2206         struct fuse_args_pages *ap = &wpa->ia.ap;
2207         struct inode *inode = data->inode;
2208         struct fuse_inode *fi = get_fuse_inode(inode);
2209         struct fuse_conn *fc = get_fuse_conn(inode);
2210         struct page *tmp_page;
2211         int err;
2212
2213         if (!data->ff) {
2214                 err = -EIO;
2215                 data->ff = fuse_write_file_get(fi);
2216                 if (!data->ff)
2217                         goto out_unlock;
2218         }
2219
2220         if (wpa && fuse_writepage_need_send(fc, &folio->page, ap, data)) {
2221                 fuse_writepages_send(data);
2222                 data->wpa = NULL;
2223         }
2224
2225         err = -ENOMEM;
2226         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2227         if (!tmp_page)
2228                 goto out_unlock;
2229
2230         /*
2231          * The page must not be redirtied until the writeout is completed
2232          * (i.e. userspace has sent a reply to the write request).  Otherwise
2233          * there could be more than one temporary page instance for each real
2234          * page.
2235          *
2236          * This is ensured by holding the page lock in page_mkwrite() while
2237          * checking fuse_page_is_writeback().  We already hold the page lock
2238          * since clear_page_dirty_for_io() and keep it held until we add the
2239          * request to the fi->writepages list and increment ap->num_pages.
2240          * After this fuse_page_is_writeback() will indicate that the page is
2241          * under writeback, so we can release the page lock.
2242          */
2243         if (data->wpa == NULL) {
2244                 err = -ENOMEM;
2245                 wpa = fuse_writepage_args_alloc();
2246                 if (!wpa) {
2247                         __free_page(tmp_page);
2248                         goto out_unlock;
2249                 }
2250                 fuse_writepage_add_to_bucket(fc, wpa);
2251
2252                 data->max_pages = 1;
2253
2254                 ap = &wpa->ia.ap;
2255                 fuse_write_args_fill(&wpa->ia, data->ff, folio_pos(folio), 0);
2256                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2257                 wpa->next = NULL;
2258                 ap->args.in_pages = true;
2259                 ap->args.end = fuse_writepage_end;
2260                 ap->num_pages = 0;
2261                 wpa->inode = inode;
2262         }
2263         folio_start_writeback(folio);
2264
2265         copy_highpage(tmp_page, &folio->page);
2266         ap->pages[ap->num_pages] = tmp_page;
2267         ap->descs[ap->num_pages].offset = 0;
2268         ap->descs[ap->num_pages].length = PAGE_SIZE;
2269         data->orig_pages[ap->num_pages] = &folio->page;
2270
2271         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2272         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2273
2274         err = 0;
2275         if (data->wpa) {
2276                 /*
2277                  * Protected by fi->lock against concurrent access by
2278                  * fuse_page_is_writeback().
2279                  */
2280                 spin_lock(&fi->lock);
2281                 ap->num_pages++;
2282                 spin_unlock(&fi->lock);
2283         } else if (fuse_writepage_add(wpa, &folio->page)) {
2284                 data->wpa = wpa;
2285         } else {
2286                 folio_end_writeback(folio);
2287         }
2288 out_unlock:
2289         folio_unlock(folio);
2290
2291         return err;
2292 }
2293
2294 static int fuse_writepages(struct address_space *mapping,
2295                            struct writeback_control *wbc)
2296 {
2297         struct inode *inode = mapping->host;
2298         struct fuse_conn *fc = get_fuse_conn(inode);
2299         struct fuse_fill_wb_data data;
2300         int err;
2301
2302         err = -EIO;
2303         if (fuse_is_bad(inode))
2304                 goto out;
2305
2306         if (wbc->sync_mode == WB_SYNC_NONE &&
2307             fc->num_background >= fc->congestion_threshold)
2308                 return 0;
2309
2310         data.inode = inode;
2311         data.wpa = NULL;
2312         data.ff = NULL;
2313
2314         err = -ENOMEM;
2315         data.orig_pages = kcalloc(fc->max_pages,
2316                                   sizeof(struct page *),
2317                                   GFP_NOFS);
2318         if (!data.orig_pages)
2319                 goto out;
2320
2321         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2322         if (data.wpa) {
2323                 WARN_ON(!data.wpa->ia.ap.num_pages);
2324                 fuse_writepages_send(&data);
2325         }
2326         if (data.ff)
2327                 fuse_file_put(data.ff, false, false);
2328
2329         kfree(data.orig_pages);
2330 out:
2331         return err;
2332 }
2333
2334 /*
2335  * It's worthy to make sure that space is reserved on disk for the write,
2336  * but how to implement it without killing performance need more thinking.
