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