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/compat.h>
  18 #include <linux/swap.h>
  19 #include <linux/falloc.h>
  20 #include <linux/uio.h>
  21
  22 static const struct file_operations fuse_direct_io_file_operations;
  23
  24 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  25                           int opcode, struct fuse_open_out *outargp)
  26 {
  27         struct fuse_open_in inarg;
  28         FUSE_ARGS(args);
  29
  30         memset(&inarg, 0, sizeof(inarg));
  31         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  32         if (!fc->atomic_o_trunc)
  33                 inarg.flags &= ~O_TRUNC;
  34         args.in.h.opcode = opcode;
  35         args.in.h.nodeid = nodeid;
  36         args.in.numargs = 1;
  37         args.in.args[0].size = sizeof(inarg);
  38         args.in.args[0].value = &inarg;
  39         args.out.numargs = 1;
  40         args.out.args[0].size = sizeof(*outargp);
  41         args.out.args[0].value = outargp;
  42
  43         return fuse_simple_request(fc, &args);
  44 }
  45
  46 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  47 {
  48         struct fuse_file *ff;
  49
  50         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
  51         if (unlikely(!ff))
  52                 return NULL;
  53
  54         ff->fc = fc;
  55         ff->reserved_req = fuse_request_alloc(0);
  56         if (unlikely(!ff->reserved_req)) {
  57                 kfree(ff);
  58                 return NULL;
  59         }
  60
  61         INIT_LIST_HEAD(&ff->write_entry);
  62         mutex_init(&ff->readdir.lock);
  63         refcount_set(&ff->count, 1);
  64         RB_CLEAR_NODE(&ff->polled_node);
  65         init_waitqueue_head(&ff->poll_wait);
  66
  67         spin_lock(&fc->lock);
  68         ff->kh = ++fc->khctr;
  69         spin_unlock(&fc->lock);
  70
  71         return ff;
  72 }
  73
  74 void fuse_file_free(struct fuse_file *ff)
  75 {
  76         fuse_request_free(ff->reserved_req);
  77         mutex_destroy(&ff->readdir.lock);
  78         kfree(ff);
  79 }
  80
  81 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  82 {
  83         refcount_inc(&ff->count);
  84         return ff;
  85 }
  86
  87 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  88 {
  89         iput(req->misc.release.inode);
  90 }
  91
  92 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
  93 {
  94         if (refcount_dec_and_test(&ff->count)) {
  95                 struct fuse_req *req = ff->reserved_req;
  96
  97                 if (ff->fc->no_open && !isdir) {
  98                         /*
  99                          * Drop the release request when client does not
 100                          * implement 'open'
 101                          */
 102                         __clear_bit(FR_BACKGROUND, &req->flags);
 103                         iput(req->misc.release.inode);
 104                         fuse_put_request(ff->fc, req);
 105                 } else if (sync) {
 106                         __set_bit(FR_FORCE, &req->flags);
 107                         __clear_bit(FR_BACKGROUND, &req->flags);
 108                         fuse_request_send(ff->fc, req);
 109                         iput(req->misc.release.inode);
 110                         fuse_put_request(ff->fc, req);
 111                 } else {
 112                         req->end = fuse_release_end;
 113                         __set_bit(FR_BACKGROUND, &req->flags);
 114                         fuse_request_send_background(ff->fc, req);
 115                 }
 116                 kfree(ff);
 117         }
 118 }
 119
 120 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 121                  bool isdir)
 122 {
 123         struct fuse_file *ff;
 124         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 125
 126         ff = fuse_file_alloc(fc);
 127         if (!ff)
 128                 return -ENOMEM;
 129
 130         ff->fh = 0;
 131         ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
 132         if (!fc->no_open || isdir) {
 133                 struct fuse_open_out outarg;
 134                 int err;
 135
 136                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 137                 if (!err) {
 138                         ff->fh = outarg.fh;
 139                         ff->open_flags = outarg.open_flags;
 140
 141                 } else if (err != -ENOSYS || isdir) {
 142                         fuse_file_free(ff);
 143                         return err;
 144                 } else {
 145                         fc->no_open = 1;
 146                 }
 147         }
 148
 149         if (isdir)
 150                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 151
 152         ff->nodeid = nodeid;
 153         file->private_data = ff;
 154
 155         return 0;
 156 }
 157 EXPORT_SYMBOL_GPL(fuse_do_open);
 158
 159 static void fuse_link_write_file(struct file *file)
 160 {
 161         struct inode *inode = file_inode(file);
 162         struct fuse_conn *fc = get_fuse_conn(inode);
 163         struct fuse_inode *fi = get_fuse_inode(inode);
 164         struct fuse_file *ff = file->private_data;
 165         /*
 166          * file may be written through mmap, so chain it onto the
 167          * inodes's write_file list
 168          */
 169         spin_lock(&fc->lock);
 170         if (list_empty(&ff->write_entry))
 171                 list_add(&ff->write_entry, &fi->write_files);
 172         spin_unlock(&fc->lock);
 173 }
 174
 175 void fuse_finish_open(struct inode *inode, struct file *file)
 176 {
 177         struct fuse_file *ff = file->private_data;
 178         struct fuse_conn *fc = get_fuse_conn(inode);
 179
 180         if (ff->open_flags & FOPEN_DIRECT_IO)
 181                 file->f_op = &fuse_direct_io_file_operations;
 182         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 183                 invalidate_inode_pages2(inode->i_mapping);
 184         if (ff->open_flags & FOPEN_NONSEEKABLE)
 185                 nonseekable_open(inode, file);
 186         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 187                 struct fuse_inode *fi = get_fuse_inode(inode);
 188
 189                 spin_lock(&fc->lock);
 190                 fi->attr_version = ++fc->attr_version;
 191                 i_size_write(inode, 0);
 192                 spin_unlock(&fc->lock);
 193                 fuse_invalidate_attr(inode);
 194                 if (fc->writeback_cache)
 195                         file_update_time(file);
 196         }
 197         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 198                 fuse_link_write_file(file);
 199 }
 200
 201 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 202 {
 203         struct fuse_conn *fc = get_fuse_conn(inode);
 204         int err;
 205         bool lock_inode = (file->f_flags & O_TRUNC) &&
 206                           fc->atomic_o_trunc &&
 207                           fc->writeback_cache;
 208
 209         err = generic_file_open(inode, file);
 210         if (err)
 211                 return err;
 212
 213         if (lock_inode)
 214                 inode_lock(inode);
 215
 216         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 217
 218         if (!err)
 219                 fuse_finish_open(inode, file);
 220
 221         if (lock_inode)
 222                 inode_unlock(inode);
 223
 224         return err;
 225 }
 226
 227 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 228 {
 229         struct fuse_conn *fc = ff->fc;
 230         struct fuse_req *req = ff->reserved_req;
 231         struct fuse_release_in *inarg = &req->misc.release.in;
 232
 233         spin_lock(&fc->lock);
 234         list_del(&ff->write_entry);
 235         if (!RB_EMPTY_NODE(&ff->polled_node))
 236                 rb_erase(&ff->polled_node, &fc->polled_files);
 237         spin_unlock(&fc->lock);
 238
 239         wake_up_interruptible_all(&ff->poll_wait);
 240
 241         inarg->fh = ff->fh;
 242         inarg->flags = flags;
 243         req->in.h.opcode = opcode;
 244         req->in.h.nodeid = ff->nodeid;
 245         req->in.numargs = 1;
 246         req->in.args[0].size = sizeof(struct fuse_release_in);
 247         req->in.args[0].value = inarg;
 248 }
 249
 250 void fuse_release_common(struct file *file, bool isdir)
 251 {
 252         struct fuse_file *ff = file->private_data;
 253         struct fuse_req *req = ff->reserved_req;
 254         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 255
 256         fuse_prepare_release(ff, file->f_flags, opcode);
 257
 258         if (ff->flock) {
 259                 struct fuse_release_in *inarg = &req->misc.release.in;
 260                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 261                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 262                                                        (fl_owner_t) file);
 263         }
 264         /* Hold inode until release is finished */
 265         req->misc.release.inode = igrab(file_inode(file));
 266
 267         /*
 268          * Normally this will send the RELEASE request, however if
 269          * some asynchronous READ or WRITE requests are outstanding,
 270          * the sending will be delayed.
 271          *
 272          * Make the release synchronous if this is a fuseblk mount,
 273          * synchronous RELEASE is allowed (and desirable) in this case
 274          * because the server can be trusted not to screw up.
 275          */
 276         fuse_file_put(ff, ff->fc->destroy_req != NULL, isdir);
 277 }
 278
 279 static int fuse_open(struct inode *inode, struct file *file)
 280 {
 281         return fuse_open_common(inode, file, false);
 282 }
 283
 284 static int fuse_release(struct inode *inode, struct file *file)
 285 {
 286         struct fuse_conn *fc = get_fuse_conn(inode);
 287
 288         /* see fuse_vma_close() for !writeback_cache case */
 289         if (fc->writeback_cache)
 290                 write_inode_now(inode, 1);
 291
 292         fuse_release_common(file, false);
 293
 294         /* return value is ignored by VFS */
 295         return 0;
 296 }
 297
 298 void fuse_sync_release(struct fuse_file *ff, int flags)
 299 {
 300         WARN_ON(refcount_read(&ff->count) > 1);
 301         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 302         /*
 303          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 304          * synchronous, we are fine with not doing igrab() here"
 305          */
 306         fuse_file_put(ff, true, false);
 307 }
 308 EXPORT_SYMBOL_GPL(fuse_sync_release);
 309
 310 /*
 311  * Scramble the ID space with XTEA, so that the value of the files_struct
 312  * pointer is not exposed to userspace.
 313  */
 314 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 315 {
 316         u32 *k = fc->scramble_key;
 317         u64 v = (unsigned long) id;
 318         u32 v0 = v;
 319         u32 v1 = v >> 32;
 320         u32 sum = 0;
 321         int i;
 322
 323         for (i = 0; i < 32; i++) {
 324                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 325                 sum += 0x9E3779B9;
 326                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 327         }
 328
 329         return (u64) v0 + ((u64) v1 << 32);
 330 }
 331
 332 /*
 333  * Check if any page in a range is under writeback
 334  *
 335  * This is currently done by walking the list of writepage requests
 336  * for the inode, which can be pretty inefficient.
 337  */
 338 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 339                                    pgoff_t idx_to)
 340 {
 341         struct fuse_conn *fc = get_fuse_conn(inode);
 342         struct fuse_inode *fi = get_fuse_inode(inode);
 343         struct fuse_req *req;
 344         bool found = false;
 345
 346         spin_lock(&fc->lock);
 347         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 348                 pgoff_t curr_index;
 349
 350                 BUG_ON(req->inode != inode);
 351                 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
 352                 if (idx_from < curr_index + req->num_pages &&
 353                     curr_index <= idx_to) {
 354                         found = true;
 355                         break;
 356                 }
 357         }
 358         spin_unlock(&fc->lock);
 359
 360         return found;
 361 }
 362
 363 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 364 {
 365         return fuse_range_is_writeback(inode, index, index);
 366 }
 367
 368 /*
 369  * Wait for page writeback to be completed.
 370  *
 371  * Since fuse doesn't rely on the VM writeback tracking, this has to
 372  * use some other means.
 373  */
 374 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 375 {
 376         struct fuse_inode *fi = get_fuse_inode(inode);
 377
 378         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 379         return 0;
 380 }
 381
 382 /*
 383  * Wait for all pending writepages on the inode to finish.
 384  *
 385  * This is currently done by blocking further writes with FUSE_NOWRITE
 386  * and waiting for all sent writes to complete.
 387  *
 388  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 389  * could conflict with truncation.