2337  */
2338 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2339                 loff_t pos, unsigned len, struct page **pagep, void **fsdata)
2340 {
2341         pgoff_t index = pos >> PAGE_SHIFT;
2342         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2343         struct page *page;
2344         loff_t fsize;
2345         int err = -ENOMEM;
2346
2347         WARN_ON(!fc->writeback_cache);
2348
2349         page = grab_cache_page_write_begin(mapping, index);
2350         if (!page)
2351                 goto error;
2352
2353         fuse_wait_on_page_writeback(mapping->host, page->index);
2354
2355         if (PageUptodate(page) || len == PAGE_SIZE)
2356                 goto success;
2357         /*
2358          * Check if the start this page comes after the end of file, in which
2359          * case the readpage can be optimized away.
2360          */
2361         fsize = i_size_read(mapping->host);
2362         if (fsize <= (pos & PAGE_MASK)) {
2363                 size_t off = pos & ~PAGE_MASK;
2364                 if (off)
2365                         zero_user_segment(page, 0, off);
2366                 goto success;
2367         }
2368         err = fuse_do_readpage(file, page);
2369         if (err)
2370                 goto cleanup;
2371 success:
2372         *pagep = page;
2373         return 0;
2374
2375 cleanup:
2376         unlock_page(page);
2377         put_page(page);
2378 error:
2379         return err;
2380 }
2381
2382 static int fuse_write_end(struct file *file, struct address_space *mapping,
2383                 loff_t pos, unsigned len, unsigned copied,
2384                 struct page *page, void *fsdata)
2385 {
2386         struct inode *inode = page->mapping->host;
2387
2388         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2389         if (!copied)
2390                 goto unlock;
2391
2392         pos += copied;
2393         if (!PageUptodate(page)) {
2394                 /* Zero any unwritten bytes at the end of the page */
2395                 size_t endoff = pos & ~PAGE_MASK;
2396                 if (endoff)
2397                         zero_user_segment(page, endoff, PAGE_SIZE);
2398                 SetPageUptodate(page);
2399         }
2400
2401         if (pos > inode->i_size)
2402                 i_size_write(inode, pos);
2403
2404         set_page_dirty(page);
2405
2406 unlock:
2407         unlock_page(page);
2408         put_page(page);
2409
2410         return copied;
2411 }
2412
2413 static int fuse_launder_folio(struct folio *folio)
2414 {
2415         int err = 0;
2416         if (folio_clear_dirty_for_io(folio)) {
2417                 struct inode *inode = folio->mapping->host;
2418
2419                 /* Serialize with pending writeback for the same page */
2420                 fuse_wait_on_page_writeback(inode, folio->index);
2421                 err = fuse_writepage_locked(&folio->page);
2422                 if (!err)
2423                         fuse_wait_on_page_writeback(inode, folio->index);
2424         }
2425         return err;
2426 }
2427
2428 /*
2429  * Write back dirty data/metadata now (there may not be any suitable
2430  * open files later for data)
2431  */
2432 static void fuse_vma_close(struct vm_area_struct *vma)
2433 {
2434         int err;
2435
2436         err = write_inode_now(vma->vm_file->f_mapping->host, 1);
2437         mapping_set_error(vma->vm_file->f_mapping, err);
2438 }
2439
2440 /*
2441  * Wait for writeback against this page to complete before allowing it
2442  * to be marked dirty again, and hence written back again, possibly
2443  * before the previous writepage completed.
2444  *
2445  * Block here, instead of in ->writepage(), so that the userspace fs
2446  * can only block processes actually operating on the filesystem.
