fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/sched/signal.h>
  16 #include <linux/module.h>
  17 #include <linux/swap.h>
  18 #include <linux/falloc.h>
  19 #include <linux/uio.h>
  20 #include <linux/fs.h>
  21 #include <linux/filelock.h>
  22 #include <linux/splice.h>
  23 #include <linux/task_io_accounting_ops.h>
  24
  25 static int fuse_send_open(struct fuse_mount *fm, u64 nodeid,
  26                           unsigned int open_flags, int opcode,
  27                           struct fuse_open_out *outargp)
  28 {
  29         struct fuse_open_in inarg;
  30         FUSE_ARGS(args);
  31
  32         memset(&inarg, 0, sizeof(inarg));
  33         inarg.flags = open_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  34         if (!fm->fc->atomic_o_trunc)
  35                 inarg.flags &= ~O_TRUNC;
  36
  37         if (fm->fc->handle_killpriv_v2 &&
  38             (inarg.flags & O_TRUNC) && !capable(CAP_FSETID)) {
  39                 inarg.open_flags |= FUSE_OPEN_KILL_SUIDGID;
  40         }
  41
  42         args.opcode = opcode;
  43         args.nodeid = nodeid;
  44         args.in_numargs = 1;
  45         args.in_args[0].size = sizeof(inarg);
  46         args.in_args[0].value = &inarg;
  47         args.out_numargs = 1;
  48         args.out_args[0].size = sizeof(*outargp);
  49         args.out_args[0].value = outargp;
  50
  51         return fuse_simple_request(fm, &args);
  52 }
  53
  54 struct fuse_file *fuse_file_alloc(struct fuse_mount *fm, bool release)
  55 {
  56         struct fuse_file *ff;
  57
  58         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
  59         if (unlikely(!ff))
  60                 return NULL;
  61
  62         ff->fm = fm;
  63         if (release) {
  64                 ff->args = kzalloc(sizeof(*ff->args), GFP_KERNEL_ACCOUNT);
  65                 if (!ff->args) {
  66                         kfree(ff);
  67                         return NULL;
  68                 }
  69         }
  70
  71         INIT_LIST_HEAD(&ff->write_entry);
  72         refcount_set(&ff->count, 1);
  73         RB_CLEAR_NODE(&ff->polled_node);
  74         init_waitqueue_head(&ff->poll_wait);
  75
  76         ff->kh = atomic64_inc_return(&fm->fc->khctr);
  77
  78         return ff;
  79 }
  80
  81 void fuse_file_free(struct fuse_file *ff)
  82 {
  83         kfree(ff->args);
  84         kfree(ff);
  85 }
  86
  87 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  88 {
  89         refcount_inc(&ff->count);
  90         return ff;
  91 }
  92
  93 static void fuse_release_end(struct fuse_mount *fm, struct fuse_args *args,
  94                              int error)
  95 {
  96         struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
  97
  98         iput(ra->inode);
  99         kfree(ra);
 100 }
 101
 102 static void fuse_file_put(struct fuse_file *ff, bool sync)
 103 {
 104         if (refcount_dec_and_test(&ff->count)) {
 105                 struct fuse_release_args *ra = &ff->args->release_args;
 106                 struct fuse_args *args = (ra ? &ra->args : NULL);
 107
 108                 if (ra && ra->inode)
 109                         fuse_file_io_release(ff, ra->inode);
 110
 111                 if (!args) {
 112                         /* Do nothing when server does not implement 'open' */
 113                 } else if (sync) {
 114                         fuse_simple_request(ff->fm, args);
 115                         fuse_release_end(ff->fm, args, 0);
 116                 } else {
 117                         args->end = fuse_release_end;
 118                         if (fuse_simple_background(ff->fm, args,
 119                                                    GFP_KERNEL | __GFP_NOFAIL))
 120                                 fuse_release_end(ff->fm, args, -ENOTCONN);
 121                 }
 122                 kfree(ff);
 123         }
 124 }
 125
 126 struct fuse_file *fuse_file_open(struct fuse_mount *fm, u64 nodeid,
 127                                  unsigned int open_flags, bool isdir)
 128 {
 129         struct fuse_conn *fc = fm->fc;
 130         struct fuse_file *ff;
 131         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 132         bool open = isdir ? !fc->no_opendir : !fc->no_open;
 133
 134         ff = fuse_file_alloc(fm, open);
 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 (open) {
 142                 /* Store outarg for fuse_finish_open() */
 143                 struct fuse_open_out *outargp = &ff->args->open_outarg;
 144                 int err;
 145
 146                 err = fuse_send_open(fm, nodeid, open_flags, opcode, outargp);
 147                 if (!err) {
 148                         ff->fh = outargp->fh;
 149                         ff->open_flags = outargp->open_flags;
 150                 } else if (err != -ENOSYS) {
 151                         fuse_file_free(ff);
 152                         return ERR_PTR(err);
 153                 } else {
 154                         /* No release needed */
 155                         kfree(ff->args);
 156                         ff->args = NULL;
 157                         if (isdir)
 158                                 fc->no_opendir = 1;
 159                         else
 160                                 fc->no_open = 1;
 161                 }
 162         }
 163
 164         if (isdir)
 165                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 166
 167         ff->nodeid = nodeid;
 168
 169         return ff;
 170 }
 171
 172 int fuse_do_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
 173                  bool isdir)
 174 {
 175         struct fuse_file *ff = fuse_file_open(fm, nodeid, file->f_flags, isdir);
 176
 177         if (!IS_ERR(ff))
 178                 file->private_data = ff;
 179
 180         return PTR_ERR_OR_ZERO(ff);
 181 }
 182 EXPORT_SYMBOL_GPL(fuse_do_open);
 183
 184 static void fuse_link_write_file(struct file *file)
 185 {
 186         struct inode *inode = file_inode(file);
 187         struct fuse_inode *fi = get_fuse_inode(inode);
 188         struct fuse_file *ff = file->private_data;
 189         /*
 190          * file may be written through mmap, so chain it onto the
 191          * inodes's write_file list
 192          */
 193         spin_lock(&fi->lock);
 194         if (list_empty(&ff->write_entry))
 195                 list_add(&ff->write_entry, &fi->write_files);
 196         spin_unlock(&fi->lock);
 197 }
 198
 199 int fuse_finish_open(struct inode *inode, struct file *file)
 200 {
 201         struct fuse_file *ff = file->private_data;
 202         struct fuse_conn *fc = get_fuse_conn(inode);
 203         int err;
 204
 205         err = fuse_file_io_open(file, inode);
 206         if (err)
 207                 return err;
 208
 209         if (ff->open_flags & FOPEN_STREAM)
 210                 stream_open(inode, file);
 211         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 212                 nonseekable_open(inode, file);
 213
 214         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 215                 fuse_link_write_file(file);
 216
 217         return 0;
 218 }
 219
 220 static void fuse_truncate_update_attr(struct inode *inode, struct file *file)
 221 {
 222         struct fuse_conn *fc = get_fuse_conn(inode);
 223         struct fuse_inode *fi = get_fuse_inode(inode);
 224
 225         spin_lock(&fi->lock);
 226         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 227         i_size_write(inode, 0);
 228         spin_unlock(&fi->lock);
 229         file_update_time(file);
 230         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
 231 }
 232
 233 static int fuse_open(struct inode *inode, struct file *file)
 234 {
 235         struct fuse_mount *fm = get_fuse_mount(inode);
 236         struct fuse_inode *fi = get_fuse_inode(inode);
 237         struct fuse_conn *fc = fm->fc;
 238         struct fuse_file *ff;
 239         int err;
 240         bool is_truncate = (file->f_flags & O_TRUNC) && fc->atomic_o_trunc;
 241         bool is_wb_truncate = is_truncate && fc->writeback_cache;
 242         bool dax_truncate = is_truncate && FUSE_IS_DAX(inode);
 243
 244         if (fuse_is_bad(inode))
 245                 return -EIO;
 246
 247         err = generic_file_open(inode, file);
 248         if (err)
 249                 return err;
 250
 251         if (is_wb_truncate || dax_truncate)
 252                 inode_lock(inode);
 253
 254         if (dax_truncate) {
 255                 filemap_invalidate_lock(inode->i_mapping);
 256                 err = fuse_dax_break_layouts(inode, 0, 0);
 257                 if (err)
 258                         goto out_inode_unlock;
 259         }
 260
 261         if (is_wb_truncate || dax_truncate)
 262                 fuse_set_nowrite(inode);
 263
 264         err = fuse_do_open(fm, get_node_id(inode), file, false);
 265         if (!err) {
 266                 ff = file->private_data;
 267                 err = fuse_finish_open(inode, file);
 268                 if (err)
 269                         fuse_sync_release(fi, ff, file->f_flags);
 270                 else if (is_truncate)
 271                         fuse_truncate_update_attr(inode, file);
 272         }
 273
 274         if (is_wb_truncate || dax_truncate)
 275                 fuse_release_nowrite(inode);
 276         if (!err) {
 277                 if (is_truncate)
 278                         truncate_pagecache(inode, 0);
 279                 else if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 280                         invalidate_inode_pages2(inode->i_mapping);
 281         }
 282         if (dax_truncate)
 283                 filemap_invalidate_unlock(inode->i_mapping);
 284 out_inode_unlock:
 285         if (is_wb_truncate || dax_truncate)
 286                 inode_unlock(inode);
 287
 288         return err;
 289 }
 290
 291 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 292                                  unsigned int flags, int opcode, bool sync)
 293 {
 294         struct fuse_conn *fc = ff->fm->fc;
 295         struct fuse_release_args *ra = &ff->args->release_args;
 296
 297         if (fuse_file_passthrough(ff))
 298                 fuse_passthrough_release(ff, fuse_inode_backing(fi));
 299
 300         /* Inode is NULL on error path of fuse_create_open() */
 301         if (likely(fi)) {
 302                 spin_lock(&fi->lock);
 303                 list_del(&ff->write_entry);
 304                 spin_unlock(&fi->lock);
 305         }
 306         spin_lock(&fc->lock);
 307         if (!RB_EMPTY_NODE(&ff->polled_node))
 308                 rb_erase(&ff->polled_node, &fc->polled_files);
 309         spin_unlock(&fc->lock);
 310
 311         wake_up_interruptible_all(&ff->poll_wait);
 312
 313         if (!ra)
 314                 return;
 315
 316         /* ff->args was used for open outarg */
 317         memset(ff->args, 0, sizeof(*ff->args));
 318         ra->inarg.fh = ff->fh;
 319         ra->inarg.flags = flags;
 320         ra->args.in_numargs = 1;
 321         ra->args.in_args[0].size = sizeof(struct fuse_release_in);
 322         ra->args.in_args[0].value = &ra->inarg;
 323         ra->args.opcode = opcode;
 324         ra->args.nodeid = ff->nodeid;
 325         ra->args.force = true;
 326         ra->args.nocreds = true;
 327
 328         /*
 329          * Hold inode until release is finished.
 330          * From fuse_sync_release() the refcount is 1 and everything's
 331          * synchronous, so we are fine with not doing igrab() here.
 332          */
 333         ra->inode = sync ? NULL : igrab(&fi->inode);
 334 }
 335
 336 void fuse_file_release(struct inode *inode, struct fuse_file *ff,
 337                        unsigned int open_flags, fl_owner_t id, bool isdir)
 338 {
 339         struct fuse_inode *fi = get_fuse_inode(inode);
 340         struct fuse_release_args *ra = &ff->args->release_args;
 341         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 342
 343         fuse_prepare_release(fi, ff, open_flags, opcode, false);
 344
 345         if (ra && ff->flock) {
 346                 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 347                 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fm->fc, id);
 348         }
 349
 350         /*
 351          * Normally this will send the RELEASE request, however if
 352          * some asynchronous READ or WRITE requests are outstanding,
 353          * the sending will be delayed.
 354          *
 355          * Make the release synchronous if this is a fuseblk mount,
 356          * synchronous RELEASE is allowed (and desirable) in this case
 357          * because the server can be trusted not to screw up.
 358          */
 359         fuse_file_put(ff, ff->fm->fc->destroy);
 360 }
 361
 362 void fuse_release_common(struct file *file, bool isdir)
 363 {
 364         fuse_file_release(file_inode(file), file->private_data, file->f_flags,
 365                           (fl_owner_t) file, isdir);
 366 }
 367
 368 static int fuse_release(struct inode *inode, struct file *file)
 369 {
 370         struct fuse_conn *fc = get_fuse_conn(inode);
 371
 372         /*
 373          * Dirty pages might remain despite write_inode_now() call from
 374          * fuse_flush() due to writes racing with the close.
 375          */
 376         if (fc->writeback_cache)
 377                 write_inode_now(inode, 1);
 378
 379         fuse_release_common(file, false);
 380
 381         /* return value is ignored by VFS */
 382         return 0;
 383 }
 384
 385 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff,
 386                        unsigned int flags)
 387 {
 388         WARN_ON(refcount_read(&ff->count) > 1);
 389         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE, true);
 390         fuse_file_put(ff, true);
 391 }
 392 EXPORT_SYMBOL_GPL(fuse_sync_release);
 393
 394 /*
 395  * Scramble the ID space with XTEA, so that the value of the files_struct
 396  * pointer is not exposed to userspace.