 390  */
 391 static void fuse_sync_writes(struct inode *inode)
 392 {
 393         fuse_set_nowrite(inode);
 394         fuse_release_nowrite(inode);
 395 }
 396
 397 static int fuse_flush(struct file *file, fl_owner_t id)
 398 {
 399         struct inode *inode = file_inode(file);
 400         struct fuse_conn *fc = get_fuse_conn(inode);
 401         struct fuse_file *ff = file->private_data;
 402         struct fuse_req *req;
 403         struct fuse_flush_in inarg;
 404         int err;
 405
 406         if (is_bad_inode(inode))
 407                 return -EIO;
 408
 409         if (fc->no_flush)
 410                 return 0;
 411
 412         err = write_inode_now(inode, 1);
 413         if (err)
 414                 return err;
 415
 416         inode_lock(inode);
 417         fuse_sync_writes(inode);
 418         inode_unlock(inode);
 419
 420         err = filemap_check_errors(file->f_mapping);
 421         if (err)
 422                 return err;
 423
 424         req = fuse_get_req_nofail_nopages(fc, file);
 425         memset(&inarg, 0, sizeof(inarg));
 426         inarg.fh = ff->fh;
 427         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 428         req->in.h.opcode = FUSE_FLUSH;
 429         req->in.h.nodeid = get_node_id(inode);
 430         req->in.numargs = 1;
 431         req->in.args[0].size = sizeof(inarg);
 432         req->in.args[0].value = &inarg;
 433         __set_bit(FR_FORCE, &req->flags);
 434         fuse_request_send(fc, req);
 435         err = req->out.h.error;
 436         fuse_put_request(fc, req);
 437         if (err == -ENOSYS) {
 438                 fc->no_flush = 1;
 439                 err = 0;
 440         }
 441         return err;
 442 }
 443
 444 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 445                       int datasync, int opcode)
 446 {
 447         struct inode *inode = file->f_mapping->host;
 448         struct fuse_conn *fc = get_fuse_conn(inode);
 449         struct fuse_file *ff = file->private_data;
 450         FUSE_ARGS(args);
 451         struct fuse_fsync_in inarg;
 452
 453         memset(&inarg, 0, sizeof(inarg));
 454         inarg.fh = ff->fh;
 455         inarg.fsync_flags = datasync ? 1 : 0;
 456         args.in.h.opcode = opcode;
 457         args.in.h.nodeid = get_node_id(inode);
 458         args.in.numargs = 1;
 459         args.in.args[0].size = sizeof(inarg);
 460         args.in.args[0].value = &inarg;
 461         return fuse_simple_request(fc, &args);
 462 }
 463
 464 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 465                       int datasync)
 466 {
 467         struct inode *inode = file->f_mapping->host;
 468         struct fuse_conn *fc = get_fuse_conn(inode);
 469         int err;
 470
 471         if (is_bad_inode(inode))
 472                 return -EIO;
 473
 474         inode_lock(inode);
 475
 476         /*
 477          * Start writeback against all dirty pages of the inode, then
 478          * wait for all outstanding writes, before sending the FSYNC
 479          * request.
 480          */
 481         err = file_write_and_wait_range(file, start, end);
 482         if (err)
 483                 goto out;
 484
 485         fuse_sync_writes(inode);
 486
 487         /*
 488          * Due to implementation of fuse writeback
 489          * file_write_and_wait_range() does not catch errors.
 490          * We have to do this directly after fuse_sync_writes()
 491          */
 492         err = file_check_and_advance_wb_err(file);
 493         if (err)
 494                 goto out;
 495
 496         err = sync_inode_metadata(inode, 1);
 497         if (err)
 498                 goto out;
 499
 500         if (fc->no_fsync)
 501                 goto out;
 502
 503         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 504         if (err == -ENOSYS) {
 505                 fc->no_fsync = 1;
 506                 err = 0;
 507         }
 508 out:
 509         inode_unlock(inode);
 510
 511         return err;
 512 }
 513
 514 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 515                     size_t count, int opcode)
 516 {
 517         struct fuse_read_in *inarg = &req->misc.read.in;
 518         struct fuse_file *ff = file->private_data;
 519
 520         inarg->fh = ff->fh;
 521         inarg->offset = pos;
 522         inarg->size = count;
 523         inarg->flags = file->f_flags;
 524         req->in.h.opcode = opcode;
 525         req->in.h.nodeid = ff->nodeid;
 526         req->in.numargs = 1;
 527         req->in.args[0].size = sizeof(struct fuse_read_in);
 528         req->in.args[0].value = inarg;
 529         req->out.argvar = 1;
 530         req->out.numargs = 1;
 531         req->out.args[0].size = count;
 532 }
 533
 534 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
 535 {
 536         unsigned i;
 537
 538         for (i = 0; i < req->num_pages; i++) {
 539                 struct page *page = req->pages[i];
 540                 if (should_dirty)
 541                         set_page_dirty_lock(page);
 542                 put_page(page);
 543         }
 544 }
 545
 546 static void fuse_io_release(struct kref *kref)
 547 {
 548         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 549 }
 550
 551 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 552 {
 553         if (io->err)
 554                 return io->err;
 555
 556         if (io->bytes >= 0 && io->write)
 557                 return -EIO;
 558
 559         return io->bytes < 0 ? io->size : io->bytes;
 560 }
 561
 562 /**
 563  * In case of short read, the caller sets 'pos' to the position of
 564  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 565  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 566  *
 567  * An example:
 568  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 569  * both submitted asynchronously. The first of them was ACKed by userspace as
 570  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 571  * second request was ACKed as short, e.g. only 1K was read, resulting in
 572  * pos == 33K.
 573  *
 574  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 575  * will be equal to the length of the longest contiguous fragment of
 576  * transferred data starting from the beginning of IO request.
 577  */
 578 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 579 {
 580         int left;
 581
 582         spin_lock(&io->lock);
 583         if (err)
 584                 io->err = io->err ? : err;
 585         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 586                 io->bytes = pos;
 587
 588         left = --io->reqs;
 589         if (!left && io->blocking)
 590                 complete(io->done);
 591         spin_unlock(&io->lock);
 592
 593         if (!left && !io->blocking) {
 594                 ssize_t res = fuse_get_res_by_io(io);
 595
 596                 if (res >= 0) {
 597                         struct inode *inode = file_inode(io->iocb->ki_filp);
 598                         struct fuse_conn *fc = get_fuse_conn(inode);
 599                         struct fuse_inode *fi = get_fuse_inode(inode);
 600
 601                         spin_lock(&fc->lock);
 602                         fi->attr_version = ++fc->attr_version;
 603                         spin_unlock(&fc->lock);
 604                 }
 605
 606                 io->iocb->ki_complete(io->iocb, res, 0);
 607         }
 608
 609         kref_put(&io->refcnt, fuse_io_release);
 610 }
 611
 612 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 613 {
 614         struct fuse_io_priv *io = req->io;
 615         ssize_t pos = -1;
 616
 617         fuse_release_user_pages(req, io->should_dirty);
 618
 619         if (io->write) {
 620                 if (req->misc.write.in.size != req->misc.write.out.size)
 621                         pos = req->misc.write.in.offset - io->offset +
 622                                 req->misc.write.out.size;
 623         } else {
 624                 if (req->misc.read.in.size != req->out.args[0].size)
 625                         pos = req->misc.read.in.offset - io->offset +
 626                                 req->out.args[0].size;
 627         }
 628
 629         fuse_aio_complete(io, req->out.h.error, pos);
 630 }
 631
 632 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 633                 size_t num_bytes, struct fuse_io_priv *io)
 634 {
 635         spin_lock(&io->lock);
 636         kref_get(&io->refcnt);
 637         io->size += num_bytes;
 638         io->reqs++;
 639         spin_unlock(&io->lock);
 640
 641         req->io = io;
 642         req->end = fuse_aio_complete_req;
 643
 644         __fuse_get_request(req);
 645         fuse_request_send_background(fc, req);
 646
 647         return num_bytes;
 648 }
 649
 650 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 651                              loff_t pos, size_t count, fl_owner_t owner)
 652 {
 653         struct file *file = io->iocb->ki_filp;
 654         struct fuse_file *ff = file->private_data;
 655         struct fuse_conn *fc = ff->fc;
 656
 657         fuse_read_fill(req, file, pos, count, FUSE_READ);
 658         if (owner != NULL) {
 659                 struct fuse_read_in *inarg = &req->misc.read.in;
 660
 661                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 662                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 663         }
 664
 665         if (io->async)
 666                 return fuse_async_req_send(fc, req, count, io);
 667
 668         fuse_request_send(fc, req);
 669         return req->out.args[0].size;
 670 }
 671
 672 static void fuse_read_update_size(struct inode *inode, loff_t size,
 673                                   u64 attr_ver)
 674 {
 675         struct fuse_conn *fc = get_fuse_conn(inode);
 676         struct fuse_inode *fi = get_fuse_inode(inode);
 677
 678         spin_lock(&fc->lock);
 679         if (attr_ver == fi->attr_version && size < inode->i_size &&
 680             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 681                 fi->attr_version = ++fc->attr_version;
 682                 i_size_write(inode, size);
 683         }
 684         spin_unlock(&fc->lock);
 685 }
 686
 687 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 688                             u64 attr_ver)
 689 {
 690         size_t num_read = req->out.args[0].size;
 691         struct fuse_conn *fc = get_fuse_conn(inode);
 692
 693         if (fc->writeback_cache) {
 694                 /*
 695                  * A hole in a file. Some data after the hole are in page cache,
 696                  * but have not reached the client fs yet. So, the hole is not
 697                  * present there.
 698                  */
 699                 int i;
 700                 int start_idx = num_read >> PAGE_SHIFT;
 701                 size_t off = num_read & (PAGE_SIZE - 1);
 702
 703                 for (i = start_idx; i < req->num_pages; i++) {
 704                         zero_user_segment(req->pages[i], off, PAGE_SIZE);
 705                         off = 0;
 706                 }
 707         } else {
 708                 loff_t pos = page_offset(req->pages[0]) + num_read;
 709                 fuse_read_update_size(inode, pos, attr_ver);
 710         }
 711 }
 712
 713 static int fuse_do_readpage(struct file *file, struct page *page)
 714 {
 715         struct kiocb iocb;
 716         struct fuse_io_priv io;
 717         struct inode *inode = page->mapping->host;
 718         struct fuse_conn *fc = get_fuse_conn(inode);
 719         struct fuse_req *req;
 720         size_t num_read;
 721         loff_t pos = page_offset(page);
 722         size_t count = PAGE_SIZE;
 723         u64 attr_ver;
 724         int err;
 725
 726         /*
 727          * Page writeback can extend beyond the lifetime of the
 728          * page-cache page, so make sure we read a properly synced
 729          * page.
