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 int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  23                           int opcode, struct fuse_open_out *outargp)
  24 {
  25         struct fuse_open_in inarg;
  26         FUSE_ARGS(args);
  27
  28         memset(&inarg, 0, sizeof(inarg));
  29         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  30         if (!fc->atomic_o_trunc)
  31                 inarg.flags &= ~O_TRUNC;
  32         args.in.h.opcode = opcode;
  33         args.in.h.nodeid = nodeid;
  34         args.in.numargs = 1;
  35         args.in.args[0].size = sizeof(inarg);
  36         args.in.args[0].value = &inarg;
  37         args.out.numargs = 1;
  38         args.out.args[0].size = sizeof(*outargp);
  39         args.out.args[0].value = outargp;
  40
  41         return fuse_simple_request(fc, &args);
  42 }
  43
  44 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  45 {
  46         struct fuse_file *ff;
  47
  48         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
  49         if (unlikely(!ff))
  50                 return NULL;
  51
  52         ff->fc = fc;
  53         ff->reserved_req = fuse_request_alloc(0);
  54         if (unlikely(!ff->reserved_req)) {
  55                 kfree(ff);
  56                 return NULL;
  57         }
  58
  59         INIT_LIST_HEAD(&ff->write_entry);
  60         mutex_init(&ff->readdir.lock);
  61         refcount_set(&ff->count, 1);
  62         RB_CLEAR_NODE(&ff->polled_node);
  63         init_waitqueue_head(&ff->poll_wait);
  64
  65         ff->kh = atomic64_inc_return(&fc->khctr);
  66
  67         return ff;
  68 }
  69
  70 void fuse_file_free(struct fuse_file *ff)
  71 {
  72         fuse_request_free(ff->reserved_req);
  73         mutex_destroy(&ff->readdir.lock);
  74         kfree(ff);
  75 }
  76
  77 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  78 {
  79         refcount_inc(&ff->count);
  80         return ff;
  81 }
  82
  83 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  84 {
  85         iput(req->misc.release.inode);
  86 }
  87
  88 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
  89 {
  90         if (refcount_dec_and_test(&ff->count)) {
  91                 struct fuse_req *req = ff->reserved_req;
  92
  93                 if (isdir ? ff->fc->no_opendir : ff->fc->no_open) {
  94                         /*
  95                          * Drop the release request when client does not
  96                          * implement 'open'
  97                          */
  98                         __clear_bit(FR_BACKGROUND, &req->flags);
  99                         iput(req->misc.release.inode);
 100                         fuse_put_request(ff->fc, req);
 101                 } else if (sync) {
 102                         __set_bit(FR_FORCE, &req->flags);
 103                         __clear_bit(FR_BACKGROUND, &req->flags);
 104                         fuse_request_send(ff->fc, req);
 105                         iput(req->misc.release.inode);
 106                         fuse_put_request(ff->fc, req);
 107                 } else {
 108                         req->end = fuse_release_end;
 109                         __set_bit(FR_BACKGROUND, &req->flags);
 110                         fuse_request_send_background(ff->fc, req);
 111                 }
 112                 kfree(ff);
 113         }
 114 }
 115
 116 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 117                  bool isdir)
 118 {
 119         struct fuse_file *ff;
 120         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 121
 122         ff = fuse_file_alloc(fc);
 123         if (!ff)
 124                 return -ENOMEM;
 125
 126         ff->fh = 0;
 127         /* Default for no-open */
 128         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
 129         if (isdir ? !fc->no_opendir : !fc->no_open) {
 130                 struct fuse_open_out outarg;
 131                 int err;
 132
 133                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 134                 if (!err) {
 135                         ff->fh = outarg.fh;
 136                         ff->open_flags = outarg.open_flags;
 137
 138                 } else if (err != -ENOSYS) {
 139                         fuse_file_free(ff);
 140                         return err;
 141                 } else {
 142                         if (isdir)
 143                                 fc->no_opendir = 1;
 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_inode *fi = get_fuse_inode(inode);
 163         struct fuse_file *ff = file->private_data;
 164         /*
 165          * file may be written through mmap, so chain it onto the
 166          * inodes's write_file list
 167          */
 168         spin_lock(&fi->lock);
 169         if (list_empty(&ff->write_entry))
 170                 list_add(&ff->write_entry, &fi->write_files);
 171         spin_unlock(&fi->lock);
 172 }
 173
 174 void fuse_finish_open(struct inode *inode, struct file *file)
 175 {
 176         struct fuse_file *ff = file->private_data;
 177         struct fuse_conn *fc = get_fuse_conn(inode);
 178
 179         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 180                 invalidate_inode_pages2(inode->i_mapping);
 181         if (ff->open_flags & FOPEN_STREAM)
 182                 stream_open(inode, file);
 183         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 184                 nonseekable_open(inode, file);
 185         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 186                 struct fuse_inode *fi = get_fuse_inode(inode);
 187
 188                 spin_lock(&fi->lock);
 189                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 190                 i_size_write(inode, 0);
 191                 spin_unlock(&fi->lock);
 192                 fuse_invalidate_attr(inode);
 193                 if (fc->writeback_cache)
 194                         file_update_time(file);
 195         }
 196         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 197                 fuse_link_write_file(file);
 198 }
 199
 200 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 201 {
 202         struct fuse_conn *fc = get_fuse_conn(inode);
 203         int err;
 204         bool lock_inode = (file->f_flags & O_TRUNC) &&
 205                           fc->atomic_o_trunc &&
 206                           fc->writeback_cache;
 207
 208         err = generic_file_open(inode, file);
 209         if (err)
 210                 return err;
 211
 212         if (lock_inode)
 213                 inode_lock(inode);
 214
 215         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 216
 217         if (!err)
 218                 fuse_finish_open(inode, file);
 219
 220         if (lock_inode)
 221                 inode_unlock(inode);
 222
 223         return err;
 224 }
 225
 226 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 227                                  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         /* Inode is NULL on error path of fuse_create_open() */
 234         if (likely(fi)) {
 235                 spin_lock(&fi->lock);
 236                 list_del(&ff->write_entry);
 237                 spin_unlock(&fi->lock);
 238         }
 239         spin_lock(&fc->lock);
 240         if (!RB_EMPTY_NODE(&ff->polled_node))
 241                 rb_erase(&ff->polled_node, &fc->polled_files);
 242         spin_unlock(&fc->lock);
 243
 244         wake_up_interruptible_all(&ff->poll_wait);
 245
 246         inarg->fh = ff->fh;
 247         inarg->flags = flags;
 248         req->in.h.opcode = opcode;
 249         req->in.h.nodeid = ff->nodeid;
 250         req->in.numargs = 1;
 251         req->in.args[0].size = sizeof(struct fuse_release_in);
 252         req->in.args[0].value = inarg;
 253 }
 254
 255 void fuse_release_common(struct file *file, bool isdir)
 256 {
 257         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 258         struct fuse_file *ff = file->private_data;
 259         struct fuse_req *req = ff->reserved_req;
 260         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 261
 262         fuse_prepare_release(fi, ff, file->f_flags, opcode);
 263
 264         if (ff->flock) {
 265                 struct fuse_release_in *inarg = &req->misc.release.in;
 266                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 267                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 268                                                        (fl_owner_t) file);
 269         }
 270         /* Hold inode until release is finished */
 271         req->misc.release.inode = igrab(file_inode(file));
 272
 273         /*
 274          * Normally this will send the RELEASE request, however if
 275          * some asynchronous READ or WRITE requests are outstanding,
 276          * the sending will be delayed.
 277          *
 278          * Make the release synchronous if this is a fuseblk mount,
 279          * synchronous RELEASE is allowed (and desirable) in this case
 280          * because the server can be trusted not to screw up.
 281          */
 282         fuse_file_put(ff, ff->fc->destroy_req != NULL, isdir);
 283 }
 284
 285 static int fuse_open(struct inode *inode, struct file *file)
 286 {
 287         return fuse_open_common(inode, file, false);
 288 }
 289
 290 static int fuse_release(struct inode *inode, struct file *file)
 291 {
 292         struct fuse_conn *fc = get_fuse_conn(inode);
 293
 294         /* see fuse_vma_close() for !writeback_cache case */
 295         if (fc->writeback_cache)
 296                 write_inode_now(inode, 1);
 297
 298         fuse_release_common(file, false);
 299
 300         /* return value is ignored by VFS */
 301         return 0;
 302 }
 303
 304 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
 305 {
 306         WARN_ON(refcount_read(&ff->count) > 1);
 307         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 308         /*
 309          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 310          * synchronous, we are fine with not doing igrab() here"
 311          */
 312         fuse_file_put(ff, true, false);
 313 }
 314 EXPORT_SYMBOL_GPL(fuse_sync_release);
 315
 316 /*
 317  * Scramble the ID space with XTEA, so that the value of the files_struct
 318  * pointer is not exposed to userspace.
 319  */
 320 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 321 {
 322         u32 *k = fc->scramble_key;
 323         u64 v = (unsigned long) id;
 324         u32 v0 = v;
 325         u32 v1 = v >> 32;
 326         u32 sum = 0;
 327         int i;
 328
 329         for (i = 0; i < 32; i++) {
 330                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 331                 sum += 0x9E3779B9;
 332                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 333         }
 334
 335         return (u64) v0 + ((u64) v1 << 32);
 336 }
 337
 338 static struct fuse_req *fuse_find_writeback(struct fuse_inode *fi,
 339                                             pgoff_t idx_from, pgoff_t idx_to)
 340 {
 341         struct fuse_req *req;
 342
 343         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 344                 pgoff_t curr_index;
 345
 346                 WARN_ON(get_fuse_inode(req->inode) != fi);
 347                 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
 348                 if (idx_from < curr_index + req->num_pages &&
 349                     curr_index <= idx_to) {
 350                         return req;
 351                 }
 352         }
 353         return NULL;
 354 }
 355
 356 /*
 357  * Check if any page in a range is under writeback
 358  *
 359  * This is currently done by walking the list of writepage requests
 360  * for the inode, which can be pretty inefficient.
 361  */
 362 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 363                                    pgoff_t idx_to)
 364 {
 365         struct fuse_inode *fi = get_fuse_inode(inode);
 366         bool found;
 367
 368         spin_lock(&fi->lock);
 369         found = fuse_find_writeback(fi, idx_from, idx_to);
 370         spin_unlock(&fi->lock);
 371
 372         return found;
 373 }
 374
 375 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 376 {
 377         return fuse_range_is_writeback(inode, index, index);
 378 }
 379
 380 /*
 381  * Wait for page writeback to be completed.
 382  *
 383  * Since fuse doesn't rely on the VM writeback tracking, this has to
 384  * use some other means.
 385  */
 386 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 387 {
 388         struct fuse_inode *fi = get_fuse_inode(inode);
 389
 390         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 391         return 0;
 392 }
 393
 394 /*
 395  * Wait for all pending writepages on the inode to finish.
 396  *
 397  * This is currently done by blocking further writes with FUSE_NOWRITE
 398  * and waiting for all sent writes to complete.
 399  *
 400  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 401  * could conflict with truncation.