 397  */
 398 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 399 {
 400         u32 *k = fc->scramble_key;
 401         u64 v = (unsigned long) id;
 402         u32 v0 = v;
 403         u32 v1 = v >> 32;
 404         u32 sum = 0;
 405         int i;
 406
 407         for (i = 0; i < 32; i++) {
 408                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 409                 sum += 0x9E3779B9;
 410                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 411         }
 412
 413         return (u64) v0 + ((u64) v1 << 32);
 414 }
 415
 416 struct fuse_writepage_args {
 417         struct fuse_io_args ia;
 418         struct rb_node writepages_entry;
 419         struct list_head queue_entry;
 420         struct fuse_writepage_args *next;
 421         struct inode *inode;
 422         struct fuse_sync_bucket *bucket;
 423 };
 424
 425 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
 426                                             pgoff_t idx_from, pgoff_t idx_to)
 427 {
 428         struct rb_node *n;
 429
 430         n = fi->writepages.rb_node;
 431
 432         while (n) {
 433                 struct fuse_writepage_args *wpa;
 434                 pgoff_t curr_index;
 435
 436                 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
 437                 WARN_ON(get_fuse_inode(wpa->inode) != fi);
 438                 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
 439                 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
 440                         n = n->rb_right;
 441                 else if (idx_to < curr_index)
 442                         n = n->rb_left;
 443                 else
 444                         return wpa;
 445         }
 446         return NULL;
 447 }
 448
 449 /*
 450  * Check if any page in a range is under writeback
 451  *
 452  * This is currently done by walking the list of writepage requests
 453  * for the inode, which can be pretty inefficient.
 454  */
 455 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 456                                    pgoff_t idx_to)
 457 {
 458         struct fuse_inode *fi = get_fuse_inode(inode);
 459         bool found;
 460
 461         spin_lock(&fi->lock);
 462         found = fuse_find_writeback(fi, idx_from, idx_to);
 463         spin_unlock(&fi->lock);
 464
 465         return found;
 466 }
 467
 468 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 469 {
 470         return fuse_range_is_writeback(inode, index, index);
 471 }
 472
 473 /*
 474  * Wait for page writeback to be completed.
 475  *
 476  * Since fuse doesn't rely on the VM writeback tracking, this has to
 477  * use some other means.
 478  */
 479 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 480 {
 481         struct fuse_inode *fi = get_fuse_inode(inode);
 482
 483         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 484 }
 485
 486 /*
 487  * Wait for all pending writepages on the inode to finish.
 488  *
 489  * This is currently done by blocking further writes with FUSE_NOWRITE
 490  * and waiting for all sent writes to complete.
 491  *
 492  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 493  * could conflict with truncation.
 494  */
 495 static void fuse_sync_writes(struct inode *inode)
 496 {
 497         fuse_set_nowrite(inode);
 498         fuse_release_nowrite(inode);
 499 }
 500
 501 static int fuse_flush(struct file *file, fl_owner_t id)
 502 {
 503         struct inode *inode = file_inode(file);
 504         struct fuse_mount *fm = get_fuse_mount(inode);
 505         struct fuse_file *ff = file->private_data;
 506         struct fuse_flush_in inarg;
 507         FUSE_ARGS(args);
 508         int err;
 509
 510         if (fuse_is_bad(inode))
 511                 return -EIO;
 512
 513         if (ff->open_flags & FOPEN_NOFLUSH && !fm->fc->writeback_cache)
 514                 return 0;
 515
 516         err = write_inode_now(inode, 1);
 517         if (err)
 518                 return err;
 519
 520         inode_lock(inode);
 521         fuse_sync_writes(inode);
 522         inode_unlock(inode);
 523
 524         err = filemap_check_errors(file->f_mapping);
 525         if (err)
 526                 return err;
 527
 528         err = 0;
 529         if (fm->fc->no_flush)
 530                 goto inval_attr_out;
 531
 532         memset(&inarg, 0, sizeof(inarg));
 533         inarg.fh = ff->fh;
 534         inarg.lock_owner = fuse_lock_owner_id(fm->fc, id);
 535         args.opcode = FUSE_FLUSH;
 536         args.nodeid = get_node_id(inode);
 537         args.in_numargs = 1;
 538         args.in_args[0].size = sizeof(inarg);
 539         args.in_args[0].value = &inarg;
 540         args.force = true;
 541
 542         err = fuse_simple_request(fm, &args);
 543         if (err == -ENOSYS) {
 544                 fm->fc->no_flush = 1;
 545                 err = 0;
 546         }
 547
 548 inval_attr_out:
 549         /*
 550          * In memory i_blocks is not maintained by fuse, if writeback cache is
 551          * enabled, i_blocks from cached attr may not be accurate.
 552          */
 553         if (!err && fm->fc->writeback_cache)
 554                 fuse_invalidate_attr_mask(inode, STATX_BLOCKS);
 555         return err;
 556 }
 557
 558 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 559                       int datasync, int opcode)
 560 {
 561         struct inode *inode = file->f_mapping->host;
 562         struct fuse_mount *fm = get_fuse_mount(inode);
 563         struct fuse_file *ff = file->private_data;
 564         FUSE_ARGS(args);
 565         struct fuse_fsync_in inarg;
 566
 567         memset(&inarg, 0, sizeof(inarg));
 568         inarg.fh = ff->fh;
 569         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 570         args.opcode = opcode;
 571         args.nodeid = get_node_id(inode);
 572         args.in_numargs = 1;
 573         args.in_args[0].size = sizeof(inarg);
 574         args.in_args[0].value = &inarg;
 575         return fuse_simple_request(fm, &args);
 576 }
 577
 578 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 579                       int datasync)
 580 {
 581         struct inode *inode = file->f_mapping->host;
 582         struct fuse_conn *fc = get_fuse_conn(inode);
 583         int err;
 584
 585         if (fuse_is_bad(inode))
 586                 return -EIO;
 587
 588         inode_lock(inode);
 589
 590         /*
 591          * Start writeback against all dirty pages of the inode, then
 592          * wait for all outstanding writes, before sending the FSYNC
 593          * request.
 594          */
 595         err = file_write_and_wait_range(file, start, end);
 596         if (err)
 597                 goto out;
 598
 599         fuse_sync_writes(inode);
 600
 601         /*
 602          * Due to implementation of fuse writeback
 603          * file_write_and_wait_range() does not catch errors.
 604          * We have to do this directly after fuse_sync_writes()
 605          */
 606         err = file_check_and_advance_wb_err(file);
 607         if (err)
 608                 goto out;
 609
 610         err = sync_inode_metadata(inode, 1);
 611         if (err)
 612                 goto out;
 613
 614         if (fc->no_fsync)
 615                 goto out;
 616
 617         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 618         if (err == -ENOSYS) {
 619                 fc->no_fsync = 1;
 620                 err = 0;
 621         }
 622 out:
 623         inode_unlock(inode);
 624
 625         return err;
 626 }
 627
 628 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
 629                          size_t count, int opcode)
 630 {
 631         struct fuse_file *ff = file->private_data;
 632         struct fuse_args *args = &ia->ap.args;
 633
 634         ia->read.in.fh = ff->fh;
 635         ia->read.in.offset = pos;
 636         ia->read.in.size = count;
 637         ia->read.in.flags = file->f_flags;
 638         args->opcode = opcode;
 639         args->nodeid = ff->nodeid;
 640         args->in_numargs = 1;
 641         args->in_args[0].size = sizeof(ia->read.in);
 642         args->in_args[0].value = &ia->read.in;
 643         args->out_argvar = true;
 644         args->out_numargs = 1;
 645         args->out_args[0].size = count;
 646 }
 647
 648 static void fuse_release_user_pages(struct fuse_args_pages *ap,
 649                                     bool should_dirty)
 650 {
 651         unsigned int i;
 652
 653         for (i = 0; i < ap->num_pages; i++) {
 654                 if (should_dirty)
 655                         set_page_dirty_lock(ap->pages[i]);
 656                 if (ap->args.is_pinned)
 657                         unpin_user_page(ap->pages[i]);
 658         }
 659 }
 660
 661 static void fuse_io_release(struct kref *kref)
 662 {
 663         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 664 }
 665
 666 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 667 {
 668         if (io->err)
 669                 return io->err;
 670
 671         if (io->bytes >= 0 && io->write)
 672                 return -EIO;
 673
 674         return io->bytes < 0 ? io->size : io->bytes;
 675 }
 676
 677 /*
 678  * In case of short read, the caller sets 'pos' to the position of
 679  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 680  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 681  *
 682  * An example:
 683  * User requested DIO read of 64K. It was split into two 32K fuse requests,
 684  * both submitted asynchronously. The first of them was ACKed by userspace as
 685  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 686  * second request was ACKed as short, e.g. only 1K was read, resulting in
 687  * pos == 33K.
 688  *
 689  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 690  * will be equal to the length of the longest contiguous fragment of
 691  * transferred data starting from the beginning of IO request.
 692  */
 693 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 694 {
 695         int left;
 696
 697         spin_lock(&io->lock);
 698         if (err)
 699                 io->err = io->err ? : err;
 700         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 701                 io->bytes = pos;
 702
 703         left = --io->reqs;
 704         if (!left && io->blocking)
 705                 complete(io->done);
 706         spin_unlock(&io->lock);
 707
 708         if (!left && !io->blocking) {
 709                 ssize_t res = fuse_get_res_by_io(io);
 710
 711                 if (res >= 0) {
 712                         struct inode *inode = file_inode(io->iocb->ki_filp);
 713                         struct fuse_conn *fc = get_fuse_conn(inode);
 714                         struct fuse_inode *fi = get_fuse_inode(inode);
 715
 716                         spin_lock(&fi->lock);
 717                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 718                         spin_unlock(&fi->lock);
 719                 }
 720
 721                 io->iocb->ki_complete(io->iocb, res);
 722         }
 723
 724         kref_put(&io->refcnt, fuse_io_release);
 725 }
 726
 727 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
 728                                           unsigned int npages)
 729 {
 730         struct fuse_io_args *ia;
 731
 732         ia = kzalloc(sizeof(*ia), GFP_KERNEL);
 733         if (ia) {
 734                 ia->io = io;
 735                 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
 736                                                 &ia->ap.descs);
 737                 if (!ia->ap.pages) {
 738                         kfree(ia);
 739                         ia = NULL;
 740                 }
 741         }
 742         return ia;
 743 }
 744
 745 static void fuse_io_free(struct fuse_io_args *ia)
 746 {
 747         kfree(ia->ap.pages);
 748         kfree(ia);
 749 }
 750
 751 static void fuse_aio_complete_req(struct fuse_mount *fm, struct fuse_args *args,
 752                                   int err)
 753 {
 754         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 755         struct fuse_io_priv *io = ia->io;
 756         ssize_t pos = -1;
 757
 758         fuse_release_user_pages(&ia->ap, io->should_dirty);
 759
 760         if (err) {
 761                 /* Nothing */
 762         } else if (io->write) {
 763                 if (ia->write.out.size > ia->write.in.size) {
 764                         err = -EIO;
 765                 } else if (ia->write.in.size != ia->write.out.size) {
 766                         pos = ia->write.in.offset - io->offset +
 767                                 ia->write.out.size;
 768                 }
 769         } else {
 770                 u32 outsize = args->out_args[0].size;
 771
 772                 if (ia->read.in.size != outsize)
 773                         pos = ia->read.in.offset - io->offset + outsize;
 774         }
 775
 776         fuse_aio_complete(io, err, pos);
 777         fuse_io_free(ia);
 778 }
 779
 780 static ssize_t fuse_async_req_send(struct fuse_mount *fm,
 781                                    struct fuse_io_args *ia, size_t num_bytes)
 782 {
 783         ssize_t err;
 784         struct fuse_io_priv *io = ia->io;
 785
 786         spin_lock(&io->lock);
 787         kref_get(&io->refcnt);
 788         io->size += num_bytes;
 789         io->reqs++;
 790         spin_unlock(&io->lock);
 791
 792         ia->ap.args.end = fuse_aio_complete_req;
 793         ia->ap.args.may_block = io->should_dirty;
 794         err = fuse_simple_background(fm, &ia->ap.args, GFP_KERNEL);
 795         if (err)
 796                 fuse_aio_complete_req(fm, &ia->ap.args, err);
 797
 798         return num_bytes;
 799 }
 800
 801 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
 802                               fl_owner_t owner)
 803 {
 804         struct file *file = ia->io->iocb->ki_filp;
 805         struct fuse_file *ff = file->private_data;
 806         struct fuse_mount *fm = ff->fm;
 807
 808         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 809         if (owner != NULL) {
 810                 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
 811                 ia->read.in.lock_owner = fuse_lock_owner_id(fm->fc, owner);
 812         }
 813
 814         if (ia->io->async)
 815                 return fuse_async_req_send(fm, ia, count);
 816
 817         return fuse_simple_request(fm, &ia->ap.args);
 818 }
 819
 820 static void fuse_read_update_size(struct inode *inode, loff_t size,
 821                                   u64 attr_ver)
 822 {
 823         struct fuse_conn *fc = get_fuse_conn(inode);
 824         struct fuse_inode *fi = get_fuse_inode(inode);
 825
 826         spin_lock(&fi->lock);
 827         if (attr_ver >= fi->attr_version && size < inode->i_size &&
 828             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 829                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 830                 i_size_write(inode, size);
 831         }
 832         spin_unlock(&fi->lock);
 833 }
 834
 835 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
 836                             struct fuse_args_pages *ap)
 837 {
 838         struct fuse_conn *fc = get_fuse_conn(inode);
 839
 840         /*
 841          * If writeback_cache is enabled, a short read means there's a hole in
 842          * the file.  Some data after the hole is in page cache, but has not
 843          * reached the client fs yet.  So the hole is not present there.
 844          */
 845         if (!fc->writeback_cache) {
 846                 loff_t pos = page_offset(ap->pages[0]) + num_read;
 847                 fuse_read_update_size(inode, pos, attr_ver);
 848         }
 849 }
 850
 851 static int fuse_do_readpage(struct file *file, struct page *page)
 852 {
 853         struct inode *inode = page->mapping->host;
 854         struct fuse_mount *fm = get_fuse_mount(inode);
 855         loff_t pos = page_offset(page);
 856         struct fuse_page_desc desc = { .length = PAGE_SIZE };
 857         struct fuse_io_args ia = {
 858                 .ap.args.page_zeroing = true,
 859                 .ap.args.out_pages = true,
 860                 .ap.num_pages = 1,
 861                 .ap.pages = &page,
 862                 .ap.descs = &desc,
 863         };
 864         ssize_t res;
 865         u64 attr_ver;
 866
 867         /*
 868          * Page writeback can extend beyond the lifetime of the
 869          * page-cache page, so make sure we read a properly synced
 870          * page.