2447  *
2448  * Otherwise unprivileged userspace fs would be able to block
2449  * unrelated:
2450  *
2451  * - page migration
2452  * - sync(2)
2453  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2454  */
2455 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2456 {
2457         struct page *page = vmf->page;
2458         struct inode *inode = file_inode(vmf->vma->vm_file);
2459
2460         file_update_time(vmf->vma->vm_file);
2461         lock_page(page);
2462         if (page->mapping != inode->i_mapping) {
2463                 unlock_page(page);
2464                 return VM_FAULT_NOPAGE;
2465         }
2466
2467         fuse_wait_on_page_writeback(inode, page->index);
2468         return VM_FAULT_LOCKED;
2469 }
2470
2471 static const struct vm_operations_struct fuse_file_vm_ops = {
2472         .close          = fuse_vma_close,
2473         .fault          = filemap_fault,
2474         .map_pages      = filemap_map_pages,
2475         .page_mkwrite   = fuse_page_mkwrite,
2476 };
2477
2478 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2479 {
2480         struct fuse_file *ff = file->private_data;
2481
2482         /* DAX mmap is superior to direct_io mmap */
2483         if (FUSE_IS_DAX(file_inode(file)))
2484                 return fuse_dax_mmap(file, vma);
2485
2486         if (ff->open_flags & FOPEN_DIRECT_IO) {
2487                 /* Can't provide the coherency needed for MAP_SHARED */
2488                 if (vma->vm_flags & VM_MAYSHARE)
2489                         return -ENODEV;
2490
2491                 invalidate_inode_pages2(file->f_mapping);
2492
2493                 return generic_file_mmap(file, vma);
2494         }
2495
2496         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2497                 fuse_link_write_file(file);
2498
2499         file_accessed(file);
2500         vma->vm_ops = &fuse_file_vm_ops;
2501         return 0;
2502 }
2503
2504 static int convert_fuse_file_lock(struct fuse_conn *fc,
2505                                   const struct fuse_file_lock *ffl,
2506                                   struct file_lock *fl)
2507 {
2508         switch (ffl->type) {
2509         case F_UNLCK:
2510                 break;
2511
2512         case F_RDLCK:
2513         case F_WRLCK:
2514                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2515                     ffl->end < ffl->start)
2516                         return -EIO;
2517
2518                 fl->fl_start = ffl->start;
2519                 fl->fl_end = ffl->end;
2520
2521                 /*
2522                  * Convert pid into init's pid namespace.  The locks API will
2523                  * translate it into the caller's pid namespace.
2524                  */
2525                 rcu_read_lock();
2526                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2527                 rcu_read_unlock();
2528                 break;
2529
2530         default:
2531                 return -EIO;
2532         }
2533         fl->fl_type = ffl->type;
2534         return 0;
2535 }
2536
2537 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2538                          const struct file_lock *fl, int opcode, pid_t pid,
2539                          int flock, struct fuse_lk_in *inarg)
2540 {
2541         struct inode *inode = file_inode(file);
2542         struct fuse_conn *fc = get_fuse_conn(inode);
2543         struct fuse_file *ff = file->private_data;
2544
2545         memset(inarg, 0, sizeof(*inarg));
2546         inarg->fh = ff->fh;
2547         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2548         inarg->lk.start = fl->fl_start;
2549         inarg->lk.end = fl->fl_end;
2550         inarg->lk.type = fl->fl_type;
2551         inarg->lk.pid = pid;
2552         if (flock)
2553                 inarg->lk_flags |= FUSE_LK_FLOCK;
2554         args->opcode = opcode;
2555         args->nodeid = get_node_id(inode);
2556         args->in_numargs = 1;
2557         args->in_args[0].size = sizeof(*inarg);
2558         args->in_args[0].value = inarg;
2559 }
2560
2561 static int fuse_getlk(struct file *file, struct file_lock *fl)
2562 {
2563         struct inode *inode = file_inode(file);
2564         struct fuse_mount *fm = get_fuse_mount(inode);
2565         FUSE_ARGS(args);
2566         struct fuse_lk_in inarg;
2567         struct fuse_lk_out outarg;
2568         int err;
2569
2570         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2571         args.out_numargs = 1;
2572         args.out_args[0].size = sizeof(outarg);
2573         args.out_args[0].value = &outarg;
2574         err = fuse_simple_request(fm, &args);
2575         if (!err)
2576                 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2577
2578         return err;
2579 }
2580
2581 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2582 {
2583         struct inode *inode = file_inode(file);
2584         struct fuse_mount *fm = get_fuse_mount(inode);
2585         FUSE_ARGS(args);
2586         struct fuse_lk_in inarg;
2587         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2588         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2589         pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2590         int err;
2591
2592         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2593                 /* NLM needs asynchronous locks, which we don't support yet */
2594                 return -ENOLCK;
2595         }
2596
2597         /* Unlock on close is handled by the flush method */
2598         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2599                 return 0;