 402  */
 403 static void fuse_sync_writes(struct inode *inode)
 404 {
 405         fuse_set_nowrite(inode);
 406         fuse_release_nowrite(inode);
 407 }
 408
 409 static int fuse_flush(struct file *file, fl_owner_t id)
 410 {
 411         struct inode *inode = file_inode(file);
 412         struct fuse_conn *fc = get_fuse_conn(inode);
 413         struct fuse_file *ff = file->private_data;
 414         struct fuse_req *req;
 415         struct fuse_flush_in inarg;
 416         int err;
 417
 418         if (is_bad_inode(inode))
 419                 return -EIO;
 420
 421         if (fc->no_flush)
 422                 return 0;
 423
 424         err = write_inode_now(inode, 1);
 425         if (err)
 426                 return err;
 427
 428         inode_lock(inode);
 429         fuse_sync_writes(inode);
 430         inode_unlock(inode);
 431
 432         err = filemap_check_errors(file->f_mapping);
 433         if (err)
 434                 return err;
 435
 436         req = fuse_get_req_nofail_nopages(fc, file);
 437         memset(&inarg, 0, sizeof(inarg));
 438         inarg.fh = ff->fh;
 439         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 440         req->in.h.opcode = FUSE_FLUSH;
 441         req->in.h.nodeid = get_node_id(inode);
 442         req->in.numargs = 1;
 443         req->in.args[0].size = sizeof(inarg);
 444         req->in.args[0].value = &inarg;
 445         __set_bit(FR_FORCE, &req->flags);
 446         fuse_request_send(fc, req);
 447         err = req->out.h.error;
 448         fuse_put_request(fc, req);
 449         if (err == -ENOSYS) {
 450                 fc->no_flush = 1;
 451                 err = 0;
 452         }
 453         return err;
 454 }
 455
 456 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 457                       int datasync, int opcode)
 458 {
 459         struct inode *inode = file->f_mapping->host;
 460         struct fuse_conn *fc = get_fuse_conn(inode);
 461         struct fuse_file *ff = file->private_data;
 462         FUSE_ARGS(args);
 463         struct fuse_fsync_in inarg;
 464
 465         memset(&inarg, 0, sizeof(inarg));
 466         inarg.fh = ff->fh;
 467         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 468         args.in.h.opcode = opcode;
 469         args.in.h.nodeid = get_node_id(inode);
 470         args.in.numargs = 1;
 471         args.in.args[0].size = sizeof(inarg);
 472         args.in.args[0].value = &inarg;
 473         return fuse_simple_request(fc, &args);
 474 }
 475
 476 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 477                       int datasync)
 478 {
 479         struct inode *inode = file->f_mapping->host;
 480         struct fuse_conn *fc = get_fuse_conn(inode);
 481         int err;
 482
 483         if (is_bad_inode(inode))
 484                 return -EIO;
 485
 486         inode_lock(inode);
 487
 488         /*
 489          * Start writeback against all dirty pages of the inode, then
 490          * wait for all outstanding writes, before sending the FSYNC
 491          * request.
 492          */
 493         err = file_write_and_wait_range(file, start, end);
 494         if (err)
 495                 goto out;
 496
 497         fuse_sync_writes(inode);
 498
 499         /*
 500          * Due to implementation of fuse writeback
 501          * file_write_and_wait_range() does not catch errors.
 502          * We have to do this directly after fuse_sync_writes()
 503          */
 504         err = file_check_and_advance_wb_err(file);
 505         if (err)
 506                 goto out;
 507
 508         err = sync_inode_metadata(inode, 1);
 509         if (err)
 510                 goto out;
 511
 512         if (fc->no_fsync)
 513                 goto out;
 514
 515         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 516         if (err == -ENOSYS) {
 517                 fc->no_fsync = 1;
 518                 err = 0;
 519         }
 520 out:
 521         inode_unlock(inode);
 522
 523         return err;
 524 }
 525
 526 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 527                     size_t count, int opcode)
 528 {
 529         struct fuse_read_in *inarg = &req->misc.read.in;
 530         struct fuse_file *ff = file->private_data;
 531
 532         inarg->fh = ff->fh;
 533         inarg->offset = pos;
 534         inarg->size = count;
 535         inarg->flags = file->f_flags;
 536         req->in.h.opcode = opcode;
 537         req->in.h.nodeid = ff->nodeid;
 538         req->in.numargs = 1;
 539         req->in.args[0].size = sizeof(struct fuse_read_in);
 540         req->in.args[0].value = inarg;
 541         req->out.argvar = 1;
 542         req->out.numargs = 1;
 543         req->out.args[0].size = count;
 544 }
 545
 546 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
 547 {
 548         unsigned i;
 549
 550         for (i = 0; i < req->num_pages; i++) {
 551                 struct page *page = req->pages[i];
 552                 if (should_dirty)
 553                         set_page_dirty_lock(page);
 554                 put_page(page);
 555         }
 556 }
 557
 558 static void fuse_io_release(struct kref *kref)
 559 {
 560         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 561 }
 562
 563 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 564 {
 565         if (io->err)
 566                 return io->err;
 567
 568         if (io->bytes >= 0 && io->write)
 569                 return -EIO;
 570
 571         return io->bytes < 0 ? io->size : io->bytes;
 572 }
 573
 574 /**
 575  * In case of short read, the caller sets 'pos' to the position of
 576  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 577  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 578  *
 579  * An example:
 580  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 581  * both submitted asynchronously. The first of them was ACKed by userspace as
 582  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 583  * second request was ACKed as short, e.g. only 1K was read, resulting in
 584  * pos == 33K.
 585  *
 586  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 587  * will be equal to the length of the longest contiguous fragment of
 588  * transferred data starting from the beginning of IO request.
 589  */
 590 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 591 {
 592         int left;
 593
 594         spin_lock(&io->lock);
 595         if (err)
 596                 io->err = io->err ? : err;
 597         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 598                 io->bytes = pos;
 599
 600         left = --io->reqs;
 601         if (!left && io->blocking)
 602                 complete(io->done);
 603         spin_unlock(&io->lock);
 604
 605         if (!left && !io->blocking) {
 606                 ssize_t res = fuse_get_res_by_io(io);
 607
 608                 if (res >= 0) {
 609                         struct inode *inode = file_inode(io->iocb->ki_filp);
 610                         struct fuse_conn *fc = get_fuse_conn(inode);
 611                         struct fuse_inode *fi = get_fuse_inode(inode);
 612
 613                         spin_lock(&fi->lock);
 614                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 615                         spin_unlock(&fi->lock);
 616                 }
 617
 618                 io->iocb->ki_complete(io->iocb, res, 0);
 619         }
 620
 621         kref_put(&io->refcnt, fuse_io_release);
 622 }
 623
 624 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 625 {
 626         struct fuse_io_priv *io = req->io;
 627         ssize_t pos = -1;
 628
 629         fuse_release_user_pages(req, io->should_dirty);
 630
 631         if (io->write) {
 632                 if (req->misc.write.in.size != req->misc.write.out.size)
 633                         pos = req->misc.write.in.offset - io->offset +
 634                                 req->misc.write.out.size;
 635         } else {
 636                 if (req->misc.read.in.size != req->out.args[0].size)
 637                         pos = req->misc.read.in.offset - io->offset +
 638                                 req->out.args[0].size;
 639         }
 640
 641         fuse_aio_complete(io, req->out.h.error, pos);
 642 }
 643
 644 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 645                 size_t num_bytes, struct fuse_io_priv *io)
 646 {
 647         spin_lock(&io->lock);
 648         kref_get(&io->refcnt);
 649         io->size += num_bytes;
 650         io->reqs++;
 651         spin_unlock(&io->lock);
 652
 653         req->io = io;
 654         req->end = fuse_aio_complete_req;
 655
 656         __fuse_get_request(req);
 657         fuse_request_send_background(fc, req);
 658
 659         return num_bytes;
 660 }
 661
 662 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 663                              loff_t pos, size_t count, fl_owner_t owner)
 664 {
 665         struct file *file = io->iocb->ki_filp;
 666         struct fuse_file *ff = file->private_data;
 667         struct fuse_conn *fc = ff->fc;
 668
 669         fuse_read_fill(req, file, pos, count, FUSE_READ);
 670         if (owner != NULL) {
 671                 struct fuse_read_in *inarg = &req->misc.read.in;
 672
 673                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 674                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 675         }
 676
 677         if (io->async)
 678                 return fuse_async_req_send(fc, req, count, io);
 679
 680         fuse_request_send(fc, req);
 681         return req->out.args[0].size;
 682 }
 683
 684 static void fuse_read_update_size(struct inode *inode, loff_t size,
 685                                   u64 attr_ver)
 686 {
 687         struct fuse_conn *fc = get_fuse_conn(inode);
 688         struct fuse_inode *fi = get_fuse_inode(inode);
 689
 690         spin_lock(&fi->lock);
 691         if (attr_ver == fi->attr_version && size < inode->i_size &&
 692             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 693                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 694                 i_size_write(inode, size);
 695         }
 696         spin_unlock(&fi->lock);
 697 }
 698
 699 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 700                             u64 attr_ver)
 701 {
 702         size_t num_read = req->out.args[0].size;
 703         struct fuse_conn *fc = get_fuse_conn(inode);
 704
 705         if (fc->writeback_cache) {
 706                 /*
 707                  * A hole in a file. Some data after the hole are in page cache,
 708                  * but have not reached the client fs yet. So, the hole is not
 709                  * present there.
 710                  */
 711                 int i;
 712                 int start_idx = num_read >> PAGE_SHIFT;
 713                 size_t off = num_read & (PAGE_SIZE - 1);
 714
 715                 for (i = start_idx; i < req->num_pages; i++) {
 716                         zero_user_segment(req->pages[i], off, PAGE_SIZE);
 717                         off = 0;
 718                 }
 719         } else {
 720                 loff_t pos = page_offset(req->pages[0]) + num_read;
 721                 fuse_read_update_size(inode, pos, attr_ver);
 722         }
 723 }
 724
 725 static int fuse_do_readpage(struct file *file, struct page *page)
 726 {
 727         struct kiocb iocb;
 728         struct fuse_io_priv io;
 729         struct inode *inode = page->mapping->host;
 730         struct fuse_conn *fc = get_fuse_conn(inode);
 731         struct fuse_req *req;
 732         size_t num_read;
 733         loff_t pos = page_offset(page);
 734         size_t count = PAGE_SIZE;
 735         u64 attr_ver;
 736         int err;
 737
 738         /*
 739          * Page writeback can extend beyond the lifetime of the
 740          * page-cache page, so make sure we read a properly synced
 741          * page.
 742          */
 743         fuse_wait_on_page_writeback(inode, page->index);
 744
 745         req = fuse_get_req(fc, 1);
 746         if (IS_ERR(req))
 747                 return PTR_ERR(req);
 748
 749         attr_ver = fuse_get_attr_version(fc);
 750
 751         req->out.page_zeroing = 1;
 752         req->out.argpages = 1;
 753         req->num_pages = 1;
 754         req->pages[0] = page;
 755         req->page_descs[0].length = count;
 756         init_sync_kiocb(&iocb, file);
 757         io = (struct fuse_io_priv) FUSE_IO_PRIV_SYNC(&iocb);
 758         num_read = fuse_send_read(req, &io, pos, count, NULL);
 759         err = req->out.h.error;
 760
 761         if (!err) {
 762                 /*
 763                  * Short read means EOF.  If file size is larger, truncate it
 764                  */
 765                 if (num_read < count)
 766                         fuse_short_read(req, inode, attr_ver);
 767
 768                 SetPageUptodate(page);
 769         }
 770
 771         fuse_put_request(fc, req);
 772
 773         return err;
 774 }
 775
 776 static int fuse_readpage(struct file *file, struct page *page)
 777 {
 778         struct inode *inode = page->mapping->host;
 779         int err;
 780
 781         err = -EIO;
 782         if (is_bad_inode(inode))
 783                 goto out;
 784
 785         err = fuse_do_readpage(file, page);
 786         fuse_invalidate_atime(inode);
 787  out:
 788         unlock_page(page);
 789         return err;
 790 }
 791
 792 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 793 {
 794         int i;
 795         size_t count = req->misc.read.in.size;
 796         size_t num_read = req->out.args[0].size;
 797         struct address_space *mapping = NULL;
 798
 799         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 800                 mapping = req->pages[i]->mapping;
 801
 802         if (mapping) {
 803                 struct inode *inode = mapping->host;
 804
 805                 /*
 806                  * Short read means EOF. If file size is larger, truncate it
 807                  */
 808                 if (!req->out.h.error && num_read < count)
 809                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 810
 811                 fuse_invalidate_atime(inode);
 812         }
 813
 814         for (i = 0; i < req->num_pages; i++) {
 815                 struct page *page = req->pages[i];
 816                 if (!req->out.h.error)
 817                         SetPageUptodate(page);
 818                 else
 819                         SetPageError(page);
 820                 unlock_page(page);
 821                 put_page(page);
 822         }
 823         if (req->ff)
 824                 fuse_file_put(req->ff, false, false);
 825 }
 826
 827 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 828 {
 829         struct fuse_file *ff = file->private_data;
 830         struct fuse_conn *fc = ff->fc;
 831         loff_t pos = page_offset(req->pages[0]);
 832         size_t count = req->num_pages << PAGE_SHIFT;
 833
 834         req->out.argpages = 1;
 835         req->out.page_zeroing = 1;
 836         req->out.page_replace = 1;
 837         fuse_read_fill(req, file, pos, count, FUSE_READ);
 838         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 839         if (fc->async_read) {
 840                 req->ff = fuse_file_get(ff);
 841                 req->end = fuse_readpages_end;
 842                 fuse_request_send_background(fc, req);
 843         } else {
 844                 fuse_request_send(fc, req);
 845                 fuse_readpages_end(fc, req);
 846                 fuse_put_request(fc, req);
 847         }
 848 }
 849
 850 struct fuse_fill_data {
 851         struct fuse_req *req;
 852         struct file *file;
 853         struct inode *inode;
 854         unsigned nr_pages;
 855 };
 856
 857 static int fuse_readpages_fill(void *_data, struct page *page)
 858 {
 859         struct fuse_fill_data *data = _data;
 860         struct fuse_req *req = data->req;
 861         struct inode *inode = data->inode;
 862         struct fuse_conn *fc = get_fuse_conn(inode);
 863
 864         fuse_wait_on_page_writeback(inode, page->index);
 865
 866         if (req->num_pages &&
 867             (req->num_pages == fc->max_pages ||
 868              (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
 869              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 870                 unsigned int nr_alloc = min_t(unsigned int, data->nr_pages,
 871                                               fc->max_pages);
 872                 fuse_send_readpages(req, data->file);
 873                 if (fc->async_read)
 874                         req = fuse_get_req_for_background(fc, nr_alloc);
 875                 else
 876                         req = fuse_get_req(fc, nr_alloc);
 877
 878                 data->req = req;
 879                 if (IS_ERR(req)) {
 880                         unlock_page(page);
 881                         return PTR_ERR(req);
 882                 }
 883         }
 884
 885         if (WARN_ON(req->num_pages >= req->max_pages)) {
 886                 unlock_page(page);
 887                 fuse_put_request(fc, req);
 888                 return -EIO;
 889         }
 890
 891         get_page(page);
 892         req->pages[req->num_pages] = page;
 893         req->page_descs[req->num_pages].length = PAGE_SIZE;
 894         req->num_pages++;
 895         data->nr_pages--;
 896         return 0;
 897 }
 898
 899 static int fuse_readpages(struct file *file, struct address_space *mapping,
 900                           struct list_head *pages, unsigned nr_pages)
 901 {
 902         struct inode *inode = mapping->host;
 903         struct fuse_conn *fc = get_fuse_conn(inode);
 904         struct fuse_fill_data data;
 905         int err;
 906         unsigned int nr_alloc = min_t(unsigned int, nr_pages, fc->max_pages);
 907
 908         err = -EIO;
 909         if (is_bad_inode(inode))
 910                 goto out;
 911
 912         data.file = file;
 913         data.inode = inode;
 914         if (fc->async_read)
 915                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 916         else
 917                 data.req = fuse_get_req(fc, nr_alloc);
 918         data.nr_pages = nr_pages;
 919         err = PTR_ERR(data.req);
 920         if (IS_ERR(data.req))
 921                 goto out;
 922
 923         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 924         if (!err) {
 925                 if (data.req->num_pages)
 926                         fuse_send_readpages(data.req, file);
 927                 else
 928                         fuse_put_request(fc, data.req);
 929         }
 930 out:
 931         return err;
 932 }
 933
 934 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
 935 {
 936         struct inode *inode = iocb->ki_filp->f_mapping->host;
 937         struct fuse_conn *fc = get_fuse_conn(inode);
 938
 939         /*
 940          * In auto invalidate mode, always update attributes on read.