 871          */
 872         fuse_wait_on_page_writeback(inode, page->index);
 873
 874         attr_ver = fuse_get_attr_version(fm->fc);
 875
 876         /* Don't overflow end offset */
 877         if (pos + (desc.length - 1) == LLONG_MAX)
 878                 desc.length--;
 879
 880         fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
 881         res = fuse_simple_request(fm, &ia.ap.args);
 882         if (res < 0)
 883                 return res;
 884         /*
 885          * Short read means EOF.  If file size is larger, truncate it
 886          */
 887         if (res < desc.length)
 888                 fuse_short_read(inode, attr_ver, res, &ia.ap);
 889
 890         SetPageUptodate(page);
 891
 892         return 0;
 893 }
 894
 895 static int fuse_read_folio(struct file *file, struct folio *folio)
 896 {
 897         struct page *page = &folio->page;
 898         struct inode *inode = page->mapping->host;
 899         int err;
 900
 901         err = -EIO;
 902         if (fuse_is_bad(inode))
 903                 goto out;
 904
 905         err = fuse_do_readpage(file, page);
 906         fuse_invalidate_atime(inode);
 907  out:
 908         unlock_page(page);
 909         return err;
 910 }
 911
 912 static void fuse_readpages_end(struct fuse_mount *fm, struct fuse_args *args,
 913                                int err)
 914 {
 915         int i;
 916         struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
 917         struct fuse_args_pages *ap = &ia->ap;
 918         size_t count = ia->read.in.size;
 919         size_t num_read = args->out_args[0].size;
 920         struct address_space *mapping = NULL;
 921
 922         for (i = 0; mapping == NULL && i < ap->num_pages; i++)
 923                 mapping = ap->pages[i]->mapping;
 924
 925         if (mapping) {
 926                 struct inode *inode = mapping->host;
 927
 928                 /*
 929                  * Short read means EOF. If file size is larger, truncate it
 930                  */
 931                 if (!err && num_read < count)
 932                         fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
 933
 934                 fuse_invalidate_atime(inode);
 935         }
 936
 937         for (i = 0; i < ap->num_pages; i++) {
 938                 struct page *page = ap->pages[i];
 939
 940                 if (!err)
 941                         SetPageUptodate(page);
 942                 else
 943                         SetPageError(page);
 944                 unlock_page(page);
 945                 put_page(page);
 946         }
 947         if (ia->ff)
 948                 fuse_file_put(ia->ff, false);
 949
 950         fuse_io_free(ia);
 951 }
 952
 953 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
 954 {
 955         struct fuse_file *ff = file->private_data;
 956         struct fuse_mount *fm = ff->fm;
 957         struct fuse_args_pages *ap = &ia->ap;
 958         loff_t pos = page_offset(ap->pages[0]);
 959         size_t count = ap->num_pages << PAGE_SHIFT;
 960         ssize_t res;
 961         int err;
 962
 963         ap->args.out_pages = true;
 964         ap->args.page_zeroing = true;
 965         ap->args.page_replace = true;
 966
 967         /* Don't overflow end offset */
 968         if (pos + (count - 1) == LLONG_MAX) {
 969                 count--;
 970                 ap->descs[ap->num_pages - 1].length--;
 971         }
 972         WARN_ON((loff_t) (pos + count) < 0);
 973
 974         fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
 975         ia->read.attr_ver = fuse_get_attr_version(fm->fc);
 976         if (fm->fc->async_read) {
 977                 ia->ff = fuse_file_get(ff);
 978                 ap->args.end = fuse_readpages_end;
 979                 err = fuse_simple_background(fm, &ap->args, GFP_KERNEL);
 980                 if (!err)
 981                         return;
 982         } else {
 983                 res = fuse_simple_request(fm, &ap->args);
 984                 err = res < 0 ? res : 0;
 985         }
 986         fuse_readpages_end(fm, &ap->args, err);
 987 }
 988
 989 static void fuse_readahead(struct readahead_control *rac)
 990 {
 991         struct inode *inode = rac->mapping->host;
 992         struct fuse_conn *fc = get_fuse_conn(inode);
 993         unsigned int i, max_pages, nr_pages = 0;
 994
 995         if (fuse_is_bad(inode))
 996                 return;
 997
 998         max_pages = min_t(unsigned int, fc->max_pages,
 999                         fc->max_read / PAGE_SIZE);
1000
1001         for (;;) {
1002                 struct fuse_io_args *ia;
1003                 struct fuse_args_pages *ap;
1004
1005                 if (fc->num_background >= fc->congestion_threshold &&
1006                     rac->ra->async_size >= readahead_count(rac))
1007                         /*
1008                          * Congested and only async pages left, so skip the
1009                          * rest.
1010                          */
1011                         break;
1012
1013                 nr_pages = readahead_count(rac) - nr_pages;
1014                 if (nr_pages > max_pages)
1015                         nr_pages = max_pages;
1016                 if (nr_pages == 0)
1017                         break;
1018                 ia = fuse_io_alloc(NULL, nr_pages);
1019                 if (!ia)
1020                         return;
1021                 ap = &ia->ap;
1022                 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
1023                 for (i = 0; i < nr_pages; i++) {
1024                         fuse_wait_on_page_writeback(inode,
1025                                                     readahead_index(rac) + i);
1026                         ap->descs[i].length = PAGE_SIZE;
1027                 }
1028                 ap->num_pages = nr_pages;
1029                 fuse_send_readpages(ia, rac->file);
1030         }
1031 }
1032
1033 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
1034 {
1035         struct inode *inode = iocb->ki_filp->f_mapping->host;
1036         struct fuse_conn *fc = get_fuse_conn(inode);
1037
1038         /*
1039          * In auto invalidate mode, always update attributes on read.
1040          * Otherwise, only update if we attempt to read past EOF (to ensure
1041          * i_size is up to date).
1042          */
1043         if (fc->auto_inval_data ||
1044             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
1045                 int err;
1046                 err = fuse_update_attributes(inode, iocb->ki_filp, STATX_SIZE);
1047                 if (err)
1048                         return err;
1049         }
1050
1051         return generic_file_read_iter(iocb, to);
1052 }
1053
1054 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1055                                  loff_t pos, size_t count)
1056 {
1057         struct fuse_args *args = &ia->ap.args;
1058
1059         ia->write.in.fh = ff->fh;
1060         ia->write.in.offset = pos;
1061         ia->write.in.size = count;
1062         args->opcode = FUSE_WRITE;
1063         args->nodeid = ff->nodeid;
1064         args->in_numargs = 2;
1065         if (ff->fm->fc->minor < 9)
1066                 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1067         else
1068                 args->in_args[0].size = sizeof(ia->write.in);
1069         args->in_args[0].value = &ia->write.in;
1070         args->in_args[1].size = count;
1071         args->out_numargs = 1;
1072         args->out_args[0].size = sizeof(ia->write.out);
1073         args->out_args[0].value = &ia->write.out;
1074 }
1075
1076 static unsigned int fuse_write_flags(struct kiocb *iocb)
1077 {
1078         unsigned int flags = iocb->ki_filp->f_flags;
1079
1080         if (iocb_is_dsync(iocb))
1081                 flags |= O_DSYNC;
1082         if (iocb->ki_flags & IOCB_SYNC)
1083                 flags |= O_SYNC;
1084
1085         return flags;
1086 }
1087
1088 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1089                                size_t count, fl_owner_t owner)
1090 {
1091         struct kiocb *iocb = ia->io->iocb;
1092         struct file *file = iocb->ki_filp;
1093         struct fuse_file *ff = file->private_data;
1094         struct fuse_mount *fm = ff->fm;
1095         struct fuse_write_in *inarg = &ia->write.in;
1096         ssize_t err;
1097
1098         fuse_write_args_fill(ia, ff, pos, count);
1099         inarg->flags = fuse_write_flags(iocb);
1100         if (owner != NULL) {
1101                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1102                 inarg->lock_owner = fuse_lock_owner_id(fm->fc, owner);
1103         }
1104
1105         if (ia->io->async)
1106                 return fuse_async_req_send(fm, ia, count);
1107
1108         err = fuse_simple_request(fm, &ia->ap.args);
1109         if (!err && ia->write.out.size > count)
1110                 err = -EIO;
1111
1112         return err ?: ia->write.out.size;
1113 }
1114
1115 bool fuse_write_update_attr(struct inode *inode, loff_t pos, ssize_t written)
1116 {
1117         struct fuse_conn *fc = get_fuse_conn(inode);
1118         struct fuse_inode *fi = get_fuse_inode(inode);
1119         bool ret = false;
1120
1121         spin_lock(&fi->lock);
1122         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1123         if (written > 0 && pos > inode->i_size) {
1124                 i_size_write(inode, pos);
1125                 ret = true;
1126         }
1127         spin_unlock(&fi->lock);
1128
1129         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
1130
1131         return ret;
1132 }
1133
1134 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1135                                      struct kiocb *iocb, struct inode *inode,
1136                                      loff_t pos, size_t count)
1137 {
1138         struct fuse_args_pages *ap = &ia->ap;
1139         struct file *file = iocb->ki_filp;
1140         struct fuse_file *ff = file->private_data;
1141         struct fuse_mount *fm = ff->fm;
1142         unsigned int offset, i;
1143         bool short_write;
1144         int err;
1145
1146         for (i = 0; i < ap->num_pages; i++)
1147                 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1148
1149         fuse_write_args_fill(ia, ff, pos, count);
1150         ia->write.in.flags = fuse_write_flags(iocb);
1151         if (fm->fc->handle_killpriv_v2 && !capable(CAP_FSETID))
1152                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1153
1154         err = fuse_simple_request(fm, &ap->args);
1155         if (!err && ia->write.out.size > count)
1156                 err = -EIO;
1157
1158         short_write = ia->write.out.size < count;
1159         offset = ap->descs[0].offset;
1160         count = ia->write.out.size;
1161         for (i = 0; i < ap->num_pages; i++) {
1162                 struct page *page = ap->pages[i];
1163
1164                 if (err) {
1165                         ClearPageUptodate(page);
1166                 } else {
1167                         if (count >= PAGE_SIZE - offset)
1168                                 count -= PAGE_SIZE - offset;
1169                         else {
1170                                 if (short_write)
1171                                         ClearPageUptodate(page);
1172                                 count = 0;
1173                         }
1174                         offset = 0;
1175                 }
1176                 if (ia->write.page_locked && (i == ap->num_pages - 1))
1177                         unlock_page(page);
1178                 put_page(page);
1179         }
1180
1181         return err;
1182 }
1183
1184 static ssize_t fuse_fill_write_pages(struct fuse_io_args *ia,
1185                                      struct address_space *mapping,
1186                                      struct iov_iter *ii, loff_t pos,
1187                                      unsigned int max_pages)
1188 {
1189         struct fuse_args_pages *ap = &ia->ap;
1190         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1191         unsigned offset = pos & (PAGE_SIZE - 1);
1192         size_t count = 0;
1193         int err;
1194
1195         ap->args.in_pages = true;
1196         ap->descs[0].offset = offset;
1197
1198         do {
1199                 size_t tmp;
1200                 struct page *page;
1201                 pgoff_t index = pos >> PAGE_SHIFT;
1202                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1203                                      iov_iter_count(ii));
1204
1205                 bytes = min_t(size_t, bytes, fc->max_write - count);
1206
1207  again:
1208                 err = -EFAULT;
1209                 if (fault_in_iov_iter_readable(ii, bytes))
1210                         break;
1211
1212                 err = -ENOMEM;
1213                 page = grab_cache_page_write_begin(mapping, index);
1214                 if (!page)
1215                         break;
1216
1217                 if (mapping_writably_mapped(mapping))
1218                         flush_dcache_page(page);
1219
1220                 tmp = copy_page_from_iter_atomic(page, offset, bytes, ii);
1221                 flush_dcache_page(page);
1222
1223                 if (!tmp) {
1224                         unlock_page(page);
1225                         put_page(page);
1226                         goto again;
1227                 }
1228
1229                 err = 0;
1230                 ap->pages[ap->num_pages] = page;
1231                 ap->descs[ap->num_pages].length = tmp;
1232                 ap->num_pages++;
1233
1234                 count += tmp;
1235                 pos += tmp;
1236                 offset += tmp;
1237                 if (offset == PAGE_SIZE)
1238                         offset = 0;
1239
1240                 /* If we copied full page, mark it uptodate */
1241                 if (tmp == PAGE_SIZE)
1242                         SetPageUptodate(page);
1243
1244                 if (PageUptodate(page)) {
1245                         unlock_page(page);
1246                 } else {
1247                         ia->write.page_locked = true;
1248                         break;
1249                 }
1250                 if (!fc->big_writes)
1251                         break;
1252         } while (iov_iter_count(ii) && count < fc->max_write &&
1253                  ap->num_pages < max_pages && offset == 0);
1254
1255         return count > 0 ? count : err;
1256 }
1257
1258 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1259                                      unsigned int max_pages)
1260 {
1261         return min_t(unsigned int,
1262                      ((pos + len - 1) >> PAGE_SHIFT) -
1263                      (pos >> PAGE_SHIFT) + 1,
1264                      max_pages);
1265 }
1266
1267 static ssize_t fuse_perform_write(struct kiocb *iocb, struct iov_iter *ii)
1268 {
1269         struct address_space *mapping = iocb->ki_filp->f_mapping;
1270         struct inode *inode = mapping->host;
1271         struct fuse_conn *fc = get_fuse_conn(inode);
1272         struct fuse_inode *fi = get_fuse_inode(inode);
1273         loff_t pos = iocb->ki_pos;
1274         int err = 0;
1275         ssize_t res = 0;
1276
1277         if (inode->i_size < pos + iov_iter_count(ii))
1278                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1279
1280         do {
1281                 ssize_t count;
1282                 struct fuse_io_args ia = {};
1283                 struct fuse_args_pages *ap = &ia.ap;
1284                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1285                                                       fc->max_pages);
1286
1287                 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1288                 if (!ap->pages) {
1289                         err = -ENOMEM;
1290                         break;
1291                 }
1292
1293                 count = fuse_fill_write_pages(&ia, mapping, ii, pos, nr_pages);
1294                 if (count <= 0) {
1295                         err = count;
1296                 } else {
1297                         err = fuse_send_write_pages(&ia, iocb, inode,
1298                                                     pos, count);
1299                         if (!err) {
1300                                 size_t num_written = ia.write.out.size;
1301
1302                                 res += num_written;
1303                                 pos += num_written;
1304
1305                                 /* break out of the loop on short write */
1306                                 if (num_written != count)
1307                                         err = -EIO;
1308                         }
1309                 }
1310                 kfree(ap->pages);
1311         } while (!err && iov_iter_count(ii));
1312
1313         fuse_write_update_attr(inode, pos, res);
1314         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1315
1316         if (!res)
1317                 return err;
1318         iocb->ki_pos += res;
1319         return res;
1320 }
1321
1322 static bool fuse_io_past_eof(struct kiocb *iocb, struct iov_iter *iter)
1323 {
1324         struct inode *inode = file_inode(iocb->ki_filp);
1325
1326         return iocb->ki_pos + iov_iter_count(iter) > i_size_read(inode);
1327 }
1328
1329 /*
1330  * @return true if an exclusive lock for direct IO writes is needed
1331  */
1332 static bool fuse_dio_wr_exclusive_lock(struct kiocb *iocb, struct iov_iter *from)
1333 {
1334         struct file *file = iocb->ki_filp;
1335         struct fuse_file *ff = file->private_data;
1336         struct inode *inode = file_inode(iocb->ki_filp);
1337         struct fuse_inode *fi = get_fuse_inode(inode);
1338
1339         /* Server side has to advise that it supports parallel dio writes. */
1340         if (!(ff->open_flags & FOPEN_PARALLEL_DIRECT_WRITES))
1341                 return true;
1342
1343         /*
1344          * Append will need to know the eventual EOF - always needs an
1345          * exclusive lock.