2600
2601         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2602         err = fuse_simple_request(fm, &args);
2603
2604         /* locking is restartable */
2605         if (err == -EINTR)
2606                 err = -ERESTARTSYS;
2607
2608         return err;
2609 }
2610
2611 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2612 {
2613         struct inode *inode = file_inode(file);
2614         struct fuse_conn *fc = get_fuse_conn(inode);
2615         int err;
2616
2617         if (cmd == F_CANCELLK) {
2618                 err = 0;
2619         } else if (cmd == F_GETLK) {
2620                 if (fc->no_lock) {
2621                         posix_test_lock(file, fl);
2622                         err = 0;
2623                 } else
2624                         err = fuse_getlk(file, fl);
2625         } else {
2626                 if (fc->no_lock)
2627                         err = posix_lock_file(file, fl, NULL);
2628                 else
2629                         err = fuse_setlk(file, fl, 0);
2630         }
2631         return err;
2632 }
2633
2634 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2635 {
2636         struct inode *inode = file_inode(file);
2637         struct fuse_conn *fc = get_fuse_conn(inode);
2638         int err;
2639
2640         if (fc->no_flock) {
2641                 err = locks_lock_file_wait(file, fl);
2642         } else {
2643                 struct fuse_file *ff = file->private_data;
2644
2645                 /* emulate flock with POSIX locks */
2646                 ff->flock = true;
2647                 err = fuse_setlk(file, fl, 1);
2648         }
2649
2650         return err;
2651 }
2652
2653 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2654 {
2655         struct inode *inode = mapping->host;
2656         struct fuse_mount *fm = get_fuse_mount(inode);
2657         FUSE_ARGS(args);
2658         struct fuse_bmap_in inarg;
2659         struct fuse_bmap_out outarg;
2660         int err;
2661
2662         if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2663                 return 0;
2664
2665         memset(&inarg, 0, sizeof(inarg));
2666         inarg.block = block;
2667         inarg.blocksize = inode->i_sb->s_blocksize;
2668         args.opcode = FUSE_BMAP;
2669         args.nodeid = get_node_id(inode);
2670         args.in_numargs = 1;
2671         args.in_args[0].size = sizeof(inarg);
2672         args.in_args[0].value = &inarg;
2673         args.out_numargs = 1;
2674         args.out_args[0].size = sizeof(outarg);
2675         args.out_args[0].value = &outarg;
2676         err = fuse_simple_request(fm, &args);
2677         if (err == -ENOSYS)
2678                 fm->fc->no_bmap = 1;
2679
2680         return err ? 0 : outarg.block;
2681 }
2682
2683 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2684 {
2685         struct inode *inode = file->f_mapping->host;
2686         struct fuse_mount *fm = get_fuse_mount(inode);
2687         struct fuse_file *ff = file->private_data;
2688         FUSE_ARGS(args);
2689         struct fuse_lseek_in inarg = {
2690                 .fh = ff->fh,
2691                 .offset = offset,
2692                 .whence = whence
2693         };
2694         struct fuse_lseek_out outarg;
2695         int err;
2696
2697         if (fm->fc->no_lseek)
2698                 goto fallback;
2699
2700         args.opcode = FUSE_LSEEK;
2701         args.nodeid = ff->nodeid;
2702         args.in_numargs = 1;
2703         args.in_args[0].size = sizeof(inarg);
2704         args.in_args[0].value = &inarg;
2705         args.out_numargs = 1;
2706         args.out_args[0].size = sizeof(outarg);
2707         args.out_args[0].value = &outarg;
2708         err = fuse_simple_request(fm, &args);
2709         if (err) {
2710                 if (err == -ENOSYS) {
2711                         fm->fc->no_lseek = 1;
2712                         goto fallback;
2713                 }
2714                 return err;
2715         }
2716
2717         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2718
2719 fallback:
2720         err = fuse_update_attributes(inode, file, STATX_SIZE);
2721         if (!err)
2722                 return generic_file_llseek(file, offset, whence);
2723         else
2724                 return err;
2725 }
2726
2727 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2728 {
2729         loff_t retval;
2730         struct inode *inode = file_inode(file);
2731
2732         switch (whence) {
2733         case SEEK_SET:
2734         case SEEK_CUR:
2735                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2736                 retval = generic_file_llseek(file, offset, whence);
2737                 break;
2738         case SEEK_END:
2739                 inode_lock(inode);
2740                 retval = fuse_update_attributes(inode, file, STATX_SIZE);
2741                 if (!retval)
2742                         retval = generic_file_llseek(file, offset, whence);
2743                 inode_unlock(inode);
2744                 break;
2745         case SEEK_HOLE:
2746         case SEEK_DATA:
2747                 inode_lock(inode);
2748                 retval = fuse_lseek(file, offset, whence);
2749                 inode_unlock(inode);
2750                 break;
2751         default:
2752                 retval = -EINVAL;
2753         }
2754
2755         return retval;
2756 }
2757
2758 /*
2759  * All files which have been polled are linked to RB tree
2760  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2761  * find the matching one.