 941          * Otherwise, only update if we attempt to read past EOF (to ensure
 942          * i_size is up to date).
 943          */
 944         if (fc->auto_inval_data ||
 945             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 946                 int err;
 947                 err = fuse_update_attributes(inode, iocb->ki_filp);
 948                 if (err)
 949                         return err;
 950         }
 951
 952         return generic_file_read_iter(iocb, to);
 953 }
 954
 955 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 956                             loff_t pos, size_t count)
 957 {
 958         struct fuse_write_in *inarg = &req->misc.write.in;
 959         struct fuse_write_out *outarg = &req->misc.write.out;
 960
 961         inarg->fh = ff->fh;
 962         inarg->offset = pos;
 963         inarg->size = count;
 964         req->in.h.opcode = FUSE_WRITE;
 965         req->in.h.nodeid = ff->nodeid;
 966         req->in.numargs = 2;
 967         if (ff->fc->minor < 9)
 968                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 969         else
 970                 req->in.args[0].size = sizeof(struct fuse_write_in);
 971         req->in.args[0].value = inarg;
 972         req->in.args[1].size = count;
 973         req->out.numargs = 1;
 974         req->out.args[0].size = sizeof(struct fuse_write_out);
 975         req->out.args[0].value = outarg;
 976 }
 977
 978 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 979                               loff_t pos, size_t count, fl_owner_t owner)
 980 {
 981         struct kiocb *iocb = io->iocb;
 982         struct file *file = iocb->ki_filp;
 983         struct fuse_file *ff = file->private_data;
 984         struct fuse_conn *fc = ff->fc;
 985         struct fuse_write_in *inarg = &req->misc.write.in;
 986
 987         fuse_write_fill(req, ff, pos, count);
 988         inarg->flags = file->f_flags;
 989         if (iocb->ki_flags & IOCB_DSYNC)
 990                 inarg->flags |= O_DSYNC;
 991         if (iocb->ki_flags & IOCB_SYNC)
 992                 inarg->flags |= O_SYNC;
 993         if (owner != NULL) {
 994                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 995                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 996         }
 997
 998         if (io->async)
 999                 return fuse_async_req_send(fc, req, count, io);
1000
1001         fuse_request_send(fc, req);
1002         return req->misc.write.out.size;
1003 }
1004
1005 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1006 {
1007         struct fuse_conn *fc = get_fuse_conn(inode);
1008         struct fuse_inode *fi = get_fuse_inode(inode);
1009         bool ret = false;
1010
1011         spin_lock(&fi->lock);
1012         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1013         if (pos > inode->i_size) {
1014                 i_size_write(inode, pos);
1015                 ret = true;
1016         }
1017         spin_unlock(&fi->lock);
1018
1019         return ret;
1020 }
1021
1022 static size_t fuse_send_write_pages(struct fuse_req *req, struct kiocb *iocb,
1023                                     struct inode *inode, loff_t pos,
1024                                     size_t count)
1025 {
1026         size_t res;
1027         unsigned offset;
1028         unsigned i;
1029         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1030
1031         for (i = 0; i < req->num_pages; i++)
1032                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1033
1034         res = fuse_send_write(req, &io, pos, count, NULL);
1035
1036         offset = req->page_descs[0].offset;
1037         count = res;
1038         for (i = 0; i < req->num_pages; i++) {
1039                 struct page *page = req->pages[i];
1040
1041                 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1042                         SetPageUptodate(page);
1043
1044                 if (count > PAGE_SIZE - offset)
1045                         count -= PAGE_SIZE - offset;
1046                 else
1047                         count = 0;
1048                 offset = 0;
1049
1050                 unlock_page(page);
1051                 put_page(page);
1052         }
1053
1054         return res;
1055 }
1056
1057 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1058                                struct address_space *mapping,
1059                                struct iov_iter *ii, loff_t pos)
1060 {
1061         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1062         unsigned offset = pos & (PAGE_SIZE - 1);
1063         size_t count = 0;
1064         int err;
1065
1066         req->in.argpages = 1;
1067         req->page_descs[0].offset = offset;
1068
1069         do {
1070                 size_t tmp;
1071                 struct page *page;
1072                 pgoff_t index = pos >> PAGE_SHIFT;
1073                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1074                                      iov_iter_count(ii));
1075
1076                 bytes = min_t(size_t, bytes, fc->max_write - count);
1077
1078  again:
1079                 err = -EFAULT;
1080                 if (iov_iter_fault_in_readable(ii, bytes))
1081                         break;
1082
1083                 err = -ENOMEM;
1084                 page = grab_cache_page_write_begin(mapping, index, 0);
1085                 if (!page)
1086                         break;
1087
1088                 if (mapping_writably_mapped(mapping))
1089                         flush_dcache_page(page);
1090
1091                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1092                 flush_dcache_page(page);
1093
1094                 iov_iter_advance(ii, tmp);
1095                 if (!tmp) {
1096                         unlock_page(page);
1097                         put_page(page);
1098                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1099                         goto again;
1100                 }
1101
1102                 err = 0;
1103                 req->pages[req->num_pages] = page;
1104                 req->page_descs[req->num_pages].length = tmp;
1105                 req->num_pages++;
1106
1107                 count += tmp;
1108                 pos += tmp;
1109                 offset += tmp;
1110                 if (offset == PAGE_SIZE)
1111                         offset = 0;
1112
1113                 if (!fc->big_writes)
1114                         break;
1115         } while (iov_iter_count(ii) && count < fc->max_write &&
1116                  req->num_pages < req->max_pages && offset == 0);
1117
1118         return count > 0 ? count : err;
1119 }
1120
1121 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1122                                      unsigned int max_pages)
1123 {
1124         return min_t(unsigned int,
1125                      ((pos + len - 1) >> PAGE_SHIFT) -
1126                      (pos >> PAGE_SHIFT) + 1,
1127                      max_pages);
1128 }
1129
1130 static ssize_t fuse_perform_write(struct kiocb *iocb,
1131                                   struct address_space *mapping,
1132                                   struct iov_iter *ii, loff_t pos)
1133 {
1134         struct inode *inode = mapping->host;
1135         struct fuse_conn *fc = get_fuse_conn(inode);
1136         struct fuse_inode *fi = get_fuse_inode(inode);
1137         int err = 0;
1138         ssize_t res = 0;
1139
1140         if (inode->i_size < pos + iov_iter_count(ii))
1141                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1142
1143         do {
1144                 struct fuse_req *req;
1145                 ssize_t count;
1146                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1147                                                       fc->max_pages);
1148
1149                 req = fuse_get_req(fc, nr_pages);
1150                 if (IS_ERR(req)) {
1151                         err = PTR_ERR(req);
1152                         break;
1153                 }
1154
1155                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1156                 if (count <= 0) {
1157                         err = count;
1158                 } else {
1159                         size_t num_written;
1160
1161                         num_written = fuse_send_write_pages(req, iocb, inode,
1162                                                             pos, count);
1163                         err = req->out.h.error;
1164                         if (!err) {
1165                                 res += num_written;
1166                                 pos += num_written;
1167
1168                                 /* break out of the loop on short write */
1169                                 if (num_written != count)
1170                                         err = -EIO;
1171                         }
1172                 }
1173                 fuse_put_request(fc, req);
1174         } while (!err && iov_iter_count(ii));
1175
1176         if (res > 0)
1177                 fuse_write_update_size(inode, pos);
1178
1179         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1180         fuse_invalidate_attr(inode);
1181
1182         return res > 0 ? res : err;
1183 }
1184
1185 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1186 {
1187         struct file *file = iocb->ki_filp;
1188         struct address_space *mapping = file->f_mapping;
1189         ssize_t written = 0;
1190         ssize_t written_buffered = 0;
1191         struct inode *inode = mapping->host;
1192         ssize_t err;
1193         loff_t endbyte = 0;
1194
1195         if (get_fuse_conn(inode)->writeback_cache) {
1196                 /* Update size (EOF optimization) and mode (SUID clearing) */
1197                 err = fuse_update_attributes(mapping->host, file);
1198                 if (err)
1199                         return err;
1200
1201                 return generic_file_write_iter(iocb, from);
1202         }
1203
1204         inode_lock(inode);
1205
1206         /* We can write back this queue in page reclaim */
1207         current->backing_dev_info = inode_to_bdi(inode);
1208
1209         err = generic_write_checks(iocb, from);
1210         if (err <= 0)
1211                 goto out;
1212
1213         err = file_remove_privs(file);
1214         if (err)
1215                 goto out;
1216
1217         err = file_update_time(file);
1218         if (err)
1219                 goto out;
1220
1221         if (iocb->ki_flags & IOCB_DIRECT) {
1222                 loff_t pos = iocb->ki_pos;
1223                 written = generic_file_direct_write(iocb, from);
1224                 if (written < 0 || !iov_iter_count(from))
1225                         goto out;
1226
1227                 pos += written;
1228
1229                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1230                 if (written_buffered < 0) {
1231                         err = written_buffered;
1232                         goto out;
1233                 }
1234                 endbyte = pos + written_buffered - 1;
1235
1236                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1237                                                    endbyte);
1238                 if (err)
1239                         goto out;
1240
1241                 invalidate_mapping_pages(file->f_mapping,
1242                                          pos >> PAGE_SHIFT,
1243                                          endbyte >> PAGE_SHIFT);
1244
1245                 written += written_buffered;
1246                 iocb->ki_pos = pos + written_buffered;
1247         } else {
1248                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1249                 if (written >= 0)
1250                         iocb->ki_pos += written;
1251         }
1252 out:
1253         current->backing_dev_info = NULL;
1254         inode_unlock(inode);
1255         if (written > 0)
1256                 written = generic_write_sync(iocb, written);
1257
1258         return written ? written : err;
1259 }
1260
1261 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1262                 unsigned index, unsigned nr_pages)
1263 {
1264         int i;
1265
1266         for (i = index; i < index + nr_pages; i++)
1267                 req->page_descs[i].length = PAGE_SIZE -
1268                         req->page_descs[i].offset;
1269 }
1270
1271 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1272 {
1273         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1274 }
1275
1276 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1277                                         size_t max_size)
1278 {
1279         return min(iov_iter_single_seg_count(ii), max_size);
1280 }
1281
1282 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1283                                size_t *nbytesp, int write)
1284 {
1285         size_t nbytes = 0;  /* # bytes already packed in req */
1286         ssize_t ret = 0;
1287
1288         /* Special case for kernel I/O: can copy directly into the buffer */
1289         if (iov_iter_is_kvec(ii)) {
1290                 unsigned long user_addr = fuse_get_user_addr(ii);
1291                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1292
1293                 if (write)
1294                         req->in.args[1].value = (void *) user_addr;
1295                 else
1296                         req->out.args[0].value = (void *) user_addr;
1297
1298                 iov_iter_advance(ii, frag_size);
1299                 *nbytesp = frag_size;