1346          */
1347         if (iocb->ki_flags & IOCB_APPEND)
1348                 return true;
1349
1350         /* shared locks are not allowed with parallel page cache IO */
1351         if (test_bit(FUSE_I_CACHE_IO_MODE, &fi->state))
1352                 return false;
1353
1354         /* Parallel dio beyond EOF is not supported, at least for now. */
1355         if (fuse_io_past_eof(iocb, from))
1356                 return true;
1357
1358         return false;
1359 }
1360
1361 static void fuse_dio_lock(struct kiocb *iocb, struct iov_iter *from,
1362                           bool *exclusive)
1363 {
1364         struct inode *inode = file_inode(iocb->ki_filp);
1365         struct fuse_inode *fi = get_fuse_inode(inode);
1366
1367         *exclusive = fuse_dio_wr_exclusive_lock(iocb, from);
1368         if (*exclusive) {
1369                 inode_lock(inode);
1370         } else {
1371                 inode_lock_shared(inode);
1372                 /*
1373                  * New parallal dio allowed only if inode is not in caching
1374                  * mode and denies new opens in caching mode. This check
1375                  * should be performed only after taking shared inode lock.
1376                  * Previous past eof check was without inode lock and might
1377                  * have raced, so check it again.
1378                  */
1379                 if (fuse_io_past_eof(iocb, from) ||
1380                     fuse_inode_uncached_io_start(fi, NULL) != 0) {
1381                         inode_unlock_shared(inode);
1382                         inode_lock(inode);
1383                         *exclusive = true;
1384                 }
1385         }
1386 }
1387
1388 static void fuse_dio_unlock(struct kiocb *iocb, bool exclusive)
1389 {
1390         struct inode *inode = file_inode(iocb->ki_filp);
1391         struct fuse_inode *fi = get_fuse_inode(inode);
1392
1393         if (exclusive) {
1394                 inode_unlock(inode);
1395         } else {
1396                 /* Allow opens in caching mode after last parallel dio end */
1397                 fuse_inode_uncached_io_end(fi);
1398                 inode_unlock_shared(inode);
1399         }
1400 }
1401
1402 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1403 {
1404         struct file *file = iocb->ki_filp;
1405         struct address_space *mapping = file->f_mapping;
1406         ssize_t written = 0;
1407         struct inode *inode = mapping->host;
1408         ssize_t err, count;
1409         struct fuse_conn *fc = get_fuse_conn(inode);
1410
1411         if (fc->writeback_cache) {
1412                 /* Update size (EOF optimization) and mode (SUID clearing) */
1413                 err = fuse_update_attributes(mapping->host, file,
1414                                              STATX_SIZE | STATX_MODE);
1415                 if (err)
1416                         return err;
1417
1418                 if (fc->handle_killpriv_v2 &&
1419                     setattr_should_drop_suidgid(&nop_mnt_idmap,
1420                                                 file_inode(file))) {
1421                         goto writethrough;
1422                 }
1423
1424                 return generic_file_write_iter(iocb, from);
1425         }
1426
1427 writethrough:
1428         inode_lock(inode);
1429
1430         err = count = generic_write_checks(iocb, from);
1431         if (err <= 0)
1432                 goto out;
1433
1434         task_io_account_write(count);
1435
1436         err = file_remove_privs(file);
1437         if (err)
1438                 goto out;
1439
1440         err = file_update_time(file);
1441         if (err)
1442                 goto out;
1443
1444         if (iocb->ki_flags & IOCB_DIRECT) {
1445                 written = generic_file_direct_write(iocb, from);
1446                 if (written < 0 || !iov_iter_count(from))
1447                         goto out;
1448                 written = direct_write_fallback(iocb, from, written,
1449                                 fuse_perform_write(iocb, from));
1450         } else {
1451                 written = fuse_perform_write(iocb, from);
1452         }
1453 out:
1454         inode_unlock(inode);
1455         if (written > 0)
1456                 written = generic_write_sync(iocb, written);
1457
1458         return written ? written : err;
1459 }
1460
1461 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1462 {
1463         return (unsigned long)iter_iov(ii)->iov_base + ii->iov_offset;
1464 }
1465
1466 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1467                                         size_t max_size)
1468 {
1469         return min(iov_iter_single_seg_count(ii), max_size);
1470 }
1471
1472 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1473                                size_t *nbytesp, int write,
1474                                unsigned int max_pages)
1475 {
1476         size_t nbytes = 0;  /* # bytes already packed in req */
1477         ssize_t ret = 0;
1478
1479         /* Special case for kernel I/O: can copy directly into the buffer */
1480         if (iov_iter_is_kvec(ii)) {
1481                 unsigned long user_addr = fuse_get_user_addr(ii);
1482                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1483
1484                 if (write)
1485                         ap->args.in_args[1].value = (void *) user_addr;
1486                 else
1487                         ap->args.out_args[0].value = (void *) user_addr;
1488
1489                 iov_iter_advance(ii, frag_size);
1490                 *nbytesp = frag_size;
1491                 return 0;
1492         }
1493
1494         while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1495                 unsigned npages;
1496                 size_t start;
1497                 struct page **pt_pages;
1498
1499                 pt_pages = &ap->pages[ap->num_pages];
1500                 ret = iov_iter_extract_pages(ii, &pt_pages,
1501                                              *nbytesp - nbytes,
1502                                              max_pages - ap->num_pages,
1503                                              0, &start);
1504                 if (ret < 0)
1505                         break;
1506
1507                 nbytes += ret;
1508
1509                 ret += start;
1510                 npages = DIV_ROUND_UP(ret, PAGE_SIZE);
1511
1512                 ap->descs[ap->num_pages].offset = start;
1513                 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1514
1515                 ap->num_pages += npages;
1516                 ap->descs[ap->num_pages - 1].length -=
1517                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1518         }
1519
1520         ap->args.is_pinned = iov_iter_extract_will_pin(ii);
1521         ap->args.user_pages = true;
1522         if (write)
1523                 ap->args.in_pages = true;
1524         else
1525                 ap->args.out_pages = true;
1526
1527         *nbytesp = nbytes;
1528
1529         return ret < 0 ? ret : 0;
1530 }
1531
1532 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1533                        loff_t *ppos, int flags)
1534 {
1535         int write = flags & FUSE_DIO_WRITE;
1536         int cuse = flags & FUSE_DIO_CUSE;
1537         struct file *file = io->iocb->ki_filp;
1538         struct address_space *mapping = file->f_mapping;
1539         struct inode *inode = mapping->host;
1540         struct fuse_file *ff = file->private_data;
1541         struct fuse_conn *fc = ff->fm->fc;
1542         size_t nmax = write ? fc->max_write : fc->max_read;
1543         loff_t pos = *ppos;
1544         size_t count = iov_iter_count(iter);
1545         pgoff_t idx_from = pos >> PAGE_SHIFT;
1546         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1547         ssize_t res = 0;
1548         int err = 0;
1549         struct fuse_io_args *ia;
1550         unsigned int max_pages;
1551         bool fopen_direct_io = ff->open_flags & FOPEN_DIRECT_IO;
1552
1553         max_pages = iov_iter_npages(iter, fc->max_pages);
1554         ia = fuse_io_alloc(io, max_pages);
1555         if (!ia)
1556                 return -ENOMEM;
1557
1558         if (fopen_direct_io && fc->direct_io_allow_mmap) {
1559                 res = filemap_write_and_wait_range(mapping, pos, pos + count - 1);
1560                 if (res) {
1561                         fuse_io_free(ia);
1562                         return res;
1563                 }
1564         }
1565         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1566                 if (!write)
1567                         inode_lock(inode);
1568                 fuse_sync_writes(inode);
1569                 if (!write)
1570                         inode_unlock(inode);
1571         }
1572
1573         if (fopen_direct_io && write) {
1574                 res = invalidate_inode_pages2_range(mapping, idx_from, idx_to);
1575                 if (res) {
1576                         fuse_io_free(ia);
1577                         return res;
1578                 }
1579         }
1580
1581         io->should_dirty = !write && user_backed_iter(iter);
1582         while (count) {
1583                 ssize_t nres;
1584                 fl_owner_t owner = current->files;
1585                 size_t nbytes = min(count, nmax);
1586
1587                 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1588                                           max_pages);
1589                 if (err && !nbytes)
1590                         break;
1591
1592                 if (write) {
1593                         if (!capable(CAP_FSETID))
1594                                 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1595
1596                         nres = fuse_send_write(ia, pos, nbytes, owner);
1597                 } else {
1598                         nres = fuse_send_read(ia, pos, nbytes, owner);
1599                 }
1600
1601                 if (!io->async || nres < 0) {
1602                         fuse_release_user_pages(&ia->ap, io->should_dirty);
1603                         fuse_io_free(ia);
1604                 }
1605                 ia = NULL;
1606                 if (nres < 0) {
1607                         iov_iter_revert(iter, nbytes);
1608                         err = nres;
1609                         break;
1610                 }
1611                 WARN_ON(nres > nbytes);
1612
1613                 count -= nres;
1614                 res += nres;
1615                 pos += nres;
1616                 if (nres != nbytes) {
1617                         iov_iter_revert(iter, nbytes - nres);
1618                         break;
1619                 }
1620                 if (count) {
1621                         max_pages = iov_iter_npages(iter, fc->max_pages);
1622                         ia = fuse_io_alloc(io, max_pages);
1623                         if (!ia)
1624                                 break;
1625                 }
1626         }
1627         if (ia)
1628                 fuse_io_free(ia);
1629         if (res > 0)
1630                 *ppos = pos;
1631
1632         return res > 0 ? res : err;
1633 }
1634 EXPORT_SYMBOL_GPL(fuse_direct_io);
1635
1636 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1637                                   struct iov_iter *iter,
1638                                   loff_t *ppos)
1639 {
1640         ssize_t res;
1641         struct inode *inode = file_inode(io->iocb->ki_filp);
1642
1643         res = fuse_direct_io(io, iter, ppos, 0);
1644
1645         fuse_invalidate_atime(inode);
1646
1647         return res;
1648 }
1649
1650 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1651
1652 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1653 {
1654         ssize_t res;
1655
1656         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1657                 res = fuse_direct_IO(iocb, to);
1658         } else {
1659                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1660
1661                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1662         }
1663
1664         return res;
1665 }
1666
1667 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1668 {
1669         struct inode *inode = file_inode(iocb->ki_filp);
1670         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1671         ssize_t res;
1672         bool exclusive;
1673
1674         fuse_dio_lock(iocb, from, &exclusive);
1675         res = generic_write_checks(iocb, from);
1676         if (res > 0) {
1677                 task_io_account_write(res);
1678                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1679                         res = fuse_direct_IO(iocb, from);
1680                 } else {
1681                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1682                                              FUSE_DIO_WRITE);
1683                         fuse_write_update_attr(inode, iocb->ki_pos, res);