2762  */
2763 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2764                                               struct rb_node **parent_out)
2765 {
2766         struct rb_node **link = &fc->polled_files.rb_node;
2767         struct rb_node *last = NULL;
2768
2769         while (*link) {
2770                 struct fuse_file *ff;
2771
2772                 last = *link;
2773                 ff = rb_entry(last, struct fuse_file, polled_node);
2774
2775                 if (kh < ff->kh)
2776                         link = &last->rb_left;
2777                 else if (kh > ff->kh)
2778                         link = &last->rb_right;
2779                 else
2780                         return link;
2781         }
2782
2783         if (parent_out)
2784                 *parent_out = last;
2785         return link;
2786 }
2787
2788 /*
2789  * The file is about to be polled.  Make sure it's on the polled_files
2790  * RB tree.  Note that files once added to the polled_files tree are
2791  * not removed before the file is released.  This is because a file
2792  * polled once is likely to be polled again.
2793  */
2794 static void fuse_register_polled_file(struct fuse_conn *fc,
2795                                       struct fuse_file *ff)
2796 {
2797         spin_lock(&fc->lock);
2798         if (RB_EMPTY_NODE(&ff->polled_node)) {
2799                 struct rb_node **link, *parent;
2800
2801                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2802                 BUG_ON(*link);
2803                 rb_link_node(&ff->polled_node, parent, link);
2804                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2805         }
2806         spin_unlock(&fc->lock);
2807 }
2808
2809 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2810 {
2811         struct fuse_file *ff = file->private_data;
2812         struct fuse_mount *fm = ff->fm;
2813         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2814         struct fuse_poll_out outarg;
2815         FUSE_ARGS(args);
2816         int err;
2817
2818         if (fm->fc->no_poll)
2819                 return DEFAULT_POLLMASK;
2820
2821         poll_wait(file, &ff->poll_wait, wait);
2822         inarg.events = mangle_poll(poll_requested_events(wait));
2823
2824         /*
2825          * Ask for notification iff there's someone waiting for it.
2826          * The client may ignore the flag and always notify.
2827          */
2828         if (waitqueue_active(&ff->poll_wait)) {
2829                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2830                 fuse_register_polled_file(fm->fc, ff);
2831         }
2832
2833         args.opcode = FUSE_POLL;
2834         args.nodeid = ff->nodeid;
2835         args.in_numargs = 1;
2836         args.in_args[0].size = sizeof(inarg);
2837         args.in_args[0].value = &inarg;
2838         args.out_numargs = 1;
2839         args.out_args[0].size = sizeof(outarg);
2840         args.out_args[0].value = &outarg;
2841         err = fuse_simple_request(fm, &args);
2842
2843         if (!err)
2844                 return demangle_poll(outarg.revents);
2845         if (err == -ENOSYS) {
2846                 fm->fc->no_poll = 1;
2847                 return DEFAULT_POLLMASK;
2848         }
2849         return EPOLLERR;
2850 }
2851 EXPORT_SYMBOL_GPL(fuse_file_poll);
2852
2853 /*
2854  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2855  * wakes up the poll waiters.