1300                 return 0;
1301         }
1302
1303         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1304                 unsigned npages;
1305                 size_t start;
1306                 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1307                                         *nbytesp - nbytes,
1308                                         req->max_pages - req->num_pages,
1309                                         &start);
1310                 if (ret < 0)
1311                         break;
1312
1313                 iov_iter_advance(ii, ret);
1314                 nbytes += ret;
1315
1316                 ret += start;
1317                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1318
1319                 req->page_descs[req->num_pages].offset = start;
1320                 fuse_page_descs_length_init(req, req->num_pages, npages);
1321
1322                 req->num_pages += npages;
1323                 req->page_descs[req->num_pages - 1].length -=
1324                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1325         }
1326
1327         if (write)
1328                 req->in.argpages = 1;
1329         else
1330                 req->out.argpages = 1;
1331
1332         *nbytesp = nbytes;
1333
1334         return ret < 0 ? ret : 0;
1335 }
1336
1337 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1338                        loff_t *ppos, int flags)
1339 {
1340         int write = flags & FUSE_DIO_WRITE;
1341         int cuse = flags & FUSE_DIO_CUSE;
1342         struct file *file = io->iocb->ki_filp;
1343         struct inode *inode = file->f_mapping->host;
1344         struct fuse_file *ff = file->private_data;
1345         struct fuse_conn *fc = ff->fc;
1346         size_t nmax = write ? fc->max_write : fc->max_read;
1347         loff_t pos = *ppos;
1348         size_t count = iov_iter_count(iter);
1349         pgoff_t idx_from = pos >> PAGE_SHIFT;
1350         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1351         ssize_t res = 0;
1352         struct fuse_req *req;
1353         int err = 0;
1354
1355         if (io->async)
1356                 req = fuse_get_req_for_background(fc, iov_iter_npages(iter,
1357                                                                 fc->max_pages));
1358         else
1359                 req = fuse_get_req(fc, iov_iter_npages(iter, fc->max_pages));
1360         if (IS_ERR(req))
1361                 return PTR_ERR(req);
1362
1363         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1364                 if (!write)
1365                         inode_lock(inode);
1366                 fuse_sync_writes(inode);
1367                 if (!write)
1368                         inode_unlock(inode);
1369         }
1370
1371         io->should_dirty = !write && iter_is_iovec(iter);
1372         while (count) {
1373                 size_t nres;
1374                 fl_owner_t owner = current->files;
1375                 size_t nbytes = min(count, nmax);
1376                 err = fuse_get_user_pages(req, iter, &nbytes, write);
1377                 if (err && !nbytes)
1378                         break;
1379
1380                 if (write) {
1381                         if (!capable(CAP_FSETID)) {
1382                                 struct fuse_write_in *inarg;
1383
1384                                 inarg = &req->misc.write.in;
1385                                 inarg->write_flags |= FUSE_WRITE_KILL_PRIV;
1386                         }
1387                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1388                 } else {
1389                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1390                 }
1391
1392                 if (!io->async)
1393                         fuse_release_user_pages(req, io->should_dirty);
1394                 if (req->out.h.error) {
1395                         err = req->out.h.error;
1396                         break;
1397                 } else if (nres > nbytes) {
1398                         res = 0;
1399                         err = -EIO;
1400                         break;
1401                 }
1402                 count -= nres;
1403                 res += nres;
1404                 pos += nres;
1405                 if (nres != nbytes)
1406                         break;
1407                 if (count) {
1408                         fuse_put_request(fc, req);
1409                         if (io->async)
1410                                 req = fuse_get_req_for_background(fc,
1411                                         iov_iter_npages(iter, fc->max_pages));
1412                         else
1413                                 req = fuse_get_req(fc, iov_iter_npages(iter,
1414                                                                 fc->max_pages));
1415                         if (IS_ERR(req))
1416                                 break;
1417                 }
1418         }
1419         if (!IS_ERR(req))
1420                 fuse_put_request(fc, req);
1421         if (res > 0)
1422                 *ppos = pos;
1423
1424         return res > 0 ? res : err;
1425 }
1426 EXPORT_SYMBOL_GPL(fuse_direct_io);
1427
1428 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1429                                   struct iov_iter *iter,
1430                                   loff_t *ppos)
1431 {
1432         ssize_t res;
1433         struct inode *inode = file_inode(io->iocb->ki_filp);
1434
1435         res = fuse_direct_io(io, iter, ppos, 0);
1436
1437         fuse_invalidate_atime(inode);
1438
1439         return res;
1440 }
1441
1442 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1443
1444 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1445 {
1446         ssize_t res;
1447
1448         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1449                 res = fuse_direct_IO(iocb, to);
1450         } else {
1451                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1452
1453                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1454         }
1455
1456         return res;
1457 }
1458
1459 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1460 {
1461         struct inode *inode = file_inode(iocb->ki_filp);
1462         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1463         ssize_t res;
1464
1465         /* Don't allow parallel writes to the same file */
1466         inode_lock(inode);
1467         res = generic_write_checks(iocb, from);
1468         if (res > 0) {
1469                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1470                         res = fuse_direct_IO(iocb, from);
1471                 } else {
1472                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1473                                              FUSE_DIO_WRITE);
1474                 }
1475         }
1476         fuse_invalidate_attr(inode);
1477         if (res > 0)
1478                 fuse_write_update_size(inode, iocb->ki_pos);
1479         inode_unlock(inode);
1480
1481         return res;
1482 }
1483
1484 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1485 {
1486         struct file *file = iocb->ki_filp;
1487         struct fuse_file *ff = file->private_data;
1488
1489         if (is_bad_inode(file_inode(file)))
1490                 return -EIO;
1491
1492         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1493                 return fuse_cache_read_iter(iocb, to);
1494         else
1495                 return fuse_direct_read_iter(iocb, to);
1496 }
1497
1498 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1499 {
1500         struct file *file = iocb->ki_filp;
1501         struct fuse_file *ff = file->private_data;
1502
1503         if (is_bad_inode(file_inode(file)))
1504                 return -EIO;
1505
1506         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1507                 return fuse_cache_write_iter(iocb, from);
1508         else
1509                 return fuse_direct_write_iter(iocb, from);
1510 }
1511
1512 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1513 {
1514         int i;
1515
1516         for (i = 0; i < req->num_pages; i++)
1517                 __free_page(req->pages[i]);
1518
1519         if (req->ff)
1520                 fuse_file_put(req->ff, false, false);
1521 }
1522
1523 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1524 {
1525         struct inode *inode = req->inode;
1526         struct fuse_inode *fi = get_fuse_inode(inode);
1527         struct backing_dev_info *bdi = inode_to_bdi(inode);
1528         int i;
1529
1530         list_del(&req->writepages_entry);
1531         for (i = 0; i < req->num_pages; i++) {
1532                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1533                 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1534                 wb_writeout_inc(&bdi->wb);
1535         }
1536         wake_up(&fi->page_waitq);
1537 }
1538
1539 /* Called under fi->lock, may release and reacquire it */
1540 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1541                                 loff_t size)
1542 __releases(fi->lock)
1543 __acquires(fi->lock)
1544 {
1545         struct fuse_req *aux, *next;
1546         struct fuse_inode *fi = get_fuse_inode(req->inode);
1547         struct fuse_write_in *inarg = &req->misc.write.in;
1548         __u64 data_size = req->num_pages * PAGE_SIZE;
1549         bool queued;
1550
1551         if (inarg->offset + data_size <= size) {
1552                 inarg->size = data_size;
1553         } else if (inarg->offset < size) {
1554                 inarg->size = size - inarg->offset;
1555         } else {
1556                 /* Got truncated off completely */
1557                 goto out_free;
1558         }
1559
1560         req->in.args[1].size = inarg->size;
1561         queued = fuse_request_queue_background(fc, req);
1562         /* Fails on broken connection only */
1563         if (unlikely(!queued))
1564                 goto out_free;
1565
1566         fi->writectr++;
1567         return;
1568
1569  out_free:
1570         fuse_writepage_finish(fc, req);
1571         spin_unlock(&fi->lock);
1572
1573         /* After fuse_writepage_finish() aux request list is private */
1574         for (aux = req->misc.write.next; aux; aux = next) {
1575                 next = aux->misc.write.next;
1576                 aux->misc.write.next = NULL;
1577                 fuse_writepage_free(fc, aux);
1578                 fuse_put_request(fc, aux);
1579         }
1580
1581         fuse_writepage_free(fc, req);
1582         fuse_put_request(fc, req);
1583         spin_lock(&fi->lock);
1584 }
1585
1586 /*
1587  * If fi->writectr is positive (no truncate or fsync going on) send
1588  * all queued writepage requests.
1589  *
1590  * Called with fi->lock
1591  */
1592 void fuse_flush_writepages(struct inode *inode)
1593 __releases(fi->lock)
1594 __acquires(fi->lock)
1595 {
1596         struct fuse_conn *fc = get_fuse_conn(inode);
1597         struct fuse_inode *fi = get_fuse_inode(inode);
1598         loff_t crop = i_size_read(inode);
1599         struct fuse_req *req;
1600
1601         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1602                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1603                 list_del_init(&req->list);
1604                 fuse_send_writepage(fc, req, crop);
1605         }
1606 }
1607
1608 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1609 {
1610         struct inode *inode = req->inode;
1611         struct fuse_inode *fi = get_fuse_inode(inode);
1612
1613         mapping_set_error(inode->i_mapping, req->out.h.error);
1614         spin_lock(&fi->lock);
1615         while (req->misc.write.next) {
1616                 struct fuse_conn *fc = get_fuse_conn(inode);
1617                 struct fuse_write_in *inarg = &req->misc.write.in;
1618                 struct fuse_req *next = req->misc.write.next;
1619                 req->misc.write.next = next->misc.write.next;
1620                 next->misc.write.next = NULL;
1621                 next->ff = fuse_file_get(req->ff);
1622                 list_add(&next->writepages_entry, &fi->writepages);
1623
1624                 /*
1625                  * Skip fuse_flush_writepages() to make it easy to crop requests
1626                  * based on primary request size.
1627                  *
1628                  * 1st case (trivial): there are no concurrent activities using
1629                  * fuse_set/release_nowrite.  Then we're on safe side because
1630                  * fuse_flush_writepages() would call fuse_send_writepage()
1631                  * anyway.
1632                  *
1633                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1634                  * now for completion of all in-flight requests.  This happens
1635                  * rarely and no more than once per page, so this should be
1636                  * okay.
1637                  *
1638                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1639                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1640                  * that fuse_set_nowrite returned implies that all in-flight
1641                  * requests were completed along with all of their secondary
1642                  * requests.  Further primary requests are blocked by negative
1643                  * writectr.  Hence there cannot be any in-flight requests and
1644                  * no invocations of fuse_writepage_end() while we're in
1645                  * fuse_set_nowrite..fuse_release_nowrite section.