1684                 }
1685         }
1686         fuse_dio_unlock(iocb, exclusive);
1687
1688         return res;
1689 }
1690
1691 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1692 {
1693         struct file *file = iocb->ki_filp;
1694         struct fuse_file *ff = file->private_data;
1695         struct inode *inode = file_inode(file);
1696
1697         if (fuse_is_bad(inode))
1698                 return -EIO;
1699
1700         if (FUSE_IS_DAX(inode))
1701                 return fuse_dax_read_iter(iocb, to);
1702
1703         /* FOPEN_DIRECT_IO overrides FOPEN_PASSTHROUGH */
1704         if (ff->open_flags & FOPEN_DIRECT_IO)
1705                 return fuse_direct_read_iter(iocb, to);
1706         else if (fuse_file_passthrough(ff))
1707                 return fuse_passthrough_read_iter(iocb, to);
1708         else
1709                 return fuse_cache_read_iter(iocb, to);
1710 }
1711
1712 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1713 {
1714         struct file *file = iocb->ki_filp;
1715         struct fuse_file *ff = file->private_data;
1716         struct inode *inode = file_inode(file);
1717
1718         if (fuse_is_bad(inode))
1719                 return -EIO;
1720
1721         if (FUSE_IS_DAX(inode))
1722                 return fuse_dax_write_iter(iocb, from);
1723
1724         /* FOPEN_DIRECT_IO overrides FOPEN_PASSTHROUGH */
1725         if (ff->open_flags & FOPEN_DIRECT_IO)
1726                 return fuse_direct_write_iter(iocb, from);
1727         else if (fuse_file_passthrough(ff))
1728                 return fuse_passthrough_write_iter(iocb, from);
1729         else
1730                 return fuse_cache_write_iter(iocb, from);
1731 }
1732
1733 static ssize_t fuse_splice_read(struct file *in, loff_t *ppos,
1734                                 struct pipe_inode_info *pipe, size_t len,
1735                                 unsigned int flags)
1736 {
1737         struct fuse_file *ff = in->private_data;
1738
1739         /* FOPEN_DIRECT_IO overrides FOPEN_PASSTHROUGH */
1740         if (fuse_file_passthrough(ff) && !(ff->open_flags & FOPEN_DIRECT_IO))
1741                 return fuse_passthrough_splice_read(in, ppos, pipe, len, flags);
1742         else
1743                 return filemap_splice_read(in, ppos, pipe, len, flags);
1744 }
1745
1746 static ssize_t fuse_splice_write(struct pipe_inode_info *pipe, struct file *out,
1747                                  loff_t *ppos, size_t len, unsigned int flags)
1748 {
1749         struct fuse_file *ff = out->private_data;
1750
1751         /* FOPEN_DIRECT_IO overrides FOPEN_PASSTHROUGH */
1752         if (fuse_file_passthrough(ff) && !(ff->open_flags & FOPEN_DIRECT_IO))
1753                 return fuse_passthrough_splice_write(pipe, out, ppos, len, flags);
1754         else
1755                 return iter_file_splice_write(pipe, out, ppos, len, flags);
1756 }
1757
1758 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1759 {
1760         struct fuse_args_pages *ap = &wpa->ia.ap;
1761         int i;
1762
1763         if (wpa->bucket)
1764                 fuse_sync_bucket_dec(wpa->bucket);
1765
1766         for (i = 0; i < ap->num_pages; i++)
1767                 __free_page(ap->pages[i]);
1768
1769         if (wpa->ia.ff)
1770                 fuse_file_put(wpa->ia.ff, false);
1771
1772         kfree(ap->pages);
1773         kfree(wpa);
1774 }
1775
1776 static void fuse_writepage_finish(struct fuse_mount *fm,
1777                                   struct fuse_writepage_args *wpa)
1778 {
1779         struct fuse_args_pages *ap = &wpa->ia.ap;
1780         struct inode *inode = wpa->inode;
1781         struct fuse_inode *fi = get_fuse_inode(inode);
1782         struct backing_dev_info *bdi = inode_to_bdi(inode);
1783         int i;
1784
1785         for (i = 0; i < ap->num_pages; i++) {
1786                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1787                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1788                 wb_writeout_inc(&bdi->wb);
1789         }
1790         wake_up(&fi->page_waitq);
1791 }
1792
1793 /* Called under fi->lock, may release and reacquire it */
1794 static void fuse_send_writepage(struct fuse_mount *fm,
1795                                 struct fuse_writepage_args *wpa, loff_t size)
1796 __releases(fi->lock)
1797 __acquires(fi->lock)
1798 {
1799         struct fuse_writepage_args *aux, *next;
1800         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1801         struct fuse_write_in *inarg = &wpa->ia.write.in;
1802         struct fuse_args *args = &wpa->ia.ap.args;
1803         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1804         int err;
1805
1806         fi->writectr++;
1807         if (inarg->offset + data_size <= size) {
1808                 inarg->size = data_size;
1809         } else if (inarg->offset < size) {
1810                 inarg->size = size - inarg->offset;
1811         } else {
1812                 /* Got truncated off completely */
1813                 goto out_free;
1814         }
1815
1816         args->in_args[1].size = inarg->size;
1817         args->force = true;
1818         args->nocreds = true;
1819
1820         err = fuse_simple_background(fm, args, GFP_ATOMIC);
1821         if (err == -ENOMEM) {
1822                 spin_unlock(&fi->lock);
1823                 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1824                 spin_lock(&fi->lock);
1825         }
1826
1827         /* Fails on broken connection only */
1828         if (unlikely(err))
1829                 goto out_free;
1830
1831         return;
1832
1833  out_free:
1834         fi->writectr--;
1835         rb_erase(&wpa->writepages_entry, &fi->writepages);
1836         fuse_writepage_finish(fm, wpa);
1837         spin_unlock(&fi->lock);
1838
1839         /* After fuse_writepage_finish() aux request list is private */
1840         for (aux = wpa->next; aux; aux = next) {
1841                 next = aux->next;
1842                 aux->next = NULL;
1843                 fuse_writepage_free(aux);
1844         }
1845
1846         fuse_writepage_free(wpa);
1847         spin_lock(&fi->lock);
1848 }
1849
1850 /*
1851  * If fi->writectr is positive (no truncate or fsync going on) send
1852  * all queued writepage requests.
1853  *
1854  * Called with fi->lock
1855  */
1856 void fuse_flush_writepages(struct inode *inode)
1857 __releases(fi->lock)
1858 __acquires(fi->lock)
1859 {
1860         struct fuse_mount *fm = get_fuse_mount(inode);
1861         struct fuse_inode *fi = get_fuse_inode(inode);
1862         loff_t crop = i_size_read(inode);
1863         struct fuse_writepage_args *wpa;
1864
1865         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1866                 wpa = list_entry(fi->queued_writes.next,
1867                                  struct fuse_writepage_args, queue_entry);
1868                 list_del_init(&wpa->queue_entry);
1869                 fuse_send_writepage(fm, wpa, crop);
1870         }
1871 }
1872
1873 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1874                                                 struct fuse_writepage_args *wpa)
1875 {
1876         pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1877         pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1878         struct rb_node **p = &root->rb_node;
1879         struct rb_node  *parent = NULL;
1880
1881         WARN_ON(!wpa->ia.ap.num_pages);
1882         while (*p) {
1883                 struct fuse_writepage_args *curr;
1884                 pgoff_t curr_index;
1885
1886                 parent = *p;
1887                 curr = rb_entry(parent, struct fuse_writepage_args,
1888                                 writepages_entry);
1889                 WARN_ON(curr->inode != wpa->inode);
1890                 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1891
1892                 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1893                         p = &(*p)->rb_right;
1894                 else if (idx_to < curr_index)
1895                         p = &(*p)->rb_left;
1896                 else
1897                         return curr;
1898         }
1899
1900         rb_link_node(&wpa->writepages_entry, parent, p);
1901         rb_insert_color(&wpa->writepages_entry, root);
1902         return NULL;
1903 }
1904
1905 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1906 {
1907         WARN_ON(fuse_insert_writeback(root, wpa));
1908 }
1909
1910 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1911                                int error)
1912 {
1913         struct fuse_writepage_args *wpa =
1914                 container_of(args, typeof(*wpa), ia.ap.args);
1915         struct inode *inode = wpa->inode;
1916         struct fuse_inode *fi = get_fuse_inode(inode);
1917         struct fuse_conn *fc = get_fuse_conn(inode);
1918
1919         mapping_set_error(inode->i_mapping, error);
1920         /*
1921          * A writeback finished and this might have updated mtime/ctime on
1922          * server making local mtime/ctime stale.  Hence invalidate attrs.
1923          * Do this only if writeback_cache is not enabled.  If writeback_cache
1924          * is enabled, we trust local ctime/mtime.
1925          */
1926         if (!fc->writeback_cache)
1927                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODIFY);
1928         spin_lock(&fi->lock);
1929         rb_erase(&wpa->writepages_entry, &fi->writepages);
1930         while (wpa->next) {
1931                 struct fuse_mount *fm = get_fuse_mount(inode);
1932                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1933                 struct fuse_writepage_args *next = wpa->next;
1934
1935                 wpa->next = next->next;
1936                 next->next = NULL;
1937                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1938                 tree_insert(&fi->writepages, next);
1939
1940                 /*
1941                  * Skip fuse_flush_writepages() to make it easy to crop requests
1942                  * based on primary request size.
1943                  *
1944                  * 1st case (trivial): there are no concurrent activities using
1945                  * fuse_set/release_nowrite.  Then we're on safe side because
1946                  * fuse_flush_writepages() would call fuse_send_writepage()
1947                  * anyway.
1948                  *
1949                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1950                  * now for completion of all in-flight requests.  This happens
1951                  * rarely and no more than once per page, so this should be
1952                  * okay.
1953                  *
1954                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1955                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1956                  * that fuse_set_nowrite returned implies that all in-flight
1957                  * requests were completed along with all of their secondary
1958                  * requests.  Further primary requests are blocked by negative
1959                  * writectr.  Hence there cannot be any in-flight requests and
1960                  * no invocations of fuse_writepage_end() while we're in
1961                  * fuse_set_nowrite..fuse_release_nowrite section.
1962                  */
1963                 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1964         }
1965         fi->writectr--;
1966         fuse_writepage_finish(fm, wpa);
1967         spin_unlock(&fi->lock);
1968         fuse_writepage_free(wpa);
1969 }
1970
1971 static struct fuse_file *__fuse_write_file_get(struct fuse_inode *fi)
1972 {
1973         struct fuse_file *ff;
1974
1975         spin_lock(&fi->lock);
1976         ff = list_first_entry_or_null(&fi->write_files, struct fuse_file,
1977                                       write_entry);
1978         if (ff)
1979                 fuse_file_get(ff);
1980         spin_unlock(&fi->lock);
1981
1982         return ff;
1983 }
1984
1985 static struct fuse_file *fuse_write_file_get(struct fuse_inode *fi)
1986 {
1987         struct fuse_file *ff = __fuse_write_file_get(fi);
1988         WARN_ON(!ff);
1989         return ff;
1990 }
1991
1992 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1993 {
1994         struct fuse_inode *fi = get_fuse_inode(inode);
1995         struct fuse_file *ff;
1996         int err;
1997
1998         /*
1999          * Inode is always written before the last reference is dropped and
2000          * hence this should not be reached from reclaim.
2001          *
2002          * Writing back the inode from reclaim can deadlock if the request
2003          * processing itself needs an allocation.  Allocations triggering
2004          * reclaim while serving a request can't be prevented, because it can
2005          * involve any number of unrelated userspace processes.