 730          */
 731         fuse_wait_on_page_writeback(inode, page->index);
 732
 733         req = fuse_get_req(fc, 1);
 734         if (IS_ERR(req))
 735                 return PTR_ERR(req);
 736
 737         attr_ver = fuse_get_attr_version(fc);
 738
 739         req->out.page_zeroing = 1;
 740         req->out.argpages = 1;
 741         req->num_pages = 1;
 742         req->pages[0] = page;
 743         req->page_descs[0].length = count;
 744         init_sync_kiocb(&iocb, file);
 745         io = (struct fuse_io_priv) FUSE_IO_PRIV_SYNC(&iocb);
 746         num_read = fuse_send_read(req, &io, pos, count, NULL);
 747         err = req->out.h.error;
 748
 749         if (!err) {
 750                 /*
 751                  * Short read means EOF.  If file size is larger, truncate it
 752                  */
 753                 if (num_read < count)
 754                         fuse_short_read(req, inode, attr_ver);
 755
 756                 SetPageUptodate(page);
 757         }
 758
 759         fuse_put_request(fc, req);
 760
 761         return err;
 762 }
 763
 764 static int fuse_readpage(struct file *file, struct page *page)
 765 {
 766         struct inode *inode = page->mapping->host;
 767         int err;
 768
 769         err = -EIO;
 770         if (is_bad_inode(inode))
 771                 goto out;
 772
 773         err = fuse_do_readpage(file, page);
 774         fuse_invalidate_atime(inode);
 775  out:
 776         unlock_page(page);
 777         return err;
 778 }
 779
 780 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 781 {
 782         int i;
 783         size_t count = req->misc.read.in.size;
 784         size_t num_read = req->out.args[0].size;
 785         struct address_space *mapping = NULL;
 786
 787         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 788                 mapping = req->pages[i]->mapping;
 789
 790         if (mapping) {
 791                 struct inode *inode = mapping->host;
 792
 793                 /*
 794                  * Short read means EOF. If file size is larger, truncate it
 795                  */
 796                 if (!req->out.h.error && num_read < count)
 797                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 798
 799                 fuse_invalidate_atime(inode);
 800         }
 801
 802         for (i = 0; i < req->num_pages; i++) {
 803                 struct page *page = req->pages[i];
 804                 if (!req->out.h.error)
 805                         SetPageUptodate(page);
 806                 else
 807                         SetPageError(page);
 808                 unlock_page(page);
 809                 put_page(page);
 810         }
 811         if (req->ff)
 812                 fuse_file_put(req->ff, false, false);
 813 }
 814
 815 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 816 {
 817         struct fuse_file *ff = file->private_data;
 818         struct fuse_conn *fc = ff->fc;
 819         loff_t pos = page_offset(req->pages[0]);
 820         size_t count = req->num_pages << PAGE_SHIFT;
 821
 822         req->out.argpages = 1;
 823         req->out.page_zeroing = 1;
 824         req->out.page_replace = 1;
 825         fuse_read_fill(req, file, pos, count, FUSE_READ);
 826         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 827         if (fc->async_read) {
 828                 req->ff = fuse_file_get(ff);
 829                 req->end = fuse_readpages_end;
 830                 fuse_request_send_background(fc, req);
 831         } else {
 832                 fuse_request_send(fc, req);
 833                 fuse_readpages_end(fc, req);
 834                 fuse_put_request(fc, req);
 835         }
 836 }
 837
 838 struct fuse_fill_data {
 839         struct fuse_req *req;
 840         struct file *file;
 841         struct inode *inode;
 842         unsigned nr_pages;
 843 };
 844
 845 static int fuse_readpages_fill(void *_data, struct page *page)
 846 {
 847         struct fuse_fill_data *data = _data;
 848         struct fuse_req *req = data->req;
 849         struct inode *inode = data->inode;
 850         struct fuse_conn *fc = get_fuse_conn(inode);
 851
 852         fuse_wait_on_page_writeback(inode, page->index);
 853
 854         if (req->num_pages &&
 855             (req->num_pages == fc->max_pages ||
 856              (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
 857              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 858                 unsigned int nr_alloc = min_t(unsigned int, data->nr_pages,
 859                                               fc->max_pages);
 860                 fuse_send_readpages(req, data->file);
 861                 if (fc->async_read)
 862                         req = fuse_get_req_for_background(fc, nr_alloc);
 863                 else
 864                         req = fuse_get_req(fc, nr_alloc);
 865
 866                 data->req = req;
 867                 if (IS_ERR(req)) {
 868                         unlock_page(page);
 869                         return PTR_ERR(req);
 870                 }
 871         }
 872
 873         if (WARN_ON(req->num_pages >= req->max_pages)) {
 874                 unlock_page(page);
 875                 fuse_put_request(fc, req);
 876                 return -EIO;
 877         }
 878
 879         get_page(page);
 880         req->pages[req->num_pages] = page;
 881         req->page_descs[req->num_pages].length = PAGE_SIZE;
 882         req->num_pages++;
 883         data->nr_pages--;
 884         return 0;
 885 }
 886
 887 static int fuse_readpages(struct file *file, struct address_space *mapping,
 888                           struct list_head *pages, unsigned nr_pages)
 889 {
 890         struct inode *inode = mapping->host;
 891         struct fuse_conn *fc = get_fuse_conn(inode);
 892         struct fuse_fill_data data;
 893         int err;
 894         unsigned int nr_alloc = min_t(unsigned int, nr_pages, fc->max_pages);
 895
 896         err = -EIO;
 897         if (is_bad_inode(inode))
 898                 goto out;
 899
 900         data.file = file;
 901         data.inode = inode;
 902         if (fc->async_read)
 903                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 904         else
 905                 data.req = fuse_get_req(fc, nr_alloc);
 906         data.nr_pages = nr_pages;
 907         err = PTR_ERR(data.req);
 908         if (IS_ERR(data.req))
 909                 goto out;
 910
 911         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 912         if (!err) {
 913                 if (data.req->num_pages)
 914                         fuse_send_readpages(data.req, file);
 915                 else
 916                         fuse_put_request(fc, data.req);
 917         }
 918 out:
 919         return err;
 920 }
 921
 922 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 923 {
 924         struct inode *inode = iocb->ki_filp->f_mapping->host;
 925         struct fuse_conn *fc = get_fuse_conn(inode);
 926
 927         /*
 928          * In auto invalidate mode, always update attributes on read.
 929          * Otherwise, only update if we attempt to read past EOF (to ensure
 930          * i_size is up to date).
 931          */
 932         if (fc->auto_inval_data ||
 933             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 934                 int err;
 935                 err = fuse_update_attributes(inode, iocb->ki_filp);
 936                 if (err)
 937                         return err;
 938         }
 939
 940         return generic_file_read_iter(iocb, to);
 941 }
 942
 943 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 944                             loff_t pos, size_t count)
 945 {
 946         struct fuse_write_in *inarg = &req->misc.write.in;
 947         struct fuse_write_out *outarg = &req->misc.write.out;
 948
 949         inarg->fh = ff->fh;
 950         inarg->offset = pos;
 951         inarg->size = count;
 952         req->in.h.opcode = FUSE_WRITE;
 953         req->in.h.nodeid = ff->nodeid;
 954         req->in.numargs = 2;
 955         if (ff->fc->minor < 9)
 956                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 957         else
 958                 req->in.args[0].size = sizeof(struct fuse_write_in);
 959         req->in.args[0].value = inarg;
 960         req->in.args[1].size = count;
 961         req->out.numargs = 1;
 962         req->out.args[0].size = sizeof(struct fuse_write_out);
 963         req->out.args[0].value = outarg;
 964 }
 965
 966 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 967                               loff_t pos, size_t count, fl_owner_t owner)
 968 {
 969         struct kiocb *iocb = io->iocb;
 970         struct file *file = iocb->ki_filp;
 971         struct fuse_file *ff = file->private_data;
 972         struct fuse_conn *fc = ff->fc;
 973         struct fuse_write_in *inarg = &req->misc.write.in;
 974
 975         fuse_write_fill(req, ff, pos, count);
 976         inarg->flags = file->f_flags;
 977         if (iocb->ki_flags & IOCB_DSYNC)
 978                 inarg->flags |= O_DSYNC;
 979         if (iocb->ki_flags & IOCB_SYNC)
 980                 inarg->flags |= O_SYNC;
 981         if (owner != NULL) {
 982                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 983                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 984         }
 985
 986         if (io->async)
 987                 return fuse_async_req_send(fc, req, count, io);
 988
 989         fuse_request_send(fc, req);
 990         return req->misc.write.out.size;
 991 }
 992
 993 bool fuse_write_update_size(struct inode *inode, loff_t pos)
 994 {
 995         struct fuse_conn *fc = get_fuse_conn(inode);
 996         struct fuse_inode *fi = get_fuse_inode(inode);
 997         bool ret = false;
 998
 999         spin_lock(&fc->lock);
1000         fi->attr_version = ++fc->attr_version;
1001         if (pos > inode->i_size) {
1002                 i_size_write(inode, pos);
1003                 ret = true;
1004         }
1005         spin_unlock(&fc->lock);
1006
1007         return ret;
1008 }
1009
1010 static size_t fuse_send_write_pages(struct fuse_req *req, struct kiocb *iocb,
1011                                     struct inode *inode, loff_t pos,
1012                                     size_t count)
1013 {
1014         size_t res;
1015         unsigned offset;
1016         unsigned i;
1017         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1018
1019         for (i = 0; i < req->num_pages; i++)
1020                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1021
1022         res = fuse_send_write(req, &io, pos, count, NULL);
1023
1024         offset = req->page_descs[0].offset;
1025         count = res;
1026         for (i = 0; i < req->num_pages; i++) {
1027                 struct page *page = req->pages[i];
1028
1029                 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1030                         SetPageUptodate(page);
1031
1032                 if (count > PAGE_SIZE - offset)
1033                         count -= PAGE_SIZE - offset;
1034                 else
1035                         count = 0;
1036                 offset = 0;
1037
1038                 unlock_page(page);
1039                 put_page(page);
1040         }
1041
1042         return res;
1043 }
1044
1045 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1046                                struct address_space *mapping,
1047                                struct iov_iter *ii, loff_t pos)
1048 {
1049         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1050         unsigned offset = pos & (PAGE_SIZE - 1);
1051         size_t count = 0;
1052         int err;
1053
1054         req->in.argpages = 1;
1055         req->page_descs[0].offset = offset;
1056
1057         do {
1058                 size_t tmp;
1059                 struct page *page;
1060                 pgoff_t index = pos >> PAGE_SHIFT;
1061                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1062                                      iov_iter_count(ii));
1063
1064                 bytes = min_t(size_t, bytes, fc->max_write - count);
1065
1066  again:
1067                 err = -EFAULT;
1068                 if (iov_iter_fault_in_readable(ii, bytes))
1069                         break;
1070
1071                 err = -ENOMEM;
1072                 page = grab_cache_page_write_begin(mapping, index, 0);
1073                 if (!page)
1074                         break;
1075
1076                 if (mapping_writably_mapped(mapping))
1077                         flush_dcache_page(page);
1078
1079                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1080                 flush_dcache_page(page);
1081
1082                 iov_iter_advance(ii, tmp);
1083                 if (!tmp) {
1084                         unlock_page(page);
1085                         put_page(page);
1086                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1087                         goto again;
1088                 }
1089
1090                 err = 0;
1091                 req->pages[req->num_pages] = page;
1092                 req->page_descs[req->num_pages].length = tmp;
1093                 req->num_pages++;
1094
1095                 count += tmp;
1096                 pos += tmp;
1097                 offset += tmp;
1098                 if (offset == PAGE_SIZE)
1099                         offset = 0;
1100
1101                 if (!fc->big_writes)
1102                         break;
1103         } while (iov_iter_count(ii) && count < fc->max_write &&
1104                  req->num_pages < req->max_pages && offset == 0);
1105
1106         return count > 0 ? count : err;
1107 }
1108
1109 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1110                                      unsigned int max_pages)
1111 {
1112         return min_t(unsigned int,
1113                      ((pos + len - 1) >> PAGE_SHIFT) -
1114                      (pos >> PAGE_SHIFT) + 1,
1115                      max_pages);
1116 }
1117
1118 static ssize_t fuse_perform_write(struct kiocb *iocb,
1119                                   struct address_space *mapping,
1120                                   struct iov_iter *ii, loff_t pos)
1121 {
1122         struct inode *inode = mapping->host;
1123         struct fuse_conn *fc = get_fuse_conn(inode);
1124         struct fuse_inode *fi = get_fuse_inode(inode);
1125         int err = 0;
1126         ssize_t res = 0;
1127
1128         if (is_bad_inode(inode))
1129                 return -EIO;
1130
1131         if (inode->i_size < pos + iov_iter_count(ii))
1132                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1133
1134         do {
1135                 struct fuse_req *req;
1136                 ssize_t count;
1137                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1138                                                       fc->max_pages);
1139
1140                 req = fuse_get_req(fc, nr_pages);
1141                 if (IS_ERR(req)) {
1142                         err = PTR_ERR(req);
1143                         break;
1144                 }
1145
1146                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1147                 if (count <= 0) {
1148                         err = count;
1149                 } else {
1150                         size_t num_written;
1151
1152                         num_written = fuse_send_write_pages(req, iocb, inode,
1153                                                             pos, count);
1154                         err = req->out.h.error;
1155                         if (!err) {
1156                                 res += num_written;
1157                                 pos += num_written;
1158
1159                                 /* break out of the loop on short write */
1160                                 if (num_written != count)
1161                                         err = -EIO;
1162                         }
1163                 }