2856  */
2857 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2858                             struct fuse_notify_poll_wakeup_out *outarg)
2859 {
2860         u64 kh = outarg->kh;
2861         struct rb_node **link;
2862
2863         spin_lock(&fc->lock);
2864
2865         link = fuse_find_polled_node(fc, kh, NULL);
2866         if (*link) {
2867                 struct fuse_file *ff;
2868
2869                 ff = rb_entry(*link, struct fuse_file, polled_node);
2870                 wake_up_interruptible_sync(&ff->poll_wait);
2871         }
2872
2873         spin_unlock(&fc->lock);
2874         return 0;
2875 }
2876
2877 static void fuse_do_truncate(struct file *file)
2878 {
2879         struct inode *inode = file->f_mapping->host;
2880         struct iattr attr;
2881
2882         attr.ia_valid = ATTR_SIZE;
2883         attr.ia_size = i_size_read(inode);
2884
2885         attr.ia_file = file;
2886         attr.ia_valid |= ATTR_FILE;
2887
2888         fuse_do_setattr(file_dentry(file), &attr, file);
2889 }
2890
2891 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2892 {
2893         return round_up(off, fc->max_pages << PAGE_SHIFT);
2894 }
2895
2896 static ssize_t
2897 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2898 {
2899         DECLARE_COMPLETION_ONSTACK(wait);
2900         ssize_t ret = 0;
2901         struct file *file = iocb->ki_filp;
2902         struct fuse_file *ff = file->private_data;
2903         loff_t pos = 0;
2904         struct inode *inode;
2905         loff_t i_size;
2906         size_t count = iov_iter_count(iter), shortened = 0;
2907         loff_t offset = iocb->ki_pos;
2908         struct fuse_io_priv *io;
2909
2910         pos = offset;
2911         inode = file->f_mapping->host;
2912         i_size = i_size_read(inode);
2913
2914         if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
2915                 return 0;
2916
2917         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2918         if (!io)
2919                 return -ENOMEM;
2920         spin_lock_init(&io->lock);
2921         kref_init(&io->refcnt);
2922         io->reqs = 1;
2923         io->bytes = -1;
2924         io->size = 0;
2925         io->offset = offset;
2926         io->write = (iov_iter_rw(iter) == WRITE);
2927         io->err = 0;
2928         /*
2929          * By default, we want to optimize all I/Os with async request
2930          * submission to the client filesystem if supported.
2931          */
2932         io->async = ff->fm->fc->async_dio;
2933         io->iocb = iocb;
2934         io->blocking = is_sync_kiocb(iocb);
2935
2936         /* optimization for short read */
2937         if (io->async && !io->write && offset + count > i_size) {
2938                 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
2939                 shortened = count - iov_iter_count(iter);
2940                 count -= shortened;
2941         }
2942
2943         /*
2944          * We cannot asynchronously extend the size of a file.
2945          * In such case the aio will behave exactly like sync io.
2946          */
2947         if ((offset + count > i_size) && io->write)
2948                 io->blocking = true;
2949
2950         if (io->async && io->blocking) {
2951                 /*
2952                  * Additional reference to keep io around after
2953                  * calling fuse_aio_complete()
2954                  */
2955                 kref_get(&io->refcnt);
2956                 io->done = &wait;
2957         }
2958
2959         if (iov_iter_rw(iter) == WRITE) {
2960                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2961                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
2962         } else {
2963                 ret = __fuse_direct_read(io, iter, &pos);
2964         }
2965         iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
2966
2967         if (io->async) {
2968                 bool blocking = io->blocking;
2969
2970                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2971
2972                 /* we have a non-extending, async request, so return */
2973                 if (!blocking)
2974                         return -EIOCBQUEUED;
2975
2976                 wait_for_completion(&wait);
2977                 ret = fuse_get_res_by_io(io);
2978         }
2979
2980         kref_put(&io->refcnt, fuse_io_release);
2981
2982         if (iov_iter_rw(iter) == WRITE) {
2983                 fuse_write_update_attr(inode, pos, ret);
2984                 /* For extending writes we already hold exclusive lock */
2985                 if (ret < 0 && offset + count > i_size)
2986                         fuse_do_truncate(file);
2987         }
2988
2989         return ret;
2990 }
2991
2992 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
2993 {
2994         int err = filemap_write_and_wait_range(inode->i_mapping, start, LLONG_MAX);
2995
2996         if (!err)
2997                 fuse_sync_writes(inode);
2998
2999         return err;
3000 }
3001
3002 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3003                                 loff_t length)
3004 {
3005         struct fuse_file *ff = file->private_data;
3006         struct inode *inode = file_inode(file);
3007         struct fuse_inode *fi = get_fuse_inode(inode);
3008         struct fuse_mount *fm = ff->fm;
3009         FUSE_ARGS(args);
3010         struct fuse_fallocate_in inarg = {
3011                 .fh = ff->fh,
3012                 .offset = offset,