2006          */
2007         WARN_ON(wbc->for_reclaim);
2008
2009         ff = __fuse_write_file_get(fi);
2010         err = fuse_flush_times(inode, ff);
2011         if (ff)
2012                 fuse_file_put(ff, false);
2013
2014         return err;
2015 }
2016
2017 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
2018 {
2019         struct fuse_writepage_args *wpa;
2020         struct fuse_args_pages *ap;
2021
2022         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
2023         if (wpa) {
2024                 ap = &wpa->ia.ap;
2025                 ap->num_pages = 0;
2026                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
2027                 if (!ap->pages) {
2028                         kfree(wpa);
2029                         wpa = NULL;
2030                 }
2031         }
2032         return wpa;
2033
2034 }
2035
2036 static void fuse_writepage_add_to_bucket(struct fuse_conn *fc,
2037                                          struct fuse_writepage_args *wpa)
2038 {
2039         if (!fc->sync_fs)
2040                 return;
2041
2042         rcu_read_lock();
2043         /* Prevent resurrection of dead bucket in unlikely race with syncfs */
2044         do {
2045                 wpa->bucket = rcu_dereference(fc->curr_bucket);
2046         } while (unlikely(!atomic_inc_not_zero(&wpa->bucket->count)));
2047         rcu_read_unlock();
2048 }
2049
2050 static int fuse_writepage_locked(struct folio *folio)
2051 {
2052         struct address_space *mapping = folio->mapping;
2053         struct inode *inode = mapping->host;
2054         struct fuse_conn *fc = get_fuse_conn(inode);
2055         struct fuse_inode *fi = get_fuse_inode(inode);
2056         struct fuse_writepage_args *wpa;
2057         struct fuse_args_pages *ap;
2058         struct folio *tmp_folio;
2059         int error = -ENOMEM;
2060
2061         folio_start_writeback(folio);
2062
2063         wpa = fuse_writepage_args_alloc();
2064         if (!wpa)
2065                 goto err;
2066         ap = &wpa->ia.ap;
2067
2068         tmp_folio = folio_alloc(GFP_NOFS | __GFP_HIGHMEM, 0);
2069         if (!tmp_folio)
2070                 goto err_free;
2071
2072         error = -EIO;
2073         wpa->ia.ff = fuse_write_file_get(fi);
2074         if (!wpa->ia.ff)
2075                 goto err_nofile;
2076
2077         fuse_writepage_add_to_bucket(fc, wpa);
2078         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, folio_pos(folio), 0);
2079
2080         folio_copy(tmp_folio, folio);
2081         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2082         wpa->next = NULL;
2083         ap->args.in_pages = true;
2084         ap->num_pages = 1;
2085         ap->pages[0] = &tmp_folio->page;
2086         ap->descs[0].offset = 0;
2087         ap->descs[0].length = PAGE_SIZE;
2088         ap->args.end = fuse_writepage_end;
2089         wpa->inode = inode;
2090
2091         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2092         node_stat_add_folio(tmp_folio, NR_WRITEBACK_TEMP);
2093
2094         spin_lock(&fi->lock);
2095         tree_insert(&fi->writepages, wpa);
2096         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2097         fuse_flush_writepages(inode);
2098         spin_unlock(&fi->lock);
2099
2100         folio_end_writeback(folio);
2101
2102         return 0;
2103
2104 err_nofile:
2105         folio_put(tmp_folio);
2106 err_free:
2107         kfree(wpa);
2108 err:
2109         mapping_set_error(folio->mapping, error);
2110         folio_end_writeback(folio);
2111         return error;
2112 }
2113
2114 struct fuse_fill_wb_data {
2115         struct fuse_writepage_args *wpa;
2116         struct fuse_file *ff;
2117         struct inode *inode;
2118         struct page **orig_pages;
2119         unsigned int max_pages;
2120 };
2121
2122 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
2123 {
2124         struct fuse_args_pages *ap = &data->wpa->ia.ap;
2125         struct fuse_conn *fc = get_fuse_conn(data->inode);
2126         struct page **pages;
2127         struct fuse_page_desc *descs;
2128         unsigned int npages = min_t(unsigned int,
2129                                     max_t(unsigned int, data->max_pages * 2,
2130                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
2131                                     fc->max_pages);
2132         WARN_ON(npages <= data->max_pages);
2133
2134         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
2135         if (!pages)
2136                 return false;
2137
2138         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
2139         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
2140         kfree(ap->pages);
2141         ap->pages = pages;
2142         ap->descs = descs;
2143         data->max_pages = npages;
2144
2145         return true;
2146 }
2147
2148 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
2149 {
2150         struct fuse_writepage_args *wpa = data->wpa;
2151         struct inode *inode = data->inode;
2152         struct fuse_inode *fi = get_fuse_inode(inode);
2153         int num_pages = wpa->ia.ap.num_pages;
2154         int i;
2155
2156         wpa->ia.ff = fuse_file_get(data->ff);
2157         spin_lock(&fi->lock);
2158         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2159         fuse_flush_writepages(inode);
2160         spin_unlock(&fi->lock);
2161
2162         for (i = 0; i < num_pages; i++)
2163                 end_page_writeback(data->orig_pages[i]);
2164 }
2165
2166 /*
2167  * Check under fi->lock if the page is under writeback, and insert it onto the
2168  * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2169  * one already added for a page at this offset.  If there's none, then insert
2170  * this new request onto the auxiliary list, otherwise reuse the existing one by
2171  * swapping the new temp page with the old one.
2172  */
2173 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2174                                struct page *page)
2175 {
2176         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2177         struct fuse_writepage_args *tmp;
2178         struct fuse_writepage_args *old_wpa;
2179         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2180
2181         WARN_ON(new_ap->num_pages != 0);
2182         new_ap->num_pages = 1;
2183
2184         spin_lock(&fi->lock);
2185         old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2186         if (!old_wpa) {
2187                 spin_unlock(&fi->lock);
2188                 return true;
2189         }
2190
2191         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2192                 pgoff_t curr_index;
2193
2194                 WARN_ON(tmp->inode != new_wpa->inode);
2195                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2196                 if (curr_index == page->index) {
2197                         WARN_ON(tmp->ia.ap.num_pages != 1);
2198                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2199                         break;
2200                 }
2201         }
2202
2203         if (!tmp) {
2204                 new_wpa->next = old_wpa->next;
2205                 old_wpa->next = new_wpa;
2206         }
2207
2208         spin_unlock(&fi->lock);
2209
2210         if (tmp) {
2211                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2212
2213                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2214                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2215                 wb_writeout_inc(&bdi->wb);
2216                 fuse_writepage_free(new_wpa);
2217         }
2218
2219         return false;
2220 }
2221
2222 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2223                                      struct fuse_args_pages *ap,
2224                                      struct fuse_fill_wb_data *data)
2225 {
2226         WARN_ON(!ap->num_pages);
2227
2228         /*
2229          * Being under writeback is unlikely but possible.  For example direct
2230          * read to an mmaped fuse file will set the page dirty twice; once when
2231          * the pages are faulted with get_user_pages(), and then after the read
2232          * completed.
2233          */
2234         if (fuse_page_is_writeback(data->inode, page->index))
2235                 return true;
2236
2237         /* Reached max pages */
2238         if (ap->num_pages == fc->max_pages)
2239                 return true;
2240
2241         /* Reached max write bytes */
2242         if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2243                 return true;
2244
2245         /* Discontinuity */
2246         if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2247                 return true;
2248
2249         /* Need to grow the pages array?  If so, did the expansion fail? */
2250         if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2251                 return true;
2252
2253         return false;
2254 }
2255
2256 static int fuse_writepages_fill(struct folio *folio,
2257                 struct writeback_control *wbc, void *_data)
2258 {
2259         struct fuse_fill_wb_data *data = _data;
2260         struct fuse_writepage_args *wpa = data->wpa;
2261         struct fuse_args_pages *ap = &wpa->ia.ap;
2262         struct inode *inode = data->inode;
2263         struct fuse_inode *fi = get_fuse_inode(inode);
2264         struct fuse_conn *fc = get_fuse_conn(inode);
2265         struct page *tmp_page;
2266         int err;
2267
2268         if (!data->ff) {
2269                 err = -EIO;
2270                 data->ff = fuse_write_file_get(fi);
2271                 if (!data->ff)
2272                         goto out_unlock;
2273         }
2274
2275         if (wpa && fuse_writepage_need_send(fc, &folio->page, ap, data)) {
2276                 fuse_writepages_send(data);
2277                 data->wpa = NULL;
2278         }
2279
2280         err = -ENOMEM;
2281         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2282         if (!tmp_page)
2283                 goto out_unlock;
2284
2285         /*
2286          * The page must not be redirtied until the writeout is completed
2287          * (i.e. userspace has sent a reply to the write request).  Otherwise
2288          * there could be more than one temporary page instance for each real
2289          * page.
2290          *
2291          * This is ensured by holding the page lock in page_mkwrite() while
2292          * checking fuse_page_is_writeback().  We already hold the page lock
2293          * since clear_page_dirty_for_io() and keep it held until we add the
2294          * request to the fi->writepages list and increment ap->num_pages.
2295          * After this fuse_page_is_writeback() will indicate that the page is
2296          * under writeback, so we can release the page lock.
2297          */
2298         if (data->wpa == NULL) {
2299                 err = -ENOMEM;
2300                 wpa = fuse_writepage_args_alloc();
2301                 if (!wpa) {
2302                         __free_page(tmp_page);
2303                         goto out_unlock;
2304                 }
2305                 fuse_writepage_add_to_bucket(fc, wpa);
2306
2307                 data->max_pages = 1;
2308
2309                 ap = &wpa->ia.ap;
2310                 fuse_write_args_fill(&wpa->ia, data->ff, folio_pos(folio), 0);
2311                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2312                 wpa->next = NULL;
2313                 ap->args.in_pages = true;
2314                 ap->args.end = fuse_writepage_end;
2315                 ap->num_pages = 0;
2316                 wpa->inode = inode;
2317         }
2318         folio_start_writeback(folio);
2319
2320         copy_highpage(tmp_page, &folio->page);
2321         ap->pages[ap->num_pages] = tmp_page;
2322         ap->descs[ap->num_pages].offset = 0;
2323         ap->descs[ap->num_pages].length = PAGE_SIZE;
2324         data->orig_pages[ap->num_pages] = &folio->page;
2325
2326         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2327         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2328
2329         err = 0;
2330         if (data->wpa) {
2331                 /*
2332                  * Protected by fi->lock against concurrent access by
2333                  * fuse_page_is_writeback().
2334                  */
2335                 spin_lock(&fi->lock);
2336                 ap->num_pages++;
2337                 spin_unlock(&fi->lock);
2338         } else if (fuse_writepage_add(wpa, &folio->page)) {
2339                 data->wpa = wpa;
2340         } else {
2341                 folio_end_writeback(folio);
2342         }
2343 out_unlock:
2344         folio_unlock(folio);
2345
2346         return err;
2347 }
2348
2349 static int fuse_writepages(struct address_space *mapping,
2350                            struct writeback_control *wbc)
2351 {
2352         struct inode *inode = mapping->host;
2353         struct fuse_conn *fc = get_fuse_conn(inode);
2354         struct fuse_fill_wb_data data;
2355         int err;
2356
2357         err = -EIO;
2358         if (fuse_is_bad(inode))
2359                 goto out;
2360
2361         if (wbc->sync_mode == WB_SYNC_NONE &&
2362             fc->num_background >= fc->congestion_threshold)
2363                 return 0;
2364
2365         data.inode = inode;
2366         data.wpa = NULL;
2367         data.ff = NULL;
2368
2369         err = -ENOMEM;
2370         data.orig_pages = kcalloc(fc->max_pages,
2371                                   sizeof(struct page *),
2372                                   GFP_NOFS);
2373         if (!data.orig_pages)
2374                 goto out;
2375
2376         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2377         if (data.wpa) {
2378                 WARN_ON(!data.wpa->ia.ap.num_pages);
2379                 fuse_writepages_send(&data);
2380         }
2381         if (data.ff)
2382                 fuse_file_put(data.ff, false);
2383
2384         kfree(data.orig_pages);
2385 out:
2386         return err;
2387 }
2388
2389 /*
2390  * It's worthy to make sure that space is reserved on disk for the write,
2391  * but how to implement it without killing performance need more thinking.
2392  */
2393 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2394                 loff_t pos, unsigned len, struct page **pagep, void **fsdata)
2395 {
2396         pgoff_t index = pos >> PAGE_SHIFT;
2397         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2398         struct page *page;
2399         loff_t fsize;
2400         int err = -ENOMEM;
2401
2402         WARN_ON(!fc->writeback_cache);
2403
2404         page = grab_cache_page_write_begin(mapping, index);
2405         if (!page)
2406                 goto error;
2407
2408         fuse_wait_on_page_writeback(mapping->host, page->index);
2409
2410         if (PageUptodate(page) || len == PAGE_SIZE)
2411                 goto success;
2412         /*
2413          * Check if the start this page comes after the end of file, in which
2414          * case the readpage can be optimized away.
2415          */
2416         fsize = i_size_read(mapping->host);
2417         if (fsize <= (pos & PAGE_MASK)) {
2418                 size_t off = pos & ~PAGE_MASK;
2419                 if (off)
2420                         zero_user_segment(page, 0, off);
2421                 goto success;
2422         }
2423         err = fuse_do_readpage(file, page);
2424         if (err)
2425                 goto cleanup;
2426 success:
2427         *pagep = page;
2428         return 0;
2429
2430 cleanup:
2431         unlock_page(page);
2432         put_page(page);
2433 error:
2434         return err;
2435 }
2436
2437 static int fuse_write_end(struct file *file, struct address_space *mapping,
2438                 loff_t pos, unsigned len, unsigned copied,
2439                 struct page *page, void *fsdata)
2440 {
2441         struct inode *inode = page->mapping->host;
2442
2443         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2444         if (!copied)
2445                 goto unlock;
2446
2447         pos += copied;
2448         if (!PageUptodate(page)) {
2449                 /* Zero any unwritten bytes at the end of the page */
2450                 size_t endoff = pos & ~PAGE_MASK;
2451                 if (endoff)
2452                         zero_user_segment(page, endoff, PAGE_SIZE);
2453                 SetPageUptodate(page);
2454         }
2455
2456         if (pos > inode->i_size)
2457                 i_size_write(inode, pos);
2458
2459         set_page_dirty(page);
2460
2461 unlock:
2462         unlock_page(page);
2463         put_page(page);
2464
2465         return copied;
2466 }
2467
2468 static int fuse_launder_folio(struct folio *folio)
2469 {
2470         int err = 0;
2471         if (folio_clear_dirty_for_io(folio)) {
2472                 struct inode *inode = folio->mapping->host;
2473
2474                 /* Serialize with pending writeback for the same page */
2475                 fuse_wait_on_page_writeback(inode, folio->index);
2476                 err = fuse_writepage_locked(folio);
2477                 if (!err)
2478                         fuse_wait_on_page_writeback(inode, folio->index);
2479         }
2480         return err;
2481 }
2482
2483 /*
2484  * Write back dirty data/metadata now (there may not be any suitable
2485  * open files later for data)
2486  */
2487 static void fuse_vma_close(struct vm_area_struct *vma)
2488 {
2489         int err;
2490
2491         err = write_inode_now(vma->vm_file->f_mapping->host, 1);
2492         mapping_set_error(vma->vm_file->f_mapping, err);
2493 }
2494
2495 /*
2496  * Wait for writeback against this page to complete before allowing it
2497  * to be marked dirty again, and hence written back again, possibly
2498  * before the previous writepage completed.
2499  *
2500  * Block here, instead of in ->writepage(), so that the userspace fs
2501  * can only block processes actually operating on the filesystem.