1164                 fuse_put_request(fc, req);
1165         } while (!err && iov_iter_count(ii));
1166
1167         if (res > 0)
1168                 fuse_write_update_size(inode, pos);
1169
1170         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1171         fuse_invalidate_attr(inode);
1172
1173         return res > 0 ? res : err;
1174 }
1175
1176 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1177 {
1178         struct file *file = iocb->ki_filp;
1179         struct address_space *mapping = file->f_mapping;
1180         ssize_t written = 0;
1181         ssize_t written_buffered = 0;
1182         struct inode *inode = mapping->host;
1183         ssize_t err;
1184         loff_t endbyte = 0;
1185
1186         if (get_fuse_conn(inode)->writeback_cache) {
1187                 /* Update size (EOF optimization) and mode (SUID clearing) */
1188                 err = fuse_update_attributes(mapping->host, file);
1189                 if (err)
1190                         return err;
1191
1192                 return generic_file_write_iter(iocb, from);
1193         }
1194
1195         inode_lock(inode);
1196
1197         /* We can write back this queue in page reclaim */
1198         current->backing_dev_info = inode_to_bdi(inode);
1199
1200         err = generic_write_checks(iocb, from);
1201         if (err <= 0)
1202                 goto out;
1203
1204         err = file_remove_privs(file);
1205         if (err)
1206                 goto out;
1207
1208         err = file_update_time(file);
1209         if (err)
1210                 goto out;
1211
1212         if (iocb->ki_flags & IOCB_DIRECT) {
1213                 loff_t pos = iocb->ki_pos;
1214                 written = generic_file_direct_write(iocb, from);
1215                 if (written < 0 || !iov_iter_count(from))
1216                         goto out;
1217
1218                 pos += written;
1219
1220                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1221                 if (written_buffered < 0) {
1222                         err = written_buffered;
1223                         goto out;
1224                 }
1225                 endbyte = pos + written_buffered - 1;
1226
1227                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1228                                                    endbyte);
1229                 if (err)
1230                         goto out;
1231
1232                 invalidate_mapping_pages(file->f_mapping,
1233                                          pos >> PAGE_SHIFT,
1234                                          endbyte >> PAGE_SHIFT);
1235
1236                 written += written_buffered;
1237                 iocb->ki_pos = pos + written_buffered;
1238         } else {
1239                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1240                 if (written >= 0)
1241                         iocb->ki_pos += written;
1242         }
1243 out:
1244         current->backing_dev_info = NULL;
1245         inode_unlock(inode);
1246         if (written > 0)
1247                 written = generic_write_sync(iocb, written);
1248
1249         return written ? written : err;
1250 }
1251
1252 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1253                 unsigned index, unsigned nr_pages)
1254 {
1255         int i;
1256
1257         for (i = index; i < index + nr_pages; i++)
1258                 req->page_descs[i].length = PAGE_SIZE -
1259                         req->page_descs[i].offset;
1260 }
1261
1262 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1263 {
1264         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1265 }
1266
1267 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1268                                         size_t max_size)
1269 {
1270         return min(iov_iter_single_seg_count(ii), max_size);
1271 }
1272
1273 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1274                                size_t *nbytesp, int write)
1275 {
1276         size_t nbytes = 0;  /* # bytes already packed in req */
1277         ssize_t ret = 0;
1278
1279         /* Special case for kernel I/O: can copy directly into the buffer */
1280         if (iov_iter_is_kvec(ii)) {
1281                 unsigned long user_addr = fuse_get_user_addr(ii);
1282                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1283
1284                 if (write)
1285                         req->in.args[1].value = (void *) user_addr;
1286                 else
1287                         req->out.args[0].value = (void *) user_addr;
1288
1289                 iov_iter_advance(ii, frag_size);
1290                 *nbytesp = frag_size;
1291                 return 0;
1292         }
1293
1294         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1295                 unsigned npages;
1296                 size_t start;
1297                 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1298                                         *nbytesp - nbytes,
1299                                         req->max_pages - req->num_pages,
1300                                         &start);
1301                 if (ret < 0)
1302                         break;
1303
1304                 iov_iter_advance(ii, ret);
1305                 nbytes += ret;
1306
1307                 ret += start;
1308                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1309
1310                 req->page_descs[req->num_pages].offset = start;
1311                 fuse_page_descs_length_init(req, req->num_pages, npages);
1312
1313                 req->num_pages += npages;
1314                 req->page_descs[req->num_pages - 1].length -=
1315                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1316         }
1317
1318         if (write)
1319                 req->in.argpages = 1;
1320         else
1321                 req->out.argpages = 1;
1322
1323         *nbytesp = nbytes;
1324
1325         return ret < 0 ? ret : 0;
1326 }
1327
1328 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1329                        loff_t *ppos, int flags)
1330 {
1331         int write = flags & FUSE_DIO_WRITE;
1332         int cuse = flags & FUSE_DIO_CUSE;
1333         struct file *file = io->iocb->ki_filp;
1334         struct inode *inode = file->f_mapping->host;
1335         struct fuse_file *ff = file->private_data;
1336         struct fuse_conn *fc = ff->fc;
1337         size_t nmax = write ? fc->max_write : fc->max_read;
1338         loff_t pos = *ppos;
1339         size_t count = iov_iter_count(iter);
1340         pgoff_t idx_from = pos >> PAGE_SHIFT;
1341         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1342         ssize_t res = 0;
1343         struct fuse_req *req;
1344         int err = 0;
1345
1346         if (io->async)
1347                 req = fuse_get_req_for_background(fc, iov_iter_npages(iter,
1348                                                                 fc->max_pages));
1349         else
1350                 req = fuse_get_req(fc, iov_iter_npages(iter, fc->max_pages));
1351         if (IS_ERR(req))
1352                 return PTR_ERR(req);
1353
1354         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1355                 if (!write)
1356                         inode_lock(inode);
1357                 fuse_sync_writes(inode);
1358                 if (!write)
1359                         inode_unlock(inode);
1360         }
1361
1362         io->should_dirty = !write && iter_is_iovec(iter);
1363         while (count) {
1364                 size_t nres;
1365                 fl_owner_t owner = current->files;
1366                 size_t nbytes = min(count, nmax);
1367                 err = fuse_get_user_pages(req, iter, &nbytes, write);
1368                 if (err && !nbytes)
1369                         break;
1370
1371                 if (write)
1372                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1373                 else
1374                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1375
1376                 if (!io->async)
1377                         fuse_release_user_pages(req, io->should_dirty);
1378                 if (req->out.h.error) {
1379                         err = req->out.h.error;
1380                         break;
1381                 } else if (nres > nbytes) {
1382                         res = 0;
1383                         err = -EIO;
1384                         break;
1385                 }
1386                 count -= nres;
1387                 res += nres;
1388                 pos += nres;
1389                 if (nres != nbytes)
1390                         break;
1391                 if (count) {
1392                         fuse_put_request(fc, req);
1393                         if (io->async)
1394                                 req = fuse_get_req_for_background(fc,
1395                                         iov_iter_npages(iter, fc->max_pages));
1396                         else
1397                                 req = fuse_get_req(fc, iov_iter_npages(iter,
1398                                                                 fc->max_pages));
1399                         if (IS_ERR(req))
1400                                 break;
1401                 }
1402         }
1403         if (!IS_ERR(req))
1404                 fuse_put_request(fc, req);
1405         if (res > 0)
1406                 *ppos = pos;
1407
1408         return res > 0 ? res : err;
1409 }
1410 EXPORT_SYMBOL_GPL(fuse_direct_io);
1411
1412 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1413                                   struct iov_iter *iter,
1414                                   loff_t *ppos)
1415 {
1416         ssize_t res;
1417         struct inode *inode = file_inode(io->iocb->ki_filp);
1418
1419         if (is_bad_inode(inode))
1420                 return -EIO;
1421
1422         res = fuse_direct_io(io, iter, ppos, 0);
1423
1424         fuse_invalidate_atime(inode);
1425
1426         return res;
1427 }
1428
1429 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1430 {
1431         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1432         return __fuse_direct_read(&io, to, &iocb->ki_pos);
1433 }
1434
1435 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1436 {
1437         struct inode *inode = file_inode(iocb->ki_filp);
1438         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1439         ssize_t res;
1440
1441         if (is_bad_inode(inode))
1442                 return -EIO;
1443
1444         /* Don't allow parallel writes to the same file */
1445         inode_lock(inode);
1446         res = generic_write_checks(iocb, from);
1447         if (res > 0)
1448                 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1449         fuse_invalidate_attr(inode);
1450         if (res > 0)
1451                 fuse_write_update_size(inode, iocb->ki_pos);
1452         inode_unlock(inode);
1453
1454         return res;
1455 }
1456
1457 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1458 {
1459         int i;
1460
1461         for (i = 0; i < req->num_pages; i++)
1462                 __free_page(req->pages[i]);
1463
1464         if (req->ff)
1465                 fuse_file_put(req->ff, false, false);
1466 }
1467
1468 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1469 {
1470         struct inode *inode = req->inode;
1471         struct fuse_inode *fi = get_fuse_inode(inode);
1472         struct backing_dev_info *bdi = inode_to_bdi(inode);
1473         int i;
1474
1475         list_del(&req->writepages_entry);
1476         for (i = 0; i < req->num_pages; i++) {
1477                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1478                 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1479                 wb_writeout_inc(&bdi->wb);
1480         }
1481         wake_up(&fi->page_waitq);
1482 }
1483
1484 /* Called under fc->lock, may release and reacquire it */
1485 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1486                                 loff_t size)
1487 __releases(fc->lock)
1488 __acquires(fc->lock)
1489 {
1490         struct fuse_inode *fi = get_fuse_inode(req->inode);
1491         struct fuse_write_in *inarg = &req->misc.write.in;
1492         __u64 data_size = req->num_pages * PAGE_SIZE;
1493         bool queued;
1494
1495         if (!fc->connected)
1496                 goto out_free;
1497
1498         if (inarg->offset + data_size <= size) {
1499                 inarg->size = data_size;
1500         } else if (inarg->offset < size) {
1501                 inarg->size = size - inarg->offset;
1502         } else {
1503                 /* Got truncated off completely */
1504                 goto out_free;
1505         }
1506
1507         req->in.args[1].size = inarg->size;
1508         fi->writectr++;
1509         queued = fuse_request_queue_background(fc, req);
1510         WARN_ON(!queued);
1511         return;
1512
1513  out_free:
1514         fuse_writepage_finish(fc, req);
1515         spin_unlock(&fc->lock);
1516         fuse_writepage_free(fc, req);
1517         fuse_put_request(fc, req);
1518         spin_lock(&fc->lock);
1519 }
1520
1521 /*
1522  * If fi->writectr is positive (no truncate or fsync going on) send
1523  * all queued writepage requests.
1524  *
1525  * Called with fc->lock
1526  */
1527 void fuse_flush_writepages(struct inode *inode)
1528 __releases(fc->lock)
1529 __acquires(fc->lock)
1530 {
1531         struct fuse_conn *fc = get_fuse_conn(inode);
1532         struct fuse_inode *fi = get_fuse_inode(inode);
1533         size_t crop = i_size_read(inode);
1534         struct fuse_req *req;
1535
1536         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1537                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1538                 list_del_init(&req->list);
1539                 fuse_send_writepage(fc, req, crop);
1540         }
1541 }
1542
1543 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1544 {
1545         struct inode *inode = req->inode;
1546         struct fuse_inode *fi = get_fuse_inode(inode);
1547
1548         mapping_set_error(inode->i_mapping, req->out.h.error);
1549         spin_lock(&fc->lock);
1550         while (req->misc.write.next) {
1551                 struct fuse_conn *fc = get_fuse_conn(inode);
1552                 struct fuse_write_in *inarg = &req->misc.write.in;
1553                 struct fuse_req *next = req->misc.write.next;
1554                 req->misc.write.next = next->misc.write.next;
1555                 next->misc.write.next = NULL;
1556                 next->ff = fuse_file_get(req->ff);
1557                 list_add(&next->writepages_entry, &fi->writepages);
1558
1559                 /*
1560                  * Skip fuse_flush_writepages() to make it easy to crop requests
1561                  * based on primary request size.
1562                  *
1563                  * 1st case (trivial): there are no concurrent activities using
1564                  * fuse_set/release_nowrite.  Then we're on safe side because
1565                  * fuse_flush_writepages() would call fuse_send_writepage()
1566                  * anyway.
1567                  *
1568                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1569                  * now for completion of all in-flight requests.  This happens
1570                  * rarely and no more than once per page, so this should be
1571                  * okay.
1572                  *
1573                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1574                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1575                  * that fuse_set_nowrite returned implies that all in-flight
1576                  * requests were completed along with all of their secondary
1577                  * requests.  Further primary requests are blocked by negative
1578                  * writectr.  Hence there cannot be any in-flight requests and
1579                  * no invocations of fuse_writepage_end() while we're in
1580                  * fuse_set_nowrite..fuse_release_nowrite section.