3013                 .length = length,
3014                 .mode = mode
3015         };
3016         int err;
3017         bool block_faults = FUSE_IS_DAX(inode) &&
3018                 (!(mode & FALLOC_FL_KEEP_SIZE) ||
3019                  (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)));
3020
3021         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3022                      FALLOC_FL_ZERO_RANGE))
3023                 return -EOPNOTSUPP;
3024
3025         if (fm->fc->no_fallocate)
3026                 return -EOPNOTSUPP;
3027
3028         inode_lock(inode);
3029         if (block_faults) {
3030                 filemap_invalidate_lock(inode->i_mapping);
3031                 err = fuse_dax_break_layouts(inode, 0, 0);
3032                 if (err)
3033                         goto out;
3034         }
3035
3036         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)) {
3037                 loff_t endbyte = offset + length - 1;
3038
3039                 err = fuse_writeback_range(inode, offset, endbyte);
3040                 if (err)
3041                         goto out;
3042         }
3043
3044         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3045             offset + length > i_size_read(inode)) {
3046                 err = inode_newsize_ok(inode, offset + length);
3047                 if (err)
3048                         goto out;
3049         }
3050
3051         err = file_modified(file);
3052         if (err)
3053                 goto out;
3054
3055         if (!(mode & FALLOC_FL_KEEP_SIZE))
3056                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3057
3058         args.opcode = FUSE_FALLOCATE;
3059         args.nodeid = ff->nodeid;
3060         args.in_numargs = 1;
3061         args.in_args[0].size = sizeof(inarg);
3062         args.in_args[0].value = &inarg;
3063         err = fuse_simple_request(fm, &args);
3064         if (err == -ENOSYS) {
3065                 fm->fc->no_fallocate = 1;
3066                 err = -EOPNOTSUPP;
3067         }
3068         if (err)
3069                 goto out;
3070
3071         /* we could have extended the file */
3072         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3073                 if (fuse_write_update_attr(inode, offset + length, length))
3074                         file_update_time(file);
3075         }
3076
3077         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE))
3078                 truncate_pagecache_range(inode, offset, offset + length - 1);
3079
3080         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
3081
3082 out:
3083         if (!(mode & FALLOC_FL_KEEP_SIZE))
3084                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3085
3086         if (block_faults)
3087                 filemap_invalidate_unlock(inode->i_mapping);
3088
3089         inode_unlock(inode);
3090
3091         fuse_flush_time_update(inode);
3092
3093         return err;
3094 }
3095
3096 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3097                                       struct file *file_out, loff_t pos_out,
3098                                       size_t len, unsigned int flags)
3099 {
3100         struct fuse_file *ff_in = file_in->private_data;
3101         struct fuse_file *ff_out = file_out->private_data;
3102         struct inode *inode_in = file_inode(file_in);
3103         struct inode *inode_out = file_inode(file_out);
3104         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3105         struct fuse_mount *fm = ff_in->fm;
3106         struct fuse_conn *fc = fm->fc;
3107         FUSE_ARGS(args);
3108         struct fuse_copy_file_range_in inarg = {
3109                 .fh_in = ff_in->fh,
3110                 .off_in = pos_in,
3111                 .nodeid_out = ff_out->nodeid,
3112                 .fh_out = ff_out->fh,
3113                 .off_out = pos_out,
3114                 .len = len,
3115                 .flags = flags
3116         };
3117         struct fuse_write_out outarg;
3118         ssize_t err;
3119         /* mark unstable when write-back is not used, and file_out gets
3120          * extended */
3121         bool is_unstable = (!fc->writeback_cache) &&
3122                            ((pos_out + len) > inode_out->i_size);
3123
3124         if (fc->no_copy_file_range)
3125                 return -EOPNOTSUPP;
3126
3127         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3128                 return -EXDEV;
3129
3130         inode_lock(inode_in);
3131         err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3132         inode_unlock(inode_in);
3133         if (err)
3134                 return err;
3135
3136         inode_lock(inode_out);
3137
3138         err = file_modified(file_out);
3139         if (err)
3140                 goto out;
3141
3142         /*
3143          * Write out dirty pages in the destination file before sending the COPY
3144          * request to userspace.  After the request is completed, truncate off
3145          * pages (including partial ones) from the cache that have been copied,
3146          * since these contain stale data at that point.
3147          *
3148          * This should be mostly correct, but if the COPY writes to partial
3149          * pages (at the start or end) and the parts not covered by the COPY are
3150          * written through a memory map after calling fuse_writeback_range(),
3151          * then these partial page modifications will be lost on truncation.