2502  *
2503  * Otherwise unprivileged userspace fs would be able to block
2504  * unrelated:
2505  *
2506  * - page migration
2507  * - sync(2)
2508  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2509  */
2510 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2511 {
2512         struct page *page = vmf->page;
2513         struct inode *inode = file_inode(vmf->vma->vm_file);
2514
2515         file_update_time(vmf->vma->vm_file);
2516         lock_page(page);
2517         if (page->mapping != inode->i_mapping) {
2518                 unlock_page(page);
2519                 return VM_FAULT_NOPAGE;
2520         }
2521
2522         fuse_wait_on_page_writeback(inode, page->index);
2523         return VM_FAULT_LOCKED;
2524 }
2525
2526 static const struct vm_operations_struct fuse_file_vm_ops = {
2527         .close          = fuse_vma_close,
2528         .fault          = filemap_fault,
2529         .map_pages      = filemap_map_pages,
2530         .page_mkwrite   = fuse_page_mkwrite,
2531 };
2532
2533 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2534 {
2535         struct fuse_file *ff = file->private_data;
2536         struct fuse_conn *fc = ff->fm->fc;
2537         struct inode *inode = file_inode(file);
2538         int rc;
2539
2540         /* DAX mmap is superior to direct_io mmap */
2541         if (FUSE_IS_DAX(inode))
2542                 return fuse_dax_mmap(file, vma);
2543
2544         /*
2545          * If inode is in passthrough io mode, because it has some file open
2546          * in passthrough mode, either mmap to backing file or fail mmap,
2547          * because mixing cached mmap and passthrough io mode is not allowed.
2548          */
2549         if (fuse_file_passthrough(ff))
2550                 return fuse_passthrough_mmap(file, vma);
2551         else if (fuse_inode_backing(get_fuse_inode(inode)))
2552                 return -ENODEV;
2553
2554         /*
2555          * FOPEN_DIRECT_IO handling is special compared to O_DIRECT,
2556          * as does not allow MAP_SHARED mmap without FUSE_DIRECT_IO_ALLOW_MMAP.
2557          */
2558         if (ff->open_flags & FOPEN_DIRECT_IO) {
2559                 /*
2560                  * Can't provide the coherency needed for MAP_SHARED
2561                  * if FUSE_DIRECT_IO_ALLOW_MMAP isn't set.
2562                  */
2563                 if ((vma->vm_flags & VM_MAYSHARE) && !fc->direct_io_allow_mmap)
2564                         return -ENODEV;
2565
2566                 invalidate_inode_pages2(file->f_mapping);
2567
2568                 if (!(vma->vm_flags & VM_MAYSHARE)) {
2569                         /* MAP_PRIVATE */
2570                         return generic_file_mmap(file, vma);
2571                 }
2572
2573                 /*
2574                  * First mmap of direct_io file enters caching inode io mode.
2575                  * Also waits for parallel dio writers to go into serial mode
2576                  * (exclusive instead of shared lock).
2577                  * After first mmap, the inode stays in caching io mode until
2578                  * the direct_io file release.
2579                  */
2580                 rc = fuse_file_cached_io_open(inode, ff);
2581                 if (rc)
2582                         return rc;
2583         }
2584
2585         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2586                 fuse_link_write_file(file);
2587
2588         file_accessed(file);
2589         vma->vm_ops = &fuse_file_vm_ops;
2590         return 0;
2591 }
2592
2593 static int convert_fuse_file_lock(struct fuse_conn *fc,
2594                                   const struct fuse_file_lock *ffl,
2595                                   struct file_lock *fl)
2596 {
2597         switch (ffl->type) {
2598         case F_UNLCK:
2599                 break;
2600
2601         case F_RDLCK:
2602         case F_WRLCK:
2603                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2604                     ffl->end < ffl->start)
2605                         return -EIO;
2606
2607                 fl->fl_start = ffl->start;
2608                 fl->fl_end = ffl->end;
2609
2610                 /*
2611                  * Convert pid into init's pid namespace.  The locks API will
2612                  * translate it into the caller's pid namespace.
2613                  */
2614                 rcu_read_lock();
2615                 fl->c.flc_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2616                 rcu_read_unlock();
2617                 break;
2618
2619         default:
2620                 return -EIO;
2621         }
2622         fl->c.flc_type = ffl->type;
2623         return 0;
2624 }
2625
2626 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2627                          const struct file_lock *fl, int opcode, pid_t pid,
2628                          int flock, struct fuse_lk_in *inarg)
2629 {
2630         struct inode *inode = file_inode(file);
2631         struct fuse_conn *fc = get_fuse_conn(inode);
2632         struct fuse_file *ff = file->private_data;
2633
2634         memset(inarg, 0, sizeof(*inarg));
2635         inarg->fh = ff->fh;
2636         inarg->owner = fuse_lock_owner_id(fc, fl->c.flc_owner);
2637         inarg->lk.start = fl->fl_start;
2638         inarg->lk.end = fl->fl_end;
2639         inarg->lk.type = fl->c.flc_type;
2640         inarg->lk.pid = pid;
2641         if (flock)
2642                 inarg->lk_flags |= FUSE_LK_FLOCK;
2643         args->opcode = opcode;
2644         args->nodeid = get_node_id(inode);
2645         args->in_numargs = 1;
2646         args->in_args[0].size = sizeof(*inarg);
2647         args->in_args[0].value = inarg;
2648 }
2649
2650 static int fuse_getlk(struct file *file, struct file_lock *fl)
2651 {
2652         struct inode *inode = file_inode(file);
2653         struct fuse_mount *fm = get_fuse_mount(inode);
2654         FUSE_ARGS(args);
2655         struct fuse_lk_in inarg;
2656         struct fuse_lk_out outarg;
2657         int err;
2658
2659         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2660         args.out_numargs = 1;
2661         args.out_args[0].size = sizeof(outarg);
2662         args.out_args[0].value = &outarg;
2663         err = fuse_simple_request(fm, &args);
2664         if (!err)
2665                 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2666
2667         return err;
2668 }
2669
2670 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2671 {
2672         struct inode *inode = file_inode(file);
2673         struct fuse_mount *fm = get_fuse_mount(inode);
2674         FUSE_ARGS(args);
2675         struct fuse_lk_in inarg;
2676         int opcode = (fl->c.flc_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2677         struct pid *pid = fl->c.flc_type != F_UNLCK ? task_tgid(current) : NULL;
2678         pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2679         int err;
2680
2681         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2682                 /* NLM needs asynchronous locks, which we don't support yet */
2683                 return -ENOLCK;
2684         }
2685
2686         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2687         err = fuse_simple_request(fm, &args);
2688
2689         /* locking is restartable */
2690         if (err == -EINTR)
2691                 err = -ERESTARTSYS;
2692
2693         return err;
2694 }
2695
2696 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2697 {
2698         struct inode *inode = file_inode(file);
2699         struct fuse_conn *fc = get_fuse_conn(inode);
2700         int err;
2701
2702         if (cmd == F_CANCELLK) {
2703                 err = 0;
2704         } else if (cmd == F_GETLK) {
2705                 if (fc->no_lock) {
2706                         posix_test_lock(file, fl);
2707                         err = 0;
2708                 } else
2709                         err = fuse_getlk(file, fl);
2710         } else {
2711                 if (fc->no_lock)
2712                         err = posix_lock_file(file, fl, NULL);
2713                 else
2714                         err = fuse_setlk(file, fl, 0);
2715         }
2716         return err;
2717 }
2718
2719 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2720 {
2721         struct inode *inode = file_inode(file);
2722         struct fuse_conn *fc = get_fuse_conn(inode);
2723         int err;
2724
2725         if (fc->no_flock) {
2726                 err = locks_lock_file_wait(file, fl);
2727         } else {
2728                 struct fuse_file *ff = file->private_data;
2729
2730                 /* emulate flock with POSIX locks */
2731                 ff->flock = true;
2732                 err = fuse_setlk(file, fl, 1);
2733         }
2734
2735         return err;
2736 }
2737
2738 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2739 {
2740         struct inode *inode = mapping->host;
2741         struct fuse_mount *fm = get_fuse_mount(inode);
2742         FUSE_ARGS(args);
2743         struct fuse_bmap_in inarg;
2744         struct fuse_bmap_out outarg;
2745         int err;
2746
2747         if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2748                 return 0;
2749
2750         memset(&inarg, 0, sizeof(inarg));
2751         inarg.block = block;
2752         inarg.blocksize = inode->i_sb->s_blocksize;
2753         args.opcode = FUSE_BMAP;
2754         args.nodeid = get_node_id(inode);
2755         args.in_numargs = 1;
2756         args.in_args[0].size = sizeof(inarg);
2757         args.in_args[0].value = &inarg;
2758         args.out_numargs = 1;
2759         args.out_args[0].size = sizeof(outarg);
2760         args.out_args[0].value = &outarg;
2761         err = fuse_simple_request(fm, &args);
2762         if (err == -ENOSYS)
2763                 fm->fc->no_bmap = 1;
2764
2765         return err ? 0 : outarg.block;
2766 }
2767
2768 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2769 {
2770         struct inode *inode = file->f_mapping->host;
2771         struct fuse_mount *fm = get_fuse_mount(inode);
2772         struct fuse_file *ff = file->private_data;
2773         FUSE_ARGS(args);
2774         struct fuse_lseek_in inarg = {
2775                 .fh = ff->fh,
2776                 .offset = offset,
2777                 .whence = whence
2778         };
2779         struct fuse_lseek_out outarg;
2780         int err;
2781
2782         if (fm->fc->no_lseek)
2783                 goto fallback;
2784
2785         args.opcode = FUSE_LSEEK;
2786         args.nodeid = ff->nodeid;
2787         args.in_numargs = 1;
2788         args.in_args[0].size = sizeof(inarg);
2789         args.in_args[0].value = &inarg;
2790         args.out_numargs = 1;
2791         args.out_args[0].size = sizeof(outarg);
2792         args.out_args[0].value = &outarg;
2793         err = fuse_simple_request(fm, &args);
2794         if (err) {
2795                 if (err == -ENOSYS) {
2796                         fm->fc->no_lseek = 1;
2797                         goto fallback;
2798                 }
2799                 return err;
2800         }
2801
2802         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2803
2804 fallback:
2805         err = fuse_update_attributes(inode, file, STATX_SIZE);
2806         if (!err)
2807                 return generic_file_llseek(file, offset, whence);
2808         else
2809                 return err;
2810 }
2811
2812 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2813 {
2814         loff_t retval;
2815         struct inode *inode = file_inode(file);
2816
2817         switch (whence) {
2818         case SEEK_SET:
2819         case SEEK_CUR:
2820                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2821                 retval = generic_file_llseek(file, offset, whence);
2822                 break;
2823         case SEEK_END:
2824                 inode_lock(inode);
2825                 retval = fuse_update_attributes(inode, file, STATX_SIZE);
2826                 if (!retval)
2827                         retval = generic_file_llseek(file, offset, whence);
2828                 inode_unlock(inode);
2829                 break;
2830         case SEEK_HOLE:
2831         case SEEK_DATA:
2832                 inode_lock(inode);
2833                 retval = fuse_lseek(file, offset, whence);
2834                 inode_unlock(inode);
2835                 break;
2836         default:
2837                 retval = -EINVAL;
2838         }
2839
2840         return retval;
2841 }
2842
2843 /*
2844  * All files which have been polled are linked to RB tree
2845  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2846  * find the matching one.
2847  */
2848 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2849                                               struct rb_node **parent_out)
2850 {
2851         struct rb_node **link = &fc->polled_files.rb_node;
2852         struct rb_node *last = NULL;
2853
2854         while (*link) {
2855                 struct fuse_file *ff;
2856
2857                 last = *link;
2858                 ff = rb_entry(last, struct fuse_file, polled_node);
2859
2860                 if (kh < ff->kh)
2861                         link = &last->rb_left;
2862                 else if (kh > ff->kh)
2863                         link = &last->rb_right;
2864                 else
2865                         return link;
2866         }
2867
2868         if (parent_out)
2869                 *parent_out = last;
2870         return link;
2871 }
2872
2873 /*
2874  * The file is about to be polled.  Make sure it's on the polled_files
2875  * RB tree.  Note that files once added to the polled_files tree are
2876  * not removed before the file is released.  This is because a file
2877  * polled once is likely to be polled again.
2878  */
2879 static void fuse_register_polled_file(struct fuse_conn *fc,
2880                                       struct fuse_file *ff)
2881 {
2882         spin_lock(&fc->lock);
2883         if (RB_EMPTY_NODE(&ff->polled_node)) {
2884                 struct rb_node **link, *parent;
2885
2886                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2887                 BUG_ON(*link);
2888                 rb_link_node(&ff->polled_node, parent, link);
2889                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2890         }
2891         spin_unlock(&fc->lock);
2892 }
2893
2894 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2895 {
2896         struct fuse_file *ff = file->private_data;
2897         struct fuse_mount *fm = ff->fm;
2898         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2899         struct fuse_poll_out outarg;
2900         FUSE_ARGS(args);
2901         int err;
2902
2903         if (fm->fc->no_poll)
2904                 return DEFAULT_POLLMASK;
2905
2906         poll_wait(file, &ff->poll_wait, wait);
2907         inarg.events = mangle_poll(poll_requested_events(wait));
2908
2909         /*
2910          * Ask for notification iff there's someone waiting for it.
2911          * The client may ignore the flag and always notify.
2912          */
2913         if (waitqueue_active(&ff->poll_wait)) {
2914                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2915                 fuse_register_polled_file(fm->fc, ff);
2916         }
2917
2918         args.opcode = FUSE_POLL;
2919         args.nodeid = ff->nodeid;
2920         args.in_numargs = 1;
2921         args.in_args[0].size = sizeof(inarg);
2922         args.in_args[0].value = &inarg;
2923         args.out_numargs = 1;
2924         args.out_args[0].size = sizeof(outarg);
2925         args.out_args[0].value = &outarg;
2926         err = fuse_simple_request(fm, &args);
2927
2928         if (!err)
2929                 return demangle_poll(outarg.revents);
2930         if (err == -ENOSYS) {
2931                 fm->fc->no_poll = 1;
2932                 return DEFAULT_POLLMASK;
2933         }
2934         return EPOLLERR;
2935 }
2936 EXPORT_SYMBOL_GPL(fuse_file_poll);
2937
2938 /*
2939  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2940  * wakes up the poll waiters.