1581                  */
1582                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1583         }
1584         fi->writectr--;
1585         fuse_writepage_finish(fc, req);
1586         spin_unlock(&fc->lock);
1587         fuse_writepage_free(fc, req);
1588 }
1589
1590 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1591                                                struct fuse_inode *fi)
1592 {
1593         struct fuse_file *ff = NULL;
1594
1595         spin_lock(&fc->lock);
1596         if (!list_empty(&fi->write_files)) {
1597                 ff = list_entry(fi->write_files.next, struct fuse_file,
1598                                 write_entry);
1599                 fuse_file_get(ff);
1600         }
1601         spin_unlock(&fc->lock);
1602
1603         return ff;
1604 }
1605
1606 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1607                                              struct fuse_inode *fi)
1608 {
1609         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1610         WARN_ON(!ff);
1611         return ff;
1612 }
1613
1614 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1615 {
1616         struct fuse_conn *fc = get_fuse_conn(inode);
1617         struct fuse_inode *fi = get_fuse_inode(inode);
1618         struct fuse_file *ff;
1619         int err;
1620
1621         ff = __fuse_write_file_get(fc, fi);
1622         err = fuse_flush_times(inode, ff);
1623         if (ff)
1624                 fuse_file_put(ff, false, false);
1625
1626         return err;
1627 }
1628
1629 static int fuse_writepage_locked(struct page *page)
1630 {
1631         struct address_space *mapping = page->mapping;
1632         struct inode *inode = mapping->host;
1633         struct fuse_conn *fc = get_fuse_conn(inode);
1634         struct fuse_inode *fi = get_fuse_inode(inode);
1635         struct fuse_req *req;
1636         struct page *tmp_page;
1637         int error = -ENOMEM;
1638
1639         set_page_writeback(page);
1640
1641         req = fuse_request_alloc_nofs(1);
1642         if (!req)
1643                 goto err;
1644
1645         /* writeback always goes to bg_queue */
1646         __set_bit(FR_BACKGROUND, &req->flags);
1647         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1648         if (!tmp_page)
1649                 goto err_free;
1650
1651         error = -EIO;
1652         req->ff = fuse_write_file_get(fc, fi);
1653         if (!req->ff)
1654                 goto err_nofile;
1655
1656         fuse_write_fill(req, req->ff, page_offset(page), 0);
1657
1658         copy_highpage(tmp_page, page);
1659         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1660         req->misc.write.next = NULL;
1661         req->in.argpages = 1;
1662         req->num_pages = 1;
1663         req->pages[0] = tmp_page;
1664         req->page_descs[0].offset = 0;
1665         req->page_descs[0].length = PAGE_SIZE;
1666         req->end = fuse_writepage_end;
1667         req->inode = inode;
1668
1669         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1670         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1671
1672         spin_lock(&fc->lock);
1673         list_add(&req->writepages_entry, &fi->writepages);
1674         list_add_tail(&req->list, &fi->queued_writes);
1675         fuse_flush_writepages(inode);
1676         spin_unlock(&fc->lock);
1677
1678         end_page_writeback(page);
1679
1680         return 0;
1681
1682 err_nofile:
1683         __free_page(tmp_page);
1684 err_free:
1685         fuse_request_free(req);
1686 err:
1687         mapping_set_error(page->mapping, error);
1688         end_page_writeback(page);
1689         return error;
1690 }
1691
1692 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1693 {
1694         int err;
1695
1696         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1697                 /*
1698                  * ->writepages() should be called for sync() and friends.  We
1699                  * should only get here on direct reclaim and then we are
1700                  * allowed to skip a page which is already in flight
1701                  */
1702                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1703
1704                 redirty_page_for_writepage(wbc, page);
1705                 return 0;
1706         }
1707
1708         err = fuse_writepage_locked(page);
1709         unlock_page(page);
1710
1711         return err;
1712 }
1713
1714 struct fuse_fill_wb_data {
1715         struct fuse_req *req;
1716         struct fuse_file *ff;
1717         struct inode *inode;
1718         struct page **orig_pages;
1719 };
1720
1721 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1722 {
1723         struct fuse_req *req = data->req;
1724         struct inode *inode = data->inode;
1725         struct fuse_conn *fc = get_fuse_conn(inode);
1726         struct fuse_inode *fi = get_fuse_inode(inode);
1727         int num_pages = req->num_pages;
1728         int i;
1729
1730         req->ff = fuse_file_get(data->ff);
1731         spin_lock(&fc->lock);
1732         list_add_tail(&req->list, &fi->queued_writes);
1733         fuse_flush_writepages(inode);
1734         spin_unlock(&fc->lock);
1735
1736         for (i = 0; i < num_pages; i++)
1737                 end_page_writeback(data->orig_pages[i]);
1738 }
1739
1740 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1741                                      struct page *page)
1742 {
1743         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1744         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1745         struct fuse_req *tmp;
1746         struct fuse_req *old_req;
1747         bool found = false;
1748         pgoff_t curr_index;
1749
1750         BUG_ON(new_req->num_pages != 0);
1751
1752         spin_lock(&fc->lock);
1753         list_del(&new_req->writepages_entry);
1754         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1755                 BUG_ON(old_req->inode != new_req->inode);
1756                 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1757                 if (curr_index <= page->index &&
1758                     page->index < curr_index + old_req->num_pages) {
1759                         found = true;
1760                         break;
1761                 }
1762         }
1763         if (!found) {
1764                 list_add(&new_req->writepages_entry, &fi->writepages);
1765                 goto out_unlock;
1766         }
1767
1768         new_req->num_pages = 1;
1769         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1770                 BUG_ON(tmp->inode != new_req->inode);
1771                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1772                 if (tmp->num_pages == 1 &&
1773                     curr_index == page->index) {
1774                         old_req = tmp;
1775                 }
1776         }
1777
1778         if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1779                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1780
1781                 copy_highpage(old_req->pages[0], page);
1782                 spin_unlock(&fc->lock);
1783
1784                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1785                 dec_node_page_state(page, NR_WRITEBACK_TEMP);
1786                 wb_writeout_inc(&bdi->wb);
1787                 fuse_writepage_free(fc, new_req);
1788                 fuse_request_free(new_req);
1789                 goto out;
1790         } else {
1791                 new_req->misc.write.next = old_req->misc.write.next;
1792                 old_req->misc.write.next = new_req;
1793         }
1794 out_unlock:
1795         spin_unlock(&fc->lock);
1796 out:
1797         return found;
1798 }
1799
1800 static int fuse_writepages_fill(struct page *page,
1801                 struct writeback_control *wbc, void *_data)
1802 {
1803         struct fuse_fill_wb_data *data = _data;
1804         struct fuse_req *req = data->req;
1805         struct inode *inode = data->inode;
1806         struct fuse_conn *fc = get_fuse_conn(inode);
1807         struct page *tmp_page;
1808         bool is_writeback;
1809         int err;
1810
1811         if (!data->ff) {
1812                 err = -EIO;
1813                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1814                 if (!data->ff)
1815                         goto out_unlock;
1816         }
1817
1818         /*
1819          * Being under writeback is unlikely but possible.  For example direct
1820          * read to an mmaped fuse file will set the page dirty twice; once when
1821          * the pages are faulted with get_user_pages(), and then after the read
1822          * completed.
1823          */
1824         is_writeback = fuse_page_is_writeback(inode, page->index);
1825
1826         if (req && req->num_pages &&
1827             (is_writeback || req->num_pages == fc->max_pages ||
1828              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1829              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1830                 fuse_writepages_send(data);
1831                 data->req = NULL;
1832         } else if (req && req->num_pages == req->max_pages) {
1833                 if (!fuse_req_realloc_pages(fc, req, GFP_NOFS)) {
1834                         fuse_writepages_send(data);
1835                         req = data->req = NULL;
1836                 }
1837         }
1838
1839         err = -ENOMEM;
1840         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1841         if (!tmp_page)
1842                 goto out_unlock;
1843
1844         /*
1845          * The page must not be redirtied until the writeout is completed
1846          * (i.e. userspace has sent a reply to the write request).  Otherwise
1847          * there could be more than one temporary page instance for each real
1848          * page.
1849          *
1850          * This is ensured by holding the page lock in page_mkwrite() while
1851          * checking fuse_page_is_writeback().  We already hold the page lock
1852          * since clear_page_dirty_for_io() and keep it held until we add the
1853          * request to the fi->writepages list and increment req->num_pages.
1854          * After this fuse_page_is_writeback() will indicate that the page is
1855          * under writeback, so we can release the page lock.
1856          */
1857         if (data->req == NULL) {
1858                 struct fuse_inode *fi = get_fuse_inode(inode);
1859
1860                 err = -ENOMEM;
1861                 req = fuse_request_alloc_nofs(FUSE_REQ_INLINE_PAGES);
1862                 if (!req) {
1863                         __free_page(tmp_page);
1864                         goto out_unlock;
1865                 }
1866
1867                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1868                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1869                 req->misc.write.next = NULL;
1870                 req->in.argpages = 1;
1871                 __set_bit(FR_BACKGROUND, &req->flags);
1872                 req->num_pages = 0;
1873                 req->end = fuse_writepage_end;
1874                 req->inode = inode;
1875
1876                 spin_lock(&fc->lock);
1877                 list_add(&req->writepages_entry, &fi->writepages);
1878                 spin_unlock(&fc->lock);
1879
1880                 data->req = req;
1881         }
1882         set_page_writeback(page);
1883
1884         copy_highpage(tmp_page, page);
1885         req->pages[req->num_pages] = tmp_page;
1886         req->page_descs[req->num_pages].offset = 0;
1887         req->page_descs[req->num_pages].length = PAGE_SIZE;
1888
1889         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1890         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1891
1892         err = 0;
1893         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1894                 end_page_writeback(page);
1895                 data->req = NULL;
1896                 goto out_unlock;
1897         }
1898         data->orig_pages[req->num_pages] = page;
1899
1900         /*
1901          * Protected by fc->lock against concurrent access by
1902          * fuse_page_is_writeback().
1903          */
1904         spin_lock(&fc->lock);
1905         req->num_pages++;
1906         spin_unlock(&fc->lock);
1907
1908 out_unlock:
1909         unlock_page(page);
1910
1911         return err;
1912 }
1913
1914 static int fuse_writepages(struct address_space *mapping,
1915                            struct writeback_control *wbc)
1916 {
1917         struct inode *inode = mapping->host;
1918         struct fuse_conn *fc = get_fuse_conn(inode);
1919         struct fuse_fill_wb_data data;
1920         int err;
1921
1922         err = -EIO;
1923         if (is_bad_inode(inode))
1924                 goto out;
1925
1926         data.inode = inode;
1927         data.req = NULL;
1928         data.ff = NULL;
1929
1930         err = -ENOMEM;
1931         data.orig_pages = kcalloc(fc->max_pages,
1932                                   sizeof(struct page *),
1933                                   GFP_NOFS);
1934         if (!data.orig_pages)
1935                 goto out;
1936
1937         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1938         if (data.req) {
1939                 /* Ignore errors if we can write at least one page */
1940                 BUG_ON(!data.req->num_pages);
1941                 fuse_writepages_send(&data);
1942                 err = 0;
1943         }
1944         if (data.ff)
1945                 fuse_file_put(data.ff, false, false);
1946
1947         kfree(data.orig_pages);
1948 out:
1949         return err;
1950 }
1951
1952 /*
1953  * It's worthy to make sure that space is reserved on disk for the write,
1954  * but how to implement it without killing performance need more thinking.
1955  */
1956 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1957                 loff_t pos, unsigned len, unsigned flags,
1958                 struct page **pagep, void **fsdata)
1959 {
1960         pgoff_t index = pos >> PAGE_SHIFT;
1961         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1962         struct page *page;
1963         loff_t fsize;
1964         int err = -ENOMEM;
1965
1966         WARN_ON(!fc->writeback_cache);
1967
1968         page = grab_cache_page_write_begin(mapping, index, flags);
1969         if (!page)
1970                 goto error;
1971
1972         fuse_wait_on_page_writeback(mapping->host, page->index);
1973
1974         if (PageUptodate(page) || len == PAGE_SIZE)
1975                 goto success;
1976         /*
1977          * Check if the start this page comes after the end of file, in which
1978          * case the readpage can be optimized away.