1646                  */
1647                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1648         }
1649         fi->writectr--;
1650         fuse_writepage_finish(fc, req);
1651         spin_unlock(&fi->lock);
1652         fuse_writepage_free(fc, req);
1653 }
1654
1655 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1656                                                struct fuse_inode *fi)
1657 {
1658         struct fuse_file *ff = NULL;
1659
1660         spin_lock(&fi->lock);
1661         if (!list_empty(&fi->write_files)) {
1662                 ff = list_entry(fi->write_files.next, struct fuse_file,
1663                                 write_entry);
1664                 fuse_file_get(ff);
1665         }
1666         spin_unlock(&fi->lock);
1667
1668         return ff;
1669 }
1670
1671 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1672                                              struct fuse_inode *fi)
1673 {
1674         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1675         WARN_ON(!ff);
1676         return ff;
1677 }
1678
1679 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1680 {
1681         struct fuse_conn *fc = get_fuse_conn(inode);
1682         struct fuse_inode *fi = get_fuse_inode(inode);
1683         struct fuse_file *ff;
1684         int err;
1685
1686         ff = __fuse_write_file_get(fc, fi);
1687         err = fuse_flush_times(inode, ff);
1688         if (ff)
1689                 fuse_file_put(ff, false, false);
1690
1691         return err;
1692 }
1693
1694 static int fuse_writepage_locked(struct page *page)
1695 {
1696         struct address_space *mapping = page->mapping;
1697         struct inode *inode = mapping->host;
1698         struct fuse_conn *fc = get_fuse_conn(inode);
1699         struct fuse_inode *fi = get_fuse_inode(inode);
1700         struct fuse_req *req;
1701         struct page *tmp_page;
1702         int error = -ENOMEM;
1703
1704         set_page_writeback(page);
1705
1706         req = fuse_request_alloc_nofs(1);
1707         if (!req)
1708                 goto err;
1709
1710         /* writeback always goes to bg_queue */
1711         __set_bit(FR_BACKGROUND, &req->flags);
1712         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1713         if (!tmp_page)
1714                 goto err_free;
1715
1716         error = -EIO;
1717         req->ff = fuse_write_file_get(fc, fi);
1718         if (!req->ff)
1719                 goto err_nofile;
1720
1721         fuse_write_fill(req, req->ff, page_offset(page), 0);
1722
1723         copy_highpage(tmp_page, page);
1724         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1725         req->misc.write.next = NULL;
1726         req->in.argpages = 1;
1727         req->num_pages = 1;
1728         req->pages[0] = tmp_page;
1729         req->page_descs[0].offset = 0;
1730         req->page_descs[0].length = PAGE_SIZE;
1731         req->end = fuse_writepage_end;
1732         req->inode = inode;
1733
1734         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1735         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1736
1737         spin_lock(&fi->lock);
1738         list_add(&req->writepages_entry, &fi->writepages);
1739         list_add_tail(&req->list, &fi->queued_writes);
1740         fuse_flush_writepages(inode);
1741         spin_unlock(&fi->lock);
1742
1743         end_page_writeback(page);
1744
1745         return 0;
1746
1747 err_nofile:
1748         __free_page(tmp_page);
1749 err_free:
1750         fuse_request_free(req);
1751 err:
1752         mapping_set_error(page->mapping, error);
1753         end_page_writeback(page);
1754         return error;
1755 }
1756
1757 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1758 {
1759         int err;
1760
1761         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1762                 /*
1763                  * ->writepages() should be called for sync() and friends.  We
1764                  * should only get here on direct reclaim and then we are
1765                  * allowed to skip a page which is already in flight
1766                  */
1767                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1768
1769                 redirty_page_for_writepage(wbc, page);
1770                 return 0;
1771         }
1772
1773         err = fuse_writepage_locked(page);
1774         unlock_page(page);
1775
1776         return err;
1777 }
1778
1779 struct fuse_fill_wb_data {
1780         struct fuse_req *req;
1781         struct fuse_file *ff;
1782         struct inode *inode;
1783         struct page **orig_pages;
1784 };
1785
1786 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1787 {
1788         struct fuse_req *req = data->req;
1789         struct inode *inode = data->inode;
1790         struct fuse_inode *fi = get_fuse_inode(inode);
1791         int num_pages = req->num_pages;
1792         int i;
1793
1794         req->ff = fuse_file_get(data->ff);
1795         spin_lock(&fi->lock);
1796         list_add_tail(&req->list, &fi->queued_writes);
1797         fuse_flush_writepages(inode);
1798         spin_unlock(&fi->lock);
1799
1800         for (i = 0; i < num_pages; i++)
1801                 end_page_writeback(data->orig_pages[i]);
1802 }
1803
1804 /*
1805  * First recheck under fi->lock if the offending offset is still under
1806  * writeback.  If yes, then iterate auxiliary write requests, to see if there's
1807  * one already added for a page at this offset.  If there's none, then insert
1808  * this new request onto the auxiliary list, otherwise reuse the existing one by
1809  * copying the new page contents over to the old temporary page.
1810  */
1811 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1812                                      struct page *page)
1813 {
1814         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1815         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1816         struct fuse_req *tmp;
1817         struct fuse_req *old_req;
1818
1819         WARN_ON(new_req->num_pages != 0);
1820
1821         spin_lock(&fi->lock);
1822         list_del(&new_req->writepages_entry);
1823         old_req = fuse_find_writeback(fi, page->index, page->index);
1824         if (!old_req) {
1825                 list_add(&new_req->writepages_entry, &fi->writepages);
1826                 spin_unlock(&fi->lock);
1827                 return false;
1828         }
1829
1830         new_req->num_pages = 1;
1831         for (tmp = old_req->misc.write.next; tmp; tmp = tmp->misc.write.next) {
1832                 pgoff_t curr_index;
1833
1834                 WARN_ON(tmp->inode != new_req->inode);
1835                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1836                 if (curr_index == page->index) {
1837                         WARN_ON(tmp->num_pages != 1);
1838                         WARN_ON(!test_bit(FR_PENDING, &tmp->flags));
1839                         swap(tmp->pages[0], new_req->pages[0]);
1840                         break;
1841                 }
1842         }
1843
1844         if (!tmp) {
1845                 new_req->misc.write.next = old_req->misc.write.next;
1846                 old_req->misc.write.next = new_req;
1847         }
1848
1849         spin_unlock(&fi->lock);
1850
1851         if (tmp) {
1852                 struct backing_dev_info *bdi = inode_to_bdi(new_req->inode);
1853
1854                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1855                 dec_node_page_state(new_req->pages[0], NR_WRITEBACK_TEMP);
1856                 wb_writeout_inc(&bdi->wb);
1857                 fuse_writepage_free(fc, new_req);
1858                 fuse_request_free(new_req);
1859         }
1860
1861         return true;
1862 }
1863
1864 static int fuse_writepages_fill(struct page *page,
1865                 struct writeback_control *wbc, void *_data)
1866 {
1867         struct fuse_fill_wb_data *data = _data;
1868         struct fuse_req *req = data->req;
1869         struct inode *inode = data->inode;
1870         struct fuse_inode *fi = get_fuse_inode(inode);
1871         struct fuse_conn *fc = get_fuse_conn(inode);
1872         struct page *tmp_page;
1873         bool is_writeback;
1874         int err;
1875
1876         if (!data->ff) {
1877                 err = -EIO;
1878                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1879                 if (!data->ff)
1880                         goto out_unlock;
1881         }
1882
1883         /*
1884          * Being under writeback is unlikely but possible.  For example direct
1885          * read to an mmaped fuse file will set the page dirty twice; once when
1886          * the pages are faulted with get_user_pages(), and then after the read
1887          * completed.
1888          */
1889         is_writeback = fuse_page_is_writeback(inode, page->index);
1890
1891         if (req && req->num_pages &&
1892             (is_writeback || req->num_pages == fc->max_pages ||
1893              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1894              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1895                 fuse_writepages_send(data);
1896                 data->req = NULL;
1897         } else if (req && req->num_pages == req->max_pages) {
1898                 if (!fuse_req_realloc_pages(fc, req, GFP_NOFS)) {
1899                         fuse_writepages_send(data);
1900                         req = data->req = NULL;
1901                 }
1902         }
1903
1904         err = -ENOMEM;
1905         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1906         if (!tmp_page)
1907                 goto out_unlock;
1908
1909         /*
1910          * The page must not be redirtied until the writeout is completed
1911          * (i.e. userspace has sent a reply to the write request).  Otherwise
1912          * there could be more than one temporary page instance for each real
1913          * page.
1914          *
1915          * This is ensured by holding the page lock in page_mkwrite() while
1916          * checking fuse_page_is_writeback().  We already hold the page lock
1917          * since clear_page_dirty_for_io() and keep it held until we add the
1918          * request to the fi->writepages list and increment req->num_pages.
1919          * After this fuse_page_is_writeback() will indicate that the page is
1920          * under writeback, so we can release the page lock.
1921          */
1922         if (data->req == NULL) {
1923                 struct fuse_inode *fi = get_fuse_inode(inode);
1924
1925                 err = -ENOMEM;
1926                 req = fuse_request_alloc_nofs(FUSE_REQ_INLINE_PAGES);
1927                 if (!req) {
1928                         __free_page(tmp_page);
1929                         goto out_unlock;
1930                 }
1931
1932                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1933                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1934                 req->misc.write.next = NULL;
1935                 req->in.argpages = 1;
1936                 __set_bit(FR_BACKGROUND, &req->flags);
1937                 req->num_pages = 0;
1938                 req->end = fuse_writepage_end;
1939                 req->inode = inode;
1940
1941                 spin_lock(&fi->lock);
1942                 list_add(&req->writepages_entry, &fi->writepages);
1943                 spin_unlock(&fi->lock);
1944
1945                 data->req = req;
1946         }
1947         set_page_writeback(page);
1948
1949         copy_highpage(tmp_page, page);
1950         req->pages[req->num_pages] = tmp_page;
1951         req->page_descs[req->num_pages].offset = 0;
1952         req->page_descs[req->num_pages].length = PAGE_SIZE;
1953
1954         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1955         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1956
1957         err = 0;
1958         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1959                 end_page_writeback(page);
1960                 data->req = NULL;
1961                 goto out_unlock;
1962         }
1963         data->orig_pages[req->num_pages] = page;
1964
1965         /*
1966          * Protected by fi->lock against concurrent access by
1967          * fuse_page_is_writeback().
1968          */
1969         spin_lock(&fi->lock);
1970         req->num_pages++;
1971         spin_unlock(&fi->lock);
1972
1973 out_unlock:
1974         unlock_page(page);
1975
1976         return err;
1977 }
1978
1979 static int fuse_writepages(struct address_space *mapping,
1980                            struct writeback_control *wbc)
1981 {
1982         struct inode *inode = mapping->host;
1983         struct fuse_conn *fc = get_fuse_conn(inode);
1984         struct fuse_fill_wb_data data;
1985         int err;
1986
1987         err = -EIO;
1988         if (is_bad_inode(inode))
1989                 goto out;
1990
1991         data.inode = inode;
1992         data.req = NULL;
1993         data.ff = NULL;
1994
1995         err = -ENOMEM;
1996         data.orig_pages = kcalloc(fc->max_pages,
1997                                   sizeof(struct page *),
1998                                   GFP_NOFS);
1999         if (!data.orig_pages)
2000                 goto out;
2001
2002         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2003         if (data.req) {
2004                 /* Ignore errors if we can write at least one page */
2005                 BUG_ON(!data.req->num_pages);
2006                 fuse_writepages_send(&data);
2007                 err = 0;
2008         }
2009         if (data.ff)
2010                 fuse_file_put(data.ff, false, false);
2011
2012         kfree(data.orig_pages);
2013 out:
2014         return err;
2015 }
2016
2017 /*
2018  * It's worthy to make sure that space is reserved on disk for the write,
2019  * but how to implement it without killing performance need more thinking.
2020  */
2021 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2022                 loff_t pos, unsigned len, unsigned flags,
2023                 struct page **pagep, void **fsdata)
2024 {
2025         pgoff_t index = pos >> PAGE_SHIFT;
2026         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2027         struct page *page;
2028         loff_t fsize;
2029         int err = -ENOMEM;
2030
2031         WARN_ON(!fc->writeback_cache);
2032
2033         page = grab_cache_page_write_begin(mapping, index, flags);
2034         if (!page)
2035                 goto error;
2036
2037         fuse_wait_on_page_writeback(mapping->host, page->index);
2038
2039         if (PageUptodate(page) || len == PAGE_SIZE)
2040                 goto success;
2041         /*
2042          * Check if the start this page comes after the end of file, in which
2043          * case the readpage can be optimized away.
