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