2941  */
2942 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2943                             struct fuse_notify_poll_wakeup_out *outarg)
2944 {
2945         u64 kh = outarg->kh;
2946         struct rb_node **link;
2947
2948         spin_lock(&fc->lock);
2949
2950         link = fuse_find_polled_node(fc, kh, NULL);
2951         if (*link) {
2952                 struct fuse_file *ff;
2953
2954                 ff = rb_entry(*link, struct fuse_file, polled_node);
2955                 wake_up_interruptible_sync(&ff->poll_wait);
2956         }
2957
2958         spin_unlock(&fc->lock);
2959         return 0;
2960 }
2961
2962 static void fuse_do_truncate(struct file *file)
2963 {
2964         struct inode *inode = file->f_mapping->host;
2965         struct iattr attr;
2966
2967         attr.ia_valid = ATTR_SIZE;
2968         attr.ia_size = i_size_read(inode);
2969
2970         attr.ia_file = file;
2971         attr.ia_valid |= ATTR_FILE;
2972
2973         fuse_do_setattr(file_dentry(file), &attr, file);
2974 }
2975
2976 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2977 {
2978         return round_up(off, fc->max_pages << PAGE_SHIFT);
2979 }
2980
2981 static ssize_t
2982 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2983 {
2984         DECLARE_COMPLETION_ONSTACK(wait);
2985         ssize_t ret = 0;
2986         struct file *file = iocb->ki_filp;
2987         struct fuse_file *ff = file->private_data;
2988         loff_t pos = 0;
2989         struct inode *inode;
2990         loff_t i_size;
2991         size_t count = iov_iter_count(iter), shortened = 0;
2992         loff_t offset = iocb->ki_pos;
2993         struct fuse_io_priv *io;
2994
2995         pos = offset;
2996         inode = file->f_mapping->host;
2997         i_size = i_size_read(inode);
2998
2999         if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
3000                 return 0;
3001
3002         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
3003         if (!io)
3004                 return -ENOMEM;
3005         spin_lock_init(&io->lock);
3006         kref_init(&io->refcnt);
3007         io->reqs = 1;
3008         io->bytes = -1;
3009         io->size = 0;
3010         io->offset = offset;
3011         io->write = (iov_iter_rw(iter) == WRITE);
3012         io->err = 0;
3013         /*
3014          * By default, we want to optimize all I/Os with async request
3015          * submission to the client filesystem if supported.
3016          */
3017         io->async = ff->fm->fc->async_dio;
3018         io->iocb = iocb;
3019         io->blocking = is_sync_kiocb(iocb);
3020
3021         /* optimization for short read */
3022         if (io->async && !io->write && offset + count > i_size) {
3023                 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
3024                 shortened = count - iov_iter_count(iter);
3025                 count -= shortened;
3026         }
3027
3028         /*
3029          * We cannot asynchronously extend the size of a file.
3030          * In such case the aio will behave exactly like sync io.
3031          */
3032         if ((offset + count > i_size) && io->write)
3033                 io->blocking = true;
3034
3035         if (io->async && io->blocking) {
3036                 /*
3037                  * Additional reference to keep io around after
3038                  * calling fuse_aio_complete()
3039                  */
3040                 kref_get(&io->refcnt);
3041                 io->done = &wait;
3042         }
3043
3044         if (iov_iter_rw(iter) == WRITE) {
3045                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
3046                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
3047         } else {
3048                 ret = __fuse_direct_read(io, iter, &pos);
3049         }
3050         iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
3051
3052         if (io->async) {
3053                 bool blocking = io->blocking;
3054
3055                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3056
3057                 /* we have a non-extending, async request, so return */
3058                 if (!blocking)
3059                         return -EIOCBQUEUED;
3060
3061                 wait_for_completion(&wait);
3062                 ret = fuse_get_res_by_io(io);
3063         }
3064
3065         kref_put(&io->refcnt, fuse_io_release);
3066
3067         if (iov_iter_rw(iter) == WRITE) {
3068                 fuse_write_update_attr(inode, pos, ret);
3069                 /* For extending writes we already hold exclusive lock */
3070                 if (ret < 0 && offset + count > i_size)
3071                         fuse_do_truncate(file);
3072         }
3073
3074         return ret;
3075 }
3076
3077 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3078 {
3079         int err = filemap_write_and_wait_range(inode->i_mapping, start, LLONG_MAX);
3080
3081         if (!err)
3082                 fuse_sync_writes(inode);
3083
3084         return err;
3085 }
3086
3087 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3088                                 loff_t length)
3089 {
3090         struct fuse_file *ff = file->private_data;
3091         struct inode *inode = file_inode(file);
3092         struct fuse_inode *fi = get_fuse_inode(inode);
3093         struct fuse_mount *fm = ff->fm;
3094         FUSE_ARGS(args);
3095         struct fuse_fallocate_in inarg = {
3096                 .fh = ff->fh,
3097                 .offset = offset,
3098                 .length = length,
3099                 .mode = mode
3100         };
3101         int err;
3102         bool block_faults = FUSE_IS_DAX(inode) &&
3103                 (!(mode & FALLOC_FL_KEEP_SIZE) ||
3104                  (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)));
3105
3106         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3107                      FALLOC_FL_ZERO_RANGE))
3108                 return -EOPNOTSUPP;
3109
3110         if (fm->fc->no_fallocate)
3111                 return -EOPNOTSUPP;
3112
3113         inode_lock(inode);
3114         if (block_faults) {
3115                 filemap_invalidate_lock(inode->i_mapping);
3116                 err = fuse_dax_break_layouts(inode, 0, 0);
3117                 if (err)
3118                         goto out;
3119         }
3120
3121         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)) {
3122                 loff_t endbyte = offset + length - 1;
3123
3124                 err = fuse_writeback_range(inode, offset, endbyte);
3125                 if (err)
3126                         goto out;
3127         }
3128
3129         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3130             offset + length > i_size_read(inode)) {
3131                 err = inode_newsize_ok(inode, offset + length);
3132                 if (err)
3133                         goto out;
3134         }
3135
3136         err = file_modified(file);
3137         if (err)
3138                 goto out;
3139
3140         if (!(mode & FALLOC_FL_KEEP_SIZE))
3141                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3142
3143         args.opcode = FUSE_FALLOCATE;
3144         args.nodeid = ff->nodeid;
3145         args.in_numargs = 1;
3146         args.in_args[0].size = sizeof(inarg);
3147         args.in_args[0].value = &inarg;
3148         err = fuse_simple_request(fm, &args);
3149         if (err == -ENOSYS) {
3150                 fm->fc->no_fallocate = 1;
3151                 err = -EOPNOTSUPP;
3152         }
3153         if (err)
3154                 goto out;
3155
3156         /* we could have extended the file */
3157         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3158                 if (fuse_write_update_attr(inode, offset + length, length))
3159                         file_update_time(file);
3160         }
3161
3162         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE))
3163                 truncate_pagecache_range(inode, offset, offset + length - 1);
3164
3165         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
3166
3167 out:
3168         if (!(mode & FALLOC_FL_KEEP_SIZE))
3169                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3170
3171         if (block_faults)
3172                 filemap_invalidate_unlock(inode->i_mapping);
3173
3174         inode_unlock(inode);
3175
3176         fuse_flush_time_update(inode);
3177
3178         return err;
3179 }
3180
3181 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3182                                       struct file *file_out, loff_t pos_out,
3183                                       size_t len, unsigned int flags)
3184 {
3185         struct fuse_file *ff_in = file_in->private_data;
3186         struct fuse_file *ff_out = file_out->private_data;
3187         struct inode *inode_in = file_inode(file_in);
3188         struct inode *inode_out = file_inode(file_out);
3189         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3190         struct fuse_mount *fm = ff_in->fm;
3191         struct fuse_conn *fc = fm->fc;
3192         FUSE_ARGS(args);
3193         struct fuse_copy_file_range_in inarg = {
3194                 .fh_in = ff_in->fh,
3195                 .off_in = pos_in,
3196                 .nodeid_out = ff_out->nodeid,
3197                 .fh_out = ff_out->fh,
3198                 .off_out = pos_out,
3199                 .len = len,
3200                 .flags = flags
3201         };
3202         struct fuse_write_out outarg;
3203         ssize_t err;
3204         /* mark unstable when write-back is not used, and file_out gets
3205          * extended */
3206         bool is_unstable = (!fc->writeback_cache) &&
3207                            ((pos_out + len) > inode_out->i_size);
3208
3209         if (fc->no_copy_file_range)
3210                 return -EOPNOTSUPP;
3211
3212         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3213                 return -EXDEV;
3214
3215         inode_lock(inode_in);
3216         err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3217         inode_unlock(inode_in);
3218         if (err)
3219                 return err;
3220
3221         inode_lock(inode_out);
3222
3223         err = file_modified(file_out);
3224         if (err)
3225                 goto out;
3226
3227         /*
3228          * Write out dirty pages in the destination file before sending the COPY
3229          * request to userspace.  After the request is completed, truncate off
3230          * pages (including partial ones) from the cache that have been copied,
3231          * since these contain stale data at that point.
3232          *
3233          * This should be mostly correct, but if the COPY writes to partial
3234          * pages (at the start or end) and the parts not covered by the COPY are
3235          * written through a memory map after calling fuse_writeback_range(),
3236          * then these partial page modifications will be lost on truncation.
3237          *
3238          * It is unlikely that someone would rely on such mixed style
3239          * modifications.  Yet this does give less guarantees than if the
3240          * copying was performed with write(2).
3241          *
3242          * To fix this a mapping->invalidate_lock could be used to prevent new
3243          * faults while the copy is ongoing.
3244          */
3245         err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3246         if (err)
3247                 goto out;
3248
3249         if (is_unstable)
3250                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3251
3252         args.opcode = FUSE_COPY_FILE_RANGE;
3253         args.nodeid = ff_in->nodeid;
3254         args.in_numargs = 1;
3255         args.in_args[0].size = sizeof(inarg);
3256         args.in_args[0].value = &inarg;
3257         args.out_numargs = 1;
3258         args.out_args[0].size = sizeof(outarg);
3259         args.out_args[0].value = &outarg;
3260         err = fuse_simple_request(fm, &args);
3261         if (err == -ENOSYS) {
3262                 fc->no_copy_file_range = 1;
3263                 err = -EOPNOTSUPP;
3264         }
3265         if (err)
3266                 goto out;
3267
3268         truncate_inode_pages_range(inode_out->i_mapping,
3269                                    ALIGN_DOWN(pos_out, PAGE_SIZE),
3270                                    ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3271
3272         file_update_time(file_out);
3273         fuse_write_update_attr(inode_out, pos_out + outarg.size, outarg.size);
3274
3275         err = outarg.size;
3276 out:
3277         if (is_unstable)
3278                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3279
3280         inode_unlock(inode_out);
3281         file_accessed(file_in);
3282
3283         fuse_flush_time_update(inode_out);
3284
3285         return err;
3286 }
3287
3288 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3289                                     struct file *dst_file, loff_t dst_off,
3290                                     size_t len, unsigned int flags)
3291 {
3292         ssize_t ret;
3293
3294         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3295                                      len, flags);
3296
3297         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3298                 ret = splice_copy_file_range(src_file, src_off, dst_file,
3299                                              dst_off, len);
3300         return ret;
3301 }
3302
3303 static const struct file_operations fuse_file_operations = {
3304         .llseek         = fuse_file_llseek,
3305         .read_iter      = fuse_file_read_iter,
3306         .write_iter     = fuse_file_write_iter,
3307         .mmap           = fuse_file_mmap,
3308         .open           = fuse_open,
3309         .flush          = fuse_flush,
3310         .release        = fuse_release,
3311         .fsync          = fuse_fsync,
3312         .lock           = fuse_file_lock,
3313         .get_unmapped_area = thp_get_unmapped_area,
3314         .flock          = fuse_file_flock,
3315         .splice_read    = fuse_splice_read,
3316         .splice_write   = fuse_splice_write,
3317         .unlocked_ioctl = fuse_file_ioctl,
3318         .compat_ioctl   = fuse_file_compat_ioctl,
3319         .poll           = fuse_file_poll,
3320         .fallocate      = fuse_file_fallocate,
3321         .copy_file_range = fuse_copy_file_range,
3322 };
3323
3324 static const struct address_space_operations fuse_file_aops  = {
3325         .read_folio     = fuse_read_folio,
3326         .readahead      = fuse_readahead,
3327         .writepages     = fuse_writepages,
3328         .launder_folio  = fuse_launder_folio,
3329         .dirty_folio    = filemap_dirty_folio,
3330         .migrate_folio  = filemap_migrate_folio,
3331         .bmap           = fuse_bmap,
3332         .direct_IO      = fuse_direct_IO,
3333         .write_begin    = fuse_write_begin,
3334         .write_end      = fuse_write_end,
3335 };
3336
3337 void fuse_init_file_inode(struct inode *inode, unsigned int flags)
3338 {
3339         struct fuse_inode *fi = get_fuse_inode(inode);
3340
3341         inode->i_fop = &fuse_file_operations;
3342         inode->i_data.a_ops = &fuse_file_aops;
3343
3344         INIT_LIST_HEAD(&fi->write_files);
3345         INIT_LIST_HEAD(&fi->queued_writes);
3346         fi->writectr = 0;
3347         fi->iocachectr = 0;
3348         init_waitqueue_head(&fi->page_waitq);
3349         init_waitqueue_head(&fi->direct_io_waitq);
3350         fi->writepages = RB_ROOT;
3351
3352         if (IS_ENABLED(CONFIG_FUSE_DAX))
3353                 fuse_dax_inode_init(inode, flags);
3354 }