1979          */
1980         fsize = i_size_read(mapping->host);
1981         if (fsize <= (pos & PAGE_MASK)) {
1982                 size_t off = pos & ~PAGE_MASK;
1983                 if (off)
1984                         zero_user_segment(page, 0, off);
1985                 goto success;
1986         }
1987         err = fuse_do_readpage(file, page);
1988         if (err)
1989                 goto cleanup;
1990 success:
1991         *pagep = page;
1992         return 0;
1993
1994 cleanup:
1995         unlock_page(page);
1996         put_page(page);
1997 error:
1998         return err;
1999 }
2000
2001 static int fuse_write_end(struct file *file, struct address_space *mapping,
2002                 loff_t pos, unsigned len, unsigned copied,
2003                 struct page *page, void *fsdata)
2004 {
2005         struct inode *inode = page->mapping->host;
2006
2007         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2008         if (!copied)
2009                 goto unlock;
2010
2011         if (!PageUptodate(page)) {
2012                 /* Zero any unwritten bytes at the end of the page */
2013                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2014                 if (endoff)
2015                         zero_user_segment(page, endoff, PAGE_SIZE);
2016                 SetPageUptodate(page);
2017         }
2018
2019         fuse_write_update_size(inode, pos + copied);
2020         set_page_dirty(page);
2021
2022 unlock:
2023         unlock_page(page);
2024         put_page(page);
2025
2026         return copied;
2027 }
2028
2029 static int fuse_launder_page(struct page *page)
2030 {
2031         int err = 0;
2032         if (clear_page_dirty_for_io(page)) {
2033                 struct inode *inode = page->mapping->host;
2034                 err = fuse_writepage_locked(page);
2035                 if (!err)
2036                         fuse_wait_on_page_writeback(inode, page->index);
2037         }
2038         return err;
2039 }
2040
2041 /*
2042  * Write back dirty pages now, because there may not be any suitable
2043  * open files later
2044  */
2045 static void fuse_vma_close(struct vm_area_struct *vma)
2046 {
2047         filemap_write_and_wait(vma->vm_file->f_mapping);
2048 }
2049
2050 /*
2051  * Wait for writeback against this page to complete before allowing it
2052  * to be marked dirty again, and hence written back again, possibly
2053  * before the previous writepage completed.
2054  *
2055  * Block here, instead of in ->writepage(), so that the userspace fs
2056  * can only block processes actually operating on the filesystem.
2057  *
2058  * Otherwise unprivileged userspace fs would be able to block
2059  * unrelated:
2060  *
2061  * - page migration
2062  * - sync(2)
2063  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2064  */
2065 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2066 {
2067         struct page *page = vmf->page;
2068         struct inode *inode = file_inode(vmf->vma->vm_file);
2069
2070         file_update_time(vmf->vma->vm_file);
2071         lock_page(page);
2072         if (page->mapping != inode->i_mapping) {
2073                 unlock_page(page);
2074                 return VM_FAULT_NOPAGE;
2075         }
2076
2077         fuse_wait_on_page_writeback(inode, page->index);
2078         return VM_FAULT_LOCKED;
2079 }
2080
2081 static const struct vm_operations_struct fuse_file_vm_ops = {
2082         .close          = fuse_vma_close,
2083         .fault          = filemap_fault,
2084         .map_pages      = filemap_map_pages,
2085         .page_mkwrite   = fuse_page_mkwrite,
2086 };
2087
2088 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2089 {
2090         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2091                 fuse_link_write_file(file);
2092
2093         file_accessed(file);
2094         vma->vm_ops = &fuse_file_vm_ops;
2095         return 0;
2096 }
2097
2098 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2099 {
2100         /* Can't provide the coherency needed for MAP_SHARED */
2101         if (vma->vm_flags & VM_MAYSHARE)
2102                 return -ENODEV;
2103
2104         invalidate_inode_pages2(file->f_mapping);
2105
2106         return generic_file_mmap(file, vma);
2107 }
2108
2109 static int convert_fuse_file_lock(struct fuse_conn *fc,
2110                                   const struct fuse_file_lock *ffl,
2111                                   struct file_lock *fl)
2112 {
2113         switch (ffl->type) {
2114         case F_UNLCK:
2115                 break;
2116
2117         case F_RDLCK:
2118         case F_WRLCK:
2119                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2120                     ffl->end < ffl->start)
2121                         return -EIO;
2122
2123                 fl->fl_start = ffl->start;
2124                 fl->fl_end = ffl->end;
2125
2126                 /*
2127                  * Convert pid into init's pid namespace.  The locks API will
2128                  * translate it into the caller's pid namespace.
2129                  */
2130                 rcu_read_lock();
2131                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2132                 rcu_read_unlock();
2133                 break;
2134
2135         default:
2136                 return -EIO;
2137         }
2138         fl->fl_type = ffl->type;
2139         return 0;
2140 }
2141
2142 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2143                          const struct file_lock *fl, int opcode, pid_t pid,
2144                          int flock, struct fuse_lk_in *inarg)
2145 {
2146         struct inode *inode = file_inode(file);
2147         struct fuse_conn *fc = get_fuse_conn(inode);
2148         struct fuse_file *ff = file->private_data;
2149
2150         memset(inarg, 0, sizeof(*inarg));
2151         inarg->fh = ff->fh;
2152         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2153         inarg->lk.start = fl->fl_start;
2154         inarg->lk.end = fl->fl_end;
2155         inarg->lk.type = fl->fl_type;
2156         inarg->lk.pid = pid;
2157         if (flock)
2158                 inarg->lk_flags |= FUSE_LK_FLOCK;
2159         args->in.h.opcode = opcode;
2160         args->in.h.nodeid = get_node_id(inode);
2161         args->in.numargs = 1;
2162         args->in.args[0].size = sizeof(*inarg);
2163         args->in.args[0].value = inarg;
2164 }
2165
2166 static int fuse_getlk(struct file *file, struct file_lock *fl)
2167 {
2168         struct inode *inode = file_inode(file);
2169         struct fuse_conn *fc = get_fuse_conn(inode);
2170         FUSE_ARGS(args);
2171         struct fuse_lk_in inarg;
2172         struct fuse_lk_out outarg;
2173         int err;
2174
2175         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2176         args.out.numargs = 1;
2177         args.out.args[0].size = sizeof(outarg);
2178         args.out.args[0].value = &outarg;
2179         err = fuse_simple_request(fc, &args);
2180         if (!err)
2181                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2182
2183         return err;
2184 }
2185
2186 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2187 {
2188         struct inode *inode = file_inode(file);
2189         struct fuse_conn *fc = get_fuse_conn(inode);
2190         FUSE_ARGS(args);
2191         struct fuse_lk_in inarg;
2192         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2193         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2194         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2195         int err;
2196
2197         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2198                 /* NLM needs asynchronous locks, which we don't support yet */
2199                 return -ENOLCK;
2200         }
2201
2202         /* Unlock on close is handled by the flush method */
2203         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2204                 return 0;
2205
2206         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2207         err = fuse_simple_request(fc, &args);
2208
2209         /* locking is restartable */
2210         if (err == -EINTR)
2211                 err = -ERESTARTSYS;
2212
2213         return err;
2214 }
2215
2216 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2217 {
2218         struct inode *inode = file_inode(file);
2219         struct fuse_conn *fc = get_fuse_conn(inode);
2220         int err;
2221
2222         if (cmd == F_CANCELLK) {
2223                 err = 0;
2224         } else if (cmd == F_GETLK) {
2225                 if (fc->no_lock) {
2226                         posix_test_lock(file, fl);
2227                         err = 0;
2228                 } else
2229                         err = fuse_getlk(file, fl);
2230         } else {
2231                 if (fc->no_lock)
2232                         err = posix_lock_file(file, fl, NULL);
2233                 else
2234                         err = fuse_setlk(file, fl, 0);
2235         }
2236         return err;
2237 }
2238
2239 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2240 {
2241         struct inode *inode = file_inode(file);
2242         struct fuse_conn *fc = get_fuse_conn(inode);
2243         int err;
2244
2245         if (fc->no_flock) {
2246                 err = locks_lock_file_wait(file, fl);
2247         } else {
2248                 struct fuse_file *ff = file->private_data;
2249
2250                 /* emulate flock with POSIX locks */
2251                 ff->flock = true;
2252                 err = fuse_setlk(file, fl, 1);
2253         }
2254
2255         return err;
2256 }
2257
2258 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2259 {
2260         struct inode *inode = mapping->host;
2261         struct fuse_conn *fc = get_fuse_conn(inode);
2262         FUSE_ARGS(args);
2263         struct fuse_bmap_in inarg;
2264         struct fuse_bmap_out outarg;
2265         int err;
2266
2267         if (!inode->i_sb->s_bdev || fc->no_bmap)
2268                 return 0;
2269
2270         memset(&inarg, 0, sizeof(inarg));
2271         inarg.block = block;
2272         inarg.blocksize = inode->i_sb->s_blocksize;
2273         args.in.h.opcode = FUSE_BMAP;
2274         args.in.h.nodeid = get_node_id(inode);
2275         args.in.numargs = 1;
2276         args.in.args[0].size = sizeof(inarg);
2277         args.in.args[0].value = &inarg;
2278         args.out.numargs = 1;
2279         args.out.args[0].size = sizeof(outarg);
2280         args.out.args[0].value = &outarg;
2281         err = fuse_simple_request(fc, &args);
2282         if (err == -ENOSYS)
2283                 fc->no_bmap = 1;
2284
2285         return err ? 0 : outarg.block;
2286 }
2287
2288 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2289 {
2290         struct inode *inode = file->f_mapping->host;
2291         struct fuse_conn *fc = get_fuse_conn(inode);
2292         struct fuse_file *ff = file->private_data;
2293         FUSE_ARGS(args);
2294         struct fuse_lseek_in inarg = {
2295                 .fh = ff->fh,
2296                 .offset = offset,
2297                 .whence = whence
2298         };
2299         struct fuse_lseek_out outarg;
2300         int err;
2301
2302         if (fc->no_lseek)
2303                 goto fallback;
2304
2305         args.in.h.opcode = FUSE_LSEEK;
2306         args.in.h.nodeid = ff->nodeid;
2307         args.in.numargs = 1;
2308         args.in.args[0].size = sizeof(inarg);
2309         args.in.args[0].value = &inarg;
2310         args.out.numargs = 1;
2311         args.out.args[0].size = sizeof(outarg);
2312         args.out.args[0].value = &outarg;
2313         err = fuse_simple_request(fc, &args);
2314         if (err) {
2315                 if (err == -ENOSYS) {
2316                         fc->no_lseek = 1;
2317                         goto fallback;
2318                 }
2319                 return err;
2320         }
2321
2322         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2323
2324 fallback:
2325         err = fuse_update_attributes(inode, file);
2326         if (!err)
2327                 return generic_file_llseek(file, offset, whence);
2328         else
2329                 return err;
2330 }
2331
2332 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2333 {
2334         loff_t retval;
2335         struct inode *inode = file_inode(file);
2336
2337         switch (whence) {
2338         case SEEK_SET:
2339         case SEEK_CUR:
2340                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2341                 retval = generic_file_llseek(file, offset, whence);
2342                 break;
2343         case SEEK_END:
2344                 inode_lock(inode);
2345                 retval = fuse_update_attributes(inode, file);
2346                 if (!retval)
2347                         retval = generic_file_llseek(file, offset, whence);
2348                 inode_unlock(inode);
2349                 break;
2350         case SEEK_HOLE:
2351         case SEEK_DATA:
2352                 inode_lock(inode);
2353                 retval = fuse_lseek(file, offset, whence);
2354                 inode_unlock(inode);
2355                 break;
2356         default:
2357                 retval = -EINVAL;
2358         }
2359
2360         return retval;
2361 }
2362
2363 /*
2364  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2365  * ABI was defined to be 'struct iovec' which is different on 32bit
2366  * and 64bit.  Fortunately we can determine which structure the server
2367  * used from the size of the reply.