3152          *
3153          * It is unlikely that someone would rely on such mixed style
3154          * modifications.  Yet this does give less guarantees than if the
3155          * copying was performed with write(2).
3156          *
3157          * To fix this a mapping->invalidate_lock could be used to prevent new
3158          * faults while the copy is ongoing.
3159          */
3160         err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3161         if (err)
3162                 goto out;
3163
3164         if (is_unstable)
3165                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3166
3167         args.opcode = FUSE_COPY_FILE_RANGE;
3168         args.nodeid = ff_in->nodeid;
3169         args.in_numargs = 1;
3170         args.in_args[0].size = sizeof(inarg);
3171         args.in_args[0].value = &inarg;
3172         args.out_numargs = 1;
3173         args.out_args[0].size = sizeof(outarg);
3174         args.out_args[0].value = &outarg;
3175         err = fuse_simple_request(fm, &args);
3176         if (err == -ENOSYS) {
3177                 fc->no_copy_file_range = 1;
3178                 err = -EOPNOTSUPP;
3179         }
3180         if (err)
3181                 goto out;
3182
3183         truncate_inode_pages_range(inode_out->i_mapping,
3184                                    ALIGN_DOWN(pos_out, PAGE_SIZE),
3185                                    ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3186
3187         file_update_time(file_out);
3188         fuse_write_update_attr(inode_out, pos_out + outarg.size, outarg.size);
3189
3190         err = outarg.size;
3191 out:
3192         if (is_unstable)
3193                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3194
3195         inode_unlock(inode_out);
3196         file_accessed(file_in);
3197
3198         fuse_flush_time_update(inode_out);
3199
3200         return err;
3201 }
3202
3203 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3204                                     struct file *dst_file, loff_t dst_off,
3205                                     size_t len, unsigned int flags)
3206 {
3207         ssize_t ret;
3208
3209         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3210                                      len, flags);
3211
3212         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3213                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3214                                               dst_off, len, flags);
3215         return ret;
3216 }
3217
3218 static const struct file_operations fuse_file_operations = {
3219         .llseek         = fuse_file_llseek,
3220         .read_iter      = fuse_file_read_iter,
3221         .write_iter     = fuse_file_write_iter,
3222         .mmap           = fuse_file_mmap,
3223         .open           = fuse_open,
3224         .flush          = fuse_flush,
3225         .release        = fuse_release,
3226         .fsync          = fuse_fsync,
3227         .lock           = fuse_file_lock,
3228         .get_unmapped_area = thp_get_unmapped_area,
3229         .flock          = fuse_file_flock,
3230         .splice_read    = filemap_splice_read,
3231         .splice_write   = iter_file_splice_write,
3232         .unlocked_ioctl = fuse_file_ioctl,
3233         .compat_ioctl   = fuse_file_compat_ioctl,
3234         .poll           = fuse_file_poll,
3235         .fallocate      = fuse_file_fallocate,
3236         .copy_file_range = fuse_copy_file_range,
3237 };
3238
3239 static const struct address_space_operations fuse_file_aops  = {
3240         .read_folio     = fuse_read_folio,
3241         .readahead      = fuse_readahead,
3242         .writepage      = fuse_writepage,
3243         .writepages     = fuse_writepages,
3244         .launder_folio  = fuse_launder_folio,
3245         .dirty_folio    = filemap_dirty_folio,
3246         .bmap           = fuse_bmap,
3247         .direct_IO      = fuse_direct_IO,
3248         .write_begin    = fuse_write_begin,
3249         .write_end      = fuse_write_end,
3250 };
3251
3252 void fuse_init_file_inode(struct inode *inode, unsigned int flags)
3253 {
3254         struct fuse_inode *fi = get_fuse_inode(inode);
3255
3256         inode->i_fop = &fuse_file_operations;
3257         inode->i_data.a_ops = &fuse_file_aops;
3258
3259         INIT_LIST_HEAD(&fi->write_files);
3260         INIT_LIST_HEAD(&fi->queued_writes);
3261         fi->writectr = 0;
3262         init_waitqueue_head(&fi->page_waitq);
3263         fi->writepages = RB_ROOT;
3264
3265         if (IS_ENABLED(CONFIG_FUSE_DAX))
3266                 fuse_dax_inode_init(inode, flags);
3267 }