2044          */
2045         fsize = i_size_read(mapping->host);
2046         if (fsize <= (pos & PAGE_MASK)) {
2047                 size_t off = pos & ~PAGE_MASK;
2048                 if (off)
2049                         zero_user_segment(page, 0, off);
2050                 goto success;
2051         }
2052         err = fuse_do_readpage(file, page);
2053         if (err)
2054                 goto cleanup;
2055 success:
2056         *pagep = page;
2057         return 0;
2058
2059 cleanup:
2060         unlock_page(page);
2061         put_page(page);
2062 error:
2063         return err;
2064 }
2065
2066 static int fuse_write_end(struct file *file, struct address_space *mapping,
2067                 loff_t pos, unsigned len, unsigned copied,
2068                 struct page *page, void *fsdata)
2069 {
2070         struct inode *inode = page->mapping->host;
2071
2072         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2073         if (!copied)
2074                 goto unlock;
2075
2076         if (!PageUptodate(page)) {
2077                 /* Zero any unwritten bytes at the end of the page */
2078                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2079                 if (endoff)
2080                         zero_user_segment(page, endoff, PAGE_SIZE);
2081                 SetPageUptodate(page);
2082         }
2083
2084         fuse_write_update_size(inode, pos + copied);
2085         set_page_dirty(page);
2086
2087 unlock:
2088         unlock_page(page);
2089         put_page(page);
2090
2091         return copied;
2092 }
2093
2094 static int fuse_launder_page(struct page *page)
2095 {
2096         int err = 0;
2097         if (clear_page_dirty_for_io(page)) {
2098                 struct inode *inode = page->mapping->host;
2099                 err = fuse_writepage_locked(page);
2100                 if (!err)
2101                         fuse_wait_on_page_writeback(inode, page->index);
2102         }
2103         return err;
2104 }
2105
2106 /*
2107  * Write back dirty pages now, because there may not be any suitable
2108  * open files later
2109  */
2110 static void fuse_vma_close(struct vm_area_struct *vma)
2111 {
2112         filemap_write_and_wait(vma->vm_file->f_mapping);
2113 }
2114
2115 /*
2116  * Wait for writeback against this page to complete before allowing it
2117  * to be marked dirty again, and hence written back again, possibly
2118  * before the previous writepage completed.
2119  *
2120  * Block here, instead of in ->writepage(), so that the userspace fs
2121  * can only block processes actually operating on the filesystem.
2122  *
2123  * Otherwise unprivileged userspace fs would be able to block
2124  * unrelated:
2125  *
2126  * - page migration
2127  * - sync(2)
2128  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2129  */
2130 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2131 {
2132         struct page *page = vmf->page;
2133         struct inode *inode = file_inode(vmf->vma->vm_file);
2134
2135         file_update_time(vmf->vma->vm_file);
2136         lock_page(page);
2137         if (page->mapping != inode->i_mapping) {
2138                 unlock_page(page);
2139                 return VM_FAULT_NOPAGE;
2140         }
2141
2142         fuse_wait_on_page_writeback(inode, page->index);
2143         return VM_FAULT_LOCKED;
2144 }
2145
2146 static const struct vm_operations_struct fuse_file_vm_ops = {
2147         .close          = fuse_vma_close,
2148         .fault          = filemap_fault,
2149         .map_pages      = filemap_map_pages,
2150         .page_mkwrite   = fuse_page_mkwrite,
2151 };
2152
2153 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2154 {
2155         struct fuse_file *ff = file->private_data;
2156
2157         if (ff->open_flags & FOPEN_DIRECT_IO) {
2158                 /* Can't provide the coherency needed for MAP_SHARED */
2159                 if (vma->vm_flags & VM_MAYSHARE)
2160                         return -ENODEV;
2161
2162                 invalidate_inode_pages2(file->f_mapping);
2163
2164                 return generic_file_mmap(file, vma);
2165         }
2166
2167         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2168                 fuse_link_write_file(file);
2169
2170         file_accessed(file);
2171         vma->vm_ops = &fuse_file_vm_ops;
2172         return 0;
2173 }
2174
2175 static int convert_fuse_file_lock(struct fuse_conn *fc,
2176                                   const struct fuse_file_lock *ffl,
2177                                   struct file_lock *fl)
2178 {
2179         switch (ffl->type) {
2180         case F_UNLCK:
2181                 break;
2182
2183         case F_RDLCK:
2184         case F_WRLCK:
2185                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2186                     ffl->end < ffl->start)
2187                         return -EIO;
2188
2189                 fl->fl_start = ffl->start;
2190                 fl->fl_end = ffl->end;
2191
2192                 /*
2193                  * Convert pid into init's pid namespace.  The locks API will
2194                  * translate it into the caller's pid namespace.
2195                  */
2196                 rcu_read_lock();
2197                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2198                 rcu_read_unlock();
2199                 break;
2200
2201         default:
2202                 return -EIO;
2203         }
2204         fl->fl_type = ffl->type;
2205         return 0;
2206 }
2207
2208 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2209                          const struct file_lock *fl, int opcode, pid_t pid,
2210                          int flock, struct fuse_lk_in *inarg)
2211 {
2212         struct inode *inode = file_inode(file);
2213         struct fuse_conn *fc = get_fuse_conn(inode);
2214         struct fuse_file *ff = file->private_data;
2215
2216         memset(inarg, 0, sizeof(*inarg));
2217         inarg->fh = ff->fh;
2218         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2219         inarg->lk.start = fl->fl_start;
2220         inarg->lk.end = fl->fl_end;
2221         inarg->lk.type = fl->fl_type;
2222         inarg->lk.pid = pid;
2223         if (flock)
2224                 inarg->lk_flags |= FUSE_LK_FLOCK;
2225         args->in.h.opcode = opcode;
2226         args->in.h.nodeid = get_node_id(inode);
2227         args->in.numargs = 1;
2228         args->in.args[0].size = sizeof(*inarg);
2229         args->in.args[0].value = inarg;
2230 }
2231
2232 static int fuse_getlk(struct file *file, struct file_lock *fl)
2233 {
2234         struct inode *inode = file_inode(file);
2235         struct fuse_conn *fc = get_fuse_conn(inode);
2236         FUSE_ARGS(args);
2237         struct fuse_lk_in inarg;
2238         struct fuse_lk_out outarg;
2239         int err;
2240
2241         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2242         args.out.numargs = 1;
2243         args.out.args[0].size = sizeof(outarg);
2244         args.out.args[0].value = &outarg;
2245         err = fuse_simple_request(fc, &args);
2246         if (!err)
2247                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2248
2249         return err;
2250 }
2251
2252 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2253 {
2254         struct inode *inode = file_inode(file);
2255         struct fuse_conn *fc = get_fuse_conn(inode);
2256         FUSE_ARGS(args);
2257         struct fuse_lk_in inarg;
2258         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2259         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2260         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2261         int err;
2262
2263         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2264                 /* NLM needs asynchronous locks, which we don't support yet */
2265                 return -ENOLCK;
2266         }
2267
2268         /* Unlock on close is handled by the flush method */
2269         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2270                 return 0;
2271
2272         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2273         err = fuse_simple_request(fc, &args);
2274
2275         /* locking is restartable */
2276         if (err == -EINTR)
2277                 err = -ERESTARTSYS;
2278
2279         return err;
2280 }
2281
2282 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2283 {
2284         struct inode *inode = file_inode(file);
2285         struct fuse_conn *fc = get_fuse_conn(inode);
2286         int err;
2287
2288         if (cmd == F_CANCELLK) {
2289                 err = 0;
2290         } else if (cmd == F_GETLK) {
2291                 if (fc->no_lock) {
2292                         posix_test_lock(file, fl);
2293                         err = 0;
2294                 } else
2295                         err = fuse_getlk(file, fl);
2296         } else {
2297                 if (fc->no_lock)
2298                         err = posix_lock_file(file, fl, NULL);
2299                 else
2300                         err = fuse_setlk(file, fl, 0);
2301         }
2302         return err;
2303 }
2304
2305 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2306 {
2307         struct inode *inode = file_inode(file);
2308         struct fuse_conn *fc = get_fuse_conn(inode);
2309         int err;
2310
2311         if (fc->no_flock) {
2312                 err = locks_lock_file_wait(file, fl);
2313         } else {
2314                 struct fuse_file *ff = file->private_data;
2315
2316                 /* emulate flock with POSIX locks */
2317                 ff->flock = true;
2318                 err = fuse_setlk(file, fl, 1);
2319         }
2320
2321         return err;
2322 }
2323
2324 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2325 {
2326         struct inode *inode = mapping->host;
2327         struct fuse_conn *fc = get_fuse_conn(inode);
2328         FUSE_ARGS(args);
2329         struct fuse_bmap_in inarg;
2330         struct fuse_bmap_out outarg;
2331         int err;
2332
2333         if (!inode->i_sb->s_bdev || fc->no_bmap)
2334                 return 0;
2335
2336         memset(&inarg, 0, sizeof(inarg));
2337         inarg.block = block;
2338         inarg.blocksize = inode->i_sb->s_blocksize;
2339         args.in.h.opcode = FUSE_BMAP;
2340         args.in.h.nodeid = get_node_id(inode);
2341         args.in.numargs = 1;
2342         args.in.args[0].size = sizeof(inarg);
2343         args.in.args[0].value = &inarg;
2344         args.out.numargs = 1;
2345         args.out.args[0].size = sizeof(outarg);
2346         args.out.args[0].value = &outarg;
2347         err = fuse_simple_request(fc, &args);
2348         if (err == -ENOSYS)
2349                 fc->no_bmap = 1;
2350
2351         return err ? 0 : outarg.block;
2352 }
2353
2354 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2355 {
2356         struct inode *inode = file->f_mapping->host;
2357         struct fuse_conn *fc = get_fuse_conn(inode);
2358         struct fuse_file *ff = file->private_data;
2359         FUSE_ARGS(args);
2360         struct fuse_lseek_in inarg = {
2361                 .fh = ff->fh,
2362                 .offset = offset,
2363                 .whence = whence
2364         };
2365         struct fuse_lseek_out outarg;
2366         int err;
2367
2368         if (fc->no_lseek)
2369                 goto fallback;
2370
2371         args.in.h.opcode = FUSE_LSEEK;
2372         args.in.h.nodeid = ff->nodeid;
2373         args.in.numargs = 1;
2374         args.in.args[0].size = sizeof(inarg);
2375         args.in.args[0].value = &inarg;
2376         args.out.numargs = 1;
2377         args.out.args[0].size = sizeof(outarg);
2378         args.out.args[0].value = &outarg;
2379         err = fuse_simple_request(fc, &args);
2380         if (err) {
2381                 if (err == -ENOSYS) {
2382                         fc->no_lseek = 1;
2383                         goto fallback;
2384                 }
2385                 return err;
2386         }
2387
2388         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2389
2390 fallback:
2391         err = fuse_update_attributes(inode, file);
2392         if (!err)
2393                 return generic_file_llseek(file, offset, whence);
2394         else
2395                 return err;
2396 }
2397
2398 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2399 {
2400         loff_t retval;
2401         struct inode *inode = file_inode(file);
2402
2403         switch (whence) {
2404         case SEEK_SET:
2405         case SEEK_CUR:
2406                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2407                 retval = generic_file_llseek(file, offset, whence);
2408                 break;
2409         case SEEK_END:
2410                 inode_lock(inode);
2411                 retval = fuse_update_attributes(inode, file);
2412                 if (!retval)
2413                         retval = generic_file_llseek(file, offset, whence);
2414                 inode_unlock(inode);
2415                 break;
2416         case SEEK_HOLE:
2417         case SEEK_DATA:
2418                 inode_lock(inode);
2419                 retval = fuse_lseek(file, offset, whence);
2420                 inode_unlock(inode);
2421                 break;
2422         default:
2423                 retval = -EINVAL;
2424         }
2425
2426         return retval;
2427 }
2428
2429 /*
2430  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2431  * ABI was defined to be 'struct iovec' which is different on 32bit
2432  * and 64bit.  Fortunately we can determine which structure the server
2433  * used from the size of the reply.