2368  */
2369 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2370                                      size_t transferred, unsigned count,
2371                                      bool is_compat)
2372 {
2373 #ifdef CONFIG_COMPAT
2374         if (count * sizeof(struct compat_iovec) == transferred) {
2375                 struct compat_iovec *ciov = src;
2376                 unsigned i;
2377
2378                 /*
2379                  * With this interface a 32bit server cannot support
2380                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2381                  * requests
2382                  */
2383                 if (!is_compat)
2384                         return -EINVAL;
2385
2386                 for (i = 0; i < count; i++) {
2387                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2388                         dst[i].iov_len = ciov[i].iov_len;
2389                 }
2390                 return 0;
2391         }
2392 #endif
2393
2394         if (count * sizeof(struct iovec) != transferred)
2395                 return -EIO;
2396
2397         memcpy(dst, src, transferred);
2398         return 0;
2399 }
2400
2401 /* Make sure iov_length() won't overflow */
2402 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2403                                  size_t count)
2404 {
2405         size_t n;
2406         u32 max = fc->max_pages << PAGE_SHIFT;
2407
2408         for (n = 0; n < count; n++, iov++) {
2409                 if (iov->iov_len > (size_t) max)
2410                         return -ENOMEM;
2411                 max -= iov->iov_len;
2412         }
2413         return 0;
2414 }
2415
2416 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2417                                  void *src, size_t transferred, unsigned count,
2418                                  bool is_compat)
2419 {
2420         unsigned i;
2421         struct fuse_ioctl_iovec *fiov = src;
2422
2423         if (fc->minor < 16) {
2424                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2425                                                  count, is_compat);
2426         }
2427
2428         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2429                 return -EIO;
2430
2431         for (i = 0; i < count; i++) {
2432                 /* Did the server supply an inappropriate value? */
2433                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2434                     fiov[i].len != (unsigned long) fiov[i].len)
2435                         return -EIO;
2436
2437                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2438                 dst[i].iov_len = (size_t) fiov[i].len;
2439
2440 #ifdef CONFIG_COMPAT
2441                 if (is_compat &&
2442                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2443                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2444                         return -EIO;
2445 #endif
2446         }
2447
2448         return 0;
2449 }
2450
2451
2452 /*
2453  * For ioctls, there is no generic way to determine how much memory
2454  * needs to be read and/or written.  Furthermore, ioctls are allowed
2455  * to dereference the passed pointer, so the parameter requires deep
2456  * copying but FUSE has no idea whatsoever about what to copy in or
2457  * out.
2458  *
2459  * This is solved by allowing FUSE server to retry ioctl with
2460  * necessary in/out iovecs.  Let's assume the ioctl implementation
2461  * needs to read in the following structure.
2462  *
2463  * struct a {
2464  *      char    *buf;
2465  *      size_t  buflen;
2466  * }
2467  *
2468  * On the first callout to FUSE server, inarg->in_size and
2469  * inarg->out_size will be NULL; then, the server completes the ioctl
2470  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2471  * the actual iov array to
2472  *
2473  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2474  *
2475  * which tells FUSE to copy in the requested area and retry the ioctl.
2476  * On the second round, the server has access to the structure and
2477  * from that it can tell what to look for next, so on the invocation,
2478  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2479  *
2480  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2481  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2482  *
2483  * FUSE will copy both struct a and the pointed buffer from the
2484  * process doing the ioctl and retry ioctl with both struct a and the
2485  * buffer.
2486  *
2487  * This time, FUSE server has everything it needs and completes ioctl
2488  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2489  *
2490  * Copying data out works the same way.
2491  *
2492  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2493  * automatically initializes in and out iovs by decoding @cmd with
2494  * _IOC_* macros and the server is not allowed to request RETRY.  This
2495  * limits ioctl data transfers to well-formed ioctls and is the forced
2496  * behavior for all FUSE servers.
2497  */
2498 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2499                    unsigned int flags)
2500 {
2501         struct fuse_file *ff = file->private_data;
2502         struct fuse_conn *fc = ff->fc;
2503         struct fuse_ioctl_in inarg = {
2504                 .fh = ff->fh,
2505                 .cmd = cmd,
2506                 .arg = arg,
2507                 .flags = flags
2508         };
2509         struct fuse_ioctl_out outarg;
2510         struct fuse_req *req = NULL;
2511         struct page **pages = NULL;
2512         struct iovec *iov_page = NULL;
2513         struct iovec *in_iov = NULL, *out_iov = NULL;
2514         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2515         size_t in_size, out_size, transferred, c;
2516         int err, i;
2517         struct iov_iter ii;
2518
2519 #if BITS_PER_LONG == 32
2520         inarg.flags |= FUSE_IOCTL_32BIT;
2521 #else
2522         if (flags & FUSE_IOCTL_COMPAT)
2523                 inarg.flags |= FUSE_IOCTL_32BIT;
2524 #endif
2525
2526         /* assume all the iovs returned by client always fits in a page */
2527         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2528
2529         err = -ENOMEM;
2530         pages = kcalloc(fc->max_pages, sizeof(pages[0]), GFP_KERNEL);
2531         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2532         if (!pages || !iov_page)
2533                 goto out;
2534
2535         /*
2536          * If restricted, initialize IO parameters as encoded in @cmd.
2537          * RETRY from server is not allowed.
2538          */
2539         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2540                 struct iovec *iov = iov_page;
2541
2542                 iov->iov_base = (void __user *)arg;
2543                 iov->iov_len = _IOC_SIZE(cmd);
2544
2545                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2546                         in_iov = iov;
2547                         in_iovs = 1;
2548                 }
2549
2550                 if (_IOC_DIR(cmd) & _IOC_READ) {
2551                         out_iov = iov;
2552                         out_iovs = 1;
2553                 }
2554         }
2555
2556  retry:
2557         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2558         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2559
2560         /*
2561          * Out data can be used either for actual out data or iovs,
2562          * make sure there always is at least one page.
2563          */
2564         out_size = max_t(size_t, out_size, PAGE_SIZE);
2565         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2566
2567         /* make sure there are enough buffer pages and init request with them */
2568         err = -ENOMEM;
2569         if (max_pages > fc->max_pages)
2570                 goto out;
2571         while (num_pages < max_pages) {
2572                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2573                 if (!pages[num_pages])
2574                         goto out;
2575                 num_pages++;
2576         }
2577
2578         req = fuse_get_req(fc, num_pages);
2579         if (IS_ERR(req)) {
2580                 err = PTR_ERR(req);
2581                 req = NULL;
2582                 goto out;
2583         }
2584         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2585         req->num_pages = num_pages;
2586         fuse_page_descs_length_init(req, 0, req->num_pages);
2587
2588         /* okay, let's send it to the client */
2589         req->in.h.opcode = FUSE_IOCTL;
2590         req->in.h.nodeid = ff->nodeid;
2591         req->in.numargs = 1;
2592         req->in.args[0].size = sizeof(inarg);
2593         req->in.args[0].value = &inarg;
2594         if (in_size) {
2595                 req->in.numargs++;
2596                 req->in.args[1].size = in_size;
2597                 req->in.argpages = 1;
2598
2599                 err = -EFAULT;
2600                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2601                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2602                         c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2603                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2604                                 goto out;
2605                 }
2606         }
2607
2608         req->out.numargs = 2;
2609         req->out.args[0].size = sizeof(outarg);
2610         req->out.args[0].value = &outarg;
2611         req->out.args[1].size = out_size;
2612         req->out.argpages = 1;
2613         req->out.argvar = 1;
2614
2615         fuse_request_send(fc, req);
2616         err = req->out.h.error;
2617         transferred = req->out.args[1].size;
2618         fuse_put_request(fc, req);
2619         req = NULL;
2620         if (err)
2621                 goto out;
2622
2623         /* did it ask for retry? */
2624         if (outarg.flags & FUSE_IOCTL_RETRY) {
2625                 void *vaddr;
2626
2627                 /* no retry if in restricted mode */
2628                 err = -EIO;
2629                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2630                         goto out;
2631
2632                 in_iovs = outarg.in_iovs;
2633                 out_iovs = outarg.out_iovs;
2634
2635                 /*
2636                  * Make sure things are in boundary, separate checks
2637                  * are to protect against overflow.
2638                  */
2639                 err = -ENOMEM;
2640                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2641                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2642                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2643                         goto out;
2644
2645                 vaddr = kmap_atomic(pages[0]);
2646                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2647                                             transferred, in_iovs + out_iovs,
2648                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2649                 kunmap_atomic(vaddr);
2650                 if (err)
2651                         goto out;
2652
2653                 in_iov = iov_page;
2654                 out_iov = in_iov + in_iovs;
2655
2656                 err = fuse_verify_ioctl_iov(fc, in_iov, in_iovs);
2657                 if (err)
2658                         goto out;
2659
2660                 err = fuse_verify_ioctl_iov(fc, out_iov, out_iovs);
2661                 if (err)
2662                         goto out;
2663
2664                 goto retry;
2665         }
2666
2667         err = -EIO;
2668         if (transferred > inarg.out_size)
2669                 goto out;
2670
2671         err = -EFAULT;
2672         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2673         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2674                 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2675                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2676                         goto out;
2677         }
2678         err = 0;
2679  out:
2680         if (req)
2681                 fuse_put_request(fc, req);
2682         free_page((unsigned long) iov_page);
2683         while (num_pages)
2684                 __free_page(pages[--num_pages]);
2685         kfree(pages);
2686
2687         return err ? err : outarg.result;
2688 }
2689 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2690
2691 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2692                        unsigned long arg, unsigned int flags)
2693 {
2694         struct inode *inode = file_inode(file);
2695         struct fuse_conn *fc = get_fuse_conn(inode);
2696
2697         if (!fuse_allow_current_process(fc))
2698                 return -EACCES;
2699
2700         if (is_bad_inode(inode))
2701                 return -EIO;
2702
2703         return fuse_do_ioctl(file, cmd, arg, flags);
2704 }
2705
2706 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2707                             unsigned long arg)
2708 {
2709         return fuse_ioctl_common(file, cmd, arg, 0);
2710 }
2711
2712 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2713                                    unsigned long arg)
2714 {
2715         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2716 }
2717
2718 /*
2719  * All files which have been polled are linked to RB tree
2720  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2721  * find the matching one.
2722  */
2723 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2724                                               struct rb_node **parent_out)
2725 {
2726         struct rb_node **link = &fc->polled_files.rb_node;
2727         struct rb_node *last = NULL;
2728
2729         while (*link) {
2730                 struct fuse_file *ff;
2731
2732                 last = *link;
2733                 ff = rb_entry(last, struct fuse_file, polled_node);
2734
2735                 if (kh < ff->kh)
2736                         link = &last->rb_left;
2737                 else if (kh > ff->kh)
2738                         link = &last->rb_right;
2739                 else
2740                         return link;
2741         }
2742
2743         if (parent_out)
2744                 *parent_out = last;
2745         return link;
2746 }
2747
2748 /*
2749  * The file is about to be polled.  Make sure it's on the polled_files
2750  * RB tree.  Note that files once added to the polled_files tree are
2751  * not removed before the file is released.  This is because a file
2752  * polled once is likely to be polled again.
2753  */
2754 static void fuse_register_polled_file(struct fuse_conn *fc,
2755                                       struct fuse_file *ff)
2756 {
2757         spin_lock(&fc->lock);
2758         if (RB_EMPTY_NODE(&ff->polled_node)) {
2759                 struct rb_node **link, *uninitialized_var(parent);
2760
2761                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2762                 BUG_ON(*link);
2763                 rb_link_node(&ff->polled_node, parent, link);
2764                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2765         }
2766         spin_unlock(&fc->lock);
2767 }
2768
2769 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2770 {
2771         struct fuse_file *ff = file->private_data;
2772         struct fuse_conn *fc = ff->fc;
2773         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2774         struct fuse_poll_out outarg;
2775         FUSE_ARGS(args);
2776         int err;
2777
2778         if (fc->no_poll)
2779                 return DEFAULT_POLLMASK;
2780
2781         poll_wait(file, &ff->poll_wait, wait);
2782         inarg.events = mangle_poll(poll_requested_events(wait));
2783
2784         /*
2785          * Ask for notification iff there's someone waiting for it.
2786          * The client may ignore the flag and always notify.
2787          */
2788         if (waitqueue_active(&ff->poll_wait)) {
2789                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2790                 fuse_register_polled_file(fc, ff);
2791         }
2792
2793         args.in.h.opcode = FUSE_POLL;
2794         args.in.h.nodeid = ff->nodeid;
2795         args.in.numargs = 1;
2796         args.in.args[0].size = sizeof(inarg);
2797         args.in.args[0].value = &inarg;
2798         args.out.numargs = 1;
2799         args.out.args[0].size = sizeof(outarg);
2800         args.out.args[0].value = &outarg;
2801         err = fuse_simple_request(fc, &args);
2802
2803         if (!err)
2804                 return demangle_poll(outarg.revents);
2805         if (err == -ENOSYS) {
2806                 fc->no_poll = 1;
2807                 return DEFAULT_POLLMASK;
2808         }
2809         return EPOLLERR;
2810 }
2811 EXPORT_SYMBOL_GPL(fuse_file_poll);
2812
2813 /*
2814  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2815  * wakes up the poll waiters.