2434  */
2435 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2436                                      size_t transferred, unsigned count,
2437                                      bool is_compat)
2438 {
2439 #ifdef CONFIG_COMPAT
2440         if (count * sizeof(struct compat_iovec) == transferred) {
2441                 struct compat_iovec *ciov = src;
2442                 unsigned i;
2443
2444                 /*
2445                  * With this interface a 32bit server cannot support
2446                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2447                  * requests
2448                  */
2449                 if (!is_compat)
2450                         return -EINVAL;
2451
2452                 for (i = 0; i < count; i++) {
2453                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2454                         dst[i].iov_len = ciov[i].iov_len;
2455                 }
2456                 return 0;
2457         }
2458 #endif
2459
2460         if (count * sizeof(struct iovec) != transferred)
2461                 return -EIO;
2462
2463         memcpy(dst, src, transferred);
2464         return 0;
2465 }
2466
2467 /* Make sure iov_length() won't overflow */
2468 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2469                                  size_t count)
2470 {
2471         size_t n;
2472         u32 max = fc->max_pages << PAGE_SHIFT;
2473
2474         for (n = 0; n < count; n++, iov++) {
2475                 if (iov->iov_len > (size_t) max)
2476                         return -ENOMEM;
2477                 max -= iov->iov_len;
2478         }
2479         return 0;
2480 }
2481
2482 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2483                                  void *src, size_t transferred, unsigned count,
2484                                  bool is_compat)
2485 {
2486         unsigned i;
2487         struct fuse_ioctl_iovec *fiov = src;
2488
2489         if (fc->minor < 16) {
2490                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2491                                                  count, is_compat);
2492         }
2493
2494         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2495                 return -EIO;
2496
2497         for (i = 0; i < count; i++) {
2498                 /* Did the server supply an inappropriate value? */
2499                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2500                     fiov[i].len != (unsigned long) fiov[i].len)
2501                         return -EIO;
2502
2503                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2504                 dst[i].iov_len = (size_t) fiov[i].len;
2505
2506 #ifdef CONFIG_COMPAT
2507                 if (is_compat &&
2508                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2509                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2510                         return -EIO;
2511 #endif
2512         }
2513
2514         return 0;
2515 }
2516
2517
2518 /*
2519  * For ioctls, there is no generic way to determine how much memory
2520  * needs to be read and/or written.  Furthermore, ioctls are allowed
2521  * to dereference the passed pointer, so the parameter requires deep
2522  * copying but FUSE has no idea whatsoever about what to copy in or
2523  * out.
2524  *
2525  * This is solved by allowing FUSE server to retry ioctl with
2526  * necessary in/out iovecs.  Let's assume the ioctl implementation
2527  * needs to read in the following structure.
2528  *
2529  * struct a {
2530  *      char    *buf;
2531  *      size_t  buflen;
2532  * }
2533  *
2534  * On the first callout to FUSE server, inarg->in_size and
2535  * inarg->out_size will be NULL; then, the server completes the ioctl
2536  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2537  * the actual iov array to
2538  *
2539  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2540  *
2541  * which tells FUSE to copy in the requested area and retry the ioctl.
2542  * On the second round, the server has access to the structure and
2543  * from that it can tell what to look for next, so on the invocation,
2544  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2545  *
2546  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2547  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2548  *
2549  * FUSE will copy both struct a and the pointed buffer from the
2550  * process doing the ioctl and retry ioctl with both struct a and the
2551  * buffer.
2552  *
2553  * This time, FUSE server has everything it needs and completes ioctl
2554  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2555  *
2556  * Copying data out works the same way.
2557  *
2558  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2559  * automatically initializes in and out iovs by decoding @cmd with
2560  * _IOC_* macros and the server is not allowed to request RETRY.  This
2561  * limits ioctl data transfers to well-formed ioctls and is the forced
2562  * behavior for all FUSE servers.
2563  */
2564 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2565                    unsigned int flags)
2566 {
2567         struct fuse_file *ff = file->private_data;
2568         struct fuse_conn *fc = ff->fc;
2569         struct fuse_ioctl_in inarg = {
2570                 .fh = ff->fh,
2571                 .cmd = cmd,
2572                 .arg = arg,
2573                 .flags = flags
2574         };
2575         struct fuse_ioctl_out outarg;
2576         struct fuse_req *req = NULL;
2577         struct page **pages = NULL;
2578         struct iovec *iov_page = NULL;
2579         struct iovec *in_iov = NULL, *out_iov = NULL;
2580         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2581         size_t in_size, out_size, transferred, c;
2582         int err, i;
2583         struct iov_iter ii;
2584
2585 #if BITS_PER_LONG == 32
2586         inarg.flags |= FUSE_IOCTL_32BIT;
2587 #else
2588         if (flags & FUSE_IOCTL_COMPAT) {
2589                 inarg.flags |= FUSE_IOCTL_32BIT;
2590 #ifdef CONFIG_X86_X32
2591                 if (in_x32_syscall())
2592                         inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2593 #endif
2594         }
2595 #endif
2596
2597         /* assume all the iovs returned by client always fits in a page */
2598         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2599
2600         err = -ENOMEM;
2601         pages = kcalloc(fc->max_pages, sizeof(pages[0]), GFP_KERNEL);
2602         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2603         if (!pages || !iov_page)
2604                 goto out;
2605
2606         /*
2607          * If restricted, initialize IO parameters as encoded in @cmd.
2608          * RETRY from server is not allowed.
2609          */
2610         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2611                 struct iovec *iov = iov_page;
2612
2613                 iov->iov_base = (void __user *)arg;
2614                 iov->iov_len = _IOC_SIZE(cmd);
2615
2616                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2617                         in_iov = iov;
2618                         in_iovs = 1;
2619                 }
2620
2621                 if (_IOC_DIR(cmd) & _IOC_READ) {
2622                         out_iov = iov;
2623                         out_iovs = 1;
2624                 }
2625         }
2626
2627  retry:
2628         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2629         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2630
2631         /*
2632          * Out data can be used either for actual out data or iovs,
2633          * make sure there always is at least one page.
2634          */
2635         out_size = max_t(size_t, out_size, PAGE_SIZE);
2636         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2637
2638         /* make sure there are enough buffer pages and init request with them */
2639         err = -ENOMEM;
2640         if (max_pages > fc->max_pages)
2641                 goto out;
2642         while (num_pages < max_pages) {
2643                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2644                 if (!pages[num_pages])
2645                         goto out;
2646                 num_pages++;
2647         }
2648
2649         req = fuse_get_req(fc, num_pages);
2650         if (IS_ERR(req)) {
2651                 err = PTR_ERR(req);
2652                 req = NULL;
2653                 goto out;
2654         }
2655         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2656         req->num_pages = num_pages;
2657         fuse_page_descs_length_init(req, 0, req->num_pages);
2658
2659         /* okay, let's send it to the client */
2660         req->in.h.opcode = FUSE_IOCTL;
2661         req->in.h.nodeid = ff->nodeid;
2662         req->in.numargs = 1;
2663         req->in.args[0].size = sizeof(inarg);
2664         req->in.args[0].value = &inarg;
2665         if (in_size) {
2666                 req->in.numargs++;
2667                 req->in.args[1].size = in_size;
2668                 req->in.argpages = 1;
2669
2670                 err = -EFAULT;
2671                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2672                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2673                         c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2674                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2675                                 goto out;
2676                 }
2677         }
2678
2679         req->out.numargs = 2;
2680         req->out.args[0].size = sizeof(outarg);
2681         req->out.args[0].value = &outarg;
2682         req->out.args[1].size = out_size;
2683         req->out.argpages = 1;
2684         req->out.argvar = 1;
2685
2686         fuse_request_send(fc, req);
2687         err = req->out.h.error;
2688         transferred = req->out.args[1].size;
2689         fuse_put_request(fc, req);
2690         req = NULL;
2691         if (err)
2692                 goto out;
2693
2694         /* did it ask for retry? */
2695         if (outarg.flags & FUSE_IOCTL_RETRY) {
2696                 void *vaddr;
2697
2698                 /* no retry if in restricted mode */
2699                 err = -EIO;
2700                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2701                         goto out;
2702
2703                 in_iovs = outarg.in_iovs;
2704                 out_iovs = outarg.out_iovs;
2705
2706                 /*
2707                  * Make sure things are in boundary, separate checks
2708                  * are to protect against overflow.
2709                  */
2710                 err = -ENOMEM;
2711                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2712                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2713                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2714                         goto out;
2715
2716                 vaddr = kmap_atomic(pages[0]);
2717                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2718                                             transferred, in_iovs + out_iovs,
2719                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2720                 kunmap_atomic(vaddr);
2721                 if (err)
2722                         goto out;
2723
2724                 in_iov = iov_page;
2725                 out_iov = in_iov + in_iovs;
2726
2727                 err = fuse_verify_ioctl_iov(fc, in_iov, in_iovs);
2728                 if (err)
2729                         goto out;
2730
2731                 err = fuse_verify_ioctl_iov(fc, out_iov, out_iovs);
2732                 if (err)
2733                         goto out;
2734
2735                 goto retry;
2736         }
2737
2738         err = -EIO;
2739         if (transferred > inarg.out_size)
2740                 goto out;
2741
2742         err = -EFAULT;
2743         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2744         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2745                 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2746                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2747                         goto out;
2748         }
2749         err = 0;
2750  out:
2751         if (req)
2752                 fuse_put_request(fc, req);
2753         free_page((unsigned long) iov_page);
2754         while (num_pages)
2755                 __free_page(pages[--num_pages]);
2756         kfree(pages);
2757
2758         return err ? err : outarg.result;
2759 }
2760 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2761
2762 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2763                        unsigned long arg, unsigned int flags)
2764 {
2765         struct inode *inode = file_inode(file);
2766         struct fuse_conn *fc = get_fuse_conn(inode);
2767
2768         if (!fuse_allow_current_process(fc))
2769                 return -EACCES;
2770
2771         if (is_bad_inode(inode))
2772                 return -EIO;
2773
2774         return fuse_do_ioctl(file, cmd, arg, flags);
2775 }
2776
2777 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2778                             unsigned long arg)
2779 {
2780         return fuse_ioctl_common(file, cmd, arg, 0);
2781 }
2782
2783 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2784                                    unsigned long arg)
2785 {
2786         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2787 }
2788
2789 /*
2790  * All files which have been polled are linked to RB tree
2791  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2792  * find the matching one.
2793  */
2794 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2795                                               struct rb_node **parent_out)
2796 {
2797         struct rb_node **link = &fc->polled_files.rb_node;
2798         struct rb_node *last = NULL;
2799
2800         while (*link) {
2801                 struct fuse_file *ff;
2802
2803                 last = *link;
2804                 ff = rb_entry(last, struct fuse_file, polled_node);
2805
2806                 if (kh < ff->kh)
2807                         link = &last->rb_left;
2808                 else if (kh > ff->kh)
2809                         link = &last->rb_right;
2810                 else
2811                         return link;
2812         }
2813
2814         if (parent_out)
2815                 *parent_out = last;
2816         return link;
2817 }
2818
2819 /*
2820  * The file is about to be polled.  Make sure it's on the polled_files
2821  * RB tree.  Note that files once added to the polled_files tree are
2822  * not removed before the file is released.  This is because a file
2823  * polled once is likely to be polled again.
2824  */
2825 static void fuse_register_polled_file(struct fuse_conn *fc,
2826                                       struct fuse_file *ff)
2827 {
2828         spin_lock(&fc->lock);
2829         if (RB_EMPTY_NODE(&ff->polled_node)) {
2830                 struct rb_node **link, *uninitialized_var(parent);
2831
2832                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2833                 BUG_ON(*link);
2834                 rb_link_node(&ff->polled_node, parent, link);
2835                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2836         }
2837         spin_unlock(&fc->lock);
2838 }
2839
2840 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2841 {
2842         struct fuse_file *ff = file->private_data;
2843         struct fuse_conn *fc = ff->fc;
2844         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2845         struct fuse_poll_out outarg;
2846         FUSE_ARGS(args);
2847         int err;
2848
2849         if (fc->no_poll)
2850                 return DEFAULT_POLLMASK;
2851
2852         poll_wait(file, &ff->poll_wait, wait);
2853         inarg.events = mangle_poll(poll_requested_events(wait));
2854
2855         /*
2856          * Ask for notification iff there's someone waiting for it.
2857          * The client may ignore the flag and always notify.
2858          */
2859         if (waitqueue_active(&ff->poll_wait)) {
2860                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2861                 fuse_register_polled_file(fc, ff);
2862         }
2863
2864         args.in.h.opcode = FUSE_POLL;
2865         args.in.h.nodeid = ff->nodeid;
2866         args.in.numargs = 1;
2867         args.in.args[0].size = sizeof(inarg);
2868         args.in.args[0].value = &inarg;
2869         args.out.numargs = 1;
2870         args.out.args[0].size = sizeof(outarg);
2871         args.out.args[0].value = &outarg;
2872         err = fuse_simple_request(fc, &args);
2873
2874         if (!err)
2875                 return demangle_poll(outarg.revents);
2876         if (err == -ENOSYS) {
2877                 fc->no_poll = 1;
2878                 return DEFAULT_POLLMASK;
2879         }
2880         return EPOLLERR;
2881 }
2882 EXPORT_SYMBOL_GPL(fuse_file_poll);
2883
2884 /*
2885  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2886  * wakes up the poll waiters.
2887  */
2888 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2889                             struct fuse_notify_poll_wakeup_out *outarg)
2890 {
2891         u64 kh = outarg->kh;
2892         struct rb_node **link;
2893
2894         spin_lock(&fc->lock);
2895
2896         link = fuse_find_polled_node(fc, kh, NULL);
2897         if (*link) {
2898                 struct fuse_file *ff;
2899
2900                 ff = rb_entry(*link, struct fuse_file, polled_node);
2901                 wake_up_interruptible_sync(&ff->poll_wait);
2902         }
2903
2904         spin_unlock(&fc->lock);
2905         return 0;
2906 }
2907
2908 static void fuse_do_truncate(struct file *file)
2909 {
2910         struct inode *inode = file->f_mapping->host;
2911         struct iattr attr;
2912
2913         attr.ia_valid = ATTR_SIZE;
2914         attr.ia_size = i_size_read(inode);
2915
2916         attr.ia_file = file;
2917         attr.ia_valid |= ATTR_FILE;
2918
2919         fuse_do_setattr(file_dentry(file), &attr, file);
2920 }
2921
2922 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2923 {
2924         return round_up(off, fc->max_pages << PAGE_SHIFT);
2925 }
2926
2927 static ssize_t
2928 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2929 {
2930         DECLARE_COMPLETION_ONSTACK(wait);
2931         ssize_t ret = 0;
2932         struct file *file = iocb->ki_filp;
2933         struct fuse_file *ff = file->private_data;
2934         bool async_dio = ff->fc->async_dio;
2935         loff_t pos = 0;
2936         struct inode *inode;
2937         loff_t i_size;
2938         size_t count = iov_iter_count(iter);
2939         loff_t offset = iocb->ki_pos;
2940         struct fuse_io_priv *io;
2941
2942         pos = offset;
2943         inode = file->f_mapping->host;
2944         i_size = i_size_read(inode);
2945
2946         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2947                 return 0;
2948
2949         /* optimization for short read */
2950         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2951                 if (offset >= i_size)
2952                         return 0;
2953                 iov_iter_truncate(iter, fuse_round_up(ff->fc, i_size - offset));
2954                 count = iov_iter_count(iter);
2955         }
2956
2957         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2958         if (!io)
2959                 return -ENOMEM;
2960         spin_lock_init(&io->lock);
2961         kref_init(&io->refcnt);
2962         io->reqs = 1;
2963         io->bytes = -1;
2964         io->size = 0;
2965         io->offset = offset;
2966         io->write = (iov_iter_rw(iter) == WRITE);
2967         io->err = 0;
2968         /*
2969          * By default, we want to optimize all I/Os with async request
2970          * submission to the client filesystem if supported.
2971          */
2972         io->async = async_dio;
2973         io->iocb = iocb;
2974         io->blocking = is_sync_kiocb(iocb);
2975
2976         /*
2977          * We cannot asynchronously extend the size of a file.
2978          * In such case the aio will behave exactly like sync io.
2979          */
2980         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2981                 io->blocking = true;
2982
2983         if (io->async && io->blocking) {
2984                 /*
2985                  * Additional reference to keep io around after
2986                  * calling fuse_aio_complete()
2987                  */
2988                 kref_get(&io->refcnt);
2989                 io->done = &wait;
2990         }
2991
2992         if (iov_iter_rw(iter) == WRITE) {
2993                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2994                 fuse_invalidate_attr(inode);
2995         } else {
2996                 ret = __fuse_direct_read(io, iter, &pos);
2997         }
2998
2999         if (io->async) {
3000                 bool blocking = io->blocking;
3001
3002                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3003
3004                 /* we have a non-extending, async request, so return */
3005                 if (!blocking)
3006                         return -EIOCBQUEUED;
3007
3008                 wait_for_completion(&wait);
3009                 ret = fuse_get_res_by_io(io);
3010         }
3011
3012         kref_put(&io->refcnt, fuse_io_release);
3013
3014         if (iov_iter_rw(iter) == WRITE) {
3015                 if (ret > 0)
3016                         fuse_write_update_size(inode, pos);
3017                 else if (ret < 0 && offset + count > i_size)
3018                         fuse_do_truncate(file);
3019         }
3020
3021         return ret;
3022 }
3023
3024 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3025 {
3026         int err = filemap_write_and_wait_range(inode->i_mapping, start, end);
3027
3028         if (!err)
3029                 fuse_sync_writes(inode);
3030
3031         return err;
3032 }
3033
3034 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3035                                 loff_t length)
3036 {
3037         struct fuse_file *ff = file->private_data;
3038         struct inode *inode = file_inode(file);
3039         struct fuse_inode *fi = get_fuse_inode(inode);
3040         struct fuse_conn *fc = ff->fc;
3041         FUSE_ARGS(args);
3042         struct fuse_fallocate_in inarg = {
3043                 .fh = ff->fh,
3044                 .offset = offset,
3045                 .length = length,
3046                 .mode = mode
3047         };
3048         int err;
3049         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
3050                            (mode & FALLOC_FL_PUNCH_HOLE);
3051
3052         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3053                 return -EOPNOTSUPP;
3054
3055         if (fc->no_fallocate)
3056                 return -EOPNOTSUPP;
3057
3058         if (lock_inode) {
3059                 inode_lock(inode);
3060                 if (mode & FALLOC_FL_PUNCH_HOLE) {
3061                         loff_t endbyte = offset + length - 1;
3062
3063                         err = fuse_writeback_range(inode, offset, endbyte);
3064                         if (err)
3065                                 goto out;
3066                 }
3067         }
3068
3069         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3070             offset + length > i_size_read(inode)) {
3071                 err = inode_newsize_ok(inode, offset + length);
3072                 if (err)
3073                         goto out;
3074         }
3075
3076         if (!(mode & FALLOC_FL_KEEP_SIZE))
3077                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3078
3079         args.in.h.opcode = FUSE_FALLOCATE;
3080         args.in.h.nodeid = ff->nodeid;
3081         args.in.numargs = 1;
3082         args.in.args[0].size = sizeof(inarg);
3083         args.in.args[0].value = &inarg;
3084         err = fuse_simple_request(fc, &args);
3085         if (err == -ENOSYS) {
3086                 fc->no_fallocate = 1;
3087                 err = -EOPNOTSUPP;
3088         }
3089         if (err)
3090                 goto out;
3091
3092         /* we could have extended the file */
3093         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3094                 bool changed = fuse_write_update_size(inode, offset + length);
3095
3096                 if (changed && fc->writeback_cache)
3097                         file_update_time(file);
3098         }
3099
3100         if (mode & FALLOC_FL_PUNCH_HOLE)
3101                 truncate_pagecache_range(inode, offset, offset + length - 1);
3102
3103         fuse_invalidate_attr(inode);
3104
3105 out:
3106         if (!(mode & FALLOC_FL_KEEP_SIZE))
3107                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3108
3109         if (lock_inode)
3110                 inode_unlock(inode);
3111
3112         return err;
3113 }
3114
3115 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3116                                       struct file *file_out, loff_t pos_out,
3117                                       size_t len, unsigned int flags)
3118 {
3119         struct fuse_file *ff_in = file_in->private_data;
3120         struct fuse_file *ff_out = file_out->private_data;
3121         struct inode *inode_in = file_inode(file_in);
3122         struct inode *inode_out = file_inode(file_out);
3123         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3124         struct fuse_conn *fc = ff_in->fc;
3125         FUSE_ARGS(args);
3126         struct fuse_copy_file_range_in inarg = {
3127                 .fh_in = ff_in->fh,
3128                 .off_in = pos_in,
3129                 .nodeid_out = ff_out->nodeid,
3130                 .fh_out = ff_out->fh,
3131                 .off_out = pos_out,
3132                 .len = len,
3133                 .flags = flags
3134         };
3135         struct fuse_write_out outarg;
3136         ssize_t err;
3137         /* mark unstable when write-back is not used, and file_out gets
3138          * extended */
3139         bool is_unstable = (!fc->writeback_cache) &&
3140                            ((pos_out + len) > inode_out->i_size);
3141
3142         if (fc->no_copy_file_range)
3143                 return -EOPNOTSUPP;
3144
3145         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3146                 return -EXDEV;
3147
3148         if (fc->writeback_cache) {
3149                 inode_lock(inode_in);
3150                 err = fuse_writeback_range(inode_in, pos_in, pos_in + len);
3151                 inode_unlock(inode_in);
3152                 if (err)
3153                         return err;
3154         }
3155
3156         inode_lock(inode_out);
3157
3158         err = file_modified(file_out);
3159         if (err)
3160                 goto out;
3161
3162         if (fc->writeback_cache) {
3163                 err = fuse_writeback_range(inode_out, pos_out, pos_out + len);
3164                 if (err)
3165                         goto out;
3166         }
3167
3168         if (is_unstable)
3169                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3170
3171         args.in.h.opcode = FUSE_COPY_FILE_RANGE;
3172         args.in.h.nodeid = ff_in->nodeid;
3173         args.in.numargs = 1;
3174         args.in.args[0].size = sizeof(inarg);
3175         args.in.args[0].value = &inarg;
3176         args.out.numargs = 1;
3177         args.out.args[0].size = sizeof(outarg);
3178         args.out.args[0].value = &outarg;
3179         err = fuse_simple_request(fc, &args);
3180         if (err == -ENOSYS) {
3181                 fc->no_copy_file_range = 1;
3182                 err = -EOPNOTSUPP;
3183         }
3184         if (err)
3185                 goto out;
3186
3187         if (fc->writeback_cache) {
3188                 fuse_write_update_size(inode_out, pos_out + outarg.size);
3189                 file_update_time(file_out);
3190         }
3191
3192         fuse_invalidate_attr(inode_out);
3193
3194         err = outarg.size;
3195 out:
3196         if (is_unstable)
3197                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3198
3199         inode_unlock(inode_out);
3200         file_accessed(file_in);
3201
3202         return err;
3203 }
3204
3205 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3206                                     struct file *dst_file, loff_t dst_off,
3207                                     size_t len, unsigned int flags)
3208 {
3209         ssize_t ret;
3210
3211         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3212                                      len, flags);
3213
3214         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3215                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3216                                               dst_off, len, flags);
3217         return ret;
3218 }
3219
3220 static const struct file_operations fuse_file_operations = {
3221         .llseek         = fuse_file_llseek,
3222         .read_iter      = fuse_file_read_iter,
3223         .write_iter     = fuse_file_write_iter,
3224         .mmap           = fuse_file_mmap,
3225         .open           = fuse_open,
3226         .flush          = fuse_flush,
3227         .release        = fuse_release,
3228         .fsync          = fuse_fsync,
3229         .lock           = fuse_file_lock,
3230         .flock          = fuse_file_flock,
3231         .splice_read    = generic_file_splice_read,
3232         .splice_write   = iter_file_splice_write,
3233         .unlocked_ioctl = fuse_file_ioctl,
3234         .compat_ioctl   = fuse_file_compat_ioctl,
3235         .poll           = fuse_file_poll,
3236         .fallocate      = fuse_file_fallocate,
3237         .copy_file_range = fuse_copy_file_range,
3238 };
3239
3240 static const struct address_space_operations fuse_file_aops  = {
3241         .readpage       = fuse_readpage,
3242         .writepage      = fuse_writepage,
3243         .writepages     = fuse_writepages,
3244         .launder_page   = fuse_launder_page,
3245         .readpages      = fuse_readpages,
3246         .set_page_dirty = __set_page_dirty_nobuffers,
3247         .bmap           = fuse_bmap,
3248         .direct_IO      = fuse_direct_IO,
3249         .write_begin    = fuse_write_begin,
3250         .write_end      = fuse_write_end,
3251 };
3252
3253 void fuse_init_file_inode(struct inode *inode)
3254 {
3255         struct fuse_inode *fi = get_fuse_inode(inode);
3256
3257         inode->i_fop = &fuse_file_operations;
3258         inode->i_data.a_ops = &fuse_file_aops;
3259
3260         INIT_LIST_HEAD(&fi->write_files);
3261         INIT_LIST_HEAD(&fi->queued_writes);
3262         fi->writectr = 0;
3263         init_waitqueue_head(&fi->page_waitq);
3264         INIT_LIST_HEAD(&fi->writepages);
3265 }