2816  */
2817 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2818                             struct fuse_notify_poll_wakeup_out *outarg)
2819 {
2820         u64 kh = outarg->kh;
2821         struct rb_node **link;
2822
2823         spin_lock(&fc->lock);
2824
2825         link = fuse_find_polled_node(fc, kh, NULL);
2826         if (*link) {
2827                 struct fuse_file *ff;
2828
2829                 ff = rb_entry(*link, struct fuse_file, polled_node);
2830                 wake_up_interruptible_sync(&ff->poll_wait);
2831         }
2832
2833         spin_unlock(&fc->lock);
2834         return 0;
2835 }
2836
2837 static void fuse_do_truncate(struct file *file)
2838 {
2839         struct inode *inode = file->f_mapping->host;
2840         struct iattr attr;
2841
2842         attr.ia_valid = ATTR_SIZE;
2843         attr.ia_size = i_size_read(inode);
2844
2845         attr.ia_file = file;
2846         attr.ia_valid |= ATTR_FILE;
2847
2848         fuse_do_setattr(file_dentry(file), &attr, file);
2849 }
2850
2851 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2852 {
2853         return round_up(off, fc->max_pages << PAGE_SHIFT);
2854 }
2855
2856 static ssize_t
2857 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2858 {
2859         DECLARE_COMPLETION_ONSTACK(wait);
2860         ssize_t ret = 0;
2861         struct file *file = iocb->ki_filp;
2862         struct fuse_file *ff = file->private_data;
2863         bool async_dio = ff->fc->async_dio;
2864         loff_t pos = 0;
2865         struct inode *inode;
2866         loff_t i_size;
2867         size_t count = iov_iter_count(iter);
2868         loff_t offset = iocb->ki_pos;
2869         struct fuse_io_priv *io;
2870
2871         pos = offset;
2872         inode = file->f_mapping->host;
2873         i_size = i_size_read(inode);
2874
2875         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2876                 return 0;
2877
2878         /* optimization for short read */
2879         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2880                 if (offset >= i_size)
2881                         return 0;
2882                 iov_iter_truncate(iter, fuse_round_up(ff->fc, i_size - offset));
2883                 count = iov_iter_count(iter);
2884         }
2885
2886         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2887         if (!io)
2888                 return -ENOMEM;
2889         spin_lock_init(&io->lock);
2890         kref_init(&io->refcnt);
2891         io->reqs = 1;
2892         io->bytes = -1;
2893         io->size = 0;
2894         io->offset = offset;
2895         io->write = (iov_iter_rw(iter) == WRITE);
2896         io->err = 0;
2897         /*
2898          * By default, we want to optimize all I/Os with async request
2899          * submission to the client filesystem if supported.
2900          */
2901         io->async = async_dio;
2902         io->iocb = iocb;
2903         io->blocking = is_sync_kiocb(iocb);
2904
2905         /*
2906          * We cannot asynchronously extend the size of a file.
2907          * In such case the aio will behave exactly like sync io.
2908          */
2909         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2910                 io->blocking = true;
2911
2912         if (io->async && io->blocking) {
2913                 /*
2914                  * Additional reference to keep io around after
2915                  * calling fuse_aio_complete()
2916                  */
2917                 kref_get(&io->refcnt);
2918                 io->done = &wait;
2919         }
2920
2921         if (iov_iter_rw(iter) == WRITE) {
2922                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2923                 fuse_invalidate_attr(inode);
2924         } else {
2925                 ret = __fuse_direct_read(io, iter, &pos);
2926         }
2927
2928         if (io->async) {
2929                 bool blocking = io->blocking;
2930
2931                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2932
2933                 /* we have a non-extending, async request, so return */
2934                 if (!blocking)
2935                         return -EIOCBQUEUED;
2936
2937                 wait_for_completion(&wait);
2938                 ret = fuse_get_res_by_io(io);
2939         }
2940
2941         kref_put(&io->refcnt, fuse_io_release);
2942
2943         if (iov_iter_rw(iter) == WRITE) {
2944                 if (ret > 0)
2945                         fuse_write_update_size(inode, pos);
2946                 else if (ret < 0 && offset + count > i_size)
2947                         fuse_do_truncate(file);
2948         }
2949
2950         return ret;
2951 }
2952
2953 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2954                                 loff_t length)
2955 {
2956         struct fuse_file *ff = file->private_data;
2957         struct inode *inode = file_inode(file);
2958         struct fuse_inode *fi = get_fuse_inode(inode);
2959         struct fuse_conn *fc = ff->fc;
2960         FUSE_ARGS(args);
2961         struct fuse_fallocate_in inarg = {
2962                 .fh = ff->fh,
2963                 .offset = offset,
2964                 .length = length,
2965                 .mode = mode
2966         };
2967         int err;
2968         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2969                            (mode & FALLOC_FL_PUNCH_HOLE);
2970
2971         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2972                 return -EOPNOTSUPP;
2973
2974         if (fc->no_fallocate)
2975                 return -EOPNOTSUPP;
2976
2977         if (lock_inode) {
2978                 inode_lock(inode);
2979                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2980                         loff_t endbyte = offset + length - 1;
2981                         err = filemap_write_and_wait_range(inode->i_mapping,
2982                                                            offset, endbyte);
2983                         if (err)
2984                                 goto out;
2985
2986                         fuse_sync_writes(inode);
2987                 }
2988         }
2989
2990         if (!(mode & FALLOC_FL_KEEP_SIZE))
2991                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2992
2993         args.in.h.opcode = FUSE_FALLOCATE;
2994         args.in.h.nodeid = ff->nodeid;
2995         args.in.numargs = 1;
2996         args.in.args[0].size = sizeof(inarg);
2997         args.in.args[0].value = &inarg;
2998         err = fuse_simple_request(fc, &args);
2999         if (err == -ENOSYS) {
3000                 fc->no_fallocate = 1;
3001                 err = -EOPNOTSUPP;
3002         }
3003         if (err)
3004                 goto out;
3005
3006         /* we could have extended the file */
3007         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3008                 bool changed = fuse_write_update_size(inode, offset + length);
3009
3010                 if (changed && fc->writeback_cache)
3011                         file_update_time(file);
3012         }
3013
3014         if (mode & FALLOC_FL_PUNCH_HOLE)
3015                 truncate_pagecache_range(inode, offset, offset + length - 1);
3016
3017         fuse_invalidate_attr(inode);
3018
3019 out:
3020         if (!(mode & FALLOC_FL_KEEP_SIZE))
3021                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3022
3023         if (lock_inode)
3024                 inode_unlock(inode);
3025
3026         return err;
3027 }
3028
3029 static ssize_t fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3030                                     struct file *file_out, loff_t pos_out,
3031                                     size_t len, unsigned int flags)
3032 {
3033         struct fuse_file *ff_in = file_in->private_data;
3034         struct fuse_file *ff_out = file_out->private_data;
3035         struct inode *inode_out = file_inode(file_out);
3036         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3037         struct fuse_conn *fc = ff_in->fc;
3038         FUSE_ARGS(args);
3039         struct fuse_copy_file_range_in inarg = {
3040                 .fh_in = ff_in->fh,
3041                 .off_in = pos_in,
3042                 .nodeid_out = ff_out->nodeid,
3043                 .fh_out = ff_out->fh,
3044                 .off_out = pos_out,
3045                 .len = len,
3046                 .flags = flags
3047         };
3048         struct fuse_write_out outarg;
3049         ssize_t err;
3050         /* mark unstable when write-back is not used, and file_out gets
3051          * extended */
3052         bool is_unstable = (!fc->writeback_cache) &&
3053                            ((pos_out + len) > inode_out->i_size);
3054
3055         if (fc->no_copy_file_range)
3056                 return -EOPNOTSUPP;
3057
3058         inode_lock(inode_out);
3059
3060         if (fc->writeback_cache) {
3061                 err = filemap_write_and_wait_range(inode_out->i_mapping,
3062                                                    pos_out, pos_out + len);
3063                 if (err)
3064                         goto out;
3065
3066                 fuse_sync_writes(inode_out);
3067         }
3068
3069         if (is_unstable)
3070                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3071
3072         args.in.h.opcode = FUSE_COPY_FILE_RANGE;
3073         args.in.h.nodeid = ff_in->nodeid;
3074         args.in.numargs = 1;
3075         args.in.args[0].size = sizeof(inarg);
3076         args.in.args[0].value = &inarg;
3077         args.out.numargs = 1;
3078         args.out.args[0].size = sizeof(outarg);
3079         args.out.args[0].value = &outarg;
3080         err = fuse_simple_request(fc, &args);
3081         if (err == -ENOSYS) {
3082                 fc->no_copy_file_range = 1;
3083                 err = -EOPNOTSUPP;
3084         }
3085         if (err)
3086                 goto out;
3087
3088         if (fc->writeback_cache) {
3089                 fuse_write_update_size(inode_out, pos_out + outarg.size);
3090                 file_update_time(file_out);
3091         }
3092
3093         fuse_invalidate_attr(inode_out);
3094
3095         err = outarg.size;
3096 out:
3097         if (is_unstable)
3098                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3099
3100         inode_unlock(inode_out);
3101
3102         return err;
3103 }
3104
3105 static const struct file_operations fuse_file_operations = {
3106         .llseek         = fuse_file_llseek,
3107         .read_iter      = fuse_file_read_iter,
3108         .write_iter     = fuse_file_write_iter,
3109         .mmap           = fuse_file_mmap,
3110         .open           = fuse_open,
3111         .flush          = fuse_flush,
3112         .release        = fuse_release,
3113         .fsync          = fuse_fsync,
3114         .lock           = fuse_file_lock,
3115         .flock          = fuse_file_flock,
3116         .splice_read    = generic_file_splice_read,
3117         .unlocked_ioctl = fuse_file_ioctl,
3118         .compat_ioctl   = fuse_file_compat_ioctl,
3119         .poll           = fuse_file_poll,
3120         .fallocate      = fuse_file_fallocate,
3121         .copy_file_range = fuse_copy_file_range,
3122 };
3123
3124 static const struct file_operations fuse_direct_io_file_operations = {
3125         .llseek         = fuse_file_llseek,
3126         .read_iter      = fuse_direct_read_iter,
3127         .write_iter     = fuse_direct_write_iter,
3128         .mmap           = fuse_direct_mmap,
3129         .open           = fuse_open,
3130         .flush          = fuse_flush,
3131         .release        = fuse_release,
3132         .fsync          = fuse_fsync,
3133         .lock           = fuse_file_lock,
3134         .flock          = fuse_file_flock,
3135         .unlocked_ioctl = fuse_file_ioctl,
3136         .compat_ioctl   = fuse_file_compat_ioctl,
3137         .poll           = fuse_file_poll,
3138         .fallocate      = fuse_file_fallocate,
3139         /* no splice_read */
3140 };
3141
3142 static const struct address_space_operations fuse_file_aops  = {
3143         .readpage       = fuse_readpage,
3144         .writepage      = fuse_writepage,
3145         .writepages     = fuse_writepages,
3146         .launder_page   = fuse_launder_page,
3147         .readpages      = fuse_readpages,
3148         .set_page_dirty = __set_page_dirty_nobuffers,
3149         .bmap           = fuse_bmap,
3150         .direct_IO      = fuse_direct_IO,
3151         .write_begin    = fuse_write_begin,
3152         .write_end      = fuse_write_end,
3153 };
3154
3155 void fuse_init_file_inode(struct inode *inode)
3156 {
3157         struct fuse_inode *fi = get_fuse_inode(inode);
3158
3159         inode->i_fop = &fuse_file_operations;
3160         inode->i_data.a_ops = &fuse_file_aops;
3161
3162         INIT_LIST_HEAD(&fi->write_files);
3163         INIT_LIST_HEAD(&fi->queued_writes);
3164         fi->writectr = 0;
3165         init_waitqueue_head(&fi->page_waitq);
3166         INIT_LIST_HEAD(&fi->writepages);
3167 }